xref: /f-stack/dpdk/drivers/net/enic/base/vnic_wq.h (revision 2d9fd380)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2008-2017 Cisco Systems, Inc.  All rights reserved.
3  * Copyright 2007 Nuova Systems, Inc.  All rights reserved.
4  */
5 
6 #ifndef _VNIC_WQ_H_
7 #define _VNIC_WQ_H_
8 
9 
10 #include "vnic_dev.h"
11 #include "vnic_cq.h"
12 #include <rte_memzone.h>
13 
14 /* Work queue control */
15 struct vnic_wq_ctrl {
16 	uint64_t ring_base;			/* 0x00 */
17 	uint32_t ring_size;			/* 0x08 */
18 	uint32_t pad0;
19 	uint32_t posted_index;			/* 0x10 */
20 	uint32_t pad1;
21 	uint32_t cq_index;			/* 0x18 */
22 	uint32_t pad2;
23 	uint32_t enable;			/* 0x20 */
24 	uint32_t pad3;
25 	uint32_t running;			/* 0x28 */
26 	uint32_t pad4;
27 	uint32_t fetch_index;			/* 0x30 */
28 	uint32_t pad5;
29 	uint32_t dca_value;			/* 0x38 */
30 	uint32_t pad6;
31 	uint32_t error_interrupt_enable;	/* 0x40 */
32 	uint32_t pad7;
33 	uint32_t error_interrupt_offset;	/* 0x48 */
34 	uint32_t pad8;
35 	uint32_t error_status;			/* 0x50 */
36 	uint32_t pad9;
37 };
38 
39 struct vnic_wq {
40 	unsigned int index;
41 	uint64_t tx_offload_notsup_mask;
42 	struct vnic_dev *vdev;
43 	struct vnic_wq_ctrl __iomem *ctrl;              /* memory-mapped */
44 	struct vnic_dev_ring ring;
45 	struct rte_mbuf **bufs;
46 	unsigned int head_idx;
47 	unsigned int cq_pend;
48 	unsigned int tail_idx;
49 	unsigned int socket_id;
50 	const struct rte_memzone *cqmsg_rz;
51 	uint16_t last_completed_index;
52 	uint64_t offloads;
53 };
54 
vnic_wq_desc_avail(struct vnic_wq * wq)55 static inline unsigned int vnic_wq_desc_avail(struct vnic_wq *wq)
56 {
57 	/* how many does SW own? */
58 	return wq->ring.desc_avail;
59 }
60 
vnic_wq_desc_used(struct vnic_wq * wq)61 static inline unsigned int vnic_wq_desc_used(struct vnic_wq *wq)
62 {
63 	/* how many does HW own? */
64 	return wq->ring.desc_count - wq->ring.desc_avail - 1;
65 }
66 
67 #define PI_LOG2_CACHE_LINE_SIZE        5
68 #define PI_INDEX_BITS            12
69 #define PI_INDEX_MASK ((1U << PI_INDEX_BITS) - 1)
70 #define PI_PREFETCH_LEN_MASK ((1U << PI_LOG2_CACHE_LINE_SIZE) - 1)
71 #define PI_PREFETCH_LEN_OFF 16
72 #define PI_PREFETCH_ADDR_BITS 43
73 #define PI_PREFETCH_ADDR_MASK ((1ULL << PI_PREFETCH_ADDR_BITS) - 1)
74 #define PI_PREFETCH_ADDR_OFF 21
75 
76 /** How many cache lines are touched by buffer (addr, len). */
num_cache_lines_touched(dma_addr_t addr,unsigned int len)77 static inline unsigned int num_cache_lines_touched(dma_addr_t addr,
78 							unsigned int len)
79 {
80 	const unsigned long mask = PI_PREFETCH_LEN_MASK;
81 	const unsigned long laddr = (unsigned long)addr;
82 	unsigned long lines, equiv_len;
83 	/* A. If addr is aligned, our solution is just to round up len to the
84 	next boundary.
85 
86 	e.g. addr = 0, len = 48
87 	+--------------------+
88 	|XXXXXXXXXXXXXXXXXXXX|    32-byte cacheline a
89 	+--------------------+
90 	|XXXXXXXXXX          |    cacheline b
91 	+--------------------+
92 
93 	B. If addr is not aligned, however, we may use an extra
94 	cacheline.  e.g. addr = 12, len = 22
95 
96 	+--------------------+
97 	|       XXXXXXXXXXXXX|
98 	+--------------------+
99 	|XX                  |
100 	+--------------------+
101 
102 	Our solution is to make the problem equivalent to case A
103 	above by adding the empty space in the first cacheline to the length:
104 	unsigned long len;
105 
106 	+--------------------+
107 	|eeeeeeeXXXXXXXXXXXXX|    "e" is empty space, which we add to len
108 	+--------------------+
109 	|XX                  |
110 	+--------------------+
111 
112 	*/
113 	equiv_len = len + (laddr & mask);
114 
115 	/* Now we can just round up this len to the next 32-byte boundary. */
116 	lines = (equiv_len + mask) & (~mask);
117 
118 	/* Scale bytes -> cachelines. */
119 	return lines >> PI_LOG2_CACHE_LINE_SIZE;
120 }
121 
vnic_cached_posted_index(dma_addr_t addr,unsigned int len,unsigned int index)122 static inline uint64_t vnic_cached_posted_index(dma_addr_t addr,
123 						unsigned int len,
124 						unsigned int index)
125 {
126 	unsigned int num_cache_lines = num_cache_lines_touched(addr, len);
127 	/* Wish we could avoid a branch here.  We could have separate
128 	 * vnic_wq_post() and vinc_wq_post_inline(), the latter
129 	 * only supporting < 1k (2^5 * 2^5) sends, I suppose.  This would
130 	 * eliminate the if (eop) branch as well.
131 	 */
132 	if (num_cache_lines > PI_PREFETCH_LEN_MASK)
133 		num_cache_lines = 0;
134 	return (index & PI_INDEX_MASK) |
135 	((num_cache_lines & PI_PREFETCH_LEN_MASK) << PI_PREFETCH_LEN_OFF) |
136 		(((addr >> PI_LOG2_CACHE_LINE_SIZE) &
137 	PI_PREFETCH_ADDR_MASK) << PI_PREFETCH_ADDR_OFF);
138 }
139 
140 static inline uint32_t
buf_idx_incr(uint32_t n_descriptors,uint32_t idx)141 buf_idx_incr(uint32_t n_descriptors, uint32_t idx)
142 {
143 	idx++;
144 	if (unlikely(idx == n_descriptors))
145 		idx = 0;
146 	return idx;
147 }
148 
149 void vnic_wq_free(struct vnic_wq *wq);
150 int vnic_wq_alloc(struct vnic_dev *vdev, struct vnic_wq *wq, unsigned int index,
151 	unsigned int desc_count, unsigned int desc_size);
152 void vnic_wq_init_start(struct vnic_wq *wq, unsigned int cq_index,
153 	unsigned int fetch_index, unsigned int posted_index,
154 	unsigned int error_interrupt_enable,
155 	unsigned int error_interrupt_offset);
156 void vnic_wq_init(struct vnic_wq *wq, unsigned int cq_index,
157 	unsigned int error_interrupt_enable,
158 	unsigned int error_interrupt_offset);
159 void vnic_wq_error_out(struct vnic_wq *wq, unsigned int error);
160 unsigned int vnic_wq_error_status(struct vnic_wq *wq);
161 void vnic_wq_enable(struct vnic_wq *wq);
162 int vnic_wq_disable(struct vnic_wq *wq);
163 void vnic_wq_clean(struct vnic_wq *wq,
164 		   void (*buf_clean)(struct rte_mbuf **buf));
165 #endif /* _VNIC_WQ_H_ */
166