On Thu, Aug 11, 2022 at 03:22:38PM -0700, Linus Torvalds wrote: > On Thu, Aug 11, 2022 at 3:04 PM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote: > > > > FWIW, I wonder if we should do > > if (READ_ONCE(dentry->d_parent) != parent) > > continue; > > before grabbing ->d_lock (and repeat the check after grabbing it, > > It kind of makes sense. We already do that d_name.hash check outside > of the lock, so we already have that "we might race with a rename" > situation. > > That said, I do think __d_lookup_rcu() is the more important of the two. > > Here's a recreation of that patch I mentioned where the OP_COMPARE is > moved out of the loop. Just for fun, look at how much better the code > generation is for the common case when you don't have the call messing > up the clobbered registers etc. > > Entirely untested, and I might have messed something up, but I suspect > this is a much bigger deal than whether d_same_name() is inlined or > not in the non-RCU path. Looks sane at the first pass, but right now I'm really half-asleep - 5 hours of sleep tonight, and about the same yesterday ;-/ I'll try to grab at least an hour or two and reread it once I'm more or less awake...
