[ruby-core:82809] [Ruby trunk Bug#13856] MinGW / mswin intermittent failure in test/socket/test_socket.rb

Issue #13856 has been updated by h.shirosaki (Hiroshi Shirosaki).

ruby -v set to ruby 2.5.0dev (2017-09-11 trunk 59829) [x64-mingw32]

Segmentation fault is caused by `RBASIC_CLASS(err)` access in hook_before_rewind() in vm.c:1667.
`err` is not a valid pointer.

Here is gdb output.

~~~
Run options: "--ruby=./miniruby.exe -I../snapshot/lib -I. -I.ext/common  ../snapshot/tool/runruby.rb --extout=.ext --debugger -- --disable-gems" --excludes-dir=../snapshot/test/excludes --name=!/memory_leak/ -v -ntest_closed_read

# Running tests:

[1/1] TestSocket#test_closed_read[New Thread 16292.0x331c]
[New Thread 16292.0x57f4]
[Thread 16292.0x331c exited with code 0]

Thread 9 received signal SIGSEGV, Segmentation fault.
[Switching to Thread 16292.0x57f4]
0x00000000680aad1a in hook_before_rewind (th=th@entry=0x33671d0, will_finish_vm_exec=will_finish_vm_exec@entry=0, state=6, err=err@entry=0xc0000241,
    cfp=<optimized out>) at ../snapshot/vm.c:1667
1667        if (state == TAG_RAISE && RBASIC_CLASS(err) == rb_eSysStackError) {
(gdb) bt
#0  0x00000000680aad1a in hook_before_rewind (th=th@entry=0x33671d0, will_finish_vm_exec=will_finish_vm_exec@entry=0, state=6, err=err@entry=0xc0000241,
    cfp=<optimized out>) at ../snapshot/vm.c:1667
#1  0x00000000680b6ffd in vm_exec (th=0x31e000, th@entry=0x33671d0) at ../snapshot/vm.c:2003
...
~~~

`err` seems to come from rb_threadptr_execute_interrupts() in thread.c.
Adding volatile as the following suppresses segfault on my test.
But this patch sometimes causes another error (TypeError: exception class/object expected).
I don't know the reason.


```patch
diff --git a/thread.c b/thread.c
index d706ee469b..fd1ee78933 100644
--- a/thread.c
+++ b/thread.c
@@ -2058,7 +2058,7 @@ rb_threadptr_execute_interrupts(rb_thread_t *th, int blocking_timing)
 
 	/* exception from another thread */
 	if (pending_interrupt && rb_threadptr_pending_interrupt_active_p(th)) {
-	    VALUE err = rb_threadptr_pending_interrupt_deque(th, blocking_timing ? INTERRUPT_ON_BLOCKING : INTERRUPT_NONE);
+	    volatile VALUE err = rb_threadptr_pending_interrupt_deque(th, blocking_timing ? INTERRUPT_ON_BLOCKING : INTERRUPT_NONE);
 	    thread_debug("rb_thread_execute_interrupts: %"PRIdVALUE"\n", err);
 
 	    if (err == Qundef) {
```

```
  1) Failure:
TestSocket#test_closed_read [C:/Users/h.shirosaki/work/rubyinstaller2-packages/mingw-w64-ruby25/src/snapshot/test/socket/test_socket.rb:543]:
[ruby-core:35203].
[IOError] exception expected, not.
Class: <TypeError>
Message: <"exception class/object expected">
---Backtrace---
C:/Users/h.shirosaki/work/rubyinstaller2-packages/mingw-w64-ruby25/src/snapshot/test/socket/test_socket.rb:537:in `readline'
C:/Users/h.shirosaki/work/rubyinstaller2-packages/mingw-w64-ruby25/src/snapshot/test/socket/test_socket.rb:537:in `block in test_closed_read'
---------------
```

----------------------------------------
Bug #13856: MinGW / mswin intermittent failure in test/socket/test_socket.rb
https://bugs.ruby-lang.org/issues/13856#change-66688

* Author: MSP-Greg (Greg L)
* Status: Open
* Priority: Normal
* Assignee: 
* Target version: 
* ruby -v: ruby 2.5.0dev (2017-09-11 trunk 59829) [x64-mingw32]
* Backport: 2.2: UNKNOWN, 2.3: UNKNOWN, 2.4: UNKNOWN
----------------------------------------
Recently, there have been two Appveyor fails, [first](https://ci.appveyor.com/project/ruby/ruby/build/1.0.5036) and [second](https://ci.appveyor.com/project/ruby/ruby/build/1.0.5078).

Both listed the following as the error:

```
running file: C:/projects/ruby/test/socket/test_socket.rb

Some worker was crashed. It seems ruby interpreter's bug
or, a bug of test/unit/parallel.rb. try again without -j
option.
```

I have a similar intermittent error as a silent segfault in MinGW builds.  I've attached the patch file I use, it patches the [TestSocket#test_closed_read](https://github.com/ruby/ruby/blob/ed740754bd96e845bc86f070b90267efc1ad5136/test/socket/test_socket.rb#L528-L548) method, and simply adds `sock.autoclose = false` after `sock` is created.

I haven't done much Ruby socket coding, so I've never looked into the issue, but the test passes and is stable with the patch.  For all I know, the patch may 'end-around' the whole point of the test.

I don't know if this helps identify the real issue or not. If not, feel free to close.

---Files--------------------------------
segv-test-socket-test_socket.rb.patch (782 Bytes)


-- 
https://bugs.ruby-lang.org/

Unsubscribe: <mailto:[e-mail]?subject=unsubscribe>
<http://lists.ruby-lang.org/cgi-bin/mailman/options/ruby-core>