Question about inode64

Question about inode64

[Su Yue]

Hi, XFS developers

Recently, I was told that some 32bit applications don't work
on large XFS filesystems. And it is caused by ino overflow in syscalls.

If I don't misunderstand libxfs/xfs_format, XFS' absolute inode number
consists of agno_log
-agblklog-inopblog bits.
So significative bits = ROUND(log2(Disk size / AG size)) + ROUND(log2(AG
size / BLK size)) + ROUND(log2(BLK size / inode size)). Right?
Does it mean that it's feasible to avoid ino overflow by control of disk
size, AG size, block size, inode size when mkfs even the disk will be mount
with option inode64.

Thanks,
Su

Re: Question about inode64

[Eric Sandeen]

On 5/10/18 12:31 PM, Su Yue wrote:
> Hi, XFS developers
> 
> Recently, I was told that some 32bit applications don't work
> on large XFS filesystems. And it is caused by ino overflow in syscalls.
> 
> If I don't misunderstand libxfs/xfs_format, XFS' absolute inode number
> consists of agno_log
> -agblklog-inopblog bits.
> So significative bits = ROUND(log2(Disk size / AG size)) + ROUND(log2(AG
> size / BLK size)) + ROUND(log2(BLK size / inode size)). Right?

> Does it mean that it's feasible to avoid ino overflow by control of disk
> size, AG size, block size, inode size when mkfs even the disk will be mount
> with option inode64.

Yes, inode number is an encoded disk location, and the encoding depends
on various other bits of filesystem geometry.

Note that "inode64" only means that it will /allow/ some inodes to exceed
32 bits, it does not mean that every inode number /will/ be > 32 bits.

But most people don't want to constrain their filesystem geometry around
inode number sizing; if you require 32-bit inodes, just use the inode32 mount
option.

Or better yet, fix the application.  :)

-Eric

> Thanks,
> Su

Re: Question about inode64

[Su Yue]

Thanks a lot!

After one-night sleep, my brain is clear to figure out
inode bits and inode64 mount option with your explanation.
And your suggestions are helpful too. : )

Su

On Fri, May 11, 2018 at 5:20 AM Eric Sandeen <sandeen@sandeen.net> wrote:

> On 5/10/18 12:31 PM, Su Yue wrote:
> > Hi, XFS developers
> >
> > Recently, I was told that some 32bit applications don't work
> > on large XFS filesystems. And it is caused by ino overflow in syscalls.
> >
> > If I don't misunderstand libxfs/xfs_format, XFS' absolute inode number
> > consists of agno_log
> > -agblklog-inopblog bits.
> > So significative bits = ROUND(log2(Disk size / AG size)) + ROUND(log2(AG
> > size / BLK size)) + ROUND(log2(BLK size / inode size)). Right?

> > Does it mean that it's feasible to avoid ino overflow by control of disk
> > size, AG size, block size, inode size when mkfs even the disk will be
mount
> > with option inode64.

> Yes, inode number is an encoded disk location, and the encoding depends
> on various other bits of filesystem geometry.

> Note that "inode64" only means that it will /allow/ some inodes to exceed
> 32 bits, it does not mean that every inode number /will/ be > 32 bits.

> But most people don't want to constrain their filesystem geometry around
> inode number sizing; if you require 32-bit inodes, just use the inode32
mount
> option.

> Or better yet, fix the application.  :)

> -Eric

> > Thanks,
> > Su