On Wed, Nov 20, 2024 at 12:13 PM Christian Brauner <brauner@xxxxxxxxxx> wrote: > > On Wed, Nov 20, 2024 at 11:42:33AM +0100, Mateusz Guzik wrote: > > Interestingly even __read_seqcount_begin (used *twice* in path_init()) > > is missing one. I sent a patch to fix it long time ago but the > > recipient did not respond > > I snatched it. Thanks. But I have to say having *two* counters to check for each lookup is bothering me and making me wonder if they could be unified (or another counter added to cover for either of those?)? No clue about feasibility, is there a known showstopper? Both are defined like so: __cacheline_aligned_in_smp DEFINE_SEQLOCK(mount_lock); __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock); Suppose nothing can be done to only look at one counter on lookup. In that case how about combining the suckers into one cacheline at least? Sure, this will result in new bounces for threads modifying these, but this is relatively infrequent compared to how often lookups performed and with these slapped together there will be only one line spent on it, instead of two. Just RFC'ing it here. -- Mateusz Guzik <mjguzik gmail.com>
