checkpoint/restore ping_group_range sysctl

criu/net.c

@@ -2128,6 +2128,79 @@
 }
 #endif
 
+static const char *ipv4_sysctl_entries[] = {
+	"ping_group_range",
+};
+
+#define IPV4_SYSCTL_BASE "net/ipv4"
+#define IPV4_SYSCTL_FMT IPV4_SYSCTL_BASE"/%s"
+#define MAX_IPV4_SYSCTL_OPT 32
+#define MAX_IPV4_SYSCTL_PATH (sizeof(IPV4_SYSCTL_FMT) + MAX_IPV4_SYSCTL_OPT - 2)
+#define MAX_STR_IPV4_SYSCTL_LEN 200
+
+static int ipv4_sysctls_op(SysctlEntry ***rsysctl, size_t *pn, int op)
+{
+	int i, ret = -1, flags = 0;
+	char path[ARRAY_SIZE(ipv4_sysctl_entries)][MAX_IPV4_SYSCTL_PATH] = {};
+	struct sysctl_req req[ARRAY_SIZE(ipv4_sysctl_entries)] = {};
+	SysctlEntry **sysctl = *rsysctl;
+	size_t n = *pn;
+
+	if (n != ARRAY_SIZE(ipv4_sysctl_entries)) {
+		pr_err("unix: Unexpected entries in sysctlig (%zu %zu)\n", n, ARRAY_SIZE(ipv4_sysctl_entries));
+		return -EINVAL;
+	}
+
+	if (opts.weak_sysctls || op == CTL_READ)
+		flags = CTL_FLAGS_OPTIONAL;
+
+	for (i = 0; i < n; i++) {
+		snprintf(path[i], MAX_IPV4_SYSCTL_PATH, IPV4_SYSCTL_FMT, ipv4_sysctl_entries[i]);
+		req[i].name = path[i];
+		req[i].flags = flags;
+
+		switch (sysctl[i]->type) {
+		case SYSCTL_TYPE__CTL_STR:
+			req[i].type = CTL_STR(MAX_STR_IPV4_SYSCTL_LEN);
+
+			/* skip write if have no value */
+			if (op == CTL_WRITE && !sysctl[i]->sarg)
+				continue;
+
+			req[i].arg = sysctl[i]->sarg;
+			break;
+		default:
+			pr_err("ipv4: Unknown sysctl type %d\n", sysctl[i]->type);
+			return -1;
+		}
+	}
+
+	ret = sysctl_op(req, n, op, CLONE_NEWNET);
+	if (ret < 0) {
+		pr_err("unix: Failed to %s %s/<sysctls>\n", (op == CTL_READ) ? "read" : "write", IPV4_SYSCTL_BASE);
+		return -1;
+	}
+
+	if (op == CTL_READ) {
+		bool has_entries = false;
+
+		for (i = 0; i < n; i++) {
+			if (req[i].flags & CTL_FLAGS_HAS) {
+				sysctl[i]->has_iarg = true;
+				if (!has_entries)
+					has_entries = true;
+			}
+		}
+
+		if (!has_entries) {
+			*pn = 0;
+			*rsysctl = NULL;
+		}
+	}
+
+	return 0;
+}
+
 static int dump_netns_conf(struct ns_id *ns, struct cr_imgset *fds)
 {
 	void *buf, *o_buf;
@@ -2142,17 +2215,26 @@
 	int size6 = ARRAY_SIZE(devconfs6);
 	char def_stable_secret[MAX_STR_CONF_LEN + 1] = {};
 	char all_stable_secret[MAX_STR_CONF_LEN + 1] = {};
+	SysctlEntry *ipv4_sysctls = NULL;
+	size_t ipv4_sysctl_size = ARRAY_SIZE(ipv4_sysctl_entries);
+	char ping_group_range[MAX_STR_IPV4_SYSCTL_LEN + 1] = {};
 	NetnsId *ids;
 	struct netns_id *p;
 
 	i = 0;
 	list_for_each_entry(p, &ns->net.ids, node)
 		i++;
 
+	/*
+	 * Here we allocate one single big buffer for storing multiple arrays
+	 * of protobuf entries and pointers to entries in it and we later use
+	 * xptr_pull_s to claim a part of this buffer of proper size for each
+	 * particular array. Next we read data from sysctl files to those
+	 * arrays and then finally save them into images.
+	 */
 	o_buf = buf = xmalloc(i * (sizeof(NetnsId *) + sizeof(NetnsId)) +
-			      size4 * (sizeof(SysctlEntry *) + sizeof(SysctlEntry)) * 2 +
-			      size6 * (sizeof(SysctlEntry *) + sizeof(SysctlEntry)) * 2 +
-			      sizex * (sizeof(SysctlEntry *) + sizeof(SysctlEntry)));
+			      (2 * size4 + 2 * size6 + sizex + ipv4_sysctl_size) *
+			      (sizeof(SysctlEntry *) + sizeof(SysctlEntry)));
 	if (!buf)
 		goto out;
 
@@ -2217,6 +2299,21 @@
 		netns.unix_conf[i]->type = SYSCTL_TYPE__CTL_32;
 	}
 
+	netns.n_ipv4_sysctl = ipv4_sysctl_size;
+	netns.ipv4_sysctl = xptr_pull_s(&buf, ipv4_sysctl_size * sizeof(SysctlEntry *));
+	ipv4_sysctls = xptr_pull_s(&buf, ipv4_sysctl_size * sizeof(SysctlEntry));
+	for (i = 0; i < ipv4_sysctl_size; i++) {
+		sysctl_entry__init(&ipv4_sysctls[i]);
+		netns.ipv4_sysctl[i] = &ipv4_sysctls[i];
+		if (!strcmp(ipv4_sysctl_entries[i], "ping_group_range")) {
+			netns.ipv4_sysctl[i]->type = SYSCTL_TYPE__CTL_STR;
+			netns.ipv4_sysctl[i]->sarg = ping_group_range;
+		} else {
+			/* Need to handle this case when we have more sysctls */
+			BUG();
+		}
+	}
+
 	ret = ipv4_conf_op("default", netns.def_conf4, size4, CTL_READ, NULL);
 	if (ret < 0)
 		goto err_free;
@@ -2235,6 +2332,10 @@
 	if (ret < 0)
 		goto err_free;
 
+	ret = ipv4_sysctls_op(&netns.ipv4_sysctl, &netns.n_ipv4_sysctl, CTL_READ);
+	if (ret < 0)
+		goto err_free;
+
 	ret = pb_write_one(img_from_set(fds, CR_FD_NETNS), &netns, PB_NETNS);
 err_free:
 	xfree(o_buf);
@@ -2587,6 +2688,12 @@
 			goto out;
 	}
 
+	if ((netns)->ipv4_sysctl) {
+		ret = ipv4_sysctls_op(&(netns)->ipv4_sysctl, &(netns)->n_ipv4_sysctl, CTL_WRITE);
+		if (ret)
+			goto out;
+	}
+
 	ns->net.netns = netns;
 out:
 	return ret;

images/netdev.proto

@@ -74,4 +74,5 @@ message netns_entry {
 	repeated netns_id nsids		= 7;
 	optional string	ext_key		= 8;
 	repeated sysctl_entry unix_conf	= 9;
+	repeated sysctl_entry ipv4_sysctl = 10;
 }

test/zdtm/lib/sysctl.c

@@ -3,6 +3,49 @@
 #include "zdtmtst.h"
 #include "sysctl.h"
 
+int sysctl_read_str(const char *name, char *data, size_t size)
+{
+	int fd, ret;
+
+	fd = open(name, O_RDONLY);
+	if (fd < 0) {
+		pr_perror("Can't open %s", name);
+		return -1;
+	}
+
+	ret = read(fd, data, size - 1);
+	if (ret < 0) {
+		pr_perror("Can't read %s", name);
+		close(fd);
+		return -1;
+	}
+	data[ret] = '\0';
+	close(fd);
+
+	return 0;
+}
+
+int sysctl_write_str(const char *name, char *data)
+{
+	int fd, ret;
+
+	fd = open(name, O_WRONLY);
+	if (fd < 0) {
+		pr_perror("Can't open %s", name);
+		return -1;
+	}
+
+	ret = write(fd, data, strlen(data));
+	if (ret < 0) {
+		pr_perror("Can't write %s into %s", data, name);
+		close(fd);
+		return -1;
+	}
+	close(fd);
+
+	return 0;
+}
+
 int sysctl_read_int(const char *name, int *data)
 {
 	int fd;

test/zdtm/lib/sysctl.h

@@ -3,5 +3,7 @@
 
 extern int sysctl_read_int(const char *name, int *data);
 extern int sysctl_write_int(const char *name, int val);
+extern int sysctl_read_str(const char *name, char *data, size_t size);
+extern int sysctl_write_str(const char *name, char *data);
 
 #endif

test/zdtm/static/netns_sub_sysctl.c

@@ -3,18 +3,33 @@
 #include "zdtmtst.h"
 #include "sysctl.h"
 
-const char *test_doc = "Check dump and restore a net.unix.max_dgram_qlen sysctl parameter in subns";
+const char *test_doc = "Check dump and restore of sysctls in subns";
 const char *test_author = "Alexander Mikhalitsyn <alexander@mihalicyn.com>";
 
+#define MAX_STR_SYSCTL_LEN 200
+
+enum {
+	SYSCTL_INT,
+	SYSCTL_STR,
+};
+
 typedef struct {
 	const char *path;
+	int type;
 	int old;
 	int new;
+	char s_old[MAX_STR_SYSCTL_LEN];
+	char s_new[MAX_STR_SYSCTL_LEN];
 } sysctl_opt_t;
 
 #define CONF_UNIX_BASE "/proc/sys/net/unix"
+#define IPV4_SYSCTL_BASE "/proc/sys/net/ipv4"
 
-static sysctl_opt_t net_unix_params[] = { { CONF_UNIX_BASE "/max_dgram_qlen", 0, 0 }, { NULL, 0, 0 } };
+static sysctl_opt_t net_unix_params[] = {
+	{CONF_UNIX_BASE "/max_dgram_qlen", SYSCTL_INT},
+	{IPV4_SYSCTL_BASE "/ping_group_range", SYSCTL_STR, 0, 0, "40000\t50000\n"},
+	{NULL, 0, 0}
+};
 
 int main(int argc, char **argv)
 {
@@ -23,24 +38,43 @@
 	test_init(argc, argv);
 
 	for (p = net_unix_params; p->path != NULL; p++) {
-		p->old = (((unsigned)lrand48()) % 1023) + 1;
-		if (sysctl_write_int(p->path, p->old)) {
-			pr_perror("Can't change %s", p->path);
-			return -1;
+		if (p->type == SYSCTL_INT) {
+			p->old = (((unsigned)lrand48()) % 1023) + 1;
+			if (sysctl_write_int(p->path, p->old)) {
+				pr_perror("Can't change %s", p->path);
+				return -1;
+			}
+		} else if (p->type == SYSCTL_STR) {
+			if (sysctl_write_str(p->path, p->s_old)) {
+				pr_perror("Can't change %s", p->path);
+				return -1;
+			}
 		}
 	}
 
 	test_daemon();
 	test_waitsig();
 
 	for (p = net_unix_params; p->path != NULL; p++) {
-		if (sysctl_read_int(p->path, &p->new))
-			ret = 1;
+		if (p->type == SYSCTL_INT) {
+			if (sysctl_read_int(p->path, &p->new))
+				ret = 1;
 
-		if (p->old != p->new) {
-			errno = EINVAL;
-			pr_perror("%s changed: %d ---> %d", p->path, p->old, p->new);
-			ret = 1;
+			if (p->old != p->new) {
+				errno = EINVAL;
+				pr_perror("%s changed: %d ---> %d", p->path, p->old, p->new);
+				ret = 1;
+			}
+		} else if (p->type == SYSCTL_STR) {
+			if (sysctl_read_str(p->path, p->s_new, MAX_STR_SYSCTL_LEN)) {
+				ret = 1;
+			} else {
+				if (strcmp(p->s_old, p->s_new)) {
+					errno = EINVAL;
+					pr_perror("%s changed: %s ---> %s", p->path, p->s_old, p->s_new);
+					ret = 1;
+				}
+			}
 		}
 	}