xref: /linux/tools/sched_ext/scx_central.c (revision 6315d93541f8a5f77c5ef5c4f25233e66d189603)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3  * Copyright (c) 2022 Meta Platforms, Inc. and affiliates.
4  * Copyright (c) 2022 Tejun Heo <tj@kernel.org>
5  * Copyright (c) 2022 David Vernet <dvernet@meta.com>
6  */
7 #define _GNU_SOURCE
8 #include <sched.h>
9 #include <stdio.h>
10 #include <unistd.h>
11 #include <inttypes.h>
12 #include <signal.h>
13 #include <assert.h>
14 #include <libgen.h>
15 #include <bpf/bpf.h>
16 #include <scx/common.h>
17 #include "scx_central.bpf.skel.h"
18 
19 const char help_fmt[] =
20 "A central FIFO sched_ext scheduler.\n"
21 "\n"
22 "See the top-level comment in .bpf.c for more details.\n"
23 "\n"
24 "Usage: %s [-s SLICE_US] [-c CPU]\n"
25 "\n"
26 "  -s SLICE_US   Override slice duration\n"
27 "  -c CPU        Override the central CPU (default: 0)\n"
28 "  -v            Print libbpf debug messages\n"
29 "  -h            Display this help and exit\n";
30 
31 static bool verbose;
32 static volatile int exit_req;
33 
34 static int libbpf_print_fn(enum libbpf_print_level level, const char *format, va_list args)
35 {
36 	if (level == LIBBPF_DEBUG && !verbose)
37 		return 0;
38 	return vfprintf(stderr, format, args);
39 }
40 
41 static void sigint_handler(int dummy)
42 {
43 	exit_req = 1;
44 }
45 
46 int main(int argc, char **argv)
47 {
48 	struct scx_central *skel;
49 	struct bpf_link *link;
50 	__u64 seq = 0, ecode;
51 	__s32 opt;
52 	cpu_set_t *cpuset;
53 
54 	libbpf_set_print(libbpf_print_fn);
55 	signal(SIGINT, sigint_handler);
56 	signal(SIGTERM, sigint_handler);
57 restart:
58 	skel = SCX_OPS_OPEN(central_ops, scx_central);
59 
60 	skel->rodata->central_cpu = 0;
61 	skel->rodata->nr_cpu_ids = libbpf_num_possible_cpus();
62 	skel->rodata->slice_ns = __COMPAT_ENUM_OR_ZERO("scx_public_consts", "SCX_SLICE_DFL");
63 
64 	assert(skel->rodata->nr_cpu_ids <= INT32_MAX);
65 
66 	while ((opt = getopt(argc, argv, "s:c:pvh")) != -1) {
67 		switch (opt) {
68 		case 's':
69 			skel->rodata->slice_ns = strtoull(optarg, NULL, 0) * 1000;
70 			break;
71 		case 'c': {
72 			u32 central_cpu = strtoul(optarg, NULL, 0);
73 			if (central_cpu >= skel->rodata->nr_cpu_ids) {
74 				fprintf(stderr, "invalid central CPU id value, %u given (%u max)\n", central_cpu, skel->rodata->nr_cpu_ids);
75 				return -1;
76 			}
77 			skel->rodata->central_cpu = (s32)central_cpu;
78 			break;
79 		}
80 		case 'v':
81 			verbose = true;
82 			break;
83 		default:
84 			fprintf(stderr, help_fmt, basename(argv[0]));
85 			return opt != 'h';
86 		}
87 	}
88 
89 	/* Resize arrays so their element count is equal to cpu count. */
90 	RESIZE_ARRAY(skel, data, cpu_gimme_task, skel->rodata->nr_cpu_ids);
91 	RESIZE_ARRAY(skel, data, cpu_started_at, skel->rodata->nr_cpu_ids);
92 
93 	SCX_OPS_LOAD(skel, central_ops, scx_central, uei);
94 
95 	/*
96 	 * Affinitize the loading thread to the central CPU, as:
97 	 * - That's where the BPF timer is first invoked in the BPF program.
98 	 * - We probably don't want this user space component to take up a core
99 	 *   from a task that would benefit from avoiding preemption on one of
100 	 *   the tickless cores.
101 	 *
102 	 * Until BPF supports pinning the timer, it's not guaranteed that it
103 	 * will always be invoked on the central CPU. In practice, this
104 	 * suffices the majority of the time.
105 	 */
106 	cpuset = CPU_ALLOC(skel->rodata->nr_cpu_ids);
107 	SCX_BUG_ON(!cpuset, "Failed to allocate cpuset");
108 	CPU_ZERO_S(CPU_ALLOC_SIZE(skel->rodata->nr_cpu_ids), cpuset);
109 	CPU_SET(skel->rodata->central_cpu, cpuset);
110 	SCX_BUG_ON(sched_setaffinity(0, sizeof(*cpuset), cpuset),
111 		   "Failed to affinitize to central CPU %d (max %d)",
112 		   skel->rodata->central_cpu, skel->rodata->nr_cpu_ids - 1);
113 	CPU_FREE(cpuset);
114 
115 	link = SCX_OPS_ATTACH(skel, central_ops, scx_central);
116 
117 	if (!skel->data->timer_pinned)
118 		printf("WARNING : BPF_F_TIMER_CPU_PIN not available, timer not pinned to central\n");
119 
120 	while (!exit_req && !UEI_EXITED(skel, uei)) {
121 		printf("[SEQ %llu]\n", seq++);
122 		printf("total   :%10" PRIu64 "    local:%10" PRIu64 "   queued:%10" PRIu64 "  lost:%10" PRIu64 "\n",
123 		       skel->bss->nr_total,
124 		       skel->bss->nr_locals,
125 		       skel->bss->nr_queued,
126 		       skel->bss->nr_lost_pids);
127 		printf("timer   :%10" PRIu64 " dispatch:%10" PRIu64 " mismatch:%10" PRIu64 " retry:%10" PRIu64 "\n",
128 		       skel->bss->nr_timers,
129 		       skel->bss->nr_dispatches,
130 		       skel->bss->nr_mismatches,
131 		       skel->bss->nr_retries);
132 		printf("overflow:%10" PRIu64 "\n",
133 		       skel->bss->nr_overflows);
134 		fflush(stdout);
135 		sleep(1);
136 	}
137 
138 	bpf_link__destroy(link);
139 	ecode = UEI_REPORT(skel, uei);
140 	scx_central__destroy(skel);
141 
142 	if (UEI_ECODE_RESTART(ecode))
143 		goto restart;
144 	return 0;
145 }
146