projects / linux-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
io_uring/rsrc: refactor io_rsrc_node_switch
[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
2933ae6e 207struct 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{
2933ae6e
PB		 234 struct io_rsrc_node *node = ctx->rsrc_node;
235 struct io_rsrc_node *backup = io_rsrc_node_alloc(ctx);
73572984 236
2933ae6e
PB		 237 if (WARN_ON_ONCE(!backup))
238 return;
73572984 239
2933ae6e
PB		 240 data_to_kill->refs++;
241 node->rsrc_data = data_to_kill;
242 list_add_tail(&node->node, &ctx->rsrc_ref_list);
243 /* put master ref */
244 io_put_rsrc_node(ctx, node);
245 ctx->rsrc_node = backup;
73572984
JA		 246}
247
248int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
249{
528407b1
PB		 250 if (io_alloc_cache_empty(&ctx->rsrc_node_cache)) {
251 struct io_rsrc_node *node = kzalloc(sizeof(*node), GFP_KERNEL);
252
253 if (!node)
254 return -ENOMEM;
255 io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache);
256 }
257 return 0;
73572984
JA		 258}
259
260__cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data,
261 struct io_ring_ctx *ctx)
262{
263 int ret;
264
265 /* As we may drop ->uring_lock, other task may have started quiesce */
266 if (data->quiesce)
267 return -ENXIO;
77e3202a
PB		 268 ret = io_rsrc_node_switch_start(ctx);
269 if (ret)
270 return ret;
271 io_rsrc_node_switch(ctx, data);
272
757ef468
PB		 273 /* kill initial ref */
274 if (io_put_rsrc_data_ref(data))
77e3202a 275 return 0;
73572984
JA		 276
277 data->quiesce = true;
77e3202a 278 mutex_unlock(&ctx->uring_lock);
73572984 279 do {
ef67fcb4 280 ret = io_run_task_work_sig(ctx);
77e3202a 281 if (ret < 0) {
77e3202a 282 mutex_lock(&ctx->uring_lock);
757ef468
PB		 283 if (!data->refs) {
284 ret = 0;
285 } else {
286 /* restore the master reference */
287 data->refs++;
288 }
77e3202a
PB		 289 break;
290 }
73572984
JA		 291 ret = wait_for_completion_interruptible(&data->done);
292 if (!ret) {
293 mutex_lock(&ctx->uring_lock);
757ef468 294 if (!data->refs)
73572984 295 break;
0ced756f
PB		 296 /*
297 * it has been revived by another thread while
298 * we were unlocked
299 */
300 mutex_unlock(&ctx->uring_lock);
73572984 301 }
77e3202a 302 } while (1);
73572984
JA		 303 data->quiesce = false;
304
305 return ret;
306}
307
308static void io_free_page_table(void **table, size_t size)
309{
310 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
311
312 for (i = 0; i < nr_tables; i++)
313 kfree(table[i]);
314 kfree(table);
315}
316
317static void io_rsrc_data_free(struct io_rsrc_data *data)
318{
319 size_t size = data->nr * sizeof(data->tags[0][0]);
320
321 if (data->tags)
322 io_free_page_table((void **)data->tags, size);
323 kfree(data);
324}
325
326static __cold void **io_alloc_page_table(size_t size)
327{
328 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
329 size_t init_size = size;
330 void **table;
331
332 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
333 if (!table)
334 return NULL;
335
336 for (i = 0; i < nr_tables; i++) {
337 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
338
339 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
340 if (!table[i]) {
341 io_free_page_table(table, init_size);
342 return NULL;
343 }
344 size -= this_size;
345 }
346 return table;
347}
348
349__cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx,
350 rsrc_put_fn *do_put, u64 __user *utags,
351 unsigned nr, struct io_rsrc_data **pdata)
352{
353 struct io_rsrc_data *data;
6acd352d 354 int ret = 0;
73572984
JA		 355 unsigned i;
356
357 data = kzalloc(sizeof(*data), GFP_KERNEL);
358 if (!data)
359 return -ENOMEM;
360 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
361 if (!data->tags) {
362 kfree(data);
363 return -ENOMEM;
364 }
365
366 data->nr = nr;
367 data->ctx = ctx;
368 data->do_put = do_put;
757ef468 369 data->refs = 1;
73572984
JA		 370 if (utags) {
371 ret = -EFAULT;
372 for (i = 0; i < nr; i++) {
373 u64 *tag_slot = io_get_tag_slot(data, i);
374
375 if (copy_from_user(tag_slot, &utags[i],
376 sizeof(*tag_slot)))
377 goto fail;
378 }
379 }
73572984
JA		 380 init_completion(&data->done);
381 *pdata = data;
382 return 0;
383fail:
384 io_rsrc_data_free(data);
385 return ret;
386}
387
388static int __io_sqe_files_update(struct io_ring_ctx *ctx,
389 struct io_uring_rsrc_update2 *up,
390 unsigned nr_args)
391{
392 u64 __user *tags = u64_to_user_ptr(up->tags);
393 __s32 __user *fds = u64_to_user_ptr(up->data);
394 struct io_rsrc_data *data = ctx->file_data;
395 struct io_fixed_file *file_slot;
396 struct file *file;
397 int fd, i, err = 0;
398 unsigned int done;
399 bool needs_switch = false;
400
401 if (!ctx->file_data)
402 return -ENXIO;
403 if (up->offset + nr_args > ctx->nr_user_files)
404 return -EINVAL;
405
406 for (done = 0; done < nr_args; done++) {
407 u64 tag = 0;
408
409 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
410 copy_from_user(&fd, &fds[done], sizeof(fd))) {
411 err = -EFAULT;
412 break;
413 }
414 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
415 err = -EINVAL;
416 break;
417 }
418 if (fd == IORING_REGISTER_FILES_SKIP)
419 continue;
420
421 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
422 file_slot = io_fixed_file_slot(&ctx->file_table, i);
423
424 if (file_slot->file_ptr) {
425 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
426 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file);
427 if (err)
428 break;
429 file_slot->file_ptr = 0;
430 io_file_bitmap_clear(&ctx->file_table, i);
431 needs_switch = true;
432 }
433 if (fd != -1) {
434 file = fget(fd);
435 if (!file) {
436 err = -EBADF;
437 break;
438 }
439 /*
440 * Don't allow io_uring instances to be registered. If
441 * UNIX isn't enabled, then this causes a reference
442 * cycle and this instance can never get freed. If UNIX
443 * is enabled we'll handle it just fine, but there's
444 * still no point in allowing a ring fd as it doesn't
445 * support regular read/write anyway.
446 */
447 if (io_is_uring_fops(file)) {
448 fput(file);
449 err = -EBADF;
450 break;
451 }
452 err = io_scm_file_account(ctx, file);
453 if (err) {
454 fput(file);
455 break;
456 }
457 *io_get_tag_slot(data, i) = tag;
458 io_fixed_file_set(file_slot, file);
459 io_file_bitmap_set(&ctx->file_table, i);
460 }
461 }
462
463 if (needs_switch)
464 io_rsrc_node_switch(ctx, data);
465 return done ? done : err;
466}
467
468static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
469 struct io_uring_rsrc_update2 *up,
470 unsigned int nr_args)
471{
472 u64 __user *tags = u64_to_user_ptr(up->tags);
473 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
474 struct page *last_hpage = NULL;
475 bool needs_switch = false;
476 __u32 done;
477 int i, err;
478
479 if (!ctx->buf_data)
480 return -ENXIO;
481 if (up->offset + nr_args > ctx->nr_user_bufs)
482 return -EINVAL;
483
484 for (done = 0; done < nr_args; done++) {
485 struct io_mapped_ubuf *imu;
486 int offset = up->offset + done;
487 u64 tag = 0;
488
489 err = io_copy_iov(ctx, &iov, iovs, done);
490 if (err)
491 break;
492 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
493 err = -EFAULT;
494 break;
495 }
496 err = io_buffer_validate(&iov);
497 if (err)
498 break;
499 if (!iov.iov_base && tag) {
500 err = -EINVAL;
501 break;
502 }
503 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
504 if (err)
505 break;
506
507 i = array_index_nospec(offset, ctx->nr_user_bufs);
508 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
509 err = io_queue_rsrc_removal(ctx->buf_data, i,
510 ctx->rsrc_node, ctx->user_bufs[i]);
511 if (unlikely(err)) {
512 io_buffer_unmap(ctx, &imu);
513 break;
514 }
5ff4fdff 515 ctx->user_bufs[i] = ctx->dummy_ubuf;
73572984
JA		 516 needs_switch = true;
517 }
518
519 ctx->user_bufs[i] = imu;
520 *io_get_tag_slot(ctx->buf_data, offset) = tag;
521 }
522
523 if (needs_switch)
524 io_rsrc_node_switch(ctx, ctx->buf_data);
525 return done ? done : err;
526}
527
528static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
529 struct io_uring_rsrc_update2 *up,
530 unsigned nr_args)
531{
532 __u32 tmp;
533 int err;
534
786788a8
PB		 535 lockdep_assert_held(&ctx->uring_lock);
536
73572984
JA		 537 if (check_add_overflow(up->offset, nr_args, &tmp))
538 return -EOVERFLOW;
539 err = io_rsrc_node_switch_start(ctx);
540 if (err)
541 return err;
542
543 switch (type) {
544 case IORING_RSRC_FILE:
545 return __io_sqe_files_update(ctx, up, nr_args);
546 case IORING_RSRC_BUFFER:
547 return __io_sqe_buffers_update(ctx, up, nr_args);
548 }
549 return -EINVAL;
550}
551
552int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
553 unsigned nr_args)
554{
555 struct io_uring_rsrc_update2 up;
556
557 if (!nr_args)
558 return -EINVAL;
559 memset(&up, 0, sizeof(up));
560 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
561 return -EFAULT;
562 if (up.resv || up.resv2)
563 return -EINVAL;
564 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
565}
566
567int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
568 unsigned size, unsigned type)
569{
570 struct io_uring_rsrc_update2 up;
571
572 if (size != sizeof(up))
573 return -EINVAL;
574 if (copy_from_user(&up, arg, sizeof(up)))
575 return -EFAULT;
576 if (!up.nr || up.resv || up.resv2)
577 return -EINVAL;
578 return __io_register_rsrc_update(ctx, type, &up, up.nr);
579}
580
581__cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
582 unsigned int size, unsigned int type)
583{
584 struct io_uring_rsrc_register rr;
585
586 /* keep it extendible */
587 if (size != sizeof(rr))
588 return -EINVAL;
589
590 memset(&rr, 0, sizeof(rr));
591 if (copy_from_user(&rr, arg, size))
592 return -EFAULT;
593 if (!rr.nr || rr.resv2)
594 return -EINVAL;
595 if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE)
596 return -EINVAL;
597
598 switch (type) {
599 case IORING_RSRC_FILE:
600 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
601 break;
602 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
603 rr.nr, u64_to_user_ptr(rr.tags));
604 case IORING_RSRC_BUFFER:
605 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
606 break;
607 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
608 rr.nr, u64_to_user_ptr(rr.tags));
609 }
610 return -EINVAL;
611}
612
d9808ceb 613int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
73572984 614{
f2ccb5ae 615 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
73572984
JA		 616
617 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
618 return -EINVAL;
619 if (sqe->rw_flags || sqe->splice_fd_in)
620 return -EINVAL;
621
622 up->offset = READ_ONCE(sqe->off);
623 up->nr_args = READ_ONCE(sqe->len);
624 if (!up->nr_args)
625 return -EINVAL;
626 up->arg = READ_ONCE(sqe->addr);
627 return 0;
628}
629
630static int io_files_update_with_index_alloc(struct io_kiocb *req,
631 unsigned int issue_flags)
632{
f2ccb5ae 633 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
73572984
JA		 634 __s32 __user *fds = u64_to_user_ptr(up->arg);
635 unsigned int done;
636 struct file *file;
637 int ret, fd;
638
639 if (!req->ctx->file_data)
640 return -ENXIO;
641
642 for (done = 0; done < up->nr_args; done++) {
643 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
644 ret = -EFAULT;
645 break;
646 }
647
648 file = fget(fd);
649 if (!file) {
650 ret = -EBADF;
651 break;
652 }
653 ret = io_fixed_fd_install(req, issue_flags, file,
654 IORING_FILE_INDEX_ALLOC);
655 if (ret < 0)
656 break;
657 if (copy_to_user(&fds[done], &ret, sizeof(ret))) {
f110ed84 658 __io_close_fixed(req->ctx, issue_flags, ret);
73572984
JA		 659 ret = -EFAULT;
660 break;
661 }
662 }
663
664 if (done)
665 return done;
666 return ret;
667}
668
d9808ceb 669int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
73572984 670{
f2ccb5ae 671 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
73572984
JA		 672 struct io_ring_ctx *ctx = req->ctx;
673 struct io_uring_rsrc_update2 up2;
674 int ret;
675
676 up2.offset = up->offset;
677 up2.data = up->arg;
678 up2.nr = 0;
679 up2.tags = 0;
680 up2.resv = 0;
681 up2.resv2 = 0;
682
683 if (up->offset == IORING_FILE_INDEX_ALLOC) {
684 ret = io_files_update_with_index_alloc(req, issue_flags);
685 } else {
686 io_ring_submit_lock(ctx, issue_flags);
687 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
688 &up2, up->nr_args);
689 io_ring_submit_unlock(ctx, issue_flags);
690 }
691
692 if (ret < 0)
693 req_set_fail(req);
694 io_req_set_res(req, ret, 0);
695 return IOU_OK;
696}
697
698int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
699 struct io_rsrc_node *node, void *rsrc)
700{
701 u64 *tag_slot = io_get_tag_slot(data, idx);
702 struct io_rsrc_put *prsrc;
ff7c75ec 703 bool inline_item = true;
73572984 704
ff7c75ec
PB		 705 if (!node->inline_items) {
706 prsrc = &node->item;
707 node->inline_items++;
708 } else {
709 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
710 if (!prsrc)
711 return -ENOMEM;
712 inline_item = false;
713 }
73572984
JA		 714
715 prsrc->tag = *tag_slot;
716 *tag_slot = 0;
717 prsrc->rsrc = rsrc;
ff7c75ec
PB		 718 if (!inline_item)
719 list_add(&prsrc->list, &node->item_list);
73572984
JA		 720 return 0;
721}
722
723void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
724{
73572984
JA		 725 int i;
726
727 for (i = 0; i < ctx->nr_user_files; i++) {
728 struct file *file = io_file_from_index(&ctx->file_table, i);
729
38eddb2c
PB		 730 /* skip scm accounted files, they'll be freed by ->ring_sock */
731 if (!file || io_file_need_scm(file))
73572984
JA		 732 continue;
733 io_file_bitmap_clear(&ctx->file_table, i);
734 fput(file);
735 }
73572984
JA		 736
737#if defined(CONFIG_UNIX)
738 if (ctx->ring_sock) {
739 struct sock *sock = ctx->ring_sock->sk;
740 struct sk_buff *skb;
741
742 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
743 kfree_skb(skb);
744 }
745#endif
746 io_free_file_tables(&ctx->file_table);
02a4d923 747 io_file_table_set_alloc_range(ctx, 0, 0);
73572984
JA		 748 io_rsrc_data_free(ctx->file_data);
749 ctx->file_data = NULL;
750 ctx->nr_user_files = 0;
751}
752
753int io_sqe_files_unregister(struct io_ring_ctx *ctx)
754{
755 unsigned nr = ctx->nr_user_files;
756 int ret;
757
758 if (!ctx->file_data)
759 return -ENXIO;
760
761 /*
762 * Quiesce may unlock ->uring_lock, and while it's not held
763 * prevent new requests using the table.
764 */
765 ctx->nr_user_files = 0;
766 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
767 ctx->nr_user_files = nr;
768 if (!ret)
769 __io_sqe_files_unregister(ctx);
770 return ret;
771}
772
773/*
774 * Ensure the UNIX gc is aware of our file set, so we are certain that
775 * the io_uring can be safely unregistered on process exit, even if we have
776 * loops in the file referencing. We account only files that can hold other
777 * files because otherwise they can't form a loop and so are not interesting
778 * for GC.
779 */
780int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file)
781{
782#if defined(CONFIG_UNIX)
783 struct sock *sk = ctx->ring_sock->sk;
784 struct sk_buff_head *head = &sk->sk_receive_queue;
785 struct scm_fp_list *fpl;
786 struct sk_buff *skb;
787
788 if (likely(!io_file_need_scm(file)))
789 return 0;
790
791 /*
792 * See if we can merge this file into an existing skb SCM_RIGHTS
793 * file set. If there's no room, fall back to allocating a new skb
794 * and filling it in.
795 */
796 spin_lock_irq(&head->lock);
797 skb = skb_peek(head);
798 if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD)
799 __skb_unlink(skb, head);
800 else
801 skb = NULL;
802 spin_unlock_irq(&head->lock);
803
804 if (!skb) {
805 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
806 if (!fpl)
807 return -ENOMEM;
808
809 skb = alloc_skb(0, GFP_KERNEL);
810 if (!skb) {
811 kfree(fpl);
812 return -ENOMEM;
813 }
814
815 fpl->user = get_uid(current_user());
816 fpl->max = SCM_MAX_FD;
817 fpl->count = 0;
818
819 UNIXCB(skb).fp = fpl;
820 skb->sk = sk;
0091bfc8 821 skb->scm_io_uring = 1;
73572984
JA		 822 skb->destructor = unix_destruct_scm;
823 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
824 }
825
826 fpl = UNIXCB(skb).fp;
827 fpl->fp[fpl->count++] = get_file(file);
828 unix_inflight(fpl->user, file);
829 skb_queue_head(head, skb);
830 fput(file);
831#endif
832 return 0;
833}
834
835static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
836{
837 struct file *file = prsrc->file;
838#if defined(CONFIG_UNIX)
839 struct sock *sock = ctx->ring_sock->sk;
840 struct sk_buff_head list, *head = &sock->sk_receive_queue;
841 struct sk_buff *skb;
842 int i;
843
844 if (!io_file_need_scm(file)) {
845 fput(file);
846 return;
847 }
848
849 __skb_queue_head_init(&list);
850
851 /*
852 * Find the skb that holds this file in its SCM_RIGHTS. When found,
853 * remove this entry and rearrange the file array.
854 */
855 skb = skb_dequeue(head);
856 while (skb) {
857 struct scm_fp_list *fp;
858
859 fp = UNIXCB(skb).fp;
860 for (i = 0; i < fp->count; i++) {
861 int left;
862
863 if (fp->fp[i] != file)
864 continue;
865
866 unix_notinflight(fp->user, fp->fp[i]);
867 left = fp->count - 1 - i;
868 if (left) {
869 memmove(&fp->fp[i], &fp->fp[i + 1],
870 left * sizeof(struct file *));
871 }
872 fp->count--;
873 if (!fp->count) {
874 kfree_skb(skb);
875 skb = NULL;
876 } else {
877 __skb_queue_tail(&list, skb);
878 }
879 fput(file);
880 file = NULL;
881 break;
882 }
883
884 if (!file)
885 break;
886
887 __skb_queue_tail(&list, skb);
888
889 skb = skb_dequeue(head);
890 }
891
892 if (skb_peek(&list)) {
893 spin_lock_irq(&head->lock);
894 while ((skb = __skb_dequeue(&list)) != NULL)
895 __skb_queue_tail(head, skb);
896 spin_unlock_irq(&head->lock);
897 }
898#else
899 fput(file);
900#endif
901}
902
903int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
904 unsigned nr_args, u64 __user *tags)
905{
906 __s32 __user *fds = (__s32 __user *) arg;
907 struct file *file;
908 int fd, ret;
909 unsigned i;
910
911 if (ctx->file_data)
912 return -EBUSY;
913 if (!nr_args)
914 return -EINVAL;
915 if (nr_args > IORING_MAX_FIXED_FILES)
916 return -EMFILE;
917 if (nr_args > rlimit(RLIMIT_NOFILE))
918 return -EMFILE;
73572984
JA		 919 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
920 &ctx->file_data);
921 if (ret)
922 return ret;
923
924 if (!io_alloc_file_tables(&ctx->file_table, nr_args)) {
925 io_rsrc_data_free(ctx->file_data);
926 ctx->file_data = NULL;
927 return -ENOMEM;
928 }
929
930 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
931 struct io_fixed_file *file_slot;
932
933 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) {
934 ret = -EFAULT;
935 goto fail;
936 }
937 /* allow sparse sets */
938 if (!fds || fd == -1) {
939 ret = -EINVAL;
940 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
941 goto fail;
942 continue;
943 }
944
945 file = fget(fd);
946 ret = -EBADF;
947 if (unlikely(!file))
948 goto fail;
949
950 /*
951 * Don't allow io_uring instances to be registered. If UNIX
952 * isn't enabled, then this causes a reference cycle and this
953 * instance can never get freed. If UNIX is enabled we'll
954 * handle it just fine, but there's still no point in allowing
955 * a ring fd as it doesn't support regular read/write anyway.
956 */
957 if (io_is_uring_fops(file)) {
958 fput(file);
959 goto fail;
960 }
961 ret = io_scm_file_account(ctx, file);
962 if (ret) {
963 fput(file);
964 goto fail;
965 }
966 file_slot = io_fixed_file_slot(&ctx->file_table, i);
967 io_fixed_file_set(file_slot, file);
968 io_file_bitmap_set(&ctx->file_table, i);
969 }
970
6e73dffb
PB		 971 /* default it to the whole table */
972 io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files);
73572984
JA		 973 return 0;
974fail:
975 __io_sqe_files_unregister(ctx);
976 return ret;
977}
978
979static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
980{
981 io_buffer_unmap(ctx, &prsrc->buf);
982 prsrc->buf = NULL;
983}
984
985void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
986{
987 unsigned int i;
988
989 for (i = 0; i < ctx->nr_user_bufs; i++)
990 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
991 kfree(ctx->user_bufs);
992 io_rsrc_data_free(ctx->buf_data);
993 ctx->user_bufs = NULL;
994 ctx->buf_data = NULL;
995 ctx->nr_user_bufs = 0;
996}
997
998int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
999{
1000 unsigned nr = ctx->nr_user_bufs;
1001 int ret;
1002
1003 if (!ctx->buf_data)
1004 return -ENXIO;
1005
1006 /*
1007 * Quiesce may unlock ->uring_lock, and while it's not held
1008 * prevent new requests using the table.
1009 */
1010 ctx->nr_user_bufs = 0;
1011 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
1012 ctx->nr_user_bufs = nr;
1013 if (!ret)
1014 __io_sqe_buffers_unregister(ctx);
1015 return ret;
1016}
1017
1018/*
1019 * Not super efficient, but this is just a registration time. And we do cache
1020 * the last compound head, so generally we'll only do a full search if we don't
1021 * match that one.
1022 *
1023 * We check if the given compound head page has already been accounted, to
1024 * avoid double accounting it. This allows us to account the full size of the
1025 * page, not just the constituent pages of a huge page.
1026 */
1027static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
1028 int nr_pages, struct page *hpage)
1029{
1030 int i, j;
1031
1032 /* check current page array */
1033 for (i = 0; i < nr_pages; i++) {
1034 if (!PageCompound(pages[i]))
1035 continue;
1036 if (compound_head(pages[i]) == hpage)
1037 return true;
1038 }
1039
1040 /* check previously registered pages */
1041 for (i = 0; i < ctx->nr_user_bufs; i++) {
1042 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
1043
1044 for (j = 0; j < imu->nr_bvecs; j++) {
1045 if (!PageCompound(imu->bvec[j].bv_page))
1046 continue;
1047 if (compound_head(imu->bvec[j].bv_page) == hpage)
1048 return true;
1049 }
1050 }
1051
1052 return false;
1053}
1054
1055static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
1056 int nr_pages, struct io_mapped_ubuf *imu,
1057 struct page **last_hpage)
1058{
1059 int i, ret;
1060
1061 imu->acct_pages = 0;
1062 for (i = 0; i < nr_pages; i++) {
1063 if (!PageCompound(pages[i])) {
1064 imu->acct_pages++;
1065 } else {
1066 struct page *hpage;
1067
1068 hpage = compound_head(pages[i]);
1069 if (hpage == *last_hpage)
1070 continue;
1071 *last_hpage = hpage;
1072 if (headpage_already_acct(ctx, pages, i, hpage))
1073 continue;
1074 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
1075 }
1076 }
1077
1078 if (!imu->acct_pages)
1079 return 0;
1080
1081 ret = io_account_mem(ctx, imu->acct_pages);
1082 if (ret)
1083 imu->acct_pages = 0;
1084 return ret;
1085}
1086
1087struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages)
1088{
1089 unsigned long start, end, nr_pages;
1090 struct vm_area_struct **vmas = NULL;
1091 struct page **pages = NULL;
1092 int i, pret, ret = -ENOMEM;
1093
1094 end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1095 start = ubuf >> PAGE_SHIFT;
1096 nr_pages = end - start;
1097
1098 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
1099 if (!pages)
1100 goto done;
1101
1102 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
1103 GFP_KERNEL);
1104 if (!vmas)
1105 goto done;
1106
1107 ret = 0;
1108 mmap_read_lock(current->mm);
1109 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
1110 pages, vmas);
1111 if (pret == nr_pages) {
edd47826
PB		 1112 struct file *file = vmas[0]->vm_file;
1113
73572984
JA		 1114 /* don't support file backed memory */
1115 for (i = 0; i < nr_pages; i++) {
edd47826
PB		 1116 if (vmas[i]->vm_file != file) {
1117 ret = -EINVAL;
1118 break;
1119 }
1120 if (!file)
73572984 1121 continue;
edd47826 1122 if (!vma_is_shmem(vmas[i]) && !is_file_hugepages(file)) {
73572984
JA		 1123 ret = -EOPNOTSUPP;
1124 break;
1125 }
1126 }
1127 *npages = nr_pages;
1128 } else {
1129 ret = pret < 0 ? pret : -EFAULT;
1130 }
1131 mmap_read_unlock(current->mm);
1132 if (ret) {
1133 /*
1134 * if we did partial map, or found file backed vmas,
1135 * release any pages we did get
1136 */
1137 if (pret > 0)
1138 unpin_user_pages(pages, pret);
1139 goto done;
1140 }
1141 ret = 0;
1142done:
1143 kvfree(vmas);
1144 if (ret < 0) {
1145 kvfree(pages);
1146 pages = ERR_PTR(ret);
1147 }
1148 return pages;
1149}
1150
1151static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
1152 struct io_mapped_ubuf **pimu,
1153 struct page **last_hpage)
1154{
1155 struct io_mapped_ubuf *imu = NULL;
1156 struct page **pages = NULL;
1157 unsigned long off;
1158 size_t size;
1159 int ret, nr_pages, i;
977bc873 1160 struct folio *folio = NULL;
73572984 1161
5ff4fdff
PB		 1162 *pimu = ctx->dummy_ubuf;
1163 if (!iov->iov_base)
73572984 1164 return 0;
73572984 1165
73572984 1166 ret = -ENOMEM;
73572984
JA		 1167 pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len,
1168 &nr_pages);
1169 if (IS_ERR(pages)) {
1170 ret = PTR_ERR(pages);
1171 pages = NULL;
1172 goto done;
1173 }
1174
57bebf80
PB		 1175 /* If it's a huge page, try to coalesce them into a single bvec entry */
1176 if (nr_pages > 1) {
1177 folio = page_folio(pages[0]);
1178 for (i = 1; i < nr_pages; i++) {
1179 if (page_folio(pages[i]) != folio) {
1180 folio = NULL;
1181 break;
1182 }
1183 }
1184 if (folio) {
d2acf789
PB		 1185 /*
1186 * The pages are bound to the folio, it doesn't
1187 * actually unpin them but drops all but one reference,
1188 * which is usually put down by io_buffer_unmap().
1189 * Note, needs a better helper.
1190 */
1191 unpin_user_pages(&pages[1], nr_pages - 1);
57bebf80
PB		 1192 nr_pages = 1;
1193 }
1194 }
1195
73572984
JA		 1196 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
1197 if (!imu)
1198 goto done;
1199
1200 ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage);
1201 if (ret) {
1202 unpin_user_pages(pages, nr_pages);
1203 goto done;
1204 }
1205
1206 off = (unsigned long) iov->iov_base & ~PAGE_MASK;
1207 size = iov->iov_len;
57bebf80
PB		 1208 /* store original address for later verification */
1209 imu->ubuf = (unsigned long) iov->iov_base;
1210 imu->ubuf_end = imu->ubuf + iov->iov_len;
1211 imu->nr_bvecs = nr_pages;
1212 *pimu = imu;
1213 ret = 0;
1214
1215 if (folio) {
1216 bvec_set_page(&imu->bvec[0], pages[0], size, off);
1217 goto done;
1218 }
73572984
JA		 1219 for (i = 0; i < nr_pages; i++) {
1220 size_t vec_len;
1221
1222 vec_len = min_t(size_t, size, PAGE_SIZE - off);
cc342a21 1223 bvec_set_page(&imu->bvec[i], pages[i], vec_len, off);
73572984
JA		 1224 off = 0;
1225 size -= vec_len;
1226 }
73572984
JA		 1227done:
1228 if (ret)
1229 kvfree(imu);
1230 kvfree(pages);
1231 return ret;
1232}
1233
1234static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
1235{
1236 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
1237 return ctx->user_bufs ? 0 : -ENOMEM;
1238}
1239
1240int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
1241 unsigned int nr_args, u64 __user *tags)
1242{
1243 struct page *last_hpage = NULL;
1244 struct io_rsrc_data *data;
1245 int i, ret;
1246 struct iovec iov;
1247
1248 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
1249
1250 if (ctx->user_bufs)
1251 return -EBUSY;
1252 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
1253 return -EINVAL;
73572984
JA		 1254 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
1255 if (ret)
1256 return ret;
1257 ret = io_buffers_map_alloc(ctx, nr_args);
1258 if (ret) {
1259 io_rsrc_data_free(data);
1260 return ret;
1261 }
1262
1263 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
1264 if (arg) {
1265 ret = io_copy_iov(ctx, &iov, arg, i);
1266 if (ret)
1267 break;
1268 ret = io_buffer_validate(&iov);
1269 if (ret)
1270 break;
1271 } else {
1272 memset(&iov, 0, sizeof(iov));
1273 }
1274
1275 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
1276 ret = -EINVAL;
1277 break;
1278 }
1279
1280 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
1281 &last_hpage);
1282 if (ret)
1283 break;
1284 }
1285
1286 WARN_ON_ONCE(ctx->buf_data);
1287
1288 ctx->buf_data = data;
1289 if (ret)
1290 __io_sqe_buffers_unregister(ctx);
73572984
JA		 1291 return ret;
1292}
c059f785
PB		 1293
1294int io_import_fixed(int ddir, struct iov_iter *iter,
1295 struct io_mapped_ubuf *imu,
1296 u64 buf_addr, size_t len)
1297{
1298 u64 buf_end;
1299 size_t offset;
1300
1301 if (WARN_ON_ONCE(!imu))
1302 return -EFAULT;
1303 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
1304 return -EFAULT;
1305 /* not inside the mapped region */
1306 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
1307 return -EFAULT;
1308
1309 /*
6bf65a1b 1310 * Might not be a start of buffer, set size appropriately
c059f785
PB		 1311 * and advance us to the beginning.
1312 */
1313 offset = buf_addr - imu->ubuf;
1314 iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len);
1315
1316 if (offset) {
1317 /*
1318 * Don't use iov_iter_advance() here, as it's really slow for
1319 * using the latter parts of a big fixed buffer - it iterates
1320 * over each segment manually. We can cheat a bit here, because
1321 * we know that:
1322 *
1323 * 1) it's a BVEC iter, we set it up
1324 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1325 * first and last bvec
1326 *
1327 * So just find our index, and adjust the iterator afterwards.
1328 * If the offset is within the first bvec (or the whole first
1329 * bvec, just use iov_iter_advance(). This makes it easier
1330 * since we can just skip the first segment, which may not
1331 * be PAGE_SIZE aligned.
1332 */
1333 const struct bio_vec *bvec = imu->bvec;
1334
1335 if (offset <= bvec->bv_len) {
57bebf80
PB		 1336 /*
1337 * Note, huge pages buffers consists of one large
1338 * bvec entry and should always go this way. The other
1339 * branch doesn't expect non PAGE_SIZE'd chunks.
1340 */
b000ae0e
PB		 1341 iter->bvec = bvec;
1342 iter->nr_segs = bvec->bv_len;
1343 iter->count -= offset;
1344 iter->iov_offset = offset;
c059f785
PB		 1345 } else {
1346 unsigned long seg_skip;
1347
1348 /* skip first vec */
1349 offset -= bvec->bv_len;
1350 seg_skip = 1 + (offset >> PAGE_SHIFT);
1351
1352 iter->bvec = bvec + seg_skip;
1353 iter->nr_segs -= seg_skip;
1354 iter->count -= bvec->bv_len + offset;
1355 iter->iov_offset = offset & ~PAGE_MASK;
1356 }
1357 }
1358
1359 return 0;
1360}
Linux 6.x block layer and io_uring tree(s)