More MIPS stat.h breakage...?

More MIPS stat.h breakage...?

[Rich Felker]

I thought all of these issues were addressed a long time ago, but I
think we still may have at least one bits-header bug on MIPS big
endian: stat.h. The kernel version of this structure has:

struct stat64 {
        unsigned long   st_dev;
        unsigned long   st_pad0[3];     /* Reserved for st_dev expansion  */
        ...

whereas st_dev needs to be dev_t, which is 64-bit.

We're handling this now simply by using the following definition:

struct stat
{
        dev_t st_dev;
        long __st_padding1[2];
        ...

i.e. incorporating one of the padding slots into st_dev. But this only
yields the correct value for little endian. For big endian, I think
device numbers are appearing in the upper 32 bits rather than the
lower ones. This renders the major, minor, etc. macros incorrect.

I think the cleanest fix for this is probably to add an arch-specific,
nop-by-default function __stat_fixup or similar that gets called
whenever any stat-family syscall succeeds in order to patch up the
results as needed. Alternatively, we clould do the patchup at the
syscall level for mips' syscall_arch.h, much like how x32 works.

Anyone have strong opinions on how this should be handled?

Rich

Re: More MIPS stat.h breakage...?

[Isaac Dunham]

On Wed, Jul 16, 2014 at 12:15:56AM -0400, Rich Felker wrote:
> I thought all of these issues were addressed a long time ago, but I
> think we still may have at least one bits-header bug on MIPS big
> endian: stat.h. The kernel version of this structure has:
> 
> struct stat64 {
>         unsigned long   st_dev;
>         unsigned long   st_pad0[3];     /* Reserved for st_dev expansion  */
>         ...
> 
> whereas st_dev needs to be dev_t, which is 64-bit.
> 
> We're handling this now simply by using the following definition:
> 
> struct stat
> {
>         dev_t st_dev;
>         long __st_padding1[2];
>         ...
> 
> i.e. incorporating one of the padding slots into st_dev. But this only
> yields the correct value for little endian. For big endian, I think
> device numbers are appearing in the upper 32 bits rather than the
> lower ones. This renders the major, minor, etc. macros incorrect.
> 
> I think the cleanest fix for this is probably to add an arch-specific,
> nop-by-default function __stat_fixup or similar that gets called
> whenever any stat-family syscall succeeds in order to patch up the
> results as needed. Alternatively, we clould do the patchup at the
> syscall level for mips' syscall_arch.h, much like how x32 works.
> 
> Anyone have strong opinions on how this should be handled?

I seem to remember some comment about trying to work on kernels that predate
stat64().  __stat_fixup would be a logical way to handle that, should that
still be intended.
The most precise information I can find is that stat64() dates to 2.4.18;
I'm not sure about the accuracy of that.

But I don't have a strong opinion; I just noticed a similarity to an idea
that was mentioned some time back.
Also, I'm not aware of any reasons 2.4.1x might be a major target
for musl. ;-)

Thanks,
Isaac Dunham

Re: More MIPS stat.h breakage...?

[Rich Felker]

On Wed, Jul 16, 2014 at 04:40:06AM -0700, Isaac Dunham wrote:
> I seem to remember some comment about trying to work on kernels that predate
> stat64().  __stat_fixup would be a logical way to handle that, should that
> still be intended.

The idea of __stat_fixup is that it would patch a broken struct layout
after a successful stat-family call, not make a new syscall. For a new
syscall it would have to be separate workaround code in each variant
(stat, lstat, fstat, fstatat) to do the appropriate workaround for
each. This is a completely different scope.

> The most precise information I can find is that stat64() dates to 2.4.18;
> I'm not sure about the accuracy of that.
> 
> But I don't have a strong opinion; I just noticed a similarity to an idea
> that was mentioned some time back.
> Also, I'm not aware of any reasons 2.4.1x might be a major target
> for musl. ;-)

I don't think there's any viable way to support kernels that old even
partially without backporting some essential features.

Rich