Re: mm: delay rmap removal until after TLB flush

[Linus Torvalds]

On Thu, Nov 3, 2022 at 9:54 AM Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> But again, those changes would have made the patch bigger, which I
> didn't want at this point (and 'release_pages()' would need that
> clean-in-place anyway, unless we changed *that* too and made the whole
> page encoding be something widely available).

And just to clarify: this is not just me trying to expand the reach of my patch.

I'd suggest people look at mlock_pagevec(), and realize that LRU_PAGE
and NEW_PAGE are both *exactly* the same kind of "encoded_page" bits
that TLB_ZAP_RMAP is.

Except the mlock code does *not* show that in the type system, and
instead just passes a "struct page **" array around in pvec->pages,
and then you'd just better know that "oh, it's not *really* just a
page pointer".

So I really think that the "array of encoded page pointers" thing is a
generic notion that we *already* have.

It's just that we've done it disgustingly in the past, and I didn't
want to do that disgusting thing again.

So I would hope that the nasty things that the mlock code would some
day use the same page pointer encoding logic to actually make the
whole "this is not a page pointer that you can use directly, it has
low bits set for flags" very explicit.

I am *not* sure if then the actual encoded bits would be unified.
Probably not - you might have very different and distinct uses of the
encode_page() thing where the bits mean different things in different
contexts.

Anyway, this is me just explaining the thinking behind it all. The
page bit encoding is a very generic thing (well, "very generic" in
this case means "has at least one other independent user"), explaining
the very generic naming.

But at the same time, the particular _patch_ was meant to be very targeted.

So slightly schizophrenic name choices as a result.

             Linus