On Wed, 22 May 2024 at 10:14, Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Wed, May 22, 2024 at 09:00:03AM -0700, Linus Torvalds wrote:
> > Of course, if you do billions of lookups of different files that do
> > not exist in the same directory, I suspect you just have yourself to
> > blame, so the "lots of negative lookups" load doesn't sound
> > particularly realistic.
>
> Oh no. We have real customers that this hits and it's not even stupid.
Oh, negative dentries exist, and yes, they are a major feature on some
loads. Exactly because of the kinds of situations you describe.
In fact, that's the _point_. Things like PATH lookup require negative
dentries for good performance, because otherwise all the missed cases
would force a lookup all the way out to the filesystem.
So having thousands of negative dentries is normal and expected. And
it will grow for bigger machines and loads, of course.
That said, I don't think we've had people on real loads complain about
them being in the hundreds of millions like Yafang's case.
> plan9 handles this so much better because it has that union-mount stuff
> instead of search paths. So it creates one dentry that tells it which of
> those directories it actually exists in. But we're stuck with unix-style
> search paths, so life is pain.
I suspect you are just not as aware of the downsides of the plan9 models.
People tend to think plan9 was great. It had huge and fundamental
design mistakes.
Union mounts à la plan9 aren't hugely wonderful, and there's a reason
overlayfs does things differently (not that that is hugely wonderful
either).
Linus
