1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2017 Intel Corporation
3  */
4 
5 #ifndef _DIST_PRIV_H_
6 #define _DIST_PRIV_H_
7 
8 /**
9  * @file
10  * RTE distributor
11  *
12  * The distributor is a component which is designed to pass packets
13  * one-at-a-time to workers, with dynamic load balancing.
14  */
15 
16 #define NO_FLAGS 0
17 #define RTE_DISTRIB_PREFIX "DT_"
18 
19 /*
20  * We will use the bottom four bits of pointer for flags, shifting out
21  * the top four bits to make room (since a 64-bit pointer actually only uses
22  * 48 bits). An arithmetic-right-shift will then appropriately restore the
23  * original pointer value with proper sign extension into the top bits.
24  */
25 #define RTE_DISTRIB_FLAG_BITS 4
26 #define RTE_DISTRIB_FLAGS_MASK (0x0F)
27 #define RTE_DISTRIB_NO_BUF 0       /**< empty flags: no buffer requested */
28 #define RTE_DISTRIB_GET_BUF (1)    /**< worker requests a buffer, returns old */
29 #define RTE_DISTRIB_RETURN_BUF (2) /**< worker returns a buffer, no request */
30 #define RTE_DISTRIB_VALID_BUF (4)  /**< set if bufptr contains ptr */
31 
32 #define RTE_DISTRIB_BACKLOG_SIZE 8
33 #define RTE_DISTRIB_BACKLOG_MASK (RTE_DISTRIB_BACKLOG_SIZE - 1)
34 
35 #define RTE_DISTRIB_MAX_RETURNS 128
36 #define RTE_DISTRIB_RETURNS_MASK (RTE_DISTRIB_MAX_RETURNS - 1)
37 
38 /**
39  * Maximum number of workers allowed.
40  * Be aware of increasing the limit, because it is limited by how we track
41  * in-flight tags. See in_flight_bitmask and rte_distributor_process
42  */
43 #define RTE_DISTRIB_MAX_WORKERS 64
44 
45 #define RTE_DISTRIBUTOR_NAMESIZE 32 /**< Length of name for instance */
46 
47 /**
48  * Buffer structure used to pass the pointer data between cores. This is cache
49  * line aligned, but to improve performance and prevent adjacent cache-line
50  * prefetches of buffers for other workers, e.g. when worker 1's buffer is on
51  * the next cache line to worker 0, we pad this out to three cache lines.
52  * Only 64-bits of the memory is actually used though.
53  */
54 union rte_distributor_buffer_single {
55 	volatile int64_t bufptr64;
56 	char pad[RTE_CACHE_LINE_SIZE*3];
57 } __rte_cache_aligned;
58 
59 /*
60  * Transfer up to 8 mbufs at a time to/from workers, and
61  * flow matching algorithm optimized for 8 flow IDs at a time
62  */
63 #define RTE_DIST_BURST_SIZE 8
64 
65 struct rte_distributor_backlog {
66 	unsigned int start;
67 	unsigned int count;
68 	int64_t pkts[RTE_DIST_BURST_SIZE] __rte_cache_aligned;
69 	uint16_t *tags; /* will point to second cacheline of inflights */
70 } __rte_cache_aligned;
71 
72 
73 struct rte_distributor_returned_pkts {
74 	unsigned int start;
75 	unsigned int count;
76 	struct rte_mbuf *mbufs[RTE_DISTRIB_MAX_RETURNS];
77 };
78 
79 struct rte_distributor_single {
80 	TAILQ_ENTRY(rte_distributor_single) next;    /**< Next in list. */
81 
82 	char name[RTE_DISTRIBUTOR_NAMESIZE];  /**< Name of the ring. */
83 	unsigned int num_workers;             /**< Number of workers polling */
84 
85 	uint32_t in_flight_tags[RTE_DISTRIB_MAX_WORKERS];
86 		/**< Tracks the tag being processed per core */
87 	uint64_t in_flight_bitmask;
88 		/**< on/off bits for in-flight tags.
89 		 * Note that if RTE_DISTRIB_MAX_WORKERS is larger than 64 then
90 		 * the bitmask has to expand.
91 		 */
92 
93 	struct rte_distributor_backlog backlog[RTE_DISTRIB_MAX_WORKERS];
94 
95 	union rte_distributor_buffer_single bufs[RTE_DISTRIB_MAX_WORKERS];
96 
97 	struct rte_distributor_returned_pkts returns;
98 };
99 
100 /* All different signature compare functions */
101 enum rte_distributor_match_function {
102 	RTE_DIST_MATCH_SCALAR = 0,
103 	RTE_DIST_MATCH_VECTOR,
104 	RTE_DIST_NUM_MATCH_FNS
105 };
106 
107 /**
108  * Buffer structure used to pass the pointer data between cores. This is cache
109  * line aligned, but to improve performance and prevent adjacent cache-line
110  * prefetches of buffers for other workers, e.g. when worker 1's buffer is on
111  * the next cache line to worker 0, we pad this out to two cache lines.
112  * We can pass up to 8 mbufs at a time in one cacheline.
113  * There is a separate cacheline for returns in the burst API.
114  */
115 struct rte_distributor_buffer {
116 	volatile int64_t bufptr64[RTE_DIST_BURST_SIZE]
117 		__rte_cache_aligned; /* <= outgoing to worker */
118 
119 	int64_t pad1 __rte_cache_aligned;    /* <= one cache line  */
120 
121 	volatile int64_t retptr64[RTE_DIST_BURST_SIZE]
122 		__rte_cache_aligned; /* <= incoming from worker */
123 
124 	int64_t pad2 __rte_cache_aligned;    /* <= one cache line  */
125 
126 	int count __rte_cache_aligned;       /* <= number of current mbufs */
127 };
128 
129 struct rte_distributor {
130 	TAILQ_ENTRY(rte_distributor) next;    /**< Next in list. */
131 
132 	char name[RTE_DISTRIBUTOR_NAMESIZE];  /**< Name of the ring. */
133 	unsigned int num_workers;             /**< Number of workers polling */
134 	unsigned int alg_type;                /**< Number of alg types */
135 
136 	/**>
137 	 * First cache line in the this array are the tags inflight
138 	 * on the worker core. Second cache line are the backlog
139 	 * that are going to go to the worker core.
140 	 */
141 	uint16_t in_flight_tags[RTE_DISTRIB_MAX_WORKERS][RTE_DIST_BURST_SIZE*2]
142 			__rte_cache_aligned;
143 
144 	struct rte_distributor_backlog backlog[RTE_DISTRIB_MAX_WORKERS]
145 			__rte_cache_aligned;
146 
147 	struct rte_distributor_buffer bufs[RTE_DISTRIB_MAX_WORKERS];
148 
149 	struct rte_distributor_returned_pkts returns;
150 
151 	enum rte_distributor_match_function dist_match_fn;
152 
153 	struct rte_distributor_single *d_single;
154 
155 	uint8_t active[RTE_DISTRIB_MAX_WORKERS];
156 	uint8_t activesum;
157 };
158 
159 void
160 find_match_scalar(struct rte_distributor *d,
161 			uint16_t *data_ptr,
162 			uint16_t *output_ptr);
163 
164 void
165 find_match_vec(struct rte_distributor *d,
166 			uint16_t *data_ptr,
167 			uint16_t *output_ptr);
168 
169 #endif /* _DIST_PRIV_H_ */
170