Enclave debugging with gdb

[Niederb]

Trying to debug the enclave I get the same error as described here: intel/linux-sgx#533
According to the answer we would have to install these depencencies in order for debug to work:

• libsgx-enclave-common-dbgsym
• libsgx-urts-dbgsym

[haerdib]

Installed on devsgx02. Does it work now?

[Niederb]

That was quick 😄
It certainly helped. There is now more stack information, but it seems the part I'm interested in is still hidden in '__morestack()'. I will do some more googling...

#3  0x00007fa97fdac153 in CEnclave::ocall(unsigned int, _ocall_table_t const*, void*) () from /usr/lib/x86_64-linux-gnu/libsgx_urts.so
#4  0x00007fa97fda926d in stack_sticker () from /usr/lib/x86_64-linux-gnu/libsgx_urts.so
#5  0x00007fa97fda91a6 in __morestack () from /usr/lib/x86_64-linux-gnu/libsgx_urts.so
#6  0x00007fa97fdac47b in CEnclave::ecall(int, void const*, void*, bool) () from /usr/lib/x86_64-linux-gnu/libsgx_urts.so
#7  0x00007fa97fdae542 in _sgx_ecall(unsigned long, int, void const*, void*, bool) () from /usr/lib/x86_64-linux-gnu/libsgx_urts.so

[Niederb]

• Attaching to a process with gdb is now possible
  • I still cannot debug the enclave
• It is possible with the "remote development" feature of CLion to debug the CLI process of the worker
• I started documenting my findings here:
  • https://tools.scs.ch/wiki/display/SCSBLOCKCHAIN/Debugging+tips+and+tricks

[Niederb]

I work now in a VM:

• I managed to get Intel VTune working for profiling (still not in the enclave though).
• Also I managed to attach to a worker process for debugging. I had to enable the debug symbols in release mode for that to work:

[profile.release]
debug = true

[Niederb]

Ok, I got the debugger to work now. Here are the steps:

• We need to produce debug symbols in rust even in release builds
• The sgx-dbg debugger has some magic implemented that automatically loads the symbols (with add-symbol-file) out of Enclave.signed.so. This did not work in our setup. Therefore I had to add the linker flag -lsgx_urts_sim.
  • https://github.com/intel/linux-sgx/blob/fdb3b8b4695b40afab5a3212cf69eb8d8db7cfb1/sdk/debugger_interface/linux/gdb-sgx-plugin/load_symbol_cmd.py#L92
  • https://github.com/intel/linux-sgx/blob/fdb3b8b4695b40afab5a3212cf69eb8d8db7cfb1/sdk/debugger_interface/linux/gdb-sgx-plugin/gdb_sgx_plugin.py#L719
  • Also check the makefile from a teaclave example. They also have this flag: https://github.com/apache/incubator-teaclave-sgx-sdk/blob/master/samplecode/helloworld/Makefile
• When enabling the debug symbols we get a compilation error '__assert_fail'. Therefore I added a dummy implementation for now. See these related issues:
  • 1.1.4 enclave linker error undefined reference to `__assert_fail' without release apache/incubator-teaclave-sgx-sdk#373
  • ring fail if debug info is enabled apache/incubator-teaclave-sgx-sdk#44

[Niederb]

• I think the next step would be to reproduce my setup on another machine
  • I tried a lot of different things so it is possible that some needed steps are missing
  • See Test remote debugging with CLion #856
• Also see issue Debug build with Optimizations #883
• The branch should not be merged
  • Then we can think about how to bring the necessary changes onto master
  • The changes are small anyway so we can also just re-create them if merging becomes complicated