spinlock alignement issue in aarch64 causing so3 to crash

[AndreCostaaa]

Encountered this bug when using the virt64 config where so3 would suddenly crash when starting a new process.

After some digging, I found that the crash happens when we try to lock a mutex. Basically, whenever the new process would call printf, that causes a mutex lock which in turn tries to lock a spinlock that is misaligned :

In user space:

printf -> vprintf -> FLOCK -> __lockfile -> mutex_lock -> sys_mutex_lock(1)

When we get to kernel space:

do_mutex_lock -> mutex_lock -> spin_lock_irqsave -> spin_trylock

And then so3 would crash:

Taking a look at what ESR 0x96000061 :

We see that the problem is with the alignment of a memory write which must mean the call "stxr %w0, %2, [%1]\n" here:

so3/so3/arch/arm64/include/asm/spinlock.h

Lines 36 to 47 in 9fd2100

	static inline int spin_trylock(spinlock_t *lock)
	{
	uint32_t tmp;
	__asm__ __volatile__(" ldaxr %w0, [%1]\n"
	" tbnz %w0, #0, 1f\n"
	" stxr %w0, %2, [%1]\n"
	"1:"
	: "=&r"(tmp)
	: "r"(&lock->lock), "r"(1)
	: "cc");

So after adding a test to check the memory alignment of the spinlock inside the mutex_init function i found this :

So basically, there's a memory misalignment with the new spinlocks created when creating a new process. I stopped investigating here and i'm still not sure why this is happening considering the fact that we're telling the compiler to align the spinlocks:

so3/so3/include/spinlock.h

Lines 24 to 29 in 9fd2100

	/*
	* On Aarch64, the field has to be 64-bit aligned apparently.
	*/
	typedef struct {
	__attribute__((aligned(8))) volatile uint32_t lock;
	} spinlock_t;