Re: [linus:master] [crypto] 40b9969796: UBSAN:unsigned-integer-overflow_in_lib/crypto/chacha20poly1305-selftest.c

Re: [linus:master] [crypto] 40b9969796: UBSAN:unsigned-integer-overflow_in_lib/crypto/chacha20poly1305-selftest.c

[Kees Cook]

On Tue, May 27, 2025 at 11:14:27PM -0700, Eric Biggers wrote:
> [+Kees and linux-hardening]
> 
> On Wed, May 28, 2025 at 01:15:05PM +0800, kernel test robot wrote:
> > 
> > 
> > Hello,
> > 
> > by this commit, the config has below diff:
> > 
> > --- /pkg/linux/x86_64-randconfig-101-20250522/clang-20/d469eaed223fa485eabebd3bcd05ddd3c891f54e/.config 2025-05-23 23:44:56.781716572 +0800
> > +++ /pkg/linux/x86_64-randconfig-101-20250522/clang-20/40b9969796bfa49ed1b0f7ddc254f48cb2ac6d2c/.config 2025-05-24 02:08:29.858605300 +0800
> > @@ -4837,7 +4837,8 @@ CONFIG_CRYPTO_ACOMP2=y
> >  CONFIG_CRYPTO_MANAGER=y
> >  CONFIG_CRYPTO_MANAGER2=y
> >  # CONFIG_CRYPTO_USER is not set
> > -CONFIG_CRYPTO_MANAGER_DISABLE_TESTS=y
> > +CONFIG_CRYPTO_SELFTESTS=y
> > +# CONFIG_CRYPTO_MANAGER_EXTRA_TESTS is not set
> >  # CONFIG_CRYPTO_NULL is not set
> >  CONFIG_CRYPTO_PCRYPT=m
> >  CONFIG_CRYPTO_CRYPTD=y
> > 
> > it seems tests are enabled then we observe the UBSAN issues
> > 
> > d469eaed223fa485 40b9969796bfa49ed1b0f7ddc25
> > ---------------- ---------------------------
> >        fail:runs  %reproduction    fail:runs
> >            |             |             |
> >            :6          100%           6:6     dmesg.UBSAN:unsigned-integer-overflow_in_lib/crypto/chacha20poly1305-selftest.c
> >            :6          100%           6:6     dmesg.UBSAN:unsigned-integer-overflow_in_lib/crypto/chacha20poly1305.c
> > 
> > it's hard for bot to apply this commit to previous commits in bisect, so we just
> > make out below report FYI that we observe UBSAN issues in boot tests.
> > 
> > 
> > kernel test robot noticed "UBSAN:unsigned-integer-overflow_in_lib/crypto/chacha20poly1305-selftest.c" on:
> > 
> > commit: 40b9969796bfa49ed1b0f7ddc254f48cb2ac6d2c ("crypto: testmgr - replace CRYPTO_MANAGER_DISABLE_TESTS with CRYPTO_SELFTESTS")
> > https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git master
> > 
> > [test failed on linux-next/master 176e917e010cb7dcc605f11d2bc33f304292482b]
> > 
> > in testcase: boot
> > 
> > config: x86_64-randconfig-101-20250522
> > compiler: clang-20
> > test machine: qemu-system-x86_64 -enable-kvm -cpu SandyBridge -smp 2 -m 16G
> > 
> > (please refer to attached dmesg/kmsg for entire log/backtrace)
> > 
> > 
> > 
> > If you fix the issue in a separate patch/commit (i.e. not just a new version of
> > the same patch/commit), kindly add following tags
> > | Reported-by: kernel test robot <oliver.sang@intel.com>
> > | Closes: https://lore.kernel.org/oe-lkp/202505281024.f42beaa7-lkp@intel.com
> > 
> > 
> > [   12.442846][    T1] ------------[ cut here ]------------
> > [   12.443261][    T1] UBSAN: unsigned-integer-overflow in lib/crypto/chacha20poly1305-selftest.c:8854:47
> > [   12.444084][    T1] 16 - 114 cannot be represented in type 'size_t' (aka 'unsigned long')
> > [   12.444682][    T1] CPU: 1 UID: 0 PID: 1 Comm: swapper/0 Tainted: G                T   6.15.0-rc5-00342-g40b9969796bf #1 VOLUNTARY
> 
> This issue predates the blamed commit, and it's specific to
> CONFIG_UBSAN_INTEGER_WRAP which was recently introduced.
> 
> CONFIG_UBSAN_INTEGER_WRAP apparently requires clang 20.
> 
> To try to reproduce this, I built clang from the release/20.x branch, then built
> a kernel with CONFIG_UBSAN_INTEGER_WRAP=y.  When booting that kernel, there are
> many UBSAN reports:
> 
>     [    0.000000] UBSAN: negation-overflow in lib/sort.c:199:36
> 
>     [    0.000000] UBSAN: negation-overflow in lib/sort.c:185:14
> 
>     [    0.276708] UBSAN: unsigned-integer-overflow in ./include/linux/min_heap.h:329:24
> 
>     [    0.277376] UBSAN: negation-overflow in ./include/linux/min_heap.h:260:42
> 
>     [    0.871191] UBSAN: unsigned-integer-overflow in lib/crypto/chacha20poly1305-selftest.c:8854:47
> 
>     [    0.890856] UBSAN: unsigned-integer-overflow in lib/crypto/chacha20poly1305-selftest.c:8851:47
> 
>     [    0.910455] UBSAN: unsigned-integer-overflow in lib/crypto/chacha20poly1305.c:260:57
> 
>     [    1.105542] UBSAN: unsigned-integer-overflow in lib/zstd/compress/zstd_compress_sequences.c:334:21
> 
>     [    1.113539] UBSAN: unsigned-integer-overflow in lib/zstd/compress/huf_compress.c:889:23
> 
>     [    1.114597] UBSAN: unsigned-integer-overflow in lib/lz4/lz4_compress.c:294:9
> 
> So I did get the chacha20poly1305 ones, but they're hardly unique.
> 
> If this new sanitizer is going to move forward, is there any sort of plan or
> guide for how to update code to be compatible with it?  Specifically considering
> common situations where unsigned wraparound (which is defined behavior in C) can
> be intentionally relied on, like calculating the distance from the next N-byte
> boundary.  What are the best practices now?

Hi, yes, this is still under development. I tried to make it hard to
enable accidentally (not via COMPILE_TEST, not UBSAN-default, etc), but
we (still) don't have a way to disable configs for randconfigs. :(

We're hoping to see Clang 21 with the more versatile Overflow Behavior Types:
https://discourse.llvm.org/t/rfc-v2-clang-introduce-overflowbehaviortypes-for-wrapping-and-non-wrapping-arithmetic/86507

and our current testing is showing many fewer false positives. (Having
run syzkaller for weeks now.)

> Documentation/dev-tools/ubsan.rst says nothing about this and only mentions
> "undefined behavior", which this is not.

Right -- this will get extensive documentation before we move it out of
its development phase.

I'm not sure how to enforce "don't enable this unless you're developing
the Overflow Behavior Types" with current Kconfig, given the randconfig
gap... I have some memory of Arnd doing something special with his
randconfigs to avoid these kinds of things, but I can't find it now.

-Kees

-- 
Kees Cook

Re: [linus:master] [crypto] 40b9969796: UBSAN:unsigned-integer-overflow_in_lib/crypto/chacha20poly1305-selftest.c

[Arnd Bergmann]

On Wed, May 28, 2025, at 18:45, Kees Cook wrote:
> On Tue, May 27, 2025 at 11:14:27PM -0700, Eric Biggers wrote:
>> On Wed, May 28, 2025 at 01:15:05PM +0800, kernel test robot wrote:

> I'm not sure how to enforce "don't enable this unless you're developing
> the Overflow Behavior Types" with current Kconfig, given the randconfig
> gap... I have some memory of Arnd doing something special with his
> randconfigs to avoid these kinds of things, but I can't find it now.
>

The main thing I do on the randconfig builds to avoid obscure issues
is to force CONFIG_COMPILE_TEST=y, but that only works for build
testing, not actually running it.

      Arnd

Re: [linus:master] [crypto] 40b9969796: UBSAN:unsigned-integer-overflow_in_lib/crypto/chacha20poly1305-selftest.c

[Jann Horn]

On Wed, May 28, 2025 at 6:46 PM Kees Cook <kees@kernel.org> wrote:
> On Tue, May 27, 2025 at 11:14:27PM -0700, Eric Biggers wrote:
> > If this new sanitizer is going to move forward, is there any sort of plan or
> > guide for how to update code to be compatible with it?  Specifically considering
> > common situations where unsigned wraparound (which is defined behavior in C) can
> > be intentionally relied on, like calculating the distance from the next N-byte
> > boundary.  What are the best practices now?
>
> Hi, yes, this is still under development. I tried to make it hard to
> enable accidentally (not via COMPILE_TEST, not UBSAN-default, etc), but
> we (still) don't have a way to disable configs for randconfigs. :(
>
> We're hoping to see Clang 21 with the more versatile Overflow Behavior Types:
> https://discourse.llvm.org/t/rfc-v2-clang-introduce-overflowbehaviortypes-for-wrapping-and-non-wrapping-arithmetic/86507
>
> and our current testing is showing many fewer false positives. (Having
> run syzkaller for weeks now.)
>
> > Documentation/dev-tools/ubsan.rst says nothing about this and only mentions
> > "undefined behavior", which this is not.
>
> Right -- this will get extensive documentation before we move it out of
> its development phase.
>
> I'm not sure how to enforce "don't enable this unless you're developing
> the Overflow Behavior Types" with current Kconfig, given the randconfig
> gap... I have some memory of Arnd doing something special with his
> randconfigs to avoid these kinds of things, but I can't find it now.

You could depend on CONFIG_BROKEN, the canonical "if you enable this
and stuff breaks, it's your fault" flag?

Or if you want something even more explicit, maybe something like this
would do the job, so that you have to set an environment variable for
the kernel build to unlock the feature? (To be clear, I'm not fluent
in kconfig, and this is almost certainly a terribly convoluted way of
writing this - if you like the idea you should figure out the proper
syntax.)

config VERY_BROKEN
  def_bool $(success,sh -c 'echo $ENABLE_VERY_BROKEN_STUFF | grep "^y$"')

config FOO
  depends on VERY_BROKEN

Re: [linus:master] [crypto] 40b9969796: UBSAN:unsigned-integer-overflow_in_lib/crypto/chacha20poly1305-selftest.c

[Kees Cook]

On Wed, May 28, 2025 at 07:15:18PM +0200, Jann Horn wrote:
> On Wed, May 28, 2025 at 6:46 PM Kees Cook <kees@kernel.org> wrote:
> > On Tue, May 27, 2025 at 11:14:27PM -0700, Eric Biggers wrote:
> > > If this new sanitizer is going to move forward, is there any sort of plan or
> > > guide for how to update code to be compatible with it?  Specifically considering
> > > common situations where unsigned wraparound (which is defined behavior in C) can
> > > be intentionally relied on, like calculating the distance from the next N-byte
> > > boundary.  What are the best practices now?
> >
> > Hi, yes, this is still under development. I tried to make it hard to
> > enable accidentally (not via COMPILE_TEST, not UBSAN-default, etc), but
> > we (still) don't have a way to disable configs for randconfigs. :(
> >
> > We're hoping to see Clang 21 with the more versatile Overflow Behavior Types:
> > https://discourse.llvm.org/t/rfc-v2-clang-introduce-overflowbehaviortypes-for-wrapping-and-non-wrapping-arithmetic/86507
> >
> > and our current testing is showing many fewer false positives. (Having
> > run syzkaller for weeks now.)
> >
> > > Documentation/dev-tools/ubsan.rst says nothing about this and only mentions
> > > "undefined behavior", which this is not.
> >
> > Right -- this will get extensive documentation before we move it out of
> > its development phase.
> >
> > I'm not sure how to enforce "don't enable this unless you're developing
> > the Overflow Behavior Types" with current Kconfig, given the randconfig
> > gap... I have some memory of Arnd doing something special with his
> > randconfigs to avoid these kinds of things, but I can't find it now.
> 
> You could depend on CONFIG_BROKEN, the canonical "if you enable this
> and stuff breaks, it's your fault" flag?

Yeah. Talking with Justin out of band, he suggested the same. It's
easier to carry a 1 line patch downstream while we're testing to enable
this feature, so I'll send a patch to add CONFIG_BROKEN for now.

-Kees

-- 
Kees Cook