Loading...
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
23struct kmem_cache *io_buf_cachep;
24
25struct io_provide_buf {
26 struct file *file;
27 __u64 addr;
28 __u32 len;
29 __u32 bgid;
30 __u32 nbufs;
31 __u16 bid;
32};
33
34static 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
42static 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
55bool 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
73void __io_put_kbuf(struct io_kiocb *req, int len, 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, len, &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, len, &req->ctx->io_buffers_cache);
97 }
98}
99
100static 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
120static 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
135static void __user *io_ring_buffer_select(struct io_kiocb *req, size_t *len,
136 struct io_buffer_list *bl,
137 unsigned int issue_flags)
138{
139 struct io_uring_buf_ring *br = bl->buf_ring;
140 __u16 tail, head = bl->head;
141 struct io_uring_buf *buf;
142 void __user *ret;
143
144 tail = smp_load_acquire(&br->tail);
145 if (unlikely(tail == head))
146 return NULL;
147
148 if (head + 1 == tail)
149 req->flags |= REQ_F_BL_EMPTY;
150
151 buf = io_ring_head_to_buf(br, head, bl->mask);
152 if (*len == 0 || *len > buf->len)
153 *len = buf->len;
154 req->flags |= REQ_F_BUFFER_RING | REQ_F_BUFFERS_COMMIT;
155 req->buf_list = bl;
156 req->buf_index = buf->bid;
157 ret = u64_to_user_ptr(buf->addr);
158
159 if (issue_flags & IO_URING_F_UNLOCKED || !io_file_can_poll(req)) {
160 /*
161 * If we came in unlocked, we have no choice but to consume the
162 * buffer here, otherwise nothing ensures that the buffer won't
163 * get used by others. This does mean it'll be pinned until the
164 * IO completes, coming in unlocked means we're being called from
165 * io-wq context and there may be further retries in async hybrid
166 * mode. For the locked case, the caller must call commit when
167 * the transfer completes (or if we get -EAGAIN and must poll of
168 * retry).
169 */
170 io_kbuf_commit(req, bl, *len, 1);
171 req->buf_list = NULL;
172 }
173 return ret;
174}
175
176void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
177 unsigned int issue_flags)
178{
179 struct io_ring_ctx *ctx = req->ctx;
180 struct io_buffer_list *bl;
181 void __user *ret = NULL;
182
183 io_ring_submit_lock(req->ctx, issue_flags);
184
185 bl = io_buffer_get_list(ctx, req->buf_index);
186 if (likely(bl)) {
187 if (bl->flags & IOBL_BUF_RING)
188 ret = io_ring_buffer_select(req, len, bl, issue_flags);
189 else
190 ret = io_provided_buffer_select(req, len, bl);
191 }
192 io_ring_submit_unlock(req->ctx, issue_flags);
193 return ret;
194}
195
196/* cap it at a reasonable 256, will be one page even for 4K */
197#define PEEK_MAX_IMPORT 256
198
199static int io_ring_buffers_peek(struct io_kiocb *req, struct buf_sel_arg *arg,
200 struct io_buffer_list *bl)
201{
202 struct io_uring_buf_ring *br = bl->buf_ring;
203 struct iovec *iov = arg->iovs;
204 int nr_iovs = arg->nr_iovs;
205 __u16 nr_avail, tail, head;
206 struct io_uring_buf *buf;
207
208 tail = smp_load_acquire(&br->tail);
209 head = bl->head;
210 nr_avail = min_t(__u16, tail - head, UIO_MAXIOV);
211 if (unlikely(!nr_avail))
212 return -ENOBUFS;
213
214 buf = io_ring_head_to_buf(br, head, bl->mask);
215 if (arg->max_len) {
216 u32 len = READ_ONCE(buf->len);
217
218 if (unlikely(!len))
219 return -ENOBUFS;
220 /*
221 * Limit incremental buffers to 1 segment. No point trying
222 * to peek ahead and map more than we need, when the buffers
223 * themselves should be large when setup with
224 * IOU_PBUF_RING_INC.
225 */
226 if (bl->flags & IOBL_INC) {
227 nr_avail = 1;
228 } else {
229 size_t needed;
230
231 needed = (arg->max_len + len - 1) / len;
232 needed = min_not_zero(needed, (size_t) PEEK_MAX_IMPORT);
233 if (nr_avail > needed)
234 nr_avail = needed;
235 }
236 }
237
238 /*
239 * only alloc a bigger array if we know we have data to map, eg not
240 * a speculative peek operation.
241 */
242 if (arg->mode & KBUF_MODE_EXPAND && nr_avail > nr_iovs && arg->max_len) {
243 iov = kmalloc_array(nr_avail, sizeof(struct iovec), GFP_KERNEL);
244 if (unlikely(!iov))
245 return -ENOMEM;
246 if (arg->mode & KBUF_MODE_FREE)
247 kfree(arg->iovs);
248 arg->iovs = iov;
249 nr_iovs = nr_avail;
250 } else if (nr_avail < nr_iovs) {
251 nr_iovs = nr_avail;
252 }
253
254 /* set it to max, if not set, so we can use it unconditionally */
255 if (!arg->max_len)
256 arg->max_len = INT_MAX;
257
258 req->buf_index = buf->bid;
259 do {
260 u32 len = buf->len;
261
262 /* truncate end piece, if needed, for non partial buffers */
263 if (len > arg->max_len) {
264 len = arg->max_len;
265 if (!(bl->flags & IOBL_INC))
266 buf->len = len;
267 }
268
269 iov->iov_base = u64_to_user_ptr(buf->addr);
270 iov->iov_len = len;
271 iov++;
272
273 arg->out_len += len;
274 arg->max_len -= len;
275 if (!arg->max_len)
276 break;
277
278 buf = io_ring_head_to_buf(br, ++head, bl->mask);
279 } while (--nr_iovs);
280
281 if (head == tail)
282 req->flags |= REQ_F_BL_EMPTY;
283
284 req->flags |= REQ_F_BUFFER_RING;
285 req->buf_list = bl;
286 return iov - arg->iovs;
287}
288
289int io_buffers_select(struct io_kiocb *req, struct buf_sel_arg *arg,
290 unsigned int issue_flags)
291{
292 struct io_ring_ctx *ctx = req->ctx;
293 struct io_buffer_list *bl;
294 int ret = -ENOENT;
295
296 io_ring_submit_lock(ctx, issue_flags);
297 bl = io_buffer_get_list(ctx, req->buf_index);
298 if (unlikely(!bl))
299 goto out_unlock;
300
301 if (bl->flags & IOBL_BUF_RING) {
302 ret = io_ring_buffers_peek(req, arg, bl);
303 /*
304 * Don't recycle these buffers if we need to go through poll.
305 * Nobody else can use them anyway, and holding on to provided
306 * buffers for a send/write operation would happen on the app
307 * side anyway with normal buffers. Besides, we already
308 * committed them, they cannot be put back in the queue.
309 */
310 if (ret > 0) {
311 req->flags |= REQ_F_BUFFERS_COMMIT | REQ_F_BL_NO_RECYCLE;
312 io_kbuf_commit(req, bl, arg->out_len, ret);
313 }
314 } else {
315 ret = io_provided_buffers_select(req, &arg->out_len, bl, arg->iovs);
316 }
317out_unlock:
318 io_ring_submit_unlock(ctx, issue_flags);
319 return ret;
320}
321
322int io_buffers_peek(struct io_kiocb *req, struct buf_sel_arg *arg)
323{
324 struct io_ring_ctx *ctx = req->ctx;
325 struct io_buffer_list *bl;
326 int ret;
327
328 lockdep_assert_held(&ctx->uring_lock);
329
330 bl = io_buffer_get_list(ctx, req->buf_index);
331 if (unlikely(!bl))
332 return -ENOENT;
333
334 if (bl->flags & IOBL_BUF_RING) {
335 ret = io_ring_buffers_peek(req, arg, bl);
336 if (ret > 0)
337 req->flags |= REQ_F_BUFFERS_COMMIT;
338 return ret;
339 }
340
341 /* don't support multiple buffer selections for legacy */
342 return io_provided_buffers_select(req, &arg->max_len, bl, arg->iovs);
343}
344
345static int __io_remove_buffers(struct io_ring_ctx *ctx,
346 struct io_buffer_list *bl, unsigned nbufs)
347{
348 unsigned i = 0;
349
350 /* shouldn't happen */
351 if (!nbufs)
352 return 0;
353
354 if (bl->flags & IOBL_BUF_RING) {
355 i = bl->buf_ring->tail - bl->head;
356 if (bl->buf_nr_pages) {
357 int j;
358
359 if (!(bl->flags & IOBL_MMAP)) {
360 for (j = 0; j < bl->buf_nr_pages; j++)
361 unpin_user_page(bl->buf_pages[j]);
362 }
363 io_pages_unmap(bl->buf_ring, &bl->buf_pages,
364 &bl->buf_nr_pages, bl->flags & IOBL_MMAP);
365 bl->flags &= ~IOBL_MMAP;
366 }
367 /* make sure it's seen as empty */
368 INIT_LIST_HEAD(&bl->buf_list);
369 bl->flags &= ~IOBL_BUF_RING;
370 return i;
371 }
372
373 /* protects io_buffers_cache */
374 lockdep_assert_held(&ctx->uring_lock);
375
376 while (!list_empty(&bl->buf_list)) {
377 struct io_buffer *nxt;
378
379 nxt = list_first_entry(&bl->buf_list, struct io_buffer, list);
380 list_move(&nxt->list, &ctx->io_buffers_cache);
381 if (++i == nbufs)
382 return i;
383 cond_resched();
384 }
385
386 return i;
387}
388
389void io_put_bl(struct io_ring_ctx *ctx, struct io_buffer_list *bl)
390{
391 if (atomic_dec_and_test(&bl->refs)) {
392 __io_remove_buffers(ctx, bl, -1U);
393 kfree_rcu(bl, rcu);
394 }
395}
396
397void io_destroy_buffers(struct io_ring_ctx *ctx)
398{
399 struct io_buffer_list *bl;
400 struct list_head *item, *tmp;
401 struct io_buffer *buf;
402 unsigned long index;
403
404 xa_for_each(&ctx->io_bl_xa, index, bl) {
405 xa_erase(&ctx->io_bl_xa, bl->bgid);
406 io_put_bl(ctx, bl);
407 }
408
409 /*
410 * Move deferred locked entries to cache before pruning
411 */
412 spin_lock(&ctx->completion_lock);
413 if (!list_empty(&ctx->io_buffers_comp))
414 list_splice_init(&ctx->io_buffers_comp, &ctx->io_buffers_cache);
415 spin_unlock(&ctx->completion_lock);
416
417 list_for_each_safe(item, tmp, &ctx->io_buffers_cache) {
418 buf = list_entry(item, struct io_buffer, list);
419 kmem_cache_free(io_buf_cachep, buf);
420 }
421}
422
423static void io_destroy_bl(struct io_ring_ctx *ctx, struct io_buffer_list *bl)
424{
425 xa_erase(&ctx->io_bl_xa, bl->bgid);
426 io_put_bl(ctx, bl);
427}
428
429int io_remove_buffers_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
430{
431 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
432 u64 tmp;
433
434 if (sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
435 sqe->splice_fd_in)
436 return -EINVAL;
437
438 tmp = READ_ONCE(sqe->fd);
439 if (!tmp || tmp > MAX_BIDS_PER_BGID)
440 return -EINVAL;
441
442 memset(p, 0, sizeof(*p));
443 p->nbufs = tmp;
444 p->bgid = READ_ONCE(sqe->buf_group);
445 return 0;
446}
447
448int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
449{
450 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
451 struct io_ring_ctx *ctx = req->ctx;
452 struct io_buffer_list *bl;
453 int ret = 0;
454
455 io_ring_submit_lock(ctx, issue_flags);
456
457 ret = -ENOENT;
458 bl = io_buffer_get_list(ctx, p->bgid);
459 if (bl) {
460 ret = -EINVAL;
461 /* can't use provide/remove buffers command on mapped buffers */
462 if (!(bl->flags & IOBL_BUF_RING))
463 ret = __io_remove_buffers(ctx, bl, p->nbufs);
464 }
465 io_ring_submit_unlock(ctx, issue_flags);
466 if (ret < 0)
467 req_set_fail(req);
468 io_req_set_res(req, ret, 0);
469 return IOU_OK;
470}
471
472int io_provide_buffers_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
473{
474 unsigned long size, tmp_check;
475 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
476 u64 tmp;
477
478 if (sqe->rw_flags || sqe->splice_fd_in)
479 return -EINVAL;
480
481 tmp = READ_ONCE(sqe->fd);
482 if (!tmp || tmp > MAX_BIDS_PER_BGID)
483 return -E2BIG;
484 p->nbufs = tmp;
485 p->addr = READ_ONCE(sqe->addr);
486 p->len = READ_ONCE(sqe->len);
487
488 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
489 &size))
490 return -EOVERFLOW;
491 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
492 return -EOVERFLOW;
493
494 size = (unsigned long)p->len * p->nbufs;
495 if (!access_ok(u64_to_user_ptr(p->addr), size))
496 return -EFAULT;
497
498 p->bgid = READ_ONCE(sqe->buf_group);
499 tmp = READ_ONCE(sqe->off);
500 if (tmp > USHRT_MAX)
501 return -E2BIG;
502 if (tmp + p->nbufs > MAX_BIDS_PER_BGID)
503 return -EINVAL;
504 p->bid = tmp;
505 return 0;
506}
507
508#define IO_BUFFER_ALLOC_BATCH 64
509
510static int io_refill_buffer_cache(struct io_ring_ctx *ctx)
511{
512 struct io_buffer *bufs[IO_BUFFER_ALLOC_BATCH];
513 int allocated;
514
515 /*
516 * Completions that don't happen inline (eg not under uring_lock) will
517 * add to ->io_buffers_comp. If we don't have any free buffers, check
518 * the completion list and splice those entries first.
519 */
520 if (!list_empty_careful(&ctx->io_buffers_comp)) {
521 spin_lock(&ctx->completion_lock);
522 if (!list_empty(&ctx->io_buffers_comp)) {
523 list_splice_init(&ctx->io_buffers_comp,
524 &ctx->io_buffers_cache);
525 spin_unlock(&ctx->completion_lock);
526 return 0;
527 }
528 spin_unlock(&ctx->completion_lock);
529 }
530
531 /*
532 * No free buffers and no completion entries either. Allocate a new
533 * batch of buffer entries and add those to our freelist.
534 */
535
536 allocated = kmem_cache_alloc_bulk(io_buf_cachep, GFP_KERNEL_ACCOUNT,
537 ARRAY_SIZE(bufs), (void **) bufs);
538 if (unlikely(!allocated)) {
539 /*
540 * Bulk alloc is all-or-nothing. If we fail to get a batch,
541 * retry single alloc to be on the safe side.
542 */
543 bufs[0] = kmem_cache_alloc(io_buf_cachep, GFP_KERNEL);
544 if (!bufs[0])
545 return -ENOMEM;
546 allocated = 1;
547 }
548
549 while (allocated)
550 list_add_tail(&bufs[--allocated]->list, &ctx->io_buffers_cache);
551
552 return 0;
553}
554
555static int io_add_buffers(struct io_ring_ctx *ctx, struct io_provide_buf *pbuf,
556 struct io_buffer_list *bl)
557{
558 struct io_buffer *buf;
559 u64 addr = pbuf->addr;
560 int i, bid = pbuf->bid;
561
562 for (i = 0; i < pbuf->nbufs; i++) {
563 if (list_empty(&ctx->io_buffers_cache) &&
564 io_refill_buffer_cache(ctx))
565 break;
566 buf = list_first_entry(&ctx->io_buffers_cache, struct io_buffer,
567 list);
568 list_move_tail(&buf->list, &bl->buf_list);
569 buf->addr = addr;
570 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
571 buf->bid = bid;
572 buf->bgid = pbuf->bgid;
573 addr += pbuf->len;
574 bid++;
575 cond_resched();
576 }
577
578 return i ? 0 : -ENOMEM;
579}
580
581int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
582{
583 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
584 struct io_ring_ctx *ctx = req->ctx;
585 struct io_buffer_list *bl;
586 int ret = 0;
587
588 io_ring_submit_lock(ctx, issue_flags);
589
590 bl = io_buffer_get_list(ctx, p->bgid);
591 if (unlikely(!bl)) {
592 bl = kzalloc(sizeof(*bl), GFP_KERNEL_ACCOUNT);
593 if (!bl) {
594 ret = -ENOMEM;
595 goto err;
596 }
597 INIT_LIST_HEAD(&bl->buf_list);
598 ret = io_buffer_add_list(ctx, bl, p->bgid);
599 if (ret) {
600 /*
601 * Doesn't need rcu free as it was never visible, but
602 * let's keep it consistent throughout.
603 */
604 kfree_rcu(bl, rcu);
605 goto err;
606 }
607 }
608 /* can't add buffers via this command for a mapped buffer ring */
609 if (bl->flags & IOBL_BUF_RING) {
610 ret = -EINVAL;
611 goto err;
612 }
613
614 ret = io_add_buffers(ctx, p, bl);
615err:
616 io_ring_submit_unlock(ctx, issue_flags);
617
618 if (ret < 0)
619 req_set_fail(req);
620 io_req_set_res(req, ret, 0);
621 return IOU_OK;
622}
623
624static int io_pin_pbuf_ring(struct io_uring_buf_reg *reg,
625 struct io_buffer_list *bl)
626{
627 struct io_uring_buf_ring *br = NULL;
628 struct page **pages;
629 int nr_pages, ret;
630
631 pages = io_pin_pages(reg->ring_addr,
632 flex_array_size(br, bufs, reg->ring_entries),
633 &nr_pages);
634 if (IS_ERR(pages))
635 return PTR_ERR(pages);
636
637 br = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
638 if (!br) {
639 ret = -ENOMEM;
640 goto error_unpin;
641 }
642
643#ifdef SHM_COLOUR
644 /*
645 * On platforms that have specific aliasing requirements, SHM_COLOUR
646 * is set and we must guarantee that the kernel and user side align
647 * nicely. We cannot do that if IOU_PBUF_RING_MMAP isn't set and
648 * the application mmap's the provided ring buffer. Fail the request
649 * if we, by chance, don't end up with aligned addresses. The app
650 * should use IOU_PBUF_RING_MMAP instead, and liburing will handle
651 * this transparently.
652 */
653 if ((reg->ring_addr | (unsigned long) br) & (SHM_COLOUR - 1)) {
654 ret = -EINVAL;
655 goto error_unpin;
656 }
657#endif
658 bl->buf_pages = pages;
659 bl->buf_nr_pages = nr_pages;
660 bl->buf_ring = br;
661 bl->flags |= IOBL_BUF_RING;
662 bl->flags &= ~IOBL_MMAP;
663 return 0;
664error_unpin:
665 unpin_user_pages(pages, nr_pages);
666 kvfree(pages);
667 vunmap(br);
668 return ret;
669}
670
671static int io_alloc_pbuf_ring(struct io_ring_ctx *ctx,
672 struct io_uring_buf_reg *reg,
673 struct io_buffer_list *bl)
674{
675 size_t ring_size;
676
677 ring_size = reg->ring_entries * sizeof(struct io_uring_buf_ring);
678
679 bl->buf_ring = io_pages_map(&bl->buf_pages, &bl->buf_nr_pages, ring_size);
680 if (IS_ERR(bl->buf_ring)) {
681 bl->buf_ring = NULL;
682 return -ENOMEM;
683 }
684
685 bl->flags |= (IOBL_BUF_RING | IOBL_MMAP);
686 return 0;
687}
688
689int io_register_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
690{
691 struct io_uring_buf_reg reg;
692 struct io_buffer_list *bl, *free_bl = NULL;
693 int ret;
694
695 lockdep_assert_held(&ctx->uring_lock);
696
697 if (copy_from_user(®, arg, sizeof(reg)))
698 return -EFAULT;
699
700 if (reg.resv[0] || reg.resv[1] || reg.resv[2])
701 return -EINVAL;
702 if (reg.flags & ~(IOU_PBUF_RING_MMAP | IOU_PBUF_RING_INC))
703 return -EINVAL;
704 if (!(reg.flags & IOU_PBUF_RING_MMAP)) {
705 if (!reg.ring_addr)
706 return -EFAULT;
707 if (reg.ring_addr & ~PAGE_MASK)
708 return -EINVAL;
709 } else {
710 if (reg.ring_addr)
711 return -EINVAL;
712 }
713
714 if (!is_power_of_2(reg.ring_entries))
715 return -EINVAL;
716
717 /* cannot disambiguate full vs empty due to head/tail size */
718 if (reg.ring_entries >= 65536)
719 return -EINVAL;
720
721 bl = io_buffer_get_list(ctx, reg.bgid);
722 if (bl) {
723 /* if mapped buffer ring OR classic exists, don't allow */
724 if (bl->flags & IOBL_BUF_RING || !list_empty(&bl->buf_list))
725 return -EEXIST;
726 io_destroy_bl(ctx, bl);
727 }
728
729 free_bl = bl = kzalloc(sizeof(*bl), GFP_KERNEL);
730 if (!bl)
731 return -ENOMEM;
732
733 if (!(reg.flags & IOU_PBUF_RING_MMAP))
734 ret = io_pin_pbuf_ring(®, bl);
735 else
736 ret = io_alloc_pbuf_ring(ctx, ®, bl);
737
738 if (!ret) {
739 bl->nr_entries = reg.ring_entries;
740 bl->mask = reg.ring_entries - 1;
741 if (reg.flags & IOU_PBUF_RING_INC)
742 bl->flags |= IOBL_INC;
743
744 io_buffer_add_list(ctx, bl, reg.bgid);
745 return 0;
746 }
747
748 kfree_rcu(free_bl, rcu);
749 return ret;
750}
751
752int io_unregister_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
753{
754 struct io_uring_buf_reg reg;
755 struct io_buffer_list *bl;
756
757 lockdep_assert_held(&ctx->uring_lock);
758
759 if (copy_from_user(®, arg, sizeof(reg)))
760 return -EFAULT;
761 if (reg.resv[0] || reg.resv[1] || reg.resv[2])
762 return -EINVAL;
763 if (reg.flags)
764 return -EINVAL;
765
766 bl = io_buffer_get_list(ctx, reg.bgid);
767 if (!bl)
768 return -ENOENT;
769 if (!(bl->flags & IOBL_BUF_RING))
770 return -EINVAL;
771
772 xa_erase(&ctx->io_bl_xa, bl->bgid);
773 io_put_bl(ctx, bl);
774 return 0;
775}
776
777int io_register_pbuf_status(struct io_ring_ctx *ctx, void __user *arg)
778{
779 struct io_uring_buf_status buf_status;
780 struct io_buffer_list *bl;
781 int i;
782
783 if (copy_from_user(&buf_status, arg, sizeof(buf_status)))
784 return -EFAULT;
785
786 for (i = 0; i < ARRAY_SIZE(buf_status.resv); i++)
787 if (buf_status.resv[i])
788 return -EINVAL;
789
790 bl = io_buffer_get_list(ctx, buf_status.buf_group);
791 if (!bl)
792 return -ENOENT;
793 if (!(bl->flags & IOBL_BUF_RING))
794 return -EINVAL;
795
796 buf_status.head = bl->head;
797 if (copy_to_user(arg, &buf_status, sizeof(buf_status)))
798 return -EFAULT;
799
800 return 0;
801}
802
803struct io_buffer_list *io_pbuf_get_bl(struct io_ring_ctx *ctx,
804 unsigned long bgid)
805{
806 struct io_buffer_list *bl;
807 bool ret;
808
809 /*
810 * We have to be a bit careful here - we're inside mmap and cannot grab
811 * the uring_lock. This means the buffer_list could be simultaneously
812 * going away, if someone is trying to be sneaky. Look it up under rcu
813 * so we know it's not going away, and attempt to grab a reference to
814 * it. If the ref is already zero, then fail the mapping. If successful,
815 * the caller will call io_put_bl() to drop the the reference at at the
816 * end. This may then safely free the buffer_list (and drop the pages)
817 * at that point, vm_insert_pages() would've already grabbed the
818 * necessary vma references.
819 */
820 rcu_read_lock();
821 bl = xa_load(&ctx->io_bl_xa, bgid);
822 /* must be a mmap'able buffer ring and have pages */
823 ret = false;
824 if (bl && bl->flags & IOBL_MMAP)
825 ret = atomic_inc_not_zero(&bl->refs);
826 rcu_read_unlock();
827
828 if (ret)
829 return bl;
830
831 return ERR_PTR(-EINVAL);
832}
833
834int io_pbuf_mmap(struct file *file, struct vm_area_struct *vma)
835{
836 struct io_ring_ctx *ctx = file->private_data;
837 loff_t pgoff = vma->vm_pgoff << PAGE_SHIFT;
838 struct io_buffer_list *bl;
839 int bgid, ret;
840
841 bgid = (pgoff & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
842 bl = io_pbuf_get_bl(ctx, bgid);
843 if (IS_ERR(bl))
844 return PTR_ERR(bl);
845
846 ret = io_uring_mmap_pages(ctx, vma, bl->buf_pages, bl->buf_nr_pages);
847 io_put_bl(ctx, bl);
848 return ret;
849}
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/io_uring.h>
11
12#include <uapi/linux/io_uring.h>
13
14#include "io_uring.h"
15#include "opdef.h"
16#include "kbuf.h"
17
18#define IO_BUFFER_LIST_BUF_PER_PAGE (PAGE_SIZE / sizeof(struct io_uring_buf))
19
20/* BIDs are addressed by a 16-bit field in a CQE */
21#define MAX_BIDS_PER_BGID (1 << 16)
22
23struct kmem_cache *io_buf_cachep;
24
25struct io_provide_buf {
26 struct file *file;
27 __u64 addr;
28 __u32 len;
29 __u32 bgid;
30 __u32 nbufs;
31 __u16 bid;
32};
33
34struct io_buf_free {
35 struct hlist_node list;
36 void *mem;
37 size_t size;
38 int inuse;
39};
40
41static inline struct io_buffer_list *__io_buffer_get_list(struct io_ring_ctx *ctx,
42 unsigned int bgid)
43{
44 return xa_load(&ctx->io_bl_xa, bgid);
45}
46
47static inline struct io_buffer_list *io_buffer_get_list(struct io_ring_ctx *ctx,
48 unsigned int bgid)
49{
50 lockdep_assert_held(&ctx->uring_lock);
51
52 return __io_buffer_get_list(ctx, bgid);
53}
54
55static int io_buffer_add_list(struct io_ring_ctx *ctx,
56 struct io_buffer_list *bl, unsigned int bgid)
57{
58 /*
59 * Store buffer group ID and finally mark the list as visible.
60 * The normal lookup doesn't care about the visibility as we're
61 * always under the ->uring_lock, but the RCU lookup from mmap does.
62 */
63 bl->bgid = bgid;
64 atomic_set(&bl->refs, 1);
65 return xa_err(xa_store(&ctx->io_bl_xa, bgid, bl, GFP_KERNEL));
66}
67
68bool io_kbuf_recycle_legacy(struct io_kiocb *req, unsigned issue_flags)
69{
70 struct io_ring_ctx *ctx = req->ctx;
71 struct io_buffer_list *bl;
72 struct io_buffer *buf;
73
74 io_ring_submit_lock(ctx, issue_flags);
75
76 buf = req->kbuf;
77 bl = io_buffer_get_list(ctx, buf->bgid);
78 list_add(&buf->list, &bl->buf_list);
79 req->flags &= ~REQ_F_BUFFER_SELECTED;
80 req->buf_index = buf->bgid;
81
82 io_ring_submit_unlock(ctx, issue_flags);
83 return true;
84}
85
86void __io_put_kbuf(struct io_kiocb *req, unsigned issue_flags)
87{
88 /*
89 * We can add this buffer back to two lists:
90 *
91 * 1) The io_buffers_cache list. This one is protected by the
92 * ctx->uring_lock. If we already hold this lock, add back to this
93 * list as we can grab it from issue as well.
94 * 2) The io_buffers_comp list. This one is protected by the
95 * ctx->completion_lock.
96 *
97 * We migrate buffers from the comp_list to the issue cache list
98 * when we need one.
99 */
100 if (issue_flags & IO_URING_F_UNLOCKED) {
101 struct io_ring_ctx *ctx = req->ctx;
102
103 spin_lock(&ctx->completion_lock);
104 __io_put_kbuf_list(req, &ctx->io_buffers_comp);
105 spin_unlock(&ctx->completion_lock);
106 } else {
107 lockdep_assert_held(&req->ctx->uring_lock);
108
109 __io_put_kbuf_list(req, &req->ctx->io_buffers_cache);
110 }
111}
112
113static void __user *io_provided_buffer_select(struct io_kiocb *req, size_t *len,
114 struct io_buffer_list *bl)
115{
116 if (!list_empty(&bl->buf_list)) {
117 struct io_buffer *kbuf;
118
119 kbuf = list_first_entry(&bl->buf_list, struct io_buffer, list);
120 list_del(&kbuf->list);
121 if (*len == 0 || *len > kbuf->len)
122 *len = kbuf->len;
123 if (list_empty(&bl->buf_list))
124 req->flags |= REQ_F_BL_EMPTY;
125 req->flags |= REQ_F_BUFFER_SELECTED;
126 req->kbuf = kbuf;
127 req->buf_index = kbuf->bid;
128 return u64_to_user_ptr(kbuf->addr);
129 }
130 return NULL;
131}
132
133static void __user *io_ring_buffer_select(struct io_kiocb *req, size_t *len,
134 struct io_buffer_list *bl,
135 unsigned int issue_flags)
136{
137 struct io_uring_buf_ring *br = bl->buf_ring;
138 __u16 tail, head = bl->head;
139 struct io_uring_buf *buf;
140
141 tail = smp_load_acquire(&br->tail);
142 if (unlikely(tail == head))
143 return NULL;
144
145 if (head + 1 == tail)
146 req->flags |= REQ_F_BL_EMPTY;
147
148 head &= bl->mask;
149 /* mmaped buffers are always contig */
150 if (bl->is_mmap || head < IO_BUFFER_LIST_BUF_PER_PAGE) {
151 buf = &br->bufs[head];
152 } else {
153 int off = head & (IO_BUFFER_LIST_BUF_PER_PAGE - 1);
154 int index = head / IO_BUFFER_LIST_BUF_PER_PAGE;
155 buf = page_address(bl->buf_pages[index]);
156 buf += off;
157 }
158 if (*len == 0 || *len > buf->len)
159 *len = buf->len;
160 req->flags |= REQ_F_BUFFER_RING;
161 req->buf_list = bl;
162 req->buf_index = buf->bid;
163
164 if (issue_flags & IO_URING_F_UNLOCKED || !io_file_can_poll(req)) {
165 /*
166 * If we came in unlocked, we have no choice but to consume the
167 * buffer here, otherwise nothing ensures that the buffer won't
168 * get used by others. This does mean it'll be pinned until the
169 * IO completes, coming in unlocked means we're being called from
170 * io-wq context and there may be further retries in async hybrid
171 * mode. For the locked case, the caller must call commit when
172 * the transfer completes (or if we get -EAGAIN and must poll of
173 * retry).
174 */
175 req->buf_list = NULL;
176 bl->head++;
177 }
178 return u64_to_user_ptr(buf->addr);
179}
180
181void __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/*
202 * Mark the given mapped range as free for reuse
203 */
204static void io_kbuf_mark_free(struct io_ring_ctx *ctx, struct io_buffer_list *bl)
205{
206 struct io_buf_free *ibf;
207
208 hlist_for_each_entry(ibf, &ctx->io_buf_list, list) {
209 if (bl->buf_ring == ibf->mem) {
210 ibf->inuse = 0;
211 return;
212 }
213 }
214
215 /* can't happen... */
216 WARN_ON_ONCE(1);
217}
218
219static int __io_remove_buffers(struct io_ring_ctx *ctx,
220 struct io_buffer_list *bl, unsigned nbufs)
221{
222 unsigned i = 0;
223
224 /* shouldn't happen */
225 if (!nbufs)
226 return 0;
227
228 if (bl->is_buf_ring) {
229 i = bl->buf_ring->tail - bl->head;
230 if (bl->is_mmap) {
231 /*
232 * io_kbuf_list_free() will free the page(s) at
233 * ->release() time.
234 */
235 io_kbuf_mark_free(ctx, bl);
236 bl->buf_ring = NULL;
237 bl->is_mmap = 0;
238 } else if (bl->buf_nr_pages) {
239 int j;
240
241 for (j = 0; j < bl->buf_nr_pages; j++)
242 unpin_user_page(bl->buf_pages[j]);
243 kvfree(bl->buf_pages);
244 bl->buf_pages = NULL;
245 bl->buf_nr_pages = 0;
246 }
247 /* make sure it's seen as empty */
248 INIT_LIST_HEAD(&bl->buf_list);
249 bl->is_buf_ring = 0;
250 return i;
251 }
252
253 /* protects io_buffers_cache */
254 lockdep_assert_held(&ctx->uring_lock);
255
256 while (!list_empty(&bl->buf_list)) {
257 struct io_buffer *nxt;
258
259 nxt = list_first_entry(&bl->buf_list, struct io_buffer, list);
260 list_move(&nxt->list, &ctx->io_buffers_cache);
261 if (++i == nbufs)
262 return i;
263 cond_resched();
264 }
265
266 return i;
267}
268
269void io_put_bl(struct io_ring_ctx *ctx, struct io_buffer_list *bl)
270{
271 if (atomic_dec_and_test(&bl->refs)) {
272 __io_remove_buffers(ctx, bl, -1U);
273 kfree_rcu(bl, rcu);
274 }
275}
276
277void io_destroy_buffers(struct io_ring_ctx *ctx)
278{
279 struct io_buffer_list *bl;
280 struct list_head *item, *tmp;
281 struct io_buffer *buf;
282 unsigned long index;
283
284 xa_for_each(&ctx->io_bl_xa, index, bl) {
285 xa_erase(&ctx->io_bl_xa, bl->bgid);
286 io_put_bl(ctx, bl);
287 }
288
289 /*
290 * Move deferred locked entries to cache before pruning
291 */
292 spin_lock(&ctx->completion_lock);
293 if (!list_empty(&ctx->io_buffers_comp))
294 list_splice_init(&ctx->io_buffers_comp, &ctx->io_buffers_cache);
295 spin_unlock(&ctx->completion_lock);
296
297 list_for_each_safe(item, tmp, &ctx->io_buffers_cache) {
298 buf = list_entry(item, struct io_buffer, list);
299 kmem_cache_free(io_buf_cachep, buf);
300 }
301}
302
303int io_remove_buffers_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
304{
305 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
306 u64 tmp;
307
308 if (sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
309 sqe->splice_fd_in)
310 return -EINVAL;
311
312 tmp = READ_ONCE(sqe->fd);
313 if (!tmp || tmp > MAX_BIDS_PER_BGID)
314 return -EINVAL;
315
316 memset(p, 0, sizeof(*p));
317 p->nbufs = tmp;
318 p->bgid = READ_ONCE(sqe->buf_group);
319 return 0;
320}
321
322int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
323{
324 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
325 struct io_ring_ctx *ctx = req->ctx;
326 struct io_buffer_list *bl;
327 int ret = 0;
328
329 io_ring_submit_lock(ctx, issue_flags);
330
331 ret = -ENOENT;
332 bl = io_buffer_get_list(ctx, p->bgid);
333 if (bl) {
334 ret = -EINVAL;
335 /* can't use provide/remove buffers command on mapped buffers */
336 if (!bl->is_buf_ring)
337 ret = __io_remove_buffers(ctx, bl, p->nbufs);
338 }
339 io_ring_submit_unlock(ctx, issue_flags);
340 if (ret < 0)
341 req_set_fail(req);
342 io_req_set_res(req, ret, 0);
343 return IOU_OK;
344}
345
346int io_provide_buffers_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
347{
348 unsigned long size, tmp_check;
349 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
350 u64 tmp;
351
352 if (sqe->rw_flags || sqe->splice_fd_in)
353 return -EINVAL;
354
355 tmp = READ_ONCE(sqe->fd);
356 if (!tmp || tmp > MAX_BIDS_PER_BGID)
357 return -E2BIG;
358 p->nbufs = tmp;
359 p->addr = READ_ONCE(sqe->addr);
360 p->len = READ_ONCE(sqe->len);
361
362 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
363 &size))
364 return -EOVERFLOW;
365 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
366 return -EOVERFLOW;
367
368 size = (unsigned long)p->len * p->nbufs;
369 if (!access_ok(u64_to_user_ptr(p->addr), size))
370 return -EFAULT;
371
372 p->bgid = READ_ONCE(sqe->buf_group);
373 tmp = READ_ONCE(sqe->off);
374 if (tmp > USHRT_MAX)
375 return -E2BIG;
376 if (tmp + p->nbufs > MAX_BIDS_PER_BGID)
377 return -EINVAL;
378 p->bid = tmp;
379 return 0;
380}
381
382#define IO_BUFFER_ALLOC_BATCH 64
383
384static int io_refill_buffer_cache(struct io_ring_ctx *ctx)
385{
386 struct io_buffer *bufs[IO_BUFFER_ALLOC_BATCH];
387 int allocated;
388
389 /*
390 * Completions that don't happen inline (eg not under uring_lock) will
391 * add to ->io_buffers_comp. If we don't have any free buffers, check
392 * the completion list and splice those entries first.
393 */
394 if (!list_empty_careful(&ctx->io_buffers_comp)) {
395 spin_lock(&ctx->completion_lock);
396 if (!list_empty(&ctx->io_buffers_comp)) {
397 list_splice_init(&ctx->io_buffers_comp,
398 &ctx->io_buffers_cache);
399 spin_unlock(&ctx->completion_lock);
400 return 0;
401 }
402 spin_unlock(&ctx->completion_lock);
403 }
404
405 /*
406 * No free buffers and no completion entries either. Allocate a new
407 * batch of buffer entries and add those to our freelist.
408 */
409
410 allocated = kmem_cache_alloc_bulk(io_buf_cachep, GFP_KERNEL_ACCOUNT,
411 ARRAY_SIZE(bufs), (void **) bufs);
412 if (unlikely(!allocated)) {
413 /*
414 * Bulk alloc is all-or-nothing. If we fail to get a batch,
415 * retry single alloc to be on the safe side.
416 */
417 bufs[0] = kmem_cache_alloc(io_buf_cachep, GFP_KERNEL);
418 if (!bufs[0])
419 return -ENOMEM;
420 allocated = 1;
421 }
422
423 while (allocated)
424 list_add_tail(&bufs[--allocated]->list, &ctx->io_buffers_cache);
425
426 return 0;
427}
428
429static int io_add_buffers(struct io_ring_ctx *ctx, struct io_provide_buf *pbuf,
430 struct io_buffer_list *bl)
431{
432 struct io_buffer *buf;
433 u64 addr = pbuf->addr;
434 int i, bid = pbuf->bid;
435
436 for (i = 0; i < pbuf->nbufs; i++) {
437 if (list_empty(&ctx->io_buffers_cache) &&
438 io_refill_buffer_cache(ctx))
439 break;
440 buf = list_first_entry(&ctx->io_buffers_cache, struct io_buffer,
441 list);
442 list_move_tail(&buf->list, &bl->buf_list);
443 buf->addr = addr;
444 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
445 buf->bid = bid;
446 buf->bgid = pbuf->bgid;
447 addr += pbuf->len;
448 bid++;
449 cond_resched();
450 }
451
452 return i ? 0 : -ENOMEM;
453}
454
455int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
456{
457 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
458 struct io_ring_ctx *ctx = req->ctx;
459 struct io_buffer_list *bl;
460 int ret = 0;
461
462 io_ring_submit_lock(ctx, issue_flags);
463
464 bl = io_buffer_get_list(ctx, p->bgid);
465 if (unlikely(!bl)) {
466 bl = kzalloc(sizeof(*bl), GFP_KERNEL_ACCOUNT);
467 if (!bl) {
468 ret = -ENOMEM;
469 goto err;
470 }
471 INIT_LIST_HEAD(&bl->buf_list);
472 ret = io_buffer_add_list(ctx, bl, p->bgid);
473 if (ret) {
474 /*
475 * Doesn't need rcu free as it was never visible, but
476 * let's keep it consistent throughout.
477 */
478 kfree_rcu(bl, rcu);
479 goto err;
480 }
481 }
482 /* can't add buffers via this command for a mapped buffer ring */
483 if (bl->is_buf_ring) {
484 ret = -EINVAL;
485 goto err;
486 }
487
488 ret = io_add_buffers(ctx, p, bl);
489err:
490 io_ring_submit_unlock(ctx, issue_flags);
491
492 if (ret < 0)
493 req_set_fail(req);
494 io_req_set_res(req, ret, 0);
495 return IOU_OK;
496}
497
498static int io_pin_pbuf_ring(struct io_uring_buf_reg *reg,
499 struct io_buffer_list *bl)
500{
501 struct io_uring_buf_ring *br;
502 struct page **pages;
503 int i, nr_pages;
504
505 pages = io_pin_pages(reg->ring_addr,
506 flex_array_size(br, bufs, reg->ring_entries),
507 &nr_pages);
508 if (IS_ERR(pages))
509 return PTR_ERR(pages);
510
511 /*
512 * Apparently some 32-bit boxes (ARM) will return highmem pages,
513 * which then need to be mapped. We could support that, but it'd
514 * complicate the code and slowdown the common cases quite a bit.
515 * So just error out, returning -EINVAL just like we did on kernels
516 * that didn't support mapped buffer rings.
517 */
518 for (i = 0; i < nr_pages; i++)
519 if (PageHighMem(pages[i]))
520 goto error_unpin;
521
522 br = page_address(pages[0]);
523#ifdef SHM_COLOUR
524 /*
525 * On platforms that have specific aliasing requirements, SHM_COLOUR
526 * is set and we must guarantee that the kernel and user side align
527 * nicely. We cannot do that if IOU_PBUF_RING_MMAP isn't set and
528 * the application mmap's the provided ring buffer. Fail the request
529 * if we, by chance, don't end up with aligned addresses. The app
530 * should use IOU_PBUF_RING_MMAP instead, and liburing will handle
531 * this transparently.
532 */
533 if ((reg->ring_addr | (unsigned long) br) & (SHM_COLOUR - 1))
534 goto error_unpin;
535#endif
536 bl->buf_pages = pages;
537 bl->buf_nr_pages = nr_pages;
538 bl->buf_ring = br;
539 bl->is_buf_ring = 1;
540 bl->is_mmap = 0;
541 return 0;
542error_unpin:
543 for (i = 0; i < nr_pages; i++)
544 unpin_user_page(pages[i]);
545 kvfree(pages);
546 return -EINVAL;
547}
548
549/*
550 * See if we have a suitable region that we can reuse, rather than allocate
551 * both a new io_buf_free and mem region again. We leave it on the list as
552 * even a reused entry will need freeing at ring release.
553 */
554static struct io_buf_free *io_lookup_buf_free_entry(struct io_ring_ctx *ctx,
555 size_t ring_size)
556{
557 struct io_buf_free *ibf, *best = NULL;
558 size_t best_dist;
559
560 hlist_for_each_entry(ibf, &ctx->io_buf_list, list) {
561 size_t dist;
562
563 if (ibf->inuse || ibf->size < ring_size)
564 continue;
565 dist = ibf->size - ring_size;
566 if (!best || dist < best_dist) {
567 best = ibf;
568 if (!dist)
569 break;
570 best_dist = dist;
571 }
572 }
573
574 return best;
575}
576
577static int io_alloc_pbuf_ring(struct io_ring_ctx *ctx,
578 struct io_uring_buf_reg *reg,
579 struct io_buffer_list *bl)
580{
581 struct io_buf_free *ibf;
582 size_t ring_size;
583 void *ptr;
584
585 ring_size = reg->ring_entries * sizeof(struct io_uring_buf_ring);
586
587 /* Reuse existing entry, if we can */
588 ibf = io_lookup_buf_free_entry(ctx, ring_size);
589 if (!ibf) {
590 ptr = io_mem_alloc(ring_size);
591 if (IS_ERR(ptr))
592 return PTR_ERR(ptr);
593
594 /* Allocate and store deferred free entry */
595 ibf = kmalloc(sizeof(*ibf), GFP_KERNEL_ACCOUNT);
596 if (!ibf) {
597 io_mem_free(ptr);
598 return -ENOMEM;
599 }
600 ibf->mem = ptr;
601 ibf->size = ring_size;
602 hlist_add_head(&ibf->list, &ctx->io_buf_list);
603 }
604 ibf->inuse = 1;
605 bl->buf_ring = ibf->mem;
606 bl->is_buf_ring = 1;
607 bl->is_mmap = 1;
608 return 0;
609}
610
611int io_register_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
612{
613 struct io_uring_buf_reg reg;
614 struct io_buffer_list *bl, *free_bl = NULL;
615 int ret;
616
617 lockdep_assert_held(&ctx->uring_lock);
618
619 if (copy_from_user(®, arg, sizeof(reg)))
620 return -EFAULT;
621
622 if (reg.resv[0] || reg.resv[1] || reg.resv[2])
623 return -EINVAL;
624 if (reg.flags & ~IOU_PBUF_RING_MMAP)
625 return -EINVAL;
626 if (!(reg.flags & IOU_PBUF_RING_MMAP)) {
627 if (!reg.ring_addr)
628 return -EFAULT;
629 if (reg.ring_addr & ~PAGE_MASK)
630 return -EINVAL;
631 } else {
632 if (reg.ring_addr)
633 return -EINVAL;
634 }
635
636 if (!is_power_of_2(reg.ring_entries))
637 return -EINVAL;
638
639 /* cannot disambiguate full vs empty due to head/tail size */
640 if (reg.ring_entries >= 65536)
641 return -EINVAL;
642
643 bl = io_buffer_get_list(ctx, reg.bgid);
644 if (bl) {
645 /* if mapped buffer ring OR classic exists, don't allow */
646 if (bl->is_buf_ring || !list_empty(&bl->buf_list))
647 return -EEXIST;
648 } else {
649 free_bl = bl = kzalloc(sizeof(*bl), GFP_KERNEL);
650 if (!bl)
651 return -ENOMEM;
652 }
653
654 if (!(reg.flags & IOU_PBUF_RING_MMAP))
655 ret = io_pin_pbuf_ring(®, bl);
656 else
657 ret = io_alloc_pbuf_ring(ctx, ®, bl);
658
659 if (!ret) {
660 bl->nr_entries = reg.ring_entries;
661 bl->mask = reg.ring_entries - 1;
662
663 io_buffer_add_list(ctx, bl, reg.bgid);
664 return 0;
665 }
666
667 kfree_rcu(free_bl, rcu);
668 return ret;
669}
670
671int io_unregister_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
672{
673 struct io_uring_buf_reg reg;
674 struct io_buffer_list *bl;
675
676 lockdep_assert_held(&ctx->uring_lock);
677
678 if (copy_from_user(®, arg, sizeof(reg)))
679 return -EFAULT;
680 if (reg.resv[0] || reg.resv[1] || reg.resv[2])
681 return -EINVAL;
682 if (reg.flags)
683 return -EINVAL;
684
685 bl = io_buffer_get_list(ctx, reg.bgid);
686 if (!bl)
687 return -ENOENT;
688 if (!bl->is_buf_ring)
689 return -EINVAL;
690
691 xa_erase(&ctx->io_bl_xa, bl->bgid);
692 io_put_bl(ctx, bl);
693 return 0;
694}
695
696int io_register_pbuf_status(struct io_ring_ctx *ctx, void __user *arg)
697{
698 struct io_uring_buf_status buf_status;
699 struct io_buffer_list *bl;
700 int i;
701
702 if (copy_from_user(&buf_status, arg, sizeof(buf_status)))
703 return -EFAULT;
704
705 for (i = 0; i < ARRAY_SIZE(buf_status.resv); i++)
706 if (buf_status.resv[i])
707 return -EINVAL;
708
709 bl = io_buffer_get_list(ctx, buf_status.buf_group);
710 if (!bl)
711 return -ENOENT;
712 if (!bl->is_buf_ring)
713 return -EINVAL;
714
715 buf_status.head = bl->head;
716 if (copy_to_user(arg, &buf_status, sizeof(buf_status)))
717 return -EFAULT;
718
719 return 0;
720}
721
722struct io_buffer_list *io_pbuf_get_bl(struct io_ring_ctx *ctx,
723 unsigned long bgid)
724{
725 struct io_buffer_list *bl;
726 bool ret;
727
728 /*
729 * We have to be a bit careful here - we're inside mmap and cannot grab
730 * the uring_lock. This means the buffer_list could be simultaneously
731 * going away, if someone is trying to be sneaky. Look it up under rcu
732 * so we know it's not going away, and attempt to grab a reference to
733 * it. If the ref is already zero, then fail the mapping. If successful,
734 * the caller will call io_put_bl() to drop the the reference at at the
735 * end. This may then safely free the buffer_list (and drop the pages)
736 * at that point, vm_insert_pages() would've already grabbed the
737 * necessary vma references.
738 */
739 rcu_read_lock();
740 bl = xa_load(&ctx->io_bl_xa, bgid);
741 /* must be a mmap'able buffer ring and have pages */
742 ret = false;
743 if (bl && bl->is_mmap)
744 ret = atomic_inc_not_zero(&bl->refs);
745 rcu_read_unlock();
746
747 if (ret)
748 return bl;
749
750 return ERR_PTR(-EINVAL);
751}
752
753/*
754 * Called at or after ->release(), free the mmap'ed buffers that we used
755 * for memory mapped provided buffer rings.
756 */
757void io_kbuf_mmap_list_free(struct io_ring_ctx *ctx)
758{
759 struct io_buf_free *ibf;
760 struct hlist_node *tmp;
761
762 hlist_for_each_entry_safe(ibf, tmp, &ctx->io_buf_list, list) {
763 hlist_del(&ibf->list);
764 io_mem_free(ibf->mem);
765 kfree(ibf);
766 }
767}