Skip to contentSkip to conversation

ublk: fix maple tree lockdep warning and unpin under spinlock

PATCH In review Ming Lei 2 participants 2 messages Apr 22

Subsystems (2)
lore
Share & tools
lore

Ming Lei

author maintainer
#0 lore raw

PATCH 1/1 ublk: fix maple tree lockdep warning and unpin under spinlock

From: Ming Lei <[email protected]>

Fix two issues in the shmem buffer maple tree usage:

1) ublk_buf_cleanup() iterates the tree with mas_for_each() without holding rcu_read_lock or mas_lock, triggering a lockdep splat on CONFIG_PROVE_RCU kernels. Add mas_lock/unlock around the iteration.

2) __ublk_ctrl_unreg_buf() calls unpin_user_pages() under mas_lock (a spinlock). unpin_user_pages can be expensive for large buffers and may take additional locks if folio refcount drops to zero. Restructure to drop mas_lock before unpinning, re-acquiring it to continue iteration.

Both functions now use the same pattern: erase under lock, drop lock, unpin and free, re-lock to continue. Extract ublk_unpin_range_pages() helper to share the page unpinning loop.

---
 drivers/block/ublk_drv.c | 63 +++++++++++++++++++++++++---------------
 1 file changed, 39 insertions(+), 24 deletions(-)
drivers/block/ublk_drv.c +39 -24
index 49fb584e392b..0a92569c0c7d 100644
@@ -5404,16 +5404,39 @@ static int ublk_ctrl_reg_buf(struct ublk_device *ub,
 	return ret;
 }
 
+static void ublk_unpin_range_pages(unsigned long base_pfn,
+				   unsigned long nr_pages)
+{
+#define UBLK_UNPIN_BATCH	32
+	struct page *pages[UBLK_UNPIN_BATCH];
+	unsigned long off;
+
+	for (off = 0; off < nr_pages; ) {
+		unsigned int batch = min_t(unsigned long,
+					   nr_pages - off, UBLK_UNPIN_BATCH);
+		unsigned int j;
+
+		for (j = 0; j < batch; j++)
+			pages[j] = pfn_to_page(base_pfn + off + j);
+		unpin_user_pages(pages, batch);
+		off += batch;
+	}
+}
+
+/*
+ * Drop mas_lock during iteration to avoid unpinning pages under spinlock.
+ * Safe because callers hold ub->mutex (via ublk_lock_buf_tree), preventing
+ * concurrent tree modifications.
+ */
 static int __ublk_ctrl_unreg_buf(struct ublk_device *ub, int buf_index)
 {
 	MA_STATE(mas, &ub->buf_tree, 0, ULONG_MAX);
 	struct ublk_buf_range *range;
-	struct page *pages[32];
 	int ret = -ENOENT;
 
 	mas_lock(&mas);
 	mas_for_each(&mas, range, ULONG_MAX) {
-		unsigned long base, nr, off;
+		unsigned long base, nr;
 
 		if (range->buf_index != buf_index)
 			continue;
@@ -5422,18 +5445,12 @@ static int __ublk_ctrl_unreg_buf(struct ublk_device *ub, int buf_index)
 		base = mas.index;
 		nr = mas.last - base + 1;
 		mas_erase(&mas);
+		mas_unlock(&mas);
 
-		for (off = 0; off < nr; ) {
-			unsigned int batch = min_t(unsigned long,
-						   nr - off, 32);
-			unsigned int j;
-
-			for (j = 0; j < batch; j++)
-				pages[j] = pfn_to_page(base + off + j);
-			unpin_user_pages(pages, batch);
-			off += batch;
-		}
+		ublk_unpin_range_pages(base, nr);
 		kfree(range);
+
+		mas_lock(&mas);
 	}
 	mas_unlock(&mas);
 
@@ -5463,29 +5480,27 @@ static int ublk_ctrl_unreg_buf(struct ublk_device *ub,
 	return ret;
 }
 
+/*
+ * Drop mas_lock during iteration to avoid unpinning pages under spinlock.
+ * Safe because this is called from device release with exclusive access.
+ */
 static void ublk_buf_cleanup(struct ublk_device *ub)
 {
 	MA_STATE(mas, &ub->buf_tree, 0, ULONG_MAX);
 	struct ublk_buf_range *range;
-	struct page *pages[32];
 
+	mas_lock(&mas);
 	mas_for_each(&mas, range, ULONG_MAX) {
 		unsigned long base = mas.index;
 		unsigned long nr = mas.last - base + 1;
-		unsigned long off;
 
-		for (off = 0; off < nr; ) {
-			unsigned int batch = min_t(unsigned long,
-						   nr - off, 32);
-			unsigned int j;
-
-			for (j = 0; j < batch; j++)
-				pages[j] = pfn_to_page(base + off + j);
-			unpin_user_pages(pages, batch);
-			off += batch;
-		}
+		mas_erase(&mas);
+		mas_unlock(&mas);
+		ublk_unpin_range_pages(base, nr);
 		kfree(range);
+		mas_lock(&mas);
 	}
+	mas_unlock(&mas);
 	mtree_destroy(&ub->buf_tree);
 	ida_destroy(&ub->buf_ida);
 }
Signature · 1 line
2.53.0
Tested-by on v1 1/1
#1 lore raw
Quoted 14 lines · Ming Lei

From: Ming Lei <[email protected]>

Fix two issues in the shmem buffer maple tree usage:

1) ublk_buf_cleanup() iterates the tree with mas_for_each() without holding rcu_read_lock or mas_lock, triggering a lockdep splat on CONFIG_PROVE_RCU kernels. Add mas_lock/unlock around the iteration.

2) __ublk_ctrl_unreg_buf() calls unpin_user_pages() under mas_lock (a spinlock). unpin_user_pages can be expensive for large buffers and may take additional locks if folio refcount drops to zero. Restructure to drop mas_lock before unpinning, re-acquiring it to continue iteration.

Quoted 8 lines

Both functions now use the same pattern: erase under lock, drop lock, unpin and free, re-lock to continue. Extract ublk_unpin_range_pages() helper to share the page unpinning loop.

I observed the same WARN at blktests test cases block/033 and ublk/*. I confirmed this patch avoids the WARN and makes the test cases pass. Thanks!

I also bisected and confirmed that the commit below is the failure trigger. Adding a Fixes tag might add a bit of value.

Fixes 5e864438e285 ("ublk: replace xarray with IDA for shmem buffer index allocation")