[Valgrind-developers] support memcheck of UserModeLinux

[John R. <jr...@Bi...>]

The attached patch to valgrind-3.3.0 is what I am using to run
UML (UserModeLinux) for x86 under memcheck on x86, thus checking
the memory accesses of a linux kernel.

-rw-rw-r-- 1 jreiser jreiser  8846 2007-12-15 11:36 valgrind-3.3.0.patch.gz
-rw-rw-r-- 1 jreiser jreiser 31383 2007-12-15 11:36 valgrind-3.3.0.patch

The changes generally are in three areas:

   Allow clone() to handle more flag cases successfully,
   and also establishing a new stack pointer in the child
   even for a fork-like clone().  These changes are
   somewhat general, although motivated by UML.

   Allow better coordination with respect to explicit stack switching;
   UML uses longjmp() for this purpose.  Perhaps after more thought
   this could be done automatically by VEX when translating
   "load stack pointer from memory."

   Enable valgrind to stop tracking the child of a clone() immediately,
   letting the child run natively.  This change is somewhat specific
   to UML and its "skas0" mode of operation.  UML has a difficult time
   dealing with virtualization of the child side of clone(); having the
   tool "let go" of most children makes it easier.

Obviously UML requires changes, too.  Although the linux kernel has
kmalloc() and kfree() subroutines, the obvious coorespondence to malloc()
and free() does not carry all the needed or desired information.
The kernel allocators already have some builtin checking, and
allocating a "page" can have non-simple implications.
There is work in annotating the linux kernel allocators
with VALGRIND_MALLOCLIKE_BLOCK, VALGRIND_FREELIKE_BLOCK,
VALGRIND_MAKE_MEM_DEFINED, and VALGRIND_MAKE_MEM_UNDEFINED.
Details will appear on the uml-devel mailing list.

-- 
John Reiser, jr...@Bi...