On Thu, 23 Nov 2023 at 07:57, Gabriel Krisman Bertazi
<gabriel@xxxxxxxxxx> wrote:
>
> The problem I found with that approach, which I originally tried, was
> preventing concurrent lookups from racing with the invalidation and
> creating more 'case-sensitive' negative dentries. Did I miss a way to
> synchronize with concurrent lookups of the children of the dentry? We
> can trivially ensure the dentry doesn't have positive children by
> holding the parent lock, but that doesn't protect from concurrent
> lookups creating negative dentries, as far as I understand.
I'd just set the "casefolded" bit, then do a RCU grace period wait,
and then invalidate all old negative dentries.
Sure, there's technically a window there where somebody could hit an
existing negative dentry that matches a casefolded name after
casefolded has been set (but before the invalidation) and the lookup
would result in a "does not exist" lookup that way.
But that seems no different from the lookup having been done before
the casefolded bit got set, so I don't think that's an _actual_
difference. If you do a lookup concurrently with the directory being
set casefolded, you get one or the other.
And no, I haven't thought about this a ton, but it seems the obvious
thing to do. Temporary stale negative dentries while the casefolded
bit is in the process of being set seems a harmless thing, exactly
because they would seem to be the same thing as if the lookup was done
before...
And yes, that "wait for RCU grace period" is a somewhat slow
operation, but how often do the casefolded bits get changed?
This is not a huge deal. I don't hate your approach, I just found it surprising.
Linus
