1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
5 #include <linux/file.h>
7 #include <linux/slab.h>
8 #include <linux/namei.h>
9 #include <linux/poll.h>
10 #include <linux/io_uring.h>
12 #include <uapi/linux/io_uring.h>
18 #define IO_BUFFER_LIST_BUF_PER_PAGE (PAGE_SIZE / sizeof(struct io_uring_buf))
22 /* BIDs are addressed by a 16-bit field in a CQE */
23 #define MAX_BIDS_PER_BGID (1 << 16)
25 struct kmem_cache
*io_buf_cachep
;
27 struct io_provide_buf
{
36 static inline struct io_buffer_list
*io_buffer_get_list(struct io_ring_ctx
*ctx
,
39 if (ctx
->io_bl
&& bgid
< BGID_ARRAY
)
40 return &ctx
->io_bl
[bgid
];
42 return xa_load(&ctx
->io_bl_xa
, bgid
);
45 static int io_buffer_add_list(struct io_ring_ctx
*ctx
,
46 struct io_buffer_list
*bl
, unsigned int bgid
)
49 if (bgid
< BGID_ARRAY
)
52 return xa_err(xa_store(&ctx
->io_bl_xa
, bgid
, bl
, GFP_KERNEL
));
55 bool io_kbuf_recycle_legacy(struct io_kiocb
*req
, unsigned issue_flags
)
57 struct io_ring_ctx
*ctx
= req
->ctx
;
58 struct io_buffer_list
*bl
;
59 struct io_buffer
*buf
;
62 * For legacy provided buffer mode, don't recycle if we already did
63 * IO to this buffer. For ring-mapped provided buffer mode, we should
64 * increment ring->head to explicitly monopolize the buffer to avoid
67 if (req
->flags
& REQ_F_PARTIAL_IO
)
70 io_ring_submit_lock(ctx
, issue_flags
);
73 bl
= io_buffer_get_list(ctx
, buf
->bgid
);
74 list_add(&buf
->list
, &bl
->buf_list
);
75 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
76 req
->buf_index
= buf
->bgid
;
78 io_ring_submit_unlock(ctx
, issue_flags
);
82 unsigned int __io_put_kbuf(struct io_kiocb
*req
, unsigned issue_flags
)
87 * We can add this buffer back to two lists:
89 * 1) The io_buffers_cache list. This one is protected by the
90 * ctx->uring_lock. If we already hold this lock, add back to this
91 * list as we can grab it from issue as well.
92 * 2) The io_buffers_comp list. This one is protected by the
93 * ctx->completion_lock.
95 * We migrate buffers from the comp_list to the issue cache list
98 if (req
->flags
& REQ_F_BUFFER_RING
) {
99 /* no buffers to recycle for this case */
100 cflags
= __io_put_kbuf_list(req
, NULL
);
101 } else if (issue_flags
& IO_URING_F_UNLOCKED
) {
102 struct io_ring_ctx
*ctx
= req
->ctx
;
104 spin_lock(&ctx
->completion_lock
);
105 cflags
= __io_put_kbuf_list(req
, &ctx
->io_buffers_comp
);
106 spin_unlock(&ctx
->completion_lock
);
108 lockdep_assert_held(&req
->ctx
->uring_lock
);
110 cflags
= __io_put_kbuf_list(req
, &req
->ctx
->io_buffers_cache
);
115 static void __user
*io_provided_buffer_select(struct io_kiocb
*req
, size_t *len
,
116 struct io_buffer_list
*bl
)
118 if (!list_empty(&bl
->buf_list
)) {
119 struct io_buffer
*kbuf
;
121 kbuf
= list_first_entry(&bl
->buf_list
, struct io_buffer
, list
);
122 list_del(&kbuf
->list
);
123 if (*len
== 0 || *len
> kbuf
->len
)
125 req
->flags
|= REQ_F_BUFFER_SELECTED
;
127 req
->buf_index
= kbuf
->bid
;
128 return u64_to_user_ptr(kbuf
->addr
);
133 static void __user
*io_ring_buffer_select(struct io_kiocb
*req
, size_t *len
,
134 struct io_buffer_list
*bl
,
135 unsigned int issue_flags
)
137 struct io_uring_buf_ring
*br
= bl
->buf_ring
;
138 struct io_uring_buf
*buf
;
139 __u16 head
= bl
->head
;
141 if (unlikely(smp_load_acquire(&br
->tail
) == head
))
145 /* mmaped buffers are always contig */
146 if (bl
->is_mmap
|| head
< IO_BUFFER_LIST_BUF_PER_PAGE
) {
147 buf
= &br
->bufs
[head
];
149 int off
= head
& (IO_BUFFER_LIST_BUF_PER_PAGE
- 1);
150 int index
= head
/ IO_BUFFER_LIST_BUF_PER_PAGE
;
151 buf
= page_address(bl
->buf_pages
[index
]);
154 if (*len
== 0 || *len
> buf
->len
)
156 req
->flags
|= REQ_F_BUFFER_RING
;
158 req
->buf_index
= buf
->bid
;
160 if (issue_flags
& IO_URING_F_UNLOCKED
|| !file_can_poll(req
->file
)) {
162 * If we came in unlocked, we have no choice but to consume the
163 * buffer here, otherwise nothing ensures that the buffer won't
164 * get used by others. This does mean it'll be pinned until the
165 * IO completes, coming in unlocked means we're being called from
166 * io-wq context and there may be further retries in async hybrid
167 * mode. For the locked case, the caller must call commit when
168 * the transfer completes (or if we get -EAGAIN and must poll of
171 req
->buf_list
= NULL
;
174 return u64_to_user_ptr(buf
->addr
);
177 void __user
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
178 unsigned int issue_flags
)
180 struct io_ring_ctx
*ctx
= req
->ctx
;
181 struct io_buffer_list
*bl
;
182 void __user
*ret
= NULL
;
184 io_ring_submit_lock(req
->ctx
, issue_flags
);
186 bl
= io_buffer_get_list(ctx
, req
->buf_index
);
189 ret
= io_ring_buffer_select(req
, len
, bl
, issue_flags
);
191 ret
= io_provided_buffer_select(req
, len
, bl
);
193 io_ring_submit_unlock(req
->ctx
, issue_flags
);
197 static __cold
int io_init_bl_list(struct io_ring_ctx
*ctx
)
201 ctx
->io_bl
= kcalloc(BGID_ARRAY
, sizeof(struct io_buffer_list
),
206 for (i
= 0; i
< BGID_ARRAY
; i
++) {
207 INIT_LIST_HEAD(&ctx
->io_bl
[i
].buf_list
);
208 ctx
->io_bl
[i
].bgid
= i
;
214 static int __io_remove_buffers(struct io_ring_ctx
*ctx
,
215 struct io_buffer_list
*bl
, unsigned nbufs
)
219 /* shouldn't happen */
224 i
= bl
->buf_ring
->tail
- bl
->head
;
226 folio_put(virt_to_folio(bl
->buf_ring
));
229 } else if (bl
->buf_nr_pages
) {
232 for (j
= 0; j
< bl
->buf_nr_pages
; j
++)
233 unpin_user_page(bl
->buf_pages
[j
]);
234 kvfree(bl
->buf_pages
);
235 bl
->buf_pages
= NULL
;
236 bl
->buf_nr_pages
= 0;
238 /* make sure it's seen as empty */
239 INIT_LIST_HEAD(&bl
->buf_list
);
244 /* protects io_buffers_cache */
245 lockdep_assert_held(&ctx
->uring_lock
);
247 while (!list_empty(&bl
->buf_list
)) {
248 struct io_buffer
*nxt
;
250 nxt
= list_first_entry(&bl
->buf_list
, struct io_buffer
, list
);
251 list_move(&nxt
->list
, &ctx
->io_buffers_cache
);
260 void io_destroy_buffers(struct io_ring_ctx
*ctx
)
262 struct io_buffer_list
*bl
;
263 struct list_head
*item
, *tmp
;
264 struct io_buffer
*buf
;
268 for (i
= 0; i
< BGID_ARRAY
; i
++) {
271 __io_remove_buffers(ctx
, &ctx
->io_bl
[i
], -1U);
274 xa_for_each(&ctx
->io_bl_xa
, index
, bl
) {
275 xa_erase(&ctx
->io_bl_xa
, bl
->bgid
);
276 __io_remove_buffers(ctx
, bl
, -1U);
280 list_for_each_safe(item
, tmp
, &ctx
->io_buffers_cache
) {
281 buf
= list_entry(item
, struct io_buffer
, list
);
282 kmem_cache_free(io_buf_cachep
, buf
);
286 int io_remove_buffers_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
288 struct io_provide_buf
*p
= io_kiocb_to_cmd(req
, struct io_provide_buf
);
291 if (sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
||
295 tmp
= READ_ONCE(sqe
->fd
);
296 if (!tmp
|| tmp
> MAX_BIDS_PER_BGID
)
299 memset(p
, 0, sizeof(*p
));
301 p
->bgid
= READ_ONCE(sqe
->buf_group
);
305 int io_remove_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
307 struct io_provide_buf
*p
= io_kiocb_to_cmd(req
, struct io_provide_buf
);
308 struct io_ring_ctx
*ctx
= req
->ctx
;
309 struct io_buffer_list
*bl
;
312 io_ring_submit_lock(ctx
, issue_flags
);
315 bl
= io_buffer_get_list(ctx
, p
->bgid
);
318 /* can't use provide/remove buffers command on mapped buffers */
320 ret
= __io_remove_buffers(ctx
, bl
, p
->nbufs
);
322 io_ring_submit_unlock(ctx
, issue_flags
);
325 io_req_set_res(req
, ret
, 0);
329 int io_provide_buffers_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
331 unsigned long size
, tmp_check
;
332 struct io_provide_buf
*p
= io_kiocb_to_cmd(req
, struct io_provide_buf
);
335 if (sqe
->rw_flags
|| sqe
->splice_fd_in
)
338 tmp
= READ_ONCE(sqe
->fd
);
339 if (!tmp
|| tmp
> MAX_BIDS_PER_BGID
)
342 p
->addr
= READ_ONCE(sqe
->addr
);
343 p
->len
= READ_ONCE(sqe
->len
);
345 if (check_mul_overflow((unsigned long)p
->len
, (unsigned long)p
->nbufs
,
348 if (check_add_overflow((unsigned long)p
->addr
, size
, &tmp_check
))
351 size
= (unsigned long)p
->len
* p
->nbufs
;
352 if (!access_ok(u64_to_user_ptr(p
->addr
), size
))
355 p
->bgid
= READ_ONCE(sqe
->buf_group
);
356 tmp
= READ_ONCE(sqe
->off
);
359 if (tmp
+ p
->nbufs
> MAX_BIDS_PER_BGID
)
365 #define IO_BUFFER_ALLOC_BATCH 64
367 static int io_refill_buffer_cache(struct io_ring_ctx
*ctx
)
369 struct io_buffer
*bufs
[IO_BUFFER_ALLOC_BATCH
];
373 * Completions that don't happen inline (eg not under uring_lock) will
374 * add to ->io_buffers_comp. If we don't have any free buffers, check
375 * the completion list and splice those entries first.
377 if (!list_empty_careful(&ctx
->io_buffers_comp
)) {
378 spin_lock(&ctx
->completion_lock
);
379 if (!list_empty(&ctx
->io_buffers_comp
)) {
380 list_splice_init(&ctx
->io_buffers_comp
,
381 &ctx
->io_buffers_cache
);
382 spin_unlock(&ctx
->completion_lock
);
385 spin_unlock(&ctx
->completion_lock
);
389 * No free buffers and no completion entries either. Allocate a new
390 * batch of buffer entries and add those to our freelist.
393 allocated
= kmem_cache_alloc_bulk(io_buf_cachep
, GFP_KERNEL_ACCOUNT
,
394 ARRAY_SIZE(bufs
), (void **) bufs
);
395 if (unlikely(!allocated
)) {
397 * Bulk alloc is all-or-nothing. If we fail to get a batch,
398 * retry single alloc to be on the safe side.
400 bufs
[0] = kmem_cache_alloc(io_buf_cachep
, GFP_KERNEL
);
407 list_add_tail(&bufs
[--allocated
]->list
, &ctx
->io_buffers_cache
);
412 static int io_add_buffers(struct io_ring_ctx
*ctx
, struct io_provide_buf
*pbuf
,
413 struct io_buffer_list
*bl
)
415 struct io_buffer
*buf
;
416 u64 addr
= pbuf
->addr
;
417 int i
, bid
= pbuf
->bid
;
419 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
420 if (list_empty(&ctx
->io_buffers_cache
) &&
421 io_refill_buffer_cache(ctx
))
423 buf
= list_first_entry(&ctx
->io_buffers_cache
, struct io_buffer
,
425 list_move_tail(&buf
->list
, &bl
->buf_list
);
427 buf
->len
= min_t(__u32
, pbuf
->len
, MAX_RW_COUNT
);
429 buf
->bgid
= pbuf
->bgid
;
435 return i
? 0 : -ENOMEM
;
438 int io_provide_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
440 struct io_provide_buf
*p
= io_kiocb_to_cmd(req
, struct io_provide_buf
);
441 struct io_ring_ctx
*ctx
= req
->ctx
;
442 struct io_buffer_list
*bl
;
445 io_ring_submit_lock(ctx
, issue_flags
);
447 if (unlikely(p
->bgid
< BGID_ARRAY
&& !ctx
->io_bl
)) {
448 ret
= io_init_bl_list(ctx
);
453 bl
= io_buffer_get_list(ctx
, p
->bgid
);
455 bl
= kzalloc(sizeof(*bl
), GFP_KERNEL_ACCOUNT
);
460 INIT_LIST_HEAD(&bl
->buf_list
);
461 ret
= io_buffer_add_list(ctx
, bl
, p
->bgid
);
467 /* can't add buffers via this command for a mapped buffer ring */
473 ret
= io_add_buffers(ctx
, p
, bl
);
475 io_ring_submit_unlock(ctx
, issue_flags
);
479 io_req_set_res(req
, ret
, 0);
483 static int io_pin_pbuf_ring(struct io_uring_buf_reg
*reg
,
484 struct io_buffer_list
*bl
)
486 struct io_uring_buf_ring
*br
;
490 pages
= io_pin_pages(reg
->ring_addr
,
491 flex_array_size(br
, bufs
, reg
->ring_entries
),
494 return PTR_ERR(pages
);
497 * Apparently some 32-bit boxes (ARM) will return highmem pages,
498 * which then need to be mapped. We could support that, but it'd
499 * complicate the code and slowdown the common cases quite a bit.
500 * So just error out, returning -EINVAL just like we did on kernels
501 * that didn't support mapped buffer rings.
503 for (i
= 0; i
< nr_pages
; i
++)
504 if (PageHighMem(pages
[i
]))
507 br
= page_address(pages
[0]);
510 * On platforms that have specific aliasing requirements, SHM_COLOUR
511 * is set and we must guarantee that the kernel and user side align
512 * nicely. We cannot do that if IOU_PBUF_RING_MMAP isn't set and
513 * the application mmap's the provided ring buffer. Fail the request
514 * if we, by chance, don't end up with aligned addresses. The app
515 * should use IOU_PBUF_RING_MMAP instead, and liburing will handle
516 * this transparently.
518 if ((reg
->ring_addr
| (unsigned long) br
) & (SHM_COLOUR
- 1))
521 bl
->buf_pages
= pages
;
522 bl
->buf_nr_pages
= nr_pages
;
528 for (i
= 0; i
< nr_pages
; i
++)
529 unpin_user_page(pages
[i
]);
534 static int io_alloc_pbuf_ring(struct io_uring_buf_reg
*reg
,
535 struct io_buffer_list
*bl
)
537 gfp_t gfp
= GFP_KERNEL_ACCOUNT
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
;
541 ring_size
= reg
->ring_entries
* sizeof(struct io_uring_buf_ring
);
542 ptr
= (void *) __get_free_pages(gfp
, get_order(ring_size
));
552 int io_register_pbuf_ring(struct io_ring_ctx
*ctx
, void __user
*arg
)
554 struct io_uring_buf_reg reg
;
555 struct io_buffer_list
*bl
, *free_bl
= NULL
;
558 if (copy_from_user(®
, arg
, sizeof(reg
)))
561 if (reg
.resv
[0] || reg
.resv
[1] || reg
.resv
[2])
563 if (reg
.flags
& ~IOU_PBUF_RING_MMAP
)
565 if (!(reg
.flags
& IOU_PBUF_RING_MMAP
)) {
568 if (reg
.ring_addr
& ~PAGE_MASK
)
575 if (!is_power_of_2(reg
.ring_entries
))
578 /* cannot disambiguate full vs empty due to head/tail size */
579 if (reg
.ring_entries
>= 65536)
582 if (unlikely(reg
.bgid
< BGID_ARRAY
&& !ctx
->io_bl
)) {
583 int ret
= io_init_bl_list(ctx
);
588 bl
= io_buffer_get_list(ctx
, reg
.bgid
);
590 /* if mapped buffer ring OR classic exists, don't allow */
591 if (bl
->is_mapped
|| !list_empty(&bl
->buf_list
))
594 free_bl
= bl
= kzalloc(sizeof(*bl
), GFP_KERNEL
);
599 if (!(reg
.flags
& IOU_PBUF_RING_MMAP
))
600 ret
= io_pin_pbuf_ring(®
, bl
);
602 ret
= io_alloc_pbuf_ring(®
, bl
);
605 bl
->nr_entries
= reg
.ring_entries
;
606 bl
->mask
= reg
.ring_entries
- 1;
608 io_buffer_add_list(ctx
, bl
, reg
.bgid
);
616 int io_unregister_pbuf_ring(struct io_ring_ctx
*ctx
, void __user
*arg
)
618 struct io_uring_buf_reg reg
;
619 struct io_buffer_list
*bl
;
621 if (copy_from_user(®
, arg
, sizeof(reg
)))
623 if (reg
.resv
[0] || reg
.resv
[1] || reg
.resv
[2])
628 bl
= io_buffer_get_list(ctx
, reg
.bgid
);
634 __io_remove_buffers(ctx
, bl
, -1U);
635 if (bl
->bgid
>= BGID_ARRAY
) {
636 xa_erase(&ctx
->io_bl_xa
, bl
->bgid
);
642 void *io_pbuf_get_address(struct io_ring_ctx
*ctx
, unsigned long bgid
)
644 struct io_buffer_list
*bl
;
646 bl
= io_buffer_get_list(ctx
, bgid
);
647 if (!bl
|| !bl
->is_mmap
)