Re: [PATCH] x86/tsx: Use MSR_TSX_CTRL to clear CPUID bits

[Pawan Gupta]

On 15.02.2022 17:28, Pawan Gupta wrote:
> On 16.02.2022 00:49, Andrew Cooper wrote:
> > On 16/02/2022 00:39, Pawan Gupta wrote:
> > > On 15.02.2022 20:33, Borislav Petkov wrote:
> > > > On Tue, Feb 15, 2022 at 10:19:31AM -0800, Pawan Gupta wrote:
> > > > > I admit it has gotten complicated with so many bits associated with
> > > > > TSX.
> > > >
> > > > Yah, and looka here:
> > > >
> > > > https://github.com/andyhhp/xen/commit/ad9f7c3b2e0df38ad6d54f4769d4dccf765fbcee
> > > >
> > > >
> > > > It seems it isn't complicated enough. ;-\
> > > >
> > > > Andy just made me aware of this thing where you guys have added a new
> > > > MSR bit:
> > > >
> > > > MSR_MCU_OPT_CTRL[1] which is called something like
> > > > MCU_OPT_CTRL_RTM_ALLOW or so.
> > >
> > > RTM_ALLOW bit was added to MSR_MCU_OPT_CTRL, but its not set by default,
> > > and it is *not* recommended to be used in production deployments [1]:
> > >
> > >   Although MSR 0x122 (TSX_CTRL) and MSR 0x123 (IA32_MCU_OPT_CTRL) can be
> > >   used to reenable Intel TSX for development, doing so is not recommended
> > >   for production deployments. In particular, applying MD_CLEAR flows for
> > >   mitigation of the Intel TSX Asynchronous Abort (TAA) transient
> > > execution
> > >   attack may not be effective on these processors when Intel TSX is
> > >   enabled with updated microcode. The processors continue to be mitigated
> > >   against TAA when Intel TSX is disabled.
> >
> > The purpose of setting RTM_ALLOW isn't to enable TSX per say.
> >
> > The purpose is to make MSR_TSX_CTRL.RTM_DISABLE behaves consistently on
> > all hardware, which reduces the complexity and invasiveness of dealing
> > with this special case, because the TAA workaround will still turn TSX
> > off by default.
> >
> > The configuration you don't want to be running with is RTM_ALLOW &&
> > !RTM_DISABLE, because that is "still vulnerable to TSX Async Abort".
>
> I am not sure how a system can end up with RTM_ALLOW && !RTM_DISABLE? I
> have no problem taking care of this case, I am just debating why we need
> it.
>
> One way to get to this state is BIOS sets RTM_ALLOW (dont know why?) and
> linux cmdline has tsx=on.

If RTM_ALLOW is set for any reason, we can still deny tsx=on request.
Below patch should do that (I haven't tested it yet).

Alternatively, we can reset RTM_ALLOW, which will set
CPUID.RTM_ALWAYS_ABORT and may require re-enumeration of CPU features or
otherwise setup_force_cpu_cap(X86_FEATURE_RTM_ALWAYS_ABORT) should also
work.

---
diff --git a/arch/x86/include/asm/msr-index.h b/arch/x86/include/asm/msr-index.h
index 3faf0f97edb1..2ef58bcfb1e4 100644
--- a/arch/x86/include/asm/msr-index.h
+++ b/arch/x86/include/asm/msr-index.h
@@ -131,6 +131,8 @@
 /* SRBDS support */
 #define MSR_IA32_MCU_OPT_CTRL		0x00000123
 #define RNGDS_MITG_DIS			BIT(0)
+#define MCU_OPT_CTRL_RTM_ALLOW		BIT(1)
+#define MCU_OPT_CTRL_RTM_LOCKED		BIT(2)
 
 #define MSR_IA32_SYSENTER_CS		0x00000174
 #define MSR_IA32_SYSENTER_ESP		0x00000175
diff --git a/arch/x86/kernel/cpu/tsx.c b/arch/x86/kernel/cpu/tsx.c
index 2835fa89fc6f..8ac085ac597f 100644
--- a/arch/x86/kernel/cpu/tsx.c
+++ b/arch/x86/kernel/cpu/tsx.c
@@ -142,6 +142,29 @@ void tsx_clear_cpuid(void)
 	}
 }
 
+/*
+ * When the microcode released in Feb 2022 is applied, TSX will be disabled by
+ * default on some processors. MSR 0x122 (TSX_CTRL) and MSR 0x123
+ * (IA32_MCU_OPT_CTRL) can be used to re-enable TSX for development, doing so is
+ * not recommended for production deployments. In particular, applying MD_CLEAR
+ * flows for mitigation of the Intel TSX Asynchronous Abort (TAA) transient
+ * execution attack may not be effective on these processors when TSX is
+ * enabled with updated microcode.
+ *
+ * Detect if this unsafe TSX development mode is enabled.
+ */
+static bool tsx_is_unsafe(void)
+{
+	u64 mcu_opt_ctrl;
+
+	if (!boot_cpu_has(X86_FEATURE_SRBDS_CTRL))
+		return false;
+
+	rdmsrl(MSR_IA32_MCU_OPT_CTRL, mcu_opt_ctrl);
+
+	return !!(mcu_opt_ctrl & MCU_OPT_CTRL_RTM_ALLOW);
+}
+
 void __init tsx_init(void)
 {
 	char arg[5] = {};
@@ -163,6 +186,13 @@ void __init tsx_init(void)
 	if (!tsx_ctrl_is_supported()) {
 		tsx_ctrl_state = TSX_CTRL_NOT_SUPPORTED;
 		return;
+	} else if (tsx_is_unsafe()) {
+		/* Do not allow tsx=on, when TSX is unsafe to use */
+		tsx_ctrl_state = TSX_CTRL_DISABLE;
+		tsx_disable();
+		setup_clear_cpu_cap(X86_FEATURE_RTM);
+		setup_clear_cpu_cap(X86_FEATURE_HLE);
+		return;
 	}
 
 	ret = cmdline_find_option(boot_command_line, "tsx", arg, sizeof(arg));