On Wed, May 29, 2019 at 04:00:23PM -0700, Kees Cook wrote:
> On Wed, May 29, 2019 at 03:20:20PM -0700, Andrew Morton wrote:
> > On Thu, 30 May 2019 00:37:08 +0300 Alexey Dobriyan <adobri...@gmail.com>
> > wrote:
> >
> > > AT_RANDOM content is always misaligned on x86_64:
> > >
> > > $ LD_SHOW_AUXV=1 /bin/true | grep AT_RANDOM
> > > AT_RANDOM: 0x7fff02101019
> > >
> > > glibc copies first few bytes for stack protector stuff, aligned
> > > access should be slightly faster.
> >
> > I just don't understand the implications of this. Is there
> > (badly-behaved) userspace out there which makes assumptions about the
> > current alignment?
> >
> > How much faster, anyway? How frequently is the AT_RANDOM record
> > accessed?
> >
> > I often have questions such as these about your performance/space
> > tweaks :(. Please try to address them as a matter of course when
> > preparing changelogs?
> >
> > And let's Cc Kees, who wrote the thing.
> >
> > > --- a/fs/binfmt_elf.c
> > > +++ b/fs/binfmt_elf.c
> > > @@ -144,11 +144,15 @@ static int padzero(unsigned long elf_bss)
> > > #define STACK_ALLOC(sp, len) ({ \
> > > elf_addr_t __user *old_sp = (elf_addr_t __user *)sp; sp += len; \
> > > old_sp; })
> > > +#define STACK_ALIGN(sp, align) \
> > > + ((typeof(sp))(((unsigned long)sp + (int)align - 1) & ~((int)align - 1)))
> >
> > I suspect plain old ALIGN() could be used here.
> >
> > > #else
> > > #define STACK_ADD(sp, items) ((elf_addr_t __user *)(sp) - (items))
> > > #define STACK_ROUND(sp, items) \
> > > (((unsigned long) (sp - items)) &~ 15UL)
> > > #define STACK_ALLOC(sp, len) ({ sp -= len ; sp; })
> > > +#define STACK_ALIGN(sp, align) \
> > > + ((typeof(sp))((unsigned long)sp & ~((int)align - 1)))
> >
> > And maybe there's a helper which does this, dunno.
> >
> > > #endif
> > >
> > > #ifndef ELF_BASE_PLATFORM
> > > @@ -217,6 +221,12 @@ create_elf_tables(struct linux_binprm *bprm, struct
> > > elfhdr *exec,
> > > return -EFAULT;
> > > }
> > >
> > > + /*
> > > + * glibc copies first bytes for stack protector purposes
> > > + * which are misaligned on x86_64 because strlen("x86_64") + 1 == 7.
> > > + */
> > > + p = STACK_ALIGN(p, sizeof(long));
> > > +
>
> I have no objection to eating some bytes here. Though perhaps things could
> just
> be reordered to leave all the aligned things together and put all the
> strings later?
There should be no bytes wasted in fact. Auxv array is aligned and
whole stack is aligned once more at 16 bytes. On x86_64 AT_RANDOM content
and "x86_64" AT_PLATFORM string are put higher, so that 1 byte doesn't
change anything.
