xref: /openbmc/linux/net/xdp/xsk_buff_pool.c (revision 842ed298)
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 	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 
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 
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 
122 static 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 
194 int xp_assign_dev(struct xsk_buff_pool *pool, struct net_device *dev,
195 		  u16 queue_id, u16 flags)
196 {
197 	return __xp_assign_dev(pool, dev, queue_id, flags);
198 }
199 
200 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_umem *umem,
201 			 struct net_device *dev, u16 queue_id)
202 {
203 	u16 flags;
204 
205 	/* One fill and completion ring required for each queue id. */
206 	if (!pool->fq || !pool->cq)
207 		return -EINVAL;
208 
209 	flags = umem->zc ? XDP_ZEROCOPY : XDP_COPY;
210 	if (pool->uses_need_wakeup)
211 		flags |= XDP_USE_NEED_WAKEUP;
212 
213 	return __xp_assign_dev(pool, dev, queue_id, flags);
214 }
215 
216 void xp_clear_dev(struct xsk_buff_pool *pool)
217 {
218 	if (!pool->netdev)
219 		return;
220 
221 	xp_disable_drv_zc(pool);
222 	xsk_clear_pool_at_qid(pool->netdev, pool->queue_id);
223 	dev_put(pool->netdev);
224 	pool->netdev = NULL;
225 }
226 
227 static void xp_release_deferred(struct work_struct *work)
228 {
229 	struct xsk_buff_pool *pool = container_of(work, struct xsk_buff_pool,
230 						  work);
231 
232 	rtnl_lock();
233 	xp_clear_dev(pool);
234 	rtnl_unlock();
235 
236 	if (pool->fq) {
237 		xskq_destroy(pool->fq);
238 		pool->fq = NULL;
239 	}
240 
241 	if (pool->cq) {
242 		xskq_destroy(pool->cq);
243 		pool->cq = NULL;
244 	}
245 
246 	xdp_put_umem(pool->umem, false);
247 	xp_destroy(pool);
248 }
249 
250 void xp_get_pool(struct xsk_buff_pool *pool)
251 {
252 	refcount_inc(&pool->users);
253 }
254 
255 bool xp_put_pool(struct xsk_buff_pool *pool)
256 {
257 	if (!pool)
258 		return false;
259 
260 	if (refcount_dec_and_test(&pool->users)) {
261 		INIT_WORK(&pool->work, xp_release_deferred);
262 		schedule_work(&pool->work);
263 		return true;
264 	}
265 
266 	return false;
267 }
268 
269 static struct xsk_dma_map *xp_find_dma_map(struct xsk_buff_pool *pool)
270 {
271 	struct xsk_dma_map *dma_map;
272 
273 	list_for_each_entry(dma_map, &pool->umem->xsk_dma_list, list) {
274 		if (dma_map->netdev == pool->netdev)
275 			return dma_map;
276 	}
277 
278 	return NULL;
279 }
280 
281 static struct xsk_dma_map *xp_create_dma_map(struct device *dev, struct net_device *netdev,
282 					     u32 nr_pages, struct xdp_umem *umem)
283 {
284 	struct xsk_dma_map *dma_map;
285 
286 	dma_map = kzalloc(sizeof(*dma_map), GFP_KERNEL);
287 	if (!dma_map)
288 		return NULL;
289 
290 	dma_map->dma_pages = kvcalloc(nr_pages, sizeof(*dma_map->dma_pages), GFP_KERNEL);
291 	if (!dma_map->dma_pages) {
292 		kfree(dma_map);
293 		return NULL;
294 	}
295 
296 	dma_map->netdev = netdev;
297 	dma_map->dev = dev;
298 	dma_map->dma_need_sync = false;
299 	dma_map->dma_pages_cnt = nr_pages;
300 	refcount_set(&dma_map->users, 1);
301 	list_add(&dma_map->list, &umem->xsk_dma_list);
302 	return dma_map;
303 }
304 
305 static void xp_destroy_dma_map(struct xsk_dma_map *dma_map)
306 {
307 	list_del(&dma_map->list);
308 	kvfree(dma_map->dma_pages);
309 	kfree(dma_map);
310 }
311 
312 static void __xp_dma_unmap(struct xsk_dma_map *dma_map, unsigned long attrs)
313 {
314 	dma_addr_t *dma;
315 	u32 i;
316 
317 	for (i = 0; i < dma_map->dma_pages_cnt; i++) {
318 		dma = &dma_map->dma_pages[i];
319 		if (*dma) {
320 			dma_unmap_page_attrs(dma_map->dev, *dma, PAGE_SIZE,
321 					     DMA_BIDIRECTIONAL, attrs);
322 			*dma = 0;
323 		}
324 	}
325 
326 	xp_destroy_dma_map(dma_map);
327 }
328 
329 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs)
330 {
331 	struct xsk_dma_map *dma_map;
332 
333 	if (pool->dma_pages_cnt == 0)
334 		return;
335 
336 	dma_map = xp_find_dma_map(pool);
337 	if (!dma_map) {
338 		WARN(1, "Could not find dma_map for device");
339 		return;
340 	}
341 
342 	if (!refcount_dec_and_test(&dma_map->users))
343 		return;
344 
345 	__xp_dma_unmap(dma_map, attrs);
346 	kvfree(pool->dma_pages);
347 	pool->dma_pages_cnt = 0;
348 	pool->dev = NULL;
349 }
350 EXPORT_SYMBOL(xp_dma_unmap);
351 
352 static void xp_check_dma_contiguity(struct xsk_dma_map *dma_map)
353 {
354 	u32 i;
355 
356 	for (i = 0; i < dma_map->dma_pages_cnt - 1; i++) {
357 		if (dma_map->dma_pages[i] + PAGE_SIZE == dma_map->dma_pages[i + 1])
358 			dma_map->dma_pages[i] |= XSK_NEXT_PG_CONTIG_MASK;
359 		else
360 			dma_map->dma_pages[i] &= ~XSK_NEXT_PG_CONTIG_MASK;
361 	}
362 }
363 
364 static int xp_init_dma_info(struct xsk_buff_pool *pool, struct xsk_dma_map *dma_map)
365 {
366 	pool->dma_pages = kvcalloc(dma_map->dma_pages_cnt, sizeof(*pool->dma_pages), GFP_KERNEL);
367 	if (!pool->dma_pages)
368 		return -ENOMEM;
369 
370 	pool->dev = dma_map->dev;
371 	pool->dma_pages_cnt = dma_map->dma_pages_cnt;
372 	pool->dma_need_sync = dma_map->dma_need_sync;
373 	memcpy(pool->dma_pages, dma_map->dma_pages,
374 	       pool->dma_pages_cnt * sizeof(*pool->dma_pages));
375 
376 	return 0;
377 }
378 
379 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
380 	       unsigned long attrs, struct page **pages, u32 nr_pages)
381 {
382 	struct xsk_dma_map *dma_map;
383 	dma_addr_t dma;
384 	int err;
385 	u32 i;
386 
387 	dma_map = xp_find_dma_map(pool);
388 	if (dma_map) {
389 		err = xp_init_dma_info(pool, dma_map);
390 		if (err)
391 			return err;
392 
393 		refcount_inc(&dma_map->users);
394 		return 0;
395 	}
396 
397 	dma_map = xp_create_dma_map(dev, pool->netdev, nr_pages, pool->umem);
398 	if (!dma_map)
399 		return -ENOMEM;
400 
401 	for (i = 0; i < dma_map->dma_pages_cnt; i++) {
402 		dma = dma_map_page_attrs(dev, pages[i], 0, PAGE_SIZE,
403 					 DMA_BIDIRECTIONAL, attrs);
404 		if (dma_mapping_error(dev, dma)) {
405 			__xp_dma_unmap(dma_map, attrs);
406 			return -ENOMEM;
407 		}
408 		if (dma_need_sync(dev, dma))
409 			dma_map->dma_need_sync = true;
410 		dma_map->dma_pages[i] = dma;
411 	}
412 
413 	if (pool->unaligned)
414 		xp_check_dma_contiguity(dma_map);
415 
416 	err = xp_init_dma_info(pool, dma_map);
417 	if (err) {
418 		__xp_dma_unmap(dma_map, attrs);
419 		return err;
420 	}
421 
422 	return 0;
423 }
424 EXPORT_SYMBOL(xp_dma_map);
425 
426 static bool xp_addr_crosses_non_contig_pg(struct xsk_buff_pool *pool,
427 					  u64 addr)
428 {
429 	return xp_desc_crosses_non_contig_pg(pool, addr, pool->chunk_size);
430 }
431 
432 static bool xp_check_unaligned(struct xsk_buff_pool *pool, u64 *addr)
433 {
434 	*addr = xp_unaligned_extract_addr(*addr);
435 	if (*addr >= pool->addrs_cnt ||
436 	    *addr + pool->chunk_size > pool->addrs_cnt ||
437 	    xp_addr_crosses_non_contig_pg(pool, *addr))
438 		return false;
439 	return true;
440 }
441 
442 static bool xp_check_aligned(struct xsk_buff_pool *pool, u64 *addr)
443 {
444 	*addr = xp_aligned_extract_addr(pool, *addr);
445 	return *addr < pool->addrs_cnt;
446 }
447 
448 static struct xdp_buff_xsk *__xp_alloc(struct xsk_buff_pool *pool)
449 {
450 	struct xdp_buff_xsk *xskb;
451 	u64 addr;
452 	bool ok;
453 
454 	if (pool->free_heads_cnt == 0)
455 		return NULL;
456 
457 	xskb = pool->free_heads[--pool->free_heads_cnt];
458 
459 	for (;;) {
460 		if (!xskq_cons_peek_addr_unchecked(pool->fq, &addr)) {
461 			pool->fq->queue_empty_descs++;
462 			xp_release(xskb);
463 			return NULL;
464 		}
465 
466 		ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
467 		     xp_check_aligned(pool, &addr);
468 		if (!ok) {
469 			pool->fq->invalid_descs++;
470 			xskq_cons_release(pool->fq);
471 			continue;
472 		}
473 		break;
474 	}
475 	xskq_cons_release(pool->fq);
476 
477 	xskb->orig_addr = addr;
478 	xskb->xdp.data_hard_start = pool->addrs + addr + pool->headroom;
479 	if (pool->dma_pages_cnt) {
480 		xskb->frame_dma = (pool->dma_pages[addr >> PAGE_SHIFT] &
481 				   ~XSK_NEXT_PG_CONTIG_MASK) +
482 				  (addr & ~PAGE_MASK);
483 		xskb->dma = xskb->frame_dma + pool->headroom +
484 			    XDP_PACKET_HEADROOM;
485 	}
486 	return xskb;
487 }
488 
489 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool)
490 {
491 	struct xdp_buff_xsk *xskb;
492 
493 	if (!pool->free_list_cnt) {
494 		xskb = __xp_alloc(pool);
495 		if (!xskb)
496 			return NULL;
497 	} else {
498 		pool->free_list_cnt--;
499 		xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk,
500 					free_list_node);
501 		list_del(&xskb->free_list_node);
502 	}
503 
504 	xskb->xdp.data = xskb->xdp.data_hard_start + XDP_PACKET_HEADROOM;
505 	xskb->xdp.data_meta = xskb->xdp.data;
506 
507 	if (pool->dma_need_sync) {
508 		dma_sync_single_range_for_device(pool->dev, xskb->dma, 0,
509 						 pool->frame_len,
510 						 DMA_BIDIRECTIONAL);
511 	}
512 	return &xskb->xdp;
513 }
514 EXPORT_SYMBOL(xp_alloc);
515 
516 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count)
517 {
518 	if (pool->free_list_cnt >= count)
519 		return true;
520 	return xskq_cons_has_entries(pool->fq, count - pool->free_list_cnt);
521 }
522 EXPORT_SYMBOL(xp_can_alloc);
523 
524 void xp_free(struct xdp_buff_xsk *xskb)
525 {
526 	xskb->pool->free_list_cnt++;
527 	list_add(&xskb->free_list_node, &xskb->pool->free_list);
528 }
529 EXPORT_SYMBOL(xp_free);
530 
531 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
532 {
533 	addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
534 	return pool->addrs + addr;
535 }
536 EXPORT_SYMBOL(xp_raw_get_data);
537 
538 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
539 {
540 	addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
541 	return (pool->dma_pages[addr >> PAGE_SHIFT] &
542 		~XSK_NEXT_PG_CONTIG_MASK) +
543 		(addr & ~PAGE_MASK);
544 }
545 EXPORT_SYMBOL(xp_raw_get_dma);
546 
547 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb)
548 {
549 	dma_sync_single_range_for_cpu(xskb->pool->dev, xskb->dma, 0,
550 				      xskb->pool->frame_len, DMA_BIDIRECTIONAL);
551 }
552 EXPORT_SYMBOL(xp_dma_sync_for_cpu_slow);
553 
554 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
555 				 size_t size)
556 {
557 	dma_sync_single_range_for_device(pool->dev, dma, 0,
558 					 size, DMA_BIDIRECTIONAL);
559 }
560 EXPORT_SYMBOL(xp_dma_sync_for_device_slow);
561