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 <net/xdp.h>
11
12 struct xsk_buff_pool;
13 struct xdp_rxq_info;
14 struct xsk_queue;
15 struct xdp_desc;
16 struct xdp_umem;
17 struct xdp_sock;
18 struct device;
19 struct page;
20
21 struct xdp_buff_xsk {
22 struct xdp_buff xdp;
23 dma_addr_t dma;
24 dma_addr_t frame_dma;
25 struct xsk_buff_pool *pool;
26 bool unaligned;
27 u64 orig_addr;
28 struct list_head free_list_node;
29 };
30
31 struct xsk_dma_map {
32 dma_addr_t *dma_pages;
33 struct device *dev;
34 struct net_device *netdev;
35 refcount_t users;
36 struct list_head list; /* Protected by the RTNL_LOCK */
37 u32 dma_pages_cnt;
38 bool dma_need_sync;
39 };
40
41 struct xsk_buff_pool {
42 /* Members only used in the control path first. */
43 struct device *dev;
44 struct net_device *netdev;
45 struct list_head xsk_tx_list;
46 /* Protects modifications to the xsk_tx_list */
47 spinlock_t xsk_tx_list_lock;
48 refcount_t users;
49 struct xdp_umem *umem;
50 struct work_struct work;
51 struct list_head free_list;
52 u32 heads_cnt;
53 u16 queue_id;
54
55 /* Data path members as close to free_heads at the end as possible. */
56 struct xsk_queue *fq ____cacheline_aligned_in_smp;
57 struct xsk_queue *cq;
58 /* For performance reasons, each buff pool has its own array of dma_pages
59 * even when they are identical.
60 */
61 dma_addr_t *dma_pages;
62 struct xdp_buff_xsk *heads;
63 u64 chunk_mask;
64 u64 addrs_cnt;
65 u32 free_list_cnt;
66 u32 dma_pages_cnt;
67 u32 free_heads_cnt;
68 u32 headroom;
69 u32 chunk_size;
70 u32 frame_len;
71 u8 cached_need_wakeup;
72 bool uses_need_wakeup;
73 bool dma_need_sync;
74 bool unaligned;
75 void *addrs;
76 /* Mutual exclusion of the completion ring in the SKB mode. Two cases to protect:
77 * NAPI TX thread and sendmsg error paths in the SKB destructor callback and when
78 * sockets share a single cq when the same netdev and queue id is shared.
79 */
80 spinlock_t cq_lock;
81 struct xdp_buff_xsk *free_heads[];
82 };
83
84 /* AF_XDP core. */
85 struct xsk_buff_pool *xp_create_and_assign_umem(struct xdp_sock *xs,
86 struct xdp_umem *umem);
87 int xp_assign_dev(struct xsk_buff_pool *pool, struct net_device *dev,
88 u16 queue_id, u16 flags);
89 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs,
90 struct net_device *dev, u16 queue_id);
91 void xp_destroy(struct xsk_buff_pool *pool);
92 void xp_release(struct xdp_buff_xsk *xskb);
93 void xp_get_pool(struct xsk_buff_pool *pool);
94 bool xp_put_pool(struct xsk_buff_pool *pool);
95 void xp_clear_dev(struct xsk_buff_pool *pool);
96 void xp_add_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs);
97 void xp_del_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs);
98
99 /* AF_XDP, and XDP core. */
100 void xp_free(struct xdp_buff_xsk *xskb);
101
102 /* AF_XDP ZC drivers, via xdp_sock_buff.h */
103 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq);
104 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
105 unsigned long attrs, struct page **pages, u32 nr_pages);
106 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs);
107 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool);
108 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count);
109 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr);
110 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr);
xp_get_dma(struct xdp_buff_xsk * xskb)111 static inline dma_addr_t xp_get_dma(struct xdp_buff_xsk *xskb)
112 {
113 return xskb->dma;
114 }
115
xp_get_frame_dma(struct xdp_buff_xsk * xskb)116 static inline dma_addr_t xp_get_frame_dma(struct xdp_buff_xsk *xskb)
117 {
118 return xskb->frame_dma;
119 }
120
121 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb);
xp_dma_sync_for_cpu(struct xdp_buff_xsk * xskb)122 static inline void xp_dma_sync_for_cpu(struct xdp_buff_xsk *xskb)
123 {
124 xp_dma_sync_for_cpu_slow(xskb);
125 }
126
127 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
128 size_t size);
xp_dma_sync_for_device(struct xsk_buff_pool * pool,dma_addr_t dma,size_t size)129 static inline void xp_dma_sync_for_device(struct xsk_buff_pool *pool,
130 dma_addr_t dma, size_t size)
131 {
132 if (!pool->dma_need_sync)
133 return;
134
135 xp_dma_sync_for_device_slow(pool, dma, size);
136 }
137
138 /* Masks for xdp_umem_page flags.
139 * The low 12-bits of the addr will be 0 since this is the page address, so we
140 * can use them for flags.
141 */
142 #define XSK_NEXT_PG_CONTIG_SHIFT 0
143 #define XSK_NEXT_PG_CONTIG_MASK BIT_ULL(XSK_NEXT_PG_CONTIG_SHIFT)
144
xp_desc_crosses_non_contig_pg(struct xsk_buff_pool * pool,u64 addr,u32 len)145 static inline bool xp_desc_crosses_non_contig_pg(struct xsk_buff_pool *pool,
146 u64 addr, u32 len)
147 {
148 bool cross_pg = (addr & (PAGE_SIZE - 1)) + len > PAGE_SIZE;
149
150 if (likely(!cross_pg))
151 return false;
152
153 return pool->dma_pages_cnt &&
154 !(pool->dma_pages[addr >> PAGE_SHIFT] & XSK_NEXT_PG_CONTIG_MASK);
155 }
156
xp_aligned_extract_addr(struct xsk_buff_pool * pool,u64 addr)157 static inline u64 xp_aligned_extract_addr(struct xsk_buff_pool *pool, u64 addr)
158 {
159 return addr & pool->chunk_mask;
160 }
161
xp_unaligned_extract_addr(u64 addr)162 static inline u64 xp_unaligned_extract_addr(u64 addr)
163 {
164 return addr & XSK_UNALIGNED_BUF_ADDR_MASK;
165 }
166
xp_unaligned_extract_offset(u64 addr)167 static inline u64 xp_unaligned_extract_offset(u64 addr)
168 {
169 return addr >> XSK_UNALIGNED_BUF_OFFSET_SHIFT;
170 }
171
xp_unaligned_add_offset_to_addr(u64 addr)172 static inline u64 xp_unaligned_add_offset_to_addr(u64 addr)
173 {
174 return xp_unaligned_extract_addr(addr) +
175 xp_unaligned_extract_offset(addr);
176 }
177
178 #endif /* XSK_BUFF_POOL_H_ */
179