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
13struct xsk_buff_pool;
14struct xdp_rxq_info;
15struct xsk_queue;
16struct xdp_desc;
17struct xdp_umem;
18struct xdp_sock;
19struct device;
20struct page;
21
22#define XSK_PRIV_MAX 24
23
24struct xdp_buff_xsk {
25	struct xdp_buff xdp;
26	u8 cb[XSK_PRIV_MAX];
27	dma_addr_t dma;
28	dma_addr_t frame_dma;
29	struct xsk_buff_pool *pool;
30	u64 orig_addr;
31	struct list_head free_list_node;
32	struct list_head xskb_list_node;
33};
34
35#define XSK_CHECK_PRIV_TYPE(t) BUILD_BUG_ON(sizeof(t) > offsetofend(struct xdp_buff_xsk, cb))
36
37struct xsk_dma_map {
38	dma_addr_t *dma_pages;
39	struct device *dev;
40	struct net_device *netdev;
41	refcount_t users;
42	struct list_head list; /* Protected by the RTNL_LOCK */
43	u32 dma_pages_cnt;
44	bool dma_need_sync;
45};
46
47struct 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 cached_need_wakeup;
81	bool uses_need_wakeup;
82	bool dma_need_sync;
83	bool unaligned;
84	void *addrs;
85	/* Mutual exclusion of the completion ring in the SKB mode. Two cases to protect:
86	 * NAPI TX thread and sendmsg error paths in the SKB destructor callback and when
87	 * sockets share a single cq when the same netdev and queue id is shared.
88	 */
89	spinlock_t cq_lock;
90	struct xdp_buff_xsk *free_heads[];
91};
92
93/* Masks for xdp_umem_page flags.
94 * The low 12-bits of the addr will be 0 since this is the page address, so we
95 * can use them for flags.
96 */
97#define XSK_NEXT_PG_CONTIG_SHIFT 0
98#define XSK_NEXT_PG_CONTIG_MASK BIT_ULL(XSK_NEXT_PG_CONTIG_SHIFT)
99
100/* AF_XDP core. */
101struct xsk_buff_pool *xp_create_and_assign_umem(struct xdp_sock *xs,
102						struct xdp_umem *umem);
103int xp_assign_dev(struct xsk_buff_pool *pool, struct net_device *dev,
104		  u16 queue_id, u16 flags);
105int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs,
106			 struct net_device *dev, u16 queue_id);
107int xp_alloc_tx_descs(struct xsk_buff_pool *pool, struct xdp_sock *xs);
108void xp_destroy(struct xsk_buff_pool *pool);
109void xp_get_pool(struct xsk_buff_pool *pool);
110bool xp_put_pool(struct xsk_buff_pool *pool);
111void xp_clear_dev(struct xsk_buff_pool *pool);
112void xp_add_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs);
113void xp_del_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs);
114
115/* AF_XDP, and XDP core. */
116void xp_free(struct xdp_buff_xsk *xskb);
117
118static inline void xp_init_xskb_addr(struct xdp_buff_xsk *xskb, struct xsk_buff_pool *pool,
119				     u64 addr)
120{
121	xskb->orig_addr = addr;
122	xskb->xdp.data_hard_start = pool->addrs + addr + pool->headroom;
123}
124
125static inline void xp_init_xskb_dma(struct xdp_buff_xsk *xskb, struct xsk_buff_pool *pool,
126				    dma_addr_t *dma_pages, u64 addr)
127{
128	xskb->frame_dma = (dma_pages[addr >> PAGE_SHIFT] & ~XSK_NEXT_PG_CONTIG_MASK) +
129		(addr & ~PAGE_MASK);
130	xskb->dma = xskb->frame_dma + pool->headroom + XDP_PACKET_HEADROOM;
131}
132
133/* AF_XDP ZC drivers, via xdp_sock_buff.h */
134void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq);
135int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
136	       unsigned long attrs, struct page **pages, u32 nr_pages);
137void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs);
138struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool);
139u32 xp_alloc_batch(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max);
140bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count);
141void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr);
142dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr);
143static inline dma_addr_t xp_get_dma(struct xdp_buff_xsk *xskb)
144{
145	return xskb->dma;
146}
147
148static inline dma_addr_t xp_get_frame_dma(struct xdp_buff_xsk *xskb)
149{
150	return xskb->frame_dma;
151}
152
153void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb);
154static inline void xp_dma_sync_for_cpu(struct xdp_buff_xsk *xskb)
155{
156	xp_dma_sync_for_cpu_slow(xskb);
157}
158
159void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
160				 size_t size);
161static inline void xp_dma_sync_for_device(struct xsk_buff_pool *pool,
162					  dma_addr_t dma, size_t size)
163{
164	if (!pool->dma_need_sync)
165		return;
166
167	xp_dma_sync_for_device_slow(pool, dma, size);
168}
169
170/* Masks for xdp_umem_page flags.
171 * The low 12-bits of the addr will be 0 since this is the page address, so we
172 * can use them for flags.
173 */
174#define XSK_NEXT_PG_CONTIG_SHIFT 0
175#define XSK_NEXT_PG_CONTIG_MASK BIT_ULL(XSK_NEXT_PG_CONTIG_SHIFT)
176
177static inline bool xp_desc_crosses_non_contig_pg(struct xsk_buff_pool *pool,
178						 u64 addr, u32 len)
179{
180	bool cross_pg = (addr & (PAGE_SIZE - 1)) + len > PAGE_SIZE;
181
182	if (likely(!cross_pg))
183		return false;
184
185	return pool->dma_pages &&
186	       !(pool->dma_pages[addr >> PAGE_SHIFT] & XSK_NEXT_PG_CONTIG_MASK);
187}
188
189static inline bool xp_mb_desc(struct xdp_desc *desc)
190{
191	return desc->options & XDP_PKT_CONTD;
192}
193
194static inline u64 xp_aligned_extract_addr(struct xsk_buff_pool *pool, u64 addr)
195{
196	return addr & pool->chunk_mask;
197}
198
199static inline u64 xp_unaligned_extract_addr(u64 addr)
200{
201	return addr & XSK_UNALIGNED_BUF_ADDR_MASK;
202}
203
204static inline u64 xp_unaligned_extract_offset(u64 addr)
205{
206	return addr >> XSK_UNALIGNED_BUF_OFFSET_SHIFT;
207}
208
209static inline u64 xp_unaligned_add_offset_to_addr(u64 addr)
210{
211	return xp_unaligned_extract_addr(addr) +
212		xp_unaligned_extract_offset(addr);
213}
214
215static inline u32 xp_aligned_extract_idx(struct xsk_buff_pool *pool, u64 addr)
216{
217	return xp_aligned_extract_addr(pool, addr) >> pool->chunk_shift;
218}
219
220static inline void xp_release(struct xdp_buff_xsk *xskb)
221{
222	if (xskb->pool->unaligned)
223		xskb->pool->free_heads[xskb->pool->free_heads_cnt++] = xskb;
224}
225
226static inline u64 xp_get_handle(struct xdp_buff_xsk *xskb)
227{
228	u64 offset = xskb->xdp.data - xskb->xdp.data_hard_start;
229
230	offset += xskb->pool->headroom;
231	if (!xskb->pool->unaligned)
232		return xskb->orig_addr + offset;
233	return xskb->orig_addr + (offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT);
234}
235
236#endif /* XSK_BUFF_POOL_H_ */
237