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)
40 kvfree(pool->tx_descs);
45 int xp_alloc_tx_descs(struct xsk_buff_pool *pool, struct xdp_sock *xs)
47 pool->tx_descs = kvcalloc(xs->tx->nentries, sizeof(*pool->tx_descs),
55 struct xsk_buff_pool *xp_create_and_assign_umem(struct xdp_sock *xs,
56 struct xdp_umem *umem)
58 struct xsk_buff_pool *pool;
59 struct xdp_buff_xsk *xskb;
62 pool = kvzalloc(struct_size(pool, free_heads, umem->chunks),
67 pool->heads = kvcalloc(umem->chunks, sizeof(*pool->heads), GFP_KERNEL);
72 if (xp_alloc_tx_descs(pool, xs))
75 pool->chunk_mask = ~((u64)umem->chunk_size - 1);
76 pool->addrs_cnt = umem->size;
77 pool->heads_cnt = umem->chunks;
78 pool->free_heads_cnt = umem->chunks;
79 pool->headroom = umem->headroom;
80 pool->chunk_size = umem->chunk_size;
81 pool->unaligned = umem->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG;
82 pool->frame_len = umem->chunk_size - umem->headroom -
85 pool->addrs = umem->addrs;
86 INIT_LIST_HEAD(&pool->free_list);
87 INIT_LIST_HEAD(&pool->xsk_tx_list);
88 spin_lock_init(&pool->xsk_tx_list_lock);
89 spin_lock_init(&pool->cq_lock);
90 refcount_set(&pool->users, 1);
92 pool->fq = xs->fq_tmp;
93 pool->cq = xs->cq_tmp;
95 for (i = 0; i < pool->free_heads_cnt; i++) {
96 xskb = &pool->heads[i];
98 xskb->xdp.frame_sz = umem->chunk_size - umem->headroom;
99 pool->free_heads[i] = xskb;
109 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq)
113 for (i = 0; i < pool->heads_cnt; i++)
114 pool->heads[i].xdp.rxq = rxq;
116 EXPORT_SYMBOL(xp_set_rxq_info);
118 static void xp_disable_drv_zc(struct xsk_buff_pool *pool)
120 struct netdev_bpf bpf;
125 if (pool->umem->zc) {
126 bpf.command = XDP_SETUP_XSK_POOL;
128 bpf.xsk.queue_id = pool->queue_id;
130 err = pool->netdev->netdev_ops->ndo_bpf(pool->netdev, &bpf);
133 WARN(1, "Failed to disable zero-copy!\n");
137 int xp_assign_dev(struct xsk_buff_pool *pool,
138 struct net_device *netdev, u16 queue_id, u16 flags)
140 bool force_zc, force_copy;
141 struct netdev_bpf bpf;
146 force_zc = flags & XDP_ZEROCOPY;
147 force_copy = flags & XDP_COPY;
149 if (force_zc && force_copy)
152 if (xsk_get_pool_from_qid(netdev, queue_id))
155 pool->netdev = netdev;
156 pool->queue_id = queue_id;
157 err = xsk_reg_pool_at_qid(netdev, pool, queue_id);
161 if (flags & XDP_USE_NEED_WAKEUP)
162 pool->uses_need_wakeup = true;
163 /* Tx needs to be explicitly woken up the first time. Also
164 * for supporting drivers that do not implement this
165 * feature. They will always have to call sendto() or poll().
167 pool->cached_need_wakeup = XDP_WAKEUP_TX;
172 /* For copy-mode, we are done. */
175 if (!netdev->netdev_ops->ndo_bpf ||
176 !netdev->netdev_ops->ndo_xsk_wakeup) {
181 bpf.command = XDP_SETUP_XSK_POOL;
183 bpf.xsk.queue_id = queue_id;
185 err = netdev->netdev_ops->ndo_bpf(netdev, &bpf);
189 if (!pool->dma_pages) {
190 WARN(1, "Driver did not DMA map zero-copy buffers");
194 pool->umem->zc = true;
198 xp_disable_drv_zc(pool);
201 err = 0; /* fallback to copy mode */
203 xsk_clear_pool_at_qid(netdev, queue_id);
209 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs,
210 struct net_device *dev, u16 queue_id)
213 struct xdp_umem *umem = umem_xs->umem;
215 /* One fill and completion ring required for each queue id. */
216 if (!pool->fq || !pool->cq)
219 flags = umem->zc ? XDP_ZEROCOPY : XDP_COPY;
220 if (umem_xs->pool->uses_need_wakeup)
221 flags |= XDP_USE_NEED_WAKEUP;
223 return xp_assign_dev(pool, dev, queue_id, flags);
226 void xp_clear_dev(struct xsk_buff_pool *pool)
231 xp_disable_drv_zc(pool);
232 xsk_clear_pool_at_qid(pool->netdev, pool->queue_id);
233 dev_put(pool->netdev);
237 static void xp_release_deferred(struct work_struct *work)
239 struct xsk_buff_pool *pool = container_of(work, struct xsk_buff_pool,
247 xskq_destroy(pool->fq);
252 xskq_destroy(pool->cq);
256 xdp_put_umem(pool->umem, false);
260 void xp_get_pool(struct xsk_buff_pool *pool)
262 refcount_inc(&pool->users);
265 bool xp_put_pool(struct xsk_buff_pool *pool)
270 if (refcount_dec_and_test(&pool->users)) {
271 INIT_WORK(&pool->work, xp_release_deferred);
272 schedule_work(&pool->work);
279 static struct xsk_dma_map *xp_find_dma_map(struct xsk_buff_pool *pool)
281 struct xsk_dma_map *dma_map;
283 list_for_each_entry(dma_map, &pool->umem->xsk_dma_list, list) {
284 if (dma_map->netdev == pool->netdev)
291 static struct xsk_dma_map *xp_create_dma_map(struct device *dev, struct net_device *netdev,
292 u32 nr_pages, struct xdp_umem *umem)
294 struct xsk_dma_map *dma_map;
296 dma_map = kzalloc(sizeof(*dma_map), GFP_KERNEL);
300 dma_map->dma_pages = kvcalloc(nr_pages, sizeof(*dma_map->dma_pages), GFP_KERNEL);
301 if (!dma_map->dma_pages) {
306 dma_map->netdev = netdev;
308 dma_map->dma_need_sync = false;
309 dma_map->dma_pages_cnt = nr_pages;
310 refcount_set(&dma_map->users, 1);
311 list_add(&dma_map->list, &umem->xsk_dma_list);
315 static void xp_destroy_dma_map(struct xsk_dma_map *dma_map)
317 list_del(&dma_map->list);
318 kvfree(dma_map->dma_pages);
322 static void __xp_dma_unmap(struct xsk_dma_map *dma_map, unsigned long attrs)
327 for (i = 0; i < dma_map->dma_pages_cnt; i++) {
328 dma = &dma_map->dma_pages[i];
330 *dma &= ~XSK_NEXT_PG_CONTIG_MASK;
331 dma_unmap_page_attrs(dma_map->dev, *dma, PAGE_SIZE,
332 DMA_BIDIRECTIONAL, attrs);
337 xp_destroy_dma_map(dma_map);
340 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs)
342 struct xsk_dma_map *dma_map;
344 if (pool->dma_pages_cnt == 0)
347 dma_map = xp_find_dma_map(pool);
349 WARN(1, "Could not find dma_map for device");
353 if (!refcount_dec_and_test(&dma_map->users))
356 __xp_dma_unmap(dma_map, attrs);
357 kvfree(pool->dma_pages);
358 pool->dma_pages_cnt = 0;
361 EXPORT_SYMBOL(xp_dma_unmap);
363 static void xp_check_dma_contiguity(struct xsk_dma_map *dma_map)
367 for (i = 0; i < dma_map->dma_pages_cnt - 1; i++) {
368 if (dma_map->dma_pages[i] + PAGE_SIZE == dma_map->dma_pages[i + 1])
369 dma_map->dma_pages[i] |= XSK_NEXT_PG_CONTIG_MASK;
371 dma_map->dma_pages[i] &= ~XSK_NEXT_PG_CONTIG_MASK;
375 static int xp_init_dma_info(struct xsk_buff_pool *pool, struct xsk_dma_map *dma_map)
377 pool->dma_pages = kvcalloc(dma_map->dma_pages_cnt, sizeof(*pool->dma_pages), GFP_KERNEL);
378 if (!pool->dma_pages)
381 pool->dev = dma_map->dev;
382 pool->dma_pages_cnt = dma_map->dma_pages_cnt;
383 pool->dma_need_sync = dma_map->dma_need_sync;
384 memcpy(pool->dma_pages, dma_map->dma_pages,
385 pool->dma_pages_cnt * sizeof(*pool->dma_pages));
390 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
391 unsigned long attrs, struct page **pages, u32 nr_pages)
393 struct xsk_dma_map *dma_map;
398 dma_map = xp_find_dma_map(pool);
400 err = xp_init_dma_info(pool, dma_map);
404 refcount_inc(&dma_map->users);
408 dma_map = xp_create_dma_map(dev, pool->netdev, nr_pages, pool->umem);
412 for (i = 0; i < dma_map->dma_pages_cnt; i++) {
413 dma = dma_map_page_attrs(dev, pages[i], 0, PAGE_SIZE,
414 DMA_BIDIRECTIONAL, attrs);
415 if (dma_mapping_error(dev, dma)) {
416 __xp_dma_unmap(dma_map, attrs);
419 if (dma_need_sync(dev, dma))
420 dma_map->dma_need_sync = true;
421 dma_map->dma_pages[i] = dma;
425 xp_check_dma_contiguity(dma_map);
427 err = xp_init_dma_info(pool, dma_map);
429 __xp_dma_unmap(dma_map, attrs);
435 EXPORT_SYMBOL(xp_dma_map);
437 static bool xp_addr_crosses_non_contig_pg(struct xsk_buff_pool *pool,
440 return xp_desc_crosses_non_contig_pg(pool, addr, pool->chunk_size);
443 static bool xp_check_unaligned(struct xsk_buff_pool *pool, u64 *addr)
445 *addr = xp_unaligned_extract_addr(*addr);
446 if (*addr >= pool->addrs_cnt ||
447 *addr + pool->chunk_size > pool->addrs_cnt ||
448 xp_addr_crosses_non_contig_pg(pool, *addr))
453 static bool xp_check_aligned(struct xsk_buff_pool *pool, u64 *addr)
455 *addr = xp_aligned_extract_addr(pool, *addr);
456 return *addr < pool->addrs_cnt;
459 static struct xdp_buff_xsk *__xp_alloc(struct xsk_buff_pool *pool)
461 struct xdp_buff_xsk *xskb;
465 if (pool->free_heads_cnt == 0)
468 xskb = pool->free_heads[--pool->free_heads_cnt];
471 if (!xskq_cons_peek_addr_unchecked(pool->fq, &addr)) {
472 pool->fq->queue_empty_descs++;
477 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
478 xp_check_aligned(pool, &addr);
480 pool->fq->invalid_descs++;
481 xskq_cons_release(pool->fq);
486 xskq_cons_release(pool->fq);
488 xskb->orig_addr = addr;
489 xskb->xdp.data_hard_start = pool->addrs + addr + pool->headroom;
490 if (pool->dma_pages_cnt) {
491 xskb->frame_dma = (pool->dma_pages[addr >> PAGE_SHIFT] &
492 ~XSK_NEXT_PG_CONTIG_MASK) +
494 xskb->dma = xskb->frame_dma + pool->headroom +
500 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool)
502 struct xdp_buff_xsk *xskb;
504 if (!pool->free_list_cnt) {
505 xskb = __xp_alloc(pool);
509 pool->free_list_cnt--;
510 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk,
512 list_del(&xskb->free_list_node);
515 xskb->xdp.data = xskb->xdp.data_hard_start + XDP_PACKET_HEADROOM;
516 xskb->xdp.data_meta = xskb->xdp.data;
518 if (pool->dma_need_sync) {
519 dma_sync_single_range_for_device(pool->dev, xskb->dma, 0,
525 EXPORT_SYMBOL(xp_alloc);
527 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count)
529 if (pool->free_list_cnt >= count)
531 return xskq_cons_has_entries(pool->fq, count - pool->free_list_cnt);
533 EXPORT_SYMBOL(xp_can_alloc);
535 void xp_free(struct xdp_buff_xsk *xskb)
537 xskb->pool->free_list_cnt++;
538 list_add(&xskb->free_list_node, &xskb->pool->free_list);
540 EXPORT_SYMBOL(xp_free);
542 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
544 addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
545 return pool->addrs + addr;
547 EXPORT_SYMBOL(xp_raw_get_data);
549 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
551 addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
552 return (pool->dma_pages[addr >> PAGE_SHIFT] &
553 ~XSK_NEXT_PG_CONTIG_MASK) +
556 EXPORT_SYMBOL(xp_raw_get_dma);
558 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb)
560 dma_sync_single_range_for_cpu(xskb->pool->dev, xskb->dma, 0,
561 xskb->pool->frame_len, DMA_BIDIRECTIONAL);
563 EXPORT_SYMBOL(xp_dma_sync_for_cpu_slow);
565 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
568 dma_sync_single_range_for_device(pool->dev, dma, 0,
569 size, DMA_BIDIRECTIONAL);
571 EXPORT_SYMBOL(xp_dma_sync_for_device_slow);