Re: [RFC][PATCH 3/3] sched: User Mode Concurency Groups

[Peter Zijlstra]

On Tue, Dec 14, 2021 at 09:44:48PM +0100, Peter Zijlstra wrote:

> The big assumption this whole thing is build on is that
> pin_user_pages() preserves user mappings in so far that
> pagefault_disable() will never generate EFAULT (unless the user does
> munmap() in which case it can keep the pieces).
> 
> shrink_page_list() does page_maybe_dma_pinned() before try_to_unmap()
> and as such seems to respect this constraint.
> 
> unmap_and_move() however seems willing to unmap otherwise pinned (and
> hence unmigratable) pages. This might need fixing.

AFAICT this should mostly do,.. I still need to check if
get_user_pages_fast() is itself sufficient to avoid all races or if we
need to strengthen/augment that too.

---
 mm/migrate.c  | 10 +++++++++-
 mm/mprotect.c |  6 ++++++
 2 files changed, 15 insertions(+), 1 deletion(-)

diff --git a/mm/migrate.c b/mm/migrate.c
index cf25b00f03c8..3850b33c64eb 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -1472,7 +1472,15 @@ int migrate_pages(struct list_head *from, new_page_t get_new_page,
 			nr_subpages = thp_nr_pages(page);
 			cond_resched();
 
-			if (PageHuge(page))
+			/*
+			 * If the page has a pin then expected_page_refs() will
+			 * not match and the whole migration will fail later
+			 * anyway, fail early and preserve the mappings.
+			 */
+			if (page_maybe_dma_pinned(page))
+				rc = -EAGAIN;
+
+			else if (PageHuge(page))
 				rc = unmap_and_move_huge_page(get_new_page,
 						put_new_page, private, page,
 						pass > 2, mode, reason,
diff --git a/mm/mprotect.c b/mm/mprotect.c
index 9a105fce5aeb..093db725d39f 100644
--- a/mm/mprotect.c
+++ b/mm/mprotect.c
@@ -105,6 +105,12 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 				if (page_is_file_lru(page) && PageDirty(page))
 					continue;
 
+				/*
+				 * Can't migrate pinned pages, avoid touching them.
+				 */
+				if (page_maybe_dma_pinned(page))
+					continue;
+
 				/*
 				 * Don't mess with PTEs if page is already on the node
 				 * a single-threaded process is running on.