[PATCH 0/2] SEV-SNP guest support fixes

[PATCH 0/2] SEV-SNP guest support fixes

[Lendacky, Thomas]

This patch series provides some fixes around AP creation:

- An erratum on AMD hardware requires that a VMSA not be aligned on a
  2MB boundary. To work around this issue, allocate 2 pages of memory
  and using the page that is not 2MB aligned and freeing the other.

- When parking APs after exiting boot services, the current SNP support
  will perform an allocation that will not be reflected in memory map
  being supplied to the OS. Instead of allocating new VMSAs each time,
  re-use the current VMSA.

BZ: https://bugzilla.tianocore.org/show_bug.cgi?id=4353

---

These patches are based on commit:
f80f052277c8 ("OvmfPkg/RiscVVirt: Add Stack HOB")

Cc:	Eric Dong <eric.dong@intel.com>
Cc:	Ray Ni <ray.ni@intel.com>
Cc:	Rahul Kumar <rahul1.kumar@intel.com>
Cc:	Gerd Hoffmann <kraxel@redhat.com>
Cc:	Michael Roth <michael.roth@amd.com>
Cc:	Ashish Kalra <Ashish.Kalra@amd.com>

Tom Lendacky (2):
  UefiCpuPkg/MpInitLib: Ensure SEV-SNP VMSA allocations are not 2MB
    aligned
  UefiCpuPkg/MpInitLib: Reuse VMSA allocation to avoid unreserved
    allocation

 UefiCpuPkg/Library/MpInitLib/X64/AmdSev.c | 224 ++++++++++++++--------
 1 file changed, 144 insertions(+), 80 deletions(-)

-- 
2.39.1

[PATCH 1/2] UefiCpuPkg/MpInitLib: Ensure SEV-SNP VMSA allocations are not 2MB aligned

[Lendacky, Thomas]

BZ: https://bugzilla.tianocore.org/show_bug.cgi?id=4353

Due to an erratum, an SEV-SNP VMSA cannot be 2MB aligned. To work around
this issue, allocate two pages instead of one. Because of the way that
page allocation is implemented, always try to use the second page. If the
second page is not 2MB aligned, free the first page and use the second
page. If the second page is 2MB aligned, free the second page and use the
first page. Freeing in this way reduces holes in the memory map.

Fixes: 06544455d0d4 ("UefiCpuPkg/MpInitLib: Use SEV-SNP AP Creation ...")
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
---
 UefiCpuPkg/Library/MpInitLib/X64/AmdSev.c | 24 +++++++++++++++++++++--
 1 file changed, 22 insertions(+), 2 deletions(-)

diff --git a/UefiCpuPkg/Library/MpInitLib/X64/AmdSev.c b/UefiCpuPkg/Library/MpInitLib/X64/AmdSev.c
index bfda1e19030d..7abdda3e1c7e 100644
--- a/UefiCpuPkg/Library/MpInitLib/X64/AmdSev.c
+++ b/UefiCpuPkg/Library/MpInitLib/X64/AmdSev.c
@@ -13,6 +13,8 @@
 #include <Register/Amd/Fam17Msr.h>
 #include <Register/Amd/Ghcb.h>
 
+#define IS_ALIGNED(x, y)  ((((UINTN)(x) & (y - 1)) == 0))
+
 /**
   Create an SEV-SNP AP save area (VMSA) for use in running the vCPU.
 
@@ -27,6 +29,7 @@ SevSnpCreateSaveArea (
   UINT32          ApicId
   )
 {
+  UINT8                     *Pages;
   SEV_ES_SAVE_AREA          *SaveArea;
   IA32_CR0                  ApCr0;
   IA32_CR0                  ResetCr0;
@@ -44,12 +47,29 @@ SevSnpCreateSaveArea (
 
   //
   // Allocate a single page for the SEV-ES Save Area and initialize it.
+  // Due to an erratum that prevents a VMSA being on a 2MB boundary,
+  // allocate an extra page to work around the issue.
   //
-  SaveArea = AllocateReservedPages (1);
-  if (!SaveArea) {
+  Pages = AllocateReservedPages (2);
+  if (!Pages) {
     return;
   }
 
+  //
+  // Since page allocation works by allocating downward in the address space,
+  // try to always free the first (lower address) page to limit possible holes
+  // in the memory map. So, if the address of the second page is 2MB aligned,
+  // then use the first page and free the second page. Otherwise, free the
+  // first page and use the second page.
+  //
+  if (IS_ALIGNED (Pages + EFI_PAGE_SIZE, SIZE_2MB)) {
+    SaveArea = (SEV_ES_SAVE_AREA *)Pages;
+    FreePages (Pages + EFI_PAGE_SIZE, 1);
+  } else {
+    SaveArea = (SEV_ES_SAVE_AREA *)(Pages + EFI_PAGE_SIZE);
+    FreePages (Pages, 1);
+  }
+
   ZeroMem (SaveArea, EFI_PAGE_SIZE);
 
   //
-- 
2.39.1

[PATCH 2/2] UefiCpuPkg/MpInitLib: Reuse VMSA allocation to avoid unreserved allocation

[Lendacky, Thomas]

https://bugzilla.tianocore.org/show_bug.cgi?id=4353

When parking the APs on exiting from UEFI, a new page allocation is made.
This allocation, however, does not end up being marked reserved in the
memory map supplied to the OS. To avoid this, re-use the VMSA by clearing
the VMSA RMP flag, updating the page contents and re-setting the VMSA RMP
flag.

Fixes: 06544455d0d4 ("UefiCpuPkg/MpInitLib: Use SEV-SNP AP Creation ...")
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
---
 UefiCpuPkg/Library/MpInitLib/X64/AmdSev.c | 234 +++++++++++++---------
 1 file changed, 139 insertions(+), 95 deletions(-)

diff --git a/UefiCpuPkg/Library/MpInitLib/X64/AmdSev.c b/UefiCpuPkg/Library/MpInitLib/X64/AmdSev.c
index 7abdda3e1c7e..ae88bbbfd828 100644
--- a/UefiCpuPkg/Library/MpInitLib/X64/AmdSev.c
+++ b/UefiCpuPkg/Library/MpInitLib/X64/AmdSev.c
@@ -16,58 +16,158 @@
 #define IS_ALIGNED(x, y)  ((((UINTN)(x) & (y - 1)) == 0))
 
 /**
-  Create an SEV-SNP AP save area (VMSA) for use in running the vCPU.
+  Perform the requested AP Creation action.
 
-  @param[in]  CpuMpData        Pointer to CPU MP Data
-  @param[in]  CpuData          Pointer to CPU AP Data
+  @param[in]  SaveArea         Pointer to VM save area (VMSA)
   @param[in]  ApicId           APIC ID of the vCPU
+  @param[in]  Action           AP action to perform
+
+  @retval  TRUE   Action completed successfully
+  @retval  FALSE  Action did not complete successfully
 **/
-VOID
-SevSnpCreateSaveArea (
-  IN CPU_MP_DATA  *CpuMpData,
-  IN CPU_AP_DATA  *CpuData,
-  UINT32          ApicId
+STATIC
+BOOLEAN
+SevSnpPerformApAction (
+  IN SEV_ES_SAVE_AREA  *SaveArea,
+  IN UINT32            ApicId,
+  IN UINTN             Action
   )
 {
-  UINT8                     *Pages;
-  SEV_ES_SAVE_AREA          *SaveArea;
-  IA32_CR0                  ApCr0;
-  IA32_CR0                  ResetCr0;
-  IA32_CR4                  ApCr4;
-  IA32_CR4                  ResetCr4;
-  UINTN                     StartIp;
-  UINT8                     SipiVector;
-  UINT32                    RmpAdjustStatus;
-  UINT64                    VmgExitStatus;
   MSR_SEV_ES_GHCB_REGISTER  Msr;
   GHCB                      *Ghcb;
   BOOLEAN                   InterruptState;
   UINT64                    ExitInfo1;
   UINT64                    ExitInfo2;
+  UINT64                    VmgExitStatus;
+  UINT32                    RmpAdjustStatus;
 
-  //
-  // Allocate a single page for the SEV-ES Save Area and initialize it.
-  // Due to an erratum that prevents a VMSA being on a 2MB boundary,
-  // allocate an extra page to work around the issue.
-  //
-  Pages = AllocateReservedPages (2);
-  if (!Pages) {
-    return;
+  if (Action == SVM_VMGEXIT_SNP_AP_CREATE) {
+    //
+    // To turn the page into a recognized VMSA page, issue RMPADJUST:
+    //   Target VMPL but numerically higher than current VMPL
+    //   Target PermissionMask is not used
+    //
+    RmpAdjustStatus = SevSnpRmpAdjust (
+                        (EFI_PHYSICAL_ADDRESS)(UINTN)SaveArea,
+                        TRUE
+                        );
+    if (RmpAdjustStatus != 0) {
+      DEBUG ((DEBUG_INFO, "SEV-SNP: RMPADJUST failed for VMSA creation\n"));
+      ASSERT (FALSE);
+
+      return FALSE;
+    }
+  }
+
+  ExitInfo1  = (UINT64)ApicId << 32;
+  ExitInfo1 |= Action;
+  ExitInfo2  = (UINT64)(UINTN)SaveArea;
+
+  Msr.GhcbPhysicalAddress = AsmReadMsr64 (MSR_SEV_ES_GHCB);
+  Ghcb                    = Msr.Ghcb;
+
+  CcExitVmgInit (Ghcb, &InterruptState);
+
+  if (Action == SVM_VMGEXIT_SNP_AP_CREATE) {
+    Ghcb->SaveArea.Rax = SaveArea->SevFeatures;
+    CcExitVmgSetOffsetValid (Ghcb, GhcbRax);
   }
 
-  //
-  // Since page allocation works by allocating downward in the address space,
-  // try to always free the first (lower address) page to limit possible holes
-  // in the memory map. So, if the address of the second page is 2MB aligned,
-  // then use the first page and free the second page. Otherwise, free the
-  // first page and use the second page.
-  //
-  if (IS_ALIGNED (Pages + EFI_PAGE_SIZE, SIZE_2MB)) {
-    SaveArea = (SEV_ES_SAVE_AREA *)Pages;
-    FreePages (Pages + EFI_PAGE_SIZE, 1);
+  VmgExitStatus = CcExitVmgExit (
+                    Ghcb,
+                    SVM_EXIT_SNP_AP_CREATION,
+                    ExitInfo1,
+                    ExitInfo2
+                    );
+
+  CcExitVmgDone (Ghcb, InterruptState);
+
+  if (VmgExitStatus != 0) {
+    DEBUG ((DEBUG_INFO, "SEV-SNP: AP Destroy failed\n"));
+    ASSERT (FALSE);
+
+    return FALSE;
+  }
+
+  if (Action == SVM_VMGEXIT_SNP_AP_DESTROY) {
+    //
+    // Make the current VMSA not runnable and accessible to be
+    // reprogrammed.
+    //
+    RmpAdjustStatus = SevSnpRmpAdjust (
+                        (EFI_PHYSICAL_ADDRESS)(UINTN)SaveArea,
+                        FALSE
+                        );
+    if (RmpAdjustStatus != 0) {
+      DEBUG ((DEBUG_INFO, "SEV-SNP: RMPADJUST failed for VMSA reset\n"));
+      ASSERT (FALSE);
+
+      return FALSE;
+    }
+  }
+
+  return TRUE;
+}
+
+/**
+  Create an SEV-SNP AP save area (VMSA) for use in running the vCPU.
+
+  @param[in]  CpuMpData        Pointer to CPU MP Data
+  @param[in]  CpuData          Pointer to CPU AP Data
+  @param[in]  ApicId           APIC ID of the vCPU
+**/
+VOID
+SevSnpCreateSaveArea (
+  IN CPU_MP_DATA  *CpuMpData,
+  IN CPU_AP_DATA  *CpuData,
+  UINT32          ApicId
+  )
+{
+  UINT8             *Pages;
+  SEV_ES_SAVE_AREA  *SaveArea;
+  IA32_CR0          ApCr0;
+  IA32_CR0          ResetCr0;
+  IA32_CR4          ApCr4;
+  IA32_CR4          ResetCr4;
+  UINTN             StartIp;
+  UINT8             SipiVector;
+
+  if (CpuData->SevEsSaveArea == NULL) {
+    //
+    // Allocate a single page for the SEV-ES Save Area and initialize it.
+    // Due to an erratum that prevents a VMSA being on a 2MB boundary,
+    // allocate an extra page to work around the issue.
+    //
+    Pages = AllocateReservedPages (2);
+    if (!Pages) {
+      return;
+    }
+
+    //
+    // Since page allocation works by allocating downward in the address space,
+    // try to always free the first (lower address) page to limit possible holes
+    // in the memory map. So, if the address of the second page is 2MB aligned,
+    // then use the first page and free the second page. Otherwise, free the
+    // first page and use the second page.
+    //
+    if (IS_ALIGNED (Pages + EFI_PAGE_SIZE, SIZE_2MB)) {
+      SaveArea = (SEV_ES_SAVE_AREA *)Pages;
+      FreePages (Pages + EFI_PAGE_SIZE, 1);
+    } else {
+      SaveArea = (SEV_ES_SAVE_AREA *)(Pages + EFI_PAGE_SIZE);
+      FreePages (Pages, 1);
+    }
+
+    CpuData->SevEsSaveArea = SaveArea;
   } else {
-    SaveArea = (SEV_ES_SAVE_AREA *)(Pages + EFI_PAGE_SIZE);
-    FreePages (Pages, 1);
+    SaveArea = CpuData->SevEsSaveArea;
+
+    //
+    // Tell the hypervisor to not use the current VMSA
+    //
+    if (!SevSnpPerformApAction (SaveArea, ApicId, SVM_VMGEXIT_SNP_AP_DESTROY)) {
+      return;
+    }
   }
 
   ZeroMem (SaveArea, EFI_PAGE_SIZE);
@@ -152,63 +252,7 @@ SevSnpCreateSaveArea (
   SaveArea->Vmpl        = 0;
   SaveArea->SevFeatures = AsmReadMsr64 (MSR_SEV_STATUS) >> 2;
 
-  //
-  // To turn the page into a recognized VMSA page, issue RMPADJUST:
-  //   Target VMPL but numerically higher than current VMPL
-  //   Target PermissionMask is not used
-  //
-  RmpAdjustStatus = SevSnpRmpAdjust (
-                      (EFI_PHYSICAL_ADDRESS)(UINTN)SaveArea,
-                      TRUE
-                      );
-  ASSERT (RmpAdjustStatus == 0);
-
-  ExitInfo1  = (UINT64)ApicId << 32;
-  ExitInfo1 |= SVM_VMGEXIT_SNP_AP_CREATE;
-  ExitInfo2  = (UINT64)(UINTN)SaveArea;
-
-  Msr.GhcbPhysicalAddress = AsmReadMsr64 (MSR_SEV_ES_GHCB);
-  Ghcb                    = Msr.Ghcb;
-
-  CcExitVmgInit (Ghcb, &InterruptState);
-  Ghcb->SaveArea.Rax = SaveArea->SevFeatures;
-  CcExitVmgSetOffsetValid (Ghcb, GhcbRax);
-  VmgExitStatus = CcExitVmgExit (
-                    Ghcb,
-                    SVM_EXIT_SNP_AP_CREATION,
-                    ExitInfo1,
-                    ExitInfo2
-                    );
-  CcExitVmgDone (Ghcb, InterruptState);
-
-  ASSERT (VmgExitStatus == 0);
-  if (VmgExitStatus != 0) {
-    RmpAdjustStatus = SevSnpRmpAdjust (
-                        (EFI_PHYSICAL_ADDRESS)(UINTN)SaveArea,
-                        FALSE
-                        );
-    if (RmpAdjustStatus == 0) {
-      FreePages (SaveArea, 1);
-    } else {
-      DEBUG ((DEBUG_INFO, "SEV-SNP: RMPADJUST failed, leaking VMSA page\n"));
-    }
-
-    SaveArea = NULL;
-  }
-
-  if (CpuData->SevEsSaveArea) {
-    RmpAdjustStatus = SevSnpRmpAdjust (
-                        (EFI_PHYSICAL_ADDRESS)(UINTN)CpuData->SevEsSaveArea,
-                        FALSE
-                        );
-    if (RmpAdjustStatus == 0) {
-      FreePages (CpuData->SevEsSaveArea, 1);
-    } else {
-      DEBUG ((DEBUG_INFO, "SEV-SNP: RMPADJUST failed, leaking VMSA page\n"));
-    }
-  }
-
-  CpuData->SevEsSaveArea = SaveArea;
+  SevSnpPerformApAction (SaveArea, ApicId, SVM_VMGEXIT_SNP_AP_CREATE);
 }
 
 /**
-- 
2.39.1

Re: [PATCH 0/2] SEV-SNP guest support fixes

[Lendacky, Thomas]

On 3/10/23 11:03, Tom Lendacky wrote:
> This patch series provides some fixes around AP creation:
> 
> - An erratum on AMD hardware requires that a VMSA not be aligned on a
>    2MB boundary. To work around this issue, allocate 2 pages of memory
>    and using the page that is not 2MB aligned and freeing the other.
> 
> - When parking APs after exiting boot services, the current SNP support
>    will perform an allocation that will not be reflected in memory map
>    being supplied to the OS. Instead of allocating new VMSAs each time,
>    re-use the current VMSA.
> 
> BZ: https://bugzilla.tianocore.org/show_bug.cgi?id=4353
> 
> ---
> 
> These patches are based on commit:
> f80f052277c8 ("OvmfPkg/RiscVVirt: Add Stack HOB")

Specified the wrong commit here...  should be:

9b94ebb0c826 ("DynamicTablesPkg: Add SMBIOS String table helper library")

Also, the first patch has a #define for IS_ALIGNED, which Gerd is trying 
to address with another patch series. Once his series is merged, I'll 
re-submit with the removal of the #define, but wanted to get this out for 
review now.

Thanks,
Tom

> 
> Cc:	Eric Dong <eric.dong@intel.com>
> Cc:	Ray Ni <ray.ni@intel.com>
> Cc:	Rahul Kumar <rahul1.kumar@intel.com>
> Cc:	Gerd Hoffmann <kraxel@redhat.com>
> Cc:	Michael Roth <michael.roth@amd.com>
> Cc:	Ashish Kalra <Ashish.Kalra@amd.com>
> 
> Tom Lendacky (2):
>    UefiCpuPkg/MpInitLib: Ensure SEV-SNP VMSA allocations are not 2MB
>      aligned
>    UefiCpuPkg/MpInitLib: Reuse VMSA allocation to avoid unreserved
>      allocation
> 
>   UefiCpuPkg/Library/MpInitLib/X64/AmdSev.c | 224 ++++++++++++++--------
>   1 file changed, 144 insertions(+), 80 deletions(-)
> 

Re: [PATCH 1/2] UefiCpuPkg/MpInitLib: Ensure SEV-SNP VMSA allocations are not 2MB aligned

[Gerd Hoffmann]

  Hi,

>    // Allocate a single page for the SEV-ES Save Area and initialize it.
> +  // Due to an erratum that prevents a VMSA being on a 2MB boundary,
> +  // allocate an extra page to work around the issue.

A reference to the erratum (web link or erratum id) would be nice here.
Also swapping the order of the two patches might simplify the other
patch which happens to shuffle around the code this patch has just
added.

take care,
  Gerd

Re: [PATCH 1/2] UefiCpuPkg/MpInitLib: Ensure SEV-SNP VMSA allocations are not 2MB aligned

[Ni, Ray]

> 
> +#define IS_ALIGNED(x, y)  ((((UINTN)(x) & (y - 1)) == 0))

1. Can you use the existing macro ALIGN_POINTER() defined in Base.h?

Re: [PATCH 1/2] UefiCpuPkg/MpInitLib: Ensure SEV-SNP VMSA allocations are not 2MB aligned

[Ni, Ray]

Agree with both comments😊

> -----Original Message-----
> From: Gerd Hoffmann <kraxel@redhat.com>
> Sent: Monday, March 13, 2023 4:00 PM
> To: Tom Lendacky <thomas.lendacky@amd.com>
> Cc: devel@edk2.groups.io; Dong, Eric <eric.dong@intel.com>; Ni, Ray
> <ray.ni@intel.com>; Kumar, Rahul R <rahul.r.kumar@intel.com>; Michael
> Roth <michael.roth@amd.com>; Ashish Kalra <Ashish.Kalra@amd.com>
> Subject: Re: [PATCH 1/2] UefiCpuPkg/MpInitLib: Ensure SEV-SNP VMSA
> allocations are not 2MB aligned
> 
>   Hi,
> 
> >    // Allocate a single page for the SEV-ES Save Area and initialize it.
> > +  // Due to an erratum that prevents a VMSA being on a 2MB boundary,
> > +  // allocate an extra page to work around the issue.
> 
> A reference to the erratum (web link or erratum id) would be nice here.
> Also swapping the order of the two patches might simplify the other
> patch which happens to shuffle around the code this patch has just
> added.
> 
> take care,
>   Gerd

Re: [PATCH 1/2] UefiCpuPkg/MpInitLib: Ensure SEV-SNP VMSA allocations are not 2MB aligned

[Gerd Hoffmann]

On Mon, Mar 13, 2023 at 08:28:57AM +0000, Ni, Ray wrote:
> > 
> > +#define IS_ALIGNED(x, y)  ((((UINTN)(x) & (y - 1)) == 0))
> 
> 1. Can you use the existing macro ALIGN_POINTER() defined in Base.h?

Having copies of this all over the tree is indeed a bad idea.

See https://edk2.groups.io/g/devel/message/100695 which adds this
and a few more commonly used macros to Base.h.  Can we get that
reviewed and merged please?

thanks & take care,
  Gerd

Re: [edk2-devel] [PATCH 1/2] UefiCpuPkg/MpInitLib: Ensure SEV-SNP VMSA allocations are not 2MB aligned

[Ni, Ray]

That depends on review from other package maintainers.

> -----Original Message-----
> From: devel@edk2.groups.io <devel@edk2.groups.io> On Behalf Of Gerd
> Hoffmann
> Sent: Monday, March 13, 2023 4:46 PM
> To: Ni, Ray <ray.ni@intel.com>
> Cc: Tom Lendacky <thomas.lendacky@amd.com>; devel@edk2.groups.io;
> Dong, Eric <eric.dong@intel.com>; Kumar, Rahul R
> <rahul.r.kumar@intel.com>; Michael Roth <michael.roth@amd.com>; Ashish
> Kalra <Ashish.Kalra@amd.com>
> Subject: Re: [edk2-devel] [PATCH 1/2] UefiCpuPkg/MpInitLib: Ensure SEV-
> SNP VMSA allocations are not 2MB aligned
> 
> On Mon, Mar 13, 2023 at 08:28:57AM +0000, Ni, Ray wrote:
> > >
> > > +#define IS_ALIGNED(x, y)  ((((UINTN)(x) & (y - 1)) == 0))
> >
> > 1. Can you use the existing macro ALIGN_POINTER() defined in Base.h?
> 
> Having copies of this all over the tree is indeed a bad idea.
> 
> See https://edk2.groups.io/g/devel/message/100695 which adds this
> and a few more commonly used macros to Base.h.  Can we get that
> reviewed and merged please?
> 
> thanks & take care,
>   Gerd
> 
> 
> 
> 
> 

Re: [PATCH 1/2] UefiCpuPkg/MpInitLib: Ensure SEV-SNP VMSA allocations are not 2MB aligned

[Lendacky, Thomas]

On 3/13/23 03:28, Ni, Ray wrote:
>>
>> +#define IS_ALIGNED(x, y)  ((((UINTN)(x) & (y - 1)) == 0))
> 
> 1. Can you use the existing macro ALIGN_POINTER() defined in Base.h?

See my reply to the cover letter where I say I want to replace the usage 
with Gerd's definitions/updates series (but wanted general feedback now on 
the series).

Thanks,
Tom

Re: [PATCH 1/2] UefiCpuPkg/MpInitLib: Ensure SEV-SNP VMSA allocations are not 2MB aligned

[Lendacky, Thomas]

On 3/13/23 03:00, Gerd Hoffmann wrote:
>    Hi,
> 
>>     // Allocate a single page for the SEV-ES Save Area and initialize it.
>> +  // Due to an erratum that prevents a VMSA being on a 2MB boundary,
>> +  // allocate an extra page to work around the issue.
> 
> A reference to the erratum (web link or erratum id) would be nice here.
> Also swapping the order of the two patches might simplify the other
> patch which happens to shuffle around the code this patch has just
> added.

Will do.

Thanks,
Tom

> 
> take care,
>    Gerd
>