On Thu, Oct 25, 2012 at 04:32:52PM +0800, Daniel J Blueman wrote:The AMD Northbridge initialisation code and EDAC assume the Northbridge IDs
are contiguous, which no longer holds on federated systems with multiple
HyperTransport fabrics and multiple PCI domains, eg on Numascale's
Numaconnect systems with NumaChip.
Address this assumption by searching the Northbridge ID array, rather than
directly indexing it, using the upper bits for the PCI domain.
RFC->v2: Correct array initialisation
v2->v3: Add Boris's neater linked list approach
Todo:
1. fix kobject/sysfs oops (see http://quora.org/2012/16-server-boot.txt later)
2. reorder amd64_edac.c or add amd64_per_family_init/pci_get_related_function
forward declarations, based on feedback
Signed-off-by: Daniel J Blueman <daniel@xxxxxxxxxxxxxxxxxx>
This patch contains code from both of us and thus needs both our SOBs:
Signed-off-by: Borislav Petkov <bp@xxxxxxxxx>
---
arch/x86/include/asm/amd_nb.h | 63 +++++++++++++++-
arch/x86/include/asm/numachip/numachip.h | 22 ++++++
arch/x86/kernel/amd_gart_64.c | 8 +-
arch/x86/kernel/amd_nb.c | 85 ++++++++++++---------
arch/x86/pci/numachip.c | 121 ++++++++++++++++++++++++++++++
drivers/char/agp/amd64-agp.c | 12 +--
drivers/edac/amd64_edac.c | 34 +++++----
drivers/edac/amd64_edac.h | 6 --
8 files changed, 283 insertions(+), 68 deletions(-)
create mode 100644 arch/x86/include/asm/numachip/numachip.h
create mode 100644 arch/x86/pci/numachip.c
diff --git a/arch/x86/include/asm/amd_nb.h b/arch/x86/include/asm/amd_nb.h
index b3341e9..6a27226 100644
--- a/arch/x86/include/asm/amd_nb.h
+++ b/arch/x86/include/asm/amd_nb.h
@@ -4,6 +4,8 @@
#include <linux/ioport.h>
#include <linux/pci.h>
+#define NUM_POSSIBLE_NBS 8
+
struct amd_nb_bus_dev_range {
u8 bus;
u8 dev_base;
@@ -51,12 +53,22 @@ struct amd_northbridge {
struct pci_dev *link;
struct amd_l3_cache l3_cache;
struct threshold_bank *bank4;
+ u16 node;
+ struct list_head nbl;
};
struct amd_northbridge_info {
u16 num;
u64 flags;
- struct amd_northbridge *nb;
+
+ /*
+ * The first 8 elems are for fast lookup of NB descriptors on single-
+ * system setups, i.e. "normal" boxes. The nb_list, OTOH, is list of
+ * additional NB descriptors which exist on confederate systems
+ * like using Numascale's Numaconnect/NumaChip.
+ */
+ struct amd_northbridge *nbs[NUM_POSSIBLE_NBS];
+ struct list_head nb_list;
};
extern struct amd_northbridge_info amd_northbridges;
@@ -78,7 +90,54 @@ static inline bool amd_nb_has_feature(unsigned feature)
static inline struct amd_northbridge *node_to_amd_nb(int node)
{
- return (node < amd_northbridges.num) ? &amd_northbridges.nb[node] : NULL;
+ struct amd_northbridge_info *nbi = &amd_northbridges;
+ struct amd_northbridge *nb;
+ int i;
+
+ /* Quick search for first domain */
+ if (node < NUM_POSSIBLE_NBS) {
+ if (node < nbi->num)
+ return nbi->nbs[node];
+ else
+ return NULL;
+ }
Why change that here from what I had before?
nbi->nbs[node] will either return a valid descriptor or NULL because it
is statically allocated in amd_northbridge_info.
So why add a conditional where you clearly don't need it?
+ /* Search for NBs from later domains in array */
+ for (i = 0; i < NUM_POSSIBLE_NBS; i++)
+ if (nbi->nbs[i]->node == node)
+ return nbi->nbs[i];
And then this is not needed.
+
+ list_for_each_entry(nb, &nbi->nb_list, nbl)
+ if (node == nb->node)
+ return nb;
And why change the list_for_each_entry_safe variant? It is not needed
now but who knows what code changes where in the future.
+
+ return NULL;
+}
+
+static inline struct amd_northbridge *index_to_amd_nb(int index)
+{
+ struct amd_northbridge_info *nbi = &amd_northbridges;
+ struct amd_northbridge *nb;
+ int count = NUM_POSSIBLE_NBS;
+
+ if (index < NUM_POSSIBLE_NBS) {
+ if (index < nbi->num)
+ return nbi->nbs[index];
+ else
+ return NULL;
+ }
+
+ list_for_each_entry(nb, &nbi->nb_list, nbl) {
+ if (count++ == index)
+ return nb;
+ }
+
+ return NULL;
+}
Huh, what do we need that function for? node should be equal to index
for the first 8 and then we use the linked list. What's up?