Re: [PATCH v9 02/15] xen: introduce generic non-atomic test_*bit()

[Jan Beulich <jbeulich@xxxxxxxx>]

On 16.05.2024 12:34, Oleksii K. wrote:
> On Thu, 2024-05-16 at 09:04 +0200, Jan Beulich wrote:
>> (Later) Wait, maybe I've finally figured it: You use
>> arch__test_and_*()
>> because those underlie __test_and_*(), but arch_test_bit() because
>> there's
>> solely test_bit() (same for the generic_* naming).
> Yes, that what I meant.
> 
>>  I guess I can accept
>> that then, despite the slight anomaly you point out, resulting in the
>> question towards 3 underscores in a row. To clarify, my thinking was
>> more
>> towards there not possibly being generic forms of test_and_*() (i.e.
>> no
>> possible set of arch_test_and_*() or generic_test_and_*()), thus
>> using
>> double inner underscores in arch__test_*() and generic__test_*() to
>> signify that those are purely internal functions, which aren't
>> supposed to
>> be called directly.
> I understand your point regarding functions that start with "__".
> For example, __test_and_clear_bit() is used not only internally (in
> terms of architecture code) but also in common code, so it is not
> strictly internal. I may have misunderstood what "internal function"
> means in this context.
> 
> I thought that, at least for bit operations, "__bit_operation" means
> that the bit operation is non-atomic and can be reordered, which
> implies that it's not a good idea to use it in common code without
> additional steps.

Correct, up to the comma; those may very well be used in common code,
provided non-atomic forms indeed suffice. But in my reply I didn't talk
about double-underscore-prefixes in names of involved functions. I
talked about inner double underscores.

> Anyway, I am not sure I understand which approach I should use in this
> patch. You mentioned that possibly test_and_() can't have a generic
> form, meaning it won't be a set of arch_test_and_() functions.
> 
> So, can I rename arch__test_() and generic__test_() to arch_test_() and
> generic_test_(), respectively, and use the renamed functions in
> _test_and*() in xen/bitops.h? Is my understanding correct?

You could. You could also stick to what you have now - as said, I can
accept that with the worked out explanation. Or you could switch to
using arch__test_bit() and generic__test_bit(), thus having the double
inner underscores identify "internal to the implementation" functions.
My preference would be in backwards order of what I have just enumerated
as possible options. I wonder whether really no-one else has any opinion
here ...

Jan