On 2024/10/17 4:47, Theodore Ts'o wrote:
On Wed, Oct 16, 2024 at 04:02:40PM +0800, Baokun Li wrote:
As server clusters get larger and larger, server maintenance becomes very
difficult. Therefore, timely detection of problems (i.e. online scanning,
similar to e2fsck -fn) and timely and non-stop fixing of problems (i.e.
online fsck, similar to e2fsck -a) have always been the requirements of
our customers. Thus online fsck has been on our TODO list, and it's really
time to start doing it. 😀
As far as online scaning is concerned, if you are using LVM, we can
use a combination of dm-snapshot and e2fsck -fn --- that is what the
e2scrub command automates.
Yes, e2scrub is very nice, but it has too many limitations. We have some
customers who don't use lvm.
Online fsck is much harder, since it would require back pointers to do
this efficienctly.
Indeed, our rough plan is to first implement isolation of abnormal file
system resources, so that the system can continue to run normally even
when there is an error; then implement online scanning, so that the
maintainer can see the health report at any time; and finally implement
the most difficult online repair.
To do this, a general way of solving this would
involve a generalized some kind of b-tree or b+tree where changes are
managed via jbd2. This could be used so that (for example) if we had
a tree which maps block ranges to an inode number, then given a block
number, we can figure out which inode "owns" that block. The harder
part is those objects that have multiple forward pointers --- for
example an inode might have multiple hard links to multiple
directories, so we need to handle this somehow.
We do need to establish the mapping of physical blocks to inodes and inodes
to parent dir. By tree managed by jbd2 do you mean updating the tree when
committing to journal? Or are updates to the tree logged to journal?
If we had the jbd2-aware b+tree, we could also use this add support
for reflink/clone, which would also be cool.
Yeah, reflink is pretty cool, we can try it out when the others are done.
If this is something that your team really weants to work on, what I'd
suggest is to create a rough design of what the journaled b+tree would
look like, and then implement it first, since this is the prerequisite
for a huge number of advanced file system features. Implementation
should be done in a way that makes it easy for the code to be usable
both in the kernel and in e2fsprogs, since life will be much easier if
we have e2fsck and debugfs support for the new file system data
structures from the very beginning of the development.
Thank you for your suggestion! This is really key to the development. We'll
discuss the overall design internally before consulting the community.
If your company is willing to invest in the engineering effort to do
this, great! But I have to point out that an alternative approach
that you should consider is whether XFS might be a closer match for
some of your customers' needs. The advantage of ext4 is that it is
much simpler and easier to understand that XFS.
The XFS maintainability enhancement is something my colleague is working
on. But we have a fair number of downstream customers who prefer ext4, so
it's worth investing the time to do that.
But as we add these
new features, ext4 will get more complex. And so one of the design
considerations we should keep in mind is to keep ext4 as simple and
miantainable as possible, even as we add new functionality.
Cheers,
- Ted
Of course! we will keep the code as simple and maintainable as possible.
Thanks again for your input! 😉
Cheers,
Baokun
