Rob Herring [off-list ref] writes:

On Mon, Dec 13, 2021 at 6:47 AM Michael Ellerman [off-list ref] wrote:

quoted

Rob Herring [off-list ref] writes:

quoted

Use of the of_scan_flat_dt() function predates libfdt and is discouraged
as libfdt provides a nicer set of APIs. Rework
early_init_dt_scan_memory() to be called directly and use libfdt.

...

quoted

diff --git a/arch/powerpc/kernel/prom.c b/arch/powerpc/kernel/prom.c
index 6e1a106f02eb..63762a3b75e8 100644
--- a/arch/powerpc/kernel/prom.c
+++ b/arch/powerpc/kernel/prom.c

@@ -532,19 +532,19 @@ static int  __init early_init_drmem_lmb(struct drmem_lmb *lmb,
 }
 #endif /* CONFIG_PPC_PSERIES */

-static int __init early_init_dt_scan_memory_ppc(unsigned long node,
-                                             const char *uname,
-                                             int depth, void *data)
+static int __init early_init_dt_scan_memory_ppc(void)
 {
 #ifdef CONFIG_PPC_PSERIES
-     if (depth == 1 &&
-         strcmp(uname, "ibm,dynamic-reconfiguration-memory") == 0) {
+     const void *fdt = initial_boot_params;
+     int node = fdt_path_offset(fdt, "/ibm,dynamic-reconfiguration-memory");
+
+     if (node > 0) {
              walk_drmem_lmbs_early(node, NULL, early_init_drmem_lmb);
              return 0;
      }

It's that return that is the problem.

Now that early_init_dt_scan_memory_ppc() is only called once, that
return causes us to skip scanning regular memory nodes if there is an
"ibm,dynamic-reconfiguration-memory" property present.

So the fix is just:

diff --git a/arch/powerpc/kernel/prom.c b/arch/powerpc/kernel/prom.c
index 1098de3b172f..125661e5fcf3 100644
--- a/arch/powerpc/kernel/prom.c
+++ b/arch/powerpc/kernel/prom.c

@@ -538,10 +538,8 @@ static int __init early_init_dt_scan_memory_ppc(void)
 	const void *fdt = initial_boot_params;
 	int node = fdt_path_offset(fdt, "/ibm,dynamic-reconfiguration-memory");
 
-	if (node > 0) {
+	if (node > 0)
 		walk_drmem_lmbs_early(node, NULL, early_init_drmem_lmb);
-		return 0;
-	}
 #endif
 	
 	return early_init_dt_scan_memory();

The only thing I see is now there is an assumption that 'memory' nodes
are off the root node only. Before they could be anywhere.

I don't know of any machines where that would be a problem. But given
all the wild and wonderful device trees out there, who really knows :)

Maybe we should continue to allow memory nodes to be anywhere, and print
a warning for any that aren't at the root. Then if no one reports any
hits for the warning we could switch to only allowing them at the root?

cheers