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 spin_lock_init(&pool->cq_lock);
75 refcount_set(&pool->users, 1);
77 pool->fq = xs->fq_tmp;
78 pool->cq = xs->cq_tmp;
80 for (i = 0; i < pool->free_heads_cnt; i++) {
81 xskb = &pool->heads[i];
83 xskb->xdp.frame_sz = umem->chunk_size - umem->headroom;
84 pool->free_heads[i] = xskb;
94 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq)
98 for (i = 0; i < pool->heads_cnt; i++)
99 pool->heads[i].xdp.rxq = rxq;
101 EXPORT_SYMBOL(xp_set_rxq_info);
103 static void xp_disable_drv_zc(struct xsk_buff_pool *pool)
105 struct netdev_bpf bpf;
110 if (pool->umem->zc) {
111 bpf.command = XDP_SETUP_XSK_POOL;
113 bpf.xsk.queue_id = pool->queue_id;
115 err = pool->netdev->netdev_ops->ndo_bpf(pool->netdev, &bpf);
118 WARN(1, "Failed to disable zero-copy!\n");
122 static int __xp_assign_dev(struct xsk_buff_pool *pool,
123 struct net_device *netdev, u16 queue_id, u16 flags)
125 bool force_zc, force_copy;
126 struct netdev_bpf bpf;
131 force_zc = flags & XDP_ZEROCOPY;
132 force_copy = flags & XDP_COPY;
134 if (force_zc && force_copy)
137 if (xsk_get_pool_from_qid(netdev, queue_id))
140 pool->netdev = netdev;
141 pool->queue_id = queue_id;
142 err = xsk_reg_pool_at_qid(netdev, pool, queue_id);
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().
152 pool->cached_need_wakeup = XDP_WAKEUP_TX;
157 /* For copy-mode, we are done. */
160 if (!netdev->netdev_ops->ndo_bpf ||
161 !netdev->netdev_ops->ndo_xsk_wakeup) {
166 bpf.command = XDP_SETUP_XSK_POOL;
168 bpf.xsk.queue_id = queue_id;
170 err = netdev->netdev_ops->ndo_bpf(netdev, &bpf);
174 if (!pool->dma_pages) {
175 WARN(1, "Driver did not DMA map zero-copy buffers");
179 pool->umem->zc = true;
183 xp_disable_drv_zc(pool);
186 err = 0; /* fallback to copy mode */
188 xsk_clear_pool_at_qid(netdev, queue_id);
194 int xp_assign_dev(struct xsk_buff_pool *pool, struct net_device *dev,
195 u16 queue_id, u16 flags)
197 return __xp_assign_dev(pool, dev, queue_id, flags);
200 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_umem *umem,
201 struct net_device *dev, u16 queue_id)
205 /* One fill and completion ring required for each queue id. */
206 if (!pool->fq || !pool->cq)
209 flags = umem->zc ? XDP_ZEROCOPY : XDP_COPY;
210 if (pool->uses_need_wakeup)
211 flags |= XDP_USE_NEED_WAKEUP;
213 return __xp_assign_dev(pool, dev, queue_id, flags);
216 void xp_clear_dev(struct xsk_buff_pool *pool)
221 xp_disable_drv_zc(pool);
222 xsk_clear_pool_at_qid(pool->netdev, pool->queue_id);
223 dev_put(pool->netdev);
227 static void xp_release_deferred(struct work_struct *work)
229 struct xsk_buff_pool *pool = container_of(work, struct xsk_buff_pool,
237 xskq_destroy(pool->fq);
242 xskq_destroy(pool->cq);
246 xdp_put_umem(pool->umem, false);
250 void xp_get_pool(struct xsk_buff_pool *pool)
252 refcount_inc(&pool->users);
255 bool xp_put_pool(struct xsk_buff_pool *pool)
260 if (refcount_dec_and_test(&pool->users)) {
261 INIT_WORK(&pool->work, xp_release_deferred);
262 schedule_work(&pool->work);
269 static struct xsk_dma_map *xp_find_dma_map(struct xsk_buff_pool *pool)
271 struct xsk_dma_map *dma_map;
273 list_for_each_entry(dma_map, &pool->umem->xsk_dma_list, list) {
274 if (dma_map->netdev == pool->netdev)
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)
284 struct xsk_dma_map *dma_map;
286 dma_map = kzalloc(sizeof(*dma_map), GFP_KERNEL);
290 dma_map->dma_pages = kvcalloc(nr_pages, sizeof(*dma_map->dma_pages), GFP_KERNEL);
291 if (!dma_map->dma_pages) {
296 dma_map->netdev = netdev;
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);
305 static void xp_destroy_dma_map(struct xsk_dma_map *dma_map)
307 list_del(&dma_map->list);
308 kvfree(dma_map->dma_pages);
312 static void __xp_dma_unmap(struct xsk_dma_map *dma_map, unsigned long attrs)
317 for (i = 0; i < dma_map->dma_pages_cnt; i++) {
318 dma = &dma_map->dma_pages[i];
320 dma_unmap_page_attrs(dma_map->dev, *dma, PAGE_SIZE,
321 DMA_BIDIRECTIONAL, attrs);
326 xp_destroy_dma_map(dma_map);
329 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs)
331 struct xsk_dma_map *dma_map;
333 if (pool->dma_pages_cnt == 0)
336 dma_map = xp_find_dma_map(pool);
338 WARN(1, "Could not find dma_map for device");
342 if (!refcount_dec_and_test(&dma_map->users))
345 __xp_dma_unmap(dma_map, attrs);
346 kvfree(pool->dma_pages);
347 pool->dma_pages_cnt = 0;
350 EXPORT_SYMBOL(xp_dma_unmap);
352 static void xp_check_dma_contiguity(struct xsk_dma_map *dma_map)
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;
360 dma_map->dma_pages[i] &= ~XSK_NEXT_PG_CONTIG_MASK;
364 static int xp_init_dma_info(struct xsk_buff_pool *pool, struct xsk_dma_map *dma_map)
366 pool->dma_pages = kvcalloc(dma_map->dma_pages_cnt, sizeof(*pool->dma_pages), GFP_KERNEL);
367 if (!pool->dma_pages)
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));
379 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
380 unsigned long attrs, struct page **pages, u32 nr_pages)
382 struct xsk_dma_map *dma_map;
387 dma_map = xp_find_dma_map(pool);
389 err = xp_init_dma_info(pool, dma_map);
393 refcount_inc(&dma_map->users);
397 dma_map = xp_create_dma_map(dev, pool->netdev, nr_pages, pool->umem);
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);
408 if (dma_need_sync(dev, dma))
409 dma_map->dma_need_sync = true;
410 dma_map->dma_pages[i] = dma;
414 xp_check_dma_contiguity(dma_map);
416 err = xp_init_dma_info(pool, dma_map);
418 __xp_dma_unmap(dma_map, attrs);
424 EXPORT_SYMBOL(xp_dma_map);
426 static bool xp_addr_crosses_non_contig_pg(struct xsk_buff_pool *pool,
429 return xp_desc_crosses_non_contig_pg(pool, addr, pool->chunk_size);
432 static bool xp_check_unaligned(struct xsk_buff_pool *pool, u64 *addr)
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))
442 static bool xp_check_aligned(struct xsk_buff_pool *pool, u64 *addr)
444 *addr = xp_aligned_extract_addr(pool, *addr);
445 return *addr < pool->addrs_cnt;
448 static struct xdp_buff_xsk *__xp_alloc(struct xsk_buff_pool *pool)
450 struct xdp_buff_xsk *xskb;
454 if (pool->free_heads_cnt == 0)
457 xskb = pool->free_heads[--pool->free_heads_cnt];
460 if (!xskq_cons_peek_addr_unchecked(pool->fq, &addr)) {
461 pool->fq->queue_empty_descs++;
466 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
467 xp_check_aligned(pool, &addr);
469 pool->fq->invalid_descs++;
470 xskq_cons_release(pool->fq);
475 xskq_cons_release(pool->fq);
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) +
483 xskb->dma = xskb->frame_dma + pool->headroom +
489 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool)
491 struct xdp_buff_xsk *xskb;
493 if (!pool->free_list_cnt) {
494 xskb = __xp_alloc(pool);
498 pool->free_list_cnt--;
499 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk,
501 list_del(&xskb->free_list_node);
504 xskb->xdp.data = xskb->xdp.data_hard_start + XDP_PACKET_HEADROOM;
505 xskb->xdp.data_meta = xskb->xdp.data;
507 if (pool->dma_need_sync) {
508 dma_sync_single_range_for_device(pool->dev, xskb->dma, 0,
514 EXPORT_SYMBOL(xp_alloc);
516 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count)
518 if (pool->free_list_cnt >= count)
520 return xskq_cons_has_entries(pool->fq, count - pool->free_list_cnt);
522 EXPORT_SYMBOL(xp_can_alloc);
524 void xp_free(struct xdp_buff_xsk *xskb)
526 xskb->pool->free_list_cnt++;
527 list_add(&xskb->free_list_node, &xskb->pool->free_list);
529 EXPORT_SYMBOL(xp_free);
531 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
533 addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
534 return pool->addrs + addr;
536 EXPORT_SYMBOL(xp_raw_get_data);
538 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
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) +
545 EXPORT_SYMBOL(xp_raw_get_dma);
547 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb)
549 dma_sync_single_range_for_cpu(xskb->pool->dev, xskb->dma, 0,
550 xskb->pool->frame_len, DMA_BIDIRECTIONAL);
552 EXPORT_SYMBOL(xp_dma_sync_for_cpu_slow);
554 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
557 dma_sync_single_range_for_device(pool->dev, dma, 0,
558 size, DMA_BIDIRECTIONAL);
560 EXPORT_SYMBOL(xp_dma_sync_for_device_slow);