KVM: arm/arm64: user_mem_abort() assorted fixes
Subsystems (3)
Marc Zyngier
maintainercover KVM: arm/arm64: user_mem_abort() assorted fixes
Alexandru recently reported an interesting issue with our handling of device mapping in user_mem_abort(), which is sligtly less than correct. The first patch of the series address this issue, and is a stable candidate.
While I was looking at this code, I spotted what I think is a potential issue when handling a poisoned page, where we can race with a VMA being removed. This second patch is mostly a RFC, as this is not my area of expertise.
Finally, the last patch is a cleanup removing an unnecessary console output.
Marc Zyngier (3):
KVM: arm/arm64: Properly handle faulting of device mappings
KVM: arm/arm64: Re-check VMA on detecting a poisoned page
KVM: arm/arm64: Drop spurious message when faulting on a non-existent
mapping
virt/kvm/arm/mmu.c | 47 +++++++++++++++++++++++++++++++++++++++------- 1 file changed, 40 insertions(+), 7 deletions(-)
Signature · 1 line
2.20.1
Marc Zyngier
maintainerPATCH 1/3 KVM: arm/arm64: Properly handle faulting of device mappings
A device mapping is normally always mapped at Stage-2, since there is very little gain in having it faulted in.
Nonetheless, it is possible to end-up in a situation where the device mapping has been removed from Stage-2 (userspace munmaped the VFIO region, and the MMU notifier did its job), but present in a userspace mapping (userpace has mapped it back at the same address). In such a situation, the device mapping will be demand-paged as the guest performs memory accesses.
This requires to be careful when dealing with mapping size, cache management, and to handle potential execution of a device mapping.
--- virt/kvm/arm/mmu.c | 21 +++++++++++++++++---- 1 file changed, 17 insertions(+), 4 deletions(-)
virt/kvm/arm/mmu.c +17 -4
@@ -38,6 +38,11 @@ static unsigned long io_map_base;#define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)#define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1) +static bool is_iomap (unsigned long flags) +{ +return flags &KVM_S2PTE_FLAG_IS_IOMAP ; +} +static bool memslot_is_logging (struct kvm_memory_slot *memslot) {return memslot->dirty_bitmap && !(memslot->flags &KVM_MEM_READONLY );
@@ -1698,6 +1703,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, vma_pagesize =vma_kernel_pagesize (vma);if (logging_active || + (vma->vm_flags &VM_PFNMAP ) || !fault_supports_stage2_huge_mapping (memslot, hva, vma_pagesize)) { force_pte = true; vma_pagesize =PAGE_SIZE ;
@@ -1760,6 +1766,9 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, writable = false; } +if (exec_fault &&is_iomap (flags)) +return -ENOEXEC ; +spin_lock (&kvm->mmu_lock);if (mmu_notifier_retry (kvm, mmu_seq)) goto out_unlock;
@@ -1781,7 +1790,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,if (writable)kvm_set_pfn_dirty (pfn); -if (fault_status !=FSC_PERM ) +if (fault_status !=FSC_PERM && !is_iomap (flags))clean_dcache_guest_page (pfn, vma_pagesize);if (exec_fault)
@@ -1948,9 +1957,8 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)if (kvm_is_error_hva (hva) || (write_fault && !writable)) {if (is_iabt) {/* Prefetch Abort on I/O address */ -kvm_inject_pabt (vcpu,kvm_vcpu_get_hfar (vcpu)); - ret =1 ; - goto out_unlock; + ret = -ENOEXEC ; + goto out; }/*
@@ -1992,6 +2000,11 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run) ret =user_mem_abort (vcpu, fault_ipa, memslot, hva, fault_status);if (ret ==0 ) ret =1 ; +out: +if (ret == -ENOEXEC ) { +kvm_inject_pabt (vcpu,kvm_vcpu_get_hfar (vcpu)); + ret =1 ; + } out_unlock:srcu_read_unlock (&vcpu->kvm->srcu, idx);return ret;
Signature · 1 line
2.20.1
Re: [PATCH 1/3] KVM: arm/arm64: Properly handle faulting of device mappings
Hi,
Quoted 77 lines · Marc Zyngier
A device mapping is normally always mapped at Stage-2, since there is very little gain in having it faulted in.
Nonetheless, it is possible to end-up in a situation where the device mapping has been removed from Stage-2 (userspace munmaped the VFIO region, and the MMU notifier did its job), but present in a userspace mapping (userpace has mapped it back at the same address). In such a situation, the device mapping will be demand-paged as the guest performs memory accesses.
This requires to be careful when dealing with mapping size, cache management, and to handle potential execution of a device mapping.
Reported-by Alexandru Elisei <[email protected]>Signed-off-by Marc Zyngier <[email protected]>--- virt/kvm/arm/mmu.c | 21 +++++++++++++++++---- 1 file changed, 17 insertions(+), 4 deletions(-)virt/kvm/arm/mmu.c +12
-4@@ -38,6 +38,11 @@ static unsigned long io_map_base;#define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)#define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1) +static bool is_iomap (unsigned long flags) +{ +return flags &KVM_S2PTE_FLAG_IS_IOMAP ; +} +static bool memslot_is_logging (struct kvm_memory_slot *memslot) {return memslot->dirty_bitmap && !(memslot->flags &KVM_MEM_READONLY );@@ -1698,6 +1703,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, vma_pagesize =vma_kernel_pagesize (vma);if (logging_active || + (vma->vm_flags &VM_PFNMAP ) || !fault_supports_stage2_huge_mapping (memslot, hva, vma_pagesize)) { force_pte = true; vma_pagesize =PAGE_SIZE ;@@ -1760,6 +1766,9 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, writable = false; } +if (exec_fault &&is_iomap (flags)) +return -ENOEXEC ; +spin_lock (&kvm->mmu_lock);if (mmu_notifier_retry (kvm, mmu_seq)) goto out_unlock;@@ -1781,7 +1790,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,if (writable)kvm_set_pfn_dirty (pfn); -if (fault_status !=FSC_PERM ) +if (fault_status !=FSC_PERM && !is_iomap (flags))clean_dcache_guest_page (pfn, vma_pagesize);if (exec_fault)@@ -1948,9 +1957,8 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)if (kvm_is_error_hva (hva) || (write_fault && !writable)) {if (is_iabt) {/* Prefetch Abort on I/O address */ -kvm_inject_pabt (vcpu,kvm_vcpu_get_hfar (vcpu)); - ret =1 ; - goto out_unlock; + ret = -ENOEXEC ; + goto out; }/* @@ -1992,6 +2000,11 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run) ret =user_mem_abort (vcpu, fault_ipa, memslot, hva, fault_status);
virt/kvm/arm/mmu.c int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run) · line 2010
Quoted 10 lines
virt/kvm/arm/mmu.c +5
-0if (ret ==0 ) ret =1 ; +out: +if (ret == -ENOEXEC ) { +kvm_inject_pabt (vcpu,kvm_vcpu_get_hfar (vcpu)); + ret =1 ; + } out_unlock:srcu_read_unlock (&vcpu->kvm->srcu, idx);return ret;
I've tested this patch using the scenario that you described in the commit message. I've also tested it with the following scenarios:
- The region is mmap'ed as backed by a VFIO device fd and the memslot is created, it is munmap'ed, then mmap'ed as anonymous.
- The region is mmap'ed as anonymous and the memslot is created, it is munmap'ed, then mmap'ed as backed by a VFIO device fd.
Everything worked as expected, so:
Just a nitpick, but stage2_set_pte has a local variable iomap which can be replaced with a call to is_iomap.
Thanks,
Alex
Marc Zyngier
maintainerRe: [PATCH 1/3] KVM: arm/arm64: Properly handle faulting of device mappings
Hi Alex,
Quoted 22 lines · Alexandru Elisei
Hi,
Quoted 19 lines · Marc Zyngier
A device mapping is normally always mapped at Stage-2, since there is very little gain in having it faulted in.
Nonetheless, it is possible to end-up in a situation where the device mapping has been removed from Stage-2 (userspace munmaped the VFIO region, and the MMU notifier did its job), but present in a userspace mapping (userpace has mapped it back at the same address). In such a situation, the device mapping will be demand-paged as the guest performs memory accesses.
This requires to be careful when dealing with mapping size, cache management, and to handle potential execution of a device mapping.
Reported-by Alexandru Elisei <[email protected]>Signed-off-by Marc Zyngier <[email protected]>--- virt/kvm/arm/mmu.c | 21 +++++++++++++++++---- 1 file changed, 17 insertions(+), 4 deletions(-)
[...]
Quoted 10 lines
I've tested this patch using the scenario that you described in the commit message. I've also tested it with the following scenarios:
- The region is mmap'ed as backed by a VFIO device fd and the memslot is created,
it is munmap'ed, then mmap'ed as anonymous.
- The region is mmap'ed as anonymous and the memslot is created, it is munmap'ed,
then mmap'ed as backed by a VFIO device fd.
Quoted 3 lines
Everything worked as expected, so:
Tested-by Alexandru Elisei <[email protected]>
Thanks for that!
Quoted 2 lines
Just a nitpick, but
stage2_set_ptehas a local variable iomap which can be replaced with a call tois_iomap.
Yeah, I noticed. I'm just trying to keep the patch relatively small so that it can be easily backported. The cleanup can come later! ;-)
Cheers,
M.
Signature · 1 line
Jazz is not dead. It just smells funny...
Re: [PATCH 1/3] KVM: arm/arm64: Properly handle faulting of device mappings
In reply to Marc ZyngierHi Marc,
Quoted 10 lines · Marc Zyngier
A device mapping is normally always mapped at Stage-2, since there is very little gain in having it faulted in.
Nonetheless, it is possible to end-up in a situation where the device mapping has been removed from Stage-2 (userspace munmaped the VFIO region, and the MMU notifier did its job), but present in a userspace mapping (userpace has mapped it back at the same address). In such a situation, the device mapping will be demand-paged as the guest performs memory accesses.
Quoted 2 lines
This requires to be careful when dealing with mapping size, cache management, and to handle potential execution of a device mapping.
Thanks,
James
Re: [PATCH 1/3] KVM: arm/arm64: Properly handle faulting of device mappings
In reply to Marc ZyngierHi Marc,
Quoted 2 lines · Marc Zyngier
A device mapping is normally always mapped at Stage-2, since there is very little gain in having it faulted in.
It is actually becoming less clear to me what the real benefits of pre-populating the stage 2 page table are, especially given that we can provoke a situation where they're faulted in anyhow. Do you recall if we had any specific case that motivated us to pre-fault in the pages?
Quoted 23 lines
Nonetheless, it is possible to end-up in a situation where the device mapping has been removed from Stage-2 (userspace munmaped the VFIO region, and the MMU notifier did its job), but present in a userspace mapping (userpace has mapped it back at the same address). In such a situation, the device mapping will be demand-paged as the guest performs memory accesses.
This requires to be careful when dealing with mapping size, cache management, and to handle potential execution of a device mapping.
Reported-by Alexandru Elisei <[email protected]>Signed-off-by Marc Zyngier <[email protected]>--- virt/kvm/arm/mmu.c | 21 +++++++++++++++++---- 1 file changed, 17 insertions(+), 4 deletions(-)virt/kvm/arm/mmu.c +0
-0
virt/kvm/arm/mmu.c static unsigned long io_map_base; · line 45
Quoted 8 lines
virt/kvm/arm/mmu.c +5
-0@@ -38,6 +38,11 @@ static unsigned long io_map_base;#define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)#define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1) +static bool is_iomap (unsigned long flags) +{ +return flags &KVM_S2PTE_FLAG_IS_IOMAP ; +} +
nit: I'm not really sure this indirection makes the code more readable, but I guess that's a matter of taste.
virt/kvm/arm/mmu.c static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, · line 1706
Quoted 8 lines
virt/kvm/arm/mmu.c +1
-0static bool memslot_is_logging (struct kvm_memory_slot *memslot) {return memslot->dirty_bitmap && !(memslot->flags &KVM_MEM_READONLY );@@ -1698,6 +1703,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, vma_pagesize =vma_kernel_pagesize (vma); if (logging_active || + (vma->vm_flags & VM_PFNMAP) ||
WHat is actually the rationale for this?
Why is a huge mapping not permitted to device memory?
Are we guaranteed that VM_PFNMAP on the vma results in device mappings? I'm not convinced this is the case, and it would be better if we can stick to a single primitive (either kvm_is_device_pfn, or VM_PFNMAP) to detect device mappings.
As a subsequent patch, I'd like to make sure that at the very least our memslot prepare function follows the exact same logic for mapping device memory as a fault-in approach does, or that we simply always fault pages in.
virt/kvm/arm/mmu.c static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, · line 1771
Quoted 10 lines
virt/kvm/arm/mmu.c +3
-0!fault_supports_stage2_huge_mapping(memslot, hva, vma_pagesize)) { force_pte = true; vma_pagesize =PAGE_SIZE ;@@ -1760,6 +1766,9 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, writable = false; } +if (exec_fault &&is_iomap (flags)) +return -ENOEXEC ; +
nit: why don't you just do this when checking kvm_is_device_pfn() and avoid having logic in two places to deal with this case?
Quoted 39 lines
virt/kvm/arm/mmu.c +8
-4spin_lock (&kvm->mmu_lock);if (mmu_notifier_retry (kvm, mmu_seq)) goto out_unlock;@@ -1781,7 +1790,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,if (writable)kvm_set_pfn_dirty (pfn); -if (fault_status !=FSC_PERM ) +if (fault_status !=FSC_PERM && !is_iomap (flags))clean_dcache_guest_page (pfn, vma_pagesize);if (exec_fault)@@ -1948,9 +1957,8 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)if (kvm_is_error_hva (hva) || (write_fault && !writable)) {if (is_iabt) {/* Prefetch Abort on I/O address */ -kvm_inject_pabt (vcpu,kvm_vcpu_get_hfar (vcpu)); - ret =1 ; - goto out_unlock; + ret = -ENOEXEC ; + goto out; }/* @@ -1992,6 +2000,11 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run) ret =user_mem_abort (vcpu, fault_ipa, memslot, hva, fault_status);if (ret ==0 ) ret =1 ; +out: +if (ret == -ENOEXEC ) { +kvm_inject_pabt (vcpu,kvm_vcpu_get_hfar (vcpu)); + ret =1 ; + } out_unlock:srcu_read_unlock (&vcpu->kvm->srcu, idx);return ret;Signature · 1 line
2.20.1
I can't seem to decide for myself if I think there's a sematic difference between trying to execute from somewhere the VMM has explicitly told us is device memory and from somewhere which we happen to have mapped with VM_PFNMAP from user space. But I also can't seem to really fault it (pun intended). Thoughts?
Thanks,
Christoffer
Marc Zyngier
maintainerRe: [PATCH 1/3] KVM: arm/arm64: Properly handle faulting of device mappings
Hi Christoffer,
Quoted 10 lines · Christoffer Dall
Hi Marc,
Quoted 2 lines · Marc Zyngier
A device mapping is normally always mapped at Stage-2, since there is very little gain in having it faulted in.
It is actually becoming less clear to me what the real benefits of pre-populating the stage 2 page table are, especially given that we can provoke a situation where they're faulted in anyhow. Do you recall if we had any specific case that motivated us to pre-fault in the pages?
It's only a minor performance optimization that was introduced by Ard in 8eef91239e57d. Which makes sense for platform devices that have a single fixed location in memory. It makes slightly less sense for PCI, where you can move things around.
Quoted 47 lines
Quoted 22 lines
Nonetheless, it is possible to end-up in a situation where the device mapping has been removed from Stage-2 (userspace munmaped the VFIO region, and the MMU notifier did its job), but present in a userspace mapping (userpace has mapped it back at the same address). In such a situation, the device mapping will be demand-paged as the guest performs memory accesses.
This requires to be careful when dealing with mapping size, cache management, and to handle potential execution of a device mapping.
Reported-by Alexandru Elisei <[email protected]>Signed-off-by Marc Zyngier <[email protected]>--- virt/kvm/arm/mmu.c | 21 +++++++++++++++++---- 1 file changed, 17 insertions(+), 4 deletions(-)virt/kvm/arm/mmu.c +0
-0virt/kvm/arm/mmu.c static unsigned long io_map_base; · line 45
Quoted 8 lines
virt/kvm/arm/mmu.c +5
-0@@ -38,6 +38,11 @@ static unsigned long io_map_base;#define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)#define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1) +static bool is_iomap (unsigned long flags) +{ +return flags &KVM_S2PTE_FLAG_IS_IOMAP ; +} +nit: I'm not really sure this indirection makes the code more readable, but I guess that's a matter of taste.
virt/kvm/arm/mmu.c static int user_mem_abort(struct kvm_vcpu · line 1707
Quoted 10 lines
virt/kvm/arm/mmu.c +0
-0static bool memslot_is_logging (struct kvm_memory_slot *memslot) { return memslot->dirty_bitmap && !(memslot->flags &
KVM_MEM_READONLY);virt/kvm/arm/mmu.c +1
-0@@ -1698,6 +1703,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, vma_pagesize =vma_kernel_pagesize (vma); if (logging_active || + (vma->vm_flags & VM_PFNMAP) ||WHat is actually the rationale for this?
Quoted 6 lines
Why is a huge mapping not permitted to device memory?
Are we guaranteed that
VM_PFNMAPon the vma results in device mappings? I'm not convinced this is the case, and it would be better if we can stick to a single primitive (eitherkvm_is_device_pfn, orVM_PFNMAP) to detect device mappings.
For now, I've tried to keep the two paths that deal with mapping devices (or rather, things that we interpret as devices) as close as possible. If we drop the "eager" mapping, then we're at liberty to restructure this in creative ways.
This includes potential huge mappings, but I'm not sure the rest of the kernel uses them for devices anyway (I need to find out).
Quoted 4 lines
As a subsequent patch, I'd like to make sure that at the very least our memslot prepare function follows the exact same logic for mapping device memory as a fault-in approach does, or that we simply always fault pages in.
As far as I can see, the two approach are now identical. Am I missing something? And yes, getting rid of the eager mapping works for me.
Quoted 10 lines
Quoted 10 lines
virt/kvm/arm/mmu.c +0
-0!fault_supports_stage2_huge_mapping(memslot, hva, vma_pagesize)) {virt/kvm/arm/mmu.c +0
-0force_pte = true; vma_pagesize =PAGE_SIZE ;@@ -1760,6 +1766,9 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, writable = false; }
Quoted 6 lines
virt/kvm/arm/mmu.c static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, · line 1771
Quoted 3 lines
virt/kvm/arm/mmu.c +3
-0+if (exec_fault &&is_iomap (flags)) +return -ENOEXEC ; +nit: why don't you just do this when checking
kvm_is_device_pfn()and avoid having logic in two places to deal with this case?
Good point. I've already sent the PR, but that could be a further cleanup.
Quoted 41 lines
Quoted 41 lines
virt/kvm/arm/mmu.c +0
-0spin_lock (&kvm->mmu_lock);if (mmu_notifier_retry (kvm, mmu_seq))virt/kvm/arm/mmu.c +1
-1goto out_unlock;@@ -1781,7 +1790,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, if (writable)kvm_set_pfn_dirty (pfn); -if (fault_status !=FSC_PERM ) +if (fault_status !=FSC_PERM && !is_iomap (flags))clean_dcache_guest_page (pfn, vma_pagesize);virt/kvm/arm/mmu.c +2
-3if (exec_fault)@@ -1948,9 +1957,8 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)if (kvm_is_error_hva (hva) || (write_fault && !writable)) {if (is_iabt) {/* Prefetch Abort on I/O address */ -kvm_inject_pabt (vcpu,kvm_vcpu_get_hfar (vcpu)); - ret =1 ; - goto out_unlock; + ret = -ENOEXEC ; + goto out; }virt/kvm/arm/mmu.c +5
-0/* @@ -1992,6 +2000,11 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run) ret =user_mem_abort (vcpu, fault_ipa, memslot, hva, fault_status);if (ret ==0 ) ret =1 ; +out: +if (ret == -ENOEXEC ) { +kvm_inject_pabt (vcpu,kvm_vcpu_get_hfar (vcpu)); + ret =1 ; + } out_unlock:srcu_read_unlock (&vcpu->kvm->srcu, idx);return ret;
--
Quoted 8 lines
Quoted 2 lines
2.20.1
I can't seem to decide for myself if I think there's a sematic difference between trying to execute from somewhere the VMM has explicitly told us is device memory and from somewhere which we happen to have mapped with
VM_PFNMAPfrom user space. But I also can't seem to really fault it (pun intended). Thoughts?
The issue is that the VMM never really tells us whether something is a device mapping or not (the only exception being the GICv2 cpuif). Even with PFNMAP, we guess it (it could well be memory that lives outside of the linear mapping). I don't see a way to lift this ambiguity.
Ideally, faulting on executing a non-mapping should be offloaded to userspace for emulation, in line with your patches that offload non-emulated data accesses. That'd be a new ABI, and I can't imagine anyone willing to deal with it.
Thanks,
M.
Signature · 1 line
Jazz is not dead. It just smells funny...
Re: [PATCH 1/3] KVM: arm/arm64: Properly handle faulting of device mappings
Quoted 14 lines · Marc Zyngier
Hi Christoffer,
Quoted 10 lines · Christoffer Dall
Hi Marc,
Quoted 2 lines · Marc Zyngier
A device mapping is normally always mapped at Stage-2, since there is very little gain in having it faulted in.
It is actually becoming less clear to me what the real benefits of pre-populating the stage 2 page table are, especially given that we can provoke a situation where they're faulted in anyhow. Do you recall if we had any specific case that motivated us to pre-fault in the pages?
Quoted 4 lines
It's only a minor performance optimization that was introduced by Ard in
8eef91239e57d. Which makes sense for platform devices that have a single fixed location in memory. It makes slightly less sense for PCI, where you can move things around.
User space could still decide to move things around in its VA map even if the device is fixed.
Anyway, I was thinking more if there was some sort of device, like a frambuffer, which for example crosses page boundaries and where it would be visible to the user that there's a sudden performance drop while operating the device over page boundaries. Anything like that?
Quoted 65 lines
Quoted 50 lines
Quoted 25 lines
Nonetheless, it is possible to end-up in a situation where the device
mapping has been removed from Stage-2 (userspace munmaped the VFIO region, and the MMU notifier did its job), but present in a userspace
mapping (userpace has mapped it back at the same address). In such a situation, the device mapping will be demand-paged as the guest performs memory accesses.This requires to be careful when dealing with mapping size, cache management, and to handle potential execution of a device mapping.
Reported-by Alexandru Elisei <[email protected]>Signed-off-by Marc Zyngier <[email protected]>--- virt/kvm/arm/mmu.c | 21 +++++++++++++++++---- 1 file changed, 17 insertions(+), 4 deletions(-)virt/kvm/arm/mmu.c +0
-0virt/kvm/arm/mmu.c static unsigned long io_map_base; · line 45
Quoted 8 lines
virt/kvm/arm/mmu.c +5
-0@@ -38,6 +38,11 @@ static unsigned long io_map_base;#define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)#define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1) +static bool is_iomap (unsigned long flags) +{ +return flags &KVM_S2PTE_FLAG_IS_IOMAP ; +} +nit: I'm not really sure this indirection makes the code more readable, but I guess that's a matter of taste.
virt/kvm/arm/mmu.c static int user_mem_abort(struct kvm_vcpu · line 1707
Quoted 10 lines
virt/kvm/arm/mmu.c +0
-0static bool memslot_is_logging (struct kvm_memory_slot *memslot) { return memslot->dirty_bitmap && !(memslot->flags &
KVM_MEM_READONLY);virt/kvm/arm/mmu.c +1
-0@@ -1698,6 +1703,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, vma_pagesize =vma_kernel_pagesize (vma); if (logging_active || + (vma->vm_flags & VM_PFNMAP) ||WHat is actually the rationale for this?
Quoted 6 lines
Why is a huge mapping not permitted to device memory?
Are we guaranteed that
VM_PFNMAPon the vma results in device mappings? I'm not convinced this is the case, and it would be better if we can stick to a single primitive (eitherkvm_is_device_pfn, orVM_PFNMAP) to detect device mappings.For now, I've tried to keep the two paths that deal with mapping devices (or rather, things that we interpret as devices) as close as possible. If we drop the "eager" mapping, then we're at liberty to restructure this in creative ways.
This includes potential huge mappings, but I'm not sure the rest of the kernel uses them for devices anyway (I need to find out).
Quoted 9 lines
Quoted 4 lines
As a subsequent patch, I'd like to make sure that at the very least our memslot prepare function follows the exact same logic for mapping device memory as a fault-in approach does, or that we simply always fault pages in.
As far as I can see, the two approach are now identical. Am I missing something?
And yes, getting rid of the eager mapping works for me.
As far as I can tell, our user_mem_abort() uses gfn_to_pfn_prot() which goes doesn a long trail which ends up at hva_to_pfn_remapped(), which might result in doing the same offset calculation that we do in kvm_arch_prepare_memory_region(), but it also considers other scenarios.
Even if we analyze all that and convince oursleves it's always all the same on arm64, the two code paths could change, leading to really hard to debug differing behavior, and nobody will actively keep the two paths in sync. I'd be fine with keeping the performance optimization if we have good grounds for that though, and using the same translation mechanism for VM_PFNMAP as user_mem_abort.
Am I missing something?
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
virt/kvm/arm/mmu.c +0
-0!fault_supports_stage2_huge_mapping(memslot, hva, vma_pagesize)) {virt/kvm/arm/mmu.c +0
-0force_pte = true; vma_pagesize =PAGE_SIZE ;@@ -1760,6 +1766,9 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, writable = false; }
Quoted 9 lines
Quoted 6 lines
virt/kvm/arm/mmu.c static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, · line 1771
Quoted 3 lines
virt/kvm/arm/mmu.c +3
-0+if (exec_fault &&is_iomap (flags)) +return -ENOEXEC ; +nit: why don't you just do this when checking
kvm_is_device_pfn()and avoid having logic in two places to deal with this case?Good point. I've already sent the PR, but that could be a further cleanup.
Sure, I can have a look when we agree on the above.
Quoted 55 lines
Quoted 41 lines
Quoted 41 lines
virt/kvm/arm/mmu.c +0
-0spin_lock (&kvm->mmu_lock);if (mmu_notifier_retry (kvm, mmu_seq))virt/kvm/arm/mmu.c +1
-1goto out_unlock;@@ -1781,7 +1790,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, if (writable)kvm_set_pfn_dirty (pfn); -if (fault_status !=FSC_PERM ) +if (fault_status !=FSC_PERM && !is_iomap (flags))clean_dcache_guest_page (pfn, vma_pagesize);virt/kvm/arm/mmu.c +2
-3if (exec_fault)@@ -1948,9 +1957,8 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)if (kvm_is_error_hva (hva) || (write_fault && !writable)) {if (is_iabt) {/* Prefetch Abort on I/O address */ -kvm_inject_pabt (vcpu,kvm_vcpu_get_hfar (vcpu)); - ret =1 ; - goto out_unlock; + ret = -ENOEXEC ; + goto out; }virt/kvm/arm/mmu.c +5
-0/* @@ -1992,6 +2000,11 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run) ret =user_mem_abort (vcpu, fault_ipa, memslot, hva, fault_status);if (ret ==0 ) ret =1 ; +out: +if (ret == -ENOEXEC ) { +kvm_inject_pabt (vcpu,kvm_vcpu_get_hfar (vcpu)); + ret =1 ; + } out_unlock:srcu_read_unlock (&vcpu->kvm->srcu, idx);return ret;
--Quoted 8 lines
Quoted 2 lines
2.20.1
I can't seem to decide for myself if I think there's a sematic difference between trying to execute from somewhere the VMM has explicitly told us is device memory and from somewhere which we happen to have mapped with
VM_PFNMAPfrom user space. But I also can't seem to really fault it (pun intended). Thoughts?The issue is that the VMM never really tells us whether something is a device mapping or not (the only exception being the GICv2 cpuif). Even with PFNMAP, we guess it (it could well be memory that lives outside of the linear mapping). I don't see a way to lift this ambiguity.
Quoted 4 lines
Ideally, faulting on executing a non-mapping should be offloaded to userspace for emulation, in line with your patches that offload non-emulated data accesses. That'd be a new ABI, and I can't imagine anyone willing to deal with it.
So what I was asking was if it makes sense to report the Prefetch Abort in the case where the VMM has already told us that it doesn't want to register anything backing the IPA (no memslot), and instead return an error to user space, so that it can make a decision (for example inject an external abort, which may have been the right thing to do in the former case as well, but that could be considered ABI now, so let's not kick that hornet's nest).
In any case, no strong feelings here, I just have a vague feeling that injecting more prefetch aborts on execute-from-some-device is not necessarily the right thing to do.
Thanks,
Christoffer
Marc Zyngier
maintainerRe: [PATCH 1/3] KVM: arm/arm64: Properly handle faulting of device mappings
Quoted 124 lines · Christoffer Dall
Quoted 14 lines · Marc Zyngier
Hi Christoffer,
Quoted 4 lines · Christoffer Dall
Hi Marc,
Quoted 1 line · Marc Zyngier
A device mapping is normally always mapped at Stage-2, since
there
Quoted 4 lines
Quoted 1 line
is very little gain in having it faulted in.
It is actually becoming less clear to me what the real benefits of pre-populating the stage 2 page table are, especially given that
we can
Quoted 1 line
provoke a situation where they're faulted in anyhow. Do you
Quoted 8 lines
recall if
Quoted 1 line
we had any specific case that motivated us to pre-fault in the
pages?
It's only a minor performance optimization that was introduced by Ard in
8eef91239e57d. Which makes sense for platform devices that have a single fixed location in memory. It makes slightly less sense for PCI, where you can move things around.User space could still decide to move things around in its VA map even if the device is fixed.
Anyway, I was thinking more if there was some sort of device, like a frambuffer, which for example crosses page boundaries and where it would be visible to the user that there's a sudden performance drop while operating the device over page boundaries. Anything like that?
Quoted 77 lines
Quoted 4 lines
Quoted 4 lines
Nonetheless, it is possible to end-up in a situation where the device
mapping has been removed from Stage-2 (userspace munmaped theVFIO
Quoted 3 lines
Quoted 3 lines
region, and the MMU notifier did its job), but present in a userspace
mapping (userpace has mapped it back at the same address). Insuch
Quoted 1 line
Quoted 1 line
a situation, the device mapping will be demand-paged as the
guest
Quoted 3 lines
Quoted 3 lines
performs memory accesses.
This requires to be careful when dealing with mapping size,
cache
Quoted 1 line
Quoted 1 line
management, and to handle potential execution of a device
mapping.
Quoted 16 lines
Quoted 16 lines
Reported-by Alexandru Elisei <[email protected]>Signed-off-by Marc Zyngier <[email protected]>--- virt/kvm/arm/mmu.c | 21 +++++++++++++++++---- 1 file changed, 17 insertions(+), 4 deletions(-)virt/kvm/arm/mmu.c +0
-0@@ -38,6 +38,11 @@ static unsigned long io_map_base;#define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)#define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1)Quoted 7 lines
virt/kvm/arm/mmu.c static unsigned long io_map_base; · line 45
Quoted 5 lines
virt/kvm/arm/mmu.c +5
-0+static bool is_iomap (unsigned long flags) +{ +return flags &KVM_S2PTE_FLAG_IS_IOMAP ; +} +nit: I'm not really sure this indirection makes the code more
readable,
Quoted 15 lines
but I guess that's a matter of taste.
virt/kvm/arm/mmu.c static int user_mem_abort(struct kvm_vcpu · line 1707
Quoted 10 lines
virt/kvm/arm/mmu.c +0
-0static bool memslot_is_logging (struct kvm_memory_slot *memslot) { return memslot->dirty_bitmap && !(memslot->flags &
KVM_MEM_READONLY);virt/kvm/arm/mmu.c +1
-0@@ -1698,6 +1703,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, vma_pagesize =vma_kernel_pagesize (vma); if (logging_active || + (vma->vm_flags & VM_PFNMAP) ||WHat is actually the rationale for this?
Quoted 3 lines
Why is a huge mapping not permitted to device memory?
Are we guaranteed that
VM_PFNMAPon the vma results in devicemappings?
Quoted 1 line
I'm not convinced this is the case, and it would be better if we
can
Quoted 1 line
stick to a single primitive (either
kvm_is_device_pfn, or
VM_PFNMAP) toQuoted 1 line
detect device mappings.
For now, I've tried to keep the two paths that deal with mapping devices (or rather, things that we interpret as devices) as close as possible. If we drop the "eager" mapping, then we're at liberty to restructure this in creative ways.
This includes potential huge mappings, but I'm not sure the rest of the kernel uses them for devices anyway (I need to find out).
Quoted 1 line
As a subsequent patch, I'd like to make sure that at the very
least our
Quoted 1 line
memslot prepare function follows the exact same logic for mapping
Quoted 9 lines
device
Quoted 1 line
memory as a fault-in approach does, or that we simply always fault
pages
Quoted 1 line
in.
As far as I can see, the two approach are now identical. Am I missing something?
And yes, getting rid of the eager mapping works for me.As far as I can tell, our
user_mem_abort()usesgfn_to_pfn_prot()which goes doesn a long trail which ends up athva_to_pfn_remapped(), which might result in doing the same offset calculation that we do inkvm_arch_prepare_memory_region(), but it also considers other scenarios.
Quoted 8 lines
Even if we analyze all that and convince oursleves it's always all the same on arm64, the two code paths could change, leading to really hard to debug differing behavior, and nobody will actively keep the two paths in sync. I'd be fine with keeping the performance optimization if we have good grounds for that though, and using the same translation mechanism for
VM_PFNMAPasuser_mem_abort.Am I missing something?
I'm not disputing any of the above. I'm only trying to keep this patch minimal so that we can easily backport it (although it is arguable that deleting code isn't that big a deal).
[...]
Quoted 26 lines
Quoted 13 lines
Quoted 3 lines
I can't seem to decide for myself if I think there's a sematic difference between trying to execute from somewhere the VMM has explicitly told us is device memory and from somewhere which we
happen
Quoted 1 line
to have mapped with
VM_PFNMAPfrom user space. But I also can'tseem to
Quoted 1 line
really fault it (pun intended). Thoughts?
The issue is that the VMM never really tells us whether something is a device mapping or not (the only exception being the GICv2 cpuif). Even with PFNMAP, we guess it (it could well be memory that lives outside of the linear mapping). I don't see a way to lift this ambiguity.
Quoted 4 lines
Ideally, faulting on executing a non-mapping should be offloaded to userspace for emulation, in line with your patches that offload non-emulated data accesses. That'd be a new ABI, and I can't imagine anyone willing to deal with it.
So what I was asking was if it makes sense to report the Prefetch Abort in the case where the VMM has already told us that it doesn't want to register anything backing the IPA (no memslot), and instead return an error to user space, so that it can make a decision (for example inject an external abort, which may have been the right thing to do in the former case as well, but that could be considered ABI now, so let's not kick that hornet's nest).
Quoted 3 lines
In any case, no strong feelings here, I just have a vague feeling that injecting more prefetch aborts on execute-from-some-device is not necessarily the right thing to do.
The ARMv8 ARM has the following stuff in B2.7.2 (Device Memory):
<quote>
Hardware does not prevent speculative instruction fetches from a memory location with any of the Device memory attributes unless the memory location is also marked as Execute-never for all Exception levels.
Note
This means that to prevent speculative instruction fetches from memory locations with Device memory attributes, any location that is assigned any Device memory type must also be marked as Execute-never for all Exception levels. Failure to mark a memory location with any Device memory attribute as Execute-never for all Exception levels is a programming error.
</quote>
and
<quote>
For instruction fetches, if branches cause the program counter to point to an area of memory with the Device attribute which is not marked as Execute-never for the current Exception level, an implementation can either:
- Treat the instruction fetch as if it were to a memory location with the Normal Non-cacheable attribute.
- Take a Permission fault.
</quote>
My reading here is that a prefetch abort is the right thing to do. What we don't do correctly is that we qualify it as an external abort instead of a permission fault (which is annoying as it requires us to find out about the S1 translation level).
Happy to revisit this once we get a S1 PTW.
M.
Signature · 1 line
Jazz is not dead. It just smells funny...
Re: [PATCH 1/3] KVM: arm/arm64: Properly handle faulting of device mappings
Quoted 139 lines · Marc Zyngier
Quoted 138 lines · Christoffer Dall
Quoted 18 lines · Marc Zyngier
Hi Christoffer,
Quoted 4 lines · Christoffer Dall
Hi Marc,
Quoted 1 line · Marc Zyngier
A device mapping is normally always mapped at Stage-2, since
there
Quoted 4 lines
Quoted 1 line
is very little gain in having it faulted in.
It is actually becoming less clear to me what the real benefits of pre-populating the stage 2 page table are, especially given that
we can
Quoted 1 line
provoke a situation where they're faulted in anyhow. Do you
recall if
Quoted 1 line
we had any specific case that motivated us to pre-fault in the
pages?
Quoted 7 lines
It's only a minor performance optimization that was introduced by Ard in
8eef91239e57d. Which makes sense for platform devices that have a single
fixed location in memory. It makes slightly less sense for PCI, where
you can move things around.User space could still decide to move things around in its VA map even if the device is fixed.
Anyway, I was thinking more if there was some sort of device, like a frambuffer, which for example crosses page boundaries and where it would be visible to the user that there's a sudden performance drop while operating the device over page boundaries. Anything like that?
Quoted 81 lines
Quoted 4 lines
Quoted 4 lines
Nonetheless, it is possible to end-up in a situation where the device
mapping has been removed from Stage-2 (userspace munmaped theVFIO
Quoted 3 lines
Quoted 3 lines
region, and the MMU notifier did its job), but present in a userspace
mapping (userpace has mapped it back at the same address). Insuch
Quoted 1 line
Quoted 1 line
a situation, the device mapping will be demand-paged as the
guest
Quoted 3 lines
Quoted 3 lines
performs memory accesses.
This requires to be careful when dealing with mapping size,
cache
Quoted 1 line
Quoted 1 line
management, and to handle potential execution of a device
mapping.
Quoted 16 lines
Quoted 16 lines
Reported-by Alexandru Elisei <[email protected]>Signed-off-by Marc Zyngier <[email protected]>--- virt/kvm/arm/mmu.c | 21 +++++++++++++++++---- 1 file changed, 17 insertions(+), 4 deletions(-)virt/kvm/arm/mmu.c +0
-0@@ -38,6 +38,11 @@ static unsigned long io_map_base;#define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)#define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1)Quoted 7 lines
virt/kvm/arm/mmu.c static unsigned long io_map_base; · line 45
Quoted 5 lines
virt/kvm/arm/mmu.c +5
-0+static bool is_iomap (unsigned long flags) +{ +return flags &KVM_S2PTE_FLAG_IS_IOMAP ; +} +nit: I'm not really sure this indirection makes the code more
readable,
Quoted 15 lines
but I guess that's a matter of taste.
virt/kvm/arm/mmu.c static int user_mem_abort(struct kvm_vcpu · line 1707
Quoted 10 lines
virt/kvm/arm/mmu.c +0
-0static bool memslot_is_logging (struct kvm_memory_slot *memslot) { return memslot->dirty_bitmap && !(memslot->flags &
KVM_MEM_READONLY);virt/kvm/arm/mmu.c +1
-0@@ -1698,6 +1703,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, vma_pagesize =vma_kernel_pagesize (vma); if (logging_active || + (vma->vm_flags & VM_PFNMAP) ||WHat is actually the rationale for this?
Quoted 3 lines
Why is a huge mapping not permitted to device memory?
Are we guaranteed that
VM_PFNMAPon the vma results in devicemappings?
Quoted 1 line
I'm not convinced this is the case, and it would be better if we
can
Quoted 1 line
stick to a single primitive (either
kvm_is_device_pfn, or
VM_PFNMAP) toQuoted 1 line
detect device mappings.
For now, I've tried to keep the two paths that deal with mapping devices
(or rather, things that we interpret as devices) as close as possible.
If we drop the "eager" mapping, then we're at liberty to restructure this in creative ways.This includes potential huge mappings, but I'm not sure the rest of the
kernel uses them for devices anyway (I need to find out).Quoted 1 line
As a subsequent patch, I'd like to make sure that at the very
least our
Quoted 1 line
memslot prepare function follows the exact same logic for mapping
device
Quoted 9 lines
Quoted 1 line
memory as a fault-in approach does, or that we simply always fault
pages
Quoted 1 line
in.
As far as I can see, the two approach are now identical. Am I missing
something?
And yes, getting rid of the eager mapping works for me.As far as I can tell, our
user_mem_abort()usesgfn_to_pfn_prot()which goes doesn a long trail which ends up athva_to_pfn_remapped(), which might result in doing the same offset calculation that we do inkvm_arch_prepare_memory_region(), but it also considers other scenarios.Even if we analyze all that and convince oursleves it's always all the same on arm64, the two code paths could change, leading to really hard to debug differing behavior, and nobody will actively keep the two paths in sync. I'd be fine with keeping the performance optimization if we have good grounds for that though, and using the same translation mechanism for
VM_PFNMAPasuser_mem_abort.
Quoted 5 lines
Quoted 1 line
Am I missing something?
I'm not disputing any of the above. I'm only trying to keep this patch minimal so that we can easily backport it (although it is arguable that deleting code isn't that big a deal).
Yes, sorry, I wasn't arguing we should change the patch, only what the direction for the future should be. Sorry for being unclear.
Quoted 65 lines
[...]
Quoted 28 lines
Quoted 15 lines
Quoted 3 lines
I can't seem to decide for myself if I think there's a sematic difference between trying to execute from somewhere the VMM has explicitly told us is device memory and from somewhere which we
happen
Quoted 1 line
to have mapped with
VM_PFNMAPfrom user space. But I also can'tseem to
Quoted 1 line
really fault it (pun intended). Thoughts?
The issue is that the VMM never really tells us whether something is a
device mapping or not (the only exception being the GICv2 cpuif). Even
with PFNMAP, we guess it (it could well be memory that lives outside of the linear mapping). I don't see a way to lift this ambiguity.Quoted 4 lines
Ideally, faulting on executing a non-mapping should be offloaded to userspace for emulation, in line with your patches that offload non-emulated data accesses. That'd be a new ABI, and I can't imagine anyone willing to deal with it.
So what I was asking was if it makes sense to report the Prefetch Abort in the case where the VMM has already told us that it doesn't want to register anything backing the IPA (no memslot), and instead return an error to user space, so that it can make a decision (for example inject an external abort, which may have been the right thing to do in the former case as well, but that could be considered ABI now, so let's not kick that hornet's nest).
Quoted 3 lines
In any case, no strong feelings here, I just have a vague feeling that injecting more prefetch aborts on execute-from-some-device is not necessarily the right thing to do.
The ARMv8 ARM has the following stuff in B2.7.2 (Device Memory):
<quote>
Hardware does not prevent speculative instruction fetches from a memory location with any of the Device
memory attributes unless the memory location is also marked as Execute-never for all Exception levels.Note
This means that to prevent speculative instruction fetches from memory locations with Device memory
attributes, any location that is assigned any Device memory type must also be marked as Execute-never for
all Exception levels. Failure to mark a memory location with any Device memory attribute as Execute-never
for all Exception levels is a programming error.
</quote>and
<quote>
For instruction fetches, if branches cause the program counter to point to an area of memory with the Device
attribute which is not marked as Execute-never for the current Exception level, an implementation can either:- Treat the instruction fetch as if it were to a memory location with the Normal Non-cacheable attribute.
Quoted 7 lines
- Take a Permission fault.
</quote>My reading here is that a prefetch abort is the right thing to do. What we don't do correctly is that we qualify it as an external abort instead of a permission fault (which is annoying as it requires us to find out about the S1 translation level).
I did not remember we have the permission fault option as a valid implementation option. In that case, never mind my ramblings.
Thanks,
Christoffer
Marc Zyngier
maintainerPATCH 2/3 KVM: arm/arm64: Re-check VMA on detecting a poisoned page
In reply to Marc ZyngierWhen we check for a poisoned page, we use the VMA to tell userspace about the looming disaster. But we pass a pointer to this VMA after having released the mmap_sem, which isn't a good idea.
Instead, re-check that we have still have a VMA, and that this VMA still points to a poisoned page. If the VMA isn't there, userspace is playing with our nerves, so lety's give it a -EFAULT (it deserves it). If the PFN isn't poisoned anymore, let's restart from the top and handle the fault again.
--- virt/kvm/arm/mmu.c | 25 +++++++++++++++++++++++-- 1 file changed, 23 insertions(+), 2 deletions(-)
virt/kvm/arm/mmu.c +23 -2
@@ -1741,9 +1741,30 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, pfn =gfn_to_pfn_prot (kvm, gfn, write_fault, &writable);if (pfn ==KVM_PFN_ERR_HWPOISON ) { -kvm_send_hwpoison_signal (hva, vma); -return 0 ; +/* +* Search for the VMA again, as it may have been +* removed in the interval... +*/ +down_read (¤t->mm->mmap_sem); + vma =find_vma_intersection (current->mm, hva, hva +1 ); +if (vma) { +/* +* Recheck for a poisoned page. If something changed +* behind our back, don't do a thing and take the +* fault again. +*/ + pfn =gfn_to_pfn_prot (kvm, gfn, write_fault, &writable); +if (pfn ==KVM_PFN_ERR_HWPOISON ) +kvm_send_hwpoison_signal (hva, vma); + + ret =0 ; + }else { + ret = -EFAULT ; + } +up_read (¤t->mm->mmap_sem); +return ret; } +if (is_error_noslot_pfn (pfn))return -EFAULT ;
Signature · 1 line
2.20.1
Marc Zyngier
maintainerRe: [PATCH 2/3] KVM: arm/arm64: Re-check VMA on detecting a poisoned page
Quoted 42 lines · Marc Zyngier
When we check for a poisoned page, we use the VMA to tell userspace about the looming disaster. But we pass a pointer to this VMA after having released the
mmap_sem, which isn't a good idea.Instead, re-check that we have still have a VMA, and that this VMA still points to a poisoned page. If the VMA isn't there, userspace is playing with our nerves, so lety's give it a
-EFAULT(it deserves it). If the PFN isn't poisoned anymore, let's restart from the top and handle the fault again.Signed-off-by Marc Zyngier <[email protected]>--- virt/kvm/arm/mmu.c | 25 +++++++++++++++++++++++-- 1 file changed, 23 insertions(+), 2 deletions(-)virt/kvm/arm/mmu.c +16
-2@@ -1741,9 +1741,30 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, pfn =gfn_to_pfn_prot (kvm, gfn, write_fault, &writable);if (pfn ==KVM_PFN_ERR_HWPOISON ) { -kvm_send_hwpoison_signal (hva, vma); -return 0 ; +/* +* Search for the VMA again, as it may have been +* removed in the interval... +*/ +down_read (¤t->mm->mmap_sem); + vma =find_vma_intersection (current->mm, hva, hva +1 ); +if (vma) { +/* +* Recheck for a poisoned page. If something changed +* behind our back, don't do a thing and take the +* fault again. +*/ + pfn =gfn_to_pfn_prot (kvm, gfn, write_fault, &writable); +if (pfn ==KVM_PFN_ERR_HWPOISON ) +kvm_send_hwpoison_signal (hva, vma); +
virt/kvm/arm/mmu.c static int user_mem_abort(struct kvm_vcpu · line 1770
Quoted 10 lines
virt/kvm/arm/mmu.c +7
-0+ ret =0 ; + }else { + ret = -EFAULT ; + } +up_read (¤t->mm->mmap_sem); +return ret; } +if (is_error_noslot_pfn (pfn))return -EFAULT ;
Revisiting this, I wonder if we're not better off just holding the mmap_sem for a bit longer. Something like:
virt/kvm/arm/mmu.c +7 -3
@@ -1719,13 +1719,13 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,if (vma_pagesize ==PMD_SIZE || (vma_pagesize ==PUD_SIZE &&kvm_stage2_has_pmd (kvm))) gfn = (fault_ipa &huge_page_mask (hstate_vma (vma))) >>PAGE_SHIFT ; -up_read (¤t->mm->mmap_sem); -/* We need minimum second+third level pages */ ret =mmu_topup_memory_cache (memcache,kvm_mmu_cache_min_pages (kvm),KVM_NR_MEM_OBJS ); -if (ret) +if (ret) { +up_read (¤t->mm->mmap_sem);return ret; + } mmu_seq = vcpu->kvm->mmu_notifier_seq;/*
@@ -1742,8 +1742,12 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, pfn =gfn_to_pfn_prot (kvm, gfn, write_fault, &writable);if (pfn ==KVM_PFN_ERR_HWPOISON ) {kvm_send_hwpoison_signal (hva, vma); +up_read (¤t->mm->mmap_sem);return 0 ; } + +up_read (¤t->mm->mmap_sem); +if (is_error_noslot_pfn (pfn))return -EFAULT ;
James, what do you think?
M.
Signature · 1 line
Jazz is not dead. It just smells funny...
Re: [PATCH 2/3] KVM: arm/arm64: Re-check VMA on detecting a poisoned page
Hi Marc,
Quoted 4 lines · Marc Zyngier
Quoted 3 lines · Marc Zyngier
When we check for a poisoned page, we use the VMA to tell userspace about the looming disaster. But we pass a pointer to this VMA after having released the
mmap_sem, which isn't a good idea.
Sounds like a bug! The vma-size might not match the poisoned pfn.
Quoted 9 lines
Quoted 9 lines
Instead, re-check that we have still have a VMA, and that this VMA still points to a poisoned page. If the VMA isn't there, userspace is playing with our nerves, so lety's give it a
-EFAULT(it deserves it). If the PFN isn't poisoned anymore, let's restart from the top and handle the fault again.virt/kvm/arm/mmu.c | 25 +++++++++++++++++++++++--
1 file changed, 23 insertions(+), 2 deletions(-)
... yeah ...
Quoted 79 lines
Quoted 27 lines
virt/kvm/arm/mmu.c +16
-2@@ -1741,9 +1741,30 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable); if (pfn == KVM_PFN_ERR_HWPOISON) { - kvm_send_hwpoison_signal(hva, vma); - return 0; + /* + * Search for the VMA again, as it may have been + * removed in the interval... + */ + down_read(¤t->mm->mmap_sem); + vma = find_vma_intersection(current->mm, hva, hva + 1); + if (vma) { + /* + * Recheck for a poisoned page. If something changed + * behind our back, don't do a thing and take the + * fault again. + */ + pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable); + if (pfn == KVM_PFN_ERR_HWPOISON) + kvm_send_hwpoison_signal(hva, vma); +virt/kvm/arm/mmu.c static int user_mem_abort(struct kvm_vcpu · line 1770
Quoted 10 lines
virt/kvm/arm/mmu.c +7
-0+ ret = 0; + } else { + ret = -EFAULT; + } + up_read(¤t->mm->mmap_sem); + return ret; } + if (is_error_noslot_pfn(pfn)) return -EFAULT;Revisiting this, I wonder if we're not better off just holding the
mmap_semfor a bit longer. Something like:virt/kvm/arm/mmu.c +0
-0@@ -1719,13 +1719,13 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,if (
vma_pagesize==PMD_SIZE||
(vma_pagesize==PUD_SIZE&&kvm_stage2_has_pmd(kvm))) gfn = (fault_ipa& huge_page_mask(hstate_vma(vma))) >>PAGE_SHIFT; -up_read(¤t->mm->mmap_sem);
-
/* We need minimum second+third level pages */
ret =mmu_topup_memory_cache(memcache, kvm_mmu_cache_min_pages(kvm), KVM_NR_MEM_OBJS);
- if (ret)
+ if (ret) {
+up_read(¤t->mm->mmap_sem);
return ret;
+ }
mmu_seq=vcpu->kvm->mmu_notifier_seq;
/*virt/kvm/arm/mmu.c +4
-0@@ -1742,8 +1742,12 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable); if (pfn == KVM_PFN_ERR_HWPOISON) { kvm_send_hwpoison_signal(hva, vma); + up_read(¤t->mm->mmap_sem); return 0; } + + up_read(¤t->mm->mmap_sem); + if (is_error_noslot_pfn(pfn)) return -EFAULT;
Quoted 1 line
James, what do you think?
(allocating from a kmemcache while holding current's mmap_sem. I don't want to think about it!)
Can we be lazier? We want the VMA to get the size of the poisoned mapping correct in the signal. The bug is that this could change when we drop the lock, before queuing the signal, so we report hwpoison on old-vmas:pfn with new-vmas:size.
Can't it equally change when we drop the lock after queuing the signal? Any time before the thread returns to user-space to take the signal gives us a stale value.
I think all that matters is the size goes with the pfn that was poisoned. If we look the vma up by hva again, we have to check if the pfn has changed too... (which you are doing)
Can we stash the size in the existing mmap_sem region, and use that in kvm_send_hwpoison_signal()? We know it matches the pfn we saw as poisoned.
The vma could be changed before/after we send the signal, but user-space can't know which. This is user-spaces' problem for messing with the memslots while a vpcu is running.
How about (untested):
-------------------------%<-------------------------
virt/kvm/arm/mmu.c +0 -0
@@ -1591,16 +1591,8 @@ static void invalidate_icache_guest_page(kvm_pfn_t pfn, unsigned
long size)
__invalidate_icache_guest_page(pfn, size);
}
-static void kvm_send_hwpoison_signal(unsigned long address,
- struct vm_area_struct *vma)
+static void kvm_send_hwpoison_signal(unsigned long address, short lsb)
{
- short lsb;
-
- if (is_vm_hugetlb_page(vma))
- lsb = huge_page_shift(hstate_vma(vma));
- else
- lsb = PAGE_SHIFT;
-
send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
}
@@ -1673,6 +1665,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
struct kvm *kvm = vcpu->kvm;
struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
struct vm_area_struct *vma;
+ short stage1_vma_size;
kvm_pfn_t pfn;
pgprot_t mem_type = PAGE_S2;
bool logging_active = memslot_is_logging(memslot);@@ -1703,6 +1696,12 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
vma_pagesize = PAGE_SIZE;
}+ /* For signals due to hwpoison, we need to use the stage1 size */
+ if (is_vm_hugetlb_page(vma))
+ stage1_vma_size = huge_page_shift(hstate_vma(vma));
+ else
+ stage1_vma_size = PAGE_SHIFT;
+
/*
* The stage2 has a minimum of 2 level table (For arm64 see
* kvm_arm_setup_stage2()). Hence, we are guaranteed that we can
@@ -1735,7 +1734,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
if (pfn == KVM_PFN_ERR_HWPOISON) {
- kvm_send_hwpoison_signal(hva, vma);
+ kvm_send_hwpoison_signal(hva, stage1_vma_size);
return 0;
}
if (is_error_noslot_pfn(pfn))-------------------------%<-------------------------
Its possible this could even be the original output of vma_kernel_pagesize()... (Punit supplied the original huge_page_shift(hstate_vma()) runes...)
Thanks,
James
Marc Zyngier
maintainerRe: [PATCH 2/3] KVM: arm/arm64: Re-check VMA on detecting a poisoned page
Hi James,
Quoted 23 lines · James Morse
Hi Marc,
Quoted 4 lines · Marc Zyngier
Quoted 3 lines · Marc Zyngier
When we check for a poisoned page, we use the VMA to tell userspace about the looming disaster. But we pass a pointer to this VMA after having released the
mmap_sem, which isn't a good idea.Sounds like a bug! The vma-size might not match the poisoned pfn.
Quoted 9 lines
Quoted 9 lines
Instead, re-check that we have still have a VMA, and that this VMA still points to a poisoned page. If the VMA isn't there, userspace is playing with our nerves, so lety's give it a
-EFAULT(it deserves it). If the PFN isn't poisoned anymore, let's restart from the top and handle the fault again.virt/kvm/arm/mmu.c | 25 +++++++++++++++++++++++--
1 file changed, 23 insertions(+), 2 deletions(-)... yeah ...
[...]
Quoted 67 lines
How about (untested):
-------------------------%<-------------------------virt/kvm/arm/mmu.c +0
-0@@ -1591,16 +1591,8 @@ static void
invalidate_icache_guest_page(kvm_pfn_t pfn, unsigned long size)__invalidate_icache_guest_page(pfn, size);}
virt/kvm/arm/mmu.c +1
-2-static void kvm_send_hwpoison_signal (unsigned long address, -struct vm_area_struct *vma) +static void kvm_send_hwpoison_signal(unsigned long address, shortlsb)
{virt/kvm/arm/mmu.c +0
-7-short lsb; - -if (is_vm_hugetlb_page (vma)) - lsb =huge_page_shift (hstate_vma (vma)); -else - lsb =PAGE_SHIFT ; - send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb,current);
}virt/kvm/arm/mmu.c +1
-0@@ -1673,6 +1665,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, struct kvm *kvm = vcpu->kvm;struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;struct vm_area_struct *vma; +short stage1_vma_size;kvm_pfn_t pfn;pgprot_t mem_type = PAGE_S2; bool logging_active = memslot_is_logging(memslot);virt/kvm/arm/mmu.c +1
-0@@ -1703,6 +1696,12 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, vma_pagesize =PAGE_SIZE ; } +/* For signals due to hwpoison, we need to use the stage1 size */
virt/kvm/arm/mmu.c +5
-0+if (is_vm_hugetlb_page (vma)) + stage1_vma_size =huge_page_shift (hstate_vma (vma)); +else + stage1_vma_size =PAGE_SHIFT ; +/* * The stage2 has a minimum of 2 level table (For arm64 see * kvm_arm_setup_stage2()). Hence, we are guaranteed that we can
virt/kvm/arm/mmu.c +1
-1@@ -1735,7 +1734,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, pfn =gfn_to_pfn_prot (kvm, gfn, write_fault, &writable);if (pfn ==KVM_PFN_ERR_HWPOISON ) { -kvm_send_hwpoison_signal (hva, vma); +kvm_send_hwpoison_signal (hva, stage1_vma_size);return 0 ; }if (
is_error_noslot_pfn(pfn))
-------------------------%<-------------------------
Quoted 3 lines
Its possible this could even be the original output of
vma_kernel_pagesize()... (Punit supplied the original huge_page_shift(hstate_vma()) runes...)
I'd be happy with something along these lines. Any chance you could a proper patch?
Thanks,
M.
Signature · 1 line
Jazz is not dead. It just smells funny...
Re: [PATCH 2/3] KVM: arm/arm64: Re-check VMA on detecting a poisoned page
In reply to James MorseHi James,
Quoted 4 lines · James Morse
Hi Marc,
Quoted 1 line · Marc Zyngier
On 2019-12-11 16:56, Marc Zyngier wrote:
[...]
Quoted 11 lines
(allocating from a kmemcache while holding current's
mmap_sem. I don't want to think about it!)Can we be lazier? We want the VMA to get the size of the poisoned mapping correct in the signal. The bug is that this could change when we drop the lock, before queuing the signal, so we report hwpoison on old-vmas:pfn with new-vmas:size.
Can't it equally change when we drop the lock after queuing the signal? Any time before the thread returns to user-space to take the signal gives us a stale value.
Quoted 9 lines
I think all that matters is the size goes with the pfn that was poisoned. If we look the vma up by hva again, we have to check if the pfn has changed too... (which you are doing)
Can we stash the size in the existing
mmap_semregion, and use that inkvm_send_hwpoison_signal()? We know it matches the pfn we saw as poisoned.The vma could be changed before/after we send the signal, but user-space can't know which. This is user-spaces' problem for messing with the memslots while a vpcu is running.
(I should clearly have expanded this thread before I replied to the original patch...)
Quoted 36 lines
How about (untested):
-------------------------%<-------------------------virt/kvm/arm/mmu.c +2
-9@@ -1591,16 +1591,8 @@ static void invalidate_icache_guest_page(kvm_pfn_t pfn, unsigned long size)__invalidate_icache_guest_page (pfn ,size ); } -static void kvm_send_hwpoison_signal (unsigned long address, -struct vm_area_struct *vma) +static void kvm_send_hwpoison_signal (unsigned long address,short lsb) { -short lsb; - -if (is_vm_hugetlb_page (vma)) - lsb =huge_page_shift (hstate_vma (vma)); -else - lsb =PAGE_SHIFT ; -send_sig_mceerr (BUS_MCEERR_AR , (void __user *)address, lsb, current); }@@ -1673,6 +1665,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,struct kvm *kvm = vcpu->kvm;struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;struct vm_area_struct *vma; +short stage1_vma_size;kvm_pfn_t pfn;pgprot_t mem_type =PAGE_S2 ;bool logging_active =memslot_is_logging (memslot);
virt/kvm/arm/mmu.c static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, · line 1704
Quoted 9 lines
virt/kvm/arm/mmu.c +6
-0@@ -1703,6 +1696,12 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, vma_pagesize =PAGE_SIZE ; } +/* For signals due to hwpoison, we need to use the stage1 size */ +if (is_vm_hugetlb_page (vma)) + stage1_vma_size =huge_page_shift (hstate_vma (vma)); +else + stage1_vma_size =PAGE_SHIFT ; +
But (see my patch) as far as I can tell, this is already what we have in vma_pagesize, and do we really have to provide the stage 1 size to user space if the fault happened within a smaller boundary? Isn't that just providing more precise information to the user?
Thanks,
Christoffer
Re: [PATCH 2/3] KVM: arm/arm64: Re-check VMA on detecting a poisoned page
In reply to Marc ZyngierQuoted 55 lines · Marc Zyngier
When we check for a poisoned page, we use the VMA to tell userspace about the looming disaster. But we pass a pointer to this VMA after having released the
mmap_sem, which isn't a good idea.Instead, re-check that we have still have a VMA, and that this VMA still points to a poisoned page. If the VMA isn't there, userspace is playing with our nerves, so lety's give it a
-EFAULT(it deserves it). If the PFN isn't poisoned anymore, let's restart from the top and handle the fault again.Signed-off-by Marc Zyngier <[email protected]>--- virt/kvm/arm/mmu.c | 25 +++++++++++++++++++++++-- 1 file changed, 23 insertions(+), 2 deletions(-)virt/kvm/arm/mmu.c +23
-2@@ -1741,9 +1741,30 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, pfn =gfn_to_pfn_prot (kvm, gfn, write_fault, &writable);if (pfn ==KVM_PFN_ERR_HWPOISON ) { -kvm_send_hwpoison_signal (hva, vma); -return 0 ; +/* +* Search for the VMA again, as it may have been +* removed in the interval... +*/ +down_read (¤t->mm->mmap_sem); + vma =find_vma_intersection (current->mm, hva, hva +1 ); +if (vma) { +/* +* Recheck for a poisoned page. If something changed +* behind our back, don't do a thing and take the +* fault again. +*/ + pfn =gfn_to_pfn_prot (kvm, gfn, write_fault, &writable); +if (pfn ==KVM_PFN_ERR_HWPOISON ) +kvm_send_hwpoison_signal (hva, vma); + + ret =0 ; + }else { + ret = -EFAULT ; + } +up_read (¤t->mm->mmap_sem); +return ret; } +if (is_error_noslot_pfn (pfn))return -EFAULT ;Signature · 1 line
2.20.1
If I read this code correctly, then all we use the VMA for is to find the page size used by the MMU to back the VMA, which we've already established in the vma_pagesize (and possibly adjusted to something more accurate based on our constraints in stage 2 which generated the error), so all we need is the size and a way to convert that into a shift.
Not being 100% confident about the semantics of the lsb bit we pass to user space (is it indicating the size of the mapping which caused the error or the size of the mapping where user space could potentially trigger an error?), or wheter we care enough at that level, could we consider something like the following instead?
virt/kvm/arm/mmu.c +3 -9
@@ -1592,15 +1592,9 @@ static void invalidate_icache_guest_page(kvm_pfn_t pfn, unsigned long size) }static void kvm_send_hwpoison_signal (unsigned long address, -struct vm_area_struct *vma) +unsigned long vma_pagesize) { -short lsb; - -if (is_vm_hugetlb_page (vma)) - lsb =huge_page_shift (hstate_vma (vma)); -else - lsb =PAGE_SHIFT ; - +short lsb =__ffs (vma_pagesize);send_sig_mceerr (BUS_MCEERR_AR , (void __user *)address, lsb, current); }
@@ -1735,7 +1729,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, pfn =gfn_to_pfn_prot (kvm, gfn, write_fault, &writable);if (pfn ==KVM_PFN_ERR_HWPOISON ) { -kvm_send_hwpoison_signal (hva, vma); +kvm_send_hwpoison_signal (hva, vma_pagesize);return 0 ; }if (is_error_noslot_pfn (pfn))
Thanks,
Christoffer
Re: [PATCH 2/3] KVM: arm/arm64: Re-check VMA on detecting a poisoned page
Hi Christoffer,
Quoted 17 lines · Christoffer Dall
Quoted 9 lines · Marc Zyngier
When we check for a poisoned page, we use the VMA to tell userspace about the looming disaster. But we pass a pointer to this VMA after having released the
mmap_sem, which isn't a good idea.Instead, re-check that we have still have a VMA, and that this VMA still points to a poisoned page. If the VMA isn't there, userspace is playing with our nerves, so lety's give it a
-EFAULT(it deserves it). If the PFN isn't poisoned anymore, let's restart from the top and handle the fault again.If I read this code correctly, then all we use the VMA for is to find the page size used by the MMU to back the VMA, which we've already established in the
vma_pagesize(and possibly adjusted to something more accurate based on our constraints in stage 2 which generated the error), so all we need is the size and a way to convert that into a shift.
Quoted 3 lines
Not being 100% confident about the semantics of the lsb bit we pass to user space (is it indicating the size of the mapping which caused the error or the size of the mapping where user space could potentially
Its the size of the hole that has opened up in the address map. The error was very likely to be much smaller, but all we can do is unmap pages.
Transparent huge-pages are split up to minimise the impact. This code is for hugetlbfs (?), whose pages are dedicated for that use, so don't get split.
arch/arm64/mm/fault.c::do_page_fault() has:
| lsb = PAGE_SHIFT; | if (fault & VM_FAULT_HWPOISON_LARGE) | lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault)); | | arm64_force_sig_mceerr(BUS_MCEERR_AR, (void __user *)addr, lsb,
(I assume its a shift because bytes in the signal union are precious)
Quoted 26 lines
trigger an error?), or wheter we care enough at that level, could we consider something like the following instead?
virt/kvm/arm/mmu.c +2
-8@@ -1592,15 +1592,9 @@ static void invalidate_icache_guest_page(kvm_pfn_t pfn, unsigned long size) }static void kvm_send_hwpoison_signal (unsigned long address, -struct vm_area_struct *vma) +unsigned long vma_pagesize) { -short lsb; - -if (is_vm_hugetlb_page (vma)) - lsb =huge_page_shift (hstate_vma (vma)); -else - lsb =PAGE_SHIFT ; - +short lsb =__ffs (vma_pagesize);send_sig_mceerr (BUS_MCEERR_AR , (void __user *)address, lsb, current); }
virt/kvm/arm/mmu.c static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, · line 1735
Quoted 8 lines
virt/kvm/arm/mmu.c +1
-1@@ -1735,7 +1729,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, pfn =gfn_to_pfn_prot (kvm, gfn, write_fault, &writable);if (pfn ==KVM_PFN_ERR_HWPOISON ) { -kvm_send_hwpoison_signal (hva, vma); +kvm_send_hwpoison_signal (hva, vma_pagesize);return 0 ; }if (is_error_noslot_pfn (pfn))
This changes the meaning, vma_pagesize is a value like 4K, not a shift like 12.
But yes, caching the shift value under the mmap_sem and passing it in is the right-thing-to-do(tm). I have a patch which I'll post, once I remember how to test this thing!
Thanks,
James
Marc Zyngier
maintainerPATCH 3/3 KVM: arm/arm64: Drop spurious message when faulting on a non-existent mapping
In reply to Marc ZyngierShould userspace unmap memory whilst the guest is running, we exit with a -EFAULT, but also having spat a useless message on the console.
Get rid of it.
--- virt/kvm/arm/mmu.c | 1 - 1 file changed, 1 deletion(-)
virt/kvm/arm/mmu.c +0 -1
@@ -1696,7 +1696,6 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,down_read (¤t->mm->mmap_sem); vma =find_vma_intersection (current->mm, hva, hva +1 );if (unlikely (!vma)) { -kvm_err ("Failed to find VMA for hva 0x%lx\n" , hva);up_read (¤t->mm->mmap_sem);return -EFAULT ; }
Signature · 1 line
2.20.1
Re: [PATCH 3/3] KVM: arm/arm64: Drop spurious message when faulting on a non-existent mapping
Quoted 4 lines · Marc Zyngier
Should userspace unmap memory whilst the guest is running, we exit with a
-EFAULT, but also having spat a useless message on the console.Get rid of it.
Quoted 21 lines
Signed-off-by Marc Zyngier <[email protected]>--- virt/kvm/arm/mmu.c | 1 - 1 file changed, 1 deletion(-)virt/kvm/arm/mmu.c +0
-1@@ -1696,7 +1696,6 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,down_read (¤t->mm->mmap_sem); vma =find_vma_intersection (current->mm, hva, hva +1 );if (unlikely (!vma)) { -kvm_err ("Failed to find VMA for hva 0x%lx\n" , hva);up_read (¤t->mm->mmap_sem);return -EFAULT ; }Signature · 1 line
2.20.1