Re: "Switch !O_CREAT case to use of do_last()" causes segfault in glibc

[Al Viro <viro@ZenIV.linux.org.uk>]

On Sat, Mar 06, 2010 at 10:20:41AM -0800, walt wrote:
> On 03/06/2010 10:03 AM, Al Viro wrote:
> 
> >_Really_ interesting; it doesn't look like an oops - smells like an attempt
> >to do opendir() that fails for some reason, goes unnoticed and resulting
> >FILE * (i.e. NULL) is fed to readdir()?
> >
> >What does it attempt to open?
> 
> Ah, this may help:
> 
> open("/usr/share/zoneinfo/", O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC) = 47
> open("/usr/share/zoneinfo/MST7MDT", O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC) = 48
> open("/usr/share/zoneinfo/MST7MDT/", O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC) = -1 ENOTDIR (Not a directory)
> --- SIGSEGV (Segmentation fault) @ 0 (0) ---

Oho...  And what do you have in /usr/share/zoneinfo?  I.e. what is MST7MDT?

BTW, icu is buggy, regardless of apparent bug introduced in the kernel -
common/putil.c:searchForTZFile() does

    DIR* dirp = opendir(path);
...
    while((dirEntry = readdir(dirp)) != NULL) {

with no references to dirp in between (no checks for errors, etc.), so it'll
segfault on any number of situations (e.g. directory in question being
unreadable for you).