X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fptlrpc%2Fimport.c;h=ef0df00e8744aaed3d27914a3329d5f3aee44c15;hp=95eb543dc182e71d7d0da1f844c4d31a096e72c7;hb=108339f1543fb006f4ddd16830e7266df0b46723;hpb=4a51b7a884e536da4e6a3386e8d25a2bd2b445a2 diff --git a/lustre/ptlrpc/import.c b/lustre/ptlrpc/import.c index 95eb543..ef0df00 100644 --- a/lustre/ptlrpc/import.c +++ b/lustre/ptlrpc/import.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. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2015, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -39,10 +39,8 @@ */ #define DEBUG_SUBSYSTEM S_RPC -#ifndef __KERNEL__ -# include -#endif +#include #include #include #include @@ -59,9 +57,25 @@ struct ptlrpc_connect_async_args { int pcaa_initial_connect; }; +/** + * Updates import \a imp current state to provided \a state value + * Helper function. Must be called under imp_lock. + */ static void __import_set_state(struct obd_import *imp, enum lustre_imp_state state) { + switch (state) { + case LUSTRE_IMP_CLOSED: + case LUSTRE_IMP_NEW: + case LUSTRE_IMP_DISCON: + case LUSTRE_IMP_CONNECTING: + break; + case LUSTRE_IMP_REPLAY_WAIT: + imp->imp_replay_state = LUSTRE_IMP_REPLAY_LOCKS; + break; + default: + imp->imp_replay_state = LUSTRE_IMP_REPLAY; + } imp->imp_state = state; imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state; imp->imp_state_hist[imp->imp_state_hist_idx].ish_time = @@ -82,13 +96,19 @@ do { \ } \ } while(0) -#define IMPORT_SET_STATE(imp, state) \ -do { \ - cfs_spin_lock(&imp->imp_lock); \ - IMPORT_SET_STATE_NOLOCK(imp, state); \ - cfs_spin_unlock(&imp->imp_lock); \ +#define IMPORT_SET_STATE(imp, state) \ +do { \ + spin_lock(&imp->imp_lock); \ + IMPORT_SET_STATE_NOLOCK(imp, state); \ + spin_unlock(&imp->imp_lock); \ } while(0) +void ptlrpc_import_enter_resend(struct obd_import *imp) +{ + IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER); +} +EXPORT_SYMBOL(ptlrpc_import_enter_resend); + static int ptlrpc_connect_interpret(const struct lu_env *env, struct ptlrpc_request *request, @@ -102,20 +122,19 @@ int ptlrpc_import_recovery_state_machine(struct obd_import *imp); * though. */ int ptlrpc_init_import(struct obd_import *imp) { - cfs_spin_lock(&imp->imp_lock); + spin_lock(&imp->imp_lock); - imp->imp_generation++; - imp->imp_state = LUSTRE_IMP_NEW; + imp->imp_generation++; + imp->imp_state = LUSTRE_IMP_NEW; - cfs_spin_unlock(&imp->imp_lock); + spin_unlock(&imp->imp_lock); - return 0; + return 0; } EXPORT_SYMBOL(ptlrpc_init_import); #define UUID_STR "_UUID" -static void deuuidify(char *uuid, const char *prefix, char **uuid_start, - int *uuid_len) +void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len) { *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix)) ? uuid : uuid + strlen(prefix); @@ -130,7 +149,8 @@ static void deuuidify(char *uuid, const char *prefix, char **uuid_start, *uuid_len -= strlen(UUID_STR); } -/* Returns true if import was FULL, false if import was already not +/** + * Returns true if import was FULL, false if import was already not * connected. * @imp - import to be disconnected * @conn_cnt - connection count (epoch) of the request that timed out @@ -142,9 +162,9 @@ static void deuuidify(char *uuid, const char *prefix, char **uuid_start, */ int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt) { - int rc = 0; + int rc = 0; - cfs_spin_lock(&imp->imp_lock); + spin_lock(&imp->imp_lock); if (imp->imp_state == LUSTRE_IMP_FULL && (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) { @@ -155,30 +175,29 @@ int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt) &target_start, &target_len); if (imp->imp_replayable) { - LCONSOLE_WARN("%s: Connection to service %.*s via nid " - "%s was lost; in progress operations using this " - "service will wait for recovery to complete.\n", + LCONSOLE_WARN("%s: Connection to %.*s (at %s) was " + "lost; in progress operations using this " + "service will wait for recovery to complete\n", imp->imp_obd->obd_name, target_len, target_start, libcfs_nid2str(imp->imp_connection->c_peer.nid)); } else { - LCONSOLE_ERROR_MSG(0x166, "%s: Connection to service " - "%.*s via nid %s was lost; in progress " - "operations using this service will fail.\n", + LCONSOLE_ERROR_MSG(0x166, "%s: Connection to " + "%.*s (at %s) was lost; in progress " + "operations using this service will fail\n", imp->imp_obd->obd_name, target_len, target_start, libcfs_nid2str(imp->imp_connection->c_peer.nid)); } - ptlrpc_deactivate_timeouts(imp); IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON); - cfs_spin_unlock(&imp->imp_lock); + spin_unlock(&imp->imp_lock); - if (obd_dump_on_timeout) - libcfs_debug_dumplog(); + if (obd_dump_on_timeout) + libcfs_debug_dumplog(); - obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON); - rc = 1; - } else { - cfs_spin_unlock(&imp->imp_lock); + obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON); + rc = 1; + } else { + spin_unlock(&imp->imp_lock); CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n", imp->imp_client->cli_name, imp, (imp->imp_state == LUSTRE_IMP_FULL && @@ -193,18 +212,18 @@ int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt) /* Must be called with imp_lock held! */ static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp) { - ENTRY; - LASSERT_SPIN_LOCKED(&imp->imp_lock); + ENTRY; + assert_spin_locked(&imp->imp_lock); - CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd)); - imp->imp_invalid = 1; - imp->imp_generation++; - cfs_spin_unlock(&imp->imp_lock); + CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd)); + imp->imp_invalid = 1; + imp->imp_generation++; + spin_unlock(&imp->imp_lock); - ptlrpc_abort_inflight(imp); - obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE); + ptlrpc_abort_inflight(imp); + obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE); - EXIT; + EXIT; } /* @@ -213,9 +232,10 @@ static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp) */ void ptlrpc_deactivate_import(struct obd_import *imp) { - cfs_spin_lock(&imp->imp_lock); - ptlrpc_deactivate_and_unlock_import(imp); + spin_lock(&imp->imp_lock); + ptlrpc_deactivate_and_unlock_import(imp); } +EXPORT_SYMBOL(ptlrpc_deactivate_import); static unsigned int ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now) @@ -243,21 +263,21 @@ ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now) static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp) { - time_t now = cfs_time_current_sec(); - cfs_list_t *tmp, *n; - struct ptlrpc_request *req; - unsigned int timeout = 0; - - cfs_spin_lock(&imp->imp_lock); - cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) { - req = cfs_list_entry(tmp, struct ptlrpc_request, rq_list); - timeout = max(ptlrpc_inflight_deadline(req, now), timeout); - } - cfs_spin_unlock(&imp->imp_lock); - return timeout; + time_t now = cfs_time_current_sec(); + struct list_head *tmp, *n; + struct ptlrpc_request *req; + unsigned int timeout = 0; + + spin_lock(&imp->imp_lock); + list_for_each_safe(tmp, n, &imp->imp_sending_list) { + req = list_entry(tmp, struct ptlrpc_request, rq_list); + timeout = max(ptlrpc_inflight_deadline(req, now), timeout); + } + spin_unlock(&imp->imp_lock); + return timeout; } -/* +/** * This function will invalidate the import, if necessary, then block * for all the RPC completions, and finally notify the obd to * invalidate its state (ie cancel locks, clear pending requests, @@ -265,25 +285,19 @@ static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp) */ void ptlrpc_invalidate_import(struct obd_import *imp) { - cfs_list_t *tmp, *n; - struct ptlrpc_request *req; - struct l_wait_info lwi; - unsigned int timeout; - int rc; - - cfs_atomic_inc(&imp->imp_inval_count); + struct list_head *tmp, *n; + struct ptlrpc_request *req; + struct l_wait_info lwi; + unsigned int timeout; + int rc; - /* - * If this is an invalid MGC connection, then don't bother - * waiting for imp_inflight to drop to 0. - */ - if (imp->imp_invalid && imp->imp_recon_bk &&!imp->imp_obd->obd_no_recov) - goto out; + atomic_inc(&imp->imp_inval_count); - if (!imp->imp_invalid || imp->imp_obd->obd_no_recov) - ptlrpc_deactivate_import(imp); + if (!imp->imp_invalid || imp->imp_obd->obd_no_recov) + ptlrpc_deactivate_import(imp); - LASSERT(imp->imp_invalid); + CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2); + LASSERT(imp->imp_invalid); /* Wait forever until inflight == 0. We really can't do it another * way because in some cases we need to wait for very long reply @@ -307,93 +321,109 @@ void ptlrpc_invalidate_import(struct obd_import *imp) CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n", timeout); - /* Wait for all requests to error out and call completion - * callbacks. Cap it at obd_timeout -- these should all - * have been locally cancelled by ptlrpc_abort_inflight. */ - lwi = LWI_TIMEOUT_INTERVAL( - cfs_timeout_cap(cfs_time_seconds(timeout)), - (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2, - NULL, NULL); - rc = l_wait_event(imp->imp_recovery_waitq, - (cfs_atomic_read(&imp->imp_inflight) == 0), - &lwi); - if (rc) { - const char *cli_tgt = obd2cli_tgt(imp->imp_obd); - - CERROR("%s: rc = %d waiting for callback (%d != 0)\n", - cli_tgt, rc, - cfs_atomic_read(&imp->imp_inflight)); - - cfs_spin_lock(&imp->imp_lock); - if (cfs_atomic_read(&imp->imp_inflight) == 0) { - int count = cfs_atomic_read(&imp->imp_unregistering); - - /* We know that "unregistering" rpcs only can - * survive in sending or delaying lists (they - * maybe waiting for long reply unlink in - * sluggish nets). Let's check this. If there - * is no inflight and unregistering != 0, this - * is bug. */ - LASSERTF(count == 0, "Some RPCs are still " - "unregistering: %d\n", count); - - /* Let's save one loop as soon as inflight have - * dropped to zero. No new inflights possible at - * this point. */ - rc = 0; - } else { - cfs_list_for_each_safe(tmp, n, - &imp->imp_sending_list) { - req = cfs_list_entry(tmp, - struct ptlrpc_request, - rq_list); - DEBUG_REQ(D_ERROR, req, - "still on sending list"); - } - cfs_list_for_each_safe(tmp, n, - &imp->imp_delayed_list) { - req = cfs_list_entry(tmp, - struct ptlrpc_request, - rq_list); - DEBUG_REQ(D_ERROR, req, - "still on delayed list"); - } - - CERROR("%s: RPCs in \"%s\" phase found (%d). " - "Network is sluggish? Waiting them " - "to error out.\n", cli_tgt, - ptlrpc_phase2str(RQ_PHASE_UNREGISTERING), - cfs_atomic_read(&imp-> - imp_unregistering)); - } - cfs_spin_unlock(&imp->imp_lock); - } - } while (rc != 0); - - /* - * Let's additionally check that no new rpcs added to import in - * "invalidate" state. - */ - LASSERT(cfs_atomic_read(&imp->imp_inflight) == 0); -out: - obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE); - sptlrpc_import_flush_all_ctx(imp); - - cfs_atomic_dec(&imp->imp_inval_count); - cfs_waitq_broadcast(&imp->imp_recovery_waitq); + /* Wait for all requests to error out and call completion + * callbacks. Cap it at obd_timeout -- these should all + * have been locally cancelled by ptlrpc_abort_inflight. */ + lwi = LWI_TIMEOUT_INTERVAL( + cfs_timeout_cap(cfs_time_seconds(timeout)), + (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2, + NULL, NULL); + rc = l_wait_event(imp->imp_recovery_waitq, + (atomic_read(&imp->imp_inflight) == 0), + &lwi); + if (rc) { + const char *cli_tgt = obd2cli_tgt(imp->imp_obd); + + CERROR("%s: rc = %d waiting for callback (%d != 0)\n", + cli_tgt, rc, atomic_read(&imp->imp_inflight)); + + spin_lock(&imp->imp_lock); + if (atomic_read(&imp->imp_inflight) == 0) { + int count = atomic_read(&imp->imp_unregistering); + + /* We know that "unregistering" rpcs only can + * survive in sending or delaying lists (they + * maybe waiting for long reply unlink in + * sluggish nets). Let's check this. If there + * is no inflight and unregistering != 0, this + * is bug. */ + LASSERTF(count == 0, "Some RPCs are still " + "unregistering: %d\n", count); + + /* Let's save one loop as soon as inflight have + * dropped to zero. No new inflights possible at + * this point. */ + rc = 0; + } else { + list_for_each_safe(tmp, n, + &imp->imp_sending_list) { + req = list_entry(tmp, + struct ptlrpc_request, + rq_list); + DEBUG_REQ(D_ERROR, req, + "still on sending list"); + } + list_for_each_safe(tmp, n, + &imp->imp_delayed_list) { + req = list_entry(tmp, + struct ptlrpc_request, + rq_list); + DEBUG_REQ(D_ERROR, req, + "still on delayed list"); + } + + CERROR("%s: Unregistering RPCs found (%d). " + "Network is sluggish? Waiting them " + "to error out.\n", cli_tgt, + atomic_read(&imp->imp_unregistering)); + } + spin_unlock(&imp->imp_lock); + } + } while (rc != 0); + + /* + * Let's additionally check that no new rpcs added to import in + * "invalidate" state. + */ + LASSERT(atomic_read(&imp->imp_inflight) == 0); + obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE); + sptlrpc_import_flush_all_ctx(imp); + + atomic_dec(&imp->imp_inval_count); + wake_up_all(&imp->imp_recovery_waitq); } +EXPORT_SYMBOL(ptlrpc_invalidate_import); /* unset imp_invalid */ void ptlrpc_activate_import(struct obd_import *imp) { - struct obd_device *obd = imp->imp_obd; + struct obd_device *obd = imp->imp_obd; + + spin_lock(&imp->imp_lock); + if (imp->imp_deactive != 0) { + spin_unlock(&imp->imp_lock); + return; + } + + imp->imp_invalid = 0; + spin_unlock(&imp->imp_lock); + obd_import_event(obd, imp, IMP_EVENT_ACTIVE); +} +EXPORT_SYMBOL(ptlrpc_activate_import); + +void ptlrpc_pinger_force(struct obd_import *imp) +{ + CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd), + ptlrpc_import_state_name(imp->imp_state)); + + spin_lock(&imp->imp_lock); + imp->imp_force_verify = 1; + spin_unlock(&imp->imp_lock); - cfs_spin_lock(&imp->imp_lock); - imp->imp_invalid = 0; - ptlrpc_activate_timeouts(imp); - cfs_spin_unlock(&imp->imp_lock); - obd_import_event(obd, imp, IMP_EVENT_ACTIVE); + if (imp->imp_state != LUSTRE_IMP_CONNECTING) + ptlrpc_pinger_wake_up(); } +EXPORT_SYMBOL(ptlrpc_pinger_force); void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt) { @@ -411,75 +441,86 @@ void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt) ptlrpc_deactivate_import(imp); } - CDEBUG(D_HA, "%s: waking up pinger\n", - obd2cli_tgt(imp->imp_obd)); - - cfs_spin_lock(&imp->imp_lock); - imp->imp_force_verify = 1; - cfs_spin_unlock(&imp->imp_lock); - - ptlrpc_pinger_wake_up(); - } - EXIT; + ptlrpc_pinger_force(imp); + } + EXIT; } int ptlrpc_reconnect_import(struct obd_import *imp) { - ptlrpc_set_import_discon(imp, 0); - /* Force a new connect attempt */ - ptlrpc_invalidate_import(imp); - /* Do a fresh connect next time by zeroing the handle */ - ptlrpc_disconnect_import(imp, 1); - /* Wait for all invalidate calls to finish */ - if (cfs_atomic_read(&imp->imp_inval_count) > 0) { - int rc; - struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL); - rc = l_wait_event(imp->imp_recovery_waitq, - (cfs_atomic_read(&imp->imp_inval_count) == 0), - &lwi); - if (rc) - CERROR("Interrupted, inval=%d\n", - cfs_atomic_read(&imp->imp_inval_count)); - } - - /* Allow reconnect attempts */ - imp->imp_obd->obd_no_recov = 0; - /* Remove 'invalid' flag */ - ptlrpc_activate_import(imp); - /* Attempt a new connect */ - ptlrpc_recover_import(imp, NULL); - return 0; +#ifdef ENABLE_PINGER + struct l_wait_info lwi; + int secs = cfs_time_seconds(obd_timeout); + int rc; + + ptlrpc_pinger_force(imp); + + CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n", + obd2cli_tgt(imp->imp_obd), secs); + + lwi = LWI_TIMEOUT(secs, NULL, NULL); + rc = l_wait_event(imp->imp_recovery_waitq, + !ptlrpc_import_in_recovery(imp), &lwi); + CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd), + ptlrpc_import_state_name(imp->imp_state)); + return rc; +#else + ptlrpc_set_import_discon(imp, 0); + /* Force a new connect attempt */ + ptlrpc_invalidate_import(imp); + /* Do a fresh connect next time by zeroing the handle */ + ptlrpc_disconnect_import(imp, 1); + /* Wait for all invalidate calls to finish */ + if (atomic_read(&imp->imp_inval_count) > 0) { + struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL); + int rc; + + rc = l_wait_event(imp->imp_recovery_waitq, + (atomic_read(&imp->imp_inval_count) == 0), + &lwi); + if (rc) + CERROR("Interrupted, inval=%d\n", + atomic_read(&imp->imp_inval_count)); + } + + /* Allow reconnect attempts */ + imp->imp_obd->obd_no_recov = 0; + /* Remove 'invalid' flag */ + ptlrpc_activate_import(imp); + /* Attempt a new connect */ + ptlrpc_recover_import(imp, NULL, 0); + return 0; +#endif } - EXPORT_SYMBOL(ptlrpc_reconnect_import); +/** + * Connection on import \a imp is changed to another one (if more than one is + * present). We typically chose connection that we have not tried to connect to + * the longest + */ static int import_select_connection(struct obd_import *imp) { struct obd_import_conn *imp_conn = NULL, *conn; struct obd_export *dlmexp; - int tried_all = 1; + char *target_start; + int target_len, tried_all = 1; ENTRY; - cfs_spin_lock(&imp->imp_lock); + spin_lock(&imp->imp_lock); - if (cfs_list_empty(&imp->imp_conn_list)) { - CERROR("%s: no connections available\n", - imp->imp_obd->obd_name); - cfs_spin_unlock(&imp->imp_lock); - RETURN(-EINVAL); - } + if (list_empty(&imp->imp_conn_list)) { + CERROR("%s: no connections available\n", + imp->imp_obd->obd_name); + spin_unlock(&imp->imp_lock); + RETURN(-EINVAL); + } - cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) { - CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n", + list_for_each_entry(conn, &imp->imp_conn_list, oic_item) { + CDEBUG(D_HA, "%s: connect to NID %s last attempt %llu\n", imp->imp_obd->obd_name, libcfs_nid2str(conn->oic_conn->c_peer.nid), conn->oic_last_attempt); - /* Don't thrash connections */ - if (cfs_time_before_64(cfs_time_current_64(), - conn->oic_last_attempt + - cfs_time_seconds(CONNECTION_SWITCH_MIN))) { - continue; - } /* If we have not tried this connection since the last successful attempt, go with this one */ @@ -514,19 +555,17 @@ static int import_select_connection(struct obd_import *imp) we do finally connect. (FIXME: really we should wait for all network state associated with the last connection attempt to drain before trying to reconnect on it.) */ - if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item) && - !imp->imp_recon_bk /* not retrying */) { - if (at_get(&imp->imp_at.iat_net_latency) < - CONNECTION_SWITCH_MAX) { - at_measured(&imp->imp_at.iat_net_latency, - at_get(&imp->imp_at.iat_net_latency) + - CONNECTION_SWITCH_INC); - } - LASSERT(imp_conn->oic_last_attempt); - CWARN("%s: tried all connections, increasing latency to %ds\n", - imp->imp_obd->obd_name, - at_get(&imp->imp_at.iat_net_latency)); - } + if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) { + struct adaptive_timeout *at = &imp->imp_at.iat_net_latency; + if (at_get(at) < CONNECTION_SWITCH_MAX) { + at_measured(at, at_get(at) + CONNECTION_SWITCH_INC); + if (at_get(at) > CONNECTION_SWITCH_MAX) + at_reset(at, CONNECTION_SWITCH_MAX); + } + LASSERT(imp_conn->oic_last_attempt); + CDEBUG(D_HA, "%s: tried all connections, increasing latency " + "to %ds\n", imp->imp_obd->obd_name, at_get(at)); + } imp_conn->oic_last_attempt = cfs_time_current_64(); @@ -543,10 +582,17 @@ static int import_select_connection(struct obd_import *imp) class_export_put(dlmexp); if (imp->imp_conn_current != imp_conn) { - if (imp->imp_conn_current) - CDEBUG(D_HA, "Changing connection for %s to %s/%s\n", - imp->imp_obd->obd_name, imp_conn->oic_uuid.uuid, + if (imp->imp_conn_current) { + deuuidify(obd2cli_tgt(imp->imp_obd), NULL, + &target_start, &target_len); + + CDEBUG(D_HA, "%s: Connection changing to" + " %.*s (at %s)\n", + imp->imp_obd->obd_name, + target_len, target_start, libcfs_nid2str(imp_conn->oic_conn->c_peer.nid)); + } + imp->imp_conn_current = imp_conn; } @@ -554,9 +600,9 @@ static int import_select_connection(struct obd_import *imp) imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid, libcfs_nid2str(imp_conn->oic_conn->c_peer.nid)); - cfs_spin_unlock(&imp->imp_lock); + spin_unlock(&imp->imp_lock); - RETURN(0); + RETURN(0); } /* @@ -564,275 +610,528 @@ static int import_select_connection(struct obd_import *imp) */ static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno) { - struct ptlrpc_request *req; - cfs_list_t *tmp; - - if (cfs_list_empty(&imp->imp_replay_list)) - return 0; - tmp = imp->imp_replay_list.next; - req = cfs_list_entry(tmp, struct ptlrpc_request, rq_replay_list); - *transno = req->rq_transno; - if (req->rq_transno == 0) { - DEBUG_REQ(D_ERROR, req, "zero transno in replay"); - LBUG(); - } - - return 1; + struct ptlrpc_request *req; + struct list_head *tmp; + + /* The requests in committed_list always have smaller transnos than + * the requests in replay_list */ + if (!list_empty(&imp->imp_committed_list)) { + tmp = imp->imp_committed_list.next; + req = list_entry(tmp, struct ptlrpc_request, rq_replay_list); + *transno = req->rq_transno; + if (req->rq_transno == 0) { + DEBUG_REQ(D_ERROR, req, "zero transno in committed_list"); + LBUG(); + } + return 1; + } + if (!list_empty(&imp->imp_replay_list)) { + tmp = imp->imp_replay_list.next; + req = list_entry(tmp, struct ptlrpc_request, rq_replay_list); + *transno = req->rq_transno; + if (req->rq_transno == 0) { + DEBUG_REQ(D_ERROR, req, "zero transno in replay_list"); + LBUG(); + } + return 1; + } + return 0; } -int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid) +/** + * Attempt to (re)connect import \a imp. This includes all preparations, + * initializing CONNECT RPC request and passing it to ptlrpcd for + * actual sending. + * Returns 0 on success or error code. + */ +int ptlrpc_connect_import(struct obd_import *imp) { - struct obd_device *obd = imp->imp_obd; - int initial_connect = 0; - int set_transno = 0; - __u64 committed_before_reconnect = 0; - struct ptlrpc_request *request; - char *bufs[] = { NULL, - obd2cli_tgt(imp->imp_obd), - obd->obd_uuid.uuid, - (char *)&imp->imp_dlm_handle, - (char *)&imp->imp_connect_data }; - struct ptlrpc_connect_async_args *aa; - int rc; - ENTRY; - - cfs_spin_lock(&imp->imp_lock); - if (imp->imp_state == LUSTRE_IMP_CLOSED) { - cfs_spin_unlock(&imp->imp_lock); - CERROR("can't connect to a closed import\n"); - RETURN(-EINVAL); - } else if (imp->imp_state == LUSTRE_IMP_FULL) { - cfs_spin_unlock(&imp->imp_lock); - CERROR("already connected\n"); - RETURN(0); - } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) { - cfs_spin_unlock(&imp->imp_lock); - CERROR("already connecting\n"); - RETURN(-EALREADY); - } - - IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING); - - imp->imp_conn_cnt++; - imp->imp_resend_replay = 0; - - if (!lustre_handle_is_used(&imp->imp_remote_handle)) - initial_connect = 1; - else - committed_before_reconnect = imp->imp_peer_committed_transno; + struct obd_device *obd = imp->imp_obd; + int initial_connect = 0; + int set_transno = 0; + __u64 committed_before_reconnect = 0; + struct ptlrpc_request *request; + char *bufs[] = { NULL, + obd2cli_tgt(imp->imp_obd), + obd->obd_uuid.uuid, + (char *)&imp->imp_dlm_handle, + (char *)&imp->imp_connect_data }; + struct ptlrpc_connect_async_args *aa; + int rc; + ENTRY; + + spin_lock(&imp->imp_lock); + if (imp->imp_state == LUSTRE_IMP_CLOSED) { + spin_unlock(&imp->imp_lock); + CERROR("can't connect to a closed import\n"); + RETURN(-EINVAL); + } else if (imp->imp_state == LUSTRE_IMP_FULL) { + spin_unlock(&imp->imp_lock); + CERROR("already connected\n"); + RETURN(0); + } else if (imp->imp_state == LUSTRE_IMP_CONNECTING || + imp->imp_connected) { + spin_unlock(&imp->imp_lock); + CERROR("already connecting\n"); + RETURN(-EALREADY); + } + + IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING); + + imp->imp_conn_cnt++; + imp->imp_resend_replay = 0; + + if (!lustre_handle_is_used(&imp->imp_remote_handle)) + initial_connect = 1; + else + committed_before_reconnect = imp->imp_peer_committed_transno; + + set_transno = ptlrpc_first_transno(imp, + &imp->imp_connect_data.ocd_transno); + spin_unlock(&imp->imp_lock); + + rc = import_select_connection(imp); + if (rc) + GOTO(out, rc); + + rc = sptlrpc_import_sec_adapt(imp, NULL, NULL); + if (rc) + GOTO(out, rc); + + /* Reset connect flags to the originally requested flags, in case + * the server is updated on-the-fly we will get the new features. */ + imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig; + imp->imp_connect_data.ocd_connect_flags2 = imp->imp_connect_flags2_orig; + /* Reset ocd_version each time so the server knows the exact versions */ + imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE; + imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; + imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18; + + rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd, + &obd->obd_uuid, &imp->imp_connect_data, NULL); + if (rc) + GOTO(out, rc); + + request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT); + if (request == NULL) + GOTO(out, rc = -ENOMEM); + + rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION, + imp->imp_connect_op, bufs, NULL); + if (rc) { + ptlrpc_request_free(request); + GOTO(out, rc); + } + + /* Report the rpc service time to the server so that it knows how long + * to wait for clients to join recovery */ + lustre_msg_set_service_time(request->rq_reqmsg, + at_timeout2est(request->rq_timeout)); + + /* The amount of time we give the server to process the connect req. + * import_select_connection will increase the net latency on + * repeated reconnect attempts to cover slow networks. + * We override/ignore the server rpc completion estimate here, + * which may be large if this is a reconnect attempt */ + request->rq_timeout = INITIAL_CONNECT_TIMEOUT; + lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout); + + request->rq_no_resend = request->rq_no_delay = 1; + request->rq_send_state = LUSTRE_IMP_CONNECTING; + /* Allow a slightly larger reply for future growth compatibility */ + req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER, + sizeof(struct obd_connect_data)+16*sizeof(__u64)); + ptlrpc_request_set_replen(request); + request->rq_interpret_reply = ptlrpc_connect_interpret; + + CLASSERT(sizeof(*aa) <= sizeof(request->rq_async_args)); + aa = ptlrpc_req_async_args(request); + memset(aa, 0, sizeof *aa); + + aa->pcaa_peer_committed = committed_before_reconnect; + aa->pcaa_initial_connect = initial_connect; + + if (aa->pcaa_initial_connect) { + spin_lock(&imp->imp_lock); + imp->imp_replayable = 1; + spin_unlock(&imp->imp_lock); + lustre_msg_add_op_flags(request->rq_reqmsg, + MSG_CONNECT_INITIAL); + } + + if (set_transno) + lustre_msg_add_op_flags(request->rq_reqmsg, + MSG_CONNECT_TRANSNO); + + DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)", + request->rq_timeout); + ptlrpcd_add_req(request); + rc = 0; +out: + if (rc != 0) + IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON); - set_transno = ptlrpc_first_transno(imp, - &imp->imp_connect_data.ocd_transno); - cfs_spin_unlock(&imp->imp_lock); + RETURN(rc); +} +EXPORT_SYMBOL(ptlrpc_connect_import); - if (new_uuid) { - struct obd_uuid uuid; +static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp) +{ + int force_verify; - obd_str2uuid(&uuid, new_uuid); - rc = import_set_conn_priority(imp, &uuid); - if (rc) - GOTO(out, rc); - } + spin_lock(&imp->imp_lock); + force_verify = imp->imp_force_verify != 0; + spin_unlock(&imp->imp_lock); - rc = import_select_connection(imp); - if (rc) - GOTO(out, rc); - - /* last in connection list */ - if (imp->imp_conn_current->oic_item.next == &imp->imp_conn_list) { - if (imp->imp_initial_recov_bk && initial_connect) { - CDEBUG(D_HA, "Last connection attempt (%d) for %s\n", - imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd)); - /* Don't retry if connect fails */ - rc = 0; - obd_set_info_async(obd->obd_self_export, - sizeof(KEY_INIT_RECOV), - KEY_INIT_RECOV, - sizeof(rc), &rc, NULL); - } - if (imp->imp_recon_bk) { - CDEBUG(D_HA, "Last reconnection attempt (%d) for %s\n", - imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd)); - cfs_spin_lock(&imp->imp_lock); - imp->imp_last_recon = 1; - cfs_spin_unlock(&imp->imp_lock); - } - } + if (force_verify) + ptlrpc_pinger_wake_up(); +} - rc = sptlrpc_import_sec_adapt(imp, NULL, 0); - if (rc) - GOTO(out, rc); - - /* Reset connect flags to the originally requested flags, in case - * the server is updated on-the-fly we will get the new features. */ - imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig; - imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; - - rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd, - &obd->obd_uuid, &imp->imp_connect_data, NULL); - if (rc) - GOTO(out, rc); - - request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT); - if (request == NULL) - GOTO(out, rc = -ENOMEM); - - rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION, - imp->imp_connect_op, bufs, NULL); - if (rc) { - ptlrpc_request_free(request); - GOTO(out, rc); - } +static int ptlrpc_busy_reconnect(int rc) +{ + return (rc == -EBUSY) || (rc == -EAGAIN); +} - /* Report the rpc service time to the server so that it knows how long - * to wait for clients to join recovery */ - lustre_msg_set_service_time(request->rq_reqmsg, - at_timeout2est(request->rq_timeout)); - - /* The amount of time we give the server to process the connect req. - * import_select_connection will increase the net latency on - * repeated reconnect attempts to cover slow networks. - * We override/ignore the server rpc completion estimate here, - * which may be large if this is a reconnect attempt */ - request->rq_timeout = INITIAL_CONNECT_TIMEOUT; - lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout); - -#ifndef __KERNEL__ - lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT); +static int ptlrpc_connect_set_flags(struct obd_import *imp, + struct obd_connect_data *ocd, + __u64 old_connect_flags, + struct obd_export *exp, int init_connect) +{ + static bool warned; + struct client_obd *cli = &imp->imp_obd->u.cli; + + if ((imp->imp_connect_flags_orig & OBD_CONNECT_IBITS) && + !(ocd->ocd_connect_flags & OBD_CONNECT_IBITS)) { + LCONSOLE_WARN("%s: MDS %s does not support ibits " + "lock, either very old or invalid: " + "requested %#llx, replied %#llx\n", + imp->imp_obd->obd_name, + imp->imp_connection->c_remote_uuid.uuid, + imp->imp_connect_flags_orig, + ocd->ocd_connect_flags); + return -EPROTO; + } + + spin_lock(&imp->imp_lock); + list_del(&imp->imp_conn_current->oic_item); + list_add(&imp->imp_conn_current->oic_item, + &imp->imp_conn_list); + imp->imp_last_success_conn = + imp->imp_conn_current->oic_last_attempt; + + spin_unlock(&imp->imp_lock); + + + if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) && + (ocd->ocd_version > LUSTRE_VERSION_CODE + + LUSTRE_VERSION_OFFSET_WARN || + ocd->ocd_version < LUSTRE_VERSION_CODE - + LUSTRE_VERSION_OFFSET_WARN)) { + /* Sigh, some compilers do not like #ifdef in the middle + of macro arguments */ + const char *older = "older than client. " + "Consider upgrading server"; + const char *newer = "newer than client. " + "Consider recompiling application"; + + LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) " + "is much %s (%s)\n", + obd2cli_tgt(imp->imp_obd), + OBD_OCD_VERSION_MAJOR(ocd->ocd_version), + OBD_OCD_VERSION_MINOR(ocd->ocd_version), + OBD_OCD_VERSION_PATCH(ocd->ocd_version), + OBD_OCD_VERSION_FIX(ocd->ocd_version), + ocd->ocd_version > LUSTRE_VERSION_CODE ? + newer : older, LUSTRE_VERSION_STRING); + warned = true; + } + +#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0) + /* Check if server has LU-1252 fix applied to not always swab + * the IR MNE entries. Do this only once per connection. 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 servers. For newer servers, + * the client will do MNE swabbing only as needed. LU-1644 */ + if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) && + !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) && + OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 && + OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 && + OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 && + strcmp(imp->imp_obd->obd_type->typ_name, + LUSTRE_MGC_NAME) == 0)) + imp->imp_need_mne_swab = 1; + else /* clear if server was upgraded since last connect */ + imp->imp_need_mne_swab = 0; #endif - lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER); - - request->rq_no_resend = request->rq_no_delay = 1; - request->rq_send_state = LUSTRE_IMP_CONNECTING; - /* Allow a slightly larger reply for future growth compatibility */ - req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER, - sizeof(struct obd_connect_data)+16*sizeof(__u64)); - ptlrpc_request_set_replen(request); - request->rq_interpret_reply = ptlrpc_connect_interpret; - - CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args)); - aa = ptlrpc_req_async_args(request); - memset(aa, 0, sizeof *aa); - - aa->pcaa_peer_committed = committed_before_reconnect; - aa->pcaa_initial_connect = initial_connect; - - if (aa->pcaa_initial_connect) { - cfs_spin_lock(&imp->imp_lock); - imp->imp_replayable = 1; - cfs_spin_unlock(&imp->imp_lock); - lustre_msg_add_op_flags(request->rq_reqmsg, - MSG_CONNECT_INITIAL); - } - if (set_transno) - lustre_msg_add_op_flags(request->rq_reqmsg, - MSG_CONNECT_TRANSNO); - - DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)", - request->rq_timeout); - ptlrpcd_add_req(request, PSCOPE_OTHER); - rc = 0; -out: - if (rc != 0) { - IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON); - } - - RETURN(rc); + if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) { + /* We sent to the server ocd_cksum_types with bits set + * for algorithms we understand. The server masked off + * the checksum types it doesn't support */ + if ((ocd->ocd_cksum_types & + cksum_types_supported_client()) == 0) { + LCONSOLE_WARN("The negotiation of the checksum " + "alogrithm to use with server %s " + "failed (%x/%x), disabling " + "checksums\n", + obd2cli_tgt(imp->imp_obd), + ocd->ocd_cksum_types, + cksum_types_supported_client()); + cli->cl_checksum = 0; + cli->cl_supp_cksum_types = OBD_CKSUM_ADLER; + } else { + cli->cl_supp_cksum_types = ocd->ocd_cksum_types; + } + } else { + /* The server does not support OBD_CONNECT_CKSUM. + * Enforce ADLER for backward compatibility*/ + cli->cl_supp_cksum_types = OBD_CKSUM_ADLER; + } + cli->cl_cksum_type = cksum_type_select(cli->cl_supp_cksum_types); + + if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) + cli->cl_max_pages_per_rpc = + min(ocd->ocd_brw_size >> PAGE_CACHE_SHIFT, + cli->cl_max_pages_per_rpc); + else if (imp->imp_connect_op == MDS_CONNECT || + imp->imp_connect_op == MGS_CONNECT) + cli->cl_max_pages_per_rpc = 1; + + LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) && + (cli->cl_max_pages_per_rpc > 0)); + + client_adjust_max_dirty(cli); + + /* Update client max modify RPCs in flight with value returned + * by the server */ + if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) + cli->cl_max_mod_rpcs_in_flight = min( + cli->cl_max_mod_rpcs_in_flight, + ocd->ocd_maxmodrpcs); + else + cli->cl_max_mod_rpcs_in_flight = 1; + + /* Reset ns_connect_flags only for initial connect. It might be + * changed in while using FS and if we reset it in reconnect + * this leads to losing user settings done before such as + * disable lru_resize, etc. */ + if (old_connect_flags != exp_connect_flags(exp) || init_connect) { + CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server " + "flags: %#llx\n", imp->imp_obd->obd_name, + ocd->ocd_connect_flags); + imp->imp_obd->obd_namespace->ns_connect_flags = + ocd->ocd_connect_flags; + imp->imp_obd->obd_namespace->ns_orig_connect_flags = + ocd->ocd_connect_flags; + } + + if (ocd->ocd_connect_flags & OBD_CONNECT_AT) + /* We need a per-message support flag, because + * a. we don't know if the incoming connect reply + * supports AT or not (in reply_in_callback) + * until we unpack it. + * b. failovered server means export and flags are gone + * (in ptlrpc_send_reply). + * Can only be set when we know AT is supported at + * both ends */ + imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT; + else + imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; + + imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18; + + return 0; } -EXPORT_SYMBOL(ptlrpc_connect_import); -static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp) +/** + * Add all replay requests back to unreplied list before start replay, + * so that we can make sure the known replied XID is always increased + * only even if when replaying requests. + */ +static void ptlrpc_prepare_replay(struct obd_import *imp) { -#ifdef __KERNEL__ - struct obd_import_conn *imp_conn; -#endif - int wake_pinger = 0; - - ENTRY; + struct ptlrpc_request *req; - cfs_spin_lock(&imp->imp_lock); - if (cfs_list_empty(&imp->imp_conn_list)) - GOTO(unlock, 0); - -#ifdef __KERNEL__ - imp_conn = cfs_list_entry(imp->imp_conn_list.prev, - struct obd_import_conn, - oic_item); - - /* XXX: When the failover node is the primary node, it is possible - * to have two identical connections in imp_conn_list. We must - * compare not conn's pointers but NIDs, otherwise we can defeat - * connection throttling. (See bug 14774.) */ - if (imp->imp_conn_current->oic_conn->c_peer.nid != - imp_conn->oic_conn->c_peer.nid) { - ptlrpc_ping_import_soon(imp); - wake_pinger = 1; - } -#else - /* liblustre has no pinger thread, so we wakeup pinger anyway */ - wake_pinger = 1; -#endif + if (imp->imp_state != LUSTRE_IMP_REPLAY || + imp->imp_resend_replay) + return; - unlock: - cfs_spin_unlock(&imp->imp_lock); + /* If the server was restart during repaly, the requests may + * have been added to the unreplied list in former replay. */ + spin_lock(&imp->imp_lock); - if (wake_pinger) - ptlrpc_pinger_wake_up(); + list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) { + if (list_empty(&req->rq_unreplied_list)) + ptlrpc_add_unreplied(req); + } - EXIT; -} + list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) { + if (list_empty(&req->rq_unreplied_list)) + ptlrpc_add_unreplied(req); + } -static int ptlrpc_busy_reconnect(int rc) -{ - return (rc == -EBUSY) || (rc == -EAGAIN); + imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp); + spin_unlock(&imp->imp_lock); } - +/** + * interpret_reply callback for connect RPCs. + * Looks into returned status of connect operation and decides + * what to do with the import - i.e enter recovery, promote it to + * full state for normal operations of disconnect it due to an error. + */ static int ptlrpc_connect_interpret(const struct lu_env *env, struct ptlrpc_request *request, void *data, int rc) { struct ptlrpc_connect_async_args *aa = data; struct obd_import *imp = request->rq_import; - struct client_obd *cli = &imp->imp_obd->u.cli; struct lustre_handle old_hdl; __u64 old_connect_flags; int msg_flags; - ENTRY; - - cfs_spin_lock(&imp->imp_lock); - if (imp->imp_state == LUSTRE_IMP_CLOSED) { - cfs_spin_unlock(&imp->imp_lock); - RETURN(0); - } - - if (rc) { - /* if this reconnect to busy export - not need select new target - * for connecting*/ - imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc); - cfs_spin_unlock(&imp->imp_lock); - GOTO(out, rc); - } - - LASSERT(imp->imp_conn_current); - - msg_flags = lustre_msg_get_op_flags(request->rq_repmsg); + struct obd_connect_data *ocd; + struct obd_export *exp; + int ret; + ENTRY; + + spin_lock(&imp->imp_lock); + if (imp->imp_state == LUSTRE_IMP_CLOSED) { + imp->imp_connect_tried = 1; + spin_unlock(&imp->imp_lock); + RETURN(0); + } + + if (rc) { + /* if this reconnect to busy export - not need select new target + * for connecting*/ + imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc); + spin_unlock(&imp->imp_lock); + ptlrpc_maybe_ping_import_soon(imp); + GOTO(out, rc); + } + + /* LU-7558: indicate that we are interpretting connect reply, + * pltrpc_connect_import() will not try to reconnect until + * interpret will finish. */ + imp->imp_connected = 1; + spin_unlock(&imp->imp_lock); + + LASSERT(imp->imp_conn_current); + + msg_flags = lustre_msg_get_op_flags(request->rq_repmsg); + + ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA, + RCL_SERVER); + /* server replied obd_connect_data is always bigger */ + ocd = req_capsule_server_sized_get(&request->rq_pill, + &RMF_CONNECT_DATA, ret); + + if (ocd == NULL) { + CERROR("%s: no connect data from server\n", + imp->imp_obd->obd_name); + rc = -EPROTO; + GOTO(out, rc); + } + + spin_lock(&imp->imp_lock); /* All imports are pingable */ imp->imp_pingable = 1; imp->imp_force_reconnect = 0; - - if (aa->pcaa_initial_connect) { - if (msg_flags & MSG_CONNECT_REPLAYABLE) { - imp->imp_replayable = 1; - cfs_spin_unlock(&imp->imp_lock); - CDEBUG(D_HA, "connected to replayable target: %s\n", - obd2cli_tgt(imp->imp_obd)); - } else { - imp->imp_replayable = 0; - cfs_spin_unlock(&imp->imp_lock); - } + imp->imp_force_verify = 0; + + imp->imp_connect_data = *ocd; + + CDEBUG(D_HA, "%s: connect to target with instance %u\n", + imp->imp_obd->obd_name, ocd->ocd_instance); + exp = class_conn2export(&imp->imp_dlm_handle); + + spin_unlock(&imp->imp_lock); + + if (!exp) { + /* This could happen if export is cleaned during the + connect attempt */ + CERROR("%s: missing export after connect\n", + imp->imp_obd->obd_name); + GOTO(out, rc = -ENODEV); + } + + /* check that server granted subset of flags we asked for. */ + if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) != + ocd->ocd_connect_flags) { + CERROR("%s: Server didn't grant requested subset of flags: " + "asked=%#llx granted=%#llx\n", + imp->imp_obd->obd_name, imp->imp_connect_flags_orig, + ocd->ocd_connect_flags); + GOTO(out, rc = -EPROTO); + } + + if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) != + ocd->ocd_connect_flags2) { + CERROR("%s: Server didn't grant requested subset of flags2: " + "asked=%#llx granted=%#llx\n", + imp->imp_obd->obd_name, imp->imp_connect_flags2_orig, + ocd->ocd_connect_flags2); + GOTO(out, rc = -EPROTO); + } + + if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) && + (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) && + (imp->imp_connect_flags_orig & OBD_CONNECT_FID) && + (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) { + __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version); + __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version); + __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version); + + /* We do not support the MDT-MDT interoperations with + * different version MDT because of protocol changes. */ + if (unlikely(major != LUSTRE_MAJOR || + minor != LUSTRE_MINOR || + abs(patch - LUSTRE_PATCH) > 3)) { + LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the " + "connection to different version MDT " + "(%d.%d.%d.%d) %s\n", + imp->imp_obd->obd_name, imp, LUSTRE_MAJOR, + LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX, + major, minor, patch, + OBD_OCD_VERSION_FIX(ocd->ocd_version), + imp->imp_connection->c_remote_uuid.uuid); + + GOTO(out, rc = -EPROTO); + } + } + + old_connect_flags = exp_connect_flags(exp); + exp->exp_connect_data = *ocd; + imp->imp_obd->obd_self_export->exp_connect_data = *ocd; + + /* The net statistics after (re-)connect is not valid anymore, + * because may reflect other routing, etc. */ + at_reinit(&imp->imp_at.iat_net_latency, 0, 0); + ptlrpc_at_adj_net_latency(request, + lustre_msg_get_service_time(request->rq_repmsg)); + + /* Import flags should be updated before waking import at FULL state */ + rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp, + aa->pcaa_initial_connect); + class_export_put(exp); + if (rc != 0) + GOTO(out, rc); + + obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD); + + if (aa->pcaa_initial_connect) { + spin_lock(&imp->imp_lock); + if (msg_flags & MSG_CONNECT_REPLAYABLE) { + imp->imp_replayable = 1; + spin_unlock(&imp->imp_lock); + CDEBUG(D_HA, "connected to replayable target: %s\n", + obd2cli_tgt(imp->imp_obd)); + } else { + imp->imp_replayable = 0; + spin_unlock(&imp->imp_lock); + } /* if applies, adjust the imp->imp_msg_magic here * according to reply flags */ @@ -848,13 +1147,11 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, obd2cli_tgt(imp->imp_obd)); IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS); } else { - IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL); - ptlrpc_activate_import(imp); + IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL); + ptlrpc_activate_import(imp); } GOTO(finish, rc = 0); - } else { - cfs_spin_unlock(&imp->imp_lock); } /* Determine what recovery state to move the import to. */ @@ -862,10 +1159,11 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, memset(&old_hdl, 0, sizeof(old_hdl)); if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg), sizeof (old_hdl))) { - CERROR("%s@%s didn't like our handle "LPX64 - ", failed\n", obd2cli_tgt(imp->imp_obd), - imp->imp_connection->c_remote_uuid.uuid, - imp->imp_dlm_handle.cookie); + LCONSOLE_WARN("Reconnect to %s (at @%s) failed due " + "bad handle %#llx\n", + obd2cli_tgt(imp->imp_obd), + imp->imp_connection->c_remote_uuid.uuid, + imp->imp_dlm_handle.cookie); GOTO(out, rc = -ENOTCONN); } @@ -882,15 +1180,27 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, * eviction. If it is in recovery - we are safe to * participate since we can reestablish all of our state * with server again */ - CDEBUG(level,"%s@%s changed server handle from " - LPX64" to "LPX64"%s\n", - obd2cli_tgt(imp->imp_obd), - imp->imp_connection->c_remote_uuid.uuid, - imp->imp_remote_handle.cookie, - lustre_msg_get_handle(request->rq_repmsg)-> - cookie, - (MSG_CONNECT_RECOVERING & msg_flags) ? - " but is still in recovery" : ""); + if ((MSG_CONNECT_RECOVERING & msg_flags)) { + CDEBUG(level,"%s@%s changed server handle from " + "%#llx to %#llx" + " but is still in recovery\n", + obd2cli_tgt(imp->imp_obd), + imp->imp_connection->c_remote_uuid.uuid, + imp->imp_remote_handle.cookie, + lustre_msg_get_handle( + request->rq_repmsg)->cookie); + } else { + LCONSOLE_WARN("Evicted from %s (at %s) " + "after server handle changed from " + "%#llx to %#llx\n", + obd2cli_tgt(imp->imp_obd), + imp->imp_connection-> \ + c_remote_uuid.uuid, + imp->imp_remote_handle.cookie, + lustre_msg_get_handle( + request->rq_repmsg)->cookie); + } + imp->imp_remote_handle = *lustre_msg_get_handle(request->rq_repmsg); @@ -915,11 +1225,11 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd)); - cfs_spin_lock(&imp->imp_lock); - imp->imp_resend_replay = 1; - cfs_spin_unlock(&imp->imp_lock); + spin_lock(&imp->imp_lock); + imp->imp_resend_replay = 1; + spin_unlock(&imp->imp_lock); - IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY); + IMPORT_SET_STATE(imp, imp->imp_replay_state); } else { IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER); } @@ -928,6 +1238,7 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, imp->imp_remote_handle = *lustre_msg_get_handle(request->rq_repmsg); imp->imp_last_replay_transno = 0; + imp->imp_replay_cursor = &imp->imp_committed_list; IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY); } else { DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags" @@ -943,10 +1254,11 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, "after reconnect. We should LBUG right here.\n"); } - if (lustre_msg_get_last_committed(request->rq_repmsg) < + if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 && + lustre_msg_get_last_committed(request->rq_repmsg) < aa->pcaa_peer_committed) { - CERROR("%s went back in time (transno "LPD64 - " was previously committed, server now claims "LPD64 + CERROR("%s went back in time (transno %lld" + " was previously committed, server now claims %lld" ")! See https://bugzilla.lustre.org/show_bug.cgi?" "id=9646\n", obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed, @@ -954,180 +1266,30 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, } finish: - rc = ptlrpc_import_recovery_state_machine(imp); - if (rc != 0) { - if (rc == -ENOTCONN) { - CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;" - "invalidating and reconnecting\n", - obd2cli_tgt(imp->imp_obd), - imp->imp_connection->c_remote_uuid.uuid); - ptlrpc_connect_import(imp, NULL); - RETURN(0); - } - } else { - struct obd_connect_data *ocd; - struct obd_export *exp; - int ret; - ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA, - RCL_SERVER); - /* server replied obd_connect_data is always bigger */ - ocd = req_capsule_server_sized_get(&request->rq_pill, - &RMF_CONNECT_DATA, ret); - - cfs_spin_lock(&imp->imp_lock); - cfs_list_del(&imp->imp_conn_current->oic_item); - cfs_list_add(&imp->imp_conn_current->oic_item, - &imp->imp_conn_list); - imp->imp_last_success_conn = - imp->imp_conn_current->oic_last_attempt; - - if (ocd == NULL) { - cfs_spin_unlock(&imp->imp_lock); - CERROR("Wrong connect data from server\n"); - rc = -EPROTO; - GOTO(out, rc); - } - - imp->imp_connect_data = *ocd; - - exp = class_conn2export(&imp->imp_dlm_handle); - cfs_spin_unlock(&imp->imp_lock); - - /* check that server granted subset of flags we asked for. */ - LASSERTF((ocd->ocd_connect_flags & - imp->imp_connect_flags_orig) == - ocd->ocd_connect_flags, LPX64" != "LPX64, - imp->imp_connect_flags_orig, ocd->ocd_connect_flags); - - if (!exp) { - /* This could happen if export is cleaned during the - connect attempt */ - CERROR("Missing export for %s\n", - imp->imp_obd->obd_name); - GOTO(out, rc = -ENODEV); - } - old_connect_flags = exp->exp_connect_flags; - exp->exp_connect_flags = ocd->ocd_connect_flags; - imp->imp_obd->obd_self_export->exp_connect_flags = - ocd->ocd_connect_flags; - class_export_put(exp); - - obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD); - - if (!ocd->ocd_ibits_known && - ocd->ocd_connect_flags & OBD_CONNECT_IBITS) - CERROR("Inodebits aware server returned zero compatible" - " bits?\n"); - - if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) && - (ocd->ocd_version > LUSTRE_VERSION_CODE + - LUSTRE_VERSION_OFFSET_WARN || - ocd->ocd_version < LUSTRE_VERSION_CODE - - LUSTRE_VERSION_OFFSET_WARN)) { - /* Sigh, some compilers do not like #ifdef in the middle - of macro arguments */ -#ifdef __KERNEL__ - const char *older = - "older. Consider upgrading this client"; -#else - const char *older = - "older. Consider recompiling this application"; -#endif - const char *newer = "newer than client version"; - - LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) " - "is much %s (%s)\n", - obd2cli_tgt(imp->imp_obd), - OBD_OCD_VERSION_MAJOR(ocd->ocd_version), - OBD_OCD_VERSION_MINOR(ocd->ocd_version), - OBD_OCD_VERSION_PATCH(ocd->ocd_version), - OBD_OCD_VERSION_FIX(ocd->ocd_version), - ocd->ocd_version > LUSTRE_VERSION_CODE ? - newer : older, LUSTRE_VERSION_STRING); - } - - if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) { - /* We sent to the server ocd_cksum_types with bits set - * for algorithms we understand. The server masked off - * the checksum types it doesn't support */ - if ((ocd->ocd_cksum_types & OBD_CKSUM_ALL) == 0) { - LCONSOLE_WARN("The negotiation of the checksum " - "alogrithm to use with server %s " - "failed (%x/%x), disabling " - "checksums\n", - obd2cli_tgt(imp->imp_obd), - ocd->ocd_cksum_types, - OBD_CKSUM_ALL); - cli->cl_checksum = 0; - cli->cl_supp_cksum_types = OBD_CKSUM_CRC32; - cli->cl_cksum_type = OBD_CKSUM_CRC32; - } else { - cli->cl_supp_cksum_types = ocd->ocd_cksum_types; - - if (ocd->ocd_cksum_types & OSC_DEFAULT_CKSUM) - cli->cl_cksum_type = OSC_DEFAULT_CKSUM; - else if (ocd->ocd_cksum_types & OBD_CKSUM_ADLER) - cli->cl_cksum_type = OBD_CKSUM_ADLER; - else - cli->cl_cksum_type = OBD_CKSUM_CRC32; - } - } else { - /* The server does not support OBD_CONNECT_CKSUM. - * Enforce CRC32 for backward compatibility*/ - cli->cl_supp_cksum_types = OBD_CKSUM_CRC32; - cli->cl_cksum_type = OBD_CKSUM_CRC32; - } - - if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) { - cli->cl_max_pages_per_rpc = - ocd->ocd_brw_size >> CFS_PAGE_SHIFT; - } - - /* Reset ns_connect_flags only for initial connect. It might be - * changed in while using FS and if we reset it in reconnect - * this leads to losing user settings done before such as - * disable lru_resize, etc. */ - if (old_connect_flags != exp->exp_connect_flags || - aa->pcaa_initial_connect) { - CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server " - "flags: "LPX64"\n", imp->imp_obd->obd_name, - ocd->ocd_connect_flags); - imp->imp_obd->obd_namespace->ns_connect_flags = - ocd->ocd_connect_flags; - imp->imp_obd->obd_namespace->ns_orig_connect_flags = - ocd->ocd_connect_flags; - } - - if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) && - (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2)) - /* We need a per-message support flag, because - a. we don't know if the incoming connect reply - supports AT or not (in reply_in_callback) - until we unpack it. - b. failovered server means export and flags are gone - (in ptlrpc_send_reply). - Can only be set when we know AT is supported at - both ends */ - imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT; - else - imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; - - LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) && - (cli->cl_max_pages_per_rpc > 0)); - } + ptlrpc_prepare_replay(imp); + rc = ptlrpc_import_recovery_state_machine(imp); + if (rc == -ENOTCONN) { + CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;" + "invalidating and reconnecting\n", + obd2cli_tgt(imp->imp_obd), + imp->imp_connection->c_remote_uuid.uuid); + ptlrpc_connect_import(imp); + spin_lock(&imp->imp_lock); + imp->imp_connected = 0; + imp->imp_connect_tried = 1; + spin_unlock(&imp->imp_lock); + RETURN(0); + } out: + spin_lock(&imp->imp_lock); + imp->imp_connected = 0; + imp->imp_connect_tried = 1; + spin_unlock(&imp->imp_lock); + if (rc != 0) { IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON); - cfs_spin_lock(&imp->imp_lock); - if (aa->pcaa_initial_connect && !imp->imp_initial_recov && - (request->rq_import_generation == imp->imp_generation)) - ptlrpc_deactivate_and_unlock_import(imp); - else - cfs_spin_unlock(&imp->imp_lock); - - if ((imp->imp_recon_bk && imp->imp_last_recon) || - (rc == -EACCES)) { + if (rc == -EACCES) { /* * Give up trying to reconnect * EACCES means client has no permission for connection @@ -1168,82 +1330,93 @@ out: RETURN(-EPROTO); } - ptlrpc_maybe_ping_import_soon(imp); + ptlrpc_maybe_ping_import_soon(imp); - CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n", - obd2cli_tgt(imp->imp_obd), - (char *)imp->imp_connection->c_remote_uuid.uuid, rc); - } - - cfs_spin_lock(&imp->imp_lock); - imp->imp_last_recon = 0; - cfs_spin_unlock(&imp->imp_lock); + CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n", + obd2cli_tgt(imp->imp_obd), + (char *)imp->imp_connection->c_remote_uuid.uuid, rc); + } - cfs_waitq_broadcast(&imp->imp_recovery_waitq); - RETURN(rc); + wake_up_all(&imp->imp_recovery_waitq); + RETURN(rc); } +/** + * interpret callback for "completed replay" RPCs. + * \see signal_completed_replay + */ static int completed_replay_interpret(const struct lu_env *env, struct ptlrpc_request *req, void * data, int rc) { - ENTRY; - cfs_atomic_dec(&req->rq_import->imp_replay_inflight); - if (req->rq_status == 0 && - !req->rq_import->imp_vbr_failed) { - ptlrpc_import_recovery_state_machine(req->rq_import); - } else { - if (req->rq_import->imp_vbr_failed) { - CDEBUG(D_WARNING, - "%s: version recovery fails, reconnecting\n", - req->rq_import->imp_obd->obd_name); - } else { - CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, " - "reconnecting\n", - req->rq_import->imp_obd->obd_name, - req->rq_status); - } - ptlrpc_connect_import(req->rq_import, NULL); - } - - RETURN(0); + ENTRY; + atomic_dec(&req->rq_import->imp_replay_inflight); + if (req->rq_status == 0 && + !req->rq_import->imp_vbr_failed) { + ptlrpc_import_recovery_state_machine(req->rq_import); + } else { + if (req->rq_import->imp_vbr_failed) { + CDEBUG(D_WARNING, + "%s: version recovery fails, reconnecting\n", + req->rq_import->imp_obd->obd_name); + } else { + CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, " + "reconnecting\n", + req->rq_import->imp_obd->obd_name, + req->rq_status); + } + ptlrpc_connect_import(req->rq_import); + } + + RETURN(0); } +/** + * Let server know that we have no requests to replay anymore. + * Achieved by just sending a PING request + */ static int signal_completed_replay(struct obd_import *imp) { - struct ptlrpc_request *req; - ENTRY; - - LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0); - cfs_atomic_inc(&imp->imp_replay_inflight); - - req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION, - OBD_PING); - if (req == NULL) { - cfs_atomic_dec(&imp->imp_replay_inflight); - RETURN(-ENOMEM); - } - - ptlrpc_request_set_replen(req); - req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT; - lustre_msg_add_flags(req->rq_reqmsg, - MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE); - if (AT_OFF) - req->rq_timeout *= 3; - req->rq_interpret_reply = completed_replay_interpret; - - ptlrpcd_add_req(req, PSCOPE_OTHER); - RETURN(0); + struct ptlrpc_request *req; + ENTRY; + + if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY))) + RETURN(0); + + LASSERT(atomic_read(&imp->imp_replay_inflight) == 0); + atomic_inc(&imp->imp_replay_inflight); + + req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION, + OBD_PING); + if (req == NULL) { + atomic_dec(&imp->imp_replay_inflight); + RETURN(-ENOMEM); + } + + ptlrpc_request_set_replen(req); + req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT; + lustre_msg_add_flags(req->rq_reqmsg, + MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE); + if (AT_OFF) + req->rq_timeout *= 3; + req->rq_interpret_reply = completed_replay_interpret; + + ptlrpcd_add_req(req); + RETURN(0); } -#ifdef __KERNEL__ +/** + * In kernel code all import invalidation happens in its own + * separate thread, so that whatever application happened to encounter + * a problem could still be killed or otherwise continue + */ static int ptlrpc_invalidate_import_thread(void *data) { struct obd_import *imp = data; ENTRY; - cfs_daemonize_ctxt("ll_imp_inval"); + unshare_fs_struct(); CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n", imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd), @@ -1262,8 +1435,27 @@ static int ptlrpc_invalidate_import_thread(void *data) class_import_put(imp); RETURN(0); } -#endif +/** + * This is the state machine for client-side recovery on import. + * + * Typicaly we have two possibly paths. If we came to server and it is not + * in recovery, we just enter IMP_EVICTED state, invalidate our import + * state and reconnect from scratch. + * If we came to server that is in recovery, we enter IMP_REPLAY import state. + * We go through our list of requests to replay and send them to server one by + * one. + * After sending all request from the list we change import state to + * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server + * and also all the locks we don't yet have and wait for server to grant us. + * After that we send a special "replay completed" request and change import + * state to IMP_REPLAY_WAIT. + * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER + * state and resend all requests from sending list. + * After that we promote import to FULL state and send all delayed requests + * and import is fully operational after that. + * + */ int ptlrpc_import_recovery_state_machine(struct obd_import *imp) { int rc = 0; @@ -1278,75 +1470,71 @@ int ptlrpc_import_recovery_state_machine(struct obd_import *imp) /* Don't care about MGC eviction */ if (strcmp(imp->imp_obd->obd_type->typ_name, LUSTRE_MGC_NAME) != 0) { - LCONSOLE_ERROR_MSG(0x167, "This client was evicted by " - "%.*s; in progress operations using " - "this service will fail.\n", - target_len, target_start); + LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted " + "by %.*s; in progress operations " + "using this service will fail.\n", + imp->imp_obd->obd_name, target_len, + target_start); } CDEBUG(D_HA, "evicted from %s@%s; invalidating\n", obd2cli_tgt(imp->imp_obd), imp->imp_connection->c_remote_uuid.uuid); /* reset vbr_failed flag upon eviction */ - cfs_spin_lock(&imp->imp_lock); - imp->imp_vbr_failed = 0; - cfs_spin_unlock(&imp->imp_lock); - -#ifdef __KERNEL__ - /* bug 17802: XXX client_disconnect_export vs connect request - * race. if client will evicted at this time, we start - * invalidate thread without reference to import and import can - * be freed at same time. */ - class_import_get(imp); - rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp, - CLONE_VM | CLONE_FILES); - if (rc < 0) { - class_import_put(imp); - CERROR("error starting invalidate thread: %d\n", rc); - } else { - rc = 0; - } - RETURN(rc); -#else - ptlrpc_invalidate_import(imp); - - IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER); -#endif + spin_lock(&imp->imp_lock); + imp->imp_vbr_failed = 0; + spin_unlock(&imp->imp_lock); + + { + struct task_struct *task; + /* bug 17802: XXX client_disconnect_export vs connect request + * race. if client is evicted at this time then we start + * invalidate thread without reference to import and import can + * be freed at same time. */ + class_import_get(imp); + task = kthread_run(ptlrpc_invalidate_import_thread, imp, + "ll_imp_inval"); + if (IS_ERR(task)) { + class_import_put(imp); + CERROR("error starting invalidate thread: %d\n", rc); + rc = PTR_ERR(task); + } else { + rc = 0; + } + RETURN(rc); + } } - if (imp->imp_state == LUSTRE_IMP_REPLAY) { - CDEBUG(D_HA, "replay requested by %s\n", - obd2cli_tgt(imp->imp_obd)); - rc = ptlrpc_replay_next(imp, &inflight); - if (inflight == 0 && - cfs_atomic_read(&imp->imp_replay_inflight) == 0) { - IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS); - rc = ldlm_replay_locks(imp); - if (rc) - GOTO(out, rc); - } - rc = 0; - } - - if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) { - if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) { - IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT); - rc = signal_completed_replay(imp); - if (rc) - GOTO(out, rc); - } - - } - - if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) { - if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) { - IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER); - } - } + if (imp->imp_state == LUSTRE_IMP_REPLAY) { + CDEBUG(D_HA, "replay requested by %s\n", + obd2cli_tgt(imp->imp_obd)); + rc = ptlrpc_replay_next(imp, &inflight); + if (inflight == 0 && + atomic_read(&imp->imp_replay_inflight) == 0) { + IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS); + rc = ldlm_replay_locks(imp); + if (rc) + GOTO(out, rc); + } + rc = 0; + } + + if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) { + if (atomic_read(&imp->imp_replay_inflight) == 0) { + IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT); + rc = signal_completed_replay(imp); + if (rc) + GOTO(out, rc); + } + } + + if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) { + if (atomic_read(&imp->imp_replay_inflight) == 0) { + IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER); + } + } if (imp->imp_state == LUSTRE_IMP_RECOVER) { - CDEBUG(D_HA, "reconnected to %s@%s\n", - obd2cli_tgt(imp->imp_obd), - imp->imp_connection->c_remote_uuid.uuid); + struct ptlrpc_connection *conn = imp->imp_connection; rc = ptlrpc_resend(imp); if (rc) @@ -1354,48 +1542,53 @@ int ptlrpc_import_recovery_state_machine(struct obd_import *imp) IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL); ptlrpc_activate_import(imp); - deuuidify(obd2cli_tgt(imp->imp_obd), NULL, - &target_start, &target_len); - LCONSOLE_INFO("%s: Connection restored to service %.*s " - "using nid %s.\n", imp->imp_obd->obd_name, - target_len, target_start, - libcfs_nid2str(imp->imp_connection->c_peer.nid)); + LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n", + imp->imp_obd->obd_name, + obd_uuid2str(&conn->c_remote_uuid), + libcfs_nid2str(imp->imp_connection->c_peer.nid)); } - if (imp->imp_state == LUSTRE_IMP_FULL) { - cfs_waitq_broadcast(&imp->imp_recovery_waitq); - ptlrpc_wake_delayed(imp); - } + if (imp->imp_state == LUSTRE_IMP_FULL) { + wake_up_all(&imp->imp_recovery_waitq); + ptlrpc_wake_delayed(imp); + } out: - RETURN(rc); + RETURN(rc); } int ptlrpc_disconnect_import(struct obd_import *imp, int noclose) { - struct ptlrpc_request *req; - int rq_opc, rc = 0; - int nowait = imp->imp_obd->obd_force; - ENTRY; - - if (nowait) - GOTO(set_state, rc); - - switch (imp->imp_connect_op) { - case OST_CONNECT: rq_opc = OST_DISCONNECT; break; - case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break; - case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break; - default: - CERROR("don't know how to disconnect from %s (connect_op %d)\n", - obd2cli_tgt(imp->imp_obd), imp->imp_connect_op); - RETURN(-EINVAL); - } + struct ptlrpc_request *req; + int rq_opc, rc = 0; + ENTRY; + + if (imp->imp_obd->obd_force) + GOTO(set_state, rc); + + switch (imp->imp_connect_op) { + case OST_CONNECT: + rq_opc = OST_DISCONNECT; + break; + case MDS_CONNECT: + rq_opc = MDS_DISCONNECT; + break; + case MGS_CONNECT: + rq_opc = MGS_DISCONNECT; + break; + default: + rc = -EINVAL; + CERROR("%s: don't know how to disconnect from %s " + "(connect_op %d): rc = %d\n", + imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd), + imp->imp_connect_op, rc); + RETURN(rc); + } if (ptlrpc_import_in_recovery(imp)) { struct l_wait_info lwi; cfs_duration_t timeout; - if (AT_OFF) { if (imp->imp_server_timeout) timeout = cfs_time_seconds(obd_timeout / 2); @@ -1415,11 +1608,10 @@ int ptlrpc_disconnect_import(struct obd_import *imp, int noclose) } - cfs_spin_lock(&imp->imp_lock); - if (imp->imp_state != LUSTRE_IMP_FULL) - GOTO(out, 0); - - cfs_spin_unlock(&imp->imp_lock); + spin_lock(&imp->imp_lock); + if (imp->imp_state != LUSTRE_IMP_FULL) + GOTO(out, rc); + spin_unlock(&imp->imp_lock); req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION, rq_opc); @@ -1429,17 +1621,10 @@ int ptlrpc_disconnect_import(struct obd_import *imp, int noclose) * if the client doesn't know the server is gone yet. */ req->rq_no_resend = 1; -#ifndef CRAY_XT3 /* We want client umounts to happen quickly, no matter the server state... */ req->rq_timeout = min_t(int, req->rq_timeout, INITIAL_CONNECT_TIMEOUT); -#else - /* ... but we always want liblustre clients to nicely - disconnect, so only use the adaptive value. */ - if (AT_OFF) - req->rq_timeout = obd_timeout / 3; -#endif IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING); req->rq_send_state = LUSTRE_IMP_CONNECTING; @@ -1449,40 +1634,44 @@ int ptlrpc_disconnect_import(struct obd_import *imp, int noclose) } set_state: - cfs_spin_lock(&imp->imp_lock); + spin_lock(&imp->imp_lock); out: - if (noclose) - IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON); - else - IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED); - memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle)); - /* Try all connections in the future - bz 12758 */ - imp->imp_last_recon = 0; - cfs_spin_unlock(&imp->imp_lock); - - RETURN(rc); + if (noclose) + IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON); + else + IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED); + memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle)); + spin_unlock(&imp->imp_lock); + + if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN) + rc = 0; + RETURN(rc); } +EXPORT_SYMBOL(ptlrpc_disconnect_import); void ptlrpc_cleanup_imp(struct obd_import *imp) { - ENTRY; + ENTRY; - cfs_spin_lock(&imp->imp_lock); - IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED); - imp->imp_generation++; - cfs_spin_unlock(&imp->imp_lock); - ptlrpc_abort_inflight(imp); + spin_lock(&imp->imp_lock); + IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED); + imp->imp_generation++; + spin_unlock(&imp->imp_lock); + ptlrpc_abort_inflight(imp); - EXIT; + EXIT; } /* Adaptive Timeout utils */ extern unsigned int at_min, at_max, at_history; -/* Bin into timeslices using AT_BINS bins. - This gives us a max of the last binlimit*AT_BINS secs without the storage, - but still smoothing out a return to normalcy from a slow response. - (E.g. remember the maximum latency in each minute of the last 4 minutes.) */ +/* Update at_current with the specified value (bounded by at_min and at_max), + * as well as the AT history "bins". + * - Bin into timeslices using AT_BINS bins. + * - This gives us a max of the last at_history seconds without the storage, + * but still smoothing out a return to normalcy from a slow response. + * - (E.g. remember the maximum latency in each minute of the last 4 minutes.) + */ int at_measured(struct adaptive_timeout *at, unsigned int val) { unsigned int old = at->at_current; @@ -1499,7 +1688,7 @@ int at_measured(struct adaptive_timeout *at, unsigned int val) drop to 0, and because 0 could mean an error */ return 0; - cfs_spin_lock(&at->at_lock); + spin_lock(&at->at_lock); if (unlikely(at->at_binstart == 0)) { /* Special case to remove default from history */ @@ -1555,7 +1744,7 @@ int at_measured(struct adaptive_timeout *at, unsigned int val) /* if we changed, report the old value */ old = (at->at_current != old) ? old : 0; - cfs_spin_unlock(&at->at_lock); + spin_unlock(&at->at_lock); return old; } @@ -1574,7 +1763,7 @@ int import_at_get_index(struct obd_import *imp, int portal) } /* Not found in list, add it under a lock */ - cfs_spin_lock(&imp->imp_lock); + spin_lock(&imp->imp_lock); /* Check unused under lock */ for (; i < IMP_AT_MAX_PORTALS; i++) { @@ -1590,6 +1779,6 @@ int import_at_get_index(struct obd_import *imp, int portal) at->iat_portal[i] = portal; out: - cfs_spin_unlock(&imp->imp_lock); - return i; + spin_unlock(&imp->imp_lock); + return i; }