Jason Thorpe writes:
> Async metadata writes do give "better performance", but IMO are the
> wrong way to get better performance.  If you crash before your metadata
> has been written, there is no consistent information from which you can
> repair the file system.  Basically, it's bloody dangerous, especially if
> you consider your data important.

The right thing, really, is for someone to finally do two needed fixes
to the system.

1) We need to be able to express block I/O dependencies so that we can
   do "nearly" as well as async metadata flushes but with guaranteed
   safety.
2) We need a better data structure than a linear list for long
   directories -- this makes a difference particularly for things like
   news server performance on expires and such, as, for that matter,
   does 1).

Perry