Whamcloud - gitweb
LU-8066 sysfs: make ping sysfs file read and writable
[fs/lustre-release.git] / lustre / utils / liblustreapi.c
index 351990d..8f69ac6 100644 (file)
@@ -68,6 +68,7 @@
 #include <unistd.h>
 #endif
 #include <poll.h>
+#include <time.h>
 
 #include <libcfs/util/ioctl.h>
 #include <libcfs/util/param.h>
@@ -762,9 +763,7 @@ retry_open:
                        errmsg = strerror(errno);
 
                llapi_err_noerrno(LLAPI_MSG_ERROR,
-                                 "error on ioctl %#jx for '%s' (%d): %s",
-                                 (uintmax_t)LL_IOC_LOV_SETSTRIPE, name, fd,
-                                 errmsg);
+                                 "setstripe error for '%s': %s", name, errmsg);
 
                close(fd);
                fd = rc;
@@ -969,8 +968,8 @@ int llapi_dir_set_default_lmv(const char *name,
                        errmsg = strerror(errno);
 
                llapi_err_noerrno(LLAPI_MSG_ERROR,
-                                 "error on LL_IOC_LMV_SETSTRIPE '%s' (%d): %s",
-                                 name, fd, errmsg);
+                                 "default dirstripe error on '%s': %s",
+                                 name, errmsg);
        }
        close(fd);
        return rc;
@@ -1004,7 +1003,7 @@ int llapi_dir_create(const char *name, mode_t mode,
                     const struct llapi_stripe_param *param)
 {
        struct lmv_user_md *lmu = NULL;
-       size_t lmu_size = sizeof(*lmu);
+       size_t lmu_size;
        struct obd_ioctl_data data = { 0 };
        char rawbuf[8192];
        char *buf = rawbuf;
@@ -1018,9 +1017,10 @@ int llapi_dir_create(const char *name, mode_t mode,
        if (rc)
                return rc;
 
-       if (param->lsp_is_specific)
-               lmu_size = lmv_user_md_size(param->lsp_stripe_count,
-                                           LMV_USER_MAGIC_SPECIFIC);
+       lmu_size = lmv_user_md_size(param->lsp_stripe_count,
+                                   param->lsp_is_specific ?
+                                        LMV_USER_MAGIC_SPECIFIC :
+                                        LMV_USER_MAGIC);
 
        lmu = calloc(1, lmu_size);
        if (lmu == NULL)
@@ -1066,13 +1066,13 @@ int llapi_dir_create(const char *name, mode_t mode,
 
        if (ioctl(fd, LL_IOC_LMV_SETSTRIPE, buf)) {
                char *errmsg = "stripe already set";
+
                rc = -errno;
                if (errno != EEXIST && errno != EALREADY)
                        errmsg = strerror(errno);
 
                llapi_err_noerrno(LLAPI_MSG_ERROR,
-                                 "error on LL_IOC_LMV_SETSTRIPE '%s' (%d): %s",
-                                 name, fd, errmsg);
+                                 "dirstripe error on '%s': %s", name, errmsg);
        }
        close(fd);
 out:
@@ -1344,26 +1344,36 @@ int llapi_search_rootpath(char *pathname, const char *fsname)
 
 int llapi_getname(const char *path, char *buf, size_t size)
 {
-        struct obd_uuid uuid_buf;
-        char *uuid = uuid_buf.uuid;
-        int rc, nr;
+       struct obd_uuid uuid_buf;
+       char *uuid = uuid_buf.uuid;
+       char *cfg_instance;
+       int rc, len, fsname_len;
 
-        memset(&uuid_buf, 0, sizeof(uuid_buf));
-        rc = llapi_file_get_lov_uuid(path, &uuid_buf);
-        if (rc)
-                return rc;
+       memset(&uuid_buf, 0, sizeof(uuid_buf));
+       rc = llapi_file_get_lov_uuid(path, &uuid_buf);
+       if (rc)
+               return rc;
 
-        /* We want to turn lustre-clilov-ffff88002738bc00 into
-         * lustre-ffff88002738bc00. */
+       /*
+        * We want to turn testfs-clilov-ffff88002738bc00 into
+        * testfs-ffff88002738bc00 in a portable way that doesn't depend
+        * on what is after "-clilov-" as it may change in the future.
+        * Unfortunately, the "fsname" part may contain a dash, so we
+        * can't just skip to the first dash, and the "instance" may be a
+        * UUID in the future, so we can't necessarily go to the last dash.
+        */
+       cfg_instance = strstr(uuid, "-clilov-");
+       if (!cfg_instance)
+               return -EINVAL;
 
-        nr = snprintf(buf, size, "%.*s-%s",
-                      (int) (strlen(uuid) - 24), uuid,
-                      uuid + strlen(uuid) - 16);
+       fsname_len = cfg_instance - uuid;
+       cfg_instance += strlen("-clilov-");
+       len = snprintf(buf, size, "%.*s-%s", fsname_len, uuid, cfg_instance);
 
-        if (nr >= size)
-                rc = -ENAMETOOLONG;
+       if (len >= size)
+               rc = -ENAMETOOLONG;
 
-        return rc;
+       return rc;
 }
 
 /**
@@ -1643,6 +1653,10 @@ static int common_param_init(struct find_param *param, char *path)
        if (lum_size < 0)
                return lum_size;
 
+       /* migrate has fp_lmv_md initialized outside */
+       if (param->fp_migrate)
+               return 0;
+
        if (lum_size < PATH_MAX + 1)
                lum_size = PATH_MAX + 1;
 
@@ -1658,25 +1672,27 @@ static int common_param_init(struct find_param *param, char *path)
        param->fp_lmv_stripe_count = 256;
        param->fp_lmv_md = calloc(1,
                                  lmv_user_md_size(param->fp_lmv_stripe_count,
-                                                  LMV_MAGIC_V1));
+                                                  LMV_USER_MAGIC_SPECIFIC));
        if (param->fp_lmv_md == NULL) {
                llapi_error(LLAPI_MSG_ERROR, -ENOMEM,
                            "error: allocation of %d bytes for ioctl",
                            lmv_user_md_size(param->fp_lmv_stripe_count,
-                                            LMV_MAGIC_V1));
+                                            LMV_USER_MAGIC_SPECIFIC));
                return -ENOMEM;
        }
 
        param->fp_got_uuids = 0;
        param->fp_obd_indexes = NULL;
        param->fp_obd_index = OBD_NOT_FOUND;
-       if (!param->fp_migrate)
-               param->fp_mdt_index = OBD_NOT_FOUND;
+       param->fp_mdt_index = OBD_NOT_FOUND;
        return 0;
 }
 
 static void find_param_fini(struct find_param *param)
 {
+       if (param->fp_migrate)
+               return;
+
        if (param->fp_obd_indexes)
                free(param->fp_obd_indexes);
 
@@ -1736,13 +1752,14 @@ again:
 
                free(param->fp_lmv_md);
                param->fp_lmv_stripe_count = stripe_count;
-               lmv_size = lmv_user_md_size(stripe_count, LMV_MAGIC_V1);
+               lmv_size = lmv_user_md_size(stripe_count,
+                                           LMV_USER_MAGIC_SPECIFIC);
                param->fp_lmv_md = malloc(lmv_size);
                if (param->fp_lmv_md == NULL) {
                        llapi_error(LLAPI_MSG_ERROR, -ENOMEM,
                                    "error: allocation of %d bytes for ioctl",
                                    lmv_user_md_size(param->fp_lmv_stripe_count,
-                                                    LMV_MAGIC_V1));
+                                                    LMV_USER_MAGIC_SPECIFIC));
                        return -ENOMEM;
                }
                goto again;
@@ -1918,8 +1935,10 @@ static int llapi_semantic_traverse(char *path, int size, DIR *parent,
                        rc = 0;
                        if (sem_init) {
                                rc = sem_init(path, d, NULL, data, dent);
-                               if (rc < 0 && ret == 0)
+                               if (rc < 0 && ret == 0) {
                                        ret = rc;
+                                       break;
+                               }
                        }
                        if (sem_fini && rc == 0)
                                sem_fini(path, d, NULL, data, dent);
@@ -2475,7 +2494,8 @@ enum lov_dump_flags {
 
 static void lov_dump_user_lmm_header(struct lov_user_md *lum, char *path,
                                     struct lov_user_ost_data_v1 *objects,
-                                    int verbose, int depth, char *pool_name,
+                                    enum llapi_layout_verbose verbose,
+                                    int depth, char *pool_name,
                                     enum lov_dump_flags flags)
 {
        bool is_dir = flags & LDF_IS_DIR;
@@ -2546,8 +2566,8 @@ static void lov_dump_user_lmm_header(struct lov_user_md *lum, char *path,
                                     seq, oid, ver);
        }
 
-       if (verbose & VERBOSE_COUNT) {
-               if (verbose & ~VERBOSE_COUNT)
+       if (verbose & VERBOSE_STRIPE_COUNT) {
+               if (verbose & ~VERBOSE_STRIPE_COUNT)
                        llapi_printf(LLAPI_MSG_NORMAL, "%s%sstripe_count:  ",
                                     space, prefix);
                if (is_dir) {
@@ -2580,9 +2600,9 @@ static void lov_dump_user_lmm_header(struct lov_user_md *lum, char *path,
                        separator = "\n";
        }
 
-       if (verbose & VERBOSE_SIZE) {
+       if (verbose & VERBOSE_STRIPE_SIZE) {
                llapi_printf(LLAPI_MSG_NORMAL, "%s", separator);
-               if (verbose & ~VERBOSE_SIZE)
+               if (verbose & ~VERBOSE_STRIPE_SIZE)
                        llapi_printf(LLAPI_MSG_NORMAL, "%s%sstripe_size:   ",
                                     space, prefix);
                if (is_dir && !is_raw && lum->lmm_stripe_size == 0) {
@@ -2605,9 +2625,9 @@ static void lov_dump_user_lmm_header(struct lov_user_md *lum, char *path,
                        separator = "\n";
        }
 
-       if ((verbose & VERBOSE_LAYOUT)) {
+       if ((verbose & VERBOSE_PATTERN)) {
                llapi_printf(LLAPI_MSG_NORMAL, "%s", separator);
-               if (verbose & ~VERBOSE_LAYOUT)
+               if (verbose & ~VERBOSE_PATTERN)
                        llapi_printf(LLAPI_MSG_NORMAL, "%s%spattern:       ",
                                     space, prefix);
                if (lov_pattern_supported(lum->lmm_pattern))
@@ -2628,9 +2648,9 @@ static void lov_dump_user_lmm_header(struct lov_user_md *lum, char *path,
                separator = "\n";
        }
 
-       if (verbose & VERBOSE_OFFSET) {
+       if (verbose & VERBOSE_STRIPE_OFFSET) {
                llapi_printf(LLAPI_MSG_NORMAL, "%s", separator);
-               if (verbose & ~VERBOSE_OFFSET)
+               if (verbose & ~VERBOSE_STRIPE_OFFSET)
                        llapi_printf(LLAPI_MSG_NORMAL, "%s%sstripe_offset: ",
                                     space, prefix);
                if (is_dir || skip_objs)
@@ -2668,7 +2688,8 @@ static void lov_dump_user_lmm_header(struct lov_user_md *lum, char *path,
 void lov_dump_user_lmm_v1v3(struct lov_user_md *lum, char *pool_name,
                            struct lov_user_ost_data_v1 *objects,
                            char *path, int obdindex, int depth,
-                           int header, enum lov_dump_flags flags)
+                           enum llapi_layout_verbose verbose,
+                           enum lov_dump_flags flags)
 {
        bool is_dir = flags & LDF_IS_DIR;
        bool indent = flags & LDF_INDENT;
@@ -2689,10 +2710,10 @@ void lov_dump_user_lmm_v1v3(struct lov_user_md *lum, char *pool_name,
        if (!obdstripe)
                return;
 
-       lov_dump_user_lmm_header(lum, path, objects, header, depth, pool_name,
+       lov_dump_user_lmm_header(lum, path, objects, verbose, depth, pool_name,
                                 flags);
 
-       if (!is_dir && !skip_objs && (header & VERBOSE_OBJID) &&
+       if (!is_dir && !skip_objs && (verbose & VERBOSE_OBJID) &&
            !(lum->lmm_pattern & LOV_PATTERN_F_RELEASED ||
              lov_pattern(lum->lmm_pattern) == LOV_PATTERN_MDT)) {
                char *space = "      - ";
@@ -2747,7 +2768,8 @@ void lov_dump_user_lmm_v1v3(struct lov_user_md *lum, char *pool_name,
 }
 
 void lmv_dump_user_lmm(struct lmv_user_md *lum, char *pool_name,
-                      char *path, int obdindex, int depth, int verbose,
+                      char *path, int obdindex, int depth,
+                      enum llapi_layout_verbose verbose,
                       enum lov_dump_flags flags)
 {
        struct lmv_user_mds_data *objects = lum->lum_objects;
@@ -2782,8 +2804,8 @@ void lmv_dump_user_lmm(struct lmv_user_md *lum, char *pool_name,
        /* show all information default */
        if (!verbose) {
                if (lum->lum_magic == LMV_USER_MAGIC)
-                       verbose = VERBOSE_POOL | VERBOSE_COUNT |
-                                 VERBOSE_OFFSET | VERBOSE_HASH_TYPE;
+                       verbose = VERBOSE_POOL | VERBOSE_STRIPE_COUNT |
+                                 VERBOSE_STRIPE_OFFSET | VERBOSE_HASH_TYPE;
                else
                        verbose = VERBOSE_OBJID;
        }
@@ -2791,21 +2813,21 @@ void lmv_dump_user_lmm(struct lmv_user_md *lum, char *pool_name,
        if (depth && path && ((verbose != VERBOSE_OBJID)))
                llapi_printf(LLAPI_MSG_NORMAL, "%s%s\n", prefix, path);
 
-       if (verbose & VERBOSE_COUNT) {
+       if (verbose & VERBOSE_STRIPE_COUNT) {
                llapi_printf(LLAPI_MSG_NORMAL, "%s", separator);
-               if (verbose & ~VERBOSE_COUNT)
+               if (verbose & ~VERBOSE_STRIPE_COUNT)
                        llapi_printf(LLAPI_MSG_NORMAL, "lmv_stripe_count: ");
                llapi_printf(LLAPI_MSG_NORMAL, "%u",
                             (int)lum->lum_stripe_count);
-               if ((verbose & VERBOSE_OFFSET) && !yaml)
+               if ((verbose & VERBOSE_STRIPE_OFFSET) && !yaml)
                        separator = " ";
                else
                        separator = "\n";
        }
 
-       if (verbose & VERBOSE_OFFSET) {
+       if (verbose & VERBOSE_STRIPE_OFFSET) {
                llapi_printf(LLAPI_MSG_NORMAL, "%s", separator);
-               if (verbose & ~VERBOSE_OFFSET)
+               if (verbose & ~VERBOSE_STRIPE_OFFSET)
                        llapi_printf(LLAPI_MSG_NORMAL, "lmv_stripe_offset: ");
                llapi_printf(LLAPI_MSG_NORMAL, "%d",
                             (int)lum->lum_stripe_offset);
@@ -2816,7 +2838,8 @@ void lmv_dump_user_lmm(struct lmv_user_md *lum, char *pool_name,
        }
 
        if (verbose & VERBOSE_HASH_TYPE) {
-               unsigned int type = lum->lum_hash_type;
+               unsigned int type = lum->lum_hash_type & LMV_HASH_TYPE_MASK;
+               unsigned int flags = lum->lum_hash_type & ~LMV_HASH_TYPE_MASK;
 
                llapi_printf(LLAPI_MSG_NORMAL, "%s", separator);
                if (verbose & ~VERBOSE_HASH_TYPE)
@@ -2825,9 +2848,18 @@ void lmv_dump_user_lmm(struct lmv_user_md *lum, char *pool_name,
                        llapi_printf(LLAPI_MSG_NORMAL, "%s",
                                     mdt_hash_name[type]);
                else
-                       llapi_printf(LLAPI_MSG_NORMAL, "%d",
-                                    (int)type);
+                       llapi_printf(LLAPI_MSG_NORMAL, "%#x", type);
+
+               if (flags & LMV_HASH_FLAG_MIGRATION)
+                       llapi_printf(LLAPI_MSG_NORMAL, ",migrating");
+               if (flags & LMV_HASH_FLAG_DEAD)
+                       llapi_printf(LLAPI_MSG_NORMAL, ",dead");
+               if (flags & LMV_HASH_FLAG_BAD_TYPE)
+                       llapi_printf(LLAPI_MSG_NORMAL, ",bad_type");
+               if (flags & LMV_HASH_FLAG_LOST_LMV)
+                       llapi_printf(LLAPI_MSG_NORMAL, ",lost_lmv");
                separator = "\n";
+
        }
 
        if (verbose & VERBOSE_OBJID && lum->lum_magic != LMV_USER_MAGIC) {
@@ -2866,7 +2898,7 @@ static void lov_dump_comp_v1_header(struct find_param *param, char *path,
 {
        struct lov_comp_md_v1 *comp_v1 = (void *)&param->fp_lmd->lmd_lmm;
        int depth = param->fp_max_depth;
-       int verbose = param->fp_verbose;
+       enum llapi_layout_verbose verbose = param->fp_verbose;
        bool yaml = flags & LDF_YAML;
 
        if (depth && path && ((verbose != VERBOSE_OBJID) ||
@@ -2951,7 +2983,7 @@ static void lov_dump_comp_v1_entry(struct find_param *param,
        struct lov_comp_md_v1 *comp_v1 = (void *)&param->fp_lmd->lmd_lmm;
        struct lov_comp_md_entry_v1 *entry;
        char *separator = "";
-       int verbose = param->fp_verbose;
+       enum llapi_layout_verbose verbose = param->fp_verbose;
        bool yaml = flags & LDF_YAML;
 
        entry = &comp_v1->lcm_entries[index];
@@ -2994,6 +3026,26 @@ static void lov_dump_comp_v1_entry(struct find_param *param,
                lcme_flags2str(entry->lcme_flags);
                separator = "\n";
        }
+       /* print snapshot timestamp if its a nosync comp */
+       if ((verbose & VERBOSE_COMP_FLAGS) &&
+           (entry->lcme_flags & LCME_FL_NOSYNC)) {
+               llapi_printf(LLAPI_MSG_NORMAL, "%s", separator);
+               if (verbose & ~VERBOSE_COMP_FLAGS)
+                       llapi_printf(LLAPI_MSG_NORMAL,
+                                    "%4slcme_timestamp:      ", " ");
+               if (yaml) {
+                       llapi_printf(LLAPI_MSG_NORMAL, "%llu",
+                                                       entry->lcme_timestamp);
+               } else {
+                       time_t stamp = entry->lcme_timestamp;
+                       char *date_str = asctime(localtime(&stamp));
+
+                       date_str[strlen(date_str) - 1] = '\0';
+                       llapi_printf(LLAPI_MSG_NORMAL, "'%s'", date_str);
+               }
+
+               separator = "\n";
+       }
 
        if (verbose & VERBOSE_COMP_START) {
                llapi_printf(LLAPI_MSG_NORMAL, "%s", separator);
@@ -3112,7 +3164,8 @@ print_last_init_comp(struct find_param *param)
 
        /* print specific component info */
        if (param->fp_check_comp_id || param->fp_check_comp_flags ||
-           param->fp_check_comp_start || param->fp_check_comp_end)
+           param->fp_check_comp_start || param->fp_check_comp_end ||
+           param->fp_check_mirror_id || param->fp_check_mirror_index)
                return false;
 
        return true;
@@ -3203,6 +3256,8 @@ static void lov_dump_comp_v1(struct find_param *param, char *path,
        int obdindex = param->fp_obd_index;
        int i, j, match;
        bool obdstripe = false;
+       __u16 mirror_index = 0;
+       __u16 mirror_id = 0;
 
        if (obdindex != OBD_NOT_FOUND) {
                for (i = 0; !(flags & LDF_IS_DIR) && !obdstripe &&
@@ -3263,6 +3318,32 @@ static void lov_dump_comp_v1(struct find_param *param, char *path,
                                continue;
                }
 
+               if (param->fp_check_mirror_index) {
+                       if (mirror_id != mirror_id_of(entry->lcme_id)) {
+                               mirror_index++;
+                               mirror_id = mirror_id_of(entry->lcme_id);
+                       }
+
+                       match = find_value_cmp(mirror_index,
+                                              param->fp_mirror_index,
+                                              param->fp_mirror_index_sign,
+                                              param->fp_exclude_mirror_index,
+                                              1, 0);
+                       if (match == -1)
+                               continue;
+               } else if (param->fp_check_mirror_id) {
+                       if (mirror_id != mirror_id_of(entry->lcme_id))
+                               mirror_id = mirror_id_of(entry->lcme_id);
+
+                       match = find_value_cmp(mirror_id,
+                                              param->fp_mirror_id,
+                                              param->fp_mirror_id_sign,
+                                              param->fp_exclude_mirror_id,
+                                              1, 0);
+                       if (match == -1)
+                               continue;
+               }
+
                if (print_last_init_comp(param)) {
                        /**
                         * if part of stripe info is needed, we'd print only
@@ -3334,7 +3415,7 @@ static void lov_dump_comp_v1(struct find_param *param, char *path,
 
 static inline bool has_any_comp_options(struct find_param *param)
 {
-       int verbose = param->fp_verbose;
+       enum llapi_layout_verbose verbose = param->fp_verbose;
 
        if (param->fp_check_comp_id || param->fp_check_comp_count ||
            param->fp_check_comp_start || param->fp_check_comp_end ||
@@ -3445,6 +3526,7 @@ static void llapi_lov_dump_user_lmm(struct find_param *param, char *path,
        switch (magic) {
        case LOV_USER_MAGIC_V1:
        case LOV_USER_MAGIC_V3:
+       case LOV_USER_MAGIC_SPECIFIC:
                lov_dump_plain_user_lmm(param, path, flags);
                break;
        case LMV_MAGIC_V1:
@@ -4337,20 +4419,24 @@ out:
 static int cb_migrate_mdt_init(char *path, DIR *parent, DIR **dirp,
                               void *param_data, struct dirent64 *de)
 {
-       struct find_param       *param = (struct find_param *)param_data;
-       DIR                     *tmp_parent = parent;
-       char                    raw[MAX_IOC_BUFLEN] = {'\0'};
-       char                    *rawbuf = raw;
-       struct obd_ioctl_data   data = { 0 };
-       int                     fd;
-       int                     ret;
-       char                    *path_copy;
-       char                    *filename;
-       bool                    retry = false;
+       struct find_param *param = (struct find_param *)param_data;
+       struct lmv_user_md *lmu = param->fp_lmv_md;
+       DIR *tmp_parent = parent;
+       char raw[MAX_IOC_BUFLEN] = {'\0'};
+       char *rawbuf = raw;
+       struct obd_ioctl_data data = { 0 };
+       int fd;
+       int ret;
+       char *path_copy;
+       char *filename;
+       bool retry = false;
 
        if (parent == NULL && dirp == NULL)
                return -EINVAL;
 
+       if (!lmu)
+               return -EINVAL;
+
        if (dirp != NULL)
                closedir(*dirp);
 
@@ -4369,10 +4455,12 @@ static int cb_migrate_mdt_init(char *path, DIR *parent, DIR **dirp,
 
        path_copy = strdup(path);
        filename = basename(path_copy);
+
        data.ioc_inlbuf1 = (char *)filename;
        data.ioc_inllen1 = strlen(filename) + 1;
-       data.ioc_inlbuf2 = (char *)&param->fp_mdt_index;
-       data.ioc_inllen2 = sizeof(param->fp_mdt_index);
+       data.ioc_inlbuf2 = (char *)lmu;
+       data.ioc_inllen2 = lmv_user_md_size(lmu->lum_stripe_count,
+                                           lmu->lum_magic);
        ret = llapi_ioctl_pack(&data, &rawbuf, sizeof(raw));
        if (ret != 0) {
                llapi_error(LLAPI_MSG_ERROR, ret,
@@ -4392,14 +4480,21 @@ migrate:
                        sync();
                        retry = true;
                        goto migrate;
+               } else if (errno == EALREADY) {
+                       if (param->fp_verbose & VERBOSE_DETAIL)
+                               fprintf(stdout,
+                                       "%s was migrated to MDT%d already\n",
+                                       path, lmu->lum_stripe_offset);
+                       ret = 0;
+               } else {
+                       ret = -errno;
+                       fprintf(stderr, "%s migrate failed: %s (%d)\n",
+                               path, strerror(-ret), ret);
+                       goto out;
                }
-               ret = -errno;
-               fprintf(stderr, "%s migrate failed: %s (%d)\n",
-                       path, strerror(-ret), ret);
-               goto out;
        } else if (param->fp_verbose & VERBOSE_DETAIL) {
-               fprintf(stdout, "migrate %s to MDT%d\n",
-                       path, param->fp_mdt_index);
+               fprintf(stdout, "migrate %s to MDT%d stripe count %d\n",
+                       path, lmu->lum_stripe_offset, lmu->lum_stripe_count);
        }
 
 out:
@@ -4425,9 +4520,41 @@ out:
        return ret;
 }
 
+/* dir migration finished, shrink its stripes */
+static int cb_migrate_mdt_fini(char *path, DIR *parent, DIR **dirp, void *data,
+                              struct dirent64 *de)
+{
+       struct find_param *param = data;
+       struct lmv_user_md *lmu = param->fp_lmv_md;
+       int lmulen = lmv_user_md_size(lmu->lum_stripe_count, lmu->lum_magic);
+       int ret = 0;
+
+       if (de && de->d_type != DT_DIR)
+               goto out;
+
+       if (*dirp) {
+               /*
+                * close it before setxattr because the latter may destroy the
+                * original object, and cause close fail.
+                */
+               ret = closedir(*dirp);
+               *dirp = NULL;
+               if (ret)
+                       goto out;
+       }
+
+       ret = setxattr(path, XATTR_NAME_LMV, lmu, lmulen, 0);
+       if (ret == -EALREADY)
+               ret = 0;
+out:
+       cb_common_fini(path, parent, dirp, data, de);
+       return ret;
+}
+
 int llapi_migrate_mdt(char *path, struct find_param *param)
 {
-       return param_callback(path, cb_migrate_mdt_init, cb_common_fini, param);
+       return param_callback(path, cb_migrate_mdt_init, cb_migrate_mdt_fini,
+                             param);
 }
 
 int llapi_mv(char *path, struct find_param *param)
@@ -4694,26 +4821,32 @@ int llapi_obd_statfs(char *path, __u32 type, __u32 index,
 
 int llapi_ping(char *obd_type, char *obd_name)
 {
+       int flags = O_RDONLY;
+       char buf[1] = { 0 };
        glob_t path;
-       char buf[1];
        int rc, fd;
 
        rc = cfs_get_param_paths(&path, "%s/%s/ping",
                                obd_type, obd_name);
        if (rc != 0)
                return -errno;
-
-       fd = open(path.gl_pathv[0], O_WRONLY);
+retry_open:
+       fd = open(path.gl_pathv[0], flags);
        if (fd < 0) {
+               if (errno == EACCES && flags == O_RDONLY) {
+                       flags = O_WRONLY;
+                       goto retry_open;
+               }
                rc = -errno;
                llapi_error(LLAPI_MSG_ERROR, rc, "error opening %s",
                            path.gl_pathv[0]);
                goto failed;
        }
 
-       /* The purpose is to send a byte as a ping, whatever this byte is. */
-       /* coverity[uninit_use_in_call] */
-       rc = write(fd, buf, 1);
+       if (flags == O_RDONLY)
+               rc = read(fd, buf, sizeof(buf));
+       else
+               rc = write(fd, buf, sizeof(buf));
        if (rc < 0)
                rc = -errno;
        close(fd);
@@ -5169,75 +5302,120 @@ int llapi_get_ost_layout_version(int fd, __u32 *layout_version)
 }
 
 /*
- * Create a file without any name open it for read/write
+ * Create a file without any name and open it for read/write
  *
  * - file is created as if it were a standard file in the given \a directory
  * - file does not appear in \a directory and mtime does not change because
  *   the filename is handled specially by the Lustre MDS.
- * - file is removed at final close
- * - file modes are rw------- since it doesn't make sense to have a read-only
- *   or write-only file that cannot be opened again.
- * - if user wants another mode it must use fchmod() on the open file, no
- *   security problems arise because it cannot be opened by another process.
+ * - file is destroyed at final close
  *
  * \param[in]  directory       directory from which to inherit layout/MDT idx
- * \param[in]  idx             MDT index on which the file is created,
+ * \param[in]  mdt_idx         MDT index on which the file is created,
  *                             \a idx == -1 means no specific MDT is requested
- * \param[in]  open_flags      standard open(2) flags
+ * \param[in]  mode            standard open(2) mode
+ * \param[in]  stripe_param    stripe parameters. May be NULL.
  *
- * \retval     0 on success.
+ * \retval     a file descriptor on success.
  * \retval     -errno on error.
  */
-int llapi_create_volatile_idx(char *directory, int idx, int open_flags)
+int llapi_create_volatile_param(const char *directory, int mdt_idx,
+                               int open_flags, mode_t mode,
+                               const struct llapi_stripe_param *stripe_param)
 {
-       char    file_path[PATH_MAX];
-       char    filename[PATH_MAX];
-       int     saved_errno = errno;
-       int     fd;
-       int     rnumber;
-       int     rc;
+       char file_path[PATH_MAX];
+       int saved_errno = errno;
+       int fd;
+       unsigned int rnumber;
+       int rc;
 
        do {
                rnumber = random();
-               if (idx == -1)
-                       snprintf(filename, sizeof(filename),
-                                LUSTRE_VOLATILE_HDR"::%.4X", rnumber);
+               if (mdt_idx == -1)
+                       rc = snprintf(file_path, sizeof(file_path),
+                                     "%s/" LUSTRE_VOLATILE_HDR "::%.4X",
+                                     directory, rnumber);
                else
-                       snprintf(filename, sizeof(filename),
-                                LUSTRE_VOLATILE_HDR":%.4X:%.4X", idx, rnumber);
+                       rc = snprintf(file_path, sizeof(file_path),
+                                     "%s/" LUSTRE_VOLATILE_HDR ":%.4X:%.4X",
+                                     directory, mdt_idx, rnumber);
+
+               if (rc < 0 || rc >= sizeof(file_path))
+                       return -ENAMETOOLONG;
 
-               rc = snprintf(file_path, sizeof(file_path),
-                             "%s/%s", directory, filename);
-               if (rc >= sizeof(file_path))
-                       return -E2BIG;
+               /*
+                * Either open O_WRONLY or O_RDWR, creating RDONLY
+                * is non-sensical here
+                */
+               if ((open_flags & O_ACCMODE) == O_RDONLY)
+                       open_flags = O_RDWR | (open_flags & ~O_ACCMODE);
 
-               fd = open(file_path,
-                         O_RDWR | O_CREAT | O_EXCL | O_NOFOLLOW | open_flags,
-                         S_IRUSR | S_IWUSR);
-       } while (fd < 0 && errno == EEXIST);
+               open_flags |= O_CREAT | O_EXCL | O_NOFOLLOW;
+
+               if (stripe_param != NULL) {
+                       fd = llapi_file_open_param(file_path, open_flags,
+                                                  mode, stripe_param);
+                       if (fd < 0)
+                               rc = fd;
+               } else {
+                       fd = open(file_path, open_flags, mode);
+                       if (fd < 0)
+                               rc = -errno;
+               }
+       } while (fd < 0 && rc == -EEXIST);
 
        if (fd < 0) {
-               llapi_error(LLAPI_MSG_ERROR, errno,
+               llapi_error(LLAPI_MSG_ERROR, rc,
                            "Cannot create volatile file '%s' in '%s'",
-                           filename + LUSTRE_VOLATILE_HDR_LEN,
+                           file_path + strlen(directory) + 1 +
+                           LUSTRE_VOLATILE_HDR_LEN,
                            directory);
-               return -errno;
+               return rc;
        }
 
-       /* Unlink file in case this wasn't a Lustre filesystem and the
-        * magic volatile filename wasn't handled as intended. The
-        * effect is the same. If volatile open was supported then we
-        * expect unlink() to return -ENOENT. */
+       /*
+        * Unlink file in case this wasn't a Lustre filesystem and the magic
+        * volatile filename wasn't handled as intended. The effect is the
+        * same. If volatile open was supported then we expect unlink() to
+        * return -ENOENT.
+        */
        (void)unlink(file_path);
 
-       /* Since we are returning successfully we restore errno (and
-        * mask out possible EEXIST from open() and ENOENT from
-        * unlink(). */
+       /*
+        * Since we are returning successfully we restore errno (and
+        * mask out possible EEXIST from open() and ENOENT from unlink().
+        */
        errno = saved_errno;
 
        return fd;
 }
 
+/*
+ * Create a file without any name open it for read/write
+ *
+ * - file is created as if it were a standard file in the given \a directory
+ * - file does not appear in \a directory and mtime does not change because
+ *   the filename is handled specially by the Lustre MDS.
+ * - file is removed at final close
+ * - file modes are rw------- since it doesn't make sense to have a read-only
+ *   or write-only file that cannot be opened again.
+ * - if user wants another mode it must use fchmod() on the open file, no
+ *   security problems arise because it cannot be opened by another process.
+ *
+ * \param[in]  directory       directory from which to inherit layout/MDT idx
+ * \param[in]  idx             MDT index on which the file is created,
+ *                             \a idx == -1 means no specific MDT is requested
+ * \param[in]  open_flags      standard open(2) flags
+ *
+ * \retval     a file descriptor on success.
+ * \retval     -errno on error.
+ */
+int llapi_create_volatile_idx(const char *directory, int mdt_idx,
+                             int open_flags)
+{
+       return llapi_create_volatile_param(directory, mdt_idx, open_flags,
+                                          S_IRUSR | S_IWUSR, NULL);
+}
+
 /**
  * Swap the layouts between 2 file descriptors
  * the 2 files must be open for writing