projects / linux-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
io_uring/rsrc: zero node's rsrc data on alloc
[linux-block.git] / io_uring / rsrc.c
CommitLineData
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
19struct io_rsrc_update {
20 struct file *file;
21 u64 arg;
22 u32 nr_args;
23 u32 offset;
24};
25
26static 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
757ef468
PB		 34static inline bool io_put_rsrc_data_ref(struct io_rsrc_data *rsrc_data)
35{
36 return !--rsrc_data->refs;
37}
38
6a9ce66f 39int __io_account_mem(struct user_struct *user, unsigned long nr_pages)
73572984
JA		 40{
41 unsigned long page_limit, cur_pages, new_pages;
42
6a9ce66f
PB		 43 if (!nr_pages)
44 return 0;
45
73572984
JA		 46 /* Don't allow more pages than we can safely lock */
47 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
48
4ccc6db0 49 cur_pages = atomic_long_read(&user->locked_vm);
73572984 50 do {
73572984
JA		 51 new_pages = cur_pages + nr_pages;
52 if (new_pages > page_limit)
53 return -ENOMEM;
4ccc6db0
UB		 54 } while (!atomic_long_try_cmpxchg(&user->locked_vm,
55 &cur_pages, new_pages));
73572984
JA		 56 return 0;
57}
58
59static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
60{
61 if (ctx->user)
62 __io_unaccount_mem(ctx->user, nr_pages);
63
64 if (ctx->mm_account)
65 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
66}
67
68static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
69{
70 int ret;
71
72 if (ctx->user) {
73 ret = __io_account_mem(ctx->user, nr_pages);
74 if (ret)
75 return ret;
76 }
77
78 if (ctx->mm_account)
79 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
80
81 return 0;
82}
83
84static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
85 void __user *arg, unsigned index)
86{
87 struct iovec __user *src;
88
89#ifdef CONFIG_COMPAT
90 if (ctx->compat) {
91 struct compat_iovec __user *ciovs;
92 struct compat_iovec ciov;
93
94 ciovs = (struct compat_iovec __user *) arg;
95 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
96 return -EFAULT;
97
98 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
99 dst->iov_len = ciov.iov_len;
100 return 0;
101 }
102#endif
103 src = (struct iovec __user *) arg;
104 if (copy_from_user(dst, &src[index], sizeof(*dst)))
105 return -EFAULT;
106 return 0;
107}
108
109static int io_buffer_validate(struct iovec *iov)
110{
111 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
112
113 /*
114 * Don't impose further limits on the size and buffer
115 * constraints here, we'll -EINVAL later when IO is
116 * submitted if they are wrong.
117 */
118 if (!iov->iov_base)
119 return iov->iov_len ? -EFAULT : 0;
120 if (!iov->iov_len)
121 return -EFAULT;
122
123 /* arbitrary limit, but we need something */
124 if (iov->iov_len > SZ_1G)
125 return -EFAULT;
126
127 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
128 return -EOVERFLOW;
129
130 return 0;
131}
132
133static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
134{
135 struct io_mapped_ubuf *imu = *slot;
136 unsigned int i;
137
138 if (imu != ctx->dummy_ubuf) {
139 for (i = 0; i < imu->nr_bvecs; i++)
140 unpin_user_page(imu->bvec[i].bv_page);
141 if (imu->acct_pages)
142 io_unaccount_mem(ctx, imu->acct_pages);
143 kvfree(imu);
144 }
145 *slot = NULL;
146}
147
ff7c75ec
PB		 148static void io_rsrc_put_work_one(struct io_rsrc_data *rsrc_data,
149 struct io_rsrc_put *prsrc)
150{
151 struct io_ring_ctx *ctx = rsrc_data->ctx;
152
36b9818a
PB		 153 if (prsrc->tag)
154 io_post_aux_cqe(ctx, prsrc->tag, 0, 0);
ff7c75ec
PB		 155 rsrc_data->do_put(ctx, prsrc);
156}
157
73572984
JA		 158static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
159{
160 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
73572984
JA		 161 struct io_rsrc_put *prsrc, *tmp;
162
ff7c75ec
PB		 163 if (ref_node->inline_items)
164 io_rsrc_put_work_one(rsrc_data, &ref_node->item);
165
c824986c 166 list_for_each_entry_safe(prsrc, tmp, &ref_node->item_list, list) {
73572984 167 list_del(&prsrc->list);
ff7c75ec 168 io_rsrc_put_work_one(rsrc_data, prsrc);
73572984
JA		 169 kfree(prsrc);
170 }
171
9eae8655 172 io_rsrc_node_destroy(rsrc_data->ctx, ref_node);
757ef468 173 if (io_put_rsrc_data_ref(rsrc_data))
73572984
JA		 174 complete(&rsrc_data->done);
175}
176
73572984
JA		 177void io_wait_rsrc_data(struct io_rsrc_data *data)
178{
757ef468 179 if (data && !io_put_rsrc_data_ref(data))
73572984
JA		 180 wait_for_completion(&data->done);
181}
182
9eae8655 183void io_rsrc_node_destroy(struct io_ring_ctx *ctx, struct io_rsrc_node *node)
73572984 184{
9eae8655
PB		 185 if (!io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache))
186 kfree(node);
73572984
JA		 187}
188
ef8ae64f
PB		 189void io_rsrc_node_ref_zero(struct io_rsrc_node *node)
190 __must_hold(&node->rsrc_data->ctx->uring_lock)
73572984 191{
73572984 192 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
73572984 193
73572984 194 node->done = true;
73572984
JA		 195 while (!list_empty(&ctx->rsrc_ref_list)) {
196 node = list_first_entry(&ctx->rsrc_ref_list,
197 struct io_rsrc_node, node);
198 /* recycle ref nodes in order */
199 if (!node->done)
200 break;
73572984 201
36b9818a
PB		 202 list_del(&node->node);
203 __io_rsrc_put_work(node);
d34b1b0b 204 }
73572984
JA		 205}
206
9eae8655 207static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
73572984
JA		 208{
209 struct io_rsrc_node *ref_node;
9eae8655
PB		 210 struct io_cache_entry *entry;
211
212 entry = io_alloc_cache_get(&ctx->rsrc_node_cache);
213 if (entry) {
214 ref_node = container_of(entry, struct io_rsrc_node, cache);
215 } else {
216 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
217 if (!ref_node)
218 return NULL;
219 }
73572984 220
13c22396 221 ref_node->rsrc_data = NULL;
ef8ae64f 222 ref_node->refs = 1;
73572984 223 INIT_LIST_HEAD(&ref_node->node);
c824986c 224 INIT_LIST_HEAD(&ref_node->item_list);
73572984 225 ref_node->done = false;
ff7c75ec 226 ref_node->inline_items = 0;
73572984
JA		 227 return ref_node;
228}
229
230void io_rsrc_node_switch(struct io_ring_ctx *ctx,
231 struct io_rsrc_data *data_to_kill)
232 __must_hold(&ctx->uring_lock)
233{
528407b1 234 WARN_ON_ONCE(io_alloc_cache_empty(&ctx->rsrc_node_cache));
73572984
JA		 235 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
236
73572984
JA		 237 if (data_to_kill) {
238 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
239
240 rsrc_node->rsrc_data = data_to_kill;
73572984 241 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
73572984 242
757ef468 243 data_to_kill->refs++;
b8fb5b4f 244 /* put master ref */
1f2c8f61 245 io_put_rsrc_node(ctx, rsrc_node);
73572984
JA		 246 ctx->rsrc_node = NULL;
247 }
248
528407b1
PB		 249 if (!ctx->rsrc_node)
250 ctx->rsrc_node = io_rsrc_node_alloc(ctx);
73572984
JA		 251}
252
253int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
254{
528407b1
PB		 255 if (io_alloc_cache_empty(&ctx->rsrc_node_cache)) {
256 struct io_rsrc_node *node = kzalloc(sizeof(*node), GFP_KERNEL);
257
258 if (!node)
259 return -ENOMEM;
260 io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache);
261 }
262 return 0;
73572984
JA		 263}
264
265__cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data,
266 struct io_ring_ctx *ctx)
267{
268 int ret;
269
270 /* As we may drop ->uring_lock, other task may have started quiesce */
271 if (data->quiesce)
272 return -ENXIO;
77e3202a
PB		 273 ret = io_rsrc_node_switch_start(ctx);
274 if (ret)
275 return ret;
276 io_rsrc_node_switch(ctx, data);
277
757ef468
PB		 278 /* kill initial ref */
279 if (io_put_rsrc_data_ref(data))
77e3202a 280 return 0;
73572984
JA		 281
282 data->quiesce = true;
77e3202a 283 mutex_unlock(&ctx->uring_lock);
73572984 284 do {
ef67fcb4 285 ret = io_run_task_work_sig(ctx);
77e3202a 286 if (ret < 0) {
77e3202a 287 mutex_lock(&ctx->uring_lock);
757ef468
PB		 288 if (!data->refs) {
289 ret = 0;
290 } else {
291 /* restore the master reference */
292 data->refs++;
293 }
77e3202a
PB		 294 break;
295 }
73572984
JA		 296 ret = wait_for_completion_interruptible(&data->done);
297 if (!ret) {
298 mutex_lock(&ctx->uring_lock);
757ef468 299 if (!data->refs)
73572984 300 break;
0ced756f
PB		 301 /*
302 * it has been revived by another thread while
303 * we were unlocked
304 */
305 mutex_unlock(&ctx->uring_lock);
73572984 306 }
77e3202a 307 } while (1);
73572984
JA		 308 data->quiesce = false;
309
310 return ret;
311}
312
313static void io_free_page_table(void **table, size_t size)
314{
315 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
316
317 for (i = 0; i < nr_tables; i++)
318 kfree(table[i]);
319 kfree(table);
320}
321
322static void io_rsrc_data_free(struct io_rsrc_data *data)
323{
324 size_t size = data->nr * sizeof(data->tags[0][0]);
325
326 if (data->tags)
327 io_free_page_table((void **)data->tags, size);
328 kfree(data);
329}
330
331static __cold void **io_alloc_page_table(size_t size)
332{
333 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
334 size_t init_size = size;
335 void **table;
336
337 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
338 if (!table)
339 return NULL;
340
341 for (i = 0; i < nr_tables; i++) {
342 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
343
344 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
345 if (!table[i]) {
346 io_free_page_table(table, init_size);
347 return NULL;
348 }
349 size -= this_size;
350 }
351 return table;
352}
353
354__cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx,
355 rsrc_put_fn *do_put, u64 __user *utags,
356 unsigned nr, struct io_rsrc_data **pdata)
357{
358 struct io_rsrc_data *data;
6acd352d 359 int ret = 0;
73572984
JA		 360 unsigned i;
361
362 data = kzalloc(sizeof(*data), GFP_KERNEL);
363 if (!data)
364 return -ENOMEM;
365 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
366 if (!data->tags) {
367 kfree(data);
368 return -ENOMEM;
369 }
370
371 data->nr = nr;
372 data->ctx = ctx;
373 data->do_put = do_put;
757ef468 374 data->refs = 1;
73572984
JA		 375 if (utags) {
376 ret = -EFAULT;
377 for (i = 0; i < nr; i++) {
378 u64 *tag_slot = io_get_tag_slot(data, i);
379
380 if (copy_from_user(tag_slot, &utags[i],
381 sizeof(*tag_slot)))
382 goto fail;
383 }
384 }
73572984
JA		 385 init_completion(&data->done);
386 *pdata = data;
387 return 0;
388fail:
389 io_rsrc_data_free(data);
390 return ret;
391}
392
393static int __io_sqe_files_update(struct io_ring_ctx *ctx,
394 struct io_uring_rsrc_update2 *up,
395 unsigned nr_args)
396{
397 u64 __user *tags = u64_to_user_ptr(up->tags);
398 __s32 __user *fds = u64_to_user_ptr(up->data);
399 struct io_rsrc_data *data = ctx->file_data;
400 struct io_fixed_file *file_slot;
401 struct file *file;
402 int fd, i, err = 0;
403 unsigned int done;
404 bool needs_switch = false;
405
406 if (!ctx->file_data)
407 return -ENXIO;
408 if (up->offset + nr_args > ctx->nr_user_files)
409 return -EINVAL;
410
411 for (done = 0; done < nr_args; done++) {
412 u64 tag = 0;
413
414 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
415 copy_from_user(&fd, &fds[done], sizeof(fd))) {
416 err = -EFAULT;
417 break;
418 }
419 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
420 err = -EINVAL;
421 break;
422 }
423 if (fd == IORING_REGISTER_FILES_SKIP)
424 continue;
425
426 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
427 file_slot = io_fixed_file_slot(&ctx->file_table, i);
428
429 if (file_slot->file_ptr) {
430 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
431 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file);
432 if (err)
433 break;
434 file_slot->file_ptr = 0;
435 io_file_bitmap_clear(&ctx->file_table, i);
436 needs_switch = true;
437 }
438 if (fd != -1) {
439 file = fget(fd);
440 if (!file) {
441 err = -EBADF;
442 break;
443 }
444 /*
445 * Don't allow io_uring instances to be registered. If
446 * UNIX isn't enabled, then this causes a reference
447 * cycle and this instance can never get freed. If UNIX
448 * is enabled we'll handle it just fine, but there's
449 * still no point in allowing a ring fd as it doesn't
450 * support regular read/write anyway.
451 */
452 if (io_is_uring_fops(file)) {
453 fput(file);
454 err = -EBADF;
455 break;
456 }
457 err = io_scm_file_account(ctx, file);
458 if (err) {
459 fput(file);
460 break;
461 }
462 *io_get_tag_slot(data, i) = tag;
463 io_fixed_file_set(file_slot, file);
464 io_file_bitmap_set(&ctx->file_table, i);
465 }
466 }
467
468 if (needs_switch)
469 io_rsrc_node_switch(ctx, data);
470 return done ? done : err;
471}
472
473static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
474 struct io_uring_rsrc_update2 *up,
475 unsigned int nr_args)
476{
477 u64 __user *tags = u64_to_user_ptr(up->tags);
478 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
479 struct page *last_hpage = NULL;
480 bool needs_switch = false;
481 __u32 done;
482 int i, err;
483
484 if (!ctx->buf_data)
485 return -ENXIO;
486 if (up->offset + nr_args > ctx->nr_user_bufs)
487 return -EINVAL;
488
489 for (done = 0; done < nr_args; done++) {
490 struct io_mapped_ubuf *imu;
491 int offset = up->offset + done;
492 u64 tag = 0;
493
494 err = io_copy_iov(ctx, &iov, iovs, done);
495 if (err)
496 break;
497 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
498 err = -EFAULT;
499 break;
500 }
501 err = io_buffer_validate(&iov);
502 if (err)
503 break;
504 if (!iov.iov_base && tag) {
505 err = -EINVAL;
506 break;
507 }
508 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
509 if (err)
510 break;
511
512 i = array_index_nospec(offset, ctx->nr_user_bufs);
513 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
514 err = io_queue_rsrc_removal(ctx->buf_data, i,
515 ctx->rsrc_node, ctx->user_bufs[i]);
516 if (unlikely(err)) {
517 io_buffer_unmap(ctx, &imu);
518 break;
519 }
5ff4fdff 520 ctx->user_bufs[i] = ctx->dummy_ubuf;
73572984
JA		 521 needs_switch = true;
522 }
523
524 ctx->user_bufs[i] = imu;
525 *io_get_tag_slot(ctx->buf_data, offset) = tag;
526 }
527
528 if (needs_switch)
529 io_rsrc_node_switch(ctx, ctx->buf_data);
530 return done ? done : err;
531}
532
533static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
534 struct io_uring_rsrc_update2 *up,
535 unsigned nr_args)
536{
537 __u32 tmp;
538 int err;
539
786788a8
PB		 540 lockdep_assert_held(&ctx->uring_lock);
541
73572984
JA		 542 if (check_add_overflow(up->offset, nr_args, &tmp))
543 return -EOVERFLOW;
544 err = io_rsrc_node_switch_start(ctx);
545 if (err)
546 return err;
547
548 switch (type) {
549 case IORING_RSRC_FILE:
550 return __io_sqe_files_update(ctx, up, nr_args);
551 case IORING_RSRC_BUFFER:
552 return __io_sqe_buffers_update(ctx, up, nr_args);
553 }
554 return -EINVAL;
555}
556
557int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
558 unsigned nr_args)
559{
560 struct io_uring_rsrc_update2 up;
561
562 if (!nr_args)
563 return -EINVAL;
564 memset(&up, 0, sizeof(up));
565 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
566 return -EFAULT;
567 if (up.resv || up.resv2)
568 return -EINVAL;
569 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
570}
571
572int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
573 unsigned size, unsigned type)
574{
575 struct io_uring_rsrc_update2 up;
576
577 if (size != sizeof(up))
578 return -EINVAL;
579 if (copy_from_user(&up, arg, sizeof(up)))
580 return -EFAULT;
581 if (!up.nr || up.resv || up.resv2)
582 return -EINVAL;
583 return __io_register_rsrc_update(ctx, type, &up, up.nr);
584}
585
586__cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
587 unsigned int size, unsigned int type)
588{
589 struct io_uring_rsrc_register rr;
590
591 /* keep it extendible */
592 if (size != sizeof(rr))
593 return -EINVAL;
594
595 memset(&rr, 0, sizeof(rr));
596 if (copy_from_user(&rr, arg, size))
597 return -EFAULT;
598 if (!rr.nr || rr.resv2)
599 return -EINVAL;
600 if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE)
601 return -EINVAL;
602
603 switch (type) {
604 case IORING_RSRC_FILE:
605 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
606 break;
607 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
608 rr.nr, u64_to_user_ptr(rr.tags));
609 case IORING_RSRC_BUFFER:
610 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
611 break;
612 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
613 rr.nr, u64_to_user_ptr(rr.tags));
614 }
615 return -EINVAL;
616}
617
d9808ceb 618int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
73572984 619{
f2ccb5ae 620 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
73572984
JA		 621
622 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
623 return -EINVAL;
624 if (sqe->rw_flags || sqe->splice_fd_in)
625 return -EINVAL;
626
627 up->offset = READ_ONCE(sqe->off);
628 up->nr_args = READ_ONCE(sqe->len);
629 if (!up->nr_args)
630 return -EINVAL;
631 up->arg = READ_ONCE(sqe->addr);
632 return 0;
633}
634
635static int io_files_update_with_index_alloc(struct io_kiocb *req,
636 unsigned int issue_flags)
637{
f2ccb5ae 638 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
73572984
JA		 639 __s32 __user *fds = u64_to_user_ptr(up->arg);
640 unsigned int done;
641 struct file *file;
642 int ret, fd;
643
644 if (!req->ctx->file_data)
645 return -ENXIO;
646
647 for (done = 0; done < up->nr_args; done++) {
648 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
649 ret = -EFAULT;
650 break;
651 }
652
653 file = fget(fd);
654 if (!file) {
655 ret = -EBADF;
656 break;
657 }
658 ret = io_fixed_fd_install(req, issue_flags, file,
659 IORING_FILE_INDEX_ALLOC);
660 if (ret < 0)
661 break;
662 if (copy_to_user(&fds[done], &ret, sizeof(ret))) {
f110ed84 663 __io_close_fixed(req->ctx, issue_flags, ret);
73572984
JA		 664 ret = -EFAULT;
665 break;
666 }
667 }
668
669 if (done)
670 return done;
671 return ret;
672}
673
d9808ceb 674int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
73572984 675{
f2ccb5ae 676 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
73572984
JA		 677 struct io_ring_ctx *ctx = req->ctx;
678 struct io_uring_rsrc_update2 up2;
679 int ret;
680
681 up2.offset = up->offset;
682 up2.data = up->arg;
683 up2.nr = 0;
684 up2.tags = 0;
685 up2.resv = 0;
686 up2.resv2 = 0;
687
688 if (up->offset == IORING_FILE_INDEX_ALLOC) {
689 ret = io_files_update_with_index_alloc(req, issue_flags);
690 } else {
691 io_ring_submit_lock(ctx, issue_flags);
692 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
693 &up2, up->nr_args);
694 io_ring_submit_unlock(ctx, issue_flags);
695 }
696
697 if (ret < 0)
698 req_set_fail(req);
699 io_req_set_res(req, ret, 0);
700 return IOU_OK;
701}
702
703int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
704 struct io_rsrc_node *node, void *rsrc)
705{
706 u64 *tag_slot = io_get_tag_slot(data, idx);
707 struct io_rsrc_put *prsrc;
ff7c75ec 708 bool inline_item = true;
73572984 709
ff7c75ec
PB		 710 if (!node->inline_items) {
711 prsrc = &node->item;
712 node->inline_items++;
713 } else {
714 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
715 if (!prsrc)
716 return -ENOMEM;
717 inline_item = false;
718 }
73572984
JA		 719
720 prsrc->tag = *tag_slot;
721 *tag_slot = 0;
722 prsrc->rsrc = rsrc;
ff7c75ec
PB		 723 if (!inline_item)
724 list_add(&prsrc->list, &node->item_list);
73572984
JA		 725 return 0;
726}
727
728void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
729{
73572984
JA		 730 int i;
731
732 for (i = 0; i < ctx->nr_user_files; i++) {
733 struct file *file = io_file_from_index(&ctx->file_table, i);
734
38eddb2c
PB		 735 /* skip scm accounted files, they'll be freed by ->ring_sock */
736 if (!file || io_file_need_scm(file))
73572984
JA		 737 continue;
738 io_file_bitmap_clear(&ctx->file_table, i);
739 fput(file);
740 }
73572984
JA		 741
742#if defined(CONFIG_UNIX)
743 if (ctx->ring_sock) {
744 struct sock *sock = ctx->ring_sock->sk;
745 struct sk_buff *skb;
746
747 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
748 kfree_skb(skb);
749 }
750#endif
751 io_free_file_tables(&ctx->file_table);
02a4d923 752 io_file_table_set_alloc_range(ctx, 0, 0);
73572984
JA		 753 io_rsrc_data_free(ctx->file_data);
754 ctx->file_data = NULL;
755 ctx->nr_user_files = 0;
756}
757
758int io_sqe_files_unregister(struct io_ring_ctx *ctx)
759{
760 unsigned nr = ctx->nr_user_files;
761 int ret;
762
763 if (!ctx->file_data)
764 return -ENXIO;
765
766 /*
767 * Quiesce may unlock ->uring_lock, and while it's not held
768 * prevent new requests using the table.
769 */
770 ctx->nr_user_files = 0;
771 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
772 ctx->nr_user_files = nr;
773 if (!ret)
774 __io_sqe_files_unregister(ctx);
775 return ret;
776}
777
778/*
779 * Ensure the UNIX gc is aware of our file set, so we are certain that
780 * the io_uring can be safely unregistered on process exit, even if we have
781 * loops in the file referencing. We account only files that can hold other
782 * files because otherwise they can't form a loop and so are not interesting
783 * for GC.
784 */
785int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file)
786{
787#if defined(CONFIG_UNIX)
788 struct sock *sk = ctx->ring_sock->sk;
789 struct sk_buff_head *head = &sk->sk_receive_queue;
790 struct scm_fp_list *fpl;
791 struct sk_buff *skb;
792
793 if (likely(!io_file_need_scm(file)))
794 return 0;
795
796 /*
797 * See if we can merge this file into an existing skb SCM_RIGHTS
798 * file set. If there's no room, fall back to allocating a new skb
799 * and filling it in.
800 */
801 spin_lock_irq(&head->lock);
802 skb = skb_peek(head);
803 if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD)
804 __skb_unlink(skb, head);
805 else
806 skb = NULL;
807 spin_unlock_irq(&head->lock);
808
809 if (!skb) {
810 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
811 if (!fpl)
812 return -ENOMEM;
813
814 skb = alloc_skb(0, GFP_KERNEL);
815 if (!skb) {
816 kfree(fpl);
817 return -ENOMEM;
818 }
819
820 fpl->user = get_uid(current_user());
821 fpl->max = SCM_MAX_FD;
822 fpl->count = 0;
823
824 UNIXCB(skb).fp = fpl;
825 skb->sk = sk;
0091bfc8 826 skb->scm_io_uring = 1;
73572984
JA		 827 skb->destructor = unix_destruct_scm;
828 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
829 }
830
831 fpl = UNIXCB(skb).fp;
832 fpl->fp[fpl->count++] = get_file(file);
833 unix_inflight(fpl->user, file);
834 skb_queue_head(head, skb);
835 fput(file);
836#endif
837 return 0;
838}
839
840static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
841{
842 struct file *file = prsrc->file;
843#if defined(CONFIG_UNIX)
844 struct sock *sock = ctx->ring_sock->sk;
845 struct sk_buff_head list, *head = &sock->sk_receive_queue;
846 struct sk_buff *skb;
847 int i;
848
849 if (!io_file_need_scm(file)) {
850 fput(file);
851 return;
852 }
853
854 __skb_queue_head_init(&list);
855
856 /*
857 * Find the skb that holds this file in its SCM_RIGHTS. When found,
858 * remove this entry and rearrange the file array.
859 */
860 skb = skb_dequeue(head);
861 while (skb) {
862 struct scm_fp_list *fp;
863
864 fp = UNIXCB(skb).fp;
865 for (i = 0; i < fp->count; i++) {
866 int left;
867
868 if (fp->fp[i] != file)
869 continue;
870
871 unix_notinflight(fp->user, fp->fp[i]);
872 left = fp->count - 1 - i;
873 if (left) {
874 memmove(&fp->fp[i], &fp->fp[i + 1],
875 left * sizeof(struct file *));
876 }
877 fp->count--;
878 if (!fp->count) {
879 kfree_skb(skb);
880 skb = NULL;
881 } else {
882 __skb_queue_tail(&list, skb);
883 }
884 fput(file);
885 file = NULL;
886 break;
887 }
888
889 if (!file)
890 break;
891
892 __skb_queue_tail(&list, skb);
893
894 skb = skb_dequeue(head);
895 }
896
897 if (skb_peek(&list)) {
898 spin_lock_irq(&head->lock);
899 while ((skb = __skb_dequeue(&list)) != NULL)
900 __skb_queue_tail(head, skb);
901 spin_unlock_irq(&head->lock);
902 }
903#else
904 fput(file);
905#endif
906}
907
908int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
909 unsigned nr_args, u64 __user *tags)
910{
911 __s32 __user *fds = (__s32 __user *) arg;
912 struct file *file;
913 int fd, ret;
914 unsigned i;
915
916 if (ctx->file_data)
917 return -EBUSY;
918 if (!nr_args)
919 return -EINVAL;
920 if (nr_args > IORING_MAX_FIXED_FILES)
921 return -EMFILE;
922 if (nr_args > rlimit(RLIMIT_NOFILE))
923 return -EMFILE;
924 ret = io_rsrc_node_switch_start(ctx);
925 if (ret)
926 return ret;
927 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
928 &ctx->file_data);
929 if (ret)
930 return ret;
931
932 if (!io_alloc_file_tables(&ctx->file_table, nr_args)) {
933 io_rsrc_data_free(ctx->file_data);
934 ctx->file_data = NULL;
935 return -ENOMEM;
936 }
937
938 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
939 struct io_fixed_file *file_slot;
940
941 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) {
942 ret = -EFAULT;
943 goto fail;
944 }
945 /* allow sparse sets */
946 if (!fds || fd == -1) {
947 ret = -EINVAL;
948 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
949 goto fail;
950 continue;
951 }
952
953 file = fget(fd);
954 ret = -EBADF;
955 if (unlikely(!file))
956 goto fail;
957
958 /*
959 * Don't allow io_uring instances to be registered. If UNIX
960 * isn't enabled, then this causes a reference cycle and this
961 * instance can never get freed. If UNIX is enabled we'll
962 * handle it just fine, but there's still no point in allowing
963 * a ring fd as it doesn't support regular read/write anyway.
964 */
965 if (io_is_uring_fops(file)) {
966 fput(file);
967 goto fail;
968 }
969 ret = io_scm_file_account(ctx, file);
970 if (ret) {
971 fput(file);
972 goto fail;
973 }
974 file_slot = io_fixed_file_slot(&ctx->file_table, i);
975 io_fixed_file_set(file_slot, file);
976 io_file_bitmap_set(&ctx->file_table, i);
977 }
978
6e73dffb
PB		 979 /* default it to the whole table */
980 io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files);
73572984
JA		 981 io_rsrc_node_switch(ctx, NULL);
982 return 0;
983fail:
984 __io_sqe_files_unregister(ctx);
985 return ret;
986}
987
988static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
989{
990 io_buffer_unmap(ctx, &prsrc->buf);
991 prsrc->buf = NULL;
992}
993
994void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
995{
996 unsigned int i;
997
998 for (i = 0; i < ctx->nr_user_bufs; i++)
999 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
1000 kfree(ctx->user_bufs);
1001 io_rsrc_data_free(ctx->buf_data);
1002 ctx->user_bufs = NULL;
1003 ctx->buf_data = NULL;
1004 ctx->nr_user_bufs = 0;
1005}
1006
1007int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
1008{
1009 unsigned nr = ctx->nr_user_bufs;
1010 int ret;
1011
1012 if (!ctx->buf_data)
1013 return -ENXIO;
1014
1015 /*
1016 * Quiesce may unlock ->uring_lock, and while it's not held
1017 * prevent new requests using the table.
1018 */
1019 ctx->nr_user_bufs = 0;
1020 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
1021 ctx->nr_user_bufs = nr;
1022 if (!ret)
1023 __io_sqe_buffers_unregister(ctx);
1024 return ret;
1025}
1026
1027/*
1028 * Not super efficient, but this is just a registration time. And we do cache
1029 * the last compound head, so generally we'll only do a full search if we don't
1030 * match that one.
1031 *
1032 * We check if the given compound head page has already been accounted, to
1033 * avoid double accounting it. This allows us to account the full size of the
1034 * page, not just the constituent pages of a huge page.
1035 */
1036static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
1037 int nr_pages, struct page *hpage)
1038{
1039 int i, j;
1040
1041 /* check current page array */
1042 for (i = 0; i < nr_pages; i++) {
1043 if (!PageCompound(pages[i]))
1044 continue;
1045 if (compound_head(pages[i]) == hpage)
1046 return true;
1047 }
1048
1049 /* check previously registered pages */
1050 for (i = 0; i < ctx->nr_user_bufs; i++) {
1051 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
1052
1053 for (j = 0; j < imu->nr_bvecs; j++) {
1054 if (!PageCompound(imu->bvec[j].bv_page))
1055 continue;
1056 if (compound_head(imu->bvec[j].bv_page) == hpage)
1057 return true;
1058 }
1059 }
1060
1061 return false;
1062}
1063
1064static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
1065 int nr_pages, struct io_mapped_ubuf *imu,
1066 struct page **last_hpage)
1067{
1068 int i, ret;
1069
1070 imu->acct_pages = 0;
1071 for (i = 0; i < nr_pages; i++) {
1072 if (!PageCompound(pages[i])) {
1073 imu->acct_pages++;
1074 } else {
1075 struct page *hpage;
1076
1077 hpage = compound_head(pages[i]);
1078 if (hpage == *last_hpage)
1079 continue;
1080 *last_hpage = hpage;
1081 if (headpage_already_acct(ctx, pages, i, hpage))
1082 continue;
1083 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
1084 }
1085 }
1086
1087 if (!imu->acct_pages)
1088 return 0;
1089
1090 ret = io_account_mem(ctx, imu->acct_pages);
1091 if (ret)
1092 imu->acct_pages = 0;
1093 return ret;
1094}
1095
1096struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages)
1097{
1098 unsigned long start, end, nr_pages;
1099 struct vm_area_struct **vmas = NULL;
1100 struct page **pages = NULL;
1101 int i, pret, ret = -ENOMEM;
1102
1103 end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1104 start = ubuf >> PAGE_SHIFT;
1105 nr_pages = end - start;
1106
1107 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
1108 if (!pages)
1109 goto done;
1110
1111 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
1112 GFP_KERNEL);
1113 if (!vmas)
1114 goto done;
1115
1116 ret = 0;
1117 mmap_read_lock(current->mm);
1118 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
1119 pages, vmas);
1120 if (pret == nr_pages) {
edd47826
PB		 1121 struct file *file = vmas[0]->vm_file;
1122
73572984
JA		 1123 /* don't support file backed memory */
1124 for (i = 0; i < nr_pages; i++) {
edd47826
PB		 1125 if (vmas[i]->vm_file != file) {
1126 ret = -EINVAL;
1127 break;
1128 }
1129 if (!file)
73572984 1130 continue;
edd47826 1131 if (!vma_is_shmem(vmas[i]) && !is_file_hugepages(file)) {
73572984
JA		 1132 ret = -EOPNOTSUPP;
1133 break;
1134 }
1135 }
1136 *npages = nr_pages;
1137 } else {
1138 ret = pret < 0 ? pret : -EFAULT;
1139 }
1140 mmap_read_unlock(current->mm);
1141 if (ret) {
1142 /*
1143 * if we did partial map, or found file backed vmas,
1144 * release any pages we did get
1145 */
1146 if (pret > 0)
1147 unpin_user_pages(pages, pret);
1148 goto done;
1149 }
1150 ret = 0;
1151done:
1152 kvfree(vmas);
1153 if (ret < 0) {
1154 kvfree(pages);
1155 pages = ERR_PTR(ret);
1156 }
1157 return pages;
1158}
1159
1160static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
1161 struct io_mapped_ubuf **pimu,
1162 struct page **last_hpage)
1163{
1164 struct io_mapped_ubuf *imu = NULL;
1165 struct page **pages = NULL;
1166 unsigned long off;
1167 size_t size;
1168 int ret, nr_pages, i;
977bc873 1169 struct folio *folio = NULL;
73572984 1170
5ff4fdff
PB		 1171 *pimu = ctx->dummy_ubuf;
1172 if (!iov->iov_base)
73572984 1173 return 0;
73572984 1174
73572984 1175 ret = -ENOMEM;
73572984
JA		 1176 pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len,
1177 &nr_pages);
1178 if (IS_ERR(pages)) {
1179 ret = PTR_ERR(pages);
1180 pages = NULL;
1181 goto done;
1182 }
1183
57bebf80
PB		 1184 /* If it's a huge page, try to coalesce them into a single bvec entry */
1185 if (nr_pages > 1) {
1186 folio = page_folio(pages[0]);
1187 for (i = 1; i < nr_pages; i++) {
1188 if (page_folio(pages[i]) != folio) {
1189 folio = NULL;
1190 break;
1191 }
1192 }
1193 if (folio) {
d2acf789
PB		 1194 /*
1195 * The pages are bound to the folio, it doesn't
1196 * actually unpin them but drops all but one reference,
1197 * which is usually put down by io_buffer_unmap().
1198 * Note, needs a better helper.
1199 */
1200 unpin_user_pages(&pages[1], nr_pages - 1);
57bebf80
PB		 1201 nr_pages = 1;
1202 }
1203 }
1204
73572984
JA		 1205 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
1206 if (!imu)
1207 goto done;
1208
1209 ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage);
1210 if (ret) {
1211 unpin_user_pages(pages, nr_pages);
1212 goto done;
1213 }
1214
1215 off = (unsigned long) iov->iov_base & ~PAGE_MASK;
1216 size = iov->iov_len;
57bebf80
PB		 1217 /* store original address for later verification */
1218 imu->ubuf = (unsigned long) iov->iov_base;
1219 imu->ubuf_end = imu->ubuf + iov->iov_len;
1220 imu->nr_bvecs = nr_pages;
1221 *pimu = imu;
1222 ret = 0;
1223
1224 if (folio) {
1225 bvec_set_page(&imu->bvec[0], pages[0], size, off);
1226 goto done;
1227 }
73572984
JA		 1228 for (i = 0; i < nr_pages; i++) {
1229 size_t vec_len;
1230
1231 vec_len = min_t(size_t, size, PAGE_SIZE - off);
cc342a21 1232 bvec_set_page(&imu->bvec[i], pages[i], vec_len, off);
73572984
JA		 1233 off = 0;
1234 size -= vec_len;
1235 }
73572984
JA		 1236done:
1237 if (ret)
1238 kvfree(imu);
1239 kvfree(pages);
1240 return ret;
1241}
1242
1243static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
1244{
1245 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
1246 return ctx->user_bufs ? 0 : -ENOMEM;
1247}
1248
1249int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
1250 unsigned int nr_args, u64 __user *tags)
1251{
1252 struct page *last_hpage = NULL;
1253 struct io_rsrc_data *data;
1254 int i, ret;
1255 struct iovec iov;
1256
1257 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
1258
1259 if (ctx->user_bufs)
1260 return -EBUSY;
1261 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
1262 return -EINVAL;
1263 ret = io_rsrc_node_switch_start(ctx);
1264 if (ret)
1265 return ret;
1266 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
1267 if (ret)
1268 return ret;
1269 ret = io_buffers_map_alloc(ctx, nr_args);
1270 if (ret) {
1271 io_rsrc_data_free(data);
1272 return ret;
1273 }
1274
1275 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
1276 if (arg) {
1277 ret = io_copy_iov(ctx, &iov, arg, i);
1278 if (ret)
1279 break;
1280 ret = io_buffer_validate(&iov);
1281 if (ret)
1282 break;
1283 } else {
1284 memset(&iov, 0, sizeof(iov));
1285 }
1286
1287 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
1288 ret = -EINVAL;
1289 break;
1290 }
1291
1292 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
1293 &last_hpage);
1294 if (ret)
1295 break;
1296 }
1297
1298 WARN_ON_ONCE(ctx->buf_data);
1299
1300 ctx->buf_data = data;
1301 if (ret)
1302 __io_sqe_buffers_unregister(ctx);
1303 else
1304 io_rsrc_node_switch(ctx, NULL);
1305 return ret;
1306}
c059f785
PB		 1307
1308int io_import_fixed(int ddir, struct iov_iter *iter,
1309 struct io_mapped_ubuf *imu,
1310 u64 buf_addr, size_t len)
1311{
1312 u64 buf_end;
1313 size_t offset;
1314
1315 if (WARN_ON_ONCE(!imu))
1316 return -EFAULT;
1317 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
1318 return -EFAULT;
1319 /* not inside the mapped region */
1320 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
1321 return -EFAULT;
1322
1323 /*
6bf65a1b 1324 * Might not be a start of buffer, set size appropriately
c059f785
PB		 1325 * and advance us to the beginning.
1326 */
1327 offset = buf_addr - imu->ubuf;
1328 iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len);
1329
1330 if (offset) {
1331 /*
1332 * Don't use iov_iter_advance() here, as it's really slow for
1333 * using the latter parts of a big fixed buffer - it iterates
1334 * over each segment manually. We can cheat a bit here, because
1335 * we know that:
1336 *
1337 * 1) it's a BVEC iter, we set it up
1338 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1339 * first and last bvec
1340 *
1341 * So just find our index, and adjust the iterator afterwards.
1342 * If the offset is within the first bvec (or the whole first
1343 * bvec, just use iov_iter_advance(). This makes it easier
1344 * since we can just skip the first segment, which may not
1345 * be PAGE_SIZE aligned.
1346 */
1347 const struct bio_vec *bvec = imu->bvec;
1348
1349 if (offset <= bvec->bv_len) {
57bebf80
PB		 1350 /*
1351 * Note, huge pages buffers consists of one large
1352 * bvec entry and should always go this way. The other
1353 * branch doesn't expect non PAGE_SIZE'd chunks.
1354 */
b000ae0e
PB		 1355 iter->bvec = bvec;
1356 iter->nr_segs = bvec->bv_len;
1357 iter->count -= offset;
1358 iter->iov_offset = offset;
c059f785
PB		 1359 } else {
1360 unsigned long seg_skip;
1361
1362 /* skip first vec */
1363 offset -= bvec->bv_len;
1364 seg_skip = 1 + (offset >> PAGE_SHIFT);
1365
1366 iter->bvec = bvec + seg_skip;
1367 iter->nr_segs -= seg_skip;
1368 iter->count -= bvec->bv_len + offset;
1369 iter->iov_offset = offset & ~PAGE_MASK;
1370 }
1371 }
1372
1373 return 0;
1374}
Linux 6.x block layer and io_uring tree(s)