inode64 and 64bit kernel with 32bit userspace

inode64 and 64bit kernel with 32bit userspace

[Matthias Schniedermeyer]

Hi


The more or less simple question is:
Is the requirement for 32bit programs to support 64bit inodes the same 
as LFS(Large File Support)?

IOW if a programs was compiled with FILE_OFFSET_BITS=64 (if i remember 
that name correctly) should it work?
So that all programs compiled in the last several years in a halfway 
recent distribution should be able to cope with 64bit inodes?

The program i care about most is rsync and it appears to work correctly 
when i tested it with 2 hard-linked files (with an inode >2^32) and the 
option to preserve hard links. And "find" and "stat" also worked, 
otherwise i couldn't have found the file to test it with. ;-)



-- 

Matthias

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: inode64 and 64bit kernel with 32bit userspace

[Eric Sandeen]

On 2/18/13 3:43 AM, Matthias Schniedermeyer wrote:
> Hi
> 
> 
> The more or less simple question is:
> Is the requirement for 32bit programs to support 64bit inodes the same 
> as LFS(Large File Support)?
> 
> IOW if a programs was compiled with FILE_OFFSET_BITS=64 (if i remember 
> that name correctly) should it work?

I think so (I don't know where the formal documentation is, 
http://users.suse.com/~aj/linux_lfs.html is an old but still good
over view I think).  From open(2):

       EOVERFLOW
              (stat())  path refers to a file whose size cannot be represented
              in the type       off_t.  This can  occur  when  an  application
              compiled  on  a  32-bit  platform without -D_FILE_OFFSET_BITS=64
              calls stat() on a file whose size exceeds (2<<31)-1 bits.

EOVERFLOW can happen if the inode nubmer doesn't fit in a (32-bit)
stat struct as well.

> So that all programs compiled in the last several years in a halfway 
> recent distribution should be able to cope with 64bit inodes?

The syscalls it makes should be able to, at least.  What it does after
that is up to the program.

(i.e. you might do a nice stat64 and then try to stuff the size or the
inode number into a 32-bit variable; the -D_FILE_OFFSET_BITS won't help
in that case.)

I did a test long ago at :

http://sandeen.net/wordpress/computers/the-world-wants-32-bit-inodes/

and the results weren't too encouraging; I should try it again on a
more recent distro, and see how things look.

-Eric

> The program i care about most is rsync and it appears to work correctly 
> when i tested it with 2 hard-linked files (with an inode >2^32) and the 
> option to preserve hard links. And "find" and "stat" also worked, 
> otherwise i couldn't have found the file to test it with. ;-)
> 
> 
> 

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: inode64 and 64bit kernel with 32bit userspace

[Matthias Schniedermeyer]

On 18.02.2013 09:12, Eric Sandeen wrote:
> On 2/18/13 3:43 AM, Matthias Schniedermeyer wrote:
> > Hi
> > 
> > 
> > The more or less simple question is:
> > Is the requirement for 32bit programs to support 64bit inodes the same 
> > as LFS(Large File Support)?
> > 
> > IOW if a programs was compiled with FILE_OFFSET_BITS=64 (if i remember 
> > that name correctly) should it work?
> 
> I think so (I don't know where the formal documentation is, 
> http://users.suse.com/~aj/linux_lfs.html is an old but still good
> over view I think).  From open(2):
> 
>        EOVERFLOW
>               (stat())  path refers to a file whose size cannot be represented
>               in the type       off_t.  This can  occur  when  an  application
>               compiled  on  a  32-bit  platform without -D_FILE_OFFSET_BITS=64
>               calls stat() on a file whose size exceeds (2<<31)-1 bits.
> 
> EOVERFLOW can happen if the inode nubmer doesn't fit in a (32-bit)
> stat struct as well.

I've looked into /usr/include/sys/stat.h

And i see this:
# ifndef __ino_t_defined
#  ifndef __USE_FILE_OFFSET64
typedef __ino_t ino_t;
#  else
typedef __ino64_t ino_t;
#  endif
#  define __ino_t_defined
# endif

So ino_t really is __ino64_t when compiled with the LFS option, which 
answers my original question. :-)

Besides i don't have that many programs that (should) care about inodes. 
Of the top of my head i care about rsync/perl/find/ln/ls, which 
apparently work correctly.




-- 

Matthias

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: inode64 and 64bit kernel with 32bit userspace

[Eric Sandeen]

On 2/18/13 10:20 AM, Matthias Schniedermeyer wrote:
> On 18.02.2013 09:12, Eric Sandeen wrote:
>> On 2/18/13 3:43 AM, Matthias Schniedermeyer wrote:
>>> Hi
>>>
>>>
>>> The more or less simple question is:
>>> Is the requirement for 32bit programs to support 64bit inodes the same 
>>> as LFS(Large File Support)?
>>>
>>> IOW if a programs was compiled with FILE_OFFSET_BITS=64 (if i remember 
>>> that name correctly) should it work?
>>
>> I think so (I don't know where the formal documentation is, 
>> http://users.suse.com/~aj/linux_lfs.html is an old but still good
>> over view I think).  From open(2):
>>
>>        EOVERFLOW
>>               (stat())  path refers to a file whose size cannot be represented
>>               in the type       off_t.  This can  occur  when  an  application
>>               compiled  on  a  32-bit  platform without -D_FILE_OFFSET_BITS=64
>>               calls stat() on a file whose size exceeds (2<<31)-1 bits.
>>
>> EOVERFLOW can happen if the inode nubmer doesn't fit in a (32-bit)
>> stat struct as well.
> 
> I've looked into /usr/include/sys/stat.h
> 
> And i see this:
> # ifndef __ino_t_defined
> #  ifndef __USE_FILE_OFFSET64
> typedef __ino_t ino_t;
> #  else
> typedef __ino64_t ino_t;
> #  endif
> #  define __ino_t_defined
> # endif
> 
> So ino_t really is __ino64_t when compiled with the LFS option, which 
> answers my original question. :-)
> 
> Besides i don't have that many programs that (should) care about inodes. 
> Of the top of my head i care about rsync/perl/find/ln/ls, which 
> apparently work correctly.

find cares, ls cares . . . but I would assume that they get it right :)

I'm getting a box set up w/ 32-bit F18, I want to re-run my test over it
and see if things have improved at all in 5 years.  :)

-Eric

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs