CHECK(leader_state_cache_.is_lock_free()) fails in gcc 7 #1189
Comments
|
Output of yb-master.raju-ubuntu1804.raju.log.INFO.20190402-234308.21926 Log file created at: 2019/04/02 23:43:08 |
|
cc - @bmatican |
|
@spolitov does https://en.cppreference.com/w/cpp/atomic/atomic/is_lock_free
|
|
Moving comments from #1246 here: From @d-uspenskiy : It looks like g++-7 generates non-lock-free code for 128 bits atomic structures https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80878 as a result https://github.com/YugaByte/yugabyte-db/blob/master/src/yb/consensus/replica_state.cc#L111 Some description about changes in g++-7 |
|
Two more potentially relevant links: http://gcc.1065356.n8.nabble.com/GCC-in-line-assembly-and-the-removal-of-mcx16-td1378660.html |
mbautin
changed the title
|
Could we just use |
|
From an offline discussion with @spolitov:
|
|
The check should be added for 16 bytes atomics, so to revisit we could use the following command: It will give all suspicious usages of std::atomic, that should be checked. |
|
BTW here is the patch that @rajukumaryb has been using locally: https://gist.githubusercontent.com/mbautin/bbb02cbf60df635b8a92a548789d54ed/raw |
|
@d-uspenskiy: we should be able to upgrade to gcc 7.x (the default version on Ubuntu 18.04) when we fix the lock-free/atomic-related issues mentioned above. |
Slightly updated regex: |
Summary:
For std::atomic `gcc-7` generates non lock-free code for 16 bytes structures even in case `-mcx16` flags is
used for compilation.
Solution is to substitute `std::atomic` with `boost::atomic` for all 16 bytes structures used in YB code
Also false positive `array-bounds` error (due to agressive `-O3` optimization in `release` build) is
fixed in `slice_transform_test.cc`
```
In file included from src/yb/rocksdb/slice_transform.h:34:0,
from src/yb/rocksdb/util/slice_transform_test.cc:24:
src/yb/util/slice.h: In member function ‘virtual void rocksdb::SliceTransformTest_CapPrefixTransform_Test::TestBody()’:
src/yb/util/slice.h:64:57: error: array subscript is below array bounds [-Werror=array-bounds]
Slice(const uint8_t* d, size_t n) : begin_(d), end_(d + n) {}
```
Using `Slice()` explicitly instead of implicit `Slice(const char* s)` constructor for `""` (empty string) solves the issue.
Problem is very specific for this particular place. There is no issue with creation of `Slice` from `""` all the time.
Test Plan: build and run YB cluster in `debug` and `release` mode with using gcc-7 (`ubuntu` 16+ have gcc-7 as a default compiler)
Reviewers: raju, mikhail, sergei
Reviewed By: sergei
Subscribers: ybase, bogdan
Differential Revision: https://phabricator.dev.yugabyte.com/D6800
Summary: D6777 and D6800 was made in parallel, so one 16 bytes std::atomic left in LockFreeStack. Changed it to boost::atomic. Test Plan: Jenkins Reviewers: dmitry, bogdan Reviewed By: bogdan Subscribers: ybase Differential Revision: https://phabricator.dev.yugabyte.com/D6811
…cture) Summary: Due to gcc 7 issue with lock free structures described in #1189 the following check fails https://github.com/yugabyte/yugabyte-db/blob/master/src/yb/tablet/operations/operation_driver.cc#L93 Solution: boost::atomic should be used instead of std::atomic Test Plan: Manual. Reinit db on centos build ``` ./yb_build.sh reinitdb ``` Reviewers: sergei Reviewed By: sergei Subscribers: ybase Differential Revision: https://phabricator.dev.yugabyte.com/D7362
Original description of the issue as filed by @rajukumaryb :
Master process does not start on Ubuntu 18.04 following a "destroy + create" sequence
The text was updated successfully, but these errors were encountered: