On 04/16/2012 10:55 AM, Zheng Liu wrote:
Hi Allison,
Currently I am trying to reduce the lock contention of direct I/O in ext4
because it is a bottleneck. A trivial idea is that a new fucntion is
defined to replace the generic_file_aio_write, which do some write
operations with acquiring i_data_sem lock in inode.
I know that you are trying to implement extent tree, and I have seen your patch
set '[PATCH] Rename delayed extents to status extents'. After extent tree is
made, the implementation of direct I/O without i_mutex and range lock is
straightforward and it is better than my trivial idea. I think that
maybe I can borrow you works. So could you please share me your schedule
and/or other information?
Last month on ext4 workshop, we discuss the extent tree, range lock and I/O
tree. Obviously, I/O tree is used to store I/O operations, which can track
delay allocation, do unwritten->written conversion and implement range lock.
It is very useful for ext4 and I am interested in this proposal. I know that
you have begun to do some works. So would you like to tell me the status of
extent tree? I don't know whether or not there has some things that I can be
involved. If you have some advices or there is something that I can help,
please let me know. Thank you and looking forward your reply.
Regards,
Zheng
Hi Zheng,
Well, I can share with you what I have done so far with Yongqiang's
delayed extent tree, but since I was moved to the ganesha project,
extent locks are no longer a business priority now. I've done some work
on it on my own time, but I have not been able to keep up pace with it.
If someone else has the hours to push it faster than I can at this
point, I would certainly be understanding of that.
Basically though the plan was to modify Yongqiang's delayed extent tree
to track allocated extents as well as delayed extents. And then add the
extent locks on top of that once that's working. My idea was to add a
type member to the extents so that we have have "delayed" "allocated"
and "hole" extents. I've renamed the delayed extent scheme to "status
extents" because it seemed more appropriate. That part I can send out,
but I was still in the middle of coding and debugging allocated extents
and extent locks. Because the tree was originally written to merge
things as much as possible, there is some rewrite that is needed. I was
working on making the add and remove routines more like "type replace"
routines sense once the extent is locked, we can really only change the
type, and we cannot merge dissimilar types or extents locked by other
processes. I hope that makes sense, please let me know if you need more
clarification. I will send out the status extent set, sense that part
is stable.
Allison Henderson
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
