1 // SPDX-License-Identifier: GPL-2.0
2
3 #include <net/xsk_buff_pool.h>
4 #include <net/xdp_sock.h>
5 #include <net/xdp_sock_drv.h>
6
7 #include "xsk_queue.h"
8 #include "xdp_umem.h"
9 #include "xsk.h"
10
11 void xp_add_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs)
12 {
13 unsigned long flags;
14
15 if (!xs->tx)
16 return;
17
18 spin_lock_irqsave(&pool->xsk_tx_list_lock, flags);
19 list_add_rcu(&xs->tx_list, &pool->xsk_tx_list);
20 spin_unlock_irqrestore(&pool->xsk_tx_list_lock, flags);
21 }
22
23 void xp_del_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs)
24 {
25 unsigned long flags;
26
27 if (!xs->tx)
28 return;
29
30 spin_lock_irqsave(&pool->xsk_tx_list_lock, flags);
31 list_del_rcu(&xs->tx_list);
32 spin_unlock_irqrestore(&pool->xsk_tx_list_lock, flags);
33 }
34
35 void xp_destroy(struct xsk_buff_pool *pool)
36 {
37 if (!pool)
38 return;
39
40 kvfree(pool->tx_descs);
41 kvfree(pool->heads);
42 kvfree(pool);
43 }
44
45 int xp_alloc_tx_descs(struct xsk_buff_pool *pool, struct xdp_sock *xs)
46 {
47 pool->tx_descs = kvcalloc(xs->tx->nentries, sizeof(*pool->tx_descs),
48 GFP_KERNEL);
49 if (!pool->tx_descs)
50 return -ENOMEM;
51
52 return 0;
53 }
54
55 struct xsk_buff_pool *xp_create_and_assign_umem(struct xdp_sock *xs,
56 struct xdp_umem *umem)
57 {
58 bool unaligned = umem->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG;
59 struct xsk_buff_pool *pool;
60 struct xdp_buff_xsk *xskb;
61 u32 i, entries;
62
63 entries = unaligned ? umem->chunks : 0;
64 pool = kvzalloc(struct_size(pool, free_heads, entries), GFP_KERNEL);
65 if (!pool)
66 goto out;
67
68 pool->heads = kvcalloc(umem->chunks, sizeof(*pool->heads), GFP_KERNEL);
69 if (!pool->heads)
70 goto out;
71
72 if (xs->tx)
73 if (xp_alloc_tx_descs(pool, xs))
74 goto out;
75
76 pool->chunk_mask = ~((u64)umem->chunk_size - 1);
77 pool->addrs_cnt = umem->size;
78 pool->heads_cnt = umem->chunks;
79 pool->free_heads_cnt = umem->chunks;
80 pool->headroom = umem->headroom;
81 pool->chunk_size = umem->chunk_size;
82 pool->chunk_shift = ffs(umem->chunk_size) - 1;
83 pool->unaligned = unaligned;
84 pool->frame_len = umem->chunk_size - umem->headroom -
85 XDP_PACKET_HEADROOM;
86 pool->umem = umem;
87 pool->addrs = umem->addrs;
88 pool->tx_metadata_len = umem->tx_metadata_len;
89 pool->tx_sw_csum = umem->flags & XDP_UMEM_TX_SW_CSUM;
90 INIT_LIST_HEAD(&pool->free_list);
91 INIT_LIST_HEAD(&pool->xskb_list);
92 INIT_LIST_HEAD(&pool->xsk_tx_list);
93 spin_lock_init(&pool->xsk_tx_list_lock);
94 spin_lock_init(&pool->cq_lock);
95 refcount_set(&pool->users, 1);
96
97 pool->fq = xs->fq_tmp;
98 pool->cq = xs->cq_tmp;
99
100 for (i = 0; i < pool->free_heads_cnt; i++) {
101 xskb = &pool->heads[i];
102 xskb->pool = pool;
103 xskb->xdp.frame_sz = umem->chunk_size - umem->headroom;
104 INIT_LIST_HEAD(&xskb->free_list_node);
105 INIT_LIST_HEAD(&xskb->xskb_list_node);
106 if (pool->unaligned)
107 pool->free_heads[i] = xskb;
108 else
109 xp_init_xskb_addr(xskb, pool, i * pool->chunk_size);
110 }
111
112 return pool;
113
114 out:
115 xp_destroy(pool);
116 return NULL;
117 }
118
119 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq)
120 {
121 u32 i;
122
123 for (i = 0; i < pool->heads_cnt; i++)
124 pool->heads[i].xdp.rxq = rxq;
125 }
126 EXPORT_SYMBOL(xp_set_rxq_info);
127
128 void xp_fill_cb(struct xsk_buff_pool *pool, struct xsk_cb_desc *desc)
129 {
130 u32 i;
131
132 for (i = 0; i < pool->heads_cnt; i++) {
133 struct xdp_buff_xsk *xskb = &pool->heads[i];
134
135 memcpy(xskb->cb + desc->off, desc->src, desc->bytes);
136 }
137 }
138 EXPORT_SYMBOL(xp_fill_cb);
139
140 static void xp_disable_drv_zc(struct xsk_buff_pool *pool)
141 {
142 struct netdev_bpf bpf;
143 int err;
144
145 ASSERT_RTNL();
146
147 if (pool->umem->zc) {
148 bpf.command = XDP_SETUP_XSK_POOL;
149 bpf.xsk.pool = NULL;
150 bpf.xsk.queue_id = pool->queue_id;
151
152 err = pool->netdev->netdev_ops->ndo_bpf(pool->netdev, &bpf);
153
154 if (err)
155 WARN(1, "Failed to disable zero-copy!\n");
156 }
157 }
158
159 #define NETDEV_XDP_ACT_ZC (NETDEV_XDP_ACT_BASIC | \
160 NETDEV_XDP_ACT_REDIRECT | \
161 NETDEV_XDP_ACT_XSK_ZEROCOPY)
162
163 int xp_assign_dev(struct xsk_buff_pool *pool,
164 struct net_device *netdev, u16 queue_id, u16 flags)
165 {
166 bool force_zc, force_copy;
167 struct netdev_bpf bpf;
168 int err = 0;
169
170 ASSERT_RTNL();
171
172 force_zc = flags & XDP_ZEROCOPY;
173 force_copy = flags & XDP_COPY;
174
175 if (force_zc && force_copy)
176 return -EINVAL;
177
178 if (xsk_get_pool_from_qid(netdev, queue_id))
179 return -EBUSY;
180
181 pool->netdev = netdev;
182 pool->queue_id = queue_id;
183 err = xsk_reg_pool_at_qid(netdev, pool, queue_id);
184 if (err)
185 return err;
186
187 if (flags & XDP_USE_SG)
188 pool->umem->flags |= XDP_UMEM_SG_FLAG;
189
190 if (flags & XDP_USE_NEED_WAKEUP)
191 pool->uses_need_wakeup = true;
192 /* Tx needs to be explicitly woken up the first time. Also
193 * for supporting drivers that do not implement this
194 * feature. They will always have to call sendto() or poll().
195 */
196 pool->cached_need_wakeup = XDP_WAKEUP_TX;
197
198 dev_hold(netdev);
199
200 if (force_copy)
201 /* For copy-mode, we are done. */
202 return 0;
203
204 if ((netdev->xdp_features & NETDEV_XDP_ACT_ZC) != NETDEV_XDP_ACT_ZC) {
205 err = -EOPNOTSUPP;
206 goto err_unreg_pool;
207 }
208
209 if (netdev->xdp_zc_max_segs == 1 && (flags & XDP_USE_SG)) {
210 err = -EOPNOTSUPP;
211 goto err_unreg_pool;
212 }
213
214 bpf.command = XDP_SETUP_XSK_POOL;
215 bpf.xsk.pool = pool;
216 bpf.xsk.queue_id = queue_id;
217
218 err = netdev->netdev_ops->ndo_bpf(netdev, &bpf);
219 if (err)
220 goto err_unreg_pool;
221
222 if (!pool->dma_pages) {
223 WARN(1, "Driver did not DMA map zero-copy buffers");
224 err = -EINVAL;
225 goto err_unreg_xsk;
226 }
227 pool->umem->zc = true;
228 return 0;
229
230 err_unreg_xsk:
231 xp_disable_drv_zc(pool);
232 err_unreg_pool:
233 if (!force_zc)
234 err = 0; /* fallback to copy mode */
235 if (err) {
236 xsk_clear_pool_at_qid(netdev, queue_id);
237 dev_put(netdev);
238 }
239 return err;
240 }
241
242 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs,
243 struct net_device *dev, u16 queue_id)
244 {
245 u16 flags;
246 struct xdp_umem *umem = umem_xs->umem;
247
248 /* One fill and completion ring required for each queue id. */
249 if (!pool->fq || !pool->cq)
250 return -EINVAL;
251
252 flags = umem->zc ? XDP_ZEROCOPY : XDP_COPY;
253 if (umem_xs->pool->uses_need_wakeup)
254 flags |= XDP_USE_NEED_WAKEUP;
255
256 return xp_assign_dev(pool, dev, queue_id, flags);
257 }
258
259 void xp_clear_dev(struct xsk_buff_pool *pool)
260 {
261 if (!pool->netdev)
262 return;
263
264 xp_disable_drv_zc(pool);
265 xsk_clear_pool_at_qid(pool->netdev, pool->queue_id);
266 dev_put(pool->netdev);
267 pool->netdev = NULL;
268 }
269
270 static void xp_release_deferred(struct work_struct *work)
271 {
272 struct xsk_buff_pool *pool = container_of(work, struct xsk_buff_pool,
273 work);
274
275 rtnl_lock();
276 xp_clear_dev(pool);
277 rtnl_unlock();
278
279 if (pool->fq) {
280 xskq_destroy(pool->fq);
281 pool->fq = NULL;
282 }
283
284 if (pool->cq) {
285 xskq_destroy(pool->cq);
286 pool->cq = NULL;
287 }
288
289 xdp_put_umem(pool->umem, false);
290 xp_destroy(pool);
291 }
292
293 void xp_get_pool(struct xsk_buff_pool *pool)
294 {
295 refcount_inc(&pool->users);
296 }
297
298 bool xp_put_pool(struct xsk_buff_pool *pool)
299 {
300 if (!pool)
301 return false;
302
303 if (refcount_dec_and_test(&pool->users)) {
304 INIT_WORK(&pool->work, xp_release_deferred);
305 schedule_work(&pool->work);
306 return true;
307 }
308
309 return false;
310 }
311
312 static struct xsk_dma_map *xp_find_dma_map(struct xsk_buff_pool *pool)
313 {
314 struct xsk_dma_map *dma_map;
315
316 list_for_each_entry(dma_map, &pool->umem->xsk_dma_list, list) {
317 if (dma_map->netdev == pool->netdev)
318 return dma_map;
319 }
320
321 return NULL;
322 }
323
324 static struct xsk_dma_map *xp_create_dma_map(struct device *dev, struct net_device *netdev,
325 u32 nr_pages, struct xdp_umem *umem)
326 {
327 struct xsk_dma_map *dma_map;
328
329 dma_map = kzalloc(sizeof(*dma_map), GFP_KERNEL);
330 if (!dma_map)
331 return NULL;
332
333 dma_map->dma_pages = kvcalloc(nr_pages, sizeof(*dma_map->dma_pages), GFP_KERNEL);
334 if (!dma_map->dma_pages) {
335 kfree(dma_map);
336 return NULL;
337 }
338
339 dma_map->netdev = netdev;
340 dma_map->dev = dev;
341 dma_map->dma_pages_cnt = nr_pages;
342 refcount_set(&dma_map->users, 1);
343 list_add(&dma_map->list, &umem->xsk_dma_list);
344 return dma_map;
345 }
346
347 static void xp_destroy_dma_map(struct xsk_dma_map *dma_map)
348 {
349 list_del(&dma_map->list);
350 kvfree(dma_map->dma_pages);
351 kfree(dma_map);
352 }
353
354 static void __xp_dma_unmap(struct xsk_dma_map *dma_map, unsigned long attrs)
355 {
356 dma_addr_t *dma;
357 u32 i;
358
359 for (i = 0; i < dma_map->dma_pages_cnt; i++) {
360 dma = &dma_map->dma_pages[i];
361 if (*dma) {
362 *dma &= ~XSK_NEXT_PG_CONTIG_MASK;
363 dma_unmap_page_attrs(dma_map->dev, *dma, PAGE_SIZE,
364 DMA_BIDIRECTIONAL, attrs);
365 *dma = 0;
366 }
367 }
368
369 xp_destroy_dma_map(dma_map);
370 }
371
372 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs)
373 {
374 struct xsk_dma_map *dma_map;
375
376 if (!pool->dma_pages)
377 return;
378
379 dma_map = xp_find_dma_map(pool);
380 if (!dma_map) {
381 WARN(1, "Could not find dma_map for device");
382 return;
383 }
384
385 if (!refcount_dec_and_test(&dma_map->users))
386 return;
387
388 __xp_dma_unmap(dma_map, attrs);
389 kvfree(pool->dma_pages);
390 pool->dma_pages = NULL;
391 pool->dma_pages_cnt = 0;
392 pool->dev = NULL;
393 }
394 EXPORT_SYMBOL(xp_dma_unmap);
395
396 static void xp_check_dma_contiguity(struct xsk_dma_map *dma_map)
397 {
398 u32 i;
399
400 for (i = 0; i < dma_map->dma_pages_cnt - 1; i++) {
401 if (dma_map->dma_pages[i] + PAGE_SIZE == dma_map->dma_pages[i + 1])
402 dma_map->dma_pages[i] |= XSK_NEXT_PG_CONTIG_MASK;
403 else
404 dma_map->dma_pages[i] &= ~XSK_NEXT_PG_CONTIG_MASK;
405 }
406 }
407
408 static int xp_init_dma_info(struct xsk_buff_pool *pool, struct xsk_dma_map *dma_map)
409 {
410 if (!pool->unaligned) {
411 u32 i;
412
413 for (i = 0; i < pool->heads_cnt; i++) {
414 struct xdp_buff_xsk *xskb = &pool->heads[i];
415
416 xp_init_xskb_dma(xskb, pool, dma_map->dma_pages, xskb->orig_addr);
417 }
418 }
419
420 pool->dma_pages = kvcalloc(dma_map->dma_pages_cnt, sizeof(*pool->dma_pages), GFP_KERNEL);
421 if (!pool->dma_pages)
422 return -ENOMEM;
423
424 pool->dev = dma_map->dev;
425 pool->dma_pages_cnt = dma_map->dma_pages_cnt;
426 memcpy(pool->dma_pages, dma_map->dma_pages,
427 pool->dma_pages_cnt * sizeof(*pool->dma_pages));
428
429 return 0;
430 }
431
432 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
433 unsigned long attrs, struct page **pages, u32 nr_pages)
434 {
435 struct xsk_dma_map *dma_map;
436 dma_addr_t dma;
437 int err;
438 u32 i;
439
440 dma_map = xp_find_dma_map(pool);
441 if (dma_map) {
442 err = xp_init_dma_info(pool, dma_map);
443 if (err)
444 return err;
445
446 refcount_inc(&dma_map->users);
447 return 0;
448 }
449
450 dma_map = xp_create_dma_map(dev, pool->netdev, nr_pages, pool->umem);
451 if (!dma_map)
452 return -ENOMEM;
453
454 for (i = 0; i < dma_map->dma_pages_cnt; i++) {
455 dma = dma_map_page_attrs(dev, pages[i], 0, PAGE_SIZE,
456 DMA_BIDIRECTIONAL, attrs);
457 if (dma_mapping_error(dev, dma)) {
458 __xp_dma_unmap(dma_map, attrs);
459 return -ENOMEM;
460 }
461 dma_map->dma_pages[i] = dma;
462 }
463
464 if (pool->unaligned)
465 xp_check_dma_contiguity(dma_map);
466
467 err = xp_init_dma_info(pool, dma_map);
468 if (err) {
469 __xp_dma_unmap(dma_map, attrs);
470 return err;
471 }
472
473 return 0;
474 }
475 EXPORT_SYMBOL(xp_dma_map);
476
477 static bool xp_addr_crosses_non_contig_pg(struct xsk_buff_pool *pool,
478 u64 addr)
479 {
480 return xp_desc_crosses_non_contig_pg(pool, addr, pool->chunk_size);
481 }
482
483 static bool xp_check_unaligned(struct xsk_buff_pool *pool, u64 *addr)
484 {
485 *addr = xp_unaligned_extract_addr(*addr);
486 if (*addr >= pool->addrs_cnt ||
487 *addr + pool->chunk_size > pool->addrs_cnt ||
488 xp_addr_crosses_non_contig_pg(pool, *addr))
489 return false;
490 return true;
491 }
492
493 static bool xp_check_aligned(struct xsk_buff_pool *pool, u64 *addr)
494 {
495 *addr = xp_aligned_extract_addr(pool, *addr);
496 return *addr < pool->addrs_cnt;
497 }
498
499 static struct xdp_buff_xsk *__xp_alloc(struct xsk_buff_pool *pool)
500 {
501 struct xdp_buff_xsk *xskb;
502 u64 addr;
503 bool ok;
504
505 if (pool->free_heads_cnt == 0)
506 return NULL;
507
508 for (;;) {
509 if (!xskq_cons_peek_addr_unchecked(pool->fq, &addr)) {
510 pool->fq->queue_empty_descs++;
511 return NULL;
512 }
513
514 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
515 xp_check_aligned(pool, &addr);
516 if (!ok) {
517 pool->fq->invalid_descs++;
518 xskq_cons_release(pool->fq);
519 continue;
520 }
521 break;
522 }
523
524 if (pool->unaligned) {
525 xskb = pool->free_heads[--pool->free_heads_cnt];
526 xp_init_xskb_addr(xskb, pool, addr);
527 if (pool->dma_pages)
528 xp_init_xskb_dma(xskb, pool, pool->dma_pages, addr);
529 } else {
530 xskb = &pool->heads[xp_aligned_extract_idx(pool, addr)];
531 }
532
533 xskq_cons_release(pool->fq);
534 return xskb;
535 }
536
537 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool)
538 {
539 struct xdp_buff_xsk *xskb;
540
541 if (!pool->free_list_cnt) {
542 xskb = __xp_alloc(pool);
543 if (!xskb)
544 return NULL;
545 } else {
546 pool->free_list_cnt--;
547 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk,
548 free_list_node);
549 list_del_init(&xskb->free_list_node);
550 }
551
552 xskb->xdp.data = xskb->xdp.data_hard_start + XDP_PACKET_HEADROOM;
553 xskb->xdp.data_meta = xskb->xdp.data;
554 xskb->xdp.flags = 0;
555
556 if (pool->dev)
557 xp_dma_sync_for_device(pool, xskb->dma, pool->frame_len);
558
559 return &xskb->xdp;
560 }
561 EXPORT_SYMBOL(xp_alloc);
562
563 static u32 xp_alloc_new_from_fq(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max)
564 {
565 u32 i, cached_cons, nb_entries;
566
567 if (max > pool->free_heads_cnt)
568 max = pool->free_heads_cnt;
569 max = xskq_cons_nb_entries(pool->fq, max);
570
571 cached_cons = pool->fq->cached_cons;
572 nb_entries = max;
573 i = max;
574 while (i--) {
575 struct xdp_buff_xsk *xskb;
576 u64 addr;
577 bool ok;
578
579 __xskq_cons_read_addr_unchecked(pool->fq, cached_cons++, &addr);
580
581 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
582 xp_check_aligned(pool, &addr);
583 if (unlikely(!ok)) {
584 pool->fq->invalid_descs++;
585 nb_entries--;
586 continue;
587 }
588
589 if (pool->unaligned) {
590 xskb = pool->free_heads[--pool->free_heads_cnt];
591 xp_init_xskb_addr(xskb, pool, addr);
592 if (pool->dma_pages)
593 xp_init_xskb_dma(xskb, pool, pool->dma_pages, addr);
594 } else {
595 xskb = &pool->heads[xp_aligned_extract_idx(pool, addr)];
596 }
597
598 *xdp = &xskb->xdp;
599 xdp++;
600 }
601
602 xskq_cons_release_n(pool->fq, max);
603 return nb_entries;
604 }
605
606 static u32 xp_alloc_reused(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 nb_entries)
607 {
608 struct xdp_buff_xsk *xskb;
609 u32 i;
610
611 nb_entries = min_t(u32, nb_entries, pool->free_list_cnt);
612
613 i = nb_entries;
614 while (i--) {
615 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk, free_list_node);
616 list_del_init(&xskb->free_list_node);
617
618 *xdp = &xskb->xdp;
619 xdp++;
620 }
621 pool->free_list_cnt -= nb_entries;
622
623 return nb_entries;
624 }
625
626 static u32 xp_alloc_slow(struct xsk_buff_pool *pool, struct xdp_buff **xdp,
627 u32 max)
628 {
629 int i;
630
631 for (i = 0; i < max; i++) {
632 struct xdp_buff *buff;
633
634 buff = xp_alloc(pool);
635 if (unlikely(!buff))
636 return i;
637 *xdp = buff;
638 xdp++;
639 }
640
641 return max;
642 }
643
644 u32 xp_alloc_batch(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max)
645 {
646 u32 nb_entries1 = 0, nb_entries2;
647
648 if (unlikely(pool->dev && dma_dev_need_sync(pool->dev)))
649 return xp_alloc_slow(pool, xdp, max);
650
651 if (unlikely(pool->free_list_cnt)) {
652 nb_entries1 = xp_alloc_reused(pool, xdp, max);
653 if (nb_entries1 == max)
654 return nb_entries1;
655
656 max -= nb_entries1;
657 xdp += nb_entries1;
658 }
659
660 nb_entries2 = xp_alloc_new_from_fq(pool, xdp, max);
661 if (!nb_entries2)
662 pool->fq->queue_empty_descs++;
663
664 return nb_entries1 + nb_entries2;
665 }
666 EXPORT_SYMBOL(xp_alloc_batch);
667
668 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count)
669 {
670 if (pool->free_list_cnt >= count)
671 return true;
672 return xskq_cons_has_entries(pool->fq, count - pool->free_list_cnt);
673 }
674 EXPORT_SYMBOL(xp_can_alloc);
675
676 void xp_free(struct xdp_buff_xsk *xskb)
677 {
678 if (!list_empty(&xskb->free_list_node))
679 return;
680
681 xskb->pool->free_list_cnt++;
682 list_add(&xskb->free_list_node, &xskb->pool->free_list);
683 }
684 EXPORT_SYMBOL(xp_free);
685
686 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
687 {
688 addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
689 return pool->addrs + addr;
690 }
691 EXPORT_SYMBOL(xp_raw_get_data);
692
693 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
694 {
695 addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
696 return (pool->dma_pages[addr >> PAGE_SHIFT] &
697 ~XSK_NEXT_PG_CONTIG_MASK) +
698 (addr & ~PAGE_MASK);
699 }
700 EXPORT_SYMBOL(xp_raw_get_dma);
701
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.