Re: [PATCH 2/2] x86: Add Support for Paging-Write Feature

[Tamas K Lengyel <tamas@xxxxxxxxxxxxx>]

On Thu, Dec 19, 2024 at 1:17 PM Petr Beneš <w1benny@xxxxxxxxx> wrote:
>
> From: Petr Beneš <w1benny@xxxxxxxxx>
>
> This patch introduces a new XENMEM_access_r_pw permission. Functionally, it 
> is similar to XENMEM_access_r, but for processors with 
> TERTIARY_EXEC_EPT_PAGING_WRITE support (Intel 12th Gen/Alder Lake and later), 
> it also permits the CPU to write to the page during guest page-table walks 
> (e.g., updating A/D bits) without triggering an EPT violation.
>
> This behavior works by both enabling the EPT paging-write feature and setting 
> the EPT paging-write flag in the EPT leaf entry.
>
> This feature provides a significant performance boost for introspection tools 
> that monitor guest page-table updates. Previously, every page-table 
> modification by the guest—including routine updates like setting A/D 
> bits—triggered an EPT violation, adding unnecessary overhead. The new 
> XENMEM_access_r_pw permission allows these "uninteresting" updates to occur 
> without EPT violations, improving efficiency.
>
> Additionally, this feature simplifies the handling of race conditions in 
> scenarios where an introspection tool:
>
> - Sets an "invisible breakpoint" in the altp2m view for a function F
> - Monitors guest page-table updates to track whether the page containing F is 
> paged out
> - Encounters a cleared Access (A) bit on the page containing F while the 
> guest is about to execute the breakpoint
>
> In the current implementation:
>
> - If xc_monitor_inguest_pagefault() is enabled, the introspection tool must 
> emulate both the breakpoint and the setting of the Access bit.
> - If xc_monitor_inguest_pagefault() is disabled, Xen handles the EPT 
> violation without notifying the introspection tool, setting the Access bit 
> and emulating the instruction. However, Xen fetches the instruction from the 
> default view instead of the altp2m view, potentially causing the breakpoint 
> to be missed.

I'm glad to see this mystery resolved! \o/

> diff --git a/xen/arch/x86/mm/p2m-ept.c b/xen/arch/x86/mm/p2m-ept.c
> index 21728397f9..5ad78ae4b5 100644
> --- a/xen/arch/x86/mm/p2m-ept.c
> +++ b/xen/arch/x86/mm/p2m-ept.c
> @@ -176,6 +176,10 @@ static void ept_p2m_type_to_flags(const struct 
> p2m_domain *p2m,
>              break;
>          case p2m_access_rwx:
>              break;
> +        case p2m_access_r_pw:
> +            entry->w = entry->x = 0;
> +            entry->pw = !!cpu_has_vmx_ept_paging_write;

I don't see ept_entry_t having a pw field. What's the deal there?

Tamas