-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
* GPL HEADER START
*
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
/*
* Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2011, 2012, Whamcloud, Inc.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
* Lustre is a trademark of Sun Microsystems, Inc.
*/
-#ifndef EXPORT_SYMTAB
-# define EXPORT_SYMTAB
-#endif
#define DEBUG_SUBSYSTEM S_LDLM
#ifdef __KERNEL__
# include <liblustre.h>
#endif
#include <obd.h>
-#include <lustre_mds.h>
+#include <obd_class.h>
#include <lustre_dlm.h>
#include <lustre_net.h>
#include <lustre_sec.h>
{
return import_set_conn(imp, uuid, priority, 1);
}
+EXPORT_SYMBOL(client_import_add_conn);
int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
{
CERROR("connection %s not found\n", uuid->uuid);
RETURN(rc);
}
+EXPORT_SYMBOL(client_import_del_conn);
+
+/**
+ * Find conn uuid by peer nid. @peer is a server nid. This function is used
+ * to find a conn uuid of @imp which can reach @peer.
+ */
+int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
+ struct obd_uuid *uuid)
+{
+ struct obd_import_conn *conn;
+ int rc = -ENOENT;
+ ENTRY;
+
+ cfs_spin_lock(&imp->imp_lock);
+ cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
+ /* check if conn uuid does have this peer nid */
+ if (class_check_uuid(&conn->oic_uuid, peer)) {
+ *uuid = conn->oic_uuid;
+ rc = 0;
+ break;
+ }
+ }
+ cfs_spin_unlock(&imp->imp_lock);
+ RETURN(rc);
+}
+EXPORT_SYMBOL(client_import_find_conn);
void client_destroy_import(struct obd_import *imp)
{
sptlrpc_import_sec_put(imp);
class_import_put(imp);
}
+EXPORT_SYMBOL(client_destroy_import);
+
+/**
+ * check whether the osc is on MDT or not
+ * In the config log,
+ * osc on MDT
+ * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
+ * osc on client
+ * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
+ *
+ **/
+static int osc_on_mdt(char *obdname)
+{
+ char *ptr;
+
+ ptr = strrchr(obdname, '-');
+ if (ptr == NULL)
+ return 0;
+
+ if (strncmp(ptr + 1, "MDT", 3) == 0)
+ return 1;
+
+ return 0;
+}
/* configure an RPC client OBD device
*
struct obd_uuid server_uuid;
int rq_portal, rp_portal, connect_op;
char *name = obddev->obd_type->typ_name;
+ ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
int rc;
ENTRY;
/* In a more perfect world, we would hang a ptlrpc_client off of
* obd_type and just use the values from there. */
- if (!strcmp(name, LUSTRE_OSC_NAME)) {
+ if (!strcmp(name, LUSTRE_OSC_NAME) ||
+ (!strcmp(name, LUSTRE_OSP_NAME) &&
+ !is_osp_on_ost(lustre_cfg_buf(lcfg, 0)))) {
rq_portal = OST_REQUEST_PORTAL;
rp_portal = OSC_REPLY_PORTAL;
connect_op = OST_CONNECT;
cli->cl_sp_me = LUSTRE_SP_CLI;
cli->cl_sp_to = LUSTRE_SP_OST;
- } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
+ ns_type = LDLM_NS_TYPE_OSC;
+
+ } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
+ (!strcmp(name, LUSTRE_OSP_NAME) &&
+ is_osp_on_ost(lustre_cfg_buf(lcfg, 0)))) {
rq_portal = MDS_REQUEST_PORTAL;
rp_portal = MDC_REPLY_PORTAL;
connect_op = MDS_CONNECT;
cli->cl_sp_me = LUSTRE_SP_CLI;
cli->cl_sp_to = LUSTRE_SP_MDT;
+ ns_type = LDLM_NS_TYPE_MDC;
+
} else if (!strcmp(name, LUSTRE_MGC_NAME)) {
rq_portal = MGS_REQUEST_PORTAL;
rp_portal = MGC_REPLY_PORTAL;
cli->cl_sp_me = LUSTRE_SP_MGC;
cli->cl_sp_to = LUSTRE_SP_MGS;
cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
+ ns_type = LDLM_NS_TYPE_MGC;
+
+ } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
+ rq_portal = OST_REQUEST_PORTAL;
+ rp_portal = OSC_REPLY_PORTAL;
+ connect_op = OST_CONNECT;
+ cli->cl_sp_me = LUSTRE_SP_CLI;
+ cli->cl_sp_to = LUSTRE_SP_OST;
+ ns_type = LDLM_NS_TYPE_OSC;
+
} else {
CERROR("unknown client OBD type \"%s\", can't setup\n",
name);
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);
+ cfs_atomic_set(&cli->cl_pending_w_pages, 0);
+ cfs_atomic_set(&cli->cl_pending_r_pages, 0);
cli->cl_r_in_flight = 0;
cli->cl_w_in_flight = 0;
cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
+
+ /* lru for osc. */
+ CFS_INIT_LIST_HEAD(&cli->cl_lru_osc);
+ cfs_atomic_set(&cli->cl_lru_shrinkers, 0);
+ cfs_atomic_set(&cli->cl_lru_busy, 0);
+ cfs_atomic_set(&cli->cl_lru_in_list, 0);
+ CFS_INIT_LIST_HEAD(&cli->cl_lru_list);
+ client_obd_list_lock_init(&cli->cl_lru_list_lock);
+
cfs_waitq_init(&cli->cl_destroy_waitq);
cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
#ifdef ENABLE_CHECKSUM
} else if (cfs_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;
+ if (osc_on_mdt(obddev->obd_name))
+ cli->cl_max_rpcs_in_flight = MDS_OSC_MAX_RIF_DEFAULT;
+ else
+ cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
}
-
rc = ldlm_get_ref();
if (rc) {
CERROR("ldlm_get_ref failed: %d\n", rc);
obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
LDLM_NAMESPACE_CLIENT,
- LDLM_NAMESPACE_GREEDY);
+ LDLM_NAMESPACE_GREEDY,
+ ns_type);
if (obddev->obd_namespace == NULL) {
CERROR("Unable to create client namespace - %s\n",
obddev->obd_name);
RETURN(rc);
}
+EXPORT_SYMBOL(client_obd_setup);
int client_obd_cleanup(struct obd_device *obddev)
{
ldlm_namespace_free_post(obddev->obd_namespace);
obddev->obd_namespace = NULL;
+ LASSERT(obddev->u.cli.cl_import == NULL);
+
ldlm_put_ref();
RETURN(0);
}
+EXPORT_SYMBOL(client_obd_cleanup);
/* ->o_connect() method for client side (OSC and MDC and MGC) */
int client_connect_import(const struct lu_env *env,
imp->imp_connect_flags_orig = data->ocd_connect_flags;
}
- rc = ptlrpc_connect_import(imp, NULL);
+ rc = ptlrpc_connect_import(imp);
if (rc != 0) {
LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
GOTO(out_ldlm, rc);
return rc;
}
+EXPORT_SYMBOL(client_connect_import);
int client_disconnect_export(struct obd_export *exp)
{
ptlrpc_invalidate_import(imp);
- if (imp->imp_rq_pool) {
- ptlrpc_free_rq_pool(imp->imp_rq_pool);
- imp->imp_rq_pool = NULL;
- }
- client_destroy_import(imp);
- cli->cl_import = NULL;
-
EXIT;
out_disconnect:
RETURN(rc);
}
+EXPORT_SYMBOL(client_disconnect_export);
+#ifdef HAVE_SERVER_SUPPORT
int server_disconnect_export(struct obd_export *exp)
{
int rc;
struct ptlrpc_reply_state *rs =
cfs_list_entry(exp->exp_outstanding_replies.next,
struct ptlrpc_reply_state, rs_exp_list);
- struct ptlrpc_service *svc = rs->rs_service;
+ struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
- cfs_spin_lock(&svc->srv_lock);
- cfs_list_del_init(&rs->rs_exp_list);
- cfs_spin_lock(&rs->rs_lock);
- ptlrpc_schedule_difficult_reply(rs);
- cfs_spin_unlock(&rs->rs_lock);
- cfs_spin_unlock(&svc->srv_lock);
- }
- cfs_spin_unlock(&exp->exp_lock);
+ cfs_spin_lock(&svcpt->scp_rep_lock);
- /* release nid stat refererence */
- lprocfs_exp_cleanup(exp);
+ cfs_list_del_init(&rs->rs_exp_list);
+ cfs_spin_lock(&rs->rs_lock);
+ ptlrpc_schedule_difficult_reply(rs);
+ cfs_spin_unlock(&rs->rs_lock);
- RETURN(rc);
+ cfs_spin_unlock(&svcpt->scp_rep_lock);
+ }
+ cfs_spin_unlock(&exp->exp_lock);
+
+ RETURN(rc);
}
+EXPORT_SYMBOL(server_disconnect_export);
/* --------------------------------------------------------------------------
* from old lib/target.c
struct obd_uuid *cluuid)
{
ENTRY;
+
if (exp->exp_connection && exp->exp_imp_reverse) {
struct lustre_handle *hdl;
+ struct obd_device *target;
+
hdl = &exp->exp_imp_reverse->imp_remote_handle;
+ target = exp->exp_obd;
+
/* Might be a re-connect after a partition. */
if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
- CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
- cluuid->uuid);
+ if (target->obd_recovering) {
+ int timeout = cfs_duration_sec(cfs_time_sub(
+ cfs_timer_deadline(
+ &target->obd_recovery_timer),
+ cfs_time_current()));
+
+ LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
+ "ing, waiting for %d clients in recov"
+ "ery for %d:%.02d\n", target->obd_name,
+ obd_uuid2str(&exp->exp_client_uuid),
+ obd_export_nid2str(exp),
+ target->obd_max_recoverable_clients,
+ timeout / 60, timeout % 60);
+ } else {
+ LCONSOLE_WARN("%s: Client %s (at %s) "
+ "reconnecting\n", target->obd_name,
+ obd_uuid2str(&exp->exp_client_uuid),
+ obd_export_nid2str(exp));
+ }
+
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, "
- "handle mismatch (ours "LPX64", theirs "
- LPX64")\n", cluuid->uuid,
- exp->exp_connection->c_remote_uuid.uuid,
- hdl->cookie, conn->cookie);
+ LCONSOLE_WARN("%s: already connected client %s (at %s) "
+ "with handle "LPX64". Rejecting client "
+ "with the same UUID trying to reconnect "
+ "with handle "LPX64"\n", target->obd_name,
+ obd_uuid2str(&exp->exp_client_uuid),
+ obd_export_nid2str(exp),
+ hdl->cookie, conn->cookie);
memset(conn, 0, sizeof *conn);
/* target_handle_connect() treats EALREADY and
* -EALREADY differently. -EALREADY is an error
}
EXPORT_SYMBOL(target_client_add_cb);
+#ifdef __KERNEL__
static void
-target_start_and_reset_recovery_timer(struct obd_device *obd,
- struct ptlrpc_request *req,
- int new_client);
+check_and_start_recovery_timer(struct obd_device *obd,
+ struct ptlrpc_request *req, int new_client);
+#else
+static inline void
+check_and_start_recovery_timer(struct obd_device *obd,
+ struct ptlrpc_request *req, int new_client)
+{
+}
+#endif
int target_handle_connect(struct ptlrpc_request *req)
{
- struct obd_device *target, *targref = NULL;
+ struct obd_device *target = NULL, *targref = NULL;
struct obd_export *export = NULL;
struct obd_import *revimp;
+ struct obd_import *tmp_imp = NULL;
struct lustre_handle conn;
struct lustre_handle *tmp;
struct obd_uuid tgtuuid;
struct obd_uuid remote_uuid;
char *str;
int rc = 0;
- int mds_conn = 0;
+ char *target_start;
+ int target_len;
+ bool mds_conn = false, lw_client = false;
struct obd_connect_data *data, *tmpdata;
+ int size, tmpsize;
lnet_nid_t *client_nid = NULL;
- ENTRY;
+ ENTRY;
OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
if (!target)
target = class_name2obd(str);
- if (!target || target->obd_stopping || !target->obd_set_up) {
- 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);
- }
+ if (!target) {
+ deuuidify(str, NULL, &target_start, &target_len);
+ LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available for "
+ "connect (no target)\n", str);
+ GOTO(out, rc = -ENODEV);
+ }
+
+ cfs_spin_lock(&target->obd_dev_lock);
+ if (target->obd_stopping || !target->obd_set_up) {
+ cfs_spin_unlock(&target->obd_dev_lock);
+
+ deuuidify(str, NULL, &target_start, &target_len);
+ LCONSOLE_ERROR_MSG(0x137, "%.*s: Not available for connect "
+ "from %s (%s)\n", target_len, target_start,
+ libcfs_nid2str(req->rq_peer.nid),
+ (target->obd_stopping ?
+ "stopping" : "not set up"));
+ GOTO(out, rc = -ENODEV);
+ }
if (target->obd_no_conn) {
- LCONSOLE_WARN("%s: temporarily refusing client connection "
+ cfs_spin_unlock(&target->obd_dev_lock);
+
+ LCONSOLE_WARN("%s: Temporarily refusing client connection "
"from %s\n", target->obd_name,
libcfs_nid2str(req->rq_peer.nid));
GOTO(out, rc = -EAGAIN);
Really, class_uuid2obd should take the ref. */
targref = class_incref(target, __FUNCTION__, cfs_current());
+ target->obd_conn_inprogress++;
+ cfs_spin_unlock(&target->obd_dev_lock);
str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
if (str == NULL) {
conn = *tmp;
+ size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
+ RCL_CLIENT);
data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
if (!data)
GOTO(out, rc = -EPROTO);
if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
(data->ocd_connect_flags & OBD_CONNECT_MDS))
- mds_conn = 1;
+ mds_conn = true;
+
+ if ((data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0)
+ lw_client = true;
/* lctl gets a backstage, all-access pass. */
if (obd_uuid_equals(&cluuid, &target->obd_uuid))
goto no_export;
/* we've found an export in the hash */
+
+ cfs_spin_lock(&export->exp_lock);
+
if (export->exp_connecting) { /* bug 9635, et. al. */
- CWARN("%s: exp %p already connecting\n",
- export->exp_obd->obd_name, export);
+ cfs_spin_unlock(&export->exp_lock);
+ LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
+ export->exp_obd->obd_name, export,
+ libcfs_nid2str(req->rq_peer.nid));
class_export_put(export);
export = NULL;
rc = -EALREADY;
} else if (mds_conn && export->exp_connection) {
+ cfs_spin_unlock(&export->exp_lock);
if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
/* mds reconnected after failover */
- CWARN("%s: received MDS connection from NID %s,"
- " removing former export from NID %s\n",
+ LCONSOLE_WARN("%s: Received MDS connection from "
+ "%s, removing former export from %s\n",
target->obd_name, libcfs_nid2str(req->rq_peer.nid),
libcfs_nid2str(export->exp_connection->c_peer.nid));
else
/* new mds connection from the same nid */
- CWARN("%s: received new MDS connection from NID %s,"
- " removing former export from same NID\n",
+ LCONSOLE_WARN("%s: Received new MDS connection from "
+ "%s, removing former export from same NID\n",
target->obd_name, libcfs_nid2str(req->rq_peer.nid));
class_fail_export(export);
class_export_put(export);
req->rq_peer.nid != export->exp_connection->c_peer.nid &&
(lustre_msg_get_op_flags(req->rq_reqmsg) &
MSG_CONNECT_INITIAL)) {
+ cfs_spin_unlock(&export->exp_lock);
/* in mds failover we have static uuid but nid can be
* changed*/
- 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));
+ LCONSOLE_WARN("%s: Client %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));
rc = -EALREADY;
class_export_put(export);
export = NULL;
} else {
- cfs_spin_lock(&export->exp_lock);
- export->exp_connecting = 1;
- cfs_spin_unlock(&export->exp_lock);
- class_export_put(export);
- LASSERT(export->exp_obd == target);
+ export->exp_connecting = 1;
+ cfs_spin_unlock(&export->exp_lock);
+ LASSERT(export->exp_obd == target);
- rc = target_handle_reconnect(&conn, export, &cluuid);
- }
+ rc = target_handle_reconnect(&conn, export, &cluuid);
+ }
/* If we found an export, we already unlocked. */
if (!export) {
OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
} else if (req->rq_export == NULL &&
cfs_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, cfs_atomic_read(&export->exp_refcount));
+ LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
+ "still busy with %d references\n",
+ target->obd_name, cluuid.uuid,
+ libcfs_nid2str(req->rq_peer.nid),
+ cfs_atomic_read(&export->exp_refcount));
GOTO(out, rc = -EBUSY);
} else if (req->rq_export != NULL &&
(cfs_atomic_read(&export->exp_rpc_count) > 1)) {
/* the current connect rpc has increased exp_rpc_count */
- CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
- target->obd_name, cluuid.uuid,
- libcfs_nid2str(req->rq_peer.nid),
- export, cfs_atomic_read(&export->exp_rpc_count) - 1);
+ LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
+ "still busy with %d active RPCs\n",
+ target->obd_name, cluuid.uuid,
+ libcfs_nid2str(req->rq_peer.nid),
+ cfs_atomic_read(&export->exp_rpc_count) - 1);
cfs_spin_lock(&export->exp_lock);
if (req->rq_export->exp_conn_cnt <
lustre_msg_get_conn_cnt(req->rq_reqmsg))
cfs_spin_unlock(&export->exp_lock);
GOTO(out, rc = -EBUSY);
} else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
- 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);
+ if (!strstr(cluuid.uuid, "mdt"))
+ LCONSOLE_WARN("%s: Rejecting reconnect from the "
+ "known client %s (at %s) because it "
+ "is indicating it is a new client",
+ target->obd_name, cluuid.uuid,
+ libcfs_nid2str(req->rq_peer.nid));
GOTO(out, rc = -EALREADY);
} else {
OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
GOTO(out, rc);
}
- CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
+ CDEBUG(D_HA, "%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)cfs_time_current_sec(),
export ? (long)export->exp_last_request_time : 0);
- /* Tell the client if we're in recovery. */
- if (target->obd_recovering) {
- lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
- /* If this is the first time a client connects,
- reset the recovery timer */
- if (rc == 0)
- target_start_and_reset_recovery_timer(target, req,
- !export);
- }
+ /* If this is the first time a client connects, reset the recovery
+ * timer. Discard lightweight connections which might be local */
+ if (!lw_client && rc == 0 && target->obd_recovering)
+ check_and_start_recovery_timer(target, req, export == NULL);
/* We want to handle EALREADY but *not* -EALREADY from
* target_handle_reconnect(), return reconnection state in a flag */
client_nid = &req->rq_peer.nid;
if (export == NULL) {
- if (target->obd_recovering) {
+ /* allow lightweight connections during recovery */
+ if (target->obd_recovering && !lw_client) {
cfs_time_t t;
-
- t = cfs_timer_deadline(&target->obd_recovery_timer);
- t = cfs_time_sub(t, cfs_time_current());
- CERROR("%s: denying connection for new client %s (%s): "
- "%d clients in recovery for "CFS_TIME_T"s\n",
- target->obd_name,
- libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
- cfs_atomic_read(&target-> \
- obd_lock_replay_clients),
- cfs_duration_sec(t));
+ int c; /* connected */
+ int i; /* in progress */
+ int k; /* known */
+
+ c = cfs_atomic_read(&target->obd_connected_clients);
+ i = cfs_atomic_read(&target->obd_lock_replay_clients);
+ k = target->obd_max_recoverable_clients;
+ t = cfs_timer_deadline(&target->obd_recovery_timer);
+ t = cfs_time_sub(t, cfs_time_current());
+ t = cfs_duration_sec(t);
+ LCONSOLE_WARN("%s: Denying connection for new client "
+ "%s (at %s), waiting for all %d known "
+ "clients (%d recovered, %d in progress, "
+ "and %d unseen) to recover in %d:%.02d\n",
+ target->obd_name, cluuid.uuid,
+ libcfs_nid2str(req->rq_peer.nid), k,
+ c - i, i, k - c, (int)t / 60,
+ (int)t % 60);
rc = -EBUSY;
} else {
dont_check_exports:
} else {
rc = obd_reconnect(req->rq_svc_thread->t_env,
export, target, &cluuid, data, client_nid);
- if (rc == 0)
- /* prevous done via class_conn2export */
- class_export_get(export);
}
if (rc)
GOTO(out, rc);
+
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
+ /* 2.2.0 clients always swab nidtbl entries due to a bug, so server
+ * will do the swabbing for if the client is using the same endianness.
+ *
+ * This fixup is version-limited, because we don't want to carry the
+ * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we need
+ * interop with unpatched 2.2 clients. For newer clients, servers
+ * will never do MNE swabbing, let the client handle that. LU-1644 */
+ export->exp_need_mne_swab = !ptlrpc_req_need_swab(req) &&
+ !(data->ocd_connect_flags & OBD_CONNECT_MNE_SWAB);
+#else
+#warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and exp_need_mne_swab"
+#endif
+
+ LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
+ data->ocd_instance = target->u.obt.obt_instance;
+
/* Return only the parts of obd_connect_data that we understand, so the
* client knows that we don't understand the rest. */
if (data) {
- tmpdata = req_capsule_server_get(&req->rq_pill,
- &RMF_CONNECT_DATA);
- //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
- *tmpdata = *data;
+ tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
+ RCL_SERVER);
+ tmpdata = req_capsule_server_get(&req->rq_pill,
+ &RMF_CONNECT_DATA);
+ /* Don't use struct assignment here, because the client reply
+ * buffer may be smaller/larger than the local struct
+ * obd_connect_data. */
+ memcpy(tmpdata, data, min(tmpsize, size));
}
/* If all else goes well, this is our RPC return code. */
if (req->rq_export != NULL)
class_export_put(req->rq_export);
- req->rq_export = export;
+ /* request takes one export refcount */
+ req->rq_export = class_export_get(export);
cfs_spin_lock(&export->exp_lock);
if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
cfs_spin_unlock(&export->exp_lock);
- CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
+ CDEBUG(D_RPCTRACE, "%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));
&export->exp_nid_hash);
}
- cfs_spin_lock(&target->obd_recovery_task_lock);
- if (target->obd_recovering && !export->exp_in_recovery) {
+ if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
+ int has_transno;
+ __u64 transno = data->ocd_transno;
+
cfs_spin_lock(&export->exp_lock);
+ /* possible race with class_disconnect_stale_exports,
+ * export may be already in the eviction process */
+ if (export->exp_failed) {
+ cfs_spin_unlock(&export->exp_lock);
+ GOTO(out, rc = -ENODEV);
+ }
export->exp_in_recovery = 1;
export->exp_req_replay_needed = 1;
export->exp_lock_replay_needed = 1;
cfs_spin_unlock(&export->exp_lock);
- if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
- && (data->ocd_transno == 0))
+
+ has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
+ MSG_CONNECT_TRANSNO);
+ if (has_transno && transno == 0)
CWARN("Connect with zero transno!\n");
- 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++;
+ if (has_transno && transno > 0 &&
+ transno < target->obd_next_recovery_transno &&
+ transno > target->obd_last_committed) {
+ /* another way is to use cmpxchg() so it will be
+ * lock free */
+ cfs_spin_lock(&target->obd_recovery_task_lock);
+ if (transno < target->obd_next_recovery_transno)
+ target->obd_next_recovery_transno = transno;
+ cfs_spin_unlock(&target->obd_recovery_task_lock);
+ }
+
cfs_atomic_inc(&target->obd_req_replay_clients);
cfs_atomic_inc(&target->obd_lock_replay_clients);
- if (target->obd_connected_clients ==
+ if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
target->obd_max_recoverable_clients)
cfs_waitq_signal(&target->obd_next_transno_waitq);
}
- cfs_spin_unlock(&target->obd_recovery_task_lock);
- tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
- conn = *tmp;
- if (export->exp_imp_reverse != NULL) {
- /* destroyed import can be still referenced in ctxt */
- obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
- KEY_REVIMP_UPD, 0, NULL, NULL);
+ /* Tell the client we're in recovery, when client is involved in it. */
+ if (target->obd_recovering && !lw_client)
+ lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
- client_destroy_import(export->exp_imp_reverse);
- }
+ tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
+ conn = *tmp;
- /* 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);
- }
+ /* for the rest part, we return -ENOTCONN in case of errors
+ * in order to let client initialize connection again.
+ */
+ revimp = class_new_import(target);
+ if (revimp == NULL) {
+ CERROR("fail to alloc new reverse import.\n");
+ GOTO(out, rc = -ENOTCONN);
+ }
+
+ cfs_spin_lock(&export->exp_lock);
+ if (export->exp_imp_reverse != NULL) {
+ /* destroyed import can be still referenced in ctxt */
+ obd_set_info_async(req->rq_svc_thread->t_env, export,
+ sizeof(KEY_REVIMP_UPD), KEY_REVIMP_UPD,
+ 0, NULL, NULL);
+ tmp_imp = export->exp_imp_reverse;
+ }
+ export->exp_imp_reverse = revimp;
+ cfs_spin_unlock(&export->exp_lock);
revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
revimp->imp_client = &export->exp_obd->obd_ldlm_client;
else
revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
- rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
- if (rc) {
- CERROR("Failed to get sec for reverse import: %d\n", rc);
- export->exp_imp_reverse = NULL;
- class_destroy_import(revimp);
- }
+ rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
+ if (rc) {
+ CERROR("Failed to get sec for reverse import: %d\n", rc);
+ cfs_spin_lock(&export->exp_lock);
+ export->exp_imp_reverse = NULL;
+ cfs_spin_unlock(&export->exp_lock);
+ class_destroy_import(revimp);
+ }
+
+ class_import_put(revimp);
- class_import_put(revimp);
out:
+ if (tmp_imp != NULL)
+ client_destroy_import(tmp_imp);
if (export) {
cfs_spin_lock(&export->exp_lock);
export->exp_connecting = 0;
cfs_spin_unlock(&export->exp_lock);
+
+ class_export_put(export);
}
- if (targref)
+ if (targref) {
+ cfs_spin_lock(&target->obd_dev_lock);
+ target->obd_conn_inprogress--;
+ cfs_spin_unlock(&target->obd_dev_lock);
+
class_decref(targref, __FUNCTION__, cfs_current());
+ }
if (rc)
req->rq_status = rc;
RETURN(rc);
}
+EXPORT_SYMBOL(target_handle_connect);
int target_handle_disconnect(struct ptlrpc_request *req)
{
RETURN(0);
}
+EXPORT_SYMBOL(target_handle_disconnect);
void target_destroy_export(struct obd_export *exp)
{
- /* exports created from last_rcvd data, and "fake"
- exports created by lctl don't have an import */
- if (exp->exp_imp_reverse != NULL)
- client_destroy_import(exp->exp_imp_reverse);
-
- LASSERT(cfs_atomic_read(&exp->exp_locks_count) == 0);
- LASSERT(cfs_atomic_read(&exp->exp_rpc_count) == 0);
- LASSERT(cfs_atomic_read(&exp->exp_cb_count) == 0);
- LASSERT(cfs_atomic_read(&exp->exp_replay_count) == 0);
+ struct obd_import *imp = NULL;
+ /* exports created from last_rcvd data, and "fake"
+ exports created by lctl don't have an import */
+ cfs_spin_lock(&exp->exp_lock);
+ if (exp->exp_imp_reverse != NULL) {
+ imp = exp->exp_imp_reverse;
+ exp->exp_imp_reverse = NULL;
+ }
+ cfs_spin_unlock(&exp->exp_lock);
+ if (imp != NULL)
+ client_destroy_import(imp);
+
+ LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
+ LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
+ LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
+ LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
}
+EXPORT_SYMBOL(target_destroy_export);
/*
* Recovery functions
class_export_rpc_get(req->rq_export);
LASSERT(cfs_list_empty(&req->rq_list));
CFS_INIT_LIST_HEAD(&req->rq_replay_list);
+
/* increase refcount to keep request in queue */
- LASSERT(cfs_atomic_read(&req->rq_refcount));
cfs_atomic_inc(&req->rq_refcount);
/** let export know it has replays to be handled */
cfs_atomic_inc(&req->rq_export->exp_replay_count);
- /* release service thread while request is queued
- * we are moving the request from active processing
- * to waiting on the replay queue */
- ptlrpc_server_active_request_dec(req);
}
static void target_request_copy_put(struct ptlrpc_request *req)
{
LASSERT(cfs_list_empty(&req->rq_replay_list));
- LASSERT(cfs_atomic_read(&req->rq_export->exp_replay_count) > 0);
+ LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
+
cfs_atomic_dec(&req->rq_export->exp_replay_count);
class_export_rpc_put(req->rq_export);
- /* ptlrpc_server_drop_request() assumes the request is active */
- ptlrpc_server_active_request_inc(req);
ptlrpc_server_drop_request(req);
}
static void target_finish_recovery(struct obd_device *obd)
{
ENTRY;
- LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
- obd->obd_name);
+
+ /* only log a recovery message when recovery has occurred */
+ if (obd->obd_recovery_start) {
+ time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
+ obd->obd_recovery_start);
+ LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
+ "%d recovered and %d %s evicted.\n", obd->obd_name,
+ (int)elapsed_time / 60, (int)elapsed_time % 60,
+ obd->obd_max_recoverable_clients,
+ cfs_atomic_read(&obd->obd_connected_clients),
+ obd->obd_stale_clients,
+ obd->obd_stale_clients == 1 ? "was" : "were");
+ }
ldlm_reprocess_all_ns(obd->obd_namespace);
cfs_spin_lock(&obd->obd_recovery_task_lock);
target_request_copy_put(req);
}
}
-#endif
/* Called from a cleanup function if the device is being cleaned up
forcefully. The exports should all have been disconnected already,
EXIT;
}
+EXPORT_SYMBOL(target_cleanup_recovery);
/* obd_recovery_task_lock should be held */
void target_cancel_recovery_timer(struct obd_device *obd)
CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
cfs_timer_disarm(&obd->obd_recovery_timer);
}
+EXPORT_SYMBOL(target_cancel_recovery_timer);
-/* extend = 1 means require at least "duration" seconds left in the timer,
- extend = 0 means set the total duration (start_recovery_timer) */
-static void reset_recovery_timer(struct obd_device *obd, int duration,
- int extend)
+static void target_start_recovery_timer(struct obd_device *obd)
{
- cfs_time_t now = cfs_time_current_sec();
- cfs_duration_t left;
+ if (obd->obd_recovery_start != 0)
+ return;
- cfs_spin_lock(&obd->obd_recovery_task_lock);
+ cfs_spin_lock(&obd->obd_dev_lock);
if (!obd->obd_recovering || obd->obd_abort_recovery) {
- cfs_spin_unlock(&obd->obd_recovery_task_lock);
+ cfs_spin_unlock(&obd->obd_dev_lock);
return;
}
- left = cfs_time_sub(obd->obd_recovery_end, now);
+ LASSERT(obd->obd_recovery_timeout != 0);
- if (extend && (duration > left))
- obd->obd_recovery_timeout += duration - left;
- else if (!extend && (duration > obd->obd_recovery_timeout))
- /* Track the client's largest expected replay time */
- obd->obd_recovery_timeout = duration;
+ if (obd->obd_recovery_start != 0) {
+ cfs_spin_unlock(&obd->obd_dev_lock);
+ return;
+ }
- /* Hard limit of obd_recovery_time_hard which should not happen */
- if (obd->obd_recovery_timeout > obd->obd_recovery_time_hard)
- obd->obd_recovery_timeout = obd->obd_recovery_time_hard;
+ cfs_timer_arm(&obd->obd_recovery_timer,
+ cfs_time_shift(obd->obd_recovery_timeout));
+ obd->obd_recovery_start = cfs_time_current_sec();
+ cfs_spin_unlock(&obd->obd_dev_lock);
- obd->obd_recovery_end = obd->obd_recovery_start +
- obd->obd_recovery_timeout;
- if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
- cfs_time_before(now, obd->obd_recovery_end)) {
- left = cfs_time_sub(obd->obd_recovery_end, now);
- cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
- }
- cfs_spin_unlock(&obd->obd_recovery_task_lock);
- CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
- obd->obd_name, (unsigned)left);
+ LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
+ "or until %d client%s reconnect%s\n",
+ obd->obd_name,
+ obd->obd_recovery_timeout / 60,
+ obd->obd_recovery_timeout % 60,
+ obd->obd_max_recoverable_clients,
+ (obd->obd_max_recoverable_clients == 1) ? "" : "s",
+ (obd->obd_max_recoverable_clients == 1) ? "s": "");
}
-static void check_and_start_recovery_timer(struct obd_device *obd)
+/**
+ * extend recovery window.
+ *
+ * if @extend is true, extend recovery window to have @drt remaining at least;
+ * otherwise, make sure the recovery timeout value is not less than @drt.
+ */
+static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
{
- cfs_spin_lock(&obd->obd_recovery_task_lock);
- if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
- cfs_spin_unlock(&obd->obd_recovery_task_lock);
+ cfs_time_t now;
+ cfs_time_t end;
+ cfs_duration_t left;
+ int to;
+
+ cfs_spin_lock(&obd->obd_dev_lock);
+ if (!obd->obd_recovering || obd->obd_abort_recovery) {
+ cfs_spin_unlock(&obd->obd_dev_lock);
return;
}
- CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
- obd->obd_recovery_start = cfs_time_current_sec();
- cfs_spin_unlock(&obd->obd_recovery_task_lock);
+ LASSERT(obd->obd_recovery_start != 0);
- reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
+ now = cfs_time_current_sec();
+ to = obd->obd_recovery_timeout;
+ end = obd->obd_recovery_start + to;
+ left = cfs_time_sub(end, now);
+
+ if (extend && (drt > left)) {
+ to += drt - left;
+ } else if (!extend && (drt > to)) {
+ to = drt;
+ /* reduce drt by already passed time */
+ drt -= obd->obd_recovery_timeout - left;
+ }
+
+ if (to > obd->obd_recovery_time_hard)
+ to = obd->obd_recovery_time_hard;
+ if (obd->obd_recovery_timeout < to) {
+ obd->obd_recovery_timeout = to;
+ cfs_timer_arm(&obd->obd_recovery_timer,
+ cfs_time_shift(drt));
+ }
+ cfs_spin_unlock(&obd->obd_dev_lock);
+
+ CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
+ obd->obd_name, (unsigned)drt);
}
/* Reset the timer with each new client connection */
*/
static void
-target_start_and_reset_recovery_timer(struct obd_device *obd,
- struct ptlrpc_request *req,
- int new_client)
+check_and_start_recovery_timer(struct obd_device *obd,
+ struct ptlrpc_request *req,
+ int new_client)
{
int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
+ struct obd_device_target *obt = &obd->u.obt;
+ struct lustre_sb_info *lsi;
if (!new_client && service_time)
/* Teach server about old server's estimates, as first guess
* at how long new requests will take. */
- at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
+ at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
service_time);
- check_and_start_recovery_timer(obd);
+ target_start_recovery_timer(obd);
/* convert the service time to rpc timeout,
* reuse service_time to limit stack usage */
/* We expect other clients to timeout within service_time, then try
* to reconnect, then try the failover server. The max delay between
* connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
- service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
- INITIAL_CONNECT_TIMEOUT);
+ service_time += 2 * INITIAL_CONNECT_TIMEOUT;
+
+ LASSERT(obt->obt_magic == OBT_MAGIC);
+ lsi = s2lsi(obt->obt_sb);
+ if (!(lsi->lsi_flags | LDD_F_IR_CAPABLE))
+ service_time += 2 * (CONNECTION_SWITCH_MAX +
+ CONNECTION_SWITCH_INC);
if (service_time > obd->obd_recovery_timeout && !new_client)
- reset_recovery_timer(obd, service_time, 0);
+ extend_recovery_timer(obd, service_time, false);
}
-#ifdef __KERNEL__
-
/** Health checking routines */
static inline int exp_connect_healthy(struct obd_export *exp)
{
/** Checking routines for recovery */
static int check_for_clients(struct obd_device *obd)
{
+ unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
+
if (obd->obd_abort_recovery || obd->obd_recovery_expired)
return 1;
- LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
+ LASSERT(clnts <= obd->obd_max_recoverable_clients);
if (obd->obd_no_conn == 0 &&
- obd->obd_connected_clients + obd->obd_stale_clients ==
- obd->obd_max_recoverable_clients)
+ clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
return 1;
return 0;
}
req_transno = 0;
}
- connected = obd->obd_connected_clients;
+ connected = cfs_atomic_read(&obd->obd_connected_clients);
completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
queue_len = obd->obd_requests_queued_for_recovery;
next_transno = obd->obd_next_recovery_transno;
} else if (obd->obd_recovery_expired) {
obd->obd_recovery_expired = 0;
/** If some clients died being recovered, evict them */
- CDEBUG(D_WARNING,
- "recovery is timed out, evict stale exports\n");
+ LCONSOLE_WARN("%s: recovery is timed out, "
+ "evict stale exports\n", obd->obd_name);
/** evict cexports with no replay in queue, they are stalled */
class_disconnect_stale_exports(obd, health_check);
/** continue with VBR */
* reset timer, recovery will proceed with versions now,
* timeout is set just to handle reconnection delays
*/
- reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
+ extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
/** Wait for recovery events again, after evicting bad clients */
goto repeat;
}
lu_context_fini(&req->rq_recov_session);
/* don't reset timer for final stage */
if (!exp_finished(req->rq_export)) {
+ int to = obd_timeout;
+
/**
* Add request timeout to the recovery time so next request from
* this client may come in recovery time
*/
- reset_recovery_timer(class_exp2obd(req->rq_export),
- AT_OFF ? obd_timeout :
- lustre_msg_get_timeout(req->rq_reqmsg), 1);
+ if (!AT_OFF) {
+ struct ptlrpc_service_part *svcpt;
+
+ svcpt = req->rq_rqbd->rqbd_svcpt;
+ /* If the server sent early reply for this request,
+ * the client will recalculate the timeout according to
+ * current server estimate service time, so we will
+ * use the maxium timeout here for waiting the client
+ * sending the next req */
+ to = max((int)at_est2timeout(
+ at_get(&svcpt->scp_at_estimate)),
+ (int)lustre_msg_get_timeout(req->rq_reqmsg));
+ /* Add net_latency (see ptlrpc_replay_req) */
+ to += lustre_msg_get_service_time(req->rq_reqmsg);
+ }
+ extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
}
reqcopy_put:
RETURN(rc);
struct target_recovery_data *trd = &obd->obd_recovery_data;
unsigned long delta;
unsigned long flags;
- struct lu_env env;
- struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
+ struct lu_env *env;
+ struct ptlrpc_thread *thread = NULL;
int rc = 0;
ENTRY;
RECALC_SIGPENDING;
SIGNAL_MASK_UNLOCK(current, flags);
- rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
- if (rc)
+ OBD_ALLOC_PTR(thread);
+ if (thread == NULL)
+ RETURN(-ENOMEM);
+
+ OBD_ALLOC_PTR(env);
+ if (env == NULL) {
+ OBD_FREE_PTR(thread);
+ RETURN(-ENOMEM);
+ }
+
+ rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
+ if (rc) {
+ OBD_FREE_PTR(thread);
+ OBD_FREE_PTR(env);
RETURN(rc);
+ }
- thread->t_env = &env;
+ thread->t_env = env;
thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
- env.le_ctx.lc_thread = thread;
+ env->le_ctx.lc_thread = thread;
thread->t_data = NULL;
+ thread->t_watchdog = NULL;
- CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
+ CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
cfs_curproc_pid());
trd->trd_processing_task = cfs_curproc_pid();
target_finish_recovery(obd);
- lu_context_fini(&env.le_ctx);
+ lu_context_fini(&env->le_ctx);
trd->trd_processing_task = 0;
cfs_complete(&trd->trd_finishing);
+
+ OBD_FREE_PTR(thread);
+ OBD_FREE_PTR(env);
RETURN(rc);
}
cfs_init_completion(&trd->trd_finishing);
trd->trd_recovery_handler = handler;
- if (cfs_kernel_thread(target_recovery_thread, lut, 0) > 0) {
+ if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
cfs_wait_for_completion(&trd->trd_starting);
LASSERT(obd->obd_recovering != 0);
} else
cfs_wait_for_completion(&trd->trd_finishing);
}
}
+EXPORT_SYMBOL(target_stop_recovery_thread);
void target_recovery_fini(struct obd_device *obd)
{
" after %lds (%d clients connected)\n",
obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
cfs_time_current_sec()- obd->obd_recovery_start,
- obd->obd_connected_clients);
+ cfs_atomic_read(&obd->obd_connected_clients));
obd->obd_recovery_expired = 1;
cfs_waitq_signal(&obd->obd_next_transno_waitq);
obd->obd_recovery_start = 0;
obd->obd_recovery_end = 0;
- /* both values can be get from mount data already */
- if (obd->obd_recovery_timeout == 0)
- obd->obd_recovery_timeout = OBD_RECOVERY_TIME_SOFT;
- if (obd->obd_recovery_time_hard == 0)
- obd->obd_recovery_time_hard = OBD_RECOVERY_TIME_HARD;
cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
target_start_recovery_thread(lut, handler);
}
EXPORT_SYMBOL(target_recovery_init);
-#endif
+#endif /* __KERNEL__ */
static int target_process_req_flags(struct obd_device *obd,
struct ptlrpc_request *req)
LASSERT(exp != NULL);
if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
/* client declares he's ready to replay locks */
+ cfs_spin_lock(&exp->exp_lock);
if (exp->exp_req_replay_needed) {
- LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) >
- 0);
- cfs_spin_lock(&exp->exp_lock);
exp->exp_req_replay_needed = 0;
cfs_spin_unlock(&exp->exp_lock);
+
+ LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
cfs_atomic_dec(&obd->obd_req_replay_clients);
+ } else {
+ cfs_spin_unlock(&exp->exp_lock);
}
}
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 */
+ cfs_spin_lock(&exp->exp_lock);
if (exp->exp_lock_replay_needed) {
- LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) >
- 0);
- cfs_spin_lock(&exp->exp_lock);
exp->exp_lock_replay_needed = 0;
cfs_spin_unlock(&exp->exp_lock);
+
+ LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
cfs_atomic_dec(&obd->obd_lock_replay_clients);
+ } else {
+ cfs_spin_unlock(&exp->exp_lock);
}
}
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. */
+ * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
if (!transno) {
CFS_INIT_LIST_HEAD(&req->rq_list);
* Also, a resent, replayed request that has already been
* handled will pass through here and be processed immediately.
*/
- CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n",
- obd->obd_next_recovery_transno, transno);
+ CDEBUG(D_HA, "Next recovery transno: "LPU64
+ ", current: "LPU64", replaying\n",
+ obd->obd_next_recovery_transno, transno);
cfs_spin_lock(&obd->obd_recovery_task_lock);
if (transno < obd->obd_next_recovery_transno) {
/* Processing the queue right now, don't re-add. */
cfs_waitq_signal(&obd->obd_next_transno_waitq);
RETURN(0);
}
+EXPORT_SYMBOL(target_queue_recovery_request);
+
+int target_handle_ping(struct ptlrpc_request *req)
+{
+ obd_ping(req->rq_svc_thread->t_env, req->rq_export);
+ return req_capsule_server_pack(&req->rq_pill);
+}
+EXPORT_SYMBOL(target_handle_ping);
+
+void target_committed_to_req(struct ptlrpc_request *req)
+{
+ struct obd_export *exp = req->rq_export;
+
+ if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
+ lustre_msg_set_last_committed(req->rq_repmsg,
+ exp->exp_last_committed);
+ else
+ DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
+ "%d)", exp->exp_obd->obd_no_transno,
+ req->rq_repmsg == NULL);
+
+ CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
+ exp->exp_last_committed, req->rq_transno, req->rq_xid);
+}
+EXPORT_SYMBOL(target_committed_to_req);
+
+#endif /* HAVE_SERVER_SUPPORT */
/**
* Packs current SLV and Limit into \a req.
RETURN(0);
}
+EXPORT_SYMBOL(target_pack_pool_reply);
int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
{
}
if (unlikely(rc)) {
- DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
+ DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
req->rq_status = rc;
return (ptlrpc_send_error(req, 1));
} else {
void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
{
+ struct ptlrpc_service_part *svcpt;
int netrc;
struct ptlrpc_reply_state *rs;
- struct obd_device *obd;
struct obd_export *exp;
- struct ptlrpc_service *svc;
ENTRY;
if (req->rq_no_reply) {
return;
}
- svc = req->rq_rqbd->rqbd_service;
+ svcpt = req->rq_rqbd->rqbd_svcpt;
rs = req->rq_reply_state;
if (rs == NULL || !rs->rs_difficult) {
/* no notifiers */
/* must be an export if locks saved */
LASSERT (req->rq_export != NULL);
/* req/reply consistent */
- LASSERT (rs->rs_service == svc);
+ LASSERT(rs->rs_svcpt == svcpt);
/* "fresh" reply */
LASSERT (!rs->rs_scheduled);
LASSERT (cfs_list_empty(&rs->rs_exp_list));
exp = class_export_get (req->rq_export);
- obd = exp->exp_obd;
/* disable reply scheduling while I'm setting up */
rs->rs_scheduled = 1;
rs->rs_xid = req->rq_xid;
rs->rs_transno = req->rq_transno;
rs->rs_export = exp;
- rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
+ rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
netrc = target_send_reply_msg (req, rc, fail_id);
- cfs_spin_lock(&svc->srv_lock);
+ cfs_spin_lock(&svcpt->scp_rep_lock);
- cfs_atomic_inc(&svc->srv_n_difficult_replies);
+ cfs_atomic_inc(&svcpt->scp_nreps_difficult);
if (netrc != 0) {
/* error sending: reply is off the net. Also we need +1
* reply_out_callback leaves alone) */
rs->rs_on_net = 0;
ptlrpc_rs_addref(rs);
- cfs_atomic_inc (&svc->srv_outstanding_replies);
}
cfs_spin_lock(&rs->rs_lock);
CDEBUG(D_HA, "Schedule reply immediately\n");
ptlrpc_dispatch_difficult_reply(rs);
} else {
- cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
- rs->rs_scheduled = 0; /* allow notifier to schedule */
- }
- cfs_spin_unlock(&rs->rs_lock);
- cfs_spin_unlock(&svc->srv_lock);
- EXIT;
-}
-
-int target_handle_ping(struct ptlrpc_request *req)
-{
- obd_ping(req->rq_export);
- return req_capsule_server_pack(&req->rq_pill);
-}
-
-void target_committed_to_req(struct ptlrpc_request *req)
-{
- struct obd_export *exp = req->rq_export;
-
- if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
- lustre_msg_set_last_committed(req->rq_repmsg,
- exp->exp_last_committed);
- else
- DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
- "%d)", exp->exp_obd->obd_no_transno,
- req->rq_repmsg == NULL);
-
- CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
- exp->exp_last_committed, req->rq_transno, req->rq_xid);
-}
-EXPORT_SYMBOL(target_committed_to_req);
-
-int target_handle_qc_callback(struct ptlrpc_request *req)
-{
- struct obd_quotactl *oqctl;
- struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
-
- oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
- if (oqctl == NULL) {
- CERROR("Can't unpack obd_quotactl\n");
- RETURN(-EPROTO);
- }
-
- cli->cl_qchk_stat = oqctl->qc_stat;
-
- return 0;
+ cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active);
+ rs->rs_scheduled = 0; /* allow notifier to schedule */
+ }
+ cfs_spin_unlock(&rs->rs_lock);
+ cfs_spin_unlock(&svcpt->scp_rep_lock);
+ EXIT;
}
-
-#ifdef HAVE_QUOTA_SUPPORT
-int target_handle_dqacq_callback(struct ptlrpc_request *req)
-{
-#ifdef __KERNEL__
- struct obd_device *obd = req->rq_export->exp_obd;
- struct obd_device *master_obd = NULL, *lov_obd = NULL;
- struct obd_device_target *obt;
- struct lustre_quota_ctxt *qctxt;
- struct qunit_data *qdata = NULL;
- int rc = 0;
- ENTRY;
-
- if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
- RETURN(rc);
-
- rc = req_capsule_server_pack(&req->rq_pill);
- if (rc) {
- CERROR("packing reply failed!: rc = %d\n", rc);
- RETURN(rc);
- }
-
- LASSERT(req->rq_export);
-
- qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
- if (IS_ERR(qdata)) {
- rc = PTR_ERR(qdata);
- CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
- req->rq_status = rc;
- GOTO(out, rc);
- }
-
- /* we use the observer */
- if (obd_pin_observer(obd, &lov_obd) ||
- obd_pin_observer(lov_obd, &master_obd)) {
- CERROR("Can't find the observer, it is recovering\n");
- req->rq_status = -EAGAIN;
- GOTO(out, rc);
- }
-
- obt = &master_obd->u.obt;
- qctxt = &obt->obt_qctxt;
-
- if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
- /* quota_type has not been processed yet, return EAGAIN
- * until we know whether or not quotas are supposed to
- * be enabled */
- CDEBUG(D_QUOTA, "quota_type not processed yet, return "
- "-EAGAIN\n");
- req->rq_status = -EAGAIN;
- GOTO(out, rc);
- }
-
- cfs_down_read(&obt->obt_rwsem);
- if (qctxt->lqc_lqs_hash == NULL) {
- cfs_up_read(&obt->obt_rwsem);
- /* quota_type has not been processed yet, return EAGAIN
- * until we know whether or not quotas are supposed to
- * be enabled */
- CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
- "-EAGAIN\n");
- req->rq_status = -EAGAIN;
- GOTO(out, rc);
- }
-
- LASSERT(qctxt->lqc_handler);
- rc = qctxt->lqc_handler(master_obd, qdata,
- lustre_msg_get_opc(req->rq_reqmsg));
- cfs_up_read(&obt->obt_rwsem);
- if (rc && rc != -EDQUOT)
- CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
- "dqacq/dqrel failed! (rc:%d)\n", rc);
- req->rq_status = rc;
-
- rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
- if (rc < 0) {
- CERROR("Can't pack qunit_data(rc: %d)\n", rc);
- GOTO(out, rc);
- }
-
- /* Block the quota req. b=14840 */
- OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
- EXIT;
-
-out:
- if (master_obd)
- obd_unpin_observer(lov_obd);
- if (lov_obd)
- obd_unpin_observer(obd);
-
- rc = ptlrpc_reply(req);
- return rc;
-#else
- return 0;
-#endif /* !__KERNEL__ */
-}
-#endif /* HAVE_QUOTA_SUPPORT */
+EXPORT_SYMBOL(target_send_reply);
ldlm_mode_t lck_compat_array[] = {
[LCK_EX] LCK_COMPAT_EX,
if (((int)error) < 0) /* cast to signed type */
result = error; /* as ldlm_error_t can be unsigned */
else {
- CERROR("Invalid DLM result code: %i\n", error);
+ CERROR("Invalid DLM result code: %d\n", error);
result = -EPROTO;
}
}
CERROR("dumping locks for export %p,"
"ignore if the unmount doesn't hang\n", exp);
cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
- ldlm_lock_dump(D_ERROR, lock, 0);
+ LDLM_ERROR(lock, "lock:");
}
cfs_spin_unlock(&exp->exp_locks_list_guard);
}
#endif
+
+#ifdef HAVE_SERVER_SUPPORT
+static int target_bulk_timeout(void *data)
+{
+ ENTRY;
+ /* We don't fail the connection here, because having the export
+ * killed makes the (vital) call to commitrw very sad.
+ */
+ RETURN(1);
+}
+
+static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
+{
+ return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
+}
+
+int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
+ struct l_wait_info *lwi)
+{
+ struct ptlrpc_request *req = desc->bd_req;
+ int rc = 0;
+ ENTRY;
+
+ /* Check if there is eviction in progress, and if so, wait for
+ * it to finish */
+ if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
+ *lwi = LWI_INTR(NULL, NULL);
+ rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
+ !cfs_atomic_read(&exp->exp_obd->
+ obd_evict_inprogress),
+ lwi);
+ }
+
+ /* Check if client was evicted or tried to reconnect already */
+ if (exp->exp_failed || exp->exp_abort_active_req) {
+ rc = -ENOTCONN;
+ } else {
+ if (desc->bd_type == BULK_PUT_SINK)
+ rc = sptlrpc_svc_wrap_bulk(req, desc);
+ if (rc == 0)
+ rc = ptlrpc_start_bulk_transfer(desc);
+ }
+
+ if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
+ ptlrpc_abort_bulk(desc);
+ } else if (rc == 0) {
+ time_t start = cfs_time_current_sec();
+ do {
+ long timeoutl = req->rq_deadline - cfs_time_current_sec();
+ cfs_duration_t timeout = timeoutl <= 0 ?
+ CFS_TICK : cfs_time_seconds(timeoutl);
+ *lwi = LWI_TIMEOUT_INTERVAL(timeout,
+ cfs_time_seconds(1),
+ target_bulk_timeout,
+ desc);
+ rc = l_wait_event(desc->bd_waitq,
+ !ptlrpc_server_bulk_active(desc) ||
+ exp->exp_failed ||
+ exp->exp_abort_active_req,
+ lwi);
+ LASSERT(rc == 0 || rc == -ETIMEDOUT);
+ /* Wait again if we changed deadline */
+ } while ((rc == -ETIMEDOUT) &&
+ (req->rq_deadline > cfs_time_current_sec()));
+
+ if (rc == -ETIMEDOUT) {
+ DEBUG_REQ(D_ERROR, req,
+ "timeout on bulk %s after %ld%+lds",
+ bulk2type(desc),
+ req->rq_deadline - start,
+ cfs_time_current_sec() -
+ req->rq_deadline);
+ ptlrpc_abort_bulk(desc);
+ } else if (exp->exp_failed) {
+ DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
+ bulk2type(desc));
+ rc = -ENOTCONN;
+ ptlrpc_abort_bulk(desc);
+ } else if (exp->exp_abort_active_req) {
+ DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
+ bulk2type(desc));
+ /* we don't reply anyway */
+ rc = -ETIMEDOUT;
+ ptlrpc_abort_bulk(desc);
+ } else if (!desc->bd_success ||
+ desc->bd_nob_transferred != desc->bd_nob) {
+ DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
+ desc->bd_success ?
+ "truncated" : "network error on",
+ bulk2type(desc),
+ desc->bd_nob_transferred,
+ desc->bd_nob);
+ /* XXX should this be a different errno? */
+ rc = -ETIMEDOUT;
+ } else if (desc->bd_type == BULK_GET_SINK) {
+ rc = sptlrpc_svc_unwrap_bulk(req, desc);
+ }
+ } else {
+ DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
+ bulk2type(desc), rc);
+ }
+
+ RETURN(rc);
+}
+EXPORT_SYMBOL(target_bulk_io);
+
+#endif /* HAVE_SERVER_SUPPORT */