gdb/656: gdb_realpath crashes on GNU/Hurd

[marcus at gnu dot org]

>Number:         656
>Category:       gdb
>Synopsis:       gdb_realpath crashes on GNU/Hurd
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Sat Aug 17 14:18:00 PDT 2002
>Closed-Date:
>Last-Modified:
>Originator:     marcus@gnu.org
>Release:        5.2.1
>Organization:
>Environment:
GNU/Hurd
>Description:
Hi,
this is another crash on the GNU/Hurd which exists for some longer time and I forgot about it.

gdb_realpath was at some time changed to prefer realpath over canonicalize_file_name.  This is bad!
First, please note that the definition of realpath() itself is bogus, and its interface is not really usable in POSIX.  This is because the definition of PATH_MAX is optional, and even pathconf ("/", _PC_PATH_MAX) can return -1, meaning that there is no limit.
This is what happens on the Hurd.  Now, the realpath preferring code does an uncehcked alloca on the return value of pathconf.  This alloca's -1 byte (haha) and then subsequently crashes in the xstrdup function.
This is since this change:
2002-03-14  Richard Henderson  <rth@redhat.com>
        (gdb_realpath): Prefer realpath if available and usable.
>How-To-Repeat:
For example, the "file" command will crash gdb with a segmentation fault.
>Fix:
Don't use realpath!  It's broken.  It can not be made work.  When using realpath on the Hurd, there is always the danger of a buffer overflow.
Instead, use canonicalize_file_name, or realpath with a buffer of NULL (which is the recommended interface for a future POSIX version).  That's why we have this feature, and this is why the code originally preferred it.

Thanks!
>Release-Note:
>Audit-Trail:
>Unformatted: