Date: Tue, 15 Mar 2005 09:28:29 -0500 From: Stephan Uphoff <ups@tree.com> To: Jeff Roberson <jroberson@chesapeake.net> Cc: arch@freebsd.org Subject: Re: Freeing vnodes. Message-ID: <1110896909.29804.39143.camel@palm> In-Reply-To: <20050315003915.C20708@mail.chesapeake.net> References: <20050314213038.V20708@mail.chesapeake.net> <1110856553.29804.37784.camel@palm> <20050315003915.C20708@mail.chesapeake.net>
next in thread | previous in thread | raw e-mail | index | archive | help
On Tue, 2005-03-15 at 00:39, Jeff Roberson wrote: > On Mon, 14 Mar 2005, Stephan Uphoff wrote: > > > On Mon, 2005-03-14 at 21:38, Jeff Roberson wrote: > > > I have a patch at http://www.chesapeake.net/~jroberson/freevnodes.diff > > > that allows us to start reclaiming vnodes from the free list and release > > > their memory. It also changes the semantics of wantfreevnodes, and makes > > > getnewvnode() much prettier. > > > > > > The changes attempt to keep some number of vnodes, currently 2.5% of > > > desiredvnodes, that are free in memory. Free vnodes are vnodes which > > > have no references or pages in memory. For example, if an application > > > simply stat's a vnode, it will end up on the free list at the end of the > > > operation. The algorithm that is currently in place will immediately > > > recycle these vnodes once there is enough pressure, which will cause us to > > > do a full lookup and reread the inode, etc. as soon as it is stat'd again. > > > > > > This also removes the recycling from the getnewvnode() path. Instead, it > > > is done by a new helper function that is called from vnlru_proc(). This > > > function just frees vnodes from the head of the list until we reach our > > > wantfreevnodes target. > > > > > > I haven't perf tested this yet, but I have a box that is doing a > > > buildworld with a fairly constant freevnodes count which shows that vnodes > > > are actually being uma_zfree'd. > > > > > > Comments? Anyone willing to do some perf tests for me? > > > > > > Thanks, > > > Jeff > > > > Just looked at the raw diff and might have missed it - how are the > > parent directory "name" cache entries ( vnode fields v_dd, v_ddid) > > handled? > > Just as they were before, by calling cache_purge. This purges the fields of the vnode that will be recycled. I am worried about the v_dd,v_ddid fields of a directory B that has the to be released vnode A as parent. (Obviously in this case there is no namecache entry with the vnode A as the directory (nc_dvp)) Right now A is type stable - but if A is released, access to B->v_dd may cause a page fault. Stephan
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?1110896909.29804.39143.camel>