Bug #747

5.2.1: test suite failure in 'test_chunk_internet_checksum' on big-endian hosts

Added by Anonymous almost 8 years ago. Updated almost 8 years ago.

Target version:
Start date:
Due date:
Estimated time:
Affected version:



I updated the Debian packages of strongswan to version 5.2.1, and the test suite fails on big-endian architectures in the new 'test_chunk_internet_checksum' test, with the following:

Running case 'chunk_internet_checksum': -
Failure in 'test_chunk_internet_checksum': 0x442e != ntohs(sum) (17454 != 17455) (suites/test_chunk.c:799, i = 0)

Full build logs are here:
- powerpc:
- s390x:


Associated revisions

Revision eb218ff8 (diff)
Added by Tobias Brunner almost 8 years ago

chunk: Fix internet checksum calculation on big-endian systems

ntohs() might be defined as noop (#define ntohs(x) (x)) so we have
to manually shorten the negated value (gets promoted to an int).

Fixes #747.

Revision e7c582e6 (diff)
Added by Tobias Brunner almost 8 years ago

unit-tests: Fix internet checksum tests on big-endian systems

We actually need to do a byte-swap, which ntohs() only does on
little-endian systems.

Fixes #747.


#1 Updated by Tobias Brunner almost 8 years ago

  • Status changed from New to Feedback
  • Assignee set to Tobias Brunner

Running case 'chunk_internet_checksum': -
Failure in 'test_chunk_internet_checksum': 0x442e != ntohs(sum) (17454 != 17455) (suites/test_chunk.c:799, i = 0)

I ran the test suite in QEMU (mips) and got the same result (17455 instead of 17454). I tried to replicate the problem in a simple test program, but wasn't able to. That is, until I included arpa/inet.h (ntohs() might be a compiler built-in as it worked fine without that include).

Further analysis showed that the problem was caused by the definition of ntohs() and friends in netinet/in.h. On big-endian systems they are defined as noops (e.g. #define ntohs(x) (x)), so there is no function call (or cast) involved that would shorten the argument (or result) of ntohs() to 16-bit. On the first line here

u_int16_t chunk_internet_checksum_inc(chunk_t data, u_int16_t checksum)
    u_int32_t sum = ntohs(~checksum);

checksum is promoted to 32-bit when it is negated, so with an initial value of 0xffff and without shortening to 16-bit we end up with sum = 0xffff0000. If ntohs() was a function call or if there was a cast to 16-bit everything would be fine, but since that doesn't happen the calculation later is incorrect.

There was also a bug in the unit test on big-endian systems. I pushed fixes to the inet-checksum branch.

#2 Updated by Anonymous almost 8 years ago

Thanks. I included both commits in the 5.2.1-2 version of the package, and the libstrongswan test suite now passes.

#3 Updated by Tobias Brunner almost 8 years ago

  • Tracker changed from Issue to Bug
  • Status changed from Feedback to Closed
  • Target version set to 5.2.2
  • Resolution set to Fixed

Also available in: Atom PDF