Re: [PATCH] mke2fs: add make_hugefile feature

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Tue, Jan 21, 2014 at 06:41:41PM -0500, Theodore Ts'o wrote:
> On Tue, Jan 21, 2014 at 11:23:56AM -0800, Darrick J. Wong wrote:
> > 
> > I'd adapt fuse2fs to use a ext2fs_fallocate function, if one existed.  That
> > said, it would have to be more general than this, handling the case where there
> > might be blocks mapped.
> > 
> > Hey Ted, if I wrote an ext2fs_fallocate, would you adapt this for it?
> 
> Sure, if there was an ext2fs_fallocate() which is as CPU efficient as
> what I currently have, and which produces an extent tree which is as
> efficient, I'd certainly use it.  

I was planning to reuse as much of your mke2fs patch as I could get away with;
it looks pretty straightforward.

But first, which patches are in your tree that haven't been pushed to
kernel.org?  It looks like my 'compute block_uninit bitmaps at load time' patch
is in your tree... ;)

> % /usr/bin/time ./mke2fs -F -q  -T hugefile /tmp/foo.img 4T
> 0.09user 0.28system 0:00.37elapsed 99%CPU (0avgtext+0avgdata 2160maxresident)k
> 0inputs+0outputs (0major+592minor)pagefaults 0swaps
> 
> % debugfs -R "stat /database/storage" /tmp/foo.img | head -10
> debugfs 1.42.9 (28-Dec-2013)
> Inode: 13   Type: regular    Mode:  0600   Flags: 0x80000
> Generation: 0    Version: 0x00000000
> User:   120   Group:    50   Size: 4397641728000
> File ACL: 0    Directory ACL: 0
> Links: 1   Blockcount: 8589145544
>      ...
> 
> So I can currently do the equivalent of fallocating close to 4TB in
> less than a half a second.  This includes searching for free blocks
> from the allocation bitmap, and setting up the extent tree, and making
> sure all of the block group checksums are correct.  (And oh yes, that
> includes making the file system and setting up all of the block group
> descriptors as well.)
> 
> Try doing this via the existing interfaces, and see what you get.
> It's not pretty.  :-)

<giggle>

> The one area where we still have some improvements to do after all of
> my changes to libext2fs is that ext2fs_extent_node_split() needs to be
> taught that if the current location of the extent handle is at the
> very end of the file, then when instead of splitting the last leaf
> block 50/50, we should split the node unevenly, by creating a new leaf
> block that has only a single entry in it --- the last extent of the
> file.  That we can use close to 100% of the extent block's space,
> instead of currently where we averaging around 170 extents per 4k
> block, instead of 340.

Hmm, I can look into that.  I don't think that's a big patch.

--D
> 
> Cheers,
> 
> 						- Ted
> --
> To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html