[haiku-gsoc] Re: [Haiku ARM port] Debugging double spinlock acquiring in dprintf_args
- From: Arvind S Raj <sraj.arvind@xxxxxxxxx>
- To: haiku-gsoc <haiku-gsoc@xxxxxxxxxxxxx>
- Date: Tue, 12 Aug 2014 20:12:07 +0530
On 27 July 2014 21:51, Paweł Dziepak <pdziepak@xxxxxxxxxxx> wrote:
> Interrupts are not a problem here, we are using LL/SC instructions so the
> atomic functions are indeed always atomic (even with interrupts enabled and
> on multiprocessor systems).
>
> Currently, atomic_get_and_set() looks like this:
>
> miss4:
> ldrex r12, [r0]
> strex r3, r1, [r0]
> teq r3, #0
> bne miss4
> bx lr
>
> The content of r12 is not moved to r0 at the end of the function so the
> return value is wrong and the spinlock implementation wrongly assumes that
> the lock is already taken. There is obviously also a problem with memory
> ordering but that won't be noticeable on single processor system since this
> function call acts as compiler barrier.
> That's why I think we should implement atomics
> like this:
>
> int32_t atomic_get_and_set(int32_t* ptr, int32_t value)
> {
> auto& obj = *reinterpret_cast<std::atomic<int32_t>*>(ptr);
> return obj.exchange(value);
> }
>
> The produced assembly with -O2 for ARMv7 is very good:
>
> dmb sy
> .L5:
> ldrex r3, [r0]
> strex r2, r1, [r0]
> cmp r2, #0
> bne .L5
> dmb sy
> mov r0, r3
> bx lr
I've created a ticket(#11121) which has a patch that sets the return
value correctly. This makes the code largely similar to the
disassembly of the C++ function pdziepak wrote earlier in the thread
except that the DMB instructions aren't present. I suppose this would
suffice for now?
Other related posts:
