xref: /openbmc/linux/net/xdp/xsk_buff_pool.c (revision 07d9a767)
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 		dev_put(netdev);
191 	}
192 	return err;
193 }
194 
195 int xp_assign_dev(struct xsk_buff_pool *pool, struct net_device *dev,
196 		  u16 queue_id, u16 flags)
197 {
198 	return __xp_assign_dev(pool, dev, queue_id, flags);
199 }
200 
201 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_umem *umem,
202 			 struct net_device *dev, u16 queue_id)
203 {
204 	u16 flags;
205 
206 	/* One fill and completion ring required for each queue id. */
207 	if (!pool->fq || !pool->cq)
208 		return -EINVAL;
209 
210 	flags = umem->zc ? XDP_ZEROCOPY : XDP_COPY;
211 	if (pool->uses_need_wakeup)
212 		flags |= XDP_USE_NEED_WAKEUP;
213 
214 	return __xp_assign_dev(pool, dev, queue_id, flags);
215 }
216 
217 void xp_clear_dev(struct xsk_buff_pool *pool)
218 {
219 	if (!pool->netdev)
220 		return;
221 
222 	xp_disable_drv_zc(pool);
223 	xsk_clear_pool_at_qid(pool->netdev, pool->queue_id);
224 	dev_put(pool->netdev);
225 	pool->netdev = NULL;
226 }
227 
228 static void xp_release_deferred(struct work_struct *work)
229 {
230 	struct xsk_buff_pool *pool = container_of(work, struct xsk_buff_pool,
231 						  work);
232 
233 	rtnl_lock();
234 	xp_clear_dev(pool);
235 	rtnl_unlock();
236 
237 	if (pool->fq) {
238 		xskq_destroy(pool->fq);
239 		pool->fq = NULL;
240 	}
241 
242 	if (pool->cq) {
243 		xskq_destroy(pool->cq);
244 		pool->cq = NULL;
245 	}
246 
247 	xdp_put_umem(pool->umem, false);
248 	xp_destroy(pool);
249 }
250 
251 void xp_get_pool(struct xsk_buff_pool *pool)
252 {
253 	refcount_inc(&pool->users);
254 }
255 
256 bool xp_put_pool(struct xsk_buff_pool *pool)
257 {
258 	if (!pool)
259 		return false;
260 
261 	if (refcount_dec_and_test(&pool->users)) {
262 		INIT_WORK(&pool->work, xp_release_deferred);
263 		schedule_work(&pool->work);
264 		return true;
265 	}
266 
267 	return false;
268 }
269 
270 static struct xsk_dma_map *xp_find_dma_map(struct xsk_buff_pool *pool)
271 {
272 	struct xsk_dma_map *dma_map;
273 
274 	list_for_each_entry(dma_map, &pool->umem->xsk_dma_list, list) {
275 		if (dma_map->netdev == pool->netdev)
276 			return dma_map;
277 	}
278 
279 	return NULL;
280 }
281 
282 static struct xsk_dma_map *xp_create_dma_map(struct device *dev, struct net_device *netdev,
283 					     u32 nr_pages, struct xdp_umem *umem)
284 {
285 	struct xsk_dma_map *dma_map;
286 
287 	dma_map = kzalloc(sizeof(*dma_map), GFP_KERNEL);
288 	if (!dma_map)
289 		return NULL;
290 
291 	dma_map->dma_pages = kvcalloc(nr_pages, sizeof(*dma_map->dma_pages), GFP_KERNEL);
292 	if (!dma_map->dma_pages) {
293 		kfree(dma_map);
294 		return NULL;
295 	}
296 
297 	dma_map->netdev = netdev;
298 	dma_map->dev = dev;
299 	dma_map->dma_need_sync = false;
300 	dma_map->dma_pages_cnt = nr_pages;
301 	refcount_set(&dma_map->users, 1);
302 	list_add(&dma_map->list, &umem->xsk_dma_list);
303 	return dma_map;
304 }
305 
306 static void xp_destroy_dma_map(struct xsk_dma_map *dma_map)
307 {
308 	list_del(&dma_map->list);
309 	kvfree(dma_map->dma_pages);
310 	kfree(dma_map);
311 }
312 
313 static void __xp_dma_unmap(struct xsk_dma_map *dma_map, unsigned long attrs)
314 {
315 	dma_addr_t *dma;
316 	u32 i;
317 
318 	for (i = 0; i < dma_map->dma_pages_cnt; i++) {
319 		dma = &dma_map->dma_pages[i];
320 		if (*dma) {
321 			dma_unmap_page_attrs(dma_map->dev, *dma, PAGE_SIZE,
322 					     DMA_BIDIRECTIONAL, attrs);
323 			*dma = 0;
324 		}
325 	}
326 
327 	xp_destroy_dma_map(dma_map);
328 }
329 
330 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs)
331 {
332 	struct xsk_dma_map *dma_map;
333 
334 	if (pool->dma_pages_cnt == 0)
335 		return;
336 
337 	dma_map = xp_find_dma_map(pool);
338 	if (!dma_map) {
339 		WARN(1, "Could not find dma_map for device");
340 		return;
341 	}
342 
343 	if (!refcount_dec_and_test(&dma_map->users))
344 		return;
345 
346 	__xp_dma_unmap(dma_map, attrs);
347 	kvfree(pool->dma_pages);
348 	pool->dma_pages_cnt = 0;
349 	pool->dev = NULL;
350 }
351 EXPORT_SYMBOL(xp_dma_unmap);
352 
353 static void xp_check_dma_contiguity(struct xsk_dma_map *dma_map)
354 {
355 	u32 i;
356 
357 	for (i = 0; i < dma_map->dma_pages_cnt - 1; i++) {
358 		if (dma_map->dma_pages[i] + PAGE_SIZE == dma_map->dma_pages[i + 1])
359 			dma_map->dma_pages[i] |= XSK_NEXT_PG_CONTIG_MASK;
360 		else
361 			dma_map->dma_pages[i] &= ~XSK_NEXT_PG_CONTIG_MASK;
362 	}
363 }
364 
365 static int xp_init_dma_info(struct xsk_buff_pool *pool, struct xsk_dma_map *dma_map)
366 {
367 	pool->dma_pages = kvcalloc(dma_map->dma_pages_cnt, sizeof(*pool->dma_pages), GFP_KERNEL);
368 	if (!pool->dma_pages)
369 		return -ENOMEM;
370 
371 	pool->dev = dma_map->dev;
372 	pool->dma_pages_cnt = dma_map->dma_pages_cnt;
373 	pool->dma_need_sync = dma_map->dma_need_sync;
374 	memcpy(pool->dma_pages, dma_map->dma_pages,
375 	       pool->dma_pages_cnt * sizeof(*pool->dma_pages));
376 
377 	return 0;
378 }
379 
380 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
381 	       unsigned long attrs, struct page **pages, u32 nr_pages)
382 {
383 	struct xsk_dma_map *dma_map;
384 	dma_addr_t dma;
385 	int err;
386 	u32 i;
387 
388 	dma_map = xp_find_dma_map(pool);
389 	if (dma_map) {
390 		err = xp_init_dma_info(pool, dma_map);
391 		if (err)
392 			return err;
393 
394 		refcount_inc(&dma_map->users);
395 		return 0;
396 	}
397 
398 	dma_map = xp_create_dma_map(dev, pool->netdev, nr_pages, pool->umem);
399 	if (!dma_map)
400 		return -ENOMEM;
401 
402 	for (i = 0; i < dma_map->dma_pages_cnt; i++) {
403 		dma = dma_map_page_attrs(dev, pages[i], 0, PAGE_SIZE,
404 					 DMA_BIDIRECTIONAL, attrs);
405 		if (dma_mapping_error(dev, dma)) {
406 			__xp_dma_unmap(dma_map, attrs);
407 			return -ENOMEM;
408 		}
409 		if (dma_need_sync(dev, dma))
410 			dma_map->dma_need_sync = true;
411 		dma_map->dma_pages[i] = dma;
412 	}
413 
414 	if (pool->unaligned)
415 		xp_check_dma_contiguity(dma_map);
416 
417 	err = xp_init_dma_info(pool, dma_map);
418 	if (err) {
419 		__xp_dma_unmap(dma_map, attrs);
420 		return err;
421 	}
422 
423 	return 0;
424 }
425 EXPORT_SYMBOL(xp_dma_map);
426 
427 static bool xp_addr_crosses_non_contig_pg(struct xsk_buff_pool *pool,
428 					  u64 addr)
429 {
430 	return xp_desc_crosses_non_contig_pg(pool, addr, pool->chunk_size);
431 }
432 
433 static bool xp_check_unaligned(struct xsk_buff_pool *pool, u64 *addr)
434 {
435 	*addr = xp_unaligned_extract_addr(*addr);
436 	if (*addr >= pool->addrs_cnt ||
437 	    *addr + pool->chunk_size > pool->addrs_cnt ||
438 	    xp_addr_crosses_non_contig_pg(pool, *addr))
439 		return false;
440 	return true;
441 }
442 
443 static bool xp_check_aligned(struct xsk_buff_pool *pool, u64 *addr)
444 {
445 	*addr = xp_aligned_extract_addr(pool, *addr);
446 	return *addr < pool->addrs_cnt;
447 }
448 
449 static struct xdp_buff_xsk *__xp_alloc(struct xsk_buff_pool *pool)
450 {
451 	struct xdp_buff_xsk *xskb;
452 	u64 addr;
453 	bool ok;
454 
455 	if (pool->free_heads_cnt == 0)
456 		return NULL;
457 
458 	xskb = pool->free_heads[--pool->free_heads_cnt];
459 
460 	for (;;) {
461 		if (!xskq_cons_peek_addr_unchecked(pool->fq, &addr)) {
462 			pool->fq->queue_empty_descs++;
463 			xp_release(xskb);
464 			return NULL;
465 		}
466 
467 		ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
468 		     xp_check_aligned(pool, &addr);
469 		if (!ok) {
470 			pool->fq->invalid_descs++;
471 			xskq_cons_release(pool->fq);
472 			continue;
473 		}
474 		break;
475 	}
476 	xskq_cons_release(pool->fq);
477 
478 	xskb->orig_addr = addr;
479 	xskb->xdp.data_hard_start = pool->addrs + addr + pool->headroom;
480 	if (pool->dma_pages_cnt) {
481 		xskb->frame_dma = (pool->dma_pages[addr >> PAGE_SHIFT] &
482 				   ~XSK_NEXT_PG_CONTIG_MASK) +
483 				  (addr & ~PAGE_MASK);
484 		xskb->dma = xskb->frame_dma + pool->headroom +
485 			    XDP_PACKET_HEADROOM;
486 	}
487 	return xskb;
488 }
489 
490 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool)
491 {
492 	struct xdp_buff_xsk *xskb;
493 
494 	if (!pool->free_list_cnt) {
495 		xskb = __xp_alloc(pool);
496 		if (!xskb)
497 			return NULL;
498 	} else {
499 		pool->free_list_cnt--;
500 		xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk,
501 					free_list_node);
502 		list_del(&xskb->free_list_node);
503 	}
504 
505 	xskb->xdp.data = xskb->xdp.data_hard_start + XDP_PACKET_HEADROOM;
506 	xskb->xdp.data_meta = xskb->xdp.data;
507 
508 	if (pool->dma_need_sync) {
509 		dma_sync_single_range_for_device(pool->dev, xskb->dma, 0,
510 						 pool->frame_len,
511 						 DMA_BIDIRECTIONAL);
512 	}
513 	return &xskb->xdp;
514 }
515 EXPORT_SYMBOL(xp_alloc);
516 
517 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count)
518 {
519 	if (pool->free_list_cnt >= count)
520 		return true;
521 	return xskq_cons_has_entries(pool->fq, count - pool->free_list_cnt);
522 }
523 EXPORT_SYMBOL(xp_can_alloc);
524 
525 void xp_free(struct xdp_buff_xsk *xskb)
526 {
527 	xskb->pool->free_list_cnt++;
528 	list_add(&xskb->free_list_node, &xskb->pool->free_list);
529 }
530 EXPORT_SYMBOL(xp_free);
531 
532 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
533 {
534 	addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
535 	return pool->addrs + addr;
536 }
537 EXPORT_SYMBOL(xp_raw_get_data);
538 
539 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
540 {
541 	addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
542 	return (pool->dma_pages[addr >> PAGE_SHIFT] &
543 		~XSK_NEXT_PG_CONTIG_MASK) +
544 		(addr & ~PAGE_MASK);
545 }
546 EXPORT_SYMBOL(xp_raw_get_dma);
547 
548 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb)
549 {
550 	dma_sync_single_range_for_cpu(xskb->pool->dev, xskb->dma, 0,
551 				      xskb->pool->frame_len, DMA_BIDIRECTIONAL);
552 }
553 EXPORT_SYMBOL(xp_dma_sync_for_cpu_slow);
554 
555 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
556 				 size_t size)
557 {
558 	dma_sync_single_range_for_device(pool->dev, dma, 0,
559 					 size, DMA_BIDIRECTIONAL);
560 }
561 EXPORT_SYMBOL(xp_dma_sync_for_device_slow);
562