Re: [PATCH 5/7] kiothrottled: throttle buffered (writeback) IO

[Andrea Righi <righi.andrea@gmail.com>]

On Thu, Apr 23, 2009 at 03:53:51PM +0800, Gui Jianfeng wrote:
> Andrea Righi wrote:
> > Together with cgroup_io_throttle() the kiothrottled kernel thread
> > represents the core of the io-throttle subsystem.
> > 
> > All the writeback IO requests that need to be throttled are not
> > dispatched immediately in submit_bio(). Instead, they are added into an
> > rbtree by iothrottle_make_request() and processed asynchronously by
> > kiothrottled.
> > 
> > A deadline is associated to each request depending on the bandwidth
> > usage of the cgroup it belongs. When a request is inserted into the
> > rbtree kiothrottled is awakened. This thread selects all the requests
> > with an expired deadline and submit the bunch of selected requests to
> > the underlying block devices using generic_make_request().
> 
>   Hi Andrea,
> 
>   What if an user issues "sync", will the bios still be buffered in the rb-tree?
>   Do we need to flush the whole tree?

Good question. From The sync(2) man page:

	According to the standard specification (e.g., POSIX.1-2001), sync()
	schedules the writes, but may return before the actual writing is done.
	However, since  version  1.3.20  Linux does  actually  wait.  (This
	still  does not guarantee data integrity: modern disks have large
	caches.)

It is not completely wrong looking at the standard. The writes are
actually scheduled, but pending in the rbtree. Anyway, if we immediately
dispatch them anyone can evade the IO controller simply issuing a lot of
sync while doing IO. OTOH dispatching the requests respecting the max
rate for each cgroup can cause the sync to wait for all the others' BW
limitations.

Honestly I don't have a good answer for this. Opinions?

-Andrea