Re: Softlockup when test shmem swapout-swapin and compaction

[Matthew Wilcox <willy@infradead.org>]

On Wed, Mar 05, 2025 at 01:44:29PM -0500, Zi Yan wrote:
> OK, it is probably still caused by the __folio_migrate_mapping() bug, since
> writing to sibling entries of a multi-index entry breaks the multi-index entry.
> Thank Matthew for doing the experiments.

Here's what I did:

diff --git a/lib/test_xarray.c b/lib/test_xarray.c
index 0e865bab4a10..4f38db416ff6 100644
--- a/lib/test_xarray.c
+++ b/lib/test_xarray.c
@@ -645,6 +645,26 @@ static noinline void check_multi_store_3(struct xarray *xa, unsigned long index,

        xa_destroy(xa);
 }
+
+static noinline void check_multi_store_4(struct xarray *xa, unsigned long index,
+               unsigned int order)
+{
+       XA_STATE(xas, xa, index);
+       int i;
+
+       xa_store_order(xa, index, order, xa_mk_value(0), GFP_KERNEL);
+       xa_dump(xa);
+
+       xas_lock(&xas);
+       for (i = 0; i < (1 << order); i++) {
+               xas_store(&xas, xa_mk_value(index));
+               xas_next(&xas);
+       }
+       xas_unlock(&xas);
+       xa_dump(xa);
+
+       xa_destroy(xa);
+}
 #endif

 static noinline void check_multi_store(struct xarray *xa)
@@ -724,6 +744,7 @@ static noinline void check_multi_store(struct xarray *xa)
                check_multi_store_3(xa, 0, i);
                check_multi_store_3(xa, 1UL << i, i);
        }
+       check_multi_store_4(xa, 16, 4);
 #endif
 }


The xa_dump before shows sibling entries:

$ ./tools/testing/radix-tree/xarray
xarray: 0x55b2335be180x head 0x516004c75b82x flags 0 marks 0 0 0
0-63: node 0x516004c75b80x max 0 parent (nil)x shift 0 count 16 values 16 array 0x55b2335be180x list 0x516004c75b98x 0x516004c75b98x marks 0 0 0
16: value 0 (0x0) [0x1x]
17: sibling (slot 16)
18: sibling (slot 16)
19: sibling (slot 16)
[...]

And then after shows them turned into normal entries:

xarray: 0x55b2335be180x head 0x516004c75b82x flags 0 marks 0 0 0
0-63: node 0x516004c75b80x max 0 parent (nil)x shift 0 count 16 values 31 array 0x55b2335be180x list 0x516004c75b98x 0x516004c75b98x marks 0 0 0
16: value 16 (0x10) [0x21x]
17: value 16 (0x10) [0x21x]
18: value 16 (0x10) [0x21x]
19: value 16 (0x10) [0x21x]

so I understand why this took a long time to show up.  For most uses,
you can't tell the difference between these situations.