The callbacks need to be invoked in different threads to avoid clogging
	the IO pool.

	Fixes bug #639656.

Pass the listener as it's own async state so that we avoid potential null derefs in the callback. Also wrap the BeginAccept and EndAccept calls in their own try/catches in case they throw an exception.

Fix possible state corruption in the case where Upgradeable wasn't reactivating its read bit after a downgrade and set events accordingly. Fix possible race where a write process wannabe could step in the middle of Upgradeable actually upgrading to write mode.

Conflicts:

	mcs/class/System.Core/System.Threading/ReaderWriterLockSlim.cs

Rework TCP listening model, and fix AcceptTcpClient() to not drop timed out requests that actually steals next requests in the air.

Turn Is*LockHeld properties into thread-specific values in ReaderWriterLockSlim. Add corresponding unit tests.

The patch removes a small race condition where a boolean flag is set after acquiring a lock to
indicate to code executing in the finally {} block that it should release the lock. The boolean
variable is now removed and the lock is released unconditionally. It carries a potential to
throw an exception when the lock is not held, but it's better than to fail to release it and
lead the application to a deadlock.

	Ditto.

	EpollEvent has to be one uint32 plus one int64, i.e. 12 bytes long.
	Manos libres now.

It could cause a deadlock when a writer tested the rwlock value before RwRead was removed and another Upgradeable operation happened at the same time.

	* AssemblyBuilder.cs (GetTypes): Raise ReflectionTypeLoadException
	if any type was not finished.

	* AssemblyBuilderTest.cs: Add test for GetTypes () and incomplete
	TypeBuilders'.

	Fixes #640288

	Not only it didn't fix anything but it broke a test.

	Why on earth this wasn't found earlier? $20 reward to whoever gives a
	plausible explanation.

	* threads.c (mono_thread_suspend_all_other_threads): wait_data
	must not be malloc'd since it holds managed pointers. Alloc
	it on stack since the struct is about 512/1024 bytes long.

	* threads.c (mono_thread_manage): Ditto.

	* threads.c (mono_threads_request_thread_dump): Ditto.

MONO_PATH does not work for subsequent dlls, so have to workaround it by
using full path for dlls in 3.5 profile which does not have it's own
mscorlib

This error was found by Gonzalo when he was looking at another problem
when thread IDs overflowed ints.

Currently we use gpointers in the io-layer for the threadid, but we were
only exposing the value as an int32.   The good news is that the code is
not really used right now, but could cause problems in the future as we
extend Process.

	* Type.cs (GetTypeCodeImpl): If the type is a TypeBuilder
	call into specific code.

	* TypeBuilder.cs (GetTypeCodeInternal): New method that
	handles enums correctly.

	This change is required because calling into the runtime
	with incomplete TypeBuilders cause the runtime to crash.

The embedding code looks at the executable name, and builds the embedding profile
from that, however the default install has a /usr/bin/mono symlink pointing to
/Library/Frameworks/Mono.framework.  We now resolve the symlink before allowing
the embedding code to proceed.

	Ditto.
	This should fix MD problems when SHM is disabled.