1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2019 Intel Corporation
5 #ifndef _RTE_STACK_LF_C11_H_
6 #define _RTE_STACK_LF_C11_H_
8 #include <rte_branch_prediction.h>
9 #include <rte_prefetch.h>
11 static __rte_always_inline unsigned int
12 __rte_stack_lf_count(struct rte_stack *s)
14 /* stack_lf_push() and stack_lf_pop() do not update the list's contents
15 * and stack_lf->len atomically, which can cause the list to appear
16 * shorter than it actually is if this function is called while other
17 * threads are modifying the list.
19 * However, given the inherently approximate nature of the get_count
20 * callback -- even if the list and its size were updated atomically,
21 * the size could change between when get_count executes and when the
22 * value is returned to the caller -- this is acceptable.
24 * The stack_lf->len updates are placed such that the list may appear to
25 * have fewer elements than it does, but will never appear to have more
26 * elements. If the mempool is near-empty to the point that this is a
27 * concern, the user should consider increasing the mempool size.
29 return (unsigned int)__atomic_load_n(&s->stack_lf.used.len,
33 static __rte_always_inline void
34 __rte_stack_lf_push_elems(struct rte_stack_lf_list *list,
35 struct rte_stack_lf_elem *first,
36 struct rte_stack_lf_elem *last,
39 #ifndef RTE_ARCH_X86_64
45 struct rte_stack_lf_head old_head;
48 old_head = list->head;
51 struct rte_stack_lf_head new_head;
53 /* Use an acquire fence to establish a synchronized-with
54 * relationship between the list->head load and store-release
55 * operations (as part of the rte_atomic128_cmp_exchange()).
57 __atomic_thread_fence(__ATOMIC_ACQUIRE);
59 /* Swing the top pointer to the first element in the list and
60 * make the last element point to the old top.
63 new_head.cnt = old_head.cnt + 1;
65 last->next = old_head.top;
67 /* Use the release memmodel to ensure the writes to the LF LIFO
68 * elements are visible before the head pointer write.
70 success = rte_atomic128_cmp_exchange(
71 (rte_int128_t *)&list->head,
72 (rte_int128_t *)&old_head,
73 (rte_int128_t *)&new_head,
76 } while (success == 0);
78 /* Ensure the stack modifications are not reordered with respect
79 * to the LIFO len update.
81 __atomic_add_fetch(&list->len, num, __ATOMIC_RELEASE);
85 static __rte_always_inline struct rte_stack_lf_elem *
86 __rte_stack_lf_pop_elems(struct rte_stack_lf_list *list,
89 struct rte_stack_lf_elem **last)
91 #ifndef RTE_ARCH_X86_64
92 RTE_SET_USED(obj_table);
99 struct rte_stack_lf_head old_head;
103 /* Reserve num elements, if available */
104 len = __atomic_load_n(&list->len, __ATOMIC_ACQUIRE);
107 /* Does the list contain enough elements? */
108 if (unlikely(len < num))
111 /* len is updated on failure */
112 if (__atomic_compare_exchange_n(&list->len,
119 /* If a torn read occurs, the CAS will fail and set old_head to the
120 * correct/latest value.
122 old_head = list->head;
124 /* Pop num elements */
126 struct rte_stack_lf_head new_head;
127 struct rte_stack_lf_elem *tmp;
130 /* Use the acquire memmodel to ensure the reads to the LF LIFO
131 * elements are properly ordered with respect to the head
134 __atomic_thread_fence(__ATOMIC_ACQUIRE);
136 rte_prefetch0(old_head.top);
140 /* Traverse the list to find the new head. A next pointer will
141 * either point to another element or NULL; if a thread
142 * encounters a pointer that has already been popped, the CAS
145 for (i = 0; i < num && tmp != NULL; i++) {
146 rte_prefetch0(tmp->next);
148 obj_table[i] = tmp->data;
154 /* If NULL was encountered, the list was modified while
155 * traversing it. Retry.
161 new_head.cnt = old_head.cnt + 1;
163 success = rte_atomic128_cmp_exchange(
164 (rte_int128_t *)&list->head,
165 (rte_int128_t *)&old_head,
166 (rte_int128_t *)&new_head,
169 } while (success == 0);
175 #endif /* _RTE_STACK_LF_C11_H_ */