mm/hmm: properly handle migration pmd

Before this patch migration pmd entry (!pmd_present()) would have been
treated as a bad entry (pmd_bad() returns true on migration pmd entry).
The outcome was that device driver would believe that the range covered by
the pmd was bad and would either SIGBUS or simply kill all the device's
threads (each device driver decide how to react when the device tries to
access poisonnous or invalid range of memory).

This patch explicitly handle the case of migration pmd entry which are non
present pmd entry and either wait for the migration to finish or report
empty range (when device is just trying to pre- fill a range of virtual
address and thus do not want to wait or trigger page fault).

Link: http://lkml.kernel.org/r/20181019160442.18723-5-jglisse@redhat.com
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Jérôme Glisse <jglisse@redhat.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Ralph Campbell <rcampbell@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 34 insertions, 6 deletions

diff --git a/mm/hmm.c b/mm/hmm.c
index 63ca3b5be306..a3d532ff7c3d 100644
--- a/mm/hmm.c
+++ b/mm/hmm.c
@@ -577,22 +577,42 @@ static int hmm_vma_walk_pmd(pmd_t *pmdp,
 {
 	struct hmm_vma_walk *hmm_vma_walk = walk->private;
 	struct hmm_range *range = hmm_vma_walk->range;
+	struct vm_area_struct *vma = walk->vma;
 	uint64_t *pfns = range->pfns;
 	unsigned long addr = start, i;
 	pte_t *ptep;
+	pmd_t pmd;
 
-	i = (addr - range->start) >> PAGE_SHIFT;
 
 again:
-	if (pmd_none(*pmdp))
+	pmd = READ_ONCE(*pmdp);
+	if (pmd_none(pmd))
 		return hmm_vma_walk_hole(start, end, walk);
 
-	if (pmd_huge(*pmdp) && (range->vma->vm_flags & VM_HUGETLB))
+	if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB))
 		return hmm_pfns_bad(start, end, walk);
 
-	if (pmd_devmap(*pmdp) || pmd_trans_huge(*pmdp)) {
-		pmd_t pmd;
+	if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
+		bool fault, write_fault;
+		unsigned long npages;
+		uint64_t *pfns;
+
+		i = (addr - range->start) >> PAGE_SHIFT;
+		npages = (end - addr) >> PAGE_SHIFT;
+		pfns = &range->pfns[i];
+
+		hmm_range_need_fault(hmm_vma_walk, pfns, npages,
+				     0, &fault, &write_fault);
+		if (fault || write_fault) {
+			hmm_vma_walk->last = addr;
+			pmd_migration_entry_wait(vma->vm_mm, pmdp);
+			return -EAGAIN;
+		}
+		return 0;
+	} else if (!pmd_present(pmd))
+		return hmm_pfns_bad(start, end, walk);
 
+	if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
 		/*
 		 * No need to take pmd_lock here, even if some other threads
 		 * is splitting the huge pmd we will get that event through
@@ -607,13 +627,21 @@ again:
 		if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 			goto again;
 
+		i = (addr - range->start) >> PAGE_SHIFT;
 		return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd);
 	}
 
-	if (pmd_bad(*pmdp))
+	/*
+	 * We have handled all the valid case above ie either none, migration,
+	 * huge or transparent huge. At this point either it is a valid pmd
+	 * entry pointing to pte directory or it is a bad pmd that will not
+	 * recover.
+	 */
+	if (pmd_bad(pmd))
 		return hmm_pfns_bad(start, end, walk);
 
 	ptep = pte_offset_map(pmdp, addr);
+	i = (addr - range->start) >> PAGE_SHIFT;
 	for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
 		int r;