Re: mlock in libc

[Marcus Brinkmann]

Hi,

shouldn't be on libc-alpha, I guess.  It's just noise for most of the people
there.

On Sun, Nov 03, 2002 at 11:29:33AM -0500, James A Morrison wrote:
>   I was looking at why gpg fails to build on the Hurd.  I discovered mlock
> is what is broken.

Without really going into your patch, here is some background noise from the
past.  Here is what Roland wrote a long time ago:

"This might be a known kernel problem, I'm just a bit sketchy on the details.
 That is, the Hurd code that does wiring (libshouldbeinlibc/wire.c) makes
 the pages writable and pokes them before doing vm_wire.

 It seems like a breach of the supposed kernel interface of vm_wire that you
 should have to do this (vm_wire should just force all the copies it needs
 to), but that is the way the kernel is.  At least I think so, but I am not
 entirely sure I am right about the situation in the kernel."

He also suggested:

"It would be cool if you could step through vm_wire using gdb on oskit-mach
 and figure out where specifically it fails."

Thomas replied to the above first paragaph:

"This is because there were kernel bugs if you get a pagein on wired
 memory.  That is, the kernel assumes that wired memory is *already* in
 core.  The little poking routine is to do our best to make sure we
 don't have that problem."

> The problem arises when the actual protection on the
> memory is less relaxed than the protection passed to vm_wire.  What does work
> is using VM_PROT_DEFAULT, because memory allocated with malloc seems to use
> VM_PROT_DEFAULT.  However, I think the best solution is to call vm_region
> and actually get the proper permissions for the pages.
> 
>  Here are a couple of patches that checks vm_region and acts more like 
> wire_segment_internal in [hurd]/libshouldbeinlibc.  The first patch relaxes
> the permissions on the memory region to max_protection so that the wired
> region can have its protections relaxed to that level at any time.  The
> second does not because perhaps something could write to a read-only part
> of memory between the calls to vm_protect.
> 
> 2002-11-03  James Morrison  <ja2morri@uwaterloo.ca>
> 
>       * sysdeps/mach/hurd/mlock.c (mlock): Get the actual protections on a
>       region of memory before wiring it.
> 
> Index: mlock.c
> ===================================================================
> RCS file: /cvs/glibc/libc/sysdeps/mach/hurd/mlock.c,v
> retrieving revision 1.2
> diff -u -r1.2 mlock.c
> --- mlock.c     6 Jul 2001 04:55:57 -0000       1.2
> +++ mlock.c     3 Nov 2002 16:18:09 -0000
> @@ -39,8 +39,19 @@
> 
>    page = trunc_page ((vm_address_t) addr);
>    len = round_page ((vm_address_t) addr + len) - page;
> -  err = __vm_wire (hostpriv, __mach_task_self (), page, len,
> -                  VM_PROT_ALL); /* XXX ? */
> +  {
> +    vm_prot_t cur, max;
> +    vm_inherit_t inher;
> +    boolean_t shared;
> +    mach_port_t obj;
> +    vm_offset_t offs;
> +
> +    __vm_region (mach_task_self (), &page, &len, &cur, &max, &inher, &shared,
> +               &obj, &offs);
> +    __vm_protect (mach_task_self (), page, len, 0, max);
> +    err = __vm_wire (hostpriv, __mach_task_self (), page, len, max);
> +    __vm_protect (mach_task_self (), page, len, 0, cur);
> +  }
>    __mach_port_deallocate (__mach_task_self (), hostpriv);
> 
>    return err ? __hurd_fail (err) : 0;
> 
> Index: mlock.c
> ===================================================================
> RCS file: /cvs/glibc/libc/sysdeps/mach/hurd/mlock.c,v
> retrieving revision 1.2
> diff -u -r1.2 mlock.c
> --- mlock.c     6 Jul 2001 04:55:57 -0000       1.2
> +++ mlock.c     3 Nov 2002 16:25:44 -0000
> @@ -39,8 +39,18 @@
> 
>    page = trunc_page ((vm_address_t) addr);
>    len = round_page ((vm_address_t) addr + len) - page;
> -  err = __vm_wire (hostpriv, __mach_task_self (), page, len,
> -                  VM_PROT_ALL); /* XXX ? */
> +  {
> +    vm_prot_t cur, max;
> +    vm_inherit_t inher;
> +    boolean_t shared;
> +    mach_port_t obj;
> +    vm_offset_t offs;
> +
> +    __vm_region (mach_task_self (), &page, &len, &cur, &max, &inher, &shared,
> +               &obj, &offs);
> +    err = __vm_wire (hostpriv, __mach_task_self (), page, len, max);
> +    __mach_port_deallocate (mach_task_self (), obj);
> +  }
>    __mach_port_deallocate (__mach_task_self (), hostpriv);
> 
>    return err ? __hurd_fail (err) : 0;
> 

-- 
`Rhubarb is no Egyptian god.' GNU      http://www.gnu.org    marcus@gnu.org
Marcus Brinkmann              The Hurd http://www.gnu.org/software/hurd/
Marcus.Brinkmann@ruhr-uni-bochum.de
http://www.marcus-brinkmann.de/