mcexec_sys_umount(): remove debug print

lseek() on directories under /sys filesystem that are part of an
overlayed filesystem behave differently than in the original /sys.
This causes segfault in libnuma when discovering topology
information. The patch fakes return value as it is supposed to be,
which also fixes the Intel MPI 2017 MPI_Init() crash.

arch/x86/tools/mcreboot-smp-x86.sh.in

@@ -396,12 +396,15 @@ if [ "$enable_mcoverlay" == "yes" ]; then
 	fi
 	# TODO: How de we revert this in case of failure??
 	mount --make-rprivate /sys
+
 	rm -rf /tmp/mcos/mcos0_sys/setup_complete
 
 	# Hide NUMA related files which are outside the LWK partition
 	for cpuid in `find /sys/devices/system/cpu/* -maxdepth 0 -name "cpu[0123456789]*" -printf "%f "`; do
 		if [ ! -e "/sys/devices/virtual/mcos/mcos0/sys/devices/system/cpu/$cpuid" ]; then
 			rm -rf /tmp/mcos/mcos0_sys/devices/system/cpu/$cpuid
+			rm -rf /tmp/mcos/mcos0_sys/bus/cpu/devices/$cpuid
+			rm -rf /tmp/mcos/mcos0_sys/bus/cpu/drivers/processor/$cpuid
 		else
 			for nodeid in `find /sys/devices/system/cpu/$cpuid/* -maxdepth 0 -name "node[0123456789]*" -printf "%f "`; do
 				if [ ! -e "/sys/devices/virtual/mcos/mcos0/sys/devices/system/cpu/$cpuid/$nodeid" ]; then
@@ -412,7 +415,8 @@ if [ "$enable_mcoverlay" == "yes" ]; then
 	done
 	for nodeid in `find /sys/devices/system/node/* -maxdepth 0 -name "node[0123456789]*" -printf "%f "`; do
 		if [ ! -e "/sys/devices/virtual/mcos/mcos0/sys/devices/system/node/$nodeid" ]; then
-			rm -rf /tmp/mcos/mcos0_sys/devices/system/node/$nodeid
+			rm -rf /tmp/mcos/mcos0_sys/devices/system/node/$nodeid/*
+			rm -rf /tmp/mcos/mcos0_sys/bus/node/devices/$nodeid
 		else
 			# Delete non-existent symlinks
 			for cpuid in `find /sys/devices/system/node/$nodeid/* -maxdepth 0 -name "cpu[0123456789]*" -printf "%f "`; do
@@ -424,6 +428,7 @@ if [ "$enable_mcoverlay" == "yes" ]; then
 			rm -f /tmp/mcos/mcos0_sys/devices/system/node/$nodeid/memory*
 		fi
 	done
+	rm -f /tmp/mcos/mcos0_sys/devices/system/node/has_*
 	for cpuid in `find /sys/bus/cpu/devices/* -maxdepth 0 -name "cpu[0123456789]*" -printf "%f "`; do
 		if [ ! -e "/sys/devices/virtual/mcos/mcos0/sys/bus/cpu/devices/$cpuid" ]; then
 			rm -rf /tmp/mcos/mcos0_sys/bus/cpu/devices/$cpuid

configure

@@ -3117,6 +3117,31 @@ _ACEOF
   fi
 
 
+  { $as_echo "$as_me:${as_lineno-$LINENO}: checking System.map for symbol sys_umount" >&5
+$as_echo_n "checking System.map for symbol sys_umount... " >&6; }
+  mcctrl_addr=`eval $MCCTRL_LINUX_SYMTAB_CMD | grep " sys_umount\$" | cut -d\  -f1`
+  if test -z $mcctrl_addr; then
+    { $as_echo "$as_me:${as_lineno-$LINENO}: result: not found" >&5
+$as_echo "not found" >&6; }
+  else
+    mcctrl_result=$mcctrl_addr
+    mcctrl_addr="0x$mcctrl_addr"
+
+      if `eval $MCCTRL_LINUX_SYMTAB_CMD | grep " __ksymtab_sys_umount\$" >/dev/null`; then
+        mcctrl_result="exported"
+        mcctrl_addr="0"
+      fi
+
+    { $as_echo "$as_me:${as_lineno-$LINENO}: result: $mcctrl_result" >&5
+$as_echo "$mcctrl_result" >&6; }
+
+cat >>confdefs.h <<_ACEOF
+#define MCCTRL_KSYM_sys_umount $mcctrl_addr
+_ACEOF
+
+  fi
+
+
   { $as_echo "$as_me:${as_lineno-$LINENO}: checking System.map for symbol sys_unshare" >&5
 $as_echo_n "checking System.map for symbol sys_unshare... " >&6; }
   mcctrl_addr=`eval $MCCTRL_LINUX_SYMTAB_CMD | grep " sys_unshare\$" | cut -d\  -f1`

configure.ac

@@ -221,6 +221,7 @@ AC_DEFUN([MCCTRL_FIND_KSYM],[
 ])
 
 MCCTRL_FIND_KSYM([sys_mount])
+MCCTRL_FIND_KSYM([sys_umount])
 MCCTRL_FIND_KSYM([sys_unshare])
 MCCTRL_FIND_KSYM([zap_page_range])
 MCCTRL_FIND_KSYM([vdso_image_64])

executer/config.h.in

@@ -51,6 +51,9 @@
 /* Define to address of kernel symbol sys_readlink, or 0 if exported */
 #undef MCCTRL_KSYM_sys_readlink
 
+/* Define to address of kernel symbol sys_umount, or 0 if exported */
+#undef MCCTRL_KSYM_sys_umount
+
 /* Define to address of kernel symbol sys_unshare, or 0 if exported */
 #undef MCCTRL_KSYM_sys_unshare
 

executer/include/uprotocol.h

@@ -41,6 +41,7 @@
 #define MCEXEC_UP_NEW_PROCESS	 0x30a02909
 #define MCEXEC_UP_GET_CRED	 0x30a0290a
 #define MCEXEC_UP_GET_CREDV	 0x30a0290b
+#define MCEXEC_UP_GET_NODES  0x30a0290c
 
 #define MCEXEC_UP_PREPARE_DMA    0x30a02910
 #define MCEXEC_UP_FREE_DMA       0x30a02911
@@ -49,7 +50,8 @@
 #define MCEXEC_UP_CLOSE_EXEC     0x30a02913
 
 #define MCEXEC_UP_SYS_MOUNT      0x30a02914
-#define MCEXEC_UP_SYS_UNSHARE    0x30a02915
+#define MCEXEC_UP_SYS_UMOUNT     0x30a02915
+#define MCEXEC_UP_SYS_UNSHARE    0x30a02916
 
 #define MCEXEC_UP_DEBUG_LOG     0x40000000
 
@@ -196,6 +198,10 @@ struct sys_mount_desc {
 	void *data;
 };
 
+struct sys_umount_desc {
+	char *dir_name;
+};
+
 struct sys_unshare_desc {
 	unsigned long unshare_flags;
 };

executer/kernel/mcctrl/control.c

@@ -66,7 +66,18 @@ int (*mcctrl_sys_mount)(char *dev_name,char *dir_name, char *type, unsigned long
         (int_star_fn_char_char_char_ulong_void_t)
         MCCTRL_KSYM_sys_mount;
 #else // exported
-int (*mcctrl_sys_mount)(char *dev_name,char *dir_name, char *type, unsigned long flags, void *data) = NULL;
+int (*mcctrl_sys_mount)(char *dev_name,char *dir_name, char *type, unsigned long flags, void *data) = sys_mount;
+#endif
+#endif
+
+#ifdef MCCTRL_KSYM_sys_umount
+#if MCCTRL_KSYM_sys_umount
+typedef int (*int_fn_char_star_int_t)(char *, int);
+int (*mcctrl_sys_umount)(char *dir_name, int flags) =
+        (int_fn_char_star_int_t)
+        MCCTRL_KSYM_sys_umount;
+#else // exported
+int (*mcctrl_sys_umount)(char *dir_name, int flags) = sys_umount;
 #endif
 #endif
 
@@ -439,6 +450,16 @@ static long mcexec_get_cpu(ihk_os_t os)
 	return info->n_cpus;
 }
 
+static long mcexec_get_nodes(ihk_os_t os)
+{
+	struct mcctrl_usrdata *usrdata = ihk_host_os_get_usrdata(os);
+
+	if (!usrdata || !usrdata->mem_info)
+		return -EINVAL;
+
+	return usrdata->mem_info->n_numa_nodes;
+}
+
 int mcctrl_add_per_proc_data(struct mcctrl_usrdata *ud, int pid, 
 	struct mcctrl_per_proc_data *ppd)
 {
@@ -1154,7 +1175,7 @@ long mcexec_sys_mount(struct sys_mount_desc *__user arg)
 	cap_raise(promoted->cap_effective, CAP_SYS_ADMIN);
 	original = override_creds(promoted);
 
-#if MCCTRL_KSYM_sys_mount
+#ifdef MCCTRL_KSYM_sys_mount
 	ret = mcctrl_sys_mount(desc.dev_name, desc.dir_name, desc.type,
 		desc.flags, desc.data);
 #else
@@ -1167,6 +1188,36 @@ long mcexec_sys_mount(struct sys_mount_desc *__user arg)
 	return ret;
 }
 
+long mcexec_sys_umount(struct sys_mount_desc *__user arg)
+{
+	struct sys_umount_desc desc;
+	struct cred *promoted;
+	const struct cred *original;
+	int ret;
+
+	if (copy_from_user(&desc, arg, sizeof(desc))) {
+		return -EFAULT;
+	}
+
+	promoted = prepare_creds();
+	if (!promoted) {
+		return -ENOMEM;
+	}
+	cap_raise(promoted->cap_effective, CAP_SYS_ADMIN);
+	original = override_creds(promoted);
+
+#ifdef MCCTRL_KSYM_sys_umount
+	ret = mcctrl_sys_umount(desc.dir_name, MNT_FORCE);
+#else
+	ret = -EFAULT;
+#endif
+
+	revert_creds(original);
+	put_cred(promoted);
+
+	return ret;
+}
+
 long mcexec_sys_unshare(struct sys_unshare_desc *__user arg)
 {
 	struct sys_unshare_desc desc;
@@ -1225,6 +1276,9 @@ long __mcctrl_control(ihk_os_t os, unsigned int req, unsigned long arg,
 	case MCEXEC_UP_GET_CPU:
 		return mcexec_get_cpu(os);
 
+	case MCEXEC_UP_GET_NODES:
+		return mcexec_get_nodes(os);
+
 	case MCEXEC_UP_STRNCPY_FROM_USER:
 		return mcexec_strncpy_from_user(os, 
 				(struct strncpy_from_user_desc *)arg);
@@ -1254,6 +1308,9 @@ long __mcctrl_control(ihk_os_t os, unsigned int req, unsigned long arg,
 	case MCEXEC_UP_SYS_MOUNT:
 		return mcexec_sys_mount((struct sys_mount_desc *)arg);
 
+	case MCEXEC_UP_SYS_UMOUNT:
+		return mcexec_sys_umount((struct sys_mount_desc *)arg);
+
 	case MCEXEC_UP_SYS_UNSHARE:
 		return mcexec_sys_unshare((struct sys_unshare_desc *)arg);
 

executer/kernel/mcctrl/driver.c

@@ -60,6 +60,7 @@ static struct ihk_os_user_call_handler mcctrl_uchs[] = {
 	{ .request = MCEXEC_UP_LOAD_SYSCALL, .func = mcctrl_ioctl },
 	{ .request = MCEXEC_UP_SEND_SIGNAL, .func = mcctrl_ioctl },
 	{ .request = MCEXEC_UP_GET_CPU, .func = mcctrl_ioctl },
+	{ .request = MCEXEC_UP_GET_NODES, .func = mcctrl_ioctl },
 	{ .request = MCEXEC_UP_STRNCPY_FROM_USER, .func = mcctrl_ioctl },
 	{ .request = MCEXEC_UP_NEW_PROCESS, .func = mcctrl_ioctl },
 	{ .request = MCEXEC_UP_PREPARE_DMA, .func = mcctrl_ioctl },
@@ -69,6 +70,7 @@ static struct ihk_os_user_call_handler mcctrl_uchs[] = {
 	{ .request = MCEXEC_UP_GET_CRED, .func = mcctrl_ioctl },
 	{ .request = MCEXEC_UP_GET_CREDV, .func = mcctrl_ioctl },
 	{ .request = MCEXEC_UP_SYS_MOUNT, .func = mcctrl_ioctl },
+	{ .request = MCEXEC_UP_SYS_UMOUNT, .func = mcctrl_ioctl },
 	{ .request = MCEXEC_UP_SYS_UNSHARE, .func = mcctrl_ioctl },
 	{ .request = MCEXEC_UP_DEBUG_LOG, .func = mcctrl_ioctl },
 };

executer/kernel/mcctrl/mcctrl.h

@@ -281,6 +281,7 @@ struct mcctrl_usrdata {
 	unsigned long cpu_online[CPU_LONGS];
 	struct ihk_cpu_info *cpu_info;
 	struct ihk_mem_info *mem_info;
+	nodemask_t numa_online;
 	struct list_head cpu_topology_list;
 	struct list_head node_topology_list;
 };

executer/kernel/mcctrl/sysfs_files.c

@@ -92,27 +92,19 @@ void setup_local_snooping_samples(ihk_os_t os)
 
 void setup_local_snooping_files(ihk_os_t os)
 {
-	struct ihk_cpu_info *info;
 	struct mcctrl_usrdata *udp = ihk_host_os_get_usrdata(os);
 	struct sysfsm_bitmap_param param;
 	static unsigned long cpu_offline = 0x0;
 	int i;
 	int error;
 
-	info = ihk_os_get_cpu_info(os);
-	if (!info) {
-		eprintk("mcctrl:ihk_os_get_cpu_info failed.\n");
-		return;
-	}
-
 	memset(udp->cpu_online, 0, sizeof(udp->cpu_online));
-	for (i = 0; i < info->n_cpus; i++) {
+	for (i = 0; i < udp->cpu_info->n_cpus; i++) {
-		udp->cpu_online[i / BITS_PER_LONG] = 
+		set_bit(i, udp->cpu_online);
-			udp->cpu_online[i / BITS_PER_LONG] | (1 << (i % BITS_PER_LONG));
 	}
 
 	param.nbits = CPU_LONGS * BITS_PER_LONG;
-	param.ptr = udp->cpu_online;
+	param.ptr = &udp->cpu_online;
 	dprintk("mcctrl:setup_local_snooping_files: CPU_LONGS=%d, BITS_PER_LONG=%d\n", 
 		CPU_LONGS, BITS_PER_LONG);
 
@@ -666,6 +658,7 @@ out:
 static int setup_node_files(struct mcctrl_usrdata *udp)
 {
 	int error;
+	int node;
 	struct node_topology *p;
 	struct sysfsm_bitmap_param param;
 
@@ -677,10 +670,21 @@ static int setup_node_files(struct mcctrl_usrdata *udp)
 		goto out;
 	}
 
+	memset(&udp->numa_online, 0, sizeof(udp->numa_online));
+	for (node = 0; node < udp->mem_info->n_numa_nodes; ++node) {
+		node_set(node, udp->numa_online);
+	}
+
+	param.nbits = MAX_NUMNODES;
+	param.ptr = &udp->numa_online;
+	sysfsm_createf(udp->os, SYSFS_SNOOPING_OPS_pbl, &param, 0444,
+			"/sys/devices/system/node/online");
+	sysfsm_createf(udp->os, SYSFS_SNOOPING_OPS_pbl, &param, 0444,
+			"/sys/devices/system/node/possible");
+
 	list_for_each_entry(p, &udp->node_topology_list, chain) {
 		struct sysfs_handle handle;
 		int cpu;
-		int node;
 		size_t offset = 0;
 		param.nbits = nr_cpumask_bits;
 		param.ptr = &p->cpumap;
@@ -697,8 +701,6 @@ static int setup_node_files(struct mcctrl_usrdata *udp)
 						mckernel_numa_2_linux_numa(udp, node)
 						));
 		}
-		offset += snprintf(&p->mckernel_numa_distance_s[offset],
-				NODE_DISTANCE_S_SIZE - offset, "%s", "\n");
 
 		sysfsm_createf(udp->os, SYSFS_SNOOPING_OPS_s,
 				p->mckernel_numa_distance_s, 0444,

executer/user/mcexec.c

@@ -41,6 +41,7 @@
 #include <sys/mman.h>
 #include <asm/unistd.h>
 #include <sched.h>
+#include <dirent.h>
 
 #include <termios.h>
 #include <sys/ioctl.h>
@@ -1148,75 +1149,41 @@ void init_worker_threads(int fd)
 
 #ifdef ENABLE_MCOVERLAYFS
 #define READ_BUFSIZE 1024
-static int isunshare(void)
+static int find_mount_prefix(char *prefix)
 {
-	int err = 0;
+	FILE *fp;
-	int ret;
+	char *line = NULL;
-	int fd;
+	size_t len = 0;
+	ssize_t read;
 	char proc_path[PATH_MAX];
-	ssize_t len_read;
+	int ret = 0;
-	char buf_read[READ_BUFSIZE + 1];
-	char *buf_read_off;
-	char *buf_find;
-	char buf_cmp[READ_BUFSIZE + 1];
-	char *buf_cmp_off;
-	ssize_t len_copy;
 
 	snprintf(proc_path, sizeof(proc_path), "/proc/%d/mounts", getpid());
-	fd = open(proc_path, O_RDONLY);
+
-	if (fd < 0) {
+	fp = fopen(proc_path, "r");
-		fprintf(stderr, "Error: Failed to open %s.\n", proc_path);
+	if (fp == NULL) {
 		return -1;
 	}
 
-	buf_cmp_off = buf_cmp;
+	while ((read = getline(&line, &len, fp)) != -1) {
-	while (1) {
+		if (strlen(line) < strlen(prefix))
-		len_read = read(fd, buf_read, READ_BUFSIZE);
+			continue;
-		if (len_read == -1) {
-			fprintf(stderr, "Error: Failed to read.\n");
-			err = -1;
-			break;
-		}
 
-		buf_read_off = buf_read;
+		if (!strncmp(line, prefix, strlen(prefix))) {
-		while (1) {
+			ret = 1;
-			if ((len_read - (buf_read_off - buf_read)) <= 0) {
-				break;
-			}
-			buf_find = memchr(buf_read_off, '\n', 
-				len_read - (buf_read_off - buf_read));
-			if (buf_find) {
-				len_copy = buf_find - buf_read_off;	
-			} else {
-				len_copy = len_read - (buf_read_off - buf_read);
-			}
-			memcpy(buf_cmp_off, buf_read_off, len_copy);
-			*(buf_cmp_off + len_copy) = '\0';
-
-			if (buf_find) {
-				buf_read_off = buf_read_off + len_copy + 1;
-				buf_cmp_off = buf_cmp;
-				ret = strncmp(buf_cmp, "mcoverlay /proc ", 16);
-				if (!ret) {
-					err = 1;
-					break;
-				}
-			} else {
-				buf_read_off = buf_read_off + len_copy;
-				buf_cmp_off = buf_cmp_off + len_copy;
 			break;
 		}
 	}
 
-		if (err == 1 || len_read == 0) {
+	if (line)
-			break;
+		free(line);
-		}
+
+	return ret;
 }
 
-	close(fd);
+static int isunshare(void)
-
+{
-	__dprintf("err=%d\n", err);
+	return find_mount_prefix("mcoverlay /proc ");
-	return err;
 }
 #endif // ENABLE_MCOVERLAYFS
 
@@ -1415,6 +1382,7 @@ int main(int argc, char **argv)
 	if (error == 0) {
 		struct sys_unshare_desc unshare_desc;
 		struct sys_mount_desc mount_desc;
+		struct sys_umount_desc umount_desc;
 
 		memset(&unshare_desc, '\0', sizeof unshare_desc);
 		memset(&mount_desc, '\0', sizeof mount_desc);
@@ -1426,6 +1394,53 @@ int main(int argc, char **argv)
 			return 1;
 		}
 
+		/*
+		 * Umount cgroup filesystems that may expose invalid NUMA
+		 * information
+		 */
+		if (find_mount_prefix("cgroup /sys/fs/cgroup/cpu,cpuacct")) {
+			umount_desc.dir_name = "/sys/fs/cgroup/cpu,cpuacct";
+
+			if (ioctl(fd, MCEXEC_UP_SYS_UMOUNT,
+						(unsigned long)&umount_desc) != 0) {
+				fprintf(stderr,
+						"WARNING: Failed to umount cgroup/cpu,cpuacct. (%s)\n",
+						strerror(errno));
+			}
+		}
+		else if (find_mount_prefix("cgroup /sys/fs/cgroup/cpu")) {
+			umount_desc.dir_name = "/sys/fs/cgroup/cpu";
+
+			if (ioctl(fd, MCEXEC_UP_SYS_UMOUNT,
+						(unsigned long)&umount_desc) != 0) {
+				fprintf(stderr,
+						"WARNING: Failed to umount cgroup/cpu. (%s)\n",
+						strerror(errno));
+			}
+		}
+
+		if (find_mount_prefix("cgroup /sys/fs/cgroup/cpuset")) {
+			umount_desc.dir_name = "/sys/fs/cgroup/cpuset";
+
+			if (ioctl(fd, MCEXEC_UP_SYS_UMOUNT,
+						(unsigned long)&umount_desc) != 0) {
+				fprintf(stderr,
+						"WARNING: Failed to umount cgroup/cpuset. (%s)\n",
+						strerror(errno));
+			}
+		}
+
+		if (find_mount_prefix("cgroup /sys/fs/cgroup/memory")) {
+			umount_desc.dir_name = "/sys/fs/cgroup/memory/";
+
+			if (ioctl(fd, MCEXEC_UP_SYS_UMOUNT,
+						(unsigned long)&umount_desc) != 0) {
+				fprintf(stderr,
+						"WARNING: Failed to umount cgroup/memory. (%s)\n",
+						strerror(errno));
+			}
+		}
+
 		sprintf(mcos_procdir, "/tmp/mcos/mcos%d_proc", mcosid);
 		mount_desc.dev_name = mcos_procdir;
 		mount_desc.dir_name = "/proc";
@@ -1686,6 +1701,97 @@ do_generic_syscall(
 		ret = -errno;
 	}
 
+	/* Overlayfs /sys/X directory lseek() problem work around */
+	if (w->sr.number == __NR_lseek && ret == -EINVAL) {
+		char proc_path[512];
+		char path[512];
+		struct stat sb;
+
+		sprintf(proc_path, "/proc/self/fd/%d", (int)w->sr.args[0]);
+
+		/* Get filename */
+		if (readlink(proc_path, path, sizeof(path)) < 0) {
+			fprintf(stderr, "%s: error: readlink() failed for %s\n",
+				__FUNCTION__, proc_path);
+			goto out;
+		}
+
+		/* Not in /sys? */
+		if (strncmp(path, "/sys/", 5))
+			goto out;
+
+		/* Stat */
+		if (stat(path, &sb) < 0) {
+			fprintf(stderr, "%s: error stat() failed for %s\n",
+				__FUNCTION__, path);
+			goto out;
+		}
+
+		/* Not dir? */
+		if ((sb.st_mode & S_IFMT) != S_IFDIR)
+			goto out;
+
+		ret = 0;
+	}
+	/* Fake that nodeX in /sys/devices/system/node do not exist,
+	 * where X >= number of LWK NUMA nodes */
+	else if (w->sr.number == __NR_getdents && ret > 0) {
+		struct linux_dirent {
+			long           d_ino;
+			off_t          d_off;
+			unsigned short d_reclen;
+			char           d_name[];
+		};
+		struct linux_dirent *d;
+		char *buf = (char *)w->sr.args[1];
+		int bpos = 0;
+		int nodes,len;
+		char proc_path[PATH_MAX];
+		char path[PATH_MAX];
+
+		sprintf(proc_path, "/proc/self/fd/%d", (int)w->sr.args[0]);
+
+		/* Get filename */
+		if ((len = readlink(proc_path, path, sizeof(path))) < 0) {
+			fprintf(stderr, "%s: error: readlink() failed for %s\n",
+				__FUNCTION__, proc_path);
+			goto out;
+		}
+		path[len] = 0;
+
+		/* Not /sys/devices/system/node ? */
+		if (strcmp(path, "/sys/devices/system/node"))
+			goto out;
+
+		nodes = ioctl(fd, MCEXEC_UP_GET_NODES, 0);
+		if (nodes == -1) {
+			goto out;
+		}
+
+		d = (struct linux_dirent *) (buf + bpos);
+		for (bpos = 0; bpos < ret; ) {
+			int nodeid, tmp_reclen;
+			d = (struct linux_dirent *) (buf + bpos);
+
+			if (sscanf(d->d_name, "node%d", &nodeid) != 1) {
+				bpos += d->d_reclen;
+				continue;
+			}
+
+			if (nodeid >= nodes) {
+				tmp_reclen = d->d_reclen;
+				memmove(buf + bpos,
+						buf + bpos + tmp_reclen,
+						ret - bpos - tmp_reclen);
+				ret -= tmp_reclen;
+				continue;
+			}
+
+			bpos += d->d_reclen;
+		}
+	}
+
+out:
 	__dprintf("do_generic_syscall(%ld):%ld (%#lx)\n", w->sr.number, ret, ret);
 	return ret;
 }

kernel/ap.c

@@ -209,7 +209,7 @@ cpu_sysfs_setup(void)
 	/* setup table */
 	info = kmalloc(sizeof(*info) * num_processors, IHK_MC_AP_CRITICAL);
 	for (cpu = 0; cpu < num_processors; ++cpu) {
-		info[cpu].online = 10+cpu;
+		info[cpu].online = 1;
 	}
 	fake_cpu_infos = info;
 

kernel/include/process.h

@@ -369,6 +369,13 @@ struct vm_range {
 	int padding;
 };
 
+struct vm_range_numa_policy {
+	struct list_head list;
+	unsigned long start, end;
+	DECLARE_BITMAP(numa_mask, PROCESS_NUMA_MASK_BITS);
+	int numa_mem_policy;
+};
+
 struct vm_regions {
 	unsigned long vm_start, vm_end;
 	unsigned long text_start, text_end;
@@ -660,6 +667,8 @@ struct process_vm {
 	long currss;
 	DECLARE_BITMAP(numa_mask, PROCESS_NUMA_MASK_BITS);
 	int numa_mem_policy;
+	/* Protected by memory_range_lock */
+	struct list_head vm_range_numa_policy_list;
 };
 
 static inline int has_cap_ipc_lock(struct thread *th)

kernel/process.c

@@ -210,6 +210,7 @@ init_process_vm(struct process *owner, struct address_space *asp, struct process
 
 	ihk_atomic_set(&vm->refcount, 1);
 	INIT_LIST_HEAD(&vm->vm_range_list);
+	INIT_LIST_HEAD(&vm->vm_range_numa_policy_list);
 	vm->address_space = asp;
 	vm->proc = owner;
 	vm->exiting = 0;
@@ -2483,6 +2484,7 @@ void sched_init(void)
 	ihk_mc_init_context(&idle_thread->ctx, NULL, idle);
 	ihk_mc_spinlock_init(&idle_thread->vm->memory_range_lock);
 	INIT_LIST_HEAD(&idle_thread->vm->vm_range_list);
+	INIT_LIST_HEAD(&idle_thread->vm->vm_range_numa_policy_list);
 	idle_thread->proc->pid = 0;
 	idle_thread->tid = ihk_mc_get_processor_id();
 

kernel/syscall.c

@@ -7081,7 +7081,332 @@ out:
 
 SYSCALL_DECLARE(mbind)
 {
-	return -ENOSYS;
+	unsigned long addr = ihk_mc_syscall_arg0(ctx);
+	unsigned long len = ihk_mc_syscall_arg1(ctx);
+	int mode = ihk_mc_syscall_arg2(ctx);
+	unsigned long *nodemask =
+		(unsigned long *)ihk_mc_syscall_arg3(ctx);
+	unsigned long maxnode = ihk_mc_syscall_arg4(ctx);
+	unsigned flags = ihk_mc_syscall_arg5(ctx);
+	struct process_vm *vm = cpu_local_var(current)->vm;
+	unsigned long nodemask_bits = 0;
+	int mode_flags = 0;
+	int error = 0;
+	int bit;
+	struct vm_range *range;
+	struct vm_range_numa_policy *range_policy, *range_policy_iter;
+	struct vm_range_numa_policy *range_policy_next = NULL;
+	DECLARE_BITMAP(numa_mask, PROCESS_NUMA_MASK_BITS);
+
+	/* Validate arguments */
+	if (addr & ~PAGE_MASK) {
+		return -EINVAL;
+	}
+
+	len = (len + PAGE_SIZE - 1) & PAGE_MASK;
+	if (addr + len < addr || addr == (addr + len)) {
+		return -EINVAL;
+	}
+
+	memset(numa_mask, 0, sizeof(numa_mask));
+
+	if (maxnode) {
+		nodemask_bits = ALIGN(maxnode, 8);
+		if (maxnode > (PAGE_SIZE << 3)) {
+			dkprintf("%s: ERROR: nodemask_bits bigger than PAGE_SIZE bits\n",
+				__FUNCTION__);
+			error = -EINVAL;
+			goto out;
+		}
+
+		if (nodemask_bits > PROCESS_NUMA_MASK_BITS) {
+			dkprintf("%s: WARNING: process NUMA mask bits is insufficient\n",
+				__FUNCTION__);
+			nodemask_bits = PROCESS_NUMA_MASK_BITS;
+		}
+	}
+
+	if ((mode & MPOL_F_STATIC_NODES) && (mode & MPOL_F_RELATIVE_NODES)) {
+		dkprintf("%s: error: MPOL_F_STATIC_NODES & MPOL_F_RELATIVE_NODES\n",
+				__FUNCTION__);
+		error = -EINVAL;
+		goto out;
+	}
+
+	if ((flags & MPOL_MF_STRICT) && (flags & MPOL_MF_MOVE)) {
+		dkprintf("%s: error: MPOL_MF_STRICT & MPOL_MF_MOVE\n",
+				__FUNCTION__);
+		/*
+		 * XXX: man page claims the correct error code is EIO,
+		 * but LTP tests for EINVAL.
+		 */
+		error = -EINVAL;
+		goto out;
+	}
+
+	mode_flags = (mode & (MPOL_F_STATIC_NODES | MPOL_F_RELATIVE_NODES));
+	mode &= ~(MPOL_F_STATIC_NODES | MPOL_F_RELATIVE_NODES);
+
+	if (mode_flags & MPOL_F_RELATIVE_NODES) {
+		/* Not supported.. */
+		dkprintf("%s: error: MPOL_F_RELATIVE_NODES not supported\n",
+				__FUNCTION__);
+		error = -EINVAL;
+		goto out;
+	}
+
+	switch (mode) {
+		case MPOL_DEFAULT:
+			if (nodemask && nodemask_bits) {
+				error = copy_from_user(numa_mask, nodemask,
+						(nodemask_bits >> 3));
+				if (error) {
+					dkprintf("%s: error: copy_from_user numa_mask\n",
+							__FUNCTION__);
+					error = -EFAULT;
+					goto out;
+				}
+
+				if (!bitmap_empty(numa_mask, nodemask_bits)) {
+					dkprintf("%s: ERROR: nodemask not empty for MPOL_DEFAULT\n",
+							__FUNCTION__);
+					error = -EINVAL;
+					goto out;
+				}
+			}
+			break;
+
+		case MPOL_BIND:
+		case MPOL_INTERLEAVE:
+		case MPOL_PREFERRED:
+			/* Special case for MPOL_PREFERRED with empty nodemask */
+			if (mode == MPOL_PREFERRED && !nodemask) {
+				error = 0;
+				break;
+			}
+
+			if (flags & MPOL_MF_STRICT) {
+				error = -EIO;
+				goto out;
+			}
+
+			error = copy_from_user(numa_mask, nodemask,
+					(nodemask_bits >> 3));
+			if (error) {
+				error = -EFAULT;
+				goto out;
+			}
+
+			if (!nodemask || bitmap_empty(numa_mask, nodemask_bits)) {
+				dkprintf("%s: ERROR: nodemask not specified\n",
+						__FUNCTION__);
+				error = -EINVAL;
+				goto out;
+			}
+
+			/* Verify NUMA mask */
+			for_each_set_bit(bit, numa_mask, maxnode) {
+				if (bit >= ihk_mc_get_nr_numa_nodes()) {
+					dkprintf("%s: %d is bigger than # of NUMA nodes\n",
+						__FUNCTION__, bit);
+					error = -EINVAL;
+					goto out;
+				}
+			}
+
+			break;
+
+		default:
+			error = -EINVAL;
+			goto out;
+	}
+
+	/* Validate address range */
+	ihk_mc_spinlock_lock_noirq(&vm->memory_range_lock);
+
+	range = lookup_process_memory_range(vm, addr, addr + len);
+	if (!range) {
+		dkprintf("%s: ERROR: range is invalid\n", __FUNCTION__);
+		error = -EFAULT;
+		goto unlock_out;
+	}
+
+	/* Do the actual policy setting */
+	switch (mode) {
+	/*
+	 * Man page claims MPOL_DEFAULT should remove any range specific
+	 * policies so that process wise policy will be used. LTP on the
+	 * other hand seems to test if MPOL_DEFAULT is set as a range policy.
+	 * MPOL_DEFAULT thus behaves the same as the rest of the policies
+	 * for now.
+	 */
+#if 0
+		case MPOL_DEFAULT:
+			/* Delete or adjust any overlapping range settings */
+			list_for_each_entry_safe(range_policy_iter, range_policy_next,
+					&vm->vm_range_numa_policy_list, list) {
+				int keep = 0;
+				unsigned long orig_end = range_policy_iter->end;
+
+				if (range_policy_iter->end < addr ||
+					range_policy_iter->start > addr + len) {
+					continue;
+				}
+
+				/* Do we need to keep the front? */
+				if (range_policy_iter->start < addr) {
+					range_policy_iter->end = addr;
+					keep = 1;
+				}
+
+				/* Do we need to keep the end? */
+				if (orig_end > addr + len) {
+					/* Are we keeping front already? */
+					if (keep) {
+						/* Add a new entry after */
+						range_policy = kmalloc(sizeof(*range_policy),
+								IHK_MC_AP_NOWAIT);
+						if (!range_policy) {
+							kprintf("%s: error allocating range_policy\n",
+								__FUNCTION__);
+							error = -ENOMEM;
+							goto unlock_out;
+						}
+
+						memcpy(range_policy, range_policy_iter,
+								sizeof(*range_policy));
+						range_policy->start = addr + len;
+						range_policy->end = orig_end;
+						list_add(&range_policy->list,
+								&range_policy_iter->list);
+					}
+					else {
+						range_policy_iter->start = addr + len;
+						keep = 1;
+					}
+				}
+
+				if (!keep) {
+					list_del(&range_policy_iter->list);
+					kfree(range_policy_iter);
+				}
+			}
+
+			break;
+#endif
+		case MPOL_DEFAULT:
+		case MPOL_BIND:
+		case MPOL_INTERLEAVE:
+		case MPOL_PREFERRED:
+			/* Adjust any overlapping range settings and add new one */
+			range_policy_next = NULL;
+			list_for_each_entry(range_policy_iter,
+					&vm->vm_range_numa_policy_list, list) {
+				int adjusted = 0;
+				unsigned long orig_end = range_policy_iter->end;
+
+				if (range_policy_iter->end < addr)
+					continue;
+
+				/* Special case of entirely overlapping */
+				if (range_policy_iter->start == addr &&
+						range_policy_iter->end == addr + len) {
+					range_policy = range_policy_iter;
+					goto mbind_update_only;
+				}
+
+				/* Overlapping partially? */
+				if (range_policy_iter->start < addr) {
+					orig_end = range_policy_iter->end;
+					range_policy_iter->end = addr;
+					adjusted = 1;
+				}
+
+				/* Do we need to keep the end? */
+				if (orig_end > addr + len) {
+					if (adjusted) {
+						/* Add a new entry after */
+						range_policy = kmalloc(sizeof(*range_policy),
+								IHK_MC_AP_NOWAIT);
+						if (!range_policy) {
+							dkprintf("%s: error allocating range_policy\n",
+									__FUNCTION__);
+							error = -ENOMEM;
+							goto unlock_out;
+						}
+
+						memcpy(range_policy, range_policy_iter,
+								sizeof(*range_policy));
+						range_policy->start = addr + len;
+						range_policy->end = orig_end;
+						list_add(&range_policy->list,
+								&range_policy_iter->list);
+						range_policy_next = range_policy;
+						break;
+					}
+					else {
+						range_policy_iter->start = addr + len;
+						range_policy_next = range_policy_iter;
+						break;
+					}
+				}
+
+				/* Next one in ascending address order? */
+				if (range_policy_iter->start >= addr + len) {
+					range_policy_next = range_policy_iter;
+					break;
+				}
+			}
+
+			/* Add a new entry */
+			range_policy = kmalloc(sizeof(*range_policy),
+					IHK_MC_AP_NOWAIT);
+			if (!range_policy) {
+				dkprintf("%s: error allocating range_policy\n",
+						__FUNCTION__);
+				error = -ENOMEM;
+				goto unlock_out;
+			}
+
+			memset(range_policy, 0, sizeof(*range_policy));
+			range_policy->start = addr;
+			range_policy->end = addr + len;
+
+			if (range_policy_next) {
+				list_add_tail(&range_policy->list,
+						&range_policy_next->list);
+			}
+			else {
+				list_add_tail(&range_policy->list,
+						&vm->vm_range_numa_policy_list);
+			}
+
+mbind_update_only:
+			if (mode == MPOL_DEFAULT) {
+				memset(range_policy->numa_mask, 0, sizeof(numa_mask));
+				for (bit = 0; bit < ihk_mc_get_nr_numa_nodes(); ++bit) {
+					set_bit(bit, range_policy->numa_mask);
+				}
+			}
+			else {
+				memcpy(range_policy->numa_mask, &numa_mask,
+					sizeof(numa_mask));
+			}
+			range_policy->numa_mem_policy = mode;
+
+			break;
+
+		default:
+			error = -EINVAL;
+			goto out;
+	}
+
+	error = 0;
+
+unlock_out:
+	ihk_mc_spinlock_unlock_noirq(&vm->memory_range_lock);
+out:
+	return error;
 } /* sys_mbind() */
 
 SYSCALL_DECLARE(set_mempolicy)
@@ -7094,6 +7419,8 @@ SYSCALL_DECLARE(set_mempolicy)
 	struct process_vm *vm = cpu_local_var(current)->vm;
 	int error = 0;
 	int bit, valid_mask;
+	struct vm_range_numa_policy *range_policy_iter;
+	struct vm_range_numa_policy *range_policy_next = NULL;
 	DECLARE_BITMAP(numa_mask, PROCESS_NUMA_MASK_BITS);
 
 	memset(numa_mask, 0, sizeof(numa_mask));
@@ -7108,7 +7435,7 @@ SYSCALL_DECLARE(set_mempolicy)
 		}
 
 		if (nodemask_bits > PROCESS_NUMA_MASK_BITS) {
-			kprintf("%s: WARNING: process NUMA mask bits is insufficient\n",
+			dkprintf("%s: WARNING: process NUMA mask bits is insufficient\n",
 				__FUNCTION__);
 			nodemask_bits = PROCESS_NUMA_MASK_BITS;
 		}
@@ -7137,7 +7464,14 @@ SYSCALL_DECLARE(set_mempolicy)
 				set_bit(bit, vm->numa_mask);
 			}
 
-			/* TODO: delete all mbind() specified regions */
+			/* Delete all range settings */
+			ihk_mc_spinlock_lock_noirq(&vm->memory_range_lock);
+			list_for_each_entry_safe(range_policy_iter, range_policy_next,
+					&vm->vm_range_numa_policy_list, list) {
+				list_del(&range_policy_iter->list);
+				kfree(range_policy_iter);
+			}
+			ihk_mc_spinlock_unlock_noirq(&vm->memory_range_lock);
 
 			vm->numa_mem_policy = mode;
 			error = 0;
@@ -7224,16 +7558,25 @@ SYSCALL_DECLARE(get_mempolicy)
 	unsigned long addr = ihk_mc_syscall_arg3(ctx);
 	unsigned long flags = ihk_mc_syscall_arg4(ctx);
 	struct process_vm *vm = cpu_local_var(current)->vm;
-	int error;
+	struct vm_range_numa_policy *range_policy = NULL;
+	int error = 0;
+	int policy;
 
-	if (((flags & MPOL_F_ADDR) && !addr) ||
+	if ((!(flags & MPOL_F_ADDR) && addr) ||
-		(!(flags & MPOL_F_ADDR) && addr) ||
 		(flags & ~(MPOL_F_ADDR | MPOL_F_NODE | MPOL_F_MEMS_ALLOWED)) ||
 		((flags & MPOL_F_NODE) && !(flags & MPOL_F_ADDR) &&
 		 vm->numa_mem_policy == MPOL_INTERLEAVE)) {
 		return -EINVAL;
 	}
 
+	/*
+	 * XXX: man page claims the correct error code is EINVAL,
+	 * but LTP tests for EFAULT.
+	 */
+	if ((flags & MPOL_F_ADDR) && !addr) {
+		return -EFAULT;
+	}
+
 	if (maxnode) {
 		if (maxnode < ihk_mc_get_nr_numa_nodes()) {
 			return -EINVAL;
@@ -7247,18 +7590,62 @@ SYSCALL_DECLARE(get_mempolicy)
 		}
 	}
 
+	/* Special case of MPOL_F_MEMS_ALLOWED */
+	if (flags == MPOL_F_MEMS_ALLOWED) {
+		if (nodemask) {
+			error = copy_to_user(nodemask,
+					cpu_local_var(current)->vm->numa_mask,
+					(nodemask_bits >> 3));
+			if (error) {
+				error = -EFAULT;
+			}
+		}
+
+		goto out;
+	}
+
+	/* Address range specific? */
+	if (flags & MPOL_F_ADDR) {
+		struct vm_range_numa_policy *range_policy_iter;
+		struct vm_range *range;
+
+		ihk_mc_spinlock_lock_noirq(&vm->memory_range_lock);
+		range = lookup_process_memory_range(vm, addr, addr + 1);
+		if (!range) {
+			dkprintf("%s: ERROR: range is invalid\n", __FUNCTION__);
+			error = -EFAULT;
+			ihk_mc_spinlock_unlock_noirq(&vm->memory_range_lock);
+			goto out;
+		}
+
+		list_for_each_entry(range_policy_iter,
+					&vm->vm_range_numa_policy_list, list) {
+			if (range_policy_iter->start > addr ||
+				range_policy_iter->end <= addr) {
+				continue;
+			}
+
+			range_policy = range_policy_iter;
+			break;
+		}
+		ihk_mc_spinlock_unlock_noirq(&vm->memory_range_lock);
+	}
+
+	/* Return policy */
+	policy = range_policy ? range_policy->numa_mem_policy :
+		vm->numa_mem_policy;
+
 	if (mode) {
-		error = copy_to_user(mode,
+		error = copy_to_user(mode, &policy, sizeof(int));
-				&cpu_local_var(current)->vm->numa_mem_policy,
-				sizeof(int));
 		if (error) {
 			error = -EFAULT;
 			goto out;
 		}
 	}
 
-	if (nodemask) {
+	if (nodemask && (policy != MPOL_DEFAULT)) {
 		error = copy_to_user(nodemask,
+				range_policy ? range_policy->numa_mask :
 				cpu_local_var(current)->vm->numa_mask,
 				(nodemask_bits >> 3));
 		if (error) {