Re: [Lse-tech] Re: [PATCH] cpusets+hotplug+preepmt broken

[Dinakar Guniguntala <dino@in.ibm.com>]

On Wed, May 11, 2005 at 02:51:56PM -0500, Nathan Lynch wrote:

> This trace is what should be fixed -- we're trying to schedule while
> the machine is "stopped" (all cpus except for one spinning with
> interrupts off).  I'm not too familiar with the cpusets code but I
> would like to stay away from nesting these semaphores if at all
> possible.

Vatsa pointed out another scenario where cpusets+hotplug is currently
broken. attach_task in cpuset.c is called without holding the hotplug 
lock and it is possible to call set_cpus_allowed for a task with no 
online cpus. 
Given this I think the patch I sent first is the most appropriate
patch. In addition we also need to take hotplug lock in the cpusets
code whenever we are modifying cpus_allowed of a task. IOW make cpusets 
and hotplug operations completly exclusive to each other. The same 
applies to memory hotplug code once it gets in.

However on the downside this would mean 
1. A lot of nested locks (mostly in cpuset_common_file_write)
2. Taking of hotplug (cpu now and later memory) locks for operations
   that may just be updating a flag


	-Dinakar