Skip to contentSkip to conversation

Fixups for fcf-protection

PATCH In review Jason Andryuk 4 participants 13 messages May 12, 2020 – May 13, 2020

lore
Share & tools
lore
#0 lore raw

cover Fixups for fcf-protection

Two patches to fix building with a cf-protection toolchain. The first handles the case where the compiler fails to run with "error:
‘-mindirect-branch’ and ‘-fcf-protection’ are not compatible".

The second fixes a runtime error that prevented Xen booting in legacy mode.

I still haven't figured out exactly what is wrong with rombios and/or ipxe.

Jason Andryuk (2):
xen/x86: Disable fcf-protection when necessary to build x86/boot: Drop .note.gnu.properties in build32.lds

 xen/arch/x86/arch.mk          | 11 ++++++++++-
 xen/arch/x86/boot/build32.lds |  5 +++++
 2 files changed, 15 insertions(+), 1 deletion(-)
Signature · 1 line
2.25.1
#1 lore raw

PATCH 1/2 xen/x86: Disable fcf-protection when necessary to build

Ubuntu gcc-9 enables -fcf-protection by default, which conflicts with -mindirect-branch=extern and prevents building the hypervisor with CONFIG_INDIRECT_THUNK:
xmalloc.h:81:1: error: ‘-mindirect-branch’ and ‘-fcf-protection’ are not compatible

Detect this incompatible combination, and add -fcf-protection=none to allow the build to succeed. To actually generated the error, the compiled program must include a function.

---
 xen/arch/x86/arch.mk | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
xen/arch/x86/arch.mk +10 -1
index 2a51553edb..3aa6ce521a 100644
@@ -63,7 +63,16 @@ CFLAGS += -mno-red-zone -fpic -fno-asynchronous-unwind-tables
 CFLAGS += -mno-sse $(call cc-option,$(CC),-mskip-rax-setup)
 
 # Compile with thunk-extern, indirect-branch-register if avaiable.
-CFLAGS-$(CONFIG_INDIRECT_THUNK) += -mindirect-branch=thunk-extern
+# Some versions of gcc error: "‘-mindirect-branch’ and ‘-fcf-protection’ are
+# not compatible".  For those, we need to disable cf-protection with
+# -fcf-protection=none
+cc-mindirect-branch = $(shell if test -n "`echo 'void foo(void) {};' | \
+      LANG=C $(CC) -mindirect-branch=thunk-extern -S -o /dev/null -x c - 2>&1 | \
+      grep -- '-mindirect-branch.*-fcf-protection.*are not compatible' -`"; \
+    then echo "-mindirect-branch=thunk-extern -fcf-protection=none"; \
+    else echo "-mindirect-branch=thunk-extern"; fi ;)
+
+CFLAGS-$(CONFIG_INDIRECT_THUNK) += $(call cc-mindirect-branch)
 CFLAGS-$(CONFIG_INDIRECT_THUNK) += -mindirect-branch-register
 CFLAGS-$(CONFIG_INDIRECT_THUNK) += -fno-jump-tables
 
Signature · 1 line
2.25.1
#2 lore raw

PATCH 2/2 x86/boot: Drop .note.gnu.properties in build32.lds

In reply to Jason Andryuk

reloc.S and cmdline.S as arrays of executable bytes for inclusion in head.S generated from compiled object files. Object files generated by an -fcf-protection toolchain include a .note.gnu.property section. The way reloc.S and cmdline.S are generated, the bytes of .note.gnu.property become the start of the .S files. When head.S calls reloc or cmdline_parse_early, those note bytes are executed instead of the intended .text section. This results in an early crash in reloc.

Discard the .note.gnu.property section when linking to avoid the extra bytes.

Stefan Bader also noticed that build32.mk requires -fcf-protection=none or else the hypervisor will not boot.
https://bugs.launchpad.net/ubuntu/+source/gcc-9/+bug/1863260

---
 xen/arch/x86/boot/build32.lds | 5 +++++
 1 file changed, 5 insertions(+)
xen/arch/x86/boot/build32.lds +5 -0
index da35aee910..0f69765579 100644
@@ -48,5 +48,10 @@ SECTIONS
          * Please check build32.mk for more details.
          */
         /* *(.got.plt) */
+        /*
+         * The note will end up before the .text section and be called
+         * incorrectly as instructions.
+         */
+        *(.note.gnu.property)
   }
 }
Signature · 1 line
2.25.1
#3 lore raw

Re: [PATCH 2/2] x86/boot: Drop .note.gnu.properties in build32.lds

Quoted 7 lines · Jason Andryuk

reloc.S and cmdline.S as arrays of executable bytes for inclusion in head.S generated from compiled object files. Object files generated by an -fcf-protection toolchain include a .note.gnu.property section. The way reloc.S and cmdline.S are generated, the bytes of .note.gnu.property become the start of the .S files. When head.S calls reloc or cmdline_parse_early, those note bytes are executed instead of the intended .text section. This results in an early crash in reloc.

I may be misremembering, but I vaguely recall some similar change suggestion. What I'm missing here is some form of statement as to whether this is legitimate tool chain behavior, or a bug, and hence whether this is a fix or a workaround.

Quoted 2 lines

Discard the .note.gnu.property section when linking to avoid the extra bytes.

If we go this route (and if, as per above, I'm misremembering, meaning we didn't reject such a change earlier on), why would we not strip .note and .note.* in one go?

Quoted 3 lines

Stefan Bader also noticed that build32.mk requires -fcf-protection=none or else the hypervisor will not boot.
https://bugs.launchpad.net/ubuntu/+source/gcc-9/+bug/1863260

How's this related to the change here?

Jan

#4 lore raw

Re: [PATCH 2/2] x86/boot: Drop .note.gnu.properties in build32.lds

Quoted 10 lines · Jan Beulich

[CAUTION - EXTERNAL EMAIL] DO NOT reply, click links, or open attachments unless you have verified the sender and know the content is safe.

Quoted 7 lines · Jason Andryuk

reloc.S and cmdline.S as arrays of executable bytes for inclusion in head.S generated from compiled object files. Object files generated by an -fcf-protection toolchain include a .note.gnu.property section. The way reloc.S and cmdline.S are generated, the bytes of .note.gnu.property become the start of the .S files. When head.S calls reloc or cmdline_parse_early, those note bytes are executed instead of the intended .text section. This results in an early crash in reloc.

Quoted 4 lines

I may be misremembering, but I vaguely recall some similar change suggestion. What I'm missing here is some form of statement as to whether this is legitimate tool chain behavior, or a bug, and hence whether this is a fix or a workaround.

The linker is free to position unreferenced sections anywhere it wishes.

It is deeply unhelpful behaviour, but neither Binutils nor Clang developers think it is something wanting fixing.

One option might be to use --orphan-handling=error so unexpected toolchain behaviour breaks the build, or in this case perhaps =discard might be better.

Quoted 10 lines
Quoted 2 lines

Discard the .note.gnu.property section when linking to avoid the extra bytes.

If we go this route (and if, as per above, I'm misremembering, meaning we didn't reject such a change earlier on), why would we not strip .note and .note.* in one go?

Quoted 3 lines

Stefan Bader also noticed that build32.mk requires -fcf-protection=none or else the hypervisor will not boot.
https://bugs.launchpad.net/ubuntu/+source/gcc-9/+bug/1863260

How's this related to the change here?

I think there is a bit of confusion as to exactly what is going on.

Ubuntu defaults -fcf-protection to enabled, which has a side effect of turning on CET, which inserts ENDBR{32,64} instructions and generates .note.gnu.properties indicating that the binary is CET-IBT compatible.

ENDBR* instructions come from the Hint Nop space so are safe on older processors, but do ultimately add to binary bloat.  It also occurs to me that it likely breaks livepath.

The reason Xen fails to boot is purely to do with the position of .note.gnu.properties, not the ENDBR* instructions.

~Andrew

#5 lore raw

Re: [PATCH 2/2] x86/boot: Drop .note.gnu.properties in build32.lds

Quoted 11 lines · Andrew Cooper
Quoted 10 lines · Jan Beulich
Quoted 2 lines · Jason Andryuk

Discard the .note.gnu.property section when linking to avoid the extra bytes.

If we go this route (and if, as per above, I'm misremembering, meaning we didn't reject such a change earlier on), why would we not strip .note and .note.* in one go?

Quoted 3 lines

Stefan Bader also noticed that build32.mk requires -fcf-protection=none or else the hypervisor will not boot.
https://bugs.launchpad.net/ubuntu/+source/gcc-9/+bug/1863260

Quoted 7 lines
Quoted 1 line

How's this related to the change here?

I think there is a bit of confusion as to exactly what is going on.

Ubuntu defaults -fcf-protection to enabled, which has a side effect of turning on CET, which inserts ENDBR{32,64} instructions and generates .note.gnu.properties indicating that the binary is CET-IBT compatible.

I.e. in principle this 2nd patch wouldn't be necessary at all if we forced -fcf-protection=none unilaterally, and provided build32.mk properly inherits CFLAGS. Discarding note sections may still be a desirable thing to do though ...

Jan

#6 lore raw

Re: [PATCH 2/2] x86/boot: Drop .note.gnu.properties in build32.lds

Quoted 17 lines · Jan Beulich

[CAUTION - EXTERNAL EMAIL] DO NOT reply, click links, or open attachments unless you have verified the sender and know the content is safe.

Quoted 14 lines · Andrew Cooper
Quoted 11 lines · Jan Beulich
Quoted 2 lines · Jason Andryuk

Discard the .note.gnu.property section when linking to avoid the extra bytes.

If we go this route (and if, as per above, I'm misremembering, meaning we didn't reject such a change earlier on), why would we not strip .note and .note.* in one go?

Quoted 3 lines

Stefan Bader also noticed that build32.mk requires -fcf-protection=none or else the hypervisor will not boot.
https://bugs.launchpad.net/ubuntu/+source/gcc-9/+bug/1863260

How's this related to the change here?

I think there is a bit of confusion as to exactly what is going on.

Quoted 7 lines
Quoted 3 lines

Ubuntu defaults -fcf-protection to enabled, which has a side effect of turning on CET, which inserts ENDBR{32,64} instructions and generates .note.gnu.properties indicating that the binary is CET-IBT compatible.

I.e. in principle this 2nd patch wouldn't be necessary at all if we forced -fcf-protection=none unilaterally, and provided build32.mk properly inherits CFLAGS. Discarding note sections may still be a desirable thing to do though ...

Even if we disable unilaterally for now, we'll still want to re-enable at some point, so this will come and bite us again.

I'm currently testing Ubuntu's default behaviour in combination with livepatch, because I bet there are further breakages to be found.

~Andrew

#7 lore raw

Re: [PATCH 2/2] x86/boot: Drop .note.gnu.properties in build32.lds

In reply to Andrew Cooper
Quoted 17 lines · Andrew Cooper
Quoted 10 lines · Jan Beulich

[CAUTION - EXTERNAL EMAIL] DO NOT reply, click links, or open attachments unless you have verified the sender and know the content is safe.

Quoted 7 lines · Jason Andryuk

reloc.S and cmdline.S as arrays of executable bytes for inclusion in head.S generated from compiled object files. Object files generated by an -fcf-protection toolchain include a .note.gnu.property section. The way reloc.S and cmdline.S are generated, the bytes of .note.gnu.property become the start of the .S files. When head.S calls reloc or cmdline_parse_early, those note bytes are executed instead of the intended .text section. This results in an early crash in reloc.

Quoted 4 lines

I may be misremembering, but I vaguely recall some similar change suggestion. What I'm missing here is some form of statement as to whether this is legitimate tool chain behavior, or a bug, and hence whether this is a fix or a workaround.

Quoted 8 lines

The linker is free to position unreferenced sections anywhere it wishes.

It is deeply unhelpful behaviour, but neither Binutils nor Clang developers think it is something wanting fixing.

One option might be to use --orphan-handling=error so unexpected toolchain behaviour breaks the build, or in this case perhaps =discard might be better.

The toolchain uses .note.gnu.property to flag object files as supporting Intel CET (Control-flow Enforcement Technology) enabled by -fcf-protection. The linker/loader uses the note to know if CET should be enabled or disabled. CET can only be enabled if the application and all libraries support it.

So it's legitimate to flag compiled objects with .note.gnu.property. The .S files generated by build32.mk are .. interesting. It seems like they should only be the runtime code & data, so we don't want the .note in there. So I guess this is a workaround for how the .S files are generated? My first attempt added -R .note.gnu.property, fyi.

I'm not familiar with the linker options Andrew references, to know how usable they are off the top of my head.

-fcf-protection=none could also be specified in CFLAGS in build32.mk to avoid generating the note.

Quoted 5 lines
Quoted 5 lines
Quoted 2 lines

Discard the .note.gnu.property section when linking to avoid the extra bytes.

If we go this route (and if, as per above, I'm misremembering, meaning we didn't reject such a change earlier on), why would we not strip .note and .note.* in one go?

Maybe? I made the conservative change since they weren't previously discarded.

Quoted 11 lines
Quoted 4 lines
Quoted 3 lines

Stefan Bader also noticed that build32.mk requires -fcf-protection=none or else the hypervisor will not boot.
https://bugs.launchpad.net/ubuntu/+source/gcc-9/+bug/1863260

How's this related to the change here?

I think there is a bit of confusion as to exactly what is going on.

Ubuntu defaults -fcf-protection to enabled, which has a side effect of turning on CET, which inserts ENDBR{32,64} instructions and generates .note.gnu.properties indicating that the binary is CET-IBT compatible.

Quoted 6 lines

ENDBR* instructions come from the Hint Nop space so are safe on older processors, but do ultimately add to binary bloat. It also occurs to me that it likely breaks livepath.

The reason Xen fails to boot is purely to do with the position of .note.gnu.properties, not the ENDBR* instructions.

Yes.

I referenced Stefan's bug since it specifically called out build32.mk as problematic even after supplying -fcf-protection=none for a hypervisor build. I was trying to give credit and reference a helpful bug entry. I don't know how Xen handles such things, but I am fine dropping it.

Regards,
Jason

#8 lore raw

Re: [PATCH 2/2] x86/boot: Drop .note.gnu.properties in build32.lds

Quoted 19 lines · Jason Andryuk
Quoted 18 lines · Andrew Cooper
Quoted 8 lines · Jan Beulich
Quoted 7 lines · Jason Andryuk

reloc.S and cmdline.S as arrays of executable bytes for inclusion in head.S generated from compiled object files. Object files generated by an -fcf-protection toolchain include a .note.gnu.property section. The way reloc.S and cmdline.S are generated, the bytes of .note.gnu.property become the start of the .S files. When head.S calls reloc or cmdline_parse_early, those note bytes are executed instead of the intended .text section. This results in an early crash in reloc.

Quoted 4 lines

I may be misremembering, but I vaguely recall some similar change suggestion. What I'm missing here is some form of statement as to whether this is legitimate tool chain behavior, or a bug, and hence whether this is a fix or a workaround.

The linker is free to position unreferenced sections anywhere it wishes.

It is deeply unhelpful behaviour, but neither Binutils nor Clang developers think it is something wanting fixing.

Quoted 8 lines
Quoted 3 lines

One option might be to use --orphan-handling=error so unexpected toolchain behaviour breaks the build, or in this case perhaps =discard might be better.

The toolchain uses .note.gnu.property to flag object files as supporting Intel CET (Control-flow Enforcement Technology) enabled by -fcf-protection. The linker/loader uses the note to know if CET should be enabled or disabled. CET can only be enabled if the application and all libraries support it.

Right, except we're a kernel here (rather than userspace), so the practicalities are different.

Quoted 4 lines

So it's legitimate to flag compiled objects with .note.gnu.property. The .S files generated by build32.mk are .. interesting. It seems like they should only be the runtime code & data, so we don't want the .note in there.

Yes...  Self-hosted relocatable 32bit code is tricky at the best of times, and this is a very good example of how not to do it.

I've got a plan to get rid of it completely, but it needs a bit more of the "switch to kbuild" series to go in first.

Quoted 32 lines

So I guess this is a workaround for how the .S files
are generated? My first attempt added -R .note.gnu.property, fyi.

I'm not familiar with the linker options Andrew references, to know how usable they are off the top of my head.

-fcf-protection=none could also be specified in CFLAGS in build32.mk to avoid generating the note.

Quoted 5 lines
Quoted 5 lines
Quoted 2 lines

Discard the .note.gnu.property section when linking to avoid the extra bytes.

If we go this route (and if, as per above, I'm misremembering, meaning we didn't reject such a change earlier on), why would we not strip .note and .note.* in one go?

Maybe? I made the conservative change since they weren't previously discarded.

Quoted 10 lines
Quoted 4 lines
Quoted 3 lines

Stefan Bader also noticed that build32.mk requires -fcf-protection=none or else the hypervisor will not boot.
https://bugs.launchpad.net/ubuntu/+source/gcc-9/+bug/1863260

How's this related to the change here?

I think there is a bit of confusion as to exactly what is going on.

Ubuntu defaults -fcf-protection to enabled, which has a side effect of turning on CET, which inserts ENDBR{32,64} instructions and generates .note.gnu.properties indicating that the binary is CET-IBT compatible.

Quoted 6 lines

ENDBR* instructions come from the Hint Nop space so are safe on older processors, but do ultimately add to binary bloat. It also occurs to me that it likely breaks livepath.

The reason Xen fails to boot is purely to do with the position of .note.gnu.properties, not the ENDBR* instructions.

Quoted 7 lines

Yes.

I referenced Stefan's bug since it specifically called out build32.mk as problematic even after supplying -fcf-protection=none for a hypervisor build. I was trying to give credit and reference a helpful bug entry. I don't know how Xen handles such things, but I am fine dropping it.

Typically a Reported-by: $PERSON <$EMAIL> tag, but frankly it would have been nice if anyone had posted any of these problems to xen-devel 6 months ago when it was first discovered to be a problem.

So far, we're at one definite (and fixed) toolchain bug, one obvious-but-not-yet-debugged toolchain bug, a robustness fix in Xen for the 32bit mess, and overriding of a system default, and thats before getting to the iPXE issues.

~Andrew

#9 lore raw
In reply to Jason Andryuk
Quoted 6 lines · Jason Andryuk

Two patches to fix building with a cf-protection toolchain. The first handles the case where the compiler fails to run with "error:
‘-mindirect-branch’ and ‘-fcf-protection’ are not compatible".

The second fixes a runtime error that prevented Xen booting in legacy mode.

That might be better than just disabling fcf-protection as well (which was done in Ubuntu lacking a better solution).

Not sure it was already hit but that additional .note section breaks the build of the emulator as generated headers become gigantic:

https://git.launchpad.net/ubuntu/+source/xen/tree/debian/patches/1001-strip-note-gnu-property.patch?h=ubuntu/focal

-Stefan

Quoted 12 lines

I still haven't figured out exactly what is wrong with rombios and/or ipxe.

Jason Andryuk (2):
xen/x86: Disable fcf-protection when necessary to build x86/boot: Drop .note.gnu.properties in build32.lds

 xen/arch/x86/arch.mk          | 11 ++++++++++-
 xen/arch/x86/boot/build32.lds |  5 +++++
 2 files changed, 15 insertions(+), 1 deletion(-)
#10 lore raw
Quoted 9 lines · Stefan Bader
Quoted 6 lines · Jason Andryuk

Two patches to fix building with a cf-protection toolchain. The first handles the case where the compiler fails to run with "error:
‘-mindirect-branch’ and ‘-fcf-protection’ are not compatible".

The second fixes a runtime error that prevented Xen booting in legacy mode.

That might be better than just disabling fcf-protection as well (which was done in Ubuntu lacking a better solution).

It is a GCC bug

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93654

Fixed for 10, and 9.4

Quoted 5 lines

Not sure it was already hit but that additional .note section breaks the build of the emulator as generated headers become gigantic:

https://git.launchpad.net/ubuntu/+source/xen/tree/debian/patches/1001-strip-note-gnu-property.patch?h=ubuntu/focal

4.6G of notes?!?  That is surely a toolchain bug.

~Andrew

#11 lore raw
Quoted 8 lines · Andrew Cooper
Quoted 7 lines · Stefan Bader
Quoted 6 lines · Jason Andryuk

Two patches to fix building with a cf-protection toolchain. The first handles the case where the compiler fails to run with "error:
‘-mindirect-branch’ and ‘-fcf-protection’ are not compatible".

The second fixes a runtime error that prevented Xen booting in legacy mode.

Quoted 8 lines
Quoted 2 lines

That might be better than just disabling fcf-protection as well (which was done in Ubuntu lacking a better solution).

It is a GCC bug

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93654

Fixed for 10, and 9.4

Ah, thanks. And yes, should have reported sooner but, as always, one runs into those things while in a hurry, with not enough time. We were glad to have something that did work somehow.

Quoted 8 lines
Quoted 6 lines

Not sure it was already hit but that additional .note section breaks the build of the emulator as generated headers become gigantic:

https://git.launchpad.net/ubuntu/+source/xen/tree/debian/patches/1001-strip-note-gnu-property.patch?h=ubuntu/focal

4.6G of notes?!?  That is surely a toolchain bug.

No, sorry if that was unclear. The .notes themselves are just about some Kb or so per object file. Problem is that each object file gets converted into a hex array header file. And this does multiply the resulting header file sizes. So the .h files generated are 4.6G in size. And there were a couple of those, all included into one .c file. Which ended in the compiler running out of memory on a 32GB system.

Quoted 3 lines

~Andrew

#12 lore raw
Quoted 7 lines · Stefan Bader
Quoted 6 lines · Andrew Cooper
Quoted 4 lines · Stefan Bader

Not sure it was already hit but that additional .note section breaks the build of the emulator as generated headers become gigantic:

https://git.launchpad.net/ubuntu/+source/xen/tree/debian/patches/1001-strip-note-gnu-property.patch?h=ubuntu/focal

Quoted 8 lines
Quoted 1 line

4.6G of notes?!?  That is surely a toolchain bug.

No, sorry if that was unclear. The .notes themselves are just about some Kb or so per object file. Problem is that each object file gets converted into a hex array header file. And this does multiply the resulting header file sizes. So the .h files generated are 4.6G in size. And there were a couple of those, all included into one .c file. Which ended in the compiler running out of memory on a 32GB system.

But as per the link above it's still 3Mb notes per object, which still seems insane.

Jan