On Wed, Mar 1, 2023 at 10:34 AM Suren Baghdasaryan [off-list ref] wrote:

On Tue, Feb 28, 2023 at 11:57 PM Hyeonggon Yoo [off-list ref] wrote:

quoted

On Wed, Mar 01, 2023 at 07:43:33AM +0000, Hyeonggon Yoo wrote:

quoted

On Mon, Feb 27, 2023 at 09:36:17AM -0800, Suren Baghdasaryan wrote:

quoted

Write-locking VMAs before isolating them ensures that page fault
handlers don't operate on isolated VMAs.

Signed-off-by: Suren Baghdasaryan <surenb@google.com>
---
 mm/mmap.c  | 1 +
 mm/nommu.c | 5 +++++
 2 files changed, 6 insertions(+)

diff --git a/mm/mmap.c b/mm/mmap.c
index 1f42b9a52b9b..f7ed357056c4 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c

@@ -2255,6 +2255,7 @@ int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
 static inline int munmap_sidetree(struct vm_area_struct *vma,
                               struct ma_state *mas_detach)
 {
+   vma_start_write(vma);
    mas_set_range(mas_detach, vma->vm_start, vma->vm_end - 1);

I may be missing something, but have few questions:

      1) Why does a writer need to both write-lock a VMA and mark the VMA detached
         when unmapping it, isn't it enough to just only write-lock a VMA?

We need to mark the VMA detached to avoid handling page fault in a
detached VMA. The possible scenario is:

lock_vma_under_rcu
  vma = mas_walk(&mas)
                                                        munmap_sidetree
                                                          vma_start_write(vma)

mas_store_gfp() // remove VMA from the tree
                                                          vma_end_write_all()
  vma_start_read(vma)
  // we locked the VMA but it is not part of the tree anymore.

So, marking the VMA locked before vma_end_write_all() and checking

Sorry, I should have said "marking the VMA *detached* before
vma_end_write_all() and checking vma->detached after vma_start_read()
helps us avoid handling faults in the detached VMA."

vma->detached after vma_start_read() helps us avoid handling faults in
the detached VMA.

quoted

quoted

      2) as VMAs that are going to be removed are already locked in vma_prepare(),
         so I think this hunk could be dropped?

After sending this just realized that I did not consider simple munmap case :)
But I still think 1) and 3) are valid question.

quoted

quoted

    if (mas_store_gfp(mas_detach, vma, GFP_KERNEL))
            return -ENOMEM;

diff --git a/mm/nommu.c b/mm/nommu.c
index 57ba243c6a37..2ab162d773e2 100644
--- a/mm/nommu.c
+++ b/mm/nommu.c

@@ -588,6 +588,7 @@ static int delete_vma_from_mm(struct vm_area_struct *vma)
                   current->pid);
            return -ENOMEM;
    }
+   vma_start_write(vma);
    cleanup_vma_from_mm(vma);

      3) I think this hunk could be dropped as Per-VMA lock depends on MMU anyway.

Ah, yes, you are right. We can safely remove the changes in nommu.c
Andrew, should I post a fixup or you can make the removal directly in
mm-unstable?
Thanks,
Suren.

quoted

quoted

Thanks,
Hyeonggon

quoted

    /* remove from the MM's tree and list */

@@ -1519,6 +1520,10 @@ void exit_mmap(struct mm_struct *mm)
     */
    mmap_write_lock(mm);
    for_each_vma(vmi, vma) {
+           /*
+            * No need to lock VMA because this is the only mm user and no
+            * page fault handled can race with it.
+            */
            cleanup_vma_from_mm(vma);
            delete_vma(mm, vma);
            cond_resched();
--

2.39.2.722.g9855ee24e9-goog

--
To unsubscribe from this group and stop receiving emails from it, send an email to kernel-team+unsubscribe@android.com.