Whamcloud - gitweb
LU-12616 obclass: fix MDS start/stop race
[fs/lustre-release.git] / lustre / osp / osp_object.c
index 0604892..336eb54 100644 (file)
@@ -431,11 +431,13 @@ static int osp_get_attr_from_reply(const struct lu_env *env,
                lustre_swab_obdo(wobdo);
 
        lustre_get_wire_obdo(NULL, lobdo, wobdo);
-       spin_lock(&obj->opo_lock);
-       la_from_obdo(&obj->opo_attr, lobdo, lobdo->o_valid);
-       if (attr != NULL)
-               *attr = obj->opo_attr;
-       spin_unlock(&obj->opo_lock);
+       if (obj) {
+               spin_lock(&obj->opo_lock);
+               la_from_obdo(&obj->opo_attr, lobdo, lobdo->o_valid);
+               spin_unlock(&obj->opo_lock);
+       }
+       if (attr)
+               la_from_obdo(attr, lobdo, lobdo->o_valid);
 
        return 0;
 }
@@ -542,7 +544,7 @@ int osp_attr_get(const struct lu_env *env, struct dt_object *dt,
        struct osp_update_request       *update;
        struct object_update_reply      *reply;
        struct ptlrpc_request           *req = NULL;
-       int                             rc = 0;
+       int                             invalidated, cache = 0, rc = 0;
        ENTRY;
 
        if (is_ost_obj(&dt->do_lu) && obj->opo_non_exist)
@@ -561,21 +563,31 @@ int osp_attr_get(const struct lu_env *env, struct dt_object *dt,
        if (IS_ERR(update))
                RETURN(PTR_ERR(update));
 
-       rc = osp_update_rpc_pack(env, attr_get, update, OUT_ATTR_GET,
+       rc = OSP_UPDATE_RPC_PACK(env, out_attr_get_pack, update,
                                 lu_object_fid(&dt->do_lu));
        if (rc != 0) {
                CERROR("%s: Insert update error "DFID": rc = %d\n",
                       dev->dd_lu_dev.ld_obd->obd_name,
                       PFID(lu_object_fid(&dt->do_lu)), rc);
 
-               GOTO(out, rc);
+               GOTO(out_req, rc);
        }
 
+       invalidated = atomic_read(&obj->opo_invalidate_seq);
+
        rc = osp_remote_sync(env, osp, update, &req);
+
+       down_read(&obj->opo_invalidate_sem);
+       if (invalidated == atomic_read(&obj->opo_invalidate_seq)) {
+               /* no invalited has came so far, we can cache the attrs */
+               cache = 1;
+       }
+
        if (rc != 0) {
                if (rc == -ENOENT) {
                        osp2lu_obj(obj)->lo_header->loh_attr &= ~LOHA_EXISTS;
-                       obj->opo_non_exist = 1;
+                       if (cache)
+                               obj->opo_non_exist = 1;
                } else {
                        CERROR("%s:osp_attr_get update error "DFID": rc = %d\n",
                               dev->dd_lu_dev.ld_obd->obd_name,
@@ -593,17 +605,22 @@ int osp_attr_get(const struct lu_env *env, struct dt_object *dt,
        if (reply == NULL || reply->ourp_magic != UPDATE_REPLY_MAGIC)
                GOTO(out, rc = -EPROTO);
 
-       rc = osp_get_attr_from_reply(env, reply, req, attr, obj, 0);
+       rc = osp_get_attr_from_reply(env, reply, req, attr,
+                                    cache ? obj : NULL, 0);
        if (rc != 0)
                GOTO(out, rc);
 
        spin_lock(&obj->opo_lock);
-       obj->opo_stale = 0;
+       if (cache)
+               obj->opo_stale = 0;
        spin_unlock(&obj->opo_lock);
 
        GOTO(out, rc);
 
 out:
+       up_read(&obj->opo_invalidate_sem);
+
+out_req:
        if (req != NULL)
                ptlrpc_req_finished(req);
 
@@ -709,8 +726,36 @@ static int osp_attr_set(const struct lu_env *env, struct dt_object *dt,
                RETURN(0);
 
        if (!is_only_remote_trans(th)) {
-               rc = osp_sync_add(env, o, MDS_SETATTR64_REC, th, attr);
-               /* XXX: send new uid/gid to OST ASAP? */
+               if (attr->la_flags & LUSTRE_SET_SYNC_FL) {
+                       struct ptlrpc_request *req = NULL;
+                       struct osp_update_request *update = NULL;
+                       struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
+
+                       update = osp_update_request_create(&osp->opd_dt_dev);
+                       if (IS_ERR(update))
+                               RETURN(PTR_ERR(update));
+
+                       rc = OSP_UPDATE_RPC_PACK(env, out_attr_set_pack, update,
+                                                lu_object_fid(&dt->do_lu),
+                                                attr);
+                       if (rc != 0) {
+                               CERROR("%s: update error "DFID": rc = %d\n",
+                                      osp->opd_obd->obd_name,
+                                      PFID(lu_object_fid(&dt->do_lu)), rc);
+
+                               osp_update_request_destroy(env, update);
+                               RETURN(rc);
+                       }
+
+                       rc = osp_remote_sync(env, osp, update, &req);
+                       if (req != NULL)
+                               ptlrpc_req_finished(req);
+
+                       osp_update_request_destroy(env, update);
+               } else {
+                       rc = osp_sync_add(env, o, MDS_SETATTR64_REC, th, attr);
+                       /* XXX: send new uid/gid to OST ASAP? */
+               }
        } else {
                struct lu_attr  *la;
 
@@ -910,7 +955,7 @@ int osp_xattr_get(const struct lu_env *env, struct dt_object *dt,
        struct object_update_reply *reply;
        struct osp_xattr_entry  *oxe    = NULL;
        const char *dname = osp_dto2name(obj);
-       int rc = 0;
+       int invalidated, rc = 0;
        ENTRY;
 
        LASSERT(buf != NULL);
@@ -930,6 +975,8 @@ int osp_xattr_get(const struct lu_env *env, struct dt_object *dt,
        if (unlikely(obj->opo_non_exist))
                RETURN(-ENOENT);
 
+       invalidated = atomic_read(&obj->opo_invalidate_seq);
+
        oxe = osp_oac_xattr_find(obj, name, false);
        if (oxe != NULL) {
                spin_lock(&obj->opo_lock);
@@ -958,17 +1005,40 @@ unlock:
        }
        update = osp_update_request_create(dev);
        if (IS_ERR(update))
-               GOTO(out, rc = PTR_ERR(update));
+               GOTO(out_req, rc = PTR_ERR(update));
 
-       rc = osp_update_rpc_pack(env, xattr_get, update, OUT_XATTR_GET,
+       rc = OSP_UPDATE_RPC_PACK(env, out_xattr_get_pack, update,
                                 lu_object_fid(&dt->do_lu), name, buf->lb_len);
        if (rc != 0) {
                CERROR("%s: Insert update error "DFID": rc = %d\n",
                       dname, PFID(lu_object_fid(&dt->do_lu)), rc);
-               GOTO(out, rc);
+               GOTO(out_req, rc);
        }
 
        rc = osp_remote_sync(env, osp, update, &req);
+
+       down_read(&obj->opo_invalidate_sem);
+       if (invalidated != atomic_read(&obj->opo_invalidate_seq)) {
+               /* invalidated has been requested, we can't cache the result */
+               if (rc < 0) {
+                       if (rc == -ENOENT)
+                               dt->do_lu.lo_header->loh_attr &= ~LOHA_EXISTS;
+                       GOTO(out, rc);
+               }
+               reply = req_capsule_server_sized_get(&req->rq_pill,
+                                                    &RMF_OUT_UPDATE_REPLY,
+                               OUT_UPDATE_REPLY_SIZE);
+               if (reply->ourp_magic != UPDATE_REPLY_MAGIC) {
+                       CERROR("%s: Wrong version %x expected %x "DFID
+                              ": rc = %d\n", dname, reply->ourp_magic,
+                              UPDATE_REPLY_MAGIC,
+                              PFID(lu_object_fid(&dt->do_lu)), -EPROTO);
+                       GOTO(out, rc = -EPROTO);
+               }
+               rc = object_update_result_data_get(reply, rbuf, 0);
+               GOTO(out, rc);
+       }
+
        if (rc < 0) {
                if (rc == -ENOENT) {
                        dt->do_lu.lo_header->loh_attr &= ~LOHA_EXISTS;
@@ -1056,6 +1126,9 @@ unlock:
        GOTO(out, rc);
 
 out:
+       up_read(&obj->opo_invalidate_sem);
+
+out_req:
        if (rc > 0 && buf->lb_buf) {
                if (unlikely(buf->lb_len < rbuf->lb_len))
                        rc = -ERANGE;
@@ -1130,10 +1203,10 @@ int osp_xattr_set(const struct lu_env *env, struct dt_object *dt,
                  const struct lu_buf *buf, const char *name, int fl,
                  struct thandle *th)
 {
-       struct osp_object       *o = dt2osp_obj(dt);
+       struct osp_object *o = dt2osp_obj(dt);
        struct osp_update_request *update;
-       struct osp_xattr_entry  *oxe;
-       int                     rc;
+       struct osp_xattr_entry *oxe;
+       int rc;
        ENTRY;
 
        update = thandle_to_osp_update_request(th);
@@ -1142,7 +1215,7 @@ int osp_xattr_set(const struct lu_env *env, struct dt_object *dt,
        CDEBUG(D_INODE, DFID" set xattr '%s' with size %zd\n",
               PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
 
-       rc = osp_update_rpc_pack(env, xattr_set, update, OUT_XATTR_SET,
+       rc = OSP_UPDATE_RPC_PACK(env, out_xattr_set_pack, update,
                                 lu_object_fid(&dt->do_lu), buf, name, fl);
        if (rc != 0)
                RETURN(rc);
@@ -1219,16 +1292,15 @@ int osp_xattr_del(const struct lu_env *env, struct dt_object *dt,
                  const char *name, struct thandle *th)
 {
        struct osp_update_request *update;
-       const struct lu_fid      *fid = lu_object_fid(&dt->do_lu);
-       struct osp_object        *o     = dt2osp_obj(dt);
-       struct osp_xattr_entry   *oxe;
-       int                       rc;
+       const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
+       struct osp_object *o = dt2osp_obj(dt);
+       struct osp_xattr_entry *oxe;
+       int rc;
 
        update = thandle_to_osp_update_request(th);
        LASSERT(update != NULL);
 
-       rc = osp_update_rpc_pack(env, xattr_del, update, OUT_XATTR_DEL,
-                                fid, name);
+       rc = OSP_UPDATE_RPC_PACK(env, out_xattr_del_pack, update, fid, name);
        if (rc != 0)
                return rc;
 
@@ -1273,12 +1345,24 @@ int osp_invalidate(const struct lu_env *env, struct dt_object *dt)
 
        CDEBUG(D_HA, "Invalidate osp_object "DFID"\n",
               PFID(lu_object_fid(&dt->do_lu)));
-       osp_obj_invalidate_cache(obj);
+
+       /* serialize attr/EA set vs. invalidation */
+       down_write(&obj->opo_invalidate_sem);
+
+       /* this should invalidate all in-flights */
+       atomic_inc(&obj->opo_invalidate_seq);
 
        spin_lock(&obj->opo_lock);
-       obj->opo_stale = 1;
+       /* do not mark new objects stale */
+       if (obj->opo_attr.la_valid)
+               obj->opo_stale = 1;
+       obj->opo_non_exist = 0;
        spin_unlock(&obj->opo_lock);
 
+       osp_obj_invalidate_cache(obj);
+
+       up_write(&obj->opo_invalidate_sem);
+
        RETURN(0);
 }
 
@@ -1350,10 +1434,8 @@ static int osp_declare_create(const struct lu_env *env, struct dt_object *dt,
 
        if (unlikely(!fid_is_zero(fid))) {
                /* replay case: caller knows fid */
-               osi->osi_off = sizeof(osi->osi_id) * d->opd_index;
-               osi->osi_lb.lb_len = sizeof(osi->osi_id);
-               osi->osi_lb.lb_buf = NULL;
-
+               osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, NULL,
+                                  d->opd_index);
                rc = dt_declare_record_write(env, d->opd_last_used_oid_file,
                                             &osi->osi_lb, osi->osi_off,
                                             local_th);
@@ -1377,9 +1459,8 @@ static int osp_declare_create(const struct lu_env *env, struct dt_object *dt,
                o->opo_reserved = 1;
 
                /* common for all OSPs file hystorically */
-               osi->osi_off = sizeof(osi->osi_id) * d->opd_index;
-               osi->osi_lb.lb_len = sizeof(osi->osi_id);
-               osi->osi_lb.lb_buf = NULL;
+               osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, NULL,
+                                  d->opd_index);
                rc = dt_declare_record_write(env, d->opd_last_used_oid_file,
                                             &osi->osi_lb, osi->osi_off,
                                             local_th);
@@ -1422,7 +1503,6 @@ static int osp_create(const struct lu_env *env, struct dt_object *dt,
        int                     rc = 0;
        struct lu_fid           *fid = &osi->osi_fid;
        struct thandle          *local_th;
-       struct lu_fid           *last_fid = &d->opd_last_used_fid;
        ENTRY;
 
        if (is_only_remote_trans(th) &&
@@ -1502,13 +1582,8 @@ static int osp_create(const struct lu_env *env, struct dt_object *dt,
 
        /* Only need update last_used oid file, seq file will only be update
         * during seq rollover */
-       if (fid_is_idif((last_fid)))
-               osi->osi_id = fid_idif_id(fid_seq(last_fid),
-                                         fid_oid(last_fid), fid_ver(last_fid));
-       else
-               osi->osi_id = fid_oid(last_fid);
        osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off,
-                          &osi->osi_id, d->opd_index);
+                          &d->opd_last_id, d->opd_index);
 
        rc = dt_record_write(env, d->opd_last_used_oid_file, &osi->osi_lb,
                             &osi->osi_off, local_th);
@@ -1544,7 +1619,9 @@ int osp_declare_destroy(const struct lu_env *env, struct dt_object *dt,
        ENTRY;
 
        LASSERT(!osp->opd_connect_mdt);
-       rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th);
+
+       if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ))
+               rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th);
 
        RETURN(rc);
 }
@@ -1578,11 +1655,14 @@ static int osp_destroy(const struct lu_env *env, struct dt_object *dt,
        o->opo_non_exist = 1;
 
        LASSERT(!osp->opd_connect_mdt);
-       /* once transaction is committed put proper command on
-        * the queue going to our OST. */
-       rc = osp_sync_add(env, o, MDS_UNLINK64_REC, th, NULL);
-       if (rc < 0)
-               RETURN(rc);
+
+       if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ)) {
+               /* once transaction is committed put proper command on
+                * the queue going to our OST. */
+               rc = osp_sync_add(env, o, MDS_UNLINK64_REC, th, NULL);
+               if (rc < 0)
+                       RETURN(rc);
+       }
 
        /* not needed in cache any more */
        set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
@@ -1611,8 +1691,7 @@ static int osp_orphan_index_insert(const struct lu_env *env,
                                   struct dt_object *dt,
                                   const struct dt_rec *rec,
                                   const struct dt_key *key,
-                                  struct thandle *handle,
-                                  int ignore_quota)
+                                  struct thandle *handle)
 {
        return -EOPNOTSUPP;
 }
@@ -1944,6 +2023,13 @@ again:
                it->ooi_pos_ent++;
                if (it->ooi_pos_ent < idxpage->lip_nr) {
                        if (it->ooi_rec_size ==
+                                       sizeof(struct lu_orphan_rec_v3)) {
+                               it->ooi_ent =
+                               (struct lu_orphan_ent_v3 *)idxpage->lip_entries+
+                                                       it->ooi_pos_ent;
+                               if (it->ooi_swab)
+                                       lustre_swab_orphan_ent_v3(it->ooi_ent);
+                       } else if (it->ooi_rec_size ==
                                        sizeof(struct lu_orphan_rec_v2)) {
                                it->ooi_ent =
                                (struct lu_orphan_ent_v2 *)idxpage->lip_entries+
@@ -2002,7 +2088,13 @@ static int osp_orphan_it_rec(const struct lu_env *env, const struct dt_it *di,
        struct osp_it *it = (struct osp_it *)di;
 
        if (likely(it->ooi_ent)) {
-               if (it->ooi_rec_size == sizeof(struct lu_orphan_rec_v2)) {
+               if (it->ooi_rec_size == sizeof(struct lu_orphan_rec_v3)) {
+                       struct lu_orphan_ent_v3 *ent =
+                               (struct lu_orphan_ent_v3 *)it->ooi_ent;
+
+                       *(struct lu_orphan_rec_v3 *)rec = ent->loe_rec;
+               } else if (it->ooi_rec_size ==
+                               sizeof(struct lu_orphan_rec_v2)) {
                        struct lu_orphan_ent_v2 *ent =
                                (struct lu_orphan_ent_v2 *)it->ooi_ent;
 
@@ -2152,6 +2244,7 @@ static int osp_object_init(const struct lu_env *env, struct lu_object *o,
        o->lo_header->loh_attr |= LOHA_REMOTE;
        INIT_LIST_HEAD(&po->opo_xattr_list);
        INIT_LIST_HEAD(&po->opo_invalidate_cb_list);
+       init_rwsem(&po->opo_invalidate_sem);
 
        if (is_ost_obj(o)) {
                po->opo_obj.do_ops = &osp_obj_ops;
@@ -2241,6 +2334,14 @@ static void osp_object_release(const struct lu_env *env, struct lu_object *o)
                d->opd_pre_reserved--;
                spin_unlock(&d->opd_pre_lock);
 
+               /*
+                * Check that osp_precreate_cleanup_orphans is not blocked
+                * due to opd_pre_reserved > 0.
+                */
+               if (unlikely(d->opd_pre_reserved == 0 &&
+                            (d->opd_pre_recovering || d->opd_pre_status)))
+                       wake_up(&d->opd_pre_waitq);
+
                /* not needed in cache any more */
                set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
        }