Re: [PATCH] pipefs unique inode numbers

[Jeff Layton <jlayton@redhat.com>]

Jeff Layton wrote:
> Bodo Eggert wrote:
>  > change pipefs to use a unique inode number equal to the memory
>  > address unless it would be truncated.
>  >
>  > Signed-Off-By: Bodo Eggert <7eggert@gmx.de>
>  > ---
>  > Tested on i386.
>  >
>  > --- 2.6.19/fs/pipe.c.ori    2007-01-30 22:02:46.000000000 +0100
>  > +++ 2.6.19/fs/pipe.c    2007-01-30 23:22:27.000000000 +0100
>  > @@ -864,6 +864,10 @@ static struct inode * get_pipe_inode(voi
>  >      inode->i_uid = current->fsuid;
>  >      inode->i_gid = current->fsgid;
>  >      inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
>  > +    /* The address of *inode is unique, so we'll get an unique inode 
> number.
>  > +     * Off cause this will not work for 32 bit inodes on 64 bit 
> systems. */
>  > +    if (sizeof(inode->i_ino) >= sizeof(struct inode*))
>  > +        inode->i_ino = (unsigned int) inode;
>  >
>  >      return inode;
>  >
>
> Also, that patch would break many 32-bit programs not compiled with large
> offsets when run in compatibility mode on a 64-bit kernel. If they were to
> do a stat on this inode, it would likely generate an EOVERFLOW error since
> the pointer address would probably not fit in a 32 bit field.
>
> That problem was the whole impetus for this set of patches.

Actually, sorry...I misread the patch. It wouldn't have that problem. My
mistake.

Still though, I considered an approach somewhat similar to this early on.
I was thinking of taking a bit-shifted inode address and hashing it to
give a unique value. If you do the math, you can discard the lower 9 bits
of the pointer, so you end up being able to use the lower 41 bits of the
pointer. So a scheme like that could work if you could guarantee that
all inode addresses wouldn't be > 2^41 apart.

The problem is, you can't guarantee that, especially in a NUMA situation.

See the thread entitled:

[RFC][PATCH] ensure i_ino uniqueness in filesystems without
permanent inode numbers (via pointer conversion)

in linux-fsdevel, ~Nov 17th for more info.

-- Jeff
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/