+/**
+ * Unlock all previously locked sparse areas for given resource.
+ *
+ * This function goes through list of locked regions, unlocking and freeing
+ * them one-by-one.
+ *
+ * \param[in] ns LDLM namespace
+ * \param[in] locked list head of regions list
+ */
+static void
+unlock_zero_regions(struct ldlm_namespace *ns, struct list_head *locked)
+{
+ struct locked_region *entry, *temp;
+
+ list_for_each_entry_safe(entry, temp, locked, list) {
+ CDEBUG(D_OTHER, "ost unlock lh=%p\n", &entry->lh);
+ tgt_extent_unlock(&entry->lh, LCK_PR);
+ list_del(&entry->list);
+ OBD_FREE_PTR(entry);
+ }
+}
+
+/**
+ * OFD request handler for OST_GET_INFO RPC.
+ *
+ * This is OFD-specific part of request handling. The OFD-specific keys are:
+ * - KEY_LAST_ID (obsolete)
+ * - KEY_FIEMAP
+ * - KEY_LAST_FID
+ *
+ * This function reads needed data from storage and fills reply with it.
+ *
+ * Note: the KEY_LAST_ID is obsolete, replaced by KEY_LAST_FID on newer MDTs,
+ * and is kept for compatibility.
+ *
+ * \param[in] tsi target session environment for this request
+ *
+ * \retval 0 if successful
+ * \retval negative value on error
+ */
+static int ofd_get_info_hdl(struct tgt_session_info *tsi)
+{
+ struct obd_export *exp = tsi->tsi_exp;
+ struct ofd_device *ofd = ofd_exp(exp);
+ struct ofd_thread_info *fti = tsi2ofd_info(tsi);
+ void *key;
+ int keylen;
+ int replylen, rc = 0;
+
+ ENTRY;
+
+ /* this common part for get_info rpc */
+ key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
+ if (key == NULL) {
+ DEBUG_REQ(D_HA, tgt_ses_req(tsi), "no get_info key");
+ RETURN(err_serious(-EPROTO));
+ }
+ keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
+ RCL_CLIENT);
+
+ if (KEY_IS(KEY_LAST_ID)) {
+ u64 *last_id;
+ struct ofd_seq *oseq;
+
+ req_capsule_extend(tsi->tsi_pill, &RQF_OST_GET_INFO_LAST_ID);
+ rc = req_capsule_server_pack(tsi->tsi_pill);
+ if (rc)
+ RETURN(err_serious(rc));
+
+ last_id = req_capsule_server_get(tsi->tsi_pill, &RMF_OBD_ID);
+
+ oseq = ofd_seq_load(tsi->tsi_env, ofd,
+ (u64)exp->exp_filter_data.fed_group);
+ if (IS_ERR(oseq))
+ rc = -EFAULT;
+ else
+ *last_id = ofd_seq_last_oid(oseq);
+ ofd_seq_put(tsi->tsi_env, oseq);
+ } else if (KEY_IS(KEY_FIEMAP)) {
+ struct ll_fiemap_info_key *fm_key;
+ struct ll_user_fiemap *fiemap;
+ struct lu_fid *fid;
+
+ req_capsule_extend(tsi->tsi_pill, &RQF_OST_GET_INFO_FIEMAP);
+
+ fm_key = req_capsule_client_get(tsi->tsi_pill, &RMF_FIEMAP_KEY);
+ rc = tgt_validate_obdo(tsi, &fm_key->oa);
+ if (rc)
+ RETURN(err_serious(rc));
+
+ fid = &fm_key->oa.o_oi.oi_fid;
+
+ CDEBUG(D_INODE, "get FIEMAP of object "DFID"\n", PFID(fid));
+
+ replylen = fiemap_count_to_size(fm_key->fiemap.fm_extent_count);
+ req_capsule_set_size(tsi->tsi_pill, &RMF_FIEMAP_VAL,
+ RCL_SERVER, replylen);
+
+ rc = req_capsule_server_pack(tsi->tsi_pill);
+ if (rc)
+ RETURN(err_serious(rc));
+
+ fiemap = req_capsule_server_get(tsi->tsi_pill, &RMF_FIEMAP_VAL);
+ if (fiemap == NULL)
+ RETURN(-ENOMEM);
+
+ *fiemap = fm_key->fiemap;
+ rc = ofd_fiemap_get(tsi->tsi_env, ofd, fid, fiemap);
+
+ /* LU-3219: Lock the sparse areas to make sure dirty
+ * flushed back from client, then call fiemap again. */
+ if (fm_key->oa.o_valid & OBD_MD_FLFLAGS &&
+ fm_key->oa.o_flags & OBD_FL_SRVLOCK) {
+ struct list_head locked;
+
+ INIT_LIST_HEAD(&locked);
+ ost_fid_build_resid(fid, &fti->fti_resid);
+ rc = lock_zero_regions(ofd->ofd_namespace,
+ &fti->fti_resid, fiemap,
+ &locked);
+ if (rc == 0 && !list_empty(&locked)) {
+ rc = ofd_fiemap_get(tsi->tsi_env, ofd, fid,
+ fiemap);
+ unlock_zero_regions(ofd->ofd_namespace,
+ &locked);
+ }
+ }
+ } else if (KEY_IS(KEY_LAST_FID)) {
+ struct ofd_device *ofd = ofd_exp(exp);
+ struct ofd_seq *oseq;
+ struct lu_fid *fid;
+ int rc;
+
+ req_capsule_extend(tsi->tsi_pill, &RQF_OST_GET_INFO_LAST_FID);
+ rc = req_capsule_server_pack(tsi->tsi_pill);
+ if (rc)
+ RETURN(err_serious(rc));
+
+ fid = req_capsule_client_get(tsi->tsi_pill, &RMF_FID);
+ if (fid == NULL)
+ RETURN(err_serious(-EPROTO));
+
+ fid_le_to_cpu(&fti->fti_ostid.oi_fid, fid);
+
+ fid = req_capsule_server_get(tsi->tsi_pill, &RMF_FID);
+ if (fid == NULL)
+ RETURN(-ENOMEM);
+
+ oseq = ofd_seq_load(tsi->tsi_env, ofd,
+ ostid_seq(&fti->fti_ostid));
+ if (IS_ERR(oseq))
+ RETURN(PTR_ERR(oseq));
+
+ rc = ostid_to_fid(fid, &oseq->os_oi,
+ ofd->ofd_lut.lut_lsd.lsd_osd_index);
+ if (rc != 0)
+ GOTO(out_put, rc);
+
+ CDEBUG(D_HA, "%s: LAST FID is "DFID"\n", ofd_name(ofd),
+ PFID(fid));
+out_put:
+ ofd_seq_put(tsi->tsi_env, oseq);
+ } else {
+ CERROR("%s: not supported key %s\n", tgt_name(tsi->tsi_tgt),
+ (char *)key);
+ rc = -EOPNOTSUPP;
+ }
+ ofd_counter_incr(tsi->tsi_exp, LPROC_OFD_STATS_GET_INFO,
+ tsi->tsi_jobid, 1);
+
+ RETURN(rc);
+}
+
+/**
+ * OFD request handler for OST_GETATTR RPC.
+ *
+ * This is OFD-specific part of request handling. It finds the OFD object
+ * by its FID, gets attributes from storage and packs result to the reply.
+ *
+ * \param[in] tsi target session environment for this request
+ *
+ * \retval 0 if successful
+ * \retval negative value on error
+ */
+static int ofd_getattr_hdl(struct tgt_session_info *tsi)
+{
+ struct ofd_thread_info *fti = tsi2ofd_info(tsi);
+ struct ofd_device *ofd = ofd_exp(tsi->tsi_exp);
+ struct ost_body *repbody;
+ struct lustre_handle lh = { 0 };
+ struct ofd_object *fo;
+ __u64 flags = 0;
+ ldlm_mode_t lock_mode = LCK_PR;
+ bool srvlock;
+ int rc;
+ ENTRY;
+
+ LASSERT(tsi->tsi_ost_body != NULL);
+
+ repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
+ if (repbody == NULL)
+ RETURN(-ENOMEM);
+
+ repbody->oa.o_oi = tsi->tsi_ost_body->oa.o_oi;
+ repbody->oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
+
+ srvlock = tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
+ tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK;
+
+ if (srvlock) {
+ if (unlikely(tsi->tsi_ost_body->oa.o_flags & OBD_FL_FLUSH))
+ lock_mode = LCK_PW;
+
+ rc = tgt_extent_lock(tsi->tsi_tgt->lut_obd->obd_namespace,
+ &tsi->tsi_resid, 0, OBD_OBJECT_EOF, &lh,
+ lock_mode, &flags);
+ if (rc != 0)
+ RETURN(rc);
+ }
+
+ fo = ofd_object_find_exists(tsi->tsi_env, ofd, &tsi->tsi_fid);
+ if (IS_ERR(fo))
+ GOTO(out, rc = PTR_ERR(fo));
+
+ rc = ofd_attr_get(tsi->tsi_env, fo, &fti->fti_attr);
+ if (rc == 0) {
+ __u64 curr_version;
+
+ 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,
+ ofd_object_child(fo));
+ if ((__s64)curr_version != -EOPNOTSUPP) {
+ repbody->oa.o_valid |= OBD_MD_FLDATAVERSION;
+ repbody->oa.o_data_version = curr_version;
+ }
+ }
+
+ ofd_object_put(tsi->tsi_env, fo);
+out:
+ if (srvlock)
+ tgt_extent_unlock(&lh, lock_mode);
+
+ ofd_counter_incr(tsi->tsi_exp, LPROC_OFD_STATS_GETATTR,
+ tsi->tsi_jobid, 1);
+
+ repbody->oa.o_valid |= OBD_MD_FLFLAGS;
+ repbody->oa.o_flags = OBD_FL_FLUSH;
+
+ RETURN(rc);
+}
+
+/**
+ * OFD request handler for OST_SETATTR RPC.
+ *
+ * This is OFD-specific part of request handling. It finds the OFD object
+ * by its FID, sets attributes from request and packs result to the reply.
+ *
+ * \param[in] tsi target session environment for this request
+ *
+ * \retval 0 if successful
+ * \retval negative value on error
+ */
+static int ofd_setattr_hdl(struct tgt_session_info *tsi)
+{
+ struct ofd_thread_info *fti = tsi2ofd_info(tsi);
+ struct ofd_device *ofd = ofd_exp(tsi->tsi_exp);
+ struct ost_body *body = tsi->tsi_ost_body;
+ struct ost_body *repbody;
+ struct ldlm_resource *res;
+ struct ofd_object *fo;
+ struct filter_fid *ff = NULL;
+ int rc = 0;
+
+ ENTRY;
+
+ LASSERT(body != NULL);
+
+ repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
+ if (repbody == NULL)
+ RETURN(-ENOMEM);
+
+ repbody->oa.o_oi = body->oa.o_oi;
+ repbody->oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
+
+ /* This would be very bad - accidentally truncating a file when
+ * changing the time or similar - bug 12203. */
+ if (body->oa.o_valid & OBD_MD_FLSIZE &&
+ body->oa.o_size != OBD_OBJECT_EOF) {
+ static char mdsinum[48];
+
+ if (body->oa.o_valid & OBD_MD_FLFID)
+ snprintf(mdsinum, sizeof(mdsinum) - 1,
+ "of parent "DFID, body->oa.o_parent_seq,
+ body->oa.o_parent_oid, 0);
+ else
+ mdsinum[0] = '\0';
+
+ CERROR("%s: setattr from %s is trying to truncate object "DFID
+ " %s\n", ofd_name(ofd), obd_export_nid2str(tsi->tsi_exp),
+ PFID(&tsi->tsi_fid), mdsinum);
+ RETURN(-EPERM);
+ }
+
+ fo = ofd_object_find_exists(tsi->tsi_env, ofd, &tsi->tsi_fid);
+ if (IS_ERR(fo))
+ GOTO(out, rc = PTR_ERR(fo));
+
+ la_from_obdo(&fti->fti_attr, &body->oa, body->oa.o_valid);
+ fti->fti_attr.la_valid &= ~LA_TYPE;
+
+ if (body->oa.o_valid & OBD_MD_FLFID) {
+ ff = &fti->fti_mds_fid;
+ ofd_prepare_fidea(ff, &body->oa);
+ }
+
+ /* setting objects attributes (including owner/group) */
+ rc = ofd_attr_set(tsi->tsi_env, fo, &fti->fti_attr, ff);
+ if (rc != 0)
+ GOTO(out_put, rc);
+
+ 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);
+ EXIT;
+out_put:
+ ofd_object_put(tsi->tsi_env, fo);
+out:
+ if (rc == 0) {
+ /* we do not call this before to avoid lu_object_find() in
+ * ->lvbo_update() holding another reference on the object.
+ * otherwise concurrent destroy can make the object unavailable
+ * for 2nd lu_object_find() waiting for the first reference
+ * to go... deadlock! */
+ res = ldlm_resource_get(ofd->ofd_namespace, NULL,
+ &tsi->tsi_resid, LDLM_EXTENT, 0);
+ if (!IS_ERR(res)) {
+ ldlm_res_lvbo_update(res, NULL, 0);
+ ldlm_resource_putref(res);
+ }
+ }
+ return rc;
+}
+
+/**
+ * Destroy OST orphans.
+ *
+ * This is part of OST_CREATE RPC handling. If there is flag OBD_FL_DELORPHAN
+ * set then we must destroy possible orphaned objects.
+ *
+ * \param[in] env execution environment
+ * \param[in] exp OBD export
+ * \param[in] ofd OFD device
+ * \param[in] oa obdo structure for reply
+ *
+ * \retval 0 if successful
+ * \retval negative value on error
+ */
+static int ofd_orphans_destroy(const struct lu_env *env,
+ struct obd_export *exp,
+ struct ofd_device *ofd, struct obdo *oa)
+{
+ struct ofd_thread_info *info = ofd_info(env);
+ struct lu_fid *fid = &info->fti_fid;
+ struct ost_id *oi = &oa->o_oi;
+ struct ofd_seq *oseq;
+ u64 seq = ostid_seq(oi);
+ u64 end_id = ostid_id(oi);
+ u64 last;
+ u64 oid;
+ int skip_orphan;
+ int rc = 0;
+
+ ENTRY;
+
+ oseq = ofd_seq_get(ofd, seq);
+ if (oseq == NULL) {
+ CERROR("%s: Can not find seq for "DOSTID"\n",
+ ofd_name(ofd), POSTID(oi));
+ RETURN(-EINVAL);
+ }
+
+ *fid = oi->oi_fid;
+ last = ofd_seq_last_oid(oseq);
+ oid = last;
+
+ LASSERT(exp != NULL);
+ skip_orphan = !!(exp_connect_flags(exp) & OBD_CONNECT_SKIP_ORPHAN);
+
+ if (OBD_FAIL_CHECK(OBD_FAIL_OST_NODESTROY))
+ goto done;
+
+ LCONSOLE(D_INFO, "%s: deleting orphan objects from "DOSTID
+ " to "DOSTID"\n", ofd_name(ofd), seq, end_id + 1, seq, last);
+
+ while (oid > end_id) {
+ rc = fid_set_id(fid, oid);
+ if (unlikely(rc != 0))
+ GOTO(out_put, rc);
+
+ rc = ofd_destroy_by_fid(env, ofd, fid, 1);
+ if (rc != 0 && rc != -ENOENT && rc != -ESTALE &&
+ likely(rc != -EREMCHG && rc != -EINPROGRESS))
+ /* this is pretty fatal... */
+ CEMERG("%s: error destroying precreated id "
+ DFID": rc = %d\n",
+ ofd_name(ofd), PFID(fid), rc);
+
+ oid--;
+ if (!skip_orphan) {
+ ofd_seq_last_oid_set(oseq, oid);
+ /* update last_id on disk periodically so that if we
+ * restart * we don't need to re-scan all of the just
+ * deleted objects. */
+ if ((oid & 511) == 0)
+ ofd_seq_last_oid_write(env, ofd, oseq);
+ }
+ }
+
+ CDEBUG(D_HA, "%s: after destroy: set last_id to "DOSTID"\n",
+ ofd_name(ofd), seq, oid);
+
+done:
+ if (!skip_orphan) {
+ ofd_seq_last_oid_set(oseq, oid);
+ rc = ofd_seq_last_oid_write(env, ofd, oseq);
+ } else {
+ /* don't reuse orphan object, return last used objid */
+ ostid_set_id(oi, last);
+ rc = 0;
+ }
+
+ GOTO(out_put, rc);
+
+out_put:
+ ofd_seq_put(env, oseq);
+ return rc;
+}
+
+/**
+ * OFD request handler for OST_CREATE RPC.
+ *
+ * This is OFD-specific part of request handling. Its main purpose is to
+ * create new data objects on OST, but it also used to destroy orphans.
+ *
+ * \param[in] tsi target session environment for this request
+ *
+ * \retval 0 if successful
+ * \retval negative value on error
+ */
+static int ofd_create_hdl(struct tgt_session_info *tsi)
+{
+ struct ptlrpc_request *req = tgt_ses_req(tsi);
+ struct ost_body *repbody;
+ const struct obdo *oa = &tsi->tsi_ost_body->oa;
+ struct obdo *rep_oa;
+ struct obd_export *exp = tsi->tsi_exp;
+ struct ofd_device *ofd = ofd_exp(exp);
+ u64 seq = ostid_seq(&oa->o_oi);
+ u64 oid = ostid_id(&oa->o_oi);
+ struct ofd_seq *oseq;
+ int rc = 0, diff;
+ int sync_trans = 0;
+
+ ENTRY;
+
+ if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
+ RETURN(-EROFS);
+
+ repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
+ if (repbody == NULL)
+ RETURN(-ENOMEM);
+
+ down_read(&ofd->ofd_lastid_rwsem);
+ /* Currently, for safe, we do not distinguish which LAST_ID is broken,
+ * we may do that in the future.
+ * Return -ENOSPC until the LAST_ID rebuilt. */
+ if (unlikely(ofd->ofd_lastid_rebuilding))
+ GOTO(out_sem, rc = -ENOSPC);
+
+ rep_oa = &repbody->oa;
+ rep_oa->o_oi = oa->o_oi;
+
+ LASSERT(seq >= FID_SEQ_OST_MDT0);
+ LASSERT(oa->o_valid & OBD_MD_FLGROUP);
+
+ CDEBUG(D_INFO, "ofd_create("DOSTID")\n", POSTID(&oa->o_oi));
+
+ oseq = ofd_seq_load(tsi->tsi_env, ofd, seq);
+ if (IS_ERR(oseq)) {
+ CERROR("%s: Can't find FID Sequence "LPX64": rc = %ld\n",
+ ofd_name(ofd), seq, PTR_ERR(oseq));
+ GOTO(out_sem, rc = -EINVAL);
+ }
+
+ if ((oa->o_valid & OBD_MD_FLFLAGS) &&
+ (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
+ "\n", ofd_name(ofd), POSTID(&oa->o_oi),
+ ofd_seq_last_oid(oseq));
+ GOTO(out_nolock, rc = -EINVAL);
+ }
+ /* Do nothing here, we re-create objects during recovery
+ * upon write replay, see ofd_preprw_write() */
+ GOTO(out_nolock, rc = 0);
+ }
+ /* former ofd_handle_precreate */
+ if ((oa->o_valid & OBD_MD_FLFLAGS) &&
+ (oa->o_flags & OBD_FL_DELORPHAN)) {
+ exp->exp_filter_data.fed_lastid_gen = ofd->ofd_lastid_gen;
+
+ /* destroy orphans */
+ if (lustre_msg_get_conn_cnt(tgt_ses_req(tsi)->rq_reqmsg) <
+ exp->exp_conn_cnt) {
+ CERROR("%s: dropping old orphan cleanup request\n",
+ ofd_name(ofd));
+ GOTO(out_nolock, rc = 0);
+ }
+ /* This causes inflight precreates to abort and drop lock */
+ 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"
+ " 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",
+ ofd_seq_last_oid(oseq), diff);
+ if (-diff > OST_MAX_PRECREATE) {
+ /* Let MDS know that we are so far ahead. */
+ ostid_set_id(&rep_oa->o_oi, ofd_seq_last_oid(oseq) + 1);
+ rc = 0;
+ } else if (diff < 0) {
+ rc = ofd_orphans_destroy(tsi->tsi_env, exp,
+ ofd, rep_oa);
+ oseq->os_destroys_in_progress = 0;
+ } else {
+ /* XXX: Used by MDS for the first time! */
+ oseq->os_destroys_in_progress = 0;
+ }
+ } else {
+ if (unlikely(exp->exp_filter_data.fed_lastid_gen !=
+ ofd->ofd_lastid_gen)) {
+ /* Keep the export ref so we can send the reply. */
+ ofd_obd_disconnect(class_export_get(exp));
+ GOTO(out_nolock, rc = -ENOTCONN);
+ }
+
+ mutex_lock(&oseq->os_create_lock);
+ if (lustre_msg_get_conn_cnt(tgt_ses_req(tsi)->rq_reqmsg) <
+ exp->exp_conn_cnt) {
+ CERROR("%s: dropping old precreate request\n",
+ ofd_name(ofd));
+ GOTO(out, rc = 0);
+ }
+ /* only precreate if seq is 0, IDIF or normal and also o_id
+ * must be specfied */
+ if ((!fid_seq_is_mdt(seq) && !fid_seq_is_norm(seq) &&
+ !fid_seq_is_idif(seq)) || oid == 0) {
+ diff = 1; /* shouldn't we create this right now? */
+ } else {
+ diff = oid - ofd_seq_last_oid(oseq);
+ /* Do sync create if the seq is about to used up */
+ if (fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq)) {
+ if (unlikely(oid >= IDIF_MAX_OID - 1))
+ sync_trans = 1;
+ } else if (fid_seq_is_norm(seq)) {
+ if (unlikely(oid >=
+ LUSTRE_DATA_SEQ_MAX_WIDTH - 1))
+ sync_trans = 1;
+ } else {
+ CERROR("%s : invalid o_seq "DOSTID"\n",
+ ofd_name(ofd), POSTID(&oa->o_oi));
+ GOTO(out, rc = -EINVAL);
+ }
+
+ if (diff < 0) {
+ /* LU-5648 */
+ CERROR("%s: invalid precreate request for "
+ DOSTID", last_id " LPU64 ". "
+ "Likely MDS last_id corruption\n",
+ ofd_name(ofd), POSTID(&oa->o_oi),
+ ofd_seq_last_oid(oseq));
+ GOTO(out, rc = -EINVAL);
+ }
+ }
+ }
+ if (diff > 0) {
+ cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT);
+ u64 next_id;
+ int created = 0;
+ int count;
+
+ if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
+ !(oa->o_flags & OBD_FL_DELORPHAN)) {
+ /* don't enforce grant during orphan recovery */
+ rc = ofd_grant_create(tsi->tsi_env,
+ ofd_obd(ofd)->obd_self_export,
+ &diff);
+ if (rc) {
+ CDEBUG(D_HA, "%s: failed to acquire grant "
+ "space for precreate (%d): rc = %d\n",
+ ofd_name(ofd), diff, rc);
+ diff = 0;
+ }
+ }
+
+ /* This can happen if a new OST is formatted and installed
+ * in place of an old one at the same index. Instead of
+ * precreating potentially millions of deleted old objects
+ * (possibly filling the OST), only precreate the last batch.
+ * LFSCK will eventually clean up any orphans. LU-14 */
+ if (diff > 5 * OST_MAX_PRECREATE) {
+ diff = OST_MAX_PRECREATE / 2;
+ LCONSOLE_WARN("%s: precreate FID "DOSTID" is over %u "
+ "larger than the LAST_ID "DOSTID", only "
+ "precreating the last %u objects.\n",
+ ofd_name(ofd), POSTID(&oa->o_oi),
+ 5 * OST_MAX_PRECREATE,
+ POSTID(&oseq->os_oi), diff);
+ ofd_seq_last_oid_set(oseq, ostid_id(&oa->o_oi) - diff);
+ }
+
+ while (diff > 0) {
+ 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),
+ count, seq, next_id);
+
+ if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
+ && cfs_time_after(jiffies, enough_time)) {
+ CDEBUG(D_HA, "%s: Slow creates, %d/%d objects"
+ " created at a rate of %d/s\n",
+ ofd_name(ofd), created, diff + created,
+ created / DISK_TIMEOUT);
+ break;
+ }
+
+ rc = ofd_precreate_objects(tsi->tsi_env, ofd, next_id,
+ oseq, count, sync_trans);
+ if (rc > 0) {
+ created += rc;
+ diff -= rc;
+ } else if (rc < 0) {
+ break;
+ }
+ }
+
+ if (diff > 0 &&
+ lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
+ LCONSOLE_WARN("%s: can't create the same count of"
+ " objects when replaying the request"
+ " (diff is %d). see LU-4621\n",
+ ofd_name(ofd), diff);
+
+ if (created > 0)
+ /* some objects got created, we can return
+ * them, even if last creation failed */
+ rc = 0;
+ else
+ CERROR("%s: unable to precreate: rc = %d\n",
+ ofd_name(ofd), rc);
+
+ if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
+ !(oa->o_flags & OBD_FL_DELORPHAN))
+ ofd_grant_commit(tsi->tsi_env,
+ ofd_obd(ofd)->obd_self_export, rc);
+
+ ostid_set_id(&rep_oa->o_oi, ofd_seq_last_oid(oseq));
+ }
+ EXIT;
+ ofd_counter_incr(exp, LPROC_OFD_STATS_CREATE,
+ tsi->tsi_jobid, 1);
+out:
+ mutex_unlock(&oseq->os_create_lock);
+out_nolock:
+ if (rc == 0) {
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 8, 53, 0)
+ struct ofd_thread_info *info = ofd_info(tsi->tsi_env);
+ struct lu_fid *fid = &info->fti_fid;
+
+ /* For compatible purpose, it needs to convert back to
+ * OST ID before put it on wire. */
+ *fid = rep_oa->o_oi.oi_fid;
+ fid_to_ostid(fid, &rep_oa->o_oi);
+#endif
+ rep_oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
+ }
+ ofd_seq_put(tsi->tsi_env, oseq);
+
+out_sem:
+ up_read(&ofd->ofd_lastid_rwsem);
+ return rc;
+}
+
+/**
+ * OFD request handler for OST_DESTROY RPC.
+ *
+ * This is OFD-specific part of request handling. It destroys data objects
+ * related to destroyed object on MDT.
+ *
+ * \param[in] tsi target session environment for this request
+ *
+ * \retval 0 if successful
+ * \retval negative value on error
+ */
+static int ofd_destroy_hdl(struct tgt_session_info *tsi)
+{
+ const struct ost_body *body = tsi->tsi_ost_body;
+ struct ost_body *repbody;
+ struct ofd_device *ofd = ofd_exp(tsi->tsi_exp);
+ struct ofd_thread_info *fti = tsi2ofd_info(tsi);
+ struct lu_fid *fid = &fti->fti_fid;
+ u64 oid;
+ u32 count;
+ int rc = 0;
+
+ ENTRY;
+
+ if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
+ RETURN(-EROFS);
+
+ /* This is old case for clients before Lustre 2.4 */
+ /* If there's a DLM request, cancel the locks mentioned in it */
+ if (req_capsule_field_present(tsi->tsi_pill, &RMF_DLM_REQ,
+ RCL_CLIENT)) {
+ struct ldlm_request *dlm;
+
+ dlm = req_capsule_client_get(tsi->tsi_pill, &RMF_DLM_REQ);
+ if (dlm == NULL)
+ RETURN(-EFAULT);
+ ldlm_request_cancel(tgt_ses_req(tsi), dlm, 0, LATF_SKIP);
+ }
+
+ *fid = body->oa.o_oi.oi_fid;
+ oid = ostid_id(&body->oa.o_oi);
+ LASSERT(oid != 0);
+
+ repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
+
+ /* check that o_misc makes sense */
+ if (body->oa.o_valid & OBD_MD_FLOBJCOUNT)
+ count = body->oa.o_misc;
+ else
+ count = 1; /* default case - single destroy */
+
+ CDEBUG(D_HA, "%s: Destroy object "DOSTID" count %d\n", ofd_name(ofd),
+ POSTID(&body->oa.o_oi), count);
+
+ while (count > 0) {
+ int lrc;
+
+ lrc = ofd_destroy_by_fid(tsi->tsi_env, ofd, fid, 0);
+ if (lrc == -ENOENT) {
+ CDEBUG(D_INODE,
+ "%s: destroying non-existent object "DFID"\n",
+ ofd_name(ofd), PFID(fid));
+ /* rewrite rc with -ENOENT only if it is 0 */
+ if (rc == 0)
+ rc = lrc;
+ } else if (lrc != 0) {
+ CERROR("%s: error destroying object "DFID": %d\n",
+ ofd_name(ofd), PFID(fid), lrc);
+ rc = lrc;
+ }
+
+ count--;
+ oid++;
+ lrc = fid_set_id(fid, oid);
+ if (unlikely(lrc != 0 && count > 0))
+ GOTO(out, rc = lrc);
+ }
+
+ ofd_counter_incr(tsi->tsi_exp, LPROC_OFD_STATS_DESTROY,
+ tsi->tsi_jobid, 1);
+
+ GOTO(out, rc);
+
+out:
+ fid_to_ostid(fid, &repbody->oa.o_oi);
+ return rc;
+}
+
+/**
+ * OFD request handler for OST_STATFS RPC.
+ *
+ * This function gets statfs data from storage as part of request
+ * processing.
+ *
+ * \param[in] tsi target session environment for this request
+ *
+ * \retval 0 if successful
+ * \retval negative value on error
+ */
+static int ofd_statfs_hdl(struct tgt_session_info *tsi)
+{
+ struct obd_statfs *osfs;
+ int rc;
+
+ ENTRY;
+
+ osfs = req_capsule_server_get(tsi->tsi_pill, &RMF_OBD_STATFS);
+
+ rc = ofd_statfs(tsi->tsi_env, tsi->tsi_exp, osfs,
+ cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS), 0);
+ if (rc != 0)
+ CERROR("%s: statfs failed: rc = %d\n",
+ tgt_name(tsi->tsi_tgt), rc);
+
+ if (OBD_FAIL_CHECK(OBD_FAIL_OST_STATFS_EINPROGRESS))
+ rc = -EINPROGRESS;
+
+ ofd_counter_incr(tsi->tsi_exp, LPROC_OFD_STATS_STATFS,
+ tsi->tsi_jobid, 1);
+
+ RETURN(rc);
+}
+
+/**
+ * OFD request handler for OST_SYNC RPC.
+ *
+ * Sync object data or all filesystem data to the disk and pack the
+ * result in reply.
+ *
+ * \param[in] tsi target session environment for this request
+ *
+ * \retval 0 if successful
+ * \retval negative value on error
+ */
+static int ofd_sync_hdl(struct tgt_session_info *tsi)
+{
+ struct ost_body *body = tsi->tsi_ost_body;
+ struct ost_body *repbody;
+ struct ofd_thread_info *fti = tsi2ofd_info(tsi);
+ struct ofd_device *ofd = ofd_exp(tsi->tsi_exp);
+ struct ofd_object *fo = NULL;
+ int rc = 0;
+
+ ENTRY;
+
+ repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
+
+ /* if no objid is specified, it means "sync whole filesystem" */
+ if (!fid_is_zero(&tsi->tsi_fid)) {
+ fo = ofd_object_find_exists(tsi->tsi_env, ofd, &tsi->tsi_fid);
+ if (IS_ERR(fo))
+ RETURN(PTR_ERR(fo));
+ }
+
+ rc = tgt_sync(tsi->tsi_env, tsi->tsi_tgt,
+ fo != NULL ? ofd_object_child(fo) : NULL,
+ repbody->oa.o_size, repbody->oa.o_blocks);
+ if (rc)
+ GOTO(put, rc);
+
+ ofd_counter_incr(tsi->tsi_exp, LPROC_OFD_STATS_SYNC,
+ tsi->tsi_jobid, 1);
+ if (fo == NULL)
+ RETURN(0);
+
+ repbody->oa.o_oi = body->oa.o_oi;
+ repbody->oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
+
+ rc = ofd_attr_get(tsi->tsi_env, fo, &fti->fti_attr);
+ if (rc == 0)
+ obdo_from_la(&repbody->oa, &fti->fti_attr,
+ OFD_VALID_FLAGS);
+ else
+ /* don't return rc from getattr */
+ rc = 0;
+ EXIT;
+put:
+ if (fo != NULL)
+ ofd_object_put(tsi->tsi_env, fo);
+ return rc;
+}
+
+/**
+ * OFD request handler for OST_PUNCH RPC.
+ *
+ * This is part of request processing. Validate request fields,
+ * punch (truncate) the given OFD object and pack reply.
+ *
+ * \param[in] tsi target session environment for this request
+ *
+ * \retval 0 if successful
+ * \retval negative value on error
+ */
+static int ofd_punch_hdl(struct tgt_session_info *tsi)
+{
+ const struct obdo *oa = &tsi->tsi_ost_body->oa;
+ struct ost_body *repbody;
+ struct ofd_thread_info *info = tsi2ofd_info(tsi);
+ struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace;
+ struct ldlm_resource *res;
+ struct ofd_object *fo;
+ struct filter_fid *ff = NULL;
+ __u64 flags = 0;
+ struct lustre_handle lh = { 0, };
+ int rc;
+ __u64 start, end;
+ bool srvlock;
+
+ ENTRY;
+
+ /* check that we do support OBD_CONNECT_TRUNCLOCK. */
+ CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
+
+ if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
+ (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
+ RETURN(err_serious(-EPROTO));
+
+ repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
+ if (repbody == NULL)
+ RETURN(err_serious(-ENOMEM));
+
+ /* punch start,end are passed in o_size,o_blocks throught wire */
+ start = oa->o_size;
+ end = oa->o_blocks;
+
+ if (end != OBD_OBJECT_EOF) /* Only truncate is supported */
+ RETURN(-EPROTO);
+
+ /* standard truncate optimization: if file body is completely
+ * destroyed, don't send data back to the server. */
+ if (start == 0)
+ flags |= LDLM_FL_AST_DISCARD_DATA;
+
+ repbody->oa.o_oi = oa->o_oi;
+ repbody->oa.o_valid = OBD_MD_FLID;
+
+ srvlock = oa->o_valid & OBD_MD_FLFLAGS &&
+ oa->o_flags & OBD_FL_SRVLOCK;
+
+ if (srvlock) {
+ rc = tgt_extent_lock(ns, &tsi->tsi_resid, start, end, &lh,
+ LCK_PW, &flags);
+ if (rc != 0)
+ RETURN(rc);
+ }
+
+ CDEBUG(D_INODE, "calling punch for object "DFID", valid = "LPX64
+ ", start = "LPD64", end = "LPD64"\n", PFID(&tsi->tsi_fid),
+ oa->o_valid, start, end);
+
+ fo = ofd_object_find_exists(tsi->tsi_env, ofd_exp(tsi->tsi_exp),
+ &tsi->tsi_fid);
+ if (IS_ERR(fo))
+ GOTO(out, rc = PTR_ERR(fo));
+
+ la_from_obdo(&info->fti_attr, oa,
+ OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
+ info->fti_attr.la_size = start;
+ info->fti_attr.la_valid |= LA_SIZE;
+
+ if (oa->o_valid & OBD_MD_FLFID) {
+ ff = &info->fti_mds_fid;
+ ofd_prepare_fidea(ff, oa);
+ }
+
+ rc = ofd_object_punch(tsi->tsi_env, fo, start, end, &info->fti_attr,
+ ff, (struct obdo *)oa);
+ if (rc)
+ GOTO(out_put, rc);
+
+ ofd_counter_incr(tsi->tsi_exp, LPROC_OFD_STATS_PUNCH,
+ tsi->tsi_jobid, 1);
+ EXIT;
+out_put:
+ ofd_object_put(tsi->tsi_env, fo);
+out:
+ if (srvlock)
+ tgt_extent_unlock(&lh, LCK_PW);
+ if (rc == 0) {
+ /* we do not call this before to avoid lu_object_find() in
+ * ->lvbo_update() holding another reference on the object.
+ * otherwise concurrent destroy can make the object unavailable
+ * for 2nd lu_object_find() waiting for the first reference
+ * to go... deadlock! */
+ res = ldlm_resource_get(ns, NULL, &tsi->tsi_resid,
+ LDLM_EXTENT, 0);
+ if (!IS_ERR(res)) {
+ ldlm_res_lvbo_update(res, NULL, 0);
+ ldlm_resource_putref(res);
+ }
+ }
+ return rc;
+}
+
+/**
+ * OFD request handler for OST_QUOTACTL RPC.
+ *
+ * This is part of request processing to validate incoming request fields,
+ * get the requested data from OSD and pack reply.
+ *
+ * \param[in] tsi target session environment for this request
+ *
+ * \retval 0 if successful
+ * \retval negative value on error
+ */
+static int ofd_quotactl(struct tgt_session_info *tsi)
+{
+ struct obd_quotactl *oqctl, *repoqc;
+ struct lu_nodemap *nodemap =
+ tsi->tsi_exp->exp_target_data.ted_nodemap;
+ int id;
+ int rc;
+
+ ENTRY;
+
+ oqctl = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_QUOTACTL);
+ if (oqctl == NULL)
+ RETURN(err_serious(-EPROTO));
+
+ repoqc = req_capsule_server_get(tsi->tsi_pill, &RMF_OBD_QUOTACTL);
+ if (repoqc == NULL)
+ RETURN(err_serious(-ENOMEM));
+
+ *repoqc = *oqctl;
+
+ id = repoqc->qc_id;
+ if (oqctl->qc_type == USRQUOTA)
+ id = nodemap_map_id(nodemap, NODEMAP_UID,
+ NODEMAP_CLIENT_TO_FS,
+ repoqc->qc_id);
+ else if (oqctl->qc_type == GRPQUOTA)
+ id = nodemap_map_id(nodemap, NODEMAP_GID,
+ NODEMAP_CLIENT_TO_FS,
+ repoqc->qc_id);
+
+ if (repoqc->qc_id != id)
+ swap(repoqc->qc_id, id);
+
+ rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, repoqc);
+
+ ofd_counter_incr(tsi->tsi_exp, LPROC_OFD_STATS_QUOTACTL,
+ tsi->tsi_jobid, 1);
+
+ if (repoqc->qc_id != id)
+ swap(repoqc->qc_id, id);
+
+ RETURN(rc);
+}
+
+/**
+ * Calculate the amount of time for lock prolongation.
+ *
+ * This is helper for ofd_prolong_extent_locks() function to get
+ * the timeout extra time.
+ *
+ * \param[in] req current request
+ *
+ * \retval amount of time to extend the timeout with
+ */
+static inline int prolong_timeout(struct ptlrpc_request *req,
+ struct ldlm_lock *lock)
+{
+ struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
+
+ 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;
+}
+
+/**
+ * Prolong lock timeout for the given extent.
+ *
+ * This function finds all locks related with incoming request and
+ * prolongs their timeout.
+ *
+ * If a client is holding a lock for a long time while it sends
+ * read or write RPCs to the OST for the object under this lock,
+ * then we don't want the OST to evict the client. Otherwise,
+ * if the network or disk is very busy then the client may not
+ * be able to make any progress to clear out dirty pages under
+ * the lock and the application will fail.
+ *
+ * Every time a Bulk Read/Write (BRW) request arrives for the object
+ * covered by the lock, extend the timeout on that lock. The RPC should
+ * contain a lock handle for the lock it is using, but this
+ * isn't handled correctly by all client versions, and the
+ * 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
+ *
+ * \retval number of prolonged locks
+ */
+static int ofd_prolong_extent_locks(struct tgt_session_info *tsi,
+ __u64 start, __u64 end)
+{
+ 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;
+
+ if (oa->o_valid & OBD_MD_FLHANDLE) {
+ /* mostly a request should be covered by only one lock, try
+ * fast path. */
+ lock = ldlm_handle2lock(&oa->o_handle);
+ 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)) {
+ /* bingo */
+ LASSERT(lock->l_export == exp);
+ lock_count = ofd_prolong_one_lock(tsi, lock,
+ &extent);
+ LDLM_LOCK_PUT(lock);
+ RETURN(lock_count);
+ }
+ 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);
+
+ 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);
+}
+
+/**
+ * Implementation of ptlrpc_hpreq_ops::hpreq_lock_match for OFD RW requests.
+ *
+ * Determine if \a lock and the lock from request \a req are equivalent
+ * by comparing their resource names, modes, and extents.
+ *
+ * It is used to give priority to read and write RPCs being done
+ * under this lock so that the client can drop the contended
+ * lock more quickly and let other clients use it. This improves
+ * overall performance in the case where the first client gets a
+ * very large lock extent that prevents other clients from
+ * submitting their writes.
+ *
+ * \param[in] req ptlrpc_request being processed
+ * \param[in] lock contended lock to match
+ *
+ * \retval 1 if lock is matched
+ * \retval 0 otherwise
+ */
+static int ofd_rw_hpreq_lock_match(struct ptlrpc_request *req,
+ struct ldlm_lock *lock)
+{
+ struct niobuf_remote *rnb;
+ struct obd_ioobj *ioo;
+ ldlm_mode_t mode;
+ struct ldlm_extent ext;
+ __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
+
+ ENTRY;
+
+ ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
+ LASSERT(ioo != NULL);
+
+ rnb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
+ LASSERT(rnb != NULL);
+
+ ext.start = rnb->rnb_offset;
+ rnb += ioo->ioo_bufcnt - 1;
+ ext.end = rnb->rnb_offset + rnb->rnb_len - 1;
+
+ LASSERT(lock->l_resource != NULL);
+ 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;
+ if (opc == OST_READ)
+ /* whereas a bulk read can be protected by either a PR or PW
+ * extent lock */
+ mode |= LCK_PR;
+
+ if (!(lock->l_granted_mode & mode))
+ RETURN(0);
+
+ RETURN(ldlm_extent_overlap(&lock->l_policy_data.l_extent, &ext));
+}
+
+/**
+ * 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.
+ *
+ * \param[in] req the incoming request
+ *
+ * \retval 1 if \a req is blocking an LDLM lock cancel
+ * \retval 0 if it is not
+ */
+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;
+
+ ENTRY;
+
+ /* Don't use tgt_ses_info() to get session info, because lock_match()
+ * can be called while request has no processing thread yet. */
+ tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
+
+ /*
+ * Use LASSERT below because malformed RPCs should have
+ * been filtered out in tgt_hpreq_handler().
+ */
+ ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
+ LASSERT(ioo != NULL);
+
+ rnb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
+ LASSERT(rnb != NULL);
+ LASSERT(!(rnb->rnb_flags & OBD_BRW_SRVLOCK));
+
+ start = rnb->rnb_offset;
+ rnb += ioo->ioo_bufcnt - 1;
+ 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);
+
+ lock_count = ofd_prolong_extent_locks(tsi, start, end);
+
+ CDEBUG(D_DLMTRACE, "%s: refreshed %u locks timeout for req %p.\n",
+ tgt_name(tsi->tsi_tgt), lock_count, req);
+
+ RETURN(lock_count > 0);
+}
+
+/**
+ * Implementation of ptlrpc_hpreq_ops::hpreq_lock_fini for OFD RW requests.
+ *
+ * Called after the request has been handled. It refreshes lock timeout again
+ * so that client has more time to send lock cancel RPC.
+ *
+ * \param[in] req request which is being processed.
+ */
+static void ofd_rw_hpreq_fini(struct ptlrpc_request *req)
+{
+ ofd_rw_hpreq_check(req);
+}
+
+/**
+ * Implementation of ptlrpc_hpreq_ops::hpreq_lock_match for OST_PUNCH request.
+ *
+ * This function checks if the given lock is the same by its resname, mode
+ * and extent as one taken from the request.
+ * It is used to give priority to punch/truncate RPCs that might lead to
+ * the fastest release of that lock when a lock is contended.
+ *
+ * \param[in] req ptlrpc_request being processed
+ * \param[in] lock contended lock to match
+ *
+ * \retval 1 if lock is matched
+ * \retval 0 otherwise
+ */
+static int ofd_punch_hpreq_lock_match(struct ptlrpc_request *req,
+ struct ldlm_lock *lock)
+{
+ struct tgt_session_info *tsi;
+
+ /* Don't use tgt_ses_info() to get session info, because lock_match()
+ * can be called while request has no processing thread yet. */
+ tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
+
+ /*
+ * Use LASSERT below because malformed RPCs should have
+ * been filtered out in tgt_hpreq_handler().
+ */
+ LASSERT(tsi->tsi_ost_body != NULL);
+ if (tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLHANDLE &&
+ tsi->tsi_ost_body->oa.o_handle.cookie == lock->l_handle.h_cookie)
+ return 1;
+
+ return 0;
+}
+
+/**
+ * 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.
+ *
+ * \param[in] req the incoming request
+ *
+ * \retval 1 if \a req is blocking an LDLM lock cancel
+ * \retval 0 if it is not
+ */
+static int ofd_punch_hpreq_check(struct ptlrpc_request *req)
+{
+ struct tgt_session_info *tsi;
+ struct obdo *oa;
+ int lock_count;
+
+ ENTRY;
+
+ /* Don't use tgt_ses_info() to get session info, because lock_match()
+ * can be called while request has no processing thread yet. */
+ tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
+ LASSERT(tsi != NULL);
+ oa = &tsi->tsi_ost_body->oa;
+
+ LASSERT(!(oa->o_valid & OBD_MD_FLFLAGS &&
+ oa->o_flags & OBD_FL_SRVLOCK));
+
+ CDEBUG(D_DLMTRACE,
+ "%s: refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
+ tgt_name(tsi->tsi_tgt), tsi->tsi_resid.name[0],
+ tsi->tsi_resid.name[1], oa->o_size, oa->o_blocks);
+
+ lock_count = ofd_prolong_extent_locks(tsi, oa->o_size, oa->o_blocks);
+
+ CDEBUG(D_DLMTRACE, "%s: refreshed %u locks timeout for req %p.\n",
+ tgt_name(tsi->tsi_tgt), lock_count, req);
+
+ RETURN(lock_count > 0);
+}
+
+/**
+ * Implementation of ptlrpc_hpreq_ops::hpreq_lock_fini for OST_PUNCH request.
+ *
+ * Called after the request has been handled. It refreshes lock timeout again
+ * so that client has more time to send lock cancel RPC.
+ *
+ * \param[in] req request which is being processed.
+ */
+static void ofd_punch_hpreq_fini(struct ptlrpc_request *req)
+{
+ ofd_punch_hpreq_check(req);
+}
+
+static struct ptlrpc_hpreq_ops ofd_hpreq_rw = {
+ .hpreq_lock_match = ofd_rw_hpreq_lock_match,
+ .hpreq_check = ofd_rw_hpreq_check,
+ .hpreq_fini = ofd_rw_hpreq_fini
+};
+
+static struct ptlrpc_hpreq_ops ofd_hpreq_punch = {
+ .hpreq_lock_match = ofd_punch_hpreq_lock_match,
+ .hpreq_check = ofd_punch_hpreq_check,
+ .hpreq_fini = ofd_punch_hpreq_fini
+};
+
+/**
+ * Assign high priority operations to an IO request.
+ *
+ * Check if the incoming request is a candidate for
+ * high-priority processing. If it is, assign it a high
+ * priority operations table.
+ *
+ * \param[in] tsi target session environment for this request
+ */
+static void ofd_hp_brw(struct tgt_session_info *tsi)
+{
+ struct niobuf_remote *rnb;
+ struct obd_ioobj *ioo;
+
+ ENTRY;
+
+ ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
+ LASSERT(ioo != NULL); /* must exist after request preprocessing */
+ if (ioo->ioo_bufcnt > 0) {
+ rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
+ LASSERT(rnb != NULL); /* must exist after request preprocessing */
+
+ /* no high priority if server lock is needed */
+ if (rnb->rnb_flags & OBD_BRW_SRVLOCK)
+ return;
+ }
+ tgt_ses_req(tsi)->rq_ops = &ofd_hpreq_rw;
+}
+
+/**
+ * Assign high priority operations to an punch request.
+ *
+ * Check if the incoming request is a candidate for
+ * high-priority processing. If it is, assign it a high
+ * priority operations table.
+ *
+ * \param[in] tsi target session environment for this request
+ */
+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)
+ return;
+ tgt_ses_req(tsi)->rq_ops = &ofd_hpreq_punch;
+}
+
+#define OBD_FAIL_OST_READ_NET OBD_FAIL_OST_BRW_NET
+#define OBD_FAIL_OST_WRITE_NET OBD_FAIL_OST_BRW_NET
+#define OST_BRW_READ OST_READ
+#define OST_BRW_WRITE OST_WRITE
+
+/**
+ * Table of OFD-specific request handlers
+ *
+ * This table contains all opcodes accepted by OFD and
+ * specifies handlers for them. The tgt_request_handler()
+ * uses such table from each target to process incoming
+ * requests.
+ */
+static struct tgt_handler ofd_tgt_handlers[] = {
+TGT_RPC_HANDLER(OST_FIRST_OPC,
+ 0, OST_CONNECT, tgt_connect,
+ &RQF_CONNECT, LUSTRE_OBD_VERSION),
+TGT_RPC_HANDLER(OST_FIRST_OPC,
+ 0, OST_DISCONNECT, tgt_disconnect,
+ &RQF_OST_DISCONNECT, LUSTRE_OBD_VERSION),
+TGT_RPC_HANDLER(OST_FIRST_OPC,
+ 0, OST_SET_INFO, ofd_set_info_hdl,
+ &RQF_OBD_SET_INFO, LUSTRE_OST_VERSION),
+TGT_OST_HDL(0, OST_GET_INFO, ofd_get_info_hdl),
+TGT_OST_HDL(HABEO_CORPUS| HABEO_REFERO, OST_GETATTR, ofd_getattr_hdl),
+TGT_OST_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR,
+ OST_SETATTR, ofd_setattr_hdl),
+TGT_OST_HDL(0 | HABEO_REFERO | MUTABOR,
+ OST_CREATE, ofd_create_hdl),
+TGT_OST_HDL(0 | HABEO_REFERO | MUTABOR,
+ OST_DESTROY, ofd_destroy_hdl),
+TGT_OST_HDL(0 | HABEO_REFERO, OST_STATFS, ofd_statfs_hdl),
+TGT_OST_HDL_HP(HABEO_CORPUS| HABEO_REFERO,
+ OST_BRW_READ, tgt_brw_read,
+ ofd_hp_brw),
+/* don't set CORPUS flag for brw_write because -ENOENT may be valid case */
+TGT_OST_HDL_HP(HABEO_CORPUS| MUTABOR, OST_BRW_WRITE, tgt_brw_write,
+ ofd_hp_brw),
+TGT_OST_HDL_HP(HABEO_CORPUS| HABEO_REFERO | MUTABOR,
+ OST_PUNCH, ofd_punch_hdl,
+ ofd_hp_punch),
+TGT_OST_HDL(HABEO_CORPUS| HABEO_REFERO, OST_SYNC, ofd_sync_hdl),
+TGT_OST_HDL(0 | HABEO_REFERO, OST_QUOTACTL, ofd_quotactl),
+};
+
+static struct tgt_opc_slice ofd_common_slice[] = {
+ {
+ .tos_opc_start = OST_FIRST_OPC,
+ .tos_opc_end = OST_LAST_OPC,
+ .tos_hs = ofd_tgt_handlers
+ },
+ {
+ .tos_opc_start = OBD_FIRST_OPC,
+ .tos_opc_end = OBD_LAST_OPC,
+ .tos_hs = tgt_obd_handlers
+ },
+ {
+ .tos_opc_start = LDLM_FIRST_OPC,
+ .tos_opc_end = LDLM_LAST_OPC,
+ .tos_hs = tgt_dlm_handlers
+ },
+ {
+ .tos_opc_start = OUT_UPDATE_FIRST_OPC,
+ .tos_opc_end = OUT_UPDATE_LAST_OPC,
+ .tos_hs = tgt_out_handlers
+ },
+ {
+ .tos_opc_start = SEQ_FIRST_OPC,
+ .tos_opc_end = SEQ_LAST_OPC,
+ .tos_hs = seq_handlers
+ },
+ {
+ .tos_opc_start = LFSCK_FIRST_OPC,
+ .tos_opc_end = LFSCK_LAST_OPC,
+ .tos_hs = tgt_lfsck_handlers
+ },
+ {
+ .tos_hs = NULL
+ }
+};
+
+/* context key constructor/destructor: ofd_key_init(), ofd_key_fini() */
+LU_KEY_INIT_FINI(ofd, struct ofd_thread_info);
+
+/**
+ * Implementation of lu_context_key::lct_key_exit.
+ *
+ * Optional method called on lu_context_exit() for all allocated
+ * keys.
+ * It is used in OFD to sanitize context values which may be re-used
+ * during another request processing by the same thread.
+ *
+ * \param[in] ctx execution context
+ * \param[in] key context key
+ * \param[in] data ofd_thread_info
+ */
+static void ofd_key_exit(const struct lu_context *ctx,
+ struct lu_context_key *key, void *data)
+{
+ struct ofd_thread_info *info = data;
+
+ info->fti_env = NULL;
+ info->fti_exp = NULL;
+
+ info->fti_xid = 0;
+ info->fti_pre_version = 0;
+ info->fti_used = 0;
+
+ memset(&info->fti_attr, 0, sizeof info->fti_attr);
+}
+
+struct lu_context_key ofd_thread_key = {
+ .lct_tags = LCT_DT_THREAD,
+ .lct_init = ofd_key_init,
+ .lct_fini = ofd_key_fini,
+ .lct_exit = ofd_key_exit
+};
+
+/**
+ * Initialize OFD device according to parameters in the config log \a cfg.
+ *
+ * This is the main starting point of OFD initialization. It fills all OFD
+ * parameters with their initial values and calls other initializing functions
+ * to set up all OFD subsystems.
+ *
+ * \param[in] env execution environment
+ * \param[in] m OFD device
+ * \param[in] ldt LU device type of OFD
+ * \param[in] cfg configuration log
+ *
+ * \retval 0 if successful
+ * \retval negative value on error
+ */
+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;
+
+ ENTRY;
+
+ obd = class_name2obd(dev);
+ if (obd == NULL) {
+ CERROR("Cannot find obd with name %s\n", dev);
+ RETURN(-ENODEV);
+ }
+
+ rc = lu_env_refill((struct lu_env *)env);
+ if (rc != 0)
+ RETURN(rc);
+
+ obd->u.obt.obt_magic = OBT_MAGIC;
+
+ m->ofd_fmd_max_num = OFD_FMD_MAX_NUM_DEFAULT;
+ m->ofd_fmd_max_age = OFD_FMD_MAX_AGE_DEFAULT;
+
+ spin_lock_init(&m->ofd_flags_lock);
+ m->ofd_raid_degraded = 0;
+ m->ofd_syncjournal = 0;
+ ofd_slc_set(m);
+ m->ofd_grant_compat_disable = 0;
+ m->ofd_soft_sync_limit = OFD_SOFT_SYNC_LIMIT_DEFAULT;
+
+ /* statfs data */
+ spin_lock_init(&m->ofd_osfs_lock);
+ m->ofd_osfs_age = cfs_time_shift_64(-1000);
+ m->ofd_osfs_unstable = 0;
+ m->ofd_statfs_inflight = 0;
+ m->ofd_osfs_inflight = 0;
+
+ /* grant data */
+ spin_lock_init(&m->ofd_grant_lock);
+ m->ofd_tot_dirty = 0;
+ m->ofd_tot_granted = 0;
+ m->ofd_tot_pending = 0;
+ m->ofd_seq_count = 0;
+ init_waitqueue_head(&m->ofd_inconsistency_thread.t_ctl_waitq);
+ INIT_LIST_HEAD(&m->ofd_inconsistency_list);
+ spin_lock_init(&m->ofd_inconsistency_lock);
+
+ spin_lock_init(&m->ofd_batch_lock);
+ init_rwsem(&m->ofd_lastid_rwsem);
+
+ m->ofd_dt_dev.dd_lu_dev.ld_ops = &ofd_lu_ops;
+ m->ofd_dt_dev.dd_lu_dev.ld_obd = obd;
+ /* set this lu_device to obd, because error handling need it */
+ obd->obd_lu_dev = &m->ofd_dt_dev.dd_lu_dev;
+
+ rc = ofd_procfs_init(m);
+ if (rc) {
+ CERROR("Can't init ofd lprocfs, rc %d\n", rc);
+ RETURN(rc);
+ }
+
+ /* No connection accepted until configurations will finish */
+ spin_lock(&obd->obd_dev_lock);
+ obd->obd_no_conn = 1;
+ spin_unlock(&obd->obd_dev_lock);
+ obd->obd_replayable = 1;
+ if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
+ char *str = lustre_cfg_string(cfg, 4);
+
+ if (strchr(str, 'n')) {
+ CWARN("%s: recovery disabled\n", obd->obd_name);
+ obd->obd_replayable = 0;
+ }
+ }
+
+ info = ofd_info_init(env, NULL);
+ if (info == NULL)
+ RETURN(-EFAULT);
+
+ rc = ofd_stack_init(env, m, cfg);
+ if (rc) {
+ CERROR("Can't init device stack, rc %d\n", rc);
+ GOTO(err_fini_proc, rc);
+ }
+
+ ofd_procfs_add_brw_stats_symlink(m);
+
+ /* populate cached statfs data */
+ osfs = &ofd_info(env)->fti_u.osfs;
+ rc = ofd_statfs_internal(env, m, osfs, 0, NULL);
+ if (rc != 0) {
+ 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)) {
+ 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;
+
+ 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;
+
+ snprintf(info->fti_u.name, sizeof(info->fti_u.name), "%s-%s",
+ "filter"/*LUSTRE_OST_NAME*/, obd->obd_uuid.uuid);
+ m->ofd_namespace = ldlm_namespace_new(obd, info->fti_u.name,
+ LDLM_NAMESPACE_SERVER,
+ LDLM_NAMESPACE_GREEDY,
+ LDLM_NS_TYPE_OST);