Chris Mason wrote:
On Sun, Jun 24, 2007 at 03:46:13AM +0200, Nick Piggin wrote:
Rewrite the buffer layer.
Overall, I like the basic concepts, but it is hard to track the locking
rules. Could you please write them up?
Yeah I will do that.
Thanks for taking a look. One thing I am thinking about is to get
rid of the unmap_underlying_metadata calls from the generic code.
I found they were required for minix to prevent corruption, however
I don't know exactly what metadata is interfering here (maybe it
is indirect blocks or something?). Anyway, I think I will make it
a requirement that the filesystem has to already handle this before
returning a newly allocated block -- they can probably do it more
efficiently and we avoid the extra work on every block allocation.
I like the way you split out the assoc_buffers from the main fsblock
code, but the list setup is still something of a wart. It also provides
poor ordering of blocks for writeback.
Yeah, I didn't know how much effort to put in here because I don't
know whether modern filesystems are going to need to implement their
own management of this stuff or not.
I haven't actually instrumented this in something like ext2 to see
how much IO comes from the assoc buffers...
I think it makes sense to replace the assoc_buffers list head with a
radix tree sorted by block number. mark_buffer_dirty_inode would up the
reference count and put it into the radix, the various flushing routines
would walk the radix etc.
If you wanted to be able to drop the reference count once the block was
written you could have a back pointer to the appropriate inode.
I was actually thinking about a radix-tree :) One annoyance is that
unsigned long != sector_t :P rbtree would probably be OK.
--
SUSE Labs, Novell Inc.
-
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html