The Cavium Octeon CPU uses a special sync instruction for implementing
wmb, and due to a CPU bug, the instruction must appear twice. A macro
had been defined to hide this:
#define __SYNC_rpt(type) (1 + (type == __SYNC_wmb))
which was intended to evaluate to 2 for __SYNC_wmb, and 1 for any other
type of sync. However, this expression is evaluated by the assembler,
and not the compiler, and the result of '==' in the assembler is 0 or
-1, not 0 or 1 as it is in C. The net result was wmb() producing no code
at all. The simple fix in this patch is to change the '+' to '-'.
Fixes: bf92927251b3 ("MIPS: barrier: Add __SYNC() infrastructure")
Signed-off-by: Mark Tomlinson <[email protected]>
---
arch/mips/include/asm/sync.h | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/arch/mips/include/asm/sync.h b/arch/mips/include/asm/sync.h
index 7c6a1095f5..aabd097933 100644
--- a/arch/mips/include/asm/sync.h
+++ b/arch/mips/include/asm/sync.h
@@ -155,9 +155,11 @@
* effective barrier as noted by commit 6b07d38aaa52 ("MIPS: Octeon: Use
* optimized memory barrier primitives."). Here we specify that the affected
* sync instructions should be emitted twice.
+ * Note that this expression is evaluated by the assembler (not the compiler),
+ * and that the assembler evaluates '==' as 0 or -1, not 0 or 1.
*/
#ifdef CONFIG_CPU_CAVIUM_OCTEON
-# define __SYNC_rpt(type) (1 + (type == __SYNC_wmb))
+# define __SYNC_rpt(type) (1 - (type == __SYNC_wmb))
#else
# define __SYNC_rpt(type) 1
#endif
--
2.25.0
Hello,
Mark Tomlinson wrote:
> The Cavium Octeon CPU uses a special sync instruction for implementing
> wmb, and due to a CPU bug, the instruction must appear twice. A macro
> had been defined to hide this:
>
> #define __SYNC_rpt(type) (1 + (type == __SYNC_wmb))
>
> which was intended to evaluate to 2 for __SYNC_wmb, and 1 for any other
> type of sync. However, this expression is evaluated by the assembler,
> and not the compiler, and the result of '==' in the assembler is 0 or
> -1, not 0 or 1 as it is in C. The net result was wmb() producing no code
> at all. The simple fix in this patch is to change the '+' to '-'.
Applied to mips-fixes.
> commit 97e914b7de3c
> https://git.kernel.org/mips/c/97e914b7de3c
>
> Fixes: bf92927251b3 ("MIPS: barrier: Add __SYNC() infrastructure")
> Signed-off-by: Mark Tomlinson <[email protected]>
> Tested-by: Chris Packham <[email protected]>
> Signed-off-by: Paul Burton <[email protected]>
Thanks,
Paul
[ This message was auto-generated; if you believe anything is incorrect
then please email [email protected] to report it. ]
Hi Mark,
On Tue, Feb 11, 2020 at 10:42 PM Mark Tomlinson
<[email protected]> wrote:
>
> The Cavium Octeon CPU uses a special sync instruction for implementing
> wmb, and due to a CPU bug, the instruction must appear twice. A macro
> had been defined to hide this:
>
> #define __SYNC_rpt(type) (1 + (type == __SYNC_wmb))
>
> which was intended to evaluate to 2 for __SYNC_wmb, and 1 for any other
> type of sync. However, this expression is evaluated by the assembler,
> and not the compiler, and the result of '==' in the assembler is 0 or
> -1, not 0 or 1 as it is in C. The net result was wmb() producing no code
> at all. The simple fix in this patch is to change the '+' to '-'.
Isn't this particular to the assembler implementation?
Can you explicit the assembler you are using in the commit description?
Assuming we have to look at your commit in 3 years from now, we'll
wonder what assembler you were using.
Thanks,
Phil.
> Fixes: bf92927251b3 ("MIPS: barrier: Add __SYNC() infrastructure")
> Signed-off-by: Mark Tomlinson <[email protected]>
> ---
> arch/mips/include/asm/sync.h | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/arch/mips/include/asm/sync.h b/arch/mips/include/asm/sync.h
> index 7c6a1095f5..aabd097933 100644
> --- a/arch/mips/include/asm/sync.h
> +++ b/arch/mips/include/asm/sync.h
> @@ -155,9 +155,11 @@
> * effective barrier as noted by commit 6b07d38aaa52 ("MIPS: Octeon: Use
> * optimized memory barrier primitives."). Here we specify that the affected
> * sync instructions should be emitted twice.
> + * Note that this expression is evaluated by the assembler (not the compiler),
> + * and that the assembler evaluates '==' as 0 or -1, not 0 or 1.
> */
> #ifdef CONFIG_CPU_CAVIUM_OCTEON
> -# define __SYNC_rpt(type) (1 + (type == __SYNC_wmb))
> +# define __SYNC_rpt(type) (1 - (type == __SYNC_wmb))
> #else
> # define __SYNC_rpt(type) 1
> #endif
> --
> 2.25.0
>
Hi Phil,
On Mon, 2020-02-17 at 01:22 +0100, Philippe Mathieu-Daudé wrote:
> Hi Mark,
>
> On Tue, Feb 11, 2020 at 10:42 PM Mark Tomlinson
> <[email protected]> wrote:
> >
> > The Cavium Octeon CPU uses a special sync instruction for implementing
> > wmb, and due to a CPU bug, the instruction must appear twice. A macro
> > had been defined to hide this:
> >
> > #define __SYNC_rpt(type) (1 + (type == __SYNC_wmb))
> >
> > which was intended to evaluate to 2 for __SYNC_wmb, and 1 for any other
> > type of sync. However, this expression is evaluated by the assembler,
> > and not the compiler, and the result of '==' in the assembler is 0 or
> > -1, not 0 or 1 as it is in C. The net result was wmb() producing no code
> > at all. The simple fix in this patch is to change the '+' to '-'.
>
> Isn't this particular to the assembler implementation?
> Can you explicit the assembler you are using in the commit description?
> Assuming we have to look at your commit in 3 years from now, we'll
> wonder what assembler you were using.
>
> Thanks,
>
> Phil.
Yes, it is tied to the assembler. But the Linux kernel is tied to GCC,
and GCC (I believe) is tied to GNU as. I can't see the specification of
GNU as changing, since that could break anything written for it.
> > Fixes: bf92927251b3 ("MIPS: barrier: Add __SYNC() infrastructure")
> > Signed-off-by: Mark Tomlinson <[email protected]>
> > ---
> > arch/mips/include/asm/sync.h | 4 +++-
> > 1 file changed, 3 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/mips/include/asm/sync.h b/arch/mips/include/asm/sync.h
> > index 7c6a1095f5..aabd097933 100644
> > --- a/arch/mips/include/asm/sync.h
> > +++ b/arch/mips/include/asm/sync.h
> > @@ -155,9 +155,11 @@
> > * effective barrier as noted by commit 6b07d38aaa52 ("MIPS: Octeon: Use
> > * optimized memory barrier primitives."). Here we specify that the affected
> > * sync instructions should be emitted twice.
> > + * Note that this expression is evaluated by the assembler (not the compiler),
> > + * and that the assembler evaluates '==' as 0 or -1, not 0 or 1.
> > */
> > #ifdef CONFIG_CPU_CAVIUM_OCTEON
> > -# define __SYNC_rpt(type) (1 + (type == __SYNC_wmb))
> > +# define __SYNC_rpt(type) (1 - (type == __SYNC_wmb))
> > #else
> > # define __SYNC_rpt(type) 1
> > #endif
> > --
> > 2.25.0
> >
On Mon, 2020-02-17 at 17:58 +1300, Mark Tomlinson wrote:
> Hi Phil,
>
> On Mon, 2020-02-17 at 01:22 +0100, Philippe Mathieu-Daudé wrote:
> > Hi Mark,
> >
> > On Tue, Feb 11, 2020 at 10:42 PM Mark Tomlinson
> > <[email protected]> wrote:
> > >
> > > The Cavium Octeon CPU uses a special sync instruction for implementing
> > > wmb, and due to a CPU bug, the instruction must appear twice. A macro
> > > had been defined to hide this:
> > >
> > > #define __SYNC_rpt(type) (1 + (type == __SYNC_wmb))
> > >
> > > which was intended to evaluate to 2 for __SYNC_wmb, and 1 for any other
> > > type of sync. However, this expression is evaluated by the assembler,
> > > and not the compiler, and the result of '==' in the assembler is 0 or
> > > -1, not 0 or 1 as it is in C. The net result was wmb() producing no code
> > > at all. The simple fix in this patch is to change the '+' to '-'.
> >
> > Isn't this particular to the assembler implementation?
> > Can you explicit the assembler you are using in the commit description?
> > Assuming we have to look at your commit in 3 years from now, we'll
> > wonder what assembler you were using.
> >
> > Thanks,
> >
> > Phil.
>
> Yes, it is tied to the assembler. But the Linux kernel is tied to GCC,
> and GCC (I believe) is tied to GNU as. I can't see the specification of
> GNU as changing, since that could break anything written for it.
>
There is an effort underway to build the kernel with clang[1]. I'm not
sure what that ends up using for an assembler or if it'll even be able
to target mips64 anytime soon.
For reference the relevant section from the GNU as manual[2] says "A
true results has a value of -1 whereas a false result has a value of
0".
[1] - https://clangbuiltlinux.github.io/
[2] - https://sourceware.org/binutils/docs/as/Infix-Ops.html#Infix-Ops
>
> > > Fixes: bf92927251b3 ("MIPS: barrier: Add __SYNC() infrastructure")
> > > Signed-off-by: Mark Tomlinson <[email protected]>
> > > ---
> > > arch/mips/include/asm/sync.h | 4 +++-
> > > 1 file changed, 3 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/arch/mips/include/asm/sync.h b/arch/mips/include/asm/sync.h
> > > index 7c6a1095f5..aabd097933 100644
> > > --- a/arch/mips/include/asm/sync.h
> > > +++ b/arch/mips/include/asm/sync.h
> > > @@ -155,9 +155,11 @@
> > > * effective barrier as noted by commit 6b07d38aaa52 ("MIPS: Octeon: Use
> > > * optimized memory barrier primitives."). Here we specify that the affected
> > > * sync instructions should be emitted twice.
> > > + * Note that this expression is evaluated by the assembler (not the compiler),
> > > + * and that the assembler evaluates '==' as 0 or -1, not 0 or 1.
> > > */
> > > #ifdef CONFIG_CPU_CAVIUM_OCTEON
> > > -# define __SYNC_rpt(type) (1 + (type == __SYNC_wmb))
> > > +# define __SYNC_rpt(type) (1 - (type == __SYNC_wmb))
> > > #else
> > > # define __SYNC_rpt(type) 1
> > > #endif
> > > --
> > > 2.25.0
> > >
>
>
On Mon, Feb 17, 2020 at 12:01 PM Chris Packham
<[email protected]> wrote:
>
> On Mon, 2020-02-17 at 17:58 +1300, Mark Tomlinson wrote:
> > Hi Phil,
> >
> > On Mon, 2020-02-17 at 01:22 +0100, Philippe Mathieu-Daudé wrote:
> > > Hi Mark,
> > >
> > > On Tue, Feb 11, 2020 at 10:42 PM Mark Tomlinson
> > > <[email protected]> wrote:
> > > >
> > > > The Cavium Octeon CPU uses a special sync instruction for implementing
> > > > wmb, and due to a CPU bug, the instruction must appear twice. A macro
> > > > had been defined to hide this:
> > > >
> > > > #define __SYNC_rpt(type) (1 + (type == __SYNC_wmb))
> > > >
> > > > which was intended to evaluate to 2 for __SYNC_wmb, and 1 for any other
> > > > type of sync. However, this expression is evaluated by the assembler,
> > > > and not the compiler, and the result of '==' in the assembler is 0 or
> > > > -1, not 0 or 1 as it is in C. The net result was wmb() producing no code
> > > > at all. The simple fix in this patch is to change the '+' to '-'.
> > >
> > > Isn't this particular to the assembler implementation?
> > > Can you explicit the assembler you are using in the commit description?
> > > Assuming we have to look at your commit in 3 years from now, we'll
> > > wonder what assembler you were using.
> > >
> > > Thanks,
> > >
> > > Phil.
> >
> > Yes, it is tied to the assembler. But the Linux kernel is tied to GCC,
> > and GCC (I believe) is tied to GNU as. I can't see the specification of
> > GNU as changing, since that could break anything written for it.
> >
>
> There is an effort underway to build the kernel with clang[1]. I'm not
> sure what that ends up using for an assembler or if it'll even be able
> to target mips64 anytime soon.
>
> For reference the relevant section from the GNU as manual[2] says "A
> true results has a value of -1 whereas a false result has a value of
> 0".
>
> [1] - https://clangbuiltlinux.github.io/
> [2] - https://sourceware.org/binutils/docs/as/Infix-Ops.html#Infix-Ops
Chris, thanks for CC'ing us.
Mark, we're building 32 bit MIPS kernels with Clang under CI (just
added big endian builds this morning). We're actively looking into
supporting 64b MIPS.
The kernel uses GCC by default, but supports using any compiler via
`make CC=<foo>`. There is extensive support in the kernel for
building with Clang.
GCC and Clang (when doing kernel builds, for clang we set
`-no-integrated-as`) will invoke GAS for inline assembly, but you can
set `AS=clang` for example for the out of line assembly files. If the
C source files don't contain inline assembly (or `-no-integrated-as`
wasn't set) then Clang will skip invoking the assembler and stream out
an object file.
If you're actively supporting 64b mips, and want to give a Clang build
a try, we'd appreciate the bug reports:
https://github.com/ClangBuiltLinux/linux/issues
--
Thanks,
~Nick Desaulniers
18.02.2020, 20:37, "'Nick Desaulniers' via Clang Built Linux" <[email protected]>:
>> There is an effort underway to build the kernel with clang[1]. I'm not
>> sure what that ends up using for an assembler or if it'll even be able
>> to target mips64 anytime soon.
I have a working build for MIPS64 (only mips64r6 at the moment), the
config is based on malta_defconfig and it boots in qemu. I still don't
have a matching buildroot image, but it shouldn't be a problem.
Regards,
Dima