Re: [PATCH 4/7] x86/vsyscall: Setup vsyscall to compromise LASS protection

From: Chen, Yian
Date: Wed Jan 11 2023 - 20:43:36 EST




On 1/10/2023 4:34 PM, Sohil Mehta wrote:
On 1/9/2023 9:52 PM, Yian Chen wrote:

The user can also opt-out LASS in config file to build kernel
binary.

This line is unnecessary.

Sure, this line can be dropped.

Signed-off-by: Yian Chen <yian.chen@xxxxxxxxx>
Reviewed-by: Tony Luck <tony.luck@xxxxxxxxx>
---
  Documentation/admin-guide/kernel-parameters.txt | 12 ++++++++----
  arch/x86/entry/vsyscall/vsyscall_64.c           | 14 ++++++++++++++
  2 files changed, 22 insertions(+), 4 deletions(-)

diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
index 6cfa6e3996cf..3988e0c8c175 100644
--- a/Documentation/admin-guide/kernel-parameters.txt
+++ b/Documentation/admin-guide/kernel-parameters.txt
@@ -6755,10 +6755,14 @@
              versions of glibc use these calls.  Because these
              functions are at fixed addresses, they make nice
              targets for exploits that can control RIP.
-
-            emulate     [default] Vsyscalls turn into traps and are
-                        emulated reasonably safely.  The vsyscall
-                    page is readable.

The existing documentation here is incorrect. The default vsyscall mode is actually xonly. This has been so since:
commit 625b7b7f79c6 (x86/vsyscall: Change the default vsyscall mode to xonly)

Yes, you are right. but this patch can overwrite and correct existing one. I am assuming we don't need to correct the existing document first before update it for LASS.

+            In newer versions of Intel platforms that come with

Words such as "newer" in the kernel start losing meaning very quickly. Also, this comment looks out of place in between the vsyscall sub-options.

+            LASS(Linear Address Space separation) protection,
+            vsyscall is disabled by default. Enabling vsyscall
+            via the parameter overrides LASS protection.
+
Sure, I will take out this part change.


IIUC, you are making the default mode somewhat dynamic.

vsyscall = xonly (if LASS is not enabled)
vsyscall = none (if LASS is enabled)

Yes, this looks better.

The decision to disable vsyscall doesn't happen at compile time but it is taken at runtime when the LASS feature is detected. This would make the system behavior highly platform dependent.

Instead of doing this dance, can we provide a simplified behavior to the user/admin and move the decision making to compile time?

Current strategy is to disable vsyscall by default only for LASS capable platforms. So that the dynamic decision is likely a necessary.

Option 1: A bigger hammer
Set the default vsyscall option as CONFIG_LEGACY_VSYSCALL_NONE. CONFIG_X86_LASS is set by default. Changing the compile time VSYSCALL option would disable LASS.

This means to disable vsyscall by default for all platforms, doen't matter LASS. I am not sure if we want to go that far.

Option 2: A smaller hammer
CONFIG_X86_LASS is off by default. Vsyscall default stays as CONFIG_LEGACY_VSYSCALL_XONLY. Selecting LASS automatically chooses CONFIG_LEGACY_VSYSCALL_NONE. In this case, even if the hardware doesn't support LASS, vsyscall would still remain disabled.

This turns out to disable LASS by default. Then the LASS may not be taken in the first place.

In both of these cases, any command line input to override the vsyscall behavior can disable LASS as well.


+            emulate     [default if not LASS capable] Vsyscalls
+                    turn into traps and are emulated reasonably
+                    safely.  The vsyscall page is readable.
              xonly       Vsyscalls turn into traps and are
                          emulated reasonably safely.  The vsyscall
diff --git a/arch/x86/entry/vsyscall/vsyscall_64.c b/arch/x86/entry/vsyscall/vsyscall_64.c
index 4af81df133ee..2691f26835d1 100644
--- a/arch/x86/entry/vsyscall/vsyscall_64.c
+++ b/arch/x86/entry/vsyscall/vsyscall_64.c
@@ -63,6 +63,12 @@ static int __init vsyscall_setup(char *str)
          else
              return -EINVAL;
+        if (cpu_feature_enabled(X86_FEATURE_LASS) &&
+            vsyscall_mode != NONE) {
+            setup_clear_cpu_cap(X86_FEATURE_LASS);
+            pr_warn("LASS disabled by command line enabling vsyscall\n");

A warning seems too drastic here. A pr_info() should probably suffice.

sure I will modify it to use pr_info.

+        }
+
          return 0;
      }
@@ -379,6 +385,14 @@ void __init map_vsyscall(void)
      extern char __vsyscall_page;
      unsigned long physaddr_vsyscall = __pa_symbol(&__vsyscall_page);
+    /*
+     * When LASS is on, vsyscall triggers a #GP fault,
+     * so that force vsyscall_mode to NONE.
+     */

This comment doesn't make much sense nor does it provide the full picture. Some of the reasoning from the cover letter/commit log can be duplicated here.

sure, How about a more detail inline comment as following:
+ /*
+ * When LASS protection is on, user space vsyscall triggers
+ * a #GP fault since the vsyscall page is
+ * 0xffffffffff600000-0xffffffffff601000,
+ * so that force vsyscall_mode to NONE and disable this mapping.
+ */

+    if (cpu_feature_enabled(X86_FEATURE_LASS)) {
+        vsyscall_mode = NONE;
+        return;
+    }
      /*
       * For full emulation, the page needs to exist for real.  In
       * execute-only mode, there is no PTE at all backing the vsyscall

Thanks,
Yian