On Fri, 2012-08-03 at 09:29 -0400, Jeff Layton wrote: > This patch is an RFC for dealing with a regression in write performance > when there are multiple processes doing I/O to the same pages on the > same machine. Earlier this week, I outlined the problem here: > > http://www.spinics.net/lists/linux-nfs/msg31728.html > > This patch seems to resolve the problem, but I still haven't quite > convinced myself that this approach is race-free. Thoughts? I'd prefer to replace the "pointer to lock owner" with a "pointer to open/lock stateid". By tagging the data with the specific stateid, that was in effect when we created the nfs_page, then we know when to fence off that particular READ or WRITE if a network partition causes an irrecoverable loss of the lock. I'm planning on drafting a patch for this tomorrow... Cheers Trond ��.n��������+%������w��{.n�����{��w��� jg��� �����ݢj����G���� ���j:+v���w�m������w�������h�����٥
