Re: [PATCH v2 00/19] mm: Support huge pfnmaps

[Jiaqi Yan <jiaqiyan@google.com>]

On Wed, Sep 4, 2024 at 10:00 AM Jason Gunthorpe <jgg@nvidia.com> wrote:
>
> On Wed, Sep 04, 2024 at 09:58:54AM -0700, Jiaqi Yan wrote:
> > On Wed, Sep 4, 2024 at 9:43 AM Jason Gunthorpe <jgg@nvidia.com> wrote:
> > >
> > > On Wed, Sep 04, 2024 at 09:38:22AM -0700, Jiaqi Yan wrote:
> > > > On Wed, Sep 4, 2024 at 8:52 AM Jason Gunthorpe <jgg@nvidia.com> wrote:
> > > > >
> > > > > On Thu, Aug 29, 2024 at 12:21:39PM -0700, Jiaqi Yan wrote:
> > > > >
> > > > > > I think we still want to attempt to SIGBUS userspace, regardless of
> > > > > > doing unmap_mapping_range or not.
> > > > >
> > > > > IMHO we need to eliminate this path if we actually want to keep things
> > > > > mapped.
> > > > >
> > > > > There is no way to generate the SIGBUS without poking a 4k hole in the
> > > > > 1G page, as only that 4k should get SIGBUS, every other byte of the 1G
> > > > > is clean.
> > > >
> > > > Ah, sorry I wasn't clear. The SIGBUS will be only for poisoned PFN;
> > > > clean PFNs under the same PUD/PMD for sure don't need any SIGBUS,
> > > > which is the whole purpose of not unmapping.
> > >
> > > You can't get a SIGBUS if the things are still mapped. This is why the
> > > SIGBUS flow requires poking a non-present hole around the poisoned
> > > memory.
> > >
> > > So keeping things mapped at 1G also means giving up on SIGBUS.
> >
> > SIGBUS during page fault is definitely impossible when memory is still
> > mapped, but the platform still MCE or SEA in case of poison
> > consumption, right? So I wanted to propose new code to SIGBUS (either
> > BUS_MCEERR_AR or BUS_OBJERR) as long as the platform notifies the
> > kernel in the synchronous poison consumption context, e.g. MCE on X86
> > and SEA on ARM64.
>
> So you want a SIGBUS that is delivered asynchronously instead of via
> the page fault handler? Something like that is sort of what I ment by
> "eliminate this path", though I didn't think keeping an async SIGBUS
> was an option?

Not really, I don't think an SIGBUS *async* to the poison consuming
thread is critical, at least not as useful as SIGBUS *sync* to the
poison consuming thread.

>
> Jason