Re: [PATCH 2/6] list: add new MACROs to make iterator invisiable outside the loop

From: Linus Torvalds
Date: Sun Mar 06 2022 - 13:57:47 EST


On Sun, Mar 6, 2022 at 4:19 AM Jakob Koschel <jakobkoschel@xxxxxxxxx> wrote:
>
> I guess we could apply this to list_for_each_entry() as well
> once all the uses after the loop are fixed?

I think that would be a good longer-term plan. "list_traverse()" ends
up being simpler syntactically, and has a certain level of inherent
type safety (not just the "don't expose the mis-typed head pointer
after the loop").

> I feel like this simply introduces a new set of macros
> (we would also need list_traverse_reverse(), list_traverse_continue_reverse()
> etc) and end up with a second set of macros that do pretty much
> the same as the first one.

I think that if we're happy with this, we can probably do a scripted
conversion. But I do like how it's incremental, in that we wouldn't
necessarily have to do it all in one go.

Because it's always really painful with flag-day interface changes,
which it would be to actually change the semantics of
"list_for_each_entry()" without a name change. It just makes for a lot
of pain for things that aren't in-tree yet (not just drivers that are
out-of-tree in general, but drivers in development etc).

And I really disliked the "pass the type to the list_for_each()"
macro, because of how it made the end result look more complex.

But list_traverse() looks like it would make the end result better
both from a user perspective (ie the code just looks simpler) but also
from the type safety point.

> Personally I guess I also prefer the name list_for_each_entry() over list_traverse()
> and not having two types of iterators for the same thing at the same time.

I absolutely agree with you in theory, and in many ways I like
list_for_each_entry() better as a name too (probably just because I'm
used to it).

But keeping the same name and changing how it works ends up being such
a "everything at once" thing that I don't think it's realistic.

Linus