Re: location of bioq lock

From: Luigi Rizzo <rizzo_at_icir.org>
Date: Thu, 14 Jul 2005 18:13:56 -0700
On Thu, Jul 14, 2005 at 06:46:23PM -0600, Scott Long wrote:
...
> How often are you going to be changing the scheduler at runtime?  Is

in general, almost never. However, suppose you have per-device
schedulers as phk was suggesting, then you'd really like
to decouple the devices so reconfiguring one of them does not
affect the others.

Anyways there might be a way out...
requests go from dev_strategy()
to the individual disksort routine where, during the switch,
the scheduler can 'absorb' them
while the real queue drains (assuming the subsequent device_start()
routine does not complain too much for finding the queue empty!).
Then, when the device queue becomes empty (the scheduler knows because
it is serving the bio_remove() requests) it can switch to the new one,
and resubmit old requests through dev_strategy().

This would also solve the problem of implementing non-work-conserving
schedulers. Basically it's the same approach followed in dummynet,
you steal the packets from ip_{input,output}() and resubmit them
later.

Does any of you know what are the assumptions (locks held etc.)
for calling dev_strategy() in 5.x and above ?

	cheers
	luigi

> the scheduler meant to be aware of individual devices?  Again, I'm
> not advocating that the upper layers be able to look at or manipulate
> the driver bioq's.
> 
> Scott
> _______________________________________________
> freebsd-current_at_freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-current
> To unsubscribe, send any mail to "freebsd-current-unsubscribe_at_freebsd.org"
Received on Thu Jul 14 2005 - 23:13:57 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:38:38 UTC