Fix 'zpool import' detection issues

This patch addresses multiple 'zpool import' block device
indentification problems which are most likely to occur on a
system configured to use blkid, by_vdev paths, multipath and
failover.  The symptom most commonly observed is the import
uses different path names to import the pool than would
normally be expected.

* When using blkid to identify vdevs the listed devices may
be added to the cache in any order.  In order to apply the
preferred search order heuristic a zfs_path_order() function
was added to calculate the order given full path names.

* Since it's possible to have multiple block devices with
different vdev guids which refer to the same ZPOOL_CONFIG_PATH
the slice cache must be indexed by guid and name.  By avoiding
collisions the preferred ordering can be maintaining even
when multiple block devices claim the same ZPOOL_CONFIG_PATH.
The preferred sorting by partition was never benefitial for
a Linux system and was removed as part of this change.

* When adding entries to the blkid cache avl_find/avl_insert
are used instead of avl_add because collisions are possible
and must be handled gracefully.

* For pools using multipath devices there are, at a minimum,
three devices where a vdev label may be read.  They are the
dm-* device and each underlying /dev/sd* device.  Due to the
way the block cache is implemented each of these devices may
have a different cached copy of the vdev label.  This can
result in "ghost pools" which appear to persist even after
a 'zpool labelclear' has been done to the dm-* device.  In
order to prevent this the vdev label is read with O_DIRECT
in order to bypass any caching to get the on-disk version.

* When opening a block device verify that vdev guid read from
the disk matches the expected vdev guid.  This allows for bad
labels to be filtered out.

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #5359

include/libzfs.h

@@ -64,6 +64,10 @@ extern "C" {
  */
 #define	DISK_LABEL_WAIT		(30 * 1000)  /* 30 seconds */
 
+#define	IMPORT_ORDER_PREFERRED_1	1
+#define	IMPORT_ORDER_PREFERRED_2	2
+#define	IMPORT_ORDER_SCAN_OFFSET	10
+#define	IMPORT_ORDER_DEFAULT		100
 #define	DEFAULT_IMPORT_PATH_SIZE	9
 extern char *zpool_default_import_path[DEFAULT_IMPORT_PATH_SIZE];
 
@@ -728,6 +732,7 @@ extern boolean_t zfs_bookmark_exists(const char *path);
 extern int zfs_append_partition(char *path, size_t max_len);
 extern int zfs_resolve_shortname(const char *name, char *path, size_t pathlen);
 extern int zfs_strcmp_pathname(char *name, char *cmp_name, int wholedisk);
+extern int zfs_path_order(char *path, int *order);
 
 /*
  * Mount support functions.

lib/libzfs/libzfs_import.c

@@ -1313,6 +1313,7 @@ zpool_read_label(int fd, nvlist_t **config, int *num_labels)
 	vdev_label_t *label;
 	nvlist_t *expected_config = NULL;
 	uint64_t expected_guid = 0, size;
+	int error;
 
 	*config = NULL;
 
@@ -1320,7 +1321,8 @@ zpool_read_label(int fd, nvlist_t **config, int *num_labels)
 		return (0);
 	size = P2ALIGN_TYPED(statbuf.st_size, sizeof (vdev_label_t), uint64_t);
 
-	if ((label = malloc(sizeof (vdev_label_t))) == NULL)
+	error = posix_memalign((void **)&label, PAGESIZE, sizeof (*label));
+	if (error)
 		return (-1);
 
 	for (l = 0; l < VDEV_LABELS; l++) {
@@ -1378,6 +1380,7 @@ typedef struct rdsk_node {
 	char *rn_name;			/* Full path to device */
 	int rn_order;			/* Preferred order (low to high) */
 	int rn_num_labels;		/* Number of valid labels */
+	uint64_t rn_vdev_guid;		/* Expected vdev guid when set */
 	libzfs_handle_t *rn_hdl;
 	nvlist_t *rn_config;		/* Label config */
 	avl_tree_t *rn_avl;
@@ -1386,39 +1389,29 @@ typedef struct rdsk_node {
 	boolean_t rn_labelpaths;
 } rdsk_node_t;
 
+/*
+ * Sorted by vdev guid and full path to allow for multiple entries with
+ * the same full path name.  This is required because it's possible to
+ * have multiple block devices with labels that refer to the same
+ * ZPOOL_CONFIG_PATH yet have different vdev guids.  In this case both
+ * entries need to be added to the cache.  Scenarios where this can occur
+ * include overwritten pool labels, devices which are visible from multiple
+ * hosts and multipath devices.
+ */
 static int
 slice_cache_compare(const void *arg1, const void *arg2)
 {
 	const char  *nm1 = ((rdsk_node_t *)arg1)->rn_name;
 	const char  *nm2 = ((rdsk_node_t *)arg2)->rn_name;
-	char *nm1slice, *nm2slice;
+	uint64_t guid1 = ((rdsk_node_t *)arg1)->rn_vdev_guid;
+	uint64_t guid2 = ((rdsk_node_t *)arg2)->rn_vdev_guid;
 	int rv;
 
-	/*
+	rv = AVL_CMP(guid1, guid2);
-	 * partitions one and three (slices zero and two) are the most
+	if (rv)
-	 * likely to provide results, so put those first
+		return (rv);
-	 */
-	nm1slice = strstr(nm1, "part1");
-	nm2slice = strstr(nm2, "part1");
-	if (nm1slice && !nm2slice) {
-		return (-1);
-	}
-	if (!nm1slice && nm2slice) {
-		return (1);
-	}
-	nm1slice = strstr(nm1, "part3");
-	nm2slice = strstr(nm2, "part3");
-	if (nm1slice && !nm2slice) {
-		return (-1);
-	}
-	if (!nm1slice && nm2slice) {
-		return (1);
-	}
 
-	rv = strcmp(nm1, nm2);
+	return (AVL_ISIGN(strcmp(nm1, nm2)));
-	if (rv == 0)
-		return (0);
-	return (rv > 0 ? 1 : -1);
 }
 
 static boolean_t
@@ -1506,6 +1499,7 @@ zpool_open_func(void *arg)
 	struct stat64 statbuf;
 	nvlist_t *config;
 	char *bname, *dupname;
+	uint64_t vdev_guid = 0;
 	int error;
 	int num_labels;
 	int fd;
@@ -1531,19 +1525,28 @@ zpool_open_func(void *arg)
 	    (!S_ISREG(statbuf.st_mode) && !S_ISBLK(statbuf.st_mode)))
 		return;
 
-	if ((fd = open(rn->rn_name, O_RDONLY)) < 0)
+	/*
+	 * Preferentially open using O_DIRECT to bypass the block device
+	 * cache which may be stale for multipath devices.  An EINVAL errno
+	 * indicates O_DIRECT is unsupported so fallback to just O_RDONLY.
+	 */
+	fd = open(rn->rn_name, O_RDONLY | O_DIRECT);
+	if ((fd < 0) && (errno == EINVAL))
+		fd = open(rn->rn_name, O_RDONLY);
+
+	if (fd < 0)
 		return;
 
 	/*
 	 * This file is too small to hold a zpool
 	 */
-	if (S_ISREG(statbuf.st_mode) &&
+	if (S_ISREG(statbuf.st_mode) && statbuf.st_size < SPA_MINDEVSIZE) {
-	    statbuf.st_size < SPA_MINDEVSIZE) {
 		(void) close(fd);
 		return;
 	}
 
-	if ((zpool_read_label(fd, &config, &num_labels)) != 0) {
+	error = zpool_read_label(fd, &config, &num_labels);
+	if (error != 0) {
 		(void) close(fd);
 		return;
 	}
@@ -1554,6 +1557,18 @@ zpool_open_func(void *arg)
 		return;
 	}
 
+	/*
+	 * Check that the vdev is for the expected guid.  Additional entries
+	 * are speculatively added based on the paths stored in the labels.
+	 * Entries with valid paths but incorrect guids must be removed.
+	 */
+	error = nvlist_lookup_uint64(config, ZPOOL_CONFIG_GUID, &vdev_guid);
+	if (error || (rn->rn_vdev_guid && rn->rn_vdev_guid != vdev_guid)) {
+		(void) close(fd);
+		nvlist_free(config);
+		return;
+	}
+
 	(void) close(fd);
 
 	rn->rn_config = config;
@@ -1580,9 +1595,10 @@ zpool_open_func(void *arg)
 		if (path != NULL) {
 			slice = zfs_alloc(hdl, sizeof (rdsk_node_t));
 			slice->rn_name = zfs_strdup(hdl, path);
+			slice->rn_vdev_guid = vdev_guid;
 			slice->rn_avl = rn->rn_avl;
 			slice->rn_hdl = hdl;
-			slice->rn_order = 1;
+			slice->rn_order = IMPORT_ORDER_PREFERRED_1;
 			slice->rn_labelpaths = B_FALSE;
 			mutex_enter(rn->rn_lock);
 			if (avl_find(rn->rn_avl, slice, &where)) {
@@ -1605,9 +1621,10 @@ zpool_open_func(void *arg)
 				return;
 			}
 
+			slice->rn_vdev_guid = vdev_guid;
 			slice->rn_avl = rn->rn_avl;
 			slice->rn_hdl = hdl;
-			slice->rn_order = 2;
+			slice->rn_order = IMPORT_ORDER_PREFERRED_2;
 			slice->rn_labelpaths = B_FALSE;
 			mutex_enter(rn->rn_lock);
 			if (avl_find(rn->rn_avl, slice, &where)) {
@@ -1709,10 +1726,11 @@ zpool_find_import_scan(libzfs_handle_t *hdl, kmutex_t *lock,
 				free(slice);
 				continue;
 			}
+			slice->rn_vdev_guid = 0;
 			slice->rn_lock = lock;
 			slice->rn_avl = cache;
 			slice->rn_hdl = hdl;
-			slice->rn_order = i+1;
+			slice->rn_order = i + IMPORT_ORDER_SCAN_OFFSET;
 			slice->rn_labelpaths = B_FALSE;
 			mutex_enter(lock);
 			avl_add(cache, slice);
@@ -1747,6 +1765,7 @@ zpool_find_import_blkid(libzfs_handle_t *hdl, kmutex_t *lock,
 	blkid_cache cache;
 	blkid_dev_iterate iter;
 	blkid_dev dev;
+	avl_index_t where;
 	int error;
 
 	*slice_cache = NULL;
@@ -1781,13 +1800,25 @@ zpool_find_import_blkid(libzfs_handle_t *hdl, kmutex_t *lock,
 	while (blkid_dev_next(iter, &dev) == 0) {
 		slice = zfs_alloc(hdl, sizeof (rdsk_node_t));
 		slice->rn_name = zfs_strdup(hdl, blkid_dev_devname(dev));
+		slice->rn_vdev_guid = 0;
 		slice->rn_lock = lock;
 		slice->rn_avl = *slice_cache;
 		slice->rn_hdl = hdl;
-		slice->rn_order = 100;
 		slice->rn_labelpaths = B_TRUE;
+
+		error = zfs_path_order(slice->rn_name, &slice->rn_order);
+		if (error == 0)
+			slice->rn_order += IMPORT_ORDER_SCAN_OFFSET;
+		else
+			slice->rn_order = IMPORT_ORDER_DEFAULT;
+
 		mutex_enter(lock);
-		avl_add(*slice_cache, slice);
+		if (avl_find(*slice_cache, slice, &where)) {
+			free(slice->rn_name);
+			free(slice);
+		} else {
+			avl_insert(*slice_cache, slice, where);
+		}
 		mutex_exit(lock);
 	}
 

lib/libzfs/libzfs_util.c

@@ -1136,6 +1136,45 @@ zfs_strcmp_pathname(char *name, char *cmp, int wholedisk)
 	return (0);
 }
 
+/*
+ * Given a full path to a device determine if that device appears in the
+ * import search path.  If it does return the first match and store the
+ * index in the passed 'order' variable, otherwise return an error.
+ */
+int
+zfs_path_order(char *name, int *order)
+{
+	int i = 0, error = ENOENT;
+	char *dir, *env, *envdup;
+
+	env = getenv("ZPOOL_IMPORT_PATH");
+	if (env) {
+		envdup = strdup(env);
+		dir = strtok(envdup, ":");
+		while (dir) {
+			if (strncmp(name, dir, strlen(dir)) == 0) {
+				*order = i;
+				error = 0;
+				break;
+			}
+			dir = strtok(NULL, ":");
+			i++;
+		}
+		free(envdup);
+	} else {
+		for (i = 0; i < DEFAULT_IMPORT_PATH_SIZE; i++) {
+			if (strncmp(name, zpool_default_import_path[i],
+			    strlen(zpool_default_import_path[i])) == 0) {
+				*order = i;
+				error = 0;
+				break;
+			}
+		}
+	}
+
+	return (error);
+}
+
 /*
  * Initialize the zc_nvlist_dst member to prepare for receiving an nvlist from
  * an ioctl().