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/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#define BGID_ARRAY 64
21
22struct io_provide_buf {
23 struct file *file;
24 __u64 addr;
25 __u32 len;
26 __u32 bgid;
27 __u16 nbufs;
28 __u16 bid;
29};
30
31static inline struct io_buffer_list *io_buffer_get_list(struct io_ring_ctx *ctx,
32 unsigned int bgid)
33{
34 if (ctx->io_bl && bgid < BGID_ARRAY)
35 return &ctx->io_bl[bgid];
36
37 return xa_load(&ctx->io_bl_xa, bgid);
38}
39
40static int io_buffer_add_list(struct io_ring_ctx *ctx,
41 struct io_buffer_list *bl, unsigned int bgid)
42{
43 bl->bgid = bgid;
44 if (bgid < BGID_ARRAY)
45 return 0;
46
47 return xa_err(xa_store(&ctx->io_bl_xa, bgid, bl, GFP_KERNEL));
48}
49
50void io_kbuf_recycle_legacy(struct io_kiocb *req, unsigned issue_flags)
51{
52 struct io_ring_ctx *ctx = req->ctx;
53 struct io_buffer_list *bl;
54 struct io_buffer *buf;
55
56 /*
57 * For legacy provided buffer mode, don't recycle if we already did
58 * IO to this buffer. For ring-mapped provided buffer mode, we should
59 * increment ring->head to explicitly monopolize the buffer to avoid
60 * multiple use.
61 */
62 if (req->flags & REQ_F_PARTIAL_IO)
63 return;
64
65 io_ring_submit_lock(ctx, issue_flags);
66
67 buf = req->kbuf;
68 bl = io_buffer_get_list(ctx, buf->bgid);
69 list_add(&buf->list, &bl->buf_list);
70 req->flags &= ~REQ_F_BUFFER_SELECTED;
71 req->buf_index = buf->bgid;
72
73 io_ring_submit_unlock(ctx, issue_flags);
74 return;
75}
76
77unsigned int __io_put_kbuf(struct io_kiocb *req, unsigned issue_flags)
78{
79 unsigned int cflags;
80
81 /*
82 * We can add this buffer back to two lists:
83 *
84 * 1) The io_buffers_cache list. This one is protected by the
85 * ctx->uring_lock. If we already hold this lock, add back to this
86 * list as we can grab it from issue as well.
87 * 2) The io_buffers_comp list. This one is protected by the
88 * ctx->completion_lock.
89 *
90 * We migrate buffers from the comp_list to the issue cache list
91 * when we need one.
92 */
93 if (req->flags & REQ_F_BUFFER_RING) {
94 /* no buffers to recycle for this case */
95 cflags = __io_put_kbuf_list(req, NULL);
96 } else if (issue_flags & IO_URING_F_UNLOCKED) {
97 struct io_ring_ctx *ctx = req->ctx;
98
99 spin_lock(&ctx->completion_lock);
100 cflags = __io_put_kbuf_list(req, &ctx->io_buffers_comp);
101 spin_unlock(&ctx->completion_lock);
102 } else {
103 lockdep_assert_held(&req->ctx->uring_lock);
104
105 cflags = __io_put_kbuf_list(req, &req->ctx->io_buffers_cache);
106 }
107 return cflags;
108}
109
110static void __user *io_provided_buffer_select(struct io_kiocb *req, size_t *len,
111 struct io_buffer_list *bl)
112{
113 if (!list_empty(&bl->buf_list)) {
114 struct io_buffer *kbuf;
115
116 kbuf = list_first_entry(&bl->buf_list, struct io_buffer, list);
117 list_del(&kbuf->list);
118 if (*len == 0 || *len > kbuf->len)
119 *len = kbuf->len;
120 req->flags |= REQ_F_BUFFER_SELECTED;
121 req->kbuf = kbuf;
122 req->buf_index = kbuf->bid;
123 return u64_to_user_ptr(kbuf->addr);
124 }
125 return NULL;
126}
127
128static void __user *io_ring_buffer_select(struct io_kiocb *req, size_t *len,
129 struct io_buffer_list *bl,
130 unsigned int issue_flags)
131{
132 struct io_uring_buf_ring *br = bl->buf_ring;
133 struct io_uring_buf *buf;
134 __u16 head = bl->head;
135
136 if (unlikely(smp_load_acquire(&br->tail) == head))
137 return NULL;
138
139 head &= bl->mask;
140 if (head < IO_BUFFER_LIST_BUF_PER_PAGE) {
141 buf = &br->bufs[head];
142 } else {
143 int off = head & (IO_BUFFER_LIST_BUF_PER_PAGE - 1);
144 int index = head / IO_BUFFER_LIST_BUF_PER_PAGE;
145 buf = page_address(bl->buf_pages[index]);
146 buf += off;
147 }
148 if (*len == 0 || *len > buf->len)
149 *len = buf->len;
150 req->flags |= REQ_F_BUFFER_RING;
151 req->buf_list = bl;
152 req->buf_index = buf->bid;
153
154 if (issue_flags & IO_URING_F_UNLOCKED || !file_can_poll(req->file)) {
155 /*
156 * If we came in unlocked, we have no choice but to consume the
157 * buffer here, otherwise nothing ensures that the buffer won't
158 * get used by others. This does mean it'll be pinned until the
159 * IO completes, coming in unlocked means we're being called from
160 * io-wq context and there may be further retries in async hybrid
161 * mode. For the locked case, the caller must call commit when
162 * the transfer completes (or if we get -EAGAIN and must poll of
163 * retry).
164 */
165 req->buf_list = NULL;
166 bl->head++;
167 }
168 return u64_to_user_ptr(buf->addr);
169}
170
171void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
172 unsigned int issue_flags)
173{
174 struct io_ring_ctx *ctx = req->ctx;
175 struct io_buffer_list *bl;
176 void __user *ret = NULL;
177
178 io_ring_submit_lock(req->ctx, issue_flags);
179
180 bl = io_buffer_get_list(ctx, req->buf_index);
181 if (likely(bl)) {
182 if (bl->buf_nr_pages)
183 ret = io_ring_buffer_select(req, len, bl, issue_flags);
184 else
185 ret = io_provided_buffer_select(req, len, bl);
186 }
187 io_ring_submit_unlock(req->ctx, issue_flags);
188 return ret;
189}
190
191static __cold int io_init_bl_list(struct io_ring_ctx *ctx)
192{
193 int i;
194
195 ctx->io_bl = kcalloc(BGID_ARRAY, sizeof(struct io_buffer_list),
196 GFP_KERNEL);
197 if (!ctx->io_bl)
198 return -ENOMEM;
199
200 for (i = 0; i < BGID_ARRAY; i++) {
201 INIT_LIST_HEAD(&ctx->io_bl[i].buf_list);
202 ctx->io_bl[i].bgid = i;
203 }
204
205 return 0;
206}
207
208static int __io_remove_buffers(struct io_ring_ctx *ctx,
209 struct io_buffer_list *bl, unsigned nbufs)
210{
211 unsigned i = 0;
212
213 /* shouldn't happen */
214 if (!nbufs)
215 return 0;
216
217 if (bl->buf_nr_pages) {
218 int j;
219
220 i = bl->buf_ring->tail - bl->head;
221 for (j = 0; j < bl->buf_nr_pages; j++)
222 unpin_user_page(bl->buf_pages[j]);
223 kvfree(bl->buf_pages);
224 bl->buf_pages = NULL;
225 bl->buf_nr_pages = 0;
226 /* make sure it's seen as empty */
227 INIT_LIST_HEAD(&bl->buf_list);
228 return i;
229 }
230
231 /* the head kbuf is the list itself */
232 while (!list_empty(&bl->buf_list)) {
233 struct io_buffer *nxt;
234
235 nxt = list_first_entry(&bl->buf_list, struct io_buffer, list);
236 list_del(&nxt->list);
237 if (++i == nbufs)
238 return i;
239 cond_resched();
240 }
241 i++;
242
243 return i;
244}
245
246void io_destroy_buffers(struct io_ring_ctx *ctx)
247{
248 struct io_buffer_list *bl;
249 unsigned long index;
250 int i;
251
252 for (i = 0; i < BGID_ARRAY; i++) {
253 if (!ctx->io_bl)
254 break;
255 __io_remove_buffers(ctx, &ctx->io_bl[i], -1U);
256 }
257
258 xa_for_each(&ctx->io_bl_xa, index, bl) {
259 xa_erase(&ctx->io_bl_xa, bl->bgid);
260 __io_remove_buffers(ctx, bl, -1U);
261 kfree(bl);
262 }
263
264 while (!list_empty(&ctx->io_buffers_pages)) {
265 struct page *page;
266
267 page = list_first_entry(&ctx->io_buffers_pages, struct page, lru);
268 list_del_init(&page->lru);
269 __free_page(page);
270 }
271}
272
273int io_remove_buffers_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
274{
275 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
276 u64 tmp;
277
278 if (sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
279 sqe->splice_fd_in)
280 return -EINVAL;
281
282 tmp = READ_ONCE(sqe->fd);
283 if (!tmp || tmp > USHRT_MAX)
284 return -EINVAL;
285
286 memset(p, 0, sizeof(*p));
287 p->nbufs = tmp;
288 p->bgid = READ_ONCE(sqe->buf_group);
289 return 0;
290}
291
292int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
293{
294 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
295 struct io_ring_ctx *ctx = req->ctx;
296 struct io_buffer_list *bl;
297 int ret = 0;
298
299 io_ring_submit_lock(ctx, issue_flags);
300
301 ret = -ENOENT;
302 bl = io_buffer_get_list(ctx, p->bgid);
303 if (bl) {
304 ret = -EINVAL;
305 /* can't use provide/remove buffers command on mapped buffers */
306 if (!bl->buf_nr_pages)
307 ret = __io_remove_buffers(ctx, bl, p->nbufs);
308 }
309 io_ring_submit_unlock(ctx, issue_flags);
310 if (ret < 0)
311 req_set_fail(req);
312 io_req_set_res(req, ret, 0);
313 return IOU_OK;
314}
315
316int io_provide_buffers_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
317{
318 unsigned long size, tmp_check;
319 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
320 u64 tmp;
321
322 if (sqe->rw_flags || sqe->splice_fd_in)
323 return -EINVAL;
324
325 tmp = READ_ONCE(sqe->fd);
326 if (!tmp || tmp > USHRT_MAX)
327 return -E2BIG;
328 p->nbufs = tmp;
329 p->addr = READ_ONCE(sqe->addr);
330 p->len = READ_ONCE(sqe->len);
331
332 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
333 &size))
334 return -EOVERFLOW;
335 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
336 return -EOVERFLOW;
337
338 size = (unsigned long)p->len * p->nbufs;
339 if (!access_ok(u64_to_user_ptr(p->addr), size))
340 return -EFAULT;
341
342 p->bgid = READ_ONCE(sqe->buf_group);
343 tmp = READ_ONCE(sqe->off);
344 if (tmp > USHRT_MAX)
345 return -E2BIG;
346 if (tmp + p->nbufs >= USHRT_MAX)
347 return -EINVAL;
348 p->bid = tmp;
349 return 0;
350}
351
352static int io_refill_buffer_cache(struct io_ring_ctx *ctx)
353{
354 struct io_buffer *buf;
355 struct page *page;
356 int bufs_in_page;
357
358 /*
359 * Completions that don't happen inline (eg not under uring_lock) will
360 * add to ->io_buffers_comp. If we don't have any free buffers, check
361 * the completion list and splice those entries first.
362 */
363 if (!list_empty_careful(&ctx->io_buffers_comp)) {
364 spin_lock(&ctx->completion_lock);
365 if (!list_empty(&ctx->io_buffers_comp)) {
366 list_splice_init(&ctx->io_buffers_comp,
367 &ctx->io_buffers_cache);
368 spin_unlock(&ctx->completion_lock);
369 return 0;
370 }
371 spin_unlock(&ctx->completion_lock);
372 }
373
374 /*
375 * No free buffers and no completion entries either. Allocate a new
376 * page worth of buffer entries and add those to our freelist.
377 */
378 page = alloc_page(GFP_KERNEL_ACCOUNT);
379 if (!page)
380 return -ENOMEM;
381
382 list_add(&page->lru, &ctx->io_buffers_pages);
383
384 buf = page_address(page);
385 bufs_in_page = PAGE_SIZE / sizeof(*buf);
386 while (bufs_in_page) {
387 list_add_tail(&buf->list, &ctx->io_buffers_cache);
388 buf++;
389 bufs_in_page--;
390 }
391
392 return 0;
393}
394
395static int io_add_buffers(struct io_ring_ctx *ctx, struct io_provide_buf *pbuf,
396 struct io_buffer_list *bl)
397{
398 struct io_buffer *buf;
399 u64 addr = pbuf->addr;
400 int i, bid = pbuf->bid;
401
402 for (i = 0; i < pbuf->nbufs; i++) {
403 if (list_empty(&ctx->io_buffers_cache) &&
404 io_refill_buffer_cache(ctx))
405 break;
406 buf = list_first_entry(&ctx->io_buffers_cache, struct io_buffer,
407 list);
408 list_move_tail(&buf->list, &bl->buf_list);
409 buf->addr = addr;
410 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
411 buf->bid = bid;
412 buf->bgid = pbuf->bgid;
413 addr += pbuf->len;
414 bid++;
415 cond_resched();
416 }
417
418 return i ? 0 : -ENOMEM;
419}
420
421int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
422{
423 struct io_provide_buf *p = io_kiocb_to_cmd(req, struct io_provide_buf);
424 struct io_ring_ctx *ctx = req->ctx;
425 struct io_buffer_list *bl;
426 int ret = 0;
427
428 io_ring_submit_lock(ctx, issue_flags);
429
430 if (unlikely(p->bgid < BGID_ARRAY && !ctx->io_bl)) {
431 ret = io_init_bl_list(ctx);
432 if (ret)
433 goto err;
434 }
435
436 bl = io_buffer_get_list(ctx, p->bgid);
437 if (unlikely(!bl)) {
438 bl = kzalloc(sizeof(*bl), GFP_KERNEL_ACCOUNT);
439 if (!bl) {
440 ret = -ENOMEM;
441 goto err;
442 }
443 INIT_LIST_HEAD(&bl->buf_list);
444 ret = io_buffer_add_list(ctx, bl, p->bgid);
445 if (ret) {
446 kfree(bl);
447 goto err;
448 }
449 }
450 /* can't add buffers via this command for a mapped buffer ring */
451 if (bl->buf_nr_pages) {
452 ret = -EINVAL;
453 goto err;
454 }
455
456 ret = io_add_buffers(ctx, p, bl);
457err:
458 io_ring_submit_unlock(ctx, issue_flags);
459
460 if (ret < 0)
461 req_set_fail(req);
462 io_req_set_res(req, ret, 0);
463 return IOU_OK;
464}
465
466int io_register_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
467{
468 struct io_uring_buf_ring *br;
469 struct io_uring_buf_reg reg;
470 struct io_buffer_list *bl, *free_bl = NULL;
471 struct page **pages;
472 int nr_pages;
473
474 if (copy_from_user(®, arg, sizeof(reg)))
475 return -EFAULT;
476
477 if (reg.pad || reg.resv[0] || reg.resv[1] || reg.resv[2])
478 return -EINVAL;
479 if (!reg.ring_addr)
480 return -EFAULT;
481 if (reg.ring_addr & ~PAGE_MASK)
482 return -EINVAL;
483 if (!is_power_of_2(reg.ring_entries))
484 return -EINVAL;
485
486 /* cannot disambiguate full vs empty due to head/tail size */
487 if (reg.ring_entries >= 65536)
488 return -EINVAL;
489
490 if (unlikely(reg.bgid < BGID_ARRAY && !ctx->io_bl)) {
491 int ret = io_init_bl_list(ctx);
492 if (ret)
493 return ret;
494 }
495
496 bl = io_buffer_get_list(ctx, reg.bgid);
497 if (bl) {
498 /* if mapped buffer ring OR classic exists, don't allow */
499 if (bl->buf_nr_pages || !list_empty(&bl->buf_list))
500 return -EEXIST;
501 } else {
502 free_bl = bl = kzalloc(sizeof(*bl), GFP_KERNEL);
503 if (!bl)
504 return -ENOMEM;
505 }
506
507 pages = io_pin_pages(reg.ring_addr,
508 struct_size(br, bufs, reg.ring_entries),
509 &nr_pages);
510 if (IS_ERR(pages)) {
511 kfree(free_bl);
512 return PTR_ERR(pages);
513 }
514
515 br = page_address(pages[0]);
516 bl->buf_pages = pages;
517 bl->buf_nr_pages = nr_pages;
518 bl->nr_entries = reg.ring_entries;
519 bl->buf_ring = br;
520 bl->mask = reg.ring_entries - 1;
521 io_buffer_add_list(ctx, bl, reg.bgid);
522 return 0;
523}
524
525int io_unregister_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
526{
527 struct io_uring_buf_reg reg;
528 struct io_buffer_list *bl;
529
530 if (copy_from_user(®, arg, sizeof(reg)))
531 return -EFAULT;
532 if (reg.pad || reg.resv[0] || reg.resv[1] || reg.resv[2])
533 return -EINVAL;
534
535 bl = io_buffer_get_list(ctx, reg.bgid);
536 if (!bl)
537 return -ENOENT;
538 if (!bl->buf_nr_pages)
539 return -EINVAL;
540
541 __io_remove_buffers(ctx, bl, -1U);
542 if (bl->bgid >= BGID_ARRAY) {
543 xa_erase(&ctx->io_bl_xa, bl->bgid);
544 kfree(bl);
545 }
546 return 0;
547}
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}