On 1/21/22 15:34, Jeffrey Walton wrote:
> On Fri, Jan 21, 2022 at 3:19 AM Dumitru Ceara <[email protected]> wrote:
>>
>> On 1/21/22 02:12, Jeffrey Walton wrote:
>>
>>> I am testing Open vSwitch 2.16.2. Undefined Behavior sanitizer is
>>> producing some findings.
>>>
>>> My question is, is the undefined behavior something the project would
>>> be interested in fixing?
>>>
>>
>> I think so.
>>
>>> If so, I can send over the findings and a proposed patch. If not, I
>>> can carry the patch privately.
>>>
>>
>> I had posted a patch series to take care of the issues currently
>> reported by UB Sanitizer:
>>
>> https://patchwork.ozlabs.org/project/openvswitch/list/?series=277900&state=*
>>
>> It will need a new revision and Adrian (in cc) is also working on some
>> more patches to deal with other UB (not flagged by UB Sanitizer).  Would
>> the patch you were planning to propose overlap with this work?
>>
>> In any case, I think it would be great if we could work together
>> (reviews and patches) to get tests passing with UB Sanitizer and ideally
>> running these kinds of checks automatically, in CI.  E.g., with the
>> series above applied all unit tests are also run with UB Sanitizer enabled:
>>
>> https://github.com/dceara/ovs/runs/4593681037?check_suite_focus=true
> 
> Awesome.
> 
> In case anyone else is interested...
> 
> export CFLAGS="-fsanitize=undefined
> -fno-sanitize=integer-divide-by-zero
> -fno-sanitize=float-divide-by-zero -fno-sanitize-recover=all"
> export CXXFLAGS="-fsanitize=undefined
> -fno-sanitize=integer-divide-by-zero
> -fno-sanitize=float-divide-by-zero -fno-sanitize-recover=all"
> export LDFLAGS="-fsanitize=undefined
> -fno-sanitize=integer-divide-by-zero
> -fno-sanitize=float-divide-by-zero -fno-sanitize-recover=all"
> 
> We don't need the divide-by-0 findings. They are usually false
> positives. IEEE-754 leaves it up to the implementation to trap.
> 
> Then, configure && make && make check. Once the self tests run (or hang):
> $ grep -IR 'runtime error' ./* 2>/dev/null
> ./tests/testsuite.dir/0245/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7fff61547e0c for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0262/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7fff3954c21c for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0408/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7fff3e1cf8ac for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0397/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7ffee8037d0c for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0410/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7ffd62e8585c for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0211/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7ffc5e3ecb0c for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0051/testsuite.log:+tests/test-hash.c:59:40:
> runtime error: shift exponent 64 is too large for 64-bit type 'long
> unsigned int'
> ./tests/testsuite.dir/0190/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7ffff13fe65c for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0302/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7ffc41b735bc for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0331/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7ffd6c61c2bc for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0381/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7fff9f4a0a3c for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0358/stderr:lib/hash.h:219:17: runtime error:
> load of misaligned address 0x7ffca418c60c for type 'const uint64_t',
> which requires 8 byte alignment
> ./tests/testsuite.dir/0358/testsuite.log:lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7ffca418c60c for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0247/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7fff27ad680c for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0124/testsuite.log:+lib/classifier.c:1763:5:
> runtime error: applying non-zero offset 18446744073709551592 to null
> pointer
> ./tests/testsuite.dir/0285/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7ffeb2264bec for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0384/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7fff69754d9c for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0417/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7ffc27b1a5bc for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0054/testsuite.log:+tests/test-cmap.c:106:9:
> runtime error: applying non-zero offset 18446744073709551608 to null
> pointer
> ./tests/testsuite.dir/0406/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7ffeb01ed1bc for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0366/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7fffc6c1783c for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0226/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7ffff0cdb64c for type 'const
> uint64_t', which requires 8 byte alignment
> ./tests/testsuite.dir/0053/testsuite.log:+tests/test-hindex.c:60:5:
> runtime error: applying non-zero offset 18446744073709551608 to null
> pointer
> ./tests/testsuite.dir/0426/stderr:lib/nx-match.c:2311:5: runtime
> error: applying non-zero offset 18446744073709551600 to null pointer
> ./tests/testsuite.dir/0426/testsuite.log:lib/nx-match.c:2311:5:
> runtime error: applying non-zero offset 18446744073709551600 to null
> pointer
> ./tests/testsuite.dir/0325/testsuite.log:+lib/hash.h:219:17: runtime
> error: load of misaligned address 0x7ffe609f708c for type 'const
> uint64_t', which requires 8 byte alignment
> ...
> 
> I see the findings in lib/hash.h:219 because I also compile with
> -march=native, and the machine has AVX2,AVX,SSE42,SSE41,...
> lib/hash.h:219 is due to the uint64_t* cast on a 32-bit aligned
> buffer.

Interesting, I see this now too after switching to -march=native.

Do you already have a fix for this in mind?

I don't see a way of fixing it without changing the SSE42 version of
hash_words_inline() to either:

a. implement a different hash function

or

b. switch to hashing one u32 at a time if the initial buffer is not
64-but aligned.

I'm suspecting both these approaches will affect performance.

Thanks,
Dumitru

_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to