Whamcloud - gitweb
LU-1842 quota: don't setup qsd for standalone MGS
[fs/lustre-release.git] / lustre / quota / qsd_lib.c
index 0ba8622..277791b 100644 (file)
  * GPL HEADER END
  */
 /*
- * Copyright (c) 2012 Whamcloud, Inc.
+ * Copyright (c) 2012 Intel, Inc.
  * Use is subject to license terms.
  *
- * Author: Johann Lombardi <johann@whamcloud.com>
- * Author: Niu    Yawei    <niu@whamcloud.com>
+ * Author: Johann Lombardi <johann.lombardi@intel.com>
+ * Author: Niu    Yawei    <yawei.niu@intel.com>
  */
 
 /*
  * Quota Slave Driver (QSD) management.
+ *
+ * The quota slave feature is implemented under the form of a library called
+ * QSD. Each OSD device should create a QSD instance via qsd_init() which will
+ * be used to manage quota enforcement for this device. This implies:
+ * - completing the reintegration procedure with the quota master (aka QMT, see
+ *   qmt_dev.c) to retrieve the latest quota settings and space distribution.
+ * - managing quota locks in order to be notified of configuration changes.
+ * - acquiring space from the QMT when quota space for a given user/group is
+ *   close to exhaustion.
+ * - allocating quota space to service threads for local request processing.
+ *
+ * Once the QSD instance created, the OSD device should invoke qsd_start()
+ * when recovery is completed. This notifies the QSD that we are about to
+ * process new requests on which quota should be strictly enforced.
+ * Then, qsd_op_begin/end can be used to reserve/release/pre-acquire quota space
+ * for/after each operation until shutdown where the QSD instance should be
+ * freed via qsd_fini().
  */
 
 #ifndef EXPORT_SYMTAB
 
 #define DEBUG_SUBSYSTEM S_LQUOTA
 
+#include <obd_class.h>
 #include "qsd_internal.h"
 
+cfs_mem_cache_t *upd_kmem;
+
+struct lu_kmem_descr qsd_caches[] = {
+       {
+               .ckd_cache = &upd_kmem,
+               .ckd_name  = "upd_kmem",
+               .ckd_size  = sizeof(struct qsd_upd_rec)
+       },
+       {
+               .ckd_cache = NULL
+       }
+};
+
+/* define qsd thread key */
+LU_KEY_INIT_FINI(qsd, struct qsd_thread_info);
+LU_CONTEXT_KEY_DEFINE(qsd, LCT_MD_THREAD | LCT_DT_THREAD | LCT_LOCAL);
+LU_KEY_INIT_GENERIC(qsd);
+
 /* some procfs helpers */
 static int lprocfs_qsd_rd_state(char *page, char **start, off_t off,
                                int count, int *eof, void *data)
 {
        struct qsd_instance     *qsd = (struct qsd_instance *)data;
+       char                     enabled[5];
+       int                      rc;
 
        LASSERT(qsd != NULL);
 
-       return snprintf(page, count,
-                       "target name:    %s\n"
-                       "quota enabled:  none\n",
-                       qsd->qsd_svname);
+       memset(enabled, 0, sizeof(enabled));
+       if (qsd_type_enabled(qsd, USRQUOTA))
+               strcat(enabled, "u");
+       if (qsd_type_enabled(qsd, GRPQUOTA))
+               strcat(enabled, "g");
+       if (strlen(enabled) == 0)
+               strcat(enabled, "none");
+
+       rc = snprintf(page, count,
+                     "target name:    %s\n"
+                     "pool ID:        %d\n"
+                     "type:           %s\n"
+                     "quota enabled:  %s\n"
+                     "conn to master: %s\n",
+                     qsd->qsd_svname, qsd->qsd_pool_id,
+                     qsd->qsd_is_md ? "md" : "dt", enabled,
+                     qsd->qsd_exp_valid ? "setup" : "not setup yet");
+
+       if (qsd->qsd_prepared)
+               rc +=  snprintf(page + rc, count - rc,
+                               "user uptodate:  glb[%d],slv[%d],reint[%d]\n"
+                               "group uptodate: glb[%d],slv[%d],reint[%d]\n",
+                               qsd->qsd_type_array[USRQUOTA]->qqi_glb_uptodate,
+                               qsd->qsd_type_array[USRQUOTA]->qqi_slv_uptodate,
+                               qsd->qsd_type_array[USRQUOTA]->qqi_reint,
+                               qsd->qsd_type_array[GRPQUOTA]->qqi_glb_uptodate,
+                               qsd->qsd_type_array[GRPQUOTA]->qqi_slv_uptodate,
+                               qsd->qsd_type_array[GRPQUOTA]->qqi_reint);
+       return rc;
+}
+
+static int lprocfs_qsd_rd_enabled(char *page, char **start, off_t off,
+                                 int count, int *eof, void *data)
+{
+       struct qsd_instance     *qsd = (struct qsd_instance *)data;
+       char                     enabled[5];
+       LASSERT(qsd != NULL);
+
+       memset(enabled, 0, sizeof(enabled));
+       if (qsd_type_enabled(qsd, USRQUOTA))
+               strcat(enabled, "u");
+       if (qsd_type_enabled(qsd, GRPQUOTA))
+               strcat(enabled, "g");
+       if (strlen(enabled) == 0)
+               strcat(enabled, "none");
+
+       return snprintf(page, count, "%s\n", enabled);
 }
 
 static struct lprocfs_vars lprocfs_quota_qsd_vars[] = {
        { "info", lprocfs_qsd_rd_state, 0, 0},
+       { "enabled", lprocfs_qsd_rd_enabled, 0, 0},
        { NULL }
 };
 
 /*
+ * Callback function invoked by the OSP layer when the connection to the master
+ * has been set up.
+ *
+ * \param data - is a pointer to the qsd_instance
+ *
+ * \retval - 0 on success, appropriate error on failure
+ */
+static int qsd_conn_callback(void *data)
+{
+       struct qsd_instance *qsd = (struct qsd_instance *)data;
+       int                  type;
+       ENTRY;
+
+       /* qsd_exp should now be valid */
+       LASSERT(qsd->qsd_exp);
+
+       /* grab reference on namespace */
+       ldlm_namespace_get(class_exp2obd(qsd->qsd_exp)->obd_namespace);
+       qsd->qsd_ns = class_exp2obd(qsd->qsd_exp)->obd_namespace;
+
+       cfs_write_lock(&qsd->qsd_lock);
+       /* notify that qsd_exp is now valid */
+       qsd->qsd_exp_valid = true;
+       cfs_write_unlock(&qsd->qsd_lock);
+
+       /* Now that the connection to master is setup, we can initiate the
+        * reintegration procedure for quota types which are enabled.
+        * It is worth noting that, if the qsd_instance hasn't been started
+        * already, then we can only complete the first two steps of the
+        * reintegration procedure (i.e. global lock enqueue and slave
+        * index transfer) since the space usage reconciliation (i.e.
+        * step 3) will have to wait for qsd_start() to be called */
+       for (type = USRQUOTA; type < MAXQUOTAS; type++) {
+               struct qsd_qtype_info *qqi = qsd->qsd_type_array[type];
+               cfs_waitq_signal(&qqi->qqi_reint_thread.t_ctl_waitq);
+       }
+
+       RETURN(0);
+}
+
+/*
  * Release qsd_qtype_info structure which contains data associated with a
  * given quota type. This releases the accounting objects.
  * It's called on OSD cleanup when the qsd instance is released.
@@ -80,12 +203,40 @@ static void qsd_qtype_fini(const struct lu_env *env, struct qsd_instance *qsd,
        qqi = qsd->qsd_type_array[qtype];
        qsd->qsd_type_array[qtype] = NULL;
 
+       /* all deferred work lists should be empty */
+       LASSERT(cfs_list_empty(&qqi->qqi_deferred_glb));
+       LASSERT(cfs_list_empty(&qqi->qqi_deferred_slv));
+
+       /* shutdown lquota site */
+       if (qqi->qqi_site != NULL && !IS_ERR(qqi->qqi_site)) {
+               lquota_site_free(env, qqi->qqi_site);
+               qqi->qqi_site = NULL;
+       }
+
+       /* by now, all qqi users should have gone away */
+       LASSERT(cfs_atomic_read(&qqi->qqi_ref) == 1);
+       lu_ref_fini(&qqi->qqi_reference);
+
        /* release accounting object */
        if (qqi->qqi_acct_obj != NULL && !IS_ERR(qqi->qqi_acct_obj)) {
                lu_object_put(env, &qqi->qqi_acct_obj->do_lu);
                qqi->qqi_acct_obj = NULL;
        }
 
+       /* release slv index */
+       if (qqi->qqi_slv_obj != NULL && !IS_ERR(qqi->qqi_slv_obj)) {
+               lu_object_put(env, &qqi->qqi_slv_obj->do_lu);
+               qqi->qqi_slv_obj = NULL;
+               qqi->qqi_slv_ver = 0;
+       }
+
+       /* release global index */
+       if (qqi->qqi_glb_obj != NULL && !IS_ERR(qqi->qqi_glb_obj)) {
+               lu_object_put(env, &qqi->qqi_glb_obj->do_lu);
+               qqi->qqi_glb_obj = NULL;
+               qqi->qqi_glb_ver = 0;
+       }
+
        OBD_FREE_PTR(qqi);
        EXIT;
 }
@@ -93,7 +244,8 @@ static void qsd_qtype_fini(const struct lu_env *env, struct qsd_instance *qsd,
 /*
  * Allocate and initialize a qsd_qtype_info structure for quota type \qtype.
  * This opens the accounting object and initializes the proc file.
- * It's called on OSD start when the qsd instance is created.
+ * It's called on OSD start when the qsd_prepare() is invoked on the qsd
+ * instance.
  *
  * \param env  - the environment passed by the caller
  * \param qsd  - is the qsd instance which will be in charge of the new
@@ -108,6 +260,7 @@ static int qsd_qtype_init(const struct lu_env *env, struct qsd_instance *qsd,
 {
        struct qsd_qtype_info   *qqi;
        int                      rc;
+       struct obd_uuid          uuid;
        ENTRY;
 
        LASSERT(qsd->qsd_type_array[qtype] == NULL);
@@ -117,20 +270,69 @@ static int qsd_qtype_init(const struct lu_env *env, struct qsd_instance *qsd,
        if (qqi == NULL)
                RETURN(-ENOMEM);
        qsd->qsd_type_array[qtype] = qqi;
+       cfs_atomic_set(&qqi->qqi_ref, 1); /* referenced from qsd */
 
        /* set backpointer and other parameters */
        qqi->qqi_qsd   = qsd;
        qqi->qqi_qtype = qtype;
+       lu_ref_init(&qqi->qqi_reference);
+       lquota_generate_fid(&qqi->qqi_fid, qsd->qsd_pool_id, QSD_RES_TYPE(qsd),
+                           qtype);
+       qqi->qqi_glb_uptodate = false;
+       qqi->qqi_slv_uptodate = false;
+       qqi->qqi_reint        = false;
+       cfs_waitq_init(&qqi->qqi_reint_thread.t_ctl_waitq);
+       thread_set_flags(&qqi->qqi_reint_thread, SVC_STOPPED);
+       CFS_INIT_LIST_HEAD(&qqi->qqi_deferred_glb);
+       CFS_INIT_LIST_HEAD(&qqi->qqi_deferred_slv);
+       memset(&qqi->qqi_lockh, 0, sizeof(qqi->qqi_lockh));
 
-        /* open accounting object */
-        LASSERT(qqi->qqi_acct_obj == NULL);
-       qqi->qqi_acct_obj = acct_obj_lookup(env, qsd->qsd_dev,
-                                           qtype == USRQUOTA ? ACCT_USER_OID
-                                                             : ACCT_GROUP_OID);
-       /* don't print any error message on failure in order not to confuse
-        * non-OFD user (e.g. 2.3 MDT stack) */
-       if (IS_ERR(qqi->qqi_acct_obj))
-               qqi->qqi_acct_obj = NULL;
+       /* open accounting object */
+       LASSERT(qqi->qqi_acct_obj == NULL);
+       qqi->qqi_acct_obj = acct_obj_lookup(env, qsd->qsd_dev, qtype);
+       if (qqi->qqi_acct_obj == NULL) {
+               LCONSOLE_ERROR("%s: No %s space accounting support. Please use "
+                              "tunefs.lustre --quota option to enable quota "
+                              "accounting.\n",
+                              qsd->qsd_svname, QTYPE_NAME(qtype));
+               GOTO(out, rc = -ENOENT);
+       }
+
+       /* open global index copy */
+       LASSERT(qqi->qqi_glb_obj == NULL);
+       qqi->qqi_glb_obj = lquota_disk_glb_find_create(env, qsd->qsd_dev,
+                                                      qsd->qsd_root,
+                                                      &qqi->qqi_fid, true);
+       if (IS_ERR(qqi->qqi_glb_obj)) {
+               CERROR("%s: can't open global index copy "DFID" %ld\n",
+                      qsd->qsd_svname, PFID(&qqi->qqi_fid),
+                      PTR_ERR(qqi->qqi_glb_obj));
+               GOTO(out, rc = PTR_ERR(qqi->qqi_glb_obj));
+       }
+       qqi->qqi_glb_ver = dt_version_get(env, qqi->qqi_glb_obj);
+
+       /* open slave index copy */
+       LASSERT(qqi->qqi_slv_obj == NULL);
+       obd_str2uuid(&uuid, qsd->qsd_svname);
+       qqi->qqi_slv_obj = lquota_disk_slv_find_create(env, qsd->qsd_dev,
+                                                      qsd->qsd_root,
+                                                      &qqi->qqi_fid, &uuid,
+                                                      true);
+       if (IS_ERR(qqi->qqi_slv_obj)) {
+               CERROR("%s: can't open slave index copy "DFID" %ld\n",
+                      qsd->qsd_svname, PFID(&qqi->qqi_fid),
+                      PTR_ERR(qqi->qqi_slv_obj));
+               GOTO(out, rc = PTR_ERR(qqi->qqi_slv_obj));
+       }
+       qqi->qqi_slv_ver = dt_version_get(env, qqi->qqi_slv_obj);
+
+       /* allocate site */
+       qqi->qqi_site = lquota_site_alloc(env, qqi, false, qtype, &qsd_lqe_ops);
+       if (IS_ERR(qqi->qqi_site)) {
+               CERROR("%s: can't allocate site "DFID" %ld\n", qsd->qsd_svname,
+                      PFID(&qqi->qqi_fid), PTR_ERR(qqi->qqi_site));
+               GOTO(out, rc = PTR_ERR(qqi->qqi_site));
+       }
 
        /* register proc entry for accounting object */
        rc = lprocfs_seq_create(qsd->qsd_proc,
@@ -138,11 +340,10 @@ static int qsd_qtype_init(const struct lu_env *env, struct qsd_instance *qsd,
                                0444, &lprocfs_quota_seq_fops,
                                qqi->qqi_acct_obj);
        if (rc) {
-               CWARN("%s: can't add procfs entry for accounting file %d\n",
-                     qsd->qsd_svname, rc);
+               CERROR("%s: can't add procfs entry for accounting file %d\n",
+                      qsd->qsd_svname, rc);
                GOTO(out, rc);
        }
-
        EXIT;
 out:
        if (rc)
@@ -152,7 +353,8 @@ out:
 
 /*
  * Release a qsd_instance. Companion of qsd_init(). This releases all data
- * structures associated with the quota slave.
+ * structures associated with the quota slave (on-disk objects, lquota entry
+ * tables, ...).
  * This function should be called when the OSD is shutting down.
  *
  * \param env - is the environment passed by the caller
@@ -163,16 +365,59 @@ void qsd_fini(const struct lu_env *env, struct qsd_instance *qsd)
        int     qtype;
        ENTRY;
 
+       CDEBUG(D_QUOTA, "%s: initiating QSD shutdown\n", qsd->qsd_svname);
+       cfs_write_lock(&qsd->qsd_lock);
+       qsd->qsd_stopping = true;
+       cfs_write_unlock(&qsd->qsd_lock);
+
+       /* remove from the list of fsinfo */
+       if (!cfs_list_empty(&qsd->qsd_link)) {
+               LASSERT(qsd->qsd_fsinfo != NULL);
+               cfs_down(&qsd->qsd_fsinfo->qfs_sem);
+               cfs_list_del_init(&qsd->qsd_link);
+               cfs_up(&qsd->qsd_fsinfo->qfs_sem);
+       }
+
        /* remove qsd proc entry */
-       if (qsd->qsd_proc != NULL && !IS_ERR(qsd->qsd_proc)) {
+       if (qsd->qsd_proc != NULL) {
                lprocfs_remove(&qsd->qsd_proc);
                qsd->qsd_proc = NULL;
        }
 
+       /* stop the writeback thread */
+       qsd_stop_upd_thread(qsd);
+
+       /* shutdown the reintegration threads */
+       for (qtype = USRQUOTA; qtype < MAXQUOTAS; qtype++) {
+               if (qsd->qsd_type_array[qtype] == NULL)
+                       continue;
+               qsd_stop_reint_thread(qsd->qsd_type_array[qtype]);
+       }
+
+       /* release reference on namespace */
+       if (qsd->qsd_ns != NULL) {
+               ldlm_namespace_put(qsd->qsd_ns);
+               qsd->qsd_ns = NULL;
+       }
+
        /* free per-quota type data */
        for (qtype = USRQUOTA; qtype < MAXQUOTAS; qtype++)
                qsd_qtype_fini(env, qsd, qtype);
 
+       /* deregister connection to the quota master */
+       qsd->qsd_exp_valid = false;
+       lustre_deregister_osp_item(&qsd->qsd_exp);
+
+       /* release per-filesystem information */
+       if (qsd->qsd_fsinfo != NULL)
+               qsd_put_fsinfo(qsd->qsd_fsinfo);
+
+       /* release quota root directory */
+       if (qsd->qsd_root != NULL) {
+               lu_object_put(env, &qsd->qsd_root->do_lu);
+               qsd->qsd_root = NULL;
+       }
+
        /* release reference on dt_device */
        if (qsd->qsd_dev != NULL) {
                lu_ref_del(&qsd->qsd_dev->dd_lu_dev.ld_reference, "qsd", qsd);
@@ -180,6 +425,7 @@ void qsd_fini(const struct lu_env *env, struct qsd_instance *qsd)
                qsd->qsd_dev = NULL;
        }
 
+       CDEBUG(D_QUOTA, "%s: QSD shutdown completed\n", qsd->qsd_svname);
        OBD_FREE_PTR(qsd);
        EXIT;
 }
@@ -187,8 +433,7 @@ EXPORT_SYMBOL(qsd_fini);
 
 /*
  * Create a new qsd_instance to be associated with backend osd device
- * identified by \dev. For now, this function just create procfs files which
- * dumps the accounting information
+ * identified by \dev.
  *
  * \param env    - the environment passed by the caller
  * \param svname - is the service name of the OSD device creating this instance
@@ -203,15 +448,32 @@ struct qsd_instance *qsd_init(const struct lu_env *env, char *svname,
                              struct dt_device *dev,
                              cfs_proc_dir_entry_t *osd_proc)
 {
+       struct qsd_thread_info  *qti = qsd_info(env);
        struct qsd_instance     *qsd;
-       int                      rc, qtype;
+       int                      rc, type, idx;
        ENTRY;
 
+       /* only configure qsd for MDT & OST */
+       type = server_name2index(svname, &idx, NULL);
+       if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST)
+               RETURN(NULL);
+
        /* allocate qsd instance */
        OBD_ALLOC_PTR(qsd);
        if (qsd == NULL)
                RETURN(ERR_PTR(-ENOMEM));
 
+       /* generic initializations */
+       cfs_rwlock_init(&qsd->qsd_lock);
+       CFS_INIT_LIST_HEAD(&qsd->qsd_link);
+       thread_set_flags(&qsd->qsd_upd_thread, SVC_STOPPED);
+       cfs_waitq_init(&qsd->qsd_upd_thread.t_ctl_waitq);
+       CFS_INIT_LIST_HEAD(&qsd->qsd_upd_list);
+       cfs_spin_lock_init(&qsd->qsd_adjust_lock);
+       CFS_INIT_LIST_HEAD(&qsd->qsd_adjust_list);
+       qsd->qsd_prepared = false;
+       qsd->qsd_started = false;
+
        /* copy service name */
        strncpy(qsd->qsd_svname, svname, MAX_OBD_NAME);
 
@@ -220,27 +482,225 @@ struct qsd_instance *qsd_init(const struct lu_env *env, char *svname,
        lu_ref_add(&dev->dd_lu_dev.ld_reference, "qsd", qsd);
        qsd->qsd_dev = dev;
 
+       /* we only support pool ID 0 (default data or metadata pool) for the
+        * time being. A different pool ID could be assigned to this target via
+        * the configuration log in the future */
+       qsd->qsd_pool_id  = 0;
+
+       /* get fsname from svname */
+       rc = server_name2fsname(svname, qti->qti_buf, NULL);
+       if (rc) {
+               CERROR("%s: fail to extract filesystem name\n", svname);
+               GOTO(out, rc);
+       }
+
+       /* look up quota setting for the filesystem the target belongs to */
+       qsd->qsd_fsinfo = qsd_get_fsinfo(qti->qti_buf, 1);
+       if (qsd->qsd_fsinfo == NULL) {
+               CERROR("%s: failed to locate filesystem information\n", svname);
+               GOTO(out, rc = -EINVAL);
+       }
+
+       /* add in the list of lquota_fsinfo */
+       cfs_down(&qsd->qsd_fsinfo->qfs_sem);
+       list_add_tail(&qsd->qsd_link, &qsd->qsd_fsinfo->qfs_qsd_list);
+       cfs_up(&qsd->qsd_fsinfo->qfs_sem);
+
        /* register procfs directory */
        qsd->qsd_proc = lprocfs_register(QSD_DIR, osd_proc,
                                         lprocfs_quota_qsd_vars, qsd);
        if (IS_ERR(qsd->qsd_proc)) {
                rc = PTR_ERR(qsd->qsd_proc);
+               qsd->qsd_proc = NULL;
                CERROR("%s: fail to create quota slave proc entry (%d)\n",
                       svname, rc);
                GOTO(out, rc);
         }
+       EXIT;
+out:
+       if (rc) {
+               qsd_fini(env, qsd);
+               return ERR_PTR(rc);
+       }
+       RETURN(qsd);
+}
+EXPORT_SYMBOL(qsd_init);
+
+/*
+ * Initialize on-disk structures in order to manage quota enforcement for
+ * the target associated with the qsd instance \qsd and starts the reintegration
+ * procedure for each quota type as soon as possible.
+ * The last step of the reintegration will be completed once qsd_start() is
+ * called, at which points the space reconciliation with the master will be
+ * executed.
+ * This function must be called when the server stack is fully configured,
+ * typically when ->ldo_prepare is called across the stack.
+ *
+ * \param env - the environment passed by the caller
+ * \param qsd - is qsd_instance to prepare
+ *
+ * \retval - 0 on success, appropriate error on failure
+ */
+int qsd_prepare(const struct lu_env *env, struct qsd_instance *qsd)
+{
+       struct qsd_thread_info  *qti = qsd_info(env);
+       int                      qtype, rc = 0;
+       ENTRY;
+
+       LASSERT(qsd != NULL);
+
+       cfs_read_lock(&qsd->qsd_lock);
+       if (qsd->qsd_prepared) {
+               CERROR("%s: qsd instance already prepared\n", qsd->qsd_svname);
+               rc = -EALREADY;
+       }
+       cfs_read_unlock(&qsd->qsd_lock);
+       if (rc)
+               RETURN(rc);
+
+       /* Record whether this qsd instance is managing quota enforcement for a
+        * MDT (i.e. inode quota) or OST (block quota) */
+       if (lu_device_is_md(qsd->qsd_dev->dd_lu_dev.ld_site->ls_top_dev)) {
+               qsd->qsd_is_md = true;
+               qsd->qsd_sync_threshold = LQUOTA_LEAST_QUNIT(LQUOTA_RES_MD);
+       } else {
+               qsd->qsd_sync_threshold = LQUOTA_LEAST_QUNIT(LQUOTA_RES_DT);
+       }
+
+       /* look-up on-disk directory for the quota slave */
+       qsd->qsd_root = lquota_disk_dir_find_create(env, qsd->qsd_dev, NULL,
+                                                   QSD_DIR);
+       if (IS_ERR(qsd->qsd_root)) {
+               rc = PTR_ERR(qsd->qsd_root);
+               qsd->qsd_root = NULL;
+               CERROR("%s: failed to create quota slave root dir (%d)\n",
+                      qsd->qsd_svname, rc);
+               RETURN(rc);
+       }
 
        /* initialize per-quota type data */
        for (qtype = USRQUOTA; qtype < MAXQUOTAS; qtype++) {
                rc = qsd_qtype_init(env, qsd, qtype);
                if (rc)
-                       GOTO(out, rc);
+                       RETURN(rc);
        }
-out:
+
+       /* pools successfully setup, mark the qsd as prepared */
+       cfs_write_lock(&qsd->qsd_lock);
+       qsd->qsd_prepared = true;
+       cfs_write_unlock(&qsd->qsd_lock);
+
+       /* start reintegration thread for each type, if required */
+       for (qtype = USRQUOTA; qtype < MAXQUOTAS; qtype++) {
+               struct qsd_qtype_info   *qqi = qsd->qsd_type_array[qtype];
+
+               rc = qsd_start_reint_thread(qqi);
+               if (rc) {
+                       CERROR("%s: failed to start reint thread for type %s "
+                              "(%d)\n", qsd->qsd_svname, QTYPE_NAME(qtype),
+                              rc);
+                       RETURN(rc);
+               }
+       }
+
+       /* start writeback thread */
+       rc = qsd_start_upd_thread(qsd);
        if (rc) {
-               qsd_fini(env, qsd);
-               return ERR_PTR(rc);
+               CERROR("%s: failed to start writeback thread (%d)\n",
+                      qsd->qsd_svname, rc);
+               RETURN(rc);
        }
-       RETURN(qsd);
+
+       /* generate osp name */
+       rc = tgt_name2ospname((char *)qsd->qsd_svname, qti->qti_buf);
+       if (rc) {
+               CERROR("%s: failed to generate ospname (%d)\n",
+                      qsd->qsd_svname, rc);
+               RETURN(rc);
+       }
+
+       /* the connection callback will start the reintegration
+        * procedure if quota is enabled */
+       rc = lustre_register_osp_item(qti->qti_buf, &qsd->qsd_exp,
+                                     qsd_conn_callback, (void *)qsd);
+       if (rc) {
+               CERROR("%s: fail to get connection to master (%d)\n",
+                      qsd->qsd_svname, rc);
+               RETURN(rc);
+       }
+
+       RETURN(0);
+}
+EXPORT_SYMBOL(qsd_prepare);
+
+/*
+ * Start a qsd instance. This will complete the last step of the reintegration
+ * procedure as soon as possible (provided that the master is reachable).
+ * This should be called when recovery has been completed and quota should now
+ * be enforced on every operations.
+ *
+ * \param env - the environment passed by the caller
+ * \param qsd - is the qsd instance associated with the osd device to start
+ */
+int qsd_start(const struct lu_env *env, struct qsd_instance *qsd)
+{
+       int     type, rc = 0;
+       ENTRY;
+
+       cfs_write_lock(&qsd->qsd_lock);
+       if (!qsd->qsd_prepared) {
+               CERROR("%s: can't start qsd instance since it was properly "
+                      "initialized\n", qsd->qsd_svname);
+               rc = -EFAULT;
+       } else if (qsd->qsd_started) {
+               CERROR("%s: qsd instance already started\n", qsd->qsd_svname);
+               rc = -EALREADY;
+       } else {
+               /* notify that the qsd_instance is now started */
+               qsd->qsd_started = true;
+       }
+       cfs_write_unlock(&qsd->qsd_lock);
+
+       if (rc)
+               RETURN(rc);
+
+       /* Trigger the 3rd step of reintegration: If usage > granted, acquire
+        * up to usage; If usage < granted, release down to usage.  */
+       for (type = USRQUOTA; type < MAXQUOTAS; type++) {
+               struct qsd_qtype_info   *qqi = qsd->qsd_type_array[type];
+               cfs_waitq_signal(&qqi->qqi_reint_thread.t_ctl_waitq);
+       }
+
+       RETURN(rc);
+}
+EXPORT_SYMBOL(qsd_start);
+
+void lustre_register_quota_process_config(int (*qpc)(struct lustre_cfg *lcfg));
+
+/*
+ * Global initialization performed at module load time
+ */
+int qsd_glb_init(void)
+{
+       int     rc;
+
+       rc = lu_kmem_init(qsd_caches);
+       if (rc)
+               return rc;
+
+       qsd_key_init_generic(&qsd_thread_key, NULL);
+       lu_context_key_register(&qsd_thread_key);
+       lustre_register_quota_process_config(qsd_process_config);
+
+       return 0;
+}
+
+/*
+ * Companion of qsd_glb_init() called at module unload time
+ */
+void qsd_glb_fini(void)
+{
+       lustre_register_quota_process_config(NULL);
+       lu_kmem_fini(qsd_caches);
+       lu_context_key_degister(&qsd_thread_key);
 }
-EXPORT_SYMBOL(qsd_init);