[PATCH 1/4] Makefile: add -fno-builtin-stpcpy

LLVM implemented a recent "libcall optimization" that lowers calls to
`sprintf(dest, "%s", str)` where the return value is used to
`stpcpy(dest, str) - dest`. This generally avoids the machinery involved
in parsing format strings. This optimization was introduced into
clang-12. Because the kernel does not provide an implementation of
stpcpy, we observe linkage failures for almost all targets when building
with ToT clang.

The interface is unsafe as it does not perform any bounds checking.
Disable this "libcall optimization" via `-fno-builtin-stpcpy`.

Unlike
commit 5f074f3e192f ("lib/string.c: implement a basic bcmp")
which cited failures with `-fno-builtin-*` flags being retained in LLVM
LTO, that bug seems to have been fixed by
https://reviews.llvm.org/D71193, so the above sha can now be reverted in
favor of `-fno-builtin-bcmp`.

Cc: stable@vger.kernel.org # 4.4
Link: https://bugs.llvm.org/show_bug.cgi?id=47162
Link: https://github.com/ClangBuiltLinux/linux/issues/1126
Link: https://reviews.llvm.org/D85963
Reported-by: Sami Tolvanen <samitolvanen@google.com>
Suggested-by: Dávid Bolvanský <redacted>
Suggested-by: Kees Cook <redacted>
Signed-off-by: Nick Desaulniers <ndesaulniers@google.com>
---
 Makefile | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/Makefile b/Makefile
index 9cac6fde3479..211a1b6f6478 100644
--- a/Makefile
+++ b/Makefile

@@ -959,6 +959,12 @@ ifdef CONFIG_RETPOLINE
 KBUILD_CFLAGS += $(call cc-option,-fcf-protection=none)
 endif
 
+# The compiler may "libcall optimize" certain function calls into the below
+# functions, for architectures that don't use -ffreestanding. If we don't plan
+# to provide implementations of these routines, then prevent the compiler from
+# emitting calls to what will be undefined symbols.
+KBUILD_CFLAGS	+= -fno-builtin-stpcpy
+
 # include additional Makefiles when needed
 include-y			:= scripts/Makefile.extrawarn
 include-$(CONFIG_KASAN)		+= scripts/Makefile.kasan

-- 
2.28.0.220.ged08abb693-goog

On 2020-08-17 15:02, Nick Desaulniers wrote:

LLVM implemented a recent "libcall optimization" that lowers calls to
`sprintf(dest, "%s", str)` where the return value is used to
`stpcpy(dest, str) - dest`. This generally avoids the machinery involved
in parsing format strings. This optimization was introduced into
clang-12. Because the kernel does not provide an implementation of
stpcpy, we observe linkage failures for almost all targets when building
with ToT clang.

The interface is unsafe as it does not perform any bounds checking.
Disable this "libcall optimization" via `-fno-builtin-stpcpy`.

Unlike
commit 5f074f3e192f ("lib/string.c: implement a basic bcmp")
which cited failures with `-fno-builtin-*` flags being retained in LLVM
LTO, that bug seems to have been fixed by
https://reviews.llvm.org/D71193, so the above sha can now be reverted in
favor of `-fno-builtin-bcmp`.

stpcpy() and (to a lesser degree) mempcpy() are fairly useful routines
in general. Perhaps we *should* provide them?

	-hpa

On Mon, Aug 17, 2020 at 3:31 PM H. Peter Anvin [off-list ref] wrote:

On 2020-08-17 15:02, Nick Desaulniers wrote:

quoted

LLVM implemented a recent "libcall optimization" that lowers calls to
`sprintf(dest, "%s", str)` where the return value is used to
`stpcpy(dest, str) - dest`. This generally avoids the machinery involved
in parsing format strings. This optimization was introduced into
clang-12. Because the kernel does not provide an implementation of
stpcpy, we observe linkage failures for almost all targets when building
with ToT clang.

The interface is unsafe as it does not perform any bounds checking.
Disable this "libcall optimization" via `-fno-builtin-stpcpy`.

Unlike
commit 5f074f3e192f ("lib/string.c: implement a basic bcmp")
which cited failures with `-fno-builtin-*` flags being retained in LLVM
LTO, that bug seems to have been fixed by
https://reviews.llvm.org/D71193, so the above sha can now be reverted in
favor of `-fno-builtin-bcmp`.

stpcpy() and (to a lesser degree) mempcpy() are fairly useful routines
in general. Perhaps we *should* provide them?

Sorry, I forgot to provide context of the previous thread, which is
worth a read.  To answer this question specifically (or at least for
stpcpy), the answer from the previous thread was (via Kees): "No;
please no more unbounded string.h routines":
https://lore.kernel.org/lkml/202008150921.B70721A359@keescook/
-- 
Thanks,
~Nick Desaulniers

On Tue, 18 Aug 2020 at 00:02, Nick Desaulniers [off-list ref] wrote:

LLVM implemented a recent "libcall optimization" that lowers calls to
`sprintf(dest, "%s", str)` where the return value is used to
`stpcpy(dest, str) - dest`. This generally avoids the machinery involved
in parsing format strings. This optimization was introduced into
clang-12. Because the kernel does not provide an implementation of
stpcpy, we observe linkage failures for almost all targets when building
with ToT clang.

The interface is unsafe as it does not perform any bounds checking.
Disable this "libcall optimization" via `-fno-builtin-stpcpy`.

Unlike
commit 5f074f3e192f ("lib/string.c: implement a basic bcmp")
which cited failures with `-fno-builtin-*` flags being retained in LLVM
LTO, that bug seems to have been fixed by
https://reviews.llvm.org/D71193, so the above sha can now be reverted in
favor of `-fno-builtin-bcmp`.

Cc: stable@vger.kernel.org # 4.4

Why does a fix for Clang-12 have to be backported all the way to v4.4?
How does that meet the requirements for stable patches?

On Tue, Aug 18, 2020 at 09:10:01AM +0200, Ard Biesheuvel wrote:

On Tue, 18 Aug 2020 at 00:02, Nick Desaulniers [off-list ref] wrote:

quoted

LLVM implemented a recent "libcall optimization" that lowers calls to
`sprintf(dest, "%s", str)` where the return value is used to
`stpcpy(dest, str) - dest`. This generally avoids the machinery involved
in parsing format strings. This optimization was introduced into
clang-12. Because the kernel does not provide an implementation of
stpcpy, we observe linkage failures for almost all targets when building
with ToT clang.

The interface is unsafe as it does not perform any bounds checking.
Disable this "libcall optimization" via `-fno-builtin-stpcpy`.

Unlike
commit 5f074f3e192f ("lib/string.c: implement a basic bcmp")
which cited failures with `-fno-builtin-*` flags being retained in LLVM
LTO, that bug seems to have been fixed by
https://reviews.llvm.org/D71193, so the above sha can now be reverted in
favor of `-fno-builtin-bcmp`.

Cc: stable@vger.kernel.org # 4.4

Why does a fix for Clang-12 have to be backported all the way to v4.4?
How does that meet the requirements for stable patches?

Because people like to build older kernels with new compliler versions.

And those "people" include me, who doesn't want to keep around old
compilers just because my distro moved to the latest one...

We've been doing this for the past 4+ years, for new versions of gcc,
keeping 4.4.y building properly with the bleeding edge of that compiler,
why is clang any different here?

thanks,

greg k-h

On Tue, 18 Aug 2020 at 09:25, Greg KH [off-list ref] wrote:

On Tue, Aug 18, 2020 at 09:10:01AM +0200, Ard Biesheuvel wrote:

quoted

On Tue, 18 Aug 2020 at 00:02, Nick Desaulniers [off-list ref] wrote:

quoted

LLVM implemented a recent "libcall optimization" that lowers calls to
`sprintf(dest, "%s", str)` where the return value is used to
`stpcpy(dest, str) - dest`. This generally avoids the machinery involved
in parsing format strings. This optimization was introduced into
clang-12. Because the kernel does not provide an implementation of
stpcpy, we observe linkage failures for almost all targets when building
with ToT clang.

The interface is unsafe as it does not perform any bounds checking.
Disable this "libcall optimization" via `-fno-builtin-stpcpy`.

Unlike
commit 5f074f3e192f ("lib/string.c: implement a basic bcmp")
which cited failures with `-fno-builtin-*` flags being retained in LLVM
LTO, that bug seems to have been fixed by
https://reviews.llvm.org/D71193, so the above sha can now be reverted in
favor of `-fno-builtin-bcmp`.

Cc: stable@vger.kernel.org # 4.4

Why does a fix for Clang-12 have to be backported all the way to v4.4?
How does that meet the requirements for stable patches?

Because people like to build older kernels with new compliler versions.

And those "people" include me, who doesn't want to keep around old
compilers just because my distro moved to the latest one...

We've been doing this for the past 4+ years, for new versions of gcc,
keeping 4.4.y building properly with the bleeding edge of that compiler,
why is clang any different here?

Fair enough. I am just struggling to match stable-kernel-rules.rst
with the actual practices - perhaps it is time to update that
document?

On Tue, Aug 18, 2020 at 09:29:39AM +0200, Ard Biesheuvel wrote:

On Tue, 18 Aug 2020 at 09:25, Greg KH [off-list ref] wrote:

quoted

On Tue, Aug 18, 2020 at 09:10:01AM +0200, Ard Biesheuvel wrote:

quoted

On Tue, 18 Aug 2020 at 00:02, Nick Desaulniers [off-list ref] wrote:

quoted

LLVM implemented a recent "libcall optimization" that lowers calls to
`sprintf(dest, "%s", str)` where the return value is used to
`stpcpy(dest, str) - dest`. This generally avoids the machinery involved
in parsing format strings. This optimization was introduced into
clang-12. Because the kernel does not provide an implementation of
stpcpy, we observe linkage failures for almost all targets when building
with ToT clang.

The interface is unsafe as it does not perform any bounds checking.
Disable this "libcall optimization" via `-fno-builtin-stpcpy`.

Unlike
commit 5f074f3e192f ("lib/string.c: implement a basic bcmp")
which cited failures with `-fno-builtin-*` flags being retained in LLVM
LTO, that bug seems to have been fixed by
https://reviews.llvm.org/D71193, so the above sha can now be reverted in
favor of `-fno-builtin-bcmp`.

Cc: stable@vger.kernel.org # 4.4

Why does a fix for Clang-12 have to be backported all the way to v4.4?
How does that meet the requirements for stable patches?

Because people like to build older kernels with new compliler versions.

And those "people" include me, who doesn't want to keep around old
compilers just because my distro moved to the latest one...

We've been doing this for the past 4+ years, for new versions of gcc,
keeping 4.4.y building properly with the bleeding edge of that compiler,
why is clang any different here?

Fair enough. I am just struggling to match stable-kernel-rules.rst
with the actual practices - perhaps it is time to update that
document?

The rules are tiny and simple for 99% of the issues involved.  Stuff
like "add patches to fix build failures and warnings for newer compiler
versions" are so rare (they only happen every 2 years or so), it's not
worth it.

thanks,

greg k-h

On Mon, Aug 17, 2020 at 03:31:26PM -0700, H. Peter Anvin wrote:

On 2020-08-17 15:02, Nick Desaulniers wrote:

quoted

LLVM implemented a recent "libcall optimization" that lowers calls to
`sprintf(dest, "%s", str)` where the return value is used to
`stpcpy(dest, str) - dest`. This generally avoids the machinery involved
in parsing format strings. This optimization was introduced into
clang-12. Because the kernel does not provide an implementation of
stpcpy, we observe linkage failures for almost all targets when building
with ToT clang.

The interface is unsafe as it does not perform any bounds checking.
Disable this "libcall optimization" via `-fno-builtin-stpcpy`.

Unlike
commit 5f074f3e192f ("lib/string.c: implement a basic bcmp")
which cited failures with `-fno-builtin-*` flags being retained in LLVM
LTO, that bug seems to have been fixed by
https://reviews.llvm.org/D71193, so the above sha can now be reverted in
favor of `-fno-builtin-bcmp`.

stpcpy() and (to a lesser degree) mempcpy() are fairly useful routines
in general. Perhaps we *should* provide them?

As Nick mentioned, I really don't want to expand the already bad
interfaces from libc. We have enough messes to clean up already, and I
don't want to add more. The kernel already uses a subset of C, we have
(several) separate non-libc memory allocators, we're using strscpy() and
scnprintf() widely in favor of their buggy libc counterparts, etc. We
don't need to match the libc string interfaces especially when they're
arguably bug-prone foot-guns. :)

-- 
Kees Cook

On Mon, Aug 17, 2020 at 03:02:09PM -0700, Nick Desaulniers wrote:

LLVM implemented a recent "libcall optimization" that lowers calls to
`sprintf(dest, "%s", str)` where the return value is used to
`stpcpy(dest, str) - dest`. This generally avoids the machinery involved
in parsing format strings. This optimization was introduced into
clang-12. Because the kernel does not provide an implementation of
stpcpy, we observe linkage failures for almost all targets when building
with ToT clang.

The interface is unsafe as it does not perform any bounds checking.
Disable this "libcall optimization" via `-fno-builtin-stpcpy`.

Unlike
commit 5f074f3e192f ("lib/string.c: implement a basic bcmp")
which cited failures with `-fno-builtin-*` flags being retained in LLVM
LTO, that bug seems to have been fixed by
https://reviews.llvm.org/D71193, so the above sha can now be reverted in
favor of `-fno-builtin-bcmp`.

Cc: stable@vger.kernel.org # 4.4
Link: https://bugs.llvm.org/show_bug.cgi?id=47162
Link: https://github.com/ClangBuiltLinux/linux/issues/1126
Link: https://reviews.llvm.org/D85963
Reported-by: Sami Tolvanen <samitolvanen@google.com>
Suggested-by: Dávid Bolvanský <redacted>
Suggested-by: Kees Cook <redacted>
Signed-off-by: Nick Desaulniers <ndesaulniers@google.com>

Reviewed-by: Kees Cook <redacted>

-- 
Kees Cook