Skip to contentSkip to conversation

arch/riscv: vdso: remove CFI landing pad from rt_sigreturn

PATCH In review Aurelien Jarno 3 participants 5 messages Jun 23 – Jul 15 RISC-V

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 arch/riscv: vdso: remove CFI landing pad from rt_sigreturn

When CONFIG_RISCV_USER_CFI is enabled, the CFI version of the vDSO, has a CFI landing pad instruction at the start of __vdso_rt_sigreturn. This breaks libgcc's unwinding code which matches on the first two instructions. Other unwinders that rely on similar instruction matching may also be affected.

Since __vdso_rt_sigreturn is reached as part of signal-return handling rather than via an indirect call/jump from userspace, it does not need a CFI landing pad. Remove it and restore the instruction sequence expected by existing unwinding code.

This matches what was done on arm64 in commit 9a964285572b ("arm64:
vdso: Don't prefix sigreturn trampoline with a BTI C instruction") for a similar issue.

Fixes 37f57bd3faea ("arch/riscv: compile vdso with landing pad and shadow stack note")
---
 arch/riscv/kernel/vdso/rt_sigreturn.S | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)
arch/riscv/kernel/vdso/rt_sigreturn.S +8 -1
index e82987dc37394..f6b053d3bb798 100644
@@ -7,11 +7,18 @@
 #include <asm/unistd.h>
 #include <asm/assembler.h>
 
+/* WARNING: Do NOT add a CFI landing pad at the start of this function.
+ * Unwinders such as libgcc identify the sigreturn trampoline by matching the
+ * instruction sequence. Adding a landing pad here would break unwinding from
+ * signal handlers.
+ *
+ * This trampoline is used only for signal return and not via an indirect
+ * call/jump from userspace, so adding CFI landing pad is unnecessary.
+ */
 	.text
 SYM_FUNC_START(__vdso_rt_sigreturn)
 	.cfi_startproc
 	.cfi_signal_frame
-	vdso_lpad
 	li a7, __NR_rt_sigreturn
 	ecall
 	.cfi_endproc
Signature · 1 line
2.53.0
maintainer
#1 lore raw

Hi Joel,

Quoted 11 lines · Aurelien Jarno

When CONFIG_RISCV_USER_CFI is enabled, the CFI version of the vDSO, has a CFI landing pad instruction at the start of __vdso_rt_sigreturn. This breaks libgcc's unwinding code which matches on the first two instructions. Other unwinders that rely on similar instruction matching may also be affected.

Since __vdso_rt_sigreturn is reached as part of signal-return handling rather than via an indirect call/jump from userspace, it does not need a CFI landing pad. Remove it and restore the instruction sequence expected by existing unwinding code.

Quoted 7 lines

This matches what was done on arm64 in commit 9a964285572b ("arm64:
vdso: Don't prefix sigreturn trampoline with a BTI C instruction") for a similar issue.

Fixes 37f57bd3faea ("arch/riscv: compile vdso with landing pad and shadow stack note")

I'd like to take this one. But since you're listed as the co-author, I
probably need your Signed-off-by: as well. Care to reply with it?

thanks

- Paul

#2 lore raw
Quoted 20 lines · Paul Walmsley

Hi Joel,

Quoted 15 lines · Aurelien Jarno

When CONFIG_RISCV_USER_CFI is enabled, the CFI version of the vDSO, has a CFI landing pad instruction at the start of __vdso_rt_sigreturn. This breaks libgcc's unwinding code which matches on the first two instructions. Other unwinders that rely on similar instruction matching may also be affected.

Since __vdso_rt_sigreturn is reached as part of signal-return handling rather than via an indirect call/jump from userspace, it does not need a CFI landing pad. Remove it and restore the instruction sequence expected by existing unwinding code.

This matches what was done on arm64 in commit 9a964285572b ("arm64:
vdso: Don't prefix sigreturn trampoline with a BTI C instruction") for a similar issue.

Quoted 6 lines
Quoted 3 lines
Fixes 37f57bd3faea ("arch/riscv: compile vdso with landing pad and shadow stack note")

I'd like to take this one. But since you're listed as the co-author, I
probably need your Signed-off-by: as well. Care to reply with it?

Thanks Aurelien for working with me on this one. I saw CFI was disabled in the debian kernel config and wondered why (failing glibc backtrace tests with CFI enabled), and we got to the bottom of it.

This one should go to stable IMO.

Cheers,

Joel

#3 lore raw

Hi Paul,

Quoted 28 lines · Joel Stanley
Quoted 20 lines · Paul Walmsley

Hi Joel,

Quoted 15 lines · Aurelien Jarno

When CONFIG_RISCV_USER_CFI is enabled, the CFI version of the vDSO, has a CFI landing pad instruction at the start of __vdso_rt_sigreturn. This breaks libgcc's unwinding code which matches on the first two instructions. Other unwinders that rely on similar instruction matching may also be affected.

Since __vdso_rt_sigreturn is reached as part of signal-return handling rather than via an indirect call/jump from userspace, it does not need a CFI landing pad. Remove it and restore the instruction sequence expected by existing unwinding code.

This matches what was done on arm64 in commit 9a964285572b ("arm64:
vdso: Don't prefix sigreturn trampoline with a BTI C instruction") for a similar issue.

Quoted 6 lines
Quoted 3 lines
Fixes 37f57bd3faea ("arch/riscv: compile vdso with landing pad and shadow stack note")

I'd like to take this one. But since you're listed as the co-author, I
probably need your Signed-off-by: as well. Care to reply with it?

Quoted 7 lines

Thanks Aurelien for working with me on this one. I saw CFI was disabled in the debian kernel config and wondered why (failing glibc backtrace tests with CFI enabled), and we got to the bottom of it.

This one should go to stable IMO.

Now that Joel answered with a Signed-off-by, would it be possible to get this merged for 7.2?

Thanks
Aurelien

Signature · 7 lines
Aurelien Jarno                          GPG: 4096R/1DDD8C9B
[email protected]                     http://aurel32.net

_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv
maintainer
#4 lore raw

Hi Aurelien, Joel,

Quoted 28 lines · Aurelien Jarno
Quoted 27 lines · Joel Stanley
Quoted 17 lines · Paul Walmsley
Quoted 15 lines · Aurelien Jarno

When CONFIG_RISCV_USER_CFI is enabled, the CFI version of the vDSO, has a CFI landing pad instruction at the start of __vdso_rt_sigreturn. This breaks libgcc's unwinding code which matches on the first two instructions. Other unwinders that rely on similar instruction matching may also be affected.

Since __vdso_rt_sigreturn is reached as part of signal-return handling rather than via an indirect call/jump from userspace, it does not need a CFI landing pad. Remove it and restore the instruction sequence expected by existing unwinding code.

This matches what was done on arm64 in commit 9a964285572b ("arm64:
vdso: Don't prefix sigreturn trampoline with a BTI C instruction") for a similar issue.

Quoted 6 lines
Quoted 3 lines
Fixes 37f57bd3faea ("arch/riscv: compile vdso with landing pad and shadow stack note")

I'd like to take this one. But since you're listed as the co-author, I
probably need your Signed-off-by: as well. Care to reply with it?

Quoted 8 lines
Quoted 5 lines

Thanks Aurelien for working with me on this one. I saw CFI was disabled in the debian kernel config and wondered why (failing glibc backtrace tests with CFI enabled), and we got to the bottom of it.

This one should go to stable IMO.

Now that Joel answered with a Signed-off-by, would it be possible to get this merged for 7.2?

Of course; thank you for the reminder. Thanks for the patch and the S-o-b, queued for v7.2-rc with a cc: to stable@vger.

- Paul