2.5.75 as-iosched.c & asm-generic/div64.h breakage

2.5.75 as-iosched.c & asm-generic/div64.h breakage

[Mikael Pettersson]

Compiling 2.5.75 for PPC32 results in:

  gcc -Wp,-MD,drivers/block/.as-iosched.o.d -D__KERNEL__ -Iinclude -Wall -Wstrict-prototypes -Wno-trigraphs -O2 -fno-strict-aliasing -fno-common -Iarch/ppc -msoft-float -pipe -ffixed-r2 -Wno-uninitialized -mmultiple -mstring -fomit-frame-pointer -nostdinc -iwithprefix include    -DKBUILD_BASENAME=as_iosched -DKBUILD_MODNAME=as_iosched -c -o drivers/block/as-iosched.o drivers/block/as-iosched.c
drivers/block/as-iosched.c: In function `as_update_iohist':
drivers/block/as-iosched.c:840: warning: right shift count >= width of type
drivers/block/as-iosched.c:840: warning: passing arg 1 of `__div64_32' from incompatible pointer type

The statement is "do_div(aic->seek_mean, aic->seek_samples);".

There are two problems here:
- aic->seek_mean is a sector_t, but this type is usually only u32 on 32-bitters,
  unless CONFIG_LBD is set. Thus, as-iosched.c often passes a u32 lvalue to
  do_div() where a u64 lvalue is expected.
- include/asm-generic/div64.h, which is what several 32-bit archs use now, is
  unsafe when its first parameter "n" is a u32 lvalue:
  * The "((n) >> 32) == 0" in the initial test invokes undefined behaviour.
    We want it to be false, but this isn't guaranteed.
  * The call "__div64_32(&(n), __base)" passes a u32* to a function
    expecting a u64*. Major breakage.

x86 survives this because it's custom do_div() implicitly casts "n" to u64
before pulling it apart, and updates "n" with an assignment from a u64, so
the compiler can compensate when "n" is a u32. asm-generic/div64.h should
do something similar, IMO.
  
/Mikael

Re: 2.5.75 as-iosched.c & asm-generic/div64.h breakage

[Andrew Morton]

Mikael Pettersson <mikpe@csd.uu.se> wrote:
>
> drivers/block/as-iosched.c: In function `as_update_iohist':
> drivers/block/as-iosched.c:840: warning: right shift count >= width of type
> drivers/block/as-iosched.c:840: warning: passing arg 1 of `__div64_32' from incompatible pointer type

You mean that code was in -mm for all those months and no ppc32 person
bothered testing it?  Bah.

Something like this?  (Could be sped up for 32-bit sector_t)

diff -puN drivers/block/as-iosched.c~as-do_div-fix drivers/block/as-iosched.c
--- 25/drivers/block/as-iosched.c~as-do_div-fix	Fri Jul 11 14:00:55 2003
+++ 25-akpm/drivers/block/as-iosched.c	Fri Jul 11 14:00:58 2003
@@ -836,8 +836,10 @@ static void as_update_iohist(struct as_i
 		aic->seek_samples += 256;
 		aic->seek_total += 256*seek_dist;
 		if (aic->seek_samples) {
-			aic->seek_mean = aic->seek_total + 128;
-			do_div(aic->seek_mean, aic->seek_samples);
+			u64 seek_mean = aic->seek_total + 128;
+
+			do_div(seek_mean, aic->seek_samples);
+			aic->seek_mean = seek_mean;
 		}
 		aic->seek_samples = (aic->seek_samples>>1)
 					+ (aic->seek_samples>>2);

_

Re: 2.5.75 as-iosched.c & asm-generic/div64.h breakage

[Paul Mackerras]

Andrew Morton writes:

> Mikael Pettersson <mikpe@csd.uu.se> wrote:
> >
> > drivers/block/as-iosched.c: In function `as_update_iohist':
> > drivers/block/as-iosched.c:840: warning: right shift count >= width of type
> > drivers/block/as-iosched.c:840: warning: passing arg 1 of `__div64_32' from incompatible pointer type
> 
> You mean that code was in -mm for all those months and no ppc32 person
> bothered testing it?  Bah.

We've only been getting those warnings since the changeover to the new
asm-generic/div64.h.  Settle down. :)

> Something like this?  (Could be sped up for 32-bit sector_t)
> 
> diff -puN drivers/block/as-iosched.c~as-do_div-fix drivers/block/as-iosched.c
> --- 25/drivers/block/as-iosched.c~as-do_div-fix	Fri Jul 11 14:00:55 2003
> +++ 25-akpm/drivers/block/as-iosched.c	Fri Jul 11 14:00:58 2003
> @@ -836,8 +836,10 @@ static void as_update_iohist(struct as_i
>  		aic->seek_samples += 256;
>  		aic->seek_total += 256*seek_dist;
>  		if (aic->seek_samples) {
> -			aic->seek_mean = aic->seek_total + 128;
> -			do_div(aic->seek_mean, aic->seek_samples);
> +			u64 seek_mean = aic->seek_total + 128;
> +
> +			do_div(seek_mean, aic->seek_samples);
> +			aic->seek_mean = seek_mean;
>  		}

There are several interesting aspects to this:

1. For some reason the LBD config option in drivers/block/Kconfig
   depends on X86.  (CONFIG_LBD is what makes sector_t an unsigned
   long long instead of an unsigned long).  I think the LBD option
   should be available on all 32-bit platforms.  Working out a neat
   way to tell the config system that is left as an exercise for the
   reader. :)

2. It seems to me that seek_total is bounded above by 1024 * the
   maximum seek distance.  I'm a bit concerned about that overflowing
   32 bits - AFAICS I would only need a > 2GB disk (or do I mean
   partition?) for that to happen.  Could we make seek_total be a u64
   unconditionally please?

3. I guess that adding 128 on to the seek_total is to get a
   round-to-nearest effect in the division.  In fact you need to add
   on (seek_samples >> 1) to get that effect.

Regards,
Paul.

Re: 2.5.75 as-iosched.c & asm-generic/div64.h breakage

[Nick Piggin]

Hi!

Paul Mackerras wrote:

>Andrew Morton writes:
>
>
>>Mikael Pettersson <mikpe@csd.uu.se> wrote:
>>
>>>drivers/block/as-iosched.c: In function `as_update_iohist':
>>>drivers/block/as-iosched.c:840: warning: right shift count >= width of type
>>>drivers/block/as-iosched.c:840: warning: passing arg 1 of `__div64_32' from incompatible pointer type
>>>
>>You mean that code was in -mm for all those months and no ppc32 person
>>bothered testing it?  Bah.
>>
>
>We've only been getting those warnings since the changeover to the new
>asm-generic/div64.h.  Settle down. :)
>
>
>>Something like this?  (Could be sped up for 32-bit sector_t)
>>
>>diff -puN drivers/block/as-iosched.c~as-do_div-fix drivers/block/as-iosched.c
>>--- 25/drivers/block/as-iosched.c~as-do_div-fix	Fri Jul 11 14:00:55 2003
>>+++ 25-akpm/drivers/block/as-iosched.c	Fri Jul 11 14:00:58 2003
>>@@ -836,8 +836,10 @@ static void as_update_iohist(struct as_i
>> 		aic->seek_samples += 256;
>> 		aic->seek_total += 256*seek_dist;
>> 		if (aic->seek_samples) {
>>-			aic->seek_mean = aic->seek_total + 128;
>>-			do_div(aic->seek_mean, aic->seek_samples);
>>+			u64 seek_mean = aic->seek_total + 128;
>>+
>>+			do_div(seek_mean, aic->seek_samples);
>>+			aic->seek_mean = seek_mean;
>> 		}
>>
>
>There are several interesting aspects to this:
>
>1. For some reason the LBD config option in drivers/block/Kconfig
>   depends on X86.  (CONFIG_LBD is what makes sector_t an unsigned
>   long long instead of an unsigned long).  I think the LBD option
>   should be available on all 32-bit platforms.  Working out a neat
>   way to tell the config system that is left as an exercise for the
>   reader. :)
>

Not me :P

>
>2. It seems to me that seek_total is bounded above by 1024 * the
>   maximum seek distance.  I'm a bit concerned about that overflowing
>   32 bits - AFAICS I would only need a > 2GB disk (or do I mean
>   partition?) for that to happen.  Could we make seek_total be a u64
>   unconditionally please?
>

Probably a good idea. It is only the seek distance for one
process though. But it wouldn't hurt.

>
>3. I guess that adding 128 on to the seek_total is to get a
>   round-to-nearest effect in the division.  In fact you need to add
>   on (seek_samples >> 1) to get that effect.
>

You're right. Thanks.

Re: 2.5.75 as-iosched.c & asm-generic/div64.h breakage

[Peter Chubb]

>> 
>> 1. For some reason the LBD config option in drivers/block/Kconfig
>> depends on X86.  (CONFIG_LBD is what makes sector_t an unsigned
>> long long instead of an unsigned long).  I think the LBD option
>> should be available on all 32-bit platforms.  Working out a neat
>> way to tell the config system that is left as an exercise for the
>> reader. :)
>> 

Nick> Not me :P

No 'twas me --- because I couldn't work oput any cvlean way to do it,
and, of the 32-bit plaftforms, I could test only x86.


--
Dr Peter Chubb  http://www.gelato.unsw.edu.au  peterc AT gelato.unsw.edu.au
You are lost in a maze of BitKeeper repositories,   all slightly different.

Re: 2.5.75 as-iosched.c & asm-generic/div64.h breakage

[Mikael Pettersson]

On Fri, 11 Jul 2003 14:01:58 -0700, Andrew Morton wrote:
> Mikael Pettersson <mikpe@csd.uu.se> wrote:
> >
> > drivers/block/as-iosched.c: In function `as_update_iohist':
> > drivers/block/as-iosched.c:840: warning: right shift count >= width of type
> > drivers/block/as-iosched.c:840: warning: passing arg 1 of `__div64_32' from incompatible pointer type
> 
> You mean that code was in -mm for all those months and no ppc32 person
> bothered testing it?  Bah.

Apparently. I don't have time for anything but standard kernels.

> Something like this?  (Could be sped up for 32-bit sector_t)
> 
> diff -puN drivers/block/as-iosched.c~as-do_div-fix drivers/block/as-iosched.c
> --- 25/drivers/block/as-iosched.c~as-do_div-fix	Fri Jul 11 14:00:55 2003
> +++ 25-akpm/drivers/block/as-iosched.c	Fri Jul 11 14:00:58 2003
> @@ -836,8 +836,10 @@ static void as_update_iohist(struct as_i
>  		aic->seek_samples += 256;
>  		aic->seek_total += 256*seek_dist;
>  		if (aic->seek_samples) {
> -			aic->seek_mean = aic->seek_total + 128;
> -			do_div(aic->seek_mean, aic->seek_samples);
> +			u64 seek_mean = aic->seek_total + 128;
> +
> +			do_div(seek_mean, aic->seek_samples);
> +			aic->seek_mean = seek_mean;
>  		}
>  		aic->seek_samples = (aic->seek_samples>>1)
>  					+ (aic->seek_samples>>2);

Perhaps, but it's my opinion that do_div() needs to be made more robust,
since arch-specific data abstraction means that callers in generic code
don't always know if some value is u32 or u64.

So your change should be in do_div() itself, not in its callers.

/Mikael