xref: /openbmc/linux/net/xdp/xsk_buff_pool.c (revision 835fd614)
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->heads);
41 	kvfree(pool);
42 }
43 
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 	refcount_set(&pool->users, 1);
75 
76 	pool->fq = xs->fq_tmp;
77 	pool->cq = xs->cq_tmp;
78 	xs->fq_tmp = NULL;
79 	xs->cq_tmp = NULL;
80 
81 	for (i = 0; i < pool->free_heads_cnt; i++) {
82 		xskb = &pool->heads[i];
83 		xskb->pool = pool;
84 		xskb->xdp.frame_sz = umem->chunk_size - umem->headroom;
85 		pool->free_heads[i] = xskb;
86 	}
87 
88 	return pool;
89 
90 out:
91 	xp_destroy(pool);
92 	return NULL;
93 }
94 
95 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq)
96 {
97 	u32 i;
98 
99 	for (i = 0; i < pool->heads_cnt; i++)
100 		pool->heads[i].xdp.rxq = rxq;
101 }
102 EXPORT_SYMBOL(xp_set_rxq_info);
103 
104 static void xp_disable_drv_zc(struct xsk_buff_pool *pool)
105 {
106 	struct netdev_bpf bpf;
107 	int err;
108 
109 	ASSERT_RTNL();
110 
111 	if (pool->umem->zc) {
112 		bpf.command = XDP_SETUP_XSK_POOL;
113 		bpf.xsk.pool = NULL;
114 		bpf.xsk.queue_id = pool->queue_id;
115 
116 		err = pool->netdev->netdev_ops->ndo_bpf(pool->netdev, &bpf);
117 
118 		if (err)
119 			WARN(1, "Failed to disable zero-copy!\n");
120 	}
121 }
122 
123 static int __xp_assign_dev(struct xsk_buff_pool *pool,
124 			   struct net_device *netdev, u16 queue_id, u16 flags)
125 {
126 	bool force_zc, force_copy;
127 	struct netdev_bpf bpf;
128 	int err = 0;
129 
130 	ASSERT_RTNL();
131 
132 	force_zc = flags & XDP_ZEROCOPY;
133 	force_copy = flags & XDP_COPY;
134 
135 	if (force_zc && force_copy)
136 		return -EINVAL;
137 
138 	if (xsk_get_pool_from_qid(netdev, queue_id))
139 		return -EBUSY;
140 
141 	pool->netdev = netdev;
142 	pool->queue_id = queue_id;
143 	err = xsk_reg_pool_at_qid(netdev, pool, queue_id);
144 	if (err)
145 		return err;
146 
147 	if (flags & XDP_USE_NEED_WAKEUP) {
148 		pool->uses_need_wakeup = true;
149 		/* Tx needs to be explicitly woken up the first time.
150 		 * Also for supporting drivers that do not implement this
151 		 * feature. They will always have to call sendto().
152 		 */
153 		pool->cached_need_wakeup = XDP_WAKEUP_TX;
154 	}
155 
156 	dev_hold(netdev);
157 
158 	if (force_copy)
159 		/* For copy-mode, we are done. */
160 		return 0;
161 
162 	if (!netdev->netdev_ops->ndo_bpf ||
163 	    !netdev->netdev_ops->ndo_xsk_wakeup) {
164 		err = -EOPNOTSUPP;
165 		goto err_unreg_pool;
166 	}
167 
168 	bpf.command = XDP_SETUP_XSK_POOL;
169 	bpf.xsk.pool = pool;
170 	bpf.xsk.queue_id = queue_id;
171 
172 	err = netdev->netdev_ops->ndo_bpf(netdev, &bpf);
173 	if (err)
174 		goto err_unreg_pool;
175 
176 	if (!pool->dma_pages) {
177 		WARN(1, "Driver did not DMA map zero-copy buffers");
178 		goto err_unreg_xsk;
179 	}
180 	pool->umem->zc = true;
181 	return 0;
182 
183 err_unreg_xsk:
184 	xp_disable_drv_zc(pool);
185 err_unreg_pool:
186 	if (!force_zc)
187 		err = 0; /* fallback to copy mode */
188 	if (err)
189 		xsk_clear_pool_at_qid(netdev, queue_id);
190 	return err;
191 }
192 
193 int xp_assign_dev(struct xsk_buff_pool *pool, struct net_device *dev,
194 		  u16 queue_id, u16 flags)
195 {
196 	return __xp_assign_dev(pool, dev, queue_id, flags);
197 }
198 
199 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_umem *umem,
200 			 struct net_device *dev, u16 queue_id)
201 {
202 	u16 flags;
203 
204 	/* One fill and completion ring required for each queue id. */
205 	if (!pool->fq || !pool->cq)
206 		return -EINVAL;
207 
208 	flags = umem->zc ? XDP_ZEROCOPY : XDP_COPY;
209 	if (pool->uses_need_wakeup)
210 		flags |= XDP_USE_NEED_WAKEUP;
211 
212 	return __xp_assign_dev(pool, dev, queue_id, flags);
213 }
214 
215 void xp_clear_dev(struct xsk_buff_pool *pool)
216 {
217 	if (!pool->netdev)
218 		return;
219 
220 	xp_disable_drv_zc(pool);
221 	xsk_clear_pool_at_qid(pool->netdev, pool->queue_id);
222 	dev_put(pool->netdev);
223 	pool->netdev = NULL;
224 }
225 
226 static void xp_release_deferred(struct work_struct *work)
227 {
228 	struct xsk_buff_pool *pool = container_of(work, struct xsk_buff_pool,
229 						  work);
230 
231 	rtnl_lock();
232 	xp_clear_dev(pool);
233 	rtnl_unlock();
234 
235 	if (pool->fq) {
236 		xskq_destroy(pool->fq);
237 		pool->fq = NULL;
238 	}
239 
240 	if (pool->cq) {
241 		xskq_destroy(pool->cq);
242 		pool->cq = NULL;
243 	}
244 
245 	xdp_put_umem(pool->umem);
246 	xp_destroy(pool);
247 }
248 
249 void xp_get_pool(struct xsk_buff_pool *pool)
250 {
251 	refcount_inc(&pool->users);
252 }
253 
254 void xp_put_pool(struct xsk_buff_pool *pool)
255 {
256 	if (!pool)
257 		return;
258 
259 	if (refcount_dec_and_test(&pool->users)) {
260 		INIT_WORK(&pool->work, xp_release_deferred);
261 		schedule_work(&pool->work);
262 	}
263 }
264 
265 static struct xsk_dma_map *xp_find_dma_map(struct xsk_buff_pool *pool)
266 {
267 	struct xsk_dma_map *dma_map;
268 
269 	list_for_each_entry(dma_map, &pool->umem->xsk_dma_list, list) {
270 		if (dma_map->netdev == pool->netdev)
271 			return dma_map;
272 	}
273 
274 	return NULL;
275 }
276 
277 static struct xsk_dma_map *xp_create_dma_map(struct device *dev, struct net_device *netdev,
278 					     u32 nr_pages, struct xdp_umem *umem)
279 {
280 	struct xsk_dma_map *dma_map;
281 
282 	dma_map = kzalloc(sizeof(*dma_map), GFP_KERNEL);
283 	if (!dma_map)
284 		return NULL;
285 
286 	dma_map->dma_pages = kvcalloc(nr_pages, sizeof(*dma_map->dma_pages), GFP_KERNEL);
287 	if (!dma_map->dma_pages) {
288 		kfree(dma_map);
289 		return NULL;
290 	}
291 
292 	dma_map->netdev = netdev;
293 	dma_map->dev = dev;
294 	dma_map->dma_need_sync = false;
295 	dma_map->dma_pages_cnt = nr_pages;
296 	refcount_set(&dma_map->users, 1);
297 	list_add(&dma_map->list, &umem->xsk_dma_list);
298 	return dma_map;
299 }
300 
301 static void xp_destroy_dma_map(struct xsk_dma_map *dma_map)
302 {
303 	list_del(&dma_map->list);
304 	kvfree(dma_map->dma_pages);
305 	kfree(dma_map);
306 }
307 
308 static void __xp_dma_unmap(struct xsk_dma_map *dma_map, unsigned long attrs)
309 {
310 	dma_addr_t *dma;
311 	u32 i;
312 
313 	for (i = 0; i < dma_map->dma_pages_cnt; i++) {
314 		dma = &dma_map->dma_pages[i];
315 		if (*dma) {
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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 
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