On Wed, May 11, 2016 at 02:33:01PM +0100, Colin Ian King wrote:
> On 11/05/16 14:18, Catalin Marinas wrote:
> > On Tue, May 10, 2016 at 10:50:40PM -0400, Mike Frysinger wrote:
> >> On 09 May 2016 22:40, Colin Ian King wrote:
> >>> On 09/05/16 22:31, Mike Frysinger wrote:
> >>>> On 25 Apr 2016 20:42, Colin Ian King wrote:
> >>>>> currently, the aarch64 clone() system call requires the stack to be
> >>>>> aligned at a 16 byte boundary, see arch/arm64/kernel/process.c,
> >>>>> copy_thread():
> >>>>>
> >>>>> if (stack_start) {
> >>>>> if (is_compat_thread(task_thread_info(p)))
> >>>>> childregs->compat_sp = stack_start;
> >>>>> /* 16-byte aligned stack mandatory on AArch64 */
> >>>>> else if (stack_start & 15)
> >>>>> return -EINVAL;
> >>>>> else
> >>>>> childregs->sp = stack_start;
> >>>>> }
> >>>>>
> >>>>>
> >>>>> ..and returns -EINVAL if not aligned correctly. This should be added to
> >>>>> the manual page clone(2) as it took me a while to figure out why clone()
> >>>>> was failing with -EINVAL for aarch64 but not on x86.
> >>>>
> >>>> seems weird for the kernel to be enforcing this. is it just because of
> >>>> the stated ABI ? or is there some weird requirement in the kernel itself
> >>>> that requires this ? it's not like other arches have this check, and
> >>>> there are def ABI requirements about stack alignments in C.
> >>>
> >>> The article here indicates it is an aarch64 convention:
> >>>
> >>> https://community.arm.com/groups/processors/blog/2015/11/19/using-the-stack-in-aarch32-and-aarch64
> >>
> >> that checks my point about the ABI having alignment requirements, but
> >> that doesn't mean it needs to be checked/enforced in the kernel. all
> >> the limitations i see there can be seen in other arches, but we don't
> >> have those arches do any stack alignment checking. so should we be
> >> dropping it from aarch64 ? why does it need to be special here ?
> >
> > It is not just a software ABI requirement but a hardware one. If you try
> > to access the stack with an unaligned SP value, you get a fault followed
> > by a SIGBUS delivered to the user application. We decided to enforce
> > this at the copy_thread() level, it is easier to catch such issue early
> > than debugging SIGBUS delivered to a thread.
>
> Rather than returning -EINVAL would it be more useful re-align
> stack_start to the 16 byte boundary in copy_thread as a silent but
> useful fixup?
I wouldn't silently re-align the stack, it's a significant kernel ABI
change. Even dropping -EINVAL in favour of a later SIGBUS is an ABI
change, though not sure if any user apps or libraries would be affected
(I wouldn't expect them to rely on the -EINVAL return).
It seems that musl does this alignment in its clone(2) implementation:
https://git.musl-libc.org/cgit/musl/tree/src/thread/aarch64/clone.s
IIUC, glibc does not.
> It took me a while to debug the -EINVAL on the clone() system call to
> figure out what was wrong because I didn't realize aarch64 has this
> constraint.
Would it have been easier to get a SIGBUS on the first stack access?
It's worth posting a patch removing -EINVAL on linux-arm-kernel for
wider discussion.
--
Catalin
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
