Re: [PATCH v2 4/5] mm: make vma cache SLAB_TYPESAFE_BY_RCU

[Suren Baghdasaryan <surenb@google.com>]

On Wed, Nov 13, 2024 at 7:29 AM Liam R. Howlett <Liam.Howlett@oracle.com> wrote:
>
> * Suren Baghdasaryan <surenb@google.com> [241113 10:25]:
> > On Wed, Nov 13, 2024 at 7:23 AM 'Liam R. Howlett' via kernel-team
> > <kernel-team@android.com> wrote:
> > >
> > > * Matthew Wilcox <willy@infradead.org> [241113 08:57]:
> > > > On Wed, Nov 13, 2024 at 07:38:02AM -0500, Liam R. Howlett wrote:
> > > > > > Hi, I was wondering if we actually need the detached flag. Couldn't
> > > > > > "detached" simply mean vma->vm_mm == NULL and we save 4 bytes? Do we ever
> > > > > > need a vma that's detached but still has a mm pointer? I'd hope the places
> > > > > > that set detached to false have the mm pointer around so it's not inconvenient.
> > > > >
> > > > > I think the gate vmas ruin this plan.
> > > >
> > > > But the gate VMAs aren't to be found in the VMA tree.  Used to be that
> > > > was because the VMA tree was the injective RB tree and so VMAs could
> > > > only be in one tree at a time.  We could change that now!
> > >
> > > \o/
> > >
> > > >
> > > > Anyway, we could use (void *)1 instead of NULL to indicate a "detached"
> > > > VMA if we need to distinguish between a detached VMA and a gate VMA.
> > >
> > > I was thinking a pointer to itself vma->vm_mm = vma, then a check for
> > > this, instead of null like we do today.
> >
> > The motivation for having a separate detached flag was that vma->vm_mm
> > is used when read/write locking the vma, so it has to stay valid even
> > when vma gets detached. Maybe we can be more cautious in
> > vma_start_read()/vma_start_write() about it but I don't recall if
> > those were the only places that was an issue.
>
> We have the mm form the callers though, so it could be passed in?

Let me try and see if something else blows up. When I was implementing
per-vma locks I thought about using vma->vm_mm to indicate detached
state but there were some issues that caused me reconsider.

>
> >
> > >
> > > Either way, we should make it a function so it's easier to reuse for
> > > whatever we need in the future, wdyt?
> > >
> > > To unsubscribe from this group and stop receiving emails from it, send an email to kernel-team+unsubscribe@android.com.
> > >