Re: [PATCH xfsprogs 2/2] linux.h: Define xfs_off_t as int64_t

From: Dave Chinner <david@fromorbit.com>
To: Felix Janda <felix.janda@posteo.de>
Cc: Christoph Hellwig <hch@infradead.org>, xfs@oss.sgi.com
Subject: Re: [PATCH xfsprogs 2/2] linux.h: Define xfs_off_t as int64_t
Date: Sun, 7 Aug 2016 09:18:03 +1000	[thread overview]
Message-ID: <20160806231802.GN16044@dastard> (raw)
In-Reply-To: <20160806083852.GA26156@nyan>

On Sat, Aug 06, 2016 at 10:38:52AM +0200, Felix Janda wrote:
> Dave Chinner wrote:
> > On Fri, Aug 05, 2016 at 03:09:05PM +0200, Felix Janda wrote:
> > > Dave Chinner wrote:
> > > > On Fri, Aug 05, 2016 at 10:02:41AM +0200, Felix Janda wrote:
> > > > > Dave Chinner wrote:
> > > > > > On Mon, Aug 01, 2016 at 08:54:10AM +0200, Felix Janda wrote:
> > > > > > > Christoph Hellwig wrote:
> > > > > > > > On Sat, Jul 30, 2016 at 03:37:37PM +0200, Felix Janda wrote:
> > > > > > > > > int64_t is guaranteed to have the correct size and signedness and is
> > > > > > > > > always avaible because linux.h has a <inttypes.h> include.
> > > > > > > > > 
> > > > > > > > > Fixes compilation error "unkown type name 'off64_t'" on linux when the
> > > > > > > > > public header <xfs.h> is included without _LARGEFILE64_SOURCE or
> > > > > > > > > _GNU_SOURCE defined. This bug was introduced in commit
> > > > > > > > > cb898f157f8410a03cf5f3400baa1df9e5eecd33.
> > > > > > > > 
> > > > > > > > I would much prefer to just define _LARGEFILE64_SOURCE in linux.h..
> > > > > > > 
> > > > > > > Thanks for the suggestion, but that does not work if the system header
> > > > > > > defining (or not) off64_t is included before the xfs headers.
> > > > > > 
> > > > > > Which, to me, is a build bug in whatever code is including the xfs
> > > > > > headers.  Isn't it the responsibility of the build environment to
> > > > > > ensure the dependencies of the libraries being used are correctly
> > > > > > met?
> > > > > 
> > > > > Every program using the xfs header is supposed to know that (only on
> > > > > linux) since commit cb898f157f8410a03cf5f3400baa1df9e5eecd33 it is
> > > > > necessary to define _LARGEFILE64_SOURCE or _GNU_SOURCE?
> > > > 
> > > > Which, I'd say, most already do, because anything trying to use XFS
> > > > ioctls needs to be 64 bit offset clean, even on 32 bit systems.  I
> > > > don't see any problem with requiring it when including a header
> > > > that exposes ioctl interfaces with 64 bit file size/offset fields
> > > > in them....
> > > 
> > > The easiest way to be 64bit clean is to use _FILEOFFSET_BITS=64. Then
> > > off_t is 64bit on all architectures and it is impossible to use 32bit
> > > interfaces. However the type off64_t will still not be defined...
> > > 
> > > (On the other hand, when just using _LARGEFILE64_SOURCE it is still
> > > easy to mix 32 and 64bit interfaces.)
> > 
> > Which, with library code, we are likely to see applications using.
> > 
> > If you want to clean this up, then remove the dependence on
> > _LARGEFILE64_SOURCE in the entire xfsprogs code base (e.g. it uses
> > lseek64 everywhere which requires off64_t to be defined) and instead
> > make it dependent on _FILEOFFSET_BITS=64. Then you can get rid of
> > all the uses of off64_t completely, and we can break the build if
> > _FILEOFFSET_BITS != 64 on inclusion of xfs.h.
> 
> Yes, I'd like to clean this up.
> 
> But first note that you can have both _FILE_OFFSET_BITS=64 and
> _LARGEFILE64_SOURCE. Then everything (off64_t, lseek64, ...) is
> defined and everything (off_t, lseek, ...) is 64bit.
> 
> So to clean up I would first get _FILE_OFFSET_BITS=64 defined and then
> start "removing 64" from functions/types in any order. *Before
> modifying the public headers* the sizeof(off_t)=8 check needs to
> be put into xfs.h.
> 
> Also note that there are 3 different (but equivalent) off_t types
> currently used in the code base: off64_t, loff_t and xfs_off_t.
> Should these be converted to xfs_off_t or off_t?

Not that simple. loff_t has to remain for the copy_file_range()
syscall in xfs_io. That syscall requires _GNU_SOURCE  and loff_t to
be defined from the system headers, so it can't really go away.

xfs_off_t is an internal XFS file offset definition, used by the
code in libxfs/ and shared with the kernel code, so it can't go
away, either.

So, essentially, the only code that should change is all
the code that uses off64_t - that can use off_t as that's what
all the systems that use those variables require...

> Still, doing these type conversions is going to be pretty invasive
> and is not unlikely to conflict with outstanding patches. Is now
> a good time for this? (How about the __uint -> uint, __int -> int
> conversion?)

off64_t -> off_t affects very little of the new code we have
outstanding. It mostly affects xfs_io, so there's little to worry
about in terms of merge conflicts here.

The __*int conversions are a different matter. They affect the
entire code base - they are widespread through the libxfs code so we
need to do a kernel code conversion first. Then we can propagate
that back into the libxfs code in xfsprogs, and then the rest of
xfsprogs can be done.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

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