Whamcloud - gitweb
LU-6245 libcfs: replace IS_PO2 with is_power_of_2 in server code
[fs/lustre-release.git] / lustre / ofd / ofd_dev.c
index 0a2ec19..137a05d 100644 (file)
@@ -23,7 +23,7 @@
  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2012, 2015, Intel Corporation.
+ * Copyright (c) 2012, 2016, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -1324,7 +1324,6 @@ static int ofd_getattr_hdl(struct tgt_session_info *tsi)
 
                obdo_from_la(&repbody->oa, &fti->fti_attr,
                             OFD_VALID_FLAGS | LA_UID | LA_GID);
-               tgt_drop_id(tsi->tsi_exp, &repbody->oa);
 
                /* Store object version in reply */
                curr_version = dt_version_get(tsi->tsi_env,
@@ -1420,7 +1419,6 @@ static int ofd_setattr_hdl(struct tgt_session_info *tsi)
 
        obdo_from_la(&repbody->oa, &fti->fti_attr,
                     OFD_VALID_FLAGS | LA_UID | LA_GID);
-       tgt_drop_id(tsi->tsi_exp, &repbody->oa);
 
        ofd_counter_incr(tsi->tsi_exp, LPROC_OFD_STATS_SETATTR,
                         tsi->tsi_jobid, 1);
@@ -1591,7 +1589,7 @@ static int ofd_create_hdl(struct tgt_session_info *tsi)
 
        oseq = ofd_seq_load(tsi->tsi_env, ofd, seq);
        if (IS_ERR(oseq)) {
-               CERROR("%s: Can't find FID Sequence "LPX64": rc = %ld\n",
+               CERROR("%s: Can't find FID Sequence %#llx: rc = %ld\n",
                       ofd_name(ofd), seq, PTR_ERR(oseq));
                GOTO(out_sem, rc = -EINVAL);
        }
@@ -1600,7 +1598,7 @@ static int ofd_create_hdl(struct tgt_session_info *tsi)
            (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
                if (!ofd_obd(ofd)->obd_recovering ||
                    oid > ofd_seq_last_oid(oseq)) {
-                       CERROR("%s: recreate objid "DOSTID" > last id "LPU64
+                       CERROR("%s: recreate objid "DOSTID" > last id %llu"
                               "\n", ofd_name(ofd), POSTID(&oa->o_oi),
                               ofd_seq_last_oid(oseq));
                        GOTO(out_nolock, rc = -EINVAL);
@@ -1625,13 +1623,13 @@ static int ofd_create_hdl(struct tgt_session_info *tsi)
                oseq->os_destroys_in_progress = 1;
                mutex_lock(&oseq->os_create_lock);
                if (!oseq->os_destroys_in_progress) {
-                       CERROR("%s:["LPU64"] destroys_in_progress already"
+                       CERROR("%s:[%llu] destroys_in_progress already"
                               " cleared\n", ofd_name(ofd), seq);
                        ostid_set_id(&rep_oa->o_oi, ofd_seq_last_oid(oseq));
                        GOTO(out, rc = 0);
                }
                diff = oid - ofd_seq_last_oid(oseq);
-               CDEBUG(D_HA, "ofd_last_id() = "LPU64" -> diff = %d\n",
+               CDEBUG(D_HA, "ofd_last_id() = %llu -> diff = %d\n",
                        ofd_seq_last_oid(oseq), diff);
                if (-diff > OST_MAX_PRECREATE) {
                        /* Let MDS know that we are so far ahead. */
@@ -1684,7 +1682,7 @@ static int ofd_create_hdl(struct tgt_session_info *tsi)
                        if (diff < 0) {
                                /* LU-5648 */
                                CERROR("%s: invalid precreate request for "
-                                      DOSTID", last_id " LPU64 ". "
+                                      DOSTID", last_id %llu. "
                                       "Likely MDS last_id corruption\n",
                                       ofd_name(ofd), POSTID(&oa->o_oi),
                                       ofd_seq_last_oid(oseq));
@@ -1739,8 +1737,8 @@ static int ofd_create_hdl(struct tgt_session_info *tsi)
                        next_id = ofd_seq_last_oid(oseq) + 1;
                        count = ofd_precreate_batch(ofd, diff);
 
-                       CDEBUG(D_HA, "%s: reserve %d objects in group "LPX64
-                              " at "LPU64"\n", ofd_name(ofd),
+                       CDEBUG(D_HA, "%s: reserve %d objects in group %#llx"
+                              " at %llu\n", ofd_name(ofd),
                               count, seq, next_id);
 
                        if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
@@ -2059,8 +2057,8 @@ static int ofd_punch_hdl(struct tgt_session_info *tsi)
                        RETURN(rc);
        }
 
-       CDEBUG(D_INODE, "calling punch for object "DFID", valid = "LPX64
-              ", start = "LPD64", end = "LPD64"\n", PFID(&tsi->tsi_fid),
+       CDEBUG(D_INODE, "calling punch for object "DFID", valid = %#llx"
+              ", start = %lld, end = %lld\n", PFID(&tsi->tsi_fid),
               oa->o_valid, start, end);
 
        fo = ofd_object_find_exists(tsi->tsi_env, ofd_exp(tsi->tsi_exp),
@@ -2107,6 +2105,60 @@ out:
        return rc;
 }
 
+static int ofd_ladvise_prefetch(const struct lu_env *env,
+                               struct ofd_object *fo,
+                               struct niobuf_local *lnb,
+                               __u64 start, __u64 end)
+{
+       struct ofd_thread_info  *info = ofd_info(env);
+       pgoff_t                  start_index, end_index, pages;
+       struct niobuf_remote     rnb;
+       unsigned long            nr_local;
+       int                      rc = 0;
+
+       if (end <= start)
+               RETURN(-EINVAL);
+
+       ofd_read_lock(env, fo);
+       if (!ofd_object_exists(fo))
+               GOTO(out_unlock, rc = -ENOENT);
+
+       rc = ofd_attr_get(env, fo, &info->fti_attr);
+       if (rc)
+               GOTO(out_unlock, rc);
+
+       if (end > info->fti_attr.la_size)
+               end = info->fti_attr.la_size;
+
+       if (end == 0)
+               GOTO(out_unlock, rc);
+
+       /* We need page aligned offset and length */
+       start_index = start >> PAGE_SHIFT;
+       end_index = (end - 1) >> PAGE_SHIFT;
+       pages = end_index - start_index + 1;
+       while (pages > 0) {
+               nr_local = pages <= PTLRPC_MAX_BRW_PAGES ? pages :
+                       PTLRPC_MAX_BRW_PAGES;
+               rnb.rnb_offset = start_index << PAGE_SHIFT;
+               rnb.rnb_len = nr_local << PAGE_SHIFT;
+               rc = dt_bufs_get(env, ofd_object_child(fo), &rnb, lnb, 0);
+               if (unlikely(rc < 0))
+                       break;
+               nr_local = rc;
+               rc = dt_read_prep(env, ofd_object_child(fo), lnb, nr_local);
+               dt_bufs_put(env, ofd_object_child(fo), lnb, nr_local);
+               if (unlikely(rc))
+                       break;
+               start_index += nr_local;
+               pages -= nr_local;
+       }
+
+out_unlock:
+       ofd_read_unlock(env, fo);
+       RETURN(rc);
+}
+
 /**
  * OFD request handler for OST_LADVISE RPC.
  *
@@ -2119,20 +2171,29 @@ out:
  */
 static int ofd_ladvise_hdl(struct tgt_session_info *tsi)
 {
-       struct ptlrpc_request   *req = tgt_ses_req(tsi);
-       struct obd_export       *exp = tsi->tsi_exp;
-       struct ofd_device       *ofd = ofd_exp(exp);
-       struct ost_body         *body, *repbody;
-       struct ofd_thread_info  *info;
-       struct ofd_object       *fo;
-       const struct lu_env     *env = req->rq_svc_thread->t_env;
-       int                      rc = 0;
-       struct lu_ladvise       *ladvise;
-       int                      num_advise;
-       struct ladvise_hdr      *ladvise_hdr;
-       int                      i;
+       struct ptlrpc_request *req = tgt_ses_req(tsi);
+       struct obd_export *exp = tsi->tsi_exp;
+       struct ofd_device *ofd = ofd_exp(exp);
+       struct ost_body *body, *repbody;
+       struct ofd_thread_info *info;
+       struct ofd_object *fo;
+       struct ptlrpc_thread *svc_thread = req->rq_svc_thread;
+       const struct lu_env *env = svc_thread->t_env;
+       struct tgt_thread_big_cache *tbc = svc_thread->t_data;
+       int rc = 0;
+       struct lu_ladvise *ladvise;
+       int num_advise;
+       struct ladvise_hdr *ladvise_hdr;
+       struct obd_ioobj ioo;
+       struct lustre_handle lockh = { 0 };
+       __u64 flags = 0;
+       int i;
+       struct dt_object *dob;
+       __u64 start;
+       __u64 end;
        ENTRY;
 
+       CFS_FAIL_TIMEOUT(OBD_FAIL_OST_LADVISE_PAUSE, cfs_fail_val);
        body = tsi->tsi_ost_body;
 
        if ((body->oa.o_valid & OBD_MD_FLID) != OBD_MD_FLID)
@@ -2156,7 +2217,7 @@ static int ofd_ladvise_hdl(struct tgt_session_info *tsi)
 
        num_advise = req_capsule_get_size(&req->rq_pill,
                                          &RMF_OST_LADVISE, RCL_CLIENT) /
-                    sizeof(*ladvise);
+                                         sizeof(*ladvise);
        if (num_advise < ladvise_hdr->lah_count)
                RETURN(err_serious(-EPROTO));
 
@@ -2176,9 +2237,12 @@ static int ofd_ladvise_hdl(struct tgt_session_info *tsi)
                RETURN(rc);
        }
        LASSERT(fo != NULL);
+       dob = ofd_object_child(fo);
 
        for (i = 0; i < num_advise; i++, ladvise++) {
-               if (ladvise->lla_end <= ladvise->lla_start) {
+               start = ladvise->lla_start;
+               end = ladvise->lla_end;
+               if (end <= start) {
                        rc = err_serious(-EPROTO);
                        break;
                }
@@ -2188,6 +2252,27 @@ static int ofd_ladvise_hdl(struct tgt_session_info *tsi)
                default:
                        rc = -ENOTSUPP;
                        break;
+               case LU_LADVISE_WILLREAD:
+                       if (tbc == NULL)
+                               RETURN(-ENOMEM);
+
+                       ioo.ioo_oid = body->oa.o_oi;
+                       ioo.ioo_bufcnt = 1;
+                       rc = tgt_extent_lock(exp->exp_obd->obd_namespace,
+                                            &tsi->tsi_resid, start, end - 1,
+                                            &lockh, LCK_PR, &flags);
+                       if (rc != 0)
+                               break;
+
+                       req->rq_status = ofd_ladvise_prefetch(env, fo,
+                                                             tbc->local,
+                                                             start, end);
+                       tgt_extent_unlock(&lockh, LCK_PR);
+                       break;
+               case LU_LADVISE_DONTNEED:
+                       rc = dt_ladvise(env, dob, ladvise->lla_start,
+                                       ladvise->lla_end, LU_LADVISE_DONTNEED);
+                       break;
                }
                if (rc != 0)
                        break;
@@ -2268,59 +2353,17 @@ static int ofd_quotactl(struct tgt_session_info *tsi)
  *
  * \retval             amount of time to extend the timeout with
  */
-static inline int prolong_timeout(struct ptlrpc_request *req,
-                                 struct ldlm_lock *lock)
+static inline int prolong_timeout(struct ptlrpc_request *req)
 {
        struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
+       time_t req_timeout;
 
        if (AT_OFF)
                return obd_timeout / 2;
 
-       /* We are in the middle of the process - BL AST is sent, CANCEL
-         is ahead. Take half of AT + IO process time. */
-       return at_est2timeout(at_get(&svcpt->scp_at_estimate)) +
-               (ldlm_bl_timeout(lock) >> 1);
-}
-
-/**
- * Prolong single lock timeout.
- *
- * This is supplemental function to the ofd_prolong_locks(). It prolongs
- * a single lock.
- *
- * \param[in] tsi      target session environment for this request
- * \param[in] lock     LDLM lock to prolong
- * \param[in] extent   related extent
- * \param[in] timeout  timeout value to add
- *
- * \retval             0 if lock is not suitable for prolongation
- * \retval             1 if lock was prolonged successfully
- */
-static int ofd_prolong_one_lock(struct tgt_session_info *tsi,
-                               struct ldlm_lock *lock,
-                               struct ldlm_extent *extent)
-{
-       int timeout = prolong_timeout(tgt_ses_req(tsi), lock);
-
-       if (lock->l_flags & LDLM_FL_DESTROYED) /* lock already cancelled */
-               return 0;
-
-       /* XXX: never try to grab resource lock here because we're inside
-        * exp_bl_list_lock; in ldlm_lockd.c to handle waiting list we take
-        * res lock and then exp_bl_list_lock. */
-
-       if (!(lock->l_flags & LDLM_FL_AST_SENT))
-               /* ignore locks not being cancelled */
-               return 0;
-
-       LDLM_DEBUG(lock, "refreshed for req x"LPU64" ext("LPU64"->"LPU64") "
-                        "to %ds.\n", tgt_ses_req(tsi)->rq_xid, extent->start,
-                        extent->end, timeout);
-
-       /* OK. this is a possible lock the user holds doing I/O
-        * let's refresh eviction timer for it */
-       ldlm_refresh_waiting_lock(lock, timeout);
-       return 1;
+       req_timeout = req->rq_deadline - req->rq_arrival_time.tv_sec;
+       return max_t(time_t, at_est2timeout(at_get(&svcpt->scp_at_estimate)),
+                    req_timeout);
 }
 
 /**
@@ -2343,25 +2386,26 @@ static int ofd_prolong_one_lock(struct tgt_session_info *tsi,
  * request may cover multiple locks.
  *
  * \param[in] tsi      target session environment for this request
- * \param[in] start    start of extent
- * \param[in] end      end of extent
+ * \param[in] data     struct of data to prolong locks
  *
- * \retval             number of prolonged locks
  */
-static int ofd_prolong_extent_locks(struct tgt_session_info *tsi,
-                                   __u64 start, __u64 end)
+static void ofd_prolong_extent_locks(struct tgt_session_info *tsi,
+                                   struct ldlm_prolong_args *data)
 {
-       struct obd_export       *exp = tsi->tsi_exp;
        struct obdo             *oa  = &tsi->tsi_ost_body->oa;
-       struct ldlm_extent       extent = {
-               .start = start,
-               .end = end
-       };
        struct ldlm_lock        *lock;
-       int                      lock_count = 0;
 
        ENTRY;
 
+       data->lpa_timeout = prolong_timeout(tgt_ses_req(tsi));
+       data->lpa_export = tsi->tsi_exp;
+       data->lpa_resid = tsi->tsi_resid;
+
+       CDEBUG(D_RPCTRACE, "Prolong locks for req %p with x%llu"
+              " ext(%llu->%llu)\n", tgt_ses_req(tsi),
+              tgt_ses_req(tsi)->rq_xid, data->lpa_extent.start,
+              data->lpa_extent.end);
+
        if (oa->o_valid & OBD_MD_FLHANDLE) {
                /* mostly a request should be covered by only one lock, try
                 * fast path. */
@@ -2369,42 +2413,21 @@ static int ofd_prolong_extent_locks(struct tgt_session_info *tsi,
                if (lock != NULL) {
                        /* Fast path to check if the lock covers the whole IO
                         * region exclusively. */
-                       if (lock->l_granted_mode == LCK_PW &&
-                           ldlm_extent_contain(&lock->l_policy_data.l_extent,
-                                               &extent)) {
+                       if (ldlm_extent_contain(&lock->l_policy_data.l_extent,
+                                               &data->lpa_extent)) {
                                /* bingo */
-                               LASSERT(lock->l_export == exp);
-                               lock_count = ofd_prolong_one_lock(tsi, lock,
-                                                                 &extent);
+                               LASSERT(lock->l_export == data->lpa_export);
+                               ldlm_lock_prolong_one(lock, data);
                                LDLM_LOCK_PUT(lock);
-                               RETURN(lock_count);
+                               RETURN_EXIT;
                        }
                        lock->l_last_used = cfs_time_current();
                        LDLM_LOCK_PUT(lock);
                }
        }
 
-       spin_lock_bh(&exp->exp_bl_list_lock);
-       list_for_each_entry(lock, &exp->exp_bl_list, l_exp_list) {
-               LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
-               LASSERT(lock->l_resource->lr_type == LDLM_EXTENT);
-
-               /* ignore waiting locks, no more granted locks in the list */
-               if (lock->l_granted_mode != lock->l_req_mode)
-                       break;
-
-               if (!ldlm_res_eq(&tsi->tsi_resid, &lock->l_resource->lr_name))
-                       continue;
-
-               if (!ldlm_extent_overlap(&lock->l_policy_data.l_extent,
-                                        &extent))
-                       continue;
-
-               lock_count += ofd_prolong_one_lock(tsi, lock, &extent);
-       }
-       spin_unlock_bh(&exp->exp_bl_list_lock);
-
-       RETURN(lock_count);
+       ldlm_resource_prolong(data);
+       EXIT;
 }
 
 /**
@@ -2451,8 +2474,10 @@ static int ofd_rw_hpreq_lock_match(struct ptlrpc_request *req,
        if (!ostid_res_name_eq(&ioo->ioo_oid, &lock->l_resource->lr_name))
                RETURN(0);
 
-       /* a bulk write can only hold a reference on a PW extent lock */
-       mode = LCK_PW;
+       /* a bulk write can only hold a reference on a PW extent lock
+        * or GROUP lock.
+        */
+       mode = LCK_PW | LCK_GROUP;
        if (opc == OST_READ)
                /* whereas a bulk read can be protected by either a PR or PW
                 * extent lock */
@@ -2468,20 +2493,22 @@ static int ofd_rw_hpreq_lock_match(struct ptlrpc_request *req,
  * Implementation of ptlrpc_hpreq_ops::hpreq_lock_check for OFD RW requests.
  *
  * Check for whether the given PTLRPC request (\a req) is blocking
- * an LDLM lock cancel.
+ * an LDLM lock cancel. Also checks whether the request is covered by an LDLM
+ * lock.
  *
  * \param[in] req      the incoming request
  *
  * \retval             1 if \a req is blocking an LDLM lock cancel
  * \retval             0 if it is not
+ * \retval             -ESTALE if lock is not found
  */
 static int ofd_rw_hpreq_check(struct ptlrpc_request *req)
 {
        struct tgt_session_info *tsi;
        struct obd_ioobj        *ioo;
        struct niobuf_remote    *rnb;
-       __u64                    start, end;
-       int                      lock_count;
+       int opc;
+       struct ldlm_prolong_args pa = { 0 };
 
        ENTRY;
 
@@ -2493,6 +2520,9 @@ static int ofd_rw_hpreq_check(struct ptlrpc_request *req)
         * Use LASSERT below because malformed RPCs should have
         * been filtered out in tgt_hpreq_handler().
         */
+       opc = lustre_msg_get_opc(req->rq_reqmsg);
+       LASSERT(opc == OST_READ || opc == OST_WRITE);
+
        ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
        LASSERT(ioo != NULL);
 
@@ -2500,21 +2530,28 @@ static int ofd_rw_hpreq_check(struct ptlrpc_request *req)
        LASSERT(rnb != NULL);
        LASSERT(!(rnb->rnb_flags & OBD_BRW_SRVLOCK));
 
-       start = rnb->rnb_offset;
+       pa.lpa_mode = LCK_PW | LCK_GROUP;
+       if (opc == OST_READ)
+               pa.lpa_mode |= LCK_PR;
+
+       pa.lpa_extent.start = rnb->rnb_offset;
        rnb += ioo->ioo_bufcnt - 1;
-       end = rnb->rnb_offset + rnb->rnb_len - 1;
+       pa.lpa_extent.end = rnb->rnb_offset + rnb->rnb_len - 1;
 
        DEBUG_REQ(D_RPCTRACE, req, "%s %s: refresh rw locks: "DFID
-                                  " ("LPU64"->"LPU64")\n",
-                 tgt_name(tsi->tsi_tgt), current->comm,
-                 PFID(&tsi->tsi_fid), start, end);
+                 " (%llu->%llu)\n", tgt_name(tsi->tsi_tgt),
+                 current->comm, PFID(&tsi->tsi_fid), pa.lpa_extent.start,
+                 pa.lpa_extent.end);
 
-       lock_count = ofd_prolong_extent_locks(tsi, start, end);
+       ofd_prolong_extent_locks(tsi, &pa);
 
        CDEBUG(D_DLMTRACE, "%s: refreshed %u locks timeout for req %p.\n",
-              tgt_name(tsi->tsi_tgt), lock_count, req);
+              tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
 
-       RETURN(lock_count > 0);
+       if (pa.lpa_blocks_cnt > 0)
+               RETURN(1);
+
+       RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
 }
 
 /**
@@ -2584,18 +2621,22 @@ static int ofd_punch_hpreq_lock_match(struct ptlrpc_request *req,
  * Implementation of ptlrpc_hpreq_ops::hpreq_lock_check for OST_PUNCH request.
  *
  * High-priority queue request check for whether the given punch request
- * (\a req) is blocking an LDLM lock cancel.
+ * (\a req) is blocking an LDLM lock cancel. Also checks whether the request is
+ * covered by an LDLM lock.
+ *
+
  *
  * \param[in] req      the incoming request
  *
  * \retval             1 if \a req is blocking an LDLM lock cancel
  * \retval             0 if it is not
+ * \retval             -ESTALE if lock is not found
  */
 static int ofd_punch_hpreq_check(struct ptlrpc_request *req)
 {
        struct tgt_session_info *tsi;
        struct obdo             *oa;
-       int                      lock_count;
+       struct ldlm_prolong_args pa = { 0 };
 
        ENTRY;
 
@@ -2608,17 +2649,24 @@ static int ofd_punch_hpreq_check(struct ptlrpc_request *req)
        LASSERT(!(oa->o_valid & OBD_MD_FLFLAGS &&
                  oa->o_flags & OBD_FL_SRVLOCK));
 
+       pa.lpa_mode = LCK_PW | LCK_GROUP;
+       pa.lpa_extent.start = oa->o_size;
+       pa.lpa_extent.end   = oa->o_blocks;
+
        CDEBUG(D_DLMTRACE,
-              "%s: refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
+              "%s: refresh locks: %llu/%llu (%llu->%llu)\n",
               tgt_name(tsi->tsi_tgt), tsi->tsi_resid.name[0],
-              tsi->tsi_resid.name[1], oa->o_size, oa->o_blocks);
+              tsi->tsi_resid.name[1], pa.lpa_extent.start, pa.lpa_extent.end);
 
-       lock_count = ofd_prolong_extent_locks(tsi, oa->o_size, oa->o_blocks);
+       ofd_prolong_extent_locks(tsi, &pa);
 
        CDEBUG(D_DLMTRACE, "%s: refreshed %u locks timeout for req %p.\n",
-              tgt_name(tsi->tsi_tgt), lock_count, req);
+              tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
 
-       RETURN(lock_count > 0);
+       if (pa.lpa_blocks_cnt > 0)
+               RETURN(1);
+
+       RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
 }
 
 /**
@@ -2669,7 +2717,9 @@ static void ofd_hp_brw(struct tgt_session_info *tsi)
                LASSERT(rnb != NULL); /* must exist after request preprocessing */
 
                /* no high priority if server lock is needed */
-               if (rnb->rnb_flags & OBD_BRW_SRVLOCK)
+               if (rnb->rnb_flags & OBD_BRW_SRVLOCK ||
+                   (lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg)
+                    & MSG_REPLAY))
                        return;
        }
        tgt_ses_req(tsi)->rq_ops = &ofd_hpreq_rw;
@@ -2688,8 +2738,10 @@ static void ofd_hp_punch(struct tgt_session_info *tsi)
 {
        LASSERT(tsi->tsi_ost_body != NULL); /* must exists if we are here */
        /* no high-priority if server lock is needed */
-       if (tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
-           tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK)
+       if ((tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
+            tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK) ||
+           tgt_conn_flags(tsi) & OBD_CONNECT_MDS ||
+           lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) & MSG_REPLAY)
                return;
        tgt_ses_req(tsi)->rq_ops = &ofd_hpreq_punch;
 }
@@ -2836,11 +2888,13 @@ struct lu_context_key ofd_thread_key = {
 static int ofd_init0(const struct lu_env *env, struct ofd_device *m,
                     struct lu_device_type *ldt, struct lustre_cfg *cfg)
 {
-       const char              *dev = lustre_cfg_string(cfg, 0);
-       struct ofd_thread_info  *info = NULL;
-       struct obd_device       *obd;
-       struct obd_statfs       *osfs;
-       int                      rc;
+       const char *dev = lustre_cfg_string(cfg, 0);
+       struct ofd_thread_info *info = NULL;
+       struct obd_device *obd;
+       struct obd_statfs *osfs;
+       struct lu_fid fid;
+       struct nm_config_file *nodemap_config;
+       int rc;
 
        ENTRY;
 
@@ -2865,7 +2919,6 @@ static int ofd_init0(const struct lu_env *env, struct ofd_device *m,
        ofd_slc_set(m);
        m->ofd_grant_compat_disable = 0;
        m->ofd_soft_sync_limit = OFD_SOFT_SYNC_LIMIT_DEFAULT;
-       m->ofd_brw_size = ONE_MB_BRW_SIZE;
 
        /* statfs data */
        spin_lock_init(&m->ofd_osfs_lock);
@@ -2931,13 +2984,18 @@ static int ofd_init0(const struct lu_env *env, struct ofd_device *m,
                CERROR("%s: can't get statfs data, rc %d\n", obd->obd_name, rc);
                GOTO(err_fini_stack, rc);
        }
-       if (!IS_PO2(osfs->os_bsize)) {
+       if (!is_power_of_2(osfs->os_bsize)) {
                CERROR("%s: blocksize (%d) is not a power of 2\n",
                                obd->obd_name, osfs->os_bsize);
                GOTO(err_fini_stack, rc = -EPROTO);
        }
        m->ofd_blockbits = fls(osfs->os_bsize) - 1;
 
+       if (ONE_MB_BRW_SIZE < (1U << m->ofd_blockbits))
+               m->ofd_brw_size = 1U << m->ofd_blockbits;
+       else
+               m->ofd_brw_size = ONE_MB_BRW_SIZE;
+
        m->ofd_precreate_batch = OFD_PRECREATE_BATCH_DEFAULT;
        if (osfs->os_bsize * osfs->os_blocks < OFD_PRECREATE_SMALL_FS)
                m->ofd_precreate_batch = OFD_PRECREATE_BATCH_SMALL;
@@ -2971,14 +3029,35 @@ static int ofd_init0(const struct lu_env *env, struct ofd_device *m,
        if (rc)
                GOTO(err_fini_lut, rc);
 
-       rc = ofd_start_inconsistency_verification_thread(m);
+       fid.f_seq = FID_SEQ_LOCAL_NAME;
+       fid.f_oid = 1;
+       fid.f_ver = 0;
+       rc = local_oid_storage_init(env, m->ofd_osd, &fid,
+                                   &m->ofd_los);
        if (rc != 0)
                GOTO(err_fini_fs, rc);
 
+       nodemap_config = nm_config_file_register_tgt(env, m->ofd_osd,
+                                                    m->ofd_los);
+       if (IS_ERR(nodemap_config))
+               GOTO(err_fini_los, rc = PTR_ERR(nodemap_config));
+
+       obd->u.obt.obt_nodemap_config_file = nodemap_config;
+
+       rc = ofd_start_inconsistency_verification_thread(m);
+       if (rc != 0)
+               GOTO(err_fini_nm, rc);
+
        tgt_adapt_sptlrpc_conf(&m->ofd_lut, 1);
 
        RETURN(0);
 
+err_fini_nm:
+       nm_config_file_deregister_tgt(env, obd->u.obt.obt_nodemap_config_file);
+       obd->u.obt.obt_nodemap_config_file = NULL;
+err_fini_los:
+       local_oid_storage_fini(env, m->ofd_los);
+       m->ofd_los = NULL;
 err_fini_fs:
        ofd_fs_cleanup(env, m);
 err_fini_lut:
@@ -3023,6 +3102,13 @@ static void ofd_fini(const struct lu_env *env, struct ofd_device *m)
        ofd_stop_inconsistency_verification_thread(m);
        lfsck_degister(env, m->ofd_osd);
        ofd_fs_cleanup(env, m);
+       nm_config_file_deregister_tgt(env, obd->u.obt.obt_nodemap_config_file);
+       obd->u.obt.obt_nodemap_config_file = NULL;
+
+       if (m->ofd_los != NULL) {
+               local_oid_storage_fini(env, m->ofd_los);
+               m->ofd_los = NULL;
+       }
 
        if (m->ofd_namespace != NULL) {
                ldlm_namespace_free_post(m->ofd_namespace);