criu: Add pidfd support

criu/Makefile.crtools

@@ -99,6 +99,7 @@ obj-$(CONFIG_COMPAT)	+= vdso-compat.o
 CFLAGS_REMOVE_vdso-compat.o	+= $(CFLAGS-ASAN) $(CFLAGS-GCOV)
 obj-y			+= pidfd-store.o
 obj-y			+= hugetlb.o
+obj-y			+= pidfd.o
 
 PROTOBUF_GEN := scripts/protobuf-gen.sh
 

criu/cr-restore.c

@@ -79,6 +79,7 @@
 #include "timens.h"
 #include "bpfmap.h"
 #include "apparmor.h"
+#include "pidfd.h"
 
 #include "parasite-syscall.h"
 #include "files-reg.h"
@@ -279,7 +280,7 @@ static struct collect_image_info *cinfos_files[] = {
 	&unix_sk_cinfo,	      &fifo_cinfo,     &pipe_cinfo,    &nsfile_cinfo,	    &packet_sk_cinfo,
 	&netlink_sk_cinfo,    &eventfd_cinfo,  &epoll_cinfo,   &epoll_tfd_cinfo,    &signalfd_cinfo,
 	&tunfile_cinfo,	      &timerfd_cinfo,  &inotify_cinfo, &inotify_mark_cinfo, &fanotify_cinfo,
-	&fanotify_mark_cinfo, &ext_file_cinfo, &memfd_cinfo,
+	&fanotify_mark_cinfo, &ext_file_cinfo, &memfd_cinfo,   &pidfd_cinfo
 };
 
 /* These images are required to restore namespaces */

criu/files.c

@@ -49,6 +49,7 @@
 #include "kerndat.h"
 #include "fdstore.h"
 #include "bpfmap.h"
+#include "pidfd.h"
 
 #include "protobuf.h"
 #include "util.h"
@@ -544,6 +545,8 @@ static int dump_one_file(struct pid *pid, int fd, int lfd, struct fd_opts *opts,
 			ops = &signalfd_dump_ops;
 		else if (is_timerfd_link(link))
 			ops = &timerfd_dump_ops;
+		else if (is_pidfd_link(link))
+			ops = &pidfd_dump_ops;
 #ifdef CONFIG_HAS_LIBBPF
 		else if (is_bpfmap_link(link))
 			ops = &bpfmap_dump_ops;
@@ -1778,6 +1781,9 @@ static int collect_one_file(void *o, ProtobufCMessage *base, struct cr_img *i)
 	case FD_TYPES__MEMFD:
 		ret = collect_one_file_entry(fe, fe->memfd->id, &fe->memfd->base, &memfd_cinfo);
 		break;
+	case FD_TYPES__PIDFD:
+		ret = collect_one_file_entry(fe, fe->pidfd->id, &fe->pidfd->base, &pidfd_cinfo);
+		break;
 #ifdef CONFIG_HAS_LIBBPF
 	case FD_TYPES__BPFMAP:
 		ret = collect_one_file_entry(fe, fe->bpf->id, &fe->bpf->base, &bpfmap_cinfo);

criu/image-desc.c

@@ -107,6 +107,7 @@ struct cr_fd_desc_tmpl imgset_template[CR_FD_MAX] = {
 	FD_ENTRY_F(BPFMAP_FILE,	"bpfmap-file", O_NOBUF),
 	FD_ENTRY_F(BPFMAP_DATA,	"bpfmap-data", O_NOBUF),
 	FD_ENTRY(APPARMOR,	"apparmor"),
+	FD_ENTRY(PIDFD,         "pidfd-%u"),
 
 	[CR_FD_STATS] = {
 		.fmt	= "stats-%s",

criu/include/fdinfo.h

@@ -8,6 +8,7 @@
 #include "images/signalfd.pb-c.h"
 #include "images/fsnotify.pb-c.h"
 #include "images/timerfd.pb-c.h"
+#include "images/pidfd.pb-c.h"
 
 struct fdinfo_common {
 	off64_t pos;

criu/include/image-desc.h

@@ -113,6 +113,7 @@ enum {
 	CR_FD_PIPES,
 	CR_FD_TTY_FILES,
 	CR_FD_MEMFD_FILE,
+	CR_FD_PIDFD,
 
 	CR_FD_AUTOFS,
 

criu/include/magic.h

@@ -100,6 +100,7 @@
 #define BPFMAP_FILE_MAGIC    0x57506142 /* Alapayevsk */
 #define BPFMAP_DATA_MAGIC    0x64324033 /* Arkhangelsk */
 #define APPARMOR_MAGIC	     0x59423047 /* Nikolskoye */
+#define PIDFD_MAGIC	     0x52253735 /* Livny */
 
 #define IFADDR_MAGIC	RAW_IMAGE_MAGIC
 #define ROUTE_MAGIC	RAW_IMAGE_MAGIC

criu/include/pidfd.h

@@ -0,0 +1,19 @@
+#ifndef __CR_PIDFD_H__
+#define __CR_PIDFD_H__
+
+#include <sys/stat.h>
+#include "images/pidfd.pb-c.h"
+#include <sys/syscall.h>
+#include <unistd.h>
+
+struct fd_parms;
+
+extern const struct fdtype_ops pidfd_dump_ops;
+extern struct collect_image_info pidfd_cinfo;
+extern int is_pidfd_link(char *link);
+
+static inline int pidfd_open(pid_t pid, unsigned int flags)
+{
+	return syscall(__NR_pidfd_open, pid, flags);
+}
+#endif

criu/include/protobuf-desc.h

@@ -81,6 +81,7 @@ enum {
 	PB_SK_QUEUES,
 	PB_IPCNS_MSG,
 	PB_IPCNS_MSG_ENT,
+	PB_PIDFD,
 
 	PB_MAX,
 };

criu/pidfd.c

@@ -0,0 +1,81 @@
+#include "pidfd.h"
+#include "util.h"
+
+#include "fdinfo.h"
+#include "files.h"
+#include "imgset.h"
+#include "protobuf.h"
+#include "fdinfo.pb-c.h"
+
+int is_pidfd_link(char *link)
+{
+	return is_anon_link_type(link, "[pidfd]");
+}
+
+static int dump_one_pidfd(int lfd, u32 id, const struct fd_parms *p)
+{
+	PidfdEntry tfe = PIDFD_ENTRY__INIT;
+	FileEntry fe = FILE_ENTRY__INIT;
+
+	if (parse_fdinfo(lfd, FD_TYPES__PIDFD, &tfe))
+		return -1;
+
+	tfe.id = id;
+	tfe.flags = p->flags;
+	tfe.inode = p->stat.st_ino;
+	tfe.mnt_id = p->mnt_id;
+	tfe.fown = (FownEntry *)&p->fown;
+
+	fe.type = FD_TYPES__PIDFD;
+	fe.id = tfe.id;
+	fe.pidfd = &tfe;
+
+	return pb_write_one(img_from_set(glob_imgset, CR_FD_FILES), &fe, PB_FILE);
+}
+
+const struct fdtype_ops pidfd_dump_ops = {
+	.type = FD_TYPES__PIDFD,
+	.dump = dump_one_pidfd,
+};
+
+struct pidfd_info {
+	PidfdEntry *pidfde;
+	struct file_desc d;
+};
+
+static int open_pidfd_fd(struct file_desc *d, int *new_fd)
+{
+	int fd = -1;
+	struct pidfd_info *info = container_of(d, struct pidfd_info, d);
+	PidfdEntry *pidfde = info->pidfde;
+
+	pr_info("Creating new pidfd %" PRId64 "\n", pidfde->pid);
+	fd = pidfd_open(pidfde->pid, 0);
+	if (fd < 0) {
+		pr_perror("Cannot create pidfd %" PRId64, pidfde->pid);
+		return -1;
+	}
+
+	*new_fd = fd;
+	return 0;
+}
+
+static struct file_desc_ops pidfd_desc_ops = {
+	.type = FD_TYPES__PIDFD,
+	.open = open_pidfd_fd,
+};
+
+static int collect_one_pidfd(void *o, ProtobufCMessage *msg, struct cr_img *i)
+{
+	struct pidfd_info *info = o;
+
+	info->pidfde = pb_msg(msg, PidfdEntry);
+	return file_desc_add(&info->d, info->pidfde->id, &pidfd_desc_ops);
+}
+
+struct collect_image_info pidfd_cinfo = {
+	.fd_type = CR_FD_PIDFD,
+	.pb_type = PB_PIDFD,
+	.priv_size = sizeof(struct pidfd_info),
+	.collect = collect_one_pidfd,
+};

criu/proc_parse.c

@@ -46,6 +46,7 @@
 #include "protobuf.h"
 #include "images/fdinfo.pb-c.h"
 #include "images/mnt.pb-c.h"
+#include "images/pidfd.pb-c.h"
 #include "plugin.h"
 
 #include <stdlib.h>
@@ -2160,6 +2161,24 @@ static int parse_fdinfo_pid_s(int pid, int fd, int type, void *arg)
 			entry_met = true;
 			continue;
 		}
+		if (fdinfo_field(str, "Pid") || fdinfo_field(str, "NSpid")) {
+			unsigned long long val;
+			PidfdEntry *pfd = arg;
+
+			if (type != FD_TYPES__PIDFD)
+				continue;
+			ret = sscanf(str, "%*s %lli", &val);
+			if (ret != 1)
+				goto parse_err;
+
+			if (fdinfo_field(str, "Pid"))
+				pfd->pid = val;
+			else if (fdinfo_field(str, "NSpid"))
+				pfd->nspid = val;
+
+			entry_met = true;
+			continue;
+		}
 	}
 
 	exit_code = 0;

images/Makefile

@@ -73,6 +73,7 @@ proto-obj-y	+= bpfmap-file.o
 proto-obj-y	+= bpfmap-data.o
 proto-obj-y	+= apparmor.o
 proto-obj-y	+= rseq.o
+proto-obj-y	+= pidfd.o
 
 CFLAGS		+= -iquote $(obj)/
 

images/fdinfo.proto

@@ -20,6 +20,7 @@ import "pipe.proto";
 import "tty.proto";
 import "memfd.proto";
 import "bpfmap-file.proto";
+import "pidfd.proto";
 
 enum fd_types {
 	UND		= 0;
@@ -42,6 +43,7 @@ enum fd_types {
 	TIMERFD		= 17;
 	MEMFD		= 18;
 	BPFMAP		= 19;
+	PIDFD           = 20;
 
 	/* Any number above the real used. Not stored to image */
 	CTL_TTY		= 65534;
@@ -78,4 +80,5 @@ message file_entry {
 	optional tty_file_entry		tty	= 19;
 	optional memfd_file_entry	memfd	= 20;
 	optional bpfmap_file_entry	bpf	= 21;
+	optional pidfd_entry	        pidfd	= 22;
 }

images/pidfd.proto

@@ -0,0 +1,16 @@
+// SPDX-License-Identifier: MIT
+
+syntax = "proto2";
+
+import "fown.proto";
+
+message pidfd_entry {
+	required uint32		id		= 1;
+	required uint64		pos		= 2;
+	required uint32		flags		= 3;
+	required uint32         mnt_id          = 4;
+	required uint64         inode           = 5;
+	required int64          pid             = 6;
+	required int64          nspid           = 7;
+	required fown_entry	fown		= 8;
+}

test/zdtm/static/Makefile

@@ -268,6 +268,7 @@ TST_NOFILE	:=				\
 		sigtrap				\
 		sigtrap01			\
 		change_mnt_context		\
+		pidfd00                         \
 #		jobctl00			\
 
 PKG_CONFIG ?= pkg-config

test/zdtm/static/pidfd00.c

@@ -0,0 +1,64 @@
+#include <errno.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <stdlib.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <signal.h>
+#include <string.h>
+#include <sys/syscall.h>
+#include <sys/types.h>
+#include <unistd.h>
+#include <poll.h>
+
+#include "zdtmtst.h"
+
+static int pidfd_open(pid_t pid, unsigned int flags)
+{
+	return syscall(SYS_pidfd_open, pid, flags);
+}
+
+const char *test_doc = "Pidfd ";
+const char *test_author = "Suraj Shirvankar <surajshirvankar@gmail.com>";
+
+int main(int argc, char **argv)
+{
+	int pidfd, errcode = 42;
+	int status;
+	int ret;
+	struct pollfd pollfd;
+
+	test_init(argc, argv);
+
+	pidfd = pidfd_open(1, 0);
+	if (pidfd == -1) {
+		perror("Couldnt open pidfd");
+		exit(1);
+	}
+
+	test_daemon();
+	test_waitsig();
+
+	pollfd.fd = pidfd;
+	pollfd.events = POLLIN;
+
+	ret = poll(&pollfd, 1, -1);
+	if (ret == -1) {
+		pr_perror("Poll error");
+		fail();
+	}
+
+	if (pollfd.revents & POLLIN) {
+		if (read(pidfd, &status, sizeof(status)) != sizeof(status)) {
+			fail("pidfd read error");
+		}
+		if (status == errcode) {
+			printf("Status code is %d", status);
+			pass();
+		} else {
+			fail("Exit code mismatch");
+		}
+	}
+
+	return 0;
+}