1*4882a593Smuzhiyun // SPDX-License-Identifier: GPL-2.0
2*4882a593Smuzhiyun #include <errno.h>
3*4882a593Smuzhiyun #include <unistd.h>
4*4882a593Smuzhiyun #include <stdlib.h>
5*4882a593Smuzhiyun #include <signal.h>
6*4882a593Smuzhiyun #include <sys/mman.h>
7*4882a593Smuzhiyun #include <sys/types.h>
8*4882a593Smuzhiyun #include <sys/wait.h>
9*4882a593Smuzhiyun #include <linux/string.h>
10*4882a593Smuzhiyun #include <linux/types.h>
11*4882a593Smuzhiyun #include "perf-sys.h"
12*4882a593Smuzhiyun #include "debug.h"
13*4882a593Smuzhiyun #include "tests/tests.h"
14*4882a593Smuzhiyun #include "cloexec.h"
15*4882a593Smuzhiyun #include "event.h"
16*4882a593Smuzhiyun #include <internal/lib.h> // page_size
17*4882a593Smuzhiyun #include "arch-tests.h"
18*4882a593Smuzhiyun
rdpmc(unsigned int counter)19*4882a593Smuzhiyun static u64 rdpmc(unsigned int counter)
20*4882a593Smuzhiyun {
21*4882a593Smuzhiyun unsigned int low, high;
22*4882a593Smuzhiyun
23*4882a593Smuzhiyun asm volatile("rdpmc" : "=a" (low), "=d" (high) : "c" (counter));
24*4882a593Smuzhiyun
25*4882a593Smuzhiyun return low | ((u64)high) << 32;
26*4882a593Smuzhiyun }
27*4882a593Smuzhiyun
rdtsc(void)28*4882a593Smuzhiyun static u64 rdtsc(void)
29*4882a593Smuzhiyun {
30*4882a593Smuzhiyun unsigned int low, high;
31*4882a593Smuzhiyun
32*4882a593Smuzhiyun asm volatile("rdtsc" : "=a" (low), "=d" (high));
33*4882a593Smuzhiyun
34*4882a593Smuzhiyun return low | ((u64)high) << 32;
35*4882a593Smuzhiyun }
36*4882a593Smuzhiyun
mmap_read_self(void * addr)37*4882a593Smuzhiyun static u64 mmap_read_self(void *addr)
38*4882a593Smuzhiyun {
39*4882a593Smuzhiyun struct perf_event_mmap_page *pc = addr;
40*4882a593Smuzhiyun u32 seq, idx, time_mult = 0, time_shift = 0;
41*4882a593Smuzhiyun u64 count, cyc = 0, time_offset = 0, enabled, running, delta;
42*4882a593Smuzhiyun
43*4882a593Smuzhiyun do {
44*4882a593Smuzhiyun seq = pc->lock;
45*4882a593Smuzhiyun barrier();
46*4882a593Smuzhiyun
47*4882a593Smuzhiyun enabled = pc->time_enabled;
48*4882a593Smuzhiyun running = pc->time_running;
49*4882a593Smuzhiyun
50*4882a593Smuzhiyun if (enabled != running) {
51*4882a593Smuzhiyun cyc = rdtsc();
52*4882a593Smuzhiyun time_mult = pc->time_mult;
53*4882a593Smuzhiyun time_shift = pc->time_shift;
54*4882a593Smuzhiyun time_offset = pc->time_offset;
55*4882a593Smuzhiyun }
56*4882a593Smuzhiyun
57*4882a593Smuzhiyun idx = pc->index;
58*4882a593Smuzhiyun count = pc->offset;
59*4882a593Smuzhiyun if (idx)
60*4882a593Smuzhiyun count += rdpmc(idx - 1);
61*4882a593Smuzhiyun
62*4882a593Smuzhiyun barrier();
63*4882a593Smuzhiyun } while (pc->lock != seq);
64*4882a593Smuzhiyun
65*4882a593Smuzhiyun if (enabled != running) {
66*4882a593Smuzhiyun u64 quot, rem;
67*4882a593Smuzhiyun
68*4882a593Smuzhiyun quot = (cyc >> time_shift);
69*4882a593Smuzhiyun rem = cyc & (((u64)1 << time_shift) - 1);
70*4882a593Smuzhiyun delta = time_offset + quot * time_mult +
71*4882a593Smuzhiyun ((rem * time_mult) >> time_shift);
72*4882a593Smuzhiyun
73*4882a593Smuzhiyun enabled += delta;
74*4882a593Smuzhiyun if (idx)
75*4882a593Smuzhiyun running += delta;
76*4882a593Smuzhiyun
77*4882a593Smuzhiyun quot = count / running;
78*4882a593Smuzhiyun rem = count % running;
79*4882a593Smuzhiyun count = quot * enabled + (rem * enabled) / running;
80*4882a593Smuzhiyun }
81*4882a593Smuzhiyun
82*4882a593Smuzhiyun return count;
83*4882a593Smuzhiyun }
84*4882a593Smuzhiyun
85*4882a593Smuzhiyun /*
86*4882a593Smuzhiyun * If the RDPMC instruction faults then signal this back to the test parent task:
87*4882a593Smuzhiyun */
segfault_handler(int sig __maybe_unused,siginfo_t * info __maybe_unused,void * uc __maybe_unused)88*4882a593Smuzhiyun static void segfault_handler(int sig __maybe_unused,
89*4882a593Smuzhiyun siginfo_t *info __maybe_unused,
90*4882a593Smuzhiyun void *uc __maybe_unused)
91*4882a593Smuzhiyun {
92*4882a593Smuzhiyun exit(-1);
93*4882a593Smuzhiyun }
94*4882a593Smuzhiyun
__test__rdpmc(void)95*4882a593Smuzhiyun static int __test__rdpmc(void)
96*4882a593Smuzhiyun {
97*4882a593Smuzhiyun volatile int tmp = 0;
98*4882a593Smuzhiyun u64 i, loops = 1000;
99*4882a593Smuzhiyun int n;
100*4882a593Smuzhiyun int fd;
101*4882a593Smuzhiyun void *addr;
102*4882a593Smuzhiyun struct perf_event_attr attr = {
103*4882a593Smuzhiyun .type = PERF_TYPE_HARDWARE,
104*4882a593Smuzhiyun .config = PERF_COUNT_HW_INSTRUCTIONS,
105*4882a593Smuzhiyun .exclude_kernel = 1,
106*4882a593Smuzhiyun };
107*4882a593Smuzhiyun u64 delta_sum = 0;
108*4882a593Smuzhiyun struct sigaction sa;
109*4882a593Smuzhiyun char sbuf[STRERR_BUFSIZE];
110*4882a593Smuzhiyun
111*4882a593Smuzhiyun sigfillset(&sa.sa_mask);
112*4882a593Smuzhiyun sa.sa_sigaction = segfault_handler;
113*4882a593Smuzhiyun sa.sa_flags = 0;
114*4882a593Smuzhiyun sigaction(SIGSEGV, &sa, NULL);
115*4882a593Smuzhiyun
116*4882a593Smuzhiyun fd = sys_perf_event_open(&attr, 0, -1, -1,
117*4882a593Smuzhiyun perf_event_open_cloexec_flag());
118*4882a593Smuzhiyun if (fd < 0) {
119*4882a593Smuzhiyun pr_err("Error: sys_perf_event_open() syscall returned "
120*4882a593Smuzhiyun "with %d (%s)\n", fd,
121*4882a593Smuzhiyun str_error_r(errno, sbuf, sizeof(sbuf)));
122*4882a593Smuzhiyun return -1;
123*4882a593Smuzhiyun }
124*4882a593Smuzhiyun
125*4882a593Smuzhiyun addr = mmap(NULL, page_size, PROT_READ, MAP_SHARED, fd, 0);
126*4882a593Smuzhiyun if (addr == (void *)(-1)) {
127*4882a593Smuzhiyun pr_err("Error: mmap() syscall returned with (%s)\n",
128*4882a593Smuzhiyun str_error_r(errno, sbuf, sizeof(sbuf)));
129*4882a593Smuzhiyun goto out_close;
130*4882a593Smuzhiyun }
131*4882a593Smuzhiyun
132*4882a593Smuzhiyun for (n = 0; n < 6; n++) {
133*4882a593Smuzhiyun u64 stamp, now, delta;
134*4882a593Smuzhiyun
135*4882a593Smuzhiyun stamp = mmap_read_self(addr);
136*4882a593Smuzhiyun
137*4882a593Smuzhiyun for (i = 0; i < loops; i++)
138*4882a593Smuzhiyun tmp++;
139*4882a593Smuzhiyun
140*4882a593Smuzhiyun now = mmap_read_self(addr);
141*4882a593Smuzhiyun loops *= 10;
142*4882a593Smuzhiyun
143*4882a593Smuzhiyun delta = now - stamp;
144*4882a593Smuzhiyun pr_debug("%14d: %14Lu\n", n, (long long)delta);
145*4882a593Smuzhiyun
146*4882a593Smuzhiyun delta_sum += delta;
147*4882a593Smuzhiyun }
148*4882a593Smuzhiyun
149*4882a593Smuzhiyun munmap(addr, page_size);
150*4882a593Smuzhiyun pr_debug(" ");
151*4882a593Smuzhiyun out_close:
152*4882a593Smuzhiyun close(fd);
153*4882a593Smuzhiyun
154*4882a593Smuzhiyun if (!delta_sum)
155*4882a593Smuzhiyun return -1;
156*4882a593Smuzhiyun
157*4882a593Smuzhiyun return 0;
158*4882a593Smuzhiyun }
159*4882a593Smuzhiyun
test__rdpmc(struct test * test __maybe_unused,int subtest __maybe_unused)160*4882a593Smuzhiyun int test__rdpmc(struct test *test __maybe_unused, int subtest __maybe_unused)
161*4882a593Smuzhiyun {
162*4882a593Smuzhiyun int status = 0;
163*4882a593Smuzhiyun int wret = 0;
164*4882a593Smuzhiyun int ret;
165*4882a593Smuzhiyun int pid;
166*4882a593Smuzhiyun
167*4882a593Smuzhiyun pid = fork();
168*4882a593Smuzhiyun if (pid < 0)
169*4882a593Smuzhiyun return -1;
170*4882a593Smuzhiyun
171*4882a593Smuzhiyun if (!pid) {
172*4882a593Smuzhiyun ret = __test__rdpmc();
173*4882a593Smuzhiyun
174*4882a593Smuzhiyun exit(ret);
175*4882a593Smuzhiyun }
176*4882a593Smuzhiyun
177*4882a593Smuzhiyun wret = waitpid(pid, &status, 0);
178*4882a593Smuzhiyun if (wret < 0 || status)
179*4882a593Smuzhiyun return -1;
180*4882a593Smuzhiyun
181*4882a593Smuzhiyun return 0;
182*4882a593Smuzhiyun }
183