The below referenced commit makes the same error as 1c563432588d ("mm: fix
is_pinnable_page against a cma page"), re-interpreting the logic to exclude
pinning of the zero page, which breaks device assignment with vfio.
To avoid further subtle mistakes, split the logic into discrete tests.
Suggested-by: Matthew Wilcox <[email protected]>
Suggested-by: Felix Kuehling <[email protected]>
Link: https://lore.kernel.org/all/165490039431.944052.12458624139225785964.stgit@omen
Fixes: f25cbb7a95a2 ("mm: add zone device coherent type memory support")
Signed-off-by: Alex Williamson <[email protected]>
---
include/linux/mm.h | 17 ++++++++++++++---
1 file changed, 14 insertions(+), 3 deletions(-)
diff --git a/include/linux/mm.h b/include/linux/mm.h
index 18e01474cf6b..835106a9718f 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -1544,9 +1544,20 @@ static inline bool is_longterm_pinnable_page(struct page *page)
if (mt == MIGRATE_CMA || mt == MIGRATE_ISOLATE)
return false;
#endif
- return !(is_device_coherent_page(page) ||
- is_zone_movable_page(page) ||
- is_zero_pfn(page_to_pfn(page)));
+ /*
+ * The zero page might reside in a movable zone, however it may not
+ * be migrated and can therefore be pinned. The vfio subsystem pins
+ * user mappings including the zero page for IOMMU translation.
+ */
+ if (is_zero_pfn(page_to_pfn(page)))
+ return true;
+
+ /* Coherent device memory must always allow eviction. */
+ if (is_device_coherent_page(page))
+ return false;
+
+ /* Otherwise, non-movable zone pages can be pinned. */
+ return !is_zone_movable_page(page);
}
#else
static inline bool is_longterm_pinnable_page(struct page *page)
> Subject: [PATCH v2] mm: re-allow pinning of zero pfns (again)
>
> The below referenced commit makes the same error as 1c563432588d ("mm:
> fix is_pinnable_page against a cma page"), re-interpreting the logic to
> exclude pinning of the zero page, which breaks device assignment with vfio.
>
> To avoid further subtle mistakes, split the logic into discrete tests.
>
> Suggested-by: Matthew Wilcox <[email protected]>
> Suggested-by: Felix Kuehling <[email protected]>
> Link:
> https://lore.kernel.org/all/165490039431.944052.12458624139225785964.stgit
> @omen
> Fixes: f25cbb7a95a2 ("mm: add zone device coherent type memory
> support")
> Signed-off-by: Alex Williamson <[email protected]>
> ---
Tested-by: Slawomir Laba <[email protected]>
On 8/10/22 09:53, Alex Williamson wrote:
> The below referenced commit makes the same error as 1c563432588d ("mm: fix
> is_pinnable_page against a cma page"), re-interpreting the logic to exclude
> pinning of the zero page, which breaks device assignment with vfio.
>
> To avoid further subtle mistakes, split the logic into discrete tests.
>
> Suggested-by: Matthew Wilcox <[email protected]>
> Suggested-by: Felix Kuehling <[email protected]>
> Link: https://lore.kernel.org/all/165490039431.944052.12458624139225785964.stgit@omen
> Fixes: f25cbb7a95a2 ("mm: add zone device coherent type memory support")
> Signed-off-by: Alex Williamson <[email protected]>
> ---
> include/linux/mm.h | 17 ++++++++++++++---
> 1 file changed, 14 insertions(+), 3 deletions(-)
Hi Alex,
Looks good. I'm suggesting a simpler comment, below, because
even though the VFIO folks are thinking about VFIO, here we
are deep in the mm layer and there are lots of non-VFIO callers
that may pin the zero page.
>
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index 18e01474cf6b..835106a9718f 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -1544,9 +1544,20 @@ static inline bool is_longterm_pinnable_page(struct page *page)
> if (mt == MIGRATE_CMA || mt == MIGRATE_ISOLATE)
> return false;
> #endif
> - return !(is_device_coherent_page(page) ||
> - is_zone_movable_page(page) ||
> - is_zero_pfn(page_to_pfn(page)));
> + /*
> + * The zero page might reside in a movable zone, however it may not
> + * be migrated and can therefore be pinned. The vfio subsystem pins
> + * user mappings including the zero page for IOMMU translation.
> + */
Those notes are all about (some of) the callers. But it's a simple
answer, really, so how about just this:
/* The zero page is always allowed to be pinned. */
?
> + if (is_zero_pfn(page_to_pfn(page)))
> + return true;
> +
> + /* Coherent device memory must always allow eviction. */
> + if (is_device_coherent_page(page))
> + return false;
> +
> + /* Otherwise, non-movable zone pages can be pinned. */
> + return !is_zone_movable_page(page);
> }
> #else
> static inline bool is_longterm_pinnable_page(struct page *page)
>
>
>
Reviewed-by: John Hubbard <[email protected]>
thanks,
--
John Hubbard
NVIDIA
On Sat, 27 Aug 2022 17:59:32 -0700 John Hubbard <[email protected]> wrote:
>
> /* The zero page is always allowed to be pinned. */
Wow, that's really verbose :)
--- a/include/linux/mm.h~mm-re-allow-pinning-of-zero-pfns-again-fix
+++ a/include/linux/mm.h
@@ -1544,11 +1544,7 @@ static inline bool is_longterm_pinnable_
if (mt == MIGRATE_CMA || mt == MIGRATE_ISOLATE)
return false;
#endif
- /*
- * The zero page might reside in a movable zone, however it may not
- * be migrated and can therefore be pinned. The vfio subsystem pins
- * user mappings including the zero page for IOMMU translation.
- */
+ /* The zero page may always be pinned */
if (is_zero_pfn(page_to_pfn(page)))
return true;
_
On Sat, 27 Aug 2022 17:59:32 -0700
John Hubbard <[email protected]> wrote:
> On 8/10/22 09:53, Alex Williamson wrote:
> > The below referenced commit makes the same error as 1c563432588d ("mm: fix
> > is_pinnable_page against a cma page"), re-interpreting the logic to exclude
> > pinning of the zero page, which breaks device assignment with vfio.
> >
> > To avoid further subtle mistakes, split the logic into discrete tests.
> >
> > Suggested-by: Matthew Wilcox <[email protected]>
> > Suggested-by: Felix Kuehling <[email protected]>
> > Link: https://lore.kernel.org/all/165490039431.944052.12458624139225785964.stgit@omen
> > Fixes: f25cbb7a95a2 ("mm: add zone device coherent type memory support")
> > Signed-off-by: Alex Williamson <[email protected]>
> > ---
> > include/linux/mm.h | 17 ++++++++++++++---
> > 1 file changed, 14 insertions(+), 3 deletions(-)
> Hi Alex,
>
> Looks good. I'm suggesting a simpler comment, below, because
> even though the VFIO folks are thinking about VFIO, here we
> are deep in the mm layer and there are lots of non-VFIO callers
> that may pin the zero page.
>
> >
> > diff --git a/include/linux/mm.h b/include/linux/mm.h
> > index 18e01474cf6b..835106a9718f 100644
> > --- a/include/linux/mm.h
> > +++ b/include/linux/mm.h
> > @@ -1544,9 +1544,20 @@ static inline bool is_longterm_pinnable_page(struct page *page)
> > if (mt == MIGRATE_CMA || mt == MIGRATE_ISOLATE)
> > return false;
> > #endif
> > - return !(is_device_coherent_page(page) ||
> > - is_zone_movable_page(page) ||
> > - is_zero_pfn(page_to_pfn(page)));
> > + /*
> > + * The zero page might reside in a movable zone, however it may not
> > + * be migrated and can therefore be pinned. The vfio subsystem pins
> > + * user mappings including the zero page for IOMMU translation.
> > + */
>
> Those notes are all about (some of) the callers. But it's a simple
> answer, really, so how about just this:
>
> /* The zero page is always allowed to be pinned. */
Sure. Are we looking for a re-spin with this? I see Andrew already
added this incremental change to his hotfix-unstable branch separately.
I'd hate for a comment re-spin to delay getting a fix for this problem,
that blocks any VM use cases of VFIO, into mainline any longer. Thanks,
Alex
On 8/28/22 05:37, Alex Williamson wrote:
>> Those notes are all about (some of) the callers. But it's a simple
>> answer, really, so how about just this:
>>
>> /* The zero page is always allowed to be pinned. */
>
> Sure. Are we looking for a re-spin with this? I see Andrew already
> added this incremental change to his hotfix-unstable branch separately.
> I'd hate for a comment re-spin to delay getting a fix for this problem,
> that blocks any VM use cases of VFIO, into mainline any longer. Thanks,
>
Definitely not. Andrews fixup should suffice.
thanks,
--
John Hubbard
NVIDIA