This means no special case for uninling-but-not-splitting, but allows
the entire split route to be deduplicated, simplifying things.
The main downside is that for littlefs to go from a single inlined mdir
filesystem to an mtree filesystem it requires a minimum of 2 mdir
allocations (4 blocks) in all cases. This can be avoided, but I think is
worth the tradeoff since it generally occurs once in a filesystem's
lifetime.
This does make 4 block block devices a bit awkward, but those geometries
are always going to be a bit awkward with littlefs's design. At least
this implementation avoids an unecessary B-tree node where possible...
code stack
before: 22586 2320
after: 22414 (-0.8%) 2120 (-8.6%)
I _think_, but haven't verified, the significant stack saving comes from
the fact that since there's one route through lfsr_mtree_split_,
lfsr_mtree_split_ can be inlined into lfsr_mdir_commit. This avoids the
marshalling of all its arguments for the function call, which I've
noticed can have a surprising cost.
---
Also fixed a bug where dstart was not updated with mid changes after
splits/drops. The mdir commit cleanup code has a lot of duplication now,
makes me wonder if there's a better way to structure this.
Christopher Haster
adcf9924fe