Record pthread routine address in clone(), keep helper threads on caller CPU core (workaround for Fugaku)

Change-Id: I29d1589e430dc1396558cdf3df4d068c27173612

arch/arm64/kernel/syscall.c

@@ -57,15 +57,34 @@ extern int num_processors;
 int obtain_clone_cpuid(cpu_set_t *cpu_set, int use_last)
 {
 	int min_queue_len = -1;
-	int cpu, min_cpu = -1, uti_cpu = -1;
+	int cpu, min_cpu = -1;
+#if 0
+	int uti_cpu = -1;
+#endif
 	unsigned long irqstate = 0;
 
+	int start, end, step;
+
+	if (use_last) {
+		start = num_processors - 1;
+		end = -1;
+		step = -1;
+	}
+	else {
+		start = 0;
+		end = num_processors;
+		step = 1;
+	}
+
 	if (!cpu_local_var(current)->proc->nr_processes) {
 		irqstate = ihk_mc_spinlock_lock(&runq_reservation_lock);
 	}
+	else {
+		irqstate = cpu_disable_interrupt_save();
+	}
 
 	/* Find the first allowed core with the shortest run queue */
-	for (cpu = 0; cpu < num_processors; ++cpu) {
+	for (cpu = start; cpu != end; cpu += step) {
 		struct cpu_local_var *v;
 
 		if (!CPU_ISSET(cpu, cpu_set))
@@ -76,11 +95,14 @@ int obtain_clone_cpuid(cpu_set_t *cpu_set, int use_last)
 		dkprintf("%s: cpu=%d,runq_len=%d,runq_reserved=%d\n",
 			 __func__, cpu, v->runq_len, v->runq_reserved);
 		if (min_queue_len == -1 ||
-		    v->runq_len + v->runq_reserved < min_queue_len) {
-			min_queue_len = v->runq_len + v->runq_reserved;
+		    //v->runq_len + v->runq_reserved < min_queue_len) {
+		    v->runq_len < min_queue_len) {
+			//min_queue_len = v->runq_len + v->runq_reserved;
+			min_queue_len = v->runq_len;
 			min_cpu = cpu;
 		}
 
+#if 0
 		/* Record the last tie CPU */
 		if (min_cpu != cpu &&
 		    v->runq_len + v->runq_reserved == min_queue_len) {
@@ -89,14 +111,15 @@ int obtain_clone_cpuid(cpu_set_t *cpu_set, int use_last)
 		dkprintf("%s: cpu=%d,runq_len=%d,runq_reserved=%d,min_cpu=%d,uti_cpu=%d\n",
 			 __func__, cpu, v->runq_len, v->runq_reserved,
 			 min_cpu, uti_cpu);
+#else
 
 		ihk_mc_spinlock_unlock_noirq(&v->runq_lock);
-#if 0
 		if (min_queue_len == 0)
 			break;
 #endif
 	}
 
+#if 0
 	min_cpu = use_last ? uti_cpu : min_cpu;
 	if (min_cpu != -1) {
 		if (get_cpu_local_var(min_cpu)->status != CPU_STATUS_RESERVED)
@@ -105,10 +128,16 @@ int obtain_clone_cpuid(cpu_set_t *cpu_set, int use_last)
 		__sync_fetch_and_add(&get_cpu_local_var(min_cpu)->runq_reserved,
 				     1);
 	}
+#else
+	__sync_fetch_and_add(&get_cpu_local_var(min_cpu)->runq_reserved, 1);
+#endif
 
 	if (!cpu_local_var(current)->proc->nr_processes) {
 		ihk_mc_spinlock_unlock(&runq_reservation_lock, irqstate);
 	}
+	else {
+		cpu_restore_interrupt(irqstate);
+	}
 
 	return min_cpu;
 }

executer/user/libsched_yield.c

@@ -1,6 +1,7 @@
 #define _GNU_SOURCE
 #include <dlfcn.h>
 #include <sys/time.h>
+#include <sys/syscall.h>
 #include <sched.h>
 #include <unistd.h>
 #include <stdlib.h>
@@ -27,3 +28,110 @@ int sched_yield(void)
 
 	return 0;
 }
+
+#undef pthread_create
+
+typedef int (*__pthread_create_fn)(pthread_t *thread,
+		const pthread_attr_t *attr,
+		void *(*start_routine) (void *),
+		void *arg);
+
+static __pthread_create_fn orig_pthread_create = 0;
+
+
+int pthread_create(pthread_t *thread,
+		const pthread_attr_t *attr,
+		void *(*start_routine) (void *),
+		void *arg)
+{
+	if (!orig_pthread_create) {
+		orig_pthread_create =
+			(__pthread_create_fn)dlsym(RTLD_NEXT, "pthread_create");
+	}
+
+	/* CLONE_VM and newsp == parent_tidptr impiles pthread start routine addr */
+	syscall(__NR_clone, CLONE_VM, start_routine, start_routine, 0, 0, 0);
+
+	return orig_pthread_create(thread, attr, start_routine, arg);
+}
+
+#if 0
+#define PROCMAPS_MAX_LEN	131072
+
+char *addr_to_lib(void *addr, unsigned long *offset_in_lib)
+{
+	char maps_path[PATH_MAX];
+	char buf[PROCMAPS_MAX_LEN];
+	int fd;
+	void *start, *end;
+	char perms[4];
+	unsigned long offset;
+	unsigned long dev[2];
+	int inode;
+	char path[PATH_MAX];
+	char *line;
+
+	sprintf(maps_path,"/proc/self/maps");
+	fd = open(maps_path, O_RDONLY);
+	if (fd < 0) {
+		fprintf(stderr,"error: cannot open the memory maps, %s\n",
+				strerror(errno));
+		return NULL;
+	}
+
+	memset(buf, 0, PROCMAPS_MAX_LEN);
+	read(fd, buf, PROCMAPS_MAX_LEN);
+	line = strtok(buf, "\n");
+	while (line) {
+		memset(path, 0, sizeof(path));
+		sscanf(line, "%012lx-%012lx %4s %lx %lx:%lx %d\t\t\t%[^\n]",
+			&start, &end, perms, &offset, &dev[0], &dev[1], &inode, path);
+
+		if (start <= addr && end > addr) {
+			close(fd);
+			if (offset_in_lib)
+				*offset_in_lib = (unsigned long)(addr - start);
+			return strlen(path) > 0 ? strdup(path) : NULL;
+		}
+
+		line = strtok(NULL, "\n");
+	}
+
+	close(fd);
+	return NULL;
+}
+
+int pthread_create(pthread_t *thread,
+		const pthread_attr_t *attr,
+		void *(*start_routine) (void *),
+		void *arg)
+{
+	char *lib = NULL;
+	int util_thread = 1;
+	unsigned long offset;
+
+	if (!orig_pthread_create) {
+		orig_pthread_create =
+			(__pthread_create_fn)dlsym(RTLD_NEXT, "pthread_create");
+	}
+
+	lib = addr_to_lib(start_routine, &offset);
+	if (lib)
+		printf("%s: 0x%lx is in %s @ 0x%lx\n",
+				__func__, start_routine, lib, offset);
+
+	if (lib && (strstr(lib, "iomp") || strstr(lib, "psm"))) {
+		util_thread = 0;
+	}
+
+	if (util_thread) {
+		/* McKernel util_indicate_clone() */
+		syscall(731);
+	}
+
+	if (lib)
+		free(lib);
+
+	return orig_pthread_create(thread, attr, start_routine, arg);
+}
+#endif

kernel/host.c

@@ -500,6 +500,7 @@ static int process_msg_prepare_process(unsigned long rphys)
 		ihk_mc_unmap_memory(NULL, phys, sz);
 		return -ENOMEM;
 	}
+	sprintf(thread->pthread_routine, "%s", "[main]");
 	proc = thread->proc;
 	vm = thread->vm;
 

kernel/include/process.h

@@ -608,6 +608,7 @@ struct thread {
 	// thread info
 	int cpu_id;
 	int tid;
+	char pthread_routine[PATH_MAX + 64];
 	int status;	// PS_RUNNING -> PS_EXITED (-> ZOMBIE / ptrace)
 			// |       ^       ^
 			// |       |       |
@@ -717,6 +718,7 @@ struct thread {
 	/* Syscall offload wait queue head */
 	struct waitq scd_wq;
 
+	unsigned long clone_pthread_start_routine;
 	int uti_state;
 	int mod_clone;
 	struct uti_attr *mod_clone_arg;

kernel/syscall.c

@@ -2757,6 +2757,9 @@ unsigned long do_fork(int clone_flags, unsigned long newsp,
 	const struct ihk_mc_cpu_info *cpu_info = ihk_mc_get_cpu_info();
 #endif
 	int err = 0;
+	unsigned long clone_pthread_start_routine = 0;
+	struct vm_range *range = NULL;
+	int helper_thread = 0;
 
 	dkprintf("%s,flags=%08x,newsp=%lx,ptidptr=%lx,"
 		"ctidptr=%lx,tls=%lx,curpc=%lx,cursp=%lx",
@@ -2766,6 +2769,18 @@ unsigned long do_fork(int clone_flags, unsigned long newsp,
 	dkprintf("do_fork(): stack_pointr passed in: 0x%lX, stack pointer of caller: 0x%lx\n",
 			 newsp, cursp);
 
+	/* CLONE_VM and newsp == parent_tidptr impiles pthread start routine addr */
+	if ((clone_flags & CLONE_VM) && newsp == parent_tidptr) {
+		old->clone_pthread_start_routine = parent_tidptr;
+		dkprintf("%s: clone_pthread_start_routine: 0x%lx\n", __func__,
+			old->clone_pthread_start_routine);
+		return 0;
+	}
+
+	/* Clear pthread routine addr regardless if we succeed */
+	clone_pthread_start_routine = old->clone_pthread_start_routine;
+	old->clone_pthread_start_routine = 0;
+
 	parent_cpuid = old->cpu_id;
 	if (((clone_flags & CLONE_VM) && !(clone_flags & CLONE_THREAD)) ||
 		(!(clone_flags & CLONE_VM) && (clone_flags & CLONE_THREAD))) {
@@ -2821,10 +2836,34 @@ unsigned long do_fork(int clone_flags, unsigned long newsp,
 		}
 	}
 
-	cpuid = obtain_clone_cpuid(&old->cpu_set, old->mod_clone == SPAWN_TO_REMOTE && oldproc->uti_use_last_cpu);
-	if (cpuid == -1) {
-		kprintf("do_fork,core not available\n");
-		return -EAGAIN;
+	if (clone_pthread_start_routine) {
+		ihk_rwspinlock_read_lock_noirq(&old->vm->memory_range_lock);
+		range = lookup_process_memory_range(old->vm,
+				clone_pthread_start_routine,
+				clone_pthread_start_routine + 1);
+		ihk_rwspinlock_read_unlock_noirq(&old->vm->memory_range_lock);
+
+		if (range && range->memobj && range->memobj->path) {
+			if (!strstr(range->memobj->path, "omp.so")) {
+				helper_thread = 1;
+			}
+			dkprintf("clone(): %s thread from %s\n",
+				helper_thread ? "helper" : "compute",
+				range->memobj->path);
+		}
+	}
+
+	if (helper_thread) {
+		cpuid = ihk_mc_get_processor_id();
+		//cpuid = obtain_clone_cpuid(&oldproc->cpu_set, 1);
+	}
+	else {
+		cpuid = obtain_clone_cpuid(&oldproc->cpu_set,
+				(old->mod_clone == SPAWN_TO_REMOTE && oldproc->uti_use_last_cpu));
+		if (cpuid == -1) {
+			kprintf("do_fork,core not available\n");
+			return -EAGAIN;
+		}
 	}
 
 	new = clone_thread(old, curpc,
@@ -2835,6 +2874,17 @@ unsigned long do_fork(int clone_flags, unsigned long newsp,
 		goto release_cpuid;
 	}
 
+	if (clone_pthread_start_routine &&
+		range && range->memobj && range->memobj->path) {
+
+		sprintf(new->pthread_routine, "0x%lx @ %s",
+			clone_pthread_start_routine,
+			range->memobj->path);
+	}
+	else {
+		sprintf(new->pthread_routine, "%s", "[unknown]");
+	}
+
 	newproc = new->proc;
 
 	cpu_set(cpuid, &new->vm->address_space->cpu_set,