Re: [PATCH/RESEND] recordmcount: arm: Implement make_nop
On 16 November 2016 at 14:08, Steven Rostedt wrote: > On Wed, 16 Nov 2016 11:48:38 + > Ard Biesheuvel wrote: > >> > >> > Does that allay your concerns? >> > >> >> Yes, it does. Thanks > > Does this mean I can pull this patch into my queue then? > Fine by me Thanks, Ard.
Re: [PATCH/RESEND] recordmcount: arm: Implement make_nop
On Wed, 16 Nov 2016 11:48:38 + Ard Biesheuvel wrote: > > > > Does that allay your concerns? > > > > Yes, it does. Thanks Does this mean I can pull this patch into my queue then? -- Steve
Re: [PATCH/RESEND] recordmcount: arm: Implement make_nop
On 15 November 2016 at 23:53, Stephen Boyd wrote:
> On 11/15, Ard Biesheuvel wrote:
>> On 15 November 2016 at 19:18, Stephen Boyd wrote:
>> > On 11/15, Ard Biesheuvel wrote:
>> >> On 19 October 2016 at 00:42, Stephen Boyd wrote:
>> >> >
>> >> > +static unsigned char ideal_nop4_arm_le[4] = { 0x00, 0x00, 0xa0, 0xe1
>> >> > }; /* mov r0, r0 */
>> >> > +static unsigned char ideal_nop4_arm_be[4] = { 0xe1, 0xa0, 0x00, 0x00
>> >> > }; /* mov r0, r0 */
>> >>
>> >> Shouldn't you be taking the difference between BE8 and BE32 into
>> >> account here? IIRC, BE8 uses little endian encoding for instructions.
>> >
>> > I admit I haven't tested on a pre-armv6 CPU so I haven't come
>> > across the case of a BE32 CPU. But from what I can tell that
>> > doesn't matter.
>> >
>> > According to scripts/Makefile.build, cmd_record_mcount only runs
>> > the recordmcount program if CONFIG_FTRACE_MCOUNT_RECORD=y. That
>> > config is defined as:
>> >
>> > config FTRACE_MCOUNT_RECORD
>> > def_bool y
>> > depends on DYNAMIC_FTRACE
>> > depends on HAVE_FTRACE_MCOUNT_RECORD
>> >
>> >
>> > And in arch/arm/Kconfig we see that DYNAMIC_FTRACE is selected:
>> >
>> > select HAVE_DYNAMIC_FTRACE if (!XIP_KERNEL) && !CPU_ENDIAN_BE32 &&
>> > MMU
>> >
>> > which means that FTRACE_MCOUNT_RECORD can't be set when
>> > CPU_ENDIAN_BE32 is set.
>> >
>> > Do you agree that BE32 is not a concern here?
>> >
>>
>> Yes. But that implies then that you should not be using big-endian
>> instruction encodings at all, and simply use the _le variants for both
>> LE and BE8
>
> Ok. I understand what you're getting at now.
>
> I believe the linker is the one that does the instruction endian
> swap to little endian. So everything is built as big-endian data
> and instructions in the assembler phase and then when the linker
> runs to generate the final vmlinux elf file it does the swaps to
> make instructions little endian. recordmcount runs on the object
> files and not the vmlinux file.
>
Very interesting, I did not know that.
> For example, the do_undefinstr() function in
> arch/arm/kernel/traps.c is one place we nop out. On an le host
> and an le build without this patch I see:
>
> (This is all ARM, not thumb)
>
> :
>0: e1a0c00dmov ip, sp
>4: e92dd9f0push{r4, r5, r6, r7, r8, fp, ip, lr, pc}
>8: e24cb004sub fp, ip, #4
>c: e24dd08csub sp, sp, #140; 0x8c
> 10: e52de004push{lr}; (str lr, [sp, #-4]!)
> 14: ebfebl 0 <__gnu_mcount_nc>
>
> After this patch on an le host and le build I see:
>
> :
>0: e1a0c00dmov ip, sp
>4: e92dd9f0push{r4, r5, r6, r7, r8, fp, ip, lr, pc}
>8: e24cb004sub fp, ip, #4
>c: e24dd08csub sp, sp, #140; 0x8c
> 10: e1a0nop ; (mov r0, r0)
> 14: e1a0nop ; (mov r0, r0)
>
> So far so good. Similarly, with this patch and an le host and be
> build I see:
>
> :
>0: e1a0c00dmov ip, sp
>4: e92dd9f0push{r4, r5, r6, r7, r8, fp, ip, lr, pc}
>8: e24cb004sub fp, ip, #4
>c: e24dd08csub sp, sp, #140; 0x8c
> 10: e1a0nop ; (mov r0, r0)
> 14: e1a0nop ; (mov r0, r0)
>
> but with *_le instead of *_be used a be build I see:
>
> :
>0: e1a0c00dmov ip, sp
>4: e92dd9f0push{r4, r5, r6, r7, r8, fp, ip, lr, pc}
>8: e24cb004sub fp, ip, #4
>c: e24dd08csub sp, sp, #140; 0x8c
> 10: a0e1andeq sl, r0, r1, ror #1
> 14: a0e1andeq sl, r0, r1, ror #1
>
> I confirmed this by looking at the hexdump of the .exception.text
> section for the traps.o object file and the .text section of the
> vmlinux file. Basically objcopy the .exception.text of traps.o to
> get the first few instructions of the do_undefinstr() function:
>
> $ hexdump -C traps.o
> e1 a0 c0 0d e9 2d d9 f0 e2 4c b0 04 e2 4d d0 8c
>
> And then objcopy the .text section in vmlinux and seek to the
> same function offset (there are a bunch of zeroes in front of it
> for padding):
>
> $ hexdump -C vmlinux
> ...
> 1000 0d c0 a0 e1 f0 d9 2d e9 04 b0 4c e2 8c d0 4d e2
>
> As can be seen everything is swapped from the original object
> file in big-endian to be in little endian.
>
> Does that allay your concerns?
>
Yes, it does. Thanks
Re: [PATCH/RESEND] recordmcount: arm: Implement make_nop
On 11/15, Ard Biesheuvel wrote:
> On 15 November 2016 at 19:18, Stephen Boyd wrote:
> > On 11/15, Ard Biesheuvel wrote:
> >> On 19 October 2016 at 00:42, Stephen Boyd wrote:
> >> >
> >> > +static unsigned char ideal_nop4_arm_le[4] = { 0x00, 0x00, 0xa0, 0xe1 };
> >> > /* mov r0, r0 */
> >> > +static unsigned char ideal_nop4_arm_be[4] = { 0xe1, 0xa0, 0x00, 0x00 };
> >> > /* mov r0, r0 */
> >>
> >> Shouldn't you be taking the difference between BE8 and BE32 into
> >> account here? IIRC, BE8 uses little endian encoding for instructions.
> >
> > I admit I haven't tested on a pre-armv6 CPU so I haven't come
> > across the case of a BE32 CPU. But from what I can tell that
> > doesn't matter.
> >
> > According to scripts/Makefile.build, cmd_record_mcount only runs
> > the recordmcount program if CONFIG_FTRACE_MCOUNT_RECORD=y. That
> > config is defined as:
> >
> > config FTRACE_MCOUNT_RECORD
> > def_bool y
> > depends on DYNAMIC_FTRACE
> > depends on HAVE_FTRACE_MCOUNT_RECORD
> >
> >
> > And in arch/arm/Kconfig we see that DYNAMIC_FTRACE is selected:
> >
> > select HAVE_DYNAMIC_FTRACE if (!XIP_KERNEL) && !CPU_ENDIAN_BE32 &&
> > MMU
> >
> > which means that FTRACE_MCOUNT_RECORD can't be set when
> > CPU_ENDIAN_BE32 is set.
> >
> > Do you agree that BE32 is not a concern here?
> >
>
> Yes. But that implies then that you should not be using big-endian
> instruction encodings at all, and simply use the _le variants for both
> LE and BE8
Ok. I understand what you're getting at now.
I believe the linker is the one that does the instruction endian
swap to little endian. So everything is built as big-endian data
and instructions in the assembler phase and then when the linker
runs to generate the final vmlinux elf file it does the swaps to
make instructions little endian. recordmcount runs on the object
files and not the vmlinux file.
For example, the do_undefinstr() function in
arch/arm/kernel/traps.c is one place we nop out. On an le host
and an le build without this patch I see:
(This is all ARM, not thumb)
:
0: e1a0c00dmov ip, sp
4: e92dd9f0push{r4, r5, r6, r7, r8, fp, ip, lr, pc}
8: e24cb004sub fp, ip, #4
c: e24dd08csub sp, sp, #140; 0x8c
10: e52de004push{lr}; (str lr, [sp, #-4]!)
14: ebfebl 0 <__gnu_mcount_nc>
After this patch on an le host and le build I see:
:
0: e1a0c00dmov ip, sp
4: e92dd9f0push{r4, r5, r6, r7, r8, fp, ip, lr, pc}
8: e24cb004sub fp, ip, #4
c: e24dd08csub sp, sp, #140; 0x8c
10: e1a0nop ; (mov r0, r0)
14: e1a0nop ; (mov r0, r0)
So far so good. Similarly, with this patch and an le host and be
build I see:
:
0: e1a0c00dmov ip, sp
4: e92dd9f0push{r4, r5, r6, r7, r8, fp, ip, lr, pc}
8: e24cb004sub fp, ip, #4
c: e24dd08csub sp, sp, #140; 0x8c
10: e1a0nop ; (mov r0, r0)
14: e1a0nop ; (mov r0, r0)
but with *_le instead of *_be used a be build I see:
:
0: e1a0c00dmov ip, sp
4: e92dd9f0push{r4, r5, r6, r7, r8, fp, ip, lr, pc}
8: e24cb004sub fp, ip, #4
c: e24dd08csub sp, sp, #140; 0x8c
10: a0e1andeq sl, r0, r1, ror #1
14: a0e1andeq sl, r0, r1, ror #1
I confirmed this by looking at the hexdump of the .exception.text
section for the traps.o object file and the .text section of the
vmlinux file. Basically objcopy the .exception.text of traps.o to
get the first few instructions of the do_undefinstr() function:
$ hexdump -C traps.o
e1 a0 c0 0d e9 2d d9 f0 e2 4c b0 04 e2 4d d0 8c
And then objcopy the .text section in vmlinux and seek to the
same function offset (there are a bunch of zeroes in front of it
for padding):
$ hexdump -C vmlinux
...
1000 0d c0 a0 e1 f0 d9 2d e9 04 b0 4c e2 8c d0 4d e2
As can be seen everything is swapped from the original object
file in big-endian to be in little endian.
Does that allay your concerns?
--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
Re: [PATCH/RESEND] recordmcount: arm: Implement make_nop
On 15 November 2016 at 19:18, Stephen Boyd wrote:
> On 11/15, Ard Biesheuvel wrote:
>> On 19 October 2016 at 00:42, Stephen Boyd wrote:
>> > In similar spirit to x86 and arm64 support, add a make_nop_arm()
>> > to replace calls to mcount with a nop in sections that aren't
>> > traced.
>> >
>> > Cc: Russell King
>> > Acked-by: Rabin Vincent
>> > Signed-off-by: Stephen Boyd
>> > ---
>> > scripts/recordmcount.c | 65
>> > ++
>> > 1 file changed, 65 insertions(+)
>> >
>> > diff --git a/scripts/recordmcount.c b/scripts/recordmcount.c
>> > index 5423a58d1b06..aeb34223167c 100644
>> > --- a/scripts/recordmcount.c
>> > +++ b/scripts/recordmcount.c
>> > @@ -213,6 +213,59 @@ static int make_nop_x86(void *map, size_t const
>> > offset)
>> > return 0;
>> > }
>> >
>> > +static unsigned char ideal_nop4_arm_le[4] = { 0x00, 0x00, 0xa0, 0xe1 };
>> > /* mov r0, r0 */
>> > +static unsigned char ideal_nop4_arm_be[4] = { 0xe1, 0xa0, 0x00, 0x00 };
>> > /* mov r0, r0 */
>>
>> Shouldn't you be taking the difference between BE8 and BE32 into
>> account here? IIRC, BE8 uses little endian encoding for instructions.
>
> I admit I haven't tested on a pre-armv6 CPU so I haven't come
> across the case of a BE32 CPU. But from what I can tell that
> doesn't matter.
>
> According to scripts/Makefile.build, cmd_record_mcount only runs
> the recordmcount program if CONFIG_FTRACE_MCOUNT_RECORD=y. That
> config is defined as:
>
> config FTRACE_MCOUNT_RECORD
> def_bool y
> depends on DYNAMIC_FTRACE
> depends on HAVE_FTRACE_MCOUNT_RECORD
>
>
> And in arch/arm/Kconfig we see that DYNAMIC_FTRACE is selected:
>
> select HAVE_DYNAMIC_FTRACE if (!XIP_KERNEL) && !CPU_ENDIAN_BE32 && MMU
>
> which means that FTRACE_MCOUNT_RECORD can't be set when
> CPU_ENDIAN_BE32 is set.
>
> Do you agree that BE32 is not a concern here?
>
Yes. But that implies then that you should not be using big-endian
instruction encodings at all, and simply use the _le variants for both
LE and BE8
Re: [PATCH/RESEND] recordmcount: arm: Implement make_nop
On 11/15, Ard Biesheuvel wrote:
> On 19 October 2016 at 00:42, Stephen Boyd wrote:
> > In similar spirit to x86 and arm64 support, add a make_nop_arm()
> > to replace calls to mcount with a nop in sections that aren't
> > traced.
> >
> > Cc: Russell King
> > Acked-by: Rabin Vincent
> > Signed-off-by: Stephen Boyd
> > ---
> > scripts/recordmcount.c | 65
> > ++
> > 1 file changed, 65 insertions(+)
> >
> > diff --git a/scripts/recordmcount.c b/scripts/recordmcount.c
> > index 5423a58d1b06..aeb34223167c 100644
> > --- a/scripts/recordmcount.c
> > +++ b/scripts/recordmcount.c
> > @@ -213,6 +213,59 @@ static int make_nop_x86(void *map, size_t const offset)
> > return 0;
> > }
> >
> > +static unsigned char ideal_nop4_arm_le[4] = { 0x00, 0x00, 0xa0, 0xe1 }; /*
> > mov r0, r0 */
> > +static unsigned char ideal_nop4_arm_be[4] = { 0xe1, 0xa0, 0x00, 0x00 }; /*
> > mov r0, r0 */
>
> Shouldn't you be taking the difference between BE8 and BE32 into
> account here? IIRC, BE8 uses little endian encoding for instructions.
I admit I haven't tested on a pre-armv6 CPU so I haven't come
across the case of a BE32 CPU. But from what I can tell that
doesn't matter.
According to scripts/Makefile.build, cmd_record_mcount only runs
the recordmcount program if CONFIG_FTRACE_MCOUNT_RECORD=y. That
config is defined as:
config FTRACE_MCOUNT_RECORD
def_bool y
depends on DYNAMIC_FTRACE
depends on HAVE_FTRACE_MCOUNT_RECORD
And in arch/arm/Kconfig we see that DYNAMIC_FTRACE is selected:
select HAVE_DYNAMIC_FTRACE if (!XIP_KERNEL) && !CPU_ENDIAN_BE32 && MMU
which means that FTRACE_MCOUNT_RECORD can't be set when
CPU_ENDIAN_BE32 is set.
Do you agree that BE32 is not a concern here?
--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
Re: [PATCH/RESEND] recordmcount: arm: Implement make_nop
On Tue, 15 Nov 2016 14:19:44 +
Ard Biesheuvel wrote:
> On 19 October 2016 at 00:42, Stephen Boyd wrote:
> > In similar spirit to x86 and arm64 support, add a make_nop_arm()
> > to replace calls to mcount with a nop in sections that aren't
> > traced.
> >
> > Cc: Russell King
> > Acked-by: Rabin Vincent
> > Signed-off-by: Stephen Boyd
> > ---
> > scripts/recordmcount.c | 65
> > ++
> > 1 file changed, 65 insertions(+)
> >
> > diff --git a/scripts/recordmcount.c b/scripts/recordmcount.c
> > index 5423a58d1b06..aeb34223167c 100644
> > --- a/scripts/recordmcount.c
> > +++ b/scripts/recordmcount.c
> > @@ -213,6 +213,59 @@ static int make_nop_x86(void *map, size_t const offset)
> > return 0;
> > }
> >
> > +static unsigned char ideal_nop4_arm_le[4] = { 0x00, 0x00, 0xa0, 0xe1 }; /*
> > mov r0, r0 */
> > +static unsigned char ideal_nop4_arm_be[4] = { 0xe1, 0xa0, 0x00, 0x00 }; /*
> > mov r0, r0 */
>
> Shouldn't you be taking the difference between BE8 and BE32 into
> account here? IIRC, BE8 uses little endian encoding for instructions.
>
I was just about to push this to linux-next (where I don't rebase). I'm
guessing I should hold off then.
Luckily, this was the last patch of my tree that I tested, and I can
just remove that one.
-- Steve
Re: [PATCH/RESEND] recordmcount: arm: Implement make_nop
On 19 October 2016 at 00:42, Stephen Boyd wrote:
> In similar spirit to x86 and arm64 support, add a make_nop_arm()
> to replace calls to mcount with a nop in sections that aren't
> traced.
>
> Cc: Russell King
> Acked-by: Rabin Vincent
> Signed-off-by: Stephen Boyd
> ---
> scripts/recordmcount.c | 65
> ++
> 1 file changed, 65 insertions(+)
>
> diff --git a/scripts/recordmcount.c b/scripts/recordmcount.c
> index 5423a58d1b06..aeb34223167c 100644
> --- a/scripts/recordmcount.c
> +++ b/scripts/recordmcount.c
> @@ -213,6 +213,59 @@ static int make_nop_x86(void *map, size_t const offset)
> return 0;
> }
>
> +static unsigned char ideal_nop4_arm_le[4] = { 0x00, 0x00, 0xa0, 0xe1 }; /*
> mov r0, r0 */
> +static unsigned char ideal_nop4_arm_be[4] = { 0xe1, 0xa0, 0x00, 0x00 }; /*
> mov r0, r0 */
Shouldn't you be taking the difference between BE8 and BE32 into
account here? IIRC, BE8 uses little endian encoding for instructions.
> +static unsigned char *ideal_nop4_arm;
> +
> +static unsigned char bl_mcount_arm_le[4] = { 0xfe, 0xff, 0xff, 0xeb }; /* bl
> */
> +static unsigned char bl_mcount_arm_be[4] = { 0xeb, 0xff, 0xff, 0xfe }; /* bl
> */
> +static unsigned char *bl_mcount_arm;
> +
> +static unsigned char push_arm_le[4] = { 0x04, 0xe0, 0x2d, 0xe5 }; /* push
> {lr} */
> +static unsigned char push_arm_be[4] = { 0xe5, 0x2d, 0xe0, 0x04 }; /* push
> {lr} */
> +static unsigned char *push_arm;
> +
> +static unsigned char ideal_nop2_thumb_le[2] = { 0x00, 0xbf }; /* nop */
> +static unsigned char ideal_nop2_thumb_be[2] = { 0xbf, 0x00 }; /* nop */
> +static unsigned char *ideal_nop2_thumb;
> +
> +static unsigned char push_bl_mcount_thumb_le[6] = { 0x00, 0xb5, 0xff, 0xf7,
> 0xfe, 0xff }; /* push {lr}, bl */
> +static unsigned char push_bl_mcount_thumb_be[6] = { 0xb5, 0x00, 0xf7, 0xff,
> 0xff, 0xfe }; /* push {lr}, bl */
> +static unsigned char *push_bl_mcount_thumb;
> +
> +static int make_nop_arm(void *map, size_t const offset)
> +{
> + char *ptr;
> + int cnt = 1;
> + int nop_size;
> + size_t off = offset;
> +
> + ptr = map + offset;
> + if (memcmp(ptr, bl_mcount_arm, 4) == 0) {
> + if (memcmp(ptr - 4, push_arm, 4) == 0) {
> + off -= 4;
> + cnt = 2;
> + }
> + ideal_nop = ideal_nop4_arm;
> + nop_size = 4;
> + } else if (memcmp(ptr - 2, push_bl_mcount_thumb, 6) == 0) {
> + cnt = 3;
> + nop_size = 2;
> + off -= 2;
> + ideal_nop = ideal_nop2_thumb;
> + } else
> + return -1;
> +
> + /* Convert to nop */
> + ulseek(fd_map, off, SEEK_SET);
> +
> + do {
> + uwrite(fd_map, ideal_nop, nop_size);
> + } while (--cnt > 0);
> +
> + return 0;
> +}
> +
> static unsigned char ideal_nop4_arm64[4] = {0x1f, 0x20, 0x03, 0xd5};
> static int make_nop_arm64(void *map, size_t const offset)
> {
> @@ -430,6 +483,11 @@ do_file(char const *const fname)
> w2 = w2rev;
> w8 = w8rev;
> }
> + ideal_nop4_arm = ideal_nop4_arm_le;
> + bl_mcount_arm = bl_mcount_arm_le;
> + push_arm = push_arm_le;
> + ideal_nop2_thumb = ideal_nop2_thumb_le;
> + push_bl_mcount_thumb = push_bl_mcount_thumb_le;
> break;
> case ELFDATA2MSB:
> if (*(unsigned char const *)&endian != 0) {
> @@ -438,6 +496,11 @@ do_file(char const *const fname)
> w2 = w2rev;
> w8 = w8rev;
> }
> + ideal_nop4_arm = ideal_nop4_arm_be;
> + bl_mcount_arm = bl_mcount_arm_be;
> + push_arm = push_arm_be;
> + ideal_nop2_thumb = ideal_nop2_thumb_be;
> + push_bl_mcount_thumb = push_bl_mcount_thumb_be;
> break;
> } /* end switch */
> if (memcmp(ELFMAG, ehdr->e_ident, SELFMAG) != 0
> @@ -463,6 +526,8 @@ do_file(char const *const fname)
> break;
> case EM_ARM: reltype = R_ARM_ABS32;
> altmcount = "__gnu_mcount_nc";
> +make_nop = make_nop_arm;
> +rel_type_nop = R_ARM_NONE;
> break;
> case EM_AARCH64:
> reltype = R_AARCH64_ABS64;
> --
> The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
> a Linux Foundation Collaborative Project
>
>
> ___
> linux-arm-kernel mailing list
> linux-arm-ker...@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Re: [PATCH/RESEND] recordmcount: arm: Implement make_nop
On Mon, Nov 14, 2016 at 01:36:41PM -0500, Steven Rostedt wrote: > On Tue, 18 Oct 2016 20:07:07 -0400 > Steven Rostedt wrote: > > > On Tue, 18 Oct 2016 16:42:00 -0700 > > Stephen Boyd wrote: > > > > > In similar spirit to x86 and arm64 support, add a make_nop_arm() > > > to replace calls to mcount with a nop in sections that aren't > > > traced. > > > > > > Cc: Russell King > > > Acked-by: Rabin Vincent > > > Signed-off-by: Stephen Boyd > > > > I can take this if I can get an ack from the ARM maintainers. > > Any ARM maintainer want to ack this, or take it in their tree if they > haven't already? Assuming it's been tested: Acked-by: Russell King -- RMK's Patch system: http://www.armlinux.org.uk/developer/patches/ FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up according to speedtest.net.
Re: [PATCH/RESEND] recordmcount: arm: Implement make_nop
On Tue, 18 Oct 2016 20:07:07 -0400 Steven Rostedt wrote: > On Tue, 18 Oct 2016 16:42:00 -0700 > Stephen Boyd wrote: > > > In similar spirit to x86 and arm64 support, add a make_nop_arm() > > to replace calls to mcount with a nop in sections that aren't > > traced. > > > > Cc: Russell King > > Acked-by: Rabin Vincent > > Signed-off-by: Stephen Boyd > > I can take this if I can get an ack from the ARM maintainers. Any ARM maintainer want to ack this, or take it in their tree if they haven't already? -- Steve
Re: [PATCH/RESEND] recordmcount: arm: Implement make_nop
On Tue, 18 Oct 2016 16:42:00 -0700 Stephen Boyd wrote: > In similar spirit to x86 and arm64 support, add a make_nop_arm() > to replace calls to mcount with a nop in sections that aren't > traced. > > Cc: Russell King > Acked-by: Rabin Vincent > Signed-off-by: Stephen Boyd I can take this if I can get an ack from the ARM maintainers. -- Steve
