POSIX compatibility layer does not work on 64-bit ports

[fdesbiens]

The POSIX compatibility layer cannot work on a 64-bit port. It compiles there and produces a library, but its message queues corrupt the pointer they carry, and the configuration that would fix that does not build. This came out of wiring the layer into CMake in #626.

Why it is broken as it stands

The layer passes a message by allocating a private buffer, copying the caller's data into it and putting the buffer's address into a ThreadX queue. ULONG is 32 bits on every ThreadX port, including the 64-bit ones, so on a 64-bit target that address does not fit in a message word.

tx_posix.h accounts for this with TX_64_BIT, which widens the message and splits the pointer across two words:

#ifdef TX_64_BIT
#define TX_POSIX_MESSAGE_SIZE           5
#else
#define TX_POSIX_MESSAGE_SIZE           4
#endif

with matching code in px_mq_send.c:198 and px_mq_receive.c:209. Without TX_64_BIT the send truncates the pointer and the receive casts the truncated value back. The compiler says nothing worse than a warning, which is why the target builds cleanly on AArch64 and RV64:

px_mq_send.c:205:14: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast]
px_mq_receive.c:214:27: warning: cast to pointer from integer of different size [-Wint-to-pointer-cast]

Why defining TX_64_BIT does not fix it

TX_64_BIT is never defined by any port; it is only ever tested. Defining it makes tx_api.h use the extension pointer macros, which need tx_thread_extension_ptr in the thread control block:

px_pth_create.c:235:5: error: 'TX_THREAD' has no member named 'tx_thread_extension_ptr'
px_pth_init.c:361:5:   error: 'TX_THREAD' has no member named 'tx_thread_extension_ptr'

Outside ports_smp and ports/linux, no port declares it. None of risc-v64, cortex_a53, cortex_a55 or cortex_a34 does.

What I measured

I prototyped the whole thing on risc-v64 to find out how much is actually missing.

The port change is one line. risc-v64 has an empty slot already:

#define TX_THREAD_EXTENSION_3                   VOID        *tx_thread_extension_ptr;

cortex_a53 and the other 64-bit Arm ports have a free slot too. This is the same declaration ports_smp/cortex_a72_smp already carries.

With that line and -DTX_64_BIT, the layer compiles clean. All 106 sources plus the #626 tests, riscv64-unknown-elf, -Werror, no warnings.

But it deadlocks at runtime. Built for risc-v64 and run under qemu-system-riscv64, the test produces no output at all. Attaching gdb shows the system parked in the idle loop with nothing runnable:

#0  _tx_thread_schedule_loop () at ports/risc-v64/gnu/src/tx_thread_schedule.S:77
#1  _tx_thread_system_suspend (thread_ptr=0x801012d0 <ptcb_pool>) at common/src/tx_thread_system_suspend.c:372

with these threads:

thread	state	meaning
System Timer Thread	3	suspended
test control thread	3	suspended
POSIX System Manager	5	waiting on its work queue (expected)
pthr	4	TX_SLEEP

The worker pthread is asleep although the test never sleeps, and nothing wakes it. The same BSP and QEMU invocation run the stock ThreadX RISC-V64 regression tests without trouble, so this is the layer rather than the port or the harness.

What is already correct

Worth recording, because it narrows the search. Every TX_64_BIT path I checked handles 64 bits properly:

• px_mq_send.c:198 / px_mq_receive.c:209 — the message pointer split.
• px_pth_init.c:720 / px_system_manager.c:87 — the work request pack and unpack.
• WORK_REQ_SIZE in tx_posix.h:257 — already scales by sizeof(ALIGN_TYPE)/sizeof(ULONG).

So the 64-bit paths were written plausibly; they appear never to have been executed. The remaining defect is somewhere less obvious, and there is likely more than one.

Suggested work

1. Add tx_thread_extension_ptr to the 64-bit ports. Trivial, but the generated ports need regenerating rather than hand-editing.
2. Debug the deadlock, and whatever it uncovers.
3. Re-enable the 64-bit tests. Added the POSIX compatibility layer to the CMake build, with regression tests #626 already handles both widths; test/posix/cmake/CMakeLists.txt currently stops with an explanation for risc-v64, and that becomes add_compile_options(-DTX_64_BIT).
4. Consider making the silent truncation loud, so a 64-bit build without TX_64_BIT fails instead of producing a library whose queues do not work.

Step 1 is an hour. Step 2 is the real cost: it is bring-up of code that has never run, so it is hard to bound. #626 gives somewhere to run it.