Investigate flaky test-cli-node-options

[Trott]

test-cli-node-options has been failing a lot on arm lately in CI. I assume it's the bug reported in #21383 ("make test: use after free: parallel/test-cli-node-options").

Sample failure: https://ci.nodejs.org/job/node-test-commit-arm/20786/nodes=ubuntu1604-arm64/consoleText

Host: test-packetnet-ubuntu1604-arm64-2

not ok 184 parallel/test-cli-node-options
  ---
  duration_ms: 1.266
  severity: fail
  exitcode: 1
  stack: |-
    assert.js:753
        throw newErr;
        ^
    
    AssertionError [ERR_ASSERTION]: ifError got unwanted exception: Command failed: /home/iojs/build/workspace/node-test-commit-arm/nodes/ubuntu1604-arm64/out/Release/node -e console.log("B")
    
    
    #
    # Fatal error in , line 0
    # Check failed: (perf_output_handle_) != nullptr.
    #
    #
    #
    #FailureMessage Object: 0xffffe553e558
        at ChildProcess.exithandler (child_process.js:294:12)
        at ChildProcess.emit (events.js:189:13)
        at maybeClose (internal/child_process.js:978:16)
        at Socket.stream.socket.on (internal/child_process.js:396:11)
        at Socket.emit (events.js:189:13)
        at Pipe._handle.close (net.js:612:12)
  ...

[Trott]

And again. Probably time to mark this as flaky. Will open a PR.

test-packetnet-ubuntu1604-arm64-2

https://ci.nodejs.org/job/node-test-commit-arm/20790/nodes=ubuntu1604-arm64/consoleText

not ok 175 parallel/test-cli-node-options
  ---
  duration_ms: 0.986
  severity: fail
  exitcode: 1
  stack: |-
    assert.js:753
        throw newErr;
        ^
    
    AssertionError [ERR_ASSERTION]: ifError got unwanted exception: Command failed: /home/iojs/build/workspace/node-test-commit-arm/nodes/ubuntu1604-arm64/out/Release/node -e console.log("B")
    
    
    #
    # Fatal error in , line 0
    # Check failed: (perf_output_handle_) != nullptr.
    #
    #
    #
    #FailureMessage Object: 0xffffcb5fc4d8
        at ChildProcess.exithandler (child_process.js:294:12)
        at ChildProcess.emit (events.js:189:13)
        at maybeClose (internal/child_process.js:978:16)
        at Socket.stream.socket.on (internal/child_process.js:396:11)
        at Socket.emit (events.js:189:13)
        at Pipe._handle.close (net.js:612:12)
  ...

[gireeshpunathil]

easily recreated if you run this a hundred times in freebsd

node --perf-basic-prof -e 'console.log("B")'

#
# Fatal error in , line 0
# Check failed: (perf_output_handle_) != nullptr.
#
#
#

(gdb) where
#0  0x00000000015a8eb2 in v8::base::OS::Abort ()
#1  0x0000000000f79516 in v8::internal::PerfBasicLogger::PerfBasicLogger ()
#2  0x0000000000f7fb92 in v8::internal::Logger::SetUp ()
#3  0x0000000000f5b25c in v8::internal::Isolate::Init ()
#4  0x000000000117dd3f in v8::internal::Snapshot::Initialize ()
#5  0x0000000000ace03b in v8::Isolate::Initialize ()
#6  0x00000000008fba32 in node::NewIsolate ()
#7  0x00000000008fdc88 in node::Start ()
#8  0x00000000008fbfdb in node::Start ()
#9  0x00000000008ae095 in _start ()
#10 0x00000008022f0000 in ?? ()
#11 0x0000000000000000 in ?? ()
(gdb)      

this corresponds to:

node/deps/v8/src/log.cc

Lines 294 to 296 in a6f69eb

	perf_output_handle_ =
	base::OS::FOpen(perf_dump_name.start(), base::OS::LogFileOpenMode);
	CHECK_NOT_NULL(perf_output_handle_);

truss output showed this:

24134: open("/tmp/perf-24134.map",O_WRONLY|O_CREAT|O_TRUNC,0666) ERR#13 'Permission denied'


#
# Fatal error in , line 0
# 24134: write(2,"\n\n#\n# Fatal error in , line 0"...,32) = 32 (0x20)
Check failed: (perf_output_handle_) != nullptr.24134: write(2,"Check failed: (perf_output_handl"...,47) =
 47 (0x2f)

#
#
#
#FailureMessage Object: 0x7fffffffdc4024134: write(2,"\n#\n#\n#\n#FailureMessage Objec"...,45) = 45 (0x2d
)

$ l /tmp/perf-24134.map
-rw-r--r--  1 iojs  wheel  68110 May 14  2018 /tmp/perf-24134.map
$ whoami
freebsd
$ 

so this is basically a temporary name collision. Unlikely to happen on production. One of:

• clean up the temp folder prior to runs (test / test harness)
• only one type of user run things
• retry on failure such as EACCES (PerfBasicLogger::PerfBasicLogger)

[sam-github]

Am I reading this correctly, its pid reuse which is causing the tmpfile name collision? Is it possible to configure freebsd to use a larger pid space so the reuse doesn't occur during our tests runs? And does this point to a problem with the tempfile name generation?

[gireeshpunathil]

@sam-github - yes, this is an issue with PIDs in small ranges that get re-used often, but when composed into file names that is being used by different users causes permission issues, leading to temp file generation failure.

I don't know the admin command to change the PID pattern, and do not have the permission to try that even if I know; though widening the PID even by one digit should resolve collision I guess.

[gireeshpunathil]

on a side note: though I recreated this in __freebsd__ and the issue is 100% same with the original report, the original failure was reported in arm. So I am wondering whether fixing it at the host level would be the right thing or not.

[sam-github]

Perhaps the test should use the env to set the temp directory to per-user location, like $PWD/out/Release, to avoid conflict? Perhaps in tools/test.py?

Though I'm a bit confused, are multiple users running the node tests at the same time? If there are multiple parallel test runs by the same user, permission problems shouldn't occur, though other conflicts could.

[gireeshpunathil]

I don't know; let us ask @nodejs/testing and @Trott