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
xp_add_xsk(struct xsk_buff_pool * pool,struct xdp_sock * xs)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
xp_del_xsk(struct xsk_buff_pool * pool,struct xdp_sock * xs)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
xp_destroy(struct xsk_buff_pool * pool)35 void xp_destroy(struct xsk_buff_pool *pool)
36 {
37 if (!pool)
38 return;
39
40 kvfree(pool->heads);
41 kvfree(pool);
42 }
43
xp_create_and_assign_umem(struct xdp_sock * xs,struct xdp_umem * umem)44 struct xsk_buff_pool *xp_create_and_assign_umem(struct xdp_sock *xs,
45 struct xdp_umem *umem)
46 {
47 struct xsk_buff_pool *pool;
48 struct xdp_buff_xsk *xskb;
49 u32 i;
50
51 pool = kvzalloc(struct_size(pool, free_heads, umem->chunks),
52 GFP_KERNEL);
53 if (!pool)
54 goto out;
55
56 pool->heads = kvcalloc(umem->chunks, sizeof(*pool->heads), GFP_KERNEL);
57 if (!pool->heads)
58 goto out;
59
60 pool->chunk_mask = ~((u64)umem->chunk_size - 1);
61 pool->addrs_cnt = umem->size;
62 pool->heads_cnt = umem->chunks;
63 pool->free_heads_cnt = umem->chunks;
64 pool->headroom = umem->headroom;
65 pool->chunk_size = umem->chunk_size;
66 pool->unaligned = umem->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG;
67 pool->frame_len = umem->chunk_size - umem->headroom -
68 XDP_PACKET_HEADROOM;
69 pool->umem = umem;
70 pool->addrs = umem->addrs;
71 INIT_LIST_HEAD(&pool->free_list);
72 INIT_LIST_HEAD(&pool->xsk_tx_list);
73 spin_lock_init(&pool->xsk_tx_list_lock);
74 spin_lock_init(&pool->cq_lock);
75 refcount_set(&pool->users, 1);
76
77 pool->fq = xs->fq_tmp;
78 pool->cq = xs->cq_tmp;
79
80 for (i = 0; i < pool->free_heads_cnt; i++) {
81 xskb = &pool->heads[i];
82 xskb->pool = pool;
83 xskb->xdp.frame_sz = umem->chunk_size - umem->headroom;
84 pool->free_heads[i] = xskb;
85 }
86
87 return pool;
88
89 out:
90 xp_destroy(pool);
91 return NULL;
92 }
93
xp_set_rxq_info(struct xsk_buff_pool * pool,struct xdp_rxq_info * rxq)94 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq)
95 {
96 u32 i;
97
98 for (i = 0; i < pool->heads_cnt; i++)
99 pool->heads[i].xdp.rxq = rxq;
100 }
101 EXPORT_SYMBOL(xp_set_rxq_info);
102
xp_disable_drv_zc(struct xsk_buff_pool * pool)103 static void xp_disable_drv_zc(struct xsk_buff_pool *pool)
104 {
105 struct netdev_bpf bpf;
106 int err;
107
108 ASSERT_RTNL();
109
110 if (pool->umem->zc) {
111 bpf.command = XDP_SETUP_XSK_POOL;
112 bpf.xsk.pool = NULL;
113 bpf.xsk.queue_id = pool->queue_id;
114
115 err = pool->netdev->netdev_ops->ndo_bpf(pool->netdev, &bpf);
116
117 if (err)
118 WARN(1, "Failed to disable zero-copy!\n");
119 }
120 }
121
xp_assign_dev(struct xsk_buff_pool * pool,struct net_device * netdev,u16 queue_id,u16 flags)122 int xp_assign_dev(struct xsk_buff_pool *pool,
123 struct net_device *netdev, u16 queue_id, u16 flags)
124 {
125 bool force_zc, force_copy;
126 struct netdev_bpf bpf;
127 int err = 0;
128
129 ASSERT_RTNL();
130
131 force_zc = flags & XDP_ZEROCOPY;
132 force_copy = flags & XDP_COPY;
133
134 if (force_zc && force_copy)
135 return -EINVAL;
136
137 if (xsk_get_pool_from_qid(netdev, queue_id))
138 return -EBUSY;
139
140 pool->netdev = netdev;
141 pool->queue_id = queue_id;
142 err = xsk_reg_pool_at_qid(netdev, pool, queue_id);
143 if (err)
144 return err;
145
146 if (flags & XDP_USE_NEED_WAKEUP)
147 pool->uses_need_wakeup = true;
148 /* Tx needs to be explicitly woken up the first time. Also
149 * for supporting drivers that do not implement this
150 * feature. They will always have to call sendto() or poll().
151 */
152 pool->cached_need_wakeup = XDP_WAKEUP_TX;
153
154 dev_hold(netdev);
155
156 if (force_copy)
157 /* For copy-mode, we are done. */
158 return 0;
159
160 if (!netdev->netdev_ops->ndo_bpf ||
161 !netdev->netdev_ops->ndo_xsk_wakeup) {
162 err = -EOPNOTSUPP;
163 goto err_unreg_pool;
164 }
165
166 bpf.command = XDP_SETUP_XSK_POOL;
167 bpf.xsk.pool = pool;
168 bpf.xsk.queue_id = queue_id;
169
170 err = netdev->netdev_ops->ndo_bpf(netdev, &bpf);
171 if (err)
172 goto err_unreg_pool;
173
174 if (!pool->dma_pages) {
175 WARN(1, "Driver did not DMA map zero-copy buffers");
176 err = -EINVAL;
177 goto err_unreg_xsk;
178 }
179 pool->umem->zc = true;
180 return 0;
181
182 err_unreg_xsk:
183 xp_disable_drv_zc(pool);
184 err_unreg_pool:
185 if (!force_zc)
186 err = 0; /* fallback to copy mode */
187 if (err) {
188 xsk_clear_pool_at_qid(netdev, queue_id);
189 dev_put(netdev);
190 }
191 return err;
192 }
193
xp_assign_dev_shared(struct xsk_buff_pool * pool,struct xdp_sock * umem_xs,struct net_device * dev,u16 queue_id)194 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs,
195 struct net_device *dev, u16 queue_id)
196 {
197 u16 flags;
198 struct xdp_umem *umem = umem_xs->umem;
199
200 /* One fill and completion ring required for each queue id. */
201 if (!pool->fq || !pool->cq)
202 return -EINVAL;
203
204 flags = umem->zc ? XDP_ZEROCOPY : XDP_COPY;
205 if (umem_xs->pool->uses_need_wakeup)
206 flags |= XDP_USE_NEED_WAKEUP;
207
208 return xp_assign_dev(pool, dev, queue_id, flags);
209 }
210
xp_clear_dev(struct xsk_buff_pool * pool)211 void xp_clear_dev(struct xsk_buff_pool *pool)
212 {
213 if (!pool->netdev)
214 return;
215
216 xp_disable_drv_zc(pool);
217 xsk_clear_pool_at_qid(pool->netdev, pool->queue_id);
218 dev_put(pool->netdev);
219 pool->netdev = NULL;
220 }
221
xp_release_deferred(struct work_struct * work)222 static void xp_release_deferred(struct work_struct *work)
223 {
224 struct xsk_buff_pool *pool = container_of(work, struct xsk_buff_pool,
225 work);
226
227 rtnl_lock();
228 xp_clear_dev(pool);
229 rtnl_unlock();
230
231 if (pool->fq) {
232 xskq_destroy(pool->fq);
233 pool->fq = NULL;
234 }
235
236 if (pool->cq) {
237 xskq_destroy(pool->cq);
238 pool->cq = NULL;
239 }
240
241 xdp_put_umem(pool->umem, false);
242 xp_destroy(pool);
243 }
244
xp_get_pool(struct xsk_buff_pool * pool)245 void xp_get_pool(struct xsk_buff_pool *pool)
246 {
247 refcount_inc(&pool->users);
248 }
249
xp_put_pool(struct xsk_buff_pool * pool)250 bool xp_put_pool(struct xsk_buff_pool *pool)
251 {
252 if (!pool)
253 return false;
254
255 if (refcount_dec_and_test(&pool->users)) {
256 INIT_WORK(&pool->work, xp_release_deferred);
257 schedule_work(&pool->work);
258 return true;
259 }
260
261 return false;
262 }
263
xp_find_dma_map(struct xsk_buff_pool * pool)264 static struct xsk_dma_map *xp_find_dma_map(struct xsk_buff_pool *pool)
265 {
266 struct xsk_dma_map *dma_map;
267
268 list_for_each_entry(dma_map, &pool->umem->xsk_dma_list, list) {
269 if (dma_map->netdev == pool->netdev)
270 return dma_map;
271 }
272
273 return NULL;
274 }
275
xp_create_dma_map(struct device * dev,struct net_device * netdev,u32 nr_pages,struct xdp_umem * umem)276 static struct xsk_dma_map *xp_create_dma_map(struct device *dev, struct net_device *netdev,
277 u32 nr_pages, struct xdp_umem *umem)
278 {
279 struct xsk_dma_map *dma_map;
280
281 dma_map = kzalloc(sizeof(*dma_map), GFP_KERNEL);
282 if (!dma_map)
283 return NULL;
284
285 dma_map->dma_pages = kvcalloc(nr_pages, sizeof(*dma_map->dma_pages), GFP_KERNEL);
286 if (!dma_map->dma_pages) {
287 kfree(dma_map);
288 return NULL;
289 }
290
291 dma_map->netdev = netdev;
292 dma_map->dev = dev;
293 dma_map->dma_need_sync = false;
294 dma_map->dma_pages_cnt = nr_pages;
295 refcount_set(&dma_map->users, 1);
296 list_add(&dma_map->list, &umem->xsk_dma_list);
297 return dma_map;
298 }
299
xp_destroy_dma_map(struct xsk_dma_map * dma_map)300 static void xp_destroy_dma_map(struct xsk_dma_map *dma_map)
301 {
302 list_del(&dma_map->list);
303 kvfree(dma_map->dma_pages);
304 kfree(dma_map);
305 }
306
__xp_dma_unmap(struct xsk_dma_map * dma_map,unsigned long attrs)307 static void __xp_dma_unmap(struct xsk_dma_map *dma_map, unsigned long attrs)
308 {
309 dma_addr_t *dma;
310 u32 i;
311
312 for (i = 0; i < dma_map->dma_pages_cnt; i++) {
313 dma = &dma_map->dma_pages[i];
314 if (*dma) {
315 *dma &= ~XSK_NEXT_PG_CONTIG_MASK;
316 dma_unmap_page_attrs(dma_map->dev, *dma, PAGE_SIZE,
317 DMA_BIDIRECTIONAL, attrs);
318 *dma = 0;
319 }
320 }
321
322 xp_destroy_dma_map(dma_map);
323 }
324
xp_dma_unmap(struct xsk_buff_pool * pool,unsigned long attrs)325 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs)
326 {
327 struct xsk_dma_map *dma_map;
328
329 if (pool->dma_pages_cnt == 0)
330 return;
331
332 dma_map = xp_find_dma_map(pool);
333 if (!dma_map) {
334 WARN(1, "Could not find dma_map for device");
335 return;
336 }
337
338 if (!refcount_dec_and_test(&dma_map->users))
339 return;
340
341 __xp_dma_unmap(dma_map, attrs);
342 kvfree(pool->dma_pages);
343 pool->dma_pages_cnt = 0;
344 pool->dev = NULL;
345 }
346 EXPORT_SYMBOL(xp_dma_unmap);
347
xp_check_dma_contiguity(struct xsk_dma_map * dma_map)348 static void xp_check_dma_contiguity(struct xsk_dma_map *dma_map)
349 {
350 u32 i;
351
352 for (i = 0; i < dma_map->dma_pages_cnt - 1; i++) {
353 if (dma_map->dma_pages[i] + PAGE_SIZE == dma_map->dma_pages[i + 1])
354 dma_map->dma_pages[i] |= XSK_NEXT_PG_CONTIG_MASK;
355 else
356 dma_map->dma_pages[i] &= ~XSK_NEXT_PG_CONTIG_MASK;
357 }
358 }
359
xp_init_dma_info(struct xsk_buff_pool * pool,struct xsk_dma_map * dma_map)360 static int xp_init_dma_info(struct xsk_buff_pool *pool, struct xsk_dma_map *dma_map)
361 {
362 pool->dma_pages = kvcalloc(dma_map->dma_pages_cnt, sizeof(*pool->dma_pages), GFP_KERNEL);
363 if (!pool->dma_pages)
364 return -ENOMEM;
365
366 pool->dev = dma_map->dev;
367 pool->dma_pages_cnt = dma_map->dma_pages_cnt;
368 pool->dma_need_sync = dma_map->dma_need_sync;
369 memcpy(pool->dma_pages, dma_map->dma_pages,
370 pool->dma_pages_cnt * sizeof(*pool->dma_pages));
371
372 return 0;
373 }
374
xp_dma_map(struct xsk_buff_pool * pool,struct device * dev,unsigned long attrs,struct page ** pages,u32 nr_pages)375 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
376 unsigned long attrs, struct page **pages, u32 nr_pages)
377 {
378 struct xsk_dma_map *dma_map;
379 dma_addr_t dma;
380 int err;
381 u32 i;
382
383 dma_map = xp_find_dma_map(pool);
384 if (dma_map) {
385 err = xp_init_dma_info(pool, dma_map);
386 if (err)
387 return err;
388
389 refcount_inc(&dma_map->users);
390 return 0;
391 }
392
393 dma_map = xp_create_dma_map(dev, pool->netdev, nr_pages, pool->umem);
394 if (!dma_map)
395 return -ENOMEM;
396
397 for (i = 0; i < dma_map->dma_pages_cnt; i++) {
398 dma = dma_map_page_attrs(dev, pages[i], 0, PAGE_SIZE,
399 DMA_BIDIRECTIONAL, attrs);
400 if (dma_mapping_error(dev, dma)) {
401 __xp_dma_unmap(dma_map, attrs);
402 return -ENOMEM;
403 }
404 if (dma_need_sync(dev, dma))
405 dma_map->dma_need_sync = true;
406 dma_map->dma_pages[i] = dma;
407 }
408
409 if (pool->unaligned)
410 xp_check_dma_contiguity(dma_map);
411
412 err = xp_init_dma_info(pool, dma_map);
413 if (err) {
414 __xp_dma_unmap(dma_map, attrs);
415 return err;
416 }
417
418 return 0;
419 }
420 EXPORT_SYMBOL(xp_dma_map);
421
xp_addr_crosses_non_contig_pg(struct xsk_buff_pool * pool,u64 addr)422 static bool xp_addr_crosses_non_contig_pg(struct xsk_buff_pool *pool,
423 u64 addr)
424 {
425 return xp_desc_crosses_non_contig_pg(pool, addr, pool->chunk_size);
426 }
427
xp_check_unaligned(struct xsk_buff_pool * pool,u64 * addr)428 static bool xp_check_unaligned(struct xsk_buff_pool *pool, u64 *addr)
429 {
430 *addr = xp_unaligned_extract_addr(*addr);
431 if (*addr >= pool->addrs_cnt ||
432 *addr + pool->chunk_size > pool->addrs_cnt ||
433 xp_addr_crosses_non_contig_pg(pool, *addr))
434 return false;
435 return true;
436 }
437
xp_check_aligned(struct xsk_buff_pool * pool,u64 * addr)438 static bool xp_check_aligned(struct xsk_buff_pool *pool, u64 *addr)
439 {
440 *addr = xp_aligned_extract_addr(pool, *addr);
441 return *addr < pool->addrs_cnt;
442 }
443
__xp_alloc(struct xsk_buff_pool * pool)444 static struct xdp_buff_xsk *__xp_alloc(struct xsk_buff_pool *pool)
445 {
446 struct xdp_buff_xsk *xskb;
447 u64 addr;
448 bool ok;
449
450 if (pool->free_heads_cnt == 0)
451 return NULL;
452
453 xskb = pool->free_heads[--pool->free_heads_cnt];
454
455 for (;;) {
456 if (!xskq_cons_peek_addr_unchecked(pool->fq, &addr)) {
457 pool->fq->queue_empty_descs++;
458 xp_release(xskb);
459 return NULL;
460 }
461
462 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
463 xp_check_aligned(pool, &addr);
464 if (!ok) {
465 pool->fq->invalid_descs++;
466 xskq_cons_release(pool->fq);
467 continue;
468 }
469 break;
470 }
471 xskq_cons_release(pool->fq);
472
473 xskb->orig_addr = addr;
474 xskb->xdp.data_hard_start = pool->addrs + addr + pool->headroom;
475 if (pool->dma_pages_cnt) {
476 xskb->frame_dma = (pool->dma_pages[addr >> PAGE_SHIFT] &
477 ~XSK_NEXT_PG_CONTIG_MASK) +
478 (addr & ~PAGE_MASK);
479 xskb->dma = xskb->frame_dma + pool->headroom +
480 XDP_PACKET_HEADROOM;
481 }
482 return xskb;
483 }
484
xp_alloc(struct xsk_buff_pool * pool)485 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool)
486 {
487 struct xdp_buff_xsk *xskb;
488
489 if (!pool->free_list_cnt) {
490 xskb = __xp_alloc(pool);
491 if (!xskb)
492 return NULL;
493 } else {
494 pool->free_list_cnt--;
495 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk,
496 free_list_node);
497 list_del(&xskb->free_list_node);
498 }
499
500 xskb->xdp.data = xskb->xdp.data_hard_start + XDP_PACKET_HEADROOM;
501 xskb->xdp.data_meta = xskb->xdp.data;
502
503 if (pool->dma_need_sync) {
504 dma_sync_single_range_for_device(pool->dev, xskb->dma, 0,
505 pool->frame_len,
506 DMA_BIDIRECTIONAL);
507 }
508 return &xskb->xdp;
509 }
510 EXPORT_SYMBOL(xp_alloc);
511
xp_can_alloc(struct xsk_buff_pool * pool,u32 count)512 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count)
513 {
514 if (pool->free_list_cnt >= count)
515 return true;
516 return xskq_cons_has_entries(pool->fq, count - pool->free_list_cnt);
517 }
518 EXPORT_SYMBOL(xp_can_alloc);
519
xp_free(struct xdp_buff_xsk * xskb)520 void xp_free(struct xdp_buff_xsk *xskb)
521 {
522 xskb->pool->free_list_cnt++;
523 list_add(&xskb->free_list_node, &xskb->pool->free_list);
524 }
525 EXPORT_SYMBOL(xp_free);
526
xp_raw_get_data(struct xsk_buff_pool * pool,u64 addr)527 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
528 {
529 addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
530 return pool->addrs + addr;
531 }
532 EXPORT_SYMBOL(xp_raw_get_data);
533
xp_raw_get_dma(struct xsk_buff_pool * pool,u64 addr)534 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
535 {
536 addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
537 return (pool->dma_pages[addr >> PAGE_SHIFT] &
538 ~XSK_NEXT_PG_CONTIG_MASK) +
539 (addr & ~PAGE_MASK);
540 }
541 EXPORT_SYMBOL(xp_raw_get_dma);
542
xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk * xskb)543 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb)
544 {
545 dma_sync_single_range_for_cpu(xskb->pool->dev, xskb->dma, 0,
546 xskb->pool->frame_len, DMA_BIDIRECTIONAL);
547 }
548 EXPORT_SYMBOL(xp_dma_sync_for_cpu_slow);
549
xp_dma_sync_for_device_slow(struct xsk_buff_pool * pool,dma_addr_t dma,size_t size)550 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
551 size_t size)
552 {
553 dma_sync_single_range_for_device(pool->dev, dma, 0,
554 size, DMA_BIDIRECTIONAL);
555 }
556 EXPORT_SYMBOL(xp_dma_sync_for_device_slow);
557