Commit | Line | Data |
---|---|---|
48363c84 MD |
1 | // SPDX-License-Identifier: MIT |
2 | /* | |
3 | * Copyright 2022 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
4 | */ | |
5 | ||
6 | #include <sched.h> | |
054b7b5c | 7 | #include <string.h> |
48363c84 MD |
8 | #include <stdint.h> |
9 | #include <pthread.h> | |
10 | #include <stdbool.h> | |
11 | #include <poll.h> | |
054b7b5c | 12 | #include <stdlib.h> |
5a76c31e | 13 | #include <unistd.h> |
bddcdc92 | 14 | #include <stdio.h> |
5a76c31e MD |
15 | #include <sys/syscall.h> |
16 | #include <linux/membarrier.h> | |
48363c84 MD |
17 | |
18 | #include "rcu.h" | |
054b7b5c | 19 | #include "smp.h" |
48363c84 | 20 | |
bddcdc92 MD |
21 | /* |
22 | * If both rseq (with glibc support) and membarrier system calls are | |
23 | * available, use them to replace barriers and atomics on the fast-path. | |
24 | */ | |
25 | unsigned int side_rcu_rseq_membarrier_available; | |
26 | ||
5a76c31e MD |
27 | static int |
28 | membarrier(int cmd, unsigned int flags, int cpu_id) | |
29 | { | |
30 | return syscall(__NR_membarrier, cmd, flags, cpu_id); | |
31 | } | |
32 | ||
48363c84 MD |
33 | /* active_readers is an input/output parameter. */ |
34 | static | |
35 | void check_active_readers(struct side_rcu_gp_state *gp_state, bool *active_readers) | |
36 | { | |
37 | uintptr_t sum[2] = { 0, 0 }; /* begin - end */ | |
38 | int i; | |
39 | ||
40 | for (i = 0; i < gp_state->nr_cpus; i++) { | |
41 | struct side_rcu_cpu_gp_state *cpu_state = &gp_state->percpu_state[i]; | |
42 | ||
7fb53c62 | 43 | if (active_readers[0]) { |
48363c84 | 44 | sum[0] -= __atomic_load_n(&cpu_state->count[0].end, __ATOMIC_RELAXED); |
7fb53c62 MD |
45 | sum[0] -= __atomic_load_n(&cpu_state->count[0].rseq_end, __ATOMIC_RELAXED); |
46 | } | |
47 | if (active_readers[1]) { | |
48363c84 | 48 | sum[1] -= __atomic_load_n(&cpu_state->count[1].end, __ATOMIC_RELAXED); |
7fb53c62 MD |
49 | sum[1] -= __atomic_load_n(&cpu_state->count[1].rseq_end, __ATOMIC_RELAXED); |
50 | } | |
48363c84 MD |
51 | } |
52 | ||
53 | /* | |
54 | * This memory barrier (C) pairs with either of memory barriers | |
55 | * (A) or (B) (one is sufficient). | |
56 | * | |
57 | * Read end counts before begin counts. Reading "end" before | |
58 | * "begin" counts ensures we never see an "end" without having | |
59 | * seen its associated "begin", because "begin" is always | |
60 | * incremented before "end", as guaranteed by memory barriers | |
61 | * (A) or (B). | |
62 | */ | |
bddcdc92 MD |
63 | if (side_rcu_rseq_membarrier_available) { |
64 | if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0, 0)) { | |
65 | perror("membarrier"); | |
66 | abort(); | |
67 | } | |
68 | } else { | |
69 | __atomic_thread_fence(__ATOMIC_SEQ_CST); | |
70 | } | |
48363c84 MD |
71 | |
72 | for (i = 0; i < gp_state->nr_cpus; i++) { | |
73 | struct side_rcu_cpu_gp_state *cpu_state = &gp_state->percpu_state[i]; | |
74 | ||
7fb53c62 | 75 | if (active_readers[0]) { |
48363c84 | 76 | sum[0] += __atomic_load_n(&cpu_state->count[0].begin, __ATOMIC_RELAXED); |
7fb53c62 MD |
77 | sum[0] += __atomic_load_n(&cpu_state->count[0].rseq_begin, __ATOMIC_RELAXED); |
78 | } | |
79 | if (active_readers[1]) { | |
48363c84 | 80 | sum[1] += __atomic_load_n(&cpu_state->count[1].begin, __ATOMIC_RELAXED); |
7fb53c62 MD |
81 | sum[1] += __atomic_load_n(&cpu_state->count[1].rseq_begin, __ATOMIC_RELAXED); |
82 | } | |
48363c84 MD |
83 | } |
84 | if (active_readers[0]) | |
85 | active_readers[0] = sum[0]; | |
86 | if (active_readers[1]) | |
87 | active_readers[1] = sum[1]; | |
88 | } | |
89 | ||
90 | /* | |
91 | * Wait for previous period to have no active readers. | |
92 | * | |
93 | * active_readers is an input/output parameter. | |
94 | */ | |
95 | static | |
96 | void wait_for_prev_period_readers(struct side_rcu_gp_state *gp_state, bool *active_readers) | |
97 | { | |
98 | unsigned int prev_period = gp_state->period ^ 1; | |
99 | ||
100 | /* | |
101 | * If a prior active readers scan already observed that no | |
102 | * readers are present for the previous period, there is no need | |
103 | * to scan again. | |
104 | */ | |
105 | if (!active_readers[prev_period]) | |
106 | return; | |
107 | /* | |
108 | * Wait for the sum of CPU begin/end counts to match for the | |
109 | * previous period. | |
110 | */ | |
111 | for (;;) { | |
112 | check_active_readers(gp_state, active_readers); | |
113 | if (!active_readers[prev_period]) | |
114 | break; | |
115 | /* Retry after 10ms. */ | |
116 | poll(NULL, 0, 10); | |
117 | } | |
118 | } | |
119 | ||
120 | /* | |
121 | * The grace period completes when it observes that there are no active | |
122 | * readers within each of the periods. | |
123 | * | |
124 | * The active_readers state is initially true for each period, until the | |
125 | * grace period observes that no readers are present for each given | |
126 | * period, at which point the active_readers state becomes false. | |
127 | */ | |
128 | void side_rcu_wait_grace_period(struct side_rcu_gp_state *gp_state) | |
129 | { | |
130 | bool active_readers[2] = { true, true }; | |
131 | ||
132 | /* | |
133 | * This memory barrier (D) pairs with memory barriers (A) and | |
134 | * (B) on the read-side. | |
135 | * | |
136 | * It orders prior loads and stores before the "end"/"begin" | |
137 | * reader state loads. In other words, it orders prior loads and | |
138 | * stores before observation of active readers quiescence, | |
139 | * effectively ensuring that read-side critical sections which | |
140 | * exist after the grace period completes are ordered after | |
141 | * loads and stores performed before the grace period. | |
142 | */ | |
bddcdc92 MD |
143 | if (side_rcu_rseq_membarrier_available) { |
144 | if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0, 0)) { | |
145 | perror("membarrier"); | |
146 | abort(); | |
147 | } | |
148 | } else { | |
149 | __atomic_thread_fence(__ATOMIC_SEQ_CST); | |
150 | } | |
48363c84 MD |
151 | |
152 | /* | |
153 | * First scan through all cpus, for both period. If no readers | |
154 | * are accounted for, we have observed quiescence and can | |
155 | * complete the grace period immediately. | |
156 | */ | |
157 | check_active_readers(gp_state, active_readers); | |
158 | if (!active_readers[0] && !active_readers[1]) | |
159 | goto end; | |
160 | ||
161 | pthread_mutex_lock(&gp_state->gp_lock); | |
162 | ||
163 | wait_for_prev_period_readers(gp_state, active_readers); | |
164 | /* | |
165 | * If the reader scan detected that there are no readers in the | |
166 | * current period as well, we can complete the grace period | |
167 | * immediately. | |
168 | */ | |
169 | if (!active_readers[gp_state->period]) | |
170 | goto unlock; | |
171 | ||
172 | /* Flip period: 0 -> 1, 1 -> 0. */ | |
173 | (void) __atomic_xor_fetch(&gp_state->period, 1, __ATOMIC_RELAXED); | |
174 | ||
175 | wait_for_prev_period_readers(gp_state, active_readers); | |
176 | unlock: | |
177 | pthread_mutex_unlock(&gp_state->gp_lock); | |
178 | end: | |
179 | /* | |
180 | * This memory barrier (E) pairs with memory barriers (A) and | |
181 | * (B) on the read-side. | |
182 | * | |
183 | * It orders the "end"/"begin" reader state loads before | |
184 | * following loads and stores. In other words, it orders | |
185 | * observation of active readers quiescence before following | |
186 | * loads and stores, effectively ensuring that read-side | |
187 | * critical sections which existed prior to the grace period | |
188 | * are ordered before loads and stores performed after the grace | |
189 | * period. | |
190 | */ | |
bddcdc92 MD |
191 | if (side_rcu_rseq_membarrier_available) { |
192 | if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0, 0)) { | |
193 | perror("membarrier"); | |
194 | abort(); | |
195 | } | |
196 | } else { | |
197 | __atomic_thread_fence(__ATOMIC_SEQ_CST); | |
198 | } | |
48363c84 | 199 | } |
054b7b5c MD |
200 | |
201 | void side_rcu_gp_init(struct side_rcu_gp_state *rcu_gp) | |
202 | { | |
bddcdc92 MD |
203 | bool has_membarrier = false, has_rseq = false; |
204 | ||
054b7b5c MD |
205 | memset(rcu_gp, 0, sizeof(*rcu_gp)); |
206 | rcu_gp->nr_cpus = get_possible_cpus_array_len(); | |
207 | if (!rcu_gp->nr_cpus) | |
208 | abort(); | |
209 | pthread_mutex_init(&rcu_gp->gp_lock, NULL); | |
210 | rcu_gp->percpu_state = calloc(rcu_gp->nr_cpus, sizeof(struct side_rcu_cpu_gp_state)); | |
211 | if (!rcu_gp->percpu_state) | |
212 | abort(); | |
bddcdc92 MD |
213 | if (!membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0, 0)) |
214 | has_membarrier = true; | |
215 | if (rseq_available(RSEQ_AVAILABLE_QUERY_LIBC)) | |
216 | has_rseq = true; | |
217 | if (has_membarrier && has_rseq) | |
218 | side_rcu_rseq_membarrier_available = 1; | |
054b7b5c | 219 | } |
6e46f5e6 MD |
220 | |
221 | void side_rcu_gp_exit(struct side_rcu_gp_state *rcu_gp) | |
222 | { | |
7fb53c62 | 223 | rseq_prepare_unload(); |
6e46f5e6 MD |
224 | pthread_mutex_destroy(&rcu_gp->gp_lock); |
225 | free(rcu_gp->percpu_state); | |
226 | } |