Support the clone system call.

[davidchisnall]

This implements two new LKL hooks. The first one to create an lthread
with a specific initial register state (to capture the returns-twice
behaviour of clone, along with the caller's ability to define the stack
and TLS addresses). The new thread is immediately associated with the
Linux task structure (normally, lthreads are associated with Linux tasks
lazily when they perform a system call).

The second hook destroys a thread. This is done in response to an exit
system call. This is somewhat complicated, because LKL never returns to
this thread and the thread's stack may be deallocated by the time we
exit it.

The lthread scheduler does not have an easy way of adding a mechanism to
kill a thread without that thread running. We can add one eventually,
but for now create a temporary stack that lthreads can use during
teardown and make them run the teardown from there.

Fixes #155

[davidchisnall]

Don't merge this yet: the LKL commit will change once lsds/lkl#1 is merged. I'll force-push to this branch to update the LKL submodule once that's done.

[davidchisnall]

Note that this doesn't yet fix #154.

[prp]

LGTM, some minor comments. I see CI failures though...

[davidchisnall]

CI is failing because the weak symbol for lkl_syscall isn't found. It does for me, so I'm not entirely sure what's going on here...

[SeanTAllen]

out of curiosity, given that 3.6 is rather dated. is there a "no later than X" dependency for sgx-lkl that I'm not aware of?

[prp]

We should use a consistent Alpine version in all tests. @letmaik?

[letmaik]

I opened #260. For this PR, it's fine to leave it at an arbitrary version. The maximum supported currently is 3.10.

[SeanTAllen]

I started looking into the failing access02 ltp test. Here's how to replicate.

• Start from fresh repo of the clone branch.
• make DEBUG=true
• cd tests/ltp/ltp-batch1
• edit ../batch.mk
• comment out the commands for run-hw and run-sw commands (those commands run all tests)

run-hw: $(ROOT_FS)
        #@${LTP_TEST_SCRIPT} run-hw

run-sw: $(ROOT_FS)
        #@${LTP_TEST_SCRIPT} run-sw

• make clean
• make DEBUG=true
• make run

At this point, you are prepped to run the single test for HW mode:

• SGXLKL_VERBOSE=1 SGXLKL_KERNEL_VERBOSE=0 ../../../build/sgx-lkl-run-oe --hw-debug sgxlkl-miniroot-fs.img /ltp/testcases/kernel/syscalls/access/access02

for SW mode:

SGXLKL_VERBOSE=1 SGXLKL_KERNEL_VERBOSE=0 ../../../build/sgx-lkl-run-oe --sw-debug sgxlkl-miniroot-fs.img /ltp/testcases/kernel/syscalls/access/access02

Result I get

HW:

tst_test.c:1106: INFO: Timeout per run is 0h 05m 00s
tst_test.c:1125: INFO: No fork support
access02.c:144: PASS: access(file_f, F_OK) as root behaviour is correct.
access02.c:144: PASS: access(file_f, F_OK) as nobody behaviour is correct.
access02.c:144: PASS: access(file_r, R_OK) as root behaviour is correct.
access02.c:144: PASS: access(file_r, R_OK) as nobody behaviour is correct.
access02.c:144: PASS: access(file_w, W_OK) as root behaviour is correct.
access02.c:144: PASS: access(file_w, W_OK) as nobody behaviour is correct.
bad count while changing owner
[[ SGX-LKL ]] FAIL: Kernel panic! Run DEBUG build with SGXLKL_KERNEL_VERBOSE=1 for more information. Aborting...
2020-05-19T19:52:36.000000Z [(H)ERROR] tid(0x7fe8a8ff1700) | :OE_ENCLAVE_ABORTING [/home/sean/openenclave-sgxlkl.git/host/calls.c:oe_call_enclave_function_by_table_id:91]
[ SGX-LKL ] ethread (4: 19) [ SGX-LKL ] FAIL: sgxlkl_ethread_init() failed (id=4 result=19 (OE_ENCLAVE_ABORTING))

SW:

It hangs after printing access02.c:144: PASS: access(file_w, W_OK) as nobody behaviour is correct.

---

This is somewhat different than what I am seeing in CI, so, before I proceed any further, @davidchisnall can you try recreating the above?

---

The source of the failing test is https://github.com/linux-test-project/ltp/blob/master/testcases/kernel/syscalls/access/access02.c

[davidchisnall]

Thanks @SeanTAllen. I actually see something more interesting when I try to reproduce this locally:

tst_test.c:1106: INFO: Timeout per run is 0h 05m 00s
tst_test.c:1125: INFO: No fork support
access02.c:144: PASS: access(file_f, F_OK) as root behaviour is correct.
access02.c:144: PASS: access(file_f, F_OK) as nobody behaviour is correct.
access02.c:144: PASS: access(file_r, R_OK) as root behaviour is correct.
access02.c:144: PASS: access(file_r, R_OK) as nobody behaviour is correct.
access02.c:144: PASS: access(file_w, W_OK) as root behaviour is correct.
access02.c:144: PASS: access(file_w, W_OK) as nobody behaviour is correct.
Created new host task 7f2fc02f4540 (for 7f2fbeada4c0)

So it appears that we're hitting the code path for creating a new clone'd task (for a test that shouldn't be calling clone). I'll take a look.