Description of the problem

When the application creates a socket pair and calls send() from more than one thread simultaneously, sometimes recv() at the other end of the pipe will receive invalid data and returns -13 (permission denied).

Tracing the problem leads to:

https://github.com/oscarlab/graphene/blob/345d271e663725d58ce941fc49f08ebc02c32fa2/Pal/lib/crypto/adapters/mbedtls_adapter.c#L504-L509

Here mbedtls_ssl_write() is called without any lock protection. Because writing to the same mbedTLS context is not thread-safe, it triggers a race condition inside mbedTLS and causes encrypted data to be sent prematurely.

Adding a spinlock (perhaps should be a futex?) around mbedtls_ssl_write() makes the problem go away.

This is found when testing Python asyncio, which internally uses one-byte write to signal the waiting thread. Errors like this appear in the log, which is caused by this same problem:

Exception in callback BaseSelectorEventLoop._read_from_self()
handle: <Handle BaseSelectorEventLoop._read_from_self()>
Traceback (most recent call last):
  File "/usr/lib/python3.7/asyncio/events.py", line 88, in _run
    self._context.run(self._callback, *self._args)
  File "/usr/lib/python3.7/asyncio/selector_events.py", line 119, in _read_from_self
    data = self._ssock.recv(4096)
PermissionError: [Errno 13] Permission denied

Steps to reproduce

Run the C program attached, with and without Graphene SGX

test.txt

Expected results

When running without Graphene, the program will report:

c1=100000, c2=100000, e13=0

Actual results

Running with Graphene SGX, the result shows a lot of error -13 and few successful recv():

c1=186, c2=64, e13=199750