Re: [RFC 1/5] xfs: [variant A] avoid time_t in user api

From: "Darrick J. Wong" <darrick.wong@oracle.com>
To: Christoph Hellwig <hch@lst.de>
Cc: Arnd Bergmann <arnd@arndb.de>,
	linux-xfs@vger.kernel.org, y2038@lists.linaro.org,
	Deepa Dinamani <deepa.kernel@gmail.com>,
	Dave Chinner <david@fromorbit.com>,
	Brian Foster <bfoster@redhat.com>,
	Allison Collins <allison.henderson@oracle.com>,
	linux-fsdevel@vger.kernel.org
Subject: Re: [RFC 1/5] xfs: [variant A] avoid time_t in user api
Date: Tue, 12 Nov 2019 21:06:28 -0800	[thread overview]
Message-ID: <20191113050628.GS6219@magnolia> (raw)
In-Reply-To: <20191112141600.GB10922@lst.de>

On Tue, Nov 12, 2019 at 03:16:00PM +0100, Christoph Hellwig wrote:
> On Tue, Nov 12, 2019 at 01:09:06PM +0100, Arnd Bergmann wrote:
> > However, as long as two observations are true, a much simpler solution
> > can be used:
> > 
> > 1. xfsprogs is the only user space project that has a copy of this header
> 
> We can't guarantee that.
> 
> > 2. xfsprogs already has a replacement for all three affected ioctl commands,
> >    based on the xfs_bulkstat structure to pass 64-bit timestamps
> >    regardless of the architecture
> 
> XFS_IOC_BULKSTAT replaces XFS_IOC_FSBULKSTAT directly, and can replace
> XFS_IOC_FSBULKSTAT_SINGLE indirectly, so that is easy.  Most users
> actually use the new one now through libfrog, although I found a user
> of the direct ioctl in the xfs_io tool, which could easily be fixed as
> well.

Agreed, XFS_IOC_BULKSTAT is the replacement for the two FSBULKSTAT
variants.  The only question in my mind for the old ioctls is whether we
should return EOVERFLOW if the timestamp would overflow?  Or just
truncate the results?

> XFS_IOC_SWAPEXT does not have a direct replacement, but the timestamp
> is only used to verify that the file did not change vs the previous
> stat.  So not being able to represent > 2038 times is not a real
> problem anyway.

Won't it become a problem when the tv_sec comparison in xfs_swap_extents
is type-promoted to the larger type but lacks the upper bits?

I guess we could force the check to the lower 32 bits on the assumption
that those are the most likely to change due to a file write.

I kinda want to do a SWAPEXT v5, fwiw....

> At some point we should probably look into a file system independent
> defrag ioctl anyway, at which point we can deprecate XFS_IOC_SWAPEXT.
> 
> > Based on those assumptions, changing xfs_bstime to use __kernel_long_t
> > instead of time_t in both the kernel and in xfsprogs preserves the current
> > ABI for any libc definition of time_t and solves the problem of passing
> > 64-bit timestamps to 32-bit user space.
> 
> As said above their are not entirely true, but I still think this patch
> is the right thing to do, if only to get the time_t out of the ABI..
> 
> Reviewed-by: Christoph Hellwig <hch@lst.de>

Seconded,
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>

--D