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 bool unaligned = umem->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG;
59 struct xsk_buff_pool *pool;
60 struct xdp_buff_xsk *xskb;
63 entries = unaligned ? umem->chunks : 0;
64 pool = kvzalloc(struct_size(pool, free_heads, entries), GFP_KERNEL);
68 pool->heads = kvcalloc(umem->chunks, sizeof(*pool->heads), GFP_KERNEL);
73 if (xp_alloc_tx_descs(pool, xs))
76 pool->chunk_mask = ~((u64)umem->chunk_size - 1);
77 pool->addrs_cnt = umem->size;
78 pool->heads_cnt = umem->chunks;
79 pool->free_heads_cnt = umem->chunks;
80 pool->headroom = umem->headroom;
81 pool->chunk_size = umem->chunk_size;
82 pool->chunk_shift = ffs(umem->chunk_size) - 1;
83 pool->unaligned = unaligned;
84 pool->frame_len = umem->chunk_size - umem->headroom -
87 pool->addrs = umem->addrs;
88 INIT_LIST_HEAD(&pool->free_list);
89 INIT_LIST_HEAD(&pool->xskb_list);
90 INIT_LIST_HEAD(&pool->xsk_tx_list);
91 spin_lock_init(&pool->xsk_tx_list_lock);
92 spin_lock_init(&pool->cq_lock);
93 refcount_set(&pool->users, 1);
95 pool->fq = xs->fq_tmp;
96 pool->cq = xs->cq_tmp;
98 for (i = 0; i < pool->free_heads_cnt; i++) {
99 xskb = &pool->heads[i];
101 xskb->xdp.frame_sz = umem->chunk_size - umem->headroom;
102 INIT_LIST_HEAD(&xskb->free_list_node);
103 INIT_LIST_HEAD(&xskb->xskb_list_node);
105 pool->free_heads[i] = xskb;
107 xp_init_xskb_addr(xskb, pool, i * pool->chunk_size);
117 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq)
121 for (i = 0; i < pool->heads_cnt; i++)
122 pool->heads[i].xdp.rxq = rxq;
124 EXPORT_SYMBOL(xp_set_rxq_info);
126 static void xp_disable_drv_zc(struct xsk_buff_pool *pool)
128 struct netdev_bpf bpf;
133 if (pool->umem->zc) {
134 bpf.command = XDP_SETUP_XSK_POOL;
136 bpf.xsk.queue_id = pool->queue_id;
138 err = pool->netdev->netdev_ops->ndo_bpf(pool->netdev, &bpf);
141 WARN(1, "Failed to disable zero-copy!\n");
145 #define NETDEV_XDP_ACT_ZC (NETDEV_XDP_ACT_BASIC | \
146 NETDEV_XDP_ACT_REDIRECT | \
147 NETDEV_XDP_ACT_XSK_ZEROCOPY)
149 int xp_assign_dev(struct xsk_buff_pool *pool,
150 struct net_device *netdev, u16 queue_id, u16 flags)
152 bool force_zc, force_copy;
153 struct netdev_bpf bpf;
158 force_zc = flags & XDP_ZEROCOPY;
159 force_copy = flags & XDP_COPY;
161 if (force_zc && force_copy)
164 if (xsk_get_pool_from_qid(netdev, queue_id))
167 pool->netdev = netdev;
168 pool->queue_id = queue_id;
169 err = xsk_reg_pool_at_qid(netdev, pool, queue_id);
173 if (flags & XDP_USE_NEED_WAKEUP)
174 pool->uses_need_wakeup = true;
175 /* Tx needs to be explicitly woken up the first time. Also
176 * for supporting drivers that do not implement this
177 * feature. They will always have to call sendto() or poll().
179 pool->cached_need_wakeup = XDP_WAKEUP_TX;
184 /* For copy-mode, we are done. */
187 if ((netdev->xdp_features & NETDEV_XDP_ACT_ZC) != NETDEV_XDP_ACT_ZC) {
192 if (netdev->xdp_zc_max_segs == 1 && (flags & XDP_USE_SG)) {
197 bpf.command = XDP_SETUP_XSK_POOL;
199 bpf.xsk.queue_id = queue_id;
201 err = netdev->netdev_ops->ndo_bpf(netdev, &bpf);
205 if (!pool->dma_pages) {
206 WARN(1, "Driver did not DMA map zero-copy buffers");
210 pool->umem->zc = true;
214 xp_disable_drv_zc(pool);
217 err = 0; /* fallback to copy mode */
219 xsk_clear_pool_at_qid(netdev, queue_id);
225 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs,
226 struct net_device *dev, u16 queue_id)
229 struct xdp_umem *umem = umem_xs->umem;
231 /* One fill and completion ring required for each queue id. */
232 if (!pool->fq || !pool->cq)
235 flags = umem->zc ? XDP_ZEROCOPY : XDP_COPY;
236 if (umem_xs->pool->uses_need_wakeup)
237 flags |= XDP_USE_NEED_WAKEUP;
239 return xp_assign_dev(pool, dev, queue_id, flags);
242 void xp_clear_dev(struct xsk_buff_pool *pool)
247 xp_disable_drv_zc(pool);
248 xsk_clear_pool_at_qid(pool->netdev, pool->queue_id);
249 dev_put(pool->netdev);
253 static void xp_release_deferred(struct work_struct *work)
255 struct xsk_buff_pool *pool = container_of(work, struct xsk_buff_pool,
263 xskq_destroy(pool->fq);
268 xskq_destroy(pool->cq);
272 xdp_put_umem(pool->umem, false);
276 void xp_get_pool(struct xsk_buff_pool *pool)
278 refcount_inc(&pool->users);
281 bool xp_put_pool(struct xsk_buff_pool *pool)
286 if (refcount_dec_and_test(&pool->users)) {
287 INIT_WORK(&pool->work, xp_release_deferred);
288 schedule_work(&pool->work);
295 static struct xsk_dma_map *xp_find_dma_map(struct xsk_buff_pool *pool)
297 struct xsk_dma_map *dma_map;
299 list_for_each_entry(dma_map, &pool->umem->xsk_dma_list, list) {
300 if (dma_map->netdev == pool->netdev)
307 static struct xsk_dma_map *xp_create_dma_map(struct device *dev, struct net_device *netdev,
308 u32 nr_pages, struct xdp_umem *umem)
310 struct xsk_dma_map *dma_map;
312 dma_map = kzalloc(sizeof(*dma_map), GFP_KERNEL);
316 dma_map->dma_pages = kvcalloc(nr_pages, sizeof(*dma_map->dma_pages), GFP_KERNEL);
317 if (!dma_map->dma_pages) {
322 dma_map->netdev = netdev;
324 dma_map->dma_need_sync = false;
325 dma_map->dma_pages_cnt = nr_pages;
326 refcount_set(&dma_map->users, 1);
327 list_add(&dma_map->list, &umem->xsk_dma_list);
331 static void xp_destroy_dma_map(struct xsk_dma_map *dma_map)
333 list_del(&dma_map->list);
334 kvfree(dma_map->dma_pages);
338 static void __xp_dma_unmap(struct xsk_dma_map *dma_map, unsigned long attrs)
343 for (i = 0; i < dma_map->dma_pages_cnt; i++) {
344 dma = &dma_map->dma_pages[i];
346 *dma &= ~XSK_NEXT_PG_CONTIG_MASK;
347 dma_unmap_page_attrs(dma_map->dev, *dma, PAGE_SIZE,
348 DMA_BIDIRECTIONAL, attrs);
353 xp_destroy_dma_map(dma_map);
356 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs)
358 struct xsk_dma_map *dma_map;
360 if (!pool->dma_pages)
363 dma_map = xp_find_dma_map(pool);
365 WARN(1, "Could not find dma_map for device");
369 if (!refcount_dec_and_test(&dma_map->users))
372 __xp_dma_unmap(dma_map, attrs);
373 kvfree(pool->dma_pages);
374 pool->dma_pages = NULL;
375 pool->dma_pages_cnt = 0;
378 EXPORT_SYMBOL(xp_dma_unmap);
380 static void xp_check_dma_contiguity(struct xsk_dma_map *dma_map)
384 for (i = 0; i < dma_map->dma_pages_cnt - 1; i++) {
385 if (dma_map->dma_pages[i] + PAGE_SIZE == dma_map->dma_pages[i + 1])
386 dma_map->dma_pages[i] |= XSK_NEXT_PG_CONTIG_MASK;
388 dma_map->dma_pages[i] &= ~XSK_NEXT_PG_CONTIG_MASK;
392 static int xp_init_dma_info(struct xsk_buff_pool *pool, struct xsk_dma_map *dma_map)
394 if (!pool->unaligned) {
397 for (i = 0; i < pool->heads_cnt; i++) {
398 struct xdp_buff_xsk *xskb = &pool->heads[i];
400 xp_init_xskb_dma(xskb, pool, dma_map->dma_pages, xskb->orig_addr);
404 pool->dma_pages = kvcalloc(dma_map->dma_pages_cnt, sizeof(*pool->dma_pages), GFP_KERNEL);
405 if (!pool->dma_pages)
408 pool->dev = dma_map->dev;
409 pool->dma_pages_cnt = dma_map->dma_pages_cnt;
410 pool->dma_need_sync = dma_map->dma_need_sync;
411 memcpy(pool->dma_pages, dma_map->dma_pages,
412 pool->dma_pages_cnt * sizeof(*pool->dma_pages));
417 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
418 unsigned long attrs, struct page **pages, u32 nr_pages)
420 struct xsk_dma_map *dma_map;
425 dma_map = xp_find_dma_map(pool);
427 err = xp_init_dma_info(pool, dma_map);
431 refcount_inc(&dma_map->users);
435 dma_map = xp_create_dma_map(dev, pool->netdev, nr_pages, pool->umem);
439 for (i = 0; i < dma_map->dma_pages_cnt; i++) {
440 dma = dma_map_page_attrs(dev, pages[i], 0, PAGE_SIZE,
441 DMA_BIDIRECTIONAL, attrs);
442 if (dma_mapping_error(dev, dma)) {
443 __xp_dma_unmap(dma_map, attrs);
446 if (dma_need_sync(dev, dma))
447 dma_map->dma_need_sync = true;
448 dma_map->dma_pages[i] = dma;
452 xp_check_dma_contiguity(dma_map);
454 err = xp_init_dma_info(pool, dma_map);
456 __xp_dma_unmap(dma_map, attrs);
462 EXPORT_SYMBOL(xp_dma_map);
464 static bool xp_addr_crosses_non_contig_pg(struct xsk_buff_pool *pool,
467 return xp_desc_crosses_non_contig_pg(pool, addr, pool->chunk_size);
470 static bool xp_check_unaligned(struct xsk_buff_pool *pool, u64 *addr)
472 *addr = xp_unaligned_extract_addr(*addr);
473 if (*addr >= pool->addrs_cnt ||
474 *addr + pool->chunk_size > pool->addrs_cnt ||
475 xp_addr_crosses_non_contig_pg(pool, *addr))
480 static bool xp_check_aligned(struct xsk_buff_pool *pool, u64 *addr)
482 *addr = xp_aligned_extract_addr(pool, *addr);
483 return *addr < pool->addrs_cnt;
486 static struct xdp_buff_xsk *__xp_alloc(struct xsk_buff_pool *pool)
488 struct xdp_buff_xsk *xskb;
492 if (pool->free_heads_cnt == 0)
496 if (!xskq_cons_peek_addr_unchecked(pool->fq, &addr)) {
497 pool->fq->queue_empty_descs++;
501 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
502 xp_check_aligned(pool, &addr);
504 pool->fq->invalid_descs++;
505 xskq_cons_release(pool->fq);
511 if (pool->unaligned) {
512 xskb = pool->free_heads[--pool->free_heads_cnt];
513 xp_init_xskb_addr(xskb, pool, addr);
515 xp_init_xskb_dma(xskb, pool, pool->dma_pages, addr);
517 xskb = &pool->heads[xp_aligned_extract_idx(pool, addr)];
520 xskq_cons_release(pool->fq);
524 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool)
526 struct xdp_buff_xsk *xskb;
528 if (!pool->free_list_cnt) {
529 xskb = __xp_alloc(pool);
533 pool->free_list_cnt--;
534 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk,
536 list_del_init(&xskb->free_list_node);
539 xskb->xdp.data = xskb->xdp.data_hard_start + XDP_PACKET_HEADROOM;
540 xskb->xdp.data_meta = xskb->xdp.data;
542 if (pool->dma_need_sync) {
543 dma_sync_single_range_for_device(pool->dev, xskb->dma, 0,
549 EXPORT_SYMBOL(xp_alloc);
551 static u32 xp_alloc_new_from_fq(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max)
553 u32 i, cached_cons, nb_entries;
555 if (max > pool->free_heads_cnt)
556 max = pool->free_heads_cnt;
557 max = xskq_cons_nb_entries(pool->fq, max);
559 cached_cons = pool->fq->cached_cons;
563 struct xdp_buff_xsk *xskb;
567 __xskq_cons_read_addr_unchecked(pool->fq, cached_cons++, &addr);
569 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
570 xp_check_aligned(pool, &addr);
572 pool->fq->invalid_descs++;
577 if (pool->unaligned) {
578 xskb = pool->free_heads[--pool->free_heads_cnt];
579 xp_init_xskb_addr(xskb, pool, addr);
581 xp_init_xskb_dma(xskb, pool, pool->dma_pages, addr);
583 xskb = &pool->heads[xp_aligned_extract_idx(pool, addr)];
590 xskq_cons_release_n(pool->fq, max);
594 static u32 xp_alloc_reused(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 nb_entries)
596 struct xdp_buff_xsk *xskb;
599 nb_entries = min_t(u32, nb_entries, pool->free_list_cnt);
603 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk, free_list_node);
604 list_del_init(&xskb->free_list_node);
609 pool->free_list_cnt -= nb_entries;
614 u32 xp_alloc_batch(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max)
616 u32 nb_entries1 = 0, nb_entries2;
618 if (unlikely(pool->dma_need_sync)) {
619 struct xdp_buff *buff;
622 buff = xp_alloc(pool);
628 if (unlikely(pool->free_list_cnt)) {
629 nb_entries1 = xp_alloc_reused(pool, xdp, max);
630 if (nb_entries1 == max)
637 nb_entries2 = xp_alloc_new_from_fq(pool, xdp, max);
639 pool->fq->queue_empty_descs++;
641 return nb_entries1 + nb_entries2;
643 EXPORT_SYMBOL(xp_alloc_batch);
645 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count)
647 if (pool->free_list_cnt >= count)
649 return xskq_cons_has_entries(pool->fq, count - pool->free_list_cnt);
651 EXPORT_SYMBOL(xp_can_alloc);
653 void xp_free(struct xdp_buff_xsk *xskb)
655 if (!list_empty(&xskb->free_list_node))
658 xskb->pool->free_list_cnt++;
659 list_add(&xskb->free_list_node, &xskb->pool->free_list);
661 EXPORT_SYMBOL(xp_free);
663 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
665 addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
666 return pool->addrs + addr;
668 EXPORT_SYMBOL(xp_raw_get_data);
670 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
672 addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
673 return (pool->dma_pages[addr >> PAGE_SHIFT] &
674 ~XSK_NEXT_PG_CONTIG_MASK) +
677 EXPORT_SYMBOL(xp_raw_get_dma);
679 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb)
681 dma_sync_single_range_for_cpu(xskb->pool->dev, xskb->dma, 0,
682 xskb->pool->frame_len, DMA_BIDIRECTIONAL);
684 EXPORT_SYMBOL(xp_dma_sync_for_cpu_slow);
686 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
689 dma_sync_single_range_for_device(pool->dev, dma, 0,
690 size, DMA_BIDIRECTIONAL);
692 EXPORT_SYMBOL(xp_dma_sync_for_device_slow);