Revert time.cpp and use monotonic-time to implement cpuwide_time outside baidu

src/butil/time.cpp

@@ -25,7 +25,6 @@
 #endif
 #include <string.h>                          // memmem
 #undef _GNU_SOURCE
-#include <ctype.h>
 
 #include "butil/time.h"
 
@@ -82,7 +81,8 @@ int64_t monotonic_time_ns() {
 
 namespace detail {
 
-static int64_t read_cpu_current_frequency(char* buf, ssize_t n) {
+// read_cpu_frequency() is modified from source code of glibc.
+int64_t read_cpu_frequency(bool* invariant_tsc) {
     /* We read the information from the /proc filesystem.  It contains at
        least one line like
        cpu MHz         : 497.840237
@@ -90,8 +90,17 @@ static int64_t read_cpu_current_frequency(char* buf, ssize_t n) {
        cpu MHz         : 497.841
        We search for this line and convert the number in an integer.  */
 
+    const int fd = open("/proc/cpuinfo", O_RDONLY);
+    if (fd < 0) {
+        return 0;
+    }
+
     int64_t result = 0;
+    char buf[4096];  // should be enough
+    const ssize_t n = read(fd, buf, sizeof(buf));
+    if (n > 0) {
         char *mhz = static_cast<char*>(memmem(buf, n, "cpu MHz", 7));
+
         if (mhz != NULL) {
             char *endp = buf + n;
             int seen_decpoint = 0;
@@ -118,120 +127,17 @@ static int64_t read_cpu_current_frequency(char* buf, ssize_t n) {
                 result *= 10;
             }
         }
-    return result;
-}
-
-#if defined(__x86_64__) || defined(__i386__)
-#if defined(__pic__) && defined(__i386__)
-static void __cpuid(uint32_t reg[4], uint32_t code) {
-    __asm__ volatile (
-        "mov %%ebx, %%edi\n"
-        "cpuid\n"
-        "xchg %%edi, %%ebx\n"
-        : "=a"(reg[0]), "=D"(reg[1]), "=c"(reg[2]), "=d"(reg[3])
-        : "a"(code)
-  );
-}
-#else
-static void __cpuid(uint32_t reg[4], uint32_t code) {
-    __asm__ volatile (
-        "cpuid \n\t"
-        : "=a"(reg[0]), "=b"(reg[1]), "=c"(reg[2]), "=d"(reg[3])
-        : "a"(code)
-    );
-}
-#endif
-#endif
-
-static int64_t read_cpu_frequency_by_cpuid() {
-    int64_t result = 0;
-#if defined(__x86_64__) || defined(__i386__)
-    uint32_t reg[4];
-    __cpuid(reg, 0);
-    if (reg[0] >= 0x16 && reg[1] == 0x756e6547 && reg[2] == 0x6c65746e && reg[3] == 0x49656e69) {
-        //Intel CPU only
-        __cpuid(reg, 0x16);
-        return static_cast<uint64_t>(reg[0]) * 1000000UL;
-    }
-#endif
-    return result;
-}
-
-static int64_t read_cpu_frequency_from_brand_string() {
-    int64_t result = 0;
-#if defined(__x86_64__) || defined(__i386__)
-    union {
-        char brand[48];
-        uint32_t reg[12];
-    } buf;
-    __cpuid(buf.reg, 0x80000000);
-    if (buf.reg[0] < 0x80000004) {
-        return 0;
-    }
-    __cpuid(buf.reg, 0x80000002);
-    __cpuid(buf.reg+4, 0x80000003);
-    __cpuid(buf.reg+8, 0x80000004);
-    //Get something like: Intel(R) Xeon(R) CPU E5-2620 v2 @ 2.10GHz
-    char* end = buf.brand + sizeof(buf.brand);
-    char* p = buf.brand;
-    while (p != end && *p != '@') {
-        if (*p++ == '\n') {
-            return 0;
-        }
-    }
-    while (p != end && !isdigit(*p)) {
-        p++;
-    }
-    //expect x.xxGhz
-    //FSB may be 0.10GHz or 0.133...GHz
-    if (end - p < 7 || p[1] != '.'
-        || !isdigit(p[2]) || !isdigit(p[3]) ||  p[4] != 'G') {
-        return 0;
-    }
-    result = (p[0]-'0') * 10 + (p[2]-'0');
-    int64_t last = p[3] - '0';
-    if (last == 7) {
-        last = 6;
-    }
-    for (int i = 0; i < 8; i++) {
-        result = result * 10 + last;
-    }
-#endif
-    return result;
-}
-
-// read_cpu_frequency() is modified from source code of glibc.
-int64_t read_cpu_frequency(bool* invariant_tsc) {
-    const int fd = open("/proc/cpuinfo", O_RDONLY);
-    if (fd < 0) {
-        return 0;
-    }
-
-    char buf[4096];  // should be enough
-    const ssize_t n = read(fd, buf, sizeof(buf));
-    close (fd);
-    if (n <= 0) {
-        return 0;
-    }
 
         if (invariant_tsc) {
             char* flags_pos = static_cast<char*>(memmem(buf, n, "flags", 5));
-        if (flags_pos
-            && memmem(flags_pos, buf + n - flags_pos, "constant_tsc", 12)
-            && memmem(flags_pos, buf + n - flags_pos, "nonstop_tsc", 11)) {
-            int64_t result = read_cpu_frequency_by_cpuid();
-            if (result <= 0) {
-                result = read_cpu_frequency_from_brand_string();
-            }
-            if (result > 0) {
-                *invariant_tsc = true;
-                return result;
+            *invariant_tsc = 
+                (flags_pos &&
+                 memmem(flags_pos, buf + n - flags_pos, "constant_tsc", 12) &&
+                 memmem(flags_pos, buf + n - flags_pos, "nonstop_tsc", 11));
         }
     }
-        //current frequency is usually not invariant
-        *invariant_tsc = false; 
-    }
-    return read_cpu_current_frequency(buf, n);
+    close (fd);
+    return result;
 }
 
 // Return value must be >= 0

src/butil/time.h

@@ -236,6 +236,17 @@ extern int64_t invariant_cpu_freq;
 // note: Inlining shortens time cost per-call for 15ns in a loop of many
 //       calls to this function.
 inline int64_t cpuwide_time_ns() {
+#if !defined(BAIDU_INTERNAL)
+    // nearly impossible to get the correct invariant cpu frequency on
+    // different CPU and machines. CPU-ID rarely works and frequencies
+    // in "model name" and "cpu Mhz" are both unreliable.
+    // Since clock_gettime() in newer glibc/kernel is much faster(~30ns)
+    // which is closer to the previous impl. of cpuwide_time(~10ns), we
+    // simply use the monotonic time to get rid of all related issues.
+    timespec now;
+    clock_gettime(CLOCK_MONOTONIC, &now);
+    return now.tv_sec * 1000000000L + now.tv_nsec;
+#else
     int64_t cpu_freq = detail::invariant_cpu_freq;
     if (cpu_freq > 0) {
         const uint64_t tsc = detail::clock_cycles();
@@ -253,6 +264,7 @@ inline int64_t cpuwide_time_ns() {
         detail::invariant_cpu_freq = detail::read_invariant_cpu_frequency();
         return cpuwide_time_ns();
     }
+#endif // defined(BAIDU_INTERNAL)
 }
 
 inline int64_t cpuwide_time_us() {