Re: [PATCH v4 1/3] mm: Don't pin ZERO_PAGE in pin_user_pages()

[David Hildenbrand <david@redhat.com>]

On 31.05.23 15:55, David Howells wrote:
> David Hildenbrand <david@redhat.com> wrote:
> 
>> Yes, it would be clearer if we would be using "pinned" now only for FOLL_PIN
> 
> You're not likely to get that.  "To pin" is too useful a verb that gets used
> in other contexts too.  For that reason, I think FOLL_PIN was a poor choice of
> name:-/.  I guess the English language has got somewhat overloaded.  Maybe
> FOLL_PEG? ;-)

You're probably right. FOLL_PIN and all around that is "get me an 
additional reference on the page and make sure I can DMA it without any 
undesired side-effects".

FOLL_PIN_DMA would have been clearer (and matches 
folio_maybe_dma_pinned() ) ... but then, there are some use cases where 
want the same semantics but not actually perform DMA, but simply 
read/write via the directmap (e.g., vmsplice, some io_uring cases). 
Sure, one could say that they behave like DMA: access page content at 
any time.

Saying a page is pinned (additional refcount) and having a pincount of 0 
or does indeed cause confusion.

... but once we start renaming FOLL_PIN, pincount, ... we also have to 
rename pin_user_pages() and friends, and things get nasty.

> 
>> and everything else is simply "taking a temporary reference on the page".
> 
> Excluding refs taken with pins, many refs are more permanent than pins as, so
> far as I'm aware, pins only last for the duration of an I/O operation.

I was more thinking along the lines of FOLL_GET vs. FOLL_PIN. Once we 
consider any references we might have on a page, things get more tricky 
indeed.

> 
>>>> "Note that the refcount of any zero_pages returned among the pinned pages will
>>>> not be incremented, and unpin_user_page() will similarly not decrement it."
>>> That's not really right (although it happens to be true), because we're
>>> talking primarily about the pin counter, not the refcount - and they may be
>>> separate.
>>
>> In any case (FOLL_PIN/FOLL_GET) you increment/decrement the refcount. If we
>> have a separate pincount, we increment/decrement the refcount by 1 when
>> (un)pinning.
> 
> FOLL_GET isn't relevant here - only FOLL_PIN.  Yes, as it happens, we count a
> ref if we count a pin, but that's kind of irrelevant; what matters is that the
> effect must be undone with un-PUP.

The point I was trying to make is that we always modify the refcount, 
and in some cases (FOLL_PIN on order > 0) also the pincount.

But if you define "pins" as "additional reference", we're on the same page.

> 
> It would be nice not to get a ref on the zero page in FOLL_GET, but I don't
> think we can do that yet.  Too many places assume that GUP will give them a
> ref they can release later via ordinary methods.

No we can't I'm afraid.

-- 
Thanks,

David / dhildenb