X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fptlrpc%2Fimport.c;h=5d1e6b8751b1d952bb15881b820f7b1edf600610;hb=0030ed640284e4ed5afa9187c5a907141132b083;hp=bc65d15884bc645d292d5550eb9eed02f100c68d;hpb=9c6ebfd11be1cc11629a2960c8c818492e1f0980;p=fs%2Flustre-release.git diff --git a/lustre/ptlrpc/import.c b/lustre/ptlrpc/import.c index bc65d15..5d1e6b8 100644 --- a/lustre/ptlrpc/import.c +++ b/lustre/ptlrpc/import.c @@ -27,7 +27,7 @@ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2013, Intel Corporation. + * Copyright (c) 2011, 2014, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -39,9 +39,6 @@ */ #define DEBUG_SUBSYSTEM S_RPC -#ifndef __KERNEL__ -# include -#endif #include #include @@ -66,6 +63,18 @@ struct ptlrpc_connect_async_args { 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 = @@ -132,7 +141,6 @@ void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len) UUID_STR, strlen(UUID_STR))) *uuid_len -= strlen(UUID_STR); } -EXPORT_SYMBOL(deuuidify); /** * Returns true if import was FULL, false if import was already not @@ -198,7 +206,7 @@ int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt) static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp) { ENTRY; - LASSERT(spin_is_locked(&imp->imp_lock)); + assert_spin_locked(&imp->imp_lock); CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd)); imp->imp_invalid = 1; @@ -249,13 +257,13 @@ 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 list_head *tmp, *n; struct ptlrpc_request *req; unsigned int timeout = 0; 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); + 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); @@ -270,18 +278,19 @@ 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; + struct list_head *tmp, *n; + struct ptlrpc_request *req; + struct l_wait_info lwi; + unsigned int timeout; + int rc; - cfs_atomic_inc(&imp->imp_inval_count); + 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 @@ -305,79 +314,77 @@ 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)); + /* 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 (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)); - } + 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: RPCs in \"%s\" phase found (%d). " + "Network is sluggish? Waiting them " + "to error out.\n", cli_tgt, + ptlrpc_phase2str(RQ_PHASE_UNREGISTERING), + 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(cfs_atomic_read(&imp->imp_inflight) == 0); - 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); + } + } 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); @@ -387,12 +394,30 @@ void ptlrpc_activate_import(struct obd_import *imp) 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); +static 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); + + if (imp->imp_state != LUSTRE_IMP_CONNECTING) + ptlrpc_pinger_wake_up(); +} + void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt) { ENTRY; @@ -409,45 +434,56 @@ 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)); - - spin_lock(&imp->imp_lock); - imp->imp_force_verify = 1; - spin_unlock(&imp->imp_lock); - - ptlrpc_pinger_wake_up(); + ptlrpc_pinger_force(imp); } EXIT; } -EXPORT_SYMBOL(ptlrpc_fail_import); 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)); - } +#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; + /* 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); @@ -466,14 +502,14 @@ static int import_select_connection(struct obd_import *imp) spin_lock(&imp->imp_lock); - if (cfs_list_empty(&imp->imp_conn_list)) { + 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) { + list_for_each_entry(conn, &imp->imp_conn_list, oic_item) { CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n", imp->imp_obd->obd_name, libcfs_nid2str(conn->oic_conn->c_peer.nid), @@ -567,20 +603,32 @@ 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; } /** @@ -638,7 +686,7 @@ int ptlrpc_connect_import(struct obd_import *imp) if (rc) GOTO(out, rc); - rc = sptlrpc_import_sec_adapt(imp, NULL, 0); + rc = sptlrpc_import_sec_adapt(imp, NULL, NULL); if (rc) GOTO(out, rc); @@ -679,9 +727,6 @@ int ptlrpc_connect_import(struct obd_import *imp) 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); -#endif lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER); request->rq_no_resend = request->rq_no_delay = 1; @@ -726,7 +771,6 @@ EXPORT_SYMBOL(ptlrpc_connect_import); static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp) { -#ifdef __KERNEL__ int force_verify; spin_lock(&imp->imp_lock); @@ -735,10 +779,6 @@ static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp) if (force_verify) ptlrpc_pinger_wake_up(); -#else - /* liblustre has no pinger thread, so we wakeup pinger anyway */ - ptlrpc_pinger_wake_up(); -#endif } static int ptlrpc_busy_reconnect(int rc) @@ -746,6 +786,155 @@ static int ptlrpc_busy_reconnect(int rc) return (rc == -EBUSY) || (rc == -EAGAIN); } +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 "LPX64", replied "LPX64"\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 + + 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); + + + /* 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: "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; + + if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) && + (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2)) + imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18; + else + imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18; + + return 0; +} + /** * interpret_reply callback for connect RPCs. * Looks into returned status of connect operation and decides @@ -758,7 +947,6 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, { 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; @@ -826,6 +1014,32 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, 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_IMP_RECOV) && + (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); + } + } + if (!exp) { /* This could happen if export is cleaned during the connect attempt */ @@ -836,7 +1050,19 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, 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_init(&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); @@ -866,8 +1092,8 @@ 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); @@ -948,7 +1174,7 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, 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); } @@ -985,152 +1211,15 @@ 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); - imp->imp_connect_tried = 1; - RETURN(0); - } - } else { - - 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; - - spin_unlock(&imp->imp_lock); - - 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 server " - "or downgrading client"; -#else - const char *older = "older. Consider recompiling this " - "application"; -#endif - const char *newer = "newer than client version. " - "Consider upgrading client"; - - 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 LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 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; -#else -#warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab" -#endif - - 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; - - /* 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) || - 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; - - if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) && - (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2)) - imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18; - else - imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18; - - LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) && - (cli->cl_max_pages_per_rpc > 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); + imp->imp_connect_tried = 1; + RETURN(0); + } out: imp->imp_connect_tried = 1; @@ -1178,15 +1267,15 @@ 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); - } + 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); } /** @@ -1197,26 +1286,26 @@ 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); - } + 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); + RETURN(0); } /** @@ -1225,35 +1314,34 @@ static int completed_replay_interpret(const struct lu_env *env, */ static int signal_completed_replay(struct obd_import *imp) { - struct ptlrpc_request *req; - ENTRY; + struct ptlrpc_request *req; + ENTRY; - if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY))) - RETURN(0); + if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY))) + RETURN(0); - LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0); - cfs_atomic_inc(&imp->imp_replay_inflight); + 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) { - cfs_atomic_dec(&imp->imp_replay_inflight); - RETURN(-ENOMEM); - } + 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; + 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, PDL_POLICY_ROUND, -1); - RETURN(0); + ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1); + RETURN(0); } -#ifdef __KERNEL__ /** * In kernel code all import invalidation happens in its own * separate thread, so that whatever application happened to encounter @@ -1284,7 +1372,6 @@ 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. @@ -1334,9 +1421,8 @@ int ptlrpc_import_recovery_state_machine(struct obd_import *imp) imp->imp_vbr_failed = 0; spin_unlock(&imp->imp_lock); -#ifdef __KERNEL__ { - cfs_task_t *task; + struct task_struct *task; /* 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 @@ -1353,42 +1439,36 @@ int ptlrpc_import_recovery_state_machine(struct obd_import *imp) } RETURN(rc); } -#else - ptlrpc_invalidate_import(imp); - - IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER); -#endif - } - - 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) { + 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 (cfs_atomic_read(&imp->imp_replay_inflight) == 0) { - IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER); - } - } + 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", @@ -1409,40 +1489,47 @@ int ptlrpc_import_recovery_state_machine(struct obd_import *imp) 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; + struct ptlrpc_request *req; + int rq_opc, rc = 0; + 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); - } + 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); @@ -1464,8 +1551,7 @@ int ptlrpc_disconnect_import(struct obd_import *imp, int noclose) spin_lock(&imp->imp_lock); if (imp->imp_state != LUSTRE_IMP_FULL) - GOTO(out, 0); - + GOTO(out, rc); spin_unlock(&imp->imp_lock); req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT, @@ -1498,6 +1584,8 @@ out: 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); @@ -1514,15 +1602,17 @@ void ptlrpc_cleanup_imp(struct obd_import *imp) EXIT; } -EXPORT_SYMBOL(ptlrpc_cleanup_imp); /* 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;