Ext4 Developer Interlock Call: 01-24-07 Minutes
Attendees: Mingming Cao, Suparna Bhattacharya, Dave Kleikamp, Laurent Vivier, Valerie Clement, Badari Pulavarty, Avantika Mathur
Minutes can be accessed at: http://ext4.wiki.kernel.org/index.php/Ext4_Developer%27s_Conference_Call
- Laurent plans to propose a paper on 64 bit work for ext4, describing what is missing in the current implementation for 64 bit ext4.
- We also plan to submit a proposal for a collaborative paper on ext4 in general. This paper would include an overview the features currently in ext4, performance evaluation, and planned features.
- Mingming has already submitted a proposal for a BOF on ext4
Directory Page Cache:
- Andrew Morton would like to see directories moved to a per directory pagecache, which would allow readahead on directories, and also address directory fragmentation issues.
- Badari mentioned that the reason directories were in pagecache in ext2 and not in ext3 may be the htree extension in ext3.
- We will look into how this can be added and supported in ext4
- The preallocation patches are ready to be submitted for inclusion; other than deciding on an interface.
- Mingming suggested abstracting the preallocation implementation from interface. But we will wait until the interface is decided.
- Still looking into posix_fallocate and ioctl as options for interface; this discussion will be moved to larger mailing lists, in order to reach consensus and move patches forward.
Delayed and Multiple Block Allocation Patches:
- Delayed allocation: Alex Tomas' patches have a few issues still to be resolved
- Multiple block allocation:
- This is ext4 specific code and can be pushed to mainline before delayed allocation.
- Ask Alex Tomas if he would like to shift focus to mult block allocation, over delayed allocation.
Inode Version Patches:
- the inode version patches recently resubmitted by Jean-Noel Cordenner are at the vfs level and port the i_version field to user space.
- There are concerns about the peformance with the overhead of updating the i_version each time ctime is updated.
- Amit Gud was working on an in kernel implementation for chunkfs, but there haven't been any patches posted yet. He does not seem to have much time to devote to this develpment. There was also a group of students working on chunkfs.
- chunkfs was originally based on ext2. One goal was to improve filesystem recovery. Developers looked at how to measure and compare fsck times based not only on time, but reliability and being thorough.
- Status on chunkfs development can be discussed at the upcoming filesystem workshop