Frank van der Linden wrote:
> We do something similar for flushed delayed writes (B_VFLUSH). I'm not
> sure how much of an optimization this particular one may be for us,
> but I'm very hesitant to adopt any changes like this from the
> FreeBSD code, because 1) we will get UBC when 1.5 is finally branched,
> which changes everything, and 2) the FreeBSD vfs_bio seems to have grown
> to a size that scares me, which also makes me think twice about
> adopting code from it. I'm already unhappy with the amount of
> special-cased code that softdep brought with it.
> 
> Of course, if you bring it in in a clean way, and can show the
> performance gains..

Well, I can't show the actual performance gains, it just seems
like a good thing - just like delayed writes, bqrelse() avoids
the (read-only) buffer to be reused for other stuff too soon. Maybe
it could be implemented as a wrapper around brelse() ?
It can set the B_VFSLUSH flag and pass it to brelse() probably.

Jaromir