Re: [Sbcl-devel] amd64 build failure

[Nikodemus S. <nik...@ra...>]

On 9/6/07, Andreas Fuchs <as...@bo...> wrote:

> I'm still getting the very same problem building 1.0.9.39 (the revision
> with Nikodemus' fix) unithread on x86-64/linux.
>
> Have a gdb attached to an instance trying to load the cold core, can
> assist debugging efforts tomorrow:

*sigh*

I cannot reproduce this on my boxes -- probably because I don't have a
real 64bit setup: the best I can do is building a x86-64 SBCL in a 32bit
environment,,,

> Program received signal SIGSEGV, Segmentation fault.
> create_initial_thread (initial_function=68759912201) at thread.c:423
> 423         th->os_address = spaces;
> (gdb) bac
> #0  create_initial_thread (initial_function=68759912201) at thread.c:423
> #1  0x0000000000412573 in main (argc=-187064080, argv=0x7ffff4d9a0c8,
> envp=0x7ffff4d9a0f8) at runtime.c:425
> (gdb) print spaces
> No symbol "spaces" in current context.
> (gdb) print th
> $1 = (struct thread *) 0xb88f9000
> (gdb) print th->os_address
> Cannot access memory at address 0xb88f9010

Adding good old fashioned printf debugging for "spaces" and "aligned_spaces"
might yield enlightenment.

    spaces=os_validate(0, THREAD_STRUCT_SIZE);
    if(!spaces)
        return NULL;
    /* Aligning up is safe as THREAD_STRUCT_SIZE has BACKEND_PAGE_SIZE
     * padding. */
    aligned_spaces = (void *)((((unsigned long)(char *)spaces)
                               + BACKEND_PAGE_SIZE - 1)
                              & ~(BACKEND_PAGE_SIZE - 1));

I'm still working through my first cup of coffee for the day, but I suspect
the casts here aren't 64-bit proof, causing us to lose the high bits
from spaces.

should probably be

  & ~(unsigned long)(BACKEND_PAGE_SIZE -1)

there.

Cheers,

 -- Nikodemus