[watchman] use fsevents on osx

Summary:
fseventsd uses some low level OS magic to subscribe to filesystem
changes without needing to hold an open descriptor for everything.

This diff makes us use it on OS/X.  For the most part, this diff
is pretty clear.  There are two things that are worth calling out:

* Discovered that we should have put the root number into the cookie filename
  to avoid the potential for cookie filename collisions when we recrawl and rewind
  the tick counter

* We run fsevents in a thread of its own so that we can filter out ignored items and
  test to see if events are available (there is no native "poll" functionality that I can
  find).  This is begging to be refactored and we'll take care of that in a later diff.

Test Plan: `make integration` passes

Reviewers: sid0, fugalh

Reviewed By: fugalh

CC: royw

Differential Revision: https://phabricator.fb.com/D1031495

README.markdown

@@ -12,7 +12,8 @@ also trigger actions (such as rebuilding assets) when matching files change.
 Watchman is known to compile and pass its test suite on:
 
  * Linux systems with `inotify`
- * OS X and BSDish systems (FreeBSD 9.1, OpenBSD 5.2) that have the
+ * OS X (uses FSEvents)
+ * BSDish systems (FreeBSD 9.1, OpenBSD 5.2) that have the
    `kqueue(2)` facility
  * Illumos and Solaris style systems that have `port_create(3C)`
 
@@ -1093,9 +1094,8 @@ haven't actually changed.
 ### Max OS File Descriptor Limits
 
 The default per-process descriptor limit on current versions of OS X is
-extremely low (256!).  Since kqueue() requires an open descriptor for each
-watched directory, you will very quickly run into resource limits if your trees
-are large or if you do not raise the limits in your system configuration.
+extremely low (256!).  More recent versions of watchman (2.9.2 and later)
+use FSEvents and are not so sensitive to descriptor limits.
 
 Watchman will attempt to raise its descriptor limit to match
 `kern.maxfilesperproc` when it starts up, so you shouldn't need to mess

configure.ac

@@ -1,4 +1,4 @@
-AC_INIT([watchman], [2.9.1], [], [watchman])
+AC_INIT([watchman], [2.9.2], [], [watchman])
 AC_CANONICAL_TARGET
 AM_INIT_AUTOMAKE([dist-bzip2 subdir-objects])
 
@@ -167,6 +167,9 @@ AC_CHECK_HEADERS(sys/types.h inttypes.h locale.h port.h sys/inotify.h sys/event.
 AC_CHECK_FUNCS(kqueue port_create inotify_init strtoll localeconv)
 AC_CHECK_HEADERS(valgrind/valgrind.h)
 AC_CHECK_HEADERS(execinfo.h)
+AC_CHECK_HEADERS(CoreServices/CoreServices.h, [
+  LIBS="$LIBS -framework CoreServices"
+])
 AC_CHECK_FUNCS(backtrace backtrace_symbols)
 
 # Do this after we've looked for functions

listener.c

@@ -659,7 +659,7 @@ static void wakeme(int signo)
   unused_parameter(signo);
 }
 
-#ifdef HAVE_KQUEUE
+#if defined(HAVE_KQUEUE) || defined(HAVE_FSEVENTS)
 #ifdef __OpenBSD__
 #include <sys/siginfo.h>
 #endif
@@ -691,7 +691,7 @@ bool w_start_listener(const char *path)
   hb = gimli_heartbeat_attach();
 #endif
 
-#ifdef HAVE_KQUEUE
+#if defined(HAVE_KQUEUE) || defined(HAVE_FSEVENTS)
   {
     struct rlimit limit;
     int mib[2] = { CTL_KERN,
@@ -730,12 +730,14 @@ bool w_start_listener(const char *path)
     }
 
     getrlimit(RLIMIT_NOFILE, &limit);
+#ifndef HAVE_FSEVENTS
     if (limit.rlim_cur < 10240) {
       w_log(W_LOG_ERR,
           "Your file descriptor limit is very low (%" PRIu64 "), "
           "please consult the watchman docs on raising the limits\n",
           limit.rlim_cur);
     }
+#endif
   }
 #endif