r21077: A step to minimize the diff later: This pulls in unmodified files from Samba4,

not compiled yet. This way the modifications become visible later.

ntvfs/common/notify.c -> smbd/notify_internal.c
ntvfs/sysdep/inotify.c -> smbd/notify_inotify.c

Naturally I had to disable notify again :-)

Volker
(This used to be commit cdb7d582b7397faa5926bff5783da7fef4209948)

4 files changed, 988 insertions, 177 deletions

diff --git a/source3/Makefile.in b/source3/Makefile.in
index 4b590119e9..c21cea0aec 100644
--- a/source3/Makefile.in
+++ b/source3/Makefile.in
@@ -409,7 +409,7 @@ PROFILES_OBJ = utils/profiles.o \
 OPLOCK_OBJ = smbd/oplock.o smbd/oplock_irix.o smbd/oplock_linux.o
 
 NOTIFY_OBJ = smbd/notify.o smbd/notify_hash.o smbd/notify_kernel.o \
-	smbd/notify_fam.o smbd/notify_inotify.o
+	smbd/notify_fam.o
 
 VFS_DEFAULT_OBJ = modules/vfs_default.o
 VFS_AUDIT_OBJ = modules/vfs_audit.o
diff --git a/source3/smbd/notify.c b/source3/smbd/notify.c
index d920633b0a..4c82b9fca8 100644
--- a/source3/smbd/notify.c
+++ b/source3/smbd/notify.c
@@ -529,11 +529,6 @@ BOOL init_change_notify(void)
 {
 	cnotify = NULL;
 
-#if HAVE_INOTIFY
-	if ((cnotify == NULL) && lp_kernel_change_notify()) {
-		cnotify = inotify_notify_init(smbd_event_context());
-	}
-#endif
 #if HAVE_KERNEL_CHANGE_NOTIFY
 	if (cnotify == NULL && lp_kernel_change_notify())
 		cnotify = kernel_notify_init(smbd_event_context());
diff --git a/source3/smbd/notify_inotify.c b/source3/smbd/notify_inotify.c
index 64ea62ff74..8bb0096dcc 100644
--- a/source3/smbd/notify_inotify.c
+++ b/source3/smbd/notify_inotify.c
@@ -1,27 +1,33 @@
+/* 
+   Unix SMB/CIFS implementation.
+
+   Copyright (C) Andrew Tridgell 2006
+   
+   This program is free software; you can redistribute it and/or modify
+   it under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 2 of the License, or
+   (at your option) any later version.
+   
+   This program is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+   GNU General Public License for more details.
+   
+   You should have received a copy of the GNU General Public License
+   along with this program; if not, write to the Free Software
+   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+*/
+
 /*
- * inotify change notify support
- *
- * Copyright (c) Andrew Tridgell 2006
- * Copyright (c) Volker Lendecke 2007
- *
- * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License as published by
- * the Free Software Foundation; either version 2 of the License, or
- * (at your option) any later version.
- *
- * This program is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU General Public License for more details.
- *
- * You should have received a copy of the GNU General Public License
- * along with this program; if not, write to the Free Software
- * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
- */
+  notify implementation using inotify
+*/
 
 #include "includes.h"
-
-#ifdef HAVE_INOTIFY
+#include "system/filesys.h"
+#include "ntvfs/sysdep/sys_notify.h"
+#include "lib/events/events.h"
+#include "lib/util/dlinklist.h"
+#include "libcli/raw/smb.h"
 
 #include <linux/inotify.h>
 #include <asm/unistd.h>
@@ -46,218 +52,367 @@ static int inotify_rm_watch(int fd, int wd)
 }
 #endif
 
-struct inotify_ctx {
-	struct inotify_ctx *prev, *next;
-	int wd;
-	files_struct *fsp;
-};
 
-static struct inotify_ctx *inotify_list;
+/* older glibc headers don't have these defines either */
+#ifndef IN_ONLYDIR
+#define IN_ONLYDIR 0x01000000
+#endif
+#ifndef IN_MASK_ADD
+#define IN_MASK_ADD 0x20000000
+#endif
 
-static int inotify_watch_fd;
+struct inotify_private {
+	struct sys_notify_context *ctx;
+	int fd;
+	struct watch_context *watches;
+};
 
-/*
-  map from a change notify mask to a inotify mask. Remove any bits
-  which we can handle
-*/
-static const struct {
-	uint32_t notify_mask;
-	uint32_t inotify_mask;
-} inotify_mapping[] = {
-	{FILE_NOTIFY_CHANGE_FILE_NAME,
-	 IN_CREATE|IN_DELETE|IN_MOVED_FROM|IN_MOVED_TO},
-	{FILE_NOTIFY_CHANGE_DIR_NAME,
-	 IN_CREATE|IN_DELETE|IN_MOVED_FROM|IN_MOVED_TO},
-	{FILE_NOTIFY_CHANGE_ATTRIBUTES,
-	 IN_ATTRIB|IN_MOVED_TO|IN_MOVED_FROM|IN_MODIFY},
-	{FILE_NOTIFY_CHANGE_LAST_WRITE,  IN_ATTRIB},
-	{FILE_NOTIFY_CHANGE_LAST_ACCESS, IN_ATTRIB},
-	{FILE_NOTIFY_CHANGE_EA,          IN_ATTRIB},
-	{FILE_NOTIFY_CHANGE_SECURITY,    IN_ATTRIB}
+struct watch_context {
+	struct watch_context *next, *prev;
+	struct inotify_private *in;
+	int wd;
+	sys_notify_callback_t callback;
+	void *private_data;
+	uint32_t mask; /* the inotify mask */
+	uint32_t filter; /* the windows completion filter */
+	const char *path;
 };
 
-static uint32_t inotify_map(uint32 *filter)
-{
-	int i;
-	uint32_t out=0;
-	for (i=0;i<ARRAY_SIZE(inotify_mapping);i++) {
-		if (inotify_mapping[i].notify_mask & *filter) {
-			out |= inotify_mapping[i].inotify_mask;
-			*filter &= ~inotify_mapping[i].notify_mask;
-		}
-	}
-	return out;
-}
 
-static int inotify_ctx_destructor(struct inotify_ctx *ctx)
+/*
+  destroy the inotify private context
+*/
+static int inotify_destructor(struct inotify_private *in)
 {
-	if (inotify_rm_watch(inotify_watch_fd, ctx->wd) == -1) {
-		DEBUG(0, ("inotify_rm_watch failed: %s\n", strerror(errno)));
-	}
-	DLIST_REMOVE(inotify_list, ctx);
+	close(in->fd);
 	return 0;
 }
 
-static void *inotify_add(TALLOC_CTX *mem_ctx, 
-			 struct event_context *event_ctx,
-			 files_struct *fsp, uint32 *pfilter)
-{
-	struct inotify_ctx *ctx;
-	uint32_t inotify_mask;
-	pstring fullpath;
-	uint32 filter;
-	
-	filter = *pfilter;
-	if ((filter & (FILE_NOTIFY_CHANGE_FILE
-		       |FILE_NOTIFY_CHANGE_DIR_NAME)) == 0) {
-		/*
-		 * This first implementation only looks at create/delete
-		 */
-		return NULL;
-	}
 
-	inotify_mask = inotify_map(&filter);
-
-	if (inotify_mask == 0) {
-		DEBUG(10, ("inotify_mask == 0, nothing to do\n"));
-		return NULL;
+/*
+  see if a particular event from inotify really does match a requested
+  notify event in SMB
+*/
+static BOOL filter_match(struct watch_context *w, struct inotify_event *e)
+{
+	if ((e->mask & w->mask) == 0) {
+		/* this happens because inotify_add_watch() coalesces watches on the same
+		   path, oring their masks together */
+		return False;
 	}
 
-	pstrcpy(fullpath, fsp->fsp_name);
-	if (!canonicalize_path(fsp->conn, fullpath)) {
-		DEBUG(0, ("failed to canonicalize path '%s'\n", fullpath));
-		return NULL;
+	/* SMB separates the filters for files and directories */
+	if (e->mask & IN_ISDIR) {
+		if ((w->filter & FILE_NOTIFY_CHANGE_DIR_NAME) == 0) {
+			return False;
+		}
+	} else {
+		if ((e->mask & IN_ATTRIB) &&
+		    (w->filter & (FILE_NOTIFY_CHANGE_ATTRIBUTES|
+				  FILE_NOTIFY_CHANGE_LAST_WRITE|
+				  FILE_NOTIFY_CHANGE_LAST_ACCESS|
+				  FILE_NOTIFY_CHANGE_EA|
+				  FILE_NOTIFY_CHANGE_SECURITY))) {
+			return True;
+		}
+		if ((e->mask & IN_MODIFY) && 
+		    (w->filter & FILE_NOTIFY_CHANGE_ATTRIBUTES)) {
+			return True;
+		}
+		if ((w->filter & FILE_NOTIFY_CHANGE_FILE_NAME) == 0) {
+			return False;
+		}
 	}
 
-	if (*fullpath != '/') {
-		DEBUG(0, ("canonicalized path '%s' into `%s`\n", fsp->fsp_name,
-			  fullpath));
-		DEBUGADD(0, ("but expected an absolute path\n"));
-		return NULL;
-	}
+	return True;
+}
 	
-	if (!(ctx = TALLOC_P(mem_ctx, struct inotify_ctx))) {
-		DEBUG(0, ("talloc failed\n"));
-		return NULL;
-	}
-
-	ctx->fsp = fsp;
-	ctx->wd = inotify_add_watch(inotify_watch_fd, fullpath, inotify_mask);
 
-	if (ctx->wd == -1) {
-		DEBUG(5, ("inotify_add_watch failed: %s\n", strerror(errno)));
-		TALLOC_FREE(ctx);
-		return NULL;
-	}
-
-	DLIST_ADD(inotify_list, ctx);
-	talloc_set_destructor(ctx, inotify_ctx_destructor);
 
-	*pfilter = filter;
-	return ctx;
-}
-
-static void inotify_dispatch(struct inotify_event *e)
+/*
+  dispatch one inotify event
+  
+  the cookies are used to correctly handle renames
+*/
+static void inotify_dispatch(struct inotify_private *in, 
+			     struct inotify_event *e, 
+			     uint32_t prev_cookie,
+			     struct inotify_event *e2)
 {
-	struct inotify_ctx *ctx;
+	struct watch_context *w, *next;
+	struct notify_event ne;
+
+	/* ignore extraneous events, such as unmount and IN_IGNORED events */
+	if ((e->mask & (IN_ATTRIB|IN_MODIFY|IN_CREATE|IN_DELETE|
+			IN_MOVED_FROM|IN_MOVED_TO)) == 0) {
+		return;
+	}
 
-	for (ctx = inotify_list; ctx; ctx = ctx->next) {
-		if (ctx->wd == e->wd) {
-			break;
+	/* map the inotify mask to a action. This gets complicated for
+	   renames */
+	if (e->mask & IN_CREATE) {
+		ne.action = NOTIFY_ACTION_ADDED;
+	} else if (e->mask & IN_DELETE) {
+		ne.action = NOTIFY_ACTION_REMOVED;
+	} else if (e->mask & IN_MOVED_FROM) {
+		if (e2 != NULL && e2->cookie == e->cookie) {
+			ne.action = NOTIFY_ACTION_OLD_NAME;
+		} else {
+			ne.action = NOTIFY_ACTION_REMOVED;
+		}
+	} else if (e->mask & IN_MOVED_TO) {
+		if (e->cookie == prev_cookie) {
+			ne.action = NOTIFY_ACTION_NEW_NAME;
+		} else {
+			ne.action = NOTIFY_ACTION_ADDED;
 		}
+	} else {
+		ne.action = NOTIFY_ACTION_MODIFIED;
 	}
+	ne.path = e->name;
 
-	if (ctx == NULL) {
-		/* not found */
-		return;
+	/* find any watches that have this watch descriptor */
+	for (w=in->watches;w;w=next) {
+		next = w->next;
+		if (w->wd == e->wd && filter_match(w, e)) {
+			w->callback(in->ctx, w->private_data, &ne);
+		}
 	}
 
-	if (e->mask & IN_CREATE) {
-		notify_fsp(ctx->fsp, NOTIFY_ACTION_ADDED, e->name);
+	/* SMB expects a file rename to generate three events, two for
+	   the rename and the other for a modify of the
+	   destination. Strange! */
+	if (ne.action != NOTIFY_ACTION_NEW_NAME ||
+	    (e->mask & IN_ISDIR) != 0) {
+		return;
 	}
 
-	if (e->mask & IN_DELETE) {
-		notify_fsp(ctx->fsp, NOTIFY_ACTION_REMOVED, e->name);
+	ne.action = NOTIFY_ACTION_MODIFIED;
+	e->mask = IN_ATTRIB;
+	
+	for (w=in->watches;w;w=next) {
+		next = w->next;
+		if (w->wd == e->wd && filter_match(w, e) &&
+		    !(w->filter & FILE_NOTIFY_CHANGE_CREATION)) {
+			w->callback(in->ctx, w->private_data, &ne);
+		}
 	}
 }
 
-static void inotify_callback(struct event_context *event_ctx,
-			     struct fd_event *event,
-			     uint16 flags,
-			     void *private_data)
+/*
+  called when the kernel has some events for us
+*/
+static void inotify_handler(struct event_context *ev, struct fd_event *fde,
+			    uint16_t flags, void *private_data)
 {
-	char *buf, *p;
-	int bufsize;
+	struct inotify_private *in = talloc_get_type(private_data,
+						     struct inotify_private);
+	int bufsize = 0;
+	struct inotify_event *e0, *e;
+	uint32_t prev_cookie=0;
 
 	/*
 	  we must use FIONREAD as we cannot predict the length of the
 	  filenames, and thus can't know how much to allocate
 	  otherwise
 	*/
-	if ((ioctl(inotify_watch_fd, FIONREAD, &bufsize) != 0)
-	    || (bufsize == 0)) {
+	if (ioctl(in->fd, FIONREAD, &bufsize) != 0 || 
+	    bufsize == 0) {
 		DEBUG(0,("No data on inotify fd?!\n"));
 		return;
 	}
 
-	if (!(buf = SMB_MALLOC_ARRAY(char, bufsize))) {
-		DEBUG(0, ("malloc failed\n"));
-		return;
-	}
+	e0 = e = talloc_size(in, bufsize);
+	if (e == NULL) return;
 
-	if (read(inotify_watch_fd, buf, bufsize) != bufsize) {
+	if (read(in->fd, e0, bufsize) != bufsize) {
 		DEBUG(0,("Failed to read all inotify data\n"));
-		SAFE_FREE(buf);
+		talloc_free(e0);
 		return;
 	}
 
-	p = buf;
+	/* we can get more than one event in the buffer */
+	while (bufsize >= sizeof(*e)) {
+		struct inotify_event *e2 = NULL;
+		bufsize -= e->len + sizeof(*e);
+		if (bufsize >= sizeof(*e)) {
+			e2 = (struct inotify_event *)(e->len + sizeof(*e) + (char *)e);
+		}
+		inotify_dispatch(in, e, prev_cookie, e2);
+		prev_cookie = e->cookie;
+		e = e2;
+	}
 
-	while (bufsize > sizeof(struct inotify_event)) {
-		struct inotify_event *iev = (struct inotify_event *)p;
-		size_t len = sizeof(struct inotify_event) + iev->len;
+	talloc_free(e0);
+}
 
-		if ((len > bufsize)
-		    || ((iev->len != 0) && (iev->name[iev->len-1] != '\0'))) {
-			smb_panic("invalid inotify reply\n");
-		}
+/*
+  setup the inotify handle - called the first time a watch is added on
+  this context
+*/
+static NTSTATUS inotify_setup(struct sys_notify_context *ctx)
+{
+	struct inotify_private *in;
 
-		inotify_dispatch(iev);
+	if (!lp_parm_bool(-1, "notify", "inotify", True)) {
+		return NT_STATUS_INVALID_SYSTEM_SERVICE;
+	}
 
-		p += len;
-		bufsize -= len;
+	in = talloc(ctx, struct inotify_private);
+	NT_STATUS_HAVE_NO_MEMORY(in);
+	in->fd = inotify_init();
+	if (in->fd == -1) {
+		DEBUG(0,("Failed to init inotify - %s\n", strerror(errno)));
+		talloc_free(in);
+		return map_nt_error_from_unix(errno);
 	}
+	in->ctx = ctx;
+	in->watches = NULL;
+
+	ctx->private_data = in;
+	talloc_set_destructor(in, inotify_destructor);
 
-	SAFE_FREE(buf);
-	return;
+	/* add a event waiting for the inotify fd to be readable */
+	event_add_fd(ctx->ev, in, in->fd, EVENT_FD_READ, inotify_handler, in);
+	
+	return NT_STATUS_OK;
 }
 
-static struct cnotify_fns inotify_fns =
-{
-    inotify_add,
+
+/*
+  map from a change notify mask to a inotify mask. Remove any bits
+  which we can handle
+*/
+static const struct {
+	uint32_t notify_mask;
+	uint32_t inotify_mask;
+} inotify_mapping[] = {
+	{FILE_NOTIFY_CHANGE_FILE_NAME,   IN_CREATE|IN_DELETE|IN_MOVED_FROM|IN_MOVED_TO},
+	{FILE_NOTIFY_CHANGE_DIR_NAME,    IN_CREATE|IN_DELETE|IN_MOVED_FROM|IN_MOVED_TO},
+	{FILE_NOTIFY_CHANGE_ATTRIBUTES,  IN_ATTRIB|IN_MOVED_TO|IN_MOVED_FROM|IN_MODIFY},
+	{FILE_NOTIFY_CHANGE_LAST_WRITE,  IN_ATTRIB},
+	{FILE_NOTIFY_CHANGE_LAST_ACCESS, IN_ATTRIB},
+	{FILE_NOTIFY_CHANGE_EA,          IN_ATTRIB},
+	{FILE_NOTIFY_CHANGE_SECURITY,    IN_ATTRIB}
 };
 
-struct cnotify_fns *inotify_notify_init(struct event_context *event_ctx)
+static uint32_t inotify_map(struct notify_entry *e)
+{
+	int i;
+	uint32_t out=0;
+	for (i=0;i<ARRAY_SIZE(inotify_mapping);i++) {
+		if (inotify_mapping[i].notify_mask & e->filter) {
+			out |= inotify_mapping[i].inotify_mask;
+			e->filter &= ~inotify_mapping[i].notify_mask;
+		}
+	}
+	return out;
+}
+
+/*
+  destroy a watch
+*/
+static int watch_destructor(struct watch_context *w)
 {
-	inotify_watch_fd = inotify_init();
+	struct inotify_private *in = w->in;
+	int wd = w->wd;
+	DLIST_REMOVE(w->in->watches, w);
+
+	/* only rm the watch if its the last one with this wd */
+	for (w=in->watches;w;w=w->next) {
+		if (w->wd == wd) break;
+	}
+	if (w == NULL) {
+		inotify_rm_watch(in->fd, wd);
+	}
+	return 0;
+}
+
+
+/*
+  add a watch. The watch is removed when the caller calls
+  talloc_free() on *handle
+*/
+static NTSTATUS inotify_watch(struct sys_notify_context *ctx,
+			      struct notify_entry *e,
+			      sys_notify_callback_t callback,
+			      void *private_data, 
+			      void *handle_p)
+{
+	struct inotify_private *in;
+	int wd;
+	uint32_t mask;
+	struct watch_context *w;
+	uint32_t filter = e->filter;
+	void **handle = (void **)handle_p;
+
+	/* maybe setup the inotify fd */
+	if (ctx->private_data == NULL) {
+		NTSTATUS status;
+		status = inotify_setup(ctx);
+		NT_STATUS_NOT_OK_RETURN(status);
+	}
 
-	DEBUG(10, ("inotify_notify_init called\n"));
+	in = talloc_get_type(ctx->private_data, struct inotify_private);
 
-	if (inotify_watch_fd == -1) {
-		DEBUG(0, ("inotify_init failed: %s\n", strerror(errno)));
-		return NULL;
+	mask = inotify_map(e);
+	if (mask == 0) {
+		/* this filter can't be handled by inotify */
+		return NT_STATUS_INVALID_PARAMETER;
 	}
 
-	if (event_add_fd(event_ctx, NULL, inotify_watch_fd,
-			 EVENT_FD_READ, inotify_callback,
-			 NULL) == NULL) {
-		DEBUG(0, ("event_add_fd failed\n"));
-		close(inotify_watch_fd);
-		inotify_watch_fd = -1;
-		return NULL;
+	/* using IN_MASK_ADD allows us to cope with inotify() returning the same
+	   watch descriptor for muliple watches on the same path */
+	mask |= (IN_MASK_ADD | IN_ONLYDIR);
+
+	/* get a new watch descriptor for this path */
+	wd = inotify_add_watch(in->fd, e->path, mask);
+	if (wd == -1) {
+		e->filter = filter;
+		return map_nt_error_from_unix(errno);
 	}
 
-	return &inotify_fns;
+	w = talloc(in, struct watch_context);
+	if (w == NULL) {
+		inotify_rm_watch(in->fd, wd);
+		e->filter = filter;
+		return NT_STATUS_NO_MEMORY;
+	}
+
+	w->in = in;
+	w->wd = wd;
+	w->callback = callback;
+	w->private_data = private_data;
+	w->mask = mask;
+	w->filter = filter;
+	w->path = talloc_strdup(w, e->path);
+	if (w->path == NULL) {
+		inotify_rm_watch(in->fd, wd);
+		e->filter = filter;
+		return NT_STATUS_NO_MEMORY;
+	}
+
+	(*handle) = w;
+
+	DLIST_ADD(in->watches, w);
+
+	/* the caller frees the handle to stop watching */
+	talloc_set_destructor(w, watch_destructor);
+	
+	return NT_STATUS_OK;
 }
 
-#endif
+
+static struct sys_notify_backend inotify = {
+	.name = "inotify",
+	.notify_watch = inotify_watch
+};
+
+/*
+  initialialise the inotify module
+ */
+NTSTATUS sys_notify_inotify_init(void)
+{
+	/* register ourselves as a system inotify module */
+	return sys_notify_register(&inotify);
+}
diff --git a/source3/smbd/notify_internal.c b/source3/smbd/notify_internal.c
new file mode 100644
index 0000000000..91fa8a1d78
--- /dev/null
+++ b/source3/smbd/notify_internal.c
@@ -0,0 +1,661 @@
+/* 
+   Unix SMB/CIFS implementation.
+
+   Copyright (C) Andrew Tridgell 2006
+   
+   This program is free software; you can redistribute it and/or modify
+   it under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 2 of the License, or
+   (at your option) any later version.
+   
+   This program is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+   GNU General Public License for more details.
+   
+   You should have received a copy of the GNU General Public License
+   along with this program; if not, write to the Free Software
+   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+*/
+
+/*
+  this is the change notify database. It implements mechanisms for
+  storing current change notify waiters in a tdb, and checking if a
+  given event matches any of the stored notify waiiters.
+*/
+
+#include "includes.h"
+#include "system/filesys.h"
+#include "lib/tdb/include/tdb.h"
+#include "lib/util/util_tdb.h"
+#include "messaging/messaging.h"
+#include "db_wrap.h"
+#include "lib/messaging/irpc.h"
+#include "librpc/gen_ndr/ndr_notify.h"
+#include "lib/util/dlinklist.h"
+#include "ntvfs/common/ntvfs_common.h"
+#include "ntvfs/sysdep/sys_notify.h"
+#include "cluster/cluster.h"
+
+struct notify_context {
+	struct tdb_wrap *w;
+	struct server_id server;
+	struct messaging_context *messaging_ctx;
+	struct notify_list *list;
+	struct notify_array *array;
+	int seqnum;
+	struct sys_notify_context *sys_notify_ctx;
+};
+
+
+struct notify_list {
+	struct notify_list *next, *prev;
+	void *private_data;
+	void (*callback)(void *, const struct notify_event *);
+	void *sys_notify_handle;
+	int depth;
+};
+
+#define NOTIFY_KEY "notify array"
+
+#define NOTIFY_ENABLE		"notify:enable"
+#define NOTIFY_ENABLE_DEFAULT	True
+
+static NTSTATUS notify_remove_all(struct notify_context *notify);
+static void notify_handler(struct messaging_context *msg_ctx, void *private_data, 
+			   uint32_t msg_type, struct server_id server_id, DATA_BLOB *data);
+
+/*
+  destroy the notify context
+*/
+static int notify_destructor(struct notify_context *notify)
+{
+	messaging_deregister(notify->messaging_ctx, MSG_PVFS_NOTIFY, notify);
+	notify_remove_all(notify);
+	return 0;
+}
+
+/*
+  Open up the notify.tdb database. You should close it down using
+  talloc_free(). We need the messaging_ctx to allow for notifications
+  via internal messages
+*/
+struct notify_context *notify_init(TALLOC_CTX *mem_ctx, struct server_id server, 
+				   struct messaging_context *messaging_ctx,
+				   struct event_context *ev,
+				   struct share_config *scfg)
+{
+	struct notify_context *notify;
+
+	if (share_bool_option(scfg, NOTIFY_ENABLE, NOTIFY_ENABLE_DEFAULT) != True) {
+		return NULL;
+	}
+
+	notify = talloc(mem_ctx, struct notify_context);
+	if (notify == NULL) {
+		return NULL;
+	}
+
+	notify->w = cluster_tdb_tmp_open(notify, "notify.tdb", TDB_SEQNUM);
+	if (notify->w == NULL) {
+		talloc_free(notify);
+		return NULL;
+	}
+
+	notify->server = server;
+	notify->messaging_ctx = messaging_ctx;
+	notify->list = NULL;
+	notify->array = NULL;
+	notify->seqnum = tdb_get_seqnum(notify->w->tdb);
+
+	talloc_set_destructor(notify, notify_destructor);
+
+	/* register with the messaging subsystem for the notify
+	   message type */
+	messaging_register(notify->messaging_ctx, notify, 
+			   MSG_PVFS_NOTIFY, notify_handler);
+
+	notify->sys_notify_ctx = sys_notify_context_create(scfg, notify, ev);
+
+	return notify;
+}
+
+
+/*
+  lock the notify db
+*/
+static NTSTATUS notify_lock(struct notify_context *notify)
+{
+	if (tdb_lock_bystring(notify->w->tdb, NOTIFY_KEY) != 0) {
+		return NT_STATUS_INTERNAL_DB_CORRUPTION;
+	}
+	return NT_STATUS_OK;
+}
+
+/*
+  unlock the notify db
+*/
+static void notify_unlock(struct notify_context *notify)
+{
+	tdb_unlock_bystring(notify->w->tdb, NOTIFY_KEY);
+}
+
+/*
+  load the notify array
+*/
+static NTSTATUS notify_load(struct notify_context *notify)
+{
+	TDB_DATA dbuf;
+	DATA_BLOB blob;
+	NTSTATUS status;
+	int seqnum;
+
+	seqnum = tdb_get_seqnum(notify->w->tdb);
+
+	if (seqnum == notify->seqnum && notify->array != NULL) {
+		return NT_STATUS_OK;
+	}
+
+	notify->seqnum = seqnum;
+
+	talloc_free(notify->array);
+	notify->array = talloc_zero(notify, struct notify_array);
+	NT_STATUS_HAVE_NO_MEMORY(notify->array);
+
+	dbuf = tdb_fetch_bystring(notify->w->tdb, NOTIFY_KEY);
+	if (dbuf.dptr == NULL) {
+		return NT_STATUS_OK;
+	}
+
+	blob.data = dbuf.dptr;
+	blob.length = dbuf.dsize;
+
+	status = ndr_pull_struct_blob(&blob, notify->array, notify->array, 
+				      (ndr_pull_flags_fn_t)ndr_pull_notify_array);
+	free(dbuf.dptr);
+
+	return status;
+}
+
+/*
+  compare notify entries for sorting
+*/
+static int notify_compare(const void *p1, const void *p2)
+{
+	const struct notify_entry *e1 = p1, *e2 = p2;
+	return strcmp(e1->path, e2->path);
+}
+
+/*
+  save the notify array
+*/
+static NTSTATUS notify_save(struct notify_context *notify)
+{
+	TDB_DATA dbuf;
+	DATA_BLOB blob;
+	NTSTATUS status;
+	int ret;
+	TALLOC_CTX *tmp_ctx;
+
+	/* if possible, remove some depth arrays */
+	while (notify->array->num_depths > 0 &&
+	       notify->array->depth[notify->array->num_depths-1].num_entries == 0) {
+		notify->array->num_depths--;
+	}
+
+	/* we might just be able to delete the record */
+	if (notify->array->num_depths == 0) {
+		ret = tdb_delete_bystring(notify->w->tdb, NOTIFY_KEY);
+		if (ret != 0) {
+			return NT_STATUS_INTERNAL_DB_CORRUPTION;
+		}
+		return NT_STATUS_OK;
+	}
+
+	tmp_ctx = talloc_new(notify);
+	NT_STATUS_HAVE_NO_MEMORY(tmp_ctx);
+
+	status = ndr_push_struct_blob(&blob, tmp_ctx, notify->array, 
+				      (ndr_push_flags_fn_t)ndr_push_notify_array);
+	if (!NT_STATUS_IS_OK(status)) {
+		talloc_free(tmp_ctx);
+		return status;
+	}
+
+	dbuf.dptr = blob.data;
+	dbuf.dsize = blob.length;
+		
+	ret = tdb_store_bystring(notify->w->tdb, NOTIFY_KEY, dbuf, TDB_REPLACE);
+	talloc_free(tmp_ctx);
+	if (ret != 0) {
+		return NT_STATUS_INTERNAL_DB_CORRUPTION;
+	}
+
+	return NT_STATUS_OK;
+}
+
+
+/*
+  handle incoming notify messages
+*/
+static void notify_handler(struct messaging_context *msg_ctx, void *private_data, 
+			   uint32_t msg_type, struct server_id server_id, DATA_BLOB *data)
+{
+	struct notify_context *notify = talloc_get_type(private_data, struct notify_context);
+	NTSTATUS status;
+	struct notify_event ev;
+	TALLOC_CTX *tmp_ctx = talloc_new(notify);
+	struct notify_list *listel;
+
+	if (tmp_ctx == NULL) {
+		return;
+	}
+
+	status = ndr_pull_struct_blob(data, tmp_ctx, &ev, 
+				      (ndr_pull_flags_fn_t)ndr_pull_notify_event);
+	if (!NT_STATUS_IS_OK(status)) {
+		talloc_free(tmp_ctx);
+		return;
+	}
+
+	for (listel=notify->list;listel;listel=listel->next) {
+		if (listel->private_data == ev.private_data) {
+			listel->callback(listel->private_data, &ev);
+			break;
+		}
+	}
+
+	talloc_free(tmp_ctx);	
+}
+
+/*
+  callback from sys_notify telling us about changes from the OS
+*/
+static void sys_notify_callback(struct sys_notify_context *ctx, 
+				void *ptr, struct notify_event *ev)
+{
+	struct notify_list *listel = talloc_get_type(ptr, struct notify_list);
+	ev->private_data = listel;
+	listel->callback(listel->private_data, ev);
+}
+
+/*
+  add an entry to the notify array
+*/
+static NTSTATUS notify_add_array(struct notify_context *notify, struct notify_entry *e,
+				 void *private_data, int depth)
+{
+	int i;
+	struct notify_depth *d;
+	struct notify_entry *ee;
+
+	/* possibly expand the depths array */
+	if (depth >= notify->array->num_depths) {
+		d = talloc_realloc(notify->array, notify->array->depth, 
+				   struct notify_depth, depth+1);
+		NT_STATUS_HAVE_NO_MEMORY(d);
+		for (i=notify->array->num_depths;i<=depth;i++) {
+			ZERO_STRUCT(d[i]);
+		}
+		notify->array->depth = d;
+		notify->array->num_depths = depth+1;
+	}
+	d = &notify->array->depth[depth];
+
+	/* expand the entries array */
+	ee = talloc_realloc(notify->array->depth, d->entries, struct notify_entry,
+			    d->num_entries+1);
+	NT_STATUS_HAVE_NO_MEMORY(ee);
+	d->entries = ee;
+
+	d->entries[d->num_entries] = *e;
+	d->entries[d->num_entries].private_data = private_data;
+	d->entries[d->num_entries].server = notify->server;
+	d->entries[d->num_entries].path_len = strlen(e->path);
+	d->num_entries++;
+
+	d->max_mask |= e->filter;
+	d->max_mask_subdir |= e->subdir_filter;
+
+	if (d->num_entries > 1) {
+		qsort(d->entries, d->num_entries, sizeof(d->entries[0]), notify_compare);
+	}
+
+	/* recalculate the maximum masks */
+	d->max_mask = 0;
+	d->max_mask_subdir = 0;
+
+	for (i=0;i<d->num_entries;i++) {
+		d->max_mask |= d->entries[i].filter;
+		d->max_mask_subdir |= d->entries[i].subdir_filter;
+	}
+
+	return notify_save(notify);
+}
+
+/*
+  add a notify watch. This is called when a notify is first setup on a open
+  directory handle.
+*/
+NTSTATUS notify_add(struct notify_context *notify, struct notify_entry *e0,
+		    void (*callback)(void *, const struct notify_event *), 
+		    void *private_data)
+{
+	struct notify_entry e = *e0;
+	NTSTATUS status;
+	char *tmp_path = NULL;
+	struct notify_list *listel;
+	size_t len;
+	int depth;
+
+	/* see if change notify is enabled at all */
+	if (notify == NULL) {
+		return NT_STATUS_NOT_IMPLEMENTED;
+	}
+
+	status = notify_lock(notify);
+	NT_STATUS_NOT_OK_RETURN(status);
+
+	status = notify_load(notify);
+	if (!NT_STATUS_IS_OK(status)) {
+		goto done;
+	}
+
+	/* cope with /. on the end of the path */
+	len = strlen(e.path);
+	if (len > 1 && e.path[len-1] == '.' && e.path[len-2] == '/') {
+		tmp_path = talloc_strndup(notify, e.path, len-2);
+		if (tmp_path == NULL) {
+			status = NT_STATUS_NO_MEMORY;
+			goto done;
+		}
+		e.path = tmp_path;
+	}
+
+	depth = count_chars(e.path, '/');
+
+	listel = talloc_zero(notify, struct notify_list);
+	if (listel == NULL) {
+		status = NT_STATUS_NO_MEMORY;
+		goto done;
+	}
+
+	listel->private_data = private_data;
+	listel->callback = callback;
+	listel->depth = depth;
+	DLIST_ADD(notify->list, listel);
+
+	/* ignore failures from sys_notify */
+	if (notify->sys_notify_ctx != NULL) {
+		/*
+		  this call will modify e.filter and e.subdir_filter
+		  to remove bits handled by the backend
+		*/
+		status = sys_notify_watch(notify->sys_notify_ctx, &e,
+					  sys_notify_callback, listel, 
+					  &listel->sys_notify_handle);
+		if (NT_STATUS_IS_OK(status)) {
+			talloc_steal(listel, listel->sys_notify_handle);
+		}
+	}
+
+	/* if the system notify handler couldn't handle some of the
+	   filter bits, or couldn't handle a request for recursion
+	   then we need to install it in the array used for the
+	   intra-samba notify handling */
+	if (e.filter != 0 || e.subdir_filter != 0) {
+		status = notify_add_array(notify, &e, private_data, depth);
+	}
+
+done:
+	notify_unlock(notify);
+	talloc_free(tmp_path);
+
+	return status;
+}
+
+/*
+  remove a notify watch. Called when the directory handle is closed
+*/
+NTSTATUS notify_remove(struct notify_context *notify, void *private_data)
+{
+	NTSTATUS status;
+	struct notify_list *listel;
+	int i, depth;
+	struct notify_depth *d;
+
+	/* see if change notify is enabled at all */
+	if (notify == NULL) {
+		return NT_STATUS_NOT_IMPLEMENTED;
+	}
+
+	for (listel=notify->list;listel;listel=listel->next) {
+		if (listel->private_data == private_data) {
+			DLIST_REMOVE(notify->list, listel);
+			break;
+		}
+	}
+	if (listel == NULL) {
+		return NT_STATUS_OBJECT_NAME_NOT_FOUND;
+	}
+
+	depth = listel->depth;
+
+	talloc_free(listel);
+
+	status = notify_lock(notify);
+	NT_STATUS_NOT_OK_RETURN(status);
+
+	status = notify_load(notify);
+	if (!NT_STATUS_IS_OK(status)) {
+		notify_unlock(notify);
+		return status;
+	}
+
+	if (depth >= notify->array->num_depths) {
+		notify_unlock(notify);
+		return NT_STATUS_OBJECT_NAME_NOT_FOUND;
+	}
+
+	/* we only have to search at the depth of this element */
+	d = &notify->array->depth[depth];
+
+	for (i=0;i<d->num_entries;i++) {
+		if (private_data == d->entries[i].private_data &&
+		    cluster_id_equal(&notify->server, &d->entries[i].server)) {
+			break;
+		}
+	}
+	if (i == d->num_entries) {
+		notify_unlock(notify);
+		return NT_STATUS_OBJECT_NAME_NOT_FOUND;
+	}
+
+	if (i < d->num_entries-1) {
+		memmove(&d->entries[i], &d->entries[i+1], 
+			sizeof(d->entries[i])*(d->num_entries-(i+1)));
+	}
+	d->num_entries--;
+
+	status = notify_save(notify);
+
+	notify_unlock(notify);
+
+	return status;
+}
+
+/*
+  remove all notify watches for this messaging server
+*/
+static NTSTATUS notify_remove_all(struct notify_context *notify)
+{
+	NTSTATUS status;
+	int i, depth, del_count=0;
+
+	if (notify->list == NULL) {
+		return NT_STATUS_OK;
+	}
+
+	status = notify_lock(notify);
+	NT_STATUS_NOT_OK_RETURN(status);
+
+	status = notify_load(notify);
+	if (!NT_STATUS_IS_OK(status)) {
+		notify_unlock(notify);
+		return status;
+	}
+
+	/* we have to search for all entries across all depths, looking for matches
+	   for our server id */
+	for (depth=0;depth<notify->array->num_depths;depth++) {
+		struct notify_depth *d = &notify->array->depth[depth];
+		for (i=0;i<d->num_entries;i++) {
+			if (cluster_id_equal(&notify->server, &d->entries[i].server)) {
+				if (i < d->num_entries-1) {
+					memmove(&d->entries[i], &d->entries[i+1], 
+						sizeof(d->entries[i])*(d->num_entries-(i+1)));
+				}
+				i--;
+				d->num_entries--;
+				del_count++;
+			}
+		}
+	}
+
+	if (del_count > 0) {
+		status = notify_save(notify);
+	}
+
+	notify_unlock(notify);
+
+	return status;
+}
+
+
+/*
+  send a notify message to another messaging server
+*/
+static void notify_send(struct notify_context *notify, struct notify_entry *e,
+			const char *path, uint32_t action)
+{
+	struct notify_event ev;
+	DATA_BLOB data;
+	NTSTATUS status;
+	TALLOC_CTX *tmp_ctx;
+
+	ev.action = action;
+	ev.path = path;
+	ev.private_data = e->private_data;
+
+	tmp_ctx = talloc_new(notify);
+
+	status = ndr_push_struct_blob(&data, tmp_ctx, &ev, 
+				      (ndr_push_flags_fn_t)ndr_push_notify_event);
+	if (!NT_STATUS_IS_OK(status)) {
+		talloc_free(tmp_ctx);
+		return;
+	}
+
+	status = messaging_send(notify->messaging_ctx, e->server, 
+				MSG_PVFS_NOTIFY, &data);
+	talloc_free(tmp_ctx);
+}
+
+
+/*
+  trigger a notify message for anyone waiting on a matching event
+
+  This function is called a lot, and needs to be very fast. The unusual data structure
+  and traversal is designed to be fast in the average case, even for large numbers of
+  notifies
+*/
+void notify_trigger(struct notify_context *notify,
+		    uint32_t action, uint32_t filter, const char *path)
+{
+	NTSTATUS status;
+	int depth;
+	const char *p, *next_p;
+
+	/* see if change notify is enabled at all */
+	if (notify == NULL) {
+		return;
+	}
+
+	status = notify_load(notify);
+	if (!NT_STATUS_IS_OK(status)) {
+		return;
+	}
+
+	/* loop along the given path, working with each directory depth separately */
+	for (depth=0,p=path;
+	     p && depth < notify->array->num_depths;
+	     p=next_p,depth++) {
+		int p_len = p - path;
+		int min_i, max_i, i;
+		struct notify_depth *d = &notify->array->depth[depth];
+		next_p = strchr(p+1, '/');
+
+		/* see if there are any entries at this depth */
+		if (d->num_entries == 0) continue;
+		
+		/* try to skip based on the maximum mask. If next_p is
+		 NULL then we know it will be a 'this directory'
+		 match, otherwise it must be a subdir match */
+		if (next_p != NULL) {
+			if (0 == (filter & d->max_mask_subdir)) {
+				continue;
+			}
+		} else {
+			if (0 == (filter & d->max_mask)) {
+				continue;
+			}
+		}
+
+		/* we know there is an entry here worth looking
+		 for. Use a bisection search to find the first entry
+		 with a matching path */
+		min_i = 0;
+		max_i = d->num_entries-1;
+
+		while (min_i < max_i) {
+			struct notify_entry *e;
+			int cmp;
+			i = (min_i+max_i)/2;
+			e = &d->entries[i];
+			cmp = strncmp(path, e->path, p_len);
+			if (cmp == 0) {
+				if (p_len == e->path_len) {
+					max_i = i;
+				} else {
+					max_i = i-1;
+				}
+			} else if (cmp < 0) {
+				max_i = i-1;
+			} else {
+				min_i = i+1;
+			}
+		}
+
+		if (min_i != max_i) {
+			/* none match */
+			continue;
+		}
+
+		/* we now know that the entries start at min_i */
+		for (i=min_i;i<d->num_entries;i++) {
+			struct notify_entry *e = &d->entries[i];
+			if (p_len != e->path_len ||
+			    strncmp(path, e->path, p_len) != 0) break;
+			if (next_p != NULL) {
+				if (0 == (filter & e->subdir_filter)) {
+					continue;
+				}
+			} else {
+				if (0 == (filter & e->filter)) {
+					continue;
+				}
+			}
+			notify_send(notify, e, path + e->path_len + 1, action);
+		}
+	}
+}