Re: [dm-devel] dm-table: delayed cleanup for dm_table_destroy()

[Mikulas Patocka <mpatocka redhat com>]

On Tue, 20 Mar 2012, Mike Snitzer wrote:

> I once wanted to replace all msleep(1); with cpu_relax();:
> http://www.redhat.com/archives/dm-devel/2010-September/msg00100.html
> 
> But Alasdair wasn't sure if cpu_relax() would provide the required
> delay effect:
> https://www.redhat.com/archives/dm-devel/2011-June/msg00080.html

cpu_relax() would basically kill the kernel if compiled without preempt.

cpu_relax() makes the current processor sleep for a little moment, but it 
doesn't schedule a different process.

cpu_relax() is useful in spinlocks - so that process waiting on a spinlock 
is not slowing down the other process on a CPU with hyperthreading.


There was another possibility --- replace msleep(1) with yield(). The 
problem with yield() is that if the process has readltime priority and 
calls yield(), it doesn't give CPU to a different process with a lower 
priority. So it would cause deadlock if executed with realtime priority. 
msleep(1) doesn't have this problem, so it's the best.

Mikulas