GNU make losing jobserver tokens

[Takashi Yano]

On Mon, 21 Mar 2022 15:28:17 +0100
Magnus Ihse Bursie wrote:
> Hi,
> 
> I'm working for Oracle on the OpenJDK build team. We're using GNU make 
> to build the JDK on all supported platforms. For Windows, we use Cygwin 
> as our build environment, including the Cygwin version of GNU make.
> 
> We have had a long-standing issue with make losing jobserver tokens. 
> ("long-standing" here means for years, and years, at least since GNU 
> make 4.0, up to and including the current latest version in Cygwin.)
> 
> Most runs end with something like:
> 
> make[2]: INTERNAL: Exiting with 11 jobserver tokens available; should be 
> 12!
> 
> Since the build still succeeds, and it just affects performance (and 
> typically not that much), we have not spend too much time getting to the 
> bottom of this.
> 
> Now, however, I've come across a machine where this happens repeatedly, 
> and on a much worse scale:
> 
> make[2]: INTERNAL: Exiting with 1 jobserver tokens available; should be 24!
> 
> This effectively turns the highly parallelized builds into 
> single-threaded builds, and is absolutely detrimental for performance. 
> On the flip side, this also makes for the perfect testing environment to 
> really get to the bottom of this issue.
> 
> I started out by sending a question to bug-make@gnu.org. The folks over 
> there reported that this was not a known problem with GNU make on 
> Windows in general, and that as far as they knew, the mingw port did not 
> suffer from this problem.
> 
> Instead, they suggested that it was a Cygwin-specific problem, possibly 
> related to issues with emulating Posix pipes and/or signals in Cygwin.
> 
> So, my first question is: Is this a known problem in Cygwin GNU make? 
> Are there any workarounds/fixes to get around it?
> 
> Otherwise: Any suggestions on how to go on and debug this? I am willing 
> to build and test an instrumented debug build of make, but I will need 
> assistance to find my way around the source and spot likely candidates 
> for the source of the problem.

I have tried to reproduce the issue by building OpenJDK
from source, however, I could not.

Instead, I encountered another issue.

Building OpenJDK sometimes (rarely) failed with error such as:

      0 [sig] make 5484 sig_send: error sending signal 11, pid 5484, pipe handle 0x118, nb 0, packsize 176, Win32 error 0
 124917 [main] make 5484 sig_send: error sending signal -72, pid 5484, pipe handle 0x118, nb 0, packsize 176, Win32 error 0
common/modules/GensrcModuleInfo.gmk:77: *** open: /home/yano/jdk/build/windows-x86-server-release/make-support/vardeps/make/common/modules/GensrcModuleInfo.gmk/jdk.accessibility/ALL_MODULES.vardeps: No such file or directory.  Stop.
make[2]: *** [make/Main.gmk:141: jdk.accessibility-gensrc-moduleinfo] Error 2
make[2]: *** Waiting for unfinished jobs....


I looked into this new problem and found that wait_sig() thread
crashes with segfault. It seems that accessing _main_tls causes
access violation if a signal is sent just after the process is
started.

static void WINAPI
wait_sig (VOID *)
{
  [...]
      if (!pack.mask)
	{
	  tl_entry = cygheap->find_tls (_main_tls);
	  dummy_mask = _main_tls->sigmask;       // <--- Segfault here
	  cygheap->unlock_tls (tl_entry);
	  pack.mask = &dummy_mask;
	}

I also found the following patch resolves the issue.

diff --git a/winsup/cygwin/sigproc.cc b/winsup/cygwin/sigproc.cc
index 62df96652..3824af199 100644
--- a/winsup/cygwin/sigproc.cc
+++ b/winsup/cygwin/sigproc.cc
@@ -1325,6 +1325,10 @@ wait_sig (VOID *)
   _sig_tls = &_my_tls;
   bool sig_held = false;
 
+  /* Wait for _main_tls initialization. */
+  while (!cygwin_finished_initializing)
+    Sleep (10);
+
   sigproc_printf ("entering ReadFile loop, my_readsig %p, my_sendsig %p",
 		  my_readsig, my_sendsig);
 

I guess _main_tls may not be initialized correctly until
cygwin_finished_initializing is set.

Any comments would be appreciated.

-- 
Takashi Yano <takashi.yano@nifty.ne.jp>