Patchwork [U-Boot,V2] Do not copy elf section to same adress

login
register
mail settings
Submitter Matthias Weisser
Date Jan. 19, 2011, 11:03 a.m.
Message ID <1295435020-14190-1-git-send-email-weisserm@arcor.de>
Download mbox | patch
Permalink /patch/79447/
State Changes Requested
Headers show

Comments

Matthias Weisser - Jan. 19, 2011, 11:03 a.m.
When an elf section is already at the right position (e.g. after image
decompression by bootm) there is no need to copy it. This saves some ms
when bootig an elf image.

Changes since V1
  - Fixed style issues

Signed-off-by: Matthias Weisser <weisserm@arcor.de>
---
 common/cmd_elf.c |    8 +++++---
 1 files changed, 5 insertions(+), 3 deletions(-)
Matthias Weisser - April 11, 2011, 4:17 p.m.
Hi

Am 19.01.2011 12:03, schrieb Matthias Weisser:
> When an elf section is already at the right position (e.g. after image
> decompression by bootm) there is no need to copy it. This saves some ms
> when bootig an elf image.
> 
> Changes since V1
>   - Fixed style issues
> 
> Signed-off-by: Matthias Weisser <weisserm@arcor.de>
> ---
>  common/cmd_elf.c |    8 +++++---
>  1 files changed, 5 insertions(+), 3 deletions(-)
> 
> diff --git a/common/cmd_elf.c b/common/cmd_elf.c
> index bf32612..3537769 100644
> --- a/common/cmd_elf.c
> +++ b/common/cmd_elf.c
> @@ -342,9 +342,11 @@ static unsigned long load_elf_image_shdr(unsigned long addr)
>  			memset ((void *)shdr->sh_addr, 0, shdr->sh_size);
>  		} else {
>  			image = (unsigned char *) addr + shdr->sh_offset;
> -			memcpy ((void *) shdr->sh_addr,
> -				(const void *) image,
> -				shdr->sh_size);
> +			if ((void *) shdr->sh_addr != (void *) image) {
> +				memcpy((void *) shdr->sh_addr,
> +					(const void *) image,
> +					shdr->sh_size);
> +			}
>  		}
>  		flush_cache (shdr->sh_addr, shdr->sh_size);
>  	}

Any comments on this patch? Any problems with it? I would like to see it
in mainline and I am open for any comments.

Regards,
Matthias Weißer
Wolfgang Denk - April 11, 2011, 7:59 p.m.
Dear Matthias Weisser,

In message <1295435020-14190-1-git-send-email-weisserm@arcor.de> you wrote:
> When an elf section is already at the right position (e.g. after image
> decompression by bootm) there is no need to copy it. This saves some ms
> when bootig an elf image.
> 
> Changes since V1
>   - Fixed style issues
> 
> Signed-off-by: Matthias Weisser <weisserm@arcor.de>
> ---
>  common/cmd_elf.c |    8 +++++---
>  1 files changed, 5 insertions(+), 3 deletions(-)
> 
> diff --git a/common/cmd_elf.c b/common/cmd_elf.c
> index bf32612..3537769 100644
> --- a/common/cmd_elf.c
> +++ b/common/cmd_elf.c
> @@ -342,9 +342,11 @@ static unsigned long load_elf_image_shdr(unsigned long addr)
>  			memset ((void *)shdr->sh_addr, 0, shdr->sh_size);
>  		} else {
>  			image = (unsigned char *) addr + shdr->sh_offset;
> -			memcpy ((void *) shdr->sh_addr,
> -				(const void *) image,
> -				shdr->sh_size);
> +			if ((void *) shdr->sh_addr != (void *) image) {
> +				memcpy((void *) shdr->sh_addr,
> +					(const void *) image,
> +					shdr->sh_size);
> +			}

The idea is correct, but I think the implementation is suboptimal.
Instead of fixing this use case only, the test should be moved into
the implementation of memcpy() itself so any other callers with such a
situation benefit from it, too.

While we are at it, we should do the same with bcopy() and memmove(),
too.

Thanks.

Wolfgang Denk
Matthias Weisser - April 11, 2011, 8:12 p.m.
Am 11.04.2011 21:59, schrieb Wolfgang Denk:
> Dear Matthias Weisser,
> 
> In message <1295435020-14190-1-git-send-email-weisserm@arcor.de> you wrote:
>> > When an elf section is already at the right position (e.g. after image
>> > decompression by bootm) there is no need to copy it. This saves some ms
>> > when bootig an elf image.
>> > 
>> > Changes since V1
>> >   - Fixed style issues
>> > 
>> > Signed-off-by: Matthias Weisser <weisserm@arcor.de>
>> > ---
>> >  common/cmd_elf.c |    8 +++++---
>> >  1 files changed, 5 insertions(+), 3 deletions(-)
>> > 
>> > diff --git a/common/cmd_elf.c b/common/cmd_elf.c
>> > index bf32612..3537769 100644
>> > --- a/common/cmd_elf.c
>> > +++ b/common/cmd_elf.c
>> > @@ -342,9 +342,11 @@ static unsigned long load_elf_image_shdr(unsigned long addr)
>> >  			memset ((void *)shdr->sh_addr, 0, shdr->sh_size);
>> >  		} else {
>> >  			image = (unsigned char *) addr + shdr->sh_offset;
>> > -			memcpy ((void *) shdr->sh_addr,
>> > -				(const void *) image,
>> > -				shdr->sh_size);
>> > +			if ((void *) shdr->sh_addr != (void *) image) {
>> > +				memcpy((void *) shdr->sh_addr,
>> > +					(const void *) image,
>> > +					shdr->sh_size);
>> > +			}
> The idea is correct, but I think the implementation is suboptimal.
> Instead of fixing this use case only, the test should be moved into
> the implementation of memcpy() itself so any other callers with such a
> situation benefit from it, too.

Well, I thought about that too. But decided against it as I thought it
will be a bit too intervening for a patch from me as this will hit all
boards.

I can't come up with an example where this may produce a problem but who
knows which exotic hardware is out there which expects that a memcpy
with identical src and dst addresses is executed exactly in that way.
But maybe we can ignore this and let these exotic boards come up with a
solution handling that special situation.

> While we are at it, we should do the same with bcopy() and memmove(),
> too.

Maybe I can post a patch tomorrow. The only thing which I can't handle
are architecture specific memcpy/memmove/... functions besides ARM.

Regards,
Matthias Weißer
Wolfgang Denk - April 11, 2011, 9:09 p.m.
Dear Matthias Weisser,

In message <4DA360A9.10401@arcor.de> you wrote:
>
> I can't come up with an example where this may produce a problem but who
> knows which exotic hardware is out there which expects that a memcpy
> with identical src and dst addresses is executed exactly in that way.
> But maybe we can ignore this and let these exotic boards come up with a
> solution handling that special situation.

Right.

> > While we are at it, we should do the same with bcopy() and memmove(),
> > too.
> 
> Maybe I can post a patch tomorrow. The only thing which I can't handle
> are architecture specific memcpy/memmove/... functions besides ARM.

That's fine. Optimized arch specific code probably already implements
such a shortcut anyway.

Best regards,

Wolfgang Denk

Patch

diff --git a/common/cmd_elf.c b/common/cmd_elf.c
index bf32612..3537769 100644
--- a/common/cmd_elf.c
+++ b/common/cmd_elf.c
@@ -342,9 +342,11 @@  static unsigned long load_elf_image_shdr(unsigned long addr)
 			memset ((void *)shdr->sh_addr, 0, shdr->sh_size);
 		} else {
 			image = (unsigned char *) addr + shdr->sh_offset;
-			memcpy ((void *) shdr->sh_addr,
-				(const void *) image,
-				shdr->sh_size);
+			if ((void *) shdr->sh_addr != (void *) image) {
+				memcpy((void *) shdr->sh_addr,
+					(const void *) image,
+					shdr->sh_size);
+			}
 		}
 		flush_cache (shdr->sh_addr, shdr->sh_size);
 	}