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 pool->tx_metadata_len = umem->tx_metadata_len;
89 pool->tx_sw_csum = umem->flags & XDP_UMEM_TX_SW_CSUM;
90 INIT_LIST_HEAD(&pool->free_list);
91 INIT_LIST_HEAD(&pool->xskb_list);
92 INIT_LIST_HEAD(&pool->xsk_tx_list);
93 spin_lock_init(&pool->xsk_tx_list_lock);
94 spin_lock_init(&pool->cq_lock);
95 refcount_set(&pool->users, 1);
97 pool->fq = xs->fq_tmp;
98 pool->cq = xs->cq_tmp;
100 for (i = 0; i < pool->free_heads_cnt; i++) {
101 xskb = &pool->heads[i];
103 xskb->xdp.frame_sz = umem->chunk_size - umem->headroom;
104 INIT_LIST_HEAD(&xskb->free_list_node);
105 INIT_LIST_HEAD(&xskb->xskb_list_node);
107 pool->free_heads[i] = xskb;
109 xp_init_xskb_addr(xskb, pool, i * pool->chunk_size);
119 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq)
123 for (i = 0; i < pool->heads_cnt; i++)
124 pool->heads[i].xdp.rxq = rxq;
126 EXPORT_SYMBOL(xp_set_rxq_info);
128 void xp_fill_cb(struct xsk_buff_pool *pool, struct xsk_cb_desc *desc)
132 for (i = 0; i < pool->heads_cnt; i++) {
133 struct xdp_buff_xsk *xskb = &pool->heads[i];
135 memcpy(xskb->cb + desc->off, desc->src, desc->bytes);
138 EXPORT_SYMBOL(xp_fill_cb);
140 static void xp_disable_drv_zc(struct xsk_buff_pool *pool)
142 struct netdev_bpf bpf;
147 if (pool->umem->zc) {
148 bpf.command = XDP_SETUP_XSK_POOL;
150 bpf.xsk.queue_id = pool->queue_id;
152 err = pool->netdev->netdev_ops->ndo_bpf(pool->netdev, &bpf);
155 WARN(1, "Failed to disable zero-copy!\n");
159 #define NETDEV_XDP_ACT_ZC (NETDEV_XDP_ACT_BASIC | \
160 NETDEV_XDP_ACT_REDIRECT | \
161 NETDEV_XDP_ACT_XSK_ZEROCOPY)
163 int xp_assign_dev(struct xsk_buff_pool *pool,
164 struct net_device *netdev, u16 queue_id, u16 flags)
166 bool force_zc, force_copy;
167 struct netdev_bpf bpf;
172 force_zc = flags & XDP_ZEROCOPY;
173 force_copy = flags & XDP_COPY;
175 if (force_zc && force_copy)
178 if (xsk_get_pool_from_qid(netdev, queue_id))
181 pool->netdev = netdev;
182 pool->queue_id = queue_id;
183 err = xsk_reg_pool_at_qid(netdev, pool, queue_id);
187 if (flags & XDP_USE_SG)
188 pool->umem->flags |= XDP_UMEM_SG_FLAG;
190 if (flags & XDP_USE_NEED_WAKEUP)
191 pool->uses_need_wakeup = true;
192 /* Tx needs to be explicitly woken up the first time. Also
193 * for supporting drivers that do not implement this
194 * feature. They will always have to call sendto() or poll().
196 pool->cached_need_wakeup = XDP_WAKEUP_TX;
201 /* For copy-mode, we are done. */
204 if ((netdev->xdp_features & NETDEV_XDP_ACT_ZC) != NETDEV_XDP_ACT_ZC) {
209 if (netdev->xdp_zc_max_segs == 1 && (flags & XDP_USE_SG)) {
214 bpf.command = XDP_SETUP_XSK_POOL;
216 bpf.xsk.queue_id = queue_id;
218 err = netdev->netdev_ops->ndo_bpf(netdev, &bpf);
222 if (!pool->dma_pages) {
223 WARN(1, "Driver did not DMA map zero-copy buffers");
227 pool->umem->zc = true;
231 xp_disable_drv_zc(pool);
234 err = 0; /* fallback to copy mode */
236 xsk_clear_pool_at_qid(netdev, queue_id);
242 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs,
243 struct net_device *dev, u16 queue_id)
246 struct xdp_umem *umem = umem_xs->umem;
248 /* One fill and completion ring required for each queue id. */
249 if (!pool->fq || !pool->cq)
252 flags = umem->zc ? XDP_ZEROCOPY : XDP_COPY;
253 if (umem_xs->pool->uses_need_wakeup)
254 flags |= XDP_USE_NEED_WAKEUP;
256 return xp_assign_dev(pool, dev, queue_id, flags);
259 void xp_clear_dev(struct xsk_buff_pool *pool)
264 xp_disable_drv_zc(pool);
265 xsk_clear_pool_at_qid(pool->netdev, pool->queue_id);
266 dev_put(pool->netdev);
270 static void xp_release_deferred(struct work_struct *work)
272 struct xsk_buff_pool *pool = container_of(work, struct xsk_buff_pool,
280 xskq_destroy(pool->fq);
285 xskq_destroy(pool->cq);
289 xdp_put_umem(pool->umem, false);
293 void xp_get_pool(struct xsk_buff_pool *pool)
295 refcount_inc(&pool->users);
298 bool xp_put_pool(struct xsk_buff_pool *pool)
303 if (refcount_dec_and_test(&pool->users)) {
304 INIT_WORK(&pool->work, xp_release_deferred);
305 schedule_work(&pool->work);
312 static struct xsk_dma_map *xp_find_dma_map(struct xsk_buff_pool *pool)
314 struct xsk_dma_map *dma_map;
316 list_for_each_entry(dma_map, &pool->umem->xsk_dma_list, list) {
317 if (dma_map->netdev == pool->netdev)
324 static struct xsk_dma_map *xp_create_dma_map(struct device *dev, struct net_device *netdev,
325 u32 nr_pages, struct xdp_umem *umem)
327 struct xsk_dma_map *dma_map;
329 dma_map = kzalloc(sizeof(*dma_map), GFP_KERNEL);
333 dma_map->dma_pages = kvcalloc(nr_pages, sizeof(*dma_map->dma_pages), GFP_KERNEL);
334 if (!dma_map->dma_pages) {
339 dma_map->netdev = netdev;
341 dma_map->dma_need_sync = false;
342 dma_map->dma_pages_cnt = nr_pages;
343 refcount_set(&dma_map->users, 1);
344 list_add(&dma_map->list, &umem->xsk_dma_list);
348 static void xp_destroy_dma_map(struct xsk_dma_map *dma_map)
350 list_del(&dma_map->list);
351 kvfree(dma_map->dma_pages);
355 static void __xp_dma_unmap(struct xsk_dma_map *dma_map, unsigned long attrs)
360 for (i = 0; i < dma_map->dma_pages_cnt; i++) {
361 dma = &dma_map->dma_pages[i];
363 *dma &= ~XSK_NEXT_PG_CONTIG_MASK;
364 dma_unmap_page_attrs(dma_map->dev, *dma, PAGE_SIZE,
365 DMA_BIDIRECTIONAL, attrs);
370 xp_destroy_dma_map(dma_map);
373 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs)
375 struct xsk_dma_map *dma_map;
377 if (!pool->dma_pages)
380 dma_map = xp_find_dma_map(pool);
382 WARN(1, "Could not find dma_map for device");
386 if (!refcount_dec_and_test(&dma_map->users))
389 __xp_dma_unmap(dma_map, attrs);
390 kvfree(pool->dma_pages);
391 pool->dma_pages = NULL;
392 pool->dma_pages_cnt = 0;
395 EXPORT_SYMBOL(xp_dma_unmap);
397 static void xp_check_dma_contiguity(struct xsk_dma_map *dma_map)
401 for (i = 0; i < dma_map->dma_pages_cnt - 1; i++) {
402 if (dma_map->dma_pages[i] + PAGE_SIZE == dma_map->dma_pages[i + 1])
403 dma_map->dma_pages[i] |= XSK_NEXT_PG_CONTIG_MASK;
405 dma_map->dma_pages[i] &= ~XSK_NEXT_PG_CONTIG_MASK;
409 static int xp_init_dma_info(struct xsk_buff_pool *pool, struct xsk_dma_map *dma_map)
411 if (!pool->unaligned) {
414 for (i = 0; i < pool->heads_cnt; i++) {
415 struct xdp_buff_xsk *xskb = &pool->heads[i];
417 xp_init_xskb_dma(xskb, pool, dma_map->dma_pages, xskb->orig_addr);
421 pool->dma_pages = kvcalloc(dma_map->dma_pages_cnt, sizeof(*pool->dma_pages), GFP_KERNEL);
422 if (!pool->dma_pages)
425 pool->dev = dma_map->dev;
426 pool->dma_pages_cnt = dma_map->dma_pages_cnt;
427 pool->dma_need_sync = dma_map->dma_need_sync;
428 memcpy(pool->dma_pages, dma_map->dma_pages,
429 pool->dma_pages_cnt * sizeof(*pool->dma_pages));
434 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
435 unsigned long attrs, struct page **pages, u32 nr_pages)
437 struct xsk_dma_map *dma_map;
442 dma_map = xp_find_dma_map(pool);
444 err = xp_init_dma_info(pool, dma_map);
448 refcount_inc(&dma_map->users);
452 dma_map = xp_create_dma_map(dev, pool->netdev, nr_pages, pool->umem);
456 for (i = 0; i < dma_map->dma_pages_cnt; i++) {
457 dma = dma_map_page_attrs(dev, pages[i], 0, PAGE_SIZE,
458 DMA_BIDIRECTIONAL, attrs);
459 if (dma_mapping_error(dev, dma)) {
460 __xp_dma_unmap(dma_map, attrs);
463 if (dma_need_sync(dev, dma))
464 dma_map->dma_need_sync = true;
465 dma_map->dma_pages[i] = dma;
469 xp_check_dma_contiguity(dma_map);
471 err = xp_init_dma_info(pool, dma_map);
473 __xp_dma_unmap(dma_map, attrs);
479 EXPORT_SYMBOL(xp_dma_map);
481 static bool xp_addr_crosses_non_contig_pg(struct xsk_buff_pool *pool,
484 return xp_desc_crosses_non_contig_pg(pool, addr, pool->chunk_size);
487 static bool xp_check_unaligned(struct xsk_buff_pool *pool, u64 *addr)
489 *addr = xp_unaligned_extract_addr(*addr);
490 if (*addr >= pool->addrs_cnt ||
491 *addr + pool->chunk_size > pool->addrs_cnt ||
492 xp_addr_crosses_non_contig_pg(pool, *addr))
497 static bool xp_check_aligned(struct xsk_buff_pool *pool, u64 *addr)
499 *addr = xp_aligned_extract_addr(pool, *addr);
500 return *addr < pool->addrs_cnt;
503 static struct xdp_buff_xsk *__xp_alloc(struct xsk_buff_pool *pool)
505 struct xdp_buff_xsk *xskb;
509 if (pool->free_heads_cnt == 0)
513 if (!xskq_cons_peek_addr_unchecked(pool->fq, &addr)) {
514 pool->fq->queue_empty_descs++;
518 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
519 xp_check_aligned(pool, &addr);
521 pool->fq->invalid_descs++;
522 xskq_cons_release(pool->fq);
528 if (pool->unaligned) {
529 xskb = pool->free_heads[--pool->free_heads_cnt];
530 xp_init_xskb_addr(xskb, pool, addr);
532 xp_init_xskb_dma(xskb, pool, pool->dma_pages, addr);
534 xskb = &pool->heads[xp_aligned_extract_idx(pool, addr)];
537 xskq_cons_release(pool->fq);
541 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool)
543 struct xdp_buff_xsk *xskb;
545 if (!pool->free_list_cnt) {
546 xskb = __xp_alloc(pool);
550 pool->free_list_cnt--;
551 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk,
553 list_del_init(&xskb->free_list_node);
556 xskb->xdp.data = xskb->xdp.data_hard_start + XDP_PACKET_HEADROOM;
557 xskb->xdp.data_meta = xskb->xdp.data;
560 if (pool->dma_need_sync) {
561 dma_sync_single_range_for_device(pool->dev, xskb->dma, 0,
567 EXPORT_SYMBOL(xp_alloc);
569 static u32 xp_alloc_new_from_fq(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max)
571 u32 i, cached_cons, nb_entries;
573 if (max > pool->free_heads_cnt)
574 max = pool->free_heads_cnt;
575 max = xskq_cons_nb_entries(pool->fq, max);
577 cached_cons = pool->fq->cached_cons;
581 struct xdp_buff_xsk *xskb;
585 __xskq_cons_read_addr_unchecked(pool->fq, cached_cons++, &addr);
587 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
588 xp_check_aligned(pool, &addr);
590 pool->fq->invalid_descs++;
595 if (pool->unaligned) {
596 xskb = pool->free_heads[--pool->free_heads_cnt];
597 xp_init_xskb_addr(xskb, pool, addr);
599 xp_init_xskb_dma(xskb, pool, pool->dma_pages, addr);
601 xskb = &pool->heads[xp_aligned_extract_idx(pool, addr)];
608 xskq_cons_release_n(pool->fq, max);
612 static u32 xp_alloc_reused(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 nb_entries)
614 struct xdp_buff_xsk *xskb;
617 nb_entries = min_t(u32, nb_entries, pool->free_list_cnt);
621 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk, free_list_node);
622 list_del_init(&xskb->free_list_node);
627 pool->free_list_cnt -= nb_entries;
632 u32 xp_alloc_batch(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max)
634 u32 nb_entries1 = 0, nb_entries2;
636 if (unlikely(pool->dma_need_sync)) {
637 struct xdp_buff *buff;
640 buff = xp_alloc(pool);
646 if (unlikely(pool->free_list_cnt)) {
647 nb_entries1 = xp_alloc_reused(pool, xdp, max);
648 if (nb_entries1 == max)
655 nb_entries2 = xp_alloc_new_from_fq(pool, xdp, max);
657 pool->fq->queue_empty_descs++;
659 return nb_entries1 + nb_entries2;
661 EXPORT_SYMBOL(xp_alloc_batch);
663 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count)
665 if (pool->free_list_cnt >= count)
667 return xskq_cons_has_entries(pool->fq, count - pool->free_list_cnt);
669 EXPORT_SYMBOL(xp_can_alloc);
671 void xp_free(struct xdp_buff_xsk *xskb)
673 if (!list_empty(&xskb->free_list_node))
676 xskb->pool->free_list_cnt++;
677 list_add(&xskb->free_list_node, &xskb->pool->free_list);
679 EXPORT_SYMBOL(xp_free);
681 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
683 addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
684 return pool->addrs + addr;
686 EXPORT_SYMBOL(xp_raw_get_data);
688 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
690 addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
691 return (pool->dma_pages[addr >> PAGE_SHIFT] &
692 ~XSK_NEXT_PG_CONTIG_MASK) +
695 EXPORT_SYMBOL(xp_raw_get_dma);
697 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb)
699 dma_sync_single_range_for_cpu(xskb->pool->dev, xskb->dma, 0,
700 xskb->pool->frame_len, DMA_BIDIRECTIONAL);
702 EXPORT_SYMBOL(xp_dma_sync_for_cpu_slow);
704 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
707 dma_sync_single_range_for_device(pool->dev, dma, 0,
708 size, DMA_BIDIRECTIONAL);
710 EXPORT_SYMBOL(xp_dma_sync_for_device_slow);