1 // SPDX-License-Identifier: GPL-2.0
3 #include <net/xsk_buff_pool.h>
4 #include <net/xdp_sock.h>
5 #include <net/xdp_sock_drv.h>
11 void xp_add_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs)
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);
23 void xp_del_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs)
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);
35 void xp_destroy(struct xsk_buff_pool *pool)
44 struct xsk_buff_pool *xp_create_and_assign_umem(struct xdp_sock *xs,
45 struct xdp_umem *umem)
47 struct xsk_buff_pool *pool;
48 struct xdp_buff_xsk *xskb;
51 pool = kvzalloc(struct_size(pool, free_heads, umem->chunks),
56 pool->heads = kvcalloc(umem->chunks, sizeof(*pool->heads), GFP_KERNEL);
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 -
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);
76 pool->fq = xs->fq_tmp;
77 pool->cq = xs->cq_tmp;
81 for (i = 0; i < pool->free_heads_cnt; i++) {
82 xskb = &pool->heads[i];
84 xskb->xdp.frame_sz = umem->chunk_size - umem->headroom;
85 pool->free_heads[i] = xskb;
95 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq)
99 for (i = 0; i < pool->heads_cnt; i++)
100 pool->heads[i].xdp.rxq = rxq;
102 EXPORT_SYMBOL(xp_set_rxq_info);
104 static void xp_disable_drv_zc(struct xsk_buff_pool *pool)
106 struct netdev_bpf bpf;
111 if (pool->umem->zc) {
112 bpf.command = XDP_SETUP_XSK_POOL;
114 bpf.xsk.queue_id = pool->queue_id;
116 err = pool->netdev->netdev_ops->ndo_bpf(pool->netdev, &bpf);
119 WARN(1, "Failed to disable zero-copy!\n");
123 static int __xp_assign_dev(struct xsk_buff_pool *pool,
124 struct net_device *netdev, u16 queue_id, u16 flags)
126 bool force_zc, force_copy;
127 struct netdev_bpf bpf;
132 force_zc = flags & XDP_ZEROCOPY;
133 force_copy = flags & XDP_COPY;
135 if (force_zc && force_copy)
138 if (xsk_get_pool_from_qid(netdev, queue_id))
141 pool->netdev = netdev;
142 pool->queue_id = queue_id;
143 err = xsk_reg_pool_at_qid(netdev, pool, queue_id);
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().
153 pool->cached_need_wakeup = XDP_WAKEUP_TX;
159 /* For copy-mode, we are done. */
162 if (!netdev->netdev_ops->ndo_bpf ||
163 !netdev->netdev_ops->ndo_xsk_wakeup) {
168 bpf.command = XDP_SETUP_XSK_POOL;
170 bpf.xsk.queue_id = queue_id;
172 err = netdev->netdev_ops->ndo_bpf(netdev, &bpf);
176 if (!pool->dma_pages) {
177 WARN(1, "Driver did not DMA map zero-copy buffers");
180 pool->umem->zc = true;
184 xp_disable_drv_zc(pool);
187 err = 0; /* fallback to copy mode */
189 xsk_clear_pool_at_qid(netdev, queue_id);
193 int xp_assign_dev(struct xsk_buff_pool *pool, struct net_device *dev,
194 u16 queue_id, u16 flags)
196 return __xp_assign_dev(pool, dev, queue_id, flags);
199 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_umem *umem,
200 struct net_device *dev, u16 queue_id)
204 /* One fill and completion ring required for each queue id. */
205 if (!pool->fq || !pool->cq)
208 flags = umem->zc ? XDP_ZEROCOPY : XDP_COPY;
209 if (pool->uses_need_wakeup)
210 flags |= XDP_USE_NEED_WAKEUP;
212 return __xp_assign_dev(pool, dev, queue_id, flags);
215 void xp_clear_dev(struct xsk_buff_pool *pool)
220 xp_disable_drv_zc(pool);
221 xsk_clear_pool_at_qid(pool->netdev, pool->queue_id);
222 dev_put(pool->netdev);
226 static void xp_release_deferred(struct work_struct *work)
228 struct xsk_buff_pool *pool = container_of(work, struct xsk_buff_pool,
236 xskq_destroy(pool->fq);
241 xskq_destroy(pool->cq);
245 xdp_put_umem(pool->umem);
249 void xp_get_pool(struct xsk_buff_pool *pool)
251 refcount_inc(&pool->users);
254 void xp_put_pool(struct xsk_buff_pool *pool)
259 if (refcount_dec_and_test(&pool->users)) {
260 INIT_WORK(&pool->work, xp_release_deferred);
261 schedule_work(&pool->work);
265 static struct xsk_dma_map *xp_find_dma_map(struct xsk_buff_pool *pool)
267 struct xsk_dma_map *dma_map;
269 list_for_each_entry(dma_map, &pool->umem->xsk_dma_list, list) {
270 if (dma_map->netdev == pool->netdev)
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)
280 struct xsk_dma_map *dma_map;
282 dma_map = kzalloc(sizeof(*dma_map), GFP_KERNEL);
286 dma_map->dma_pages = kvcalloc(nr_pages, sizeof(*dma_map->dma_pages), GFP_KERNEL);
287 if (!dma_map->dma_pages) {
292 dma_map->netdev = netdev;
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);
301 static void xp_destroy_dma_map(struct xsk_dma_map *dma_map)
303 list_del(&dma_map->list);
304 kvfree(dma_map->dma_pages);
308 static void __xp_dma_unmap(struct xsk_dma_map *dma_map, unsigned long attrs)
313 for (i = 0; i < dma_map->dma_pages_cnt; i++) {
314 dma = &dma_map->dma_pages[i];
316 dma_unmap_page_attrs(dma_map->dev, *dma, PAGE_SIZE,
317 DMA_BIDIRECTIONAL, attrs);
322 xp_destroy_dma_map(dma_map);
325 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs)
327 struct xsk_dma_map *dma_map;
329 if (pool->dma_pages_cnt == 0)
332 dma_map = xp_find_dma_map(pool);
334 WARN(1, "Could not find dma_map for device");
338 if (!refcount_dec_and_test(&dma_map->users))
341 __xp_dma_unmap(dma_map, attrs);
342 kvfree(pool->dma_pages);
343 pool->dma_pages_cnt = 0;
346 EXPORT_SYMBOL(xp_dma_unmap);
348 static void xp_check_dma_contiguity(struct xsk_dma_map *dma_map)
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;
356 dma_map->dma_pages[i] &= ~XSK_NEXT_PG_CONTIG_MASK;
360 static int xp_init_dma_info(struct xsk_buff_pool *pool, struct xsk_dma_map *dma_map)
362 pool->dma_pages = kvcalloc(dma_map->dma_pages_cnt, sizeof(*pool->dma_pages), GFP_KERNEL);
363 if (!pool->dma_pages)
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));
375 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
376 unsigned long attrs, struct page **pages, u32 nr_pages)
378 struct xsk_dma_map *dma_map;
383 dma_map = xp_find_dma_map(pool);
385 err = xp_init_dma_info(pool, dma_map);
389 refcount_inc(&dma_map->users);
393 dma_map = xp_create_dma_map(dev, pool->netdev, nr_pages, pool->umem);
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);
404 if (dma_need_sync(dev, dma))
405 dma_map->dma_need_sync = true;
406 dma_map->dma_pages[i] = dma;
410 xp_check_dma_contiguity(dma_map);
412 err = xp_init_dma_info(pool, dma_map);
414 __xp_dma_unmap(dma_map, attrs);
420 EXPORT_SYMBOL(xp_dma_map);
422 static bool xp_addr_crosses_non_contig_pg(struct xsk_buff_pool *pool,
425 return xp_desc_crosses_non_contig_pg(pool, addr, pool->chunk_size);
428 static bool xp_check_unaligned(struct xsk_buff_pool *pool, u64 *addr)
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))
438 static bool xp_check_aligned(struct xsk_buff_pool *pool, u64 *addr)
440 *addr = xp_aligned_extract_addr(pool, *addr);
441 return *addr < pool->addrs_cnt;
444 static struct xdp_buff_xsk *__xp_alloc(struct xsk_buff_pool *pool)
446 struct xdp_buff_xsk *xskb;
450 if (pool->free_heads_cnt == 0)
453 xskb = pool->free_heads[--pool->free_heads_cnt];
456 if (!xskq_cons_peek_addr_unchecked(pool->fq, &addr)) {
457 pool->fq->queue_empty_descs++;
462 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
463 xp_check_aligned(pool, &addr);
465 pool->fq->invalid_descs++;
466 xskq_cons_release(pool->fq);
471 xskq_cons_release(pool->fq);
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) +
479 xskb->dma = xskb->frame_dma + pool->headroom +
485 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool)
487 struct xdp_buff_xsk *xskb;
489 if (!pool->free_list_cnt) {
490 xskb = __xp_alloc(pool);
494 pool->free_list_cnt--;
495 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk,
497 list_del(&xskb->free_list_node);
500 xskb->xdp.data = xskb->xdp.data_hard_start + XDP_PACKET_HEADROOM;
501 xskb->xdp.data_meta = xskb->xdp.data;
503 if (pool->dma_need_sync) {
504 dma_sync_single_range_for_device(pool->dev, xskb->dma, 0,
510 EXPORT_SYMBOL(xp_alloc);
512 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count)
514 if (pool->free_list_cnt >= count)
516 return xskq_cons_has_entries(pool->fq, count - pool->free_list_cnt);
518 EXPORT_SYMBOL(xp_can_alloc);
520 void xp_free(struct xdp_buff_xsk *xskb)
522 xskb->pool->free_list_cnt++;
523 list_add(&xskb->free_list_node, &xskb->pool->free_list);
525 EXPORT_SYMBOL(xp_free);
527 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
529 addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
530 return pool->addrs + addr;
532 EXPORT_SYMBOL(xp_raw_get_data);
534 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
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) +
541 EXPORT_SYMBOL(xp_raw_get_dma);
543 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb)
545 dma_sync_single_range_for_cpu(xskb->pool->dev, xskb->dma, 0,
546 xskb->pool->frame_len, DMA_BIDIRECTIONAL);
548 EXPORT_SYMBOL(xp_dma_sync_for_cpu_slow);
550 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
553 dma_sync_single_range_for_device(pool->dev, dma, 0,
554 size, DMA_BIDIRECTIONAL);
556 EXPORT_SYMBOL(xp_dma_sync_for_device_slow);