Date: Thu, 11 Jul 2019 08:29:35 -0700 From: Mark Millard <marklmi@yahoo.com> To: adr <adr@SDF.ORG> Cc: Robert Crowston via freebsd-arm <freebsd-arm@freebsd.org> Subject: Re: FreeBSD arm EABI5 documentation? Message-ID: <1CB61FE0-5665-424F-8B94-ABFC06906112@yahoo.com> In-Reply-To: <AFF6488F-B2F5-486E-8F14-FAF852DB2C3C@yahoo.com> References: <alpine.NEB.2.21.1907101508370.22895@sdf.lonestar.org> <f2967859f68f1d40f260661791126956f48f4d12.camel@freebsd.org> <alpine.NEB.2.21.1907101735020.3201@sdf.lonestar.org> <da6a8f7e596da9bf9015f76798a5575908ad1be4.camel@freebsd.org> <alpine.NEB.2.21.1907101904570.15468@sdf.lonestar.org> <1788e13e706b9fdaf610e4ddd671a5ed715f9dfe.camel@freebsd.org> <alpine.NEB.2.21.1907102043040.9461@sdf.lonestar.org> <AFF6488F-B2F5-486E-8F14-FAF852DB2C3C@yahoo.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On 2019-Jul-10, at 15:18, Mark Millard <marklmi at yahoo.com> wrote:
>=20
> On 2019-Jul-10, at 13:47, adr <adr at SDF.ORG> wrote:
>=20
>> On Wed, 10 Jul 2019, Ian Lepore wrote:
>>=20
>>> That's... odd. The arm spec requires the stack to be 8-byte aligned =
at
>>> any public interface. It's hard to imagine how anything could work
>>> properly if it were not, given that the toolchains will assume that =
64-
>>> bit values are aligned at 64-bit boundaries and will thus generate
>>> instructions that require that alignment (require it even if strict
>>> alignment checking for most instructions is disabled in the control
>>> register). If you could enter a function with the stack only 4-byte
>>> aligned, how would the compiler know it's safe to use something like =
an
>>> LDREXD instruction on a local variable allocated on the stack?
>>=20
>> I have no idea. The only thing I can assure you is that in the code =
I'm
>> talking about (a forth implementation using SDL2) I've never aligned =
the
>> C stack when passing arguments to external functions. Until now!
>=20
> Curious, I looked around in NetBSD code and found the likes of:
>=20
> /* ARM-specific macro to align a stack pointer (downwards). */
> #define STACK_ALIGNBYTES (8 - 1)
>=20
> in netbsd's src/sys/arch/arm/include/param.h .
>=20
> In src/sys/arch/aarch64/include/param.h there is:
>=20
> /* AARCH64-specific macro to align a stack pointer (downwards). */
> #define STACK_ALIGNBYTES (16 - 1)
>=20
>=20
> So it appears that the kernel does have stack alignment:
> src/sys/sys/param.h has . . .
>=20
> #if defined(_KERNEL) || defined(__EXPOSE_STACK)
> . . .
> #ifndef STACK_ALIGNBYTES
> #define STACK_ALIGNBYTES __ALIGNBYTES
> #endif
> . . .
> #define STACK_ALIGN(sp, bytes) \
> ((char *)(((unsigned long)(sp)) & ~(bytes)))
> . . .
> #define STACK_LEN_ALIGN(len, bytes) (((len) + (bytes)) & =
~(bytes))
>=20
> There is code for signal delivery:
>=20
> void
> sendsig_siginfo(const ksiginfo_t *ksi, const sigset_t *mask)
> {
> . . .=09
> fp =3D getframe(l, sig, &onstack);
> =09
> /* make room on the stack */
> fp--;
> =09
> /* make the stack aligned */
> fp =3D (struct sigframe_siginfo *)STACK_ALIGN(fp, =
STACK_ALIGNBYTES);
> . . .
>=20
> AARCH64 even has for runing 32-bit code:
>=20
> static void
> netbsd32_sendsig_siginfo(const ksiginfo_t *ksi, const sigset_t *mask)
> {
> . . .
> fp =3D (struct netbsd32_sigframe_siginfo *)sp;
> fp =3D (struct netbsd32_sigframe_siginfo *)STACK_ALIGN(fp - 1, =
8);
> . . .
>=20
> (But that last looks wrong: 8 should likely be 8-1, as in
> STACK_ALIGNBYTES for 32-bit ARM.)
>=20
> There is:
>=20
> static size_t
> calcstack(struct execve_data * restrict data, const size_t gaplen)
> {
> . . .
> /* make the stack "safely" aligned */
> return STACK_LEN_ALIGN(stacklen, STACK_ALIGNBYTES);
> }
>=20
> I saw other comments mentioning ' stack "safely" aligned' when
> looking around.
>=20
> So it basically matches up with Ian's comments at the kernel vs. user
> space interface, despite the probable netbsd32_sendsig_siginfo error.
I had an exchange via Jared McNeill (who is active on NetBSD)
that involved NetBSD's criteria. Overallit went like this for
that aspect of the exchange:
QUOTE
From: Nick Hudson <nick.hudson at gmx.co.uk>
Subject: Re: netbsd32_sendsig_siginfo STACK_ALIGN use problem?
Date: July 11, 2019 at 03:59:52 PDT
To: Jared McNeill <jared.mcneill at live.ca>, Mark Millard <marklmi at =
yahoo.com>
On 11/07/2019 11:21, Jared McNeill wrote:
> Hi Mark =E2=80=94
>=20
> This isn=E2=80=99t really an area I am familiar with. Nick Hudson (in =
copy), our resident arm expert, can likely help.
>=20
> Cheers,
> Jared
>=20
>=20
>> On Jul 10, 2019, at 7:17 PM, Mark Millard <marklmi at yahoo.com> =
wrote:
>>=20
>> Curious about 32-bit arm stack alignment requirements in netbsd
>> (based on a FreeBSD thread making claims that user space allows
>> 4-byte stack alignment), I went looking around some in NetBSD
>> code. . . .
>>=20
8 byte stack alignment is a requirement of
=
http://infocenter.arm.com/help/topic/com.arm.doc.ihi0042f/IHI0042F_aapcs.p=
df
5.2.1.2 Stack constraints at a public interface
The stack must also conform to the following constraint at a public =
interface:
SP mod 8 =3D 0. The stack must be double-word aligned
I hope this helps.
END QUOTE
(I was really reporting a coding error in netbsd32_sendsig_siginfo
relative to meeting this criteria.)
=3D=3D=3D
Mark Millard
marklmi at yahoo.com
( dsl-only.net went
away in early 2018-Mar)
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?1CB61FE0-5665-424F-8B94-ABFC06906112>