Whamcloud - gitweb
LU-8929 lfsck: dumper gets current position properly
[fs/lustre-release.git] / lustre / lfsck / lfsck_lib.c
index ed2a3d8..98f7e24 100644 (file)
@@ -20,7 +20,7 @@
  * GPL HEADER END
  */
 /*
- * Copyright (c) 2013, 2015, Intel Corporation.
+ * Copyright (c) 2013, 2016, Intel Corporation.
  */
 /*
  * lustre/lfsck/lfsck_lib.c
@@ -31,7 +31,8 @@
 #define DEBUG_SUBSYSTEM S_LFSCK
 
 #include <linux/kthread.h>
-#include <libcfs/list.h>
+#include <linux/sched.h>
+#include <linux/list.h>
 #include <lu_object.h>
 #include <dt_object.h>
 #include <md_object.h>
@@ -67,21 +68,6 @@ static struct list_head lfsck_ost_orphan_list;
 static struct list_head lfsck_mdt_orphan_list;
 static DEFINE_SPINLOCK(lfsck_instance_lock);
 
-static const char *lfsck_status_names[] = {
-       [LS_INIT]               = "init",
-       [LS_SCANNING_PHASE1]    = "scanning-phase1",
-       [LS_SCANNING_PHASE2]    = "scanning-phase2",
-       [LS_COMPLETED]          = "completed",
-       [LS_FAILED]             = "failed",
-       [LS_STOPPED]            = "stopped",
-       [LS_PAUSED]             = "paused",
-       [LS_CRASHED]            = "crashed",
-       [LS_PARTIAL]            = "partial",
-       [LS_CO_FAILED]          = "co-failed",
-       [LS_CO_STOPPED]         = "co-stopped",
-       [LS_CO_PAUSED]          = "co-paused"
-};
-
 const char *lfsck_flags_names[] = {
        "scanned-once",
        "inconsistent",
@@ -100,6 +86,8 @@ const char *lfsck_param_names[] = {
        "orphan",
        "create_ostobj",
        "create_mdtobj",
+       NULL,
+       "delay_create_ostobj",
        NULL
 };
 
@@ -108,12 +96,16 @@ enum lfsck_verify_lpf_types {
        LVLT_BY_NAMEENTRY       = 1,
 };
 
-const char *lfsck_status2names(enum lfsck_status status)
+static inline void
+lfsck_reset_ltd_status(struct lfsck_tgt_desc *ltd, enum lfsck_type type)
 {
-       if (unlikely(status < 0 || status >= LS_MAX))
-               return "unknown";
-
-       return lfsck_status_names[status];
+       if (type == LFSCK_TYPE_LAYOUT) {
+               ltd->ltd_layout_status = LS_MAX;
+               ltd->ltd_layout_repaired = 0;
+       } else {
+               ltd->ltd_namespace_status = LS_MAX;
+               ltd->ltd_namespace_repaired = 0;
+       }
 }
 
 static int lfsck_tgt_descs_init(struct lfsck_tgt_descs *ltds)
@@ -201,8 +193,8 @@ static int __lfsck_add_target(const struct lu_env *env,
        if (index >= ltds->ltd_tgts_bitmap->size) {
                __u32 newsize = max((__u32)ltds->ltd_tgts_bitmap->size,
                                    (__u32)BITS_PER_LONG);
-               cfs_bitmap_t *old_bitmap = ltds->ltd_tgts_bitmap;
-               cfs_bitmap_t *new_bitmap;
+               struct cfs_bitmap *old_bitmap = ltds->ltd_tgts_bitmap;
+               struct cfs_bitmap *new_bitmap;
 
                while (newsize < index + 1)
                        newsize <<= 1;
@@ -541,6 +533,14 @@ int lfsck_find_mdt_idx_by_fid(const struct lu_env *env,
        struct lu_seq_range     *range  = &lfsck_env_info(env)->lti_range;
        int                      rc;
 
+       if (unlikely(fid_seq(fid) == FID_SEQ_LOCAL_FILE)) {
+               /* "ROOT" is always on the MDT0. */
+               if (lu_fid_eq(fid, &lfsck->li_global_root_fid))
+                       return 0;
+
+               return lfsck_dev_idx(lfsck);
+       }
+
        fld_range_set_mdt(range);
        rc = fld_server_lookup(env, ss->ss_server_fld, fid_seq(fid), range);
        if (rc == 0)
@@ -644,13 +644,9 @@ static int lfsck_create_lpf_local(const struct lu_env *env,
        int                      rc;
        ENTRY;
 
-       rc = linkea_data_new(&ldata,
-                            &lfsck_env_info(env)->lti_linkea_buf2);
-       if (rc != 0)
-               RETURN(rc);
-
        cname = lfsck_name_get_const(env, name, strlen(name));
-       rc = linkea_add_buf(&ldata, cname, lfsck_dto2fid(parent));
+       rc = linkea_links_new(&ldata, &lfsck_env_info(env)->lti_linkea_buf2,
+                             cname, lfsck_dto2fid(parent));
        if (rc != 0)
                RETURN(rc);
 
@@ -663,12 +659,30 @@ static int lfsck_create_lpf_local(const struct lu_env *env,
        if (rc != 0)
                GOTO(stop, rc);
 
+       if (!dt_try_as_dir(env, child))
+               GOTO(stop, rc = -ENOTDIR);
+
        /* 2a. increase child nlink */
        rc = dt_declare_ref_add(env, child, th);
        if (rc != 0)
                GOTO(stop, rc);
 
-       /* 3a. insert linkEA for child */
+       /* 3a. insert dot into child dir */
+       rec->rec_type = S_IFDIR;
+       rec->rec_fid = cfid;
+       rc = dt_declare_insert(env, child, (const struct dt_rec *)rec,
+                              (const struct dt_key *)dot, th);
+       if (rc != 0)
+               GOTO(stop, rc);
+
+       /* 4a. insert dotdot into child dir */
+       rec->rec_fid = &LU_LPF_FID;
+       rc = dt_declare_insert(env, child, (const struct dt_rec *)rec,
+                              (const struct dt_key *)dotdot, th);
+       if (rc != 0)
+               GOTO(stop, rc);
+
+       /* 5a. insert linkEA for child */
        lfsck_buf_init(&linkea_buf, ldata.ld_buf->lb_buf,
                       ldata.ld_leh->leh_len);
        rc = dt_declare_xattr_set(env, child, &linkea_buf,
@@ -676,7 +690,7 @@ static int lfsck_create_lpf_local(const struct lu_env *env,
        if (rc != 0)
                GOTO(stop, rc);
 
-       /* 4a. insert name into parent dir */
+       /* 6a. insert name into parent dir */
        rec->rec_type = S_IFDIR;
        rec->rec_fid = cfid;
        rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
@@ -684,12 +698,12 @@ static int lfsck_create_lpf_local(const struct lu_env *env,
        if (rc != 0)
                GOTO(stop, rc);
 
-       /* 5a. increase parent nlink */
+       /* 7a. increase parent nlink */
        rc = dt_declare_ref_add(env, parent, th);
        if (rc != 0)
                GOTO(stop, rc);
 
-       /* 6a. update bookmark */
+       /* 8a. update bookmark */
        rc = dt_declare_record_write(env, bk_obj,
                                     lfsck_buf_get(env, bk, len), 0, th);
        if (rc != 0)
@@ -700,41 +714,38 @@ static int lfsck_create_lpf_local(const struct lu_env *env,
                GOTO(stop, rc);
 
        dt_write_lock(env, child, 0);
-       /* 1b.1. create child */
+       /* 1b. create child */
        rc = dt_create(env, child, la, NULL, dof, th);
        if (rc != 0)
                GOTO(unlock, rc);
 
-       if (unlikely(!dt_try_as_dir(env, child)))
-               GOTO(unlock, rc = -ENOTDIR);
+       /* 2b. increase child nlink */
+       rc = dt_ref_add(env, child, th);
+       if (rc != 0)
+               GOTO(unlock, rc);
 
-       /* 1b.2. insert dot into child dir */
+       /* 3b. insert dot into child dir */
        rec->rec_fid = cfid;
        rc = dt_insert(env, child, (const struct dt_rec *)rec,
                       (const struct dt_key *)dot, th, 1);
        if (rc != 0)
                GOTO(unlock, rc);
 
-       /* 1b.3. insert dotdot into child dir */
+       /* 4b. insert dotdot into child dir */
        rec->rec_fid = &LU_LPF_FID;
        rc = dt_insert(env, child, (const struct dt_rec *)rec,
                       (const struct dt_key *)dotdot, th, 1);
        if (rc != 0)
                GOTO(unlock, rc);
 
-       /* 2b. increase child nlink */
-       rc = dt_ref_add(env, child, th);
-       if (rc != 0)
-               GOTO(unlock, rc);
-
-       /* 3b. insert linkEA for child. */
+       /* 5b. insert linkEA for child. */
        rc = dt_xattr_set(env, child, &linkea_buf,
                          XATTR_NAME_LINK, 0, th);
        dt_write_unlock(env, child);
        if (rc != 0)
                GOTO(stop, rc);
 
-       /* 4b. insert name into parent dir */
+       /* 6b. insert name into parent dir */
        rec->rec_fid = cfid;
        rc = dt_insert(env, parent, (const struct dt_rec *)rec,
                       (const struct dt_key *)name, th, 1);
@@ -742,7 +753,7 @@ static int lfsck_create_lpf_local(const struct lu_env *env,
                GOTO(stop, rc);
 
        dt_write_lock(env, parent, 0);
-       /* 5b. increase parent nlink */
+       /* 7b. increase parent nlink */
        rc = dt_ref_add(env, parent, th);
        dt_write_unlock(env, parent);
        if (rc != 0)
@@ -751,7 +762,7 @@ static int lfsck_create_lpf_local(const struct lu_env *env,
        bk->lb_lpf_fid = *cfid;
        lfsck_bookmark_cpu_to_le(&lfsck->li_bookmark_disk, bk);
 
-       /* 6b. update bookmark */
+       /* 8b. update bookmark */
        rc = dt_record_write(env, bk_obj,
                             lfsck_buf_get(env, bk, len), &pos, th);
 
@@ -788,13 +799,9 @@ static int lfsck_create_lpf_remote(const struct lu_env *env,
        int                      rc;
        ENTRY;
 
-       rc = linkea_data_new(&ldata,
-                            &lfsck_env_info(env)->lti_linkea_buf2);
-       if (rc != 0)
-               RETURN(rc);
-
        cname = lfsck_name_get_const(env, name, strlen(name));
-       rc = linkea_add_buf(&ldata, cname, lfsck_dto2fid(parent));
+       rc = linkea_links_new(&ldata, &lfsck_env_info(env)->lti_linkea_buf2,
+                             cname, lfsck_dto2fid(parent));
        if (rc != 0)
                RETURN(rc);
 
@@ -830,12 +837,30 @@ static int lfsck_create_lpf_remote(const struct lu_env *env,
        if (rc != 0)
                GOTO(stop, rc);
 
+       if (!dt_try_as_dir(env, child))
+               GOTO(stop, rc = -ENOTDIR);
+
        /* 2a. increase child nlink */
        rc = dt_declare_ref_add(env, child, th);
        if (rc != 0)
                GOTO(stop, rc);
 
-       /* 3a. insert linkEA for child */
+       /* 3a. insert dot into child dir */
+       rec->rec_type = S_IFDIR;
+       rec->rec_fid = cfid;
+       rc = dt_declare_insert(env, child, (const struct dt_rec *)rec,
+                              (const struct dt_key *)dot, th);
+       if (rc != 0)
+               GOTO(stop, rc);
+
+       /* 4a. insert dotdot into child dir */
+       rec->rec_fid = &LU_LPF_FID;
+       rc = dt_declare_insert(env, child, (const struct dt_rec *)rec,
+                              (const struct dt_key *)dotdot, th);
+       if (rc != 0)
+               GOTO(stop, rc);
+
+       /* 5a. insert linkEA for child */
        lfsck_buf_init(&linkea_buf, ldata.ld_buf->lb_buf,
                       ldata.ld_leh->leh_len);
        rc = dt_declare_xattr_set(env, child, &linkea_buf,
@@ -843,7 +868,7 @@ static int lfsck_create_lpf_remote(const struct lu_env *env,
        if (rc != 0)
                GOTO(stop, rc);
 
-       /* 4a. update bookmark */
+       /* 6a. update bookmark */
        rc = dt_declare_record_write(env, bk_obj,
                                     lfsck_buf_get(env, bk, len), 0, th);
        if (rc != 0)
@@ -854,15 +879,17 @@ static int lfsck_create_lpf_remote(const struct lu_env *env,
                GOTO(stop, rc);
 
        dt_write_lock(env, child, 0);
-       /* 1b.1. create child */
+       /* 1b. create child */
        rc = dt_create(env, child, la, NULL, dof, th);
        if (rc != 0)
                GOTO(unlock, rc);
 
-       if (unlikely(!dt_try_as_dir(env, child)))
-               GOTO(unlock, rc = -ENOTDIR);
+       /* 2b. increase child nlink */
+       rc = dt_ref_add(env, child, th);
+       if (rc != 0)
+               GOTO(unlock, rc);
 
-       /* 1b.2. insert dot into child dir */
+       /* 3b. insert dot into child dir */
        rec->rec_type = S_IFDIR;
        rec->rec_fid = cfid;
        rc = dt_insert(env, child, (const struct dt_rec *)rec,
@@ -870,19 +897,14 @@ static int lfsck_create_lpf_remote(const struct lu_env *env,
        if (rc != 0)
                GOTO(unlock, rc);
 
-       /* 1b.3. insert dotdot into child dir */
+       /* 4b. insert dotdot into child dir */
        rec->rec_fid = &LU_LPF_FID;
        rc = dt_insert(env, child, (const struct dt_rec *)rec,
                       (const struct dt_key *)dotdot, th, 1);
        if (rc != 0)
                GOTO(unlock, rc);
 
-       /* 2b. increase child nlink */
-       rc = dt_ref_add(env, child, th);
-       if (rc != 0)
-               GOTO(unlock, rc);
-
-       /* 3b. insert linkEA for child */
+       /* 5b. insert linkEA for child */
        rc = dt_xattr_set(env, child, &linkea_buf,
                          XATTR_NAME_LINK, 0, th);
        if (rc != 0)
@@ -891,7 +913,7 @@ static int lfsck_create_lpf_remote(const struct lu_env *env,
        bk->lb_lpf_fid = *cfid;
        lfsck_bookmark_cpu_to_le(&lfsck->li_bookmark_disk, bk);
 
-       /* 4b. update bookmark */
+       /* 6b. update bookmark */
        rc = dt_record_write(env, bk_obj,
                             lfsck_buf_get(env, bk, len), &pos, th);
 
@@ -1014,7 +1036,7 @@ static int lfsck_create_lpf(const struct lu_env *env,
                *cfid = bk->lb_lpf_fid;
        }
 
-       child = lfsck_object_find_bottom(env, lfsck, cfid);
+       child = lfsck_object_find_bottom_new(env, lfsck, cfid);
        if (IS_ERR(child))
                GOTO(unlock, rc = PTR_ERR(child));
 
@@ -1405,61 +1427,13 @@ int lfsck_verify_lpf(const struct lu_env *env, struct lfsck_instance *lfsck)
                               lfsck_lfsck2name(lfsck), rc);
        }
 
-       if (!fid_is_zero(&bk->lb_lpf_fid)) {
-               if (unlikely(!fid_is_norm(&bk->lb_lpf_fid))) {
-                       struct lu_fid tfid = bk->lb_lpf_fid;
-
-                       /* Invalid FID record in the bookmark file, reset it. */
-                       fid_zero(&bk->lb_lpf_fid);
-                       rc = lfsck_bookmark_store(env, lfsck);
-
-                       CDEBUG(D_LFSCK, "%s: reset invalid LPF fid "DFID
-                              " in the bookmark file: rc = %d\n",
-                              lfsck_lfsck2name(lfsck), PFID(&tfid), rc);
-
-                       if (rc != 0)
-                               GOTO(put, rc);
-               } else {
-                       child1 = lfsck_object_find_bottom(env, lfsck,
-                                                         &bk->lb_lpf_fid);
-                       if (IS_ERR(child1)) {
-                               child1 = NULL;
-                               goto find_child2;
-                       }
-
-                       if (unlikely(!dt_object_exists(child1) ||
-                                    dt_object_remote(child1)) ||
-                                    !S_ISDIR(lfsck_object_type(child1))) {
-                               /* Invalid FID record in the bookmark file,
-                                * reset it. */
-                               fid_zero(&bk->lb_lpf_fid);
-                               rc = lfsck_bookmark_store(env, lfsck);
-
-                               CDEBUG(D_LFSCK, "%s: reset invalid LPF fid "DFID
-                                      " in the bookmark file: rc = %d\n",
-                                      lfsck_lfsck2name(lfsck),
-                                      PFID(lfsck_dto2fid(child1)), rc);
-
-                               if (rc != 0)
-                                       GOTO(put, rc);
-
-                               lfsck_object_put(env, child1);
-                               child1 = NULL;
-                       } else if (unlikely(!dt_try_as_dir(env, child1))) {
-                               GOTO(put, rc = -ENOTDIR);
-                       }
-               }
-       }
-
-find_child2:
+       /* child2 */
        snprintf(name, 8, "MDT%04x", node);
        rc = dt_lookup(env, parent, (struct dt_rec *)cfid,
                       (const struct dt_key *)name);
        if (rc == -ENOENT) {
-               if (!fid_is_zero(&bk->lb_lpf_fid))
-                       goto check_child1;
-
-               GOTO(put, rc = 0);
+               rc = 0;
+               goto find_child1;
        }
 
        if (rc != 0)
@@ -1471,7 +1445,7 @@ find_child2:
                if (rc != 0)
                        GOTO(put, rc);
 
-               goto check_child1;
+               goto find_child1;
        }
 
        child2 = lfsck_object_find_bottom(env, lfsck, cfid);
@@ -1485,23 +1459,20 @@ find_child2:
                if (rc != 0)
                        GOTO(put, rc);
 
-               goto check_child1;
+               goto find_child1;
        }
 
-       if (unlikely(!dt_try_as_dir(env, child2)))
-               GOTO(put, rc = -ENOTDIR);
+       if (unlikely(!dt_try_as_dir(env, child2))) {
+               lfsck_object_put(env, child2);
+               child2 = NULL;
+               rc = -ENOTDIR;
+       }
 
-       if (child1 == NULL) {
-               rc = lfsck_verify_lpf_pairs(env, lfsck, child2, name,
-                                           pfid, LVLT_BY_NAMEENTRY);
-       } else if (!lu_fid_eq(cfid, &bk->lb_lpf_fid)) {
-               rc = lfsck_verify_lpf_pairs(env, lfsck, child1, name,
-                                           pfid, LVLT_BY_BOOKMARK);
-               if (!lu_fid_eq(pfid, &LU_LPF_FID))
-                       rc = lfsck_verify_lpf_pairs(env, lfsck, child2,
-                                                   name, pfid,
-                                                   LVLT_BY_NAMEENTRY);
-       } else {
+find_child1:
+       if (fid_is_zero(&bk->lb_lpf_fid))
+               goto check_child2;
+
+       if (likely(lu_fid_eq(cfid, &bk->lb_lpf_fid))) {
                if (lfsck->li_lpf_obj == NULL) {
                        lu_object_get(&child2->do_lu);
                        lfsck->li_lpf_obj = child2;
@@ -1509,14 +1480,69 @@ find_child2:
 
                cname = lfsck_name_get_const(env, name, strlen(name));
                rc = lfsck_verify_linkea(env, child2, cname, &LU_LPF_FID);
+
+               GOTO(put, rc);
        }
 
-       GOTO(put, rc);
+       if (unlikely(!fid_is_norm(&bk->lb_lpf_fid))) {
+               struct lu_fid tfid = bk->lb_lpf_fid;
 
-check_child1:
-       if (child1 != NULL)
-               rc = lfsck_verify_lpf_pairs(env, lfsck, child1, name,
-                                           pfid, LVLT_BY_BOOKMARK);
+               /* Invalid FID record in the bookmark file, reset it. */
+               fid_zero(&bk->lb_lpf_fid);
+               rc = lfsck_bookmark_store(env, lfsck);
+
+               CDEBUG(D_LFSCK, "%s: reset invalid LPF fid "DFID
+                      " in the bookmark file: rc = %d\n",
+                      lfsck_lfsck2name(lfsck), PFID(&tfid), rc);
+
+               if (rc != 0)
+                       GOTO(put, rc);
+
+               goto check_child2;
+       }
+
+       child1 = lfsck_object_find_bottom(env, lfsck, &bk->lb_lpf_fid);
+       if (IS_ERR(child1)) {
+               child1 = NULL;
+               goto check_child2;
+       }
+
+       if (unlikely(!dt_object_exists(child1) ||
+                    dt_object_remote(child1)) ||
+                    !S_ISDIR(lfsck_object_type(child1))) {
+               /* Invalid FID record in the bookmark file, reset it. */
+               fid_zero(&bk->lb_lpf_fid);
+               rc = lfsck_bookmark_store(env, lfsck);
+
+               CDEBUG(D_LFSCK, "%s: reset invalid LPF fid "DFID
+                      " in the bookmark file: rc = %d\n",
+                      lfsck_lfsck2name(lfsck),
+                      PFID(lfsck_dto2fid(child1)), rc);
+
+               if (rc != 0)
+                       GOTO(put, rc);
+
+               lfsck_object_put(env, child1);
+               child1 = NULL;
+               goto check_child2;
+       }
+
+       if (unlikely(!dt_try_as_dir(env, child1))) {
+               lfsck_object_put(env, child1);
+               child1 = NULL;
+               rc = -ENOTDIR;
+               goto check_child2;
+       }
+
+       rc = lfsck_verify_lpf_pairs(env, lfsck, child1, name, pfid,
+                                   LVLT_BY_BOOKMARK);
+       if (lu_fid_eq(pfid, &LU_LPF_FID))
+               GOTO(put, rc);
+
+check_child2:
+       if (child2 != NULL)
+               rc = lfsck_verify_lpf_pairs(env, lfsck, child2, name,
+                                           pfid, LVLT_BY_NAMEENTRY);
 
        GOTO(put, rc);
 
@@ -1714,17 +1740,14 @@ static inline int lfsck_instance_add(struct lfsck_instance *lfsck)
        return 0;
 }
 
-int lfsck_bits_dump(struct seq_file *m, int bits, const char *names[],
-                   const char *prefix)
+void lfsck_bits_dump(struct seq_file *m, int bits, const char *names[],
+                    const char *prefix)
 {
        int flag;
        int i;
        bool newline = (bits != 0 ? false : true);
-       int rc;
 
-       rc = seq_printf(m, "%s:%c", prefix, bits != 0 ? ' ' : '\n');
-       if (rc < 0)
-               return rc;
+       seq_printf(m, "%s:%c", prefix, bits != 0 ? ' ' : '\n');
 
        for (i = 0, flag = 1; bits != 0; i++, flag = 1 << i) {
                if (flag & bits) {
@@ -1733,54 +1756,43 @@ int lfsck_bits_dump(struct seq_file *m, int bits, const char *names[],
                                if (bits == 0)
                                        newline = true;
 
-                               rc = seq_printf(m, "%s%c", names[i],
-                                               newline ? '\n' : ',');
-                               if (rc < 0)
-                                       return rc;
+                               seq_printf(m, "%s%c", names[i],
+                                          newline ? '\n' : ',');
                        }
                }
        }
 
        if (!newline)
-               rc = seq_printf(m, "\n");
-
-       return rc;
+               seq_putc(m, '\n');
 }
 
-int lfsck_time_dump(struct seq_file *m, __u64 time, const char *name)
+void lfsck_time_dump(struct seq_file *m, __u64 time, const char *name)
 {
-       int rc;
-
        if (time == 0) {
-               rc = seq_printf(m, "%s_time: N/A\n", name);
-               if (rc == 0)
-                       rc = seq_printf(m, "time_since_%s: N/A\n", name);
-
-               return rc;
+               seq_printf(m, "%s_time: N/A\n", name);
+               seq_printf(m, "time_since_%s: N/A\n", name);
+       } else {
+               seq_printf(m, "%s_time: %llu\n", name, time);
+               seq_printf(m, "time_since_%s: %llu seconds\n",
+                          name, cfs_time_current_sec() - time);
        }
-
-       rc = seq_printf(m, "%s_time: "LPU64"\n", name, time);
-       if (rc == 0)
-               rc = seq_printf(m, "time_since_%s: "LPU64" seconds\n",
-                               name, cfs_time_current_sec() - time);
-
-       return rc;
 }
 
-int lfsck_pos_dump(struct seq_file *m, struct lfsck_position *pos,
-                  const char *prefix)
+void lfsck_pos_dump(struct seq_file *m, struct lfsck_position *pos,
+                   const char *prefix)
 {
        if (fid_is_zero(&pos->lp_dir_parent)) {
-               if (pos->lp_oit_cookie == 0)
-                       return seq_printf(m, "%s: N/A, N/A, N/A\n", prefix);
-
-               return seq_printf(m, "%s: "LPU64", N/A, N/A\n",
-                                 prefix, pos->lp_oit_cookie);
+               if (pos->lp_oit_cookie == 0) {
+                       seq_printf(m, "%s: N/A, N/A, N/A\n", prefix);
+                       return;
+               }
+               seq_printf(m, "%s: %llu, N/A, N/A\n",
+                          prefix, pos->lp_oit_cookie);
+       } else {
+               seq_printf(m, "%s: %llu, "DFID", %#llx\n",
+                          prefix, pos->lp_oit_cookie,
+                          PFID(&pos->lp_dir_parent), pos->lp_dir_cookie);
        }
-
-       return seq_printf(m, "%s: "LPU64", "DFID", "LPX64"\n",
-                         prefix, pos->lp_oit_cookie,
-                         PFID(&pos->lp_dir_parent), pos->lp_dir_cookie);
 }
 
 void lfsck_pos_fill(const struct lu_env *env, struct lfsck_instance *lfsck,
@@ -2122,6 +2134,11 @@ int lfsck_async_interpret_common(const struct lu_env *env,
                }
 
                if (rc != 0) {
+                       if (lr->lr_flags & LEF_QUERY_ALL) {
+                               lfsck_reset_ltd_status(ltd, com->lc_type);
+                               break;
+                       }
+
                        spin_lock(&ltds->ltd_lock);
                        list_del_init(phase_list);
                        list_del_init(list);
@@ -2136,6 +2153,12 @@ int lfsck_async_interpret_common(const struct lu_env *env,
                        CDEBUG(D_LFSCK, "%s: invalid query reply for %s: "
                               "rc = %d\n", lfsck_lfsck2name(com->lc_lfsck),
                               lad->lad_name, rc);
+
+                       if (lr->lr_flags & LEF_QUERY_ALL) {
+                               lfsck_reset_ltd_status(ltd, com->lc_type);
+                               break;
+                       }
+
                        spin_lock(&ltds->ltd_lock);
                        list_del_init(phase_list);
                        list_del_init(list);
@@ -2143,6 +2166,18 @@ int lfsck_async_interpret_common(const struct lu_env *env,
                        break;
                }
 
+               if (lr->lr_flags & LEF_QUERY_ALL) {
+                       if (com->lc_type == LFSCK_TYPE_LAYOUT) {
+                               ltd->ltd_layout_status = reply->lr_status;
+                               ltd->ltd_layout_repaired = reply->lr_repaired;
+                       } else {
+                               ltd->ltd_namespace_status = reply->lr_status;
+                               ltd->ltd_namespace_repaired =
+                                                       reply->lr_repaired;
+                       }
+                       break;
+               }
+
                switch (reply->lr_status) {
                case LS_SCANNING_PHASE1:
                        break;
@@ -2288,12 +2323,12 @@ static int lfsck_stop_notify(const struct lu_env *env,
                if (ltds == &lfsck->li_ost_descs)
                        lr->lr_flags = LEF_TO_OST;
 
+               memset(laia, 0, sizeof(*laia));
                laia->laia_com = com;
                laia->laia_ltds = ltds;
                atomic_inc(&ltd->ltd_ref);
                laia->laia_ltd = ltd;
                laia->laia_lr = lr;
-               laia->laia_shared = 0;
 
                rc = lfsck_async_request(env, ltd->ltd_exp, lr, set,
                                         lfsck_async_interpret_common,
@@ -2378,11 +2413,76 @@ int lfsck_async_request(const struct lu_env *env, struct obd_export *exp,
        if (laia->laia_com != NULL)
                lfsck_component_get(laia->laia_com);
        req->rq_interpret_reply = interpreter;
+       req->rq_allow_intr = 1;
        ptlrpc_set_add_req(set, req);
 
        return 0;
 }
 
+int lfsck_query_all(const struct lu_env *env, struct lfsck_component *com)
+{
+       struct lfsck_thread_info          *info  = lfsck_env_info(env);
+       struct lfsck_request              *lr    = &info->lti_lr;
+       struct lfsck_async_interpret_args *laia  = &info->lti_laia;
+       struct lfsck_instance             *lfsck = com->lc_lfsck;
+       struct lfsck_tgt_descs            *ltds  = &lfsck->li_mdt_descs;
+       struct lfsck_tgt_desc             *ltd;
+       struct ptlrpc_request_set         *set;
+       int                                idx;
+       int                                rc;
+       ENTRY;
+
+       memset(lr, 0, sizeof(*lr));
+       lr->lr_event = LE_QUERY;
+       lr->lr_active = com->lc_type;
+       lr->lr_flags = LEF_QUERY_ALL;
+
+       memset(laia, 0, sizeof(*laia));
+       laia->laia_com = com;
+       laia->laia_lr = lr;
+
+       set = ptlrpc_prep_set();
+       if (set == NULL)
+               RETURN(-ENOMEM);
+
+again:
+       laia->laia_ltds = ltds;
+       down_read(&ltds->ltd_rw_sem);
+       cfs_foreach_bit(ltds->ltd_tgts_bitmap, idx) {
+               ltd = lfsck_tgt_get(ltds, idx);
+               LASSERT(ltd != NULL);
+
+               laia->laia_ltd = ltd;
+               up_read(&ltds->ltd_rw_sem);
+               rc = lfsck_async_request(env, ltd->ltd_exp, lr, set,
+                                        lfsck_async_interpret_common,
+                                        laia, LFSCK_QUERY);
+               if (rc != 0) {
+                       struct lfsck_assistant_data *lad = com->lc_data;
+
+                       CDEBUG(D_LFSCK, "%s: Fail to query %s %x for stat %s: "
+                              "rc = %d\n", lfsck_lfsck2name(lfsck),
+                              (lr->lr_flags & LEF_TO_OST) ? "OST" : "MDT",
+                              ltd->ltd_index, lad->lad_name, rc);
+                       lfsck_reset_ltd_status(ltd, com->lc_type);
+                       lfsck_tgt_put(ltd);
+               }
+               down_read(&ltds->ltd_rw_sem);
+       }
+       up_read(&ltds->ltd_rw_sem);
+
+       if (com->lc_type == LFSCK_TYPE_LAYOUT && !(lr->lr_flags & LEF_TO_OST)) {
+               ltds = &lfsck->li_ost_descs;
+               lr->lr_flags |= LEF_TO_OST;
+               goto again;
+       }
+
+       rc = ptlrpc_set_wait(set);
+       ptlrpc_set_destroy(set);
+
+       RETURN(rc);
+}
+
 int lfsck_start_assistant(const struct lu_env *env, struct lfsck_component *com,
                          struct lfsck_start_param *lsp)
 {
@@ -2463,6 +2563,9 @@ void lfsck_post_generic(const struct lu_env *env,
                lad->lad_exit = 1;
        lad->lad_to_post = 1;
 
+       CDEBUG(D_LFSCK, "%s: waiting for assistant to do %s post, rc = %d\n",
+              lfsck_lfsck2name(com->lc_lfsck), lad->lad_name, *result);
+
        wake_up_all(&athread->t_ctl_waitq);
        l_wait_event(mthread->t_ctl_waitq,
                     (*result > 0 && list_empty(&lad->lad_req_list)) ||
@@ -2471,6 +2574,9 @@ void lfsck_post_generic(const struct lu_env *env,
 
        if (lad->lad_assistant_status < 0)
                *result = lad->lad_assistant_status;
+
+       CDEBUG(D_LFSCK, "%s: the assistant has done %s post, rc = %d\n",
+              lfsck_lfsck2name(com->lc_lfsck), lad->lad_name, *result);
 }
 
 int lfsck_double_scan_generic(const struct lu_env *env,
@@ -2486,12 +2592,20 @@ int lfsck_double_scan_generic(const struct lu_env *env,
        else
                lad->lad_to_double_scan = 1;
 
+       CDEBUG(D_LFSCK, "%s: waiting for assistant to do %s double_scan, "
+              "status %d\n",
+              lfsck_lfsck2name(com->lc_lfsck), lad->lad_name, status);
+
        wake_up_all(&athread->t_ctl_waitq);
        l_wait_event(mthread->t_ctl_waitq,
                     lad->lad_in_double_scan ||
                     thread_is_stopped(athread),
                     &lwi);
 
+       CDEBUG(D_LFSCK, "%s: the assistant has done %s double_scan, "
+              "status %d\n", lfsck_lfsck2name(com->lc_lfsck), lad->lad_name,
+              lad->lad_assistant_status);
+
        if (lad->lad_assistant_status < 0)
                return lad->lad_assistant_status;
 
@@ -2514,6 +2628,96 @@ void lfsck_quit_generic(const struct lu_env *env,
                     &lwi);
 }
 
+int lfsck_load_one_trace_file(const struct lu_env *env,
+                             struct lfsck_component *com,
+                             struct dt_object *parent,
+                             struct dt_object **child,
+                             const struct dt_index_features *ft,
+                             const char *name, bool reset)
+{
+       struct lfsck_instance *lfsck = com->lc_lfsck;
+       struct dt_object *obj;
+       int rc;
+       ENTRY;
+
+       if (*child != NULL) {
+               struct dt_it *it;
+               const struct dt_it_ops *iops;
+               struct lu_fid *fid = &lfsck_env_info(env)->lti_fid3;
+
+               if (!reset)
+                       RETURN(0);
+
+               obj = *child;
+               rc = obj->do_ops->do_index_try(env, obj, ft);
+               if (rc)
+                       /* unlink by force */
+                       goto unlink;
+
+               iops = &obj->do_index_ops->dio_it;
+               it = iops->init(env, obj, 0);
+               if (IS_ERR(it))
+                       /* unlink by force */
+                       goto unlink;
+
+               fid_zero(fid);
+               rc = iops->get(env, it, (const struct dt_key *)fid);
+               if (rc >= 0) {
+                       rc = iops->next(env, it);
+                       iops->put(env, it);
+               }
+               iops->fini(env, it);
+               if (rc > 0)
+                       /* "rc > 0" means the index file is empty. */
+                       RETURN(0);
+
+unlink:
+               /* The old index is not empty, remove it firstly. */
+               rc = local_object_unlink(env, lfsck->li_bottom, parent, name);
+
+               CDEBUG(D_LFSCK, "%s: unlink lfsck sub trace file %s: rc = %d\n",
+                      lfsck_lfsck2name(com->lc_lfsck), name, rc);
+
+               if (rc)
+                       RETURN(rc);
+
+               lfsck_object_put(env, *child);
+               *child = NULL;
+       }
+
+       obj = local_index_find_or_create(env, lfsck->li_los, parent, name,
+                                        S_IFREG | S_IRUGO | S_IWUSR, ft);
+       if (IS_ERR(obj))
+               RETURN(PTR_ERR(obj));
+
+       rc = obj->do_ops->do_index_try(env, obj, ft);
+       if (rc == 0)
+               *child = obj;
+
+       RETURN(rc);
+}
+
+int lfsck_load_sub_trace_files(const struct lu_env *env,
+                              struct lfsck_component *com,
+                              const struct dt_index_features *ft,
+                              const char *prefix, bool reset)
+{
+       char *name = lfsck_env_info(env)->lti_key;
+       struct lfsck_sub_trace_obj *lsto;
+       int rc;
+       int i;
+
+       for (i = 0, rc = 0, lsto = &com->lc_sub_trace_objs[0];
+            i < LFSCK_STF_COUNT && rc == 0; i++, lsto++) {
+               snprintf(name, NAME_MAX, "%s_%02d", prefix, i);
+               rc = lfsck_load_one_trace_file(env, com,
+                               com->lc_lfsck->li_lfsck_dir,
+                               &lsto->lsto_obj, ft, name, reset);
+       }
+
+       return rc;
+}
+
 /* external interfaces */
 
 int lfsck_get_speed(struct seq_file *m, struct dt_device *key)
@@ -2541,7 +2745,7 @@ int lfsck_get_speed(struct seq_file *m, struct dt_device *key)
 }
 EXPORT_SYMBOL(lfsck_get_speed);
 
-int lfsck_set_speed(struct dt_device *key, int val)
+int lfsck_set_speed(struct dt_device *key, __u32 val)
 {
        struct lu_env           env;
        struct lfsck_instance  *lfsck;
@@ -2646,7 +2850,7 @@ int lfsck_dump(struct seq_file *m, struct dt_device *key, enum lfsck_type type)
        if (likely(lfsck != NULL)) {
                com = lfsck_component_find(lfsck, type);
                if (likely(com != NULL)) {
-                       rc = com->lc_ops->lfsck_dump(&env, com, m);
+                       com->lc_ops->lfsck_dump(&env, com, m);
                        lfsck_component_put(&env, com);
                } else {
                        rc = -ENOTSUPP;
@@ -2693,10 +2897,9 @@ static int lfsck_stop_all(const struct lu_env *env,
        lr->lr_active = LFSCK_TYPES_ALL;
        lr->lr_param = stop->ls_flags;
 
-       laia->laia_com = NULL;
+       memset(laia, 0, sizeof(*laia));
        laia->laia_ltds = ltds;
        laia->laia_lr = lr;
-       laia->laia_result = 0;
        laia->laia_shared = 1;
 
        down_read(&ltds->ltd_rw_sem);
@@ -2767,10 +2970,9 @@ static int lfsck_start_all(const struct lu_env *env,
                       LSV_ASYNC_WINDOWS | LSV_CREATE_OSTOBJ |
                       LSV_CREATE_MDTOBJ;
 
-       laia->laia_com = NULL;
+       memset(laia, 0, sizeof(*laia));
        laia->laia_ltds = ltds;
        laia->laia_lr = lr;
-       laia->laia_result = 0;
        laia->laia_shared = 1;
 
        down_read(&ltds->ltd_rw_sem);
@@ -2835,6 +3037,9 @@ int lfsck_start(const struct lu_env *env, struct dt_device *key,
        struct l_wait_info               lwi    = { 0 };
        struct lfsck_thread_args        *lta;
        struct task_struct              *task;
+       struct lfsck_tgt_descs          *ltds;
+       struct lfsck_tgt_desc           *ltd;
+       __u32                            idx;
        int                              rc     = 0;
        __u16                            valid  = 0;
        __u16                            flags  = 0;
@@ -2981,6 +3186,38 @@ int lfsck_start(const struct lu_env *env, struct dt_device *key,
                }
        }
 
+       ltds = &lfsck->li_mdt_descs;
+       down_read(&ltds->ltd_rw_sem);
+       cfs_foreach_bit(ltds->ltd_tgts_bitmap, idx) {
+               ltd = lfsck_ltd2tgt(ltds, idx);
+               LASSERT(ltd != NULL);
+
+               ltd->ltd_layout_done = 0;
+               ltd->ltd_namespace_done = 0;
+               ltd->ltd_synced_failures = 0;
+               lfsck_reset_ltd_status(ltd, LFSCK_TYPE_NAMESPACE);
+               lfsck_reset_ltd_status(ltd, LFSCK_TYPE_LAYOUT);
+               list_del_init(&ltd->ltd_layout_phase_list);
+               list_del_init(&ltd->ltd_layout_list);
+               list_del_init(&ltd->ltd_namespace_phase_list);
+               list_del_init(&ltd->ltd_namespace_list);
+       }
+       up_read(&ltds->ltd_rw_sem);
+
+       ltds = &lfsck->li_ost_descs;
+       down_read(&ltds->ltd_rw_sem);
+       cfs_foreach_bit(ltds->ltd_tgts_bitmap, idx) {
+               ltd = lfsck_ltd2tgt(ltds, idx);
+               LASSERT(ltd != NULL);
+
+               ltd->ltd_layout_done = 0;
+               ltd->ltd_synced_failures = 0;
+               lfsck_reset_ltd_status(ltd, LFSCK_TYPE_LAYOUT);
+               list_del_init(&ltd->ltd_layout_phase_list);
+               list_del_init(&ltd->ltd_layout_list);
+       }
+       up_read(&ltds->ltd_rw_sem);
+
 trigger:
        lfsck->li_args_dir = LUDA_64BITHASH | LUDA_VERIFY | LUDA_TYPE;
        if (bk->lb_param & LPF_DRYRUN)
@@ -3109,6 +3346,28 @@ int lfsck_stop(const struct lu_env *env, struct dt_device *key,
        }
 
        thread_set_flags(thread, SVC_STOPPING);
+
+       if (lfsck->li_master) {
+               struct lfsck_component *com;
+               struct lfsck_assistant_data *lad;
+
+               list_for_each_entry(com, &lfsck->li_list_scan, lc_link) {
+                       lad = com->lc_data;
+                       spin_lock(&lad->lad_lock);
+                       if (lad->lad_task != NULL)
+                               force_sig(SIGINT, lad->lad_task);
+                       spin_unlock(&lad->lad_lock);
+               }
+
+               list_for_each_entry(com, &lfsck->li_list_double_scan, lc_link) {
+                       lad = com->lc_data;
+                       spin_lock(&lad->lad_lock);
+                       if (lad->lad_task != NULL)
+                               force_sig(SIGINT, lad->lad_task);
+                       spin_unlock(&lad->lad_lock);
+               }
+       }
+
        spin_unlock(&lfsck->li_lock);
 
        wake_up_all(&thread->t_ctl_waitq);
@@ -3165,8 +3424,6 @@ int lfsck_in_notify(const struct lu_env *env, struct dt_device *key,
        case LE_FID_ACCESSED:
        case LE_PEER_EXIT:
        case LE_CONDITIONAL_DESTROY:
-       case LE_SKIP_NLINK_DECLARE:
-       case LE_SKIP_NLINK:
        case LE_SET_LMV_MASTER:
        case LE_SET_LMV_SLAVE:
        case LE_PAIRS_VERIFY: {
@@ -3195,29 +3452,97 @@ int lfsck_in_notify(const struct lu_env *env, struct dt_device *key,
 EXPORT_SYMBOL(lfsck_in_notify);
 
 int lfsck_query(const struct lu_env *env, struct dt_device *key,
-               struct lfsck_request *lr)
+               struct lfsck_request *req, struct lfsck_reply *rep,
+               struct lfsck_query *que)
 {
        struct lfsck_instance  *lfsck;
        struct lfsck_component *com;
-       int                     rc;
+       int                     i;
+       int                     rc = 0;
+       __u16                   type;
        ENTRY;
 
        lfsck = lfsck_instance_find(key, true, false);
        if (unlikely(lfsck == NULL))
                RETURN(-ENXIO);
 
-       com = lfsck_component_find(lfsck, lr->lr_active);
-       if (likely(com != NULL)) {
-               rc = com->lc_ops->lfsck_query(env, com);
-               lfsck_component_put(env, com);
+       if (que != NULL) {
+               if (que->lu_types == LFSCK_TYPES_ALL)
+                       que->lu_types =
+                               LFSCK_TYPES_SUPPORTED & ~LFSCK_TYPE_SCRUB;
+
+               if (que->lu_types & ~LFSCK_TYPES_SUPPORTED) {
+                       que->lu_types &= ~LFSCK_TYPES_SUPPORTED;
+
+                       GOTO(out, rc = -ENOTSUPP);
+               }
+
+               for (i = 0, type = 1 << i; i < LFSCK_TYPE_BITS;
+                    i++, type = 1 << i) {
+                       if (!(que->lu_types & type))
+                               continue;
+
+again:
+                       com = lfsck_component_find(lfsck, type);
+                       if (unlikely(com == NULL))
+                               GOTO(out, rc = -ENOTSUPP);
+
+                       memset(que->lu_mdts_count[i], 0,
+                              sizeof(__u32) * (LS_MAX + 1));
+                       memset(que->lu_osts_count[i], 0,
+                              sizeof(__u32) * (LS_MAX + 1));
+                       que->lu_repaired[i] = 0;
+                       rc = com->lc_ops->lfsck_query(env, com, req, rep,
+                                                     que, i);
+                       lfsck_component_put(env, com);
+                       if  (rc < 0)
+                               GOTO(out, rc);
+               }
+
+               if (!(que->lu_flags & LPF_WAIT))
+                       GOTO(out, rc);
+
+               for (i = 0, type = 1 << i; i < LFSCK_TYPE_BITS;
+                    i++, type = 1 << i) {
+                       if (!(que->lu_types & type))
+                               continue;
+
+                       if (que->lu_mdts_count[i][LS_SCANNING_PHASE1] != 0 ||
+                           que->lu_mdts_count[i][LS_SCANNING_PHASE2] != 0 ||
+                           que->lu_osts_count[i][LS_SCANNING_PHASE1] != 0 ||
+                           que->lu_osts_count[i][LS_SCANNING_PHASE2] != 0) {
+                               struct l_wait_info lwi;
+
+                               /* If it is required to wait, then sleep
+                                * 3 seconds and try to query again. */
+                               lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(3),
+                                                      NULL,
+                                                      LWI_ON_SIGNAL_NOOP,
+                                                      NULL);
+                               rc = l_wait_event(lfsck->li_thread.t_ctl_waitq,
+                                                 0, &lwi);
+                               if (rc == -ETIMEDOUT)
+                                       goto again;
+                       }
+               }
        } else {
-               rc = -ENOTSUPP;
+               com = lfsck_component_find(lfsck, req->lr_active);
+               if (likely(com != NULL)) {
+                       rc = com->lc_ops->lfsck_query(env, com, req, rep,
+                                                     que, -1);
+                       lfsck_component_put(env, com);
+               } else {
+                       rc = -ENOTSUPP;
+               }
        }
 
-       lfsck_instance_put(env, lfsck);
+       GOTO(out, rc);
 
-       RETURN(rc);
+out:
+       lfsck_instance_put(env, lfsck);
+       return rc;
 }
+EXPORT_SYMBOL(lfsck_query);
 
 int lfsck_register_namespace(const struct lu_env *env, struct dt_device *key,
                             struct ldlm_namespace *ns)
@@ -3592,7 +3917,7 @@ static void __exit lfsck_exit(void)
 }
 
 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
-MODULE_DESCRIPTION("LFSCK");
+MODULE_DESCRIPTION("Lustre File System Checker");
 MODULE_VERSION(LUSTRE_VERSION_STRING);
 MODULE_LICENSE("GPL");