[Support] Use SA_ONSTACK in CrashRecoveryContext

[thejh]

Summary

The signal handlers LLVM installs in RegisterHandlers() (used for things
like printing a stack trace) use SA_ONSTACK so that on stack overflow,
the signal handler still has a stack from which it can run.

However, CrashRecoveryContext overwrites those signal handlers with ones
without SA_ONSTACK. Therefore, if the stack overflows while a
CrashRecoveryContext is active, the kernel fails to write a signal frame
and terminates the process.

Fix it by registering CrashRecoveryContext's signal handlers such that
they also run off the alternate signal stack.

You can test this by calling an endlessly recursing function directly
below the call to llvm::CrashRecoveryContext::Enable() in
clang/tools/driver/driver.cpp.

[Dinistro]

It seems that we ran into a similar issue recently, but I'm not experienced in this part to determine if that patch fixes the issue. Could someone take a quick look?

llvm/lib/Support/CrashRecoveryContext.cpp

static void CrashRecoverySignalHandler(int Signal) {
   if (CRCI)
     const_cast<CrashRecoveryContextImpl *>(CRCI)->HandleCrash(RetCode, Signal);
 }
 
 static void installExceptionOrSignalHandlers() {
   // Setup the signal handler.
   struct sigaction Handler;
   Handler.sa_handler = CrashRecoverySignalHandler;
-  Handler.sa_flags = 0;
+  Handler.sa_flags = SA_ONSTACK;
   sigemptyset(&Handler.sa_mask);
 
   for (unsigned i = 0; i != NumSignals; ++i) {
     sigaction(Signals[i], &Handler, &PrevActions[i]);
   }
 }
 
 static void uninstallExceptionOrSignalHandlers() {