On Fri, Jul 11, 2025 at 03:34:23PM +0200, Vlastimil Babka wrote:

+cc linux-api - see the description of the new behavior below

Ah yeah :) I sent on 0/10 also. Friday...

On 7/11/25 13:38, Lorenzo Stoakes wrote:

quoted

Historically we've made it a uAPI requirement that mremap() may only
operate on a single VMA at a time.

For instances where VMAs need to be resized, this makes sense, as it
becomes very difficult to determine what a user actually wants should they
indicate a desire to expand or shrink the size of multiple VMAs (truncate?
Adjust sizes individually? Some other strategy?).

However, in instances where a user is moving VMAs, it is restrictive to
disallow this.

This is especially the case when anonymous mapping remap may or may not be
mergeable depending on whether VMAs have or have not been faulted due to
anon_vma assignment and folio index alignment with vma->vm_pgoff.

Often this can result in surprising impact where a moved region is faulted,
then moved back and a user fails to observe a merge from otherwise
compatible, adjacent VMAs.

This change allows such cases to work without the user having to be
cognizant of whether a prior mremap() move or other VMA operations has
resulted in VMA fragmentation.

We only permit this for mremap() operations that do NOT change the size of
the VMA and DO specify MREMAP_MAYMOVE | MREMAP_FIXED.

Should no VMA exist in the range, -EFAULT is returned as usual.

If a VMA move spans a single VMA - then there is no functional change.

Otherwise, we place additional requirements upon VMAs:

* They must not have a userfaultfd context associated with them - this
  requires dropping the lock to notify users, and we want to perform the
  operation with the mmap write lock held throughout.

* If file-backed, they cannot have a custom get_unmapped_area handler -
  this might result in MREMAP_FIXED not being honoured, which could result
  in unexpected positioning of VMAs in the moved region.

There may be gaps in the range of VMAs that are moved:

                   X        Y                       X        Y
                 <--->     <->                    <--->     <->
         |-------|   |-----| |-----|      |-------|   |-----| |-----|
         |   A   |   |  B  | |  C  | ---> |   A'  |   |  B' | |  C' |
         |-------|   |-----| |-----|      |-------|   |-----| |-----|
        addr                           new_addr

The move will preserve the gaps between each VMA.

AFAIU "moving a gap" doesn't mean we unmap anything pre-existing where the
moved gap's range falls to, right? Worth pointing out explicitly.

quoted

Note that any failures encountered will result in a partial move. Since an
mremap() can fail at any time, this might result in only some of the VMAs
being moved.

Note that failures are very rare and typically require an out of a memory
condition or a mapping limit condition to be hit, assuming the VMAs being
moved are valid.

We don't try to assess ahead of time whether VMAs are valid according to
the multi VMA rules, as it would be rather unusual for a user to mix
uffd-enabled VMAs and/or VMAs which map unusual driver mappings that
specify custom get_unmapped_area() handlers in an aggregate operation.

So we optimise for the far, far more likely case of the operation being
entirely permissible.

Guess it's the sanest thing to do given all the cirumstances.

quoted

In the case of the move of a single VMA, the above conditions are
permitted. This makes the behaviour identical for a single VMA as before.

Signed-off-by: Lorenzo Stoakes <redacted>

Reviewed-by: Vlastimil Babka <redacted>

Some nits:

quoted

---
 mm/mremap.c | 157 +++++++++++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 150 insertions(+), 7 deletions(-)

diff --git a/mm/mremap.c b/mm/mremap.c
index 8cb08ccea6ad..59f49de0f84e 100644
--- a/mm/mremap.c
+++ b/mm/mremap.c

@@ -69,6 +69,8 @@ struct vma_remap_struct {
 	enum mremap_type remap_type;	/* expand, shrink, etc. */
 	bool mmap_locked;		/* Is mm currently write-locked? */
 	unsigned long charged;		/* If VM_ACCOUNT, # pages to account. */
+	bool seen_vma;			/* Is >1 VMA being moved? */

Seems this could be local variable of remap_move().

Yes, this is because before there _was_ some external use, but after rework
not any more. Will fix up in a fix-patch.

quoted

+	bool vmi_needs_reset;		/* Was the VMA iterator invalidated? */
 };

 static pud_t *get_old_pud(struct mm_struct *mm, unsigned long addr)

@@ -1188,6 +1190,9 @@ static int copy_vma_and_data(struct vma_remap_struct *vrm,
 		*new_vma_ptr = NULL;
 		return -ENOMEM;
 	}

A newline here?

I kinda thought it made sense to 'group' it with logic above, so this was
on purpose.

quoted

+	if (vma != vrm->vma)
+		vrm->vmi_needs_reset = true;

A comment on what this condition means wouldn't hurt? Is it when "Source vma
may have been merged into new_vma" in copy_vma(), or when not?

Sure will add in a fix-patch.