*
* Copyright (c) 2003 Cluster File Systems, Inc.
*
- * This file is part of Lustre, http://www.lustre.org.
+ * This file is part of the Lustre file system, http://www.lustre.org
+ * Lustre is a trademark of Cluster File Systems, Inc.
*
- * Lustre is free software; you can redistribute it and/or
- * modify it under the terms of version 2 of the GNU General Public
- * License as published by the Free Software Foundation.
+ * You may have signed or agreed to another license before downloading
+ * this software. If so, you are bound by the terms and conditions
+ * of that agreement, and the following does not apply to you. See the
+ * LICENSE file included with this distribution for more information.
*
- * Lustre is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
- * GNU General Public License for more details.
+ * If you did not agree to a different license, then this copy of Lustre
+ * is open source software; you can redistribute it and/or modify it
+ * under the terms of version 2 of the GNU General Public License as
+ * published by the Free Software Foundation.
*
- * You should have received a copy of the GNU General Public License
- * along with Lustre; if not, write to the Free Software
- * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ * In either case, Lustre is distributed in the hope that it will be
+ * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
+ * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * license text for more details.
*/
#ifndef EXPORT_SYMTAB
#define DEBUG_SUBSYSTEM S_LDLM
#ifdef __KERNEL__
-# include <linux/module.h>
+# include <libcfs/libcfs.h>
#else
# include <liblustre.h>
#endif
-#include <linux/obd.h>
-#include <linux/obd_ost.h> /* for LUSTRE_OSC_NAME */
-#include <linux/lustre_mds.h> /* for LUSTRE_MDC_NAME */
-#include <linux/lustre_mgmt.h>
-#include <linux/lustre_dlm.h>
-#include <linux/lustre_net.h>
-
-int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
+#include <obd.h>
+#include <lustre_mds.h>
+#include <lustre_dlm.h>
+#include <lustre_net.h>
+#include <lustre_sec.h>
+
+/* @priority: if non-zero, move the selected to the list head
+ * @create: if zero, only search in existed connections
+ */
+static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
+ int priority, int create)
+{
+ struct ptlrpc_connection *ptlrpc_conn;
+ struct obd_import_conn *imp_conn = NULL, *item;
+ int rc = 0;
+ ENTRY;
+
+ if (!create && !priority) {
+ CDEBUG(D_HA, "Nothing to do\n");
+ RETURN(-EINVAL);
+ }
+
+ ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
+ if (!ptlrpc_conn) {
+ CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
+ RETURN (-ENOENT);
+ }
+
+ if (create) {
+ OBD_ALLOC(imp_conn, sizeof(*imp_conn));
+ if (!imp_conn) {
+ GOTO(out_put, rc = -ENOMEM);
+ }
+ }
+
+ spin_lock(&imp->imp_lock);
+ list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
+ if (obd_uuid_equals(uuid, &item->oic_uuid)) {
+ if (priority) {
+ list_del(&item->oic_item);
+ list_add(&item->oic_item, &imp->imp_conn_list);
+ item->oic_last_attempt = 0;
+ }
+ CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
+ imp, imp->imp_obd->obd_name, uuid->uuid,
+ (priority ? ", moved to head" : ""));
+ spin_unlock(&imp->imp_lock);
+ GOTO(out_free, rc = 0);
+ }
+ }
+ /* not found */
+ if (create) {
+ imp_conn->oic_conn = ptlrpc_conn;
+ imp_conn->oic_uuid = *uuid;
+ item->oic_last_attempt = 0;
+ if (priority)
+ list_add(&imp_conn->oic_item, &imp->imp_conn_list);
+ else
+ list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
+ CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
+ imp, imp->imp_obd->obd_name, uuid->uuid,
+ (priority ? "head" : "tail"));
+ } else {
+ spin_unlock(&imp->imp_lock);
+ GOTO(out_free, rc = -ENOENT);
+ }
+
+ spin_unlock(&imp->imp_lock);
+ RETURN(0);
+out_free:
+ if (imp_conn)
+ OBD_FREE(imp_conn, sizeof(*imp_conn));
+out_put:
+ ptlrpc_put_connection(ptlrpc_conn);
+ RETURN(rc);
+}
+
+int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
+{
+ return import_set_conn(imp, uuid, 1, 0);
+}
+
+int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
+ int priority)
+{
+ return import_set_conn(imp, uuid, priority, 1);
+}
+
+int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
+{
+ struct obd_import_conn *imp_conn;
+ struct obd_export *dlmexp;
+ int rc = -ENOENT;
+ ENTRY;
+
+ spin_lock(&imp->imp_lock);
+ if (list_empty(&imp->imp_conn_list)) {
+ LASSERT(!imp->imp_connection);
+ GOTO(out, rc);
+ }
+
+ list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
+ if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
+ continue;
+ LASSERT(imp_conn->oic_conn);
+
+ /* is current conn? */
+ if (imp_conn == imp->imp_conn_current) {
+ LASSERT(imp_conn->oic_conn == imp->imp_connection);
+
+ if (imp->imp_state != LUSTRE_IMP_CLOSED &&
+ imp->imp_state != LUSTRE_IMP_DISCON) {
+ CERROR("can't remove current connection\n");
+ GOTO(out, rc = -EBUSY);
+ }
+
+ ptlrpc_put_connection(imp->imp_connection);
+ imp->imp_connection = NULL;
+
+ dlmexp = class_conn2export(&imp->imp_dlm_handle);
+ if (dlmexp && dlmexp->exp_connection) {
+ LASSERT(dlmexp->exp_connection ==
+ imp_conn->oic_conn);
+ ptlrpc_put_connection(dlmexp->exp_connection);
+ dlmexp->exp_connection = NULL;
+ }
+ }
+
+ list_del(&imp_conn->oic_item);
+ ptlrpc_put_connection(imp_conn->oic_conn);
+ OBD_FREE(imp_conn, sizeof(*imp_conn));
+ CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
+ imp, imp->imp_obd->obd_name, uuid->uuid);
+ rc = 0;
+ break;
+ }
+out:
+ spin_unlock(&imp->imp_lock);
+ if (rc == -ENOENT)
+ CERROR("connection %s not found\n", uuid->uuid);
+ RETURN(rc);
+}
+
+static void destroy_import(struct obd_import *imp)
+{
+ /* drop security policy instance after all rpc finished/aborted
+ * to let all busy credentials be released.
+ */
+ class_import_get(imp);
+ class_destroy_import(imp);
+ sptlrpc_import_put_sec(imp);
+ class_import_put(imp);
+}
+
+/* configure an RPC client OBD device
+ *
+ * lcfg parameters:
+ * 1 - client UUID
+ * 2 - server UUID
+ * 3 - inactive-on-startup
+ */
+int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
{
- struct ptlrpc_connection *conn;
- struct lustre_cfg* lcfg = buf;
struct client_obd *cli = &obddev->u.cli;
struct obd_import *imp;
struct obd_uuid server_uuid;
int rq_portal, rp_portal, connect_op;
char *name = obddev->obd_type->typ_name;
- char *mgmt_name = NULL;
- int rc = 0;
- struct obd_device *mgmt_obd;
- mgmtcli_register_for_events_t register_f;
+ int rc;
ENTRY;
/* In a more perfect world, we would hang a ptlrpc_client off of
rq_portal = MDS_REQUEST_PORTAL;
rp_portal = MDC_REPLY_PORTAL;
connect_op = MDS_CONNECT;
- } else if (!strcmp(name, LUSTRE_MGMTCLI_NAME)) {
- rq_portal = MGMT_REQUEST_PORTAL;
- rp_portal = MGMT_REPLY_PORTAL;
- connect_op = MGMT_CONNECT;
+ } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
+ rq_portal = MGS_REQUEST_PORTAL;
+ rp_portal = MGC_REPLY_PORTAL;
+ connect_op = MGS_CONNECT;
} else {
CERROR("unknown client OBD type \"%s\", can't setup\n",
name);
RETURN(-EINVAL);
}
- if (lcfg->lcfg_inllen1 < 1) {
+ if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
CERROR("requires a TARGET UUID\n");
RETURN(-EINVAL);
}
- if (lcfg->lcfg_inllen1 > 37) {
+ if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
CERROR("client UUID must be less than 38 characters\n");
RETURN(-EINVAL);
}
- if (lcfg->lcfg_inllen2 < 1) {
+ if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
CERROR("setup requires a SERVER UUID\n");
RETURN(-EINVAL);
}
- if (lcfg->lcfg_inllen2 > 37) {
+ if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
CERROR("target UUID must be less than 38 characters\n");
RETURN(-EINVAL);
}
sema_init(&cli->cl_sem, 1);
+ sema_init(&cli->cl_mgc_sem, 1);
+ cli->cl_sec_conf.sfc_rpc_flavor = SPTLRPC_FLVR_NULL;
+ cli->cl_sec_conf.sfc_bulk_csum = BULK_CSUM_ALG_NULL;
+ cli->cl_sec_conf.sfc_bulk_priv = BULK_PRIV_ALG_NULL;
+ cli->cl_sec_conf.sfc_flags = 0;
cli->cl_conn_count = 0;
- memcpy(server_uuid.uuid, lcfg->lcfg_inlbuf2, MIN(lcfg->lcfg_inllen2,
- sizeof(server_uuid)));
+ memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
+ min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
+ sizeof(server_uuid)));
cli->cl_dirty = 0;
- cli->cl_dirty_granted = 0;
- cli->cl_dirty_max = 64*1024*1024; /* some default */
- cli->cl_ost_can_grant = 1;
- INIT_LIST_HEAD(&cli->cl_cache_waiters);
- INIT_LIST_HEAD(&cli->cl_loi_ready_list);
- spin_lock_init(&cli->cl_loi_list_lock);
- cli->cl_brw_in_flight = 0;
+ cli->cl_avail_grant = 0;
+ /* FIXME: should limit this for the sum of all cl_dirty_max */
+ cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
+ if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
+ cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
+ CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
+ CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
+ CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
+ CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
+ client_obd_list_lock_init(&cli->cl_loi_list_lock);
+ cli->cl_r_in_flight = 0;
+ cli->cl_w_in_flight = 0;
+
spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
spin_lock_init(&cli->cl_read_page_hist.oh_lock);
spin_lock_init(&cli->cl_write_page_hist.oh_lock);
- cli->cl_max_pages_per_rpc = PTL_MD_MAX_PAGES;
- cli->cl_max_rpcs_in_flight = 8;
+ spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
+ spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
+#ifdef ENABLE_CHECKSUM
+ cli->cl_checksum = 1;
+#endif
+ atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
+
+ /* This value may be changed at connect time in
+ ptlrpc_connect_interpret. */
+ cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
+ (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
+
+ if (!strcmp(name, LUSTRE_MDC_NAME)) {
+ cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
+ } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
+ cli->cl_max_rpcs_in_flight = 2;
+ } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
+ cli->cl_max_rpcs_in_flight = 3;
+ } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
+ cli->cl_max_rpcs_in_flight = 4;
+ } else {
+ cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
+ }
- ldlm_get_ref();
+ rc = ldlm_get_ref();
if (rc) {
CERROR("ldlm_get_ref failed: %d\n", rc);
GOTO(err, rc);
}
- conn = ptlrpc_uuid_to_connection(&server_uuid);
- if (conn == NULL)
- GOTO(err_ldlm, rc = -ENOENT);
-
ptlrpc_init_client(rq_portal, rp_portal, name,
&obddev->obd_ldlm_client);
- imp = class_new_import();
- if (imp == NULL) {
- ptlrpc_put_connection(conn);
+ imp = class_new_import(obddev);
+ if (imp == NULL)
GOTO(err_ldlm, rc = -ENOENT);
- }
- imp->imp_connection = conn;
imp->imp_client = &obddev->obd_ldlm_client;
- imp->imp_obd = obddev;
imp->imp_connect_op = connect_op;
- imp->imp_generation = 0;
- INIT_LIST_HEAD(&imp->imp_pinger_chain);
- memcpy(imp->imp_target_uuid.uuid, lcfg->lcfg_inlbuf1,
- lcfg->lcfg_inllen1);
+ imp->imp_initial_recov = 1;
+ imp->imp_initial_recov_bk = 0;
+ CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
+ memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
+ LUSTRE_CFG_BUFLEN(lcfg, 1));
class_import_put(imp);
+ rc = client_import_add_conn(imp, &server_uuid, 1);
+ if (rc) {
+ CERROR("can't add initial connection\n");
+ GOTO(err_import, rc);
+ }
+
cli->cl_import = imp;
+ /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
- cli->cl_sandev = to_kdev_t(0);
- if (lcfg->lcfg_inllen3 != 0) {
- if (!strcmp(lcfg->lcfg_inlbuf3, "inactive")) {
+ if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
+ if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
name, obddev->obd_name,
- imp->imp_target_uuid.uuid);
+ cli->cl_target_uuid.uuid);
+ spin_lock(&imp->imp_lock);
imp->imp_invalid = 1;
-
- if (lcfg->lcfg_inllen4 != 0)
- mgmt_name = lcfg->lcfg_inlbuf4;
- } else {
- mgmt_name = lcfg->lcfg_inlbuf3;
+ spin_unlock(&imp->imp_lock);
}
}
- if (mgmt_name != NULL) {
- /* Register with management client if we need to. */
- CDEBUG(D_HA, "%s registering with %s for events about %s\n",
- obddev->obd_name, mgmt_name, server_uuid.uuid);
-
- mgmt_obd = class_name2obd(mgmt_name);
- if (!mgmt_obd) {
- CERROR("can't find mgmtcli %s to register\n",
- mgmt_name);
- GOTO(err_import, rc = -ENOSYS);
- }
-
- register_f = inter_module_get("mgmtcli_register_for_events");
- if (!register_f) {
- CERROR("can't i_m_g mgmtcli_register_for_events\n");
- GOTO(err_import, rc = -ENOSYS);
- }
-
- rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
- inter_module_put("mgmtcli_register_for_events");
-
- if (!rc)
- cli->cl_mgmtcli_obd = mgmt_obd;
- }
+ cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
RETURN(rc);
}
-int client_obd_cleanup(struct obd_device *obddev, int flags)
+int client_obd_cleanup(struct obd_device *obddev)
{
- struct client_obd *cli = &obddev->u.cli;
-
- if (!cli->cl_import)
- RETURN(-EINVAL);
- if (cli->cl_mgmtcli_obd) {
- mgmtcli_deregister_for_events_t dereg_f;
-
- dereg_f = inter_module_get("mgmtcli_deregister_for_events");
- dereg_f(cli->cl_mgmtcli_obd, obddev);
- inter_module_put("mgmtcli_deregister_for_events");
- }
- class_destroy_import(cli->cl_import);
- cli->cl_import = NULL;
-
- ldlm_put_ref(flags & OBD_OPT_FORCE);
-
+ ENTRY;
+ ldlm_put_ref(obddev->obd_force);
RETURN(0);
}
-int client_connect_import(struct lustre_handle *dlm_handle,
- struct obd_device *obd,
- struct obd_uuid *cluuid)
+/* ->o_connect() method for client side (OSC and MDC and MGC) */
+int client_connect_import(const struct lu_env *env,
+ struct lustre_handle *dlm_handle,
+ struct obd_device *obd, struct obd_uuid *cluuid,
+ struct obd_connect_data *data)
{
struct client_obd *cli = &obd->u.cli;
struct obd_import *imp = cli->cl_import;
struct obd_export *exp;
+ struct obd_connect_data *ocd;
int rc;
ENTRY;
- down(&cli->cl_sem);
+ mutex_down(&cli->cl_sem);
rc = class_connect(dlm_handle, obd, cluuid);
if (rc)
GOTO(out_sem, rc);
if (obd->obd_namespace != NULL)
CERROR("already have namespace!\n");
obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
- LDLM_NAMESPACE_CLIENT);
+ LDLM_NAMESPACE_CLIENT,
+ LDLM_NAMESPACE_GREEDY);
if (obd->obd_namespace == NULL)
GOTO(out_disco, rc = -ENOMEM);
imp->imp_dlm_handle = *dlm_handle;
- imp->imp_state = LUSTRE_IMP_DISCON;
+ rc = ptlrpc_init_import(imp);
+ if (rc != 0)
+ GOTO(out_ldlm, rc);
+
+ rc = sptlrpc_import_get_sec(imp, NULL, cli->cl_sec_conf.sfc_rpc_flavor,
+ cli->cl_sec_conf.sfc_flags);
+ if (rc)
+ GOTO(out_ldlm, rc);
- rc = ptlrpc_connect_import(imp);
+ ocd = &imp->imp_connect_data;
+ if (data) {
+ *ocd = *data;
+ imp->imp_connect_flags_orig = data->ocd_connect_flags;
+ }
+
+ rc = ptlrpc_connect_import(imp, NULL);
if (rc != 0) {
LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
GOTO(out_ldlm, rc);
}
+ LASSERT(exp->exp_connection);
- LASSERT (imp->imp_state == LUSTRE_IMP_FULL);
-
- exp->exp_connection = ptlrpc_connection_addref(imp->imp_connection);
-
- if (imp->imp_replayable) {
- CDEBUG(D_HA, "connected to replayable target: %s\n",
- imp->imp_target_uuid.uuid);
- ptlrpc_pinger_add_import(imp);
+ if (data) {
+ LASSERT((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
+ ocd->ocd_connect_flags);
+ data->ocd_connect_flags = ocd->ocd_connect_flags;
}
- CDEBUG(D_HA, "local import: %p, remote handle: "LPX64"\n", imp,
- imp->imp_remote_handle.cookie);
+ ptlrpc_pinger_add_import(imp);
EXIT;
obd->obd_namespace = NULL;
out_disco:
cli->cl_conn_count--;
- class_disconnect(exp, 0);
+ class_disconnect(exp);
} else {
class_export_put(exp);
}
out_sem:
- up(&cli->cl_sem);
+ mutex_up(&cli->cl_sem);
return rc;
}
-int client_disconnect_export(struct obd_export *exp, int failover)
+int client_disconnect_export(struct obd_export *exp)
{
struct obd_device *obd = class_exp2obd(exp);
- struct client_obd *cli = &obd->u.cli;
- struct obd_import *imp = cli->cl_import;
+ struct client_obd *cli;
+ struct obd_import *imp;
int rc = 0, err;
ENTRY;
RETURN(-EINVAL);
}
- down(&cli->cl_sem);
+ cli = &obd->u.cli;
+ imp = cli->cl_import;
+
+ mutex_down(&cli->cl_sem);
if (!cli->cl_conn_count) {
CERROR("disconnecting disconnected device (%s)\n",
obd->obd_name);
if (cli->cl_conn_count)
GOTO(out_no_disconnect, rc = 0);
+ /* Mark import deactivated now, so we don't try to reconnect if any
+ * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
+ * fully deactivate the import, or that would drop all requests. */
+ spin_lock(&imp->imp_lock);
+ imp->imp_deactive = 1;
+ spin_unlock(&imp->imp_lock);
+
/* Some non-replayable imports (MDS's OSCs) are pinged, so just
* delete it regardless. (It's safe to delete an import that was
* never added.) */
(void)ptlrpc_pinger_del_import(imp);
if (obd->obd_namespace != NULL) {
- /* obd_no_recov == local only */
+ /* obd_force == local only */
ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
- obd->obd_no_recov, NULL);
- ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
+ obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
+ NULL);
+ ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
obd->obd_namespace = NULL;
}
- /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
- if (obd->obd_no_recov)
- ptlrpc_set_import_active(imp, 0);
- else
- rc = ptlrpc_disconnect_import(imp);
+ if (!obd->obd_force)
+ rc = ptlrpc_disconnect_import(imp, 0);
- imp->imp_state = LUSTRE_IMP_NEW;
+ ptlrpc_invalidate_import(imp);
+ ptlrpc_free_rq_pool(imp->imp_rq_pool);
+ destroy_import(imp);
+ cli->cl_import = NULL;
EXIT;
out_no_disconnect:
- err = class_disconnect(exp, 0);
+ err = class_disconnect(exp);
if (!rc && err)
rc = err;
out_sem:
- up(&cli->cl_sem);
+ mutex_up(&cli->cl_sem);
RETURN(rc);
}
* -------------------------------------------------------------------------- */
int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
- struct obd_uuid *cluuid)
+ struct obd_uuid *cluuid, int initial_conn)
{
- if (exp->exp_connection) {
+ ENTRY;
+ if (exp->exp_connection && exp->exp_imp_reverse && !initial_conn) {
struct lustre_handle *hdl;
hdl = &exp->exp_imp_reverse->imp_remote_handle;
/* Might be a re-connect after a partition. */
if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
- CERROR("%s reconnecting\n", cluuid->uuid);
+ CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
+ cluuid->uuid);
conn->cookie = exp->exp_handle.h_cookie;
+ /* target_handle_connect() treats EALREADY and
+ * -EALREADY differently. EALREADY means we are
+ * doing a valid reconnect from the same client. */
RETURN(EALREADY);
} else {
CERROR("%s reconnecting from %s, "
exp->exp_connection->c_remote_uuid.uuid,
hdl->cookie, conn->cookie);
memset(conn, 0, sizeof *conn);
+ /* target_handle_connect() treats EALREADY and
+ * -EALREADY differently. -EALREADY is an error
+ * (same UUID, different handle). */
RETURN(-EALREADY);
}
}
conn->cookie = exp->exp_handle.h_cookie;
- CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
- cluuid->uuid, exp);
- CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
+ CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
+ cluuid->uuid, exp, conn->cookie);
RETURN(0);
}
-int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
+void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
+ int error)
+{
+ struct obd_export *exp = cb_data;
+
+ CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
+ obd->obd_name, exp->exp_client_uuid.uuid);
+
+ spin_lock(&exp->exp_lock);
+ exp->exp_need_sync = 0;
+ spin_unlock(&exp->exp_lock);
+}
+EXPORT_SYMBOL(target_client_add_cb);
+
+int target_handle_connect(struct ptlrpc_request *req)
{
- struct obd_device *target;
+ struct obd_device *target, *targref = NULL;
struct obd_export *export = NULL;
struct obd_import *revimp;
struct lustre_handle conn;
struct obd_uuid remote_uuid;
struct list_head *p;
char *str, *tmp;
- int rc = 0, abort_recovery;
+ int rc = 0;
+ int initial_conn = 0;
+ struct obd_connect_data *data;
+ int size[2] = { sizeof(struct ptlrpc_body), sizeof(*data) };
ENTRY;
- LASSERT_REQSWAB (req, 0);
- str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
+ OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
+
+ lustre_set_req_swabbed(req, REQ_REC_OFF);
+ str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF, sizeof(tgtuuid)-1);
if (str == NULL) {
- CERROR("bad target UUID for connect\n");
+ DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
GOTO(out, rc = -EINVAL);
}
obd_str2uuid (&tgtuuid, str);
target = class_uuid2obd(&tgtuuid);
- if (!target) {
+ if (!target)
target = class_name2obd(str);
- }
if (!target || target->obd_stopping || !target->obd_set_up) {
- CERROR("UUID '%s' is not available for connect\n", str);
+ LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
+ " for connect (%s)\n", str,
+ !target ? "no target" :
+ (target->obd_stopping ? "stopping" :
+ "not set up"));
GOTO(out, rc = -ENODEV);
}
- LASSERT_REQSWAB (req, 1);
- str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
+ if (target->obd_no_conn) {
+ LCONSOLE_WARN("%s: temporarily refusing client connection "
+ "from %s\n", target->obd_name,
+ libcfs_nid2str(req->rq_peer.nid));
+ GOTO(out, rc = -EAGAIN);
+ }
+
+ /* Make sure the target isn't cleaned up while we're here. Yes,
+ there's still a race between the above check and our incref here.
+ Really, class_uuid2obd should take the ref. */
+ targref = class_incref(target);
+
+ lustre_set_req_swabbed(req, REQ_REC_OFF + 1);
+ str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF + 1,
+ sizeof(cluuid) - 1);
if (str == NULL) {
- CERROR("bad client UUID for connect\n");
+ DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
GOTO(out, rc = -EINVAL);
}
obd_str2uuid (&cluuid, str);
/* XXX extract a nettype and format accordingly */
- snprintf(remote_uuid.uuid, sizeof remote_uuid,
- "NET_"LPX64"_UUID", req->rq_peer.peer_nid);
-
- spin_lock_bh(&target->obd_processing_task_lock);
- abort_recovery = target->obd_abort_recovery;
- spin_unlock_bh(&target->obd_processing_task_lock);
- if (abort_recovery)
- target_abort_recovery(target);
+ switch (sizeof(lnet_nid_t)) {
+ /* NB the casts only avoid compiler warnings */
+ case 8:
+ snprintf(remote_uuid.uuid, sizeof remote_uuid,
+ "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
+ break;
+ case 4:
+ snprintf(remote_uuid.uuid, sizeof remote_uuid,
+ "NET_%x_UUID", (__u32)req->rq_peer.nid);
+ break;
+ default:
+ LBUG();
+ }
- tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
+ tmp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn);
if (tmp == NULL)
GOTO(out, rc = -EPROTO);
memcpy(&conn, tmp, sizeof conn);
- rc = lustre_pack_reply(req, 0, NULL, NULL);
+ data = lustre_swab_reqbuf(req, REQ_REC_OFF + 3, sizeof(*data),
+ lustre_swab_connect);
+
+ if (!data)
+ GOTO(out, rc = -EPROTO);
+
+ rc = lustre_pack_reply(req, 2, size, NULL);
if (rc)
GOTO(out, rc);
+ if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
+ if (!data) {
+ DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
+ "libclient connection attempt");
+ GOTO(out, rc = -EPROTO);
+ } else if (data->ocd_version < LUSTRE_VERSION_CODE -
+ LUSTRE_VERSION_ALLOWED_OFFSET ||
+ data->ocd_version > LUSTRE_VERSION_CODE +
+ LUSTRE_VERSION_ALLOWED_OFFSET) {
+ DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
+ "libclient connection attempt",
+ data->ocd_version < LUSTRE_VERSION_CODE ?
+ "old" : "new",
+ OBD_OCD_VERSION_MAJOR(data->ocd_version),
+ OBD_OCD_VERSION_MINOR(data->ocd_version),
+ OBD_OCD_VERSION_PATCH(data->ocd_version),
+ OBD_OCD_VERSION_FIX(data->ocd_version));
+ data = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
+ offsetof(typeof(*data),
+ ocd_version) +
+ sizeof(data->ocd_version));
+ if (data) {
+ data->ocd_connect_flags = OBD_CONNECT_VERSION;
+ data->ocd_version = LUSTRE_VERSION_CODE;
+ }
+ GOTO(out, rc = -EPROTO);
+ }
+ }
+
+ if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
+ initial_conn = 1;
+
/* lctl gets a backstage, all-access pass. */
if (obd_uuid_equals(&cluuid, &target->obd_uuid))
goto dont_check_exports;
list_for_each(p, &target->obd_exports) {
export = list_entry(p, struct obd_export, exp_obd_chain);
if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
+ if (export->exp_connecting) { /* bug 9635, et. al. */
+ CWARN("%s: exp %p already connecting\n",
+ export->exp_obd->obd_name, export);
+ export = NULL;
+ rc = -EALREADY;
+ break;
+ }
+
+ /* make darn sure this is coming from the same peer
+ * if the UUIDs matched */
+ if ((export->exp_connection != NULL) &&
+ (strcmp(libcfs_nid2str(req->rq_peer.nid),
+ libcfs_nid2str(export->exp_connection->c_peer.nid)))) {
+ CWARN("%s: cookie %s seen on new NID %s when "
+ "existing NID %s is already connected\n",
+ target->obd_name, cluuid.uuid,
+ libcfs_nid2str(req->rq_peer.nid),
+ libcfs_nid2str(export->exp_connection->c_peer.nid));
+ export = NULL;
+ rc = -EALREADY;
+ break;
+ }
+
+ spin_lock(&export->exp_lock);
+ export->exp_connecting = 1;
+ spin_unlock(&export->exp_lock);
spin_unlock(&target->obd_dev_lock);
LASSERT(export->exp_obd == target);
- rc = target_handle_reconnect(&conn, export, &cluuid);
+ rc = target_handle_reconnect(&conn, export, &cluuid,
+ initial_conn);
break;
}
export = NULL;
}
+
/* If we found an export, we already unlocked. */
if (!export) {
spin_unlock(&target->obd_dev_lock);
- } else if (req->rq_reqmsg->conn_cnt == 1) {
- CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
- cluuid.uuid);
+ OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
+ } else if (req->rq_export == NULL &&
+ atomic_read(&export->exp_rpc_count) > 0) {
+ CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
+ target->obd_name, cluuid.uuid,
+ libcfs_nid2str(req->rq_peer.nid),
+ export, atomic_read(&export->exp_refcount));
+ GOTO(out, rc = -EBUSY);
+ } else if (req->rq_export != NULL &&
+ (atomic_read(&export->exp_rpc_count) > 1)) {
+ CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
+ target->obd_name, cluuid.uuid,
+ libcfs_nid2str(req->rq_peer.nid),
+ export, atomic_read(&export->exp_rpc_count));
+ GOTO(out, rc = -EBUSY);
+ } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
+ !initial_conn) {
+ CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
+ "cookies not random?\n", target->obd_name,
+ libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
GOTO(out, rc = -EALREADY);
+ } else {
+ OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
+ if (req->rq_export == NULL && initial_conn)
+ export->exp_last_request_time =
+ max(export->exp_last_request_time,
+ (time_t)CURRENT_SECONDS);
}
+ /* We want to handle EALREADY but *not* -EALREADY from
+ * target_handle_reconnect(), return reconnection state in a flag */
+ if (rc == EALREADY) {
+ lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
+ rc = 0;
+ } else if (rc) {
+ GOTO(out, rc);
+ }
/* Tell the client if we're in recovery. */
/* If this is the first client, start the recovery timer */
+ CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
+ target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
+ target->obd_recovering ? "recovering/" : "", data->ocd_transno,
+ export, (long)CURRENT_SECONDS,
+ export ? (long)export->exp_last_request_time : 0);
+
+
if (target->obd_recovering) {
lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
- target_start_recovery_timer(target, handler);
+ target_start_recovery_timer(target);
}
/* Tell the client if we support replayable requests */
if (export == NULL) {
if (target->obd_recovering) {
- CERROR("denying connection for new client %s: "
- "%d clients in recovery for %lds\n", cluuid.uuid,
+ CERROR("%s: denying connection for new client %s (%s): "
+ "%d clients in recovery for %lds\n",
+ target->obd_name,
+ libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
target->obd_recoverable_clients,
- (target->obd_recovery_timer.expires-jiffies)/HZ);
+ cfs_duration_sec(cfs_time_sub(cfs_timer_deadline(&target->obd_recovery_timer),
+ cfs_time_current())));
rc = -EBUSY;
} else {
- dont_check_exports:
- rc = obd_connect(&conn, target, &cluuid);
+dont_check_exports:
+ rc = obd_connect(req->rq_svc_thread->t_env,
+ &conn, target, &cluuid, data);
}
+ } else {
+ rc = obd_reconnect(export, target, &cluuid, data);
}
+ if (rc)
+ GOTO(out, rc);
+ /* Return only the parts of obd_connect_data that we understand, so the
+ * client knows that we don't understand the rest. */
+ if (data)
+ memcpy(lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
+ sizeof(*data)),
+ data, sizeof(*data));
/* If all else goes well, this is our RPC return code. */
req->rq_status = 0;
- if (rc && rc != EALREADY)
- GOTO(out, rc);
+ lustre_msg_set_handle(req->rq_repmsg, &conn);
- /* XXX track this all the time? */
- if (target->obd_recovering) {
- target->obd_connected_clients++;
+ /* ownership of this export ref transfers to the request AFTER we
+ * drop any previous reference the request had, but we don't want
+ * that to go to zero before we get our new export reference. */
+ export = class_conn2export(&conn);
+ if (!export) {
+ DEBUG_REQ(D_ERROR, req, "Missing export!");
+ GOTO(out, rc = -ENODEV);
}
- req->rq_repmsg->handle = conn;
-
/* If the client and the server are the same node, we will already
* have an export that really points to the client's DLM export,
* because we have a shared handles table.
if (req->rq_export != NULL)
class_export_put(req->rq_export);
- /* ownership of this export ref transfers to the request */
- export = req->rq_export = class_conn2export(&conn);
- LASSERT(export != NULL);
+ req->rq_export = export;
+
+ spin_lock(&export->exp_lock);
+ if (initial_conn) {
+ lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
+ } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
+ spin_unlock(&export->exp_lock);
+ CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
+ cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
+ export->exp_conn_cnt,
+ lustre_msg_get_conn_cnt(req->rq_reqmsg));
+
+ GOTO(out, rc = -EALREADY);
+ }
+ export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
+
+ /* request from liblustre? Don't evict it for not pinging. */
+ if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
+ export->exp_libclient = 1;
+ spin_unlock(&export->exp_lock);
+
+ spin_lock(&target->obd_dev_lock);
+ list_del_init(&export->exp_obd_chain_timed);
+ spin_unlock(&target->obd_dev_lock);
+ } else {
+ spin_unlock(&export->exp_lock);
+ }
- if (req->rq_connection != NULL)
- ptlrpc_put_connection(req->rq_connection);
if (export->exp_connection != NULL)
ptlrpc_put_connection(export->exp_connection);
- export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
+ export->exp_connection = ptlrpc_get_connection(req->rq_peer,
+ req->rq_self,
&remote_uuid);
- req->rq_connection = ptlrpc_connection_addref(export->exp_connection);
- LASSERT(export->exp_conn_cnt < req->rq_reqmsg->conn_cnt);
- export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
+ spin_lock_bh(&target->obd_processing_task_lock);
+ if (target->obd_recovering && !export->exp_in_recovery) {
+ spin_lock(&export->exp_lock);
+ export->exp_in_recovery = 1;
+ export->exp_req_replay_needed = 1;
+ export->exp_lock_replay_needed = 1;
+ spin_unlock(&export->exp_lock);
+ if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
+ && data->ocd_transno < target->obd_next_recovery_transno)
+ target->obd_next_recovery_transno = data->ocd_transno;
+ target->obd_connected_clients++;
+ /* each connected client is counted as recoverable */
+ target->obd_recoverable_clients++;
+ atomic_inc(&target->obd_req_replay_clients);
+ atomic_inc(&target->obd_lock_replay_clients);
+ if (target->obd_connected_clients ==
+ target->obd_max_recoverable_clients)
+ wake_up(&target->obd_next_transno_waitq);
+ }
+ spin_unlock_bh(&target->obd_processing_task_lock);
+ memcpy(&conn,
+ lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn),
+ sizeof conn);
- if (rc == EALREADY) {
- /* We indicate the reconnection in a flag, not an error code. */
- lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
- GOTO(out, rc = 0);
+ if (export->exp_imp_reverse != NULL) {
+ /* destroyed import can be still referenced in ctxt */
+ obd_set_info_async(export, strlen(KEY_REVIMP_UPD),
+ KEY_REVIMP_UPD, 0, NULL, NULL);
+ destroy_import(export->exp_imp_reverse);
}
- memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
- sizeof conn);
+ /* for the rest part, we return -ENOTCONN in case of errors
+ * in order to let client initialize connection again.
+ */
+ revimp = export->exp_imp_reverse = class_new_import(target);
+ if (!revimp) {
+ CERROR("fail to alloc new reverse import.\n");
+ GOTO(out, rc = -ENOTCONN);
+ }
- if (export->exp_imp_reverse != NULL)
- class_destroy_import(export->exp_imp_reverse);
- revimp = export->exp_imp_reverse = class_new_import();
- revimp->imp_connection = ptlrpc_connection_addref(req->rq_connection);
+ revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
revimp->imp_client = &export->exp_obd->obd_ldlm_client;
revimp->imp_remote_handle = conn;
- revimp->imp_obd = target;
revimp->imp_dlm_fake = 1;
revimp->imp_state = LUSTRE_IMP_FULL;
+
+ if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_NEXT_VER) {
+ revimp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
+ lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_NEXT_VER);
+ }
+
+ rc = sptlrpc_import_get_sec(revimp, req->rq_svc_ctx,
+ req->rq_sec_flavor, 0);
+ if (rc) {
+ CERROR("Failed to get sec for reverse import: %d\n", rc);
+ export->exp_imp_reverse = NULL;
+ class_destroy_import(revimp);
+ }
+
class_import_put(revimp);
out:
+ if (export) {
+ spin_lock(&export->exp_lock);
+ export->exp_connecting = 0;
+ spin_unlock(&export->exp_lock);
+ }
+ if (targref)
+ class_decref(targref);
if (rc)
req->rq_status = rc;
RETURN(rc);
int rc;
ENTRY;
- rc = lustre_pack_reply(req, 0, NULL, NULL);
+ rc = lustre_pack_reply(req, 1, NULL, NULL);
if (rc)
RETURN(rc);
- req->rq_status = obd_disconnect(req->rq_export, 0);
- req->rq_export = NULL;
+ /* keep the rq_export around so we can send the reply */
+ req->rq_status = obd_disconnect(class_export_get(req->rq_export));
+
RETURN(0);
}
/* exports created from last_rcvd data, and "fake"
exports created by lctl don't have an import */
if (exp->exp_imp_reverse != NULL)
- class_destroy_import(exp->exp_imp_reverse);
+ destroy_import(exp->exp_imp_reverse);
/* We cancel locks at disconnect time, but this will catch any locks
* granted in a race with recovery-induced disconnect. */
- ldlm_cancel_locks_for_export(exp);
+ if (exp->exp_obd->obd_namespace != NULL)
+ ldlm_cancel_locks_for_export(exp);
}
/*
* Recovery functions
*/
-void target_cancel_recovery_timer(struct obd_device *obd)
+struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
{
- del_timer(&obd->obd_recovery_timer);
+ struct ptlrpc_request *copy_req;
+ struct lustre_msg *copy_reqmsg;
+ struct ptlrpc_user_desc *udesc = NULL;
+
+ OBD_ALLOC_PTR(copy_req);
+ if (!copy_req)
+ return NULL;
+ OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
+ if (!copy_reqmsg){
+ OBD_FREE_PTR(copy_req);
+ return NULL;
+ }
+
+ if (orig_req->rq_user_desc) {
+ int ngroups = orig_req->rq_user_desc->pud_ngroups;
+
+ OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
+ if (!udesc) {
+ OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
+ OBD_FREE_PTR(copy_req);
+ return NULL;
+ }
+ memcpy(udesc, orig_req->rq_user_desc,
+ sptlrpc_user_desc_size(ngroups));
+ }
+
+ *copy_req = *orig_req;
+ memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
+ copy_req->rq_reqmsg = copy_reqmsg;
+ copy_req->rq_user_desc = udesc;
+
+ class_export_get(copy_req->rq_export);
+ CFS_INIT_LIST_HEAD(©_req->rq_list);
+ sptlrpc_svc_ctx_addref(copy_req);
+
+ if (copy_req->rq_reply_state) {
+ /* the copied req takes over the reply state */
+ orig_req->rq_reply_state = NULL;
+ /* to catch further access */
+ orig_req->rq_repmsg = NULL;
+ orig_req->rq_replen = 0;
+ }
+
+ return copy_req;
}
-static void abort_delayed_replies(struct obd_device *obd)
+void ptlrpc_free_clone( struct ptlrpc_request *req)
{
- struct ptlrpc_request *req;
- struct list_head *tmp, *n;
- list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
- req = list_entry(tmp, struct ptlrpc_request, rq_list);
- DEBUG_REQ(D_ERROR, req, "aborted:");
+ if (req->rq_reply_state) {
+ ptlrpc_rs_decref(req->rq_reply_state);
+ req->rq_reply_state = NULL;
+ }
+
+ sptlrpc_svc_ctx_decref(req);
+ class_export_put(req->rq_export);
+ list_del(&req->rq_list);
+
+ if (req->rq_user_desc) {
+ int ngroups = req->rq_user_desc->pud_ngroups;
+ OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
+ }
+ OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
+ OBD_FREE_PTR(req);
+}
+
+#ifdef __KERNEL__
+static void target_finish_recovery(struct obd_device *obd)
+{
+ ENTRY;
+ LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
+ obd->obd_name);
+
+ ldlm_reprocess_all_ns(obd->obd_namespace);
+
+ /* when recovery finished, cleanup orphans on mds and ost */
+ if (OBT(obd) && OBP(obd, postrecov)) {
+ int rc = OBP(obd, postrecov)(obd);
+ LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
+ rc < 0 ? "failed" : "complete", rc);
+ }
+
+ obd->obd_recovery_end = CURRENT_SECONDS;
+ EXIT;
+}
+
+static void abort_req_replay_queue(struct obd_device *obd)
+{
+ struct ptlrpc_request *req, *n;
+
+ list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
+ DEBUG_REQ(D_WARNING, req, "aborted:");
req->rq_status = -ENOTCONN;
- req->rq_type = PTL_RPC_MSG_ERR;
- ptlrpc_reply(req);
- list_del(&req->rq_list);
- OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
- OBD_FREE(req, sizeof *req);
+ if (ptlrpc_error(req)) {
+ DEBUG_REQ(D_ERROR, req,
+ "failed abort_req_reply; skipping");
+ }
+ ptlrpc_free_clone(req);
}
}
-static void abort_recovery_queue(struct obd_device *obd)
+static void abort_lock_replay_queue(struct obd_device *obd)
{
- struct ptlrpc_request *req;
- struct list_head *tmp, *n;
- int rc;
+ struct ptlrpc_request *req, *n;
- list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
- req = list_entry(tmp, struct ptlrpc_request, rq_list);
+ list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
DEBUG_REQ(D_ERROR, req, "aborted:");
req->rq_status = -ENOTCONN;
- req->rq_type = PTL_RPC_MSG_ERR;
- rc = lustre_pack_reply(req, 0, NULL, NULL);
- if (rc == 0) {
- ptlrpc_reply(req);
- } else {
+ if (ptlrpc_error(req)) {
DEBUG_REQ(D_ERROR, req,
- "packing failed for abort-reply; skipping");
+ "failed abort_lock_reply; skipping");
}
- list_del(&req->rq_list);
- class_export_put(req->rq_export);
- OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
- OBD_FREE(req, sizeof *req);
+ ptlrpc_free_clone(req);
}
}
+#endif
-void target_abort_recovery(void *data)
+/* Called from a cleanup function if the device is being cleaned up
+ forcefully. The exports should all have been disconnected already,
+ the only thing left to do is
+ - clear the recovery flags
+ - cancel the timer
+ - free queued requests and replies, but don't send replies
+ Because the obd_stopping flag is set, no new requests should be received.
+
+*/
+void target_cleanup_recovery(struct obd_device *obd)
{
- struct obd_device *obd = data;
- int rc;
+ struct ptlrpc_request *req, *n;
+ ENTRY;
+
+ LASSERT(obd->obd_stopping);
- CERROR("disconnecting clients and aborting recovery\n");
spin_lock_bh(&obd->obd_processing_task_lock);
if (!obd->obd_recovering) {
spin_unlock_bh(&obd->obd_processing_task_lock);
EXIT;
return;
}
-
obd->obd_recovering = obd->obd_abort_recovery = 0;
-
- wake_up(&obd->obd_next_transno_waitq);
target_cancel_recovery_timer(obd);
spin_unlock_bh(&obd->obd_processing_task_lock);
- class_disconnect_exports(obd, 0);
-
- /* when recovery was abort, cleanup orphans for mds */
- if (OBT(obd) && OBP(obd, postrecov)) {
- rc = OBP(obd, postrecov)(obd);
- if (rc >= 0)
- CWARN("Cleanup %d orphans after recovery was aborted\n", rc);
- else
- CERROR("postrecov failed %d\n", rc);
+ list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
+ LASSERT (req->rq_reply_state == 0);
+ ptlrpc_free_clone(req);
+ }
+ list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
+ LASSERT (req->rq_reply_state == 0);
+ ptlrpc_free_clone(req);
+ }
+ list_for_each_entry_safe(req, n, &obd->obd_final_req_queue, rq_list) {
+ LASSERT (req->rq_reply_state == 0);
+ ptlrpc_free_clone(req);
}
- abort_delayed_replies(obd);
- abort_recovery_queue(obd);
- ptlrpc_run_recovery_over_upcall(obd);
+ EXIT;
}
static void target_recovery_expired(unsigned long castmeharder)
{
struct obd_device *obd = (struct obd_device *)castmeharder;
- CERROR("recovery timed out, aborting\n");
+ CERROR("%s: recovery timed out, aborting\n", obd->obd_name);
spin_lock_bh(&obd->obd_processing_task_lock);
- obd->obd_abort_recovery = 1;
- wake_up(&obd->obd_next_transno_waitq);
+ if (obd->obd_recovering)
+ obd->obd_abort_recovery = 1;
+ cfs_waitq_signal(&obd->obd_next_transno_waitq);
spin_unlock_bh(&obd->obd_processing_task_lock);
}
-static void reset_recovery_timer(struct obd_device *obd)
+
+/* obd_processing_task_lock should be held */
+void target_cancel_recovery_timer(struct obd_device *obd)
{
- int recovering;
- spin_lock(&obd->obd_dev_lock);
- recovering = obd->obd_recovering;
- spin_unlock(&obd->obd_dev_lock);
+ CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
+ cfs_timer_disarm(&obd->obd_recovery_timer);
+}
- if (!recovering)
+static void reset_recovery_timer(struct obd_device *obd)
+{
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ if (!obd->obd_recovering) {
+ spin_unlock_bh(&obd->obd_processing_task_lock);
return;
- CDEBUG(D_HA, "timer will expire in %u seconds\n",
- OBD_RECOVERY_TIMEOUT / HZ);
- mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
+ }
+ cfs_timer_arm(&obd->obd_recovery_timer,
+ cfs_time_shift(OBD_RECOVERY_TIMEOUT));
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ CDEBUG(D_HA, "%s: timer will expire in %u seconds\n", obd->obd_name,
+ OBD_RECOVERY_TIMEOUT);
+ /* Only used for lprocfs_status */
+ obd->obd_recovery_end = CURRENT_SECONDS + OBD_RECOVERY_TIMEOUT;
}
/* Only start it the first time called */
-void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler)
+void target_start_recovery_timer(struct obd_device *obd)
{
spin_lock_bh(&obd->obd_processing_task_lock);
- if (obd->obd_recovery_handler) {
+ if (obd->obd_recovery_handler
+ || timer_pending((struct timer_list *)&obd->obd_recovery_timer)) {
spin_unlock_bh(&obd->obd_processing_task_lock);
return;
}
CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
- OBD_RECOVERY_TIMEOUT / HZ);
- obd->obd_recovery_handler = handler;
- obd->obd_recovery_timer.function = target_recovery_expired;
- obd->obd_recovery_timer.data = (unsigned long)obd;
- init_timer(&obd->obd_recovery_timer);
+ OBD_RECOVERY_TIMEOUT);
+ cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
spin_unlock_bh(&obd->obd_processing_task_lock);
reset_recovery_timer(obd);
}
+#ifdef __KERNEL__
static int check_for_next_transno(struct obd_device *obd)
{
- struct ptlrpc_request *req;
+ struct ptlrpc_request *req = NULL;
int wake_up = 0, connected, completed, queue_len, max;
__u64 next_transno, req_transno;
-
+ ENTRY;
spin_lock_bh(&obd->obd_processing_task_lock);
- req = list_entry(obd->obd_recovery_queue.next,
- struct ptlrpc_request, rq_list);
+
+ if (!list_empty(&obd->obd_req_replay_queue)) {
+ req = list_entry(obd->obd_req_replay_queue.next,
+ struct ptlrpc_request, rq_list);
+ req_transno = lustre_msg_get_transno(req->rq_reqmsg);
+ } else {
+ req_transno = 0;
+ }
+
max = obd->obd_max_recoverable_clients;
- req_transno = req->rq_reqmsg->transno;
connected = obd->obd_connected_clients;
- completed = max - obd->obd_recoverable_clients;
+ completed = connected - obd->obd_recoverable_clients;
queue_len = obd->obd_requests_queued_for_recovery;
next_transno = obd->obd_next_recovery_transno;
+ CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
+ "req_transno: "LPU64", next_transno: "LPU64"\n",
+ max, connected, completed, queue_len, req_transno, next_transno);
+
if (obd->obd_abort_recovery) {
CDEBUG(D_HA, "waking for aborted recovery\n");
wake_up = 1;
- } else if (!obd->obd_recovering) {
- CDEBUG(D_HA, "waking for completed recovery (?)\n");
+ } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
+ CDEBUG(D_HA, "waking for completed recovery\n");
wake_up = 1;
} else if (req_transno == next_transno) {
CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
wake_up = 1;
} else if (queue_len + completed == max) {
- CDEBUG(D_ERROR,
+ /* handle gaps occured due to lost reply. It is allowed gaps
+ * because all clients are connected and there will be resend
+ * for missed transaction */
+ LASSERTF(req_transno >= next_transno,
+ "req_transno: "LPU64", next_transno: "LPU64"\n",
+ req_transno, next_transno);
+
+ CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
"waking for skipped transno (skip: "LPD64
", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
- next_transno, queue_len, completed, max, req_transno);
+ next_transno, queue_len, completed, connected, req_transno);
obd->obd_next_recovery_transno = req_transno;
wake_up = 1;
+ } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
+ /* some clients haven't connected in time, but we can try
+ * to replay requests that demand on already committed ones
+ * also, we can replay first non-committed transation */
+ LASSERT(req_transno != 0);
+ if (req_transno == obd->obd_last_committed + 1) {
+ obd->obd_next_recovery_transno = req_transno;
+ } else if (req_transno > obd->obd_last_committed) {
+ /* can't continue recovery: have no needed transno */
+ obd->obd_abort_recovery = 1;
+ CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
+ "connected: %d, completed: %d, queue_len: %d, "
+ "req_transno: "LPU64", next_transno: "LPU64"\n",
+ max, connected, completed, queue_len,
+ req_transno, next_transno);
+ }
+ wake_up = 1;
}
+
spin_unlock_bh(&obd->obd_processing_task_lock);
- LASSERT(req->rq_reqmsg->transno >= next_transno);
return wake_up;
}
-static void process_recovery_queue(struct obd_device *obd)
+static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
{
- struct ptlrpc_request *req;
- int abort_recovery = 0;
struct l_wait_info lwi = { 0 };
- ENTRY;
+ struct ptlrpc_request *req;
- for (;;) {
- spin_lock_bh(&obd->obd_processing_task_lock);
- LASSERT(obd->obd_processing_task == current->pid);
- req = list_entry(obd->obd_recovery_queue.next,
- struct ptlrpc_request, rq_list);
+ CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
+ obd->obd_next_recovery_transno);
+ l_wait_event(obd->obd_next_transno_waitq,
+ check_for_next_transno(obd), &lwi);
- if (req->rq_reqmsg->transno != obd->obd_next_recovery_transno) {
- spin_unlock_bh(&obd->obd_processing_task_lock);
- CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
- LPD64")\n",
- obd->obd_next_recovery_transno,
- req->rq_reqmsg->transno);
- l_wait_event(obd->obd_next_transno_waitq,
- check_for_next_transno(obd), &lwi);
- spin_lock_bh(&obd->obd_processing_task_lock);
- abort_recovery = obd->obd_abort_recovery;
- spin_unlock_bh(&obd->obd_processing_task_lock);
- if (abort_recovery) {
- target_abort_recovery(obd);
- return;
- }
- continue;
- }
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ if (obd->obd_abort_recovery) {
+ req = NULL;
+ } else if (!list_empty(&obd->obd_req_replay_queue)) {
+ req = list_entry(obd->obd_req_replay_queue.next,
+ struct ptlrpc_request, rq_list);
list_del_init(&req->rq_list);
obd->obd_requests_queued_for_recovery--;
- spin_unlock_bh(&obd->obd_processing_task_lock);
+ } else {
+ req = NULL;
+ }
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ RETURN(req);
+}
+
+static int check_for_next_lock(struct obd_device *obd)
+{
+ struct ptlrpc_request *req = NULL;
+ int wake_up = 0;
+
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ if (!list_empty(&obd->obd_lock_replay_queue)) {
+ req = list_entry(obd->obd_lock_replay_queue.next,
+ struct ptlrpc_request, rq_list);
+ CDEBUG(D_HA, "waking for next lock\n");
+ wake_up = 1;
+ } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
+ CDEBUG(D_HA, "waking for completed lock replay\n");
+ wake_up = 1;
+ } else if (obd->obd_abort_recovery) {
+ CDEBUG(D_HA, "waking for aborted recovery\n");
+ wake_up = 1;
+ }
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+
+ return wake_up;
+}
+
+static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
+{
+ struct l_wait_info lwi = { 0 };
+ struct ptlrpc_request *req;
+
+ CDEBUG(D_HA, "Waiting for lock\n");
+ l_wait_event(obd->obd_next_transno_waitq,
+ check_for_next_lock(obd), &lwi);
+
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ if (obd->obd_abort_recovery) {
+ req = NULL;
+ } else if (!list_empty(&obd->obd_lock_replay_queue)) {
+ req = list_entry(obd->obd_lock_replay_queue.next,
+ struct ptlrpc_request, rq_list);
+ list_del_init(&req->rq_list);
+ } else {
+ req = NULL;
+ }
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ return req;
+}
+
+static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
+{
+ struct ptlrpc_request *req;
+
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ if (!list_empty(&obd->obd_final_req_queue)) {
+ req = list_entry(obd->obd_final_req_queue.next,
+ struct ptlrpc_request, rq_list);
+ list_del_init(&req->rq_list);
+ } else {
+ req = NULL;
+ }
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ return req;
+}
+
+static inline int req_replay_done(struct obd_export *exp)
+{
+ return (exp->exp_req_replay_needed == 0);
+}
- DEBUG_REQ(D_HA, req, "processing: ");
- (void)obd->obd_recovery_handler(req);
+static inline int lock_replay_done(struct obd_export *exp)
+{
+ return (exp->exp_lock_replay_needed == 0);
+}
+
+static inline int connect_done(struct obd_export *exp)
+{
+ return (exp->exp_in_recovery != 0);
+}
+
+static int check_for_clients(struct obd_device *obd)
+{
+ if (obd->obd_abort_recovery)
+ return 1;
+ LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
+ if (obd->obd_no_conn == 0 &&
+ obd->obd_connected_clients == obd->obd_max_recoverable_clients)
+ return 1;
+ return 0;
+}
+
+static int handle_recovery_req(struct ptlrpc_thread *thread,
+ struct ptlrpc_request *req,
+ svc_handler_t handler)
+{
+ int rc;
+ ENTRY;
+
+ rc = lu_context_init(&req->rq_session, LCT_SESSION);
+ if (rc) {
+ CERROR("Failure to initialize session: %d\n", rc);
+ return rc;
+ }
+ req->rq_session.lc_thread = thread;
+ lu_context_enter(&req->rq_session);
+ req->rq_svc_thread = thread;
+ req->rq_svc_thread->t_env->le_ses = &req->rq_session;
+
+ (void)handler(req);
+
+ lu_context_exit(&req->rq_session);
+ lu_context_fini(&req->rq_session);
+ /* don't reset timer for final stage */
+ if (!req_replay_done(req->rq_export) ||
+ !lock_replay_done(req->rq_export))
+ reset_recovery_timer(class_exp2obd(req->rq_export));
+ ptlrpc_free_clone(req);
+ RETURN(0);
+}
+
+static int target_recovery_thread(void *arg)
+{
+ struct obd_device *obd = arg;
+ struct ptlrpc_request *req;
+ struct target_recovery_data *trd = &obd->obd_recovery_data;
+ struct l_wait_info lwi = { 0 };
+ unsigned long delta;
+ unsigned long flags;
+ struct lu_env env;
+ struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
+ __u32 recov_ctx_tags = LCT_MD_THREAD;
+ int rc = 0;
+ ENTRY;
+
+ cfs_daemonize("tgt_recov");
+
+ SIGNAL_MASK_LOCK(current, flags);
+ sigfillset(¤t->blocked);
+ RECALC_SIGPENDING;
+ SIGNAL_MASK_UNLOCK(current, flags);
+
+ rc = lu_context_init(&env.le_ctx, recov_ctx_tags);
+ if (rc)
+ RETURN(rc);
+
+ thread->t_env = &env;
+ env.le_ctx.lc_thread = thread;
+
+ CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
+ current->pid);
+ trd->trd_processing_task = current->pid;
+
+ obd->obd_recovering = 1;
+ complete(&trd->trd_starting);
+
+ /* first of all, we have to know the first transno to replay */
+ obd->obd_abort_recovery = 0;
+ l_wait_event(obd->obd_next_transno_waitq,
+ check_for_clients(obd), &lwi);
+
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ target_cancel_recovery_timer(obd);
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+
+ /* If some clients haven't connected in time, evict them */
+ if (obd->obd_abort_recovery) {
+ CWARN("Some clients haven't connect in time (%d/%d),"
+ "evict them\n", obd->obd_connected_clients,
+ obd->obd_max_recoverable_clients);
+ obd->obd_abort_recovery = obd->obd_stopping;
+ class_disconnect_stale_exports(obd, connect_done);
+ }
+ /* next stage: replay requests */
+ delta = jiffies;
+ obd->obd_req_replaying = 1;
+ CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
+ atomic_read(&obd->obd_req_replay_clients),
+ obd->obd_next_recovery_transno);
+ while ((req = target_next_replay_req(obd))) {
+ LASSERT(trd->trd_processing_task == current->pid);
+ DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
+ lustre_msg_get_transno(req->rq_reqmsg),
+ libcfs_nid2str(req->rq_peer.nid));
+
+ handle_recovery_req(thread, req,
+ trd->trd_recovery_handler);
obd->obd_replayed_requests++;
- reset_recovery_timer(obd);
- /* bug 1580: decide how to properly sync() in recovery */
- //mds_fsync_super(mds->mds_sb);
- class_export_put(req->rq_export);
- OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
- OBD_FREE(req, sizeof *req);
spin_lock_bh(&obd->obd_processing_task_lock);
obd->obd_next_recovery_transno++;
- if (list_empty(&obd->obd_recovery_queue)) {
- obd->obd_processing_task = 0;
- spin_unlock_bh(&obd->obd_processing_task_lock);
- break;
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ }
+
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ target_cancel_recovery_timer(obd);
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+
+ /* If some clients haven't replayed requests in time, evict them */
+ if (obd->obd_abort_recovery) {
+ CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
+ obd->obd_abort_recovery = obd->obd_stopping;
+ class_disconnect_stale_exports(obd, req_replay_done);
+ abort_req_replay_queue(obd);
+ }
+ /* The second stage: replay locks */
+ CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
+ atomic_read(&obd->obd_lock_replay_clients));
+ while ((req = target_next_replay_lock(obd))) {
+ LASSERT(trd->trd_processing_task == current->pid);
+ DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
+ libcfs_nid2str(req->rq_peer.nid));
+ handle_recovery_req(thread, req,
+ trd->trd_recovery_handler);
+ obd->obd_replayed_locks++;
+ }
+
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ target_cancel_recovery_timer(obd);
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ /* If some clients haven't replayed requests in time, evict them */
+ if (obd->obd_abort_recovery) {
+ int stale;
+ CERROR("lock replay timed out, aborting ...\n");
+ obd->obd_abort_recovery = obd->obd_stopping;
+ stale = class_disconnect_stale_exports(obd, lock_replay_done);
+ abort_lock_replay_queue(obd);
+ }
+
+ /* We drop recoverying flag to forward all new requests
+ * to regular mds_handle() since now */
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ obd->obd_recovering = obd->obd_abort_recovery = 0;
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ /* The third stage: reply on final pings */
+ CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
+ while ((req = target_next_final_ping(obd))) {
+ LASSERT(trd->trd_processing_task == current->pid);
+ DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
+ libcfs_nid2str(req->rq_peer.nid));
+ handle_recovery_req(thread, req,
+ trd->trd_recovery_handler);
+ }
+
+ delta = (jiffies - delta) / HZ;
+ CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
+ delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
+ LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
+ LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
+ if (delta > obd_timeout * 2) {
+ CWARN("too long recovery - read logs\n");
+ libcfs_debug_dumplog();
+ }
+
+ target_finish_recovery(obd);
+
+ lu_env_fini(&env);
+ trd->trd_processing_task = 0;
+ complete(&trd->trd_finishing);
+ RETURN(rc);
+}
+
+int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
+{
+ int rc = 0;
+ struct target_recovery_data *trd = &obd->obd_recovery_data;
+
+ memset(trd, 0, sizeof(*trd));
+ init_completion(&trd->trd_starting);
+ init_completion(&trd->trd_finishing);
+ trd->trd_recovery_handler = handler;
+
+ if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
+ wait_for_completion(&trd->trd_starting);
+ LASSERT(obd->obd_recovering != 0);
+ } else
+ rc = -ECHILD;
+
+ return rc;
+}
+
+void target_stop_recovery_thread(struct obd_device *obd)
+{
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ if (obd->obd_recovery_data.trd_processing_task > 0) {
+ struct target_recovery_data *trd = &obd->obd_recovery_data;
+ CERROR("%s: Aborting recovery\n", obd->obd_name);
+ obd->obd_abort_recovery = 1;
+ wake_up(&obd->obd_next_transno_waitq);
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ wait_for_completion(&trd->trd_finishing);
+ } else {
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ }
+}
+
+void target_recovery_fini(struct obd_device *obd)
+{
+ class_disconnect_exports(obd);
+ target_stop_recovery_thread(obd);
+ target_cleanup_recovery(obd);
+}
+EXPORT_SYMBOL(target_recovery_fini);
+
+void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
+{
+ if (obd->obd_max_recoverable_clients == 0)
+ return;
+
+ CWARN("RECOVERY: service %s, %d recoverable clients, "
+ "last_transno "LPU64"\n", obd->obd_name,
+ obd->obd_max_recoverable_clients, obd->obd_last_committed);
+ obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
+ target_start_recovery_thread(obd, handler);
+ obd->obd_recovery_start = CURRENT_SECONDS;
+ /* Only used for lprocfs_status */
+ obd->obd_recovery_end = obd->obd_recovery_start + OBD_RECOVERY_TIMEOUT;
+}
+EXPORT_SYMBOL(target_recovery_init);
+
+#endif
+
+int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
+{
+ struct obd_export *exp = req->rq_export;
+ LASSERT(exp != NULL);
+ if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
+ /* client declares he's ready to replay locks */
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ if (exp->exp_req_replay_needed) {
+ LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
+ spin_lock(&exp->exp_lock);
+ exp->exp_req_replay_needed = 0;
+ spin_unlock(&exp->exp_lock);
+ atomic_dec(&obd->obd_req_replay_clients);
+ LASSERT(obd->obd_recoverable_clients > 0);
+ obd->obd_recoverable_clients--;
+ if (atomic_read(&obd->obd_req_replay_clients) == 0)
+ CDEBUG(D_HA, "all clients have replayed reqs\n");
+ wake_up(&obd->obd_next_transno_waitq);
}
spin_unlock_bh(&obd->obd_processing_task_lock);
}
- EXIT;
+ if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
+ /* client declares he's ready to complete recovery
+ * so, we put the request on th final queue */
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ if (exp->exp_lock_replay_needed) {
+ LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
+ spin_lock(&exp->exp_lock);
+ exp->exp_lock_replay_needed = 0;
+ spin_unlock(&exp->exp_lock);
+ atomic_dec(&obd->obd_lock_replay_clients);
+ if (atomic_read(&obd->obd_lock_replay_clients) == 0)
+ CDEBUG(D_HA, "all clients have replayed locks\n");
+ wake_up(&obd->obd_next_transno_waitq);
+ }
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ }
+
+ return 0;
}
int target_queue_recovery_request(struct ptlrpc_request *req,
{
struct list_head *tmp;
int inserted = 0;
- __u64 transno = req->rq_reqmsg->transno;
- struct ptlrpc_request *saved_req;
- struct lustre_msg *reqmsg;
+ __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
+
+ ENTRY;
+
+ if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
+ /* Processing the queue right now, don't re-add. */
+ RETURN(1);
+ }
+
+ target_process_req_flags(obd, req);
+
+ if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
+ /* client declares he's ready to complete recovery
+ * so, we put the request on th final queue */
+ req = ptlrpc_clone_req(req);
+ if (req == NULL)
+ RETURN(-ENOMEM);
+ DEBUG_REQ(D_HA, req, "queue final req");
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ if (obd->obd_recovering)
+ list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
+ else {
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ ptlrpc_free_clone(req);
+ if (obd->obd_stopping) {
+ RETURN(-ENOTCONN);
+ } else {
+ RETURN(1);
+ }
+ }
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ RETURN(0);
+ }
+ if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
+ /* client declares he's ready to replay locks */
+ req = ptlrpc_clone_req(req);
+ if (req == NULL)
+ RETURN(-ENOMEM);
+ DEBUG_REQ(D_HA, req, "queue lock replay req");
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ LASSERT(obd->obd_recovering);
+ /* usually due to recovery abort */
+ if (!req->rq_export->exp_in_recovery) {
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ ptlrpc_free_clone(req);
+ RETURN(-ENOTCONN);
+ }
+ LASSERT(req->rq_export->exp_lock_replay_needed);
+ list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ wake_up(&obd->obd_next_transno_waitq);
+ RETURN(0);
+ }
/* CAVEAT EMPTOR: The incoming request message has been swabbed
* (i.e. buflens etc are in my own byte order), but type-dependent
* buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
if (!transno) {
- INIT_LIST_HEAD(&req->rq_list);
+ CFS_INIT_LIST_HEAD(&req->rq_list);
DEBUG_REQ(D_HA, req, "not queueing");
- return 1;
+ RETURN(1);
}
- /* XXX If I were a real man, these LBUGs would be sane cleanups. */
- /* XXX just like the request-dup code in queue_final_reply */
- OBD_ALLOC(saved_req, sizeof *saved_req);
- if (!saved_req)
- LBUG();
- OBD_ALLOC(reqmsg, req->rq_reqlen);
- if (!reqmsg)
- LBUG();
-
spin_lock_bh(&obd->obd_processing_task_lock);
/* If we're processing the queue, we want don't want to queue this
* Also, if this request has a transno less than the one we're waiting
* for, we should process it now. It could (and currently always will)
* be an open request for a descriptor that was opened some time ago.
+ *
+ * Also, a resent, replayed request that has already been
+ * handled will pass through here and be processed immediately.
*/
- if (obd->obd_processing_task == current->pid ||
- transno < obd->obd_next_recovery_transno) {
+ CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
+ obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
+ if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
/* Processing the queue right now, don't re-add. */
LASSERT(list_empty(&req->rq_list));
spin_unlock_bh(&obd->obd_processing_task_lock);
- OBD_FREE(reqmsg, req->rq_reqlen);
- OBD_FREE(saved_req, sizeof *saved_req);
- return 1;
+ RETURN(1);
+ }
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+
+ /* A resent, replayed request that is still on the queue; just drop it.
+ The queued request will handle this. */
+ if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT|MSG_REPLAY)) ==
+ (MSG_RESENT | MSG_REPLAY)) {
+ DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
+ RETURN(0);
}
- memcpy(saved_req, req, sizeof *req);
- memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
- req = saved_req;
- req->rq_reqmsg = reqmsg;
- class_export_get(req->rq_export);
- INIT_LIST_HEAD(&req->rq_list);
+ req = ptlrpc_clone_req(req);
+ if (req == NULL)
+ RETURN(-ENOMEM);
+
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ LASSERT(obd->obd_recovering);
+ if (!req->rq_export->exp_in_recovery) {
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ ptlrpc_free_clone(req);
+ RETURN(-ENOTCONN);
+ }
+ LASSERT(req->rq_export->exp_req_replay_needed);
/* XXX O(n^2) */
- list_for_each(tmp, &obd->obd_recovery_queue) {
+ list_for_each(tmp, &obd->obd_req_replay_queue) {
struct ptlrpc_request *reqiter =
list_entry(tmp, struct ptlrpc_request, rq_list);
- if (reqiter->rq_reqmsg->transno > transno) {
+ if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
list_add_tail(&req->rq_list, &reqiter->rq_list);
inserted = 1;
break;
}
}
- if (!inserted) {
- list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
- }
+ if (!inserted)
+ list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
obd->obd_requests_queued_for_recovery++;
-
- if (obd->obd_processing_task != 0) {
- /* Someone else is processing this queue, we'll leave it to
- * them.
- */
- wake_up(&obd->obd_next_transno_waitq);
- spin_unlock_bh(&obd->obd_processing_task_lock);
- return 0;
- }
-
- /* Nobody is processing, and we know there's (at least) one to process
- * now, so we'll do the honours.
- */
- obd->obd_processing_task = current->pid;
+ wake_up(&obd->obd_next_transno_waitq);
spin_unlock_bh(&obd->obd_processing_task_lock);
+ RETURN(0);
- process_recovery_queue(obd);
- return 0;
}
struct obd_device * target_req2obd(struct ptlrpc_request *req)
return req->rq_export->exp_obd;
}
-int target_queue_final_reply(struct ptlrpc_request *req, int rc)
+static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
{
- struct obd_device *obd = target_req2obd(req);
- struct ptlrpc_request *saved_req;
- struct lustre_msg *reqmsg;
- int recovery_done = 0;
- int rc2;
+ LASSERT(exp != NULL);
+ return &exp->exp_obd->obd_namespace->ns_pool;
+}
- if (rc) {
- /* Just like ptlrpc_error, but without the sending. */
- rc = lustre_pack_reply(req, 0, NULL, NULL);
- LASSERT(rc == 0); /* XXX handle this */
- req->rq_type = PTL_RPC_MSG_ERR;
+int target_pack_pool_reply(struct ptlrpc_request *req)
+{
+ struct ldlm_pool *pl;
+ ENTRY;
+
+ if (req->rq_export == NULL) {
+ lustre_msg_set_slv(req->rq_repmsg, 0);
+ lustre_msg_set_limit(req->rq_repmsg, 0);
+ RETURN(0);
}
+
+ if (!exp_connect_lru_resize(req->rq_export))
+ RETURN(0);
+
+ pl = ldlm_exp2pl(req->rq_export);
- LASSERT(list_empty(&req->rq_list));
- /* XXX a bit like the request-dup code in queue_recovery_request */
- OBD_ALLOC(saved_req, sizeof *saved_req);
- if (!saved_req)
- LBUG();
- OBD_ALLOC(reqmsg, req->rq_reqlen);
- if (!reqmsg)
- LBUG();
- memcpy(saved_req, req, sizeof *saved_req);
- memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
- req = saved_req;
- req->rq_reqmsg = reqmsg;
- list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
+ spin_lock(&pl->pl_lock);
+ lustre_msg_set_slv(req->rq_repmsg, ldlm_pool_get_slv(pl));
+ lustre_msg_set_limit(req->rq_repmsg, ldlm_pool_get_limit(pl));
+ spin_unlock(&pl->pl_lock);
- spin_lock_bh(&obd->obd_processing_task_lock);
- --obd->obd_recoverable_clients;
- recovery_done = (obd->obd_recoverable_clients == 0);
- spin_unlock_bh(&obd->obd_processing_task_lock);
+ RETURN(0);
+}
- if (recovery_done) {
- struct list_head *tmp, *n;
- ldlm_reprocess_all_ns(req->rq_export->exp_obd->obd_namespace);
- CWARN("%s: all clients recovered, sending delayed replies\n",
- obd->obd_name);
- obd->obd_recovering = 0;
-
- /* when recovering finished, cleanup orphans for mds */
- if (OBT(obd) && OBP(obd, postrecov)) {
- rc2 = OBP(obd, postrecov)(obd);
- if (rc2 >= 0)
- CWARN("%s: all clients recovered, %d MDS orphans "
- "deleted\n", obd->obd_name, rc2);
- else
- CERROR("postrecov failed %d\n", rc2);
- }
+int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
+{
+ if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
+ obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
+ DEBUG_REQ(D_ERROR, req, "dropping reply");
+ return (-ECOMM);
+ }
- list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
- req = list_entry(tmp, struct ptlrpc_request, rq_list);
- DEBUG_REQ(D_ERROR, req, "delayed:");
- ptlrpc_reply(req);
- list_del(&req->rq_list);
- OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
- OBD_FREE(req, sizeof *req);
- }
- target_cancel_recovery_timer(obd);
+ if (unlikely(rc)) {
+ DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
+ req->rq_status = rc;
+ return (ptlrpc_error(req));
} else {
- CERROR("%s: %d recoverable clients remain\n",
- obd->obd_name, obd->obd_recoverable_clients);
- wake_up(&obd->obd_next_transno_waitq);
+ DEBUG_REQ(D_NET, req, "sending reply");
}
- return 1;
+ target_pack_pool_reply(req);
+ return (ptlrpc_send_reply(req, 1));
}
-static void ptlrpc_abort_reply (struct ptlrpc_request *req)
+void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
{
- /* On return, we must be sure that the ACK callback has either
- * happened or will not happen. Note that the SENT callback will
- * happen come what may since we successfully posted the PUT. */
- int rc;
- struct l_wait_info lwi;
- unsigned long flags;
+ int netrc;
+ struct ptlrpc_reply_state *rs;
+ struct obd_device *obd;
+ struct obd_export *exp;
+ struct ptlrpc_service *svc;
- again:
- /* serialise with ACK callback */
- spin_lock_irqsave (&req->rq_lock, flags);
- if (!req->rq_want_ack) {
- spin_unlock_irqrestore (&req->rq_lock, flags);
- /* The ACK callback has happened already. Although the
- * SENT callback might still be outstanding (yes really) we
- * don't care; this is just like normal completion. */
+ if (req->rq_no_reply)
return;
- }
- spin_unlock_irqrestore (&req->rq_lock, flags);
- /* Have a bash at unlinking the MD. This will fail until the SENT
- * callback has happened since the MD is busy from the PUT. If the
- * ACK still hasn't arrived after then, a successful unlink will
- * ensure the ACK callback never happens. */
- rc = PtlMDUnlink (req->rq_reply_md_h);
- switch (rc) {
- default:
- LBUG ();
- case PTL_OK:
- /* SENT callback happened; ACK callback preempted */
- LASSERT (req->rq_want_ack);
- spin_lock_irqsave (&req->rq_lock, flags);
- req->rq_want_ack = 0;
- spin_unlock_irqrestore (&req->rq_lock, flags);
+ svc = req->rq_rqbd->rqbd_service;
+ rs = req->rq_reply_state;
+ if (rs == NULL || !rs->rs_difficult) {
+ /* no notifiers */
+ target_send_reply_msg (req, rc, fail_id);
return;
- case PTL_INV_MD:
- return;
- case PTL_MD_INUSE:
- /* Still sending or ACK callback in progress: wait until
- * either callback has completed and try again.
- * Actually we can't wait for the SENT callback because
- * there's no state the SENT callback can touch that will
- * allow it to communicate with us! So we just wait here
- * for a short time, effectively polling for the SENT
- * callback by calling PtlMDUnlink() again, to see if it
- * has finished. Note that if the ACK does arrive, its
- * callback wakes us in short order. --eeb */
- lwi = LWI_TIMEOUT (HZ/4, NULL, NULL);
- rc = l_wait_event(req->rq_reply_waitq, !req->rq_want_ack,
- &lwi);
- CDEBUG (D_HA, "Retrying req %p: %d\n", req, rc);
- /* NB go back and test rq_want_ack with locking, to ensure
- * if ACK callback happened, it has completed stopped
- * referencing this req. */
- goto again;
}
-}
-void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
-{
- int i;
- int netrc;
- unsigned long flags;
- struct ptlrpc_req_ack_lock *ack_lock;
- struct l_wait_info lwi = { 0 };
- wait_queue_t commit_wait;
- struct obd_device *obd =
- req->rq_export ? req->rq_export->exp_obd : NULL;
- struct obd_export *exp = NULL;
-
- if (req->rq_export) {
- for (i = 0; i < REQ_MAX_ACK_LOCKS; i++) {
- if (req->rq_ack_locks[i].mode) {
- exp = req->rq_export;
- break;
- }
- }
+ /* must be an export if locks saved */
+ LASSERT (req->rq_export != NULL);
+ /* req/reply consistent */
+ LASSERT (rs->rs_service == svc);
+
+ /* "fresh" reply */
+ LASSERT (!rs->rs_scheduled);
+ LASSERT (!rs->rs_scheduled_ever);
+ LASSERT (!rs->rs_handled);
+ LASSERT (!rs->rs_on_net);
+ LASSERT (rs->rs_export == NULL);
+ LASSERT (list_empty(&rs->rs_obd_list));
+ LASSERT (list_empty(&rs->rs_exp_list));
+
+ exp = class_export_get (req->rq_export);
+ obd = exp->exp_obd;
+
+ /* disable reply scheduling onto srv_reply_queue while I'm setting up */
+ rs->rs_scheduled = 1;
+ rs->rs_on_net = 1;
+ rs->rs_xid = req->rq_xid;
+ rs->rs_transno = req->rq_transno;
+ rs->rs_export = exp;
+
+ spin_lock(&obd->obd_uncommitted_replies_lock);
+
+ if (rs->rs_transno > obd->obd_last_committed) {
+ /* not committed already */
+ list_add_tail (&rs->rs_obd_list,
+ &obd->obd_uncommitted_replies);
}
- if (exp) {
- exp->exp_outstanding_reply = req;
- spin_lock_irqsave (&req->rq_lock, flags);
- req->rq_want_ack = 1;
- spin_unlock_irqrestore (&req->rq_lock, flags);
- }
+ spin_unlock (&obd->obd_uncommitted_replies_lock);
+ spin_lock (&exp->exp_lock);
- if (!OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
- if (rc == 0) {
- DEBUG_REQ(D_NET, req, "sending reply");
- netrc = ptlrpc_reply(req);
- } else if (rc == -ENOTCONN) {
- DEBUG_REQ(D_HA, req, "processing error (%d)", rc);
- netrc = ptlrpc_error(req);
- } else {
- DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
- netrc = ptlrpc_error(req);
- }
- } else {
- obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
- DEBUG_REQ(D_ERROR, req, "dropping reply");
- if (req->rq_repmsg) {
- OBD_FREE(req->rq_repmsg, req->rq_replen);
- req->rq_repmsg = NULL;
- }
- init_waitqueue_head(&req->rq_reply_waitq);
- netrc = 0;
+ list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
+
+ spin_unlock(&exp->exp_lock);
+
+ netrc = target_send_reply_msg (req, rc, fail_id);
+
+ spin_lock(&svc->srv_lock);
+
+ svc->srv_n_difficult_replies++;
+
+ if (netrc != 0) {
+ /* error sending: reply is off the net. Also we need +1
+ * reply ref until ptlrpc_server_handle_reply() is done
+ * with the reply state (if the send was successful, there
+ * would have been +1 ref for the net, which
+ * reply_out_callback leaves alone) */
+ rs->rs_on_net = 0;
+ ptlrpc_rs_addref(rs);
+ atomic_inc (&svc->srv_outstanding_replies);
}
- /* a failed send simulates the callbacks */
- LASSERT(netrc == 0 || req->rq_want_ack == 0);
- if (exp == NULL) {
- LASSERT(req->rq_want_ack == 0);
- return;
+ if (!rs->rs_on_net || /* some notifier */
+ list_empty(&rs->rs_exp_list) || /* completed already */
+ list_empty(&rs->rs_obd_list)) {
+ list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
+ cfs_waitq_signal (&svc->srv_waitq);
+ } else {
+ list_add (&rs->rs_list, &svc->srv_active_replies);
+ rs->rs_scheduled = 0; /* allow notifier to schedule */
}
- LASSERT(obd != NULL);
- init_waitqueue_entry(&commit_wait, current);
- add_wait_queue(&obd->obd_commit_waitq, &commit_wait);
- rc = l_wait_event(req->rq_reply_waitq,
- !req->rq_want_ack || req->rq_resent ||
- req->rq_transno <= obd->obd_last_committed, &lwi);
- remove_wait_queue(&obd->obd_commit_waitq, &commit_wait);
+ spin_unlock(&svc->srv_lock);
+}
- spin_lock_irqsave (&req->rq_lock, flags);
- /* If we got here because the ACK callback ran, this acts as a
- * barrier to ensure the callback completed the wakeup. */
- spin_unlock_irqrestore (&req->rq_lock, flags);
+int target_handle_ping(struct ptlrpc_request *req)
+{
+ obd_ping(req->rq_export);
+ return lustre_pack_reply(req, 1, NULL, NULL);
+}
- /* If we committed the transno already, then we might wake up before
- * the ack arrives. We need to stop waiting for the ack before we can
- * reuse this request structure. We are guaranteed by this point that
- * this cannot abort the sending of the actual reply.*/
- ptlrpc_abort_reply(req);
+void target_committed_to_req(struct ptlrpc_request *req)
+{
+ struct obd_device *obd;
- if (req->rq_resent) {
- DEBUG_REQ(D_HA, req, "resent: not cancelling locks");
+ if (req == NULL || req->rq_export == NULL)
return;
- }
- LASSERT(rc == 0);
- DEBUG_REQ(D_HA, req, "cancelling locks for %s",
- req->rq_want_ack ? "commit" : "ack");
+ obd = req->rq_export->exp_obd;
+ if (obd == NULL)
+ return;
- exp->exp_outstanding_reply = NULL;
+ if (!obd->obd_no_transno && req->rq_repmsg != NULL)
+ lustre_msg_set_last_committed(req->rq_repmsg,
+ obd->obd_last_committed);
+ else
+ DEBUG_REQ(D_IOCTL, req, "not sending last_committed update");
- for (ack_lock = req->rq_ack_locks, i = 0;
- i < REQ_MAX_ACK_LOCKS; i++, ack_lock++) {
- if (!ack_lock->mode)
- continue;
- ldlm_lock_decref(&ack_lock->lock, ack_lock->mode);
- }
+ CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
+ obd->obd_last_committed, req->rq_transno, req->rq_xid);
}
-int target_handle_ping(struct ptlrpc_request *req)
+EXPORT_SYMBOL(target_committed_to_req);
+
+#ifdef HAVE_QUOTA_SUPPORT
+int target_handle_qc_callback(struct ptlrpc_request *req)
{
- return lustre_pack_reply(req, 0, NULL, NULL);
+ struct obd_quotactl *oqctl;
+ struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
+
+ oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
+ lustre_swab_obd_quotactl);
+ if (oqctl == NULL) {
+ CERROR("Can't unpack obd_quotactl\n");
+ RETURN(-EPROTO);
+ }
+
+ cli->cl_qchk_stat = oqctl->qc_stat;
+
+ return 0;
}
-void *ldlm_put_lock_into_req(struct ptlrpc_request *req,
- struct lustre_handle *lock, int mode)
+int target_handle_dqacq_callback(struct ptlrpc_request *req)
{
- int i;
+#ifdef __KERNEL__
+ struct obd_device *obd = req->rq_export->exp_obd;
+ struct obd_device *master_obd;
+ struct lustre_quota_ctxt *qctxt;
+ struct qunit_data *qdata;
+ void* rep;
+ struct qunit_data_old *qdata_old;
+ int rc = 0;
+ int repsize[2] = { sizeof(struct ptlrpc_body),
+ sizeof(struct qunit_data) };
+ ENTRY;
- for (i = 0; i < REQ_MAX_ACK_LOCKS; i++) {
- if (req->rq_ack_locks[i].mode)
- continue;
- memcpy(&req->rq_ack_locks[i].lock, lock, sizeof(*lock));
- req->rq_ack_locks[i].mode = mode;
- return &req->rq_ack_locks[i];
+ rc = lustre_pack_reply(req, 2, repsize, NULL);
+ if (rc) {
+ CERROR("packing reply failed!: rc = %d\n", rc);
+ RETURN(rc);
+ }
+ LASSERT(req->rq_export);
+
+ /* fixed for bug10707 */
+ if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
+ !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
+ CDEBUG(D_QUOTA, "qd_count is 64bit!\n");
+ rep = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
+ sizeof(struct qunit_data));
+ LASSERT(rep);
+ qdata = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*qdata),
+ lustre_swab_qdata);
+ } else {
+ CDEBUG(D_QUOTA, "qd_count is 32bit!\n");
+ rep = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
+ sizeof(struct qunit_data_old));
+ LASSERT(rep);
+ qdata_old = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*qdata_old),
+ lustre_swab_qdata_old);
+ qdata = lustre_quota_old_to_new(qdata_old);
}
- CERROR("no space for lock in struct ptlrpc_request\n");
- LBUG();
- return NULL;
-}
+ if (qdata == NULL) {
+ CERROR("Can't unpack qunit_data\n");
+ RETURN(-EPROTO);
+ }
+
+ /* we use the observer */
+ LASSERT(obd->obd_observer && obd->obd_observer->obd_observer);
+ master_obd = obd->obd_observer->obd_observer;
+ qctxt = &master_obd->u.obt.obt_qctxt;
+
+ LASSERT(qctxt->lqc_handler);
+ rc = qctxt->lqc_handler(master_obd, qdata,
+ lustre_msg_get_opc(req->rq_reqmsg));
+ if (rc && rc != -EDQUOT)
+ CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
+ "dqacq failed! (rc:%d)\n", rc);
+
+ /* the qd_count might be changed in lqc_handler */
+ if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
+ !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
+ memcpy(rep,qdata,sizeof(*qdata));
+ } else {
+ qdata_old = lustre_quota_new_to_old(qdata);
+ memcpy(rep,qdata_old,sizeof(*qdata_old));
+ }
+ req->rq_status = rc;
+ rc = ptlrpc_reply(req);
+
+ RETURN(rc);
+#else
+ return 0;
+#endif /* !__KERNEL__ */
+}
+#endif /* HAVE_QUOTA_SUPPORT */
+
+ldlm_mode_t lck_compat_array[] = {
+ [LCK_EX] LCK_COMPAT_EX,
+ [LCK_PW] LCK_COMPAT_PW,
+ [LCK_PR] LCK_COMPAT_PR,
+ [LCK_CW] LCK_COMPAT_CW,
+ [LCK_CR] LCK_COMPAT_CR,
+ [LCK_NL] LCK_COMPAT_NL,
+ [LCK_GROUP] LCK_COMPAT_GROUP
+};