projects / linux-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
io_uring: split out fixed file installation and removal
[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
30#define IO_RSRC_REF_BATCH 100
31
32/* only define max */
33#define IORING_MAX_FIXED_FILES (1U << 20)
34#define IORING_MAX_REG_BUFFERS (1U << 14)
35
36void io_rsrc_refs_drop(struct io_ring_ctx *ctx)
37 __must_hold(&ctx->uring_lock)
38{
39 if (ctx->rsrc_cached_refs) {
40 io_rsrc_put_node(ctx->rsrc_node, ctx->rsrc_cached_refs);
41 ctx->rsrc_cached_refs = 0;
42 }
43}
44
45static inline void __io_unaccount_mem(struct user_struct *user,
46 unsigned long nr_pages)
47{
48 atomic_long_sub(nr_pages, &user->locked_vm);
49}
50
51static inline int __io_account_mem(struct user_struct *user,
52 unsigned long nr_pages)
53{
54 unsigned long page_limit, cur_pages, new_pages;
55
56 /* Don't allow more pages than we can safely lock */
57 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
58
59 do {
60 cur_pages = atomic_long_read(&user->locked_vm);
61 new_pages = cur_pages + nr_pages;
62 if (new_pages > page_limit)
63 return -ENOMEM;
64 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
65 new_pages) != cur_pages);
66
67 return 0;
68}
69
70static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
71{
72 if (ctx->user)
73 __io_unaccount_mem(ctx->user, nr_pages);
74
75 if (ctx->mm_account)
76 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
77}
78
79static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
80{
81 int ret;
82
83 if (ctx->user) {
84 ret = __io_account_mem(ctx->user, nr_pages);
85 if (ret)
86 return ret;
87 }
88
89 if (ctx->mm_account)
90 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
91
92 return 0;
93}
94
95static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
96 void __user *arg, unsigned index)
97{
98 struct iovec __user *src;
99
100#ifdef CONFIG_COMPAT
101 if (ctx->compat) {
102 struct compat_iovec __user *ciovs;
103 struct compat_iovec ciov;
104
105 ciovs = (struct compat_iovec __user *) arg;
106 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
107 return -EFAULT;
108
109 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
110 dst->iov_len = ciov.iov_len;
111 return 0;
112 }
113#endif
114 src = (struct iovec __user *) arg;
115 if (copy_from_user(dst, &src[index], sizeof(*dst)))
116 return -EFAULT;
117 return 0;
118}
119
120static int io_buffer_validate(struct iovec *iov)
121{
122 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
123
124 /*
125 * Don't impose further limits on the size and buffer
126 * constraints here, we'll -EINVAL later when IO is
127 * submitted if they are wrong.
128 */
129 if (!iov->iov_base)
130 return iov->iov_len ? -EFAULT : 0;
131 if (!iov->iov_len)
132 return -EFAULT;
133
134 /* arbitrary limit, but we need something */
135 if (iov->iov_len > SZ_1G)
136 return -EFAULT;
137
138 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
139 return -EOVERFLOW;
140
141 return 0;
142}
143
144static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
145{
146 struct io_mapped_ubuf *imu = *slot;
147 unsigned int i;
148
149 if (imu != ctx->dummy_ubuf) {
150 for (i = 0; i < imu->nr_bvecs; i++)
151 unpin_user_page(imu->bvec[i].bv_page);
152 if (imu->acct_pages)
153 io_unaccount_mem(ctx, imu->acct_pages);
154 kvfree(imu);
155 }
156 *slot = NULL;
157}
158
159void io_rsrc_refs_refill(struct io_ring_ctx *ctx)
160 __must_hold(&ctx->uring_lock)
161{
162 ctx->rsrc_cached_refs += IO_RSRC_REF_BATCH;
163 percpu_ref_get_many(&ctx->rsrc_node->refs, IO_RSRC_REF_BATCH);
164}
165
166static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
167{
168 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
169 struct io_ring_ctx *ctx = rsrc_data->ctx;
170 struct io_rsrc_put *prsrc, *tmp;
171
172 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
173 list_del(&prsrc->list);
174
175 if (prsrc->tag) {
d245bca6 176 if (ctx->flags & IORING_SETUP_IOPOLL) {
73572984 177 mutex_lock(&ctx->uring_lock);
d245bca6 178 io_post_aux_cqe(ctx, prsrc->tag, 0, 0);
73572984 179 mutex_unlock(&ctx->uring_lock);
d245bca6
PB		 180 } else {
181 io_post_aux_cqe(ctx, prsrc->tag, 0, 0);
182 }
73572984
JA		 183 }
184
185 rsrc_data->do_put(ctx, prsrc);
186 kfree(prsrc);
187 }
188
189 io_rsrc_node_destroy(ref_node);
190 if (atomic_dec_and_test(&rsrc_data->refs))
191 complete(&rsrc_data->done);
192}
193
194void io_rsrc_put_work(struct work_struct *work)
195{
196 struct io_ring_ctx *ctx;
197 struct llist_node *node;
198
199 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
200 node = llist_del_all(&ctx->rsrc_put_llist);
201
202 while (node) {
203 struct io_rsrc_node *ref_node;
204 struct llist_node *next = node->next;
205
206 ref_node = llist_entry(node, struct io_rsrc_node, llist);
207 __io_rsrc_put_work(ref_node);
208 node = next;
209 }
210}
211
212void io_wait_rsrc_data(struct io_rsrc_data *data)
213{
214 if (data && !atomic_dec_and_test(&data->refs))
215 wait_for_completion(&data->done);
216}
217
218void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
219{
220 percpu_ref_exit(&ref_node->refs);
221 kfree(ref_node);
222}
223
224static __cold void io_rsrc_node_ref_zero(struct percpu_ref *ref)
225{
226 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
227 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
228 unsigned long flags;
229 bool first_add = false;
230 unsigned long delay = HZ;
231
232 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
233 node->done = true;
234
235 /* if we are mid-quiesce then do not delay */
236 if (node->rsrc_data->quiesce)
237 delay = 0;
238
239 while (!list_empty(&ctx->rsrc_ref_list)) {
240 node = list_first_entry(&ctx->rsrc_ref_list,
241 struct io_rsrc_node, node);
242 /* recycle ref nodes in order */
243 if (!node->done)
244 break;
245 list_del(&node->node);
246 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
247 }
248 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
249
250 if (first_add)
251 mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
252}
253
254static struct io_rsrc_node *io_rsrc_node_alloc(void)
255{
256 struct io_rsrc_node *ref_node;
257
258 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
259 if (!ref_node)
260 return NULL;
261
262 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
263 0, GFP_KERNEL)) {
264 kfree(ref_node);
265 return NULL;
266 }
267 INIT_LIST_HEAD(&ref_node->node);
268 INIT_LIST_HEAD(&ref_node->rsrc_list);
269 ref_node->done = false;
270 return ref_node;
271}
272
273void io_rsrc_node_switch(struct io_ring_ctx *ctx,
274 struct io_rsrc_data *data_to_kill)
275 __must_hold(&ctx->uring_lock)
276{
277 WARN_ON_ONCE(!ctx->rsrc_backup_node);
278 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
279
280 io_rsrc_refs_drop(ctx);
281
282 if (data_to_kill) {
283 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
284
285 rsrc_node->rsrc_data = data_to_kill;
286 spin_lock_irq(&ctx->rsrc_ref_lock);
287 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
288 spin_unlock_irq(&ctx->rsrc_ref_lock);
289
290 atomic_inc(&data_to_kill->refs);
291 percpu_ref_kill(&rsrc_node->refs);
292 ctx->rsrc_node = NULL;
293 }
294
295 if (!ctx->rsrc_node) {
296 ctx->rsrc_node = ctx->rsrc_backup_node;
297 ctx->rsrc_backup_node = NULL;
298 }
299}
300
301int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
302{
303 if (ctx->rsrc_backup_node)
304 return 0;
305 ctx->rsrc_backup_node = io_rsrc_node_alloc();
306 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
307}
308
309__cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data,
310 struct io_ring_ctx *ctx)
311{
312 int ret;
313
314 /* As we may drop ->uring_lock, other task may have started quiesce */
315 if (data->quiesce)
316 return -ENXIO;
317
318 data->quiesce = true;
319 do {
320 ret = io_rsrc_node_switch_start(ctx);
321 if (ret)
322 break;
323 io_rsrc_node_switch(ctx, data);
324
325 /* kill initial ref, already quiesced if zero */
326 if (atomic_dec_and_test(&data->refs))
327 break;
328 mutex_unlock(&ctx->uring_lock);
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);
333 if (atomic_read(&data->refs) > 0) {
334 /*
335 * it has been revived by another thread while
336 * we were unlocked
337 */
338 mutex_unlock(&ctx->uring_lock);
339 } else {
340 break;
341 }
342 }
343
344 atomic_inc(&data->refs);
345 /* wait for all works potentially completing data->done */
346 flush_delayed_work(&ctx->rsrc_put_work);
347 reinit_completion(&data->done);
348
349 ret = io_run_task_work_sig();
350 mutex_lock(&ctx->uring_lock);
351 } while (ret >= 0);
352 data->quiesce = false;
353
354 return ret;
355}
356
357static void io_free_page_table(void **table, size_t size)
358{
359 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
360
361 for (i = 0; i < nr_tables; i++)
362 kfree(table[i]);
363 kfree(table);
364}
365
366static void io_rsrc_data_free(struct io_rsrc_data *data)
367{
368 size_t size = data->nr * sizeof(data->tags[0][0]);
369
370 if (data->tags)
371 io_free_page_table((void **)data->tags, size);
372 kfree(data);
373}
374
375static __cold void **io_alloc_page_table(size_t size)
376{
377 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
378 size_t init_size = size;
379 void **table;
380
381 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
382 if (!table)
383 return NULL;
384
385 for (i = 0; i < nr_tables; i++) {
386 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
387
388 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
389 if (!table[i]) {
390 io_free_page_table(table, init_size);
391 return NULL;
392 }
393 size -= this_size;
394 }
395 return table;
396}
397
398__cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx,
399 rsrc_put_fn *do_put, u64 __user *utags,
400 unsigned nr, struct io_rsrc_data **pdata)
401{
402 struct io_rsrc_data *data;
403 int ret = -ENOMEM;
404 unsigned i;
405
406 data = kzalloc(sizeof(*data), GFP_KERNEL);
407 if (!data)
408 return -ENOMEM;
409 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
410 if (!data->tags) {
411 kfree(data);
412 return -ENOMEM;
413 }
414
415 data->nr = nr;
416 data->ctx = ctx;
417 data->do_put = do_put;
418 if (utags) {
419 ret = -EFAULT;
420 for (i = 0; i < nr; i++) {
421 u64 *tag_slot = io_get_tag_slot(data, i);
422
423 if (copy_from_user(tag_slot, &utags[i],
424 sizeof(*tag_slot)))
425 goto fail;
426 }
427 }
428
429 atomic_set(&data->refs, 1);
430 init_completion(&data->done);
431 *pdata = data;
432 return 0;
433fail:
434 io_rsrc_data_free(data);
435 return ret;
436}
437
438static int __io_sqe_files_update(struct io_ring_ctx *ctx,
439 struct io_uring_rsrc_update2 *up,
440 unsigned nr_args)
441{
442 u64 __user *tags = u64_to_user_ptr(up->tags);
443 __s32 __user *fds = u64_to_user_ptr(up->data);
444 struct io_rsrc_data *data = ctx->file_data;
445 struct io_fixed_file *file_slot;
446 struct file *file;
447 int fd, i, err = 0;
448 unsigned int done;
449 bool needs_switch = false;
450
451 if (!ctx->file_data)
452 return -ENXIO;
453 if (up->offset + nr_args > ctx->nr_user_files)
454 return -EINVAL;
455
456 for (done = 0; done < nr_args; done++) {
457 u64 tag = 0;
458
459 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
460 copy_from_user(&fd, &fds[done], sizeof(fd))) {
461 err = -EFAULT;
462 break;
463 }
464 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
465 err = -EINVAL;
466 break;
467 }
468 if (fd == IORING_REGISTER_FILES_SKIP)
469 continue;
470
471 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
472 file_slot = io_fixed_file_slot(&ctx->file_table, i);
473
474 if (file_slot->file_ptr) {
475 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
476 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file);
477 if (err)
478 break;
479 file_slot->file_ptr = 0;
480 io_file_bitmap_clear(&ctx->file_table, i);
481 needs_switch = true;
482 }
483 if (fd != -1) {
484 file = fget(fd);
485 if (!file) {
486 err = -EBADF;
487 break;
488 }
489 /*
490 * Don't allow io_uring instances to be registered. If
491 * UNIX isn't enabled, then this causes a reference
492 * cycle and this instance can never get freed. If UNIX
493 * is enabled we'll handle it just fine, but there's
494 * still no point in allowing a ring fd as it doesn't
495 * support regular read/write anyway.
496 */
497 if (io_is_uring_fops(file)) {
498 fput(file);
499 err = -EBADF;
500 break;
501 }
502 err = io_scm_file_account(ctx, file);
503 if (err) {
504 fput(file);
505 break;
506 }
507 *io_get_tag_slot(data, i) = tag;
508 io_fixed_file_set(file_slot, file);
509 io_file_bitmap_set(&ctx->file_table, i);
510 }
511 }
512
513 if (needs_switch)
514 io_rsrc_node_switch(ctx, data);
515 return done ? done : err;
516}
517
518static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
519 struct io_uring_rsrc_update2 *up,
520 unsigned int nr_args)
521{
522 u64 __user *tags = u64_to_user_ptr(up->tags);
523 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
524 struct page *last_hpage = NULL;
525 bool needs_switch = false;
526 __u32 done;
527 int i, err;
528
529 if (!ctx->buf_data)
530 return -ENXIO;
531 if (up->offset + nr_args > ctx->nr_user_bufs)
532 return -EINVAL;
533
534 for (done = 0; done < nr_args; done++) {
535 struct io_mapped_ubuf *imu;
536 int offset = up->offset + done;
537 u64 tag = 0;
538
539 err = io_copy_iov(ctx, &iov, iovs, done);
540 if (err)
541 break;
542 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
543 err = -EFAULT;
544 break;
545 }
546 err = io_buffer_validate(&iov);
547 if (err)
548 break;
549 if (!iov.iov_base && tag) {
550 err = -EINVAL;
551 break;
552 }
553 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
554 if (err)
555 break;
556
557 i = array_index_nospec(offset, ctx->nr_user_bufs);
558 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
559 err = io_queue_rsrc_removal(ctx->buf_data, i,
560 ctx->rsrc_node, ctx->user_bufs[i]);
561 if (unlikely(err)) {
562 io_buffer_unmap(ctx, &imu);
563 break;
564 }
5ff4fdff 565 ctx->user_bufs[i] = ctx->dummy_ubuf;
73572984
JA		 566 needs_switch = true;
567 }
568
569 ctx->user_bufs[i] = imu;
570 *io_get_tag_slot(ctx->buf_data, offset) = tag;
571 }
572
573 if (needs_switch)
574 io_rsrc_node_switch(ctx, ctx->buf_data);
575 return done ? done : err;
576}
577
578static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
579 struct io_uring_rsrc_update2 *up,
580 unsigned nr_args)
581{
582 __u32 tmp;
583 int err;
584
585 if (check_add_overflow(up->offset, nr_args, &tmp))
586 return -EOVERFLOW;
587 err = io_rsrc_node_switch_start(ctx);
588 if (err)
589 return err;
590
591 switch (type) {
592 case IORING_RSRC_FILE:
593 return __io_sqe_files_update(ctx, up, nr_args);
594 case IORING_RSRC_BUFFER:
595 return __io_sqe_buffers_update(ctx, up, nr_args);
596 }
597 return -EINVAL;
598}
599
600int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
601 unsigned nr_args)
602{
603 struct io_uring_rsrc_update2 up;
604
605 if (!nr_args)
606 return -EINVAL;
607 memset(&up, 0, sizeof(up));
608 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
609 return -EFAULT;
610 if (up.resv || up.resv2)
611 return -EINVAL;
612 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
613}
614
615int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
616 unsigned size, unsigned type)
617{
618 struct io_uring_rsrc_update2 up;
619
620 if (size != sizeof(up))
621 return -EINVAL;
622 if (copy_from_user(&up, arg, sizeof(up)))
623 return -EFAULT;
624 if (!up.nr || up.resv || up.resv2)
625 return -EINVAL;
626 return __io_register_rsrc_update(ctx, type, &up, up.nr);
627}
628
629__cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
630 unsigned int size, unsigned int type)
631{
632 struct io_uring_rsrc_register rr;
633
634 /* keep it extendible */
635 if (size != sizeof(rr))
636 return -EINVAL;
637
638 memset(&rr, 0, sizeof(rr));
639 if (copy_from_user(&rr, arg, size))
640 return -EFAULT;
641 if (!rr.nr || rr.resv2)
642 return -EINVAL;
643 if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE)
644 return -EINVAL;
645
646 switch (type) {
647 case IORING_RSRC_FILE:
648 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
649 break;
650 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
651 rr.nr, u64_to_user_ptr(rr.tags));
652 case IORING_RSRC_BUFFER:
653 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
654 break;
655 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
656 rr.nr, u64_to_user_ptr(rr.tags));
657 }
658 return -EINVAL;
659}
660
661int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
662{
663 struct io_rsrc_update *up = io_kiocb_to_cmd(req);
664
665 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
666 return -EINVAL;
667 if (sqe->rw_flags || sqe->splice_fd_in)
668 return -EINVAL;
669
670 up->offset = READ_ONCE(sqe->off);
671 up->nr_args = READ_ONCE(sqe->len);
672 if (!up->nr_args)
673 return -EINVAL;
674 up->arg = READ_ONCE(sqe->addr);
675 return 0;
676}
677
678static int io_files_update_with_index_alloc(struct io_kiocb *req,
679 unsigned int issue_flags)
680{
681 struct io_rsrc_update *up = io_kiocb_to_cmd(req);
682 __s32 __user *fds = u64_to_user_ptr(up->arg);
683 unsigned int done;
684 struct file *file;
685 int ret, fd;
686
687 if (!req->ctx->file_data)
688 return -ENXIO;
689
690 for (done = 0; done < up->nr_args; done++) {
691 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
692 ret = -EFAULT;
693 break;
694 }
695
696 file = fget(fd);
697 if (!file) {
698 ret = -EBADF;
699 break;
700 }
701 ret = io_fixed_fd_install(req, issue_flags, file,
702 IORING_FILE_INDEX_ALLOC);
703 if (ret < 0)
704 break;
705 if (copy_to_user(&fds[done], &ret, sizeof(ret))) {
f110ed84 706 __io_close_fixed(req->ctx, issue_flags, ret);
73572984
JA		 707 ret = -EFAULT;
708 break;
709 }
710 }
711
712 if (done)
713 return done;
714 return ret;
715}
716
717int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
718{
719 struct io_rsrc_update *up = io_kiocb_to_cmd(req);
720 struct io_ring_ctx *ctx = req->ctx;
721 struct io_uring_rsrc_update2 up2;
722 int ret;
723
724 up2.offset = up->offset;
725 up2.data = up->arg;
726 up2.nr = 0;
727 up2.tags = 0;
728 up2.resv = 0;
729 up2.resv2 = 0;
730
731 if (up->offset == IORING_FILE_INDEX_ALLOC) {
732 ret = io_files_update_with_index_alloc(req, issue_flags);
733 } else {
734 io_ring_submit_lock(ctx, issue_flags);
735 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
736 &up2, up->nr_args);
737 io_ring_submit_unlock(ctx, issue_flags);
738 }
739
740 if (ret < 0)
741 req_set_fail(req);
742 io_req_set_res(req, ret, 0);
743 return IOU_OK;
744}
745
746int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
747 struct io_rsrc_node *node, void *rsrc)
748{
749 u64 *tag_slot = io_get_tag_slot(data, idx);
750 struct io_rsrc_put *prsrc;
751
752 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
753 if (!prsrc)
754 return -ENOMEM;
755
756 prsrc->tag = *tag_slot;
757 *tag_slot = 0;
758 prsrc->rsrc = rsrc;
759 list_add(&prsrc->list, &node->rsrc_list);
760 return 0;
761}
762
763void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
764{
765#if !defined(IO_URING_SCM_ALL)
766 int i;
767
768 for (i = 0; i < ctx->nr_user_files; i++) {
769 struct file *file = io_file_from_index(&ctx->file_table, i);
770
771 if (!file)
772 continue;
773 if (io_fixed_file_slot(&ctx->file_table, i)->file_ptr & FFS_SCM)
774 continue;
775 io_file_bitmap_clear(&ctx->file_table, i);
776 fput(file);
777 }
778#endif
779
780#if defined(CONFIG_UNIX)
781 if (ctx->ring_sock) {
782 struct sock *sock = ctx->ring_sock->sk;
783 struct sk_buff *skb;
784
785 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
786 kfree_skb(skb);
787 }
788#endif
789 io_free_file_tables(&ctx->file_table);
790 io_rsrc_data_free(ctx->file_data);
791 ctx->file_data = NULL;
792 ctx->nr_user_files = 0;
793}
794
795int io_sqe_files_unregister(struct io_ring_ctx *ctx)
796{
797 unsigned nr = ctx->nr_user_files;
798 int ret;
799
800 if (!ctx->file_data)
801 return -ENXIO;
802
803 /*
804 * Quiesce may unlock ->uring_lock, and while it's not held
805 * prevent new requests using the table.
806 */
807 ctx->nr_user_files = 0;
808 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
809 ctx->nr_user_files = nr;
810 if (!ret)
811 __io_sqe_files_unregister(ctx);
812 return ret;
813}
814
815/*
816 * Ensure the UNIX gc is aware of our file set, so we are certain that
817 * the io_uring can be safely unregistered on process exit, even if we have
818 * loops in the file referencing. We account only files that can hold other
819 * files because otherwise they can't form a loop and so are not interesting
820 * for GC.
821 */
822int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file)
823{
824#if defined(CONFIG_UNIX)
825 struct sock *sk = ctx->ring_sock->sk;
826 struct sk_buff_head *head = &sk->sk_receive_queue;
827 struct scm_fp_list *fpl;
828 struct sk_buff *skb;
829
830 if (likely(!io_file_need_scm(file)))
831 return 0;
832
833 /*
834 * See if we can merge this file into an existing skb SCM_RIGHTS
835 * file set. If there's no room, fall back to allocating a new skb
836 * and filling it in.
837 */
838 spin_lock_irq(&head->lock);
839 skb = skb_peek(head);
840 if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD)
841 __skb_unlink(skb, head);
842 else
843 skb = NULL;
844 spin_unlock_irq(&head->lock);
845
846 if (!skb) {
847 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
848 if (!fpl)
849 return -ENOMEM;
850
851 skb = alloc_skb(0, GFP_KERNEL);
852 if (!skb) {
853 kfree(fpl);
854 return -ENOMEM;
855 }
856
857 fpl->user = get_uid(current_user());
858 fpl->max = SCM_MAX_FD;
859 fpl->count = 0;
860
861 UNIXCB(skb).fp = fpl;
862 skb->sk = sk;
863 skb->destructor = unix_destruct_scm;
864 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
865 }
866
867 fpl = UNIXCB(skb).fp;
868 fpl->fp[fpl->count++] = get_file(file);
869 unix_inflight(fpl->user, file);
870 skb_queue_head(head, skb);
871 fput(file);
872#endif
873 return 0;
874}
875
876static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
877{
878 struct file *file = prsrc->file;
879#if defined(CONFIG_UNIX)
880 struct sock *sock = ctx->ring_sock->sk;
881 struct sk_buff_head list, *head = &sock->sk_receive_queue;
882 struct sk_buff *skb;
883 int i;
884
885 if (!io_file_need_scm(file)) {
886 fput(file);
887 return;
888 }
889
890 __skb_queue_head_init(&list);
891
892 /*
893 * Find the skb that holds this file in its SCM_RIGHTS. When found,
894 * remove this entry and rearrange the file array.
895 */
896 skb = skb_dequeue(head);
897 while (skb) {
898 struct scm_fp_list *fp;
899
900 fp = UNIXCB(skb).fp;
901 for (i = 0; i < fp->count; i++) {
902 int left;
903
904 if (fp->fp[i] != file)
905 continue;
906
907 unix_notinflight(fp->user, fp->fp[i]);
908 left = fp->count - 1 - i;
909 if (left) {
910 memmove(&fp->fp[i], &fp->fp[i + 1],
911 left * sizeof(struct file *));
912 }
913 fp->count--;
914 if (!fp->count) {
915 kfree_skb(skb);
916 skb = NULL;
917 } else {
918 __skb_queue_tail(&list, skb);
919 }
920 fput(file);
921 file = NULL;
922 break;
923 }
924
925 if (!file)
926 break;
927
928 __skb_queue_tail(&list, skb);
929
930 skb = skb_dequeue(head);
931 }
932
933 if (skb_peek(&list)) {
934 spin_lock_irq(&head->lock);
935 while ((skb = __skb_dequeue(&list)) != NULL)
936 __skb_queue_tail(head, skb);
937 spin_unlock_irq(&head->lock);
938 }
939#else
940 fput(file);
941#endif
942}
943
944int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
945 unsigned nr_args, u64 __user *tags)
946{
947 __s32 __user *fds = (__s32 __user *) arg;
948 struct file *file;
949 int fd, ret;
950 unsigned i;
951
952 if (ctx->file_data)
953 return -EBUSY;
954 if (!nr_args)
955 return -EINVAL;
956 if (nr_args > IORING_MAX_FIXED_FILES)
957 return -EMFILE;
958 if (nr_args > rlimit(RLIMIT_NOFILE))
959 return -EMFILE;
960 ret = io_rsrc_node_switch_start(ctx);
961 if (ret)
962 return ret;
963 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
964 &ctx->file_data);
965 if (ret)
966 return ret;
967
968 if (!io_alloc_file_tables(&ctx->file_table, nr_args)) {
969 io_rsrc_data_free(ctx->file_data);
970 ctx->file_data = NULL;
971 return -ENOMEM;
972 }
973
974 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
975 struct io_fixed_file *file_slot;
976
977 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) {
978 ret = -EFAULT;
979 goto fail;
980 }
981 /* allow sparse sets */
982 if (!fds || fd == -1) {
983 ret = -EINVAL;
984 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
985 goto fail;
986 continue;
987 }
988
989 file = fget(fd);
990 ret = -EBADF;
991 if (unlikely(!file))
992 goto fail;
993
994 /*
995 * Don't allow io_uring instances to be registered. If UNIX
996 * isn't enabled, then this causes a reference cycle and this
997 * instance can never get freed. If UNIX is enabled we'll
998 * handle it just fine, but there's still no point in allowing
999 * a ring fd as it doesn't support regular read/write anyway.
1000 */
1001 if (io_is_uring_fops(file)) {
1002 fput(file);
1003 goto fail;
1004 }
1005 ret = io_scm_file_account(ctx, file);
1006 if (ret) {
1007 fput(file);
1008 goto fail;
1009 }
1010 file_slot = io_fixed_file_slot(&ctx->file_table, i);
1011 io_fixed_file_set(file_slot, file);
1012 io_file_bitmap_set(&ctx->file_table, i);
1013 }
1014
1015 io_rsrc_node_switch(ctx, NULL);
1016 return 0;
1017fail:
1018 __io_sqe_files_unregister(ctx);
1019 return ret;
1020}
1021
1022static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
1023{
1024 io_buffer_unmap(ctx, &prsrc->buf);
1025 prsrc->buf = NULL;
1026}
1027
1028void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
1029{
1030 unsigned int i;
1031
1032 for (i = 0; i < ctx->nr_user_bufs; i++)
1033 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
1034 kfree(ctx->user_bufs);
1035 io_rsrc_data_free(ctx->buf_data);
1036 ctx->user_bufs = NULL;
1037 ctx->buf_data = NULL;
1038 ctx->nr_user_bufs = 0;
1039}
1040
1041int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
1042{
1043 unsigned nr = ctx->nr_user_bufs;
1044 int ret;
1045
1046 if (!ctx->buf_data)
1047 return -ENXIO;
1048
1049 /*
1050 * Quiesce may unlock ->uring_lock, and while it's not held
1051 * prevent new requests using the table.
1052 */
1053 ctx->nr_user_bufs = 0;
1054 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
1055 ctx->nr_user_bufs = nr;
1056 if (!ret)
1057 __io_sqe_buffers_unregister(ctx);
1058 return ret;
1059}
1060
1061/*
1062 * Not super efficient, but this is just a registration time. And we do cache
1063 * the last compound head, so generally we'll only do a full search if we don't
1064 * match that one.
1065 *
1066 * We check if the given compound head page has already been accounted, to
1067 * avoid double accounting it. This allows us to account the full size of the
1068 * page, not just the constituent pages of a huge page.
1069 */
1070static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
1071 int nr_pages, struct page *hpage)
1072{
1073 int i, j;
1074
1075 /* check current page array */
1076 for (i = 0; i < nr_pages; i++) {
1077 if (!PageCompound(pages[i]))
1078 continue;
1079 if (compound_head(pages[i]) == hpage)
1080 return true;
1081 }
1082
1083 /* check previously registered pages */
1084 for (i = 0; i < ctx->nr_user_bufs; i++) {
1085 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
1086
1087 for (j = 0; j < imu->nr_bvecs; j++) {
1088 if (!PageCompound(imu->bvec[j].bv_page))
1089 continue;
1090 if (compound_head(imu->bvec[j].bv_page) == hpage)
1091 return true;
1092 }
1093 }
1094
1095 return false;
1096}
1097
1098static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
1099 int nr_pages, struct io_mapped_ubuf *imu,
1100 struct page **last_hpage)
1101{
1102 int i, ret;
1103
1104 imu->acct_pages = 0;
1105 for (i = 0; i < nr_pages; i++) {
1106 if (!PageCompound(pages[i])) {
1107 imu->acct_pages++;
1108 } else {
1109 struct page *hpage;
1110
1111 hpage = compound_head(pages[i]);
1112 if (hpage == *last_hpage)
1113 continue;
1114 *last_hpage = hpage;
1115 if (headpage_already_acct(ctx, pages, i, hpage))
1116 continue;
1117 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
1118 }
1119 }
1120
1121 if (!imu->acct_pages)
1122 return 0;
1123
1124 ret = io_account_mem(ctx, imu->acct_pages);
1125 if (ret)
1126 imu->acct_pages = 0;
1127 return ret;
1128}
1129
1130struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages)
1131{
1132 unsigned long start, end, nr_pages;
1133 struct vm_area_struct **vmas = NULL;
1134 struct page **pages = NULL;
1135 int i, pret, ret = -ENOMEM;
1136
1137 end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1138 start = ubuf >> PAGE_SHIFT;
1139 nr_pages = end - start;
1140
1141 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
1142 if (!pages)
1143 goto done;
1144
1145 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
1146 GFP_KERNEL);
1147 if (!vmas)
1148 goto done;
1149
1150 ret = 0;
1151 mmap_read_lock(current->mm);
1152 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
1153 pages, vmas);
1154 if (pret == nr_pages) {
1155 /* don't support file backed memory */
1156 for (i = 0; i < nr_pages; i++) {
1157 struct vm_area_struct *vma = vmas[i];
1158
1159 if (vma_is_shmem(vma))
1160 continue;
1161 if (vma->vm_file &&
1162 !is_file_hugepages(vma->vm_file)) {
1163 ret = -EOPNOTSUPP;
1164 break;
1165 }
1166 }
1167 *npages = nr_pages;
1168 } else {
1169 ret = pret < 0 ? pret : -EFAULT;
1170 }
1171 mmap_read_unlock(current->mm);
1172 if (ret) {
1173 /*
1174 * if we did partial map, or found file backed vmas,
1175 * release any pages we did get
1176 */
1177 if (pret > 0)
1178 unpin_user_pages(pages, pret);
1179 goto done;
1180 }
1181 ret = 0;
1182done:
1183 kvfree(vmas);
1184 if (ret < 0) {
1185 kvfree(pages);
1186 pages = ERR_PTR(ret);
1187 }
1188 return pages;
1189}
1190
1191static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
1192 struct io_mapped_ubuf **pimu,
1193 struct page **last_hpage)
1194{
1195 struct io_mapped_ubuf *imu = NULL;
1196 struct page **pages = NULL;
1197 unsigned long off;
1198 size_t size;
1199 int ret, nr_pages, i;
1200
5ff4fdff
PB		 1201 *pimu = ctx->dummy_ubuf;
1202 if (!iov->iov_base)
73572984 1203 return 0;
73572984 1204
73572984 1205 ret = -ENOMEM;
73572984
JA		 1206 pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len,
1207 &nr_pages);
1208 if (IS_ERR(pages)) {
1209 ret = PTR_ERR(pages);
1210 pages = NULL;
1211 goto done;
1212 }
1213
1214 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
1215 if (!imu)
1216 goto done;
1217
1218 ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage);
1219 if (ret) {
1220 unpin_user_pages(pages, nr_pages);
1221 goto done;
1222 }
1223
1224 off = (unsigned long) iov->iov_base & ~PAGE_MASK;
1225 size = iov->iov_len;
1226 for (i = 0; i < nr_pages; i++) {
1227 size_t vec_len;
1228
1229 vec_len = min_t(size_t, size, PAGE_SIZE - off);
1230 imu->bvec[i].bv_page = pages[i];
1231 imu->bvec[i].bv_len = vec_len;
1232 imu->bvec[i].bv_offset = off;
1233 off = 0;
1234 size -= vec_len;
1235 }
1236 /* store original address for later verification */
1237 imu->ubuf = (unsigned long) iov->iov_base;
1238 imu->ubuf_end = imu->ubuf + iov->iov_len;
1239 imu->nr_bvecs = nr_pages;
1240 *pimu = imu;
1241 ret = 0;
1242done:
1243 if (ret)
1244 kvfree(imu);
1245 kvfree(pages);
1246 return ret;
1247}
1248
1249static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
1250{
1251 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
1252 return ctx->user_bufs ? 0 : -ENOMEM;
1253}
1254
1255int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
1256 unsigned int nr_args, u64 __user *tags)
1257{
1258 struct page *last_hpage = NULL;
1259 struct io_rsrc_data *data;
1260 int i, ret;
1261 struct iovec iov;
1262
1263 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
1264
1265 if (ctx->user_bufs)
1266 return -EBUSY;
1267 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
1268 return -EINVAL;
1269 ret = io_rsrc_node_switch_start(ctx);
1270 if (ret)
1271 return ret;
1272 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
1273 if (ret)
1274 return ret;
1275 ret = io_buffers_map_alloc(ctx, nr_args);
1276 if (ret) {
1277 io_rsrc_data_free(data);
1278 return ret;
1279 }
1280
1281 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
1282 if (arg) {
1283 ret = io_copy_iov(ctx, &iov, arg, i);
1284 if (ret)
1285 break;
1286 ret = io_buffer_validate(&iov);
1287 if (ret)
1288 break;
1289 } else {
1290 memset(&iov, 0, sizeof(iov));
1291 }
1292
1293 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
1294 ret = -EINVAL;
1295 break;
1296 }
1297
1298 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
1299 &last_hpage);
1300 if (ret)
1301 break;
1302 }
1303
1304 WARN_ON_ONCE(ctx->buf_data);
1305
1306 ctx->buf_data = data;
1307 if (ret)
1308 __io_sqe_buffers_unregister(ctx);
1309 else
1310 io_rsrc_node_switch(ctx, NULL);
1311 return ret;
1312}
c059f785
PB		 1313
1314int io_import_fixed(int ddir, struct iov_iter *iter,
1315 struct io_mapped_ubuf *imu,
1316 u64 buf_addr, size_t len)
1317{
1318 u64 buf_end;
1319 size_t offset;
1320
1321 if (WARN_ON_ONCE(!imu))
1322 return -EFAULT;
1323 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
1324 return -EFAULT;
1325 /* not inside the mapped region */
1326 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
1327 return -EFAULT;
1328
1329 /*
1330 * May not be a start of buffer, set size appropriately
1331 * and advance us to the beginning.
1332 */
1333 offset = buf_addr - imu->ubuf;
1334 iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len);
1335
1336 if (offset) {
1337 /*
1338 * Don't use iov_iter_advance() here, as it's really slow for
1339 * using the latter parts of a big fixed buffer - it iterates
1340 * over each segment manually. We can cheat a bit here, because
1341 * we know that:
1342 *
1343 * 1) it's a BVEC iter, we set it up
1344 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1345 * first and last bvec
1346 *
1347 * So just find our index, and adjust the iterator afterwards.
1348 * If the offset is within the first bvec (or the whole first
1349 * bvec, just use iov_iter_advance(). This makes it easier
1350 * since we can just skip the first segment, which may not
1351 * be PAGE_SIZE aligned.
1352 */
1353 const struct bio_vec *bvec = imu->bvec;
1354
1355 if (offset <= bvec->bv_len) {
1356 iov_iter_advance(iter, offset);
1357 } else {
1358 unsigned long seg_skip;
1359
1360 /* skip first vec */
1361 offset -= bvec->bv_len;
1362 seg_skip = 1 + (offset >> PAGE_SHIFT);
1363
1364 iter->bvec = bvec + seg_skip;
1365 iter->nr_segs -= seg_skip;
1366 iter->count -= bvec->bv_len + offset;
1367 iter->iov_offset = offset & ~PAGE_MASK;
1368 }
1369 }
1370
1371 return 0;
1372}
Linux 6.x block layer and io_uring tree(s)