xref: /dpdk/drivers/net/enic/base/vnic_wq.h (revision 00ce43111dc5b364722c882cdd37d3664d87b6cc)
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 	bool admin_chan;
54 	const struct rte_memzone *admin_msg_rz;
55 	uint64_t soft_stats_tx;
56 };
57 
58 static inline unsigned int vnic_wq_desc_avail(struct vnic_wq *wq)
59 {
60 	/* how many does SW own? */
61 	return wq->ring.desc_avail;
62 }
63 
64 static inline unsigned int vnic_wq_desc_used(struct vnic_wq *wq)
65 {
66 	/* how many does HW own? */
67 	return wq->ring.desc_count - wq->ring.desc_avail - 1;
68 }
69 
70 #define PI_LOG2_CACHE_LINE_SIZE        5
71 #define PI_INDEX_BITS            12
72 #define PI_INDEX_MASK ((1U << PI_INDEX_BITS) - 1)
73 #define PI_PREFETCH_LEN_MASK ((1U << PI_LOG2_CACHE_LINE_SIZE) - 1)
74 #define PI_PREFETCH_LEN_OFF 16
75 #define PI_PREFETCH_ADDR_BITS 43
76 #define PI_PREFETCH_ADDR_MASK ((1ULL << PI_PREFETCH_ADDR_BITS) - 1)
77 #define PI_PREFETCH_ADDR_OFF 21
78 
79 /** How many cache lines are touched by buffer (addr, len). */
80 static inline unsigned int num_cache_lines_touched(dma_addr_t addr,
81 							unsigned int len)
82 {
83 	const unsigned long mask = PI_PREFETCH_LEN_MASK;
84 	const unsigned long laddr = (unsigned long)addr;
85 	unsigned long lines, equiv_len;
86 	/* A. If addr is aligned, our solution is just to round up len to the
87 	next boundary.
88 
89 	e.g. addr = 0, len = 48
90 	+--------------------+
91 	|XXXXXXXXXXXXXXXXXXXX|    32-byte cacheline a
92 	+--------------------+
93 	|XXXXXXXXXX          |    cacheline b
94 	+--------------------+
95 
96 	B. If addr is not aligned, however, we may use an extra
97 	cacheline.  e.g. addr = 12, len = 22
98 
99 	+--------------------+
100 	|       XXXXXXXXXXXXX|
101 	+--------------------+
102 	|XX                  |
103 	+--------------------+
104 
105 	Our solution is to make the problem equivalent to case A
106 	above by adding the empty space in the first cacheline to the length:
107 	unsigned long len;
108 
109 	+--------------------+
110 	|eeeeeeeXXXXXXXXXXXXX|    "e" is empty space, which we add to len
111 	+--------------------+
112 	|XX                  |
113 	+--------------------+
114 
115 	*/
116 	equiv_len = len + (laddr & mask);
117 
118 	/* Now we can just round up this len to the next 32-byte boundary. */
119 	lines = (equiv_len + mask) & (~mask);
120 
121 	/* Scale bytes -> cachelines. */
122 	return lines >> PI_LOG2_CACHE_LINE_SIZE;
123 }
124 
125 static inline uint64_t vnic_cached_posted_index(dma_addr_t addr,
126 						unsigned int len,
127 						unsigned int index)
128 {
129 	unsigned int num_cache_lines = num_cache_lines_touched(addr, len);
130 	/* Wish we could avoid a branch here.  We could have separate
131 	 * vnic_wq_post() and vinc_wq_post_inline(), the latter
132 	 * only supporting < 1k (2^5 * 2^5) sends, I suppose.  This would
133 	 * eliminate the if (eop) branch as well.
134 	 */
135 	if (num_cache_lines > PI_PREFETCH_LEN_MASK)
136 		num_cache_lines = 0;
137 	return (index & PI_INDEX_MASK) |
138 	((num_cache_lines & PI_PREFETCH_LEN_MASK) << PI_PREFETCH_LEN_OFF) |
139 		(((addr >> PI_LOG2_CACHE_LINE_SIZE) &
140 	PI_PREFETCH_ADDR_MASK) << PI_PREFETCH_ADDR_OFF);
141 }
142 
143 static inline uint32_t
144 buf_idx_incr(uint32_t n_descriptors, uint32_t idx)
145 {
146 	idx++;
147 	if (unlikely(idx == n_descriptors))
148 		idx = 0;
149 	return idx;
150 }
151 
152 void vnic_wq_free(struct vnic_wq *wq);
153 int vnic_wq_alloc(struct vnic_dev *vdev, struct vnic_wq *wq, unsigned int index,
154 	unsigned int desc_count, unsigned int desc_size);
155 int vnic_admin_wq_alloc(struct vnic_dev *vdev, struct vnic_wq *wq,
156 	unsigned int desc_count, unsigned int desc_size);
157 void vnic_wq_init_start(struct vnic_wq *wq, unsigned int cq_index,
158 	unsigned int fetch_index, unsigned int posted_index,
159 	unsigned int error_interrupt_enable,
160 	unsigned int error_interrupt_offset);
161 void vnic_wq_init(struct vnic_wq *wq, unsigned int cq_index,
162 	unsigned int error_interrupt_enable,
163 	unsigned int error_interrupt_offset);
164 void vnic_wq_error_out(struct vnic_wq *wq, unsigned int error);
165 unsigned int vnic_wq_error_status(struct vnic_wq *wq);
166 void vnic_wq_enable(struct vnic_wq *wq);
167 int vnic_wq_disable(struct vnic_wq *wq);
168 void vnic_wq_clean(struct vnic_wq *wq,
169 		   void (*buf_clean)(struct rte_mbuf **buf));
170 #endif /* _VNIC_WQ_H_ */
171