Skip to contentSkip to conversation

x86/retpoline: Add clang support

PATCH In review Guenter Roeck 2 participants 9 messages Feb 7, 2018 – Feb 14, 2018 x86

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 x86/retpoline: Add clang support

clang has its own set of compiler options for retpoline support.

---
v2: llvm has been updated to use the same thunk names as gcc.

    Tested with:
    clang version 7.0.0
	(https://git.llvm.org/git/clang.git/
		848874aed95a913fb45f363120500cebfe54e2ef)
	(https://git.llvm.org/git/llvm.git/
		3afd566557f3616881505db0d69f5d19bf55ae14)
    cross-checked with gcc 7.3.0 (x86_64-linux-gcc.br_real (Buildroot
    2018.02-rc1) 7.3.0).

Tested with 64-bit builds only; 32-bit images fail to build with clang with various unrelated errors and are difficult to test.

I had to change '+=' to '=' below since make otherwise sets RETPOLINE_CFLAGS to " ", and the subsequent ifneq would always match. This is also the reason for the "ifeq ($(RETPOLINE_CFLAGS),)". If there is another/different/better way to handle this, please let me know.

There are curently lots of warnings when building an image with clang.

./include/linux/init.h:134:6: warning:
unknown attribute 'indirect_branch' ignored

I was inclined to add "&& !defined(__clang__)" to the condition for the __noretpoline define to fix the problem, but concluded that this should be a separate patch unless it can be addressed in clang.

 arch/x86/Makefile | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
arch/x86/Makefile +6 -1
index fad55160dcb9..c1a35cea88e1 100644
@@ -232,7 +232,12 @@ KBUILD_CFLAGS += -fno-asynchronous-unwind-tables
 
 # Avoid indirect branches in kernel to deal with Spectre
 ifdef CONFIG_RETPOLINE
-    RETPOLINE_CFLAGS += $(call cc-option,-mindirect-branch=thunk-extern -mindirect-branch-register)
+    # gcc
+    RETPOLINE_CFLAGS = $(call cc-option,-mindirect-branch=thunk-extern -mindirect-branch-register)
+    ifeq ($(RETPOLINE_CFLAGS),)
+	# clang
+	RETPOLINE_CFLAGS = $(call cc-option,-mretpoline-external-thunk)
+    endif
     ifneq ($(RETPOLINE_CFLAGS),)
         KBUILD_CFLAGS += $(RETPOLINE_CFLAGS) -DRETPOLINE
     endif
Signature · 1 line
2.7.4
#1 lore raw
Quoted 34 lines · Guenter Roeck

clang has its own set of compiler options for retpoline support.

---
v2: llvm has been updated to use the same thunk names as gcc. Tested with:
    clang version 7.0.0

	(https://git.llvm.org/git/clang.git/
		848874aed95a913fb45f363120500cebfe54e2ef)
	(https://git.llvm.org/git/llvm.git/
		3afd566557f3616881505db0d69f5d19bf55ae14)

    cross-checked with gcc 7.3.0 (x86_64-linux-gcc.br_real (Buildroot 2018.02-rc1) 7.3.0).

Quoted 8 lines

Tested with 64-bit builds only; 32-bit images fail to build with clang with various unrelated errors and are difficult to test.

I had to change '+=' to '=' below since make otherwise sets RETPOLINE_CFLAGS to " ", and the subsequent ifneq would always match. This is also the reason for the "ifeq ($(RETPOLINE_CFLAGS),)". If there is another/different/better way to handle this, please let me know.

See
http://git.infradead.org/users/dwmw2/linux-retpoline.git/commitdiff/82a1f41600

Quoted 8 lines

There are curently lots of warnings when building an image with clang.

./include/linux/init.h:134:6: warning:
unknown attribute 'indirect_branch' ignored

I was inclined to add "&& !defined(__clang__)" to the condition for the __noretpoline define to fix the problem, but concluded that this should be a separate patch unless it can be addressed in clang.

... where I fixed that too as well as providing the __x86_indirect_thunk for 32-bit builds (which are broken for other reasons anyway).

We do still need the %V asm constraint modifier to work in clang though, because otherwise it still doesn't build if you have any hypervisor guest support.

#2 lore raw
Quoted 40 lines · David Woodhouse
Quoted 37 lines · Guenter Roeck

clang has its own set of compiler options for retpoline support.

---
v2: llvm has been updated to use the same thunk names as gcc. Tested with:
    clang version 7.0.0

	(https://git.llvm.org/git/clang.git/
		848874aed95a913fb45f363120500cebfe54e2ef)
	(https://git.llvm.org/git/llvm.git/
		3afd566557f3616881505db0d69f5d19bf55ae14)

    cross-checked with gcc 7.3.0 (x86_64-linux-gcc.br_real (Buildroot 2018.02-rc1) 7.3.0).

Tested with 64-bit builds only; 32-bit images fail to build with clang with various unrelated errors and are difficult to test.

Quoted 9 lines
Quoted 5 lines

I had to change '+=' to '=' below since make otherwise sets RETPOLINE_CFLAGS to " ", and the subsequent ifneq would always match. This is also the reason for the "ifeq ($(RETPOLINE_CFLAGS),)". If there is another/different/better way to handle this, please let me know.

See
http://git.infradead.org/users/dwmw2/linux-retpoline.git/commitdiff/82a1f41600

Nice.

Quoted 6 lines
Quoted 6 lines

There are curently lots of warnings when building an image with clang.

./include/linux/init.h:134:6: warning:
unknown attribute 'indirect_branch' ignored

Quoted 8 lines
Quoted 3 lines

I was inclined to add "&& !defined(__clang__)" to the condition for the __noretpoline define to fix the problem, but concluded that this should be a separate patch unless it can be addressed in clang.

... where I fixed that too as well as providing the __x86_indirect_thunk for 32-bit builds (which are broken for other reasons anyway).

Much better and more comprehensive than my patch. Let's go with yours.

Quoted 3 lines

We do still need the %V asm constraint modifier to work in clang though, because otherwise it still doesn't build if you have any hypervisor guest support.

Let's see what Chandler's feedback is. Worst case we can handle that in a separate patch; we won't be able to fix all clang problems in a single patch anyway.

Thanks,
Guenter

#3 lore raw
In reply to David Woodhouse
Quoted 40 lines · David Woodhouse
Quoted 37 lines · Guenter Roeck

clang has its own set of compiler options for retpoline support.

---
v2: llvm has been updated to use the same thunk names as gcc. Tested with:
    clang version 7.0.0

	(https://git.llvm.org/git/clang.git/
		848874aed95a913fb45f363120500cebfe54e2ef)
	(https://git.llvm.org/git/llvm.git/
		3afd566557f3616881505db0d69f5d19bf55ae14)

    cross-checked with gcc 7.3.0 (x86_64-linux-gcc.br_real (Buildroot 2018.02-rc1) 7.3.0).

Tested with 64-bit builds only; 32-bit images fail to build with clang with various unrelated errors and are difficult to test.

Quoted 9 lines
Quoted 5 lines

I had to change '+=' to '=' below since make otherwise sets RETPOLINE_CFLAGS to " ", and the subsequent ifneq would always match. This is also the reason for the "ifeq ($(RETPOLINE_CFLAGS),)". If there is another/different/better way to handle this, please let me know.

See
http://git.infradead.org/users/dwmw2/linux-retpoline.git/commitdiff/82a1f41600

You have __x86_indirect_thunk within #ifdef CONFIG_64BIT. I think it should be in the else case.

Guenter

#5 lore raw
Quoted 9 lines · Guenter Roeck

Hi David,

On Wed, Feb 07, 2018 at 03:32:11PM -0800, Guenter Roeck wrote:
[ ... ]

Quoted 3 lines
Quoted 3 lines


See
http://git.infradead.org/users/dwmw2/linux-retpoline.git/commitdiff/82a1f41600

Any chance to see a patch with this change anytime soon ?

We were waiting for a fix for
https://bugs.llvm.org/show_bug.cgi?id=36329

I tested the candidate fix tonight and reported success; the corresponding patch is at the top of my tree. Since I think we now have a consensus on what the clang build will look like, I'll post that in the morning along with the other bits I have ready to send upstream.

We're going to need the percpu.h fix too, and I'd also like to see the status of the i915 build failure you mentioned. Is there a bug filed for that already, and is it on the blocker list for 6.0? If not, why not?

We also need to resolve the asm-goto thing.

#6 lore raw
Quoted 13 lines · David Woodhouse
Quoted 9 lines · Guenter Roeck

Hi David,

On Wed, Feb 07, 2018 at 03:32:11PM -0800, Guenter Roeck wrote:
[ ... ]

Quoted 3 lines
Quoted 3 lines

See
http://git.infradead.org/users/dwmw2/linux-retpoline.git/commitdiff/82a1f41600

Any chance to see a patch with this change anytime soon ?

Quoted 8 lines

We were waiting for a fix for
https://bugs.llvm.org/show_bug.cgi?id=36329

I tested the candidate fix tonight and reported success; the corresponding patch is at the top of my tree. Since I think we now have a consensus on what the clang build will look like, I'll post that in the morning along with the other bits I have ready to send upstream.

Excellent, thanks a lot! That the retpoline.S changes for clang are no longer required is excellent news.

Quoted 5 lines

We're going to need the percpu.h fix too, and I'd also like to see the status of the i915 build failure you mentioned. Is there a bug filed for that already, and is it on the blocker list for 6.0? If not, why not?

For my part I completely forgot about it :-(. Did you expect me to file a bug ? If so, I'll need to retest and try to figure out how to do that. Let me know; I should be able to get that done tomorrow.

Thanks,
Guenter

#7 lore raw
Quoted 9 lines · Guenter Roeck
Quoted 6 lines

We're going to need the percpu.h fix too, and I'd also like to see the status of the i915 build failure you mentioned. Is there a bug filed for that already, and is it on the blocker list for 6.0? If not, why not?

For my part I completely forgot about it :-(. Did you expect me to file a bug ? If so, I'll need to retest and try to figure out how to do that. Let me know; I should be able to get that done tomorrow.

I *thought* you had mentioned that it didn't build; in that case absolutely, I expected there to be a bug filed already. As a general rule, nobody should *ever* mention that something is broken, without first ensuring the bug is filed.

Looking back, I can't now see where you said that; maybe I made it up and I only observed this failure myself. Either way, I already filed
https://bugs.llvm.org/show_bug.cgi?id=36378 after searching bugzilla and not finding it.

Thanks.

#8 lore raw
Quoted 13 lines · David Woodhouse
Quoted 9 lines · Guenter Roeck
Quoted 6 lines

We're going to need the percpu.h fix too, and I'd also like to see the status of the i915 build failure you mentioned. Is there a bug filed for that already, and is it on the blocker list for 6.0? If not, why not?

For my part I completely forgot about it :-(. Did you expect me to file a bug ? If so, I'll need to retest and try to figure out how to do that. Let me know; I should be able to get that done tomorrow.

Quoted 7 lines

I *thought* you had mentioned that it didn't build; in that case absolutely, I expected there to be a bug filed already. As a general rule, nobody should *ever* mention that something is broken, without first ensuring the bug is filed.

Looking back, I can't now see where you said that; maybe I made it up and I only observed this failure myself. Either way, I already filed

I found it in my e-mail logs, so it definitely was me. Or maybe both of us in parallel.

Quoted 3 lines

https://bugs.llvm.org/show_bug.cgi?id=36378 after searching bugzilla and not finding it.

Thanks!
Guenter