diff --git a/Makefile b/Makefile
index 3564332..ed07a4e 100644
--- a/Makefile
+++ b/Makefile
@@ -26,7 +26,7 @@ endif
 # at an arbitrary location, so build it as position independent, but link as
 # non-pie as all relocations are internal and there is no dynamic loader to
 # help.
-CFLAGS  += -Os -g -MMD -MP -march=btver2 -mno-sse -mno-mmx -fpie -fomit-frame-pointer
+CFLAGS  += -Os -g -MMD -MP -mno-sse -mno-mmx -fpie -fomit-frame-pointer
 CFLAGS  += -Iinclude -ffreestanding -fno-common -Wall -Werror
 LDFLAGS += -nostdlib -no-pie -Wl,--build-id=none,--fatal-warnings,--defsym=BITS=$(BITS)
 
diff --git a/dev.c b/dev.c
index 835901b..f71bd50 100644
--- a/dev.c
+++ b/dev.c
@@ -20,67 +20,89 @@
 #include <pci.h>
 #include <dev.h>
 
-u32 dev_locate(void)
+/*
+ * There are only 5 bits (0x00..0x1f) for PCI slot number (see definition of
+ * PCI_DEVFN) and we start at 0x18 (DEV_PCI_DEVICE), so there is hard upper
+ * limit on how many nodes can exist.
+ */
+#define MAX_CPU_NODES 8
+
+u32 dev_locate(u8 cpu_node)
 {
        return pci_locate(DEV_PCI_BUS,
-                         PCI_DEVFN(DEV_PCI_DEVICE, DEV_PCI_FUNCTION));
+                         PCI_DEVFN(DEV_PCI_DEVICE + cpu_node, DEV_PCI_FUNCTION));
 }
 
-u32 dev_read(u32 dev_cap, u32 function, u32 index)
+u32 dev_read(u8 cpu_node, u32 dev_cap, u32 function, u32 index)
 {
     u32 value;
 
     pci_write(0, DEV_PCI_BUS,
-              PCI_DEVFN(DEV_PCI_DEVICE, DEV_PCI_FUNCTION),
+              PCI_DEVFN(DEV_PCI_DEVICE + cpu_node, DEV_PCI_FUNCTION),
               dev_cap + DEV_OP_OFFSET,
               4,
               (u32)(((function & 0xff) << 8) + (index & 0xff)));
 
     pci_read(0, DEV_PCI_BUS,
-             PCI_DEVFN(DEV_PCI_DEVICE, DEV_PCI_FUNCTION),
+             PCI_DEVFN(DEV_PCI_DEVICE + cpu_node, DEV_PCI_FUNCTION),
              dev_cap + DEV_DATA_OFFSET,
              4, &value);
 
        return value;
 }
 
-void dev_write(u32 dev, u32 function, u32 index, u32 value)
+void dev_write(u8 cpu_node, u32 dev, u32 function, u32 index, u32 value)
 {
     pci_write(0, DEV_PCI_BUS,
-        PCI_DEVFN(DEV_PCI_DEVICE, DEV_PCI_FUNCTION),
+        PCI_DEVFN(DEV_PCI_DEVICE + cpu_node, DEV_PCI_FUNCTION),
         dev + DEV_OP_OFFSET,
         4,
         (u32)(((function & 0xff) << 8) + (index & 0xff)) );
 
     pci_write(0, DEV_PCI_BUS,
-        PCI_DEVFN(DEV_PCI_DEVICE, DEV_PCI_FUNCTION),
+        PCI_DEVFN(DEV_PCI_DEVICE + cpu_node, DEV_PCI_FUNCTION),
         dev + DEV_DATA_OFFSET,
         4, value);
 }
 
-void dev_disable_sl(u32 dev)
+void dev_disable_sl(u8 cpu_node, u32 dev)
 {
-    u32 dev_cr = dev_read(dev, DEV_CR, 0);
-    dev_write(dev, DEV_CR, 0, dev_cr & ~(DEV_CR_SL_DEV_EN_MASK));
+    u32 dev_cr = dev_read(cpu_node, dev, DEV_CR, 0);
+    dev_write(cpu_node, dev, DEV_CR, 0, dev_cr & ~(DEV_CR_SL_DEV_EN_MASK));
 }
 
 void disable_memory_protection(void)
 {
-       u32 dev_cap, sldev;
-
-       dev_cap = dev_locate();
-       if (dev_cap) {
-               /* Older families with remains of DEV */
-               dev_disable_sl(dev_cap);
-               return;
-       }
-
-       /* Fam 17h uses different DMA protection control register */
-       pci_read(0, MCH_PCI_BUS,
-                PCI_DEVFN(MCH_PCI_DEVICE, MCH_PCI_FUNCTION),
-                MEMPROT_CR, 4, &sldev);
-       pci_write(0, MCH_PCI_BUS,
-                 PCI_DEVFN(MCH_PCI_DEVICE, MCH_PCI_FUNCTION),
-                 MEMPROT_CR, 4, sldev & ~(MEMPROT_EN));
+    u32 dev_cap, sldev, vid_did;
+    u8 cpu_node = 0;
+
+    dev_cap = dev_locate(cpu_node);
+    if (dev_cap) {
+        /* Older families with remains of DEV */
+        do {
+            dev_disable_sl(cpu_node, dev_cap);
+
+            cpu_node++;
+            if (cpu_node == MAX_CPU_NODES)
+                break;
+
+            dev_cap = dev_locate(cpu_node);
+        } while (dev_cap);
+        return;
+    }
+
+    /* Fam 17h uses different DMA protection control register */
+    while (cpu_node < MAX_CPU_NODES &&
+           pci_read(0, MCH_PCI_BUS,
+                    PCI_DEVFN(MCH_PCI_DEVICE + cpu_node, MCH_PCI_FUNCTION),
+                    VIDDID, 4, &vid_did) == 0 &&
+           vid_did != 0xffffffffU) {
+        u8 devfn = PCI_DEVFN(MCH_PCI_DEVICE + cpu_node, MCH_PCI_FUNCTION);
+
+        pci_read(0, MCH_PCI_BUS, devfn, MEMPROT_CR, 4, &sldev);
+        pci_write(0, MCH_PCI_BUS, devfn, MEMPROT_CR, 4, sldev & ~(MEMPROT_EN));
+
+        cpu_node++;
+    }
 }
 
diff --git a/include/dev.h b/include/dev.h
index 860c139..6c52112 100644
--- a/include/dev.h
+++ b/include/dev.h
@@ -37,12 +37,13 @@
 #define MCH_PCI_DEVICE		0x18
 #define MCH_PCI_FUNCTION	0x0
 
+#define VIDDID			0
 #define MEMPROT_CR		0x384
 
 #define MEMPROT_EN		(1<<0)
 
-u32 dev_locate(void);
-void dev_disable_sl(u32 dev);
+u32 dev_locate(u8 cpu_node);
+void dev_disable_sl(u8 cpu_node, u32 dev);
 void disable_memory_protection(void);
 
 #endif /* __DEV_H__ */