On 31.01.25 18:05, Simona Vetter wrote:
On Fri, Jan 31, 2025 at 11:55:55AM +0100, David Hildenbrand wrote:
On 31.01.25 00:06, Alistair Popple wrote:
On Thu, Jan 30, 2025 at 02:03:42PM +0100, Simona Vetter wrote:
On Thu, Jan 30, 2025 at 10:58:51AM +0100, David Hildenbrand wrote:
On 30.01.25 10:51, Simona Vetter wrote:
On Wed, Jan 29, 2025 at 12:54:03PM +0100, David Hildenbrand wrote:
Let's do it just like mprotect write-upgrade or during NUMA-hinting
faults on PROT_NONE PTEs: detect if the PTE can be writable by using
can_change_pte_writable().
Set the PTE only dirty if the folio is dirty: we might not
necessarily have a write access, and setting the PTE writable doesn't
require setting the PTE dirty.
Not sure whether there's much difference in practice, since a device
exclusive access means a write, so the folio better be dirty (unless we
aborted halfway through). But then I couldn't find the code in nouveau to
do that, so now I'm confused.
That confused me as well. Requiring the PTE to be writable does not imply
that it is dirty.
So something must either set the PTE or the folio dirty.
Yeah I'm not finding that something.
( In practice, most anonymous folios are dirty most of the time ... )
And yup that's why I think it hasn't blown up yet.
If we assume that "device-exclusive entries" are always dirty, then it
doesn't make sense to set the folio dirty when creating device-exclusive
entries. We'd always have to set the PTE dirty when restoring the exclusive
pte.
I do agree with your change, I think it's correct to put this
responsibility onto drivers. It's just that nouveau seems to not be
entirely correct.
Yeah, agree it should be a driver responsibility but also can't see how nouveau
is correct there either. I might see if I can get it to blow up...
(in context of the rmap walkers) The question is, how do we consider
device-exclusive entries:
(1) dirty? Not from a CPU perspective.
(2) referenced? Not from a CPU perspective.
If the answer is always "no" to all questions, then memory notifiers must
handle it, because we'd be answering the question from the CPU point of
view.
If the answer is always "yes", there is a problem: we can only make it
clean/young by converting it to an ordinary PTE first (requiring MMU
notifiers etc.), which makes it quite nasty.
Mixed answers are not possible, because we don't know just from staring at
the entry.
I think it's the gpu's (or whatever is using it) responsibility to update
folio state while it has ptes pointing at memory. Whether that's
device-exclusive system memory or device-private migrated memory. Anything
else doesn't make sense to me conceptually.
And I don't think we can just blindly assume even for device-exclusive
mappings that they will be dirty when we convert them back to a real pte,
because we might have raced trying to set up the gpu mapping and restarted
before we even put the pte into place. Or maybe someone was real quick at
writing it back after the gpu already dropped it's pte.
I guess maybe some clear documentation in all these functions
(make_device_exclusive, hmm_range_fault, migration helpers) that it's the
drivers job to dirty pages correctly would help?
I'll add a comment to make_device_exclusive(), stating that these
entries are considered clean+old from a MM perspective, and that the
driver must update the folio when notified by MMU notifiers.
We should probably document that somewhere else as well as you suggested
separately.
--
Cheers,
David / dhildenb
