[v3,2/4] asm-generic: Provide a fncpy() implementation
diff mbox

Message ID 20170617000744.22158-3-f.fainelli@gmail.com
State New
Headers show

Commit Message

Florian Fainelli June 17, 2017, 12:07 a.m. UTC
Define a generic fncpy() implementation largely based on the ARM version
that requires an 8 bytes alignment for the destination address where to
copy this function as well as the function's own address.

Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
---
 include/asm-generic/fncpy.h | 93 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 93 insertions(+)
 create mode 100644 include/asm-generic/fncpy.h

Comments

Yury Norov June 18, 2017, 11:51 p.m. UTC | #1
Hi Florian,

Some questions and thoughts inline.

Yury

On Fri, Jun 16, 2017 at 05:07:42PM -0700, Florian Fainelli wrote:
> Define a generic fncpy() implementation largely based on the ARM version
> that requires an 8 bytes alignment for the destination address where to
> copy this function as well as the function's own address.
> 
> Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
> ---
>  include/asm-generic/fncpy.h | 93 +++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 93 insertions(+)
>  create mode 100644 include/asm-generic/fncpy.h
> 
> diff --git a/include/asm-generic/fncpy.h b/include/asm-generic/fncpy.h
> new file mode 100644
> index 000000000000..ec03b83b8535
> --- /dev/null
> +++ b/include/asm-generic/fncpy.h
> @@ -0,0 +1,93 @@
> +/*
> + * include/asm-generic/fncpy.h - helper macros for function body copying
> + *
> + * Copyright (C) 2011 Linaro Limited
> + *
> + * This program is free software; you can redistribute it and/or modify
> + * it under the terms of the GNU General Public License version 2 as
> + * published by the Free Software Foundation.
> + *
> + * This program is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
> + * GNU General Public License for more details.
> + *
> + * You should have received a copy of the GNU General Public License
> + * along with this program; if not, write to the Free Software
> + * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
> + */
> +
> +/*
> + * These macros are intended for use when there is a need to copy a low-level
> + * function body into special memory.
> + *
> + * For example, when reconfiguring the SDRAM controller, the code doing the
> + * reconfiguration may need to run from SRAM.
> + *
> + * NOTE: that the copied function body must be entirely self-contained and
> + * position-independent in order for this to work properly.
> + *
> + * NOTE: in order for embedded literals and data to get referenced correctly,
> + * the alignment of functions must be preserved when copying.  To ensure this,
> + * the source and destination addresses for fncpy() must be aligned to a
> + * multiple of 8 bytes: you will be get a BUG() if this condition is not met.
> + * You will typically need a ".align 3" directive in the assembler where the
> + * function to be copied is defined, and ensure that your allocator for the
> + * destination buffer returns 8-byte-aligned pointers.
> + *
> + * Typical usage example:
> + *
> + * extern int f(args);
> + * extern uint32_t size_of_f;
> + * int (*copied_f)(args);
> + * void *sram_buffer;
> + *
> + * copied_f = fncpy(sram_buffer, &f, size_of_f);
> + *
> + * ... later, call the function: ...
> + *
> + * copied_f(args);
> + *
> + * The size of the function to be copied can't be determined from C:
> + * this must be determined by other means, such as adding assmbler directives
> + * in the file where f is defined.
> + */
> +
> +#ifndef __ASM_FNCPY_H
> +#define __ASM_FNCPY_H
> +
> +#include <linux/types.h>
> +#include <linux/string.h>
> +
> +#include <asm/bug.h>
> +#include <asm/cacheflush.h>
> +
> +/*
> + * Minimum alignment requirement for the source and destination addresses
> + * for function copying.
> + */
> +#define FNCPY_ALIGN 8

From now this is not arm-only, and it's possible that some architectures
might want to redefine it in their arch/xxx/include/asm/fncpy.h files.
So it will be easier for them if you'll wrap FNCPY_ALIGN here with #ifdef
guards.

By the way, compiler already has an information on the proper alignment.
Maybe it's better to use it as the default value here instead of the
hardcoded value?

#ifndef FNCPY_ALIGN
#define FNCPY_ALIGN ({void foo(); __alignof__(&foo);})
#endif

> +
> +#define fncpy(dest_buf, funcp, size) ({					\

Do you really need to check types inside the macro? If not, you can
declare it as function, which is better in general, with the memcpy-like
prototype.

> +	uintptr_t __funcp_address;					\
> +	typeof(funcp) __result;						\
> +									\
> +	asm("" : "=r" (__funcp_address) : "0" (funcp));			\
> +									\
> +	/*								\
> +	 * Ensure alignment of source and destination addresses.	\
> +	 */								\
> +	BUG_ON((uintptr_t)(dest_buf) & (FNCPY_ALIGN - 1) ||		\

People don't like new BUG_ONs. Maybe it's better to use BUILD_BUG_ON()
at compile time and WARN_ON() at runtime?

> +		(__funcp_address & (FNCPY_ALIGN - 1)));			\

There is IS_ALIGNED() macro for things like this.

And I frankly don't understand the 2nd check. One can imagine the
situation when someone wants copy the function from the packed blob or
some intermediate location were the function is unaligned, and it's
impossible with the current implementation.

> +									\
> +	memcpy(dest_buf, (void const *)__funcp_address, size);		\
> +	flush_icache_range((unsigned long)(dest_buf),			\
> +		(unsigned long)(dest_buf) + (size));			\
> +									\
> +	asm("" : "=r" (__result)					\
> +		: "0" ((uintptr_t)(dest_buf)));				\
> +									\
> +	__result;							\
> +})
> +
> +#endif /* !__ASM_FNCPY_H */
> -- 
> 2.9.3
--
To unsubscribe from this list: send the line "unsubscribe linux-omap" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Yury Norov June 19, 2017, 1:11 a.m. UTC | #2
> > +/*
> > + * Minimum alignment requirement for the source and destination addresses
> > + * for function copying.
> > + */
> > +#define FNCPY_ALIGN 8
> 
> >From now this is not arm-only, and it's possible that some architectures
> might want to redefine it in their arch/xxx/include/asm/fncpy.h files.
> So it will be easier for them if you'll wrap FNCPY_ALIGN here with #ifdef
> guards.
> 
> By the way, compiler already has an information on the proper alignment.
> Maybe it's better to use it as the default value here instead of the
> hardcoded value?
> 
> #ifndef FNCPY_ALIGN
> #define FNCPY_ALIGN ({void foo(); __alignof__(&foo);})
> #endif

Ah sorry, at first it should be like this:
#define FNCPY_ALIGN ({void foo(); __alignof__(foo);})

And at second, the correct version returns 1 always.

Even if I pass falign-functions=4096 to gcc, and I see that functions
are aligned accordingly in elf file, the macro returns 1 anyway. So if 
it doesn't work, the hardcoded '8' is the only option.

Yury
--
To unsubscribe from this list: send the line "unsubscribe linux-omap" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Yury Norov June 19, 2017, 3:18 p.m. UTC | #3
On Mon, Jun 19, 2017 at 02:51:08AM +0300, Yury Norov wrote:
> Hi Florian,
> 
> Some questions and thoughts inline.
> 
> Yury
> 
> On Fri, Jun 16, 2017 at 05:07:42PM -0700, Florian Fainelli wrote:
> > Define a generic fncpy() implementation largely based on the ARM version
> > that requires an 8 bytes alignment for the destination address where to
> > copy this function as well as the function's own address.
> > 
> > Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
> > ---
> >  include/asm-generic/fncpy.h | 93 +++++++++++++++++++++++++++++++++++++++++++++
> >  1 file changed, 93 insertions(+)
> >  create mode 100644 include/asm-generic/fncpy.h

One else thing I forgot to ask - now you have the generic
implementation for fncpy(), so do you really need to save arm
version of it?

Yury
--
To unsubscribe from this list: send the line "unsubscribe linux-omap" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Florian Fainelli June 19, 2017, 5:27 p.m. UTC | #4
On 06/19/2017 08:18 AM, Yury Norov wrote:
> On Mon, Jun 19, 2017 at 02:51:08AM +0300, Yury Norov wrote:
>> Hi Florian,
>>
>> Some questions and thoughts inline.
>>
>> Yury
>>
>> On Fri, Jun 16, 2017 at 05:07:42PM -0700, Florian Fainelli wrote:
>>> Define a generic fncpy() implementation largely based on the ARM version
>>> that requires an 8 bytes alignment for the destination address where to
>>> copy this function as well as the function's own address.
>>>
>>> Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
>>> ---
>>>  include/asm-generic/fncpy.h | 93 +++++++++++++++++++++++++++++++++++++++++++++
>>>  1 file changed, 93 insertions(+)
>>>  create mode 100644 include/asm-generic/fncpy.h
> 
> One else thing I forgot to ask - now you have the generic
> implementation for fncpy(), so do you really need to save arm
> version of it?

Yes, it needs to deal with the Thumb bit, and there is no point in
making the generic implementation extremely flexible to support that
special case.

Thanks
Russell King - ARM Linux admin June 19, 2017, 5:43 p.m. UTC | #5
On Mon, Jun 19, 2017 at 06:18:18PM +0300, Yury Norov wrote:
> One else thing I forgot to ask - now you have the generic
> implementation for fncpy(), so do you really need to save arm
> version of it?

This was covered in the review of v1, which took the ARM version
and incorrectly used it as an asm-generic implementation.

I explicitly asked Florian _not_ to copy the ARM fncpy() version
to asm-generic because it has (surprise surprise) ARM specific
behaviours that do not belong in a cross-architecture generic
version.

Namely, the ARM specific behaviour that bit 0 of a code address is
used to signal whether the code should be executed as ARM code or
as Thumb code.

This behaviour has no meaning on other architectures (eg, x86)
where code addresses are not 32-bit aligned.

So, suggesting that the ARM fncpy() should be used as an asm-generic
version is completely absurd, and just because we have an asm-generic
version also does not mean ARM should use it.

Florian's approach to providing an asm-generic version, leaving the
ARM specific version is entirely correct and appropriate.

So, in answer to your question, yes, we need _both_ an ARM specific
version and an asm-generic version, where the ARM specific version is
different from the asm-generic version.  Purely because it needs
architecture specific details.
Florian Fainelli June 19, 2017, 8:58 p.m. UTC | #6
On 06/18/2017 04:51 PM, Yury Norov wrote:
> Hi Florian,
> 
> Some questions and thoughts inline.
> 
> Yury
> 
> On Fri, Jun 16, 2017 at 05:07:42PM -0700, Florian Fainelli wrote:
>> Define a generic fncpy() implementation largely based on the ARM version
>> that requires an 8 bytes alignment for the destination address where to
>> copy this function as well as the function's own address.
>>
>> Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
>> ---
>>  include/asm-generic/fncpy.h | 93 +++++++++++++++++++++++++++++++++++++++++++++
>>  1 file changed, 93 insertions(+)
>>  create mode 100644 include/asm-generic/fncpy.h
>>
>> diff --git a/include/asm-generic/fncpy.h b/include/asm-generic/fncpy.h
>> new file mode 100644
>> index 000000000000..ec03b83b8535
>> --- /dev/null
>> +++ b/include/asm-generic/fncpy.h
>> @@ -0,0 +1,93 @@
>> +/*
>> + * include/asm-generic/fncpy.h - helper macros for function body copying
>> + *
>> + * Copyright (C) 2011 Linaro Limited
>> + *
>> + * This program is free software; you can redistribute it and/or modify
>> + * it under the terms of the GNU General Public License version 2 as
>> + * published by the Free Software Foundation.
>> + *
>> + * This program is distributed in the hope that it will be useful,
>> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
>> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
>> + * GNU General Public License for more details.
>> + *
>> + * You should have received a copy of the GNU General Public License
>> + * along with this program; if not, write to the Free Software
>> + * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
>> + */
>> +
>> +/*
>> + * These macros are intended for use when there is a need to copy a low-level
>> + * function body into special memory.
>> + *
>> + * For example, when reconfiguring the SDRAM controller, the code doing the
>> + * reconfiguration may need to run from SRAM.
>> + *
>> + * NOTE: that the copied function body must be entirely self-contained and
>> + * position-independent in order for this to work properly.
>> + *
>> + * NOTE: in order for embedded literals and data to get referenced correctly,
>> + * the alignment of functions must be preserved when copying.  To ensure this,
>> + * the source and destination addresses for fncpy() must be aligned to a
>> + * multiple of 8 bytes: you will be get a BUG() if this condition is not met.
>> + * You will typically need a ".align 3" directive in the assembler where the
>> + * function to be copied is defined, and ensure that your allocator for the
>> + * destination buffer returns 8-byte-aligned pointers.
>> + *
>> + * Typical usage example:
>> + *
>> + * extern int f(args);
>> + * extern uint32_t size_of_f;
>> + * int (*copied_f)(args);
>> + * void *sram_buffer;
>> + *
>> + * copied_f = fncpy(sram_buffer, &f, size_of_f);
>> + *
>> + * ... later, call the function: ...
>> + *
>> + * copied_f(args);
>> + *
>> + * The size of the function to be copied can't be determined from C:
>> + * this must be determined by other means, such as adding assmbler directives
>> + * in the file where f is defined.
>> + */
>> +
>> +#ifndef __ASM_FNCPY_H
>> +#define __ASM_FNCPY_H
>> +
>> +#include <linux/types.h>
>> +#include <linux/string.h>
>> +
>> +#include <asm/bug.h>
>> +#include <asm/cacheflush.h>
>> +
>> +/*
>> + * Minimum alignment requirement for the source and destination addresses
>> + * for function copying.
>> + */
>> +#define FNCPY_ALIGN 8
> 
> From now this is not arm-only, and it's possible that some architectures
> might want to redefine it in their arch/xxx/include/asm/fncpy.h files.
> So it will be easier for them if you'll wrap FNCPY_ALIGN here with #ifdef
> guards.
> 
> By the way, compiler already has an information on the proper alignment.
> Maybe it's better to use it as the default value here instead of the
> hardcoded value?
> 
> #ifndef FNCPY_ALIGN
> #define FNCPY_ALIGN ({void foo(); __alignof__(&foo);})
> #endif
> 
>> +
>> +#define fncpy(dest_buf, funcp, size) ({					\
> 
> Do you really need to check types inside the macro? If not, you can
> declare it as function, which is better in general, with the memcpy-like
> prototype.
> 
>> +	uintptr_t __funcp_address;					\
>> +	typeof(funcp) __result;						\
>> +									\
>> +	asm("" : "=r" (__funcp_address) : "0" (funcp));			\
>> +									\
>> +	/*								\
>> +	 * Ensure alignment of source and destination addresses.	\
>> +	 */								\
>> +	BUG_ON((uintptr_t)(dest_buf) & (FNCPY_ALIGN - 1) ||		\
> 
> People don't like new BUG_ONs. Maybe it's better to use BUILD_BUG_ON()
> at compile time and WARN_ON() at runtime?

If you have a BUILD_BUG_ON() what's the point of the WARN_ON()?

> 
>> +		(__funcp_address & (FNCPY_ALIGN - 1)));			\
> 
> There is IS_ALIGNED() macro for things like this.

Sure, makes sense.

> 
> And I frankly don't understand the 2nd check. One can imagine the
> situation when someone wants copy the function from the packed blob or
> some intermediate location were the function is unaligned, and it's
> impossible with the current implementation.

That's a good point, I am not sure if this is historical, or if there is
a reason for that from the ARM/Linux implementation. It sounds unlikely
that the function would be unaligned though considering that you'd have
to refer to the function being copied by its symbolic name, which
assumes it's in the kernel image or a module, and highly probable that
it is also aligned.

> 
>> +									\
>> +	memcpy(dest_buf, (void const *)__funcp_address, size);		\
>> +	flush_icache_range((unsigned long)(dest_buf),			\
>> +		(unsigned long)(dest_buf) + (size));			\
>> +									\
>> +	asm("" : "=r" (__result)					\
>> +		: "0" ((uintptr_t)(dest_buf)));				\
>> +									\
>> +	__result;							\
>> +})
>> +
>> +#endif /* !__ASM_FNCPY_H */
>> -- 
>> 2.9.3
Yury Norov June 20, 2017, 2:24 p.m. UTC | #7
On Mon, Jun 19, 2017 at 01:58:53PM -0700, Florian Fainelli wrote:
> On 06/18/2017 04:51 PM, Yury Norov wrote:
> > Hi Florian,
> > 
> > Some questions and thoughts inline.
> > 
> > Yury
> > 
> > On Fri, Jun 16, 2017 at 05:07:42PM -0700, Florian Fainelli wrote:
> >> Define a generic fncpy() implementation largely based on the ARM version
> >> that requires an 8 bytes alignment for the destination address where to
> >> copy this function as well as the function's own address.
> >>
> >> Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
> >> ---
> >>  include/asm-generic/fncpy.h | 93 +++++++++++++++++++++++++++++++++++++++++++++
> >>  1 file changed, 93 insertions(+)
> >>  create mode 100644 include/asm-generic/fncpy.h
> >>
> >> diff --git a/include/asm-generic/fncpy.h b/include/asm-generic/fncpy.h
> >> new file mode 100644
> >> index 000000000000..ec03b83b8535
> >> --- /dev/null
> >> +++ b/include/asm-generic/fncpy.h
> >> @@ -0,0 +1,93 @@
> >> +/*
> >> + * include/asm-generic/fncpy.h - helper macros for function body copying
> >> + *
> >> + * Copyright (C) 2011 Linaro Limited
> >> + *
> >> + * This program is free software; you can redistribute it and/or modify
> >> + * it under the terms of the GNU General Public License version 2 as
> >> + * published by the Free Software Foundation.
> >> + *
> >> + * This program is distributed in the hope that it will be useful,
> >> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> >> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
> >> + * GNU General Public License for more details.
> >> + *
> >> + * You should have received a copy of the GNU General Public License
> >> + * along with this program; if not, write to the Free Software
> >> + * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
> >> + */
> >> +
> >> +/*
> >> + * These macros are intended for use when there is a need to copy a low-level
> >> + * function body into special memory.
> >> + *
> >> + * For example, when reconfiguring the SDRAM controller, the code doing the
> >> + * reconfiguration may need to run from SRAM.
> >> + *
> >> + * NOTE: that the copied function body must be entirely self-contained and
> >> + * position-independent in order for this to work properly.
> >> + *
> >> + * NOTE: in order for embedded literals and data to get referenced correctly,
> >> + * the alignment of functions must be preserved when copying.  To ensure this,
> >> + * the source and destination addresses for fncpy() must be aligned to a
> >> + * multiple of 8 bytes: you will be get a BUG() if this condition is not met.
> >> + * You will typically need a ".align 3" directive in the assembler where the
> >> + * function to be copied is defined, and ensure that your allocator for the
> >> + * destination buffer returns 8-byte-aligned pointers.
> >> + *
> >> + * Typical usage example:
> >> + *
> >> + * extern int f(args);
> >> + * extern uint32_t size_of_f;
> >> + * int (*copied_f)(args);
> >> + * void *sram_buffer;
> >> + *
> >> + * copied_f = fncpy(sram_buffer, &f, size_of_f);
> >> + *
> >> + * ... later, call the function: ...
> >> + *
> >> + * copied_f(args);
> >> + *
> >> + * The size of the function to be copied can't be determined from C:
> >> + * this must be determined by other means, such as adding assmbler directives
> >> + * in the file where f is defined.
> >> + */
> >> +
> >> +#ifndef __ASM_FNCPY_H
> >> +#define __ASM_FNCPY_H
> >> +
> >> +#include <linux/types.h>
> >> +#include <linux/string.h>
> >> +
> >> +#include <asm/bug.h>
> >> +#include <asm/cacheflush.h>
> >> +
> >> +/*
> >> + * Minimum alignment requirement for the source and destination addresses
> >> + * for function copying.
> >> + */
> >> +#define FNCPY_ALIGN 8
> > 
> > From now this is not arm-only, and it's possible that some architectures
> > might want to redefine it in their arch/xxx/include/asm/fncpy.h files.
> > So it will be easier for them if you'll wrap FNCPY_ALIGN here with #ifdef
> > guards.
> > 
> > By the way, compiler already has an information on the proper alignment.
> > Maybe it's better to use it as the default value here instead of the
> > hardcoded value?
> > 
> > #ifndef FNCPY_ALIGN
> > #define FNCPY_ALIGN ({void foo(); __alignof__(&foo);})
> > #endif
> > 
> >> +
> >> +#define fncpy(dest_buf, funcp, size) ({					\
> > 
> > Do you really need to check types inside the macro? If not, you can
> > declare it as function, which is better in general, with the memcpy-like
> > prototype.
> > 
> >> +	uintptr_t __funcp_address;					\
> >> +	typeof(funcp) __result;						\
> >> +									\
> >> +	asm("" : "=r" (__funcp_address) : "0" (funcp));			\
> >> +									\
> >> +	/*								\
> >> +	 * Ensure alignment of source and destination addresses.	\
> >> +	 */								\
> >> +	BUG_ON((uintptr_t)(dest_buf) & (FNCPY_ALIGN - 1) ||		\
> > 
> > People don't like new BUG_ONs. Maybe it's better to use BUILD_BUG_ON()
> > at compile time and WARN_ON() at runtime?
> 
> If you have a BUILD_BUG_ON() what's the point of the WARN_ON()?

To warn user if bad destination address comes at runtime.

> > 
> >> +		(__funcp_address & (FNCPY_ALIGN - 1)));			\
> > 
> > There is IS_ALIGNED() macro for things like this.
> 
> Sure, makes sense.
> 
> > 
> > And I frankly don't understand the 2nd check. One can imagine the
> > situation when someone wants copy the function from the packed blob or
> > some intermediate location were the function is unaligned, and it's
> > impossible with the current implementation.
> 
> That's a good point, I am not sure if this is historical, or if there is
> a reason for that from the ARM/Linux implementation. It sounds unlikely
> that the function would be unaligned though considering that you'd have
> to refer to the function being copied by its symbolic name, which
> assumes it's in the kernel image or a module, and highly probable that
> it is also aligned.
> 
> > 
> >> +									\
> >> +	memcpy(dest_buf, (void const *)__funcp_address, size);		\
> >> +	flush_icache_range((unsigned long)(dest_buf),			\
> >> +		(unsigned long)(dest_buf) + (size));			\
> >> +									\
> >> +	asm("" : "=r" (__result)					\
> >> +		: "0" ((uintptr_t)(dest_buf)));				\
> >> +									\
> >> +	__result;							\
> >> +})
> >> +
> >> +#endif /* !__ASM_FNCPY_H */
> >> -- 
> >> 2.9.3
> 
> 
> -- 
> Florian
--
To unsubscribe from this list: send the line "unsubscribe linux-omap" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Yury Norov June 20, 2017, 2:27 p.m. UTC | #8
On Mon, Jun 19, 2017 at 06:43:48PM +0100, Russell King - ARM Linux wrote:
> On Mon, Jun 19, 2017 at 06:18:18PM +0300, Yury Norov wrote:
> > One else thing I forgot to ask - now you have the generic
> > implementation for fncpy(), so do you really need to save arm
> > version of it?
> 
> This was covered in the review of v1, which took the ARM version
> and incorrectly used it as an asm-generic implementation.
> 
> I explicitly asked Florian _not_ to copy the ARM fncpy() version
> to asm-generic because it has (surprise surprise) ARM specific
> behaviours that do not belong in a cross-architecture generic
> version.
> 
> Namely, the ARM specific behaviour that bit 0 of a code address is
> used to signal whether the code should be executed as ARM code or
> as Thumb code.
> 
> This behaviour has no meaning on other architectures (eg, x86)
> where code addresses are not 32-bit aligned.
> 
> So, suggesting that the ARM fncpy() should be used as an asm-generic
> version is completely absurd, and just because we have an asm-generic
> version also does not mean ARM should use it.
> 
> Florian's approach to providing an asm-generic version, leaving the
> ARM specific version is entirely correct and appropriate.
> 
> So, in answer to your question, yes, we need _both_ an ARM specific
> version and an asm-generic version, where the ARM specific version is
> different from the asm-generic version.  Purely because it needs
> architecture specific details.

Hi Russell, Florian,

Thanks for clarifications. Thumb bit is a good reason to save arm
version, and I completely agree with you in this. Sorry that missed it
in the v1 discussion.

> I explicitly asked Florian _not_ to copy the ARM fncpy() version
> to asm-generic because it has (surprise surprise) ARM specific
> behaviours that do not belong in a cross-architecture generic
> version.

But it seems that v3 does exactly that - copies arm with very small
changes. :) Maybe there are good reasons to have arm version exactly
how it looks now, but in general case, for me, some things that
it does are not needed. I mean checking the alignment of the source and
the type of destination. And after some headscratching I became even
more convinced that for the general case it would be much preferable
to write the fncpy() as regular function in .c file, not a macro, at
least to have the corresponding symbol in binary and let the assembler
code to call it, which is very probable.

Yury
--
To unsubscribe from this list: send the line "unsubscribe linux-omap" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch
diff mbox

diff --git a/include/asm-generic/fncpy.h b/include/asm-generic/fncpy.h
new file mode 100644
index 000000000000..ec03b83b8535
--- /dev/null
+++ b/include/asm-generic/fncpy.h
@@ -0,0 +1,93 @@ 
+/*
+ * include/asm-generic/fncpy.h - helper macros for function body copying
+ *
+ * Copyright (C) 2011 Linaro Limited
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
+ */
+
+/*
+ * These macros are intended for use when there is a need to copy a low-level
+ * function body into special memory.
+ *
+ * For example, when reconfiguring the SDRAM controller, the code doing the
+ * reconfiguration may need to run from SRAM.
+ *
+ * NOTE: that the copied function body must be entirely self-contained and
+ * position-independent in order for this to work properly.
+ *
+ * NOTE: in order for embedded literals and data to get referenced correctly,
+ * the alignment of functions must be preserved when copying.  To ensure this,
+ * the source and destination addresses for fncpy() must be aligned to a
+ * multiple of 8 bytes: you will be get a BUG() if this condition is not met.
+ * You will typically need a ".align 3" directive in the assembler where the
+ * function to be copied is defined, and ensure that your allocator for the
+ * destination buffer returns 8-byte-aligned pointers.
+ *
+ * Typical usage example:
+ *
+ * extern int f(args);
+ * extern uint32_t size_of_f;
+ * int (*copied_f)(args);
+ * void *sram_buffer;
+ *
+ * copied_f = fncpy(sram_buffer, &f, size_of_f);
+ *
+ * ... later, call the function: ...
+ *
+ * copied_f(args);
+ *
+ * The size of the function to be copied can't be determined from C:
+ * this must be determined by other means, such as adding assmbler directives
+ * in the file where f is defined.
+ */
+
+#ifndef __ASM_FNCPY_H
+#define __ASM_FNCPY_H
+
+#include <linux/types.h>
+#include <linux/string.h>
+
+#include <asm/bug.h>
+#include <asm/cacheflush.h>
+
+/*
+ * Minimum alignment requirement for the source and destination addresses
+ * for function copying.
+ */
+#define FNCPY_ALIGN 8
+
+#define fncpy(dest_buf, funcp, size) ({					\
+	uintptr_t __funcp_address;					\
+	typeof(funcp) __result;						\
+									\
+	asm("" : "=r" (__funcp_address) : "0" (funcp));			\
+									\
+	/*								\
+	 * Ensure alignment of source and destination addresses.	\
+	 */								\
+	BUG_ON((uintptr_t)(dest_buf) & (FNCPY_ALIGN - 1) ||		\
+		(__funcp_address & (FNCPY_ALIGN - 1)));			\
+									\
+	memcpy(dest_buf, (void const *)__funcp_address, size);		\
+	flush_icache_range((unsigned long)(dest_buf),			\
+		(unsigned long)(dest_buf) + (size));			\
+									\
+	asm("" : "=r" (__result)					\
+		: "0" ((uintptr_t)(dest_buf)));				\
+									\
+	__result;							\
+})
+
+#endif /* !__ASM_FNCPY_H */