Message-ID: <7eb992f3-ddc0-41d3-98e3-e9020eb90cb0@intel.com>
Date: Wed, 24 Apr 2024 11:29:00 -0700
From: Dave Hansen <dave.hansen@...el.com>
To: "Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>,
 tglx@...utronix.de, mingo@...hat.com, bp@...en8.de
Cc: sathyanarayanan.kuppuswamy@...ux.intel.com, hpa@...or.com,
 seanjc@...gle.com, elena.reshetova@...el.com, rick.p.edgecombe@...el.com,
 x86@...nel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCHv2 3/4] x86/tdx: Handle PENDING_EPT_VIOLATION_V2

On 3/25/24 03:46, Kirill A. Shutemov wrote:
> PENDING_EPT_VIOLATION_V2 allows TD to control whether access to
> a pending page triggers #VE.
> 
> Kernel doesn't want to see any #VEs on accesses to private memory:
> disable such #VEs.

This is really a bare bones changelog.  The connection between "pending"
and "private" goes unmentioned.  It's not obvious when the kernel might
be exposed to one of these #VE's.

> diff --git a/arch/x86/coco/tdx/tdx.c b/arch/x86/coco/tdx/tdx.c
> index 08e2bb462ce8..860bfdd5a11d 100644
> --- a/arch/x86/coco/tdx/tdx.c
> +++ b/arch/x86/coco/tdx/tdx.c
> @@ -77,6 +77,17 @@ static inline void tdcall(u64 fn, struct tdx_module_args *args)
>  		panic("TDCALL %lld failed (Buggy TDX module!)\n", fn);
>  }
>  
> +static inline u64 tdg_vm_rd(u64 field)
> +{
> +	struct tdx_module_args args = {
> +		.rdx = field,
> +	};
> +
> +	tdcall(TDG_VM_RD, &args);
> +
> +	return args.r8;
> +}
> +
>  static inline u64 tdg_vm_wr(u64 field, u64 value, u64 mask)
>  {
>  	struct tdx_module_args args = {
> @@ -91,6 +102,28 @@ static inline u64 tdg_vm_wr(u64 field, u64 value, u64 mask)
>  	return args.r8;
>  }
>  
> +static inline u64 tdg_sys_rd(u64 field)
> +{
> +	struct tdx_module_args args = {
> +		.rdx = field,
> +	};
> +
> +	tdcall(TDG_SYS_RD, &args);
> +
> +	return args.r8;
> +}

I scratched my head for a minute on why these two identical functions
were introduced.  Then I realized that they're different TDCALLs.

Please break them out into a separate patch where you can mention this.
Also, the least we can do is a one-line comment about what they do:

/* Read metadata about how *this* VM is configured. */
static inline u64 tdg_vm_rd(u64 field)
..

/* Read metadata about the TDX system as whole. */
static inline u64 tdg_sys_rd(u64 field)

> +static bool tdcs_ctls_set(u64 mask)
> +{
> +	struct tdx_module_args args = {
> +		.rdx = TDCS_TD_CTLS,
> +		.r8 = mask,
> +		.r9 = mask,
> +	};
> +
> +	return __tdcall(TDG_VM_WR, &args) == TDX_SUCCESS;
> +}

This is acronym soup.  It's made worse by "TDCS_TD_CTLS" being
ungoogleable and not literally present in the first three TDX PDFs that
I dug in.

It's also not clear logically how this differs from tdg_vm_wr().

>  /**
>   * tdx_mcall_get_report0() - Wrapper to get TDREPORT0 (a.k.a. TDREPORT
>   *                           subtype 0) using TDG.MR.REPORT TDCALL.
> @@ -185,7 +218,8 @@ static void tdx_setup(u64 *cc_mask)
>  {
>  	struct tdx_module_args args = {};
>  	unsigned int gpa_width;
> -	u64 td_attr;
> +	u64 td_attr, features;
> +	bool sept_ve_disabled;
>  
>  	/*
>  	 * TDINFO TDX module call is used to get the TD execution environment
> @@ -206,19 +240,50 @@ static void tdx_setup(u64 *cc_mask)
>  	gpa_width = args.rcx & GENMASK(5, 0);
>  	*cc_mask = BIT_ULL(gpa_width - 1);
>  
> +	td_attr = args.rdx;
> +
>  	/* Kernel does not use NOTIFY_ENABLES and does not need random #VEs */
>  	tdg_vm_wr(TDCS_NOTIFY_ENABLES, 0, -1ULL);
>  
> +	features = tdg_sys_rd(TDCS_TDX_FEATURES0);
> +
>  	/*
>  	 * The kernel can not handle #VE's when accessing normal kernel
>  	 * memory.  Ensure that no #VE will be delivered for accesses to
>  	 * TD-private memory.  Only VMM-shared memory (MMIO) will #VE.
> +	 *
> +	 * Check if the TD is created with SEPT #VE disabled.
>  	 */
> -	td_attr = args.rdx;
> -	if (!(td_attr & ATTR_SEPT_VE_DISABLE)) {
> -		const char *msg = "TD misconfiguration: SEPT_VE_DISABLE attribute must be set.";
> +	sept_ve_disabled = td_attr & ATTR_SEPT_VE_DISABLE;
>  
> -		/* Relax SEPT_VE_DISABLE check for debug TD. */
> +	/*
> +	 * Check if flexible control of SEPT #VE is supported.
> +	 *
> +	 * The check consists of verifying if the feature is supported by the
> +	 * TDX module (the TDX_FEATURES0 check) and if the feature is enabled
> +	 * for this TD (CONFIG_FLAGS check).
> +	 *
> +	 * If flexible control is supported, disable SEPT #VE.
> +	 *
> +	 * Disable SEPT #VE regardless of ATTR_SEPT_VE_DISABLE status as
> +	 * flexible control allows software running before the kernel to
> +	 * enable it.
> +	 *
> +	 * Skip SEPT disabling for debug TD. SEPT #VE is unsafe but can be
> +	 * useful for debugging to produce a stack trace. Known to be useful
> +	 * for debugging unaccepted memory problems.
> +	 */
> +	if (features & TDX_FEATURES0_PENDING_EPT_VIOLATION_V2 &&
> +	    (tdg_vm_rd(TDCS_CONFIG_FLAGS) & TDCS_CONFIG_FLEXIBLE_PENDING_VE) &&

Can't we just check TDCS_CONFIG_FLEXIBLE_PENDING_VE alone?  Surely that
flag can't get set if !TDX_FEATURES0_PENDING_EPT_VIOLATION_V2.

In any case, this is, um, kinda messy.  Look at what a helper could do:

/*
 * Private memory #VE's are bad because the kernel has no way of
 * recovering from them and can only die. Newer TDX modules can
 * configured to inflict fewer #VE's on guests.  Do it when available.
 */
static inline int try_disable_sept_ve(td_addr)
{
	/* Does the TDX module support the #VE reduction feature? */
	if (!(features & TDX_FEATURES0_PENDING_EPT_VIOLATION_V2))
		return -ENOTSUP;

	/* Is this TD allowed to use the feature? */
	if (!(tdg_vm_rd(TDCS_CONFIG_FLAGS) &
	     TDCS_CONFIG_FLEXIBLE_PENDING_VE)))
		return -ENOTSUP;

	/* Keep #VE's enabled for splats in debugging environments: */
	if (td_attr & ATTR_DEBUG)
		return -ENOTSUP;

	if (!tdcs_ctls_set(TD_CTLS_PENDING_VE_DISABLE))
		return -ENOTSUP;

	return 0;
}

Then the code becomes:
	
	if (!try_disable_sept_ve(td_addr))
		sept_ve_disabled = true;
	else
		sept_ve_disabled = td_attr & ATTR_SEPT_VE_DISABLE;

See how it sets the state in *ONE* easy-to-understand place?

Oh, 'sept_ve_disabled' either gets disabled dynamically if it can, or it
gets read from the static state if not.

Oh, and if you ever have a comment that long with a bunch of
explanations of individual if() conditions, *PLEASE* try to break it up
in some way.  Otherwise readers just spent all their time trying to
match up the comment to the code.  It's not a fun game.

> +	    !(td_attr & ATTR_DEBUG)) {
> +		if (tdcs_ctls_set(TD_CTLS_PENDING_VE_DISABLE))
> +			sept_ve_disabled = true;
> +	}
> +
> +	if (!sept_ve_disabled) {
> +		const char *msg = "TD misconfiguration: SEPT #VE has to be disabled";
> +
> +		/* Relax SEPT #VE disable check for debug TD. */
>  		if (td_attr & ATTR_DEBUG)
>  			pr_warn("%s\n", msg);
>  		else
> diff --git a/arch/x86/include/asm/shared/tdx.h b/arch/x86/include/asm/shared/tdx.h
> index fdfd41511b02..29a61c72e4dd 100644
> --- a/arch/x86/include/asm/shared/tdx.h
> +++ b/arch/x86/include/asm/shared/tdx.h
> @@ -16,11 +16,27 @@
>  #define TDG_VP_VEINFO_GET		3
>  #define TDG_MR_REPORT			4
>  #define TDG_MEM_PAGE_ACCEPT		6
> +#define TDG_VM_RD			7
>  #define TDG_VM_WR			8
> +#define TDG_SYS_RD			11
>  
> -/* TDCS fields. To be used by TDG.VM.WR and TDG.VM.RD module calls */
> +/* TDX Global Metadata. To be used by TDG.SYS.RD */
> +#define TDCS_TDX_FEATURES0		0x0A00000300000008
> +
> +/* TDX TD-Scope Metadata. To be used by TDG.VM.WR and TDG.VM.RD */
> +#define TDCS_CONFIG_FLAGS		0x1110000300000016
> +#define TDCS_TD_CTLS			0x1110000300000017
>  #define TDCS_NOTIFY_ENABLES		0x9100000000000010
>  
> +/* TDCS_TDX_FEATURES0 bits */
> +#define TDX_FEATURES0_PENDING_EPT_VIOLATION_V2	BIT_ULL(16)
> +
> +/* TDCS_CONFIG_FLAGS bits */
> +#define TDCS_CONFIG_FLEXIBLE_PENDING_VE	BIT_ULL(1)
> +
> +/* TDCS_TD_CTLS bits */
> +#define TD_CTLS_PENDING_VE_DISABLE	BIT_ULL(0)
> +
>  /* TDX hypercall Leaf IDs */
>  #define TDVMCALL_MAP_GPA		0x10001
>  #define TDVMCALL_GET_QUOTE		0x10002

Powered by blists - more mailing lists