Re: [PATCH v6 7/8] fs/proc/task_mmu: read proc/pid/maps under per-vma lock

[Suren Baghdasaryan <surenb@google.com>]

On Tue, Jul 15, 2025 at 10:29 AM Andrii Nakryiko
<andrii.nakryiko@gmail.com> wrote:
>
> On Tue, Jul 15, 2025 at 10:21 AM Lorenzo Stoakes
> <lorenzo.stoakes@oracle.com> wrote:
> >
> > On Tue, Jul 15, 2025 at 06:10:16PM +0100, Lorenzo Stoakes wrote:
> > > > For PROCMAP_QUERY, we need priv->mm, but the newly added locked_vma
> > > > and locked_vma don't need to be persisted between ioctl calls. So we
> > > > can just add those two fields into a small struct, and for seq_file
> > > > case have it in priv, but for PROCMAP_QUERY just have it on the stack.
> > > > The code can be written to accept this struct to maintain the state,
> > > > which for PROCMAP_QUERY ioctl will be very short-lived on the stack
> > > > one.
> > > >
> > > > Would that work?
> > >
> > > Yeah that's a great idea actually, the stack would obviously give us the
> > > per-query invocation thing. Nice!
> > >
> > > I am kicking myself because I jokingly suggested (off-list) that a helper
> > > struct would be the answer to everything (I do love them) and of
> > > course... here we are :P
> >
> > Hm but actually we'd have to invert things I think, what I mean is - since
> > these fields can be updated at any time by racing threads, we can't have
> > _anything_ in the priv struct that is mutable.
> >
>
> Exactly, and I guess I was just being incomplete with just listing two
> of the fields that Suren make use of in PROCMAP_QUERY. See below.
>
> > So instead we should do something like:
> >
> > struct proc_maps_state {
> >         const struct proc_maps_private *priv;
> >         bool mmap_locked;
> >         struct vm_area_struct *locked_vma;
> >         struct vma_iterator iter;
> >         loff_t last_pos;
> > };
> >
> > static long procfs_procmap_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
> > {
> >         struct seq_file *seq = file->private_data;
> >         struct proc_maps_private *priv = seq->private;
> >         struct proc_maps_state state = {
> >                 .priv = priv,
> >         };
> >
> >         switch (cmd) {
> >         case PROCMAP_QUERY:
> >                 return do_procmap_query(state, (void __user *)arg);
>
> I guess it's a matter of preference, but I'd actually just pass
> seq->priv->mm and arg into do_procmap_query(), which will make it
> super obvious that priv is not used or mutated, and all the new stuff
> that Suren needs for lockless VMA iteration, including iter (not sure
> PROCMAP_QUERY needs last_pos, tbh), I'd just put into this new struct,
> which do_procmap_query() can keep private to itself.
>
> Ultimately, I think we are on the same page, it's just a matter of
> structuring code and types.

That sounds cleaner to me too.
I'll post a new version of my patchset today without the last patch to
keep PROCMAP_QUERY changes separate, and then a follow-up patch that
does this refactoring and changes PROCMAP_QUERY to use per-vma locks.

Thanks folks! It's good to be back from vacation with the problem
already figured out for you :)

>
> >         default:
> >                 return -ENOIOCTLCMD;
> >         }
> > }
> >
> > And then we have a stack-based thing with the bits that change and a
> > read-only pointer to the bits that must remain static. And we can enforce
> > that with const...
> >
> > We'd have to move the VMI and last_pos out too to make it const.
> >
> > Anyway the general idea should work!