* [PATCH] drivers/virt: pkvm: Zero and clean memory before sharing it with the host
@ 2026-11-16 11:05 Fuad Tabba
2026-11-16 11:13 ` Suzuki K Poulose
2026-11-16 11:28 ` Vincent Donnefort
0 siblings, 2 replies; 8+ messages in thread
From: Fuad Tabba @ 2026-11-16 11:05 UTC (permalink / raw)
To: Catalin Marinas, Will Deacon, Aneesh Kumar K.V (Arm)
Cc: Mark Rutland, Marc Zyngier, Mostafa Saleh, Vincent Donnefort,
Suzuki K Poulose, linux-arm-kernel, kvmarm, linux-coco,
linux-kernel, Fuad Tabba, stable
From: "Aneesh Kumar K.V (Arm)" <aneesh.kumar@kernel.org>
On pKVM, set_memory_decrypted() shares pages with the host in place and
without cache maintenance. The atomic pool, dma-direct and dynamic
SWIOTLB pools zero their memory only after sharing it, so unless the
page allocator has zeroed it, the host can read whatever the guest last
left in it. Zeroing first isn't enough on its own, as the zeroes may not
have been written back to memory, and a host that maps the pages
Non-cacheable then reads the old contents.
Zero the range and clean it to the PoC before sharing it, so that no
caller can hand the host stale data.
Fixes: ebc59b120c588 ("drivers/virt: pkvm: Hook up mem_encrypt API using pKVM hypercalls")
Fixes: e62decaf98e7c ("arm64/coco: Add pKVM as a CC platform")
Link: https://lore.kernel.org/r/20261116091137.94252-3-aneesh.kumar@kernel.org
Cc: stable@vger.kernel.org
Signed-off-by: Aneesh Kumar K.V (Arm) <aneesh.kumar@kernel.org>
[fuad: take the pKVM hunk of v10 patch 2 as a standalone fix, add the clean to the PoC]
Co-developed-by: Fuad Tabba <fuad.tabba@linux.dev>
Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
---
Notes:
This is the pKVM hunk of patch 2 of Aneesh's v10 [1] plus a clean to
the PoC, taken on its own because it fixes a bug that has exposed
guest memory to the host since v6.12.
Based on Linux 7.4-rc3 (aebb727447640).
[1] https://lore.kernel.org/r/20261116091137.94252-3-aneesh.kumar@kernel.org
drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c | 11 +++++++++++
1 file changed, 11 insertions(+)
diff --git a/drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c b/drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c
index 26fe9c3f22e36..81d1cd24a2eff 100644
--- a/drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c
+++ b/drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c
@@ -9,10 +9,12 @@
#include <linux/arm-smccc.h>
#include <linux/array_size.h>
+#include <linux/cacheflush.h>
#include <linux/io.h>
#include <linux/mem_encrypt.h>
#include <linux/mm.h>
#include <linux/pgtable.h>
+#include <linux/string.h>
#include <asm/hypervisor.h>
@@ -61,6 +63,15 @@ static int pkvm_set_memory_encrypted(unsigned long addr, int numpages)
static int pkvm_set_memory_decrypted(unsigned long addr, int numpages)
{
+ size_t size = numpages * PAGE_SIZE;
+
+ /*
+ * The host can read the pages as soon as they are shared, possibly
+ * through a Non-cacheable mapping. No caller relies on their contents.
+ */
+ memset((void *)addr, 0, size);
+ dcache_clean_poc(addr, addr + size);
+
return __set_memory_range(ARM_SMCCC_VENDOR_HYP_KVM_MEM_SHARE_FUNC_ID,
addr, numpages);
}
base-commit: aebb7274476403040d582c61478df8f4a7bf9cd3
--
2.47.3
^ permalink raw reply related [flat|nested] 8+ messages in thread* Re: [PATCH] drivers/virt: pkvm: Zero and clean memory before sharing it with the host
2026-11-16 11:05 [PATCH] drivers/virt: pkvm: Zero and clean memory before sharing it with the host Fuad Tabba
@ 2026-11-16 11:13 ` Suzuki K Poulose
2026-11-16 11:28 ` Vincent Donnefort
1 sibling, 0 replies; 8+ messages in thread
From: Suzuki K Poulose @ 2026-11-16 11:13 UTC (permalink / raw)
To: Fuad Tabba, Catalin Marinas, Will Deacon, Aneesh Kumar K.V (Arm)
Cc: Mark Rutland, Marc Zyngier, Mostafa Saleh, Vincent Donnefort,
linux-arm-kernel, kvmarm, linux-coco, linux-kernel, Fuad Tabba,
stable
On 16/11/2026 11:05, Fuad Tabba wrote:
> From: "Aneesh Kumar K.V (Arm)" <aneesh.kumar@kernel.org>
>
> On pKVM, set_memory_decrypted() shares pages with the host in place and
> without cache maintenance. The atomic pool, dma-direct and dynamic
> SWIOTLB pools zero their memory only after sharing it, so unless the
> page allocator has zeroed it, the host can read whatever the guest last
> left in it. Zeroing first isn't enough on its own, as the zeroes may not
> have been written back to memory, and a host that maps the pages
> Non-cacheable then reads the old contents.
>
> Zero the range and clean it to the PoC before sharing it, so that no
> caller can hand the host stale data.
>
> Fixes: ebc59b120c588 ("drivers/virt: pkvm: Hook up mem_encrypt API using pKVM hypercalls")
> Fixes: e62decaf98e7c ("arm64/coco: Add pKVM as a CC platform")
> Link: https://lore.kernel.org/r/20261116091137.94252-3-aneesh.kumar@kernel.org
> Cc: stable@vger.kernel.org
> Signed-off-by: Aneesh Kumar K.V (Arm) <aneesh.kumar@kernel.org>
> [fuad: take the pKVM hunk of v10 patch 2 as a standalone fix, add the clean to the PoC]
> Co-developed-by: Fuad Tabba <fuad.tabba@linux.dev>
> Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
Reviewed-by: Suzuki K Poulose <suzuki.poulose@arm.com>
> ---
>
> Notes:
> This is the pKVM hunk of patch 2 of Aneesh's v10 [1] plus a clean to
> the PoC, taken on its own because it fixes a bug that has exposed
> guest memory to the host since v6.12.
>
> Based on Linux 7.4-rc3 (aebb727447640).
>
> [1] https://lore.kernel.org/r/20261116091137.94252-3-aneesh.kumar@kernel.org
>
> drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c | 11 +++++++++++
> 1 file changed, 11 insertions(+)
>
> diff --git a/drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c b/drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c
> index 26fe9c3f22e36..81d1cd24a2eff 100644
> --- a/drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c
> +++ b/drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c
> @@ -9,10 +9,12 @@
>
> #include <linux/arm-smccc.h>
> #include <linux/array_size.h>
> +#include <linux/cacheflush.h>
> #include <linux/io.h>
> #include <linux/mem_encrypt.h>
> #include <linux/mm.h>
> #include <linux/pgtable.h>
> +#include <linux/string.h>
>
> #include <asm/hypervisor.h>
>
> @@ -61,6 +63,15 @@ static int pkvm_set_memory_encrypted(unsigned long addr, int numpages)
>
> static int pkvm_set_memory_decrypted(unsigned long addr, int numpages)
> {
> + size_t size = numpages * PAGE_SIZE;
> +
> + /*
> + * The host can read the pages as soon as they are shared, possibly
> + * through a Non-cacheable mapping. No caller relies on their contents.
> + */
> + memset((void *)addr, 0, size);
> + dcache_clean_poc(addr, addr + size);
> +
> return __set_memory_range(ARM_SMCCC_VENDOR_HYP_KVM_MEM_SHARE_FUNC_ID,
> addr, numpages);
> }
>
> base-commit: aebb7274476403040d582c61478df8f4a7bf9cd3
^ permalink raw reply [flat|nested] 8+ messages in thread* Re: [PATCH] drivers/virt: pkvm: Zero and clean memory before sharing it with the host
2026-11-16 11:05 [PATCH] drivers/virt: pkvm: Zero and clean memory before sharing it with the host Fuad Tabba
2026-11-16 11:13 ` Suzuki K Poulose
@ 2026-11-16 11:28 ` Vincent Donnefort
2026-11-16 12:25 ` Fuad Tabba
1 sibling, 1 reply; 8+ messages in thread
From: Vincent Donnefort @ 2026-11-16 11:28 UTC (permalink / raw)
To: Fuad Tabba
Cc: Catalin Marinas, Will Deacon, Aneesh Kumar K.V (Arm),
Mark Rutland, Marc Zyngier, Mostafa Saleh, Suzuki K Poulose,
linux-arm-kernel, kvmarm, linux-coco, linux-kernel, Fuad Tabba,
stable
On Mon, Nov 16, 2026 at 11:05:49AM +0000, Fuad Tabba wrote:
> From: "Aneesh Kumar K.V (Arm)" <aneesh.kumar@kernel.org>
>
> On pKVM, set_memory_decrypted() shares pages with the host in place and
> without cache maintenance. The atomic pool, dma-direct and dynamic
> SWIOTLB pools zero their memory only after sharing it, so unless the
> page allocator has zeroed it, the host can read whatever the guest last
> left in it. Zeroing first isn't enough on its own, as the zeroes may not
> have been written back to memory, and a host that maps the pages
> Non-cacheable then reads the old contents.
>
> Zero the range and clean it to the PoC before sharing it, so that no
> caller can hand the host stale data.
>
> Fixes: ebc59b120c588 ("drivers/virt: pkvm: Hook up mem_encrypt API using pKVM hypercalls")
> Fixes: e62decaf98e7c ("arm64/coco: Add pKVM as a CC platform")
> Link: https://lore.kernel.org/r/20261116091137.94252-3-aneesh.kumar@kernel.org
> Cc: stable@vger.kernel.org
> Signed-off-by: Aneesh Kumar K.V (Arm) <aneesh.kumar@kernel.org>
> [fuad: take the pKVM hunk of v10 patch 2 as a standalone fix, add the clean to the PoC]
> Co-developed-by: Fuad Tabba <fuad.tabba@linux.dev>
> Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
> ---
>
> Notes:
> This is the pKVM hunk of patch 2 of Aneesh's v10 [1] plus a clean to
> the PoC, taken on its own because it fixes a bug that has exposed
> guest memory to the host since v6.12.
>
> Based on Linux 7.4-rc3 (aebb727447640).
>
> [1] https://lore.kernel.org/r/20261116091137.94252-3-aneesh.kumar@kernel.org
>
> drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c | 11 +++++++++++
> 1 file changed, 11 insertions(+)
>
> diff --git a/drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c b/drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c
> index 26fe9c3f22e36..81d1cd24a2eff 100644
> --- a/drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c
> +++ b/drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c
> @@ -9,10 +9,12 @@
>
> #include <linux/arm-smccc.h>
> #include <linux/array_size.h>
> +#include <linux/cacheflush.h>
> #include <linux/io.h>
> #include <linux/mem_encrypt.h>
> #include <linux/mm.h>
> #include <linux/pgtable.h>
> +#include <linux/string.h>
>
> #include <asm/hypervisor.h>
>
> @@ -61,6 +63,15 @@ static int pkvm_set_memory_encrypted(unsigned long addr, int numpages)
>
> static int pkvm_set_memory_decrypted(unsigned long addr, int numpages)
> {
> + size_t size = numpages * PAGE_SIZE;
> +
> + /*
> + * The host can read the pages as soon as they are shared, possibly
> + * through a Non-cacheable mapping. No caller relies on their contents.
> + */
> + memset((void *)addr, 0, size);
Are we sure we want that? The guest may want to setup memory and then share it.
I am not sure about modifying this API.
> + dcache_clean_poc(addr, addr + size);
I believe the hypervisor is already doing the cache maintainance when the page
is mapped back into the host. (see stage2_map_walker_try_leaf())
--
Vincent
> +
> return __set_memory_range(ARM_SMCCC_VENDOR_HYP_KVM_MEM_SHARE_FUNC_ID,
> addr, numpages);
> }
>
> base-commit: aebb7274476403040d582c61478df8f4a7bf9cd3
> --
> 2.47.3
>
^ permalink raw reply [flat|nested] 8+ messages in thread* Re: [PATCH] drivers/virt: pkvm: Zero and clean memory before sharing it with the host
2026-11-16 11:28 ` Vincent Donnefort
@ 2026-11-16 12:25 ` Fuad Tabba
2026-11-16 14:04 ` Vincent Donnefort
0 siblings, 1 reply; 8+ messages in thread
From: Fuad Tabba @ 2026-11-16 12:25 UTC (permalink / raw)
To: Vincent Donnefort
Cc: Catalin Marinas, Will Deacon, Aneesh Kumar K.V (Arm),
Mark Rutland, Marc Zyngier, Mostafa Saleh, Suzuki K Poulose,
linux-arm-kernel, kvmarm, linux-coco, linux-kernel, stable
Hi Vincent,
On Mon, 16 Nov 2026 at 11:28, Vincent Donnefort <vdonnefort@google.com> wrote:
[...]
> > + /*
> > + * The host can read the pages as soon as they are shared, possibly
> > + * through a Non-cacheable mapping. No caller relies on their contents.
> > + */
> > + memset((void *)addr, 0, size);
>
> Are we sure we want that? The guest may want to setup memory and then share it.
> I am not sure about modifying this API.
Right! I wanted to fix the bug for the ITS property table, which is
shared before it's initialized, and totally forgot that this is the
path all sharing goes through... I'll drop this, and put that fix back
in the ITS series.
> > + dcache_clean_poc(addr, addr + size);
>
> I believe the hypervisor is already doing the cache maintainance when the page
> is mapped back into the host. (see stage2_map_walker_try_leaf())
I don't think it does on this path. host_mmu.mm_ops doesn't have a
dcache_clean_inval_poc() callback, so stage2_map_walker_try_leaf()
doesn't do the CMO for a host mapping. On the guest's side,
__pkvm_guest_share_host() only changes the software bits of the
existing mapping, and the walker returns before the CMO in that case.
Cheers,
/fuad
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [PATCH] drivers/virt: pkvm: Zero and clean memory before sharing it with the host
2026-11-16 12:25 ` Fuad Tabba
@ 2026-11-16 14:04 ` Vincent Donnefort
2026-11-16 16:32 ` Fuad Tabba
0 siblings, 1 reply; 8+ messages in thread
From: Vincent Donnefort @ 2026-11-16 14:04 UTC (permalink / raw)
To: Fuad Tabba
Cc: Catalin Marinas, Will Deacon, Aneesh Kumar K.V (Arm),
Mark Rutland, Marc Zyngier, Mostafa Saleh, Suzuki K Poulose,
linux-arm-kernel, kvmarm, linux-coco, linux-kernel, stable
On Mon, Nov 16, 2026 at 12:25:43PM +0000, Fuad Tabba wrote:
> Hi Vincent,
>
> On Mon, 16 Nov 2026 at 11:28, Vincent Donnefort <vdonnefort@google.com> wrote:
> [...]
> > > + /*
> > > + * The host can read the pages as soon as they are shared, possibly
> > > + * through a Non-cacheable mapping. No caller relies on their contents.
> > > + */
> > > + memset((void *)addr, 0, size);
> >
> > Are we sure we want that? The guest may want to setup memory and then share it.
> > I am not sure about modifying this API.
>
> Right! I wanted to fix the bug for the ITS property table, which is
> shared before it's initialized, and totally forgot that this is the
> path all sharing goes through... I'll drop this, and put that fix back
> in the ITS series.
>
> > > + dcache_clean_poc(addr, addr + size);
> >
> > I believe the hypervisor is already doing the cache maintainance when the page
> > is mapped back into the host. (see stage2_map_walker_try_leaf())
>
> I don't think it does on this path. host_mmu.mm_ops doesn't have a
> dcache_clean_inval_poc() callback, so stage2_map_walker_try_leaf()
> doesn't do the CMO for a host mapping. On the guest's side,
> __pkvm_guest_share_host() only changes the software bits of the
> existing mapping, and the walker returns before the CMO in that case.
>
> Cheers,
> /fuad
Ha yes of course, that's the host s2 we are mapping into. And on the guest side,
we are only changing the software bits, so no CMO. Actually I should have read
the comment in stage2_map_walker_try_leaf() more carefully.
That said, I wonder if that shouldn't be handled by the hypervisor, just like
hyp_poison_page() do the cache clean.
That would be slightly slower than letting the guest (who has a mapping already)
do it, but pKVM is enforcing isolation between host and guest, and it feels like
it should do it by default no matter the memory transition.
--
Vincent
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [PATCH] drivers/virt: pkvm: Zero and clean memory before sharing it with the host
2026-11-16 14:04 ` Vincent Donnefort
@ 2026-11-16 16:32 ` Fuad Tabba
2026-11-16 16:42 ` Vincent Donnefort
0 siblings, 1 reply; 8+ messages in thread
From: Fuad Tabba @ 2026-11-16 16:32 UTC (permalink / raw)
To: Vincent Donnefort
Cc: Catalin Marinas, Will Deacon, Aneesh Kumar K.V (Arm),
Mark Rutland, Marc Zyngier, Mostafa Saleh, Suzuki K Poulose,
linux-arm-kernel, kvmarm, linux-coco, linux-kernel, stable
On Mon, 16 Nov 2026 at 14:04, Vincent Donnefort <vdonnefort@google.com> wrote:
[...]
> Ha yes of course, that's the host s2 we are mapping into. And on the guest side,
> we are only changing the software bits, so no CMO. Actually I should have read
> the comment in stage2_map_walker_try_leaf() more carefully.
>
> That said, I wonder if that shouldn't be handled by the hypervisor, just like
> hyp_poison_page() do the cache clean.
>
> That would be slightly slower than letting the guest (who has a mapping already)
> do it, but pKVM is enforcing isolation between host and guest, and it feels like
> it should do it by default no matter the memory transition.
After the discussion on the related thread [1], if we have the zeroing
here, we need the clean here as well. So I think that this patch is
the right fix now.
Cheers,
/fuad
[1] https://lore.kernel.org/all/CA+EHjTywWTaKLtM6OPF_0ETv29-m8JsHdvF3BeW9Xa=h713eEg@mail.gmail.com/
>
> --
> Vincent
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [PATCH] drivers/virt: pkvm: Zero and clean memory before sharing it with the host
2026-11-16 16:32 ` Fuad Tabba
@ 2026-11-16 16:42 ` Vincent Donnefort
2026-11-16 16:46 ` Fuad Tabba
0 siblings, 1 reply; 8+ messages in thread
From: Vincent Donnefort @ 2026-11-16 16:42 UTC (permalink / raw)
To: Fuad Tabba
Cc: Catalin Marinas, Will Deacon, Aneesh Kumar K.V (Arm),
Mark Rutland, Marc Zyngier, Mostafa Saleh, Suzuki K Poulose,
linux-arm-kernel, kvmarm, linux-coco, linux-kernel, stable
On Mon, Nov 16, 2026 at 04:32:40PM +0000, Fuad Tabba wrote:
> On Mon, 16 Nov 2026 at 14:04, Vincent Donnefort <vdonnefort@google.com> wrote:
> [...]
> > Ha yes of course, that's the host s2 we are mapping into. And on the guest side,
> > we are only changing the software bits, so no CMO. Actually I should have read
> > the comment in stage2_map_walker_try_leaf() more carefully.
> >
> > That said, I wonder if that shouldn't be handled by the hypervisor, just like
> > hyp_poison_page() do the cache clean.
> >
> > That would be slightly slower than letting the guest (who has a mapping already)
> > do it, but pKVM is enforcing isolation between host and guest, and it feels like
> > it should do it by default no matter the memory transition.
>
> After the discussion on the related thread [1], if we have the zeroing
> here, we need the clean here as well. So I think that this patch is
> the right fix now.
>
> Cheers,
> /fuad
>
>
> [1] https://lore.kernel.org/all/CA+EHjTywWTaKLtM6OPF_0ETv29-m8JsHdvF3BeW9Xa=h713eEg@mail.gmail.com/
>
>
> >
> > --
> > Vincent
While I don't disagree with having the memset in CMOs in the guest, I still
believe the hypervisor should do it as well.
Linux is not the only virtualised guest and they all should be able to trust the
hypervisor to do the right thing for them, which it doesn't.
--
Vincent
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [PATCH] drivers/virt: pkvm: Zero and clean memory before sharing it with the host
2026-11-16 16:42 ` Vincent Donnefort
@ 2026-11-16 16:46 ` Fuad Tabba
0 siblings, 0 replies; 8+ messages in thread
From: Fuad Tabba @ 2026-11-16 16:46 UTC (permalink / raw)
To: Vincent Donnefort
Cc: Catalin Marinas, Will Deacon, Aneesh Kumar K.V (Arm),
Mark Rutland, Marc Zyngier, Mostafa Saleh, Suzuki K Poulose,
linux-arm-kernel, kvmarm, linux-coco, linux-kernel, stable
On Mon, 16 Nov 2026 at 16:42, Vincent Donnefort <vdonnefort@google.com> wrote:
[...]
> While I don't disagree with having the memset in CMOs in the guest, I still
> believe the hypervisor should do it as well.
>
> Linux is not the only virtualised guest and they all should be able to trust the
> hypervisor to do the right thing for them, which it doesn't.
I agree, and I think that that's the consensus [1].
This fix is for the guest part.
Cheers,
/fuad
[1] https://lore.kernel.org/all/20260923130031.GE1540250@ziepe.ca/
^ permalink raw reply [flat|nested] 8+ messages in thread
end of thread, other threads:[~2026-11-16 16:47 UTC | newest]
Thread overview: 8+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-16 11:05 [PATCH] drivers/virt: pkvm: Zero and clean memory before sharing it with the host Fuad Tabba
2026-11-16 11:13 ` Suzuki K Poulose
2026-11-16 11:28 ` Vincent Donnefort
2026-11-16 12:25 ` Fuad Tabba
2026-11-16 14:04 ` Vincent Donnefort
2026-11-16 16:32 ` Fuad Tabba
2026-11-16 16:42 ` Vincent Donnefort
2026-11-16 16:46 ` Fuad Tabba
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox