When FOLL_PIN is passed to __get_user_pages() the page list must be put back using unpin_user_pages() otherwise the page pin reference persists in a corrupted state. There are two places in the unwind of __gup_longterm_locked() that put the pages back without checking. Normally on error this function would return the partial page list making this the caller's responsibility, but in these two cases the caller is not allowed to see these pages at all. Cc: <stable@xxxxxxxxxx> Cc: Aneesh Kumar K.V <aneesh.kumar@xxxxxxxxxxxxx> Fixes: 3faa52c03f44 ("mm/gup: track FOLL_PIN pages") Reported-by: Ira Weiny <ira.weiny@xxxxxxxxx> Signed-off-by: Jason Gunthorpe <jgg@xxxxxxxxxx> --- mm/gup.c | 14 ++++++++++---- 1 file changed, 10 insertions(+), 4 deletions(-) v2: - Catch the DAX related case as well (Ira) v1: https://lore.kernel.org/r/0-v1-976effcd4468+d4-gup_cma_fix_jgg@xxxxxxxxxx Andrew, this version with a modified commit message and extra hunk replaces: mm-gup-use-unpin_user_pages-in-check_and_migrate_cma_pages.patch Thanks, Jason diff --git a/mm/gup.c b/mm/gup.c index 102877ed77a4b4..98eb8e6d2609c3 100644 --- a/mm/gup.c +++ b/mm/gup.c @@ -1647,8 +1647,11 @@ static long check_and_migrate_cma_pages(struct mm_struct *mm, /* * drop the above get_user_pages reference. */ - for (i = 0; i < nr_pages; i++) - put_page(pages[i]); + if (gup_flags & FOLL_PIN) + unpin_user_pages(pages, nr_pages); + else + for (i = 0; i < nr_pages; i++) + put_page(pages[i]); if (migrate_pages(&cma_page_list, alloc_migration_target, NULL, (unsigned long)&mtc, MIGRATE_SYNC, MR_CONTIG_RANGE)) { @@ -1728,8 +1731,11 @@ static long __gup_longterm_locked(struct mm_struct *mm, goto out; if (check_dax_vmas(vmas_tmp, rc)) { - for (i = 0; i < rc; i++) - put_page(pages[i]); + if (gup_flags & FOLL_PIN) + unpin_user_pages(pages, rc); + else + for (i = 0; i < rc; i++) + put_page(pages[i]); rc = -EOPNOTSUPP; goto out; } -- 2.28.0