Re: gcc versus clang issue for 32-bit binaries

On Wed, Jun 10, 2020 at 06:41:50PM +0200, Michael Tuexen wrote:
> Dear all,
> 
> consider the following program test.c:
> 
> #include <sys/mman.h>
> #include <stdio.h>
> 
> int 
> main(void)
> {
> 	void *p;
> 	
> 	p = mmap((void *)0x20000000, 0x1000000, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_ANON | MAP_PRIVATE | MAP_FIXED, -1, 0);
> 	printf("p= %p\n", p);
> 	return (0);
> }
> 
> On i386 the following happens:
> * when compiling it with cc and running it, it crashes.
> * when compiling it with gcc it runs fine.
> 
> On amd64 the following happens:
> * when compiling it with cc -m64 it runs fine.
> * when compiling it with cc -m32 is crashes.
> * when compiling it with gcc -m64 it runs fine.
> * when compiling it with gcc -m32 it runs fine.
> 
> So why does the above program crash when compiled for 32-bit when using clang, but runs fine when compiled with gcc.

The difference is between ld.bfd and ld.lld, which emit executables with
different entry point addresses.  cc -m32 -fuse-ld=bfd gives an
executable that does not crash.

When linked with lld, libc and ld-elf get mapped into the region
[0x20000000,0x21000000], so the program crashes when the libc.so mapping
is overwritten with that created by the mmap() call and the program
calls printf().

> I'm testing this on 32-bit and 64-bit head systems. gcc is from ports.
> 
> The reason I'm looking into it is that I want to get syzkaller working on 32-bit with clang.

Do you know why SYZ_DATA_OFFSET is hard-coded the way it is?  It looks
like it works more or less by accident, but at a glance I don't see why
it has to be a fixed mapping.