Whamcloud - gitweb
branch: HEAD
[fs/lustre-release.git] / lustre / mgc / mgc_request.c
index 37cfda5..2ab2ef3 100644 (file)
@@ -126,28 +126,42 @@ static int config_log_get(struct config_llog_data *cld)
 static void config_log_put(struct config_llog_data *cld)
 {
         ENTRY;
+
         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
                atomic_read(&cld->cld_refcount));
-        if (atomic_dec_and_test(&cld->cld_refcount)) {
-                CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
-                class_export_put(cld->cld_mgcexp);
-                spin_lock(&config_list_lock);
+        LASSERT(atomic_read(&cld->cld_refcount) > 0);
+
+        /* spinlock to make sure no item with 0 refcount in the list */
+        spin_lock(&config_list_lock);
+        if (unlikely(atomic_dec_and_test(&cld->cld_refcount))) {
                 list_del(&cld->cld_list_chain);
                 spin_unlock(&config_list_lock);
+
+                CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
+
+                if (cld->cld_sptlrpc)
+                        config_log_put(cld->cld_sptlrpc);
+                if (cld->cld_is_sptlrpc)
+                        sptlrpc_conf_log_stop(cld->cld_logname);
+
+                class_export_put(cld->cld_mgcexp);
                 OBD_FREE(cld->cld_logname, strlen(cld->cld_logname) + 1);
                 if (cld->cld_cfg.cfg_instance != NULL)
                         OBD_FREE(cld->cld_cfg.cfg_instance,
                                  strlen(cld->cld_cfg.cfg_instance) + 1);
                 OBD_FREE(cld, sizeof(*cld));
+        } else {
+                spin_unlock(&config_list_lock);
         }
+
         EXIT;
 }
 
 /* Find a config log by name */
-static struct config_llog_data *config_log_find(char *logname,
-                                               struct config_llog_instance *cfg)
+static
+struct config_llog_data *config_log_find(char *logname,
+                                         struct config_llog_instance *cfg)
 {
-        struct list_head *tmp;
         struct config_llog_data *cld;
         char *logid = logname;
         int match_instance = 0;
@@ -163,8 +177,7 @@ static struct config_llog_data *config_log_find(char *logname,
         }
 
         spin_lock(&config_list_lock);
-        list_for_each(tmp, &config_llog_list) {
-                cld = list_entry(tmp, struct config_llog_data, cld_list_chain);
+        list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
                 if (match_instance && cld->cld_cfg.cfg_instance &&
                     strcmp(logid, cld->cld_cfg.cfg_instance) == 0)
                         goto out_found;
@@ -179,56 +192,123 @@ static struct config_llog_data *config_log_find(char *logname,
 out_found:
         atomic_inc(&cld->cld_refcount);
         spin_unlock(&config_list_lock);
+        LASSERT(cld->cld_stopping == 0 || cld->cld_is_sptlrpc == 0);
         RETURN(cld);
 }
 
-/* Add this log to our list of active logs.
-   We have one active log per "mount" - client instance or servername.
-   Each instance may be at a different point in the log. */
-static int config_log_add(char *logname, struct config_llog_instance *cfg,
-                          struct super_block *sb)
+static
+struct config_llog_data *do_config_log_add(struct obd_device *obd,
+                                           char *logname,
+                                           unsigned int is_sptlrpc,
+                                           struct config_llog_instance *cfg,
+                                           struct super_block *sb)
 {
         struct config_llog_data *cld;
-        struct lustre_sb_info *lsi = s2lsi(sb);
-        int rc;
+        int                      rc;
         ENTRY;
 
-        CDEBUG(D_MGC, "adding config log %s:%s\n", logname, cfg->cfg_instance);
+        CDEBUG(D_MGC, "do adding config log %s:%s\n", logname,
+               cfg ? cfg->cfg_instance : "NULL");
 
         OBD_ALLOC(cld, sizeof(*cld));
         if (!cld)
-                RETURN(-ENOMEM);
+                RETURN(ERR_PTR(-ENOMEM));
         OBD_ALLOC(cld->cld_logname, strlen(logname) + 1);
         if (!cld->cld_logname) {
                 OBD_FREE(cld, sizeof(*cld));
-                RETURN(-ENOMEM);
+                RETURN(ERR_PTR(-ENOMEM));
         }
         strcpy(cld->cld_logname, logname);
-        cld->cld_cfg = *cfg;
+        if (cfg)
+                cld->cld_cfg = *cfg;
         cld->cld_cfg.cfg_last_idx = 0;
         cld->cld_cfg.cfg_flags = 0;
         cld->cld_cfg.cfg_sb = sb;
+        cld->cld_is_sptlrpc = is_sptlrpc;
         atomic_set(&cld->cld_refcount, 1);
 
         /* Keep the mgc around until we are done */
-        cld->cld_mgcexp = class_export_get(lsi->lsi_mgc->obd_self_export);
+        cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 
-        if (cfg->cfg_instance != NULL) {
+        if (cfg && cfg->cfg_instance != NULL) {
                 OBD_ALLOC(cld->cld_cfg.cfg_instance,
                           strlen(cfg->cfg_instance) + 1);
                 strcpy(cld->cld_cfg.cfg_instance, cfg->cfg_instance);
         }
+
+        if (is_sptlrpc) {
+                sptlrpc_conf_log_start(logname);
+                cld->cld_cfg.cfg_obdname = obd->obd_name;
+        }
+
         rc = mgc_logname2resid(logname, &cld->cld_resid);
+
         spin_lock(&config_list_lock);
         list_add(&cld->cld_list_chain, &config_llog_list);
         spin_unlock(&config_list_lock);
 
         if (rc) {
                 config_log_put(cld);
-                RETURN(rc);
+                RETURN(ERR_PTR(rc));
         }
 
-        RETURN(rc);
+        if (is_sptlrpc) {
+                rc = mgc_process_log(obd, cld);
+                if (rc)
+                        CERROR("failed processing sptlrpc log: %d\n", rc);
+        }
+
+        RETURN(cld);
+}
+
+/**
+ * Add this log to our list of active logs.
+ * We have one active log per "mount" - client instance or servername.
+ * Each instance may be at a different point in the log.
+ */
+static int config_log_add(struct obd_device *obd, char *logname,
+                          struct config_llog_instance *cfg,
+                          struct super_block *sb)
+{
+        struct config_llog_data *cld, *sptlrpc_cld;
+        char                     seclogname[20];
+        char                    *ptr;
+        ENTRY;
+
+        CDEBUG(D_MGC, "adding config log %s:%s\n", logname, cfg->cfg_instance);
+
+        /*
+         * for each regular log, the depended sptlrpc log name is
+         * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
+         */
+        ptr = strrchr(logname, '-');
+        if (ptr == NULL || ptr - logname > 8) {
+                CERROR("logname %s is too long\n", logname);
+                RETURN(-EINVAL);
+        }
+
+        memcpy(seclogname, logname, ptr - logname);
+        strcpy(seclogname + (ptr - logname), "-sptlrpc");
+
+        sptlrpc_cld = config_log_find(seclogname, NULL);
+        if (IS_ERR(sptlrpc_cld)) {
+                sptlrpc_cld = do_config_log_add(obd, seclogname, 1, NULL, NULL);
+                if (IS_ERR(sptlrpc_cld)) {
+                        CERROR("can't create sptlrpc log: %s\n", seclogname);
+                        RETURN(PTR_ERR(sptlrpc_cld));
+                }
+        }
+
+        cld = do_config_log_add(obd, logname, 0, cfg, sb);
+        if (IS_ERR(cld)) {
+                CERROR("can't create log: %s\n", logname);
+                config_log_put(sptlrpc_cld);
+                RETURN(PTR_ERR(cld));
+        }
+
+        cld->cld_sptlrpc = sptlrpc_cld;
+
+        RETURN(0);
 }
 
 DECLARE_MUTEX(llog_process_lock);
@@ -236,7 +316,7 @@ DECLARE_MUTEX(llog_process_lock);
 /* Stop watching for updates on this log. */
 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 {
-        struct config_llog_data *cld;
+        struct config_llog_data *cld, *cld_sptlrpc;
         int rc = 0;
         ENTRY;
 
@@ -250,6 +330,14 @@ static int config_log_end(char *logname, struct config_llog_instance *cfg)
         cld->cld_stopping = 1;
         up(&llog_process_lock);
 
+        spin_lock(&config_list_lock);
+        cld_sptlrpc = cld->cld_sptlrpc;
+        cld->cld_sptlrpc = NULL;
+        spin_unlock(&config_list_lock);
+
+        if (cld_sptlrpc)
+                config_log_put(cld_sptlrpc);
+
         /* drop the start ref */
         config_log_put(cld);
         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
@@ -265,14 +353,29 @@ static int config_log_end(char *logname, struct config_llog_instance *cfg)
 static int rq_state = 0;
 static cfs_waitq_t rq_waitq;
 
-static int mgc_process_log(struct obd_device *mgc,
-                           struct config_llog_data *cld);
 static int mgc_requeue_add(struct config_llog_data *cld, int later);
 
+static void do_requeue(struct config_llog_data *cld)
+{
+        LASSERT(atomic_read(&cld->cld_refcount) > 0);
+
+        if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
+                CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
+                mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
+        } else {
+                CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
+                       cld->cld_logname);
+        }
+
+        /* Whether we enqueued again or not in mgc_process_log, we're done
+         * with the ref from the old enqueue */
+        config_log_put(cld);
+}
+
 static int mgc_requeue_thread(void *data)
 {
         struct l_wait_info lwi_now, lwi_later;
-        struct config_llog_data *cld, *n;
+        struct config_llog_data *cld, *cld_next, *cld_prev;
         char name[] = "ll_cfg_requeue";
         int rc = 0;
         ENTRY;
@@ -298,22 +401,52 @@ static int mgc_requeue_thread(void *data)
                                       NULL, NULL);
                 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi_now);
 
+                /*
+                 * iterate & processing through the list. for each cld, process
+                 * its depending sptlrpc cld firstly (if any) and then itself.
+                 *
+                 * it's guaranteed any item in the list must have
+                 * reference > 0; and if cld_lostlock is set, at
+                 * least one reference is taken by the previous enqueue.
+                 *
+                 * Note: releasing a cld might lead to itself and its depended
+                 * sptlrpc cld be unlinked from the list. to safely iterate
+                 * we need to take a reference on next cld before processing.
+                 */
+                cld_prev = NULL;
+
                 spin_lock(&config_list_lock);
-                list_for_each_entry_safe(cld, n, &config_llog_list,
+                list_for_each_entry_safe(cld, cld_next, &config_llog_list,
                                          cld_list_chain) {
-                        spin_unlock(&config_list_lock);
+                        if (cld->cld_list_chain.next != &config_llog_list)
+                                atomic_inc(&cld_next->cld_refcount);
+
                         if (cld->cld_lostlock) {
-                                CDEBUG(D_MGC, "updating log %s\n",
-                                       cld->cld_logname);
+                                if (cld->cld_sptlrpc &&
+                                    cld->cld_sptlrpc->cld_lostlock) {
+                                        cld->cld_sptlrpc->cld_lostlock = 0;
+
+                                        spin_unlock(&config_list_lock);
+                                        do_requeue(cld->cld_sptlrpc);
+                                        spin_lock(&config_list_lock);
+                                        LASSERT(cld->cld_lostlock);
+                                }
+
                                 cld->cld_lostlock = 0;
-                                rc = mgc_process_log(cld->cld_mgcexp->exp_obd,
-                                                     cld);
-                                /* Whether we enqueued again or not in
-                                   mgc_process_log, we're done with the ref
-                                   from the old enqueue */
-                                config_log_put(cld);
+
+                                spin_unlock(&config_list_lock);
+                                do_requeue(cld);
+                                spin_lock(&config_list_lock);
                         }
-                        spin_lock(&config_list_lock);
+
+
+                        if (cld_prev) {
+                                spin_unlock(&config_list_lock);
+                                config_log_put(cld_prev);
+                                spin_lock(&config_list_lock);
+                        }
+
+                        cld_prev = cld_next;
                 }
                 spin_unlock(&config_list_lock);
 
@@ -339,11 +472,13 @@ static int mgc_requeue_add(struct config_llog_data *cld, int later)
         CDEBUG(D_INFO, "log %s: requeue (l=%d r=%d sp=%d st=%x)\n",
                cld->cld_logname, later, atomic_read(&cld->cld_refcount),
                cld->cld_stopping, rq_state);
+        LASSERT(atomic_read(&cld->cld_refcount) > 0);
 
         /* Hold lock for rq_state */
         spin_lock(&config_list_lock);
 
         if (cld->cld_stopping || (rq_state & RQ_STOP)) {
+                cld->cld_lostlock = 0;
                 spin_unlock(&config_list_lock);
                 config_log_put(cld);
                 RETURN(0);
@@ -599,7 +734,8 @@ static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
                         break;
                 }
                 /* Did we fail to get the lock? */
-                if (lock->l_req_mode != lock->l_granted_mode) {
+                if (lock->l_req_mode != lock->l_granted_mode &&
+                    !cld->cld_is_sptlrpc) {
                         CDEBUG(D_MGC, "log %s: original grant failed, will "
                                "requeue later\n", cld->cld_logname);
                         /* Try to re-enqueue later */
@@ -1081,22 +1217,18 @@ out:
 
 /* Get a config log from the MGS and process it.
    This func is called for both clients and servers. */
-static int mgc_process_log(struct obd_device *mgc,
-                           struct config_llog_data *cld)
+int mgc_process_log(struct obd_device *mgc,
+                    struct config_llog_data *cld)
 {
         struct llog_ctxt *ctxt, *lctxt;
         struct lustre_handle lockh;
         struct client_obd *cli = &mgc->u.cli;
         struct lvfs_run_ctxt saved;
-        struct lustre_sb_info *lsi;
+        struct lustre_sb_info *lsi = NULL;
         int rc = 0, rcl, flags = 0, must_pop = 0;
         ENTRY;
 
-        if (!cld || !cld->cld_cfg.cfg_sb) {
-                /* This should never happen */
-                CERROR("Missing cld, aborting log update\n");
-                RETURN(-EINVAL);
-        }
+        LASSERT(cld);
 
         /* I don't want mutliple processes running process_log at once --
            sounds like badness.  It actually might be fine, as long as
@@ -1111,7 +1243,8 @@ static int mgc_process_log(struct obd_device *mgc,
 
         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
 
-        lsi = s2lsi(cld->cld_cfg.cfg_sb);
+        if (cld->cld_cfg.cfg_sb)
+                lsi = s2lsi(cld->cld_cfg.cfg_sb);
 
         CDEBUG(D_MGC, "Process log %s:%s from %d\n", cld->cld_logname,
                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
@@ -1132,6 +1265,13 @@ static int mgc_process_log(struct obd_device *mgc,
 
         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
 
+        /*
+         * local copy of sptlrpc log is controlled elsewhere, don't try to
+         * read it up here.
+         */
+        if (rcl && cld->cld_is_sptlrpc)
+                goto out_pop;
+        
         /* Copy the setup log locally if we can. Don't mess around if we're
            running an MGS though (logs are already local). */
         if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) &&
@@ -1160,6 +1300,9 @@ static int mgc_process_log(struct obd_device *mgc,
                 ctxt = lctxt;
         }
 
+        if (cld->cld_is_sptlrpc)
+                sptlrpc_conf_log_update_begin(cld->cld_logname);
+
         /* logname and instance info should be the same, so use our
            copy of the instance for the update.  The cfg_last_idx will
            be updated here. */
@@ -1171,6 +1314,19 @@ out_pop:
         if (must_pop)
                 pop_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
 
+        /*
+         * update settings on existing OBDs. doing it inside
+         * of llog_process_lock so no device is attaching/detaching
+         * in parallel.
+         * the logname must be <fsname>-sptlrpc
+         */
+        if (cld->cld_is_sptlrpc && rcl == 0) {
+                sptlrpc_conf_log_update_end(cld->cld_logname);
+                class_notify_sptlrpc_conf(cld->cld_logname,
+                                          strlen(cld->cld_logname) -
+                                          strlen("-sptlrpc"));
+        }
+
         /* Now drop the lock so MGS can revoke it */
         if (!rcl) {
                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
@@ -1215,6 +1371,10 @@ static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
                 CERROR("lov_del_obd unimplemented\n");
                 rc = -ENOSYS;
                 break;
+        case LCFG_SPTLRPC_CONF: {
+                rc = sptlrpc_process_config(lcfg);
+                break;
+        }
         case LCFG_LOG_START: {
                 struct config_llog_data *cld;
                 struct config_llog_instance *cfg;
@@ -1227,7 +1387,7 @@ static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
                        cfg->cfg_last_idx);
 
                 /* We're only called through here on the initial mount */
-                rc = config_log_add(logname, cfg, sb);
+                rc = config_log_add(obd, logname, cfg, sb);
                 if (rc)
                         break;
                 cld = config_log_find(logname, cfg);