On Wed, Jul 12, 2023 at 6:42 AM Hugh Dickins [off-list ref] wrote:

Bring collapse_and_free_pmd() back into collapse_pte_mapped_thp().
It does need mmap_read_lock(), but it does not need mmap_write_lock(),
nor vma_start_write() nor i_mmap lock nor anon_vma lock.  All racing
paths are relying on pte_offset_map_lock() and pmd_lock(), so use those.

We can still have a racing userfaultfd operation at the "/* step 4:
remove page table */" point that installs a new PTE before the page
table is removed.

To reproduce, patch a delay into the kernel like this:

diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index 9a6e0d507759..27cc8dfbf3a7 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c

@@ -20,6 +20,7 @@
 #include <linux/swapops.h>
 #include <linux/shmem_fs.h>
 #include <linux/ksm.h>
+#include <linux/delay.h>

 #include <asm/tlb.h>
 #include <asm/pgalloc.h>

@@ -1617,6 +1618,11 @@ int collapse_pte_mapped_thp(struct mm_struct

*mm, unsigned long addr,
        }

        /* step 4: remove page table */
+       if (strcmp(current->comm, "DELAYME") == 0) {
+               pr_warn("%s: BEGIN DELAY INJECTION\n", __func__);
+               mdelay(5000);
+               pr_warn("%s: END DELAY INJECTION\n", __func__);
+       }

        /* Huge page lock is still held, so page table must remain empty */
        pml = pmd_lock(mm, pmd);


And then run the attached reproducer against mm/mm-everything. You
should get this in dmesg:

[  206.578096] BUG: Bad rss-counter state mm:000000000942ebea
type:MM_ANONPAGES val:1