> > > Why would these reads be generated?  If the write has not completed, is
> > > it not the case that the page must still be dirty, and should thus still
> > > be in the cache?
> > 
> > because the page is marked as PG_BUSY during delayed write,
> > VOP_GETPAGES will wait on it.  we have no way to hurry
> > the i/o in this case.
> > 
> > maybe distinguishing PG_BUSY-for-read and PG_BUSY-for-write can be
> > an alternative solution?
> 
> Won't this also help avoid reads, period, when the I/O queues get lots
> of writes backed up in them?  That seems like a very good thing.

anyway, it won't help other cases like fsync on the
delayed writed pages, etc.

YAMAMOTO Takashi