[TUHS] The evolution of Unix facilities and architecture

Chris Torek torek at torek.net
Sat May 13 11:17:02 AEST 2017

>Ah.  There's the other piece.  You start the SUID program under
>the debugger, and rather than kicking off the debugger, it simply
>starts it non-suid.  *However*, in the presence of shared text
>(either of the two cases being checked for in the other place),
>you can make changes to the text image (e.g.  put whatever code
>you want at the entry point), which will be reused the *next*
>time it is started *without* the debugger.

Right.  Some of this was not a problem with demand-paged files
(4.xBSD) since you would not share swap images.  But there's more,
or rather, *was* more, once people added PT_ATTACH and new system
calls or behavior ... specifically, the setreuid / setregid calls
from 4.2BSD or the saved setuid behavior in System V.  This is one
I personally touched.

Suppose a process starts out setuid or setgid.  This means it has
alternative privileges (maybe super-user, maybe not).  With these
it can do things like open some files or transit some directories.
Afterward, using setreuid() and setregid() in 4.2BSD, the process
can swap its real and effective IDs, or give up its effective UID
or GID entirely, to give up its privileges.  So some program
could, for instance, chdir past a "lock" directory -- this was the
MDQS trick -- and now exist in a tree that it had no access to, or
open a file with secrets, that it no longer has permission to

Once a process had given up any special privileges, though, it
could be ptraced again (via PT_ATTACH).  Now you can swap file
descriptor variables around in it, or otherwise make it do bad
things with the privileges it gained while it was setuid.  On
SysV, with "saved setuid", it is even worse: you can make the
process *regain* privileges and do whatever you want.

The fix we used was an extra bit in the process flags: "process
has had or used privileges". Once set (cleared only on exec()),
the process could no longer be ptraced except by root.  Processes
did -- and still do -- have to be careful about files they may
leave open, or other entities that survive exec().  (One should
also consider mmap() and shared regions.)


More information about the TUHS mailing list