On Tue, Aug 14, 2007 at 05:32:29AM -0700, Daniel Phillips (phillips@phunq.net) wrote:

On Tuesday 14 August 2007 04:50, Evgeniy Polyakov wrote:

quoted

On Tue, Aug 14, 2007 at 04:35:43AM -0700, Daniel Phillips 

(phillips@phunq.net) wrote:

quoted

quoted

On Tuesday 14 August 2007 04:30, Evgeniy Polyakov wrote:

quoted

quoted

And it will not solve the deadlock problem in general.  (Maybe
it works for your virtual device, but I wonder...)  If the
virtual device allocates memory during generic_make_request
then the memory needs to be throttled.

Daniel, if device process bio by itself, it has a limit and thus
it will wait in generic_make_request()

What will make it wait?

gneric_make_request() for given block device.

Not good enough, that only makes one thread wait.  Look here:

    http://lkml.org/lkml/2007/8/13/788

An unlimited number of threads can come in, each consuming resources of 
the virtual device, and violating the throttling rules.

The throttling of the virtual device must begin in generic_make_request 
and last to ->endio.  You release the throttle of the virtual device at 
the point you remap the bio to an underlying device, which you have 
convinced yourself is ok, but it is not.  You seem to miss the fact 
that whatever resources the virtual device has allocated are no longer 
protected by the throttle count *of the virtual device*, or you do not 

Because it is charged to another device. No matter how many of them are
chained, limit is applied to the last device being used.
So, if you have unlimited number of threads, each one allocates a
request, forward it down to low-level devices, each one will eventually
sleep, but yes, each one _can_ allocate _one_ request before it goes
sleeping. It is done to allow fain-grained limits, since some devices
(like locally attached disks) do not require throttling.

Here is an example with threads you mentioned:
http://article.gmane.org/gmane.linux.file-systems/17644

-- 
	Evgeniy Polyakov