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/namei.h>
9 #include <linux/poll.h>
10 #include <linux/vmalloc.h>
11 #include <linux/io_uring.h>
12
13 #include <uapi/linux/io_uring.h>
14
15 #include "io_uring.h"
16 #include "opdef.h"
17 #include "kbuf.h"
18 #include "memmap.h"
19
20 /* BIDs are addressed by a 16-bit field in a CQE */
21 #define MAX_BIDS_PER_BGID (1 << 16)
22
23 struct kmem_cache *io_buf_cachep;
24
25 struct io_provide_buf {
26 struct file *file;
27 __u64 addr;
28 __u32 len;
29 __u32 bgid;
30 __u32 nbufs;
31 __u16 bid;
32 };
33
34 static inline struct io_buffer_list *io_buffer_get_list(struct io_ring_ctx *ctx,
35 unsigned int bgid)
36 {
37 lockdep_assert_held(&ctx->uring_lock);
38
39 return xa_load(&ctx->io_bl_xa, bgid);
40 }
41
42 static int io_buffer_add_list(struct io_ring_ctx *ctx,
43 struct io_buffer_list *bl, unsigned int bgid)
44 {
45 /*
46 * Store buffer group ID and finally mark the list as visible.
47 * The normal lookup doesn't care about the visibility as we're
48 * always under the ->uring_lock, but the RCU lookup from mmap does.
49 */
50 bl->bgid = bgid;
51 atomic_set(&bl->refs, 1);
52 return xa_err(xa_store(&ctx->io_bl_xa, bgid, bl, GFP_KERNEL));
53 }
54
55 bool io_kbuf_recycle_legacy(struct io_kiocb *req, unsigned issue_flags)
56 {
57 struct io_ring_ctx *ctx = req->ctx;
58 struct io_buffer_list *bl;
59 struct io_buffer *buf;
60
61 io_ring_submit_lock(ctx, issue_flags);
62
63 buf = req->kbuf;
64 bl = io_buffer_get_list(ctx, buf->bgid);
65 list_add(&buf->list, &bl->buf_list);
66 req->flags &= ~REQ_F_BUFFER_SELECTED;
67 req->buf_index = buf->bgid;
68
69 io_ring_submit_unlock(ctx, issue_flags);
70 return true;
71 }
72
73 void __io_put_kbuf(struct io_kiocb *req, unsigned issue_flags)
74 {
75 /*
76 * We can add this buffer back to two lists:
77 *
78 * 1) The io_buffers_cache list. This one is protected by the
79 * ctx->uring_lock. If we already hold this lock, add back to this
80 * list as we can grab it from issue as well.
81 * 2) The io_buffers_comp list. This one is protected by the
82 * ctx->completion_lock.
83 *
84 * We migrate buffers from the comp_list to the issue cache list
85 * when we need one.
86 */
87 if (issue_flags & IO_URING_F_UNLOCKED) {
88 struct io_ring_ctx *ctx = req->ctx;
89
90 spin_lock(&ctx->completion_lock);
91 __io_put_kbuf_list(req, &ctx->io_buffers_comp);
92 spin_unlock(&ctx->completion_lock);
93 } else {
94 lockdep_assert_held(&req->ctx->uring_lock);
95
96 __io_put_kbuf_list(req, &req->ctx->io_buffers_cache);
97 }
98 }
99
100 static void __user *io_provided_buffer_select(struct io_kiocb *req, size_t *len,
101 struct io_buffer_list *bl)
102 {
103 if (!list_empty(&bl->buf_list)) {
104 struct io_buffer *kbuf;
105
106 kbuf = list_first_entry(&bl->buf_list, struct io_buffer, list);
107 list_del(&kbuf->list);
108 if (*len == 0 || *len > kbuf->len)
109 *len = kbuf->len;
110 if (list_empty(&bl->buf_list))
111 req->flags |= REQ_F_BL_EMPTY;
112 req->flags |= REQ_F_BUFFER_SELECTED;
113 req->kbuf = kbuf;
114 req->buf_index = kbuf->bid;
115 return u64_to_user_ptr(kbuf->addr);
116 }
117 return NULL;
118 }
119
120 static int io_provided_buffers_select(struct io_kiocb *req, size_t *len,
121 struct io_buffer_list *bl,
122 struct iovec *iov)
123 {
124 void __user *buf;
125
126 buf = io_provided_buffer_select(req, len, bl);
127 if (unlikely(!buf))
128 return -ENOBUFS;
129
130 iov[0].iov_base = buf;
131 iov[0].iov_len = *len;
132 return 1;
133 }
134
135 static struct io_uring_buf *io_ring_head_to_buf(struct io_uring_buf_ring *br,
136 __u16 head, __u16 mask)
137 {
138 return &br->bufs[head & mask];
139 }
140
141 static void __user *io_ring_buffer_select(struct io_kiocb *req, size_t *len,
142 struct io_buffer_list *bl,
143 unsigned int issue_flags)
144 {
145 struct io_uring_buf_ring *br = bl->buf_ring;
146 __u16 tail, head = bl->head;
147 struct io_uring_buf *buf;
148
149 tail = smp_load_acquire(&br->tail);
150 if (unlikely(tail == head))
151 return NULL;
152
153 if (head + 1 == tail)
154 req->flags |= REQ_F_BL_EMPTY;
155
156 buf = io_ring_head_to_buf(br, head, bl->mask);
157 if (*len == 0 || *len > buf->len)
158 *len = buf->len;
159 req->flags |= REQ_F_BUFFER_RING | REQ_F_BUFFERS_COMMIT;
160 req->buf_list = bl;
161 req->buf_index = buf->bid;
162
163 if (issue_flags & IO_URING_F_UNLOCKED || !io_file_can_poll(req)) {
164 /*
165 * If we came in unlocked, we have no choice but to consume the
166 * buffer here, otherwise nothing ensures that the buffer won't
167 * get used by others. This does mean it'll be pinned until the
168 * IO completes, coming in unlocked means we're being called from
169 * io-wq context and there may be further retries in async hybrid
170 * mode. For the locked case, the caller must call commit when
171 * the transfer completes (or if we get -EAGAIN and must poll of
172 * retry).
173 */
174 req->flags &= ~REQ_F_BUFFERS_COMMIT;
175 req->buf_list = NULL;
176 bl->head++;
177 }
178 return u64_to_user_ptr(buf->addr);
179 }
180
181 void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
182 unsigned int issue_flags)
183 {
184 struct io_ring_ctx *ctx = req->ctx;
185 struct io_buffer_list *bl;
186 void __user *ret = NULL;
187
188 io_ring_submit_lock(req->ctx, issue_flags);
189
190 bl = io_buffer_get_list(ctx, req->buf_index);
191 if (likely(bl)) {
192 if (bl->is_buf_ring)
193 ret = io_ring_buffer_select(req, len, bl, issue_flags);
194 else
195 ret = io_provided_buffer_select(req, len, bl);
196 }
197 io_ring_submit_unlock(req->ctx, issue_flags);
198 return ret;
199 }
200
201 /* cap it at a reasonable 256, will be one page even for 4K */
202 #define PEEK_MAX_IMPORT 256
203
204 static int io_ring_buffers_peek(struct io_kiocb *req, struct buf_sel_arg *arg,
205 struct io_buffer_list *bl)
206 {
207 struct io_uring_buf_ring *br = bl->buf_ring;
208 struct iovec *iov = arg->iovs;
209 int nr_iovs = arg->nr_iovs;
210 __u16 nr_avail, tail, head;
211 struct io_uring_buf *buf;
212
213 tail = smp_load_acquire(&br->tail);
214 head = bl->head;
215 nr_avail = min_t(__u16, tail - head, UIO_MAXIOV);
216 if (unlikely(!nr_avail))
217 return -ENOBUFS;
218
219 buf = io_ring_head_to_buf(br, head, bl->mask);
220 if (arg->max_len) {
221 u32 len = READ_ONCE(buf->len);
222 size_t needed;
223
224 if (unlikely(!len))
225 return -ENOBUFS;
226 needed = (arg->max_len + len - 1) / len;
227 needed = min_not_zero(needed, (size_t) PEEK_MAX_IMPORT);
228 if (nr_avail > needed)
229 nr_avail = needed;
230 }
231
232 /*
233 * only alloc a bigger array if we know we have data to map, eg not
234 * a speculative peek operation.
235 */
236 if (arg->mode & KBUF_MODE_EXPAND && nr_avail > nr_iovs && arg->max_len) {
237 iov = kmalloc_array(nr_avail, sizeof(struct iovec), GFP_KERNEL);
238 if (unlikely(!iov))
239 return -ENOMEM;
240 if (arg->mode & KBUF_MODE_FREE)
241 kfree(arg->iovs);
242 arg->iovs = iov;
243 nr_iovs = nr_avail;
244 } else if (nr_avail < nr_iovs) {
245 nr_iovs = nr_avail;
246 }
247
248 /* set it to max, if not set, so we can use it unconditionally */
249 if (!arg->max_len)
250 arg->max_len = INT_MAX;
251
252 req->buf_index = buf->bid;
253 do {
254 /* truncate end piece, if needed */
255 if (buf->len > arg->max_len)
256 buf->len = arg->max_len;
257
258 iov->iov_base = u64_to_user_ptr(buf->addr);
259 iov->iov_len = buf->len;
260 iov++;
261
262 arg->out_len += buf->len;
263 arg->max_len -= buf->len;
264 if (!arg->max_len)
265 break;
266
267 buf = io_ring_head_to_buf(br, ++head, bl->mask);
268 } while (--nr_iovs);
269
270 if (head == tail)
271 req->flags |= REQ_F_BL_EMPTY;
272
273 req->flags |= REQ_F_BUFFER_RING;
274 req->buf_list = bl;
275 return iov - arg->iovs;
276 }
277
278 int io_buffers_select(struct io_kiocb *req, struct buf_sel_arg *arg,
279 unsigned int issue_flags)
280 {
281 struct io_ring_ctx *ctx = req->ctx;
282 struct io_buffer_list *bl;
283 int ret = -ENOENT;
284
285 io_ring_submit_lock(ctx, issue_flags);
286 bl = io_buffer_get_list(ctx, req->buf_index);
287 if (unlikely(!bl))
288 goto out_unlock;
289
290 if (bl->is_buf_ring) {
291 ret = io_ring_buffers_peek(req, arg, bl);
292 /*
293 * Don't recycle these buffers if we need to go through poll.
294 * Nobody else can use them anyway, and holding on to provided
295 * buffers for a send/write operation would happen on the app
296 * side anyway with normal buffers. Besides, we already
297 * committed them, they cannot be put back in the queue.
298 */
299 if (ret > 0) {
300 req->flags |= REQ_F_BL_NO_RECYCLE;
301 req->buf_list->head += ret;
302 }
303 } else {
304 ret = io_provided_buffers_select(req, &arg->out_len, bl, arg->iovs);
305 }
306 out_unlock:
307 io_ring_submit_unlock(ctx, issue_flags);
308 return ret;
309 }
310
311 int io_buffers_peek(struct io_kiocb *req, struct buf_sel_arg *arg)
312 {
313 struct io_ring_ctx *ctx = req->ctx;
314 struct io_buffer_list *bl;
315 int ret;
316
317 lockdep_assert_held(&ctx->uring_lock);
318
319 bl = io_buffer_get_list(ctx, req->buf_index);
320 if (unlikely(!bl))
321 return -ENOENT;
322
323 if (bl->is_buf_ring) {
324 ret = io_ring_buffers_peek(req, arg, bl);
325 if (ret > 0)
326 req->flags |= REQ_F_BUFFERS_COMMIT;
327 return ret;
328 }
329
330 /* don't support multiple buffer selections for legacy */
331 return io_provided_buffers_select(req, &arg->max_len, bl, arg->iovs);
332 }
333
334 static int __io_remove_buffers(struct io_ring_ctx *ctx,
335 struct io_buffer_list *bl, unsigned nbufs)
336 {
337 unsigned i = 0;
338
339 /* shouldn't happen */
340 if (!nbufs)
341 return 0;
342
343 if (bl->is_buf_ring) {
344 i = bl->buf_ring->tail - bl->head;
345 if (bl->buf_nr_pages) {
346 int j;
347
348 if (!bl->is_mmap) {
349 for (j = 0; j < bl->buf_nr_pages; j++)
350 unpin_user_page(bl->buf_pages[j]);
351 }
352 io_pages_unmap(bl->buf_ring, &bl->buf_pages,
353 &bl->buf_nr_pages, bl->is_mmap);
354 bl->is_mmap = 0;
355 }
356 /* make sure it's seen as empty */
357 INIT_LIST_HEAD(&bl->buf_list);
358 bl->is_buf_ring = 0;
359 return i;
360 }
361
362 /* protects io_buffers_cache */
363 lockdep_assert_held(&ctx->uring_lock);
364
365 while (!list_empty(&bl->buf_list)) {
366 struct io_buffer *nxt;
367
368 nxt = list_first_entry(&bl->buf_list, struct io_buffer, list);
369 list_move(&nxt->list, &ctx->io_buffers_cache);
370 if (++i == nbufs)
371 return i;
372 cond_resched();
373 }
374
375 return i;
376 }
377
378 void io_put_bl(struct io_ring_ctx *ctx, struct io_buffer_list *bl)
379 {
380 if (atomic_dec_and_test(&bl->refs)) {
381 __io_remove_buffers(ctx, bl, -1U);
382 kfree_rcu(bl, rcu);
383 }
384 }
385
386 void io_destroy_buffers(struct io_ring_ctx *ctx)
387 {
388 struct io_buffer_list *bl;
389 struct list_head *item, *tmp;
390 struct io_buffer *buf;
391 unsigned long index;
392
393 xa_for_each(&ctx->io_bl_xa, index, bl) {
394 xa_erase(&ctx->io_bl_xa, bl->bgid);
395 io_put_bl(ctx, bl);
396 }
397
398 /*
399 * Move deferred locked entries to cache before pruning
400 */
401 spin_lock(&ctx->completion_lock);
402 if (!list_empty(&ctx->io_buffers_comp))
403 list_splice_init(&ctx->io_buffers_comp, &ctx->io_buffers_cache);
404 spin_unlock(&ctx->completion_lock);
405
406 list_for_each_safe(item, tmp, &ctx->io_buffers_cache) {
407 buf = list_entry(item, struct io_buffer, list);
408 kmem_cache_free(io_buf_cachep, buf);
409 }
410 }
411
412 int io_remove_buffers_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
413 {
414 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
415 u64 tmp;
416
417 if (sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
418 sqe->splice_fd_in)
419 return -EINVAL;
420
421 tmp = READ_ONCE(sqe->fd);
422 if (!tmp || tmp > MAX_BIDS_PER_BGID)
423 return -EINVAL;
424
425 memset(p, 0, sizeof(*p));
426 p->nbufs = tmp;
427 p->bgid = READ_ONCE(sqe->buf_group);
428 return 0;
429 }
430
431 int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
432 {
433 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
434 struct io_ring_ctx *ctx = req->ctx;
435 struct io_buffer_list *bl;
436 int ret = 0;
437
438 io_ring_submit_lock(ctx, issue_flags);
439
440 ret = -ENOENT;
441 bl = io_buffer_get_list(ctx, p->bgid);
442 if (bl) {
443 ret = -EINVAL;
444 /* can't use provide/remove buffers command on mapped buffers */
445 if (!bl->is_buf_ring)
446 ret = __io_remove_buffers(ctx, bl, p->nbufs);
447 }
448 io_ring_submit_unlock(ctx, issue_flags);
449 if (ret < 0)
450 req_set_fail(req);
451 io_req_set_res(req, ret, 0);
452 return IOU_OK;
453 }
454
455 int io_provide_buffers_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
456 {
457 unsigned long size, tmp_check;
458 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
459 u64 tmp;
460
461 if (sqe->rw_flags || sqe->splice_fd_in)
462 return -EINVAL;
463
464 tmp = READ_ONCE(sqe->fd);
465 if (!tmp || tmp > MAX_BIDS_PER_BGID)
466 return -E2BIG;
467 p->nbufs = tmp;
468 p->addr = READ_ONCE(sqe->addr);
469 p->len = READ_ONCE(sqe->len);
470
471 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
472 &size))
473 return -EOVERFLOW;
474 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
475 return -EOVERFLOW;
476
477 size = (unsigned long)p->len * p->nbufs;
478 if (!access_ok(u64_to_user_ptr(p->addr), size))
479 return -EFAULT;
480
481 p->bgid = READ_ONCE(sqe->buf_group);
482 tmp = READ_ONCE(sqe->off);
483 if (tmp > USHRT_MAX)
484 return -E2BIG;
485 if (tmp + p->nbufs > MAX_BIDS_PER_BGID)
486 return -EINVAL;
487 p->bid = tmp;
488 return 0;
489 }
490
491 #define IO_BUFFER_ALLOC_BATCH 64
492
493 static int io_refill_buffer_cache(struct io_ring_ctx *ctx)
494 {
495 struct io_buffer *bufs[IO_BUFFER_ALLOC_BATCH];
496 int allocated;
497
498 /*
499 * Completions that don't happen inline (eg not under uring_lock) will
500 * add to ->io_buffers_comp. If we don't have any free buffers, check
501 * the completion list and splice those entries first.
502 */
503 if (!list_empty_careful(&ctx->io_buffers_comp)) {
504 spin_lock(&ctx->completion_lock);
505 if (!list_empty(&ctx->io_buffers_comp)) {
506 list_splice_init(&ctx->io_buffers_comp,
507 &ctx->io_buffers_cache);
508 spin_unlock(&ctx->completion_lock);
509 return 0;
510 }
511 spin_unlock(&ctx->completion_lock);
512 }
513
514 /*
515 * No free buffers and no completion entries either. Allocate a new
516 * batch of buffer entries and add those to our freelist.
517 */
518
519 allocated = kmem_cache_alloc_bulk(io_buf_cachep, GFP_KERNEL_ACCOUNT,
520 ARRAY_SIZE(bufs), (void **) bufs);
521 if (unlikely(!allocated)) {
522 /*
523 * Bulk alloc is all-or-nothing. If we fail to get a batch,
524 * retry single alloc to be on the safe side.
525 */
526 bufs[0] = kmem_cache_alloc(io_buf_cachep, GFP_KERNEL);
527 if (!bufs[0])
528 return -ENOMEM;
529 allocated = 1;
530 }
531
532 while (allocated)
533 list_add_tail(&bufs[--allocated]->list, &ctx->io_buffers_cache);
534
535 return 0;
536 }
537
538 static int io_add_buffers(struct io_ring_ctx *ctx, struct io_provide_buf *pbuf,
539 struct io_buffer_list *bl)
540 {
541 struct io_buffer *buf;
542 u64 addr = pbuf->addr;
543 int i, bid = pbuf->bid;
544
545 for (i = 0; i < pbuf->nbufs; i++) {
546 if (list_empty(&ctx->io_buffers_cache) &&
547 io_refill_buffer_cache(ctx))
548 break;
549 buf = list_first_entry(&ctx->io_buffers_cache, struct io_buffer,
550 list);
551 list_move_tail(&buf->list, &bl->buf_list);
552 buf->addr = addr;
553 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
554 buf->bid = bid;
555 buf->bgid = pbuf->bgid;
556 addr += pbuf->len;
557 bid++;
558 cond_resched();
559 }
560
561 return i ? 0 : -ENOMEM;
562 }
563
564 int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
565 {
566 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
567 struct io_ring_ctx *ctx = req->ctx;
568 struct io_buffer_list *bl;
569 int ret = 0;
570
571 io_ring_submit_lock(ctx, issue_flags);
572
573 bl = io_buffer_get_list(ctx, p->bgid);
574 if (unlikely(!bl)) {
575 bl = kzalloc(sizeof(*bl), GFP_KERNEL_ACCOUNT);
576 if (!bl) {
577 ret = -ENOMEM;
578 goto err;
579 }
580 INIT_LIST_HEAD(&bl->buf_list);
581 ret = io_buffer_add_list(ctx, bl, p->bgid);
582 if (ret) {
583 /*
584 * Doesn't need rcu free as it was never visible, but
585 * let's keep it consistent throughout.
586 */
587 kfree_rcu(bl, rcu);
588 goto err;
589 }
590 }
591 /* can't add buffers via this command for a mapped buffer ring */
592 if (bl->is_buf_ring) {
593 ret = -EINVAL;
594 goto err;
595 }
596
597 ret = io_add_buffers(ctx, p, bl);
598 err:
599 io_ring_submit_unlock(ctx, issue_flags);
600
601 if (ret < 0)
602 req_set_fail(req);
603 io_req_set_res(req, ret, 0);
604 return IOU_OK;
605 }
606
607 static int io_pin_pbuf_ring(struct io_uring_buf_reg *reg,
608 struct io_buffer_list *bl)
609 {
610 struct io_uring_buf_ring *br = NULL;
611 struct page **pages;
612 int nr_pages, ret;
613
614 pages = io_pin_pages(reg->ring_addr,
615 flex_array_size(br, bufs, reg->ring_entries),
616 &nr_pages);
617 if (IS_ERR(pages))
618 return PTR_ERR(pages);
619
620 br = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
621 if (!br) {
622 ret = -ENOMEM;
623 goto error_unpin;
624 }
625
626 #ifdef SHM_COLOUR
627 /*
628 * On platforms that have specific aliasing requirements, SHM_COLOUR
629 * is set and we must guarantee that the kernel and user side align
630 * nicely. We cannot do that if IOU_PBUF_RING_MMAP isn't set and
631 * the application mmap's the provided ring buffer. Fail the request
632 * if we, by chance, don't end up with aligned addresses. The app
633 * should use IOU_PBUF_RING_MMAP instead, and liburing will handle
634 * this transparently.
635 */
636 if ((reg->ring_addr | (unsigned long) br) & (SHM_COLOUR - 1)) {
637 ret = -EINVAL;
638 goto error_unpin;
639 }
640 #endif
641 bl->buf_pages = pages;
642 bl->buf_nr_pages = nr_pages;
643 bl->buf_ring = br;
644 bl->is_buf_ring = 1;
645 bl->is_mmap = 0;
646 return 0;
647 error_unpin:
648 unpin_user_pages(pages, nr_pages);
649 kvfree(pages);
650 vunmap(br);
651 return ret;
652 }
653
654 static int io_alloc_pbuf_ring(struct io_ring_ctx *ctx,
655 struct io_uring_buf_reg *reg,
656 struct io_buffer_list *bl)
657 {
658 size_t ring_size;
659
660 ring_size = reg->ring_entries * sizeof(struct io_uring_buf_ring);
661
662 bl->buf_ring = io_pages_map(&bl->buf_pages, &bl->buf_nr_pages, ring_size);
663 if (IS_ERR(bl->buf_ring)) {
664 bl->buf_ring = NULL;
665 return -ENOMEM;
666 }
667
668 bl->is_buf_ring = 1;
669 bl->is_mmap = 1;
670 return 0;
671 }
672
673 int io_register_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
674 {
675 struct io_uring_buf_reg reg;
676 struct io_buffer_list *bl, *free_bl = NULL;
677 int ret;
678
679 lockdep_assert_held(&ctx->uring_lock);
680
681 if (copy_from_user(®, arg, sizeof(reg)))
682 return -EFAULT;
683
684 if (reg.resv[0] || reg.resv[1] || reg.resv[2])
685 return -EINVAL;
686 if (reg.flags & ~IOU_PBUF_RING_MMAP)
687 return -EINVAL;
688 if (!(reg.flags & IOU_PBUF_RING_MMAP)) {
689 if (!reg.ring_addr)
690 return -EFAULT;
691 if (reg.ring_addr & ~PAGE_MASK)
692 return -EINVAL;
693 } else {
694 if (reg.ring_addr)
695 return -EINVAL;
696 }
697
698 if (!is_power_of_2(reg.ring_entries))
699 return -EINVAL;
700
701 /* cannot disambiguate full vs empty due to head/tail size */
702 if (reg.ring_entries >= 65536)
703 return -EINVAL;
704
705 bl = io_buffer_get_list(ctx, reg.bgid);
706 if (bl) {
707 /* if mapped buffer ring OR classic exists, don't allow */
708 if (bl->is_buf_ring || !list_empty(&bl->buf_list))
709 return -EEXIST;
710 } else {
711 free_bl = bl = kzalloc(sizeof(*bl), GFP_KERNEL);
712 if (!bl)
713 return -ENOMEM;
714 }
715
716 if (!(reg.flags & IOU_PBUF_RING_MMAP))
717 ret = io_pin_pbuf_ring(®, bl);
718 else
719 ret = io_alloc_pbuf_ring(ctx, ®, bl);
720
721 if (!ret) {
722 bl->nr_entries = reg.ring_entries;
723 bl->mask = reg.ring_entries - 1;
724
725 io_buffer_add_list(ctx, bl, reg.bgid);
726 return 0;
727 }
728
729 kfree_rcu(free_bl, rcu);
730 return ret;
731 }
732
733 int io_unregister_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
734 {
735 struct io_uring_buf_reg reg;
736 struct io_buffer_list *bl;
737
738 lockdep_assert_held(&ctx->uring_lock);
739
740 if (copy_from_user(®, arg, sizeof(reg)))
741 return -EFAULT;
742 if (reg.resv[0] || reg.resv[1] || reg.resv[2])
743 return -EINVAL;
744 if (reg.flags)
745 return -EINVAL;
746
747 bl = io_buffer_get_list(ctx, reg.bgid);
748 if (!bl)
749 return -ENOENT;
750 if (!bl->is_buf_ring)
751 return -EINVAL;
752
753 xa_erase(&ctx->io_bl_xa, bl->bgid);
754 io_put_bl(ctx, bl);
755 return 0;
756 }
757
758 int io_register_pbuf_status(struct io_ring_ctx *ctx, void __user *arg)
759 {
760 struct io_uring_buf_status buf_status;
761 struct io_buffer_list *bl;
762 int i;
763
764 if (copy_from_user(&buf_status, arg, sizeof(buf_status)))
765 return -EFAULT;
766
767 for (i = 0; i < ARRAY_SIZE(buf_status.resv); i++)
768 if (buf_status.resv[i])
769 return -EINVAL;
770
771 bl = io_buffer_get_list(ctx, buf_status.buf_group);
772 if (!bl)
773 return -ENOENT;
774 if (!bl->is_buf_ring)
775 return -EINVAL;
776
777 buf_status.head = bl->head;
778 if (copy_to_user(arg, &buf_status, sizeof(buf_status)))
779 return -EFAULT;
780
781 return 0;
782 }
783
784 struct io_buffer_list *io_pbuf_get_bl(struct io_ring_ctx *ctx,
785 unsigned long bgid)
786 {
787 struct io_buffer_list *bl;
788 bool ret;
789
790 /*
791 * We have to be a bit careful here - we're inside mmap and cannot grab
792 * the uring_lock. This means the buffer_list could be simultaneously
793 * going away, if someone is trying to be sneaky. Look it up under rcu
794 * so we know it's not going away, and attempt to grab a reference to
795 * it. If the ref is already zero, then fail the mapping. If successful,
796 * the caller will call io_put_bl() to drop the the reference at at the
797 * end. This may then safely free the buffer_list (and drop the pages)
798 * at that point, vm_insert_pages() would've already grabbed the
799 * necessary vma references.
800 */
801 rcu_read_lock();
802 bl = xa_load(&ctx->io_bl_xa, bgid);
803 /* must be a mmap'able buffer ring and have pages */
804 ret = false;
805 if (bl && bl->is_mmap)
806 ret = atomic_inc_not_zero(&bl->refs);
807 rcu_read_unlock();
808
809 if (ret)
810 return bl;
811
812 return ERR_PTR(-EINVAL);
813 }
814
815 int io_pbuf_mmap(struct file *file, struct vm_area_struct *vma)
816 {
817 struct io_ring_ctx *ctx = file->private_data;
818 loff_t pgoff = vma->vm_pgoff << PAGE_SHIFT;
819 struct io_buffer_list *bl;
820 int bgid, ret;
821
822 bgid = (pgoff & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
823 bl = io_pbuf_get_bl(ctx, bgid);
824 if (IS_ERR(bl))
825 return PTR_ERR(bl);
826
827 ret = io_uring_mmap_pages(ctx, vma, bl->buf_pages, bl->buf_nr_pages);
828 io_put_bl(ctx, bl);
829 return ret;
830 }
831
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.