xref: /linux/include/net/xsk_buff_pool.h (revision 7fc7f25419f5a6b09199ba4b5026b94ef184fa79)
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_queue;
16 struct xdp_desc;
17 struct xdp_umem;
18 struct xdp_sock;
19 struct device;
20 struct page;
21 
22 struct xdp_buff_xsk {
23 	struct xdp_buff xdp;
24 	dma_addr_t dma;
25 	dma_addr_t frame_dma;
26 	struct xsk_buff_pool *pool;
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 	struct xdp_desc *tx_descs;
64 	u64 chunk_mask;
65 	u64 addrs_cnt;
66 	u32 free_list_cnt;
67 	u32 dma_pages_cnt;
68 	u32 free_heads_cnt;
69 	u32 headroom;
70 	u32 chunk_size;
71 	u32 chunk_shift;
72 	u32 frame_len;
73 	u8 cached_need_wakeup;
74 	bool uses_need_wakeup;
75 	bool dma_need_sync;
76 	bool unaligned;
77 	void *addrs;
78 	/* Mutual exclusion of the completion ring in the SKB mode. Two cases to protect:
79 	 * NAPI TX thread and sendmsg error paths in the SKB destructor callback and when
80 	 * sockets share a single cq when the same netdev and queue id is shared.
81 	 */
82 	spinlock_t cq_lock;
83 	struct xdp_buff_xsk *free_heads[];
84 };
85 
86 /* Masks for xdp_umem_page flags.
87  * The low 12-bits of the addr will be 0 since this is the page address, so we
88  * can use them for flags.
89  */
90 #define XSK_NEXT_PG_CONTIG_SHIFT 0
91 #define XSK_NEXT_PG_CONTIG_MASK BIT_ULL(XSK_NEXT_PG_CONTIG_SHIFT)
92 
93 /* AF_XDP core. */
94 struct xsk_buff_pool *xp_create_and_assign_umem(struct xdp_sock *xs,
95 						struct xdp_umem *umem);
96 int xp_assign_dev(struct xsk_buff_pool *pool, struct net_device *dev,
97 		  u16 queue_id, u16 flags);
98 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs,
99 			 struct net_device *dev, u16 queue_id);
100 int xp_alloc_tx_descs(struct xsk_buff_pool *pool, struct xdp_sock *xs);
101 void xp_destroy(struct xsk_buff_pool *pool);
102 void xp_get_pool(struct xsk_buff_pool *pool);
103 bool xp_put_pool(struct xsk_buff_pool *pool);
104 void xp_clear_dev(struct xsk_buff_pool *pool);
105 void xp_add_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs);
106 void xp_del_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs);
107 
108 /* AF_XDP, and XDP core. */
109 void xp_free(struct xdp_buff_xsk *xskb);
110 
111 static inline void xp_init_xskb_addr(struct xdp_buff_xsk *xskb, struct xsk_buff_pool *pool,
112 				     u64 addr)
113 {
114 	xskb->orig_addr = addr;
115 	xskb->xdp.data_hard_start = pool->addrs + addr + pool->headroom;
116 }
117 
118 static inline void xp_init_xskb_dma(struct xdp_buff_xsk *xskb, struct xsk_buff_pool *pool,
119 				    dma_addr_t *dma_pages, u64 addr)
120 {
121 	xskb->frame_dma = (dma_pages[addr >> PAGE_SHIFT] & ~XSK_NEXT_PG_CONTIG_MASK) +
122 		(addr & ~PAGE_MASK);
123 	xskb->dma = xskb->frame_dma + pool->headroom + XDP_PACKET_HEADROOM;
124 }
125 
126 /* AF_XDP ZC drivers, via xdp_sock_buff.h */
127 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq);
128 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
129 	       unsigned long attrs, struct page **pages, u32 nr_pages);
130 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs);
131 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool);
132 u32 xp_alloc_batch(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max);
133 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count);
134 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr);
135 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr);
136 static inline dma_addr_t xp_get_dma(struct xdp_buff_xsk *xskb)
137 {
138 	return xskb->dma;
139 }
140 
141 static inline dma_addr_t xp_get_frame_dma(struct xdp_buff_xsk *xskb)
142 {
143 	return xskb->frame_dma;
144 }
145 
146 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb);
147 static inline void xp_dma_sync_for_cpu(struct xdp_buff_xsk *xskb)
148 {
149 	xp_dma_sync_for_cpu_slow(xskb);
150 }
151 
152 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
153 				 size_t size);
154 static inline void xp_dma_sync_for_device(struct xsk_buff_pool *pool,
155 					  dma_addr_t dma, size_t size)
156 {
157 	if (!pool->dma_need_sync)
158 		return;
159 
160 	xp_dma_sync_for_device_slow(pool, dma, size);
161 }
162 
163 /* Masks for xdp_umem_page flags.
164  * The low 12-bits of the addr will be 0 since this is the page address, so we
165  * can use them for flags.
166  */
167 #define XSK_NEXT_PG_CONTIG_SHIFT 0
168 #define XSK_NEXT_PG_CONTIG_MASK BIT_ULL(XSK_NEXT_PG_CONTIG_SHIFT)
169 
170 static inline bool xp_desc_crosses_non_contig_pg(struct xsk_buff_pool *pool,
171 						 u64 addr, u32 len)
172 {
173 	bool cross_pg = (addr & (PAGE_SIZE - 1)) + len > PAGE_SIZE;
174 
175 	if (likely(!cross_pg))
176 		return false;
177 
178 	if (pool->dma_pages_cnt) {
179 		return !(pool->dma_pages[addr >> PAGE_SHIFT] &
180 			 XSK_NEXT_PG_CONTIG_MASK);
181 	}
182 
183 	/* skb path */
184 	return addr + len > pool->addrs_cnt;
185 }
186 
187 static inline u64 xp_aligned_extract_addr(struct xsk_buff_pool *pool, u64 addr)
188 {
189 	return addr & pool->chunk_mask;
190 }
191 
192 static inline u64 xp_unaligned_extract_addr(u64 addr)
193 {
194 	return addr & XSK_UNALIGNED_BUF_ADDR_MASK;
195 }
196 
197 static inline u64 xp_unaligned_extract_offset(u64 addr)
198 {
199 	return addr >> XSK_UNALIGNED_BUF_OFFSET_SHIFT;
200 }
201 
202 static inline u64 xp_unaligned_add_offset_to_addr(u64 addr)
203 {
204 	return xp_unaligned_extract_addr(addr) +
205 		xp_unaligned_extract_offset(addr);
206 }
207 
208 static inline u32 xp_aligned_extract_idx(struct xsk_buff_pool *pool, u64 addr)
209 {
210 	return xp_aligned_extract_addr(pool, addr) >> pool->chunk_shift;
211 }
212 
213 static inline void xp_release(struct xdp_buff_xsk *xskb)
214 {
215 	if (xskb->pool->unaligned)
216 		xskb->pool->free_heads[xskb->pool->free_heads_cnt++] = xskb;
217 }
218 
219 static inline u64 xp_get_handle(struct xdp_buff_xsk *xskb)
220 {
221 	u64 offset = xskb->xdp.data - xskb->xdp.data_hard_start;
222 
223 	offset += xskb->pool->headroom;
224 	if (!xskb->pool->unaligned)
225 		return xskb->orig_addr + offset;
226 	return xskb->orig_addr + (offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT);
227 }
228 
229 #endif /* XSK_BUFF_POOL_H_ */
230