Re: 2.6.5-rc1-mm2 and direct_read_under and wb

[Andrew Morton <akpm@osdl.org>]

Daniel McNeil <daniel@osdl.org> wrote:
>
> Andrew,
> 
> Just to confirm my tests ran overnight without errors.

good.

> So the !wbc->nonblocking lowers the cpu utilization.

yup.  It prevents balance_dirty_pages() callers from spinning over the same
pages again and again.

> What does sync_mode=WB_SYNC_ALL and nonblocking=1 mean?

WB_SYNC_ALL means "we're performing this writeout for data-integrity
purposes, as opposed to for a regular dirty memory flush".  The code paths
for these two things are deep, and are almost identical, so we pass it up
and down in a flag.

->nonblocking means "try to avoid blocking on request queues or locked
pages".  pdflush and kswapd set this because those processes must serve
many queues and cannot accord to get stuck on any one queue.  But regular
write() callers do not set ->nonblocking because we _want_ these guys to
throttle.

It's only OK to let pdflush avoid blocking on the buffer because pdflush
will take an explicit nap in background_writeout().  hm, but only if the
queue is congested - we may still have a problem there.