On Tue, Sep 09, 2025 at 10:35:15AM -0400, Liam R. Howlett wrote: > * Uladzislau Rezki <ure...@gmail.com> [250909 05:08]: > > ... > > > > > > > - call_rcu() can be slow, therefore we do not use it in the > > > > kvfree_rcu(). > > > > > > If call_rcu() is called once per 32 kfree_rcu() filling up the rcu sheaf, > > > is > > > it still too slow? > > > > > You do not know where in a queue this callback lands, in the beginning, > > in the end, etc. It is part of generic list which is processed one by > > one. It can contain thousands of callbacks. > > How does this differ from kvfree_rcu()? > > Surely if you have enough calls to kvfree_rcu(), you will end up with a > large list of frees before the end of a grace period? Our placement in > the freeing order would still be dependent on what else is using the > infrastructure in the same grace period, right? > In kfree_rcu() we use page blocks to carry pointers. Lists can be used if there is a low memory condition so a page can not be allocated or cache is empty. But this is not part of carr_rcu() track in any way.
Right regular call_rcu() puts callback into its own internal lists and they are processed one by one during list iteration. In such lists you can have hundred of thousand callback. > > How is kvfree_rcu() affected by rcu callback offloading to a specific > cpu and rcu expedite? Often these two features come into play for > certain workloads which are of concern to us. > We maintain a separate path. Offload is done after a grace period is over. It is classic way. Historically all deferred freeing was one call_rcu() per ptr. > > > > If performance is not needed then it is not an issue. But in > > kvfree_rcu() we do not use it, because of we want to offload > > fast. > > Today, I free things using call_rcu() and a custom callback so I would > think stacking 32 together would make the list shorter, but latency > would increase by waiting until there are 32. > > If we wanted to flush the kvfree_rcu() list, is it done in the same way > as the call_rcu() list, or is there a better way? > For this case we have kvfree_rcu_barrier(). It is not same as call_rcu() flushing. -- Uladzislau Rezki
