> > Something that's been discussed before, but is worth mentioning again
> > at this point - I'd just love to see a mechanism that allowed
> > really-not-very-interesting requests (the canonical example being
> > atime updates) to wait ~forever if the disk is not spun up, and then
> > to proceed when some other higher-priority request wakes up the disk
> > (or there are too many such waiting?)
> > 
> > That needs some help from the layers above, and probably below.
> 
> I'd like to see this too. IMHO, this can share the same queue as delayed
> writes. Only wake up the disk when there's too much requests in the queue, or
> we have something else to do.
> This would also prevent writes to log files to wake up the disk.

it isn't so easy.
"delayed writes" can't be delayed so much because e.g. synchronous
read requests can happen on the delayed-writed page and we have no way
to look up corresponding buf from a page.

YAMAMOTO Takashi