1 /* SPDX-License-Identifier: BSD-3-Clause
2 *
3 * Copyright (c) 2010-2020 Intel Corporation
4 * Copyright (c) 2007-2009 Kip Macy [email protected]
5 * All rights reserved.
6 * Derived from FreeBSD's bufring.h
7 * Used as BSD-3 Licensed with permission from Kip Macy.
8 */
9
10 #ifndef _RTE_RING_HTS_C11_MEM_H_
11 #define _RTE_RING_HTS_C11_MEM_H_
12
13 /**
14 * @file rte_ring_hts_c11_mem.h
15 * It is not recommended to include this file directly,
16 * include <rte_ring.h> instead.
17 * Contains internal helper functions for head/tail sync (HTS) ring mode.
18 * For more information please refer to <rte_ring_hts.h>.
19 */
20
21 /**
22 * @internal update tail with new value.
23 */
24 static __rte_always_inline void
__rte_ring_hts_update_tail(struct rte_ring_hts_headtail * ht,uint32_t old_tail,uint32_t num,uint32_t enqueue)25 __rte_ring_hts_update_tail(struct rte_ring_hts_headtail *ht, uint32_t old_tail,
26 uint32_t num, uint32_t enqueue)
27 {
28 uint32_t tail;
29
30 RTE_SET_USED(enqueue);
31
32 tail = old_tail + num;
33 __atomic_store_n(&ht->ht.pos.tail, tail, __ATOMIC_RELEASE);
34 }
35
36 /**
37 * @internal waits till tail will become equal to head.
38 * Means no writer/reader is active for that ring.
39 * Suppose to work as serialization point.
40 */
41 static __rte_always_inline void
__rte_ring_hts_head_wait(const struct rte_ring_hts_headtail * ht,union __rte_ring_hts_pos * p)42 __rte_ring_hts_head_wait(const struct rte_ring_hts_headtail *ht,
43 union __rte_ring_hts_pos *p)
44 {
45 while (p->pos.head != p->pos.tail) {
46 rte_pause();
47 p->raw = __atomic_load_n(&ht->ht.raw, __ATOMIC_ACQUIRE);
48 }
49 }
50
51 /**
52 * @internal This function updates the producer head for enqueue
53 */
54 static __rte_always_inline unsigned int
__rte_ring_hts_move_prod_head(struct rte_ring * r,unsigned int num,enum rte_ring_queue_behavior behavior,uint32_t * old_head,uint32_t * free_entries)55 __rte_ring_hts_move_prod_head(struct rte_ring *r, unsigned int num,
56 enum rte_ring_queue_behavior behavior, uint32_t *old_head,
57 uint32_t *free_entries)
58 {
59 uint32_t n;
60 union __rte_ring_hts_pos np, op;
61
62 const uint32_t capacity = r->capacity;
63
64 op.raw = __atomic_load_n(&r->hts_prod.ht.raw, __ATOMIC_ACQUIRE);
65
66 do {
67 /* Reset n to the initial burst count */
68 n = num;
69
70 /*
71 * wait for tail to be equal to head,
72 * make sure that we read prod head/tail *before*
73 * reading cons tail.
74 */
75 __rte_ring_hts_head_wait(&r->hts_prod, &op);
76
77 /*
78 * The subtraction is done between two unsigned 32bits value
79 * (the result is always modulo 32 bits even if we have
80 * *old_head > cons_tail). So 'free_entries' is always between 0
81 * and capacity (which is < size).
82 */
83 *free_entries = capacity + r->cons.tail - op.pos.head;
84
85 /* check that we have enough room in ring */
86 if (unlikely(n > *free_entries))
87 n = (behavior == RTE_RING_QUEUE_FIXED) ?
88 0 : *free_entries;
89
90 if (n == 0)
91 break;
92
93 np.pos.tail = op.pos.tail;
94 np.pos.head = op.pos.head + n;
95
96 /*
97 * this CAS(ACQUIRE, ACQUIRE) serves as a hoist barrier to prevent:
98 * - OOO reads of cons tail value
99 * - OOO copy of elems from the ring
100 */
101 } while (__atomic_compare_exchange_n(&r->hts_prod.ht.raw,
102 &op.raw, np.raw,
103 0, __ATOMIC_ACQUIRE, __ATOMIC_ACQUIRE) == 0);
104
105 *old_head = op.pos.head;
106 return n;
107 }
108
109 /**
110 * @internal This function updates the consumer head for dequeue
111 */
112 static __rte_always_inline unsigned int
__rte_ring_hts_move_cons_head(struct rte_ring * r,unsigned int num,enum rte_ring_queue_behavior behavior,uint32_t * old_head,uint32_t * entries)113 __rte_ring_hts_move_cons_head(struct rte_ring *r, unsigned int num,
114 enum rte_ring_queue_behavior behavior, uint32_t *old_head,
115 uint32_t *entries)
116 {
117 uint32_t n;
118 union __rte_ring_hts_pos np, op;
119
120 op.raw = __atomic_load_n(&r->hts_cons.ht.raw, __ATOMIC_ACQUIRE);
121
122 /* move cons.head atomically */
123 do {
124 /* Restore n as it may change every loop */
125 n = num;
126
127 /*
128 * wait for tail to be equal to head,
129 * make sure that we read cons head/tail *before*
130 * reading prod tail.
131 */
132 __rte_ring_hts_head_wait(&r->hts_cons, &op);
133
134 /* The subtraction is done between two unsigned 32bits value
135 * (the result is always modulo 32 bits even if we have
136 * cons_head > prod_tail). So 'entries' is always between 0
137 * and size(ring)-1.
138 */
139 *entries = r->prod.tail - op.pos.head;
140
141 /* Set the actual entries for dequeue */
142 if (n > *entries)
143 n = (behavior == RTE_RING_QUEUE_FIXED) ? 0 : *entries;
144
145 if (unlikely(n == 0))
146 break;
147
148 np.pos.tail = op.pos.tail;
149 np.pos.head = op.pos.head + n;
150
151 /*
152 * this CAS(ACQUIRE, ACQUIRE) serves as a hoist barrier to prevent:
153 * - OOO reads of prod tail value
154 * - OOO copy of elems from the ring
155 */
156 } while (__atomic_compare_exchange_n(&r->hts_cons.ht.raw,
157 &op.raw, np.raw,
158 0, __ATOMIC_ACQUIRE, __ATOMIC_ACQUIRE) == 0);
159
160 *old_head = op.pos.head;
161 return n;
162 }
163
164 #endif /* _RTE_RING_HTS_C11_MEM_H_ */
165