15137 ZFS should allow direct import of a root pool from a /devices path

15122 vdev_disk_preroot_fini can race against the child vdev open taskq
Reviewed by: Robert Mustacchi <rm@fingolfin.org>
Reviewed by: Toomas Soome <tsoome@me.com>
Reviewed by: Dan Cross <cross@oxidecomputer.com>
Approved by: Dan McDonald <danmcd@mnx.io>

4 files changed, 72 insertions, 6 deletions

diff --git a/usr/src/uts/common/fs/zfs/spa.c b/usr/src/uts/common/fs/zfs/spa.c
index d390e39195..d6e230fbb4 100644
--- a/usr/src/uts/common/fs/zfs/spa.c
+++ b/usr/src/uts/common/fs/zfs/spa.c
@@ -5479,6 +5479,21 @@ spa_import_rootpool(char *devpath, char *devid, uint64_t pool_guid,
 	char *pname;
 	int error;
 	const char *altdevpath = NULL;
+	const char *rdpath = NULL;
+
+	if ((rdpath = vdev_disk_preroot_force_path()) != NULL) {
+		/*
+		 * We expect to import a single-vdev pool from a specific
+		 * device such as a ramdisk device.  We don't care what the
+		 * pool label says.
+		 */
+		config = spa_generate_rootconf(rdpath, NULL, &guid, 0);
+		if (config != NULL) {
+			goto configok;
+		}
+		cmn_err(CE_NOTE, "Cannot use root disk device '%s'", rdpath);
+		return (SET_ERROR(EIO));
+	}
 
 	/*
 	 * Read the label from the boot device and generate a configuration.
@@ -5521,6 +5536,7 @@ spa_import_rootpool(char *devpath, char *devid, uint64_t pool_guid,
 		return (SET_ERROR(EIO));
 	}
 
+configok:
 	VERIFY(nvlist_lookup_string(config, ZPOOL_CONFIG_POOL_NAME,
 	    &pname) == 0);
 	VERIFY(nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
diff --git a/usr/src/uts/common/fs/zfs/sys/vdev_impl.h b/usr/src/uts/common/fs/zfs/sys/vdev_impl.h
index d1341dbcd3..d542368e7c 100644
--- a/usr/src/uts/common/fs/zfs/sys/vdev_impl.h
+++ b/usr/src/uts/common/fs/zfs/sys/vdev_impl.h
@@ -588,9 +588,10 @@ typedef struct vdev_buf {
  * Support routines used during boot from a ZFS pool
  */
 extern int vdev_disk_read_rootlabel(const char *, const char *, nvlist_t **);
-extern void vdev_disk_preroot_init(void);
+extern void vdev_disk_preroot_init(const char *);
 extern void vdev_disk_preroot_fini(void);
 extern const char *vdev_disk_preroot_lookup(uint64_t, uint64_t);
+extern const char *vdev_disk_preroot_force_path(void);
 
 #ifdef	__cplusplus
 }
diff --git a/usr/src/uts/common/fs/zfs/vdev_disk.c b/usr/src/uts/common/fs/zfs/vdev_disk.c
index 1cdc458974..cd5e80d769 100644
--- a/usr/src/uts/common/fs/zfs/vdev_disk.c
+++ b/usr/src/uts/common/fs/zfs/vdev_disk.c
@@ -301,6 +301,7 @@ vdev_disk_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 	int otyp;
 	boolean_t validate_devid = B_FALSE;
 	uint64_t capacity = 0, blksz = 0, pbsize;
+	const char *rdpath = vdev_disk_preroot_force_path();
 
 	/*
 	 * We must have a pathname, and it must be absolute.
@@ -328,7 +329,8 @@ vdev_disk_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 	/*
 	 * Allow bypassing the devid.
 	 */
-	if (vd->vdev_devid != NULL && vdev_disk_bypass_devid) {
+	if (vd->vdev_devid != NULL &&
+	    (vdev_disk_bypass_devid || rdpath != NULL)) {
 		vdev_dbgmsg(vd, "vdev_disk_open, devid %s bypassed",
 		    vd->vdev_devid);
 		spa_strfree(vd->vdev_devid);
@@ -364,6 +366,17 @@ vdev_disk_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 
 	error = EINVAL;		/* presume failure */
 
+	if (rdpath != NULL) {
+		/*
+		 * We have been asked to open only a specific root device, and
+		 * to fail otherwise.
+		 */
+		error = ldi_open_by_name((char *)rdpath, spa_mode(spa), kcred,
+		    &dvd->vd_lh, zfs_li);
+		validate_devid = B_TRUE;
+		goto rootdisk_only;
+	}
+
 	if (vd->vdev_path != NULL) {
 		if (vd->vdev_wholedisk == -1ULL) {
 			size_t len = strlen(vd->vdev_path) + 3;
@@ -501,6 +514,7 @@ vdev_disk_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
 		}
 	}
 
+rootdisk_only:
 	if (error != 0) {
 		vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
 		vdev_dbgmsg(vd, "vdev_disk_open: failed to open [error=%d]",
@@ -1158,6 +1172,7 @@ vdev_disk_read_rootlabel(const char *devpath, const char *devid,
 struct veb {
 	list_t veb_ents;
 	boolean_t veb_scanned;
+	char *veb_force_path;
 };
 
 struct veb_ent {
@@ -1269,8 +1284,22 @@ vdev_disk_preroot_lookup(uint64_t pool_guid, uint64_t vdev_guid)
 	return (path);
 }
 
+const char *
+vdev_disk_preroot_force_path(void)
+{
+	const char *force_path = NULL;
+
+	mutex_enter(&veb_lock);
+	if (veb != NULL) {
+		force_path = veb->veb_force_path;
+	}
+	mutex_exit(&veb_lock);
+
+	return (force_path);
+}
+
 void
-vdev_disk_preroot_init(void)
+vdev_disk_preroot_init(const char *force_path)
 {
 	mutex_init(&veb_lock, NULL, MUTEX_DEFAULT, NULL);
 
@@ -1279,6 +1308,9 @@ vdev_disk_preroot_init(void)
 	list_create(&veb->veb_ents, sizeof (struct veb_ent),
 	    offsetof(struct veb_ent, vebe_link));
 	veb->veb_scanned = B_FALSE;
+	if (force_path != NULL) {
+		veb->veb_force_path = spa_strdup(force_path);
+	}
 }
 
 void
@@ -1295,6 +1327,10 @@ vdev_disk_preroot_fini(void)
 			kmem_free(vebe, sizeof (*vebe));
 		}
 
+		if (veb->veb_force_path != NULL) {
+			spa_strfree(veb->veb_force_path);
+		}
+
 		kmem_free(veb, sizeof (*veb));
 		veb = NULL;
 	}
diff --git a/usr/src/uts/common/fs/zfs/zfs_vfsops.c b/usr/src/uts/common/fs/zfs/zfs_vfsops.c
index 87b4b36a7d..288dc93e3c 100644
--- a/usr/src/uts/common/fs/zfs/zfs_vfsops.c
+++ b/usr/src/uts/common/fs/zfs/zfs_vfsops.c
@@ -1767,6 +1767,7 @@ zfs_mountroot(vfs_t *vfsp, enum whymountroot why)
 	vnode_t *vp = NULL;
 	char *zfs_bootfs;
 	char *zfs_devid;
+	char *zfs_rootdisk_path;
 	uint64_t zfs_bootpool;
 	uint64_t zfs_bootvdev;
 
@@ -1806,12 +1807,19 @@ zfs_mountroot(vfs_t *vfsp, enum whymountroot why)
 		zfs_bootvdev = spa_get_bootprop_uint64("zfs-bootvdev", 0);
 
 		/*
+		 * If we have been given a root disk override path, we want to
+		 * ignore device paths from the pool configuration and use only
+		 * the specific path we were given in the boot properties.
+		 */
+		zfs_rootdisk_path = spa_get_bootprop("zfs-rootdisk-path");
+
+		/*
 		 * Initialise the early boot device rescan mechanism.  A scan
 		 * will not actually be performed unless we need to do so in
 		 * order to find the correct /devices path for a relocated
 		 * device.
 		 */
-		vdev_disk_preroot_init();
+		vdev_disk_preroot_init(zfs_rootdisk_path);
 
 		error = spa_import_rootpool(rootfs.bo_name, zfs_devid,
 		    zfs_bootpool, zfs_bootvdev);
@@ -1820,6 +1828,7 @@ zfs_mountroot(vfs_t *vfsp, enum whymountroot why)
 
 		if (error != 0) {
 			spa_free_bootprop(zfs_bootfs);
+			spa_free_bootprop(zfs_rootdisk_path);
 			vdev_disk_preroot_fini();
 			cmn_err(CE_NOTE, "spa_import_rootpool: error %d",
 			    error);
@@ -1828,6 +1837,7 @@ zfs_mountroot(vfs_t *vfsp, enum whymountroot why)
 
 		if (error = zfs_parse_bootfs(zfs_bootfs, rootfs.bo_name)) {
 			spa_free_bootprop(zfs_bootfs);
+			spa_free_bootprop(zfs_rootdisk_path);
 			vdev_disk_preroot_fini();
 			cmn_err(CE_NOTE, "zfs_parse_bootfs: error %d",
 			    error);
@@ -1835,10 +1845,12 @@ zfs_mountroot(vfs_t *vfsp, enum whymountroot why)
 		}
 
 		spa_free_bootprop(zfs_bootfs);
-		vdev_disk_preroot_fini();
+		spa_free_bootprop(zfs_rootdisk_path);
 
-		if (error = vfs_lock(vfsp))
+		if ((error = vfs_lock(vfsp)) != 0) {
+			vdev_disk_preroot_fini();
 			return (error);
+		}
 
 		if (error = zfs_domount(vfsp, rootfs.bo_name)) {
 			cmn_err(CE_NOTE, "zfs_domount: error %d", error);
@@ -1864,6 +1876,7 @@ zfs_mountroot(vfs_t *vfsp, enum whymountroot why)
 		vfs_add((struct vnode *)0, vfsp,
 		    (vfsp->vfs_flag & VFS_RDONLY) ? MS_RDONLY : 0);
 out:
+		vdev_disk_preroot_fini();
 		vfs_unlock(vfsp);
 		return (error);
 	} else if (why == ROOT_REMOUNT) {