VFS && UFS write support: possible on *BSD/Solaris, impossible on Linux?

VFS && UFS write support: possible on *BSD/Solaris, impossible on Linux?

[Belan Kumar]

Hello.

May be better call subject something like: "Shortest way to ufs write support",
but current subject give more chances to recieve answer.

First of all I try explain the current problem:
UFS as other block file systems has notition of "block",
but further to "block" notion it has "fragment" concept.
"fragment" used for preventing waste of space.
Usually fragment==block/8, and if "the rest of file" doesn't occupy
the whole block, it occupy several fragments. If file grows, at some point
"the rest of file"  will be occupy 8 fragments. And HERE IS PROBLEM,
we should allocate block and move 8 fragments to it.

On *BSD/Solaris it is simple: they read analog of buffer_head and
change analog of b_blocknr and that's all.

The current code of fs/ufs/balloc.c does the same:
----------------------
sb_bread
bh->b_blocknr =
mark_buffer_dirty (bh)
brelse (bh)
--------------------------
I suppose you guess: it doesn't work,
latter when you do sb_getblk(old_blocknr) it give to us the same block
and after that kernel hang up.


Hence question: what would be the proper solution in this situation?
Is it possible in current VFS change b_blocknr?

Re: VFS && UFS write support: possible on *BSD/Solaris, impossible on Linux?

[Jan Kara]

  Hello,

> May be better call subject something like: "Shortest way to ufs write support",
> but current subject give more chances to recieve answer.
> 
> First of all I try explain the current problem:
> UFS as other block file systems has notition of "block",
> but further to "block" notion it has "fragment" concept.
> "fragment" used for preventing waste of space.
> Usually fragment==block/8, and if "the rest of file" doesn't occupy
> the whole block, it occupy several fragments. If file grows, at some point
> "the rest of file"  will be occupy 8 fragments. And HERE IS PROBLEM,
> we should allocate block and move 8 fragments to it.
> 
> On *BSD/Solaris it is simple: they read analog of buffer_head and
> change analog of b_blocknr and that's all.
> 
> The current code of fs/ufs/balloc.c does the same:
> ----------------------
> sb_bread
> bh->b_blocknr =
> mark_buffer_dirty (bh)
> brelse (bh)
> --------------------------
> I suppose you guess: it doesn't work,
> latter when you do sb_getblk(old_blocknr) it give to us the same block
> and after that kernel hang up.
> 
> 
> Hence question: what would be the proper solution in this situation?
> Is it possible in current VFS change b_blocknr?
  I guess it is possible but certainly it's not that easy as just
changing b_blocknr. Each buffer head is attached to a page (either to a
page of a backing device mapping or to a page belonging to the inode
mapping). And if you look at e.g. __find_get_block_slow() you'll see
that we look for the buffer using that page. So you definitely need to
somehow attach the buffer to the new page at the correct location.

								Honza
-- 
Jan Kara <jack@suse.cz>
SuSE CR Labs