experimental, and likely broken, inotify support

11 files changed, 250 insertions, 43 deletions

diff --git a/ev.c b/ev.c
index 06c0ec3..0d5e4b9 100644
--- a/ev.c
+++ b/ev.c
@@ -96,6 +96,14 @@ extern "C" {
 #  endif
 # endif
 
+# ifndef EV_USE_INOTIFY
+#  if HAVE_INOTIFY_INIT && HAVE_SYS_INOTIFY_H
+#   define EV_USE_INOTIFY 1
+#  else
+#   define EV_USE_INOTIFY 0
+#  endif
+# endif
+
 #endif
 
 #include <math.h>
@@ -112,6 +120,12 @@ extern "C" {
 
 #include <signal.h>
 
+#ifdef EV_H
+# include EV_H
+#else
+# include "ev.h"
+#endif
+
 #ifndef _WIN32
 # include <sys/time.h>
 # include <sys/wait.h>
@@ -158,6 +172,10 @@ extern "C" {
 # define EV_USE_PORT 0
 #endif
 
+#ifndef EV_USE_INOTIFY
+# define EV_USE_INOTIFY 0
+#endif
+
 #ifndef EV_PID_HASHSIZE
 # if EV_MINIMAL
 #  define EV_PID_HASHSIZE 1
@@ -166,6 +184,14 @@ extern "C" {
 # endif
 #endif
 
+#ifndef EV_INOTIFY_HASHSIZE
+# if EV_MINIMAL
+#  define EV_INOTIFY_HASHSIZE 1
+# else
+#  define EV_INOTIFY_HASHSIZE 16
+# endif
+#endif
+
 /**/
 
 #ifndef CLOCK_MONOTONIC
@@ -182,18 +208,20 @@ extern "C" {
 # include <winsock.h>
 #endif
 
+#if !EV_STAT_ENABLE
+# define EV_USE_INOTIFY 0
+#endif
+
+#if EV_USE_INOTIFY
+# include <sys/inotify.h>
+#endif
+
 /**/
 
 #define MIN_TIMEJUMP  1. /* minimum timejump that gets detected (if monotonic clock available) */
 #define MAX_BLOCKTIME 59.743 /* never wait longer than this time (to detect time jumps) */
 /*#define CLEANUP_INTERVAL (MAX_BLOCKTIME * 5.) /* how often to try to free memory and re-check fds */
 
-#ifdef EV_H
-# include EV_H
-#else
-# include "ev.h"
-#endif
-
 #if __GNUC__ >= 3
 # define expect(expr,value)         __builtin_expect ((expr),(value))
 # define inline_size                static inline /* inline for codesize */
@@ -298,6 +326,13 @@ typedef struct
   int events;
 } ANPENDING;
 
+typedef struct
+{
+#if EV_USE_INOTIFY
+  WL head;
+#endif
+} ANFS;
+
 #if EV_MULTIPLICITY
 
   struct ev_loop
@@ -881,6 +916,11 @@ loop_init (EV_P_ unsigned int flags)
         flags |= ev_recommended_backends ();
 
       backend = 0;
+      backend_fd = -1;
+#if EV_USE_INOTIFY
+      fs_fd = -2;
+#endif
+
 #if EV_USE_PORT
       if (!backend && (flags & EVBACKEND_PORT  )) backend = port_init   (EV_A_ flags);
 #endif
@@ -907,6 +947,14 @@ loop_destroy (EV_P)
 {
   int i;
 
+#if EV_USE_INOTIFY
+  if (fs_fd >= 0)
+    close (fs_fd);
+#endif
+
+  if (backend_fd >= 0)
+    close (backend_fd);
+
 #if EV_USE_PORT
   if (backend == EVBACKEND_PORT  ) port_destroy   (EV_A);
 #endif
@@ -1667,6 +1715,127 @@ ev_child_stop (EV_P_ ev_child *w)
 #define DEF_STAT_INTERVAL 5.0074891
 #define MIN_STAT_INTERVAL 0.1074891
 
+void noinline stat_timer_cb (EV_P_ ev_timer *w_, int revents);
+
+#if EV_USE_INOTIFY
+# define EV_INOTIFY_BUFSIZE ((PATH_MAX + sizeof (struct inotify_event)) + 2048)
+
+static void noinline
+infy_add (EV_P_ ev_stat *w)
+{
+  w->wd = inotify_add_watch (fs_fd, w->path, IN_ATTRIB | IN_DELETE_SELF | IN_MOVE_SELF | IN_MODIFY | IN_DONT_FOLLOW | IN_MASK_ADD);
+
+  if (w->wd < 0)
+    {
+      ev_timer_start (EV_A_ &w->timer); /* this is not race-free, so we still need to recheck periodically */
+
+      /* monitor some parent directory for speedup hints */
+      if (errno == ENOENT || errno == EACCES)
+        {
+          char path [PATH_MAX];
+          strcpy (path, w->path);
+
+          do
+            {
+              int mask = IN_MASK_ADD | IN_DELETE_SELF | IN_MOVE_SELF
+                       | (errno == EACCES ? IN_ATTRIB : IN_CREATE | IN_MOVED_TO);
+
+              char *pend = strrchr (path, '/');
+
+              if (!pend)
+                break; /* whoops, no '/', complain to your admin */
+
+              *pend = 0;
+              w->wd = inotify_add_watch (fs_fd, path, IN_DELETE_SELF | IN_CREATE | IN_MOVED_TO | IN_MASK_ADD);
+            } 
+          while (w->wd < 0 && (errno == ENOENT || errno == EACCES));
+        }
+    }
+  else
+    ev_timer_stop (EV_A_ &w->timer); /* we can watch this in a race-free way */
+
+  if (w->wd >= 0)
+    wlist_add (&fs_hash [w->wd & (EV_INOTIFY_HASHSIZE - 1)].head, (WL)w);
+}
+
+static void noinline
+infy_del (EV_P_ ev_stat *w)
+{
+  WL w_;
+  int slot;
+  int wd = w->wd;
+
+  if (wd < 0)
+    return;
+
+  w->wd = -2;
+  slot = wd & (EV_INOTIFY_HASHSIZE - 1);
+  wlist_del (&fs_hash [slot].head, (WL)w);
+
+  /* remove this watcher, if others are watching it, they will rearm */
+  inotify_rm_watch (fs_fd, wd);
+}
+
+static void noinline
+infy_wd (EV_P_ int slot, int wd, struct inotify_event *ev)
+{
+  if (slot < 0)
+    /* overflow, need to check for all hahs slots */
+    for (slot = 0; slot < EV_INOTIFY_HASHSIZE; ++slot)
+      infy_wd (EV_A_ slot, wd, ev);
+  else
+    {
+      WL w_;
+
+      for (w_ = fs_hash [slot & (EV_INOTIFY_HASHSIZE - 1)].head; w_; )
+        {
+          ev_stat *w = (ev_stat *)w_;
+          w_ = w_->next; /* lets us remove this watcher and all before it */
+
+          if (w->wd == wd || wd == -1)
+            {
+              if (ev->mask & (IN_IGNORED | IN_UNMOUNT | IN_DELETE_SELF))
+                {
+                  w->wd = -1;
+                  infy_add (EV_A_ w); /* re-add, no matter what */
+                }
+
+              stat_timer_cb (EV_P_ &w->timer, 0);
+            }
+        }
+    }
+}
+
+static void
+infy_cb (EV_P_ ev_io *w, int revents)
+{
+  char buf [EV_INOTIFY_BUFSIZE];
+  struct inotify_event *ev = (struct inotify_event *)buf;
+  int ofs;
+  int len = read (fs_fd, buf, sizeof (buf));
+
+  for (ofs = 0; ofs < len; ofs += sizeof (struct inotify_event) + ev->len)
+    infy_wd (EV_A_ ev->wd, ev->wd, ev);
+}
+
+void inline_size
+infy_init (EV_P)
+{
+  if (fs_fd != -2)
+    return;
+
+  fs_fd = inotify_init ();
+
+  if (fs_fd >= 0)
+    {
+      ev_io_init (&fs_w, infy_cb, fs_fd, EV_READ);
+      ev_set_priority (&fs_w, EV_MAXPRI);
+      ev_io_start (EV_A_ &fs_w);
+    }
+}
+
+#endif
+
 void
 ev_stat_stat (EV_P_ ev_stat *w)
 {
@@ -1676,7 +1845,7 @@ ev_stat_stat (EV_P_ ev_stat *w)
     w->attr.st_nlink = 1;
 }
 
-static void
+void noinline
 stat_timer_cb (EV_P_ ev_timer *w_, int revents)
 {
   ev_stat *w = (ev_stat *)(((char *)w_) - offsetof (ev_stat, timer));
@@ -1687,7 +1856,15 @@ stat_timer_cb (EV_P_ ev_timer *w_, int revents)
   ev_stat_stat (EV_A_ w);
 
   if (memcmp (&w->prev, &w->attr, sizeof (ev_statdata)))
-    ev_feed_event (EV_A_ w, EV_STAT);
+    {
+      #if EV_USE_INOTIFY
+        infy_del (EV_A_ w);
+        infy_add (EV_A_ w);
+        ev_stat_stat (EV_A_ w); /* avoid race... */
+      #endif
+
+      ev_feed_event (EV_A_ w, EV_STAT);
+    }
 }
 
 void
@@ -1707,7 +1884,15 @@ ev_stat_start (EV_P_ ev_stat *w)
 
   ev_timer_init (&w->timer, stat_timer_cb, w->interval, w->interval);
   ev_set_priority (&w->timer, ev_priority (w));
-  ev_timer_start (EV_A_ &w->timer);
+
+#if EV_USE_INOTIFY
+  infy_init (EV_A);
+
+  if (fs_fd >= 0)
+    infy_add (EV_A_ w);
+  else
+#endif
+    ev_timer_start (EV_A_ &w->timer);
 
   ev_start (EV_A_ (W)w, 1);
 }
@@ -1719,6 +1904,9 @@ ev_stat_stop (EV_P_ ev_stat *w)
   if (expect_false (!ev_is_active (w)))
     return;
 
+#if EV_USE_INOTIFY
+  infy_del (EV_A_ w);
+#endif
   ev_timer_stop (EV_A_ &w->timer);
 
   ev_stop (EV_A_ (W)w);
diff --git a/ev.h b/ev.h
index db79520..bd023a7 100644
--- a/ev.h
+++ b/ev.h
@@ -220,23 +220,25 @@ typedef struct ev_child
 
 #if EV_STAT_ENABLE
 /* st_nlink = 0 means missing file or other error */
-#ifdef _WIN32
+# ifdef _WIN32
 typedef struct _stati64 ev_statdata;
-#else
+# else
 typedef struct stat ev_statdata;
-#endif
+# endif
 
 /* invoked each time the stat data changes for a given path */
 /* revent EV_STAT */
 typedef struct ev_stat
 {
-  EV_WATCHER (ev_stat)
+  EV_WATCHER_LIST (ev_stat)
 
   ev_timer timer;     /* private */
   ev_tstamp interval; /* ro */
   const char *path;   /* ro */
   ev_statdata prev;   /* ro */
   ev_statdata attr;   /* ro */
+
+  int wd; /* wd for inotify, fd for kqueue */
 } ev_stat;
 #endif
 
@@ -426,7 +428,7 @@ void ev_once (EV_P_ int fd, int events, ev_tstamp timeout, void (*cb)(int revent
 #define ev_periodic_set(ev,at_,ival_,res_)  do { (ev)->at = (at_); (ev)->interval = (ival_); (ev)->reschedule_cb= (res_); } while (0)
 #define ev_signal_set(ev,signum_)           do { (ev)->signum = (signum_); } while (0)
 #define ev_child_set(ev,pid_)               do { (ev)->pid = (pid_); } while (0)
-#define ev_stat_set(ev,path_,interval_)     do { (ev)->path = (path_); (ev)->interval = (interval_); } while (0)
+#define ev_stat_set(ev,path_,interval_)     do { (ev)->path = (path_); (ev)->interval = (interval_); (ev)->wd = -2; } while (0)
 #define ev_idle_set(ev)                     /* nop, yes, this is a serious in-joke */
 #define ev_prepare_set(ev)                  /* nop, yes, this is a serious in-joke */
 #define ev_check_set(ev)                    /* nop, yes, this is a serious in-joke */
diff --git a/ev.pod b/ev.pod
index cac8de1..241fade 100644
--- a/ev.pod
+++ b/ev.pod
@@ -2016,6 +2016,12 @@ backend for Solaris 10 systems.
 
 reserved for future expansion, works like the USE symbols above.
 
+=item EV_USE_INOTIFY
+
+If defined to be C<1>, libev will compile in support for the Linux inotify
+interface to speed up C<ev_stat> watchers. Its actual availability will
+be detected at runtime.
+
 =item EV_H
 
 The name of the F<ev.h> header file used to include it. The default if
@@ -2080,7 +2086,15 @@ some inlining decisions, saves roughly 30% codesize of amd64.
 C<ev_child> watchers use a small hash table to distribute workload by
 pid. The default size is C<16> (or C<1> with C<EV_MINIMAL>), usually more
 than enough. If you need to manage thousands of children you might want to
-increase this value.
+increase this value (I<must> be a power of two).
+
+=item EV_INOTIFY_HASHSIZE
+
+C<ev_staz> watchers use a small hash table to distribute workload by
+inotify watch id. The default size is C<16> (or C<1> with C<EV_MINIMAL>),
+usually more than enough. If you need to manage thousands of C<ev_stat>
+watchers you might want to increase this value (I<must> be a power of
+two).
 
 =item EV_COMMON
 
@@ -2150,7 +2164,7 @@ documentation for C<ev_default_init>.
 
 =item Stopping check/prepare/idle watchers: O(1)
 
-=item Stopping an io/signal/child watcher: O(number_of_watchers_for_this_(fd/signal/pid % 16))
+=item Stopping an io/signal/child watcher: O(number_of_watchers_for_this_(fd/signal/pid % EV_PID_HASHSIZE))
 
 =item Finding the next timer per loop iteration: O(1)
 
diff --git a/ev_epoll.c b/ev_epoll.c
index 4edbecd..54a4ccb 100644
--- a/ev_epoll.c
+++ b/ev_epoll.c
@@ -79,7 +79,7 @@ epoll_poll (EV_P_ ev_tstamp timeout)
     }
 }
 
-static int
+int inline_size
 epoll_init (EV_P_ int flags)
 {
   backend_fd = epoll_create (256);
@@ -99,15 +99,13 @@ epoll_init (EV_P_ int flags)
   return EVBACKEND_EPOLL;
 }
 
-static void
+void inline_size
 epoll_destroy (EV_P)
 {
-  close (backend_fd);
-
   ev_free (epoll_events);
 }
 
-static void
+void inline_size
 epoll_fork (EV_P)
 {
   close (backend_fd);
diff --git a/ev_kqueue.c b/ev_kqueue.c
index 7a34e9b..afd3f9c 100644
--- a/ev_kqueue.c
+++ b/ev_kqueue.c
@@ -36,7 +36,7 @@
 #include <string.h>
 #include <errno.h>
 
-static void
+void inline_speed
 kqueue_change (EV_P_ int fd, int filter, int flags, int fflags)
 {
   struct kevent *ke;
@@ -142,7 +142,7 @@ kqueue_poll (EV_P_ ev_tstamp timeout)
     }
 }
 
-static int
+int inline_size
 kqueue_init (EV_P_ int flags)
 {
   struct kevent ch, ev;
@@ -168,16 +168,14 @@ kqueue_init (EV_P_ int flags)
   return EVBACKEND_KQUEUE;
 }
 
-static void
+void inline_size
 kqueue_destroy (EV_P)
 {
-  close (backend_fd);
-
   ev_free (kqueue_events);
   ev_free (kqueue_changes);
 }
 
-static void
+void inline_size
 kqueue_fork (EV_P)
 {
   close (backend_fd);
diff --git a/ev_poll.c b/ev_poll.c
index 293958f..52fe42d 100644
--- a/ev_poll.c
+++ b/ev_poll.c
@@ -1,5 +1,5 @@
 /*
- * libev epoll fd activity backend
+ * libev poll fd activity backend
  *
  * Copyright (c) 2007 Marc Alexander Lehmann <libev@schmorp.de>
  * All rights reserved.
@@ -105,7 +105,7 @@ poll_poll (EV_P_ ev_tstamp timeout)
       );
 }
 
-static int
+int inline_size
 poll_init (EV_P_ int flags)
 {
   backend_fudge  = 1e-3; /* needed to compensate for select returning early, very conservative */
@@ -118,7 +118,7 @@ poll_init (EV_P_ int flags)
   return EVBACKEND_POLL;
 }
 
-static void
+void inline_size
 poll_destroy (EV_P)
 {
   ev_free (pollidxs);
diff --git a/ev_port.c b/ev_port.c
index e6555bb..db3bae5 100644
--- a/ev_port.c
+++ b/ev_port.c
@@ -108,7 +108,7 @@ port_poll (EV_P_ ev_tstamp timeout)
     }
 }
 
-static int
+int inline_size
 port_init (EV_P_ int flags)
 {
   /* Initalize the kernel queue */
@@ -127,15 +127,13 @@ port_init (EV_P_ int flags)
   return EVBACKEND_PORT;
 }
 
-static void
+void inline_size
 port_destroy (EV_P)
 {
-  close (backend_fd);
-
   ev_free (port_events);
 }
 
-static void
+void inline_size
 port_fork (EV_P)
 {
   close (backend_fd);
diff --git a/ev_select.c b/ev_select.c
index 38a2398..1df0953 100644
--- a/ev_select.c
+++ b/ev_select.c
@@ -200,7 +200,7 @@ select_poll (EV_P_ ev_tstamp timeout)
 #endif
 }
 
-static int
+int inline_size
 select_init (EV_P_ int flags)
 {
   backend_fudge  = 1e-2; /* needed to compensate for select returning early, very conservative */
@@ -224,7 +224,7 @@ select_init (EV_P_ int flags)
   return EVBACKEND_SELECT;
 }
 
-static void
+void inline_size
 select_destroy (EV_P)
 {
   ev_free (vec_ri);
diff --git a/ev_vars.h b/ev_vars.h
index c1bfaa4..3750492 100644
--- a/ev_vars.h
+++ b/ev_vars.h
@@ -2,7 +2,7 @@
 
 VARx(ev_tstamp, now_floor) /* last time we refreshed rt_time */
 VARx(ev_tstamp, mn_now)    /* monotonic clock "now" */
-VARx(ev_tstamp, rtmn_diff)      /* difference realtime - monotonic time */
+VARx(ev_tstamp, rtmn_diff) /* difference realtime - monotonic time */
 VARx(int, backend)
 
 VARx(ev_tstamp, backend_fudge) /* assumed typical timer resolution */
@@ -11,7 +11,7 @@ VAR (backend_poll  , void (*backend_poll)(EV_P_ ev_tstamp timeout))
 VARx(int, backend_fd)
 
 VARx(int, postfork)  /* true if we need to recreate kernel state after fork */
-VARx(int, activecnt) /* number of active events */
+VARx(int, activecnt) /* total number of active events ("refcount") */
 
 #if EV_USE_SELECT || EV_GENWRAP
 VARx(void *, vec_ri)
@@ -86,5 +86,11 @@ VARx(int, forkmax)
 VARx(int, forkcnt)
 #endif
 
+#if EV_USE_INOTIFY || EV_GENWRAP
+VARx(int, fs_fd)
+VARx(ev_io, fs_w);
+VAR (fs_hash, ANFS fs_hash [EV_INOTIFY_HASHSIZE])
+#endif
+
 #undef VARx
 
diff --git a/ev_wrap.h b/ev_wrap.h
index 37812ca..f3942f9 100644
--- a/ev_wrap.h
+++ b/ev_wrap.h
@@ -54,3 +54,6 @@
 #define forks ((loop)->forks)
 #define forkmax ((loop)->forkmax)
 #define forkcnt ((loop)->forkcnt)
+#define fs_fd ((loop)->fs_fd)
+#define fs_w ((loop)->fs_w);
+#define fs_hash ((loop)->fs_hash)
diff --git a/libev.m4 b/libev.m4
index 55f94f4..39c28f4 100644
--- a/libev.m4
+++ b/libev.m4
@@ -2,14 +2,14 @@ dnl this file is part of libev, do not make local modifications
 dnl http://software.schmorp.de/pkg/libev
 
 dnl libev support 
-AC_CHECK_HEADERS(sys/epoll.h sys/event.h sys/queue.h port.h poll.h sys/select.h) 
+AC_CHECK_HEADERS(sys/inotify.h sys/epoll.h sys/event.h sys/queue.h port.h poll.h sys/select.h) 
  
-AC_CHECK_FUNCS(epoll_ctl kqueue port_create poll select)
+AC_CHECK_FUNCS(inotify_init epoll_ctl kqueue port_create poll select)
  
 AC_CHECK_FUNC(clock_gettime, [], [ 
    if test -z "$LIBEV_M4_AVOID_LIBRT"; then
-     AC_CHECK_LIB(rt, clock_gettime) 
-     AC_CHECK_FUNCS(clock_gettime)
+      AC_CHECK_LIB(rt, clock_gettime) 
+      AC_CHECK_FUNCS(clock_gettime)
    fi
 ])