/*
* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2011, 2012, Whamcloud, Inc.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#include <lustre_fsfilt.h>
#include <lustre_disk.h>
#include "mgs_internal.h"
-
+#include <lustre_param.h>
/* Establish a connection to the MGS.*/
static int mgs_connect(const struct lu_env *env,
data->ocd_version = LUSTRE_VERSION_CODE;
}
- rc = mgs_export_stats_init(obd, lexp, 0, localdata);
+ rc = mgs_export_stats_init(obd, lexp, localdata);
if (rc) {
class_disconnect(lexp);
data->ocd_version = LUSTRE_VERSION_CODE;
}
- RETURN(mgs_export_stats_init(obd, exp, 1, localdata));
+ RETURN(mgs_export_stats_init(obd, exp, localdata));
}
static int mgs_disconnect(struct obd_export *exp)
LASSERT(exp);
+ mgs_fsc_cleanup(exp);
+
class_export_get(exp);
mgs_counter_incr(exp, LPROC_MGS_DISCONNECT);
rc = server_disconnect_export(exp);
-
class_export_put(exp);
RETURN(rc);
}
GOTO(err_ops, rc = -EROFS);
}
+ obd->u.obt.obt_magic = OBT_MAGIC;
+ obd->u.obt.obt_instance = 0;
+
/* namespace for mgs llog */
- obd->obd_namespace = ldlm_namespace_new(obd ,"MGS", LDLM_NAMESPACE_SERVER,
- LDLM_NAMESPACE_MODEST);
+ obd->obd_namespace = ldlm_namespace_new(obd ,"MGS",
+ LDLM_NAMESPACE_SERVER,
+ LDLM_NAMESPACE_MODEST,
+ LDLM_NS_TYPE_MGT);
if (obd->obd_namespace == NULL)
GOTO(err_ops, rc = -ENOMEM);
/* Internal mgs setup */
mgs_init_fsdb_list(obd);
cfs_sema_init(&mgs->mgs_sem, 1);
+ mgs->mgs_start_time = cfs_time_current_sec();
/* Setup proc */
lprocfs_mgs_init_vars(&lvars);
GOTO(err_llog, rc = -ENOMEM);
}
- rc = ptlrpc_start_threads(obd, mgs->mgs_service);
+ rc = ptlrpc_start_threads(mgs->mgs_service);
if (rc)
GOTO(err_thread, rc);
ping_evictor_start();
- LCONSOLE_INFO("MGS %s started\n", obd->obd_name);
+ CDEBUG(D_INFO, "MGS %s started\n", obd->obd_name);
RETURN(0);
static int mgs_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
{
+ struct mgs_obd *mgs = &obd->u.mgs;
int rc = 0;
ENTRY;
switch (stage) {
case OBD_CLEANUP_EARLY:
+ break;
case OBD_CLEANUP_EXPORTS:
+ ping_evictor_stop();
+ ptlrpc_unregister_service(mgs->mgs_service);
+ mgs_cleanup_fsdb_list(obd);
rc = obd_llog_finish(obd, 0);
+ lproc_mgs_cleanup(obd);
break;
}
RETURN(rc);
if (mgs->mgs_sb == NULL)
RETURN(0);
- ping_evictor_stop();
-
- ptlrpc_unregister_service(mgs->mgs_service);
-
- mgs_cleanup_fsdb_list(obd);
- lproc_mgs_cleanup(obd);
mgs_fs_cleanup(obd);
server_put_mount(obd->obd_name, mgs->mgs_vfsmnt);
RETURN(0);
}
-/* similar to filter_prepare_destroy */
-static int mgs_get_cfg_lock(struct obd_device *obd, char *fsname,
- struct lustre_handle *lockh)
+static int mgs_completion_ast_config(struct ldlm_lock *lock, int flags,
+ void *cbdata)
{
- struct ldlm_res_id res_id;
- int rc, flags = 0;
ENTRY;
- rc = mgc_fsname2resid(fsname, &res_id);
- if (!rc)
- rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id,
- LDLM_PLAIN, NULL, LCK_EX,
- &flags, ldlm_blocking_ast,
- ldlm_completion_ast, NULL,
- fsname, 0, NULL, lockh);
- if (rc)
- CERROR("can't take cfg lock for %s (%d)\n", fsname, rc);
+ if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
+ LDLM_FL_BLOCK_CONV))) {
+ struct fs_db *fsdb = (struct fs_db *)lock->l_ast_data;
+ struct lustre_handle lockh;
- RETURN(rc);
+ /* clear the bit before lock put */
+ cfs_clear_bit(FSDB_REVOKING_LOCK, &fsdb->fsdb_flags);
+
+ ldlm_lock2handle(lock, &lockh);
+ ldlm_lock_decref_and_cancel(&lockh, LCK_EX);
+ }
+
+ RETURN(ldlm_completion_ast(lock, flags, cbdata));
}
-static int mgs_put_cfg_lock(struct lustre_handle *lockh)
+static int mgs_completion_ast_ir(struct ldlm_lock *lock, int flags,
+ void *cbdata)
{
ENTRY;
- ldlm_lock_decref(lockh, LCK_EX);
- RETURN(0);
+
+ if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
+ LDLM_FL_BLOCK_CONV))) {
+ struct fs_db *fsdb = (struct fs_db *)lock->l_ast_data;
+ struct lustre_handle lockh;
+
+ mgs_ir_notify_complete(fsdb);
+
+ ldlm_lock2handle(lock, &lockh);
+ ldlm_lock_decref_and_cancel(&lockh, LCK_EX);
+ }
+
+ RETURN(ldlm_completion_ast(lock, flags, cbdata));
}
-void mgs_revoke_lock(struct obd_device *obd, struct fs_db *fsdb)
+void mgs_revoke_lock(struct obd_device *obd, struct fs_db *fsdb, int type)
{
- struct lustre_handle lockh;
- int lockrc;
+ ldlm_completion_callback cp = NULL;
+ struct lustre_handle lockh = { 0 };
+ struct ldlm_res_id res_id;
+ int flags = LDLM_FL_ATOMIC_CB;
+ int rc;
+ ENTRY;
LASSERT(fsdb->fsdb_name[0] != '\0');
+ rc = mgc_fsname2resid(fsdb->fsdb_name, &res_id, type);
+ LASSERT(rc == 0);
+
+ switch (type) {
+ case CONFIG_T_CONFIG:
+ cp = mgs_completion_ast_config;
+ if (cfs_test_and_set_bit(FSDB_REVOKING_LOCK, &fsdb->fsdb_flags))
+ rc = -EALREADY;
+ break;
+ case CONFIG_T_RECOVER:
+ cp = mgs_completion_ast_ir;
+ default:
+ break;
+ }
- if (cfs_test_and_set_bit(FSDB_REVOKING_LOCK, &fsdb->fsdb_flags) == 0) {
- lockrc = mgs_get_cfg_lock(obd, fsdb->fsdb_name, &lockh);
- /* clear the bit before lock put */
- cfs_clear_bit(FSDB_REVOKING_LOCK, &fsdb->fsdb_flags);
-
- if (lockrc != ELDLM_OK)
- CERROR("lock error %d for fs %s\n",
- lockrc, fsdb->fsdb_name);
- else
- mgs_put_cfg_lock(&lockh);
+ if (!rc) {
+ LASSERT(cp != NULL);
+ rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id,
+ LDLM_PLAIN, NULL, LCK_EX, &flags,
+ ldlm_blocking_ast, cp, NULL,
+ fsdb, 0, NULL, &lockh);
+ if (rc != ELDLM_OK) {
+ CERROR("can't take cfg lock for "LPX64"/"LPX64"(%d)\n",
+ le64_to_cpu(res_id.name[0]),
+ le64_to_cpu(res_id.name[1]), rc);
+
+ if (type == CONFIG_T_CONFIG)
+ cfs_clear_bit(FSDB_REVOKING_LOCK,
+ &fsdb->fsdb_flags);
+ }
+ /* lock has been cancelled in completion_ast. */
}
+
+ RETURN_EXIT;
}
/* rc=0 means ok
RETURN(rc);
}
+/* Ensure this is not a failover node that is connecting first*/
+static int mgs_check_failover_reg(struct mgs_target_info *mti)
+{
+ lnet_nid_t nid;
+ char *ptr;
+ int i;
+
+ ptr = mti->mti_params;
+ while (class_find_param(ptr, PARAM_FAILNODE, &ptr) == 0) {
+ while (class_parse_nid(ptr, &nid, &ptr) == 0) {
+ for (i = 0; i < mti->mti_nid_count; i++) {
+ if (nid == mti->mti_nids[i]) {
+ LCONSOLE_WARN("Denying initial registra"
+ "tion attempt from nid %s"
+ ", specified as failover"
+ "\n",libcfs_nid2str(nid));
+ return -EADDRNOTAVAIL;
+ }
+ }
+ }
+ }
+ return 0;
+}
+
/* Called whenever a target starts up. Flags indicate first connect, etc. */
static int mgs_handle_target_reg(struct ptlrpc_request *req)
{
struct obd_device *obd = req->rq_export->exp_obd;
struct mgs_target_info *mti, *rep_mti;
struct fs_db *fsdb;
+ int opc;
int rc = 0;
ENTRY;
mgs_counter_incr(req->rq_export, LPROC_MGS_TARGET_REG);
mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
+
+ opc = mti->mti_flags & LDD_F_OPC_MASK;
+ if (opc == LDD_F_OPC_READY) {
+ CDEBUG(D_MGS, "fs: %s index: %d is ready to reconnect.\n",
+ mti->mti_fsname, mti->mti_stripe_index);
+ rc = mgs_ir_update(obd, mti);
+ if (rc) {
+ LASSERT(!(mti->mti_flags & LDD_F_IR_CAPABLE));
+ CERROR("Update IR return with %d(ignore and IR "
+ "disabled)\n", rc);
+ }
+ GOTO(out_nolock, rc);
+ }
+
+ /* Do not support unregistering right now. */
+ if (opc != LDD_F_OPC_REG)
+ GOTO(out_nolock, rc = -EINVAL);
+
+ CDEBUG(D_MGS, "fs: %s index: %d is registered to MGS.\n",
+ mti->mti_fsname, mti->mti_stripe_index);
+
+ if (mti->mti_flags & LDD_F_NEED_INDEX)
+ mti->mti_flags |= LDD_F_WRITECONF;
+
if (!(mti->mti_flags & (LDD_F_WRITECONF | LDD_F_UPGRADE14 |
LDD_F_UPDATE))) {
/* We're just here as a startup ping. */
if (rc <= 0)
/* Nothing wrong, or fatal error */
GOTO(out_nolock, rc);
+ } else {
+ if (!(mti->mti_flags & LDD_F_NO_PRIMNODE)
+ && (rc = mgs_check_failover_reg(mti)))
+ GOTO(out_nolock, rc);
}
OBD_FAIL_TIMEOUT(OBD_FAIL_MGS_PAUSE_TARGET_REG, 10);
}
out:
- mgs_revoke_lock(obd, fsdb);
+ mgs_revoke_lock(obd, fsdb, CONFIG_T_CONFIG);
out_nolock:
CDEBUG(D_MGS, "replying with %s, index=%d, rc=%d\n", mti->mti_svname,
mti->mti_stripe_index, rc);
+ req->rq_status = rc;
+ if (rc)
+ /* we need an error flag to tell the target what's going on,
+ * instead of just doing it by error code only. */
+ mti->mti_flags |= LDD_F_ERROR;
+
rc = req_capsule_server_pack(&req->rq_pill);
if (rc)
RETURN(rc);
RETURN(rc);
}
+static int mgs_config_read(struct ptlrpc_request *req)
+{
+ struct mgs_config_body *body;
+ int rc;
+ ENTRY;
+
+ body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
+ if (body == NULL)
+ RETURN(-EINVAL);
+
+ switch (body->mcb_type) {
+ case CONFIG_T_RECOVER:
+ rc = mgs_get_ir_logs(req);
+ break;
+
+ case CONFIG_T_CONFIG:
+ rc = -ENOTSUPP;
+ break;
+
+ default:
+ rc = -EINVAL;
+ break;
+ }
+
+ RETURN(rc);
+}
+
/*
* similar as in ost_connect_check_sptlrpc()
*/
RETURN(0);
}
+/*
+ * For old clients there is no direct way of knowing which filesystems
+ * a client is operating at the MGS side. But we need to pick up those
+ * clients so that the MGS can mark the corresponding filesystem as
+ * non-IR capable because old clients are not ready to be notified.
+ *
+ * This is why we have this _hack_ function. We detect the filesystem's
+ * name by hacking llog operation which is currently used by the clients
+ * to fetch configuration logs. At present this is fine because this is
+ * the ONLY llog operation between mgc and the MGS.
+ *
+ * If extra llog operation is going to be added, this function needs fixing.
+ *
+ * If releases prior than 2.0 are not supported, we can remove this function.
+ */
+static int mgs_handle_fslog_hack(struct ptlrpc_request *req)
+{
+ char *logname;
+ char fsname[16];
+ char *ptr;
+ int rc;
+
+ /* XXX: We suppose that llog at mgs is only used for
+ * fetching file system log */
+ logname = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
+ if (logname == NULL) {
+ CERROR("No logname, is llog on MGS used for something else?\n");
+ return -EINVAL;
+ }
+
+ ptr = strchr(logname, '-');
+ rc = (int)(ptr - logname);
+ if (ptr == NULL || rc >= sizeof(fsname)) {
+ CERROR("Invalid logname received: %s\n", logname);
+ return -EINVAL;
+ }
+
+ strncpy(fsname, logname, rc);
+ fsname[rc] = 0;
+ rc = mgs_fsc_attach(req->rq_export, fsname);
+ if (rc < 0 && rc != -EEXIST)
+ CERROR("add fs client %s returns %d\n", fsname, rc);
+
+ return rc;
+}
+
/* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
int mgs_handle(struct ptlrpc_request *req)
{
ENTRY;
req_capsule_init(&req->rq_pill, req, RCL_SERVER);
- OBD_FAIL_TIMEOUT_MS(OBD_FAIL_MGS_PAUSE_REQ, obd_fail_val);
- if (OBD_FAIL_CHECK(OBD_FAIL_MGS_ALL_REQUEST_NET))
+ CFS_FAIL_TIMEOUT_MS(OBD_FAIL_MGS_PAUSE_REQ, cfs_fail_val);
+ if (CFS_FAIL_CHECK(OBD_FAIL_MGS_ALL_REQUEST_NET))
RETURN(0);
LASSERT(current->journal_info == NULL);
if (opc != MGS_CONNECT) {
if (!class_connected_export(req->rq_export)) {
- CERROR("lustre_mgs: operation %d on unconnected MGS\n",
- opc);
+ DEBUG_REQ(D_MGS, req, "operation on unconnected MGS\n");
req->rq_status = -ENOTCONN;
GOTO(out, rc = -ENOTCONN);
}
req_capsule_set(&req->rq_pill, &RQF_MGS_SET_INFO);
rc = mgs_set_info_rpc(req);
break;
-
+ case MGS_CONFIG_READ:
+ DEBUG_REQ(D_MGS, req, "read config");
+ req_capsule_set(&req->rq_pill, &RQF_MGS_CONFIG_READ);
+ rc = mgs_config_read(req);
+ break;
case LDLM_ENQUEUE:
DEBUG_REQ(D_MGS, req, "enqueue");
req_capsule_set(&req->rq_pill, &RQF_LDLM_ENQUEUE);
DEBUG_REQ(D_MGS, req, "llog_init");
req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
rc = llog_origin_handle_create(req);
+ if (rc == 0)
+ (void)mgs_handle_fslog_hack(req);
break;
case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
DEBUG_REQ(D_MGS, req, "llog next block");
static inline int mgs_init_export(struct obd_export *exp)
{
+ struct mgs_export_data *data = &exp->u.eu_mgs_data;
+
+ /* init mgs_export_data for fsc */
+ cfs_spin_lock_init(&data->med_lock);
+ CFS_INIT_LIST_HEAD(&data->med_clients);
+
cfs_spin_lock(&exp->exp_lock);
exp->exp_connecting = 1;
cfs_spin_unlock(&exp->exp_lock);
+ /* self-export doesn't need client data and ldlm initialization */
+ if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
+ &exp->exp_client_uuid)))
+ return 0;
return ldlm_init_export(exp);
}
target_destroy_export(exp);
mgs_client_free(exp);
+
+ if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
+ &exp->exp_client_uuid)))
+ RETURN(0);
+
ldlm_destroy_export(exp);
RETURN(0);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff