diff --git a/fs/io_uring.c b/fs/io_uring.c
index 69a53c96a29eaf2e2da63d4a523c478f3f9737c4..b652e7f4b4db4fa8e4c2c6870048fdfb57c58d6f 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -7290,9 +7290,6 @@ static void io_ring_ctx_free(struct io_ring_ctx *ctx)
 	io_mem_free(ctx->sq_sqes);
 
 	percpu_ref_exit(&ctx->refs);
-	if (ctx->account_mem)
-		io_unaccount_mem(ctx->user,
-				ring_pages(ctx->sq_entries, ctx->cq_entries));
 	free_uid(ctx->user);
 	put_cred(ctx->creds);
 	kfree(ctx->cancel_hash);
@@ -7377,6 +7374,16 @@ static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
 	if (ctx->rings)
 		io_cqring_overflow_flush(ctx, true);
 	idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
+
+	/*
+	 * Do this upfront, so we won't have a grace period where the ring
+	 * is closed but resources aren't reaped yet. This can cause
+	 * spurious failure in setting up a new ring.
+	 */
+	if (ctx->account_mem)
+		io_unaccount_mem(ctx->user,
+				ring_pages(ctx->sq_entries, ctx->cq_entries));
+
 	INIT_WORK(&ctx->exit_work, io_ring_exit_work);
 	queue_work(system_wq, &ctx->exit_work);
 }