1 /* SPDX-License-Identifier: GPL-2.0 */ 2 /* Copyright(c) 2020 Intel Corporation. */ 3 4 #ifndef XSK_BUFF_POOL_H_ 5 #define XSK_BUFF_POOL_H_ 6 7 #include <linux/if_xdp.h> 8 #include <linux/types.h> 9 #include <linux/dma-mapping.h> 10 #include <linux/bpf.h> 11 #include <net/xdp.h> 12 13 struct xsk_buff_pool; 14 struct xdp_rxq_info; 15 struct xsk_cb_desc; 16 struct xsk_queue; 17 struct xdp_desc; 18 struct xdp_umem; 19 struct xdp_sock; 20 struct device; 21 struct page; 22 23 #define XSK_PRIV_MAX 24 24 25 struct xdp_buff_xsk { 26 struct xdp_buff xdp; 27 u8 cb[XSK_PRIV_MAX]; 28 dma_addr_t dma; 29 dma_addr_t frame_dma; 30 struct xsk_buff_pool *pool; 31 u64 orig_addr; 32 struct list_head list_node; 33 }; 34 35 #define XSK_CHECK_PRIV_TYPE(t) BUILD_BUG_ON(sizeof(t) > offsetofend(struct xdp_buff_xsk, cb)) 36 #define XSK_TX_COMPL_FITS(t) BUILD_BUG_ON(sizeof(struct xsk_tx_metadata_compl) > sizeof(t)) 37 38 struct xsk_dma_map { 39 dma_addr_t *dma_pages; 40 struct device *dev; 41 struct net_device *netdev; 42 refcount_t users; 43 struct list_head list; /* Protected by the RTNL_LOCK */ 44 u32 dma_pages_cnt; 45 }; 46 47 struct xsk_buff_pool { 48 /* Members only used in the control path first. */ 49 struct device *dev; 50 struct net_device *netdev; 51 struct list_head xsk_tx_list; 52 /* Protects modifications to the xsk_tx_list */ 53 spinlock_t xsk_tx_list_lock; 54 refcount_t users; 55 struct xdp_umem *umem; 56 struct work_struct work; 57 struct list_head free_list; 58 struct list_head xskb_list; 59 u32 heads_cnt; 60 u16 queue_id; 61 62 /* Data path members as close to free_heads at the end as possible. */ 63 struct xsk_queue *fq ____cacheline_aligned_in_smp; 64 struct xsk_queue *cq; 65 /* For performance reasons, each buff pool has its own array of dma_pages 66 * even when they are identical. 67 */ 68 dma_addr_t *dma_pages; 69 struct xdp_buff_xsk *heads; 70 struct xdp_desc *tx_descs; 71 u64 chunk_mask; 72 u64 addrs_cnt; 73 u32 free_list_cnt; 74 u32 dma_pages_cnt; 75 u32 free_heads_cnt; 76 u32 headroom; 77 u32 chunk_size; 78 u32 chunk_shift; 79 u32 frame_len; 80 u8 tx_metadata_len; /* inherited from umem */ 81 u8 cached_need_wakeup; 82 bool uses_need_wakeup; 83 bool unaligned; 84 bool tx_sw_csum; 85 void *addrs; 86 /* Mutual exclusion of the completion ring in the SKB mode. Two cases to protect: 87 * NAPI TX thread and sendmsg error paths in the SKB destructor callback and when 88 * sockets share a single cq when the same netdev and queue id is shared. 89 */ 90 spinlock_t cq_lock; 91 struct xdp_buff_xsk *free_heads[]; 92 }; 93 94 /* Masks for xdp_umem_page flags. 95 * The low 12-bits of the addr will be 0 since this is the page address, so we 96 * can use them for flags. 97 */ 98 #define XSK_NEXT_PG_CONTIG_SHIFT 0 99 #define XSK_NEXT_PG_CONTIG_MASK BIT_ULL(XSK_NEXT_PG_CONTIG_SHIFT) 100 101 /* AF_XDP core. */ 102 struct xsk_buff_pool *xp_create_and_assign_umem(struct xdp_sock *xs, 103 struct xdp_umem *umem); 104 int xp_assign_dev(struct xsk_buff_pool *pool, struct net_device *dev, 105 u16 queue_id, u16 flags); 106 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs, 107 struct net_device *dev, u16 queue_id); 108 int xp_alloc_tx_descs(struct xsk_buff_pool *pool, struct xdp_sock *xs); 109 void xp_destroy(struct xsk_buff_pool *pool); 110 void xp_get_pool(struct xsk_buff_pool *pool); 111 bool xp_put_pool(struct xsk_buff_pool *pool); 112 void xp_clear_dev(struct xsk_buff_pool *pool); 113 void xp_add_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs); 114 void xp_del_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs); 115 116 /* AF_XDP, and XDP core. */ 117 void xp_free(struct xdp_buff_xsk *xskb); 118 119 static inline void xp_init_xskb_addr(struct xdp_buff_xsk *xskb, struct xsk_buff_pool *pool, 120 u64 addr) 121 { 122 xskb->orig_addr = addr; 123 xskb->xdp.data_hard_start = pool->addrs + addr + pool->headroom; 124 } 125 126 static inline void xp_init_xskb_dma(struct xdp_buff_xsk *xskb, struct xsk_buff_pool *pool, 127 dma_addr_t *dma_pages, u64 addr) 128 { 129 xskb->frame_dma = (dma_pages[addr >> PAGE_SHIFT] & ~XSK_NEXT_PG_CONTIG_MASK) + 130 (addr & ~PAGE_MASK); 131 xskb->dma = xskb->frame_dma + pool->headroom + XDP_PACKET_HEADROOM; 132 } 133 134 /* AF_XDP ZC drivers, via xdp_sock_buff.h */ 135 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq); 136 void xp_fill_cb(struct xsk_buff_pool *pool, struct xsk_cb_desc *desc); 137 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev, 138 unsigned long attrs, struct page **pages, u32 nr_pages); 139 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs); 140 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool); 141 u32 xp_alloc_batch(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max); 142 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count); 143 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr); 144 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr); 145 static inline dma_addr_t xp_get_dma(struct xdp_buff_xsk *xskb) 146 { 147 return xskb->dma; 148 } 149 150 static inline dma_addr_t xp_get_frame_dma(struct xdp_buff_xsk *xskb) 151 { 152 return xskb->frame_dma; 153 } 154 155 static inline void xp_dma_sync_for_cpu(struct xdp_buff_xsk *xskb) 156 { 157 dma_sync_single_for_cpu(xskb->pool->dev, xskb->dma, 158 xskb->pool->frame_len, 159 DMA_BIDIRECTIONAL); 160 } 161 162 static inline void xp_dma_sync_for_device(struct xsk_buff_pool *pool, 163 dma_addr_t dma, size_t size) 164 { 165 dma_sync_single_for_device(pool->dev, dma, size, DMA_BIDIRECTIONAL); 166 } 167 168 /* Masks for xdp_umem_page flags. 169 * The low 12-bits of the addr will be 0 since this is the page address, so we 170 * can use them for flags. 171 */ 172 #define XSK_NEXT_PG_CONTIG_SHIFT 0 173 #define XSK_NEXT_PG_CONTIG_MASK BIT_ULL(XSK_NEXT_PG_CONTIG_SHIFT) 174 175 static inline bool xp_desc_crosses_non_contig_pg(struct xsk_buff_pool *pool, 176 u64 addr, u32 len) 177 { 178 bool cross_pg = (addr & (PAGE_SIZE - 1)) + len > PAGE_SIZE; 179 180 if (likely(!cross_pg)) 181 return false; 182 183 return pool->dma_pages && 184 !(pool->dma_pages[addr >> PAGE_SHIFT] & XSK_NEXT_PG_CONTIG_MASK); 185 } 186 187 static inline bool xp_mb_desc(struct xdp_desc *desc) 188 { 189 return desc->options & XDP_PKT_CONTD; 190 } 191 192 static inline u64 xp_aligned_extract_addr(struct xsk_buff_pool *pool, u64 addr) 193 { 194 return addr & pool->chunk_mask; 195 } 196 197 static inline u64 xp_unaligned_extract_addr(u64 addr) 198 { 199 return addr & XSK_UNALIGNED_BUF_ADDR_MASK; 200 } 201 202 static inline u64 xp_unaligned_extract_offset(u64 addr) 203 { 204 return addr >> XSK_UNALIGNED_BUF_OFFSET_SHIFT; 205 } 206 207 static inline u64 xp_unaligned_add_offset_to_addr(u64 addr) 208 { 209 return xp_unaligned_extract_addr(addr) + 210 xp_unaligned_extract_offset(addr); 211 } 212 213 static inline u32 xp_aligned_extract_idx(struct xsk_buff_pool *pool, u64 addr) 214 { 215 return xp_aligned_extract_addr(pool, addr) >> pool->chunk_shift; 216 } 217 218 static inline void xp_release(struct xdp_buff_xsk *xskb) 219 { 220 if (xskb->pool->unaligned) 221 xskb->pool->free_heads[xskb->pool->free_heads_cnt++] = xskb; 222 } 223 224 static inline u64 xp_get_handle(struct xdp_buff_xsk *xskb) 225 { 226 u64 offset = xskb->xdp.data - xskb->xdp.data_hard_start; 227 228 offset += xskb->pool->headroom; 229 if (!xskb->pool->unaligned) 230 return xskb->orig_addr + offset; 231 return xskb->orig_addr + (offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT); 232 } 233 234 static inline bool xp_tx_metadata_enabled(const struct xsk_buff_pool *pool) 235 { 236 return pool->tx_metadata_len > 0; 237 } 238 239 #endif /* XSK_BUFF_POOL_H_ */ 240