]> git.ipfire.org Git - thirdparty/kernel/stable-queue.git/commitdiff
Fixes for 4.19
authorSasha Levin <sashal@kernel.org>
Mon, 27 Apr 2020 01:14:11 +0000 (21:14 -0400)
committerSasha Levin <sashal@kernel.org>
Mon, 27 Apr 2020 01:14:11 +0000 (21:14 -0400)
Signed-off-by: Sasha Levin <sashal@kernel.org>
12 files changed:
queue-4.19/kvm-fix-compilation-on-aarch64.patch [new file with mode: 0644]
queue-4.19/kvm-fix-compilation-on-s390.patch [new file with mode: 0644]
queue-4.19/kvm-fix-compile-on-s390-part-2.patch [new file with mode: 0644]
queue-4.19/kvm-introduce-a-new-guest-mapping-api.patch [new file with mode: 0644]
queue-4.19/kvm-nvmx-always-sync-guest_bndcfgs-when-it-comes-fro.patch [new file with mode: 0644]
queue-4.19/kvm-properly-check-if-page-is-valid-in-kvm_vcpu_unma.patch [new file with mode: 0644]
queue-4.19/kvm-vmx-zero-out-all-general-purpose-registers-after.patch [new file with mode: 0644]
queue-4.19/series
queue-4.19/x86-kvm-cache-gfn-to-pfn-translation.patch [new file with mode: 0644]
queue-4.19/x86-kvm-clean-up-host-s-steal-time-structure.patch [new file with mode: 0644]
queue-4.19/x86-kvm-introduce-kvm_-un-map_gfn.patch [new file with mode: 0644]
queue-4.19/x86-kvm-make-sure-kvm_vcpu_flush_tlb-flag-is-not-mis.patch [new file with mode: 0644]

diff --git a/queue-4.19/kvm-fix-compilation-on-aarch64.patch b/queue-4.19/kvm-fix-compilation-on-aarch64.patch
new file mode 100644 (file)
index 0000000..17f5474
--- /dev/null
@@ -0,0 +1,56 @@
+From 6b8c0b11f88c59cdbde997e1828fbffe47b3d60c Mon Sep 17 00:00:00 2001
+From: Sasha Levin <sashal@kernel.org>
+Date: Fri, 17 May 2019 14:08:53 +0200
+Subject: kvm: fix compilation on aarch64
+
+From: Paolo Bonzini <pbonzini@redhat.com>
+
+commit c011d23ba046826ccf8c4a4a6c1d01c9ccaa1403 upstream.
+
+Commit e45adf665a53 ("KVM: Introduce a new guest mapping API", 2019-01-31)
+introduced a build failure on aarch64 defconfig:
+
+$ make -j$(nproc) ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- O=out defconfig \
+                Image.gz
+...
+../arch/arm64/kvm/../../../virt/kvm/kvm_main.c:
+    In function '__kvm_map_gfn':
+../arch/arm64/kvm/../../../virt/kvm/kvm_main.c:1763:9: error:
+    implicit declaration of function 'memremap'; did you mean 'memset_p'?
+../arch/arm64/kvm/../../../virt/kvm/kvm_main.c:1763:46: error:
+    'MEMREMAP_WB' undeclared (first use in this function)
+../arch/arm64/kvm/../../../virt/kvm/kvm_main.c:
+    In function 'kvm_vcpu_unmap':
+../arch/arm64/kvm/../../../virt/kvm/kvm_main.c:1795:3: error:
+    implicit declaration of function 'memunmap'; did you mean 'vm_munmap'?
+
+because these functions are declared in <linux/io.h> rather than <asm/io.h>,
+and the former was being pulled in already on x86 but not on aarch64.
+
+Reported-by: Nathan Chancellor <natechancellor@gmail.com>
+Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
+[bwh: Backported to 4.19: adjust context]
+Signed-off-by: Ben Hutchings <ben.hutchings@codethink.co.uk>
+Signed-off-by: Sasha Levin <sashal@kernel.org>
+---
+ virt/kvm/kvm_main.c | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+
+diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
+index ec1479abb29de..4a5ea263edf62 100644
+--- a/virt/kvm/kvm_main.c
++++ b/virt/kvm/kvm_main.c
+@@ -52,9 +52,9 @@
+ #include <linux/sort.h>
+ #include <linux/bsearch.h>
+ #include <linux/kthread.h>
++#include <linux/io.h>
+ #include <asm/processor.h>
+-#include <asm/io.h>
+ #include <asm/ioctl.h>
+ #include <linux/uaccess.h>
+ #include <asm/pgtable.h>
+-- 
+2.20.1
+
diff --git a/queue-4.19/kvm-fix-compilation-on-s390.patch b/queue-4.19/kvm-fix-compilation-on-s390.patch
new file mode 100644 (file)
index 0000000..ecf8edd
--- /dev/null
@@ -0,0 +1,37 @@
+From ad7ab152d6afb9a22b9e2ad364fb77797ed3e0df Mon Sep 17 00:00:00 2001
+From: Sasha Levin <sashal@kernel.org>
+Date: Mon, 20 May 2019 12:06:36 +0200
+Subject: kvm: fix compilation on s390
+
+From: Paolo Bonzini <pbonzini@redhat.com>
+
+commit d30b214d1d0addb7b2c9c78178d1501cd39a01fb upstream.
+
+s390 does not have memremap, even though in this particular case it
+would be useful.
+
+Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
+Signed-off-by: Ben Hutchings <ben.hutchings@codethink.co.uk>
+Signed-off-by: Sasha Levin <sashal@kernel.org>
+---
+ virt/kvm/kvm_main.c | 2 ++
+ 1 file changed, 2 insertions(+)
+
+diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
+index 4a5ea263edf62..f99b99b77a486 100644
+--- a/virt/kvm/kvm_main.c
++++ b/virt/kvm/kvm_main.c
+@@ -1722,8 +1722,10 @@ static int __kvm_map_gfn(struct kvm_memory_slot *slot, gfn_t gfn,
+       if (pfn_valid(pfn)) {
+               page = pfn_to_page(pfn);
+               hva = kmap(page);
++#ifdef CONFIG_HAS_IOMEM
+       } else {
+               hva = memremap(pfn_to_hpa(pfn), PAGE_SIZE, MEMREMAP_WB);
++#endif
+       }
+       if (!hva)
+-- 
+2.20.1
+
diff --git a/queue-4.19/kvm-fix-compile-on-s390-part-2.patch b/queue-4.19/kvm-fix-compile-on-s390-part-2.patch
new file mode 100644 (file)
index 0000000..e1a6d1c
--- /dev/null
@@ -0,0 +1,40 @@
+From 4f5f9ed02efbba0976cf10a6918f1d15e471b9c5 Mon Sep 17 00:00:00 2001
+From: Sasha Levin <sashal@kernel.org>
+Date: Mon, 27 May 2019 10:28:25 +0200
+Subject: kvm: fix compile on s390 part 2
+
+From: Christian Borntraeger <borntraeger@de.ibm.com>
+
+commit eb1f2f387db8c0d084581fb26e7faffde700bc8e upstream.
+
+We also need to fence the memunmap part.
+
+Fixes: e45adf665a53 ("KVM: Introduce a new guest mapping API")
+Fixes: d30b214d1d0a (kvm: fix compilation on s390)
+Cc: Michal Kubecek <mkubecek@suse.cz>
+Cc: KarimAllah Ahmed <karahmed@amazon.de>
+Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
+Signed-off-by: Ben Hutchings <ben.hutchings@codethink.co.uk>
+Signed-off-by: Sasha Levin <sashal@kernel.org>
+---
+ virt/kvm/kvm_main.c | 2 ++
+ 1 file changed, 2 insertions(+)
+
+diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
+index f99b99b77a486..5b949aa273de5 100644
+--- a/virt/kvm/kvm_main.c
++++ b/virt/kvm/kvm_main.c
+@@ -1756,8 +1756,10 @@ void kvm_vcpu_unmap(struct kvm_vcpu *vcpu, struct kvm_host_map *map,
+       if (map->page)
+               kunmap(map->page);
++#ifdef CONFIG_HAS_IOMEM
+       else
+               memunmap(map->hva);
++#endif
+       if (dirty) {
+               kvm_vcpu_mark_page_dirty(vcpu, map->gfn);
+-- 
+2.20.1
+
diff --git a/queue-4.19/kvm-introduce-a-new-guest-mapping-api.patch b/queue-4.19/kvm-introduce-a-new-guest-mapping-api.patch
new file mode 100644 (file)
index 0000000..d9401ae
--- /dev/null
@@ -0,0 +1,163 @@
+From 71b37c00da11fbc6cbc3c7b7bd2a26e5f60d65a1 Mon Sep 17 00:00:00 2001
+From: Sasha Levin <sashal@kernel.org>
+Date: Thu, 31 Jan 2019 21:24:34 +0100
+Subject: KVM: Introduce a new guest mapping API
+
+From: KarimAllah Ahmed <karahmed@amazon.de>
+
+commit e45adf665a53df0db37f784ed87c6b57ddd81885 upstream.
+
+In KVM, specially for nested guests, there is a dominant pattern of:
+
+       => map guest memory -> do_something -> unmap guest memory
+
+In addition to all this unnecessarily noise in the code due to boiler plate
+code, most of the time the mapping function does not properly handle memory
+that is not backed by "struct page". This new guest mapping API encapsulate
+most of this boiler plate code and also handles guest memory that is not
+backed by "struct page".
+
+The current implementation of this API is using memremap for memory that is
+not backed by a "struct page" which would lead to a huge slow-down if it
+was used for high-frequency mapping operations. The API does not have any
+effect on current setups where guest memory is backed by a "struct page".
+Further patches are going to also introduce a pfn-cache which would
+significantly improve the performance of the memremap case.
+
+Signed-off-by: KarimAllah Ahmed <karahmed@amazon.de>
+Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
+Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
+[bwh: Backported to 4.19 as dependency of commit 1eff70a9abd4
+ "x86/kvm: Introduce kvm_(un)map_gfn()"]
+Signed-off-by: Ben Hutchings <ben.hutchings@codethink.co.uk>
+Signed-off-by: Sasha Levin <sashal@kernel.org>
+---
+ include/linux/kvm_host.h | 28 ++++++++++++++++++
+ virt/kvm/kvm_main.c      | 64 ++++++++++++++++++++++++++++++++++++++++
+ 2 files changed, 92 insertions(+)
+
+diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
+index 0f99ecc01bc7d..bef95dba14e8b 100644
+--- a/include/linux/kvm_host.h
++++ b/include/linux/kvm_host.h
+@@ -206,6 +206,32 @@ enum {
+       READING_SHADOW_PAGE_TABLES,
+ };
++#define KVM_UNMAPPED_PAGE     ((void *) 0x500 + POISON_POINTER_DELTA)
++
++struct kvm_host_map {
++      /*
++       * Only valid if the 'pfn' is managed by the host kernel (i.e. There is
++       * a 'struct page' for it. When using mem= kernel parameter some memory
++       * can be used as guest memory but they are not managed by host
++       * kernel).
++       * If 'pfn' is not managed by the host kernel, this field is
++       * initialized to KVM_UNMAPPED_PAGE.
++       */
++      struct page *page;
++      void *hva;
++      kvm_pfn_t pfn;
++      kvm_pfn_t gfn;
++};
++
++/*
++ * Used to check if the mapping is valid or not. Never use 'kvm_host_map'
++ * directly to check for that.
++ */
++static inline bool kvm_vcpu_mapped(struct kvm_host_map *map)
++{
++      return !!map->hva;
++}
++
+ /*
+  * Sometimes a large or cross-page mmio needs to be broken up into separate
+  * exits for userspace servicing.
+@@ -711,7 +737,9 @@ struct kvm_memslots *kvm_vcpu_memslots(struct kvm_vcpu *vcpu);
+ struct kvm_memory_slot *kvm_vcpu_gfn_to_memslot(struct kvm_vcpu *vcpu, gfn_t gfn);
+ kvm_pfn_t kvm_vcpu_gfn_to_pfn_atomic(struct kvm_vcpu *vcpu, gfn_t gfn);
+ kvm_pfn_t kvm_vcpu_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn);
++int kvm_vcpu_map(struct kvm_vcpu *vcpu, gpa_t gpa, struct kvm_host_map *map);
+ struct page *kvm_vcpu_gfn_to_page(struct kvm_vcpu *vcpu, gfn_t gfn);
++void kvm_vcpu_unmap(struct kvm_vcpu *vcpu, struct kvm_host_map *map, bool dirty);
+ unsigned long kvm_vcpu_gfn_to_hva(struct kvm_vcpu *vcpu, gfn_t gfn);
+ unsigned long kvm_vcpu_gfn_to_hva_prot(struct kvm_vcpu *vcpu, gfn_t gfn, bool *writable);
+ int kvm_vcpu_read_guest_page(struct kvm_vcpu *vcpu, gfn_t gfn, void *data, int offset,
+diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
+index 4e499b78569b7..ec1479abb29de 100644
+--- a/virt/kvm/kvm_main.c
++++ b/virt/kvm/kvm_main.c
+@@ -1705,6 +1705,70 @@ struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn)
+ }
+ EXPORT_SYMBOL_GPL(gfn_to_page);
++static int __kvm_map_gfn(struct kvm_memory_slot *slot, gfn_t gfn,
++                       struct kvm_host_map *map)
++{
++      kvm_pfn_t pfn;
++      void *hva = NULL;
++      struct page *page = KVM_UNMAPPED_PAGE;
++
++      if (!map)
++              return -EINVAL;
++
++      pfn = gfn_to_pfn_memslot(slot, gfn);
++      if (is_error_noslot_pfn(pfn))
++              return -EINVAL;
++
++      if (pfn_valid(pfn)) {
++              page = pfn_to_page(pfn);
++              hva = kmap(page);
++      } else {
++              hva = memremap(pfn_to_hpa(pfn), PAGE_SIZE, MEMREMAP_WB);
++      }
++
++      if (!hva)
++              return -EFAULT;
++
++      map->page = page;
++      map->hva = hva;
++      map->pfn = pfn;
++      map->gfn = gfn;
++
++      return 0;
++}
++
++int kvm_vcpu_map(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map)
++{
++      return __kvm_map_gfn(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn, map);
++}
++EXPORT_SYMBOL_GPL(kvm_vcpu_map);
++
++void kvm_vcpu_unmap(struct kvm_vcpu *vcpu, struct kvm_host_map *map,
++                  bool dirty)
++{
++      if (!map)
++              return;
++
++      if (!map->hva)
++              return;
++
++      if (map->page)
++              kunmap(map->page);
++      else
++              memunmap(map->hva);
++
++      if (dirty) {
++              kvm_vcpu_mark_page_dirty(vcpu, map->gfn);
++              kvm_release_pfn_dirty(map->pfn);
++      } else {
++              kvm_release_pfn_clean(map->pfn);
++      }
++
++      map->hva = NULL;
++      map->page = NULL;
++}
++EXPORT_SYMBOL_GPL(kvm_vcpu_unmap);
++
+ struct page *kvm_vcpu_gfn_to_page(struct kvm_vcpu *vcpu, gfn_t gfn)
+ {
+       kvm_pfn_t pfn;
+-- 
+2.20.1
+
diff --git a/queue-4.19/kvm-nvmx-always-sync-guest_bndcfgs-when-it-comes-fro.patch b/queue-4.19/kvm-nvmx-always-sync-guest_bndcfgs-when-it-comes-fro.patch
new file mode 100644 (file)
index 0000000..e6dfe53
--- /dev/null
@@ -0,0 +1,59 @@
+From fffc3105b4d26926bc8141ec363a50ddc8c2fc37 Mon Sep 17 00:00:00 2001
+From: Sasha Levin <sashal@kernel.org>
+Date: Tue, 7 May 2019 09:06:28 -0700
+Subject: KVM: nVMX: Always sync GUEST_BNDCFGS when it comes from vmcs01
+
+From: Sean Christopherson <sean.j.christopherson@intel.com>
+
+commit 3b013a2972d5bc344d6eaa8f24fdfe268211e45f upstream.
+
+If L1 does not set VM_ENTRY_LOAD_BNDCFGS, then L1's BNDCFGS value must
+be propagated to vmcs02 since KVM always runs with VM_ENTRY_LOAD_BNDCFGS
+when MPX is supported.  Because the value effectively comes from vmcs01,
+vmcs02 must be updated even if vmcs12 is clean.
+
+Fixes: 62cf9bd8118c4 ("KVM: nVMX: Fix emulation of VM_ENTRY_LOAD_BNDCFGS")
+Cc: Liran Alon <liran.alon@oracle.com>
+Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
+Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
+[bwh: Backported to 4.19: adjust filename, context]
+Signed-off-by: Ben Hutchings <ben.hutchings@codethink.co.uk>
+Signed-off-by: Sasha Levin <sashal@kernel.org>
+---
+ arch/x86/kvm/vmx.c | 13 ++++++-------
+ 1 file changed, 6 insertions(+), 7 deletions(-)
+
+diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
+index e4d0ad06790e1..ccbddc80ad55f 100644
+--- a/arch/x86/kvm/vmx.c
++++ b/arch/x86/kvm/vmx.c
+@@ -12137,13 +12137,9 @@ static void prepare_vmcs02_full(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12)
+       set_cr4_guest_host_mask(vmx);
+-      if (kvm_mpx_supported()) {
+-              if (vmx->nested.nested_run_pending &&
+-                      (vmcs12->vm_entry_controls & VM_ENTRY_LOAD_BNDCFGS))
+-                      vmcs_write64(GUEST_BNDCFGS, vmcs12->guest_bndcfgs);
+-              else
+-                      vmcs_write64(GUEST_BNDCFGS, vmx->nested.vmcs01_guest_bndcfgs);
+-      }
++      if (kvm_mpx_supported() && vmx->nested.nested_run_pending &&
++          (vmcs12->vm_entry_controls & VM_ENTRY_LOAD_BNDCFGS))
++              vmcs_write64(GUEST_BNDCFGS, vmcs12->guest_bndcfgs);
+       if (enable_vpid) {
+               if (nested_cpu_has_vpid(vmcs12) && vmx->nested.vpid02)
+@@ -12207,6 +12203,9 @@ static int prepare_vmcs02(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12,
+               kvm_set_dr(vcpu, 7, vcpu->arch.dr7);
+               vmcs_write64(GUEST_IA32_DEBUGCTL, vmx->nested.vmcs01_debugctl);
+       }
++      if (kvm_mpx_supported() && (!vmx->nested.nested_run_pending ||
++          !(vmcs12->vm_entry_controls & VM_ENTRY_LOAD_BNDCFGS)))
++              vmcs_write64(GUEST_BNDCFGS, vmx->nested.vmcs01_guest_bndcfgs);
+       if (vmx->nested.nested_run_pending) {
+               vmcs_write32(VM_ENTRY_INTR_INFO_FIELD,
+                            vmcs12->vm_entry_intr_info_field);
+-- 
+2.20.1
+
diff --git a/queue-4.19/kvm-properly-check-if-page-is-valid-in-kvm_vcpu_unma.patch b/queue-4.19/kvm-properly-check-if-page-is-valid-in-kvm_vcpu_unma.patch
new file mode 100644 (file)
index 0000000..86071a5
--- /dev/null
@@ -0,0 +1,38 @@
+From 1291a73c2cf7bdc3be49d99524d6691e486147c9 Mon Sep 17 00:00:00 2001
+From: Sasha Levin <sashal@kernel.org>
+Date: Wed, 10 Jul 2019 11:13:13 +0200
+Subject: KVM: Properly check if "page" is valid in kvm_vcpu_unmap
+
+From: KarimAllah Ahmed <karahmed@amazon.de>
+
+commit b614c6027896ff9ad6757122e84760d938cab15e upstream.
+
+The field "page" is initialized to KVM_UNMAPPED_PAGE when it is not used
+(i.e. when the memory lives outside kernel control). So this check will
+always end up using kunmap even for memremap regions.
+
+Fixes: e45adf665a53 ("KVM: Introduce a new guest mapping API")
+Signed-off-by: KarimAllah Ahmed <karahmed@amazon.de>
+Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
+Signed-off-by: Ben Hutchings <ben.hutchings@codethink.co.uk>
+Signed-off-by: Sasha Levin <sashal@kernel.org>
+---
+ virt/kvm/kvm_main.c | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+
+diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
+index 5b949aa273de5..33b288469c70c 100644
+--- a/virt/kvm/kvm_main.c
++++ b/virt/kvm/kvm_main.c
+@@ -1754,7 +1754,7 @@ void kvm_vcpu_unmap(struct kvm_vcpu *vcpu, struct kvm_host_map *map,
+       if (!map->hva)
+               return;
+-      if (map->page)
++      if (map->page != KVM_UNMAPPED_PAGE)
+               kunmap(map->page);
+ #ifdef CONFIG_HAS_IOMEM
+       else
+-- 
+2.20.1
+
diff --git a/queue-4.19/kvm-vmx-zero-out-all-general-purpose-registers-after.patch b/queue-4.19/kvm-vmx-zero-out-all-general-purpose-registers-after.patch
new file mode 100644 (file)
index 0000000..e6afd77
--- /dev/null
@@ -0,0 +1,68 @@
+From 141266af5129133438d51e90635ff96fc23e0180 Mon Sep 17 00:00:00 2001
+From: Sasha Levin <sashal@kernel.org>
+Date: Fri, 25 Jan 2019 07:40:50 -0800
+Subject: KVM: VMX: Zero out *all* general purpose registers after VM-Exit
+
+From: Sean Christopherson <sean.j.christopherson@intel.com>
+
+commit 0e0ab73c9a0243736bcd779b30b717e23ba9a56d upstream.
+
+...except RSP, which is restored by hardware as part of VM-Exit.
+
+Paolo theorized that restoring registers from the stack after a VM-Exit
+in lieu of zeroing them could lead to speculative execution with the
+guest's values, e.g. if the stack accesses miss the L1 cache[1].
+Zeroing XORs are dirt cheap, so just be ultra-paranoid.
+
+Note that the scratch register (currently RCX) used to save/restore the
+guest state is also zeroed as its host-defined value is loaded via the
+stack, just with a MOV instead of a POP.
+
+[1] https://patchwork.kernel.org/patch/10771539/#22441255
+
+Fixes: 0cb5b30698fd ("kvm: vmx: Scrub hardware GPRs at VM-exit")
+Cc: Jim Mattson <jmattson@google.com>
+Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
+Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
+[bwh: Backported to 4.19: adjust filename, context]
+Signed-off-by: Ben Hutchings <ben.hutchings@codethink.co.uk>
+Signed-off-by: Sasha Levin <sashal@kernel.org>
+---
+ arch/x86/kvm/vmx.c | 12 ++++++++++++
+ 1 file changed, 12 insertions(+)
+
+diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
+index d37b48173e9cf..e4d0ad06790e1 100644
+--- a/arch/x86/kvm/vmx.c
++++ b/arch/x86/kvm/vmx.c
+@@ -10841,6 +10841,15 @@ static void __noclone vmx_vcpu_run(struct kvm_vcpu *vcpu)
+               "mov %%r13, %c[r13](%0) \n\t"
+               "mov %%r14, %c[r14](%0) \n\t"
+               "mov %%r15, %c[r15](%0) \n\t"
++
++              /*
++               * Clear all general purpose registers (except RSP, which is loaded by
++               * the CPU during VM-Exit) to prevent speculative use of the guest's
++               * values, even those that are saved/loaded via the stack.  In theory,
++               * an L1 cache miss when restoring registers could lead to speculative
++               * execution with the guest's values.  Zeroing XORs are dirt cheap,
++               * i.e. the extra paranoia is essentially free.
++               */
+               "xor %%r8d,  %%r8d \n\t"
+               "xor %%r9d,  %%r9d \n\t"
+               "xor %%r10d, %%r10d \n\t"
+@@ -10855,8 +10864,11 @@ static void __noclone vmx_vcpu_run(struct kvm_vcpu *vcpu)
+               "xor %%eax, %%eax \n\t"
+               "xor %%ebx, %%ebx \n\t"
++              "xor %%ecx, %%ecx \n\t"
++              "xor %%edx, %%edx \n\t"
+               "xor %%esi, %%esi \n\t"
+               "xor %%edi, %%edi \n\t"
++              "xor %%ebp, %%ebp \n\t"
+               "pop  %%" _ASM_BP "; pop  %%" _ASM_DX " \n\t"
+               ".pushsection .rodata \n\t"
+               ".global vmx_return \n\t"
+-- 
+2.20.1
+
index 431589e07e26521b7e8c00096b7f71eccc3d8c4c..b7c00f0d9ca2eec9f5f502d1e03567b75ba1ddcc 100644 (file)
@@ -39,3 +39,14 @@ net-ipv6_stub-use-ip6_dst_lookup_flow-instead-of-ip6.patch
 blktrace-protect-q-blk_trace-with-rcu.patch
 blktrace-fix-dereference-after-null-check.patch
 f2fs-fix-to-avoid-memory-leakage-in-f2fs_listxattr.patch
+kvm-vmx-zero-out-all-general-purpose-registers-after.patch
+kvm-nvmx-always-sync-guest_bndcfgs-when-it-comes-fro.patch
+kvm-introduce-a-new-guest-mapping-api.patch
+kvm-fix-compilation-on-aarch64.patch
+kvm-fix-compilation-on-s390.patch
+kvm-fix-compile-on-s390-part-2.patch
+kvm-properly-check-if-page-is-valid-in-kvm_vcpu_unma.patch
+x86-kvm-introduce-kvm_-un-map_gfn.patch
+x86-kvm-cache-gfn-to-pfn-translation.patch
+x86-kvm-make-sure-kvm_vcpu_flush_tlb-flag-is-not-mis.patch
+x86-kvm-clean-up-host-s-steal-time-structure.patch
diff --git a/queue-4.19/x86-kvm-cache-gfn-to-pfn-translation.patch b/queue-4.19/x86-kvm-cache-gfn-to-pfn-translation.patch
new file mode 100644 (file)
index 0000000..b90ca62
--- /dev/null
@@ -0,0 +1,297 @@
+From 4c093ac7fc05527876ed5e62b4874b2feae23cf9 Mon Sep 17 00:00:00 2001
+From: Sasha Levin <sashal@kernel.org>
+Date: Thu, 5 Dec 2019 01:30:51 +0000
+Subject: x86/kvm: Cache gfn to pfn translation
+
+From: Boris Ostrovsky <boris.ostrovsky@oracle.com>
+
+commit 917248144db5d7320655dbb41d3af0b8a0f3d589 upstream.
+
+__kvm_map_gfn()'s call to gfn_to_pfn_memslot() is
+* relatively expensive
+* in certain cases (such as when done from atomic context) cannot be called
+
+Stashing gfn-to-pfn mapping should help with both cases.
+
+This is part of CVE-2019-3016.
+
+Signed-off-by: Boris Ostrovsky <boris.ostrovsky@oracle.com>
+Reviewed-by: Joao Martins <joao.m.martins@oracle.com>
+Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
+Signed-off-by: Ben Hutchings <ben.hutchings@codethink.co.uk>
+Signed-off-by: Sasha Levin <sashal@kernel.org>
+---
+ arch/x86/include/asm/kvm_host.h |  1 +
+ arch/x86/kvm/x86.c              | 10 ++++
+ include/linux/kvm_host.h        |  7 ++-
+ include/linux/kvm_types.h       |  9 ++-
+ virt/kvm/kvm_main.c             | 98 ++++++++++++++++++++++++++-------
+ 5 files changed, 103 insertions(+), 22 deletions(-)
+
+diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
+index 5c99b9bfce045..ca9c7110b99dd 100644
+--- a/arch/x86/include/asm/kvm_host.h
++++ b/arch/x86/include/asm/kvm_host.h
+@@ -626,6 +626,7 @@ struct kvm_vcpu_arch {
+               u64 last_steal;
+               struct gfn_to_hva_cache stime;
+               struct kvm_steal_time steal;
++              struct gfn_to_pfn_cache cache;
+       } st;
+       u64 tsc_offset;
+diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
+index 1a6e1aa2fb297..6916f46909ab0 100644
+--- a/arch/x86/kvm/x86.c
++++ b/arch/x86/kvm/x86.c
+@@ -8634,6 +8634,9 @@ static void fx_init(struct kvm_vcpu *vcpu)
+ void kvm_arch_vcpu_free(struct kvm_vcpu *vcpu)
+ {
+       void *wbinvd_dirty_mask = vcpu->arch.wbinvd_dirty_mask;
++      struct gfn_to_pfn_cache *cache = &vcpu->arch.st.cache;
++
++      kvm_release_pfn(cache->pfn, cache->dirty, cache);
+       kvmclock_reset(vcpu);
+@@ -9298,11 +9301,18 @@ int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
+ void kvm_arch_memslots_updated(struct kvm *kvm, u64 gen)
+ {
++      struct kvm_vcpu *vcpu;
++      int i;
++
+       /*
+        * memslots->generation has been incremented.
+        * mmio generation may have reached its maximum value.
+        */
+       kvm_mmu_invalidate_mmio_sptes(kvm, gen);
++
++      /* Force re-initialization of steal_time cache */
++      kvm_for_each_vcpu(i, vcpu, kvm)
++              kvm_vcpu_kick(vcpu);
+ }
+ int kvm_arch_prepare_memory_region(struct kvm *kvm,
+diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
+index 303c1a6916cea..dabb60f907269 100644
+--- a/include/linux/kvm_host.h
++++ b/include/linux/kvm_host.h
+@@ -708,6 +708,7 @@ void kvm_set_pfn_dirty(kvm_pfn_t pfn);
+ void kvm_set_pfn_accessed(kvm_pfn_t pfn);
+ void kvm_get_pfn(kvm_pfn_t pfn);
++void kvm_release_pfn(kvm_pfn_t pfn, bool dirty, struct gfn_to_pfn_cache *cache);
+ int kvm_read_guest_page(struct kvm *kvm, gfn_t gfn, void *data, int offset,
+                       int len);
+ int kvm_read_guest_atomic(struct kvm *kvm, gpa_t gpa, void *data,
+@@ -738,10 +739,12 @@ struct kvm_memory_slot *kvm_vcpu_gfn_to_memslot(struct kvm_vcpu *vcpu, gfn_t gfn
+ kvm_pfn_t kvm_vcpu_gfn_to_pfn_atomic(struct kvm_vcpu *vcpu, gfn_t gfn);
+ kvm_pfn_t kvm_vcpu_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn);
+ int kvm_vcpu_map(struct kvm_vcpu *vcpu, gpa_t gpa, struct kvm_host_map *map);
+-int kvm_map_gfn(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map);
++int kvm_map_gfn(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map,
++              struct gfn_to_pfn_cache *cache, bool atomic);
+ struct page *kvm_vcpu_gfn_to_page(struct kvm_vcpu *vcpu, gfn_t gfn);
+ void kvm_vcpu_unmap(struct kvm_vcpu *vcpu, struct kvm_host_map *map, bool dirty);
+-int kvm_unmap_gfn(struct kvm_vcpu *vcpu, struct kvm_host_map *map, bool dirty);
++int kvm_unmap_gfn(struct kvm_vcpu *vcpu, struct kvm_host_map *map,
++                struct gfn_to_pfn_cache *cache, bool dirty, bool atomic);
+ unsigned long kvm_vcpu_gfn_to_hva(struct kvm_vcpu *vcpu, gfn_t gfn);
+ unsigned long kvm_vcpu_gfn_to_hva_prot(struct kvm_vcpu *vcpu, gfn_t gfn, bool *writable);
+ int kvm_vcpu_read_guest_page(struct kvm_vcpu *vcpu, gfn_t gfn, void *data, int offset,
+diff --git a/include/linux/kvm_types.h b/include/linux/kvm_types.h
+index 8bf259dae9f6c..a38729c8296f4 100644
+--- a/include/linux/kvm_types.h
++++ b/include/linux/kvm_types.h
+@@ -32,7 +32,7 @@ struct kvm_memslots;
+ enum kvm_mr_change;
+-#include <asm/types.h>
++#include <linux/types.h>
+ /*
+  * Address types:
+@@ -63,4 +63,11 @@ struct gfn_to_hva_cache {
+       struct kvm_memory_slot *memslot;
+ };
++struct gfn_to_pfn_cache {
++      u64 generation;
++      gfn_t gfn;
++      kvm_pfn_t pfn;
++      bool dirty;
++};
++
+ #endif /* __KVM_TYPES_H__ */
+diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
+index 8e29b2e0bf2eb..aca15bd1cc4cd 100644
+--- a/virt/kvm/kvm_main.c
++++ b/virt/kvm/kvm_main.c
+@@ -1705,27 +1705,72 @@ struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn)
+ }
+ EXPORT_SYMBOL_GPL(gfn_to_page);
++void kvm_release_pfn(kvm_pfn_t pfn, bool dirty, struct gfn_to_pfn_cache *cache)
++{
++      if (pfn == 0)
++              return;
++
++      if (cache)
++              cache->pfn = cache->gfn = 0;
++
++      if (dirty)
++              kvm_release_pfn_dirty(pfn);
++      else
++              kvm_release_pfn_clean(pfn);
++}
++
++static void kvm_cache_gfn_to_pfn(struct kvm_memory_slot *slot, gfn_t gfn,
++                               struct gfn_to_pfn_cache *cache, u64 gen)
++{
++      kvm_release_pfn(cache->pfn, cache->dirty, cache);
++
++      cache->pfn = gfn_to_pfn_memslot(slot, gfn);
++      cache->gfn = gfn;
++      cache->dirty = false;
++      cache->generation = gen;
++}
++
+ static int __kvm_map_gfn(struct kvm_memslots *slots, gfn_t gfn,
+-                       struct kvm_host_map *map)
++                       struct kvm_host_map *map,
++                       struct gfn_to_pfn_cache *cache,
++                       bool atomic)
+ {
+       kvm_pfn_t pfn;
+       void *hva = NULL;
+       struct page *page = KVM_UNMAPPED_PAGE;
+       struct kvm_memory_slot *slot = __gfn_to_memslot(slots, gfn);
++      u64 gen = slots->generation;
+       if (!map)
+               return -EINVAL;
+-      pfn = gfn_to_pfn_memslot(slot, gfn);
++      if (cache) {
++              if (!cache->pfn || cache->gfn != gfn ||
++                      cache->generation != gen) {
++                      if (atomic)
++                              return -EAGAIN;
++                      kvm_cache_gfn_to_pfn(slot, gfn, cache, gen);
++              }
++              pfn = cache->pfn;
++      } else {
++              if (atomic)
++                      return -EAGAIN;
++              pfn = gfn_to_pfn_memslot(slot, gfn);
++      }
+       if (is_error_noslot_pfn(pfn))
+               return -EINVAL;
+       if (pfn_valid(pfn)) {
+               page = pfn_to_page(pfn);
+-              hva = kmap(page);
++              if (atomic)
++                      hva = kmap_atomic(page);
++              else
++                      hva = kmap(page);
+ #ifdef CONFIG_HAS_IOMEM
+-      } else {
++      } else if (!atomic) {
+               hva = memremap(pfn_to_hpa(pfn), PAGE_SIZE, MEMREMAP_WB);
++      } else {
++              return -EINVAL;
+ #endif
+       }
+@@ -1740,20 +1785,25 @@ static int __kvm_map_gfn(struct kvm_memslots *slots, gfn_t gfn,
+       return 0;
+ }
+-int kvm_map_gfn(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map)
++int kvm_map_gfn(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map,
++              struct gfn_to_pfn_cache *cache, bool atomic)
+ {
+-      return __kvm_map_gfn(kvm_memslots(vcpu->kvm), gfn, map);
++      return __kvm_map_gfn(kvm_memslots(vcpu->kvm), gfn, map,
++                      cache, atomic);
+ }
+ EXPORT_SYMBOL_GPL(kvm_map_gfn);
+ int kvm_vcpu_map(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map)
+ {
+-      return __kvm_map_gfn(kvm_vcpu_memslots(vcpu), gfn, map);
++      return __kvm_map_gfn(kvm_vcpu_memslots(vcpu), gfn, map,
++              NULL, false);
+ }
+ EXPORT_SYMBOL_GPL(kvm_vcpu_map);
+ static void __kvm_unmap_gfn(struct kvm_memory_slot *memslot,
+-                      struct kvm_host_map *map, bool dirty)
++                      struct kvm_host_map *map,
++                      struct gfn_to_pfn_cache *cache,
++                      bool dirty, bool atomic)
+ {
+       if (!map)
+               return;
+@@ -1761,34 +1811,44 @@ static void __kvm_unmap_gfn(struct kvm_memory_slot *memslot,
+       if (!map->hva)
+               return;
+-      if (map->page != KVM_UNMAPPED_PAGE)
+-              kunmap(map->page);
++      if (map->page != KVM_UNMAPPED_PAGE) {
++              if (atomic)
++                      kunmap_atomic(map->hva);
++              else
++                      kunmap(map->page);
++      }
+ #ifdef CONFIG_HAS_IOMEM
+-      else
++      else if (!atomic)
+               memunmap(map->hva);
++      else
++              WARN_ONCE(1, "Unexpected unmapping in atomic context");
+ #endif
+-      if (dirty) {
++      if (dirty)
+               mark_page_dirty_in_slot(memslot, map->gfn);
+-              kvm_release_pfn_dirty(map->pfn);
+-      } else {
+-              kvm_release_pfn_clean(map->pfn);
+-      }
++
++      if (cache)
++              cache->dirty |= dirty;
++      else
++              kvm_release_pfn(map->pfn, dirty, NULL);
+       map->hva = NULL;
+       map->page = NULL;
+ }
+-int kvm_unmap_gfn(struct kvm_vcpu *vcpu, struct kvm_host_map *map, bool dirty)
++int kvm_unmap_gfn(struct kvm_vcpu *vcpu, struct kvm_host_map *map, 
++                struct gfn_to_pfn_cache *cache, bool dirty, bool atomic)
+ {
+-      __kvm_unmap_gfn(gfn_to_memslot(vcpu->kvm, map->gfn), map, dirty);
++      __kvm_unmap_gfn(gfn_to_memslot(vcpu->kvm, map->gfn), map,
++                      cache, dirty, atomic);
+       return 0;
+ }
+ EXPORT_SYMBOL_GPL(kvm_unmap_gfn);
+ void kvm_vcpu_unmap(struct kvm_vcpu *vcpu, struct kvm_host_map *map, bool dirty)
+ {
+-      __kvm_unmap_gfn(kvm_vcpu_gfn_to_memslot(vcpu, map->gfn), map, dirty);
++      __kvm_unmap_gfn(kvm_vcpu_gfn_to_memslot(vcpu, map->gfn), map, NULL,
++                      dirty, false);
+ }
+ EXPORT_SYMBOL_GPL(kvm_vcpu_unmap);
+-- 
+2.20.1
+
diff --git a/queue-4.19/x86-kvm-clean-up-host-s-steal-time-structure.patch b/queue-4.19/x86-kvm-clean-up-host-s-steal-time-structure.patch
new file mode 100644 (file)
index 0000000..f34f98e
--- /dev/null
@@ -0,0 +1,87 @@
+From be590373bf483005053907e113fc3a8af728eeda Mon Sep 17 00:00:00 2001
+From: Sasha Levin <sashal@kernel.org>
+Date: Fri, 6 Dec 2019 15:36:12 +0000
+Subject: x86/KVM: Clean up host's steal time structure
+
+From: Boris Ostrovsky <boris.ostrovsky@oracle.com>
+
+commit a6bd811f1209fe1c64c9f6fd578101d6436c6b6e upstream.
+
+Now that we are mapping kvm_steal_time from the guest directly we
+don't need keep a copy of it in kvm_vcpu_arch.st. The same is true
+for the stime field.
+
+This is part of CVE-2019-3016.
+
+Signed-off-by: Boris Ostrovsky <boris.ostrovsky@oracle.com>
+Reviewed-by: Joao Martins <joao.m.martins@oracle.com>
+Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
+Signed-off-by: Ben Hutchings <ben.hutchings@codethink.co.uk>
+Signed-off-by: Sasha Levin <sashal@kernel.org>
+---
+ arch/x86/include/asm/kvm_host.h |  3 +--
+ arch/x86/kvm/x86.c              | 11 +++--------
+ 2 files changed, 4 insertions(+), 10 deletions(-)
+
+diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
+index ca9c7110b99dd..33136395db8fc 100644
+--- a/arch/x86/include/asm/kvm_host.h
++++ b/arch/x86/include/asm/kvm_host.h
+@@ -622,10 +622,9 @@ struct kvm_vcpu_arch {
+       bool pvclock_set_guest_stopped_request;
+       struct {
++              u8 preempted;
+               u64 msr_val;
+               u64 last_steal;
+-              struct gfn_to_hva_cache stime;
+-              struct kvm_steal_time steal;
+               struct gfn_to_pfn_cache cache;
+       } st;
+diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
+index d77822e03ff6b..6bfc9eaf8dee0 100644
+--- a/arch/x86/kvm/x86.c
++++ b/arch/x86/kvm/x86.c
+@@ -2418,7 +2418,7 @@ static void record_steal_time(struct kvm_vcpu *vcpu)
+       if (xchg(&st->preempted, 0) & KVM_VCPU_FLUSH_TLB)
+               kvm_vcpu_flush_tlb(vcpu, false);
+-      vcpu->arch.st.steal.preempted = 0;
++      vcpu->arch.st.preempted = 0;
+       if (st->version & 1)
+               st->version += 1;  /* first time write, random junk */
+@@ -2577,11 +2577,6 @@ int kvm_set_msr_common(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
+               if (data & KVM_STEAL_RESERVED_MASK)
+                       return 1;
+-              if (kvm_gfn_to_hva_cache_init(vcpu->kvm, &vcpu->arch.st.stime,
+-                                              data & KVM_STEAL_VALID_BITS,
+-                                              sizeof(struct kvm_steal_time)))
+-                      return 1;
+-
+               vcpu->arch.st.msr_val = data;
+               if (!(data & KVM_MSR_ENABLED))
+@@ -3280,7 +3275,7 @@ static void kvm_steal_time_set_preempted(struct kvm_vcpu *vcpu)
+       if (!(vcpu->arch.st.msr_val & KVM_MSR_ENABLED))
+               return;
+-      if (vcpu->arch.st.steal.preempted)
++      if (vcpu->arch.st.preempted)
+               return;
+       if (kvm_map_gfn(vcpu, vcpu->arch.st.msr_val >> PAGE_SHIFT, &map,
+@@ -3290,7 +3285,7 @@ static void kvm_steal_time_set_preempted(struct kvm_vcpu *vcpu)
+       st = map.hva +
+               offset_in_page(vcpu->arch.st.msr_val & KVM_STEAL_VALID_BITS);
+-      st->preempted = vcpu->arch.st.steal.preempted = KVM_VCPU_PREEMPTED;
++      st->preempted = vcpu->arch.st.preempted = KVM_VCPU_PREEMPTED;
+       kvm_unmap_gfn(vcpu, &map, &vcpu->arch.st.cache, true, true);
+ }
+-- 
+2.20.1
+
diff --git a/queue-4.19/x86-kvm-introduce-kvm_-un-map_gfn.patch b/queue-4.19/x86-kvm-introduce-kvm_-un-map_gfn.patch
new file mode 100644 (file)
index 0000000..3c74541
--- /dev/null
@@ -0,0 +1,115 @@
+From 1dd52c3a1eaf0e1a74f179add6faf5e3b700236f Mon Sep 17 00:00:00 2001
+From: Sasha Levin <sashal@kernel.org>
+Date: Tue, 12 Nov 2019 16:35:06 +0000
+Subject: x86/kvm: Introduce kvm_(un)map_gfn()
+
+From: Boris Ostrovsky <boris.ostrovsky@oracle.com>
+
+commit 1eff70a9abd46f175defafd29bc17ad456f398a7 upstream.
+
+kvm_vcpu_(un)map operates on gfns from any current address space.
+In certain cases we want to make sure we are not mapping SMRAM
+and for that we can use kvm_(un)map_gfn() that we are introducing
+in this patch.
+
+This is part of CVE-2019-3016.
+
+Signed-off-by: Boris Ostrovsky <boris.ostrovsky@oracle.com>
+Reviewed-by: Joao Martins <joao.m.martins@oracle.com>
+Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
+Signed-off-by: Ben Hutchings <ben.hutchings@codethink.co.uk>
+Signed-off-by: Sasha Levin <sashal@kernel.org>
+---
+ include/linux/kvm_host.h |  2 ++
+ virt/kvm/kvm_main.c      | 29 ++++++++++++++++++++++++-----
+ 2 files changed, 26 insertions(+), 5 deletions(-)
+
+diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
+index bef95dba14e8b..303c1a6916cea 100644
+--- a/include/linux/kvm_host.h
++++ b/include/linux/kvm_host.h
+@@ -738,8 +738,10 @@ struct kvm_memory_slot *kvm_vcpu_gfn_to_memslot(struct kvm_vcpu *vcpu, gfn_t gfn
+ kvm_pfn_t kvm_vcpu_gfn_to_pfn_atomic(struct kvm_vcpu *vcpu, gfn_t gfn);
+ kvm_pfn_t kvm_vcpu_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn);
+ int kvm_vcpu_map(struct kvm_vcpu *vcpu, gpa_t gpa, struct kvm_host_map *map);
++int kvm_map_gfn(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map);
+ struct page *kvm_vcpu_gfn_to_page(struct kvm_vcpu *vcpu, gfn_t gfn);
+ void kvm_vcpu_unmap(struct kvm_vcpu *vcpu, struct kvm_host_map *map, bool dirty);
++int kvm_unmap_gfn(struct kvm_vcpu *vcpu, struct kvm_host_map *map, bool dirty);
+ unsigned long kvm_vcpu_gfn_to_hva(struct kvm_vcpu *vcpu, gfn_t gfn);
+ unsigned long kvm_vcpu_gfn_to_hva_prot(struct kvm_vcpu *vcpu, gfn_t gfn, bool *writable);
+ int kvm_vcpu_read_guest_page(struct kvm_vcpu *vcpu, gfn_t gfn, void *data, int offset,
+diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
+index 33b288469c70c..8e29b2e0bf2eb 100644
+--- a/virt/kvm/kvm_main.c
++++ b/virt/kvm/kvm_main.c
+@@ -1705,12 +1705,13 @@ struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn)
+ }
+ EXPORT_SYMBOL_GPL(gfn_to_page);
+-static int __kvm_map_gfn(struct kvm_memory_slot *slot, gfn_t gfn,
++static int __kvm_map_gfn(struct kvm_memslots *slots, gfn_t gfn,
+                        struct kvm_host_map *map)
+ {
+       kvm_pfn_t pfn;
+       void *hva = NULL;
+       struct page *page = KVM_UNMAPPED_PAGE;
++      struct kvm_memory_slot *slot = __gfn_to_memslot(slots, gfn);
+       if (!map)
+               return -EINVAL;
+@@ -1739,14 +1740,20 @@ static int __kvm_map_gfn(struct kvm_memory_slot *slot, gfn_t gfn,
+       return 0;
+ }
++int kvm_map_gfn(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map)
++{
++      return __kvm_map_gfn(kvm_memslots(vcpu->kvm), gfn, map);
++}
++EXPORT_SYMBOL_GPL(kvm_map_gfn);
++
+ int kvm_vcpu_map(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map)
+ {
+-      return __kvm_map_gfn(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn, map);
++      return __kvm_map_gfn(kvm_vcpu_memslots(vcpu), gfn, map);
+ }
+ EXPORT_SYMBOL_GPL(kvm_vcpu_map);
+-void kvm_vcpu_unmap(struct kvm_vcpu *vcpu, struct kvm_host_map *map,
+-                  bool dirty)
++static void __kvm_unmap_gfn(struct kvm_memory_slot *memslot,
++                      struct kvm_host_map *map, bool dirty)
+ {
+       if (!map)
+               return;
+@@ -1762,7 +1769,7 @@ void kvm_vcpu_unmap(struct kvm_vcpu *vcpu, struct kvm_host_map *map,
+ #endif
+       if (dirty) {
+-              kvm_vcpu_mark_page_dirty(vcpu, map->gfn);
++              mark_page_dirty_in_slot(memslot, map->gfn);
+               kvm_release_pfn_dirty(map->pfn);
+       } else {
+               kvm_release_pfn_clean(map->pfn);
+@@ -1771,6 +1778,18 @@ void kvm_vcpu_unmap(struct kvm_vcpu *vcpu, struct kvm_host_map *map,
+       map->hva = NULL;
+       map->page = NULL;
+ }
++
++int kvm_unmap_gfn(struct kvm_vcpu *vcpu, struct kvm_host_map *map, bool dirty)
++{
++      __kvm_unmap_gfn(gfn_to_memslot(vcpu->kvm, map->gfn), map, dirty);
++      return 0;
++}
++EXPORT_SYMBOL_GPL(kvm_unmap_gfn);
++
++void kvm_vcpu_unmap(struct kvm_vcpu *vcpu, struct kvm_host_map *map, bool dirty)
++{
++      __kvm_unmap_gfn(kvm_vcpu_gfn_to_memslot(vcpu, map->gfn), map, dirty);
++}
+ EXPORT_SYMBOL_GPL(kvm_vcpu_unmap);
+ struct page *kvm_vcpu_gfn_to_page(struct kvm_vcpu *vcpu, gfn_t gfn)
+-- 
+2.20.1
+
diff --git a/queue-4.19/x86-kvm-make-sure-kvm_vcpu_flush_tlb-flag-is-not-mis.patch b/queue-4.19/x86-kvm-make-sure-kvm_vcpu_flush_tlb-flag-is-not-mis.patch
new file mode 100644 (file)
index 0000000..4b87225
--- /dev/null
@@ -0,0 +1,131 @@
+From 5139b7c841f3531daf4313aab2de2b4a8ecce9d8 Mon Sep 17 00:00:00 2001
+From: Sasha Levin <sashal@kernel.org>
+Date: Thu, 5 Dec 2019 03:45:32 +0000
+Subject: x86/KVM: Make sure KVM_VCPU_FLUSH_TLB flag is not missed
+
+From: Boris Ostrovsky <boris.ostrovsky@oracle.com>
+
+commit b043138246a41064527cf019a3d51d9f015e9796 upstream.
+
+There is a potential race in record_steal_time() between setting
+host-local vcpu->arch.st.steal.preempted to zero (i.e. clearing
+KVM_VCPU_PREEMPTED) and propagating this value to the guest with
+kvm_write_guest_cached(). Between those two events the guest may
+still see KVM_VCPU_PREEMPTED in its copy of kvm_steal_time, set
+KVM_VCPU_FLUSH_TLB and assume that hypervisor will do the right
+thing. Which it won't.
+
+Instad of copying, we should map kvm_steal_time and that will
+guarantee atomicity of accesses to @preempted.
+
+This is part of CVE-2019-3016.
+
+Signed-off-by: Boris Ostrovsky <boris.ostrovsky@oracle.com>
+Reviewed-by: Joao Martins <joao.m.martins@oracle.com>
+Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
+[bwh: Backported to 4.19: No tracepoint in record_steal_time().]
+Signed-off-by: Ben Hutchings <ben.hutchings@codethink.co.uk>
+Signed-off-by: Sasha Levin <sashal@kernel.org>
+---
+ arch/x86/kvm/x86.c | 49 +++++++++++++++++++++++++++-------------------
+ 1 file changed, 29 insertions(+), 20 deletions(-)
+
+diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
+index 6916f46909ab0..d77822e03ff6b 100644
+--- a/arch/x86/kvm/x86.c
++++ b/arch/x86/kvm/x86.c
+@@ -2397,43 +2397,45 @@ static void kvm_vcpu_flush_tlb(struct kvm_vcpu *vcpu, bool invalidate_gpa)
+ static void record_steal_time(struct kvm_vcpu *vcpu)
+ {
++      struct kvm_host_map map;
++      struct kvm_steal_time *st;
++
+       if (!(vcpu->arch.st.msr_val & KVM_MSR_ENABLED))
+               return;
+-      if (unlikely(kvm_read_guest_cached(vcpu->kvm, &vcpu->arch.st.stime,
+-              &vcpu->arch.st.steal, sizeof(struct kvm_steal_time))))
++      /* -EAGAIN is returned in atomic context so we can just return. */
++      if (kvm_map_gfn(vcpu, vcpu->arch.st.msr_val >> PAGE_SHIFT,
++                      &map, &vcpu->arch.st.cache, false))
+               return;
++      st = map.hva +
++              offset_in_page(vcpu->arch.st.msr_val & KVM_STEAL_VALID_BITS);
++
+       /*
+        * Doing a TLB flush here, on the guest's behalf, can avoid
+        * expensive IPIs.
+        */
+-      if (xchg(&vcpu->arch.st.steal.preempted, 0) & KVM_VCPU_FLUSH_TLB)
++      if (xchg(&st->preempted, 0) & KVM_VCPU_FLUSH_TLB)
+               kvm_vcpu_flush_tlb(vcpu, false);
+-      if (vcpu->arch.st.steal.version & 1)
+-              vcpu->arch.st.steal.version += 1;  /* first time write, random junk */
++      vcpu->arch.st.steal.preempted = 0;
+-      vcpu->arch.st.steal.version += 1;
++      if (st->version & 1)
++              st->version += 1;  /* first time write, random junk */
+-      kvm_write_guest_cached(vcpu->kvm, &vcpu->arch.st.stime,
+-              &vcpu->arch.st.steal, sizeof(struct kvm_steal_time));
++      st->version += 1;
+       smp_wmb();
+-      vcpu->arch.st.steal.steal += current->sched_info.run_delay -
++      st->steal += current->sched_info.run_delay -
+               vcpu->arch.st.last_steal;
+       vcpu->arch.st.last_steal = current->sched_info.run_delay;
+-      kvm_write_guest_cached(vcpu->kvm, &vcpu->arch.st.stime,
+-              &vcpu->arch.st.steal, sizeof(struct kvm_steal_time));
+-
+       smp_wmb();
+-      vcpu->arch.st.steal.version += 1;
++      st->version += 1;
+-      kvm_write_guest_cached(vcpu->kvm, &vcpu->arch.st.stime,
+-              &vcpu->arch.st.steal, sizeof(struct kvm_steal_time));
++      kvm_unmap_gfn(vcpu, &map, &vcpu->arch.st.cache, true, false);
+ }
+ int kvm_set_msr_common(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
+@@ -3272,18 +3274,25 @@ void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
+ static void kvm_steal_time_set_preempted(struct kvm_vcpu *vcpu)
+ {
++      struct kvm_host_map map;
++      struct kvm_steal_time *st;
++
+       if (!(vcpu->arch.st.msr_val & KVM_MSR_ENABLED))
+               return;
+       if (vcpu->arch.st.steal.preempted)
+               return;
+-      vcpu->arch.st.steal.preempted = KVM_VCPU_PREEMPTED;
++      if (kvm_map_gfn(vcpu, vcpu->arch.st.msr_val >> PAGE_SHIFT, &map,
++                      &vcpu->arch.st.cache, true))
++              return;
++
++      st = map.hva +
++              offset_in_page(vcpu->arch.st.msr_val & KVM_STEAL_VALID_BITS);
++
++      st->preempted = vcpu->arch.st.steal.preempted = KVM_VCPU_PREEMPTED;
+-      kvm_write_guest_offset_cached(vcpu->kvm, &vcpu->arch.st.stime,
+-                      &vcpu->arch.st.steal.preempted,
+-                      offsetof(struct kvm_steal_time, preempted),
+-                      sizeof(vcpu->arch.st.steal.preempted));
++      kvm_unmap_gfn(vcpu, &map, &vcpu->arch.st.cache, true, true);
+ }
+ void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
+-- 
+2.20.1
+