On Sun, Sep 22, 2002 at 07:34:25PM +0000, Charles Hannum wrote:
> 
> While your patch may help a little, I do not think it's sufficient
> even to fully mitigate the problem you're talking about.  The problem
> is that it only affects lookup hits, not misses, and so a certain
> number of missed lookups will occur in many (most?) cases anyway.
> 
> Perhaps better would be to hook into:
> 
>         for (sp = _rtld_paths; sp != NULL; sp = sp->sp_next)
>                 if ((pathname = _rtld_search_library_path(name, namelen,
>                     sp->sp_path, sp->sp_pathlen)) != NULL) 
>                         return (pathname);
> 
> (and the _rtld_default_paths case below it), and to record both
> positive and negative hits here by file name (without the path).  This
> would allow us to do a simple table lookup the next time we see the
> file name.

That would require us to maintain file name list aside from path list,
wouldn't it?

BTW, I had a look at it again and moved the code into
_rtld_search_library_path() so that I could save calls to
_rtld_check_library() in many cases (with mozilla, the number is 418).

Index: search.c
===================================================================
RCS file: /cvsroot/basesrc/libexec/ld.elf_so/search.c,v
retrieving revision 1.10
diff -u -r1.10 search.c
--- search.c	2000/07/27 10:44:39	1.10
+++ search.c	2002/09/23 04:32:23
@@ -110,11 +110,18 @@
 	size_t dirlen;
 {
 	char *pathname;
+	Obj_Entry *obj;
 
 	pathname = xmalloc(dirlen + 1 + namelen + 1);
 	(void)strncpy(pathname, dir, dirlen);
 	pathname[dirlen] = '/';
 	strcpy(pathname + dirlen + 1, name);
+
+	for (obj = _rtld_objlist->next; obj != NULL; obj = obj->next)
+		if (strcmp(obj->path, pathname) == 0) {
+			dbg(("  %s is already on the list", pathname));
+			return pathname;
+		}
 
 	dbg(("  Trying \"%s\"", pathname));
 	if (_rtld_check_library(pathname))	/* We found it */

If there's no better solution, I'll commit it in a few days.

Jun-Young

-- 
Bang Jun-Young <junyoung@mogua.com>