[RFC 2/3] powerpc/numa: Define mapping between HW and kernel cpus

[Michael Bringmann]

Define interface to map external powerpc cpus across multiple nodes
to a range of kernel cpu values.  Mapping is intended to prevent
confusion within the kernel about the cpu+node mapping, and the
changes in configuration that may happen due to powerpc LPAR
migration or other associativity changes during the lifetime of a
system.  These interfaces will be used entirely within the powerpc
kernel code to maintain separation between the machine and kernel
contexts.

Signed-off-by: Michael Bringmann <mwb@xxxxxxxxxxxxxxxxxx>
---
 arch/powerpc/include/asm/topology.h       |   31 +++++++
 arch/powerpc/platforms/pseries/Kconfig    |   10 ++
 arch/powerpc/platforms/pseries/Makefile   |    1 
 arch/powerpc/platforms/pseries/cpuremap.c |  131 +++++++++++++++++++++++++++++
 4 files changed, 173 insertions(+)
 create mode 100644 arch/powerpc/platforms/pseries/cpuremap.c

diff --git a/arch/powerpc/include/asm/topology.h b/arch/powerpc/include/asm/topology.h
index 4621f40..db11969 100644
--- a/arch/powerpc/include/asm/topology.h
+++ b/arch/powerpc/include/asm/topology.h
@@ -135,5 +135,36 @@ static inline void shared_proc_topology_init(void) {}
 #endif
 #endif
 
+#define CPUREMAP_NO_CPU		(~0)
+#define CPUREMAP_NO_THREAD	(~0)
+
+#ifdef CONFIG_CPUREMAP
+extern int cpuremap_thread_to_cpu(int thread_index);
+		/* Return CPUREMAP_NO_CPU if not found */
+extern int cpuremap_map_cpu(int thread_index, int in_core_ndx, int node);
+		/* Return CPUREMAP_NO_CPU if fails */
+extern int cpuremap_reserve_cpu(int cpu);
+		/* Return CPUREMAP_NO_CPU if fails */
+extern int cpuremap_release_cpu(int cpu);
+		/* Return CPUREMAP_NO_CPU if fails */
+extern int cpuremap_cpu_to_thread(int cpu);
+		/* Return CPUREMAP_NO_THREAD if not found */
+extern void cpuremap_init(void);
+		/* Identify necessary constants & alloc memory at boot */
+#else
+static inline int cpuremap_thread_to_cpu(int thread_index)
+{
+	return thread_index;
+}
+static inline int cpuremap_map_cpu(int thread_index, int in_core_ndx, int node)
+{
+	return thread_index;
+}
+static inline int cpuremap_reserve_cpu(int cpu) { return cpu; }
+static inline int cpuremap_release_cpu(int cpu) { return cpu; }
+static inline int cpuremap_cpu_to_thread(int cpu) { return cpu; }
+static inline void cpuremap_init(void) {}
+#endif
+
 #endif /* __KERNEL__ */
 #endif	/* _ASM_POWERPC_TOPOLOGY_H */
diff --git a/arch/powerpc/platforms/pseries/Kconfig b/arch/powerpc/platforms/pseries/Kconfig
index 2e4bd32..c35009f 100644
--- a/arch/powerpc/platforms/pseries/Kconfig
+++ b/arch/powerpc/platforms/pseries/Kconfig
@@ -145,3 +145,13 @@ config PAPR_SCM
 	tristate "Support for the PAPR Storage Class Memory interface"
 	help
 	  Enable access to hypervisor provided storage class memory.
+          Enable access to hypervisor provided storage class memory.
+
+config CPUREMAP
+        bool "Support for mapping hw cpu+node to kernel index"
+        depends on SMP && (PPC_PSERIES)
+        ---help---
+          Say Y here to be able to remap hw cpu+node to standardized
+          kernel CPUs at runtime on Pseries machines.
+
+          Say N if you are unsure.
diff --git a/arch/powerpc/platforms/pseries/Makefile b/arch/powerpc/platforms/pseries/Makefile
index a43ec84..ad49d8e 100644
--- a/arch/powerpc/platforms/pseries/Makefile
+++ b/arch/powerpc/platforms/pseries/Makefile
@@ -13,6 +13,7 @@ obj-$(CONFIG_KEXEC_CORE)	+= kexec.o
 obj-$(CONFIG_PSERIES_ENERGY)	+= pseries_energy.o
 
 obj-$(CONFIG_HOTPLUG_CPU)	+= hotplug-cpu.o
+obj-$(CONFIG_CPUREMAP)		+= cpuremap.o
 obj-$(CONFIG_MEMORY_HOTPLUG)	+= hotplug-memory.o pmem.o
 
 obj-$(CONFIG_HVC_CONSOLE)	+= hvconsole.o
diff --git a/arch/powerpc/platforms/pseries/cpuremap.c b/arch/powerpc/platforms/pseries/cpuremap.c
new file mode 100644
index 0000000..86fdf12
--- /dev/null
+++ b/arch/powerpc/platforms/pseries/cpuremap.c
@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: GPL-2.0
+#include <linux/string.h>
+#include <linux/err.h>
+#include <linux/slab.h>
+#include <linux/of.h>
+#include <asm/prom.h>
+#include <asm/topology.h>
+
+struct cpuremap_cpu {
+	int thread_index;
+		/* Set to thread_index from ibm,ppc-interrupt-server#s arrays
+		 * Don't clear when release'ed
+		 */
+	int node;
+	bool in_use;
+		/* Set to true when reserve'ed
+		 * Don't clear when release'ed
+		*/
+};
+
+struct cpuremap_struct {
+	int num_nodes;
+	int num_cores;
+	int num_threads_per_core;
+	struct cpuremap_cpu *threads;
+} cpuremap_data;
+
+
+void cpuremap_init(void)
+{
+	int i, k;
+
+	/* Identify necessary constants & alloc memory at boot */
+	cpuremap_data.num_threads_per_core = 8;
+	cpuremap_data.num_cores = 32;
+	cpuremap_data.num_nodes =
+		nr_cpu_ids /
+		(cpuremap_data.num_threads_per_core * cpuremap_data.num_cores);
+	cpuremap_data.threads = kcalloc(nr_cpu_ids, sizeof(struct cpuremap_cpu), GFP_KERNEL);
+
+	k = cpuremap_data.num_nodes *
+		cpuremap_data.num_threads_per_core *
+		cpuremap_data.num_cores;
+	for (i = 0; i < k; k++)
+		cpuremap_data.threads[i].thread_index = CPUREMAP_NO_THREAD;
+}
+
+int cpuremap_thread_to_cpu(int thread_index)
+{
+	int i, k;
+
+	/* Return NO_CPU if not found */
+	for (i = thread_index, k = 0; k < nr_cpu_ids; k++) {
+		if (cpuremap_data.threads[i].in_use &&
+		    (cpuremap_data.threads[i].thread_index == thread_index)) {
+			cpuremap_data.threads[i].in_use = true;
+			cpuremap_data.threads[i].thread_index = thread_index;
+			return i;
+		}
+		if (i >= nr_cpu_ids)
+			i = 0;
+	}
+	return CPUREMAP_NO_CPU;
+}
+
+int cpuremap_cpu_to_thread(int cpu)
+{
+	/* Return NO_THREAD if not found */
+	if (cpuremap_data.threads[cpu].in_use)
+		return cpuremap_data.threads[cpu].thread_index;
+	return CPUREMAP_NO_THREAD;
+}
+
+int cpuremap_map_cpu(int thread_index, int in_core_ndx, int node)
+{
+	int first_thread, i, k;
+
+	/* Return NO_CPU if fails */
+	first_thread = (node *
+		(cpuremap_data.num_threads_per_core *
+		 cpuremap_data.num_cores)) + in_core_ndx;
+
+	/* Alternative 0: Compressed map of cpus+nodes+threads
+	 *   assuming that no system will be fully built out.
+	 * Alternative 1: Fully compact.  Allocate new cpu ids
+	 *   as needed.  No 'pretty' separation between nodes.
+	 * Alternative 2: Also map incoming nodes from pHyp
+	 *   to virtual nodes for purposes of new cpu ids.
+	 */
+
+	if (first_thread > nr_cpu_ids)
+		first_thread = 0 + in_core_ndx;
+	for (i = first_thread, k = 0; k < nr_cpu_ids; k++) {
+		if (!cpuremap_data.threads[i].in_use || (cpuremap_data.threads[i].thread_index == thread_index)) {
+			cpuremap_data.threads[i].thread_index = thread_index;
+			cpuremap_data.threads[i].node = node;
+			return i;
+		}
+		if (i >= nr_cpu_ids)
+			i = 0;
+	}
+	return CPUREMAP_NO_CPU;
+}
+
+int cpuremap_reserve_cpu(int cpu)
+{
+	if (!cpuremap_data.threads[cpu].in_use) {
+		cpuremap_data.threads[cpu].in_use = true;
+		return cpu;
+	}
+	return CPUREMAP_NO_CPU;
+}
+
+int cpuremap_release_cpu(int cpu)
+{
+	if (cpuremap_data.threads[cpu].in_use) {
+		cpuremap_data.threads[cpu].in_use = false;
+		return cpu;
+	}
+	return CPUREMAP_NO_CPU;
+}
+
+int cpuremap_free_cpu(int cpu)
+{
+	/* Return NO_CPU if fails */
+	if (cpuremap_data.threads[cpu].in_use) {
+		cpuremap_data.threads[cpu].in_use = false;
+		return cpu;
+	}
+	return CPUREMAP_NO_CPU;
+}