include/net/xsk_buff_pool.h

0001 /* SPDX-License-Identifier: GPL-2.0 */
0002 /* Copyright(c) 2020 Intel Corporation. */
0003
0004 #ifndef XSK_BUFF_POOL_H_
0005 #define XSK_BUFF_POOL_H_
0006
0007 #include <linux/if_xdp.h>
0008 #include <linux/types.h>
0009 #include <linux/dma-mapping.h>
0010 #include <linux/bpf.h>
0011 #include <net/xdp.h>
0012
0013 struct xsk_buff_pool;
0014 struct xdp_rxq_info;
0015 struct xsk_queue;
0016 struct xdp_desc;
0017 struct xdp_umem;
0018 struct xdp_sock;
0019 struct device;
0020 struct page;
0021
0022 struct xdp_buff_xsk {
0023     struct xdp_buff xdp;
0024     dma_addr_t dma;
0025     dma_addr_t frame_dma;
0026     struct xsk_buff_pool *pool;
0027     u64 orig_addr;
0028     struct list_head free_list_node;
0029 };
0030
0031 struct xsk_dma_map {
0032     dma_addr_t *dma_pages;
0033     struct device *dev;
0034     struct net_device *netdev;
0035     refcount_t users;
0036     struct list_head list; /* Protected by the RTNL_LOCK */
0037     u32 dma_pages_cnt;
0038     bool dma_need_sync;
0039 };
0040
0041 struct xsk_buff_pool {
0042     /* Members only used in the control path first. */
0043     struct device *dev;
0044     struct net_device *netdev;
0045     struct list_head xsk_tx_list;
0046     /* Protects modifications to the xsk_tx_list */
0047     spinlock_t xsk_tx_list_lock;
0048     refcount_t users;
0049     struct xdp_umem *umem;
0050     struct work_struct work;
0051     struct list_head free_list;
0052     u32 heads_cnt;
0053     u16 queue_id;
0054
0055     /* Data path members as close to free_heads at the end as possible. */
0056     struct xsk_queue *fq ____cacheline_aligned_in_smp;
0057     struct xsk_queue *cq;
0058     /* For performance reasons, each buff pool has its own array of dma_pages
0059      * even when they are identical.
0060      */
0061     dma_addr_t *dma_pages;
0062     struct xdp_buff_xsk *heads;
0063     struct xdp_desc *tx_descs;
0064     u64 chunk_mask;
0065     u64 addrs_cnt;
0066     u32 free_list_cnt;
0067     u32 dma_pages_cnt;
0068     u32 free_heads_cnt;
0069     u32 headroom;
0070     u32 chunk_size;
0071     u32 chunk_shift;
0072     u32 frame_len;
0073     u8 cached_need_wakeup;
0074     bool uses_need_wakeup;
0075     bool dma_need_sync;
0076     bool unaligned;
0077     void *addrs;
0078     /* Mutual exclusion of the completion ring in the SKB mode. Two cases to protect:
0079      * NAPI TX thread and sendmsg error paths in the SKB destructor callback and when
0080      * sockets share a single cq when the same netdev and queue id is shared.
0081      */
0082     spinlock_t cq_lock;
0083     struct xdp_buff_xsk *free_heads[];
0084 };
0085
0086 /* Masks for xdp_umem_page flags.
0087  * The low 12-bits of the addr will be 0 since this is the page address, so we
0088  * can use them for flags.
0089  */
0090 #define XSK_NEXT_PG_CONTIG_SHIFT 0
0091 #define XSK_NEXT_PG_CONTIG_MASK BIT_ULL(XSK_NEXT_PG_CONTIG_SHIFT)
0092
0093 /* AF_XDP core. */
0094 struct xsk_buff_pool *xp_create_and_assign_umem(struct xdp_sock *xs,
0095                         struct xdp_umem *umem);
0096 int xp_assign_dev(struct xsk_buff_pool *pool, struct net_device *dev,
0097           u16 queue_id, u16 flags);
0098 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs,
0099              struct net_device *dev, u16 queue_id);
0100 int xp_alloc_tx_descs(struct xsk_buff_pool *pool, struct xdp_sock *xs);
0101 void xp_destroy(struct xsk_buff_pool *pool);
0102 void xp_get_pool(struct xsk_buff_pool *pool);
0103 bool xp_put_pool(struct xsk_buff_pool *pool);
0104 void xp_clear_dev(struct xsk_buff_pool *pool);
0105 void xp_add_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs);
0106 void xp_del_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs);
0107
0108 /* AF_XDP, and XDP core. */
0109 void xp_free(struct xdp_buff_xsk *xskb);
0110
0111 static inline void xp_init_xskb_addr(struct xdp_buff_xsk *xskb, struct xsk_buff_pool *pool,
0112                      u64 addr)
0113 {
0114     xskb->orig_addr = addr;
0115     xskb->xdp.data_hard_start = pool->addrs + addr + pool->headroom;
0116 }
0117
0118 static inline void xp_init_xskb_dma(struct xdp_buff_xsk *xskb, struct xsk_buff_pool *pool,
0119                     dma_addr_t *dma_pages, u64 addr)
0120 {
0121     xskb->frame_dma = (dma_pages[addr >> PAGE_SHIFT] & ~XSK_NEXT_PG_CONTIG_MASK) +
0122         (addr & ~PAGE_MASK);
0123     xskb->dma = xskb->frame_dma + pool->headroom + XDP_PACKET_HEADROOM;
0124 }
0125
0126 /* AF_XDP ZC drivers, via xdp_sock_buff.h */
0127 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq);
0128 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
0129            unsigned long attrs, struct page **pages, u32 nr_pages);
0130 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs);
0131 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool);
0132 u32 xp_alloc_batch(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max);
0133 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count);
0134 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr);
0135 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr);
0136 static inline dma_addr_t xp_get_dma(struct xdp_buff_xsk *xskb)
0137 {
0138     return xskb->dma;
0139 }
0140
0141 static inline dma_addr_t xp_get_frame_dma(struct xdp_buff_xsk *xskb)
0142 {
0143     return xskb->frame_dma;
0144 }
0145
0146 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb);
0147 static inline void xp_dma_sync_for_cpu(struct xdp_buff_xsk *xskb)
0148 {
0149     xp_dma_sync_for_cpu_slow(xskb);
0150 }
0151
0152 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
0153                  size_t size);
0154 static inline void xp_dma_sync_for_device(struct xsk_buff_pool *pool,
0155                       dma_addr_t dma, size_t size)
0156 {
0157     if (!pool->dma_need_sync)
0158         return;
0159
0160     xp_dma_sync_for_device_slow(pool, dma, size);
0161 }
0162
0163 /* Masks for xdp_umem_page flags.
0164  * The low 12-bits of the addr will be 0 since this is the page address, so we
0165  * can use them for flags.
0166  */
0167 #define XSK_NEXT_PG_CONTIG_SHIFT 0
0168 #define XSK_NEXT_PG_CONTIG_MASK BIT_ULL(XSK_NEXT_PG_CONTIG_SHIFT)
0169
0170 static inline bool xp_desc_crosses_non_contig_pg(struct xsk_buff_pool *pool,
0171                          u64 addr, u32 len)
0172 {
0173     bool cross_pg = (addr & (PAGE_SIZE - 1)) + len > PAGE_SIZE;
0174
0175     if (likely(!cross_pg))
0176         return false;
0177
0178     if (pool->dma_pages_cnt) {
0179         return !(pool->dma_pages[addr >> PAGE_SHIFT] &
0180              XSK_NEXT_PG_CONTIG_MASK);
0181     }
0182
0183     /* skb path */
0184     return addr + len > pool->addrs_cnt;
0185 }
0186
0187 static inline u64 xp_aligned_extract_addr(struct xsk_buff_pool *pool, u64 addr)
0188 {
0189     return addr & pool->chunk_mask;
0190 }
0191
0192 static inline u64 xp_unaligned_extract_addr(u64 addr)
0193 {
0194     return addr & XSK_UNALIGNED_BUF_ADDR_MASK;
0195 }
0196
0197 static inline u64 xp_unaligned_extract_offset(u64 addr)
0198 {
0199     return addr >> XSK_UNALIGNED_BUF_OFFSET_SHIFT;
0200 }
0201
0202 static inline u64 xp_unaligned_add_offset_to_addr(u64 addr)
0203 {
0204     return xp_unaligned_extract_addr(addr) +
0205         xp_unaligned_extract_offset(addr);
0206 }
0207
0208 static inline u32 xp_aligned_extract_idx(struct xsk_buff_pool *pool, u64 addr)
0209 {
0210     return xp_aligned_extract_addr(pool, addr) >> pool->chunk_shift;
0211 }
0212
0213 static inline void xp_release(struct xdp_buff_xsk *xskb)
0214 {
0215     if (xskb->pool->unaligned)
0216         xskb->pool->free_heads[xskb->pool->free_heads_cnt++] = xskb;
0217 }
0218
0219 static inline u64 xp_get_handle(struct xdp_buff_xsk *xskb)
0220 {
0221     u64 offset = xskb->xdp.data - xskb->xdp.data_hard_start;
0222
0223     offset += xskb->pool->headroom;
0224     if (!xskb->pool->unaligned)
0225         return xskb->orig_addr + offset;
0226     return xskb->orig_addr + (offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT);
0227 }
0228
0229 #endif /* XSK_BUFF_POOL_H_ */