Nouveau - Mar 2021 - [PATCH v3 5/8] mm: Device exclusive memory access

On Tuesday, 2 March 2021 11:05:59 AM AEDT Jason Gunthorpe
wrote:
> On Fri, Feb 26, 2021 at 06:18:29PM +1100, Alistair Popple wrote:
> 
> > +/**
> > + * make_device_exclusive_range() - Mark a range for exclusive use by
a
device
> > + * @mm: mm_struct of assoicated target process
> > + * @start: start of the region to mark for exclusive device access
> > + * @end: end address of region
> > + * @pages: returns the pages which were successfully mark for
exclusive
acces
> > + *
> > + * Returns: number of pages successfully marked for exclusive access
> > + *
> > + * This function finds the ptes mapping page(s) to the given address 
range and
> > + * replaces them with special swap entries preventing userspace CPU 
access. On
> > + * fault these entries are replaced with the original mapping after 
calling MMU
> > + * notifiers.
> > + */
> > +int make_device_exclusive_range(struct mm_struct *mm, unsigned long 
start,
> > +				unsigned long end, struct page **pages)
> > +{
> > +	long npages = (end - start) >> PAGE_SHIFT;
> > +	long i;
> > +
> > +	npages = get_user_pages_remote(mm, start, npages,
> > +				       FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
> > +				       pages, NULL, NULL);
> > +	for (i = 0; i < npages; i++) {
> > +		if (!trylock_page(pages[i])) {
> > +			put_page(pages[i]);
> > +			pages[i] = NULL;
> > +			continue;
> > +		}
> > +
> > +		if (!try_to_protect(pages[i])) {
> 
> Isn't this racy? get_user_pages returns the ptes at an instant in
> time, they could have already been changed to something else?
Right. On it's own this does not guarantee that the page is mapped at the 
given location, only that a mapping won't get established without an mmu 
notifier callback to clear the swap entry.

The intent was a driver could use HMM or some other mechanism to keep PTEs 
synchronised if required. However I just looked at patch 8 in the series again 
and it appears I got this wrong when converting from the old migration 
approach:

+               mutex_unlock(&svmm->mutex);
+               ret = nouveau_atomic_range_fault(svmm, drm, args,
+                                               size, hmm_flags, mm);

The mutex needs to be unlocked after the range fault to ensure the PTE
hasn't
changed. But this ends up being a problem because try_to_protect() calls 
notifiers which need to take that mutex and hence deadlocks.
> I would think you'd want to switch to the swap entry atomically under
> th PTLs?
That is one approach, but the reuse of get_user_pages() to walk the page 
tables and fault/gather the pages is a nice simplification and adding a new 
FOLL flag/mode to atomically swap entries doesn't seem right.

However try_to_protect() scans the PTEs again under the PTL so checking the 
mapping of interest actually gets replaced during the rmap walk seems like a 
reasonable solution. Thanks for the comments.

 - Alistair
> Jason
>

On Tue, Mar 02, 2021 at 07:57:58PM +1100, Alistair Popple wrote:
> The intent was a driver could use HMM or some other mechanism to keep PTEs 
> synchronised if required. However I just looked at patch 8 in the series
again
> and it appears I got this wrong when converting from the old migration 
> approach:
> 
> +               mutex_unlock(&svmm->mutex);
> +               ret = nouveau_atomic_range_fault(svmm, drm, args,
> +                                               size, hmm_flags, mm);
> 
> The mutex needs to be unlocked after the range fault to ensure the PTE
hasn't
> changed. But this ends up being a problem because try_to_protect() calls 
> notifiers which need to take that mutex and hence deadlocks.
you have to check the notifier sequence under the mutex and loop
again. The mutex should only cover programming the HW to use the
pages, nothing else.
> However try_to_protect() scans the PTEs again under the PTL so checking the
> mapping of interest actually gets replaced during the rmap walk seems like
a
> reasonable solution. Thanks for the comments.
It does seem cleaner if you can manage it, the notifier will still be
needd to program the HW though

Jason