Use faster clock_gettime() instead of sysctl() to get current ticks on iOS10

base/time/time_mac.cc

Index: base/time/time_mac.cc
diff --git a/base/time/time_mac.cc b/base/time/time_mac.cc
index 0ae3a30a3251419baead95006ed1d71843e8750c..9b3f0fe519d4921005caba74e164dcb5b5b465b4 100644
--- a/base/time/time_mac.cc
+++ b/base/time/time_mac.cc
@@ -23,6 +23,11 @@
 #include "base/numerics/safe_conversions.h"
 #include "build/build_config.h"
 
+#if defined(OS_IOS)
+#include <time.h>
+#include "base/ios/ios_util.h"
+#endif
+
 namespace {
 
 #if defined(OS_MACOSX) && !defined(OS_IOS)
@@ -52,8 +57,20 @@ int64_t MachAbsoluteTimeToTicks(uint64_t mach_absolute_time) {
 }
 #endif  // defined(OS_MACOSX) && !defined(OS_IOS)
 
+// Returns monotonically growing number of ticks in milliseconds since some

[miu, 2017/05/22 22:16:14]

s/milliseconds/microseconds/

[kapishnikov, 2017/05/23 17:56:50]

Done.

+// unspecified starting point.
 int64_t ComputeCurrentTicks() {
 #if defined(OS_IOS)
+  // iOS 10 supports clock_gettime(CLOCK_MONOTONIC, ...), which is
+  // around 15 times faster than sysctl() call. Use it if possible;
+  // otherwise, fall back to sysctl().
+  if (base::ios::IsRunningOnIOS10OrLater()) {
+    struct timespec tp;
+    if (clock_gettime(CLOCK_MONOTONIC, &tp) == 0) {

[miu, 2017/05/22 22:16:14]

Hmm...Curious that the POSIX compatibility layer would be faster than a
more-direct syscall interface. What interface does the POSIX library use? Maybe
we should use that directly? Or, does the POSIX library use some kind of special
CPU ticks + periodic reconciliation logic to avoid querying the system clock?

[kapishnikov, 2017/05/23 17:56:50]

I couldn't find the sources but my guess is that it used CPU ticks + adjustments
for the time the device was in sleep. The implementation may use some calls
which are not allowed to be called directly by apps.

BTW, I found that iOS 10 also supports CLOCK_MONOTONIC_RAW clock type, which I
think is a better candidate than CLOCK_MONOTONIC, since CLOCK_MONOTONIC
frequency may be impacted by NTP time adjustments. So, I changed it to
CLOCK_MONOTONIC_RAW.

+      return tp.tv_sec * 1000000 + tp.tv_nsec / 1000;
+    }
+  }
+
   // On iOS mach_absolute_time stops while the device is sleeping. Instead use
   // now - KERN_BOOTTIME to get a time difference that is not impacted by clock
   // changes. KERN_BOOTTIME will be updated by the system whenever the system