Skip to contentSkip to conversation

x86/retpoline: Don't clobber RFLAGS during srso_safe_ret()

PATCH Applied Sean Christopherson 5 participants 7 messages Aug 11, 2023 – Aug 14, 2023

Subsystems (3)
lore
Share & tools
lore

Sean Christopherson

author reviewer
#0 lore raw

PATCH 1/1 x86/retpoline: Don't clobber RFLAGS during srso_safe_ret()

Use 'lea' instead of 'add' when adjusting %rsp in srso_safe_ret() so as to avoid clobbering flags. Drop one of the INT3 instructions to account for the LEA consuming one more byte than the ADD.

KVM's emulator makes indirect calls into a jump table of sorts, where the destination of each call is a small blob of code that performs fast emulation by executing the target instruction with fixed operands.

E.g. to emulate ADC, fastop() invokes adcb_al_dl():

adcb_al_dl:

      0xffffffff8105f5f0 <+0>:  adc    %dl,%al
      0xffffffff8105f5f2 <+2>:  jmp    0xffffffff81a39270 <__x86_return_thunk>

A major motivation for doing fast emulation is to leverage the CPU to handle consumption and manipulation of arithmetic flags, i.e. RFLAGS is both an input and output to the target of the call. fastop() collects the RFLAGS result by pushing RFLAGS onto the stack and popping them back into a variable (held in RDI in this case)

asm("push %[flags]; popf; " CALL_NOSPEC " ; pushf; pop %[flags]\n"

      0xffffffff81062be7 <+71>: mov    0xc0(%r8),%rdx
      0xffffffff81062bee <+78>: mov    0x100(%r8),%rcx
      0xffffffff81062bf5 <+85>: push   %rdi
      0xffffffff81062bf6 <+86>: popf
      0xffffffff81062bf7 <+87>: call   *%rsi
      0xffffffff81062bf9 <+89>: nop
      0xffffffff81062bfa <+90>: nop
      0xffffffff81062bfb <+91>: nop
      0xffffffff81062bfc <+92>: pushf
      0xffffffff81062bfd <+93>: pop    %rdi

and then propagating the arithmetic flags into the vCPU's emulator state:

    ctxt->eflags = (ctxt->eflags & ~EFLAGS_MASK) | (flags & EFLAGS_MASK);

      0xffffffff81062be0 <+64>:  and    $0xfffffffffffff72a,%r9
      0xffffffff81062bfe <+94>:  and    $0x8d5,%edi
      0xffffffff81062c0d <+109>: or     %rdi,%r9
      0xffffffff81062c1a <+122>: mov    %r9,0x10(%r8)

The failures can be most easily reproduced by running the "emulator" test in KVM-Unit-Tests.

If you're feeling a bit of deja vu, see commit b63f20a778c8 ("x86/retpoline: Don't clobber RFLAGS during CALL_NOSPEC on i386").

Fixes fb3bd914b3ec ("x86/srso: Add a Speculative RAS Overflow mitigation")
Cc Paolo Bonzini <[email protected]>

---

Those that fail to learn from history are doomed to repeat it. :-D

 arch/x86/lib/retpoline.S | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)
arch/x86/lib/retpoline.S +3 -4
index 2cff585f22f2..132cedbf9e57 100644
@@ -164,7 +164,7 @@ __EXPORT_THUNK(srso_untrain_ret_alias)
 /* Needs a definition for the __x86_return_thunk alternative below. */
 SYM_START(srso_safe_ret_alias, SYM_L_GLOBAL, SYM_A_NONE)
 #ifdef CONFIG_CPU_SRSO
-	add $8, %_ASM_SP
+	lea 8(%_ASM_SP), %_ASM_SP
 	UNWIND_HINT_FUNC
 #endif
 	ANNOTATE_UNRET_SAFE
@@ -239,7 +239,7 @@ __EXPORT_THUNK(zen_untrain_ret)
  * SRSO untraining sequence for Zen1/2, similar to zen_untrain_ret()
  * above. On kernel entry, srso_untrain_ret() is executed which is a
  *
- * movabs $0xccccccc308c48348,%rax
+ * movabs $0xccccc30824648d48,%rax
  *
  * and when the return thunk executes the inner label srso_safe_ret()
  * later, it is a stack manipulation and a RET which is mispredicted and
@@ -252,11 +252,10 @@ SYM_START(srso_untrain_ret, SYM_L_GLOBAL, SYM_A_NONE)
 	.byte 0x48, 0xb8
 
 SYM_INNER_LABEL(srso_safe_ret, SYM_L_GLOBAL)
-	add $8, %_ASM_SP
+	lea 8(%_ASM_SP), %_ASM_SP
 	ret
 	int3
 	int3
-	int3
 	lfence
 	call srso_safe_ret
 	int3

base-commit: 25aa0bebba72b318e71fe205bfd1236550cc9534

Signature · 1 line
2.41.0.694.ge786442a9b-goog
Tested-by on v1 1/1
#1 lore raw
Quoted 48 lines · Sean Christopherson

Use 'lea' instead of 'add' when adjusting %rsp in srso_safe_ret() so as to avoid clobbering flags. Drop one of the INT3 instructions to account for the LEA consuming one more byte than the ADD.

KVM's emulator makes indirect calls into a jump table of sorts, where the destination of each call is a small blob of code that performs fast emulation by executing the target instruction with fixed operands.

E.g. to emulate ADC, fastop() invokes adcb_al_dl():

adcb_al_dl:

      0xffffffff8105f5f0 <+0>:  adc    %dl,%al
      0xffffffff8105f5f2 <+2>:  jmp    0xffffffff81a39270 <__x86_return_thunk>

A major motivation for doing fast emulation is to leverage the CPU to handle consumption and manipulation of arithmetic flags, i.e. RFLAGS is both an input and output to the target of the call. fastop() collects the RFLAGS result by pushing RFLAGS onto the stack and popping them back into a variable (held in RDI in this case)

asm("push %[flags]; popf; " CALL_NOSPEC " ; pushf; pop %[flags]\n"

      0xffffffff81062be7 <+71>: mov    0xc0(%r8),%rdx
      0xffffffff81062bee <+78>: mov    0x100(%r8),%rcx
      0xffffffff81062bf5 <+85>: push   %rdi
      0xffffffff81062bf6 <+86>: popf
      0xffffffff81062bf7 <+87>: call   *%rsi
      0xffffffff81062bf9 <+89>: nop
      0xffffffff81062bfa <+90>: nop
      0xffffffff81062bfb <+91>: nop
      0xffffffff81062bfc <+92>: pushf
      0xffffffff81062bfd <+93>: pop    %rdi

and then propagating the arithmetic flags into the vCPU's emulator state:

    ctxt->eflags = (ctxt->eflags & ~EFLAGS_MASK) | (flags & EFLAGS_MASK);

      0xffffffff81062be0 <+64>:  and    $0xfffffffffffff72a,%r9
      0xffffffff81062bfe <+94>:  and    $0x8d5,%edi
      0xffffffff81062c0d <+109>: or     %rdi,%r9
      0xffffffff81062c1a <+122>: mov    %r9,0x10(%r8)

The failures can be most easily reproduced by running the "emulator" test in KVM-Unit-Tests.

If you're feeling a bit of deja vu, see commit b63f20a778c8 ("x86/retpoline: Don't clobber RFLAGS during CALL_NOSPEC on i386").

Quoted 7 lines
Fixes fb3bd914b3ec ("x86/srso: Add a Speculative RAS Overflow mitigation")
Cc Paolo Bonzini <[email protected]>

This resolves the issue I reported at [1].

[1]: https://lore.kernel.org/[email protected]/

Quoted 47 lines

---

Those that fail to learn from history are doomed to repeat it. :-D

 arch/x86/lib/retpoline.S | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)
arch/x86/lib/retpoline.S +3 -4
index 2cff585f22f2..132cedbf9e57 100644
@@ -164,7 +164,7 @@ __EXPORT_THUNK(srso_untrain_ret_alias)
 /* Needs a definition for the __x86_return_thunk alternative below. */
 SYM_START(srso_safe_ret_alias, SYM_L_GLOBAL, SYM_A_NONE)
 #ifdef CONFIG_CPU_SRSO
-	add $8, %_ASM_SP
+	lea 8(%_ASM_SP), %_ASM_SP
 	UNWIND_HINT_FUNC
 #endif
 	ANNOTATE_UNRET_SAFE
@@ -239,7 +239,7 @@ __EXPORT_THUNK(zen_untrain_ret)
  * SRSO untraining sequence for Zen1/2, similar to zen_untrain_ret()
  * above. On kernel entry, srso_untrain_ret() is executed which is a
  *
- * movabs $0xccccccc308c48348,%rax
+ * movabs $0xccccc30824648d48,%rax
  *
  * and when the return thunk executes the inner label srso_safe_ret()
  * later, it is a stack manipulation and a RET which is mispredicted and
@@ -252,11 +252,10 @@ SYM_START(srso_untrain_ret, SYM_L_GLOBAL, SYM_A_NONE)
 	.byte 0x48, 0xb8
 
 SYM_INNER_LABEL(srso_safe_ret, SYM_L_GLOBAL)
-	add $8, %_ASM_SP
+	lea 8(%_ASM_SP), %_ASM_SP
 	ret
 	int3
 	int3
-	int3
 	lfence
 	call srso_safe_ret
 	int3

base-commit: 25aa0bebba72b318e71fe205bfd1236550cc9534

Signature · 1 line
2.41.0.694.ge786442a9b-goog
maintainer
#2 lore raw
In reply to Sean Christopherson
Quoted 7 lines · Sean Christopherson

A major motivation for doing fast emulation is to leverage the CPU to handle consumption and manipulation of arithmetic flags, i.e. RFLAGS is both an input and output to the target of the call. fastop() collects the RFLAGS result by pushing RFLAGS onto the stack and popping them back into a variable (held in RDI in this case)

asm("push %[flags]; popf; " CALL_NOSPEC " ; pushf; pop %[flags]\n"

Right, and I've tested this countless times with gcc-built host and guest.

But Nathan's case where the host is built with gcc but the guest with clang, would trigger this. And as he confirms, that fixes it so I wonder what is the difference in code generation to make this rFLAGS corruption noticeable in that particular configuration.

Oh well, later when the fires are put out.

Thx.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
author reviewer
#3 lore raw
Quoted 9 lines · Borislav Petkov
Quoted 7 lines · Sean Christopherson

A major motivation for doing fast emulation is to leverage the CPU to handle consumption and manipulation of arithmetic flags, i.e. RFLAGS is both an input and output to the target of the call. fastop() collects the RFLAGS result by pushing RFLAGS onto the stack and popping them back into a variable (held in RDI in this case)

asm("push %[flags]; popf; " CALL_NOSPEC " ; pushf; pop %[flags]\n"

Quoted 7 lines

Right, and I've tested this countless times with gcc-built host and guest.

But Nathan's case where the host is built with gcc but the guest with clang, would trigger this. And as he confirms, that fixes it so I wonder what is the difference in code generation to make this rFLAGS corruption noticeable in that particular configuration.

Might be I/O APIC accesses? Unless things have changed, the I/O APIC code uses a struct overlay to access the I/O APIC, i.e. when doing emulated MMIO accesses. If clang generates an ADD or whatever and consumes flags, e.g. instead of a straight MOV, that would explain the problems.

#4 lore raw
In reply to Sean Christopherson
Quoted 98 lines · Sean Christopherson

Use 'lea' instead of 'add' when adjusting %rsp in srso_safe_ret() so as to avoid clobbering flags. Drop one of the INT3 instructions to account for the LEA consuming one more byte than the ADD.

KVM's emulator makes indirect calls into a jump table of sorts, where the destination of each call is a small blob of code that performs fast emulation by executing the target instruction with fixed operands.

E.g. to emulate ADC, fastop() invokes adcb_al_dl():

   adcb_al_dl:
       0xffffffff8105f5f0 <+0>:  adc    %dl,%al
       0xffffffff8105f5f2 <+2>:  jmp    0xffffffff81a39270 <__x86_return_thunk>

A major motivation for doing fast emulation is to leverage the CPU to handle consumption and manipulation of arithmetic flags, i.e. RFLAGS is both an input and output to the target of the call. fastop() collects the RFLAGS result by pushing RFLAGS onto the stack and popping them back into a variable (held in RDI in this case)

   asm("push %[flags]; popf; " CALL_NOSPEC " ; pushf; pop %[flags]\n"

       0xffffffff81062be7 <+71>: mov    0xc0(%r8),%rdx
       0xffffffff81062bee <+78>: mov    0x100(%r8),%rcx
       0xffffffff81062bf5 <+85>: push   %rdi
       0xffffffff81062bf6 <+86>: popf
       0xffffffff81062bf7 <+87>: call   *%rsi
       0xffffffff81062bf9 <+89>: nop
       0xffffffff81062bfa <+90>: nop
       0xffffffff81062bfb <+91>: nop
       0xffffffff81062bfc <+92>: pushf
       0xffffffff81062bfd <+93>: pop    %rdi

and then propagating the arithmetic flags into the vCPU's emulator state:

     ctxt->eflags = (ctxt->eflags & ~EFLAGS_MASK) | (flags & EFLAGS_MASK);

       0xffffffff81062be0 <+64>:  and    $0xfffffffffffff72a,%r9
       0xffffffff81062bfe <+94>:  and    $0x8d5,%edi
       0xffffffff81062c0d <+109>: or     %rdi,%r9
       0xffffffff81062c1a <+122>: mov    %r9,0x10(%r8)

The failures can be most easily reproduced by running the "emulator" test in KVM-Unit-Tests.

If you're feeling a bit of deja vu, see commit b63f20a778c8 ("x86/retpoline: Don't clobber RFLAGS during CALL_NOSPEC on i386").

Fixes fb3bd914b3ec ("x86/srso: Add a Speculative RAS Overflow mitigation")
Cc Paolo Bonzini <[email protected]>

---

Those that fail to learn from history are doomed to repeat it. :-D

  arch/x86/lib/retpoline.S | 7 +++----
  1 file changed, 3 insertions(+), 4 deletions(-)
arch/x86/lib/retpoline.S +3 -4
index 2cff585f22f2..132cedbf9e57 100644
@@ -164,7 +164,7 @@ __EXPORT_THUNK(srso_untrain_ret_alias)
  /* Needs a definition for the __x86_return_thunk alternative below. */
  SYM_START(srso_safe_ret_alias, SYM_L_GLOBAL, SYM_A_NONE)
  #ifdef CONFIG_CPU_SRSO
-	add $8, %_ASM_SP
+	lea 8(%_ASM_SP), %_ASM_SP
  	UNWIND_HINT_FUNC
  #endif
  	ANNOTATE_UNRET_SAFE
@@ -239,7 +239,7 @@ __EXPORT_THUNK(zen_untrain_ret)
   * SRSO untraining sequence for Zen1/2, similar to zen_untrain_ret()
   * above. On kernel entry, srso_untrain_ret() is executed which is a
   *
- * movabs $0xccccccc308c48348,%rax
+ * movabs $0xccccc30824648d48,%rax
   *
   * and when the return thunk executes the inner label srso_safe_ret()
   * later, it is a stack manipulation and a RET which is mispredicted and
@@ -252,11 +252,10 @@ SYM_START(srso_untrain_ret, SYM_L_GLOBAL, SYM_A_NONE)
  	.byte 0x48, 0xb8
 
  SYM_INNER_LABEL(srso_safe_ret, SYM_L_GLOBAL)
-	add $8, %_ASM_SP
+	lea 8(%_ASM_SP), %_ASM_SP
  	ret
  	int3
  	int3
-	int3
  	lfence
  	call srso_safe_ret
  	int3
Quoted 1 line

base-commit: 25aa0bebba72b318e71fe205bfd1236550cc9534

Don't we have the same kind of problems with __x86_return_skl ?

--Mika

author reviewer
#5 lore raw
Quoted 13 lines · Mika Penttilä
Quoted 13 lines
arch/x86/lib/retpoline.S +1 -2
@@ -252,11 +252,10 @@ SYM_START(srso_untrain_ret, SYM_L_GLOBAL, SYM_A_NONE)
  	.byte 0x48, 0xb8
  SYM_INNER_LABEL(srso_safe_ret, SYM_L_GLOBAL)
-	add $8, %_ASM_SP
+	lea 8(%_ASM_SP), %_ASM_SP
  	ret
  	int3
  	int3
-	int3
  	lfence
  	call srso_safe_ret
  	int3
 
Quoted 3 lines
Quoted 1 line

base-commit: 25aa0bebba72b318e71fe205bfd1236550cc9534

Don't we have the same kind of problems with __x86_return_skl ?

Yep, forcing that path via "retbleed=force retbleed=stuff spectre_v2=retpoline,generic" yields the same failures. I have no idea how to go about cleanly fixing that. The logic effectively requires modifying flags, the only thing I can think of is to save/restore flags across the thunk, which seems beyond gross.

Given that no one has complained about this, I think I'd vote to simply disable KVM if call depth tracking is being used.

#6 lore raw
In reply to Sean Christopherson
The following commit has been merged into the x86/urgent branch of tip:

The following commit has been merged into the x86/urgent branch of tip:

Commit-ID:     ba5ca5e5e6a1d55923e88b4a83da452166f5560e
Gitweb:        https://git.kernel.org/tip/ba5ca5e5e6a1d55923e88b4a83da452166f5560e
Author:        Sean Christopherson <[email protected]>
AuthorDate:    Fri, 11 Aug 2023 08:52:55 -07:00
Committer:     Borislav Petkov (AMD) <[email protected]>
CommitterDate: Mon, 14 Aug 2023 10:47:55 +02:00

x86/retpoline: Don't clobber RFLAGS during srso_safe_ret()

Use LEA instead of ADD when adjusting %rsp in srso_safe_ret{,_alias}() so as to avoid clobbering flags. Drop one of the INT3 instructions to account for the LEA consuming one more byte than the ADD.

KVM's emulator makes indirect calls into a jump table of sorts, where the destination of each call is a small blob of code that performs fast emulation by executing the target instruction with fixed operands.

E.g. to emulate ADC, fastop() invokes adcb_al_dl():

adcb_al_dl:

    <+0>:  adc    %dl,%al
    <+2>:  jmp    <__x86_return_thunk>

A major motivation for doing fast emulation is to leverage the CPU to handle consumption and manipulation of arithmetic flags, i.e. RFLAGS is both an input and output to the target of the call. fastop() collects the RFLAGS result by pushing RFLAGS onto the stack and popping them back into a variable (held in %rdi in this case):

asm("push %[flags]; popf; " CALL_NOSPEC " ; pushf; pop %[flags]\n"

  <+71>: mov    0xc0(%r8),%rdx
  <+78>: mov    0x100(%r8),%rcx
  <+85>: push   %rdi

<+86>: popf
<+87>: call *%rsi
<+89>: nop
<+90>: nop
<+91>: nop
<+92>: pushf
<+93>: pop %rdi

and then propagating the arithmetic flags into the vCPU's emulator state:

  ctxt->eflags = (ctxt->eflags & ~EFLAGS_MASK) | (flags & EFLAGS_MASK);
  <+64>:  and    $0xfffffffffffff72a,%r9
  <+94>:  and    $0x8d5,%edi
  <+109>: or     %rdi,%r9
  <+122>: mov    %r9,0x10(%r8)

The failures can be most easily reproduced by running the "emulator" test in KVM-Unit-Tests.

If you're feeling a bit of deja vu, see commit b63f20a778c8 ("x86/retpoline: Don't clobber RFLAGS during CALL_NOSPEC on i386").

In addition, this breaks booting of clang-compiled guest on a gcc-compiled host where the host contains the %rsp-modifying SRSO mitigations.

[ bp: Massage commit message, extend, remove addresses. ]

---
 arch/x86/lib/retpoline.S | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)
arch/x86/lib/retpoline.S +3 -4
index 2cff585..132cedb 100644
@@ -164,7 +164,7 @@ __EXPORT_THUNK(srso_untrain_ret_alias)
 /* Needs a definition for the __x86_return_thunk alternative below. */
 SYM_START(srso_safe_ret_alias, SYM_L_GLOBAL, SYM_A_NONE)
 #ifdef CONFIG_CPU_SRSO
-	add $8, %_ASM_SP
+	lea 8(%_ASM_SP), %_ASM_SP
 	UNWIND_HINT_FUNC
 #endif
 	ANNOTATE_UNRET_SAFE
@@ -239,7 +239,7 @@ __EXPORT_THUNK(zen_untrain_ret)
  * SRSO untraining sequence for Zen1/2, similar to zen_untrain_ret()
  * above. On kernel entry, srso_untrain_ret() is executed which is a
  *
- * movabs $0xccccccc308c48348,%rax
+ * movabs $0xccccc30824648d48,%rax
  *
  * and when the return thunk executes the inner label srso_safe_ret()
  * later, it is a stack manipulation and a RET which is mispredicted and
@@ -252,11 +252,10 @@ SYM_START(srso_untrain_ret, SYM_L_GLOBAL, SYM_A_NONE)
 	.byte 0x48, 0xb8
 
 SYM_INNER_LABEL(srso_safe_ret, SYM_L_GLOBAL)
-	add $8, %_ASM_SP
+	lea 8(%_ASM_SP), %_ASM_SP
 	ret
 	int3
 	int3
-	int3
 	lfence
 	call srso_safe_ret
 	int3