xref: /linux/include/net/xsk_buff_pool.h (revision a23e1966932464e1c5226cb9ac4ce1d5fc10ba22)
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 free_list_node;
33 	struct list_head xskb_list_node;
34 };
35 
36 #define XSK_CHECK_PRIV_TYPE(t) BUILD_BUG_ON(sizeof(t) > offsetofend(struct xdp_buff_xsk, cb))
37 #define XSK_TX_COMPL_FITS(t) BUILD_BUG_ON(sizeof(struct xsk_tx_metadata_compl) > sizeof(t))
38 
39 struct xsk_dma_map {
40 	dma_addr_t *dma_pages;
41 	struct device *dev;
42 	struct net_device *netdev;
43 	refcount_t users;
44 	struct list_head list; /* Protected by the RTNL_LOCK */
45 	u32 dma_pages_cnt;
46 	bool dma_need_sync;
47 };
48 
49 struct xsk_buff_pool {
50 	/* Members only used in the control path first. */
51 	struct device *dev;
52 	struct net_device *netdev;
53 	struct list_head xsk_tx_list;
54 	/* Protects modifications to the xsk_tx_list */
55 	spinlock_t xsk_tx_list_lock;
56 	refcount_t users;
57 	struct xdp_umem *umem;
58 	struct work_struct work;
59 	struct list_head free_list;
60 	struct list_head xskb_list;
61 	u32 heads_cnt;
62 	u16 queue_id;
63 
64 	/* Data path members as close to free_heads at the end as possible. */
65 	struct xsk_queue *fq ____cacheline_aligned_in_smp;
66 	struct xsk_queue *cq;
67 	/* For performance reasons, each buff pool has its own array of dma_pages
68 	 * even when they are identical.
69 	 */
70 	dma_addr_t *dma_pages;
71 	struct xdp_buff_xsk *heads;
72 	struct xdp_desc *tx_descs;
73 	u64 chunk_mask;
74 	u64 addrs_cnt;
75 	u32 free_list_cnt;
76 	u32 dma_pages_cnt;
77 	u32 free_heads_cnt;
78 	u32 headroom;
79 	u32 chunk_size;
80 	u32 chunk_shift;
81 	u32 frame_len;
82 	u8 tx_metadata_len; /* inherited from umem */
83 	u8 cached_need_wakeup;
84 	bool uses_need_wakeup;
85 	bool dma_need_sync;
86 	bool unaligned;
87 	bool tx_sw_csum;
88 	void *addrs;
89 	/* Mutual exclusion of the completion ring in the SKB mode. Two cases to protect:
90 	 * NAPI TX thread and sendmsg error paths in the SKB destructor callback and when
91 	 * sockets share a single cq when the same netdev and queue id is shared.
92 	 */
93 	spinlock_t cq_lock;
94 	struct xdp_buff_xsk *free_heads[];
95 };
96 
97 /* Masks for xdp_umem_page flags.
98  * The low 12-bits of the addr will be 0 since this is the page address, so we
99  * can use them for flags.
100  */
101 #define XSK_NEXT_PG_CONTIG_SHIFT 0
102 #define XSK_NEXT_PG_CONTIG_MASK BIT_ULL(XSK_NEXT_PG_CONTIG_SHIFT)
103 
104 /* AF_XDP core. */
105 struct xsk_buff_pool *xp_create_and_assign_umem(struct xdp_sock *xs,
106 						struct xdp_umem *umem);
107 int xp_assign_dev(struct xsk_buff_pool *pool, struct net_device *dev,
108 		  u16 queue_id, u16 flags);
109 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs,
110 			 struct net_device *dev, u16 queue_id);
111 int xp_alloc_tx_descs(struct xsk_buff_pool *pool, struct xdp_sock *xs);
112 void xp_destroy(struct xsk_buff_pool *pool);
113 void xp_get_pool(struct xsk_buff_pool *pool);
114 bool xp_put_pool(struct xsk_buff_pool *pool);
115 void xp_clear_dev(struct xsk_buff_pool *pool);
116 void xp_add_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs);
117 void xp_del_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs);
118 
119 /* AF_XDP, and XDP core. */
120 void xp_free(struct xdp_buff_xsk *xskb);
121 
122 static inline void xp_init_xskb_addr(struct xdp_buff_xsk *xskb, struct xsk_buff_pool *pool,
123 				     u64 addr)
124 {
125 	xskb->orig_addr = addr;
126 	xskb->xdp.data_hard_start = pool->addrs + addr + pool->headroom;
127 }
128 
129 static inline void xp_init_xskb_dma(struct xdp_buff_xsk *xskb, struct xsk_buff_pool *pool,
130 				    dma_addr_t *dma_pages, u64 addr)
131 {
132 	xskb->frame_dma = (dma_pages[addr >> PAGE_SHIFT] & ~XSK_NEXT_PG_CONTIG_MASK) +
133 		(addr & ~PAGE_MASK);
134 	xskb->dma = xskb->frame_dma + pool->headroom + XDP_PACKET_HEADROOM;
135 }
136 
137 /* AF_XDP ZC drivers, via xdp_sock_buff.h */
138 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq);
139 void xp_fill_cb(struct xsk_buff_pool *pool, struct xsk_cb_desc *desc);
140 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
141 	       unsigned long attrs, struct page **pages, u32 nr_pages);
142 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs);
143 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool);
144 u32 xp_alloc_batch(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max);
145 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count);
146 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr);
147 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr);
148 static inline dma_addr_t xp_get_dma(struct xdp_buff_xsk *xskb)
149 {
150 	return xskb->dma;
151 }
152 
153 static inline dma_addr_t xp_get_frame_dma(struct xdp_buff_xsk *xskb)
154 {
155 	return xskb->frame_dma;
156 }
157 
158 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb);
159 static inline void xp_dma_sync_for_cpu(struct xdp_buff_xsk *xskb)
160 {
161 	xp_dma_sync_for_cpu_slow(xskb);
162 }
163 
164 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
165 				 size_t size);
166 static inline void xp_dma_sync_for_device(struct xsk_buff_pool *pool,
167 					  dma_addr_t dma, size_t size)
168 {
169 	if (!pool->dma_need_sync)
170 		return;
171 
172 	xp_dma_sync_for_device_slow(pool, dma, size);
173 }
174 
175 /* Masks for xdp_umem_page flags.
176  * The low 12-bits of the addr will be 0 since this is the page address, so we
177  * can use them for flags.
178  */
179 #define XSK_NEXT_PG_CONTIG_SHIFT 0
180 #define XSK_NEXT_PG_CONTIG_MASK BIT_ULL(XSK_NEXT_PG_CONTIG_SHIFT)
181 
182 static inline bool xp_desc_crosses_non_contig_pg(struct xsk_buff_pool *pool,
183 						 u64 addr, u32 len)
184 {
185 	bool cross_pg = (addr & (PAGE_SIZE - 1)) + len > PAGE_SIZE;
186 
187 	if (likely(!cross_pg))
188 		return false;
189 
190 	return pool->dma_pages &&
191 	       !(pool->dma_pages[addr >> PAGE_SHIFT] & XSK_NEXT_PG_CONTIG_MASK);
192 }
193 
194 static inline bool xp_mb_desc(struct xdp_desc *desc)
195 {
196 	return desc->options & XDP_PKT_CONTD;
197 }
198 
199 static inline u64 xp_aligned_extract_addr(struct xsk_buff_pool *pool, u64 addr)
200 {
201 	return addr & pool->chunk_mask;
202 }
203 
204 static inline u64 xp_unaligned_extract_addr(u64 addr)
205 {
206 	return addr & XSK_UNALIGNED_BUF_ADDR_MASK;
207 }
208 
209 static inline u64 xp_unaligned_extract_offset(u64 addr)
210 {
211 	return addr >> XSK_UNALIGNED_BUF_OFFSET_SHIFT;
212 }
213 
214 static inline u64 xp_unaligned_add_offset_to_addr(u64 addr)
215 {
216 	return xp_unaligned_extract_addr(addr) +
217 		xp_unaligned_extract_offset(addr);
218 }
219 
220 static inline u32 xp_aligned_extract_idx(struct xsk_buff_pool *pool, u64 addr)
221 {
222 	return xp_aligned_extract_addr(pool, addr) >> pool->chunk_shift;
223 }
224 
225 static inline void xp_release(struct xdp_buff_xsk *xskb)
226 {
227 	if (xskb->pool->unaligned)
228 		xskb->pool->free_heads[xskb->pool->free_heads_cnt++] = xskb;
229 }
230 
231 static inline u64 xp_get_handle(struct xdp_buff_xsk *xskb)
232 {
233 	u64 offset = xskb->xdp.data - xskb->xdp.data_hard_start;
234 
235 	offset += xskb->pool->headroom;
236 	if (!xskb->pool->unaligned)
237 		return xskb->orig_addr + offset;
238 	return xskb->orig_addr + (offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT);
239 }
240 
241 static inline bool xp_tx_metadata_enabled(const struct xsk_buff_pool *pool)
242 {
243 	return pool->tx_metadata_len > 0;
244 }
245 
246 #endif /* XSK_BUFF_POOL_H_ */
247