Re: [PATCH RFC v2 2/2] pidfs: use maple tree

[Christian Brauner <brauner@kernel.org>]

On Fri, Dec 13, 2024 at 11:35:48AM +0100, Marek Szyprowski wrote:
> On 09.12.2024 14:46, Christian Brauner wrote:
> > So far we've been using an idr to track pidfs inodes. For some time now
> > each struct pid has a unique 64bit value that is used as the inode
> > number on 64 bit. That unique inode couldn't be used for looking up a
> > specific struct pid though.
> >
> > Now that we support file handles we need this ability while avoiding to
> > leak actual pid identifiers into userspace which can be problematic in
> > containers.
> >
> > So far I had used an idr-based mechanism where the idr is used to
> > generate a 32 bit number and each time it wraps we increment an upper
> > bit value and generate a unique 64 bit value. The lower 32 bits are used
> > to lookup the pid.
> >
> > I've been looking at the maple tree because it now has
> > mas_alloc_cyclic(). Since it uses unsigned long it would simplify the
> > 64bit implementation and its dense node mode supposedly also helps to
> > mitigate fragmentation.
> >
> > Signed-off-by: Christian Brauner <brauner@kernel.org>
> 
> This patch landed in today's linux-next as commit a2c8e88a30f7 ("pidfs: 
> use maple tree"). In my tests I found that it triggers the following 
> lockdep warning, what probably means that something has not been 
> properly initialized:

Ah, no, I think the issue that it didn't use irq{save,restore} spin lock
variants in that codepath as this is free_pid() which needs it.

I pushed a fix. Please yell if this issue persists.