Whamcloud - gitweb
LU-9285 osp: revert patches LU-8367 and LU-8973
[fs/lustre-release.git] / lustre / osp / osp_precreate.c
index a89c9f5..4d867fa 100644 (file)
@@ -577,7 +577,6 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
                RETURN(rc);
        }
 
-       LASSERT(d->opd_pre->osp_pre_delorphan_sent != 0);
        spin_lock(&d->opd_pre_lock);
        if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
                d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
@@ -783,20 +782,18 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env,
        int                      update_status = 0;
        int                      rc;
        int                      diff;
-       struct lu_fid            fid;
 
        ENTRY;
 
        /*
-        * wait for local recovery to finish, so we can cleanup orphans.
-        * orphans are all objects since "last used" (assigned).
-        * consider reserved objects as created otherwise we can get into
-        * a livelock when one blocked thread holding a reservation can
-        * block recovery. see LU-8367 for the details. in some cases this
-        * can result in gaps (i.e. leaked objects), but we've got LFSCK...
-        *
-        * do not allow new reservations because they may end up getting
-        * orphans being cleaned up below. so we block new reservations.
+        * wait for local recovery to finish, so we can cleanup orphans
+        * orphans are all objects since "last used" (assigned), but
+        * there might be objects reserved and in some cases they won't
+        * be used. we can't cleanup them till we're sure they won't be
+        * used. also can't we allow new reservations because they may
+        * end up getting orphans being cleaned up below. so we block
+        * new reservations and wait till all reserved objects either
+        * user or released.
         */
        spin_lock(&d->opd_pre_lock);
        d->opd_pre_recovering = 1;
@@ -806,12 +803,16 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env,
         * catch all osp_precreate_reserve() calls who find
         * "!opd_pre_recovering".
         */
-       l_wait_event(d->opd_pre_waitq, d->opd_recovery_completed ||
+       l_wait_event(d->opd_pre_waitq,
+                    (!d->opd_pre_reserved && d->opd_recovery_completed) ||
                     !osp_precreate_running(d) || d->opd_got_disconnected,
                     &lwi);
        if (!osp_precreate_running(d) || d->opd_got_disconnected)
                GOTO(out, rc = -EAGAIN);
 
+       CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
+              d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
+
        *last_fid = d->opd_last_used_fid;
        /* The OSP should already get the valid seq now */
        LASSERT(!fid_is_zero(last_fid));
@@ -840,24 +841,10 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env,
        if (body == NULL)
                GOTO(out, rc = -EPROTO);
 
-       body->oa.o_flags = 0;
+       body->oa.o_flags = OBD_FL_DELORPHAN;
        body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
 
-       /* unless this is the very first DELORPHAN (when we really
-        * can destroy some orphans), just tell OST to recreate
-        * missing objects in our precreate pool */
-       spin_lock(&d->opd_pre_lock);
-       if (d->opd_pre->osp_pre_delorphan_sent) {
-               fid = d->opd_pre_last_created_fid;
-       } else {
-               fid = d->opd_last_used_fid;
-               body->oa.o_flags = OBD_FL_DELORPHAN;
-       }
-       spin_unlock(&d->opd_pre_lock);
-       fid_to_ostid(&fid, &body->oa.o_oi);
-
-       CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
-              d->opd_obd->obd_name, PFID(&fid));
+       fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
 
        ptlrpc_request_set_replen(req);
 
@@ -880,10 +867,10 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env,
        ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
 
        spin_lock(&d->opd_pre_lock);
-       diff = osp_fid_diff(&fid, last_fid);
+       diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
        if (diff > 0) {
                d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
-               d->opd_pre_last_created_fid = *last_fid;
+               d->opd_pre_last_created_fid = d->opd_last_used_fid;
        } else {
                d->opd_pre_create_count = OST_MIN_PRECREATE;
                d->opd_pre_last_created_fid = *last_fid;
@@ -894,11 +881,9 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env,
         */
        LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
                LUSTRE_DATA_SEQ_MAX_WIDTH);
-       if (d->opd_pre->osp_pre_delorphan_sent == 0)
-               d->opd_pre_used_fid = d->opd_pre_last_created_fid;
+       d->opd_pre_used_fid = d->opd_pre_last_created_fid;
        d->opd_pre_create_slow = 0;
        spin_unlock(&d->opd_pre_lock);
-       d->opd_pre->osp_pre_delorphan_sent = 1;
 
        CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
               "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
@@ -1386,12 +1371,6 @@ int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
        if (d->opd_pre_max_create_count == 0)
                RETURN(-ENOBUFS);
 
-       if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_OSP_PRECREATE_WAIT)) {
-               if (d->opd_index == cfs_fail_val)
-                       OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_OSP_PRECREATE_WAIT,
-                                        obd_timeout);
-       }
-
        /*
         * wait till:
         *  - preallocation is done