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_RTS_C11_MEM_H_
11 #define _RTE_RING_RTS_C11_MEM_H_
12
13 /**
14 * @file rte_ring_rts_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 Relaxed Tail Sync (RTS) ring mode.
18 * For more information please refer to <rte_ring_rts.h>.
19 */
20
21 /**
22 * @internal This function updates tail values.
23 */
24 static __rte_always_inline void
__rte_ring_rts_update_tail(struct rte_ring_rts_headtail * ht)25 __rte_ring_rts_update_tail(struct rte_ring_rts_headtail *ht)
26 {
27 union __rte_ring_rts_poscnt h, ot, nt;
28
29 /*
30 * If there are other enqueues/dequeues in progress that
31 * might preceded us, then don't update tail with new value.
32 */
33
34 ot.raw = __atomic_load_n(&ht->tail.raw, __ATOMIC_ACQUIRE);
35
36 do {
37 /* on 32-bit systems we have to do atomic read here */
38 h.raw = __atomic_load_n(&ht->head.raw, __ATOMIC_RELAXED);
39
40 nt.raw = ot.raw;
41 if (++nt.val.cnt == h.val.cnt)
42 nt.val.pos = h.val.pos;
43
44 } while (__atomic_compare_exchange_n(&ht->tail.raw, &ot.raw, nt.raw,
45 0, __ATOMIC_RELEASE, __ATOMIC_ACQUIRE) == 0);
46 }
47
48 /**
49 * @internal This function waits till head/tail distance wouldn't
50 * exceed pre-defined max value.
51 */
52 static __rte_always_inline void
__rte_ring_rts_head_wait(const struct rte_ring_rts_headtail * ht,union __rte_ring_rts_poscnt * h)53 __rte_ring_rts_head_wait(const struct rte_ring_rts_headtail *ht,
54 union __rte_ring_rts_poscnt *h)
55 {
56 uint32_t max;
57
58 max = ht->htd_max;
59
60 while (h->val.pos - ht->tail.val.pos > max) {
61 rte_pause();
62 h->raw = __atomic_load_n(&ht->head.raw, __ATOMIC_ACQUIRE);
63 }
64 }
65
66 /**
67 * @internal This function updates the producer head for enqueue.
68 */
69 static __rte_always_inline uint32_t
__rte_ring_rts_move_prod_head(struct rte_ring * r,uint32_t num,enum rte_ring_queue_behavior behavior,uint32_t * old_head,uint32_t * free_entries)70 __rte_ring_rts_move_prod_head(struct rte_ring *r, uint32_t num,
71 enum rte_ring_queue_behavior behavior, uint32_t *old_head,
72 uint32_t *free_entries)
73 {
74 uint32_t n;
75 union __rte_ring_rts_poscnt nh, oh;
76
77 const uint32_t capacity = r->capacity;
78
79 oh.raw = __atomic_load_n(&r->rts_prod.head.raw, __ATOMIC_ACQUIRE);
80
81 do {
82 /* Reset n to the initial burst count */
83 n = num;
84
85 /*
86 * wait for prod head/tail distance,
87 * make sure that we read prod head *before*
88 * reading cons tail.
89 */
90 __rte_ring_rts_head_wait(&r->rts_prod, &oh);
91
92 /*
93 * The subtraction is done between two unsigned 32bits value
94 * (the result is always modulo 32 bits even if we have
95 * *old_head > cons_tail). So 'free_entries' is always between 0
96 * and capacity (which is < size).
97 */
98 *free_entries = capacity + r->cons.tail - oh.val.pos;
99
100 /* check that we have enough room in ring */
101 if (unlikely(n > *free_entries))
102 n = (behavior == RTE_RING_QUEUE_FIXED) ?
103 0 : *free_entries;
104
105 if (n == 0)
106 break;
107
108 nh.val.pos = oh.val.pos + n;
109 nh.val.cnt = oh.val.cnt + 1;
110
111 /*
112 * this CAS(ACQUIRE, ACQUIRE) serves as a hoist barrier to prevent:
113 * - OOO reads of cons tail value
114 * - OOO copy of elems to the ring
115 */
116 } while (__atomic_compare_exchange_n(&r->rts_prod.head.raw,
117 &oh.raw, nh.raw,
118 0, __ATOMIC_ACQUIRE, __ATOMIC_ACQUIRE) == 0);
119
120 *old_head = oh.val.pos;
121 return n;
122 }
123
124 /**
125 * @internal This function updates the consumer head for dequeue
126 */
127 static __rte_always_inline unsigned int
__rte_ring_rts_move_cons_head(struct rte_ring * r,uint32_t num,enum rte_ring_queue_behavior behavior,uint32_t * old_head,uint32_t * entries)128 __rte_ring_rts_move_cons_head(struct rte_ring *r, uint32_t num,
129 enum rte_ring_queue_behavior behavior, uint32_t *old_head,
130 uint32_t *entries)
131 {
132 uint32_t n;
133 union __rte_ring_rts_poscnt nh, oh;
134
135 oh.raw = __atomic_load_n(&r->rts_cons.head.raw, __ATOMIC_ACQUIRE);
136
137 /* move cons.head atomically */
138 do {
139 /* Restore n as it may change every loop */
140 n = num;
141
142 /*
143 * wait for cons head/tail distance,
144 * make sure that we read cons head *before*
145 * reading prod tail.
146 */
147 __rte_ring_rts_head_wait(&r->rts_cons, &oh);
148
149 /* The subtraction is done between two unsigned 32bits value
150 * (the result is always modulo 32 bits even if we have
151 * cons_head > prod_tail). So 'entries' is always between 0
152 * and size(ring)-1.
153 */
154 *entries = r->prod.tail - oh.val.pos;
155
156 /* Set the actual entries for dequeue */
157 if (n > *entries)
158 n = (behavior == RTE_RING_QUEUE_FIXED) ? 0 : *entries;
159
160 if (unlikely(n == 0))
161 break;
162
163 nh.val.pos = oh.val.pos + n;
164 nh.val.cnt = oh.val.cnt + 1;
165
166 /*
167 * this CAS(ACQUIRE, ACQUIRE) serves as a hoist barrier to prevent:
168 * - OOO reads of prod tail value
169 * - OOO copy of elems from the ring
170 */
171 } while (__atomic_compare_exchange_n(&r->rts_cons.head.raw,
172 &oh.raw, nh.raw,
173 0, __ATOMIC_ACQUIRE, __ATOMIC_ACQUIRE) == 0);
174
175 *old_head = oh.val.pos;
176 return n;
177 }
178
179 #endif /* _RTE_RING_RTS_C11_MEM_H_ */
180