Include the beginnings of a FAQ about which timer to use.

1 files changed, 96 insertions, 0 deletions

diff --git a/src/lib/time/compat_time.h b/src/lib/time/compat_time.h
index 44fab62de5..c5337e9998 100644
--- a/src/lib/time/compat_time.h
+++ b/src/lib/time/compat_time.h
@@ -15,6 +15,102 @@
  * of tens of milliseconds.
  */
 
+/* Q: Should you use monotime or monotime_coarse as your source?
+ *
+ * A: Generally, you get better precision with monotime, but better
+ * performance with monotime_coarse.
+ *
+ * Q: Should you use monotime_t or monotime_coarse_t directly? Should you use
+ *    usec? msec? "stamp units?"
+ *
+ * A: Using monotime_t and monotime_coarse_t directly is most time-efficient,
+ * since no conversion needs to happen.  But they can potentially use more
+ * memory than you would need for a usec/msec/"stamp unit" count.
+ *
+ * Converting to usec or msec on some platforms, and working with them in
+ * general, creates a risk of doing a 64-bit division.  64-bit division is
+ * expensive on 32-bit platforms, which still do exist.
+ *
+ * The "stamp unit" type is designed to give a type that is cheap to convert
+ * from monotime_coarse, has resolution of about 1-2ms, and fits nicely in a
+ * 32-bit integer.  Its downside is that it does not correspond directly
+ * to a natural unit of time.
+ *
+ * There is not much point in using "coarse usec" or "coarse nsec", since the
+ * current coarse monotime implementations give you on the order of
+ * milliseconds of precision.
+ *
+ * Q: So, what backends is monotime_coarse using?
+ *
+ * A: Generally speaking, it uses "whatever monotonic-ish time implemenation
+ * does not require a context switch."  The various implementations provide
+ * this by having a view of the current time in a read-only memory page that
+ * is updated with a frequency corresponding to the kernel's tick count.
+ *
+ * On Windows, monotime_coarse uses GetCount64() [or GetTickCount() on
+ * obsolete systems].  MSDN claims that the resolution is "typically in the
+ * range of 10-16 msec", but it has said that for years.  Storing
+ * monotime_coarse_t uses 8 bytes.
+ *
+ * On OSX/iOS, monotime_coarse uses uses mach_approximate_time() where
+ * available, and falls back to regular monotime. The precision is not
+ * documented, but the implementation is open-source: it reads from a page
+ * that the kernel updates. Storing monotime_coarse_t uses 8 bytes.
+ *
+ * On unixy systems, monotime_coarse uses clock_gettime() with
+ * CLOCK_MONOTONIC_COARSE where available, and falls back to CLOCK_MONOTONIC.
+ * It typically uses vdso tricks to read from a page that the kernel updates.
+ * Its precision fixed, but you can get it with clock_getres(): on my Linux
+ * desktop, it claims to be 1 msec, but it will depend on the system HZ
+ * setting. Storing monotime_coarse_t uses 16 bytes.
+ *
+ * [TODO: Try CLOCK_MONOTONIC_FAST on foobsd.]
+ *
+ * Q: What backends is regular monotonic time using?
+ *
+ * A: In general, regular monotime uses something that requires a system call.
+ * On platforms where system calls are cheap, you win!  Otherwise, you lose.
+ *
+ * On Windows, monotonic time uses QuereyPerformanceCounter.  Storing
+ * monotime_t costs 8 bytes.
+ *
+ * On OSX/Apple, monotonic time uses mach_absolute_time.  Storing
+ * monotime_t costs 8 bytes.
+ *
+ * On unixy systems, monotonic time uses CLOCK_MONOTONIC.  Storing
+ * monotime_t costs 16 bytes.
+ *
+ * Q: Tell me about the costs of converting to a 64-bit nsec, usec, or msec
+ *    count.
+ *
+ * A: Windows, coarse: Cheap, since it's all multiplication.
+ *
+ * Windows, precise: Expensive on 32-bit: it needs 64-bit division.
+ *
+ * Apple, all: Expensive on 32-bit: it needs 64-bit division.
+ *
+ * Unixy, all: Fairly cheap, since the only division required is dividing
+ * tv_nsec 1000, and nanoseconds-per-second fits in a 32-bit value.
+ *
+ * All, "timestamp units": Cheap everywhere: it never divides.
+ *
+ * Q: This is only somewhat related, but how much precision could I hope for
+ *    from a libevent time.?
+ *
+ * A: Actually, it's _very_ related if you're timing in order to have a
+ * timeout happen.
+ *
+ * On Windows, it uses select: you could in theory have a microsecond
+ * resolution, but it usually isn't that accurate.
+ *
+ * On OSX, iOS, and BSD, you have kqueue: You could in theory have a nanosecond
+ * resolution, but it usually isn't that accurate.
+ *
+ * On Linux, you have epoll: It has a millisecond resolution.  Some recent
+ * Libevents can also use timerfd for higher resolution if
+ * EVENT_BASE_FLAG_PRECISE_TIMER is set: Tor doesn't set that flag.
+ */
+
 #ifndef TOR_COMPAT_TIME_H
 #define TOR_COMPAT_TIME_H