Re: [patch] mmap() MAP_TEXT implementation (to use for shared libraries)

On Tue, Sep 04, 2012 at 02:49:07PM +0200, Svatopluk Kraus wrote:
> On Mon, Sep 3, 2012 at 2:46 PM, Konstantin Belousov <kostikbel_at_gmail.com> wrote:
> > On Mon, Sep 03, 2012 at 12:35:08PM +0200, Svatopluk Kraus wrote:
> >> Hi,
> >>
> >>   I found out that while the running excecutables and a dynamic linker
> >> are protected against writing (ETXTBSY), the loaded shared libraries
> >> are not protected. The libraries are mapped by mmap() in dynamic
> >> linker (rtld) and there is no way how to set VV_TEXT flag on the
> >> libraries vnodes in mmap() code.
> >>
> >>   In linux compability code \compat\linux\linux_misc.c, linux_uselib()
> >> sets VV_TEXT flags on a library vnode. In Solaris, MAP_TEXT flag
> >> exists which informs mmap() that the mapped region will be used
> >> primarily for executing instructions (for better MMU utilization).
> >> With these on mind, I propose to implement MAP_TEXT option in mmap()
> >> and in case that underlying object is a vnode, set VV_TEXT flag on it.
> >>
> >>   I already have implemented it and with rtld map_object() patch it
> >> works fine for me (of course). The rtld patch looks easy, however I'm
> >> not sure about mmap patch.
> >>
> >>   After some investigation, it looks that VV_TEXT once set on a vnode
> >> remains set until last reference on the vnode is left. So, I don't
> >> bother with VV_TEXT unset in munmap() to be consistent. The
> >> executables and dynamic linker are activated in kernel, so VV_TEXT is
> >> set before activation and cleared if something failed. Shared library
> >> activation is done in dynamic linker (i.e., in userland). It's done in
> >> steps and mmaping the library is one from them. So, I think that
> >> VV_TEXT can be set in mmap() just after everything is finished
> >> successfully.
> > This is right, the object reference counter is also used as
> > VV_TEXT counter. It is somewhat unaccurate, but in practice does
> > not cause issues.
> >
> >>
> >>   The patch itself is implemented in vm_mmap_vnode(). If I want to set
> >> VV_TEXT flag on a vnode, I need an exclusive lock. In current code,
> >> the exclusive lock flag is (mis)used as a flag for
> >> vnode_pager_update_writecount() call. (I hope that I didn't miss
> >> something.) So, the patch is bigger slightly.
> >>
> >>   I defined the MAP_TEXT flag in extented flags sections. However, I'm
> >> feeling the relation to MAP_STACK flag, but not sure if and when
> >> reserved flags (in other flags section) can be re-used.
> >>
> >>        Svata
> >>
> >>
> >>   Index: libexec/rtld-elf/map_object.c
> >> ===================================================================
> >> --- libexec/rtld-elf/map_object.c     (revision 239770)
> >> +++ libexec/rtld-elf/map_object.c     (working copy)
> >> _at__at_ -199,7 +199,8 _at__at_
> >>       data_prot = convert_prot(segs[i]->p_flags);
> >>       data_flags = convert_flags(segs[i]->p_flags) | MAP_FIXED;
> >>       if (mmap(data_addr, data_vlimit - data_vaddr, data_prot,
> >> -       data_flags | MAP_PREFAULT_READ, fd, data_offset) == (caddr_t) -1) {
> >> +       data_flags | MAP_PREFAULT_READ | MAP_TEXT, fd, data_offset) ==
> >> +         (caddr_t) -1) {
> > I am not sure that we shall mark all segments mappings with MAP_TEXT.
> > I understand the logic of the change, since we do not want data segment
> > to be changed under us. Still, having MAP_TEXT for non-text segments looks
> > strange.
> >
> 
> I agree. However, only way how to recognize a text segment is an
> executable flag set. The new patch for map_object.c is following:
> 
> Index: libexec/rtld-elf/map_object.c
> ===================================================================
> --- libexec/rtld-elf/map_object.c	(revision 239770)
> +++ libexec/rtld-elf/map_object.c	(working copy)
> _at__at_ -442,5 +442,10 _at__at_
>       */
>      if (!(elfflags & PF_W))
>  	flags |= MAP_NOCORE;
> +    /*
> +     * Executable mappings are marked "MAP_TEXT".
> +     */
> +    if (elfflags & PF_X)
> +	flags |= MAP_TEXT;
>      return flags;
>  }
> 
> 
> >>           _rtld_error("%s: mmap of data failed: %s", path,
> >>               rtld_strerror(errno));
> >>           goto error1;
> >> Index: sys/vm/vm_mmap.c
> >> ===================================================================
> >> --- sys/vm/vm_mmap.c  (revision 239770)
> >> +++ sys/vm/vm_mmap.c  (working copy)
> >> _at__at_ -1258,10 +1258,13 _at__at_
> >>       struct mount *mp;
> >>       struct ucred *cred;
> >>       int error, flags, locktype, vfslocked;
> >> +     int writeable_shared;
> >>
> >>       mp = vp->v_mount;
> >>       cred = td->td_ucred;
> >> -     if ((*maxprotp & VM_PROT_WRITE) && (*flagsp & MAP_SHARED))
> >> +     flags = *flagsp;
> >> +     writeable_shared = ((*maxprotp & VM_PROT_WRITE) && (flags & MAP_SHARED));
> >> +     if (writeable_shared || ((flags & MAP_TEXT) != 0))
> >>               locktype = LK_EXCLUSIVE;
> >>       else
> >>               locktype = LK_SHARED;
> >> _at__at_ -1271,7 +1274,6 _at__at_
> >>               return (error);
> >>       }
> >>       foff = *foffp;
> >> -     flags = *flagsp;
> >>       obj = vp->v_object;
> >>       if (vp->v_type == VREG) {
> >>               /*
> >> _at__at_ -1294,7 +1296,7 _at__at_
> >>                               return (error);
> >>                       }
> >>               }
> >> -             if (locktype == LK_EXCLUSIVE) {
> >> +             if (writeable_shared) {
> >>                       *writecounted = TRUE;
> >>                       vnode_pager_update_writecount(obj, 0, objsize);
> >>               }
> >> _at__at_ -1337,6 +1339,14 _at__at_
> >>               error = ENOMEM;
> >>               goto done;
> >>       }
> >> +     /*
> >> +      * If MAP_TEXT is announced, set VV_TEXT so no one can write
> >> +      * to the executable.
> >> +      */
> >> +     if ((flags & MAP_TEXT) != 0) {
> >> +             ASSERT_VOP_ELOCKED(vp, "vv_text");
> >> +             vp->v_vflag |= VV_TEXT;
> >> +     }
> > I do not think we want to set VV_TEXT for device vnodes.
> >
> 
> I agree too. However, my patch doesn't set VV_TEXT for device vnodes.
> Device vnodes never enter into patched part of code.
Hm, yes.

Anyway, after thinking about the patch more, I see two issues:

1. You are setting VV_TEXT without checking v_writecount. This basically
   nullifies the main reason for the patch, since existing writer can still
   write or truncate the shared library after the mapping.

2. I do not see what would prevent malicious local user from mmaping
   arbitrary file readonly with MAP_TEXT, thus blocking any modifications
   to the file. Note that this is not a problem for executables, because
   kernel only sets VV_TEXT on executables if +x permission is set and
   file is valid binary which kernel is able to execute.

   E.g. you might block log writes with VV_TEXT, or other user editing
   session or whatever, having just read access to corresponding files.

Am I wrong ?