90% is about where I'd expect problems to start happening with our
current implementation.  Performance at close to capacity is a
weakness of LFS in general, AFAIK.

Moreover, LFS is unfortunately not always stable on -current (in the
MP case).  Also, IIRC, lfsresbuf is one place with a [possibly]
problematic call to ltsleep().  I've done some testing using
decreasing values of vfs.lfs.pagetrip on -current, and notice that
locking problems manifest more and more the smaller the "page trip"
gets, increasing geometrically after it's lowered below
segsize/PAGE_SIZE.  In your case, mega-maid, e.g. the cleaner, acting
up may also be an issue.

If you use ffs+softdep and raise kern.maxvnodes, you likely will get
the better performance writing tons of small files that you're looking
for.

Regards,

--Blair