systemd zvol target

cmd/zed/zed.d/history_event-zfs-list-cacher.sh.in

@@ -74,6 +74,13 @@ PROPS="name,mountpoint,canmount,atime,relatime,devices,exec\
 ,org.openzfs.systemd:nofail,org.openzfs.systemd:ignore"
 
 "${ZFS}" list -H -t filesystem -o "${PROPS}" -r "${ZEVENT_POOL}" > "${FSLIST_TMP}"
+"${ZFS}" list -H -t volume -o "name,org.openzfs.systemd:ignore" -r "${ZEVENT_POOL}" |
+    while IFS="	" read -r name ignore ; do
+        [ "${ignore}" = "on" ] && continue
+
+        # fake "zvol" mountpoint
+        printf '%s\tzvol\n' "${name}" >> "${FSLIST_TMP}"
+done
 
 # Sort the output so that it is stable
 sort "${FSLIST_TMP}" -o "${FSLIST_TMP}"

cmd/zvol_wait/zvol_wait

@@ -1,5 +1,7 @@
 #!/bin/sh
 
+read -r default_volmode < /sys/module/zfs/parameters/zvol_volmode
+
 count_zvols() {
 	if [ -z "$zvols" ]; then
 		echo 0
@@ -25,34 +27,70 @@ filter_out_deleted_zvols() {
 	IFS="$OIFS"
 }
 
+should_process_zvol() {
+	# IFS=\t here!
+	name="${1}"
+	volmode="${2}"
+	token="${3}"
+	redacted="${4}"
+
+	# /dev links are not created for zvols with volmode = "none"
+	# or for redacted zvols.
+	[ "$volmode" = "none" ] && return 1
+	[ "$volmode" = "default" ] && [ "$default_volmode" = "3" ] &&
+	    return 1
+	[ "$redacted" = "-" ] || return 1
+
+	# We also ignore partially received zvols if it is
+	# not an incremental receive, as those won't even have a block
+	# device minor node created yet.
+	if [ "$token" != "-" ]; then
+		# Incremental receives create an invisible clone that
+		# is not automatically displayed by zfs list.
+		if ! zfs list "$name/%recv" >/dev/null 2>&1; then
+			return 1
+		fi
+	fi
+	return 0
+}
+
 list_zvols() {
-	read -r default_volmode < /sys/module/zfs/parameters/zvol_volmode
 	zfs list -t volume -H -o \
 	    name,volmode,receive_resume_token,redact_snaps |
-	    while IFS="	" read -r name volmode token redacted; do # IFS=\t here!
-
-		# /dev links are not created for zvols with volmode = "none"
-		# or for redacted zvols.
-		[ "$volmode" = "none" ] && continue
-		[ "$volmode" = "default" ] && [ "$default_volmode" = "3" ] &&
-		    continue
-		[ "$redacted" = "-" ] || continue
-
-		# We also ignore partially received zvols if it is
-		# not an incremental receive, as those won't even have a block
-		# device minor node created yet.
-		if [ "$token" != "-" ]; then
-
-			# Incremental receives create an invisible clone that
-			# is not automatically displayed by zfs list.
-			if ! zfs list "$name/%recv" >/dev/null 2>&1; then
-				continue
-			fi
-		fi
-		echo "$name"
+	    while IFS="	" read -r _a _b _c _d ; do # IFS=\t here!
+		should_process_zvol "${_a}" "${_b}" "${_c}" "${_d}" && echo "$name"
 	done
 }
 
+if [ -n "$1" ] ; then
+	devnode="/dev/zvol/${1}"
+	OLDIFS="$IFS"
+	IFS="	"
+	# shellcheck disable=SC2046
+	if ! should_process_zvol $(zfs list -H -o name,volmode,receive_resume_token,redact_snaps "$1") ; then
+		exit 0
+	fi
+	IFS="$OLDIFS"
+
+	outerloop=0
+	while [ "$outerloop" -lt 20 ] ; do
+		innerloop=0
+		while [ "$innerloop" -lt 30 ] ; do
+			[ -e "$devnode" ] && [ ! -L "$devnode" ] && exit 0
+			innerloop="$(( innerloop + 1))"
+			sleep 1
+		done
+		if ! zfs list "$1" 2>/dev/null >/dev/null ; then
+			# The dataset was deleted
+			exit 0
+		fi
+		outerloop="$(( outerloop + 1 ))"
+		sleep 1
+	done
+
+	exit 1
+fi
+
 zvols=$(list_zvols)
 zvols_count=$(count_zvols)
 if [ "$zvols_count" -eq 0 ]; then

etc/systemd/system-generators/Makefile.am

@@ -1,5 +1,7 @@
 include $(top_srcdir)/config/Rules.am
 
+AM_CPPFLAGS += -DSYSTEMDUNITDIR=\"$(systemdunitdir)\"
+
 systemdgenerator_PROGRAMS = \
 	zfs-mount-generator
 
@@ -9,6 +11,4 @@ zfs_mount_generator_SOURCES = \
 zfs_mount_generator_LDADD = \
 	$(abs_top_builddir)/lib/libzfs/libzfs.la
 
-zfs_mount_generator_LDFLAGS = -pthread
-
 include $(top_srcdir)/config/CppCheck.am

etc/systemd/system-generators/zfs-mount-generator.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 Antonio Russo <antonio.e.russo@gmail.com>
+ * Copyright (c) 2017, 2022 Antonio Russo <aerusso@aerusso.net>
  * Copyright (c) 2020 InsanePrawn <insane.prawny@gmail.com>
  *
  * Permission is hereby granted, free of charge, to any person obtaining
@@ -56,6 +56,7 @@
 #define	PROGNAME "zfs-mount-generator"
 #define	FSLIST SYSCONFDIR "/zfs/zfs-list.cache"
 #define	ZFS SBINDIR "/zfs"
+#define	ZVOL_TARGET_DIR "zfs-volumes.target.requires"
 
 #define	OUTPUT_HEADER \
 	"# Automatically generated by " PROGNAME "\n" \
@@ -194,7 +195,58 @@ fopenat(int dirfd, const char *pathname, int flags,
 }
 
 static int
-line_worker(char *line, const char *cachefile)
+line_worker_zvol(char *line, int *zvol_depdir_fd)
+{
+	char *toktmp;
+	const char *dataset = strtok_r(line, "\t", &toktmp);
+
+	if (*zvol_depdir_fd < 0) {
+		(void) mkdirat(destdir_fd, ZVOL_TARGET_DIR, 0755);
+		*zvol_depdir_fd = openat(destdir_fd, ZVOL_TARGET_DIR,
+		    O_PATH | O_DIRECTORY | O_CLOEXEC);
+		if (*zvol_depdir_fd < 0) {
+			fprintf(stderr, PROGNAME "[%d]: "
+			    "can't open %s/" ZVOL_TARGET_DIR ": %s\n",
+			    getpid(), destdir, strerror(errno));
+			return (1);
+		}
+	}
+
+	char *zvol_service = systemd_escape(dataset, "zfs-volume@", ".service");
+	if (zvol_service == NULL)
+		return (1);
+
+	int retval = 0;
+
+	if (symlinkat(SYSTEMDUNITDIR "/zfs-volume@.service",
+	    *zvol_depdir_fd, zvol_service) == -1) {
+		struct stat stbuf;
+
+		if ((fstatat(*zvol_depdir_fd, zvol_service, &stbuf,
+		    AT_SYMLINK_NOFOLLOW) == 0) &&
+		    ((stbuf.st_mode & S_IFMT) == S_IFLNK)) {
+			fprintf(stderr, PROGNAME "[%d]: %s: "
+			    "symlink at %s under %s/"
+			    ZVOL_TARGET_DIR " already exists: Skipping.\n",
+			    getpid(), dataset, zvol_service, destdir);
+		} else {
+			fprintf(stderr, PROGNAME "[%d]: %s: "
+			    "couldn't symlink at %s under "
+			    "%s/" ZVOL_TARGET_DIR ": %s\n",
+			    getpid(), dataset, zvol_service,
+			    destdir, strerror(errno));
+			retval = 1;
+		}
+	}
+
+	free(zvol_service);
+
+	return (retval);
+}
+
+
+static int
+line_worker_fs(char *line, const char *cachefile)
 {
 	int ret = 0;
 	void *tofree_all[8];
@@ -771,6 +823,7 @@ int
 main(int argc, char **argv)
 {
 	struct timespec time_init = {};
+	int zvol_depdir_fd = -1;
 	clock_gettime(CLOCK_MONOTONIC_RAW, &time_init);
 
 	{
@@ -895,13 +948,33 @@ main(int argc, char **argv)
 
 			char *canmount = line;
 			canmount += strcspn(canmount, "\t");
+
+			/* This SHOULD contain the dataset name */
+			if (canmount == line) {
+				/* Throw an error if we didn't find anything */
+				fprintf(stderr,
+				    PROGNAME "[%d]: %s: not enough tokens!\n",
+				    getpid(), cachent->d_name);
+				ret = 1;
+				continue;
+			}
+
 			canmount += strspn(canmount, "\t");
+
+			/* This is the second attribute in the zfs-list file */
+			if ((strncmp(canmount, "zvol", 4) == 0) &&
+			    ((canmount[4] == '\0') || (canmount[4] == '\t'))) {
+				/* special case of zvol */
+				ret |= line_worker_zvol(line, &zvol_depdir_fd);
+				continue;
+			}
+
 			canmount += strcspn(canmount, "\t");
 			canmount += strspn(canmount, "\t");
 			bool canmount_on = strncmp(canmount, "on", 2) == 0;
 
 			if (canmount_on)
-				ret |= line_worker(line, cachent->d_name);
+				ret |= line_worker_fs(line, cachent->d_name);
 			else {
 				if (filename == NULL)
 					filename =
@@ -928,11 +1001,14 @@ main(int argc, char **argv)
 	}
 	free(line);
 
+	if (zvol_depdir_fd >= 0)
+		(void) close(zvol_depdir_fd);
+
 	while (lines_canmount_not_on) {
 		struct line *l = lines_canmount_not_on;
 		lines_canmount_not_on = l->next;
 
-		ret |= line_worker(l->line, l->fname);
+		ret |= line_worker_fs(l->line, l->fname);
 		if (FREE_STATICS) {
 			free(l->line);
 			free(l);

etc/systemd/system/Makefile.am

@@ -10,6 +10,7 @@ systemdunit_DATA = \
 	zfs-mount.service \
 	zfs-share.service \
 	zfs-volume-wait.service \
+	zfs-volume@.service \
 	zfs-import.target \
 	zfs-volumes.target \
 	zfs.target \

etc/systemd/system/zfs-volume@.service.in

@@ -0,0 +1,10 @@
+[Unit]
+Description=Wait for ZFS volume %I
+Documentation=man:zfs-mount-generator(8)
+After=zfs-import.target
+ConditionPathIsDirectory=/sys/module/zfs
+
+[Service]
+Type=oneshot
+RemainAfterExit=yes
+ExecStart=@bindir@/zvol_wait %I