Re: [RFC PATCH 2/4] mm: shmem: Support case-insensitive file name lookups

[André Almeida]

Hi Al Viro,

Às 20:19 de 23/03/21, Al Viro escreveu:
> On Tue, Mar 23, 2021 at 04:59:39PM -0300, André Almeida wrote:
>
> > * dcache handling:
> >
> > For now, negative lookups are not inserted in the dcache, since they
> > would need to be invalidated anyway, because we can't trust missing file
> > dentries. This is bad for performance but requires some leveraging of
> > the VFS layer to fix. We can live without that for now, and so does
> > everyone else.
>
> "For now"?  Not a single practical suggestion has ever materialized.
> Pardon me, but by now I'm very sceptical about the odds of that
> ever changing.  And no, I don't have any suggestions either.

Right, I'll reword this to reflect that there's no expectation that this 
will be done, while keeping documented this performance issue.

> > The lookup() path at tmpfs creates negatives dentries, that are later
> > instantiated if the file is created. In that way, all files in tmpfs
> > have a dentry given that the filesystem exists exclusively in memory.
> > As explained above, we don't have negative dentries for casefold files,
> > so dentries are created at lookup() iff files aren't casefolded. Else,
> > the dentry is created just before being instantiated at create path.
> > At the remove path, dentries are invalidated for casefolded files.
>
> Umm...  What happens to those assertions if previously sane directory
> gets case-buggered?  You've got an ioctl for doing just that...
> Incidentally, that ioctl is obviously racy - result of that simple_empty()
> might have nothing to do with reality before it is returned to caller.
> And while we are at it, simple_empty() doesn't check a damn thing about
> negative dentries in there...

Thanks for pointing those issues. I'll move my lock at IOCTL to make 
impossible to change directory attributes and add a file there at the 
same time. About the negative dentries that existed before at that 
directory, I believe the way to solve this is by invalidating them all. 
How that sound to you?

Thanks,
	André