list_del(&trd->trd_list);
ted->ted_reply_cnt--;
- if (lut != NULL)
+ if (lut != NULL && trd->trd_index != TRD_INDEX_MEMORY)
tgt_clear_reply_slot(lut, trd->trd_index);
OBD_FREE_PTR(trd);
}
struct lu_env env;
struct ptlrpc_request *req;
__u32 start_epoch;
- struct list_head client_list;
+ LIST_HEAD(client_list);
int rc;
if (tgt->lut_obd->obd_stopping)
tgt->lut_lsd.lsd_start_epoch = start_epoch;
spin_unlock(&tgt->lut_translock);
- INIT_LIST_HEAD(&client_list);
/**
* The recovery is not yet finished and final queue can still be updated
* with resend requests. Move final list to separate one for processing
* - there is no client to recover or the recovery was aborted
*/
if (!strncmp(tgt->lut_obd->obd_type->typ_name, LUSTRE_MDT_NAME, 3) &&
- (tgt->lut_obd->obd_max_recoverable_clients == 0 ||
+ (atomic_read(&tgt->lut_obd->obd_max_recoverable_clients) == 0 ||
tgt->lut_obd->obd_abort_recovery))
tgt->lut_lsd.lsd_feature_incompat &= ~OBD_INCOMPAT_MULTI_RPCS;
}
EXPORT_SYMBOL(tgt_client_del);
-int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
+static void tgt_clean_by_tag(struct obd_export *exp, __u64 xid, __u16 tag)
+{
+ struct tg_export_data *ted = &exp->exp_target_data;
+ struct lu_target *lut = class_exp2tgt(exp);
+ struct tg_reply_data *trd, *tmp;
+
+ if (tag == 0)
+ return;
+
+ list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
+ if (trd->trd_tag != tag)
+ continue;
+
+ LASSERT(ergo(tgt_is_increasing_xid_client(exp),
+ trd->trd_reply.lrd_xid <= xid));
+
+ ted->ted_release_tag++;
+ tgt_release_reply_data(lut, ted, trd);
+ }
+}
+
+static int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
struct tg_export_data *ted, struct tg_reply_data *trd,
+ struct ptlrpc_request *req,
struct thandle *th, bool update_lrd_file)
{
struct lsd_reply_data *lrd;
ted->ted_lcd->lcd_last_transno = lrd->lrd_transno;
mutex_unlock(&ted->ted_lcd_lock);
- /* find a empty slot */
- i = tgt_find_free_reply_slot(tgt);
- if (unlikely(i < 0)) {
- CERROR("%s: couldn't find a slot for reply data: "
- "rc = %d\n", tgt_name(tgt), i);
- RETURN(i);
- }
- trd->trd_index = i;
+ if (tgt != NULL) {
+ /* find a empty slot */
+ i = tgt_find_free_reply_slot(tgt);
+ if (unlikely(i < 0)) {
+ CERROR("%s: couldn't find a slot for reply data: "
+ "rc = %d\n", tgt_name(tgt), i);
+ RETURN(i);
+ }
+ trd->trd_index = i;
- if (update_lrd_file) {
- loff_t off;
- int rc;
+ if (update_lrd_file) {
+ loff_t off;
+ int rc;
- /* write reply data to disk */
- off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
- rc = tgt_reply_data_write(env, tgt, lrd, off, th);
- if (unlikely(rc != 0)) {
- CERROR("%s: can't update %s file: rc = %d\n",
- tgt_name(tgt), REPLY_DATA, rc);
- RETURN(rc);
+ /* write reply data to disk */
+ off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
+ rc = tgt_reply_data_write(env, tgt, lrd, off, th);
+ if (unlikely(rc != 0)) {
+ CERROR("%s: can't update %s file: rc = %d\n",
+ tgt_name(tgt), REPLY_DATA, rc);
+ RETURN(rc);
+ }
}
+ } else {
+ trd->trd_index = TRD_INDEX_MEMORY;
}
+
/* add reply data to target export's reply list */
mutex_lock(&ted->ted_lcd_lock);
+ if (req != NULL) {
+ int exclude = tgt_is_increasing_xid_client(req->rq_export) ?
+ MSG_REPLAY : MSG_REPLAY|MSG_RESENT;
+
+ if (req->rq_obsolete) {
+ mutex_unlock(&ted->ted_lcd_lock);
+ RETURN(-EALREADY);
+ }
+
+ if (!(lustre_msg_get_flags(req->rq_reqmsg) & exclude))
+ tgt_clean_by_tag(req->rq_export, req->rq_xid,
+ trd->trd_tag);
+ }
list_add(&trd->trd_list, &ted->ted_reply_list);
ted->ted_reply_cnt++;
if (ted->ted_reply_cnt > ted->ted_reply_max)
CDEBUG(D_TRACE, "add reply %p: xid %llu, transno %llu, "
"tag %hu, client gen %u, slot idx %d\n",
trd, lrd->lrd_xid, lrd->lrd_transno,
- trd->trd_tag, lrd->lrd_client_gen, i);
+ trd->trd_tag, lrd->lrd_client_gen, trd->trd_index);
+
RETURN(0);
}
-EXPORT_SYMBOL(tgt_add_reply_data);
+
+int tgt_mk_reply_data(const struct lu_env *env,
+ struct lu_target *tgt,
+ struct tg_export_data *ted,
+ struct ptlrpc_request *req,
+ __u64 opdata,
+ struct thandle *th,
+ bool write_update,
+ __u64 transno)
+{
+ struct tg_reply_data *trd;
+ struct lsd_reply_data *lrd;
+ __u64 *pre_versions = NULL;
+ int rc;
+
+ OBD_ALLOC_PTR(trd);
+ if (unlikely(trd == NULL))
+ RETURN(-ENOMEM);
+
+ /* fill reply data information */
+ lrd = &trd->trd_reply;
+ lrd->lrd_transno = transno;
+ if (req != NULL) {
+ lrd->lrd_xid = req->rq_xid;
+ trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
+ lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
+ if (write_update) {
+ pre_versions = lustre_msg_get_versions(req->rq_repmsg);
+ lrd->lrd_result = th->th_result;
+ }
+ } else {
+ struct tgt_session_info *tsi;
+
+ LASSERT(env != NULL);
+ tsi = tgt_ses_info(env);
+ LASSERT(tsi->tsi_xid != 0);
+
+ lrd->lrd_xid = tsi->tsi_xid;
+ lrd->lrd_result = tsi->tsi_result;
+ lrd->lrd_client_gen = tsi->tsi_client_gen;
+ }
+
+ lrd->lrd_data = opdata;
+ if (pre_versions) {
+ trd->trd_pre_versions[0] = pre_versions[0];
+ trd->trd_pre_versions[1] = pre_versions[1];
+ trd->trd_pre_versions[2] = pre_versions[2];
+ trd->trd_pre_versions[3] = pre_versions[3];
+ }
+
+ rc = tgt_add_reply_data(env, tgt, ted, trd, req,
+ th, write_update);
+ if (rc < 0)
+ OBD_FREE_PTR(trd);
+ return rc;
+
+}
+EXPORT_SYMBOL(tgt_mk_reply_data);
/*
* last_rcvd & last_committed update callbacks
/* Target that supports multiple reply data */
if (tgt_is_multimodrpcs_client(exp)) {
- struct tg_reply_data *trd;
- struct lsd_reply_data *lrd;
- __u64 *pre_versions;
- bool write_update;
-
- OBD_ALLOC_PTR(trd);
- if (unlikely(trd == NULL))
- RETURN(-ENOMEM);
-
- /* fill reply data information */
- lrd = &trd->trd_reply;
- lrd->lrd_transno = tti->tti_transno;
- if (req != NULL) {
- lrd->lrd_xid = req->rq_xid;
- trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
- pre_versions = lustre_msg_get_versions(req->rq_repmsg);
- lrd->lrd_result = th->th_result;
- lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
- write_update = true;
- } else {
- LASSERT(tsi->tsi_xid != 0);
- lrd->lrd_xid = tsi->tsi_xid;
- lrd->lrd_result = tsi->tsi_result;
- lrd->lrd_client_gen = tsi->tsi_client_gen;
- trd->trd_tag = 0;
- pre_versions = NULL;
- write_update = false;
- }
-
- lrd->lrd_data = opdata;
- if (pre_versions) {
- trd->trd_pre_versions[0] = pre_versions[0];
- trd->trd_pre_versions[1] = pre_versions[1];
- trd->trd_pre_versions[2] = pre_versions[2];
- trd->trd_pre_versions[3] = pre_versions[3];
- }
-
- rc = tgt_add_reply_data(env, tgt, ted, trd, th, write_update);
- if (rc < 0)
- OBD_FREE_PTR(trd);
- return rc;
+ return tgt_mk_reply_data(env, tgt, ted, req, opdata, th,
+ !!(req != NULL), tti->tti_transno);
}
/* Enough for update replay, let's return */
if (tgt->lut_bottom->dd_rdonly)
RETURN(0);
- CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
- sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
+ BUILD_BUG_ON(offsetof(struct lsd_client_data, lcd_padding) +
+ sizeof(lcd->lcd_padding) != LR_CLIENT_SIZE);
OBD_ALLOC_PTR(lcd);
if (lcd == NULL)
exp->exp_connecting = 0;
exp->exp_in_recovery = 0;
spin_unlock(&exp->exp_lock);
- obd->obd_max_recoverable_clients++;
+ atomic_inc(&obd->obd_max_recoverable_clients);
if (tgt->lut_lsd.lsd_feature_incompat &
OBD_INCOMPAT_MULTI_RPCS &&
last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
/* ensure padding in the struct is the correct size */
- CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
- sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
+ BUILD_BUG_ON(offsetof(struct lr_server_data, lsd_padding) +
+ sizeof(lsd->lsd_padding) != LR_SERVER_SIZE);
rc = server_name2index(tgt_name(tgt), &index, NULL);
if (rc < 0) {
}
if (lsd->lsd_osd_index != index) {
- LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd "
- "is different with the index %d in"
- "config log, It might be disk"
- "corruption!\n", tgt_name(tgt),
+ LCONSOLE_ERROR_MSG(0x157,
+ "%s: index %d in last rcvd is different with the index %d in config log, It might be disk corruption!\n",
+ tgt_name(tgt),
lsd->lsd_osd_index, index);
RETURN(-EINVAL);
}
unsigned long reply_data_size;
int rc;
struct lsd_reply_header *lrh = NULL;
- struct lsd_client_data *lcd = NULL;
struct tg_reply_data *trd = NULL;
int idx;
loff_t off;
if (hash == NULL)
GOTO(out, rc = -ENODEV);
- OBD_ALLOC_PTR(lcd);
- if (lcd == NULL)
- GOTO(out, rc = -ENOMEM);
-
OBD_ALLOC_PTR(trd);
if (trd == NULL)
GOTO(out, rc = -ENOMEM);
/* update export last committed transation */
exp->exp_last_committed = max(exp->exp_last_committed,
lrd->lrd_transno);
+ /* Update lcd_last_transno as well for check in
+ * tgt_release_reply_data() or the latest client
+ * transno can be lost.
+ */
+ ted->ted_lcd->lcd_last_transno =
+ max(ted->ted_lcd->lcd_last_transno,
+ exp->exp_last_committed);
mutex_unlock(&ted->ted_lcd_lock);
class_export_put(exp);
out:
if (hash != NULL)
cfs_hash_putref(hash);
- if (lcd != NULL)
- OBD_FREE_PTR(lcd);
if (trd != NULL)
OBD_FREE_PTR(trd);
if (lrh != NULL)
return rc;
}
-struct tg_reply_data *tgt_lookup_reply_by_xid(struct tg_export_data *ted,
- __u64 xid)
+static int tgt_check_lookup_req(struct ptlrpc_request *req, int lookup,
+ struct tg_reply_data *trd)
{
- struct tg_reply_data *found = NULL;
- struct tg_reply_data *reply;
+ struct tg_export_data *ted = &req->rq_export->exp_target_data;
+ struct lu_target *lut = class_exp2tgt(req->rq_export);
+ __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
+ int rc = 0;
+ struct tg_reply_data *reply;
+ bool check_increasing;
+
+ if (tag == 0)
+ return 0;
+
+ check_increasing = tgt_is_increasing_xid_client(req->rq_export) &&
+ !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY);
+ if (!lookup && !check_increasing)
+ return 0;
- mutex_lock(&ted->ted_lcd_lock);
list_for_each_entry(reply, &ted->ted_reply_list, trd_list) {
- if (reply->trd_reply.lrd_xid == xid) {
- found = reply;
+ if (lookup && reply->trd_reply.lrd_xid == req->rq_xid) {
+ rc = 1;
+ if (trd != NULL)
+ *trd = *reply;
+ break;
+ } else if (check_increasing && reply->trd_tag == tag &&
+ reply->trd_reply.lrd_xid > req->rq_xid) {
+ rc = -EPROTO;
+ CERROR("%s: busy tag=%u req_xid=%llu, trd=%p: xid=%llu transno=%llu client_gen=%u slot_idx=%d: rc = %d\n",
+ tgt_name(lut), tag, req->rq_xid, trd,
+ reply->trd_reply.lrd_xid,
+ reply->trd_reply.lrd_transno,
+ reply->trd_reply.lrd_client_gen,
+ reply->trd_index, rc);
break;
}
}
- mutex_unlock(&ted->ted_lcd_lock);
- return found;
+
+ return rc;
}
-EXPORT_SYMBOL(tgt_lookup_reply_by_xid);
/* Look for a reply data matching specified request @req
* A copy is returned in @trd if the pointer is not NULL
*/
-bool tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
+int tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
{
- struct tg_export_data *ted = &req->rq_export->exp_target_data;
- struct tg_reply_data *reply;
- bool found = false;
-
- reply = tgt_lookup_reply_by_xid(ted, req->rq_xid);
- if (reply != NULL) {
- found = true;
- if (trd != NULL)
- *trd = *reply;
+ struct tg_export_data *ted = &req->rq_export->exp_target_data;
+ int found = 0;
+ bool not_replay = !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY);
+
+ mutex_lock(&ted->ted_lcd_lock);
+ if (not_replay && req->rq_xid <= req->rq_export->exp_last_xid) {
+ /* A check for the last_xid is needed here in case there is
+ * no reply data is left in the list. It may happen if another
+ * RPC on another slot increased the last_xid between our
+ * process_req_last_xid & tgt_lookup_reply calls */
+ found = -EPROTO;
+ } else {
+ found = tgt_check_lookup_req(req, 1, trd);
}
+ mutex_unlock(&ted->ted_lcd_lock);
- CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d\n",
- tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid,
- found ? 1 : 0);
+ CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d last_xid %llu\n",
+ tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid, found,
+ req->rq_export->exp_last_xid);
return found;
}
struct lu_target *lut = class_exp2tgt(exp);
struct tg_reply_data *trd, *tmp;
- mutex_lock(&ted->ted_lcd_lock);
+
list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
if (trd->trd_reply.lrd_xid > rcvd_xid)
continue;
ted->ted_release_xid++;
tgt_release_reply_data(lut, ted, trd);
}
- mutex_unlock(&ted->ted_lcd_lock);
return 0;
}
-int tgt_handle_tag(struct obd_export *exp, __u16 tag)
+int tgt_handle_tag(struct ptlrpc_request *req)
{
- struct tg_export_data *ted = &exp->exp_target_data;
- struct lu_target *lut = class_exp2tgt(exp);
- struct tg_reply_data *trd, *tmp;
-
- if (tag == 0)
- return 0;
-
- mutex_lock(&ted->ted_lcd_lock);
- list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
- if (trd->trd_tag != tag)
- continue;
- ted->ted_release_tag++;
- tgt_release_reply_data(lut, ted, trd);
- break;
- }
- mutex_unlock(&ted->ted_lcd_lock);
-
- return 0;
+ return tgt_check_lookup_req(req, 0, NULL);
}