Whamcloud - gitweb
LU-7888 obdclass: not hold global lock when lwp callback
[fs/lustre-release.git] / lustre / obdclass / obd_mount_server.c
index 148cbf3..9945ef0 100644 (file)
@@ -27,7 +27,7 @@
  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2013, 2014, Intel Corporation.
+ * Copyright (c) 2013, 2015, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
 #define PRINT_CMD CDEBUG
 #define PRINT_MASK (D_SUPER | D_CONFIG)
 
+#include <linux/types.h>
+#include <linux/selinux.h>
 #include <linux/statfs.h>
-#include <obd.h>
-#include <obd_class.h>
-#include <lustre/lustre_user.h>
 #include <linux/version.h>
-#include <lustre_ioctl.h>
-#include <lustre_log.h>
-#include <lustre_disk.h>
-#include <lustre_param.h>
 #ifdef HAVE_KERNEL_LOCKED
 #include <linux/smp_lock.h>
 #endif
-#ifdef HAVE_SELINUX_IS_ENABLED
-#include <linux/selinux.h>
-#endif
+
+#include <lustre/lustre_idl.h>
+#include <lustre/lustre_user.h>
+
+#include <llog_swab.h>
+#include <lustre_disk.h>
+#include <lustre_ioctl.h>
+#include <lustre_log.h>
+#include <lustre_param.h>
+#include <obd.h>
+#include <obd_class.h>
 
 /*********** mount lookup *********/
 
@@ -241,8 +244,8 @@ static int server_start_mgs(struct super_block *sb)
 
        if (!rc) {
                rc = lustre_start_simple(LUSTRE_MGS_OBDNAME, LUSTRE_MGS_NAME,
-                                        LUSTRE_MGS_OBDNAME, 0, 0,
-                                        lsi->lsi_osd_obdname, 0);
+                                        LUSTRE_MGS_OBDNAME, NULL, NULL,
+                                        lsi->lsi_osd_obdname, NULL);
                /* server_deregister_mount() is not called previously, for lsi
                 * and other stuff can't be freed cleanly when mgs calls
                 * server_put_mount() in error handling case (see b=17758),
@@ -264,8 +267,14 @@ static int server_stop_mgs(struct super_block *sb)
 {
        struct obd_device *obd;
        int rc;
+       struct lustre_mount_info *lmi;
        ENTRY;
 
+       /* Do not stop MGS if this device is not the running MGT */
+       lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
+       if (lmi != NULL && lmi->lmi_sb != sb)
+               RETURN(0);
+
        CDEBUG(D_MOUNT, "Stop MGS service %s\n", LUSTRE_MGS_OBDNAME);
 
        /* There better be only one MGS */
@@ -283,7 +292,8 @@ static int server_stop_mgs(struct super_block *sb)
 
 /* Since there's only one mgc per node, we have to change it's fs to get
    access to the right disk. */
-static int server_mgc_set_fs(struct obd_device *mgc, struct super_block *sb)
+static int server_mgc_set_fs(const struct lu_env *env,
+                            struct obd_device *mgc, struct super_block *sb)
 {
        struct lustre_sb_info *lsi = s2lsi(sb);
        int rc;
@@ -292,7 +302,7 @@ static int server_mgc_set_fs(struct obd_device *mgc, struct super_block *sb)
        CDEBUG(D_MOUNT, "Set mgc disk for %s\n", lsi->lsi_lmd->lmd_dev);
 
        /* cl_mgc_sem in mgc insures we sleep if the mgc_fs is busy */
-       rc = obd_set_info_async(NULL, mgc->obd_self_export,
+       rc = obd_set_info_async(env, mgc->obd_self_export,
                                sizeof(KEY_SET_FS), KEY_SET_FS,
                                sizeof(*sb), sb, NULL);
        if (rc != 0)
@@ -301,14 +311,15 @@ static int server_mgc_set_fs(struct obd_device *mgc, struct super_block *sb)
        RETURN(rc);
 }
 
-static int server_mgc_clear_fs(struct obd_device *mgc)
+static int server_mgc_clear_fs(const struct lu_env *env,
+                              struct obd_device *mgc)
 {
        int rc;
        ENTRY;
 
        CDEBUG(D_MOUNT, "Unassign mgc disk\n");
 
-       rc = obd_set_info_async(NULL, mgc->obd_self_export,
+       rc = obd_set_info_async(env, mgc->obd_self_export,
                                sizeof(KEY_CLEAR_FS), KEY_CLEAR_FS,
                                0, NULL, NULL);
        RETURN(rc);
@@ -378,7 +389,18 @@ EXPORT_SYMBOL(tgt_name2lwp_name);
 
 static struct list_head lwp_register_list =
        LIST_HEAD_INIT(lwp_register_list);
-static DEFINE_MUTEX(lwp_register_list_lock);
+static DEFINE_SPINLOCK(lwp_register_list_lock);
+
+static void lustre_put_lwp_item(struct lwp_register_item *lri)
+{
+       if (atomic_dec_and_test(&lri->lri_ref)) {
+               LASSERT(list_empty(&lri->lri_list));
+
+               if (*lri->lri_exp != NULL)
+                       class_export_put(*lri->lri_exp);
+               OBD_FREE_PTR(lri);
+       }
+}
 
 int lustre_register_lwp_item(const char *lwpname, struct obd_export **exp,
                             register_lwp_cb cb_func, void *cb_data)
@@ -395,15 +417,12 @@ int lustre_register_lwp_item(const char *lwpname, struct obd_export **exp,
        if (lri == NULL)
                RETURN(-ENOMEM);
 
-       mutex_lock(&lwp_register_list_lock);
-
        lwp = class_name2obd(lwpname);
        if (lwp != NULL && lwp->obd_set_up == 1) {
                struct obd_uuid *uuid;
 
                OBD_ALLOC_PTR(uuid);
                if (uuid == NULL) {
-                       mutex_unlock(&lwp_register_list_lock);
                        OBD_FREE_PTR(lri);
                        RETURN(-ENOMEM);
                }
@@ -417,31 +436,40 @@ int lustre_register_lwp_item(const char *lwpname, struct obd_export **exp,
        lri->lri_cb_func = cb_func;
        lri->lri_cb_data = cb_data;
        INIT_LIST_HEAD(&lri->lri_list);
+       /*
+        * Initialize the lri_ref at 2, one will be released before
+        * current function returned via lustre_put_lwp_item(), the
+        * other will be released in lustre_deregister_lwp_item().
+        */
+       atomic_set(&lri->lri_ref, 2);
+
+       spin_lock(&lwp_register_list_lock);
        list_add(&lri->lri_list, &lwp_register_list);
+       spin_unlock(&lwp_register_list_lock);
 
        if (*exp != NULL && cb_func != NULL)
                cb_func(cb_data);
+       lustre_put_lwp_item(lri);
 
-       mutex_unlock(&lwp_register_list_lock);
        RETURN(0);
 }
 EXPORT_SYMBOL(lustre_register_lwp_item);
 
 void lustre_deregister_lwp_item(struct obd_export **exp)
 {
-       struct lwp_register_item *lri, *tmp;
+       struct lwp_register_item *lri;
 
-       mutex_lock(&lwp_register_list_lock);
-       list_for_each_entry_safe(lri, tmp, &lwp_register_list, lri_list) {
+       spin_lock(&lwp_register_list_lock);
+       list_for_each_entry(lri, &lwp_register_list, lri_list) {
                if (exp == lri->lri_exp) {
-                       if (*exp)
-                               class_export_put(*exp);
-                       list_del(&lri->lri_list);
-                       OBD_FREE_PTR(lri);
-                       break;
+                       list_del_init(&lri->lri_list);
+                       spin_unlock(&lwp_register_list_lock);
+
+                       lustre_put_lwp_item(lri);
+                       return;
                }
        }
-       mutex_unlock(&lwp_register_list_lock);
+       spin_unlock(&lwp_register_list_lock);
 }
 EXPORT_SYMBOL(lustre_deregister_lwp_item);
 
@@ -472,7 +500,7 @@ struct obd_export *lustre_find_lwp_by_index(const char *dev, __u32 idx)
        list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
                char *ptr = strstr(lwp->obd_name, lwp_name);
 
-               if (ptr != NULL) {
+               if (ptr != NULL && lwp->obd_lwp_export != NULL) {
                        exp = class_export_get(lwp->obd_lwp_export);
                        break;
                }
@@ -486,23 +514,36 @@ err_lmi:
 }
 EXPORT_SYMBOL(lustre_find_lwp_by_index);
 
-static void lustre_notify_lwp_list(struct obd_export *exp)
+void lustre_notify_lwp_list(struct obd_export *exp)
 {
-       struct lwp_register_item *lri, *tmp;
+       struct lwp_register_item *lri;
        LASSERT(exp != NULL);
 
-       mutex_lock(&lwp_register_list_lock);
-       list_for_each_entry_safe(lri, tmp, &lwp_register_list, lri_list) {
+again:
+       spin_lock(&lwp_register_list_lock);
+       list_for_each_entry(lri, &lwp_register_list, lri_list) {
                if (strcmp(exp->exp_obd->obd_name, lri->lri_name))
                        continue;
                if (*lri->lri_exp != NULL)
                        continue;
                *lri->lri_exp = class_export_get(exp);
+               atomic_inc(&lri->lri_ref);
+               spin_unlock(&lwp_register_list_lock);
+
                if (lri->lri_cb_func != NULL)
                        lri->lri_cb_func(lri->lri_cb_data);
+               lustre_put_lwp_item(lri);
+
+               /* Others may have changed the list after we unlock, we have
+                * to rescan the list from the beginning. Usually, the list
+                * 'lwp_register_list' is very short, and there is 'guard'
+                * lri::lri_exp that will prevent the callback to be done
+                * repeatedly. So rescanning the list has no problem. */
+               goto again;
        }
-       mutex_unlock(&lwp_register_list_lock);
+       spin_unlock(&lwp_register_list_lock);
 }
+EXPORT_SYMBOL(lustre_notify_lwp_list);
 
 static int lustre_lwp_connect(struct obd_device *lwp)
 {
@@ -533,7 +574,8 @@ static int lustre_lwp_connect(struct obd_device *lwp)
        data->ocd_connect_flags |= OBD_CONNECT_MDS_MDS | OBD_CONNECT_FID |
                OBD_CONNECT_AT | OBD_CONNECT_LRU_RESIZE |
                OBD_CONNECT_FULL20 | OBD_CONNECT_LVB_TYPE |
-               OBD_CONNECT_LIGHTWEIGHT | OBD_CONNECT_LFSCK;
+               OBD_CONNECT_LIGHTWEIGHT | OBD_CONNECT_LFSCK |
+               OBD_CONNECT_BULK_MBITS;
        OBD_ALLOC_PTR(uuid);
        if (uuid == NULL)
                GOTO(out, rc = -ENOMEM);
@@ -553,7 +595,6 @@ static int lustre_lwp_connect(struct obd_device *lwp)
                if (unlikely(lwp->obd_lwp_export != NULL))
                        class_export_put(lwp->obd_lwp_export);
                lwp->obd_lwp_export = class_export_get(exp);
-               lustre_notify_lwp_list(exp);
        }
 
        GOTO(out, rc);
@@ -609,7 +650,7 @@ static int lustre_lwp_setup(struct lustre_cfg *lcfg, struct lustre_sb_info *lsi,
        sprintf(lwpuuid, "%s_UUID", lwpname);
        rc = lustre_start_simple(lwpname, LUSTRE_LWP_NAME,
                                 lwpuuid, lustre_cfg_string(lcfg, 1),
-                                0, 0, 0);
+                                NULL, NULL, NULL);
        if (rc) {
                CERROR("%s: setup up failed: rc %d\n", lwpname, rc);
                GOTO(out, rc);
@@ -877,7 +918,7 @@ static int lustre_disconnect_lwp(struct super_block *sb)
                /* end log first */
                cfg->cfg_instance = sb;
                rc = lustre_end_log(sb, logname, cfg);
-               if (rc != 0)
+               if (rc != 0 && rc != -ENOENT)
                        GOTO(out, rc);
 
                lsi->lsi_lwp_started = 0;
@@ -986,8 +1027,8 @@ static int lustre_start_lwp(struct super_block *sb)
        cfg->cfg_callback = client_lwp_config_process;
        cfg->cfg_instance = sb;
        rc = lustre_process_log(sb, logname, cfg);
-       if (rc == 0)
-               lsi->lsi_lwp_started = 1;
+       /* need to remove config llog from mgc */
+       lsi->lsi_lwp_started = 1;
 
        GOTO(out, rc);
 
@@ -1047,7 +1088,6 @@ int server_mti_print(const char *title, struct mgs_target_info *mti)
                  mti->mti_config_ver, mti->mti_flags);
        return 0;
 }
-EXPORT_SYMBOL(server_mti_print);
 
 /* Generate data for registration */
 static int server_lsi2mti(struct lustre_sb_info *lsi,
@@ -1164,8 +1204,8 @@ static int server_register_target(struct lustre_sb_info *lsi)
                                "rc = %d. Is the MGS running?\n",
                                lsi->lsi_svname, rc);
                } else {
-                       CERROR("%s: error registering with the MGS: rc = %d "
-                              "(not fatal)\n", lsi->lsi_svname, rc);
+                       CDEBUG(D_HA, "%s: error registering with the MGS: "
+                              "rc = %d (not fatal)\n", lsi->lsi_svname, rc);
                        /* reset the error code for non-fatal error. */
                        rc = 0;
                }
@@ -1230,7 +1270,7 @@ static int server_start_targets(struct super_block *sb)
        struct obd_device *obd;
        struct lustre_sb_info *lsi = s2lsi(sb);
        struct config_llog_instance cfg;
-       struct lu_env env;
+       struct lu_env mgc_env;
        struct lu_device *dev;
        int rc;
        ENTRY;
@@ -1245,7 +1285,7 @@ static int server_start_targets(struct super_block *sb)
                        rc = lustre_start_simple(LUSTRE_MDS_OBDNAME,
                                                 LUSTRE_MDS_NAME,
                                                 LUSTRE_MDS_OBDNAME"_uuid",
-                                                0, 0, 0, 0);
+                                                NULL, NULL, NULL, NULL);
                        if (rc) {
                                mutex_unlock(&server_start_lock);
                                CERROR("failed to start MDS: %d\n", rc);
@@ -1264,7 +1304,7 @@ static int server_start_targets(struct super_block *sb)
                        rc = lustre_start_simple(LUSTRE_OSS_OBDNAME,
                                                 LUSTRE_OSS_NAME,
                                                 LUSTRE_OSS_OBDNAME"_uuid",
-                                                0, 0, 0, 0);
+                                                NULL, NULL, NULL, NULL);
                        if (rc) {
                                mutex_unlock(&server_start_lock);
                                CERROR("failed to start OSS: %d\n", rc);
@@ -1274,11 +1314,15 @@ static int server_start_targets(struct super_block *sb)
                mutex_unlock(&server_start_lock);
        }
 
+       rc = lu_env_init(&mgc_env, LCT_MG_THREAD);
+       if (rc != 0)
+               GOTO(out_stop_service, rc);
+
        /* Set the mgc fs to our server disk.  This allows the MGC to
         * read and write configs locally, in case it can't talk to the MGS. */
-       rc = server_mgc_set_fs(lsi->lsi_mgc, sb);
+       rc = server_mgc_set_fs(&mgc_env, lsi->lsi_mgc, sb);
        if (rc)
-               GOTO(out_stop_service, rc);
+               GOTO(out_env, rc);
 
        /* Register with MGS */
        rc = server_register_target(lsi);
@@ -1330,6 +1374,8 @@ static int server_start_targets(struct super_block *sb)
        /* log has been fully processed, let clients connect */
        dev = obd->obd_lu_dev;
        if (dev && dev->ld_ops->ldo_prepare) {
+               struct lu_env env;
+
                rc = lu_env_init(&env, dev->ld_type->ldt_ctx_tags);
                if (rc == 0) {
                        struct lu_context  session_ctx;
@@ -1357,8 +1403,9 @@ static int server_start_targets(struct super_block *sb)
 
 out_mgc:
        /* Release the mgc fs for others to use */
-       server_mgc_clear_fs(lsi->lsi_mgc);
-
+       server_mgc_clear_fs(&mgc_env, lsi->lsi_mgc);
+out_env:
+       lu_env_fini(&mgc_env);
 out_stop_service:
        if (rc != 0)
                server_stop_servers(lsi->lsi_flags);
@@ -1481,9 +1528,13 @@ static void server_put_super(struct super_block *sb)
                   If there are any setup/cleanup errors, save the lov
                   name for safety cleanup later. */
                lprof = class_get_profile(lsi->lsi_svname);
-               if (lprof && lprof->lp_dt) {
-                       OBD_ALLOC(extraname, strlen(lprof->lp_dt) + 1);
-                       strcpy(extraname, lprof->lp_dt);
+               if (lprof != NULL) {
+                       if (lprof->lp_dt != NULL) {
+                               OBD_ALLOC(extraname, strlen(lprof->lp_dt) + 1);
+                               strncpy(extraname, lprof->lp_dt,
+                                       strlen(lprof->lp_dt) + 1);
+                       }
+                       class_put_profile(lprof);
                }
 
                obd = class_name2obd(lsi->lsi_svname);
@@ -1630,7 +1681,7 @@ static const struct inode_operations server_inode_operations = {
 
 static int server_fill_super_common(struct super_block *sb)
 {
-       struct inode *root = 0;
+       struct inode *root = NULL;
        ENTRY;
 
        CDEBUG(D_MOUNT, "Server sb, dev=%d\n", (int)sb->s_dev);
@@ -1705,8 +1756,6 @@ static int osd_start(struct lustre_sb_info *lsi, unsigned long mflags)
        rc = obd_connect(NULL, &lsi->lsi_osd_exp,
                         obd, &obd->obd_uuid, NULL, NULL);
 
-       OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 10);
-
        if (rc) {
                obd->obd_force = 1;
                class_manual_cleanup(obd);
@@ -1739,6 +1788,9 @@ int server_fill_super(struct super_block *sb)
        int rc;
        ENTRY;
 
+       /* to simulate target mount race */
+       OBD_RACE(OBD_FAIL_TGT_MOUNT_RACE);
+
        rc = lsi_prepare(lsi);
        if (rc)
                RETURN(rc);
@@ -1868,4 +1920,3 @@ void server_calc_timeout(struct lustre_sb_info *lsi, struct obd_device *obd)
        obd->obd_recovery_time_hard = hard;
        obd->obd_recovery_ir_factor = factor;
 }
-EXPORT_SYMBOL(server_calc_timeout);