Re: Metadata usage following device add

[Chris Webb <chris@arachsys.com>]

Kent Overstreet <kent.overstreet@gmail.com> writes:

> Good catch! I completely forgot about that comment and didn't think to
> fix it that way, the last time I was looking at that I was looking at the
> device accounting code in buckets.c. I've just pushed a fix, can you tell
> me if you see anything I missed?

Thanks! I've played with it and tried everything I can think of to break it,
but I didn't find a way to get it out of sync or inconsistent.

> This is a good opportunity to talk about the disk space accounting code.

That was a really interesting write-up of your planned transition to move
the accounting into btree keys, thanks! Makes a lot of sense as you describe
it. (I saw you mention hitting 1M snapshots on a filesystem the other day,
which is good motivation for a scalable place to put accounting info!)

I need to spend a bit of time reading the code and getting a better sense of
how the allocation system more generally works. It's still one of the more
mysterious bits of the filesystem to me.

Am I right that in the current implementation, buckets only get discarded
when the allocator thread is woken because we've run out of free buckets for
a new allocation and need to reclaim - so they're discarded and quickly
reused? I wonder if waking the allocator thread periodically (or from a
userspace ioctl?) to reclaim and discard buckets earlier might be better for
SSD health, kind of like a fancy version of a manual fstrim on ext4?

Best wishes,

Chris.