Commit | Line | Data |
---|---|---|
73572984 JA |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | #include <linux/kernel.h> | |
3 | #include <linux/errno.h> | |
4 | #include <linux/fs.h> | |
5 | #include <linux/file.h> | |
6 | #include <linux/mm.h> | |
7 | #include <linux/slab.h> | |
8 | #include <linux/nospec.h> | |
9 | #include <linux/hugetlb.h> | |
10 | #include <linux/compat.h> | |
11 | #include <linux/io_uring.h> | |
12 | ||
13 | #include <uapi/linux/io_uring.h> | |
14 | ||
73572984 JA |
15 | #include "io_uring.h" |
16 | #include "openclose.h" | |
17 | #include "rsrc.h" | |
18 | ||
19 | struct io_rsrc_update { | |
20 | struct file *file; | |
21 | u64 arg; | |
22 | u32 nr_args; | |
23 | u32 offset; | |
24 | }; | |
25 | ||
26 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov, | |
27 | struct io_mapped_ubuf **pimu, | |
28 | struct page **last_hpage); | |
29 | ||
73572984 JA |
30 | /* only define max */ |
31 | #define IORING_MAX_FIXED_FILES (1U << 20) | |
32 | #define IORING_MAX_REG_BUFFERS (1U << 14) | |
33 | ||
6a9ce66f | 34 | int __io_account_mem(struct user_struct *user, unsigned long nr_pages) |
73572984 JA |
35 | { |
36 | unsigned long page_limit, cur_pages, new_pages; | |
37 | ||
6a9ce66f PB |
38 | if (!nr_pages) |
39 | return 0; | |
40 | ||
73572984 JA |
41 | /* Don't allow more pages than we can safely lock */ |
42 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
43 | ||
4ccc6db0 | 44 | cur_pages = atomic_long_read(&user->locked_vm); |
73572984 | 45 | do { |
73572984 JA |
46 | new_pages = cur_pages + nr_pages; |
47 | if (new_pages > page_limit) | |
48 | return -ENOMEM; | |
4ccc6db0 UB |
49 | } while (!atomic_long_try_cmpxchg(&user->locked_vm, |
50 | &cur_pages, new_pages)); | |
73572984 JA |
51 | return 0; |
52 | } | |
53 | ||
54 | static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages) | |
55 | { | |
56 | if (ctx->user) | |
57 | __io_unaccount_mem(ctx->user, nr_pages); | |
58 | ||
59 | if (ctx->mm_account) | |
60 | atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm); | |
61 | } | |
62 | ||
63 | static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages) | |
64 | { | |
65 | int ret; | |
66 | ||
67 | if (ctx->user) { | |
68 | ret = __io_account_mem(ctx->user, nr_pages); | |
69 | if (ret) | |
70 | return ret; | |
71 | } | |
72 | ||
73 | if (ctx->mm_account) | |
74 | atomic64_add(nr_pages, &ctx->mm_account->pinned_vm); | |
75 | ||
76 | return 0; | |
77 | } | |
78 | ||
79 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
80 | void __user *arg, unsigned index) | |
81 | { | |
82 | struct iovec __user *src; | |
83 | ||
84 | #ifdef CONFIG_COMPAT | |
85 | if (ctx->compat) { | |
86 | struct compat_iovec __user *ciovs; | |
87 | struct compat_iovec ciov; | |
88 | ||
89 | ciovs = (struct compat_iovec __user *) arg; | |
90 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
91 | return -EFAULT; | |
92 | ||
93 | dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); | |
94 | dst->iov_len = ciov.iov_len; | |
95 | return 0; | |
96 | } | |
97 | #endif | |
98 | src = (struct iovec __user *) arg; | |
99 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
100 | return -EFAULT; | |
101 | return 0; | |
102 | } | |
103 | ||
104 | static int io_buffer_validate(struct iovec *iov) | |
105 | { | |
106 | unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1); | |
107 | ||
108 | /* | |
109 | * Don't impose further limits on the size and buffer | |
110 | * constraints here, we'll -EINVAL later when IO is | |
111 | * submitted if they are wrong. | |
112 | */ | |
113 | if (!iov->iov_base) | |
114 | return iov->iov_len ? -EFAULT : 0; | |
115 | if (!iov->iov_len) | |
116 | return -EFAULT; | |
117 | ||
118 | /* arbitrary limit, but we need something */ | |
119 | if (iov->iov_len > SZ_1G) | |
120 | return -EFAULT; | |
121 | ||
122 | if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp)) | |
123 | return -EOVERFLOW; | |
124 | ||
125 | return 0; | |
126 | } | |
127 | ||
128 | static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot) | |
129 | { | |
130 | struct io_mapped_ubuf *imu = *slot; | |
131 | unsigned int i; | |
132 | ||
133 | if (imu != ctx->dummy_ubuf) { | |
134 | for (i = 0; i < imu->nr_bvecs; i++) | |
135 | unpin_user_page(imu->bvec[i].bv_page); | |
136 | if (imu->acct_pages) | |
137 | io_unaccount_mem(ctx, imu->acct_pages); | |
138 | kvfree(imu); | |
139 | } | |
140 | *slot = NULL; | |
141 | } | |
142 | ||
ff7c75ec PB |
143 | static void io_rsrc_put_work_one(struct io_rsrc_data *rsrc_data, |
144 | struct io_rsrc_put *prsrc) | |
145 | { | |
146 | struct io_ring_ctx *ctx = rsrc_data->ctx; | |
147 | ||
148 | if (prsrc->tag) { | |
149 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
150 | mutex_lock(&ctx->uring_lock); | |
151 | io_post_aux_cqe(ctx, prsrc->tag, 0, 0); | |
152 | mutex_unlock(&ctx->uring_lock); | |
153 | } else { | |
154 | io_post_aux_cqe(ctx, prsrc->tag, 0, 0); | |
155 | } | |
156 | } | |
157 | rsrc_data->do_put(ctx, prsrc); | |
158 | } | |
159 | ||
73572984 JA |
160 | static void __io_rsrc_put_work(struct io_rsrc_node *ref_node) |
161 | { | |
162 | struct io_rsrc_data *rsrc_data = ref_node->rsrc_data; | |
73572984 JA |
163 | struct io_rsrc_put *prsrc, *tmp; |
164 | ||
ff7c75ec PB |
165 | if (ref_node->inline_items) |
166 | io_rsrc_put_work_one(rsrc_data, &ref_node->item); | |
167 | ||
c824986c | 168 | list_for_each_entry_safe(prsrc, tmp, &ref_node->item_list, list) { |
73572984 | 169 | list_del(&prsrc->list); |
ff7c75ec | 170 | io_rsrc_put_work_one(rsrc_data, prsrc); |
73572984 JA |
171 | kfree(prsrc); |
172 | } | |
173 | ||
174 | io_rsrc_node_destroy(ref_node); | |
175 | if (atomic_dec_and_test(&rsrc_data->refs)) | |
176 | complete(&rsrc_data->done); | |
177 | } | |
178 | ||
179 | void io_rsrc_put_work(struct work_struct *work) | |
180 | { | |
181 | struct io_ring_ctx *ctx; | |
182 | struct llist_node *node; | |
183 | ||
184 | ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work); | |
185 | node = llist_del_all(&ctx->rsrc_put_llist); | |
186 | ||
187 | while (node) { | |
188 | struct io_rsrc_node *ref_node; | |
189 | struct llist_node *next = node->next; | |
190 | ||
191 | ref_node = llist_entry(node, struct io_rsrc_node, llist); | |
192 | __io_rsrc_put_work(ref_node); | |
193 | node = next; | |
194 | } | |
195 | } | |
196 | ||
d34b1b0b PB |
197 | void io_rsrc_put_tw(struct callback_head *cb) |
198 | { | |
199 | struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx, | |
200 | rsrc_put_tw); | |
201 | ||
202 | io_rsrc_put_work(&ctx->rsrc_put_work.work); | |
203 | } | |
204 | ||
73572984 JA |
205 | void io_wait_rsrc_data(struct io_rsrc_data *data) |
206 | { | |
207 | if (data && !atomic_dec_and_test(&data->refs)) | |
208 | wait_for_completion(&data->done); | |
209 | } | |
210 | ||
211 | void io_rsrc_node_destroy(struct io_rsrc_node *ref_node) | |
212 | { | |
73572984 JA |
213 | kfree(ref_node); |
214 | } | |
215 | ||
ef8ae64f PB |
216 | void io_rsrc_node_ref_zero(struct io_rsrc_node *node) |
217 | __must_hold(&node->rsrc_data->ctx->uring_lock) | |
73572984 | 218 | { |
73572984 | 219 | struct io_ring_ctx *ctx = node->rsrc_data->ctx; |
73572984 JA |
220 | bool first_add = false; |
221 | unsigned long delay = HZ; | |
222 | ||
73572984 JA |
223 | node->done = true; |
224 | ||
225 | /* if we are mid-quiesce then do not delay */ | |
226 | if (node->rsrc_data->quiesce) | |
227 | delay = 0; | |
228 | ||
229 | while (!list_empty(&ctx->rsrc_ref_list)) { | |
230 | node = list_first_entry(&ctx->rsrc_ref_list, | |
231 | struct io_rsrc_node, node); | |
232 | /* recycle ref nodes in order */ | |
233 | if (!node->done) | |
234 | break; | |
235 | list_del(&node->node); | |
236 | first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist); | |
237 | } | |
73572984 | 238 | |
d34b1b0b PB |
239 | if (!first_add) |
240 | return; | |
241 | ||
242 | if (ctx->submitter_task) { | |
243 | if (!task_work_add(ctx->submitter_task, &ctx->rsrc_put_tw, | |
244 | ctx->notify_method)) | |
245 | return; | |
246 | } | |
247 | mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay); | |
73572984 JA |
248 | } |
249 | ||
250 | static struct io_rsrc_node *io_rsrc_node_alloc(void) | |
251 | { | |
252 | struct io_rsrc_node *ref_node; | |
253 | ||
254 | ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL); | |
255 | if (!ref_node) | |
256 | return NULL; | |
257 | ||
ef8ae64f | 258 | ref_node->refs = 1; |
73572984 | 259 | INIT_LIST_HEAD(&ref_node->node); |
c824986c | 260 | INIT_LIST_HEAD(&ref_node->item_list); |
73572984 | 261 | ref_node->done = false; |
ff7c75ec | 262 | ref_node->inline_items = 0; |
73572984 JA |
263 | return ref_node; |
264 | } | |
265 | ||
266 | void io_rsrc_node_switch(struct io_ring_ctx *ctx, | |
267 | struct io_rsrc_data *data_to_kill) | |
268 | __must_hold(&ctx->uring_lock) | |
269 | { | |
270 | WARN_ON_ONCE(!ctx->rsrc_backup_node); | |
271 | WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node); | |
272 | ||
73572984 JA |
273 | if (data_to_kill) { |
274 | struct io_rsrc_node *rsrc_node = ctx->rsrc_node; | |
275 | ||
276 | rsrc_node->rsrc_data = data_to_kill; | |
73572984 | 277 | list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list); |
73572984 JA |
278 | |
279 | atomic_inc(&data_to_kill->refs); | |
b8fb5b4f | 280 | /* put master ref */ |
ef8ae64f | 281 | io_put_rsrc_node(rsrc_node); |
73572984 JA |
282 | ctx->rsrc_node = NULL; |
283 | } | |
284 | ||
285 | if (!ctx->rsrc_node) { | |
286 | ctx->rsrc_node = ctx->rsrc_backup_node; | |
287 | ctx->rsrc_backup_node = NULL; | |
288 | } | |
289 | } | |
290 | ||
291 | int io_rsrc_node_switch_start(struct io_ring_ctx *ctx) | |
292 | { | |
293 | if (ctx->rsrc_backup_node) | |
294 | return 0; | |
295 | ctx->rsrc_backup_node = io_rsrc_node_alloc(); | |
296 | return ctx->rsrc_backup_node ? 0 : -ENOMEM; | |
297 | } | |
298 | ||
299 | __cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, | |
300 | struct io_ring_ctx *ctx) | |
301 | { | |
302 | int ret; | |
303 | ||
304 | /* As we may drop ->uring_lock, other task may have started quiesce */ | |
305 | if (data->quiesce) | |
306 | return -ENXIO; | |
77e3202a PB |
307 | ret = io_rsrc_node_switch_start(ctx); |
308 | if (ret) | |
309 | return ret; | |
310 | io_rsrc_node_switch(ctx, data); | |
311 | ||
312 | /* kill initial ref, already quiesced if zero */ | |
313 | if (atomic_dec_and_test(&data->refs)) | |
314 | return 0; | |
73572984 JA |
315 | |
316 | data->quiesce = true; | |
77e3202a | 317 | mutex_unlock(&ctx->uring_lock); |
73572984 | 318 | do { |
ef67fcb4 | 319 | ret = io_run_task_work_sig(ctx); |
77e3202a PB |
320 | if (ret < 0) { |
321 | atomic_inc(&data->refs); | |
322 | /* wait for all works potentially completing data->done */ | |
323 | flush_delayed_work(&ctx->rsrc_put_work); | |
324 | reinit_completion(&data->done); | |
325 | mutex_lock(&ctx->uring_lock); | |
326 | break; | |
327 | } | |
ef67fcb4 | 328 | |
73572984 JA |
329 | flush_delayed_work(&ctx->rsrc_put_work); |
330 | ret = wait_for_completion_interruptible(&data->done); | |
331 | if (!ret) { | |
332 | mutex_lock(&ctx->uring_lock); | |
0ced756f | 333 | if (atomic_read(&data->refs) <= 0) |
73572984 | 334 | break; |
0ced756f PB |
335 | /* |
336 | * it has been revived by another thread while | |
337 | * we were unlocked | |
338 | */ | |
339 | mutex_unlock(&ctx->uring_lock); | |
73572984 | 340 | } |
77e3202a | 341 | } while (1); |
73572984 JA |
342 | data->quiesce = false; |
343 | ||
344 | return ret; | |
345 | } | |
346 | ||
347 | static void io_free_page_table(void **table, size_t size) | |
348 | { | |
349 | unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); | |
350 | ||
351 | for (i = 0; i < nr_tables; i++) | |
352 | kfree(table[i]); | |
353 | kfree(table); | |
354 | } | |
355 | ||
356 | static void io_rsrc_data_free(struct io_rsrc_data *data) | |
357 | { | |
358 | size_t size = data->nr * sizeof(data->tags[0][0]); | |
359 | ||
360 | if (data->tags) | |
361 | io_free_page_table((void **)data->tags, size); | |
362 | kfree(data); | |
363 | } | |
364 | ||
365 | static __cold void **io_alloc_page_table(size_t size) | |
366 | { | |
367 | unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); | |
368 | size_t init_size = size; | |
369 | void **table; | |
370 | ||
371 | table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT); | |
372 | if (!table) | |
373 | return NULL; | |
374 | ||
375 | for (i = 0; i < nr_tables; i++) { | |
376 | unsigned int this_size = min_t(size_t, size, PAGE_SIZE); | |
377 | ||
378 | table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT); | |
379 | if (!table[i]) { | |
380 | io_free_page_table(table, init_size); | |
381 | return NULL; | |
382 | } | |
383 | size -= this_size; | |
384 | } | |
385 | return table; | |
386 | } | |
387 | ||
388 | __cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, | |
389 | rsrc_put_fn *do_put, u64 __user *utags, | |
390 | unsigned nr, struct io_rsrc_data **pdata) | |
391 | { | |
392 | struct io_rsrc_data *data; | |
6acd352d | 393 | int ret = 0; |
73572984 JA |
394 | unsigned i; |
395 | ||
396 | data = kzalloc(sizeof(*data), GFP_KERNEL); | |
397 | if (!data) | |
398 | return -ENOMEM; | |
399 | data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0])); | |
400 | if (!data->tags) { | |
401 | kfree(data); | |
402 | return -ENOMEM; | |
403 | } | |
404 | ||
405 | data->nr = nr; | |
406 | data->ctx = ctx; | |
407 | data->do_put = do_put; | |
408 | if (utags) { | |
409 | ret = -EFAULT; | |
410 | for (i = 0; i < nr; i++) { | |
411 | u64 *tag_slot = io_get_tag_slot(data, i); | |
412 | ||
413 | if (copy_from_user(tag_slot, &utags[i], | |
414 | sizeof(*tag_slot))) | |
415 | goto fail; | |
416 | } | |
417 | } | |
418 | ||
419 | atomic_set(&data->refs, 1); | |
420 | init_completion(&data->done); | |
421 | *pdata = data; | |
422 | return 0; | |
423 | fail: | |
424 | io_rsrc_data_free(data); | |
425 | return ret; | |
426 | } | |
427 | ||
428 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, | |
429 | struct io_uring_rsrc_update2 *up, | |
430 | unsigned nr_args) | |
431 | { | |
432 | u64 __user *tags = u64_to_user_ptr(up->tags); | |
433 | __s32 __user *fds = u64_to_user_ptr(up->data); | |
434 | struct io_rsrc_data *data = ctx->file_data; | |
435 | struct io_fixed_file *file_slot; | |
436 | struct file *file; | |
437 | int fd, i, err = 0; | |
438 | unsigned int done; | |
439 | bool needs_switch = false; | |
440 | ||
441 | if (!ctx->file_data) | |
442 | return -ENXIO; | |
443 | if (up->offset + nr_args > ctx->nr_user_files) | |
444 | return -EINVAL; | |
445 | ||
446 | for (done = 0; done < nr_args; done++) { | |
447 | u64 tag = 0; | |
448 | ||
449 | if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) || | |
450 | copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
451 | err = -EFAULT; | |
452 | break; | |
453 | } | |
454 | if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) { | |
455 | err = -EINVAL; | |
456 | break; | |
457 | } | |
458 | if (fd == IORING_REGISTER_FILES_SKIP) | |
459 | continue; | |
460 | ||
461 | i = array_index_nospec(up->offset + done, ctx->nr_user_files); | |
462 | file_slot = io_fixed_file_slot(&ctx->file_table, i); | |
463 | ||
464 | if (file_slot->file_ptr) { | |
465 | file = (struct file *)(file_slot->file_ptr & FFS_MASK); | |
466 | err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file); | |
467 | if (err) | |
468 | break; | |
469 | file_slot->file_ptr = 0; | |
470 | io_file_bitmap_clear(&ctx->file_table, i); | |
471 | needs_switch = true; | |
472 | } | |
473 | if (fd != -1) { | |
474 | file = fget(fd); | |
475 | if (!file) { | |
476 | err = -EBADF; | |
477 | break; | |
478 | } | |
479 | /* | |
480 | * Don't allow io_uring instances to be registered. If | |
481 | * UNIX isn't enabled, then this causes a reference | |
482 | * cycle and this instance can never get freed. If UNIX | |
483 | * is enabled we'll handle it just fine, but there's | |
484 | * still no point in allowing a ring fd as it doesn't | |
485 | * support regular read/write anyway. | |
486 | */ | |
487 | if (io_is_uring_fops(file)) { | |
488 | fput(file); | |
489 | err = -EBADF; | |
490 | break; | |
491 | } | |
492 | err = io_scm_file_account(ctx, file); | |
493 | if (err) { | |
494 | fput(file); | |
495 | break; | |
496 | } | |
497 | *io_get_tag_slot(data, i) = tag; | |
498 | io_fixed_file_set(file_slot, file); | |
499 | io_file_bitmap_set(&ctx->file_table, i); | |
500 | } | |
501 | } | |
502 | ||
503 | if (needs_switch) | |
504 | io_rsrc_node_switch(ctx, data); | |
505 | return done ? done : err; | |
506 | } | |
507 | ||
508 | static int __io_sqe_buffers_update(struct io_ring_ctx *ctx, | |
509 | struct io_uring_rsrc_update2 *up, | |
510 | unsigned int nr_args) | |
511 | { | |
512 | u64 __user *tags = u64_to_user_ptr(up->tags); | |
513 | struct iovec iov, __user *iovs = u64_to_user_ptr(up->data); | |
514 | struct page *last_hpage = NULL; | |
515 | bool needs_switch = false; | |
516 | __u32 done; | |
517 | int i, err; | |
518 | ||
519 | if (!ctx->buf_data) | |
520 | return -ENXIO; | |
521 | if (up->offset + nr_args > ctx->nr_user_bufs) | |
522 | return -EINVAL; | |
523 | ||
524 | for (done = 0; done < nr_args; done++) { | |
525 | struct io_mapped_ubuf *imu; | |
526 | int offset = up->offset + done; | |
527 | u64 tag = 0; | |
528 | ||
529 | err = io_copy_iov(ctx, &iov, iovs, done); | |
530 | if (err) | |
531 | break; | |
532 | if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) { | |
533 | err = -EFAULT; | |
534 | break; | |
535 | } | |
536 | err = io_buffer_validate(&iov); | |
537 | if (err) | |
538 | break; | |
539 | if (!iov.iov_base && tag) { | |
540 | err = -EINVAL; | |
541 | break; | |
542 | } | |
543 | err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage); | |
544 | if (err) | |
545 | break; | |
546 | ||
547 | i = array_index_nospec(offset, ctx->nr_user_bufs); | |
548 | if (ctx->user_bufs[i] != ctx->dummy_ubuf) { | |
549 | err = io_queue_rsrc_removal(ctx->buf_data, i, | |
550 | ctx->rsrc_node, ctx->user_bufs[i]); | |
551 | if (unlikely(err)) { | |
552 | io_buffer_unmap(ctx, &imu); | |
553 | break; | |
554 | } | |
5ff4fdff | 555 | ctx->user_bufs[i] = ctx->dummy_ubuf; |
73572984 JA |
556 | needs_switch = true; |
557 | } | |
558 | ||
559 | ctx->user_bufs[i] = imu; | |
560 | *io_get_tag_slot(ctx->buf_data, offset) = tag; | |
561 | } | |
562 | ||
563 | if (needs_switch) | |
564 | io_rsrc_node_switch(ctx, ctx->buf_data); | |
565 | return done ? done : err; | |
566 | } | |
567 | ||
568 | static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type, | |
569 | struct io_uring_rsrc_update2 *up, | |
570 | unsigned nr_args) | |
571 | { | |
572 | __u32 tmp; | |
573 | int err; | |
574 | ||
575 | if (check_add_overflow(up->offset, nr_args, &tmp)) | |
576 | return -EOVERFLOW; | |
577 | err = io_rsrc_node_switch_start(ctx); | |
578 | if (err) | |
579 | return err; | |
580 | ||
581 | switch (type) { | |
582 | case IORING_RSRC_FILE: | |
583 | return __io_sqe_files_update(ctx, up, nr_args); | |
584 | case IORING_RSRC_BUFFER: | |
585 | return __io_sqe_buffers_update(ctx, up, nr_args); | |
586 | } | |
587 | return -EINVAL; | |
588 | } | |
589 | ||
590 | int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg, | |
591 | unsigned nr_args) | |
592 | { | |
593 | struct io_uring_rsrc_update2 up; | |
594 | ||
595 | if (!nr_args) | |
596 | return -EINVAL; | |
597 | memset(&up, 0, sizeof(up)); | |
598 | if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update))) | |
599 | return -EFAULT; | |
600 | if (up.resv || up.resv2) | |
601 | return -EINVAL; | |
602 | return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args); | |
603 | } | |
604 | ||
605 | int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg, | |
606 | unsigned size, unsigned type) | |
607 | { | |
608 | struct io_uring_rsrc_update2 up; | |
609 | ||
610 | if (size != sizeof(up)) | |
611 | return -EINVAL; | |
612 | if (copy_from_user(&up, arg, sizeof(up))) | |
613 | return -EFAULT; | |
614 | if (!up.nr || up.resv || up.resv2) | |
615 | return -EINVAL; | |
616 | return __io_register_rsrc_update(ctx, type, &up, up.nr); | |
617 | } | |
618 | ||
619 | __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg, | |
620 | unsigned int size, unsigned int type) | |
621 | { | |
622 | struct io_uring_rsrc_register rr; | |
623 | ||
624 | /* keep it extendible */ | |
625 | if (size != sizeof(rr)) | |
626 | return -EINVAL; | |
627 | ||
628 | memset(&rr, 0, sizeof(rr)); | |
629 | if (copy_from_user(&rr, arg, size)) | |
630 | return -EFAULT; | |
631 | if (!rr.nr || rr.resv2) | |
632 | return -EINVAL; | |
633 | if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE) | |
634 | return -EINVAL; | |
635 | ||
636 | switch (type) { | |
637 | case IORING_RSRC_FILE: | |
638 | if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data) | |
639 | break; | |
640 | return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data), | |
641 | rr.nr, u64_to_user_ptr(rr.tags)); | |
642 | case IORING_RSRC_BUFFER: | |
643 | if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data) | |
644 | break; | |
645 | return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data), | |
646 | rr.nr, u64_to_user_ptr(rr.tags)); | |
647 | } | |
648 | return -EINVAL; | |
649 | } | |
650 | ||
d9808ceb | 651 | int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
73572984 | 652 | { |
f2ccb5ae | 653 | struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); |
73572984 JA |
654 | |
655 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) | |
656 | return -EINVAL; | |
657 | if (sqe->rw_flags || sqe->splice_fd_in) | |
658 | return -EINVAL; | |
659 | ||
660 | up->offset = READ_ONCE(sqe->off); | |
661 | up->nr_args = READ_ONCE(sqe->len); | |
662 | if (!up->nr_args) | |
663 | return -EINVAL; | |
664 | up->arg = READ_ONCE(sqe->addr); | |
665 | return 0; | |
666 | } | |
667 | ||
668 | static int io_files_update_with_index_alloc(struct io_kiocb *req, | |
669 | unsigned int issue_flags) | |
670 | { | |
f2ccb5ae | 671 | struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); |
73572984 JA |
672 | __s32 __user *fds = u64_to_user_ptr(up->arg); |
673 | unsigned int done; | |
674 | struct file *file; | |
675 | int ret, fd; | |
676 | ||
677 | if (!req->ctx->file_data) | |
678 | return -ENXIO; | |
679 | ||
680 | for (done = 0; done < up->nr_args; done++) { | |
681 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
682 | ret = -EFAULT; | |
683 | break; | |
684 | } | |
685 | ||
686 | file = fget(fd); | |
687 | if (!file) { | |
688 | ret = -EBADF; | |
689 | break; | |
690 | } | |
691 | ret = io_fixed_fd_install(req, issue_flags, file, | |
692 | IORING_FILE_INDEX_ALLOC); | |
693 | if (ret < 0) | |
694 | break; | |
695 | if (copy_to_user(&fds[done], &ret, sizeof(ret))) { | |
f110ed84 | 696 | __io_close_fixed(req->ctx, issue_flags, ret); |
73572984 JA |
697 | ret = -EFAULT; |
698 | break; | |
699 | } | |
700 | } | |
701 | ||
702 | if (done) | |
703 | return done; | |
704 | return ret; | |
705 | } | |
706 | ||
d9808ceb | 707 | int io_files_update(struct io_kiocb *req, unsigned int issue_flags) |
73572984 | 708 | { |
f2ccb5ae | 709 | struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); |
73572984 JA |
710 | struct io_ring_ctx *ctx = req->ctx; |
711 | struct io_uring_rsrc_update2 up2; | |
712 | int ret; | |
713 | ||
714 | up2.offset = up->offset; | |
715 | up2.data = up->arg; | |
716 | up2.nr = 0; | |
717 | up2.tags = 0; | |
718 | up2.resv = 0; | |
719 | up2.resv2 = 0; | |
720 | ||
721 | if (up->offset == IORING_FILE_INDEX_ALLOC) { | |
722 | ret = io_files_update_with_index_alloc(req, issue_flags); | |
723 | } else { | |
724 | io_ring_submit_lock(ctx, issue_flags); | |
725 | ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE, | |
726 | &up2, up->nr_args); | |
727 | io_ring_submit_unlock(ctx, issue_flags); | |
728 | } | |
729 | ||
730 | if (ret < 0) | |
731 | req_set_fail(req); | |
732 | io_req_set_res(req, ret, 0); | |
733 | return IOU_OK; | |
734 | } | |
735 | ||
736 | int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx, | |
737 | struct io_rsrc_node *node, void *rsrc) | |
738 | { | |
739 | u64 *tag_slot = io_get_tag_slot(data, idx); | |
740 | struct io_rsrc_put *prsrc; | |
ff7c75ec | 741 | bool inline_item = true; |
73572984 | 742 | |
ff7c75ec PB |
743 | if (!node->inline_items) { |
744 | prsrc = &node->item; | |
745 | node->inline_items++; | |
746 | } else { | |
747 | prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL); | |
748 | if (!prsrc) | |
749 | return -ENOMEM; | |
750 | inline_item = false; | |
751 | } | |
73572984 JA |
752 | |
753 | prsrc->tag = *tag_slot; | |
754 | *tag_slot = 0; | |
755 | prsrc->rsrc = rsrc; | |
ff7c75ec PB |
756 | if (!inline_item) |
757 | list_add(&prsrc->list, &node->item_list); | |
73572984 JA |
758 | return 0; |
759 | } | |
760 | ||
761 | void __io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
762 | { | |
73572984 JA |
763 | int i; |
764 | ||
765 | for (i = 0; i < ctx->nr_user_files; i++) { | |
766 | struct file *file = io_file_from_index(&ctx->file_table, i); | |
767 | ||
38eddb2c PB |
768 | /* skip scm accounted files, they'll be freed by ->ring_sock */ |
769 | if (!file || io_file_need_scm(file)) | |
73572984 JA |
770 | continue; |
771 | io_file_bitmap_clear(&ctx->file_table, i); | |
772 | fput(file); | |
773 | } | |
73572984 JA |
774 | |
775 | #if defined(CONFIG_UNIX) | |
776 | if (ctx->ring_sock) { | |
777 | struct sock *sock = ctx->ring_sock->sk; | |
778 | struct sk_buff *skb; | |
779 | ||
780 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
781 | kfree_skb(skb); | |
782 | } | |
783 | #endif | |
784 | io_free_file_tables(&ctx->file_table); | |
02a4d923 | 785 | io_file_table_set_alloc_range(ctx, 0, 0); |
73572984 JA |
786 | io_rsrc_data_free(ctx->file_data); |
787 | ctx->file_data = NULL; | |
788 | ctx->nr_user_files = 0; | |
789 | } | |
790 | ||
791 | int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
792 | { | |
793 | unsigned nr = ctx->nr_user_files; | |
794 | int ret; | |
795 | ||
796 | if (!ctx->file_data) | |
797 | return -ENXIO; | |
798 | ||
799 | /* | |
800 | * Quiesce may unlock ->uring_lock, and while it's not held | |
801 | * prevent new requests using the table. | |
802 | */ | |
803 | ctx->nr_user_files = 0; | |
804 | ret = io_rsrc_ref_quiesce(ctx->file_data, ctx); | |
805 | ctx->nr_user_files = nr; | |
806 | if (!ret) | |
807 | __io_sqe_files_unregister(ctx); | |
808 | return ret; | |
809 | } | |
810 | ||
811 | /* | |
812 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
813 | * the io_uring can be safely unregistered on process exit, even if we have | |
814 | * loops in the file referencing. We account only files that can hold other | |
815 | * files because otherwise they can't form a loop and so are not interesting | |
816 | * for GC. | |
817 | */ | |
818 | int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file) | |
819 | { | |
820 | #if defined(CONFIG_UNIX) | |
821 | struct sock *sk = ctx->ring_sock->sk; | |
822 | struct sk_buff_head *head = &sk->sk_receive_queue; | |
823 | struct scm_fp_list *fpl; | |
824 | struct sk_buff *skb; | |
825 | ||
826 | if (likely(!io_file_need_scm(file))) | |
827 | return 0; | |
828 | ||
829 | /* | |
830 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
831 | * file set. If there's no room, fall back to allocating a new skb | |
832 | * and filling it in. | |
833 | */ | |
834 | spin_lock_irq(&head->lock); | |
835 | skb = skb_peek(head); | |
836 | if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD) | |
837 | __skb_unlink(skb, head); | |
838 | else | |
839 | skb = NULL; | |
840 | spin_unlock_irq(&head->lock); | |
841 | ||
842 | if (!skb) { | |
843 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
844 | if (!fpl) | |
845 | return -ENOMEM; | |
846 | ||
847 | skb = alloc_skb(0, GFP_KERNEL); | |
848 | if (!skb) { | |
849 | kfree(fpl); | |
850 | return -ENOMEM; | |
851 | } | |
852 | ||
853 | fpl->user = get_uid(current_user()); | |
854 | fpl->max = SCM_MAX_FD; | |
855 | fpl->count = 0; | |
856 | ||
857 | UNIXCB(skb).fp = fpl; | |
858 | skb->sk = sk; | |
0091bfc8 | 859 | skb->scm_io_uring = 1; |
73572984 JA |
860 | skb->destructor = unix_destruct_scm; |
861 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
862 | } | |
863 | ||
864 | fpl = UNIXCB(skb).fp; | |
865 | fpl->fp[fpl->count++] = get_file(file); | |
866 | unix_inflight(fpl->user, file); | |
867 | skb_queue_head(head, skb); | |
868 | fput(file); | |
869 | #endif | |
870 | return 0; | |
871 | } | |
872 | ||
873 | static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) | |
874 | { | |
875 | struct file *file = prsrc->file; | |
876 | #if defined(CONFIG_UNIX) | |
877 | struct sock *sock = ctx->ring_sock->sk; | |
878 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
879 | struct sk_buff *skb; | |
880 | int i; | |
881 | ||
882 | if (!io_file_need_scm(file)) { | |
883 | fput(file); | |
884 | return; | |
885 | } | |
886 | ||
887 | __skb_queue_head_init(&list); | |
888 | ||
889 | /* | |
890 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
891 | * remove this entry and rearrange the file array. | |
892 | */ | |
893 | skb = skb_dequeue(head); | |
894 | while (skb) { | |
895 | struct scm_fp_list *fp; | |
896 | ||
897 | fp = UNIXCB(skb).fp; | |
898 | for (i = 0; i < fp->count; i++) { | |
899 | int left; | |
900 | ||
901 | if (fp->fp[i] != file) | |
902 | continue; | |
903 | ||
904 | unix_notinflight(fp->user, fp->fp[i]); | |
905 | left = fp->count - 1 - i; | |
906 | if (left) { | |
907 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
908 | left * sizeof(struct file *)); | |
909 | } | |
910 | fp->count--; | |
911 | if (!fp->count) { | |
912 | kfree_skb(skb); | |
913 | skb = NULL; | |
914 | } else { | |
915 | __skb_queue_tail(&list, skb); | |
916 | } | |
917 | fput(file); | |
918 | file = NULL; | |
919 | break; | |
920 | } | |
921 | ||
922 | if (!file) | |
923 | break; | |
924 | ||
925 | __skb_queue_tail(&list, skb); | |
926 | ||
927 | skb = skb_dequeue(head); | |
928 | } | |
929 | ||
930 | if (skb_peek(&list)) { | |
931 | spin_lock_irq(&head->lock); | |
932 | while ((skb = __skb_dequeue(&list)) != NULL) | |
933 | __skb_queue_tail(head, skb); | |
934 | spin_unlock_irq(&head->lock); | |
935 | } | |
936 | #else | |
937 | fput(file); | |
938 | #endif | |
939 | } | |
940 | ||
941 | int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, | |
942 | unsigned nr_args, u64 __user *tags) | |
943 | { | |
944 | __s32 __user *fds = (__s32 __user *) arg; | |
945 | struct file *file; | |
946 | int fd, ret; | |
947 | unsigned i; | |
948 | ||
949 | if (ctx->file_data) | |
950 | return -EBUSY; | |
951 | if (!nr_args) | |
952 | return -EINVAL; | |
953 | if (nr_args > IORING_MAX_FIXED_FILES) | |
954 | return -EMFILE; | |
955 | if (nr_args > rlimit(RLIMIT_NOFILE)) | |
956 | return -EMFILE; | |
957 | ret = io_rsrc_node_switch_start(ctx); | |
958 | if (ret) | |
959 | return ret; | |
960 | ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args, | |
961 | &ctx->file_data); | |
962 | if (ret) | |
963 | return ret; | |
964 | ||
965 | if (!io_alloc_file_tables(&ctx->file_table, nr_args)) { | |
966 | io_rsrc_data_free(ctx->file_data); | |
967 | ctx->file_data = NULL; | |
968 | return -ENOMEM; | |
969 | } | |
970 | ||
971 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { | |
972 | struct io_fixed_file *file_slot; | |
973 | ||
974 | if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) { | |
975 | ret = -EFAULT; | |
976 | goto fail; | |
977 | } | |
978 | /* allow sparse sets */ | |
979 | if (!fds || fd == -1) { | |
980 | ret = -EINVAL; | |
981 | if (unlikely(*io_get_tag_slot(ctx->file_data, i))) | |
982 | goto fail; | |
983 | continue; | |
984 | } | |
985 | ||
986 | file = fget(fd); | |
987 | ret = -EBADF; | |
988 | if (unlikely(!file)) | |
989 | goto fail; | |
990 | ||
991 | /* | |
992 | * Don't allow io_uring instances to be registered. If UNIX | |
993 | * isn't enabled, then this causes a reference cycle and this | |
994 | * instance can never get freed. If UNIX is enabled we'll | |
995 | * handle it just fine, but there's still no point in allowing | |
996 | * a ring fd as it doesn't support regular read/write anyway. | |
997 | */ | |
998 | if (io_is_uring_fops(file)) { | |
999 | fput(file); | |
1000 | goto fail; | |
1001 | } | |
1002 | ret = io_scm_file_account(ctx, file); | |
1003 | if (ret) { | |
1004 | fput(file); | |
1005 | goto fail; | |
1006 | } | |
1007 | file_slot = io_fixed_file_slot(&ctx->file_table, i); | |
1008 | io_fixed_file_set(file_slot, file); | |
1009 | io_file_bitmap_set(&ctx->file_table, i); | |
1010 | } | |
1011 | ||
6e73dffb PB |
1012 | /* default it to the whole table */ |
1013 | io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files); | |
73572984 JA |
1014 | io_rsrc_node_switch(ctx, NULL); |
1015 | return 0; | |
1016 | fail: | |
1017 | __io_sqe_files_unregister(ctx); | |
1018 | return ret; | |
1019 | } | |
1020 | ||
1021 | static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) | |
1022 | { | |
1023 | io_buffer_unmap(ctx, &prsrc->buf); | |
1024 | prsrc->buf = NULL; | |
1025 | } | |
1026 | ||
1027 | void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx) | |
1028 | { | |
1029 | unsigned int i; | |
1030 | ||
1031 | for (i = 0; i < ctx->nr_user_bufs; i++) | |
1032 | io_buffer_unmap(ctx, &ctx->user_bufs[i]); | |
1033 | kfree(ctx->user_bufs); | |
1034 | io_rsrc_data_free(ctx->buf_data); | |
1035 | ctx->user_bufs = NULL; | |
1036 | ctx->buf_data = NULL; | |
1037 | ctx->nr_user_bufs = 0; | |
1038 | } | |
1039 | ||
1040 | int io_sqe_buffers_unregister(struct io_ring_ctx *ctx) | |
1041 | { | |
1042 | unsigned nr = ctx->nr_user_bufs; | |
1043 | int ret; | |
1044 | ||
1045 | if (!ctx->buf_data) | |
1046 | return -ENXIO; | |
1047 | ||
1048 | /* | |
1049 | * Quiesce may unlock ->uring_lock, and while it's not held | |
1050 | * prevent new requests using the table. | |
1051 | */ | |
1052 | ctx->nr_user_bufs = 0; | |
1053 | ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx); | |
1054 | ctx->nr_user_bufs = nr; | |
1055 | if (!ret) | |
1056 | __io_sqe_buffers_unregister(ctx); | |
1057 | return ret; | |
1058 | } | |
1059 | ||
1060 | /* | |
1061 | * Not super efficient, but this is just a registration time. And we do cache | |
1062 | * the last compound head, so generally we'll only do a full search if we don't | |
1063 | * match that one. | |
1064 | * | |
1065 | * We check if the given compound head page has already been accounted, to | |
1066 | * avoid double accounting it. This allows us to account the full size of the | |
1067 | * page, not just the constituent pages of a huge page. | |
1068 | */ | |
1069 | static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages, | |
1070 | int nr_pages, struct page *hpage) | |
1071 | { | |
1072 | int i, j; | |
1073 | ||
1074 | /* check current page array */ | |
1075 | for (i = 0; i < nr_pages; i++) { | |
1076 | if (!PageCompound(pages[i])) | |
1077 | continue; | |
1078 | if (compound_head(pages[i]) == hpage) | |
1079 | return true; | |
1080 | } | |
1081 | ||
1082 | /* check previously registered pages */ | |
1083 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
1084 | struct io_mapped_ubuf *imu = ctx->user_bufs[i]; | |
1085 | ||
1086 | for (j = 0; j < imu->nr_bvecs; j++) { | |
1087 | if (!PageCompound(imu->bvec[j].bv_page)) | |
1088 | continue; | |
1089 | if (compound_head(imu->bvec[j].bv_page) == hpage) | |
1090 | return true; | |
1091 | } | |
1092 | } | |
1093 | ||
1094 | return false; | |
1095 | } | |
1096 | ||
1097 | static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages, | |
1098 | int nr_pages, struct io_mapped_ubuf *imu, | |
1099 | struct page **last_hpage) | |
1100 | { | |
1101 | int i, ret; | |
1102 | ||
1103 | imu->acct_pages = 0; | |
1104 | for (i = 0; i < nr_pages; i++) { | |
1105 | if (!PageCompound(pages[i])) { | |
1106 | imu->acct_pages++; | |
1107 | } else { | |
1108 | struct page *hpage; | |
1109 | ||
1110 | hpage = compound_head(pages[i]); | |
1111 | if (hpage == *last_hpage) | |
1112 | continue; | |
1113 | *last_hpage = hpage; | |
1114 | if (headpage_already_acct(ctx, pages, i, hpage)) | |
1115 | continue; | |
1116 | imu->acct_pages += page_size(hpage) >> PAGE_SHIFT; | |
1117 | } | |
1118 | } | |
1119 | ||
1120 | if (!imu->acct_pages) | |
1121 | return 0; | |
1122 | ||
1123 | ret = io_account_mem(ctx, imu->acct_pages); | |
1124 | if (ret) | |
1125 | imu->acct_pages = 0; | |
1126 | return ret; | |
1127 | } | |
1128 | ||
1129 | struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages) | |
1130 | { | |
1131 | unsigned long start, end, nr_pages; | |
1132 | struct vm_area_struct **vmas = NULL; | |
1133 | struct page **pages = NULL; | |
1134 | int i, pret, ret = -ENOMEM; | |
1135 | ||
1136 | end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
1137 | start = ubuf >> PAGE_SHIFT; | |
1138 | nr_pages = end - start; | |
1139 | ||
1140 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); | |
1141 | if (!pages) | |
1142 | goto done; | |
1143 | ||
1144 | vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *), | |
1145 | GFP_KERNEL); | |
1146 | if (!vmas) | |
1147 | goto done; | |
1148 | ||
1149 | ret = 0; | |
1150 | mmap_read_lock(current->mm); | |
1151 | pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM, | |
1152 | pages, vmas); | |
1153 | if (pret == nr_pages) { | |
edd47826 PB |
1154 | struct file *file = vmas[0]->vm_file; |
1155 | ||
73572984 JA |
1156 | /* don't support file backed memory */ |
1157 | for (i = 0; i < nr_pages; i++) { | |
edd47826 PB |
1158 | if (vmas[i]->vm_file != file) { |
1159 | ret = -EINVAL; | |
1160 | break; | |
1161 | } | |
1162 | if (!file) | |
73572984 | 1163 | continue; |
edd47826 | 1164 | if (!vma_is_shmem(vmas[i]) && !is_file_hugepages(file)) { |
73572984 JA |
1165 | ret = -EOPNOTSUPP; |
1166 | break; | |
1167 | } | |
1168 | } | |
1169 | *npages = nr_pages; | |
1170 | } else { | |
1171 | ret = pret < 0 ? pret : -EFAULT; | |
1172 | } | |
1173 | mmap_read_unlock(current->mm); | |
1174 | if (ret) { | |
1175 | /* | |
1176 | * if we did partial map, or found file backed vmas, | |
1177 | * release any pages we did get | |
1178 | */ | |
1179 | if (pret > 0) | |
1180 | unpin_user_pages(pages, pret); | |
1181 | goto done; | |
1182 | } | |
1183 | ret = 0; | |
1184 | done: | |
1185 | kvfree(vmas); | |
1186 | if (ret < 0) { | |
1187 | kvfree(pages); | |
1188 | pages = ERR_PTR(ret); | |
1189 | } | |
1190 | return pages; | |
1191 | } | |
1192 | ||
1193 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov, | |
1194 | struct io_mapped_ubuf **pimu, | |
1195 | struct page **last_hpage) | |
1196 | { | |
1197 | struct io_mapped_ubuf *imu = NULL; | |
1198 | struct page **pages = NULL; | |
1199 | unsigned long off; | |
1200 | size_t size; | |
1201 | int ret, nr_pages, i; | |
977bc873 | 1202 | struct folio *folio = NULL; |
73572984 | 1203 | |
5ff4fdff PB |
1204 | *pimu = ctx->dummy_ubuf; |
1205 | if (!iov->iov_base) | |
73572984 | 1206 | return 0; |
73572984 | 1207 | |
73572984 | 1208 | ret = -ENOMEM; |
73572984 JA |
1209 | pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len, |
1210 | &nr_pages); | |
1211 | if (IS_ERR(pages)) { | |
1212 | ret = PTR_ERR(pages); | |
1213 | pages = NULL; | |
1214 | goto done; | |
1215 | } | |
1216 | ||
57bebf80 PB |
1217 | /* If it's a huge page, try to coalesce them into a single bvec entry */ |
1218 | if (nr_pages > 1) { | |
1219 | folio = page_folio(pages[0]); | |
1220 | for (i = 1; i < nr_pages; i++) { | |
1221 | if (page_folio(pages[i]) != folio) { | |
1222 | folio = NULL; | |
1223 | break; | |
1224 | } | |
1225 | } | |
1226 | if (folio) { | |
d2acf789 PB |
1227 | /* |
1228 | * The pages are bound to the folio, it doesn't | |
1229 | * actually unpin them but drops all but one reference, | |
1230 | * which is usually put down by io_buffer_unmap(). | |
1231 | * Note, needs a better helper. | |
1232 | */ | |
1233 | unpin_user_pages(&pages[1], nr_pages - 1); | |
57bebf80 PB |
1234 | nr_pages = 1; |
1235 | } | |
1236 | } | |
1237 | ||
73572984 JA |
1238 | imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL); |
1239 | if (!imu) | |
1240 | goto done; | |
1241 | ||
1242 | ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage); | |
1243 | if (ret) { | |
1244 | unpin_user_pages(pages, nr_pages); | |
1245 | goto done; | |
1246 | } | |
1247 | ||
1248 | off = (unsigned long) iov->iov_base & ~PAGE_MASK; | |
1249 | size = iov->iov_len; | |
57bebf80 PB |
1250 | /* store original address for later verification */ |
1251 | imu->ubuf = (unsigned long) iov->iov_base; | |
1252 | imu->ubuf_end = imu->ubuf + iov->iov_len; | |
1253 | imu->nr_bvecs = nr_pages; | |
1254 | *pimu = imu; | |
1255 | ret = 0; | |
1256 | ||
1257 | if (folio) { | |
1258 | bvec_set_page(&imu->bvec[0], pages[0], size, off); | |
1259 | goto done; | |
1260 | } | |
73572984 JA |
1261 | for (i = 0; i < nr_pages; i++) { |
1262 | size_t vec_len; | |
1263 | ||
1264 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
cc342a21 | 1265 | bvec_set_page(&imu->bvec[i], pages[i], vec_len, off); |
73572984 JA |
1266 | off = 0; |
1267 | size -= vec_len; | |
1268 | } | |
73572984 JA |
1269 | done: |
1270 | if (ret) | |
1271 | kvfree(imu); | |
1272 | kvfree(pages); | |
1273 | return ret; | |
1274 | } | |
1275 | ||
1276 | static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args) | |
1277 | { | |
1278 | ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL); | |
1279 | return ctx->user_bufs ? 0 : -ENOMEM; | |
1280 | } | |
1281 | ||
1282 | int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg, | |
1283 | unsigned int nr_args, u64 __user *tags) | |
1284 | { | |
1285 | struct page *last_hpage = NULL; | |
1286 | struct io_rsrc_data *data; | |
1287 | int i, ret; | |
1288 | struct iovec iov; | |
1289 | ||
1290 | BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16)); | |
1291 | ||
1292 | if (ctx->user_bufs) | |
1293 | return -EBUSY; | |
1294 | if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS) | |
1295 | return -EINVAL; | |
1296 | ret = io_rsrc_node_switch_start(ctx); | |
1297 | if (ret) | |
1298 | return ret; | |
1299 | ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data); | |
1300 | if (ret) | |
1301 | return ret; | |
1302 | ret = io_buffers_map_alloc(ctx, nr_args); | |
1303 | if (ret) { | |
1304 | io_rsrc_data_free(data); | |
1305 | return ret; | |
1306 | } | |
1307 | ||
1308 | for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) { | |
1309 | if (arg) { | |
1310 | ret = io_copy_iov(ctx, &iov, arg, i); | |
1311 | if (ret) | |
1312 | break; | |
1313 | ret = io_buffer_validate(&iov); | |
1314 | if (ret) | |
1315 | break; | |
1316 | } else { | |
1317 | memset(&iov, 0, sizeof(iov)); | |
1318 | } | |
1319 | ||
1320 | if (!iov.iov_base && *io_get_tag_slot(data, i)) { | |
1321 | ret = -EINVAL; | |
1322 | break; | |
1323 | } | |
1324 | ||
1325 | ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i], | |
1326 | &last_hpage); | |
1327 | if (ret) | |
1328 | break; | |
1329 | } | |
1330 | ||
1331 | WARN_ON_ONCE(ctx->buf_data); | |
1332 | ||
1333 | ctx->buf_data = data; | |
1334 | if (ret) | |
1335 | __io_sqe_buffers_unregister(ctx); | |
1336 | else | |
1337 | io_rsrc_node_switch(ctx, NULL); | |
1338 | return ret; | |
1339 | } | |
c059f785 PB |
1340 | |
1341 | int io_import_fixed(int ddir, struct iov_iter *iter, | |
1342 | struct io_mapped_ubuf *imu, | |
1343 | u64 buf_addr, size_t len) | |
1344 | { | |
1345 | u64 buf_end; | |
1346 | size_t offset; | |
1347 | ||
1348 | if (WARN_ON_ONCE(!imu)) | |
1349 | return -EFAULT; | |
1350 | if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end))) | |
1351 | return -EFAULT; | |
1352 | /* not inside the mapped region */ | |
1353 | if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end)) | |
1354 | return -EFAULT; | |
1355 | ||
1356 | /* | |
6bf65a1b | 1357 | * Might not be a start of buffer, set size appropriately |
c059f785 PB |
1358 | * and advance us to the beginning. |
1359 | */ | |
1360 | offset = buf_addr - imu->ubuf; | |
1361 | iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len); | |
1362 | ||
1363 | if (offset) { | |
1364 | /* | |
1365 | * Don't use iov_iter_advance() here, as it's really slow for | |
1366 | * using the latter parts of a big fixed buffer - it iterates | |
1367 | * over each segment manually. We can cheat a bit here, because | |
1368 | * we know that: | |
1369 | * | |
1370 | * 1) it's a BVEC iter, we set it up | |
1371 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1372 | * first and last bvec | |
1373 | * | |
1374 | * So just find our index, and adjust the iterator afterwards. | |
1375 | * If the offset is within the first bvec (or the whole first | |
1376 | * bvec, just use iov_iter_advance(). This makes it easier | |
1377 | * since we can just skip the first segment, which may not | |
1378 | * be PAGE_SIZE aligned. | |
1379 | */ | |
1380 | const struct bio_vec *bvec = imu->bvec; | |
1381 | ||
1382 | if (offset <= bvec->bv_len) { | |
57bebf80 PB |
1383 | /* |
1384 | * Note, huge pages buffers consists of one large | |
1385 | * bvec entry and should always go this way. The other | |
1386 | * branch doesn't expect non PAGE_SIZE'd chunks. | |
1387 | */ | |
b000ae0e PB |
1388 | iter->bvec = bvec; |
1389 | iter->nr_segs = bvec->bv_len; | |
1390 | iter->count -= offset; | |
1391 | iter->iov_offset = offset; | |
c059f785 PB |
1392 | } else { |
1393 | unsigned long seg_skip; | |
1394 | ||
1395 | /* skip first vec */ | |
1396 | offset -= bvec->bv_len; | |
1397 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
1398 | ||
1399 | iter->bvec = bvec + seg_skip; | |
1400 | iter->nr_segs -= seg_skip; | |
1401 | iter->count -= bvec->bv_len + offset; | |
1402 | iter->iov_offset = offset & ~PAGE_MASK; | |
1403 | } | |
1404 | } | |
1405 | ||
1406 | return 0; | |
1407 | } |