Arm's Memory Tagging Extension (MTE) adds 4 bits of tag data to every 16
bytes of memory in the system. This along with stashing a tag within the
high bit of virtual addresses allows runtime checking of memory
accesses.
These patches add support to KVM to enable MTE within a guest. They are
based on Catalin's v9 MTE user-space support series[1].
I'd welcome feedback on the proposed user-kernel ABI. Specifically this
series currently:
1. Requires the VMM to enable MTE per-VCPU.
2. Automatically promotes (normal host) memory given to the guest to be
tag enabled (sets PG_mte_tagged), if any VCPU has MTE enabled. The
tags are cleared if the memory wasn't previously MTE enabled.
3. Doesn't provide any new methods for the VMM to access the tags on
memory.
(2) and (3) are particularly interesting from the aspect of VM migration.
The guest is able to store/retrieve data in the tags (presumably for the
purpose of tag checking, but architecturally it could be used as just
storage). This means that when migrating a guest the data needs to be
transferred (or saved/restored).
MTE tags are controlled by the same permission model as normal pages
(i.e. a read-only page has read-only tags), so the normal methods of
detecting guest changes to pages can be used. But this would also
require the tags within a page to be migrated at the same time as the
data (since the access control for tags is the same as the normal data
within a page).
(3) may be problematic and I'd welcome input from those familiar with
VMMs. User space cannot access tags unless the memory is mapped with the
PROT_MTE flag. However enabling PROT_MTE will also enable tag checking
for the user space process (assuming the VMM enables tag checking for
the process) and since the tags in memory are controlled by the guest
it's unlikely the VMM would have an appropriately tagged pointer for its
access. This means the VMM would either need to maintain two mappings of
memory (one to access tags, the other to access data) or disable tag
checking during the accesses to data.
If it's not practical to either disable tag checking in the VMM or
maintain multiple mappings then the alternatives I'm aware of are:
* Provide a KVM-specific method to extract the tags from guest memory.
This might also have benefits in terms of providing an easy way to
read bulk tag data from guest memory (since the LDGM instruction
isn't available at EL0).
* Provide support for user space setting the TCMA0 or TCMA1 bits in
TCR_EL1. These would allow the VMM to generate pointers which are not
tag checked.
Feedback is welcome, and feel free to ask questions if anything in the
above doesn't make sense.
Changes since the previous v1 posting[2]:
* Rebasing clean-ups
* sysreg visibility is now controlled based on whether the VCPU has MTE
enabled or not
[1] https://lore.kernel.org/r/[email protected]
[2] https://lore.kernel.org/r/20200713100102.53664-1-steven.price%40arm.com
Steven Price (2):
arm64: kvm: Save/restore MTE registers
arm64: kvm: Introduce MTE VCPU feature
arch/arm64/include/asm/kvm_emulate.h | 3 +++
arch/arm64/include/asm/kvm_host.h | 9 ++++++++-
arch/arm64/include/asm/sysreg.h | 3 ++-
arch/arm64/include/uapi/asm/kvm.h | 1 +
arch/arm64/kvm/hyp/include/hyp/sysreg-sr.h | 14 ++++++++++++++
arch/arm64/kvm/mmu.c | 15 +++++++++++++++
arch/arm64/kvm/reset.c | 8 ++++++++
arch/arm64/kvm/sys_regs.c | 20 +++++++++++++++-----
8 files changed, 66 insertions(+), 7 deletions(-)
--
2.20.1
(cc'ing in Eric Auger)
* Steven Price ([email protected]) wrote:
> Arm's Memory Tagging Extension (MTE) adds 4 bits of tag data to every 16
> bytes of memory in the system. This along with stashing a tag within the
> high bit of virtual addresses allows runtime checking of memory
> accesses.
>
> These patches add support to KVM to enable MTE within a guest. They are
> based on Catalin's v9 MTE user-space support series[1].
>
> I'd welcome feedback on the proposed user-kernel ABI. Specifically this
> series currently:
>
> 1. Requires the VMM to enable MTE per-VCPU.
> 2. Automatically promotes (normal host) memory given to the guest to be
> tag enabled (sets PG_mte_tagged), if any VCPU has MTE enabled. The
> tags are cleared if the memory wasn't previously MTE enabled.
> 3. Doesn't provide any new methods for the VMM to access the tags on
> memory.
>
> (2) and (3) are particularly interesting from the aspect of VM migration.
> The guest is able to store/retrieve data in the tags (presumably for the
> purpose of tag checking, but architecturally it could be used as just
> storage). This means that when migrating a guest the data needs to be
> transferred (or saved/restored).
>
> MTE tags are controlled by the same permission model as normal pages
> (i.e. a read-only page has read-only tags), so the normal methods of
> detecting guest changes to pages can be used. But this would also
> require the tags within a page to be migrated at the same time as the
> data (since the access control for tags is the same as the normal data
> within a page).
(Without understanding anything about your tag system...)
Note that during (normal, non-postcopy) migration the consistency can
be a little loose - until the guest starts running; i.e. you can send
a page that's in themiddle of being modified as long as you make sure
you send it again later so that what the guest sees on the destination
when it runs is consistent; i.e. it would be fine to send your tags
separately to your data and allow them to get a little out of sync, as
long as they caught up before the guest ran.
> (3) may be problematic and I'd welcome input from those familiar with
> VMMs. User space cannot access tags unless the memory is mapped with the
> PROT_MTE flag. However enabling PROT_MTE will also enable tag checking
> for the user space process (assuming the VMM enables tag checking for
> the process) and since the tags in memory are controlled by the guest
> it's unlikely the VMM would have an appropriately tagged pointer for its
> access. This means the VMM would either need to maintain two mappings of
> memory (one to access tags, the other to access data) or disable tag
> checking during the accesses to data.
Imagine I had a second mapping; what would it look like; how would I get
and restore the tags?
In terms of migration stream, I guess we have two ways to do this,
either it rides shotgun on the main RAM section pages, transmitting
those few extra bytes whenever we transmit a page, or you have a
separate iteratable device for RAMtags, and it just transmits those.
How you keep the two together is an interesting question.
The shotgun method sounds nasty to avoid putting special cases in the,
already hairy, RAM code.
> If it's not practical to either disable tag checking in the VMM or
> maintain multiple mappings then the alternatives I'm aware of are:
>
> * Provide a KVM-specific method to extract the tags from guest memory.
> This might also have benefits in terms of providing an easy way to
> read bulk tag data from guest memory (since the LDGM instruction
> isn't available at EL0).
> * Provide support for user space setting the TCMA0 or TCMA1 bits in
> TCR_EL1. These would allow the VMM to generate pointers which are not
> tag checked.
I guess you want the VMM to do as much tagged checked access as possible
on it's own data structures?
How do things like virtio work where the qemu or kernel is accessing
guest memory for IO?
Dave
> Feedback is welcome, and feel free to ask questions if anything in the
> above doesn't make sense.
>
> Changes since the previous v1 posting[2]:
>
> * Rebasing clean-ups
> * sysreg visibility is now controlled based on whether the VCPU has MTE
> enabled or not
>
> [1] https://lore.kernel.org/r/[email protected]
> [2] https://lore.kernel.org/r/20200713100102.53664-1-steven.price%40arm.com
>
> Steven Price (2):
> arm64: kvm: Save/restore MTE registers
> arm64: kvm: Introduce MTE VCPU feature
>
> arch/arm64/include/asm/kvm_emulate.h | 3 +++
> arch/arm64/include/asm/kvm_host.h | 9 ++++++++-
> arch/arm64/include/asm/sysreg.h | 3 ++-
> arch/arm64/include/uapi/asm/kvm.h | 1 +
> arch/arm64/kvm/hyp/include/hyp/sysreg-sr.h | 14 ++++++++++++++
> arch/arm64/kvm/mmu.c | 15 +++++++++++++++
> arch/arm64/kvm/reset.c | 8 ++++++++
> arch/arm64/kvm/sys_regs.c | 20 +++++++++++++++-----
> 8 files changed, 66 insertions(+), 7 deletions(-)
>
> --
> 2.20.1
>
--
Dr. David Alan Gilbert / [email protected] / Manchester, UK
On 07/09/2020 16:28, Dr. David Alan Gilbert wrote:
> (cc'ing in Eric Auger)
>
> * Steven Price ([email protected]) wrote:
>> Arm's Memory Tagging Extension (MTE) adds 4 bits of tag data to every 16
>> bytes of memory in the system. This along with stashing a tag within the
>> high bit of virtual addresses allows runtime checking of memory
>> accesses.
>>
>> These patches add support to KVM to enable MTE within a guest. They are
>> based on Catalin's v9 MTE user-space support series[1].
>>
>> I'd welcome feedback on the proposed user-kernel ABI. Specifically this
>> series currently:
>>
>> 1. Requires the VMM to enable MTE per-VCPU.
>> 2. Automatically promotes (normal host) memory given to the guest to be
>> tag enabled (sets PG_mte_tagged), if any VCPU has MTE enabled. The
>> tags are cleared if the memory wasn't previously MTE enabled.
>> 3. Doesn't provide any new methods for the VMM to access the tags on
>> memory.
>>
>> (2) and (3) are particularly interesting from the aspect of VM migration.
>> The guest is able to store/retrieve data in the tags (presumably for the
>> purpose of tag checking, but architecturally it could be used as just
>> storage). This means that when migrating a guest the data needs to be
>> transferred (or saved/restored).
>>
>> MTE tags are controlled by the same permission model as normal pages
>> (i.e. a read-only page has read-only tags), so the normal methods of
>> detecting guest changes to pages can be used. But this would also
>> require the tags within a page to be migrated at the same time as the
>> data (since the access control for tags is the same as the normal data
>> within a page).
>
> (Without understanding anything about your tag system...)
>
> Note that during (normal, non-postcopy) migration the consistency can
> be a little loose - until the guest starts running; i.e. you can send
> a page that's in themiddle of being modified as long as you make sure
> you send it again later so that what the guest sees on the destination
> when it runs is consistent; i.e. it would be fine to send your tags
> separately to your data and allow them to get a little out of sync, as
> long as they caught up before the guest ran.
Yes, you can obviously pro-actively send data early as you as you
appropriately deal with any potential changes that the guest might make.
I'm not very familiar with exactly how QEMU handles this, so it might
not be a problem - I just wanted to point out that we don't have
separate access permissions.
>> (3) may be problematic and I'd welcome input from those familiar with
>> VMMs. User space cannot access tags unless the memory is mapped with the
>> PROT_MTE flag. However enabling PROT_MTE will also enable tag checking
>> for the user space process (assuming the VMM enables tag checking for
>> the process) and since the tags in memory are controlled by the guest
>> it's unlikely the VMM would have an appropriately tagged pointer for its
>> access. This means the VMM would either need to maintain two mappings of
>> memory (one to access tags, the other to access data) or disable tag
>> checking during the accesses to data.
>
> Imagine I had a second mapping; what would it look like; how would I get
> and restore the tags?
At a very simple level you could do something like:
normal_mapping = mmap(..., PROT_READ | PROT_WRITE, ..., fd, 0);
mte_mapping = mmap(..., PROT_READ | PROT_WRITE | PROT_MTE, ..., fd, 0);
/* access normal mapping as normal */
normal_mapping[offset] = 0xf00 + normal_mapping[offset + 1];
/* read tag from mte_mapping */
uint64_t tag = ldg(&mte_mapping[offset]);
/* write a new tag value (8)
* NOTE: tags are stored in the top byte, hence the shift
*/
stg(0x8ULL << 56, &mte_mapping[offset]);
Where stg() and ldg() are simple wrappers around the new instructions:
stg:
STG x0, [x1]
RET
ldg:
LDG x0, [x0]
RET
> In terms of migration stream, I guess we have two ways to do this,
> either it rides shotgun on the main RAM section pages, transmitting
> those few extra bytes whenever we transmit a page, or you have a
> separate iteratable device for RAMtags, and it just transmits those.
> How you keep the two together is an interesting question.
> The shotgun method sounds nasty to avoid putting special cases in the,
> already hairy, RAM code.
As you say above it may be possible to simply let the normal RAM and
tags get out of sync. E.g. if you send all the normal RAM (marking
read-only as you go), then all the tags (not changing the permissions)
you will end up with all the pages that have remained read-only (i.e.
the guest hasn't modified) being consistent on the destination. Pages
that have been written by the guest will be inconsistent, but you were
going to have to resend those anyway.
However for post-migration copy you need to copy *both* normal RAM and
tags before resuming the guest. You might need special cases for this.
>> If it's not practical to either disable tag checking in the VMM or
>> maintain multiple mappings then the alternatives I'm aware of are:
>>
>> * Provide a KVM-specific method to extract the tags from guest memory.
>> This might also have benefits in terms of providing an easy way to
>> read bulk tag data from guest memory (since the LDGM instruction
>> isn't available at EL0).
>> * Provide support for user space setting the TCMA0 or TCMA1 bits in
>> TCR_EL1. These would allow the VMM to generate pointers which are not
>> tag checked.
>
> I guess you want the VMM to do as much tagged checked access as possible
> on it's own data structures?
Ideally yes, you would want the VMM to have checked accesses for all
it's internal data structures because that gives the maximum benefit
from MTE.
> How do things like virtio work where the qemu or kernel is accessing
> guest memory for IO?
Since virtio is effectively emulating a device it should be treated like
a device - no tag checking and no tag storage used. This would be the
obvious situation where you would use "normal_mapping" as above so tags
wouldn't be visible or checked.
Really the VMM is only interested in guest tags for the migration case
where it simply needs to preserve them. I don't expect the guest and VMM
(or hypervisor) to communicate using tagged memory.
Steve
On 09/09/2020 16:25, Andrew Jones wrote:
> On Fri, Sep 04, 2020 at 05:00:16PM +0100, Steven Price wrote:
>> Arm's Memory Tagging Extension (MTE) adds 4 bits of tag data to every 16
>> bytes of memory in the system. This along with stashing a tag within the
>> high bit of virtual addresses allows runtime checking of memory
>> accesses.
>>
>> These patches add support to KVM to enable MTE within a guest. They are
>> based on Catalin's v9 MTE user-space support series[1].
>>
>> I'd welcome feedback on the proposed user-kernel ABI. Specifically this
>> series currently:
>>
> 0. Feature probing
>
> Probably a KVM cap, rather than requiring userspace to attempt VCPU
> features one at a time with a scratch VCPU.
Ah, good point - thanks for pointing that out.
>> 1. Requires the VMM to enable MTE per-VCPU.
>
> I suppose. We're collecting many features that are enabling CPU features,
> so they map nicely to VCPU features, yet they're effectively VM features
> due to a shared resource such as an irq or memory.
Yeah this is a little weird I'll admit. The architectural feature is
described per-CPU (well "processing element"), but it makes little sense
to have it only on some CPUs and has effects on the rest of the memory
system. Given that it's theoretically possible to build e.g. a
big.LITTLE setup with only some CPUs support MTE it seemed more
future-proof to design the API to allow it even though I hope no-one
will use it.
>> 2. Automatically promotes (normal host) memory given to the guest to be
>> tag enabled (sets PG_mte_tagged), if any VCPU has MTE enabled. The
>> tags are cleared if the memory wasn't previously MTE enabled.
>
> Shouldn't this be up to the guest? Or, is this required in order for the
> guest to use tagging at all. Something like making the guest IPAs memtag
> capable, but if the guest doesn't enable tagging then there is no guest
> impact? In any case, shouldn't userspace be the one that adds PROT_MTE
> to the memory regions it wants the guest to be able to use tagging with,
> rather than KVM adding the attribute page by page?
I think I've probably explained this badly.
The guest can choose how to populate the stage 1 mapping - so can choose
which parts of memory are accessed tagged or not. However, the
hypervisor cannot restrict this in stage 2 (except by e.g. making the
memory uncached but that's obviously not great - however devices forward
to the guest can be handled like this).
Because the hypervisor cannot restrict the guest's access to the tags,
the hypervisor must assume that all memory given to the guest could have
the tags accessed. So it must (a) clear any stale data from the tags,
and (b) ensure that the tags are preserved (e.g. when swapping pages out).
Because of the above the current series automatically sets PG_mte_tagged
on the pages. Note that this doesn't change the mappings that the VMM
has (a non-PROT_MTE mapping will still not have access to the tags).
It's a shame that the stage-2 can't usefully restrict tag access, but
this matches the architectural expectation: that if MTE is supported
then all standard memory will be MTE-enabled.
>> 3. Doesn't provide any new methods for the VMM to access the tags on
>> memory.
>>
>> (2) and (3) are particularly interesting from the aspect of VM migration.
>> The guest is able to store/retrieve data in the tags (presumably for the
>> purpose of tag checking, but architecturally it could be used as just
>> storage). This means that when migrating a guest the data needs to be
>> transferred (or saved/restored).
>>
>> MTE tags are controlled by the same permission model as normal pages
>> (i.e. a read-only page has read-only tags), so the normal methods of
>> detecting guest changes to pages can be used. But this would also
>> require the tags within a page to be migrated at the same time as the
>> data (since the access control for tags is the same as the normal data
>> within a page).
>>
>> (3) may be problematic and I'd welcome input from those familiar with
>> VMMs. User space cannot access tags unless the memory is mapped with the
>> PROT_MTE flag. However enabling PROT_MTE will also enable tag checking
>> for the user space process (assuming the VMM enables tag checking for
>> the process) and since the tags in memory are controlled by the guest
>> it's unlikely the VMM would have an appropriately tagged pointer for its
>> access. This means the VMM would either need to maintain two mappings of
>> memory (one to access tags, the other to access data) or disable tag
>> checking during the accesses to data.
>
> If userspace needs to write to guest memory then it should be due to
> a device DMA or other specific hardware emulation. Those accesses can
> be done with tag checking disabled.
Yes, the question is can the VMM (sensibly) wrap the accesses with a
disable/renable tag checking for the process sequence. The alternative
at the moment is to maintain a separate (untagged) mapping for the
purpose which might present it's own problems.
>>
>> If it's not practical to either disable tag checking in the VMM or
>> maintain multiple mappings then the alternatives I'm aware of are:
>>
>> * Provide a KVM-specific method to extract the tags from guest memory.
>> This might also have benefits in terms of providing an easy way to
>> read bulk tag data from guest memory (since the LDGM instruction
>> isn't available at EL0).
>
> Maybe we need a new version of KVM_GET_DIRTY_LOG that also provides
> the tags for all addresses of each dirty page.
Certainly possible, although it seems to conflate two operations: "get
list of dirty pages", "get tags from page". It would also require a lot
of return space (size of slot/32).
>> * Provide support for user space setting the TCMA0 or TCMA1 bits in
>> TCR_EL1. These would allow the VMM to generate pointers which are not
>> tag checked.
>
> So this is necessary to allow the VMM to keep tag checking enabled for
> itself, plus map guest memory as PROT_MTE, and write to that memory when
> needed?
This is certainly one option. The architecture provides two "magic"
values (all-0s and all-1s) which can be configured using TCMAx to be
treated differently. The VMM could therefore construct pointers to
otherwise tagged memory which would be treated as untagged.
However, Catalin's user space series doesn't at the moment expose this
functionality.
Steve
On Fri, Sep 04, 2020 at 05:00:16PM +0100, Steven Price wrote:
> Arm's Memory Tagging Extension (MTE) adds 4 bits of tag data to every 16
> bytes of memory in the system. This along with stashing a tag within the
> high bit of virtual addresses allows runtime checking of memory
> accesses.
>
> These patches add support to KVM to enable MTE within a guest. They are
> based on Catalin's v9 MTE user-space support series[1].
>
> I'd welcome feedback on the proposed user-kernel ABI. Specifically this
> series currently:
>
0. Feature probing
Probably a KVM cap, rather than requiring userspace to attempt VCPU
features one at a time with a scratch VCPU.
> 1. Requires the VMM to enable MTE per-VCPU.
I suppose. We're collecting many features that are enabling CPU features,
so they map nicely to VCPU features, yet they're effectively VM features
due to a shared resource such as an irq or memory.
> 2. Automatically promotes (normal host) memory given to the guest to be
> tag enabled (sets PG_mte_tagged), if any VCPU has MTE enabled. The
> tags are cleared if the memory wasn't previously MTE enabled.
Shouldn't this be up to the guest? Or, is this required in order for the
guest to use tagging at all. Something like making the guest IPAs memtag
capable, but if the guest doesn't enable tagging then there is no guest
impact? In any case, shouldn't userspace be the one that adds PROT_MTE
to the memory regions it wants the guest to be able to use tagging with,
rather than KVM adding the attribute page by page?
> 3. Doesn't provide any new methods for the VMM to access the tags on
> memory.
>
> (2) and (3) are particularly interesting from the aspect of VM migration.
> The guest is able to store/retrieve data in the tags (presumably for the
> purpose of tag checking, but architecturally it could be used as just
> storage). This means that when migrating a guest the data needs to be
> transferred (or saved/restored).
>
> MTE tags are controlled by the same permission model as normal pages
> (i.e. a read-only page has read-only tags), so the normal methods of
> detecting guest changes to pages can be used. But this would also
> require the tags within a page to be migrated at the same time as the
> data (since the access control for tags is the same as the normal data
> within a page).
>
> (3) may be problematic and I'd welcome input from those familiar with
> VMMs. User space cannot access tags unless the memory is mapped with the
> PROT_MTE flag. However enabling PROT_MTE will also enable tag checking
> for the user space process (assuming the VMM enables tag checking for
> the process) and since the tags in memory are controlled by the guest
> it's unlikely the VMM would have an appropriately tagged pointer for its
> access. This means the VMM would either need to maintain two mappings of
> memory (one to access tags, the other to access data) or disable tag
> checking during the accesses to data.
If userspace needs to write to guest memory then it should be due to
a device DMA or other specific hardware emulation. Those accesses can
be done with tag checking disabled.
>
> If it's not practical to either disable tag checking in the VMM or
> maintain multiple mappings then the alternatives I'm aware of are:
>
> * Provide a KVM-specific method to extract the tags from guest memory.
> This might also have benefits in terms of providing an easy way to
> read bulk tag data from guest memory (since the LDGM instruction
> isn't available at EL0).
Maybe we need a new version of KVM_GET_DIRTY_LOG that also provides
the tags for all addresses of each dirty page.
> * Provide support for user space setting the TCMA0 or TCMA1 bits in
> TCR_EL1. These would allow the VMM to generate pointers which are not
> tag checked.
So this is necessary to allow the VMM to keep tag checking enabled for
itself, plus map guest memory as PROT_MTE, and write to that memory when
needed?
Thanks,
drew
>
> Feedback is welcome, and feel free to ask questions if anything in the
> above doesn't make sense.
>
> Changes since the previous v1 posting[2]:
>
> * Rebasing clean-ups
> * sysreg visibility is now controlled based on whether the VCPU has MTE
> enabled or not
>
> [1] https://lore.kernel.org/r/[email protected]
> [2] https://lore.kernel.org/r/20200713100102.53664-1-steven.price%40arm.com
>
> Steven Price (2):
> arm64: kvm: Save/restore MTE registers
> arm64: kvm: Introduce MTE VCPU feature
>
> arch/arm64/include/asm/kvm_emulate.h | 3 +++
> arch/arm64/include/asm/kvm_host.h | 9 ++++++++-
> arch/arm64/include/asm/sysreg.h | 3 ++-
> arch/arm64/include/uapi/asm/kvm.h | 1 +
> arch/arm64/kvm/hyp/include/hyp/sysreg-sr.h | 14 ++++++++++++++
> arch/arm64/kvm/mmu.c | 15 +++++++++++++++
> arch/arm64/kvm/reset.c | 8 ++++++++
> arch/arm64/kvm/sys_regs.c | 20 +++++++++++++++-----
> 8 files changed, 66 insertions(+), 7 deletions(-)
>
> --
> 2.20.1
>
> _______________________________________________
> kvmarm mailing list
> [email protected]
> https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
>
On 9/9/20 8:25 AM, Andrew Jones wrote:
>> * Provide a KVM-specific method to extract the tags from guest memory.
>> This might also have benefits in terms of providing an easy way to
>> read bulk tag data from guest memory (since the LDGM instruction
>> isn't available at EL0).
>
> Maybe we need a new version of KVM_GET_DIRTY_LOG that also provides
> the tags for all addresses of each dirty page.
KVM_GET_DIRTY_LOG just provides one bit per dirty page, no? Then VMM copies
the data out from its local address to guest memory.
There'd be no difference with or without tags, afaik. It's just about how VMM
copies the data, with or without tags.
>> * Provide support for user space setting the TCMA0 or TCMA1 bits in
>> TCR_EL1. These would allow the VMM to generate pointers which are not
>> tag checked.
>
> So this is necessary to allow the VMM to keep tag checking enabled for
> itself, plus map guest memory as PROT_MTE, and write to that memory when
> needed?
I don't see a requirement for the VMM to set TCMA0.
r~
On 9/4/20 9:00 AM, Steven Price wrote:
> 3. Doesn't provide any new methods for the VMM to access the tags on
> memory.
...
> (3) may be problematic and I'd welcome input from those familiar with
> VMMs. User space cannot access tags unless the memory is mapped with the
> PROT_MTE flag. However enabling PROT_MTE will also enable tag checking
> for the user space process (assuming the VMM enables tag checking for
> the process)...
The latest version of the kernel patches for user mte support has separate
controls for how tag check fail is reported. Including
> +- ``PR_MTE_TCF_NONE`` - *Ignore* tag check faults
That may be less than optimal once userland starts uses tags itself, e.g.
running qemu itself with an mte-aware malloc.
Independent of that, there's also the TCO bit, which can be toggled by any
piece of code that wants to disable checking locally.
However, none of that is required for accessing tags. User space can always
load/store tags via LDG/STG. That's going to be slow, though.
It's a shame that LDGM/STGM are privileged instructions. I don't understand
why that was done, since there's absolutely nothing that those insns can do
that you can't do with (up to) 16x LDG/STG.
I think it might be worth adding some sort of kernel entry point that can bulk
copy tags, e.g. page aligned quantities. But that's just a speed of migration
thing and could come later.
r~
On Wed, Sep 09, 2020 at 06:45:33PM -0700, Richard Henderson wrote:
> On 9/9/20 8:25 AM, Andrew Jones wrote:
> >> * Provide a KVM-specific method to extract the tags from guest memory.
> >> This might also have benefits in terms of providing an easy way to
> >> read bulk tag data from guest memory (since the LDGM instruction
> >> isn't available at EL0).
> >
> > Maybe we need a new version of KVM_GET_DIRTY_LOG that also provides
> > the tags for all addresses of each dirty page.
>
> KVM_GET_DIRTY_LOG just provides one bit per dirty page, no? Then VMM copies
> the data out from its local address to guest memory.
>
> There'd be no difference with or without tags, afaik. It's just about how VMM
> copies the data, with or without tags.
Right, as long as it's fast enough to do
for_each_dirty_page(page, dirty_log)
for (i = 0; i < host-page-size/16; i += 16)
append_tag(LDG(page + i))
to get all the tags for each dirty page. I understood it would be faster
to use LDGM, but we'd need a new ioctl for that. So I was proposing we
just piggyback on a new dirty-log ioctl instead.
Thanks,
drew
On Wed, Sep 09, 2020 at 05:04:15PM +0100, Steven Price wrote:
> On 09/09/2020 16:25, Andrew Jones wrote:
> > On Fri, Sep 04, 2020 at 05:00:16PM +0100, Steven Price wrote:
> > > 2. Automatically promotes (normal host) memory given to the guest to be
> > > tag enabled (sets PG_mte_tagged), if any VCPU has MTE enabled. The
> > > tags are cleared if the memory wasn't previously MTE enabled.
> >
> > Shouldn't this be up to the guest? Or, is this required in order for the
> > guest to use tagging at all. Something like making the guest IPAs memtag
> > capable, but if the guest doesn't enable tagging then there is no guest
> > impact? In any case, shouldn't userspace be the one that adds PROT_MTE
> > to the memory regions it wants the guest to be able to use tagging with,
> > rather than KVM adding the attribute page by page?
>
> I think I've probably explained this badly.
>
> The guest can choose how to populate the stage 1 mapping - so can choose
> which parts of memory are accessed tagged or not. However, the hypervisor
> cannot restrict this in stage 2 (except by e.g. making the memory uncached
> but that's obviously not great - however devices forward to the guest can be
> handled like this).
>
> Because the hypervisor cannot restrict the guest's access to the tags, the
> hypervisor must assume that all memory given to the guest could have the
> tags accessed. So it must (a) clear any stale data from the tags, and (b)
> ensure that the tags are preserved (e.g. when swapping pages out).
>
Yes, this is how I understood it.
> Because of the above the current series automatically sets PG_mte_tagged on
> the pages. Note that this doesn't change the mappings that the VMM has (a
> non-PROT_MTE mapping will still not have access to the tags).
But if userspace created the memslots with memory already set with
PROT_MTE, then this wouldn't be necessary, right? And, as long as
there's still a way to access the memory with tag checking disabled,
then it shouldn't be a problem.
> >
> > If userspace needs to write to guest memory then it should be due to
> > a device DMA or other specific hardware emulation. Those accesses can
> > be done with tag checking disabled.
>
> Yes, the question is can the VMM (sensibly) wrap the accesses with a
> disable/renable tag checking for the process sequence. The alternative at
> the moment is to maintain a separate (untagged) mapping for the purpose
> which might present it's own problems.
Hmm, so there's no easy way to disable tag checking when necessary? If we
don't map the guest ram with PROT_MTE and continue setting the attribute
in KVM, as this series does, then we don't need to worry about it tag
checking when accessing the memory, but then we can't access the tags for
migration.
>
> > >
> > > If it's not practical to either disable tag checking in the VMM or
> > > maintain multiple mappings then the alternatives I'm aware of are:
> > >
> > > * Provide a KVM-specific method to extract the tags from guest memory.
> > > This might also have benefits in terms of providing an easy way to
> > > read bulk tag data from guest memory (since the LDGM instruction
> > > isn't available at EL0).
> >
> > Maybe we need a new version of KVM_GET_DIRTY_LOG that also provides
> > the tags for all addresses of each dirty page.
>
> Certainly possible, although it seems to conflate two operations: "get list
> of dirty pages", "get tags from page". It would also require a lot of return
> space (size of slot/32).
>
It would require num-set-bits * host-page-size / 16 / 2, right?
> > > * Provide support for user space setting the TCMA0 or TCMA1 bits in
> > > TCR_EL1. These would allow the VMM to generate pointers which are not
> > > tag checked.
> >
> > So this is necessary to allow the VMM to keep tag checking enabled for
> > itself, plus map guest memory as PROT_MTE, and write to that memory when
> > needed?
>
> This is certainly one option. The architecture provides two "magic" values
> (all-0s and all-1s) which can be configured using TCMAx to be treated
> differently. The VMM could therefore construct pointers to otherwise tagged
> memory which would be treated as untagged.
>
> However, Catalin's user space series doesn't at the moment expose this
> functionality.
>
So if I understand correctly this would allow us to map the guest memory
with PAGE_MTE and still access the memory when needed. If so, then this
sounds interesting.
Thanks,
drew
On 10/09/2020 07:29, Andrew Jones wrote:
> On Wed, Sep 09, 2020 at 05:04:15PM +0100, Steven Price wrote:
>> On 09/09/2020 16:25, Andrew Jones wrote:
>>> On Fri, Sep 04, 2020 at 05:00:16PM +0100, Steven Price wrote:
>>>> 2. Automatically promotes (normal host) memory given to the guest to be
>>>> tag enabled (sets PG_mte_tagged), if any VCPU has MTE enabled. The
>>>> tags are cleared if the memory wasn't previously MTE enabled.
>>>
>>> Shouldn't this be up to the guest? Or, is this required in order for the
>>> guest to use tagging at all. Something like making the guest IPAs memtag
>>> capable, but if the guest doesn't enable tagging then there is no guest
>>> impact? In any case, shouldn't userspace be the one that adds PROT_MTE
>>> to the memory regions it wants the guest to be able to use tagging with,
>>> rather than KVM adding the attribute page by page?
>>
>> I think I've probably explained this badly.
>>
>> The guest can choose how to populate the stage 1 mapping - so can choose
>> which parts of memory are accessed tagged or not. However, the hypervisor
>> cannot restrict this in stage 2 (except by e.g. making the memory uncached
>> but that's obviously not great - however devices forward to the guest can be
>> handled like this).
>>
>> Because the hypervisor cannot restrict the guest's access to the tags, the
>> hypervisor must assume that all memory given to the guest could have the
>> tags accessed. So it must (a) clear any stale data from the tags, and (b)
>> ensure that the tags are preserved (e.g. when swapping pages out).
>>
>
> Yes, this is how I understood it.
Ok, I've obviously misunderstood your comment instead ;)
>> Because of the above the current series automatically sets PG_mte_tagged on
>> the pages. Note that this doesn't change the mappings that the VMM has (a
>> non-PROT_MTE mapping will still not have access to the tags).
>
> But if userspace created the memslots with memory already set with
> PROT_MTE, then this wouldn't be necessary, right? And, as long as
> there's still a way to access the memory with tag checking disabled,
> then it shouldn't be a problem.
Yes, so one option would be to attempt to validate that the VMM has
provided memory pages with the PG_mte_tagged bit set (e.g. by mapping
with PROT_MTE). The tricky part here is that we support KVM_CAP_SYNC_MMU
which means that the VMM can change the memory backing at any time - so
we could end up in user_mem_abort() discovering that a page doesn't have
PG_mte_tagged set - at that point there's no nice way of handling it
(other than silently upgrading the page) so the VM is dead.
So since enforcing that PG_mte_tagged is set isn't easy and provides a
hard-to-debug foot gun to the VMM I decided the better option was to let
the kernel set the bit automatically.
>>>
>>> If userspace needs to write to guest memory then it should be due to
>>> a device DMA or other specific hardware emulation. Those accesses can
>>> be done with tag checking disabled.
>>
>> Yes, the question is can the VMM (sensibly) wrap the accesses with a
>> disable/renable tag checking for the process sequence. The alternative at
>> the moment is to maintain a separate (untagged) mapping for the purpose
>> which might present it's own problems.
>
> Hmm, so there's no easy way to disable tag checking when necessary? If we
> don't map the guest ram with PROT_MTE and continue setting the attribute
> in KVM, as this series does, then we don't need to worry about it tag
> checking when accessing the memory, but then we can't access the tags for
> migration.
There's a "TCO" (Tag Check Override) bit in PSTATE which allows
disabling tag checking, so if it's reasonable to wrap accesses to the
memory you can simply set the TCO bit, perform the memory access and
then unset TCO. That would mean a single mapping with MTE enabled would
work fine. What I don't have a clue about is whether it's practical in
the VMM to wrap guest accesses like this.
>>
>>>>
>>>> If it's not practical to either disable tag checking in the VMM or
>>>> maintain multiple mappings then the alternatives I'm aware of are:
>>>>
>>>> * Provide a KVM-specific method to extract the tags from guest memory.
>>>> This might also have benefits in terms of providing an easy way to
>>>> read bulk tag data from guest memory (since the LDGM instruction
>>>> isn't available at EL0).
>>>
>>> Maybe we need a new version of KVM_GET_DIRTY_LOG that also provides
>>> the tags for all addresses of each dirty page.
>>
>> Certainly possible, although it seems to conflate two operations: "get list
>> of dirty pages", "get tags from page". It would also require a lot of return
>> space (size of slot/32).
>>
>
> It would require num-set-bits * host-page-size / 16 / 2, right?
Yes, where the worst case is all bits set which is size/32. Since you
don't know at the time of the call how many bits are going to be set I'm
not sure how you would design the API which doesn't require
preallocating the worst case.
>>>> * Provide support for user space setting the TCMA0 or TCMA1 bits in
>>>> TCR_EL1. These would allow the VMM to generate pointers which are not
>>>> tag checked.
>>>
>>> So this is necessary to allow the VMM to keep tag checking enabled for
>>> itself, plus map guest memory as PROT_MTE, and write to that memory when
>>> needed?
>>
>> This is certainly one option. The architecture provides two "magic" values
>> (all-0s and all-1s) which can be configured using TCMAx to be treated
>> differently. The VMM could therefore construct pointers to otherwise tagged
>> memory which would be treated as untagged.
>>
>> However, Catalin's user space series doesn't at the moment expose this
>> functionality.
>>
>
> So if I understand correctly this would allow us to map the guest memory
> with PAGE_MTE and still access the memory when needed. If so, then this
> sounds interesting.
Yes - you could derive a pointer which didn't perform tag checking. Note
that this also requires the rest of user space to play along (i.e.
understand that the tag value is reserved). I believe for user space we
have to use the all-0s value which means that a standard pointer
(top-byte is 0) would be unchecked.
Steve
On 10/09/2020 01:33, Richard Henderson wrote:
> On 9/4/20 9:00 AM, Steven Price wrote:
>> 3. Doesn't provide any new methods for the VMM to access the tags on
>> memory.
> ...
>> (3) may be problematic and I'd welcome input from those familiar with
>> VMMs. User space cannot access tags unless the memory is mapped with the
>> PROT_MTE flag. However enabling PROT_MTE will also enable tag checking
>> for the user space process (assuming the VMM enables tag checking for
>> the process)...
>
> The latest version of the kernel patches for user mte support has separate
> controls for how tag check fail is reported. Including
>
>> +- ``PR_MTE_TCF_NONE`` - *Ignore* tag check faults
>
> That may be less than optimal once userland starts uses tags itself, e.g.
> running qemu itself with an mte-aware malloc.
>
> Independent of that, there's also the TCO bit, which can be toggled by any
> piece of code that wants to disable checking locally.
Yes, I would expect the TCO bit is the best option for wrapping accesses
to make them unchecked.
> However, none of that is required for accessing tags. User space can always
> load/store tags via LDG/STG. That's going to be slow, though.
Yes as things stand LDG/STG is the way for user space to access tags.
Since I don't have any real hardware I can't really comment on speed.
> It's a shame that LDGM/STGM are privileged instructions. I don't understand
> why that was done, since there's absolutely nothing that those insns can do
> that you can't do with (up to) 16x LDG/STG.
It is a shame, however I suspect this is because to use those
instructions you need to know the block size held in GMID_EL1. And at
least in theory that could vary between CPUs.
> I think it might be worth adding some sort of kernel entry point that can bulk
> copy tags, e.g. page aligned quantities. But that's just a speed of migration
> thing and could come later.
When we have some real hardware it would be worth profiling this. At the
moment I've no idea whether the kernel entry overhead would make such an
interface useful from a performance perspective or not.
Steve
* Andrew Jones ([email protected]) wrote:
> On Wed, Sep 09, 2020 at 06:45:33PM -0700, Richard Henderson wrote:
> > On 9/9/20 8:25 AM, Andrew Jones wrote:
> > >> * Provide a KVM-specific method to extract the tags from guest memory.
> > >> This might also have benefits in terms of providing an easy way to
> > >> read bulk tag data from guest memory (since the LDGM instruction
> > >> isn't available at EL0).
> > >
> > > Maybe we need a new version of KVM_GET_DIRTY_LOG that also provides
> > > the tags for all addresses of each dirty page.
> >
> > KVM_GET_DIRTY_LOG just provides one bit per dirty page, no? Then VMM copies
> > the data out from its local address to guest memory.
> >
> > There'd be no difference with or without tags, afaik. It's just about how VMM
> > copies the data, with or without tags.
>
> Right, as long as it's fast enough to do
>
> for_each_dirty_page(page, dirty_log)
> for (i = 0; i < host-page-size/16; i += 16)
> append_tag(LDG(page + i))
>
> to get all the tags for each dirty page. I understood it would be faster
> to use LDGM, but we'd need a new ioctl for that. So I was proposing we
> just piggyback on a new dirty-log ioctl instead.
That feels a bad idea to me; there's a couple of different ways dirty
page checking work; lets keep extracting the tags separate.
Dave
> Thanks,
> drew
--
Dr. David Alan Gilbert / [email protected] / Manchester, UK
On 9/10/20 3:24 AM, Steven Price wrote:
> It is a shame, however I suspect this is because to use those instructions you
> need to know the block size held in GMID_EL1. And at least in theory that could
> vary between CPUs.
Which is no different from having to read DCZID_EL0 in order to implement
memset, in my opinion. But, whatever.
> When we have some real hardware it would be worth profiling this. At the moment
> I've no idea whether the kernel entry overhead would make such an interface
> useful from a performance perspective or not.
Yep.
r~
On Thu, Sep 10, 2020 at 10:21:04AM +0100, Steven Price wrote:
> On 10/09/2020 07:29, Andrew Jones wrote:
> > But if userspace created the memslots with memory already set with
> > PROT_MTE, then this wouldn't be necessary, right? And, as long as
> > there's still a way to access the memory with tag checking disabled,
> > then it shouldn't be a problem.
>
> Yes, so one option would be to attempt to validate that the VMM has provided
> memory pages with the PG_mte_tagged bit set (e.g. by mapping with PROT_MTE).
> The tricky part here is that we support KVM_CAP_SYNC_MMU which means that
> the VMM can change the memory backing at any time - so we could end up in
> user_mem_abort() discovering that a page doesn't have PG_mte_tagged set - at
> that point there's no nice way of handling it (other than silently upgrading
> the page) so the VM is dead.
>
> So since enforcing that PG_mte_tagged is set isn't easy and provides a
> hard-to-debug foot gun to the VMM I decided the better option was to let the
> kernel set the bit automatically.
>
The foot gun still exists when migration is considered, no? If userspace
is telling a guest it can use MTE on its normal memory, but then doesn't
prepare that memory correctly, or remember to migrate the tags correctly
(which requires knowing the memory has tags and knowing how to get them),
then I guess the VM is in trouble one way or another.
I feel like we should trust the VMM to ensure MTE will work on any memory
the guest could use it on, and change the action in user_mem_abort() to
abort the guest with a big error message if it sees the flag is missing.
> > > >
> > > > If userspace needs to write to guest memory then it should be due to
> > > > a device DMA or other specific hardware emulation. Those accesses can
> > > > be done with tag checking disabled.
> > >
> > > Yes, the question is can the VMM (sensibly) wrap the accesses with a
> > > disable/renable tag checking for the process sequence. The alternative at
> > > the moment is to maintain a separate (untagged) mapping for the purpose
> > > which might present it's own problems.
> >
> > Hmm, so there's no easy way to disable tag checking when necessary? If we
> > don't map the guest ram with PROT_MTE and continue setting the attribute
> > in KVM, as this series does, then we don't need to worry about it tag
> > checking when accessing the memory, but then we can't access the tags for
> > migration.
>
> There's a "TCO" (Tag Check Override) bit in PSTATE which allows disabling
> tag checking, so if it's reasonable to wrap accesses to the memory you can
> simply set the TCO bit, perform the memory access and then unset TCO. That
> would mean a single mapping with MTE enabled would work fine. What I don't
> have a clue about is whether it's practical in the VMM to wrap guest
> accesses like this.
>
At least QEMU goes through many abstractions to get to memory already.
There may already be a hook we could use, if not, it probably wouldn't
be too hard to add one (famous last words).
Thanks,
drew
On 10/09/2020 14:56, Andrew Jones wrote:
> On Thu, Sep 10, 2020 at 10:21:04AM +0100, Steven Price wrote:
>> On 10/09/2020 07:29, Andrew Jones wrote:
>>> But if userspace created the memslots with memory already set with
>>> PROT_MTE, then this wouldn't be necessary, right? And, as long as
>>> there's still a way to access the memory with tag checking disabled,
>>> then it shouldn't be a problem.
>>
>> Yes, so one option would be to attempt to validate that the VMM has provided
>> memory pages with the PG_mte_tagged bit set (e.g. by mapping with PROT_MTE).
>> The tricky part here is that we support KVM_CAP_SYNC_MMU which means that
>> the VMM can change the memory backing at any time - so we could end up in
>> user_mem_abort() discovering that a page doesn't have PG_mte_tagged set - at
>> that point there's no nice way of handling it (other than silently upgrading
>> the page) so the VM is dead.
>>
>> So since enforcing that PG_mte_tagged is set isn't easy and provides a
>> hard-to-debug foot gun to the VMM I decided the better option was to let the
>> kernel set the bit automatically.
>>
>
> The foot gun still exists when migration is considered, no? If userspace
> is telling a guest it can use MTE on its normal memory, but then doesn't
> prepare that memory correctly, or remember to migrate the tags correctly
> (which requires knowing the memory has tags and knowing how to get them),
> then I guess the VM is in trouble one way or another.
Well not all VMMs support migration, and it's only migration that is
affected by this for a simple VMM (e.g. the changes to kvmtool are
minimal for MTE). But yes fundamentally if a VMM enables MTE it needs to
know how to deal with the extra tags everywhere.
> I feel like we should trust the VMM to ensure MTE will work on any memory
> the guest could use it on, and change the action in user_mem_abort() to
> abort the guest with a big error message if it sees the flag is missing.
I'm happy to change it, if you feel this is easier to debug.
>>>>>
>>>>> If userspace needs to write to guest memory then it should be due to
>>>>> a device DMA or other specific hardware emulation. Those accesses can
>>>>> be done with tag checking disabled.
>>>>
>>>> Yes, the question is can the VMM (sensibly) wrap the accesses with a
>>>> disable/renable tag checking for the process sequence. The alternative at
>>>> the moment is to maintain a separate (untagged) mapping for the purpose
>>>> which might present it's own problems.
>>>
>>> Hmm, so there's no easy way to disable tag checking when necessary? If we
>>> don't map the guest ram with PROT_MTE and continue setting the attribute
>>> in KVM, as this series does, then we don't need to worry about it tag
>>> checking when accessing the memory, but then we can't access the tags for
>>> migration.
>>
>> There's a "TCO" (Tag Check Override) bit in PSTATE which allows disabling
>> tag checking, so if it's reasonable to wrap accesses to the memory you can
>> simply set the TCO bit, perform the memory access and then unset TCO. That
>> would mean a single mapping with MTE enabled would work fine. What I don't
>> have a clue about is whether it's practical in the VMM to wrap guest
>> accesses like this.
>>
>
> At least QEMU goes through many abstractions to get to memory already.
> There may already be a hook we could use, if not, it probably wouldn't
> be too hard to add one (famous last words).
Sounds good. My hope was that the abstractions were already in there.
Thanks,
Steve
On Thu, Sep 10, 2020 at 02:27:48PM +0100, Dr. David Alan Gilbert wrote:
> * Andrew Jones ([email protected]) wrote:
> > On Wed, Sep 09, 2020 at 06:45:33PM -0700, Richard Henderson wrote:
> > > On 9/9/20 8:25 AM, Andrew Jones wrote:
> > > >> * Provide a KVM-specific method to extract the tags from guest memory.
> > > >> This might also have benefits in terms of providing an easy way to
> > > >> read bulk tag data from guest memory (since the LDGM instruction
> > > >> isn't available at EL0).
> > > >
> > > > Maybe we need a new version of KVM_GET_DIRTY_LOG that also provides
> > > > the tags for all addresses of each dirty page.
> > >
> > > KVM_GET_DIRTY_LOG just provides one bit per dirty page, no? Then VMM copies
> > > the data out from its local address to guest memory.
> > >
> > > There'd be no difference with or without tags, afaik. It's just about how VMM
> > > copies the data, with or without tags.
> >
> > Right, as long as it's fast enough to do
> >
> > for_each_dirty_page(page, dirty_log)
> > for (i = 0; i < host-page-size/16; i += 16)
> > append_tag(LDG(page + i))
> >
> > to get all the tags for each dirty page. I understood it would be faster
> > to use LDGM, but we'd need a new ioctl for that. So I was proposing we
> > just piggyback on a new dirty-log ioctl instead.
>
> That feels a bad idea to me; there's a couple of different ways dirty
> page checking work; lets keep extracting the tags separate.
>
It's sounding like it was a premature optimization anyway. We don't yet
know if an ioctl for LDGM is worth it. Looping over LDG may work fine.
Thanks,
drew