On Wed, Feb 07, 2018 at 10:10:55AM -0800, Matthew Wilcox wrote: > On Wed, Feb 07, 2018 at 05:45:13PM +0100, Jesper Dangaard Brouer wrote: > > On Wed, 7 Feb 2018 08:57:00 -0500 > > Steven Rostedt <rost...@goodmis.org> wrote: > > > To me kvfree() is a special case and should not be used by RCU as a > > > generic function. That would make RCU and MM much more coupled than > > > necessary. > > > > For the record, I fully agree with Steve here. > > > > And being a performance "fanatic" I don't like to have the extra branch > > (and compares) in the free code path... but it's a MM-decision (and > > sometimes you should not listen to "fanatics" ;-)) > > While free_rcu() is not withut its performance requirements, I think it's > currently dominated by cache misses and not by branches. By the time RCU > gets to run callbacks, memory is certainly L1/L2 cache-cold and probably > L3 cache-cold. Also calling the callback functions is utterly impossible > for the branch predictor.
This seems to have fallen by the wayside. To get things going again, I suggest starting out by simply replacing the kfree() in __rcu_reclaim() with kvfree(). If desired, a kvfree_rcu() can also be defined as a synonym for kfree_rcu(). This gets us a very simple and small patch which provides the ability to dispose of kvmalloc() memory after a grace period. Thanx, Paul