On Wed, Sep 25, 2024 at 04:46:23PM -0700, Pawan Gupta wrote:
> On Thu, Sep 26, 2024 at 12:29:00AM +0100, Andrew Cooper wrote:
> > On 25/09/2024 11:25 pm, Pawan Gupta wrote:
> > > Robert Gill reported below #GP in 32-bit mode when dosemu software was
> > > executing vm86() system call:
> > >
> > > general protection fault: 0000 [#1] PREEMPT SMP
> > > CPU: 4 PID: 4610 Comm: dosemu.bin Not tainted 6.6.21-gentoo-x86 #1
> > > Hardware name: Dell Inc. PowerEdge 1950/0H723K, BIOS 2.7.0 10/30/2010
> > > EIP: restore_all_switch_stack+0xbe/0xcf
> > > EAX: 00000000 EBX: 00000000 ECX: 00000000 EDX: 00000000
> > > ESI: 00000000 EDI: 00000000 EBP: 00000000 ESP: ff8affdc
> > > DS: 0000 ES: 0000 FS: 0000 GS: 0033 SS: 0068 EFLAGS: 00010046
> > > CR0: 80050033 CR2: 00c2101c CR3: 04b6d000 CR4: 000406d0
> > > Call Trace:
> > > show_regs+0x70/0x78
> > > die_addr+0x29/0x70
> > > exc_general_protection+0x13c/0x348
> > > exc_bounds+0x98/0x98
> > > handle_exception+0x14d/0x14d
> > > exc_bounds+0x98/0x98
> > > restore_all_switch_stack+0xbe/0xcf
> > > exc_bounds+0x98/0x98
> > > restore_all_switch_stack+0xbe/0xcf
> > >
> > > This only happens in 32-bit mode when VERW based mitigations like MDS/RFDS
> > > are enabled. This is because segment registers with an arbitrary user value
> > > can result in #GP when executing VERW. Intel SDM vol. 2C documents the
> > > following behavior for VERW instruction:
> > >
> > > #GP(0) - If a memory operand effective address is outside the CS, DS, ES,
> > > FS, or GS segment limit.
> > >
> > > CLEAR_CPU_BUFFERS macro executes VERW instruction before returning to user
> > > space. Use %cs selector to reference VERW operand. This ensures VERW will
> > > not #GP for an arbitrary user %ds.
> > >
> > > Fixes: a0e2dab44d22 ("x86/entry_32: Add VERW just before userspace transition")
> > > Cc: stable@xxxxxxxxxxxxxxx # 5.10+
> > > Reported-by: Robert Gill <rtgill82@xxxxxxxxx>
> > > Closes: https://bugzilla.kernel.org/show_bug.cgi?id=218707
> > > Closes: https://lore.kernel.org/all/8c77ccfd-d561-45a1-8ed5-6b75212c7a58@xxxxxxxxxxxxx/
> > > Suggested-by: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
> > > Suggested-by: Brian Gerst <brgerst@xxxxxxxxx>
> > > Signed-off-by: Pawan Gupta <pawan.kumar.gupta@xxxxxxxxxxxxxxx>
> > > ---
> > > arch/x86/include/asm/nospec-branch.h | 6 ++++--
> > > 1 file changed, 4 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/arch/x86/include/asm/nospec-branch.h b/arch/x86/include/asm/nospec-branch.h
> > > index ff5f1ecc7d1e..e18a6aaf414c 100644
> > > --- a/arch/x86/include/asm/nospec-branch.h
> > > +++ b/arch/x86/include/asm/nospec-branch.h
> > > @@ -318,12 +318,14 @@
> > > /*
> > > * Macro to execute VERW instruction that mitigate transient data sampling
> > > * attacks such as MDS. On affected systems a microcode update overloaded VERW
> > > - * instruction to also clear the CPU buffers. VERW clobbers CFLAGS.ZF.
> > > + * instruction to also clear the CPU buffers. VERW clobbers CFLAGS.ZF. Using %cs
> > > + * to reference VERW operand avoids a #GP fault for an arbitrary user %ds in
> > > + * 32-bit mode.
> > > *
> > > * Note: Only the memory operand variant of VERW clears the CPU buffers.
> > > */
> > > .macro CLEAR_CPU_BUFFERS
> > > - ALTERNATIVE "", __stringify(verw _ASM_RIP(mds_verw_sel)), X86_FEATURE_CLEAR_CPU_BUF
> > > + ALTERNATIVE "", __stringify(verw %cs:_ASM_RIP(mds_verw_sel)), X86_FEATURE_CLEAR_CPU_BUF
> > > .endm
> >
> > People ought rightly to double-take at this using %cs and not %ss.
> > There is a good reason, but it needs describing explicitly. May I
> > suggest the following:
> >
> > *...
> > * In 32bit mode, the memory operand must be a %cs reference. The data
> > segments may not be usable (vm86 mode), and the stack segment may not be
> > flat (espfix32).
> > *...
>
> Thanks for the suggestion. I will include this.
>
> > .macro CLEAR_CPU_BUFFERS
> > #ifdef __x86_64__
> > ALTERNATIVE "", "verw mds_verw_sel(%rip)", X86_FEATURE_CLEAR_CPU_BUF
> > #else
> > ALTERNATIVE "", "verw %cs:mds_verw_sel", X86_FEATURE_CLEAR_CPU_BUF
> > #endif
> > .endm
> >
> > This also lets you drop _ASM_RIP(). It's a cute idea, but is more
> > confusion than it's worth, because there's no such thing in 32bit mode.
> >
> > "%cs:_ASM_RIP(mds_verw_sel)" reads as if it does nothing, because it
> > really doesn't in 64bit mode.
>
> Right, will drop _ASM_RIP() in 32-bit mode and %cs in 64-bit mode.
Its probably too soon for next version, pasting the patch here:
---8<---
diff --git a/arch/x86/include/asm/nospec-branch.h b/arch/x86/include/asm/nospec-branch.h
index e18a6aaf414c..4228a1fd2c2e 100644
--- a/arch/x86/include/asm/nospec-branch.h
+++ b/arch/x86/include/asm/nospec-branch.h
@@ -318,14 +318,21 @@
/*
* Macro to execute VERW instruction that mitigate transient data sampling
* attacks such as MDS. On affected systems a microcode update overloaded VERW
- * instruction to also clear the CPU buffers. VERW clobbers CFLAGS.ZF. Using %cs
- * to reference VERW operand avoids a #GP fault for an arbitrary user %ds in
- * 32-bit mode.
+ * instruction to also clear the CPU buffers. VERW clobbers CFLAGS.ZF.
*
* Note: Only the memory operand variant of VERW clears the CPU buffers.
*/
.macro CLEAR_CPU_BUFFERS
- ALTERNATIVE "", __stringify(verw %cs:_ASM_RIP(mds_verw_sel)), X86_FEATURE_CLEAR_CPU_BUF
+#ifdef CONFIG_X86_64
+ ALTERNATIVE "", __stringify(verw _ASM_RIP(mds_verw_sel)), X86_FEATURE_CLEAR_CPU_BUF
+#else
+ /*
+ * In 32bit mode, the memory operand must be a %cs reference. The data
+ * segments may not be usable (vm86 mode), and the stack segment may not
+ * be flat (ESPFIX32).
+ */
+ ALTERNATIVE "", __stringify(verw %cs:mds_verw_sel), X86_FEATURE_CLEAR_CPU_BUF
+#endif
.endm
#ifdef CONFIG_X86_64
