Date: Wed, 05 Jun 2013 17:52:34 +0800 From: David Xu <davidxu@freebsd.org> To: John Baldwin <jhb@freebsd.org> Cc: arch@freebsd.org Subject: Re: [PATCH] Allow atomic sets of non-overlapping CPU sets for a global cpuset Message-ID: <51AF0A62.4040206@freebsd.org> In-Reply-To: <201305311216.56558.jhb@freebsd.org> References: <201305311216.56558.jhb@freebsd.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On 2013/06/01 00:16, John Baldwin wrote: > So there's an oddity with cpuset I've run into recently at work. Suppose I > have created a new cpuset and want to change the set of CPUs for that set (say > from a mask of just CPU 1 to a mask of just CPU 2). I can't do that > atomically. I have to first set the mask to contain both the old set (CPU 1) > and the new set (CPU 2) and then change it a second time to only contain the > new set (CPU 2). The reason is that cpuset_modify() runs cpuset_testupdate() > on the set it is about to modify, so when I try to change it in a single > operation the new mask doesn't overlap with the old mask and it fails with > EDEADLK. > > % cpuset -c -l 1 /bin/sh > $ cpuset -gi > pid -1 cpuset id: 2 > $ cpuset -g > pid -1 mask: 1 > $ cpuset -l 2 -s 2 > cpuset: setaffinity: Resource deadlock avoided > > I think that the correct logic here is that we should only check descendants > of the set we are changing, but not the set we are about to change. The patch > does this and allows my test case above to work: The patch looks fine to me.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?51AF0A62.4040206>