From 2026cf8dad0b948b920352219203efbc71bba886 Mon Sep 17 00:00:00 2001
From: Balazs Gerofi <bgerofi@riken.jp>
Date: Tue, 21 Apr 2020 11:19:43 +0900
Subject: [PATCH] mcexec: explicit CPU list in partitoned execution (for
 Fujitsu's FLIB_AFFINITY_ON_PROCESS)

Change-Id: I05c11f73553de8ccb5f79083ce2115ac57e62584
---
 executer/include/uprotocol.h     |  2 +
 executer/kernel/mcctrl/control.c | 72 ++++++++++++++++++++++++++++++++
 executer/user/mcexec.c           | 18 ++++++++
 3 files changed, 92 insertions(+)

diff --git a/executer/include/uprotocol.h b/executer/include/uprotocol.h
index 81e1dce3..a277162e 100644
--- a/executer/include/uprotocol.h
+++ b/executer/include/uprotocol.h
@@ -91,6 +91,8 @@ struct program_image_section {
 
 struct get_cpu_set_arg {
 	int nr_processes;
+	char *req_cpu_list;   // Requested by user-space
+	int req_cpu_list_len; // Lenght of request string
 	int *process_rank;
 	pid_t ppid;
 	void *cpu_set;
diff --git a/executer/kernel/mcctrl/control.c b/executer/kernel/mcctrl/control.c
index d027f8dc..dac0d0de 100644
--- a/executer/kernel/mcctrl/control.c
+++ b/executer/kernel/mcctrl/control.c
@@ -620,6 +620,78 @@ static long mcexec_get_cpuset(ihk_os_t os, unsigned long arg)
 		goto put_out;
 	}
 
+	/* User requested CPU mask? */
+	if (req.req_cpu_list && req.req_cpu_list_len) {
+		char *cpu_list = NULL;
+
+		cpu_list = kmalloc(req.req_cpu_list_len, GFP_KERNEL);
+		if (!cpu_list) {
+			printk("%s: error: allocating CPU list\n", __FUNCTION__);
+			ret = -ENOMEM;
+			goto put_out;
+		}
+
+		if (copy_from_user(cpu_list,
+					req.req_cpu_list, req.req_cpu_list_len)) {
+			printk("%s: error copying CPU list request\n", __FUNCTION__);
+			kfree(cpu_list);
+			ret = -EINVAL;
+			goto put_out;
+		}
+
+		cpus_used = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
+		cpus_to_use = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
+		if (!cpus_to_use || !cpus_used) {
+			printk("%s: error: allocating CPU mask\n", __FUNCTION__);
+			ret = -ENOMEM;
+			kfree(cpu_list);
+			goto put_out;
+		}
+		memset(cpus_used, 0, sizeof(cpumask_t));
+		memset(cpus_to_use, 0, sizeof(cpumask_t));
+
+		/* Parse CPU list */
+		if (cpulist_parse(cpu_list, cpus_to_use) < 0) {
+			printk("%s: invalid CPUs requested: %s\n",
+				__FUNCTION__, cpu_list);
+			ret = -EINVAL;
+			kfree(cpu_list);
+			goto put_out;
+		}
+
+		memcpy(cpus_used, cpus_to_use, sizeof(cpumask_t));
+
+		/* Copy mask to user-space */
+		if (copy_to_user(req.cpu_set, cpus_used,
+					(req.cpu_set_size < sizeof(cpumask_t) ?
+					 req.cpu_set_size : sizeof(cpumask_t)))) {
+			printk("%s: error copying mask to user\n", __FUNCTION__);
+			ret = -EINVAL;
+			kfree(cpu_list);
+			goto put_out;
+		}
+
+		/* Copy IKC target core */
+		cpu = cpumask_next(-1, cpus_used);
+		if (copy_to_user(req.target_core, &cpu, sizeof(cpu))) {
+			printk("%s: error copying target core to user\n",
+					__FUNCTION__);
+			ret = -EINVAL;
+			kfree(cpu_list);
+			goto put_out;
+		}
+
+		/* Save in per-process structure */
+		memcpy(&ppd->cpu_set, cpus_used, sizeof(cpumask_t));
+		ppd->ikc_target_cpu = cpu;
+		printk("%s: %s -> target McKernel CPU: %d\n",
+			__func__, cpu_list, cpu);
+
+		ret = 0;
+		kfree(cpu_list);
+		goto put_out;
+	}
+
 	mutex_lock(&udp->part_exec_lock);
 	/* Find part_exec having same node_proxy */
 	list_for_each_entry_reverse(pe_itr, &udp->part_exec_list, chain) {
diff --git a/executer/user/mcexec.c b/executer/user/mcexec.c
index 834b2304..4ce21fe7 100644
--- a/executer/user/mcexec.c
+++ b/executer/user/mcexec.c
@@ -2483,6 +2483,8 @@ int main(int argc, char **argv)
 
 		CPU_ZERO(&mcexec_cpu_set);
 
+		cpu_set_arg.req_cpu_list = NULL;
+		cpu_set_arg.req_cpu_list_len = 0;
 		cpu_set_arg.cpu_set = (void *)&desc->cpu_set;
 		cpu_set_arg.cpu_set_size = sizeof(desc->cpu_set);
 		cpu_set_arg.nr_processes = nr_processes;
@@ -2494,6 +2496,16 @@ int main(int argc, char **argv)
 		cpu_set_arg.mcexec_cpu_set_size = sizeof(mcexec_cpu_set);
 		cpu_set_arg.ikc_mapped = &ikc_mapped;
 
+		/* Fugaku specific: Fujitsu CPU binding */
+		if (getenv("FLIB_AFFINITY_ON_PROCESS")) {
+			cpu_set_arg.req_cpu_list =
+				getenv("FLIB_AFFINITY_ON_PROCESS");
+			cpu_set_arg.req_cpu_list_len =
+				strlen(cpu_set_arg.req_cpu_list) + 1;
+			__dprintf("%s: requesting CPUs: %s\n",
+				__func__, cpu_set_arg.req_cpu_list);
+		}
+
 		if (ioctl(fd, MCEXEC_UP_GET_CPUSET, (void *)&cpu_set_arg) != 0) {
 			perror("getting CPU set for partitioned execution");
 			close(fd);
@@ -2502,6 +2514,12 @@ int main(int argc, char **argv)
 
 		desc->cpu = target_core;
 		desc->process_rank = process_rank;
+		/* Fugaku specific: Fujitsu node-local rank */
+		if (getenv("PLE_RANK_ON_NODE")) {
+			desc->process_rank = atoi(getenv("PLE_RANK_ON_NODE"));
+			__dprintf("%s: rank: %d, target CPU: %d\n",
+				__func__, desc->process_rank, desc->cpu);
+		}
 
 		/* Bind to CPU cores where the LWK process' IKC target maps to */
 		if (ikc_mapped && !no_bind_ikc_map) {