b=17485 Don't reset the lov_objid to last_id sent by OST

i=andrew
i=hongchao

Resetting the lov_objid values to last_id reported by
the OST in mds_lov_get_objid() unconditionally is just wrong and can cause
the same object to be allocated twice by the MDS.

The SKIP_ORPHAN feature is supposed to work as follows:
* during orphan recovery, the MDS must supply to the OST the lov_objid
  value;
* on the OST side:
  - if MDS's lov_objid > OST's last_id, the OST recreates the missing objects
    up to lov_objid and packs this value in the reply;
  - if OST's last_id > MDS's lov_objid, the OST destroys orphan objects from
    lov_objid + 1 up to last_id and does *not* change last_id. The OST packs
    in the reply the value of last_id which is from where the MDS should
    restart allocations if it does not want to reuse orphans.
* the MDS sets next_id to the value reported by the OST in the orphan recovery
  request.

If we always discard the lov_objid values as done currently in mds_lov_get_objid(),
orphan recovery is just a noop and neither clears orphans nor recreates missing
objects.

diff --git a/lustre/mds/mds_lov.c b/lustre/mds/mds_lov.c
index 29419b3..b478bd9 100644 (file)
--- a/lustre/mds/mds_lov.c
+++ b/lustre/mds/mds_lov.c
@@ -474,7 +474,6 @@ static int mds_lov_get_objid(struct obd_device * obd,
         unsigned int page;
         unsigned int off;
         obd_id *data;
-        __u64 connect_flags;
         __u32 size;
         int rc = 0;
         ENTRY;
@@ -485,14 +484,8 @@ static int mds_lov_get_objid(struct obd_device * obd,
         off = idx % OBJID_PER_PAGE();
 
         data = mds->mds_lov_page_array[page];
-        size = sizeof(__u64);
-        connect_flags = idx;
-        rc = obd_get_info(lov_exp, sizeof(KEY_CONNECT_FLAG), KEY_CONNECT_FLAG, 
-                          &size, &connect_flags, NULL);
-        if (rc)
-                GOTO(out, rc);
 
-        if (data[off] < 2 || connect_flags & OBD_CONNECT_SKIP_ORPHAN) {
+        if (data[off] < 2) {
                 /* We never read this lastid; ask the osc */
                 struct obd_id_info lastid;
 

diff --git a/lustre/obdfilter/filter.c b/lustre/obdfilter/filter.c
index 228f59c..f11472f 100644 (file)
--- a/lustre/obdfilter/filter.c
+++ b/lustre/obdfilter/filter.c
@@ -3124,7 +3124,14 @@ static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
                 filter_set_last_id(filter, id, doa.o_gr);
                 rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
         } else {
-                /* don't reuse orphan object, return last used objid */
+                /*
+                 * We have destroyed orphan objects, but don't want to reuse
+                 * them. Therefore we don't reset last_id to the last created
+                 * objects. Instead, we report back to the MDS the object id
+                 * of the last orphan, so that the MDS can restart allocating
+                 * objects from this id + 1 and thus skip the whole orphan
+                 * object id range
+                 */
                 oa->o_id = last;
                 rc = 0;
         }

diff --git a/lustre/osc/osc_create.c b/lustre/osc/osc_create.c
index 42fe4ef..535e5bf 100644 (file)
--- a/lustre/osc/osc_create.c
+++ b/lustre/osc/osc_create.c
@@ -571,6 +571,11 @@ int osc_create(struct obd_export *exp, struct obdo *oa,
 
                         ocd = &imp->imp_connect_data;
                         if (ocd->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN) {
+                                /*
+                                 * The OST reports back in oa->o_id from where
+                                 * we should restart in order to skip orphan
+                                 * objects
+                                 */
                                 CDEBUG(D_HA, "%s: Skip orphan set, reset last "
                                        "objid\n", oscc->oscc_obd->obd_name);
                                 oscc->oscc_next_id = oa->o_id + 1;