[PATCH] x86_64: Support more than 8 cores on AMD systems

Use physical mode instead of logical mode to address more CPUs.  This is also
used in the CPU hotplug case to avoid a race.

Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

2 files changed, 85 insertions, 10 deletions

diff --git a/arch/x86_64/kernel/genapic.c b/arch/x86_64/kernel/genapic.c
index 69b9c25a8fc4..30c843a5efdd 100644
--- a/arch/x86_64/kernel/genapic.c
+++ b/arch/x86_64/kernel/genapic.c
@@ -31,6 +31,7 @@ u8 x86_cpu_to_log_apicid[NR_CPUS] = { [0 ... NR_CPUS-1] = BAD_APICID };
 
 extern struct genapic apic_cluster;
 extern struct genapic apic_flat;
+extern struct genapic apic_physflat;
 
 struct genapic *genapic = &apic_flat;
 
@@ -44,12 +45,7 @@ void __init clustered_apic_check(void)
         u8 clusters, max_cluster;
         u8 id;
         u8 cluster_cnt[NUM_APIC_CLUSTERS];
-
-        if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
-                /* AMD always uses flat mode right now */
-                genapic = &apic_flat;
-                goto print;
-        }
+        int num_cpus = 0;
 
 #if defined(CONFIG_ACPI_BUS)
         /*
@@ -64,15 +60,34 @@ void __init clustered_apic_check(void)
 #endif
 
         memset(cluster_cnt, 0, sizeof(cluster_cnt));
-
         for (i = 0; i < NR_CPUS; i++) {
                 id = bios_cpu_apicid[i];
-                if (id != BAD_APICID)
-                        cluster_cnt[APIC_CLUSTERID(id)]++;
+                if (id == BAD_APICID)
+                        continue;
+                num_cpus++;
+                cluster_cnt[APIC_CLUSTERID(id)]++;
         }
 
+        /* Don't use clustered mode on AMD platforms. */
+        if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
+                genapic = &apic_physflat;
+#ifndef CONFIG_CPU_HOTPLUG
+                /* In the CPU hotplug case we cannot use broadcast mode
+                   because that opens a race when a CPU is removed.
+                   Stay at physflat mode in this case.
+                   It is bad to do this unconditionally though. Once
+                   we have ACPI platform support for CPU hotplug
+                   we should detect hotplug capablity from ACPI tables and
+                   only do this when really needed. -AK */
+                if (num_cpus <= 8)
+                        genapic = &apic_flat;
+#endif
+                goto print;
+        }
+
         clusters = 0;
         max_cluster = 0;
+
         for (i = 0; i < NUM_APIC_CLUSTERS; i++) {
                 if (cluster_cnt[i] > 0) {
                         ++clusters;
diff --git a/arch/x86_64/kernel/genapic_flat.c b/arch/x86_64/kernel/genapic_flat.c
index fdfa15f5d2ed..adc96282a9e2 100644
--- a/arch/x86_64/kernel/genapic_flat.c
+++ b/arch/x86_64/kernel/genapic_flat.c
@@ -2,7 +2,7 @@
  * Copyright 2004 James Cleverdon, IBM.
  * Subject to the GNU Public License, v.2
  *
- * Flat APIC subarch code.  Maximum 8 CPUs, logical delivery.
+ * Flat APIC subarch code.
  *
  * Hacked for x86-64 by James Cleverdon from i386 architecture code by
  * Martin Bligh, Andi Kleen, James Bottomley, John Stultz, and
@@ -119,3 +119,63 @@ struct genapic apic_flat =  {
         .cpu_mask_to_apicid = flat_cpu_mask_to_apicid,
         .phys_pkg_id = phys_pkg_id,
 };
+
+/*
+ * Physflat mode is used when there are more than 8 CPUs on a AMD system.
+ * We cannot use logical delivery in this case because the mask
+ * overflows, so use physical mode.
+ */
+
+static cpumask_t physflat_target_cpus(void)
+{
+        return cpumask_of_cpu(0);
+}
+
+static void physflat_send_IPI_mask(cpumask_t cpumask, int vector)
+{
+        send_IPI_mask_sequence(cpumask, vector);
+}
+
+static void physflat_send_IPI_allbutself(int vector)
+{
+        cpumask_t allbutme = cpu_online_map;
+        int me = get_cpu();
+        cpu_clear(me, allbutme);
+        physflat_send_IPI_mask(allbutme, vector);
+        put_cpu();
+}
+
+static void physflat_send_IPI_all(int vector)
+{
+        physflat_send_IPI_mask(cpu_online_map, vector);
+}
+
+static unsigned int physflat_cpu_mask_to_apicid(cpumask_t cpumask)
+{
+        int cpu;
+
+        /*
+         * We're using fixed IRQ delivery, can only return one phys APIC ID.
+         * May as well be the first.
+         */
+        cpu = first_cpu(cpumask);
+        if ((unsigned)cpu < NR_CPUS)
+                return x86_cpu_to_apicid[cpu];
+        else
+                return BAD_APICID;
+}
+
+struct genapic apic_physflat =  {
+        .name = "physical flat",
+        .int_delivery_mode = dest_LowestPrio,
+        .int_dest_mode = (APIC_DEST_PHYSICAL != 0),
+        .int_delivery_dest = APIC_DEST_PHYSICAL | APIC_DM_LOWEST,
+        .target_cpus = physflat_target_cpus,
+        .apic_id_registered = flat_apic_id_registered,
+        .init_apic_ldr = flat_init_apic_ldr,/*not needed, but shouldn't hurt*/
+        .send_IPI_all = physflat_send_IPI_all,
+        .send_IPI_allbutself = physflat_send_IPI_allbutself,
+        .send_IPI_mask = physflat_send_IPI_mask,
+        .cpu_mask_to_apicid = physflat_cpu_mask_to_apicid,
+        .phys_pkg_id = phys_pkg_id,
+};