On 06.09.22 15:47, Jason Gunthorpe wrote:
On Mon, Sep 05, 2022 at 09:59:47AM +0200, David Hildenbrand wrote:
That should be READ_ONCE() for the *pmdp and *ptep reads. Because this
whole lockless house of cards may fall apart if we try reading the
page table values without READ_ONCE().
I came to the conclusion that the implicit memory barrier when grabbing a
reference on the page is sufficient such that we don't need READ_ONCE here.
READ_ONCE is not about barriers or ordering, you still need the
acquire inside the atomic to make the algorithm work.
While I don't disagree with what say is, I'll refer to
Documentation/memory-barriers.txt "COMPILER BARRIER".
As discussed somewhere in this thread, if we already have an atomic RWM
that implies a full ordering, it implies a compile barrier.
READ_ONCE primarily is a marker that the data being read is unstable
and that the compiler must avoid all instability when reading it. eg
in this case the compiler could insanely double read the value, even
though the 'if' requires only a single read. This would result in
corrupt calculation.
As we have a full memory barrier + compile barrier, the compiler might
indeed do double reads and all that stuff. BUT, it has to re-read after
we incremented the refcount, and IMHO that's the important part to
detect the change.
--
Thanks,
David / dhildenb
