X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lib.c;h=47279bae1ff6d957220b03c69076b3edd97d1245;hp=3acbdd7bb2d83e50ddf50aff1a64a9377575bb53;hb=c4220f43d3aa7c938808da60a925d67904cc3ece;hpb=18d78c77953017e5a76cd10bc74a0d078217a626 diff --git a/lustre/ldlm/ldlm_lib.c b/lustre/ldlm/ldlm_lib.c index 3acbdd7..47279ba 100644 --- a/lustre/ldlm/ldlm_lib.c +++ b/lustre/ldlm/ldlm_lib.c @@ -1,6 +1,4 @@ -/* -*- 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. @@ -29,16 +27,20 @@ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2012, Whamcloud, Inc. + * Copyright (c) 2010, 2013, Intel Corporation. */ /* * 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 +/** + * This file deals with various client/target related logic including recovery. + * + * TODO: This code more logically belongs in the ptlrpc module than in ldlm and + * should be moved. + */ + #define DEBUG_SUBSYSTEM S_LDLM #ifdef __KERNEL__ @@ -47,14 +49,14 @@ # include #endif #include -#include +#include #include #include #include #include "ldlm_internal.h" -/* @priority: if non-zero, move the selected to the list head - * @create: if zero, only search in existed connections +/* @priority: If non-zero, move the selected connection to the list head. + * @create: If zero, only search in existing connections. */ static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid, int priority, int create) @@ -82,7 +84,7 @@ static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid, } } - cfs_spin_lock(&imp->imp_lock); + spin_lock(&imp->imp_lock); cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) { if (obd_uuid_equals(uuid, &item->oic_uuid)) { if (priority) { @@ -94,11 +96,11 @@ static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid, CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n", imp, imp->imp_obd->obd_name, uuid->uuid, (priority ? ", moved to head" : "")); - cfs_spin_unlock(&imp->imp_lock); + spin_unlock(&imp->imp_lock); GOTO(out_free, rc = 0); } } - /* not found */ + /* No existing import connection found for \a uuid. */ if (create) { imp_conn->oic_conn = ptlrpc_conn; imp_conn->oic_uuid = *uuid; @@ -112,11 +114,11 @@ static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid, imp, imp->imp_obd->obd_name, uuid->uuid, (priority ? "head" : "tail")); } else { - cfs_spin_unlock(&imp->imp_lock); - GOTO(out_free, rc = -ENOENT); - } + spin_unlock(&imp->imp_lock); + GOTO(out_free, rc = -ENOENT); + } - cfs_spin_unlock(&imp->imp_lock); + spin_unlock(&imp->imp_lock); RETURN(0); out_free: if (imp_conn) @@ -136,15 +138,16 @@ int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid, { 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) { - struct obd_import_conn *imp_conn; - struct obd_export *dlmexp; - int rc = -ENOENT; - ENTRY; + struct obd_import_conn *imp_conn; + struct obd_export *dlmexp; + int rc = -ENOENT; + ENTRY; - cfs_spin_lock(&imp->imp_lock); + spin_lock(&imp->imp_lock); if (cfs_list_empty(&imp->imp_conn_list)) { LASSERT(!imp->imp_connection); GOTO(out, rc); @@ -155,7 +158,6 @@ int client_import_del_conn(struct obd_import *imp, struct obd_uuid *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); @@ -186,41 +188,42 @@ int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid) break; } out: - cfs_spin_unlock(&imp->imp_lock); - if (rc == -ENOENT) - CERROR("connection %s not found\n", uuid->uuid); - RETURN(rc); + spin_unlock(&imp->imp_lock); + if (rc == -ENOENT) + 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. + * Find conn UUID by peer NID. \a peer is a server NID. This function is used + * to find a conn uuid of \a imp which can reach \a peer. */ int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer, - struct obd_uuid *uuid) + struct obd_uuid *uuid) { - struct obd_import_conn *conn; - int rc = -ENOENT; - ENTRY; + struct obd_import_conn *conn; + int rc = -ENOENT; + ENTRY; - cfs_spin_lock(&imp->imp_lock); + 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 */ + /* 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); + spin_unlock(&imp->imp_lock); + RETURN(rc); } EXPORT_SYMBOL(client_import_find_conn); void client_destroy_import(struct obd_import *imp) { - /* drop security policy instance after all rpc finished/aborted - * to let all busy contexts be released. */ + /* Drop security policy instance after all RPCs have finished/aborted + * to let all busy contexts be released. */ class_import_get(imp); class_destroy_import(imp); sptlrpc_import_sec_put(imp); @@ -228,7 +231,30 @@ void client_destroy_import(struct obd_import *imp) } EXPORT_SYMBOL(client_destroy_import); -/* configure an RPC client OBD device +/** + * Check whether or not the OSC is on MDT. + * 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. * * lcfg parameters: * 1 - client UUID @@ -248,22 +274,37 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) /* 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)) { - 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 if (!strcmp(name, LUSTRE_MDC_NAME)) { - 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; - + if (!strcmp(name, LUSTRE_OSC_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 if (!strcmp(name, LUSTRE_MDC_NAME) || + !strcmp(name, LUSTRE_LWP_NAME)) { + 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_OSP_NAME)) { + if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) { + /* OSP_on_MDT for other MDTs */ + connect_op = MDS_CONNECT; + cli->cl_sp_to = LUSTRE_SP_MDT; + ns_type = LDLM_NS_TYPE_MDC; + rq_portal = OUT_PORTAL; + } else { + /* OSP on MDT for OST */ + connect_op = OST_CONNECT; + cli->cl_sp_to = LUSTRE_SP_OST; + ns_type = LDLM_NS_TYPE_OSC; + rq_portal = OST_REQUEST_PORTAL; + } + rp_portal = OSC_REPLY_PORTAL; + cli->cl_sp_me = LUSTRE_SP_CLI; } else if (!strcmp(name, LUSTRE_MGC_NAME)) { rq_portal = MGS_REQUEST_PORTAL; rp_portal = MGC_REPLY_PORTAL; @@ -272,8 +313,7 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) cli->cl_sp_to = LUSTRE_SP_MGS; cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID; ns_type = LDLM_NS_TYPE_MGC; - - } else { + } else { CERROR("unknown client OBD type \"%s\", can't setup\n", name); RETURN(-EINVAL); @@ -299,8 +339,8 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) RETURN(-EINVAL); } - cfs_init_rwsem(&cli->cl_sem); - cfs_sema_init(&cli->cl_mgc_sem, 1); + init_rwsem(&cli->cl_sem); + sema_init(&cli->cl_mgc_sem, 1); cli->cl_conn_count = 0; memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2), min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2), @@ -308,30 +348,42 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) cli->cl_dirty = 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 > cfs_num_physpages / 8) - cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3); + /* 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 >> PAGE_CACHE_SHIFT > totalram_pages / 8) + cli->cl_dirty_max = totalram_pages << (PAGE_CACHE_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_hp_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; - - cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock); - cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock); - cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock); - 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); - cfs_waitq_init(&cli->cl_destroy_waitq); - cfs_atomic_set(&cli->cl_destroy_in_flight, 0); + 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; + + 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); + spin_lock_init(&cli->cl_read_offset_hist.oh_lock); + 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_atomic_set(&cli->cl_unstable_count, 0); + + init_waitqueue_head(&cli->cl_destroy_waitq); + cfs_atomic_set(&cli->cl_destroy_in_flight, 0); #ifdef ENABLE_CHECKSUM - /* Turn on checksumming by default. */ - cli->cl_checksum = 1; + /* Turn on checksumming by default. */ + cli->cl_checksum = 1; /* * The supported checksum types will be worked out at connect time * Set cl_chksum* to CRC32 for now to avoid returning screwed info @@ -341,23 +393,31 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) #endif cfs_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 (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) { - cli->cl_max_rpcs_in_flight = 2; - } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) { - cli->cl_max_rpcs_in_flight = 3; - } 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; + /* This value may be reduced at connect time in + * ptlrpc_connect_interpret() . We initialize it to only + * 1MB until we know what the performance looks like. + * In the future this should likely be increased. LU-1431 */ + cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES, + LNET_MTU >> PAGE_CACHE_SHIFT); + + /* set cl_chunkbits default value to PAGE_CACHE_SHIFT, + * it will be updated at OSC connection time. */ + cli->cl_chunkbits = PAGE_CACHE_SHIFT; + + if (!strcmp(name, LUSTRE_MDC_NAME)) { + cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT; + } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) { + cli->cl_max_rpcs_in_flight = 2; + } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) { + cli->cl_max_rpcs_in_flight = 3; + } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) { + cli->cl_max_rpcs_in_flight = 4; + } else { + 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); @@ -372,7 +432,6 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) GOTO(err_ldlm, rc = -ENOENT); imp->imp_client = &obddev->obd_ldlm_client; imp->imp_connect_op = connect_op; - 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); @@ -393,9 +452,9 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) CDEBUG(D_HA, "marking %s %s->%s as inactive\n", name, obddev->obd_name, cli->cl_target_uuid.uuid); - cfs_spin_lock(&imp->imp_lock); - imp->imp_deactive = 1; - cfs_spin_unlock(&imp->imp_lock); + spin_lock(&imp->imp_lock); + imp->imp_deactive = 1; + spin_unlock(&imp->imp_lock); } } @@ -421,19 +480,22 @@ err: RETURN(rc); } +EXPORT_SYMBOL(client_obd_setup); int client_obd_cleanup(struct obd_device *obddev) { - ENTRY; + ENTRY; - ldlm_namespace_free_post(obddev->obd_namespace); - obddev->obd_namespace = NULL; + ldlm_namespace_free_post(obddev->obd_namespace); + obddev->obd_namespace = NULL; - LASSERT(obddev->u.cli.cl_import == NULL); + obd_cleanup_client_import(obddev); + LASSERT(obddev->u.cli.cl_import == NULL); - ldlm_put_ref(); - RETURN(0); + 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, @@ -441,15 +503,15 @@ int client_connect_import(const struct lu_env *env, struct obd_device *obd, struct obd_uuid *cluuid, struct obd_connect_data *data, void *localdata) { - struct client_obd *cli = &obd->u.cli; - struct obd_import *imp = cli->cl_import; - struct obd_connect_data *ocd; - struct lustre_handle conn = { 0 }; - int rc; - ENTRY; + struct client_obd *cli = &obd->u.cli; + struct obd_import *imp = cli->cl_import; + struct obd_connect_data *ocd; + struct lustre_handle conn = { 0 }; + int rc; + ENTRY; *exp = NULL; - cfs_down_write(&cli->cl_sem); + down_write(&cli->cl_sem); if (cli->cl_conn_count > 0 ) GOTO(out_sem, rc = -EALREADY); @@ -498,10 +560,11 @@ out_ldlm: *exp = NULL; } out_sem: - cfs_up_write(&cli->cl_sem); + up_write(&cli->cl_sem); - return rc; + return rc; } +EXPORT_SYMBOL(client_connect_import); int client_disconnect_export(struct obd_export *exp) { @@ -520,7 +583,7 @@ int client_disconnect_export(struct obd_export *exp) cli = &obd->u.cli; imp = cli->cl_import; - cfs_down_write(&cli->cl_sem); + down_write(&cli->cl_sem); CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name, cli->cl_conn_count); @@ -534,12 +597,12 @@ int client_disconnect_export(struct obd_export *exp) if (cli->cl_conn_count) GOTO(out_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. */ - cfs_spin_lock(&imp->imp_lock); - imp->imp_deactive = 1; - cfs_spin_unlock(&imp->imp_lock); + /* 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 @@ -553,29 +616,28 @@ int client_disconnect_export(struct obd_export *exp) ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force); } - /* - * there's no need to hold sem during disconnecting an import, - * and actually it may cause deadlock in gss. - */ - cfs_up_write(&cli->cl_sem); - rc = ptlrpc_disconnect_import(imp, 0); - cfs_down_write(&cli->cl_sem); + /* There's no need to hold sem while disconnecting an import, + * and it may actually cause deadlock in GSS. */ + up_write(&cli->cl_sem); + rc = ptlrpc_disconnect_import(imp, 0); + down_write(&cli->cl_sem); ptlrpc_invalidate_import(imp); EXIT; - out_disconnect: - /* use server style - class_disconnect should be always called for - * o_disconnect */ +out_disconnect: + /* Use server style - class_disconnect should be always called for + * o_disconnect. */ err = class_disconnect(exp); if (!rc && err) rc = err; - cfs_up_write(&cli->cl_sem); + up_write(&cli->cl_sem); RETURN(rc); } +EXPORT_SYMBOL(client_disconnect_export); #ifdef HAVE_SERVER_SUPPORT int server_disconnect_export(struct obd_export *exp) @@ -583,34 +645,37 @@ int server_disconnect_export(struct obd_export *exp) int rc; ENTRY; - /* Disconnect early so that clients can't keep using export */ - rc = class_disconnect(exp); - /* close import for avoid sending any requests */ - if (exp->exp_imp_reverse) - ptlrpc_cleanup_imp(exp->exp_imp_reverse); + /* Disconnect early so that clients can't keep using export. */ + rc = class_disconnect(exp); + /* Close import to avoid sending any requests. */ + if (exp->exp_imp_reverse) + ptlrpc_cleanup_imp(exp->exp_imp_reverse); - if (exp->exp_obd->obd_namespace != NULL) - ldlm_cancel_locks_for_export(exp); + if (exp->exp_obd->obd_namespace != NULL) + ldlm_cancel_locks_for_export(exp); /* complete all outstanding replies */ - cfs_spin_lock(&exp->exp_lock); - while (!cfs_list_empty(&exp->exp_outstanding_replies)) { - 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; - - cfs_spin_lock(&svc->srv_rs_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_rs_lock); - } - cfs_spin_unlock(&exp->exp_lock); + spin_lock(&exp->exp_lock); + while (!cfs_list_empty(&exp->exp_outstanding_replies)) { + struct ptlrpc_reply_state *rs = + cfs_list_entry(exp->exp_outstanding_replies.next, + struct ptlrpc_reply_state, rs_exp_list); + struct ptlrpc_service_part *svcpt = rs->rs_svcpt; - RETURN(rc); + spin_lock(&svcpt->scp_rep_lock); + + cfs_list_del_init(&rs->rs_exp_list); + spin_lock(&rs->rs_lock); + ptlrpc_schedule_difficult_reply(rs); + spin_unlock(&rs->rs_lock); + + spin_unlock(&svcpt->scp_rep_lock); + } + spin_unlock(&exp->exp_lock); + + RETURN(rc); } +EXPORT_SYMBOL(server_disconnect_export); /* -------------------------------------------------------------------------- * from old lib/target.c @@ -657,14 +722,13 @@ static int target_handle_reconnect(struct lustre_handle *conn, * doing a valid reconnect from the same client. */ RETURN(EALREADY); } else { - LCONSOLE_WARN("%s: The server has already connected " - "client %s (at %s) with handle " LPX64 - ", rejecting a client with the same " - "uuid trying to reconnect with " - "handle " LPX64, target->obd_name, - obd_uuid2str(&exp->exp_client_uuid), - obd_export_nid2str(exp), - 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 @@ -687,10 +751,10 @@ void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data, CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n", obd->obd_name, exp->exp_client_uuid.uuid); - cfs_spin_lock(&exp->exp_lock); - exp->exp_need_sync = 0; - cfs_spin_unlock(&exp->exp_lock); - class_export_cb_put(exp); + spin_lock(&exp->exp_lock); + exp->exp_need_sync = 0; + spin_unlock(&exp->exp_lock); + class_export_cb_put(exp); } EXPORT_SYMBOL(target_client_add_cb); @@ -708,9 +772,10 @@ check_and_start_recovery_timer(struct obd_device *obd, 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; @@ -720,11 +785,11 @@ int target_handle_connect(struct ptlrpc_request *req) int rc = 0; char *target_start; int target_len; - int mds_conn = 0; + 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); @@ -739,28 +804,45 @@ int target_handle_connect(struct ptlrpc_request *req) if (!target) target = class_name2obd(str); - if (!target || target->obd_stopping || !target->obd_set_up) { - 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 ? - "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, "%s: not available for connect " + "from %s (no target). If you are running " + "an HA pair check that the target is " + "mounted on the other server.\n", str, + libcfs_nid2str(req->rq_peer.nid)); + GOTO(out, rc = -ENODEV); + } + + spin_lock(&target->obd_dev_lock); + if (target->obd_stopping || !target->obd_set_up) { + spin_unlock(&target->obd_dev_lock); + + deuuidify(str, NULL, &target_start, &target_len); + LCONSOLE_INFO("%.*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 " - "from %s\n", target->obd_name, - libcfs_nid2str(req->rq_peer.nid)); - GOTO(out, rc = -EAGAIN); - } + spin_unlock(&target->obd_dev_lock); + + CDEBUG(D_INFO, "%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, __FUNCTION__, cfs_current()); + /* 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, __FUNCTION__, current); + target->obd_conn_inprogress++; + spin_unlock(&target->obd_dev_lock); str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID); if (str == NULL) { @@ -770,9 +852,9 @@ int target_handle_connect(struct ptlrpc_request *req) obd_str2uuid(&cluuid, str); - /* XXX extract a nettype and format accordingly */ - switch (sizeof(lnet_nid_t)) { - /* NB the casts only avoid compiler warnings */ + /* XXX Extract a nettype and format accordingly. */ + 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); @@ -801,38 +883,37 @@ int target_handle_connect(struct ptlrpc_request *req) 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 = req_capsule_server_sized_get(&req->rq_pill, - &RMF_CONNECT_DATA, - 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_LIBCLIENT) { + 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 = req_capsule_server_sized_get(&req->rq_pill, + &RMF_CONNECT_DATA, + 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) && (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)) @@ -842,23 +923,28 @@ int target_handle_connect(struct ptlrpc_request *req) if (!export) goto no_export; - /* we've found an export in the hash */ - if (export->exp_connecting) { /* bug 9635, et. al. */ - 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) { + /* We've found an export in the hash. */ + + spin_lock(&export->exp_lock); + + if (export->exp_connecting) { /* bug 9635, et. al. */ + 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) { + spin_unlock(&export->exp_lock); if (req->rq_peer.nid != export->exp_connection->c_peer.nid) - /* mds reconnected after failover */ - 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 */ + /* MDS reconnected after failover. */ + 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. */ 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)); @@ -870,8 +956,8 @@ int target_handle_connect(struct ptlrpc_request *req) req->rq_peer.nid != export->exp_connection->c_peer.nid && (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)) { - /* in mds failover we have static uuid but nid can be - * changed*/ + spin_unlock(&export->exp_lock); + /* In MDS failover we have static UUID but NID can change. */ LCONSOLE_WARN("%s: Client %s seen on new nid %s when " "existing nid %s is already connected\n", target->obd_name, cluuid.uuid, @@ -882,13 +968,12 @@ int target_handle_connect(struct ptlrpc_request *req) class_export_put(export); export = NULL; } else { - cfs_spin_lock(&export->exp_lock); - export->exp_connecting = 1; - cfs_spin_unlock(&export->exp_lock); - LASSERT(export->exp_obd == target); + export->exp_connecting = 1; + 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) { @@ -902,21 +987,6 @@ no_export: 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 */ - 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)) - /* try to abort active requests */ - req->rq_export->exp_abort_active_req = 1; - cfs_spin_unlock(&export->exp_lock); - GOTO(out, rc = -EBUSY); } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) { if (!strstr(cluuid.uuid, "mdt")) LCONSOLE_WARN("%s: Rejecting reconnect from the " @@ -939,13 +1009,13 @@ no_export: export, (long)cfs_time_current_sec(), export ? (long)export->exp_last_request_time : 0); - /* If this is the first time a client connects, reset the recovery - * timer */ - if (rc == 0 && target->obd_recovering) - check_and_start_recovery_timer(target, req, export == NULL); + /* 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 */ + /* 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; @@ -953,50 +1023,68 @@ no_export: LASSERT(rc == 0); } - /* Tell the client if we support replayable requests */ + /* Tell the client if we support replayable requests. */ if (target->obd_replayable) lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE); 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()); - t = cfs_duration_sec(t); - LCONSOLE_WARN("%s: Denying connection for new client " - "%s (at %s), waiting for %d clients in " - "recovery for %d:%.02d\n", - target->obd_name, - libcfs_nid2str(req->rq_peer.nid), - cluuid.uuid, - cfs_atomic_read(&target-> \ - obd_lock_replay_clients), - (int)t / 60, (int)t % 60); + int c; /* connected */ + int i; /* in progress */ + int k; /* known */ + int s; /* stale/evicted */ + + c = cfs_atomic_read(&target->obd_connected_clients); + i = cfs_atomic_read(&target->obd_lock_replay_clients); + k = target->obd_max_recoverable_clients; + s = target->obd_stale_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 evicted) to recover in %d:%.02d\n", + target->obd_name, cluuid.uuid, + libcfs_nid2str(req->rq_peer.nid), k, + c - i, i, s, (int)t / 60, + (int)t % 60); rc = -EBUSY; } else { dont_check_exports: rc = obd_connect(req->rq_svc_thread->t_env, &export, target, &cluuid, data, client_nid); - if (rc == 0) { + if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG)) + lustre_msg_add_op_flags(req->rq_repmsg, + MSG_CONNECT_RECOVERING); + if (rc == 0) conn.cookie = export->exp_handle.h_cookie; - /* LU-1092 reconnect put export refcount in the - * end, connect needs take one here too. */ - class_export_get(export); - } } } 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; @@ -1025,16 +1113,13 @@ dont_check_exports: * XXX this will go away when shaver stops sending the "connect" handle * in the real "remote handle" field of the request --phik 24 Apr 2003 */ - if (req->rq_export != NULL) - class_export_put(req->rq_export); - - req->rq_export = export; + ptlrpc_request_change_export(req, 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); - CDEBUG(D_RPCTRACE, "%s: %s already connected at higher " - "conn_cnt: %d > %d\n", + spin_lock(&export->exp_lock); + if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) { + spin_unlock(&export->exp_lock); + CDEBUG(D_RPCTRACE, "%s: %s already connected at greater " + "or equal 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)); @@ -1043,22 +1128,21 @@ dont_check_exports: } LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0); export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg); - export->exp_abort_active_req = 0; - /* request from liblustre? Don't evict it for not pinging. */ + /* Don't evict liblustre clients for not pinging. */ if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) { export->exp_libclient = 1; - cfs_spin_unlock(&export->exp_lock); + spin_unlock(&export->exp_lock); - cfs_spin_lock(&target->obd_dev_lock); - cfs_list_del_init(&export->exp_obd_chain_timed); - cfs_spin_unlock(&target->obd_dev_lock); - } else { - cfs_spin_unlock(&export->exp_lock); - } + spin_lock(&target->obd_dev_lock); + cfs_list_del_init(&export->exp_obd_chain_timed); + spin_unlock(&target->obd_dev_lock); + } else { + spin_unlock(&export->exp_lock); + } if (export->exp_connection != NULL) { - /* Check to see if connection came from another NID */ + /* Check to see if connection came from another NID. */ if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) && !cfs_hlist_unhashed(&export->exp_nid_hash)) cfs_hash_del(export->exp_obd->obd_nid_hash, @@ -1076,18 +1160,22 @@ dont_check_exports: &export->exp_connection->c_peer.nid, &export->exp_nid_hash); } - /** - class_disconnect->class_export_recovery_cleanup() race - */ - 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); - export->exp_in_recovery = 1; - export->exp_req_replay_needed = 1; - export->exp_lock_replay_needed = 1; - cfs_spin_unlock(&export->exp_lock); + spin_lock(&export->exp_lock); + /* possible race with class_disconnect_stale_exports, + * export may be already in the eviction process */ + if (export->exp_failed) { + 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; + spin_unlock(&export->exp_lock); has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO); @@ -1097,44 +1185,40 @@ dont_check_exports: 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); + /* Another way is to use cmpxchg() to be lock-free. */ + spin_lock(&target->obd_recovery_task_lock); + if (transno < target->obd_next_recovery_transno) + target->obd_next_recovery_transno = transno; + spin_unlock(&target->obd_recovery_task_lock); } - cfs_atomic_inc(&target->obd_req_replay_clients); - cfs_atomic_inc(&target->obd_lock_replay_clients); - if (cfs_atomic_inc_return(&target->obd_connected_clients) == - target->obd_max_recoverable_clients) - cfs_waitq_signal(&target->obd_next_transno_waitq); - } + cfs_atomic_inc(&target->obd_req_replay_clients); + cfs_atomic_inc(&target->obd_lock_replay_clients); + if (cfs_atomic_inc_return(&target->obd_connected_clients) == + target->obd_max_recoverable_clients) + wake_up(&target->obd_next_transno_waitq); + } /* Tell the client we're in recovery, when client is involved in it. */ - if (target->obd_recovering) + if (target->obd_recovering && !lw_client) lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING); 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); + /* Return -ENOTCONN in case of errors to let client reconnect. */ + revimp = class_new_import(target); + if (revimp == NULL) { + CERROR("fail to alloc new reverse import.\n"); + GOTO(out, rc = -ENOTCONN); + } - client_destroy_import(export->exp_imp_reverse); - } - - /* 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); - } + spin_lock(&export->exp_lock); + if (export->exp_imp_reverse != NULL) + /* destroyed import can be still referenced in ctxt */ + tmp_imp = export->exp_imp_reverse; + export->exp_imp_reverse = revimp; + spin_unlock(&export->exp_lock); revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection); revimp->imp_client = &export->exp_obd->obd_ldlm_client; @@ -1142,44 +1226,55 @@ dont_check_exports: revimp->imp_dlm_fake = 1; revimp->imp_state = LUSTRE_IMP_FULL; - /* unknown versions will be caught in - * ptlrpc_handle_server_req_in->lustre_unpack_msg() */ + /* Unknown versions will be caught in + * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */ revimp->imp_msg_magic = req->rq_reqmsg->lm_magic; - if ((export->exp_connect_flags & OBD_CONNECT_AT) && - (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1)) - revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT; - else - revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; + if ((data->ocd_connect_flags & OBD_CONNECT_AT) && + (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1)) + revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT; + else + revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; - if ((export->exp_connect_flags & OBD_CONNECT_FULL20) && + if ((data->ocd_connect_flags & OBD_CONNECT_FULL20) && (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1)) revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18; 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); + spin_lock(&export->exp_lock); + export->exp_imp_reverse = NULL; + spin_unlock(&export->exp_lock); + class_destroy_import(revimp); + } - class_import_put(revimp); -out: - if (export) { - cfs_spin_lock(&export->exp_lock); - export->exp_connecting = 0; - cfs_spin_unlock(&export->exp_lock); + class_import_put(revimp); - class_export_put(export); - } - if (targref) - class_decref(targref, __FUNCTION__, cfs_current()); - if (rc) - req->rq_status = rc; - RETURN(rc); +out: + if (tmp_imp != NULL) + client_destroy_import(tmp_imp); + if (export) { + spin_lock(&export->exp_lock); + export->exp_connecting = 0; + spin_unlock(&export->exp_lock); + + class_export_put(export); + } + if (targref) { + spin_lock(&target->obd_dev_lock); + target->obd_conn_inprogress--; + spin_unlock(&target->obd_dev_lock); + + class_decref(targref, __func__, current); + } + if (rc) + req->rq_status = rc; + RETURN(rc); } +EXPORT_SYMBOL(target_handle_connect); int target_handle_disconnect(struct ptlrpc_request *req) { @@ -1190,37 +1285,46 @@ int target_handle_disconnect(struct ptlrpc_request *req) if (rc) RETURN(rc); - /* keep the rq_export around so we can send the reply */ + /* Keep the rq_export around so we can send the reply. */ req->rq_status = obd_disconnect(class_export_get(req->rq_export)); 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_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); + struct obd_import *imp = NULL; + /* exports created from last_rcvd data, and "fake" + exports created by lctl don't have an import */ + spin_lock(&exp->exp_lock); + if (exp->exp_imp_reverse != NULL) { + imp = exp->exp_imp_reverse; + exp->exp_imp_reverse = NULL; + } + 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 */ static void target_request_copy_get(struct ptlrpc_request *req) { - class_export_rpc_get(req->rq_export); + class_export_rpc_inc(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 */ - cfs_atomic_inc(&req->rq_refcount); - /** let export know it has replays to be handled */ + /* Increase refcount to keep request in queue. */ + cfs_atomic_inc(&req->rq_refcount); + /* Let export know it has replays to be handled. */ cfs_atomic_inc(&req->rq_export->exp_replay_count); } @@ -1230,7 +1334,7 @@ static void target_request_copy_put(struct ptlrpc_request *req) 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); + class_export_rpc_dec(req->rq_export); ptlrpc_server_drop_request(req); } @@ -1243,7 +1347,7 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) LASSERT(exp); - cfs_spin_lock(&exp->exp_lock); + spin_lock(&exp->exp_lock); cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue, rq_replay_list) { if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) { @@ -1253,7 +1357,7 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) } if (dup) { - /* we expect it with RESENT and REPLAY flags */ + /* We expect it with RESENT and REPLAY flags. */ if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY)) CERROR("invalid flags %x of resent replay\n", @@ -1263,37 +1367,40 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) &exp->exp_req_replay_queue); } - cfs_spin_unlock(&exp->exp_lock); + spin_unlock(&exp->exp_lock); return dup; } static void target_exp_dequeue_req_replay(struct ptlrpc_request *req) { - LASSERT(!cfs_list_empty(&req->rq_replay_list)); - LASSERT(req->rq_export); + LASSERT(!cfs_list_empty(&req->rq_replay_list)); + LASSERT(req->rq_export); - cfs_spin_lock(&req->rq_export->exp_lock); - cfs_list_del_init(&req->rq_replay_list); - cfs_spin_unlock(&req->rq_export->exp_lock); + spin_lock(&req->rq_export->exp_lock); + cfs_list_del_init(&req->rq_replay_list); + spin_unlock(&req->rq_export->exp_lock); } #ifdef __KERNEL__ static void target_finish_recovery(struct obd_device *obd) { - time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() - - obd->obd_recovery_start); ENTRY; - 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"); + /* 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); + spin_lock(&obd->obd_recovery_task_lock); if (!cfs_list_empty(&obd->obd_req_replay_queue) || !cfs_list_empty(&obd->obd_lock_replay_queue) || !cfs_list_empty(&obd->obd_final_req_queue)) { @@ -1304,14 +1411,14 @@ static void target_finish_recovery(struct obd_device *obd) "" : "lock ", cfs_list_empty(&obd->obd_final_req_queue) ? \ "" : "final "); - cfs_spin_unlock(&obd->obd_recovery_task_lock); - LBUG(); - } - cfs_spin_unlock(&obd->obd_recovery_task_lock); + spin_unlock(&obd->obd_recovery_task_lock); + LBUG(); + } + spin_unlock(&obd->obd_recovery_task_lock); obd->obd_recovery_end = cfs_time_current_sec(); - /* when recovery finished, cleanup orphans on mds and ost */ + /* When recovery finished, cleanup orphans on MDS and OST. */ if (OBT(obd) && OBP(obd, postrecov)) { int rc = OBP(obd, postrecov)(obd); if (rc < 0) @@ -1323,13 +1430,13 @@ static void target_finish_recovery(struct obd_device *obd) static void abort_req_replay_queue(struct obd_device *obd) { - struct ptlrpc_request *req, *n; - cfs_list_t abort_list; + struct ptlrpc_request *req, *n; + cfs_list_t abort_list; - CFS_INIT_LIST_HEAD(&abort_list); - cfs_spin_lock(&obd->obd_recovery_task_lock); - cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list); - cfs_spin_unlock(&obd->obd_recovery_task_lock); + CFS_INIT_LIST_HEAD(&abort_list); + spin_lock(&obd->obd_recovery_task_lock); + cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list); + spin_unlock(&obd->obd_recovery_task_lock); cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) { DEBUG_REQ(D_WARNING, req, "aborted:"); req->rq_status = -ENOTCONN; @@ -1344,13 +1451,13 @@ static void abort_req_replay_queue(struct obd_device *obd) static void abort_lock_replay_queue(struct obd_device *obd) { - struct ptlrpc_request *req, *n; - cfs_list_t abort_list; + struct ptlrpc_request *req, *n; + cfs_list_t abort_list; - CFS_INIT_LIST_HEAD(&abort_list); - cfs_spin_lock(&obd->obd_recovery_task_lock); - cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list); - cfs_spin_unlock(&obd->obd_recovery_task_lock); + CFS_INIT_LIST_HEAD(&abort_list); + spin_lock(&obd->obd_recovery_task_lock); + cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list); + spin_unlock(&obd->obd_recovery_task_lock); cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){ DEBUG_REQ(D_ERROR, req, "aborted:"); req->rq_status = -ENOTCONN; @@ -1378,30 +1485,30 @@ void target_cleanup_recovery(struct obd_device *obd) ENTRY; CFS_INIT_LIST_HEAD(&clean_list); - cfs_spin_lock(&obd->obd_dev_lock); - if (!obd->obd_recovering) { - cfs_spin_unlock(&obd->obd_dev_lock); - EXIT; - return; - } - obd->obd_recovering = obd->obd_abort_recovery = 0; - cfs_spin_unlock(&obd->obd_dev_lock); - - cfs_spin_lock(&obd->obd_recovery_task_lock); - target_cancel_recovery_timer(obd); - cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list); - cfs_spin_unlock(&obd->obd_recovery_task_lock); - - cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) { - LASSERT(req->rq_reply_state == 0); - target_exp_dequeue_req_replay(req); - target_request_copy_put(req); - } - - cfs_spin_lock(&obd->obd_recovery_task_lock); - cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list); - cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list); - cfs_spin_unlock(&obd->obd_recovery_task_lock); + spin_lock(&obd->obd_dev_lock); + if (!obd->obd_recovering) { + spin_unlock(&obd->obd_dev_lock); + EXIT; + return; + } + obd->obd_recovering = obd->obd_abort_recovery = 0; + spin_unlock(&obd->obd_dev_lock); + + spin_lock(&obd->obd_recovery_task_lock); + target_cancel_recovery_timer(obd); + cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list); + spin_unlock(&obd->obd_recovery_task_lock); + + cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) { + LASSERT(req->rq_reply_state == 0); + target_exp_dequeue_req_replay(req); + target_request_copy_put(req); + } + + spin_lock(&obd->obd_recovery_task_lock); + cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list); + cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list); + spin_unlock(&obd->obd_recovery_task_lock); cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){ LASSERT(req->rq_reply_state == 0); @@ -1410,6 +1517,7 @@ void target_cleanup_recovery(struct obd_device *obd) EXIT; } +EXPORT_SYMBOL(target_cleanup_recovery); /* obd_recovery_task_lock should be held */ void target_cancel_recovery_timer(struct obd_device *obd) @@ -1417,29 +1525,30 @@ 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); static void target_start_recovery_timer(struct obd_device *obd) { - if (obd->obd_recovery_start != 0) - return; + if (obd->obd_recovery_start != 0) + return; - cfs_spin_lock(&obd->obd_dev_lock); - if (!obd->obd_recovering || obd->obd_abort_recovery) { - cfs_spin_unlock(&obd->obd_dev_lock); - return; - } + spin_lock(&obd->obd_dev_lock); + if (!obd->obd_recovering || obd->obd_abort_recovery) { + spin_unlock(&obd->obd_dev_lock); + return; + } - LASSERT(obd->obd_recovery_timeout != 0); + LASSERT(obd->obd_recovery_timeout != 0); - if (obd->obd_recovery_start != 0) { - cfs_spin_unlock(&obd->obd_dev_lock); - return; - } + if (obd->obd_recovery_start != 0) { + spin_unlock(&obd->obd_dev_lock); + return; + } - 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); + cfs_timer_arm(&obd->obd_recovery_timer, + cfs_time_shift(obd->obd_recovery_timeout)); + obd->obd_recovery_start = cfs_time_current_sec(); + spin_unlock(&obd->obd_dev_lock); LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, " "or until %d client%s reconnect%s\n", @@ -1459,14 +1568,14 @@ static void target_start_recovery_timer(struct obd_device *obd) */ static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend) { - 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); + cfs_time_t now; + cfs_time_t end; + cfs_duration_t left; + int to; + + spin_lock(&obd->obd_dev_lock); + if (!obd->obd_recovering || obd->obd_abort_recovery) { + spin_unlock(&obd->obd_dev_lock); return; } LASSERT(obd->obd_recovery_start != 0); @@ -1486,15 +1595,16 @@ static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend) if (to > obd->obd_recovery_time_hard) to = obd->obd_recovery_time_hard; - if (obd->obd_recovery_timeout < to) { + if (obd->obd_recovery_timeout < to || + obd->obd_recovery_timeout == obd->obd_recovery_time_hard) { obd->obd_recovery_timeout = to; cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(drt)); } - cfs_spin_unlock(&obd->obd_dev_lock); + spin_unlock(&obd->obd_dev_lock); - CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n", - obd->obd_name, (unsigned)drt); + 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 */ @@ -1515,32 +1625,28 @@ check_and_start_recovery_timer(struct obd_device *obd, { 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); target_start_recovery_timer(obd); - /* convert the service time to rpc timeout, - * reuse service_time to limit stack usage */ - service_time = at_est2timeout(service_time); + /* Convert the service time to RPC timeout, + * and reuse service_time to limit stack usage. */ + service_time = at_est2timeout(service_time); - /* 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 */ + /* 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 * INITIAL_CONNECT_TIMEOUT; LASSERT(obt->obt_magic == OBT_MAGIC); - lsi = s2lsi(obt->obt_sb); - if (!(lsi->lsi_flags | LSI_IR_CAPABLE)) - service_time += 2 * (CONNECTION_SWITCH_MAX + - CONNECTION_SWITCH_INC); - if (service_time > obd->obd_recovery_timeout && !new_client) - extend_recovery_timer(obd, service_time, false); + service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC); + if (service_time > obd->obd_recovery_timeout && !new_client) + extend_recovery_timer(obd, service_time, false); } /** Health checking routines */ @@ -1580,20 +1686,18 @@ static int check_for_clients(struct obd_device *obd) if (obd->obd_abort_recovery || obd->obd_recovery_expired) return 1; LASSERT(clnts <= obd->obd_max_recoverable_clients); - if (obd->obd_no_conn == 0 && - clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients) - return 1; - return 0; + return (clnts + obd->obd_stale_clients == + obd->obd_max_recoverable_clients); } static int check_for_next_transno(struct obd_device *obd) { - struct ptlrpc_request *req = NULL; - int wake_up = 0, connected, completed, queue_len; - __u64 next_transno, req_transno; - ENTRY; + struct ptlrpc_request *req = NULL; + int wake_up = 0, connected, completed, queue_len; + __u64 next_transno, req_transno; + ENTRY; - cfs_spin_lock(&obd->obd_recovery_task_lock); + spin_lock(&obd->obd_recovery_task_lock); if (!cfs_list_empty(&obd->obd_req_replay_queue)) { req = cfs_list_entry(obd->obd_req_replay_queue.next, struct ptlrpc_request, rq_list); @@ -1618,13 +1722,11 @@ static int check_for_next_transno(struct obd_device *obd) } else if (obd->obd_recovery_expired) { CDEBUG(D_HA, "waking for expired recovery\n"); wake_up = 1; - } else if (cfs_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 == cfs_atomic_read(&obd->obd_req_replay_clients)) { + } else if (queue_len > 0 && + queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) { int d_lvl = D_HA; /** handle gaps occured due to lost reply or VBR */ LASSERTF(req_transno >= next_transno, @@ -1642,21 +1744,24 @@ static int check_for_next_transno(struct obd_device *obd) req_transno, obd->obd_last_committed); obd->obd_next_recovery_transno = req_transno; wake_up = 1; + } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) { + CDEBUG(D_HA, "waking for completed recovery\n"); + wake_up = 1; } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) { CDEBUG(D_HA, "accepting transno gaps is explicitly allowed" " by fail_lock, waking up ("LPD64")\n", next_transno); obd->obd_next_recovery_transno = req_transno; wake_up = 1; } - cfs_spin_unlock(&obd->obd_recovery_task_lock); - return wake_up; + spin_unlock(&obd->obd_recovery_task_lock); + return wake_up; } static int check_for_next_lock(struct obd_device *obd) { - int wake_up = 0; + int wake_up = 0; - cfs_spin_lock(&obd->obd_recovery_task_lock); + spin_lock(&obd->obd_recovery_task_lock); if (!cfs_list_empty(&obd->obd_lock_replay_queue)) { CDEBUG(D_HA, "waking for next lock\n"); wake_up = 1; @@ -1670,9 +1775,9 @@ static int check_for_next_lock(struct obd_device *obd) CDEBUG(D_HA, "waking for expired recovery\n"); wake_up = 1; } - cfs_spin_unlock(&obd->obd_recovery_task_lock); + spin_unlock(&obd->obd_recovery_task_lock); - return wake_up; + return wake_up; } /** @@ -1681,36 +1786,36 @@ static int check_for_next_lock(struct obd_device *obd) * evict dead clients via health_check */ static int target_recovery_overseer(struct obd_device *obd, - int (*check_routine)(struct obd_device *), - int (*health_check)(struct obd_export *)) + int (*check_routine)(struct obd_device *), + int (*health_check)(struct obd_export *)) { repeat: - cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd)); - if (obd->obd_abort_recovery) { - CWARN("recovery is aborted, evict exports in recovery\n"); - /** evict exports which didn't finish recovery yet */ - class_disconnect_stale_exports(obd, exp_finished); - return 1; - } 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"); - /** evict cexports with no replay in queue, they are stalled */ - class_disconnect_stale_exports(obd, health_check); - /** continue with VBR */ - cfs_spin_lock(&obd->obd_dev_lock); - obd->obd_version_recov = 1; - cfs_spin_unlock(&obd->obd_dev_lock); - /** - * reset timer, recovery will proceed with versions now, - * timeout is set just to handle reconnection delays - */ - extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true); - /** Wait for recovery events again, after evicting bad clients */ - goto repeat; - } - return 0; + wait_event(obd->obd_next_transno_waitq, check_routine(obd)); + if (obd->obd_abort_recovery) { + CWARN("recovery is aborted, evict exports in recovery\n"); + /** evict exports which didn't finish recovery yet */ + class_disconnect_stale_exports(obd, exp_finished); + return 1; + } else if (obd->obd_recovery_expired) { + obd->obd_recovery_expired = 0; + /** If some clients died being recovered, evict them */ + 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 */ + spin_lock(&obd->obd_dev_lock); + obd->obd_version_recov = 1; + spin_unlock(&obd->obd_dev_lock); + /** + * reset timer, recovery will proceed with versions now, + * timeout is set just to handle reconnection delays + */ + extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true); + /** Wait for recovery events again, after evicting bad clients */ + goto repeat; + } + return 0; } static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd) @@ -1727,21 +1832,21 @@ static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd) abort_lock_replay_queue(obd); } - cfs_spin_lock(&obd->obd_recovery_task_lock); - if (!cfs_list_empty(&obd->obd_req_replay_queue)) { - req = cfs_list_entry(obd->obd_req_replay_queue.next, - struct ptlrpc_request, rq_list); - cfs_list_del_init(&req->rq_list); - obd->obd_requests_queued_for_recovery--; - cfs_spin_unlock(&obd->obd_recovery_task_lock); - } else { - cfs_spin_unlock(&obd->obd_recovery_task_lock); - LASSERT(cfs_list_empty(&obd->obd_req_replay_queue)); - LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0); - /** evict exports failed VBR */ - class_disconnect_stale_exports(obd, exp_vbr_healthy); - } - RETURN(req); + spin_lock(&obd->obd_recovery_task_lock); + if (!cfs_list_empty(&obd->obd_req_replay_queue)) { + req = cfs_list_entry(obd->obd_req_replay_queue.next, + struct ptlrpc_request, rq_list); + cfs_list_del_init(&req->rq_list); + obd->obd_requests_queued_for_recovery--; + spin_unlock(&obd->obd_recovery_task_lock); + } else { + spin_unlock(&obd->obd_recovery_task_lock); + LASSERT(cfs_list_empty(&obd->obd_req_replay_queue)); + LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0); + /** evict exports failed VBR */ + class_disconnect_stale_exports(obd, exp_vbr_healthy); + } + RETURN(req); } static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd) @@ -1753,14 +1858,14 @@ static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd) exp_lock_replay_healthy)) abort_lock_replay_queue(obd); - cfs_spin_lock(&obd->obd_recovery_task_lock); - if (!cfs_list_empty(&obd->obd_lock_replay_queue)) { - req = cfs_list_entry(obd->obd_lock_replay_queue.next, - struct ptlrpc_request, rq_list); - cfs_list_del_init(&req->rq_list); - cfs_spin_unlock(&obd->obd_recovery_task_lock); - } else { - cfs_spin_unlock(&obd->obd_recovery_task_lock); + spin_lock(&obd->obd_recovery_task_lock); + if (!cfs_list_empty(&obd->obd_lock_replay_queue)) { + req = cfs_list_entry(obd->obd_lock_replay_queue.next, + struct ptlrpc_request, rq_list); + cfs_list_del_init(&req->rq_list); + spin_unlock(&obd->obd_recovery_task_lock); + } else { + spin_unlock(&obd->obd_recovery_task_lock); LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue)); LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0); /** evict exports failed VBR */ @@ -1771,23 +1876,23 @@ static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd) static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd) { - struct ptlrpc_request *req = NULL; - - cfs_spin_lock(&obd->obd_recovery_task_lock); - if (!cfs_list_empty(&obd->obd_final_req_queue)) { - req = cfs_list_entry(obd->obd_final_req_queue.next, - struct ptlrpc_request, rq_list); - cfs_list_del_init(&req->rq_list); - cfs_spin_unlock(&obd->obd_recovery_task_lock); - if (req->rq_export->exp_in_recovery) { - cfs_spin_lock(&req->rq_export->exp_lock); - req->rq_export->exp_in_recovery = 0; - cfs_spin_unlock(&req->rq_export->exp_lock); - } - } else { - cfs_spin_unlock(&obd->obd_recovery_task_lock); - } - return req; + struct ptlrpc_request *req = NULL; + + spin_lock(&obd->obd_recovery_task_lock); + if (!cfs_list_empty(&obd->obd_final_req_queue)) { + req = cfs_list_entry(obd->obd_final_req_queue.next, + struct ptlrpc_request, rq_list); + cfs_list_del_init(&req->rq_list); + spin_unlock(&obd->obd_recovery_task_lock); + if (req->rq_export->exp_in_recovery) { + spin_lock(&req->rq_export->exp_lock); + req->rq_export->exp_in_recovery = 0; + spin_unlock(&req->rq_export->exp_lock); + } + } else { + spin_unlock(&obd->obd_recovery_task_lock); + } + return req; } static int handle_recovery_req(struct ptlrpc_thread *thread, @@ -1804,7 +1909,7 @@ static int handle_recovery_req(struct ptlrpc_thread *thread, if (req->rq_export->exp_disconnected) GOTO(reqcopy_put, rc = 0); - rc = lu_context_init(&req->rq_recov_session, LCT_SESSION); + rc = lu_context_init(&req->rq_recov_session, LCT_SERVER_SESSION); if (rc) { CERROR("Failure to initialize session: %d\n", rc); GOTO(reqcopy_put, rc); @@ -1831,17 +1936,22 @@ static int handle_recovery_req(struct ptlrpc_thread *thread, * this client may come in recovery time */ if (!AT_OFF) { - struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service; - /* 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(&svc->srv_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); + 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 2 net_latency, one for balance rq_deadline + * (see ptl_send_rpc), one for resend the req to server, + * Note: client will pack net_latency in replay req + * (see ptlrpc_replay_req) */ + to += 2 * lustre_msg_get_service_time(req->rq_reqmsg); } extend_recovery_timer(class_exp2obd(req->rq_export), to, true); } @@ -1856,19 +1966,12 @@ static int target_recovery_thread(void *arg) struct ptlrpc_request *req; struct target_recovery_data *trd = &obd->obd_recovery_data; unsigned long delta; - unsigned long flags; struct lu_env *env; struct ptlrpc_thread *thread = NULL; int rc = 0; ENTRY; - cfs_daemonize_ctxt("tgt_recov"); - - SIGNAL_MASK_LOCK(current, flags); - sigfillset(¤t->blocked); - RECALC_SIGPENDING; - SIGNAL_MASK_UNLOCK(current, flags); - + unshare_fs_struct(); OBD_ALLOC_PTR(thread); if (thread == NULL) RETURN(-ENOMEM); @@ -1879,7 +1982,7 @@ static int target_recovery_thread(void *arg) RETURN(-ENOMEM); } - rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD); + rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD); if (rc) { OBD_FREE_PTR(thread); OBD_FREE_PTR(env); @@ -1889,17 +1992,17 @@ static int target_recovery_thread(void *arg) thread->t_env = env; thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */ env->le_ctx.lc_thread = thread; - thread->t_data = NULL; - thread->t_watchdog = NULL; + tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */ + thread->t_watchdog = NULL; - CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name, - cfs_curproc_pid()); - trd->trd_processing_task = cfs_curproc_pid(); + CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name, + current_pid()); + trd->trd_processing_task = current_pid(); - cfs_spin_lock(&obd->obd_dev_lock); - obd->obd_recovering = 1; - cfs_spin_unlock(&obd->obd_dev_lock); - cfs_complete(&trd->trd_starting); + spin_lock(&obd->obd_dev_lock); + obd->obd_recovering = 1; + spin_unlock(&obd->obd_dev_lock); + complete(&trd->trd_starting); /* first of all, we have to know the first transno to replay */ if (target_recovery_overseer(obd, check_for_clients, @@ -1913,20 +2016,20 @@ static int target_recovery_thread(void *arg) CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n", cfs_atomic_read(&obd->obd_req_replay_clients), obd->obd_next_recovery_transno); - while ((req = target_next_replay_req(obd))) { - LASSERT(trd->trd_processing_task == cfs_curproc_pid()); - DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s", - lustre_msg_get_transno(req->rq_reqmsg), - libcfs_nid2str(req->rq_peer.nid)); + 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); /** * bz18031: increase next_recovery_transno before * target_request_copy_put() will drop exp_rpc reference */ - cfs_spin_lock(&obd->obd_recovery_task_lock); - obd->obd_next_recovery_transno++; - cfs_spin_unlock(&obd->obd_recovery_task_lock); + spin_lock(&obd->obd_recovery_task_lock); + obd->obd_next_recovery_transno++; + spin_unlock(&obd->obd_recovery_task_lock); target_exp_dequeue_req_replay(req); target_request_copy_put(req); obd->obd_replayed_requests++; @@ -1937,10 +2040,10 @@ static int target_recovery_thread(void *arg) */ CDEBUG(D_INFO, "2: lock replay stage - %d clients\n", cfs_atomic_read(&obd->obd_lock_replay_clients)); - while ((req = target_next_replay_lock(obd))) { - LASSERT(trd->trd_processing_task == cfs_curproc_pid()); - DEBUG_REQ(D_HA, req, "processing lock from %s: ", - libcfs_nid2str(req->rq_peer.nid)); + while ((req = target_next_replay_lock(obd))) { + LASSERT(trd->trd_processing_task == current_pid()); + DEBUG_REQ(D_HA, req, "processing lock from %s: ", + libcfs_nid2str(req->rq_peer.nid)); handle_recovery_req(thread, req, trd->trd_recovery_handler); target_request_copy_put(req); @@ -1953,79 +2056,83 @@ static int target_recovery_thread(void *arg) */ CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n"); /** Update server last boot epoch */ - lut_boot_epoch_update(lut); + tgt_boot_epoch_update(lut); /* We drop recoverying flag to forward all new requests * to regular mds_handle() since now */ - cfs_spin_lock(&obd->obd_dev_lock); - obd->obd_recovering = obd->obd_abort_recovery = 0; - cfs_spin_unlock(&obd->obd_dev_lock); - cfs_spin_lock(&obd->obd_recovery_task_lock); - target_cancel_recovery_timer(obd); - cfs_spin_unlock(&obd->obd_recovery_task_lock); - while ((req = target_next_final_ping(obd))) { - LASSERT(trd->trd_processing_task == cfs_curproc_pid()); - DEBUG_REQ(D_HA, req, "processing final ping from %s: ", - libcfs_nid2str(req->rq_peer.nid)); + spin_lock(&obd->obd_dev_lock); + obd->obd_recovering = obd->obd_abort_recovery = 0; + spin_unlock(&obd->obd_dev_lock); + spin_lock(&obd->obd_recovery_task_lock); + target_cancel_recovery_timer(obd); + spin_unlock(&obd->obd_recovery_task_lock); + 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); target_request_copy_put(req); } - delta = (jiffies - delta) / CFS_HZ; - CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n", - delta, obd->obd_replayed_requests, obd->obd_replayed_locks); - if (delta > OBD_RECOVERY_TIME_SOFT) { - CWARN("too long recovery - read logs\n"); - libcfs_debug_dumplog(); - } + 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); + if (delta > OBD_RECOVERY_TIME_SOFT) { + CWARN("too long recovery - read logs\n"); + libcfs_debug_dumplog(); + } target_finish_recovery(obd); lu_context_fini(&env->le_ctx); trd->trd_processing_task = 0; - cfs_complete(&trd->trd_finishing); + complete(&trd->trd_finishing); - OBD_FREE_PTR(thread); - OBD_FREE_PTR(env); - RETURN(rc); + tgt_io_thread_done(thread); + OBD_FREE_PTR(thread); + OBD_FREE_PTR(env); + RETURN(rc); } static int target_start_recovery_thread(struct lu_target *lut, svc_handler_t handler) { - struct obd_device *obd = lut->lut_obd; - int rc = 0; - struct target_recovery_data *trd = &obd->obd_recovery_data; - - memset(trd, 0, sizeof(*trd)); - cfs_init_completion(&trd->trd_starting); - cfs_init_completion(&trd->trd_finishing); - trd->trd_recovery_handler = handler; - - if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) { - cfs_wait_for_completion(&trd->trd_starting); - LASSERT(obd->obd_recovering != 0); - } else - rc = -ECHILD; - - return rc; + struct obd_device *obd = lut->lut_obd; + 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 (!IS_ERR(kthread_run(target_recovery_thread, + lut, "tgt_recov"))) { + 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) { - if (obd->obd_recovery_data.trd_processing_task > 0) { - struct target_recovery_data *trd = &obd->obd_recovery_data; - /** recovery can be done but postrecovery is not yet */ - cfs_spin_lock(&obd->obd_dev_lock); - if (obd->obd_recovering) { - CERROR("%s: Aborting recovery\n", obd->obd_name); - obd->obd_abort_recovery = 1; - cfs_waitq_signal(&obd->obd_next_transno_waitq); - } - cfs_spin_unlock(&obd->obd_dev_lock); - cfs_wait_for_completion(&trd->trd_finishing); - } + if (obd->obd_recovery_data.trd_processing_task > 0) { + struct target_recovery_data *trd = &obd->obd_recovery_data; + /** recovery can be done but postrecovery is not yet */ + spin_lock(&obd->obd_dev_lock); + if (obd->obd_recovering) { + CERROR("%s: Aborting recovery\n", obd->obd_name); + obd->obd_abort_recovery = 1; + wake_up(&obd->obd_next_transno_waitq); + } + spin_unlock(&obd->obd_dev_lock); + wait_for_completion(&trd->trd_finishing); + } } +EXPORT_SYMBOL(target_stop_recovery_thread); void target_recovery_fini(struct obd_device *obd) { @@ -2037,15 +2144,15 @@ EXPORT_SYMBOL(target_recovery_fini); static void target_recovery_expired(unsigned long castmeharder) { - struct obd_device *obd = (struct obd_device *)castmeharder; - CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery" - " 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, - cfs_atomic_read(&obd->obd_connected_clients)); - - obd->obd_recovery_expired = 1; - cfs_waitq_signal(&obd->obd_next_transno_waitq); + struct obd_device *obd = (struct obd_device *)castmeharder; + CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery" + " 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, + cfs_atomic_read(&obd->obd_connected_clients)); + + obd->obd_recovery_expired = 1; + wake_up(&obd->obd_next_transno_waitq); } void target_recovery_init(struct lu_target *lut, svc_handler_t handler) @@ -2053,13 +2160,13 @@ void target_recovery_init(struct lu_target *lut, svc_handler_t handler) struct obd_device *obd = lut->lut_obd; if (obd->obd_max_recoverable_clients == 0) { /** Update server last boot epoch */ - lut_boot_epoch_update(lut); + tgt_boot_epoch_update(lut); return; } - CWARN("RECOVERY: service %s, %d recoverable clients, " - "last_transno "LPU64"\n", obd->obd_name, - obd->obd_max_recoverable_clients, obd->obd_last_committed); + CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, " + "last_transno "LPU64"\n", obd->obd_name, + obd->obd_max_recoverable_clients, obd->obd_last_committed); LASSERT(obd->obd_stopping == 0); obd->obd_next_recovery_transno = obd->obd_last_committed + 1; obd->obd_recovery_start = 0; @@ -2075,36 +2182,36 @@ EXPORT_SYMBOL(target_recovery_init); static 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 */ - cfs_spin_lock(&exp->exp_lock); - if (exp->exp_req_replay_needed) { - 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) { - 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; + 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(&exp->exp_lock); + if (exp->exp_req_replay_needed) { + exp->exp_req_replay_needed = 0; + spin_unlock(&exp->exp_lock); + + LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients); + cfs_atomic_dec(&obd->obd_req_replay_clients); + } else { + 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 */ + spin_lock(&exp->exp_lock); + if (exp->exp_lock_replay_needed) { + exp->exp_lock_replay_needed = 0; + spin_unlock(&exp->exp_lock); + + LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients); + cfs_atomic_dec(&obd->obd_lock_replay_clients); + } else { + spin_unlock(&exp->exp_lock); + } + } + return 0; } int target_queue_recovery_request(struct ptlrpc_request *req, @@ -2115,49 +2222,49 @@ int target_queue_recovery_request(struct ptlrpc_request *req, __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); - } + if (obd->obd_recovery_data.trd_processing_task == current_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 */ - target_request_copy_get(req); - DEBUG_REQ(D_HA, req, "queue final req"); - cfs_waitq_signal(&obd->obd_next_transno_waitq); - cfs_spin_lock(&obd->obd_recovery_task_lock); - if (obd->obd_recovering) { - cfs_list_add_tail(&req->rq_list, - &obd->obd_final_req_queue); - } else { - cfs_spin_unlock(&obd->obd_recovery_task_lock); - target_request_copy_put(req); - RETURN(obd->obd_stopping ? -ENOTCONN : 1); - } - cfs_spin_unlock(&obd->obd_recovery_task_lock); - RETURN(0); - } - if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) { - /* client declares he's ready to replay locks */ - target_request_copy_get(req); - DEBUG_REQ(D_HA, req, "queue lock replay req"); - cfs_waitq_signal(&obd->obd_next_transno_waitq); - cfs_spin_lock(&obd->obd_recovery_task_lock); - LASSERT(obd->obd_recovering); - /* usually due to recovery abort */ - if (!req->rq_export->exp_in_recovery) { - cfs_spin_unlock(&obd->obd_recovery_task_lock); - target_request_copy_put(req); - RETURN(-ENOTCONN); - } - LASSERT(req->rq_export->exp_lock_replay_needed); - cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue); - cfs_spin_unlock(&obd->obd_recovery_task_lock); - RETURN(0); - } + target_request_copy_get(req); + DEBUG_REQ(D_HA, req, "queue final req"); + wake_up(&obd->obd_next_transno_waitq); + spin_lock(&obd->obd_recovery_task_lock); + if (obd->obd_recovering) { + cfs_list_add_tail(&req->rq_list, + &obd->obd_final_req_queue); + } else { + spin_unlock(&obd->obd_recovery_task_lock); + target_request_copy_put(req); + RETURN(obd->obd_stopping ? -ENOTCONN : 1); + } + spin_unlock(&obd->obd_recovery_task_lock); + RETURN(0); + } + if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) { + /* client declares he's ready to replay locks */ + target_request_copy_get(req); + DEBUG_REQ(D_HA, req, "queue lock replay req"); + wake_up(&obd->obd_next_transno_waitq); + spin_lock(&obd->obd_recovery_task_lock); + LASSERT(obd->obd_recovering); + /* usually due to recovery abort */ + if (!req->rq_export->exp_in_recovery) { + spin_unlock(&obd->obd_recovery_task_lock); + target_request_copy_put(req); + RETURN(-ENOTCONN); + } + LASSERT(req->rq_export->exp_lock_replay_needed); + cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue); + spin_unlock(&obd->obd_recovery_task_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 @@ -2179,16 +2286,17 @@ int target_queue_recovery_request(struct ptlrpc_request *req, * 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); - cfs_spin_lock(&obd->obd_recovery_task_lock); - if (transno < obd->obd_next_recovery_transno) { - /* Processing the queue right now, don't re-add. */ - LASSERT(cfs_list_empty(&req->rq_list)); - cfs_spin_unlock(&obd->obd_recovery_task_lock); - RETURN(1); - } - cfs_spin_unlock(&obd->obd_recovery_task_lock); + CDEBUG(D_HA, "Next recovery transno: "LPU64 + ", current: "LPU64", replaying\n", + obd->obd_next_recovery_transno, transno); + spin_lock(&obd->obd_recovery_task_lock); + if (transno < obd->obd_next_recovery_transno) { + /* Processing the queue right now, don't re-add. */ + LASSERT(cfs_list_empty(&req->rq_list)); + spin_unlock(&obd->obd_recovery_task_lock); + RETURN(1); + } + spin_unlock(&obd->obd_recovery_task_lock); if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP)) RETURN(0); @@ -2207,7 +2315,7 @@ int target_queue_recovery_request(struct ptlrpc_request *req, } /* XXX O(n^2) */ - cfs_spin_lock(&obd->obd_recovery_task_lock); + spin_lock(&obd->obd_recovery_task_lock); LASSERT(obd->obd_recovering); cfs_list_for_each(tmp, &obd->obd_req_replay_queue) { struct ptlrpc_request *reqiter = @@ -2223,7 +2331,7 @@ int target_queue_recovery_request(struct ptlrpc_request *req, transno)) { DEBUG_REQ(D_ERROR, req, "dropping replay: transno " "has been claimed by another client"); - cfs_spin_unlock(&obd->obd_recovery_task_lock); + spin_unlock(&obd->obd_recovery_task_lock); target_exp_dequeue_req_replay(req); target_request_copy_put(req); RETURN(0); @@ -2234,16 +2342,18 @@ int target_queue_recovery_request(struct ptlrpc_request *req, cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue); obd->obd_requests_queued_for_recovery++; - cfs_spin_unlock(&obd->obd_recovery_task_lock); - cfs_waitq_signal(&obd->obd_next_transno_waitq); - RETURN(0); + spin_unlock(&obd->obd_recovery_task_lock); + wake_up(&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_export); + 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) { @@ -2272,10 +2382,8 @@ int target_pack_pool_reply(struct ptlrpc_request *req) struct obd_device *obd; ENTRY; - /* - * Check that we still have all structures alive as this may - * be some late rpc in shutdown time. - */ + /* Check that we still have all structures alive as this may + * be some late RPC at shutdown time. */ if (unlikely(!req->rq_export || !req->rq_export->exp_obd || !exp_connect_lru_resize(req->rq_export))) { lustre_msg_set_slv(req->rq_repmsg, 0); @@ -2283,18 +2391,17 @@ int target_pack_pool_reply(struct ptlrpc_request *req) RETURN(0); } - /* - * OBD is alive here as export is alive, which we checked above. - */ + /* OBD is alive here as export is alive, which we checked above. */ obd = req->rq_export->exp_obd; - cfs_read_lock(&obd->obd_pool_lock); + read_lock(&obd->obd_pool_lock); lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv); lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit); - cfs_read_unlock(&obd->obd_pool_lock); + read_unlock(&obd->obd_pool_lock); RETURN(0); } +EXPORT_SYMBOL(target_pack_pool_reply); int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id) { @@ -2316,10 +2423,10 @@ int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id) 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_export *exp; - struct ptlrpc_service *svc; ENTRY; if (req->rq_no_reply) { @@ -2327,7 +2434,7 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) 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 */ @@ -2339,7 +2446,7 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) /* 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); @@ -2358,27 +2465,27 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) 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", - rs->rs_transno, exp->exp_last_committed); - if (rs->rs_transno > exp->exp_last_committed) { - /* not committed already */ - cfs_list_add_tail(&rs->rs_obd_list, - &exp->exp_uncommitted_replies); - } - cfs_spin_unlock (&exp->exp_uncommitted_replies_lock); + spin_lock(&exp->exp_uncommitted_replies_lock); + CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n", + rs->rs_transno, exp->exp_last_committed); + if (rs->rs_transno > exp->exp_last_committed) { + /* not committed already */ + cfs_list_add_tail(&rs->rs_obd_list, + &exp->exp_uncommitted_replies); + } + spin_unlock(&exp->exp_uncommitted_replies_lock); - cfs_spin_lock(&exp->exp_lock); - cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies); - cfs_spin_unlock(&exp->exp_lock); + spin_lock(&exp->exp_lock); + cfs_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); + netrc = target_send_reply_msg(req, rc, fail_id); - cfs_spin_lock(&svc->srv_rs_lock); + 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 @@ -2390,134 +2497,22 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) ptlrpc_rs_addref(rs); } - cfs_spin_lock(&rs->rs_lock); - if (rs->rs_transno <= exp->exp_last_committed || - (!rs->rs_on_net && !rs->rs_no_ack) || - cfs_list_empty(&rs->rs_exp_list) || /* completed already */ - cfs_list_empty(&rs->rs_obd_list)) { - 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_rs_lock); - EXIT; + spin_lock(&rs->rs_lock); + if (rs->rs_transno <= exp->exp_last_committed || + (!rs->rs_on_net && !rs->rs_no_ack) || + cfs_list_empty(&rs->rs_exp_list) || /* completed already */ + cfs_list_empty(&rs->rs_obd_list)) { + CDEBUG(D_HA, "Schedule reply immediately\n"); + ptlrpc_dispatch_difficult_reply(rs); + } else { + cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active); + rs->rs_scheduled = 0; /* allow notifier to schedule */ + } + spin_unlock(&rs->rs_lock); + spin_unlock(&svcpt->scp_rep_lock); + EXIT; } - -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; -} - -#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, @@ -2611,16 +2606,17 @@ EXPORT_SYMBOL(ldlm_errno2error); #if LUSTRE_TRACKS_LOCK_EXP_REFS void ldlm_dump_export_locks(struct obd_export *exp) { - cfs_spin_lock(&exp->exp_locks_list_guard); - if (!cfs_list_empty(&exp->exp_locks_list)) { - struct ldlm_lock *lock; - - 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); - } - cfs_spin_unlock(&exp->exp_locks_list_guard); + spin_lock(&exp->exp_locks_list_guard); + if (!cfs_list_empty(&exp->exp_locks_list)) { + struct ldlm_lock *lock; + + 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_ERROR(lock, "lock:"); + } + spin_unlock(&exp->exp_locks_list_guard); } #endif @@ -2642,90 +2638,91 @@ static inline char *bulk2type(struct ptlrpc_bulk_desc *desc) 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); + struct ptlrpc_request *req = desc->bd_req; + time_t start = cfs_time_current_sec(); + int rc = 0; + + ENTRY; + + /* If there is eviction in progress, 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 reconnected already. */ + if (exp->exp_failed || + exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) { + 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) { + DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d", + bulk2type(desc), rc); + RETURN(rc); + } + + if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) { + ptlrpc_abort_bulk(desc); + RETURN(0); + } + + 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_conn_cnt > + lustre_msg_get_conn_cnt(req->rq_reqmsg), + 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_conn_cnt > + lustre_msg_get_conn_cnt(req->rq_reqmsg)) { + 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_failure || + desc->bd_nob_transferred != desc->bd_nob) { + DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)", + desc->bd_failure ? "network error on" : "truncated", + 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); + } + + RETURN(rc); } EXPORT_SYMBOL(target_bulk_io);