[PATCH] mm/mremap: allow VMAs with VM_DONTEXPAND|VM_PFNMAP when creating new mapping

[Vivek Kasireddy <vivek.kasireddy@intel.com>]

When mremap is used to create a new mapping, we should not return
-EFAULT for VMAs with VM_DONTEXPAND or VM_PFNMAP flags set because
the old VMA would neither be expanded nor shrunk in this case. This
is particularly useful when trying to create a new VMA using other
existing VMAs that have these flags set, such as the ones associated
with VFIO devices.

Specifically, there are use-cases where a VMM such as Qemu would
want to map a non-contiguous buffer associated with a VFIO device
in the following way:

    void *start, *cur;
    int i;

    start = mmap(NULL, size, PROT_NONE, MAP_SHARED, -1, 0);
    if (start == MAP_FAILED) {
        return start;
    }

    cur = start;
    for (i = 0; i < iov_cnt; i++) {
         if (mremap(iov[i].iov_base, 0, iov[i].iov_len,
             MREMAP_FIXED | MREMAP_MAYMOVE, cur) == MAP_FAILED) {
             goto err;
         }
         cur += iov[i].iov_len;
    }
    return start;

The above code currently works when mapping buffers backed by
shmem (memfd) but fails with -EFAULT when mapping VFIO backed
buffers because the VMAs associated with iov[i].iov_base addresses
have VM_DONTEXPAND and VM_PFNMAP flags set. Therefore, fix this
issue by not returning -EFAULT when a new mapping is being created.

Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Pedro Falcato <pfalcato@suse.de>
Cc: David Hildenbrand <david@kernel.org>
Cc: Akihiko Odaki <odaki@rsg.ci.i.u-tokyo.ac.jp>
Signed-off-by: Vivek Kasireddy <vivek.kasireddy@intel.com>
---
 mm/mremap.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/mm/mremap.c b/mm/mremap.c
index fdb0485ede74..d3868d941f72 100644
--- a/mm/mremap.c
+++ b/mm/mremap.c
@@ -1736,7 +1736,8 @@ static int check_prep_vma(struct vma_remap_struct *vrm)
 	if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 		return -EINVAL;
 
-	if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
+	if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP) &&
+	    !vrm_implies_new_addr(vrm))
 		return -EFAULT;
 
 	if (!mlock_future_ok(mm, vma->vm_flags, vrm->delta))
-- 
2.50.1