KVM: arm/arm64: vgic-its: Fix restoration of unmapped collections
PATCH 1/1 KVM: arm/arm64: vgic-its: Fix restoration of unmapped collections
Saving/restoring an unmapped collection is a valid scenario. For example this happens if a MAPTI command was sent, featuring an unmapped collection. At the moment the CTE fails to be restored. Only compare against the number of online vcpus if the rdist base is set.
--- virt/kvm/arm/vgic/vgic-its.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
virt/kvm/arm/vgic/vgic-its.c +2 -1
@@ -2475,7 +2475,8 @@ static int vgic_its_restore_cte(struct vgic_its *its, gpa_t gpa, int esz) target_addr = (u32)(val >>KVM_ITS_CTE_RDBASE_SHIFT ); coll_id = val &KVM_ITS_CTE_ICID_MASK ; -if (target_addr >=atomic_read (&kvm->online_vcpus)) +if (target_addr !=COLLECTION_NOT_MAPPED && + target_addr >=atomic_read (&kvm->online_vcpus))return -EINVAL ; collection =find_collection (its, coll_id);
Signature · 1 line
2.20.1
Marc Zyngier
maintainerHi Eric,
Quoted 27 lines · Eric Auger
Saving/restoring an unmapped collection is a valid scenario. For example this happens if a MAPTI command was sent, featuring an unmapped collection. At the moment the CTE fails to be restored. Only compare against the number of online vcpus if the rdist base is set.
Cc [email protected] # v4.11+Fixes ea1ad53e1e31a ("KVM: arm64: vgic-its: Collection table save/restore")Signed-off-by Eric Auger <[email protected]>--- virt/kvm/arm/vgic/vgic-its.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)virt/kvm/arm/vgic/vgic-its.c +2
-1@@ -2475,7 +2475,8 @@ static int vgic_its_restore_cte(struct vgic_its *its, gpa_t gpa, int esz) target_addr = (u32)(val >>KVM_ITS_CTE_RDBASE_SHIFT ); coll_id = val &KVM_ITS_CTE_ICID_MASK ; -if (target_addr >=atomic_read (&kvm->online_vcpus)) +if (target_addr !=COLLECTION_NOT_MAPPED && + target_addr >=atomic_read (&kvm->online_vcpus))return -EINVAL ;
Quoted 1 line
collection =
find_collection(its, coll_id);
Looks good to me. Out of curiosity, how was this spotted?
Thanks,
M.
Signature · 1 line
Jazz is not dead. It just smells funny...
Hi Marc,
Quoted 26 lines · Marc Zyngier
Hi Eric,
Quoted 23 lines · Eric Auger
Saving/restoring an unmapped collection is a valid scenario. For example this happens if a MAPTI command was sent, featuring an unmapped collection. At the moment the CTE fails to be restored. Only compare against the number of online vcpus if the rdist base is set.
Cc [email protected] # v4.11+Fixes ea1ad53e1e31a ("KVM: arm64: vgic-its: Collection tablesave/restore")
Signed-off-by Eric Auger <[email protected]>---
virt/kvm/arm/vgic/vgic-its.c | 3 ++-1 file changed, 2 insertions(+), 1 deletion(-)virt/kvm/arm/vgic/vgic-its.c +0
-0@@ -2475,7 +2475,8 @@ static int vgic_its_restore_cte(struct vgic_its *its, gpa_t gpa, int esz) target_addr = (u32)(val >> KVM_ITS_CTE_RDBASE_SHIFT); coll_id = val & KVM_ITS_CTE_ICID_MASK;
Quoted 8 lines
Quoted 6 lines
virt/kvm/arm/vgic/vgic-its.c +2
-1- if (target_addr >= atomic_read(&kvm->online_vcpus)) + if (target_addr != COLLECTION_NOT_MAPPED && + target_addr >= atomic_read(&kvm->online_vcpus)) return -EINVAL;collection =
find_collection(its, coll_id);Looks good to me. Out of curiosity, how was this spotted?
I am currently writing some kvm-unit-tests to better test ITS and its migration.
Thanks
Eric
Quoted 4 lines
Thanks,
M.
Hi Eric,
Quoted 5 lines · Eric Auger
Saving/restoring an unmapped collection is a valid scenario. For example this happens if a MAPTI command was sent, featuring an unmapped collection. At the moment the CTE fails to be restored. Only compare against the number of online vcpus if the rdist base is set.
Have you actually seen a problem and this patch fixed it? To be honest, I'm surprised to find that we can map a LPI to an unmapped collection ;) (and prevent it to be delivered to vcpu with an INT_UNMAPPED_INTERRUPT error, until someone had actually mapped the collection). After a quick glance of spec (MAPTI), just as you said, this is valid.
If Marc has no objection to this fix, please add
Thanks,
Zenghui
Quoted 14 lines
Cc [email protected] # v4.11+Fixes ea1ad53e1e31a ("KVM: arm64: vgic-its: Collection table save/restore")Signed-off-by Eric Auger <[email protected]>--- virt/kvm/arm/vgic/vgic-its.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)virt/kvm/arm/vgic/vgic-its.c +0
-0
virt/kvm/arm/vgic/vgic-its.c static int vgic_its_restore_cte(struct vgic_its *its, gpa_t gpa, int esz) · line 2482
Quoted 9 lines
virt/kvm/arm/vgic/vgic-its.c +2
-1@@ -2475,7 +2475,8 @@ static int vgic_its_restore_cte(struct vgic_its *its, gpa_t gpa, int esz) target_addr = (u32)(val >>KVM_ITS_CTE_RDBASE_SHIFT ); coll_id = val &KVM_ITS_CTE_ICID_MASK ; -if (target_addr >=atomic_read (&kvm->online_vcpus)) +if (target_addr !=COLLECTION_NOT_MAPPED && + target_addr >=atomic_read (&kvm->online_vcpus))return -EINVAL ; collection =find_collection (its, coll_id);
_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
Hi Zenghui,
Quoted 10 lines · Zenghui Yu
Hi Eric,
Quoted 5 lines · Eric Auger
Saving/restoring an unmapped collection is a valid scenario. For example this happens if a MAPTI command was sent, featuring an unmapped collection. At the moment the CTE fails to be restored. Only compare against the number of online vcpus if the rdist base is set.
Have you actually seen a problem and this patch fixed it?
It is not with a linux guest but with kvm-unit-test.
To be honest,
Quoted 8 lines
I'm surprised to find that we can map a LPI to an unmapped collection ;) (and prevent it to be delivered to vcpu with an
INT_UNMAPPED_INTERRUPTerror, until someone had actually mapped the collection). After a quick glance of spec (MAPTI), just as you said, this is valid.If Marc has no objection to this fix, please add
Reviewed-by Zenghui Yu <[email protected]>
Thank you for the review.
Eric
Quoted 29 lines
Thanks,
ZenghuiQuoted 23 lines
Cc [email protected] # v4.11+Fixes ea1ad53e1e31a ("KVM: arm64: vgic-its: Collection tablesave/restore")
Signed-off-by Eric Auger <[email protected]>---
virt/kvm/arm/vgic/vgic-its.c| 3 ++-1 file changed, 2 insertions(+), 1 deletion(-)virt/kvm/arm/vgic/vgic-its.c +2
-0@@ -2475,7 +2475,8 @@ static int vgic_its_restore_cte(struct vgic_its *its, gpa_t gpa, int esz) target_addr = (u32)(val >> KVM_ITS_CTE_RDBASE_SHIFT); coll_id = val & KVM_ITS_CTE_ICID_MASK; - if (target_addr >= atomic_read(&kvm->online_vcpus)) + if (target_addr != COLLECTION_NOT_MAPPED && + target_addr >= atomic_read(&kvm->online_vcpus)) return -EINVAL; collection = find_collection(its, coll_id);
Marc Zyngier
maintainerHi Zenghui,
Quoted 9 lines · Zenghui Yu
Hi Eric,
Quoted 5 lines · Eric Auger
Saving/restoring an unmapped collection is a valid scenario. For example this happens if a MAPTI command was sent, featuring an unmapped collection. At the moment the CTE fails to be restored. Only compare against the number of online vcpus if the rdist base is set.
Quoted 5 lines
Have you actually seen a problem and this patch fixed it? To be honest, I'm surprised to find that we can map a LPI to an unmapped collection ;) (and prevent it to be delivered to vcpu with an
INT_UNMAPPED_INTERRUPTerror, until someone had actually mapped the collection). After a quick glance of spec (MAPTI), just as you said, this is valid.
Yes, this is one of the (many) odd bits in the architecture. And there is a bizarre wording in the MAPC description when V=0:
"Behavior is unpredictable if there are interrupts that are mapped to the specified collection, with the restriction that further translation requests from that device are ignored."
It is really odd that:
- it is unpredictable to unmap the collection with mapped interrupts, but mapping interrupts to an unmapped collection is fine
- the notion of "interrupts from that device" doesn't match any of the MAPC parameters
Do you hate the GIC already? ;-)
Quoted 3 lines
If Marc has no objection to this fix, please add
Reviewed-by Zenghui Yu <[email protected]>
Thanks for that, I've applied it to the patch and will push out the update as soon as ra.kernel.org is reachable again.
M.
Signature · 1 line
Jazz is not dead. It just smells funny...
Zenghui Yu
reviewerHi Marc,
Quoted 21 lines · Marc Zyngier
Hi Zenghui,
Quoted 9 lines · Zenghui Yu
Hi Eric,
Quoted 5 lines · Eric Auger
Saving/restoring an unmapped collection is a valid scenario. For example this happens if a MAPTI command was sent, featuring an unmapped collection. At the moment the CTE fails to be restored. Only compare against the number of online vcpus if the rdist base is set.
Quoted 5 lines
Have you actually seen a problem and this patch fixed it? To be honest, I'm surprised to find that we can map a LPI to an unmapped collection ;) (and prevent it to be delivered to vcpu with an
INT_UNMAPPED_INTERRUPTerror, until someone had actually mapped the collection). After a quick glance of spec (MAPTI), just as you said, this is valid.Yes, this is one of the (many) odd bits in the architecture. And there is a bizarre wording in the MAPC description when V=0:
Quoted 8 lines
"Behavior is unpredictable if there are interrupts that are mapped to the specified collection, with the restriction that further translation requests from that device are ignored."
It is really odd that:
- it is unpredictable to unmap the collection with mapped interrupts, but mapping interrupts to an unmapped collection is fine
Yes, looks odd... Without Eric's patch, I won't even notice it.
I guess that unmapping the collection with mapped interrupts will make it difficult for Hardware to manage those interrupts' internal states, but only a guess.
Quoted 2 lines
- the notion of "interrupts from that device" doesn't match any of the MAPC parameters
Looks like a writing mistake, a better statement *might be* "further translation requests targeting that ICID are ignored"?
Quoted 1 line
Do you hate the GIC already? ;-)
Not yet! (I'd like to continue being messed with GIC and see when I will hate it :)
Quoted 7 lines
Quoted 4 lines
If Marc has no objection to this fix, please add
Reviewed-by Zenghui Yu <[email protected]>Thanks for that, I've applied it to the patch and will push out the update as soon as ra.kernel.org is reachable again.
Thanks,
Zenghui