在 2024/9/24 7:52, Andi Kleen 写道:

quoted

Thanks for the suggestions, I will experiment with a read-write lock, meanwhile,
adding the documentation and testing for the lockless scheme.

Read-write locks are usually not worth it for short critical sections,
in fact they can be slower due to cache line effects.

OK, I will start from a simple spinlock.

quoted

Sorry, I may not probably get the point clear here, and it would be very
nice if more details are provided for the concern. Do you mean it's necessary
to make the if-body excution exclusive among the CPUs? If that's the case,
I guess the test_and_put_task_slot() is the equvialent to the race condition
check. test_and_put_task_slot() uses a compare and exchange operation on the
slot_ref of utask instance. Regardless of the work type being performed by
other CPU, it will always bail out unless the slot_ref has a value of one,
indicating the utask is free to access from local CPU.

What I meant is that the typical pattern for handling races in destruction
is to detect someone else is racing and then let it do the destruction
work or reacquire the resource (so just bail out).

Agreed.

But that's not what you're doing here, in fact you're adding a
completely new code path that has different semantics? I haven't checked
all the code, but it looks dubious.

Andi, I've just sent v2. Looking forward to your feedback. Thanks.

  https://lore.kernel.org/all/20240927094549.3382916-1-liaochang1@huawei.com/ (local)

-Andi

-- 
BR
Liao, Chang