Re: [PATCH 7/7] KVM: arm/arm64: Elide CMOs when unmapping a range

From: James Morse <james.morse@arm.com>
To: Marc Zyngier <maz@kernel.org>
Cc: "Julien Thierry" <julien.thierry.kdev@gmail.com>,
	"Suzuki K Poulose" <suzuki.poulose@arm.com>,
	"James Hogan" <jhogan@kernel.org>,
	"Paul Mackerras" <paulus@ozlabs.org>,
	"Paolo Bonzini" <pbonzini@redhat.com>,
	"Radim Krčmář" <rkrcmar@redhat.com>,
	"Sean Christopherson" <sean.j.christopherson@intel.com>,
	"Vitaly Kuznetsov" <vkuznets@redhat.com>,
	"Wanpeng Li" <wanpengli@tencent.com>,
	"Jim Mattson" <jmattson@google.com>,
	"Joerg Roedel" <joro@8bytes.org>,
	linux-arm-kernel@lists.infradead.org,
	kvmarm@lists.cs.columbia.edu, linux-mips@vger.kernel.org,
	kvm-ppc@vger.kernel.org, kvm@vger.kernel.org
Subject: Re: [PATCH 7/7] KVM: arm/arm64: Elide CMOs when unmapping a range
Date: Thu, 19 Dec 2019 13:46:19 +0000	[thread overview]
Message-ID: <01c9439f-0de1-78ca-632b-f662876cc4a1@arm.com> (raw)
In-Reply-To: <de462fe29fb40fb1644e6a071e6c0c69@www.loen.fr>

Hi Marc,

On 18/12/2019 15:30, Marc Zyngier wrote:
> On 2019-12-18 15:07, James Morse wrote:
>> On 13/12/2019 18:25, Marc Zyngier wrote:
>>> If userspace issues a munmap() on a set of pages, there is no
>>> expectation that the pages are cleaned to the PoC.

>>> So let's
>>> not do more work than strictly necessary, and set the magic
>>> flag that avoids CMOs in this case.
>>
>> I think this assumes the pages went from anonymous->free, so no-one
>> cares about the contents.
>>
>> If the pages are backed by a file, won't dirty pages will still get
>> written back before the page is free? (e.g. EFI flash 'file' mmap()ed in)
> 
> I believe so. Is that a problem?

If we skipped the dcache maintenance on unmap, when the the dirty page is later reclaimed
the clean+stale lines are written back to the file. File-backed dirty pages will stick
around in the page cache in the hope someone else needs them.

This would happen for a guest:device-mapping that is written to, but is actually backed by
a mmap()d file. I think the EFI flash emulation does exactly this.

>> What if this isn't the only mapping of the page? Can't it be swapped
>> out from another VMA? (tenuous example, poor man's memory mirroring?)
> 
> Swap-out wouldn't trigger this code path, as it would use a different
> MMU notifier event (MMU_NOTIFY_CLEAR vs MMU_NOTIFY_UNMAP), I believe.

This was a half thought-through special case of the above. The sequence would be:

VM-A and VM-B both share a mapping of $page.

1. VM-A writes to $page through a device mapping
2. The kernel unmaps $page from VM-A for swap. KVM does the maintenance

3. VM-B writes to $page through a device mapping
4. VM-B exits, KVM skips the maintenance, $page may have clean+stale lines

5. Swap finds no further mappings, and writes $page and its clean+stale lines to disk.

Two VMs with a shared mapping is the 'easy' example. I think you just need a second
mapping for this to happen: it means the page isn't really free after the VM has exited.

Thanks,

James