Re: struct statfs/statfs64 in linux-generic

[Chung-Lin Tang <cltang@codesourcery.com>]

On 2013/11/8 03:55 AM, Chris Metcalf wrote:
> On 11/4/2013 2:18 AM, Chung-Lin Tang wrote:
>> Hi,
>>
>> I'm currently working on the glibc port for Altera Nios II, and per the
>> kernel upstreaming requirements, we're making it a user of linux-generic
>> in glibc.
>>
>> I've come across an issue about struct statfs/statfs64, where I've
>> listed the struct definitions in both glibc and the kernel below.
>>
>> In linux-generic, both struct statfs/statfs64 has some fields as 64-bit
>> words, padded properly if need on 32-bit targets. Effectively, they have
>> to be same-format, to pass to the statfs64 syscall uniformly.
>>
>> __SWORD_TYPE appears to be int/long on 32/64-bit targets. This means
>> that for 32-bit targets, struct statfs[64] will be made of 32-bit words,
>> with a few 64-bit fields embedded in the middle.
>>
>>
>> In the kernel however, the fields of importance in struct statfs is made
>> up of entirely '__statfs_word', defined to be 64-bit or 32-bit depending
>> on target (putting aside signedness for now).
> 
> I have to carefully swap all this stuff back into my head every time I look at this issue :-)

Thanks for caring to do that :)

> For native 32-bit (for us that means tilepro) the constraint is that the kernel's "struct statfs64" has to match the two structs in glibc's <bits/statfs.h>.  Most of the statfs types are fixed-size (fsid_t and the __u64 types), with __statfs_word/__SWORD_TYPE being the exception.  For tilepro both of those kernel and glibc types are 32-bit types.
> 
> For 32-bit userspace in a 64-bit kernel (for us that means tilegx -m32) the constraint is that the kernel's "struct compat_statfs64" match the <bits/statfs.h> types, which again, it seems that they should.  (Note that tilegx -m32 is explicitly a "compat" syscall model.)
> 
> A quick look suggests that nios2 is a pure 32-bit architecture, so more like the tilepro model.
> 
> Andrew Pinski's email was a bit confusing since __SYSCALL_SLONG_TYPE seems to be the same size as __SWORD_TYPE, and I don't see any redefinitions of it for aarch64.  I believe his usecase is the 32-bit userspace on 64-bit kernel model.  One thing to watch out for is that some architectures (like x32) actually use the 64-bit syscalls natively (no compat_statfs64) so they need to use "long long" types to match the 64-bit fields, etc., so it's a different problem.  In retrospect that might have been a cleaner model for tilegx -m32, but I suspect that ship has sailed for us.
> 
> In your followup email you write:
> 
>> Still, I'm not sure if using full 64-bit fields should be the intended
>> solution. A full 32-bit target should be able to have those fields as
>> 32-bit (which means a kernel patch to differentiate the current uses of
>> '__statfs_word' into 32/64-bit words)
> 
> I have a guess as to what the problem you're seeing is: are you looking at "struct statfs" in the kernel?  That structure is completely unused on 32-bit platforms, since it's only relevant for the 64-bit platform's sys_statfs/sys_fstatfs syscalls, whereas on 32-bit platforms you only get sys_statfs64/sys_fstatfs64.  In other words there is no syscall that fills in a structure with 32-bit "f_blocks" (for example).  See the __NR3264_statfs macro in <asm-generic/unistd.h> and how it resolves with the __SC_COMP_3264() macro, to the __SC_3264() macro, to the 32-bit version of the syscall.

I think I found my problem, which is more simpler than I thought: unlike
the majority of targets, 64-bit types are 4-byte aligned under Nios II,
which causes a 4-byte difference between the glibc/kernel structure
definitions because the kernel one doesn't have
__attribute__((aligned(8))) forced on to it.

Stack alignment is 4-bytes as well in nios2, so I'll probably be
maintaining that and using a specific nios2/bits/statfs.h header without
the double-word alignment bits.

Thanks,
Chung-Lin