Skip to contentSkip to conversation

seqlock: Allow UBSAN to fail optimizing

PATCH In review Peter Zijlstra 4 participants 7 messages Jan 29 – Feb 2 Locking primitives

lore
Share & tools
lore

Peter Zijlstra

author maintainer
#0 lore raw

PATCH 1/1 seqlock: Allow UBSAN to fail optimizing

Very similar to commit:

b94d45b6bbb4 ("seqlock: Allow KASAN to fail optimizing")

mark UBSAN as failing to optimize and not suitable for release builds.

---

Debian folks, can you please remove CONFIG_UBSAN=y from your builds, this is, per GCC devs, not a feature suitable for production!

include/linux/seqlock.h +3 -3
index c00063dffba3..436f05bf5dd7 100644
@@ -1259,14 +1259,14 @@ static __always_inline void __scoped_seqlock_cleanup(struct ss_tmp *sst)
 
 extern void __scoped_seqlock_invalid_target(void);
 
-#if (defined(CONFIG_CC_IS_GCC) && CONFIG_GCC_VERSION < 90000) || defined(CONFIG_KASAN)
+#if (defined(CONFIG_CC_IS_GCC) && CONFIG_GCC_VERSION < 90000) || defined(CONFIG_KASAN) || defined(CONFIG_UBSAN)
 /*
  * For some reason some GCC-8 architectures (nios2, alpha) have trouble
  * determining that the ss_done state is impossible in __scoped_seqlock_next()
  * below.
  *
- * Similarly KASAN is known to confuse compilers enough to break this. But we
- * don't care about code quality for KASAN builds anyway.
+ * Similarly *SAN is known to confuse compilers enough to break this. But we
+ * don't care about code quality for *SAN builds anyway.
  */
 static inline void __scoped_seqlock_bug(void) { }
 #else
#1 lore raw

Hi Peter,

[Adding Kees as well]

Quoted 28 lines · Peter Zijlstra

Very similar to commit:

b94d45b6bbb4 ("seqlock: Allow KASAN to fail optimizing")

mark UBSAN as failing to optimize and not suitable for release builds.

---

Debian folks, can you please remove CONFIG_UBSAN=y from your builds, this is, per GCC devs, not a feature suitable for production!

include/linux/seqlock.h +1 -1
index c00063dffba3..436f05bf5dd7 100644
@@ -1259,14 +1259,14 @@ static __always_inline void __scoped_seqlock_cleanup(struct ss_tmp *sst)
 
 extern void __scoped_seqlock_invalid_target(void);
 
-#if (defined(CONFIG_CC_IS_GCC) && CONFIG_GCC_VERSION < 90000) || defined(CONFIG_KASAN)
+#if (defined(CONFIG_CC_IS_GCC) && CONFIG_GCC_VERSION < 90000) || defined(CONFIG_KASAN) || defined(CONFIG_UBSAN)
 /*
  * For some reason some GCC-8 architectures (nios2, alpha) have trouble

include/linux/seqlock.h static __always_inline void __scoped_seqlock_cleanup(struct ss_tmp *sst) · line 1272

Quoted 10 lines
include/linux/seqlock.h +2 -2
  * determining that the ss_done state is impossible in __scoped_seqlock_next()
  * below.
  *
- * Similarly KASAN is known to confuse compilers enough to break this. But we
- * don't care about code quality for KASAN builds anyway.
+ * Similarly *SAN is known to confuse compilers enough to break this. But we
+ * don't care about code quality for *SAN builds anyway.
  */
 static inline void __scoped_seqlock_bug(void) { }
 #else

Kees, Peter approached the Debian kernel list above to drop CONFIG_UBSAN again, which, so I think we need to revert your 6cfadabfe015 ("Enable UBSAN_BOUNDS and UBSAN_SHIFT"):
https://salsa.debian.org/kernel-team/linux/-/commit/6cfadabfe015fa0d659fc8e3efd495cbcae3e44e

I have make a MR for our packaging for the change in
https://salsa.debian.org/kernel-team/linux/-/merge_requests/1804

Regards,
Salvatore

#2 lore raw
Quoted 7 lines · Salvatore Bonaccorso

Kees, Peter approached the Debian kernel list above to drop CONFIG_UBSAN again, which, so I think we need to revert your 6cfadabfe015 ("Enable UBSAN_BOUNDS and UBSAN_SHIFT"):
https://salsa.debian.org/kernel-team/linux/-/commit/6cfadabfe015fa0d659fc8e3efd495cbcae3e44e

I have make a MR for our packaging for the change in
https://salsa.debian.org/kernel-team/linux/-/merge_requests/1804

I am strongly opposed -- this undoes years of security flaw mitigation work and leaves Debian (and only Debian!) exposed to trivial array index overflows. The bounds sanitizer is the corner stone of memory safety for C, and is not some "experimental" feature. GCC has a long history of trouble with inlining, so this is not something unique to enabling this feature.

I replied similarly to the PR. This would be a major mistake to disable.

-Kees

Signature · 1 line
Kees Cook                                            @debian.org
author maintainer
#3 lore raw
Quoted 9 lines · Kees Cook
Quoted 7 lines · Salvatore Bonaccorso

Kees, Peter approached the Debian kernel list above to drop CONFIG_UBSAN again, which, so I think we need to revert your 6cfadabfe015 ("Enable UBSAN_BOUNDS and UBSAN_SHIFT"):
https://salsa.debian.org/kernel-team/linux/-/commit/6cfadabfe015fa0d659fc8e3efd495cbcae3e44e

I have make a MR for our packaging for the change in
https://salsa.debian.org/kernel-team/linux/-/merge_requests/1804

Quoted 8 lines

I am strongly opposed -- this undoes years of security flaw mitigation work and leaves Debian (and only Debian!) exposed to trivial array index overflows. The bounds sanitizer is the corner stone of memory safety for C, and is not some "experimental" feature. GCC has a long history of trouble with inlining, so this is not something unique to enabling this feature.

I replied similarly to the PR. This would be a major mistake to disable.

Why the heck is bounds checking part of UBSAN? The simple fix here is to get it out from CONFIG_UBSAN, so that CONFIG_UBSAN is debug only crap.

Notably, none of the UBSAN configs that tripped the optimization fail even had bounds checking enabled.

#4 lore raw
Quoted 17 lines · Peter Zijlstra
Quoted 16 lines · Kees Cook
Quoted 7 lines · Salvatore Bonaccorso

Kees, Peter approached the Debian kernel list above to drop CONFIG_UBSAN again, which, so I think we need to revert your 6cfadabfe015 ("Enable UBSAN_BOUNDS and UBSAN_SHIFT"):
https://salsa.debian.org/kernel-team/linux/-/commit/6cfadabfe015fa0d659fc8e3efd495cbcae3e44e

I have make a MR for our packaging for the change in
https://salsa.debian.org/kernel-team/linux/-/merge_requests/1804

I am strongly opposed -- this undoes years of security flaw mitigation work and leaves Debian (and only Debian!) exposed to trivial array index overflows. The bounds sanitizer is the corner stone of memory safety for C, and is not some "experimental" feature. GCC has a long history of trouble with inlining, so this is not something unique to enabling this feature.

Quoted 4 lines
Quoted 1 line

I replied similarly to the PR. This would be a major mistake to disable.

Why the heck is bounds checking part of UBSAN? The simple fix here is to get it out from CONFIG_UBSAN, so that CONFIG_UBSAN is debug only crap.

Out of bounds accesses are considered "undefined". *sigh*

But yes, now that we have the "transitional" kconfig symbols I can trivially rename CONFIG_UBSAN_BOUNDS and remove its CONFIG_UBSAN dependency.

Quoted 2 lines

Notably, none of the UBSAN configs that tripped the optimization fail even had bounds checking enabled.

Which ones tripped it? KASAN (in software tagging mode) is usually the heavy-weight one that considerably bloats code generation. I haven't seen systemic problems with -fsanitize=bounds, and I thought the weird cases (which confused the value range tracking) with -fsanitize=shift got fixed back in GCC 12 (or maybe 13).

-Kees

Signature · 1 line
Kees Cook
author maintainer
#5 lore raw
Quoted 18 lines · Kees Cook
Quoted 17 lines · Peter Zijlstra
Quoted 16 lines · Kees Cook
Quoted 7 lines · Salvatore Bonaccorso

Kees, Peter approached the Debian kernel list above to drop CONFIG_UBSAN again, which, so I think we need to revert your 6cfadabfe015 ("Enable UBSAN_BOUNDS and UBSAN_SHIFT"):
https://salsa.debian.org/kernel-team/linux/-/commit/6cfadabfe015fa0d659fc8e3efd495cbcae3e44e

I have make a MR for our packaging for the change in
https://salsa.debian.org/kernel-team/linux/-/merge_requests/1804

I am strongly opposed -- this undoes years of security flaw mitigation work and leaves Debian (and only Debian!) exposed to trivial array index overflows. The bounds sanitizer is the corner stone of memory safety for C, and is not some "experimental" feature. GCC has a long history of trouble with inlining, so this is not something unique to enabling this feature.

Quoted 6 lines
Quoted 4 lines
Quoted 1 line

I replied similarly to the PR. This would be a major mistake to disable.

Why the heck is bounds checking part of UBSAN? The simple fix here is to get it out from CONFIG_UBSAN, so that CONFIG_UBSAN is debug only crap.

Out of bounds accesses are considered "undefined". *sigh*

*sigh* indeed.

Quoted 3 lines

But yes, now that we have the "transitional" kconfig symbols I can trivially rename CONFIG_UBSAN_BOUNDS and remove its CONFIG_UBSAN dependency.

That would be great; I can sit on this patch a while, its mostly build robots and the like occasionally tripping this.

It would be good to have the compiler folks agree that bounds checking is production code though :-)

Quoted 8 lines
Quoted 2 lines

Notably, none of the UBSAN configs that tripped the optimization fail even had bounds checking enabled.

Which ones tripped it? KASAN (in software tagging mode) is usually the heavy-weight one that considerably bloats code generation. I haven't seen systemic problems with -fsanitize=bounds, and I thought the weird cases (which confused the value range tracking) with -fsanitize=shift got fixed back in GCC 12 (or maybe 13).

I'm not sure which one (I didn't care to find out, its debug nonsense and nobody cares etc.. :-).

One has:

CONFIG_ARCH_HAS_UBSAN=y
CONFIG_UBSAN=y
CONFIG_UBSAN_TRAP=y
CONFIG_CC_HAS_UBSAN_BOUNDS_STRICT=y
# CONFIG_UBSAN_BOUNDS is not set
# CONFIG_UBSAN_SHIFT is not set
CONFIG_UBSAN_DIV_ZERO=y
CONFIG_UBSAN_BOOL=y
CONFIG_UBSAN_ENUM=y

The other has:

CONFIG_ARCH_HAS_UBSAN=y
CONFIG_UBSAN=y
# CONFIG_UBSAN_TRAP is not set
CONFIG_CC_HAS_UBSAN_BOUNDS_STRICT=y
# CONFIG_UBSAN_BOUNDS is not set
CONFIG_UBSAN_SHIFT=y
CONFIG_UBSAN_DIV_ZERO=y
# CONFIG_UBSAN_UNREACHABLE is not set
CONFIG_UBSAN_BOOL=y
# CONFIG_UBSAN_ENUM is not set
CONFIG_UBSAN_ALIGNMENT=y

The common ones are DIV_ZERO and BOOL.

#6 lore raw
Quoted 76 lines · Peter Zijlstra
Quoted 18 lines · Kees Cook
Quoted 17 lines · Peter Zijlstra
Quoted 16 lines · Kees Cook
Quoted 7 lines · Salvatore Bonaccorso

Kees, Peter approached the Debian kernel list above to drop CONFIG_UBSAN again, which, so I think we need to revert your 6cfadabfe015 ("Enable UBSAN_BOUNDS and UBSAN_SHIFT"):
https://salsa.debian.org/kernel-team/linux/-/commit/6cfadabfe015fa0d659fc8e3efd495cbcae3e44e

I have make a MR for our packaging for the change in
https://salsa.debian.org/kernel-team/linux/-/merge_requests/1804

I am strongly opposed -- this undoes years of security flaw mitigation work and leaves Debian (and only Debian!) exposed to trivial array index overflows. The bounds sanitizer is the corner stone of memory safety for C, and is not some "experimental" feature. GCC has a long history of trouble with inlining, so this is not something unique to enabling this feature.

Quoted 6 lines
Quoted 4 lines
Quoted 1 line

I replied similarly to the PR. This would be a major mistake to disable.

Why the heck is bounds checking part of UBSAN? The simple fix here is to get it out from CONFIG_UBSAN, so that CONFIG_UBSAN is debug only crap.

Out of bounds accesses are considered "undefined". *sigh*

*sigh* indeed.

Quoted 3 lines

But yes, now that we have the "transitional" kconfig symbols I can trivially rename CONFIG_UBSAN_BOUNDS and remove its CONFIG_UBSAN dependency.

That would be great; I can sit on this patch a while, its mostly build robots and the like occasionally tripping this.

It would be good to have the compiler folks agree that bounds checking is production code though :-)

Quoted 8 lines
Quoted 2 lines

Notably, none of the UBSAN configs that tripped the optimization fail even had bounds checking enabled.

Which ones tripped it? KASAN (in software tagging mode) is usually the heavy-weight one that considerably bloats code generation. I haven't seen systemic problems with -fsanitize=bounds, and I thought the weird cases (which confused the value range tracking) with -fsanitize=shift got fixed back in GCC 12 (or maybe 13).

I'm not sure which one (I didn't care to find out, its debug nonsense and nobody cares etc.. :-).

One has:

CONFIG_ARCH_HAS_UBSAN=y
CONFIG_UBSAN=y
CONFIG_UBSAN_TRAP=y
CONFIG_CC_HAS_UBSAN_BOUNDS_STRICT=y
# CONFIG_UBSAN_BOUNDS is not set
# CONFIG_UBSAN_SHIFT is not set
CONFIG_UBSAN_DIV_ZERO=y
CONFIG_UBSAN_BOOL=y
CONFIG_UBSAN_ENUM=y

The other has:

CONFIG_ARCH_HAS_UBSAN=y
CONFIG_UBSAN=y
# CONFIG_UBSAN_TRAP is not set
CONFIG_CC_HAS_UBSAN_BOUNDS_STRICT=y
# CONFIG_UBSAN_BOUNDS is not set
CONFIG_UBSAN_SHIFT=y
CONFIG_UBSAN_DIV_ZERO=y
# CONFIG_UBSAN_UNREACHABLE is not set
CONFIG_UBSAN_BOOL=y
# CONFIG_UBSAN_ENUM is not set
CONFIG_UBSAN_ALIGNMENT=y

Quoted 1 line

The common ones are DIV_ZERO and BOOL.

DIV_ZERO is a known trouble-maker in Clang (though for its pathological behavior it needs CONFIG_UBSAN_TRAP). Perhaps we should drop it entirely?

-Kees

Signature · 1 line
Kees Cook