Fix hang caused by memory corruption caused by insufficient alternate stack space for signal handlers.

[bryant]

Summary

On systems that support alternate signal stack frames,, the code currently asks
for MINSIGSTKSZ + 8192 _bytes_, which translates to roughly 10 KiB. However,
this amount is far too low and results in later signal handlers writing past the
bounds of their allotted space.

On trunk builds compiled with the latest libcxx/-abi, the result is an acutely
unpleasant hang (as opposed to abort) upon false assertions.

The new size of 8 MiB was on the discerned similarity between "8192" and the
default Linux stack size of 8192 KiB:

$ ulimit -s
8192

Source of original code: https://reviews.llvm.org/rL270273

[bryant]

Fix typo.

[joerg]

I find that size to be excessive. Sure, two pages on many architectures is very small, but a more reasonable and still conservative limit would be 1MB top.

[bryant]

8 MiB => 1 MiB

[rsmith]

1MB is *huge* for a sigaltstack. What are we doing in our signal handlers that requires more than 8KB? It'd be preferable to figure out how much we're actually using, and allocate some small multiple of that.

[bryant]

I don't know of any technique to measure max stack usage outside of breaking on writes to $rsp in gdb, or binary searching for the value of AltStackSize that causes a hang. Both are terrible for large programs. If you or anyone else have suggestions to do this, I'd love to hear them.

[bryant]

So I bit the bullet and binary searched for the hang. The magic value seems to be MINSIGSTKSZ + 8192 + 512 + 256 - 128 + 64 - 32 + 16 - 8 - 4 - 2 + 1 + 1 = MINGSIGSTKSZ + 8868. Below this, hang. So I suppose 64KiB as discussed over IRC would be reasonable for the near future. 64K ought to be enough for anyone.

[bryant]

"'64K ought to be enough for anyone' -- bill gates " -- bryant.

[rsmith]

Committed as r279599.

[vitalybuka]

This patch has memory leaks:
http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux-fast/builds/15762/steps/check-llvm%20asan/logs/stdio

[vitalybuka]

it is already fixed by r279605

lib/Support/Unix/Signals.inc

 #if defined(HAVE_SIGALTSTACK)
 // Hold onto the old alternate signal stack so that it's not reported as a leak.
 // We don't make any attempt to remove our alt signal stack if we remove our
 // signal handlers; that can't be done reliably if someone else is also trying
 // to do the same thing.
 static stack_t OldAltStack;
 
 static void CreateSigAltStack() {
-  const size_t AltStackSize = MINSIGSTKSZ + 8192;
+  const size_t AltStackSize = MINSIGSTKSZ + 64 * 1024;
 
   // If we're executing on the alternate stack, or we already have an alternate
   // signal stack that we're happy with, there's nothing for us to do. Don't
   // reduce the size, some other part of the process might need a larger stack
   // than we do.
   if (sigaltstack(nullptr, &OldAltStack) != 0 ||
       OldAltStack.ss_flags & SS_ONSTACK ||
       (OldAltStack.ss_sp && OldAltStack.ss_size >= AltStackSize))