Re: [RFC5 PATCH v6 00/21] ILP32 for ARM64

Arnd Bergmann [off-list ref] writes:

In https://github.com/norov/glibc/commit/5d4290435e428267171ece871539b76e1d079d11
you are defining a struct __kernel_stat64 in the glibc. Is this the expected
way to do it? I would have thought you'd get the definition from the kernel
headers.

The problem really is that struct stat64 does not match the kernel
struct stat64.  The latter uses the broken st_ino, where the 64-bit
inode is actually stored at the end of the structure, and there is
padding *before* the 32-bit inode, not after (so you cannot overlay a
64-bit inode field in little endian mode).  You need to add the attached
file as sysdeps/unix/sysv/linux/aarch64/bits/stat.h, then you can use
the user-space struct stat64 to communicate with the *stat64 syscalls,
and __kernel_stat64 is not needed any more.

Andreas.

-- 
Andreas Schwab, SUSE Labs, schwab at suse.de
GPG Key fingerprint = 0196 BAD8 1CE9 1970 F4BE  1748 E4D4 88E3 0EEA B9D7
"And now for something completely different."
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: stat.h
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20160229/6c496eb8/attachment-0001.h>

On Monday 29 February 2016 17:00:29 Andreas Schwab wrote:

Arnd Bergmann [off-list ref] writes:

quoted

In https://github.com/norov/glibc/commit/5d4290435e428267171ece871539b76e1d079d11
you are defining a struct __kernel_stat64 in the glibc. Is this the expected
way to do it? I would have thought you'd get the definition from the kernel
headers.

The problem really is that struct stat64 does not match the kernel
struct stat64.  The latter uses the broken st_ino, where the 64-bit
inode is actually stored at the end of the structure, and there is
padding *before* the 32-bit inode, not after (so you cannot overlay a
64-bit inode field in little endian mode).  You need to add the attached
file as sysdeps/unix/sysv/linux/aarch64/bits/stat.h, then you can use
the user-space struct stat64 to communicate with the *stat64 syscalls,
and __kernel_stat64 is not needed any more.

Ok, got it. So this is an unfortunate side-effect of using the same
'stat64' as arm32 does.

If it's really that broken, should we change the kernel (once more)
to use the layout of arm64 instead? That has the 64-bit time fields
though, so it might require similar hacks, unless we first merge
support for 64-bit time_t based syscalls (which seems unlikely
at this point, sorry for the delay on my end).

Is there possibly a way to reference the implementation that we
have for stat64 on arm32 rather than duplicating it?

	Arnd