Intermittent segmentation faults on Circle CI

[bslobodin]

Yesterday we saw a number of builds fail with a seg fault, that is seemingly related to bootsnap. It was inconsistent at best, but happened on a number of unrelated builds. I'd be happy to share more logs, if you think those would help?

/home/circleci/project/vendor/bundle/ruby/2.3.0/gems/bootsnap-1.1.2/lib/bootsnap/load_path_cache/store.rb:72: [BUG] Segmentation fault at 0x00000000000033
ruby 2.3.4p301 (2017-03-30 revision 58214) [x86_64-linux]

-- Control frame information -----------------------------------------------
c:0045 p:---- s:0201 e:000200 CFUNC  :dump
c:0044 p:0078 s:0197 e:000194 METHOD /home/circleci/project/vendor/bundle/ruby/2.3.0/gems/bootsnap-1.1.2/lib/bootsnap/load_path_cache/store.rb:72
c:0043 p:0015 s:0191 e:000190 METHOD /home/circleci/project/vendor/bundle/ruby/2.3.0/gems/bootsnap-1.1.2/lib/bootsnap/load_path_cache/store.rb:53
c:0042 p:0038 s:0188 e:000186 METHOD /home/circleci/project/vendor/bundle/ruby/2.3.0/gems/bootsnap-1.1.2/lib/bootsnap/load_path_cache/store.rb:45
c:0041 p:0011 s:0184 e:000183 METHOD /home/circleci/project/vendor/bundle/ruby/2.3.0/gems/bootsnap-1.1.2/lib/bootsnap/load_path_cache/cache.rb:140
c:0040 p:0070 s:0180 e:000179 BLOCK  /home/circleci/project/vendor/bundle/ruby/2.3.0/gems/bootsnap-1.1.2/lib/bootsnap/load_path_cache/cache.rb:123 [FINISH]
c:0039 p:---- s:0178 e:000177 CFUNC  :synchronize

💾 build12502.log

[casperisfine]

So the backtrace indicates the issue is likely inside MessagePack.dump

bootsnap/lib/bootsnap/load_path_cache/store.rb

Line 72 in d84bf7a

File.binwrite(tmp, MessagePack.dump(@data))

What version of message pack are you using? Because bootsnap requires ~> 1.0, so there is many compatible versions.

[casperisfine]

Nvm, I found it in the dump: msgpack-1.1.0

[bslobodin]

We've since disabled Bootsnap in CI for our project, keeping it only to local environments. It was still segfaulting once in a while.

[ahuth]

Looks like msgpack fixed a segfault in January of 2018 (after this issue was logged). Not sure if its the same one referenced here or not, but note the changelog notes for v1.2.2.

[burke]

Haven't heard about this in ages, probably that was the fix.