facebook
Linux: fix clock_gettime and recvmmsg build problems
Replaces https://github.com/facebook/folly/pull/1514. This more minimal change makes folly compile successfully by showing where clock_gettime() is found on Linux and solving the problem created by ee1e6c7b7219a545093e07b785efb5f396fac1ef when recvmmsg is not found. Further changes to code in //folly/portability would be needed in order to allow Ninja (for example) to find these Linux syscalls at compile time. Nonetheless, clock_gettime_wrappers_test passes.
On my Linux localhost, the SmallLocks.MicroLockSlotsAsData passes. Of course, the test result could depend on the kernel version.
$ /home/alison/gitsrc/fbcode-scratch-24Jan2021/build/folly/small_locks_test [==========] Running 17 tests from 2 test suites. [----------] Global test environment set-up. [----------] 16 tests from SmallLocks [ RUN ] SmallLocks.SpinLockCorrectness [ OK ] SmallLocks.SpinLockCorrectness (7391 ms) [ RUN ] SmallLocks.PicoSpinCorrectness [ OK ] SmallLocks.PicoSpinCorrectness (25 ms) [ RUN ] SmallLocks.PicoSpinSigned [ OK ] SmallLocks.PicoSpinSigned (0 ms) [ RUN ] SmallLocks.PicoSpinLockThreadSanitizer [ SKIPPED ] SmallLocks.PicoSpinLockThreadSanitizer (0 ms) [ RUN ] SmallLocks.RegClobber [ OK ] SmallLocks.RegClobber (0 ms) [ RUN ] SmallLocks.MicroLock [ OK ] SmallLocks.MicroLock (93 ms) [ RUN ] SmallLocks.MicroLockTryLock [ OK ] SmallLocks.MicroLockTryLock (0 ms) [ RUN ] SmallLocks.MicroLockSlotsAsData [ OK ] SmallLocks.MicroLockSlotsAsData (0 ms) [ RUN ] SmallLocks.MicroSpinLockStressTestLockTwoThreads [ OK ] SmallLocks.MicroSpinLockStressTestLockTwoThreads (2000 ms) [ RUN ] SmallLocks.MicroSpinLockStressTestLockHardwareConcurrency [ OK ] SmallLocks.MicroSpinLockStressTestLockHardwareConcurrency (2000 ms) [ RUN ] SmallLocks.PicoSpinLockStressTestLockTwoThreads [ OK ] SmallLocks.PicoSpinLockStressTestLockTwoThreads (2001 ms) [ RUN ] SmallLocks.PicoSpinLockStressTestLockHardwareConcurrency [ OK ] SmallLocks.PicoSpinLockStressTestLockHardwareConcurrency (2000 ms) [ RUN ] SmallLocks.MicroSpinLockStressTestTryLockTwoThreads [ OK ] SmallLocks.MicroSpinLockStressTestTryLockTwoThreads (2000 ms) [ RUN ] SmallLocks.MicroSpinLockStressTestTryLockHardwareConcurrency [ OK ] SmallLocks.MicroSpinLockStressTestTryLockHardwareConcurrency (2001 ms) [ RUN ] SmallLocks.PicoSpinLockStressTestTryLockTwoThreads [ OK ] SmallLocks.PicoSpinLockStressTestTryLockTwoThreads (2000 ms) [ RUN ] SmallLocks.PicoSpinLockStressTestTryLockHardwareConcurrency [ OK ] SmallLocks.PicoSpinLockStressTestTryLockHardwareConcurrency (2000 ms) [----------] 16 tests from SmallLocks (23511 ms total)
[----------] 1 test from SmallLocksk [ RUN ] SmallLocksk.MicroSpinLockThreadSanitizer [ SKIPPED ] SmallLocksk.MicroSpinLockThreadSanitizer (0 ms) [----------] 1 test from SmallLocksk (0 ms total)
[----------] Global test environment tear-down [==========] 17 tests from 2 test suites ran. (23511 ms total) [ PASSED ] 15 tests. [ SKIPPED ] 2 tests, listed below: [ SKIPPED ] SmallLocks.PicoSpinLockThreadSanitizer [ SKIPPED ] SmallLocksk.MicroSpinLockThreadSanitizer
I ran an the unit tests here on this system; only Random.SecureFork fails. 36 tests are skipped, mostly because more code is needed in folly/Portability.h so that cmake finds Linux's pthread_atfork() implementation.
The .cpp changes here should have no affect, and the config changes don't make sense. If clock_gettime is not accessible via time.h, pipe2 is not in unistd.h, and accept4 is not in sys/sockets.h then I believe something has gone wrong with your linux install, or else something has failed to add the root include path when building Folly.
The .cpp changes here should have no affect, and the config changes don't make sense. If
clock_gettimeis not accessible viatime.h,pipe2is not inunistd.h, andaccept4is not insys/sockets.hthen I believe something has gone wrong with your linux install, or else something has failed to add the root include path when building Folly.
The Debian Testing release has much newer packages than Ubuntu 18.04. In particular, Ubuntu 18.04 includes linux-libc-dev version 4.15.0-135.139, while Debian ships linux-libc-dev version 5.10.12-1. The implementation of clock_gettime in 5.10.12-1 fixes the "Y2038 problem": https://sourceware.org/pipermail/libc-alpha/2020-May/113643.html The consequence is that the implementation of clock_gettime() has changed. Perhaps either portability/Time.cpp or CMake/FollyConfigChecks.cmake needs logic that detects the libc version and then chooses the right headers. Undoubtedly the same problem will occur with newer versions of Ubuntu that include the newer libc.
The problems with the other system calls are likely similar, but De I have not had a chance to analyze them yet. Debian also has newer versions of Cmake and boost than Ubuntu which might require changes to detection scripts.