[PATCH edk2 hci-8.1] OvmfPkg/PlatformInitLib: switch to 5-level paging only above 16TB #VSTOR-152672 #VSTOR-152405

[PATCH edk2 hci-8.1] OvmfPkg/PlatformInitLib: switch to 5-level paging only above 16TB #VSTOR-152672 #VSTOR-152405

[Andrey Drobyshev]

With PcdUse5LevelPageTable the 64-bit PEI decides per guest whether to
hand the OS over in 4- or 5-level paging mode.  Without a fw_cfg
override it uses 5-level paging as soon as the guest address space
(RAM plus the memory hotplug window) reaches 1TB.

That threshold is too low for us:

  * 4-level firmware paging already covers 64TB of address space, so
    5-level paging below that buys nothing and only changes what the OS
    sees at ExitBootServices;
  * on Windows that means booting the LA57 kernel (ntkrla57.exe), which
    bugchecks on AMD guests exposed to the ARCH_CAPABILITIES MSR
    (VSTOR-152405) and breaks Windows Setup.  QEMU since the commit
    7fc0e1c9f577 ("i386: hide la57 at fresh start on vz machine types
    that show ARCH_CAPABILITIES #VSTOR-152405") now hides la57 on the
    VZ machine types that expose the MSR, but the firmware shouldn't
    depend on that.

Raise the threshold to 16TB.  Ordinary guests, including those with a
large hotplug window, stay in 4-level paging.

Signed-off-by: Andrey Drobyshev <andrey.drobyshev@virtuozzo.com>
---
 OvmfPkg/Library/PlatformInitLib/MemDetect.c | 6 +++---
 OvmfPkg/RUNTIME_CONFIG.md                   | 2 +-
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/OvmfPkg/Library/PlatformInitLib/MemDetect.c b/OvmfPkg/Library/PlatformInitLib/MemDetect.c
index 937e2b77a5..7220db8ef9 100644
--- a/OvmfPkg/Library/PlatformInitLib/MemDetect.c
+++ b/OvmfPkg/Library/PlatformInitLib/MemDetect.c
@@ -967,7 +967,7 @@ Switch4Level (
    Should that not be the case the function checks the size of the address space
    needed, which is the RAM installed plus fw_cfg reservations.  The downgrade
    to 4-level paging will happen for small guests where the address space needed
-   is lower than 1TB.
+   is lower than 16TB.
 
 
    This function will also log the paging level used and the reason for that.
@@ -1002,9 +1002,9 @@ PlatformSetupPagingLevel (
              );
   switch (Status) {
     case EFI_NOT_FOUND:
-      if (PlatformInfoHob->FirstNonAddress < (1ll << 40)) {
+      if (PlatformInfoHob->FirstNonAddress < (1ll << 44)) {
         //
-        // If the highest address actually used is below 1TB switch back into
+        // If the highest address actually used is below 16TB switch back into
         // 4-level paging mode for better compatibility with older guests.
         //
         DEBUG ((DEBUG_INFO, "%a: using 4-level paging (default for small guest)\n", __func__));
diff --git a/OvmfPkg/RUNTIME_CONFIG.md b/OvmfPkg/RUNTIME_CONFIG.md
index b75a5dacad..3bef1e6e4a 100644
--- a/OvmfPkg/RUNTIME_CONFIG.md
+++ b/OvmfPkg/RUNTIME_CONFIG.md
@@ -228,7 +228,7 @@ paging levels are better for compatibility, not all OSes can handle
 levels are better for huge VMs, the address space managed by OVMF can
 be larger.
 
-Default: Use 5-level paging in case more than 1TB of memory is
+Default: Use 5-level paging in case more than 16TB of memory is
 present, 4-level paging otherwise.
 
 Usage:
-- 
2.47.1

Re: [PATCH edk2 hci-8.1] OvmfPkg/PlatformInitLib: switch to 5-level paging only above 16TB #VSTOR-152672 #VSTOR-152405

[Denis V. Lunev]

On 11/2/26 16:25, Andrey Drobyshev wrote:
> With PcdUse5LevelPageTable the 64-bit PEI decides per guest whether to
> hand the OS over in 4- or 5-level paging mode.  Without a fw_cfg
> override it uses 5-level paging as soon as the guest address space
> (RAM plus the memory hotplug window) reaches 1TB.
>
> That threshold is too low for us:
>
>   * 4-level firmware paging already covers 64TB of address space, so
>     5-level paging below that buys nothing and only changes what the OS
>     sees at ExitBootServices;
>   * on Windows that means booting the LA57 kernel (ntkrla57.exe), which
>     bugchecks on AMD guests exposed to the ARCH_CAPABILITIES MSR
>     (VSTOR-152405) and breaks Windows Setup.  QEMU since the commit
>     7fc0e1c9f577 ("i386: hide la57 at fresh start on vz machine types
>     that show ARCH_CAPABILITIES #VSTOR-152405") now hides la57 on the
>     VZ machine types that expose the MSR, but the firmware shouldn't
>     depend on that.
>
> Raise the threshold to 16TB.  Ordinary guests, including those with a
> large hotplug window, stay in 4-level paging.
>
> Signed-off-by: Andrey Drobyshev <andrey.drobyshev@virtuozzo.com>
> ---
>  OvmfPkg/Library/PlatformInitLib/MemDetect.c | 6 +++---
>  OvmfPkg/RUNTIME_CONFIG.md                   | 2 +-
>  2 files changed, 4 insertions(+), 4 deletions(-)
>
> diff --git a/OvmfPkg/Library/PlatformInitLib/MemDetect.c b/OvmfPkg/Library/PlatformInitLib/MemDetect.c
> index 937e2b77a5..7220db8ef9 100644
> --- a/OvmfPkg/Library/PlatformInitLib/MemDetect.c
> +++ b/OvmfPkg/Library/PlatformInitLib/MemDetect.c
> @@ -967,7 +967,7 @@ Switch4Level (
>     Should that not be the case the function checks the size of the address space
>
>     needed, which is the RAM installed plus fw_cfg reservations.  The downgrade
>
>     to 4-level paging will happen for small guests where the address space needed
>
> -   is lower than 1TB.
>
> +   is lower than 16TB.
>
>  
>
>  
>
>     This function will also log the paging level used and the reason for that.
>
> @@ -1002,9 +1002,9 @@ PlatformSetupPagingLevel (
>               );
>
>    switch (Status) {
>
>      case EFI_NOT_FOUND:
>
> -      if (PlatformInfoHob->FirstNonAddress < (1ll << 40)) {
>
> +      if (PlatformInfoHob->FirstNonAddress < (1ll << 44)) {
>
>          //
>
> -        // If the highest address actually used is below 1TB switch back into
>
> +        // If the highest address actually used is below 16TB switch back into
>
>          // 4-level paging mode for better compatibility with older guests.
>
>          //
>
>          DEBUG ((DEBUG_INFO, "%a: using 4-level paging (default for small guest)\n", __func__));
>
> diff --git a/OvmfPkg/RUNTIME_CONFIG.md b/OvmfPkg/RUNTIME_CONFIG.md
> index b75a5dacad..3bef1e6e4a 100644
> --- a/OvmfPkg/RUNTIME_CONFIG.md
> +++ b/OvmfPkg/RUNTIME_CONFIG.md
> @@ -228,7 +228,7 @@ paging levels are better for compatibility, not all OSes can handle
>  levels are better for huge VMs, the address space managed by OVMF can
>
>  be larger.
>
>  
>
> -Default: Use 5-level paging in case more than 1TB of memory is
>
> +Default: Use 5-level paging in case more than 16TB of memory is
>
>  present, 4-level paging otherwise.
>
>  
>
>  Usage:
>
Reviewed-by: Denis V. Lunev <den@openvz.org>