Whamcloud - gitweb
b=23595 return registration errors
[fs/lustre-release.git] / lustre / mgs / mgs_handler.c
index 97878c8..2796994 100644 (file)
@@ -26,7 +26,7 @@
  * GPL HEADER END
  */
 /*
- * Copyright  2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  */
 /*
@@ -42,7 +42,7 @@
 # define EXPORT_SYMTAB
 #endif
 #define DEBUG_SUBSYSTEM S_MGS
-#define D_MGS D_CONFIG/*|D_WARNING*/
+#define D_MGS D_CONFIG
 
 #ifdef __KERNEL__
 # include <linux/module.h>
@@ -59,7 +59,7 @@
 #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,
@@ -90,7 +90,7 @@ static int mgs_connect(const struct lu_env *env,
                 data->ocd_version = LUSTRE_VERSION_CODE;
         }
 
-        rc = mgs_client_add(obd, lexp, localdata);
+        rc = mgs_export_stats_init(obd, lexp, 0, localdata);
 
         if (rc) {
                 class_disconnect(lexp);
@@ -119,7 +119,7 @@ static int mgs_reconnect(const struct lu_env *env,
                 data->ocd_version = LUSTRE_VERSION_CODE;
         }
 
-        RETURN(0);
+        RETURN(mgs_export_stats_init(obd, exp, 1, localdata));
 }
 
 static int mgs_disconnect(struct obd_export *exp)
@@ -132,28 +132,7 @@ static int mgs_disconnect(struct obd_export *exp)
         class_export_get(exp);
         mgs_counter_incr(exp, LPROC_MGS_DISCONNECT);
 
-        /* Disconnect early so that clients can't keep using export */
-        rc = class_disconnect(exp);
-        ldlm_cancel_locks_for_export(exp);
-
-        lprocfs_exp_cleanup(exp);
-
-        /* complete all outstanding replies */
-        spin_lock(&exp->exp_lock);
-        while (!list_empty(&exp->exp_outstanding_replies)) {
-                struct ptlrpc_reply_state *rs =
-                        list_entry(exp->exp_outstanding_replies.next,
-                                   struct ptlrpc_reply_state, rs_exp_list);
-                struct ptlrpc_service *svc = rs->rs_service;
-
-                spin_lock(&svc->srv_lock);
-                list_del_init(&rs->rs_exp_list);
-                spin_lock(&rs->rs_lock);
-                ptlrpc_schedule_difficult_reply(rs);
-                spin_unlock(&rs->rs_lock);
-                spin_unlock(&svc->srv_lock);
-        }
-        spin_unlock(&exp->exp_lock);
+        rc = server_disconnect_export(exp);
 
         class_export_put(exp);
         RETURN(rc);
@@ -163,8 +142,7 @@ static int mgs_cleanup(struct obd_device *obd);
 static int mgs_handle(struct ptlrpc_request *req);
 
 static int mgs_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
-                         struct obd_device *tgt, int count,
-                         struct llog_catid *logid, struct obd_uuid *uuid)
+                         struct obd_device *tgt, int *index)
 {
         int rc;
         ENTRY;
@@ -235,7 +213,7 @@ static int mgs_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
                 GOTO(err_ns, rc);
         }
 
-        rc = obd_llog_init(obd, &obd->obd_olg, obd, 0, NULL, NULL);
+        rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
         if (rc)
                 GOTO(err_fs, rc);
 
@@ -244,13 +222,22 @@ static int mgs_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 
         /* Internal mgs setup */
         mgs_init_fsdb_list(obd);
-        sema_init(&mgs->mgs_sem, 1);
+        cfs_sema_init(&mgs->mgs_sem, 1);
+
+        /* Setup proc */
+        lprocfs_mgs_init_vars(&lvars);
+        if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
+                lproc_mgs_setup(obd);
+                rc = lprocfs_alloc_md_stats(obd, LPROC_MGS_LAST);
+                if (rc)
+                        GOTO(err_llog, rc);
+        }
 
         /* Start the service threads */
         mgs->mgs_service =
                 ptlrpc_init_svc(MGS_NBUFS, MGS_BUFSIZE, MGS_MAXREQSIZE,
                                 MGS_MAXREPSIZE, MGS_REQUEST_PORTAL,
-                                MGC_REPLY_PORTAL, 2000,
+                                MGC_REPLY_PORTAL, 2,
                                 mgs_handle, LUSTRE_MGS_NAME,
                                 obd->obd_proc_entry, target_print_req,
                                 MGS_THREADS_AUTO_MIN, MGS_THREADS_AUTO_MAX,
@@ -265,12 +252,6 @@ static int mgs_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
         if (rc)
                 GOTO(err_thread, rc);
 
-        /* Setup proc */
-        lprocfs_mgs_init_vars(&lvars);
-        if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
-                lproc_mgs_setup(obd);
-        }
-
         ping_evictor_start();
 
         LCONSOLE_INFO("MGS %s started\n", obd->obd_name);
@@ -280,6 +261,7 @@ static int mgs_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 err_thread:
         ptlrpc_unregister_service(mgs->mgs_service);
 err_llog:
+        lproc_mgs_cleanup(obd);
         obd_llog_finish(obd, 0);
 err_fs:
         /* No extra cleanup needed for llog_init_commit_thread() */
@@ -354,7 +336,7 @@ static int mgs_get_cfg_lock(struct obd_device *obd, char *fsname,
                                             LDLM_PLAIN, NULL, LCK_EX,
                                             &flags, ldlm_blocking_ast,
                                             ldlm_completion_ast, NULL,
-                                            fsname, 0, NULL, NULL, lockh);
+                                            fsname, 0, NULL, lockh);
         if (rc)
                 CERROR("can't take cfg lock for %s (%d)\n", fsname, rc);
 
@@ -368,18 +350,23 @@ static int mgs_put_cfg_lock(struct lustre_handle *lockh)
         RETURN(0);
 }
 
-static void mgs_revoke_lock(struct obd_device *obd, char *fsname,
-                            struct lustre_handle *lockh)
+void mgs_revoke_lock(struct obd_device *obd, struct fs_db *fsdb)
 {
-        int lockrc;
+        struct lustre_handle lockh;
+        int                  lockrc;
+
+        LASSERT(fsdb->fsdb_name[0] != '\0');
+
+        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 (fsname[0]) {
-                lockrc = mgs_get_cfg_lock(obd, fsname, lockh);
                 if (lockrc != ELDLM_OK)
-                        CERROR("lock error %d for fs %s\n", lockrc,
-                               fsname);
+                        CERROR("lock error %d for fs %s\n",
+                               lockrc, fsdb->fsdb_name);
                 else
-                        mgs_put_cfg_lock(lockh);
+                        mgs_put_cfg_lock(&lockh);
         }
 }
 
@@ -414,18 +401,46 @@ static int mgs_check_target(struct obd_device *obd, struct mgs_target_info *mti)
         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 lustre_handle lockh;
         struct mgs_target_info *mti, *rep_mti;
-        int rc = 0, lockrc;
+        struct fs_db *fsdb;
+        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);
+
+        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. */
@@ -436,28 +451,15 @@ static int mgs_handle_target_reg(struct ptlrpc_request *req)
                 if (rc <= 0)
                         /* Nothing wrong, or fatal error */
                         GOTO(out_nolock, rc);
-        }
-
-        /* Revoke the config lock to make sure nobody is reading. */
-        /* Although actually I think it should be alright if
-           someone was reading while we were updating the logs - if we
-           revoke at the end they will just update from where they left off. */
-        lockrc = mgs_get_cfg_lock(obd, mti->mti_fsname, &lockh);
-        if (lockrc != ELDLM_OK) {
-                LCONSOLE_ERROR_MSG(0x13d, "%s: Can't signal other nodes to "
-                                   "update their configuration (%d). Updating "
-                                   "local logs anyhow; you might have to "
-                                   "manually restart other nodes to get the "
-                                   "latest configuration.\n",
-                                   obd->obd_name, lockrc);
+        } else {
+                if ((rc = mgs_check_failover_reg(mti)))
+                        GOTO(out_nolock, rc);
         }
 
         OBD_FAIL_TIMEOUT(OBD_FAIL_MGS_PAUSE_TARGET_REG, 10);
 
-        /* Log writing contention is handled by the fsdb_sem */
-
         if (mti->mti_flags & LDD_F_WRITECONF) {
-                if (mti->mti_flags & LDD_F_SV_TYPE_MDT &&
+                if (mti->mti_flags & SVTYPE_MDT &&
                     mti->mti_stripe_index == 0) {
                         rc = mgs_erase_logs(obd, mti->mti_fsname);
                         LCONSOLE_WARN("%s: Logs for fs %s were removed by user "
@@ -465,7 +467,7 @@ static int mgs_handle_target_reg(struct ptlrpc_request *req)
                                       "in order to regenerate the logs."
                                       "\n", obd->obd_name, mti->mti_fsname);
                 } else if (mti->mti_flags &
-                           (LDD_F_SV_TYPE_OST | LDD_F_SV_TYPE_MDT)) {
+                           (SVTYPE_OST | SVTYPE_MDT)) {
                         rc = mgs_erase_log(obd, mti->mti_svname);
                         LCONSOLE_WARN("%s: Regenerating %s log by user "
                                       "request.\n",
@@ -476,9 +478,23 @@ static int mgs_handle_target_reg(struct ptlrpc_request *req)
                 mti->mti_flags &= ~LDD_F_UPGRADE14;
         }
 
+        rc = mgs_find_or_make_fsdb(obd, mti->mti_fsname, &fsdb);
+        if (rc) {
+                CERROR("Can't get db for %s: %d\n", mti->mti_fsname, rc);
+                GOTO(out_nolock, rc);
+        }
+
+        /*
+         * Log writing contention is handled by the fsdb_sem.
+         *
+         * It should be alright if someone was reading while we were
+         * updating the logs - if we revoke at the end they will just update
+         * from where they left off.
+         */
+
         /* COMPAT_146 */
         if (mti->mti_flags & LDD_F_UPGRADE14) {
-                rc = mgs_upgrade_sv_14(obd, mti);
+                rc = mgs_upgrade_sv_14(obd, mti, fsdb);
                 if (rc) {
                         CERROR("Can't upgrade from 1.4 (%d)\n", rc);
                         GOTO(out, rc);
@@ -495,7 +511,7 @@ static int mgs_handle_target_reg(struct ptlrpc_request *req)
 
                 /* create or update the target log
                    and update the client/mdt logs */
-                rc = mgs_write_log_target(obd, mti);
+                rc = mgs_write_log_target(obd, mti, fsdb);
                 if (rc) {
                         CERROR("Failed to write %s log (%d)\n",
                                mti->mti_svname, rc);
@@ -509,12 +525,12 @@ static int mgs_handle_target_reg(struct ptlrpc_request *req)
         }
 
 out:
-        /* done with log update */
-        if (lockrc == ELDLM_OK)
-                mgs_put_cfg_lock(&lockh);
+        mgs_revoke_lock(obd, fsdb);
+
 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;
         rc = req_capsule_server_pack(&req->rq_pill);
         if (rc)
                 RETURN(rc);
@@ -532,7 +548,6 @@ static int mgs_set_info_rpc(struct ptlrpc_request *req)
 {
         struct obd_device *obd = req->rq_export->exp_obd;
         struct mgs_send_param *msp, *rep_msp;
-        struct lustre_handle lockh;
         int rc;
         struct lustre_cfg_bufs bufs;
         struct lustre_cfg *lcfg;
@@ -553,9 +568,6 @@ static int mgs_set_info_rpc(struct ptlrpc_request *req)
                 RETURN(rc);
         }
 
-        /* request for update */
-        mgs_revoke_lock(obd, fsname, &lockh);
-
         lustre_cfg_free(lcfg);
 
         rc = req_capsule_server_pack(&req->rq_pill);
@@ -582,7 +594,7 @@ static int mgs_connect_check_sptlrpc(struct ptlrpc_request *req)
                 if (rc)
                         return rc;
 
-                down(&fsdb->fsdb_sem);
+                cfs_down(&fsdb->fsdb_sem);
                 if (sptlrpc_rule_set_choose(&fsdb->fsdb_srpc_gen,
                                             LUSTRE_SP_MGC, LUSTRE_SP_MGS,
                                             req->rq_peer.nid,
@@ -590,9 +602,9 @@ static int mgs_connect_check_sptlrpc(struct ptlrpc_request *req)
                         /* by defualt allow any flavors */
                         flvr.sf_rpc = SPTLRPC_FLVR_ANY;
                 }
-                up(&fsdb->fsdb_sem);
+                cfs_up(&fsdb->fsdb_sem);
 
-                spin_lock(&exp->exp_lock);
+                cfs_spin_lock(&exp->exp_lock);
 
                 exp->exp_sp_peer = req->rq_sp_from;
                 exp->exp_flvr = flvr;
@@ -605,7 +617,7 @@ static int mgs_connect_check_sptlrpc(struct ptlrpc_request *req)
                         rc = -EACCES;
                 }
 
-                spin_unlock(&exp->exp_lock);
+                cfs_spin_unlock(&exp->exp_lock);
         } else {
                 if (exp->exp_sp_peer != req->rq_sp_from) {
                         CERROR("RPC source %s doesn't match %s\n",
@@ -658,7 +670,7 @@ int mgs_handle(struct ptlrpc_request *req)
                 GOTO(out, rc = 0);
 
         if (opc != MGS_CONNECT) {
-                if (req->rq_export == NULL) {
+                if (!class_connected_export(req->rq_export)) {
                         CERROR("lustre_mgs: operation %d on unconnected MGS\n",
                                opc);
                         req->rq_status = -ENOTCONN;
@@ -774,6 +786,10 @@ out:
 
 static inline int mgs_init_export(struct obd_export *exp)
 {
+        cfs_spin_lock(&exp->exp_lock);
+        exp->exp_connecting = 1;
+        cfs_spin_unlock(&exp->exp_lock);
+
         return ldlm_init_export(exp);
 }
 
@@ -810,7 +826,6 @@ static int mgs_iocontrol_pool(struct obd_device *obd,
                               struct obd_ioctl_data *data)
 {
         int rc;
-        struct lustre_handle lockh;
         struct lustre_cfg *lcfg = NULL;
         struct llog_rec_hdr rec;
         char *fsname = NULL;
@@ -845,7 +860,7 @@ static int mgs_iocontrol_pool(struct obd_device *obd,
         if (lcfg == NULL)
                 GOTO(out_pool, rc = -ENOMEM);
 
-        if (copy_from_user(lcfg, data->ioc_pbuf1, data->ioc_plen1))
+        if (cfs_copy_from_user(lcfg, data->ioc_pbuf1, data->ioc_plen1))
                 GOTO(out_pool, rc = -EFAULT);
 
         if (lcfg->lcfg_bufcount < 2) {
@@ -897,9 +912,6 @@ static int mgs_iocontrol_pool(struct obd_device *obd,
                 GOTO(out_pool, rc);
         }
 
-        /* request for update */
-        mgs_revoke_lock(obd, fsname, &lockh);
-
 out_pool:
         if (lcfg != NULL)
                 OBD_FREE(lcfg, data->ioc_plen1);
@@ -928,7 +940,6 @@ int mgs_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
         switch (cmd) {
 
         case OBD_IOC_PARAM: {
-                struct lustre_handle lockh;
                 struct lustre_cfg *lcfg;
                 struct llog_rec_hdr rec;
                 char fsname[MTI_NAME_MAXLEN];
@@ -945,7 +956,7 @@ int mgs_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
                 OBD_ALLOC(lcfg, data->ioc_plen1);
                 if (lcfg == NULL)
                         RETURN(-ENOMEM);
-                if (copy_from_user(lcfg, data->ioc_pbuf1, data->ioc_plen1))
+                if (cfs_copy_from_user(lcfg, data->ioc_pbuf1, data->ioc_plen1))
                         GOTO(out_free, rc = -EFAULT);
 
                 if (lcfg->lcfg_bufcount < 1)
@@ -956,13 +967,6 @@ int mgs_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
                         CERROR("setparam err %d\n", rc);
                         GOTO(out_free, rc);
                 }
-
-                /* Revoke lock so everyone updates.  Should be alright if
-                   someone was already reading while we were updating the logs,
-                   so we don't really need to hold the lock while we're
-                   writing (above). */
-                mgs_revoke_lock(obd, fsname, &lockh);
-
 out_free:
                 OBD_FREE(lcfg, data->ioc_plen1);
                 RETURN(rc);