Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Fri, 30 Jul 2021 15:07:09 +0100
From:   Will Deacon <will@kernel.org>
To:     Marc Zyngier <maz@kernel.org>
Cc:     linux-arm-kernel@lists.infradead.org, kvmarm@lists.cs.columbia.edu,
        kvm@vger.kernel.org, linux-kernel@vger.kernel.org,
        qperret@google.com, dbrazdil@google.com,
        Srivatsa Vaddagiri <vatsa@codeaurora.org>,
        Shanker R Donthineni <sdonthineni@nvidia.com>,
        James Morse <james.morse@arm.com>,
        Suzuki K Poulose <suzuki.poulose@arm.com>,
        Alexandru Elisei <alexandru.elisei@arm.com>,
        kernel-team@android.com
Subject: Re: [PATCH 12/16] mm/ioremap: Add arch-specific callbacks on
 ioremap/iounmap calls
Message-ID: <20210730140709.GE23756@willie-the-truck>
References: <20210715163159.1480168-1-maz@kernel.org>
 <20210715163159.1480168-13-maz@kernel.org>
 <20210727181203.GG19173@willie-the-truck>
 <87tuked7mm.wl-maz@kernel.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <87tuked7mm.wl-maz@kernel.org>
User-Agent: Mutt/1.10.1 (2018-07-13)
Precedence: bulk

On Wed, Jul 28, 2021 at 12:01:53PM +0100, Marc Zyngier wrote:
> On Tue, 27 Jul 2021 19:12:04 +0100,
> Will Deacon <will@kernel.org> wrote:
> > 
> > On Thu, Jul 15, 2021 at 05:31:55PM +0100, Marc Zyngier wrote:
> > > Add a pair of hooks (ioremap_page_range_hook/iounmap_page_range_hook)
> > > that can be implemented by an architecture.
> > > 
> > > Signed-off-by: Marc Zyngier <maz@kernel.org>
> > > ---
> > >  include/linux/io.h |  3 +++
> > >  mm/ioremap.c       | 13 ++++++++++++-
> > >  mm/vmalloc.c       |  8 ++++++++
> > >  3 files changed, 23 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/include/linux/io.h b/include/linux/io.h
> > > index 9595151d800d..0ffc265f114c 100644
> > > --- a/include/linux/io.h
> > > +++ b/include/linux/io.h
> > > @@ -21,6 +21,9 @@ void __ioread32_copy(void *to, const void __iomem *from, size_t count);
> > >  void __iowrite64_copy(void __iomem *to, const void *from, size_t count);
> > >  
> > >  #ifdef CONFIG_MMU
> > > +void ioremap_page_range_hook(unsigned long addr, unsigned long end,
> > > +			     phys_addr_t phys_addr, pgprot_t prot);
> > > +void iounmap_page_range_hook(phys_addr_t phys_addr, size_t size);
> > >  int ioremap_page_range(unsigned long addr, unsigned long end,
> > >  		       phys_addr_t phys_addr, pgprot_t prot);
> > >  #else
> > 
> > Can we avoid these hooks by instead not registering the regions proactively
> > in the guest and moving that logic to a fault handler which runs off the
> > back of the injected data abort? From there, we could check if the faulting
> > IPA is a memory address and register it as MMIO if not.
> > 
> > Dunno, you've spent more time than me thinking about this, but just
> > wondering if you'd had a crack at doing it that way, as it _seems_ simpler
> > to my naive brain.
> 
> I thought about it, but couldn't work out whether it was always
> possible for the guest to handle these faults (first access in an
> interrupt context, for example?).

If the check is a simple pfn_valid() I think it should be ok, but yes, we'd
definitely not want to do anything more involved given that this could run
in all sorts of horrible contexts.

> Also, this changes the semantics of the protection this is supposed to
> offer: any access out of the RAM space will generate an abort, and the
> fault handler will grant MMIO forwarding for this page. Stray accesses
> that would normally be properly handled as fatal would now succeed and
> be forwarded to userspace, even if there was no emulated devices
> there.

That's true, it would offer much weaker guarantees to the guest. It's more
like a guarantee that memory never traps to the VMM. It also then wouldn't
help with the write-combine fun. It would be simpler though, but with less
functionality.

> For this to work, we'd need to work out whether there is any existing
> device mapping that actually points to this page. And whether it
> actually is supposed to be forwarded to userspace. Do we have a rmap
> for device mappings?

I don't think this would be possible given your comments above.

So let's stick with the approach you've taken. It just feels like there
should be a way to do this without introducing new hooks into the core
code. If it wasn't for pci_remap_iospace(), we could simply hook our
definition of __ioremap_caller(). Another avenue to explore would be
looking at the IO resource instead; I see x86 already uses
IORES_MAP_ENCRYPTED and IORES_MAP_SYSTEM_RAM to drive pgprot...

Will