X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fptlrpc%2Fimport.c;h=26d7b9892fd6d5b5a9a39fa3534b3ae96722c221;hp=6d798ca9bdc7f6ac6e2f62a720b89b039d2ef361;hb=7462e8cad730897f459da31886c57585654f26b8;hpb=c77e504fdac12d3be7d19a652d6c7da497018c76 diff --git a/lustre/ptlrpc/import.c b/lustre/ptlrpc/import.c index 6d798ca..26d7b98 100644 --- a/lustre/ptlrpc/import.c +++ b/lustre/ptlrpc/import.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2014, Intel Corporation. + * Copyright (c) 2011, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -41,6 +37,7 @@ #define DEBUG_SUBSYSTEM S_RPC #include +#include #include #include #include @@ -59,10 +56,10 @@ struct ptlrpc_connect_async_args { /** * Updates import \a imp current state to provided \a state value - * Helper function. Must be called under imp_lock. + * Helper function. */ -static void __import_set_state(struct obd_import *imp, - enum lustre_imp_state state) +static void import_set_state_nolock(struct obd_import *imp, + enum lustre_imp_state state) { switch (state) { case LUSTRE_IMP_CLOSED: @@ -75,44 +72,46 @@ static void __import_set_state(struct obd_import *imp, break; default: imp->imp_replay_state = LUSTRE_IMP_REPLAY; + break; + } + + /* A CLOSED import should remain so. */ + if (imp->imp_state == LUSTRE_IMP_CLOSED) + return; + + if (imp->imp_state != LUSTRE_IMP_NEW) { + CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n", + imp, obd2cli_tgt(imp->imp_obd), + ptlrpc_import_state_name(imp->imp_state), + ptlrpc_import_state_name(state)); } + 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 = - cfs_time_current_sec(); + ktime_get_real_seconds(); imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) % IMP_STATE_HIST_LEN; } -/* A CLOSED import should remain so. */ -#define IMPORT_SET_STATE_NOLOCK(imp, state) \ -do { \ - if (imp->imp_state != LUSTRE_IMP_CLOSED) { \ - CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n", \ - imp, obd2cli_tgt(imp->imp_obd), \ - ptlrpc_import_state_name(imp->imp_state), \ - ptlrpc_import_state_name(state)); \ - __import_set_state(imp, state); \ - } \ -} while(0) - -#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) +static void import_set_state(struct obd_import *imp, + enum lustre_imp_state new_state) +{ + spin_lock(&imp->imp_lock); + import_set_state_nolock(imp, new_state); + spin_unlock(&imp->imp_lock); +} void ptlrpc_import_enter_resend(struct obd_import *imp) { - IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER); + 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, - void * data, int rc); + struct ptlrpc_request *request, + void *args, int rc); int ptlrpc_import_recovery_state_machine(struct obd_import *imp); /* Only this function is allowed to change the import state when it is @@ -149,6 +148,21 @@ void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len) *uuid_len -= strlen(UUID_STR); } +/* Must be called with imp_lock held! */ +static void ptlrpc_deactivate_import_nolock(struct obd_import *imp) +{ + 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++; + + ptlrpc_abort_inflight(imp); + + EXIT; +} + /** * Returns true if import was FULL, false if import was already not * connected. @@ -159,8 +173,10 @@ void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len) * bulk requests) and if one has already caused a reconnection * (increasing the import->conn_cnt) the older failure should * not also cause a reconnection. If zero it forces a reconnect. + * @invalid - set import invalid flag */ -int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt) +int ptlrpc_set_import_discon(struct obd_import *imp, + __u32 conn_cnt, bool invalid) { int rc = 0; @@ -170,31 +186,43 @@ int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt) (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) { char *target_start; int target_len; + bool inact = false; deuuidify(obd2cli_tgt(imp->imp_obd), NULL, &target_start, &target_len); + import_set_state_nolock(imp, LUSTRE_IMP_DISCON); if (imp->imp_replayable) { 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 " - "%.*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)); - } - IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON); + obd_import_nid2str(imp)); + } else { + 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, + obd_import_nid2str(imp)); + if (invalid) { + CDEBUG(D_HA, "import %s@%s for %s not " + "replayable, auto-deactivating\n", + obd2cli_tgt(imp->imp_obd), + imp->imp_connection->c_remote_uuid.uuid, + imp->imp_obd->obd_name); + ptlrpc_deactivate_import_nolock(imp); + inact = true; + } + } spin_unlock(&imp->imp_lock); if (obd_dump_on_timeout) libcfs_debug_dumplog(); obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON); + + if (inact) + obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE); rc = 1; } else { spin_unlock(&imp->imp_lock); @@ -209,23 +237,6 @@ int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt) return rc; } -/* Must be called with imp_lock held! */ -static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp) -{ - 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++; - spin_unlock(&imp->imp_lock); - - ptlrpc_abort_inflight(imp); - obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE); - - EXIT; -} - /* * This acts as a barrier; all existing requests are rejected, and * no new requests will be accepted until the import is valid again. @@ -233,14 +244,17 @@ static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp) void ptlrpc_deactivate_import(struct obd_import *imp) { spin_lock(&imp->imp_lock); - ptlrpc_deactivate_and_unlock_import(imp); + ptlrpc_deactivate_import_nolock(imp); + spin_unlock(&imp->imp_lock); + + obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE); } EXPORT_SYMBOL(ptlrpc_deactivate_import); -static unsigned int -ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now) +static time64_t ptlrpc_inflight_deadline(struct ptlrpc_request *req, + time64_t now) { - long dl; + time64_t dl; if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) || (req->rq_phase == RQ_PHASE_BULK) || @@ -261,18 +275,15 @@ ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now) return dl - now; } -static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp) +static time64_t ptlrpc_inflight_timeout(struct obd_import *imp) { - time_t now = cfs_time_current_sec(); - struct list_head *tmp, *n; + time64_t now = ktime_get_real_seconds(); struct ptlrpc_request *req; - unsigned int timeout = 0; + time64_t 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); + list_for_each_entry(req, &imp->imp_sending_list, rq_list) timeout = max(ptlrpc_inflight_deadline(req, now), timeout); - } spin_unlock(&imp->imp_lock); return timeout; } @@ -285,10 +296,8 @@ static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp) */ void ptlrpc_invalidate_import(struct obd_import *imp) { - struct list_head *tmp, *n; struct ptlrpc_request *req; - struct l_wait_info lwi; - unsigned int timeout; + time64_t timeout; int rc; atomic_inc(&imp->imp_inval_count); @@ -296,46 +305,54 @@ void ptlrpc_invalidate_import(struct obd_import *imp) if (!imp->imp_invalid || imp->imp_obd->obd_no_recov) ptlrpc_deactivate_import(imp); + if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CONNECT_RACE)) { + OBD_RACE(OBD_FAIL_PTLRPC_CONNECT_RACE); + msleep(10 * MSEC_PER_SEC); + } 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 - * unlink. We can't do anything before that because there is really - * no guarantee that some rdma transfer is not in progress right now. */ - do { - /* Calculate max timeout for waiting on rpcs to error - * out. Use obd_timeout if calculated value is smaller - * than it. */ - if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) { - timeout = ptlrpc_inflight_timeout(imp); - timeout += timeout / 3; - - if (timeout == 0) - timeout = obd_timeout; - } else { - /* decrease the interval to increase race condition */ - timeout = 1; - } + /* 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 + * unlink. We can't do anything before that because there is really + * no guarantee that some rdma transfer is not in progress right now. + */ + do { + long timeout_jiffies; + + /* Calculate max timeout for waiting on rpcs to error + * out. Use obd_timeout if calculated value is smaller + * than it. + */ + if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) { + timeout = ptlrpc_inflight_timeout(imp); + timeout += div_u64(timeout, 3); + + if (timeout == 0) + timeout = obd_timeout; + } else { + /* decrease the interval to increase race condition */ + timeout = 1; + } - CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n", - timeout); + CDEBUG(D_RPCTRACE, "Sleeping %llds 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, - (atomic_read(&imp->imp_inflight) == 0), - &lwi); - if (rc) { + * have been locally cancelled by ptlrpc_abort_inflight. + */ + timeout_jiffies = max_t(long, cfs_time_seconds(timeout), 1); + rc = wait_event_idle_timeout( + imp->imp_recovery_waitq, + (atomic_read(&imp->imp_inflight) == 0), + timeout_jiffies); + + if (rc == 0) { 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)); + CERROR("%s: timeout waiting for callback (%d != 0)\n", + cli_tgt, atomic_read(&imp->imp_inflight)); spin_lock(&imp->imp_lock); if (atomic_read(&imp->imp_inflight) == 0) { @@ -353,34 +370,27 @@ void ptlrpc_invalidate_import(struct obd_import *imp) /* Let's save one loop as soon as inflight have * dropped to zero. No new inflights possible at * this point. */ - rc = 0; + rc = 1; } else { - list_for_each_safe(tmp, n, - &imp->imp_sending_list) { - req = list_entry(tmp, - struct ptlrpc_request, - rq_list); + list_for_each_entry(req, &imp->imp_sending_list, + 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); + list_for_each_entry(req, &imp->imp_delayed_list, + rq_list) { DEBUG_REQ(D_ERROR, req, "still on delayed list"); } - CERROR("%s: RPCs in \"%s\" phase found (%d). " - "Network is sluggish? Waiting them " + CERROR("%s: Unregistering RPCs found (%d). " + "Network is sluggish? Waiting for 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); + } while (rc == 0); /* * Let's additionally check that no new rpcs added to import in @@ -391,22 +401,28 @@ void ptlrpc_invalidate_import(struct obd_import *imp) sptlrpc_import_flush_all_ctx(imp); atomic_dec(&imp->imp_inval_count); - wake_up_all(&imp->imp_recovery_waitq); + wake_up(&imp->imp_recovery_waitq); } EXPORT_SYMBOL(ptlrpc_invalidate_import); /* unset imp_invalid */ -void ptlrpc_activate_import(struct obd_import *imp) +void ptlrpc_activate_import(struct obd_import *imp, bool set_state_full) { struct obd_device *obd = imp->imp_obd; spin_lock(&imp->imp_lock); if (imp->imp_deactive != 0) { + LASSERT(imp->imp_state != LUSTRE_IMP_FULL); + if (imp->imp_state != LUSTRE_IMP_DISCON) + import_set_state_nolock(imp, LUSTRE_IMP_DISCON); spin_unlock(&imp->imp_lock); return; } + if (set_state_full) + import_set_state_nolock(imp, LUSTRE_IMP_FULL); imp->imp_invalid = 0; + spin_unlock(&imp->imp_lock); obd_import_event(obd, imp, IMP_EVENT_ACTIVE); } @@ -428,57 +444,50 @@ EXPORT_SYMBOL(ptlrpc_pinger_force); void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt) { - ENTRY; - - LASSERT(!imp->imp_dlm_fake); + ENTRY; - if (ptlrpc_set_import_discon(imp, conn_cnt)) { - if (!imp->imp_replayable) { - CDEBUG(D_HA, "import %s@%s for %s not replayable, " - "auto-deactivating\n", - obd2cli_tgt(imp->imp_obd), - imp->imp_connection->c_remote_uuid.uuid, - imp->imp_obd->obd_name); - ptlrpc_deactivate_import(imp); - } + LASSERT(!imp->imp_dlm_fake); + if (ptlrpc_set_import_discon(imp, conn_cnt, true)) ptlrpc_pinger_force(imp); - } + EXIT; } int ptlrpc_reconnect_import(struct obd_import *imp) { -#ifdef ENABLE_PINGER - struct l_wait_info lwi; - int secs = cfs_time_seconds(obd_timeout); +#ifdef CONFIG_LUSTRE_FS_PINGER + long timeout_jiffies = 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); + obd2cli_tgt(imp->imp_obd), obd_timeout); - lwi = LWI_TIMEOUT(secs, NULL, NULL); - rc = l_wait_event(imp->imp_recovery_waitq, - !ptlrpc_import_in_recovery(imp), &lwi); + rc = wait_event_idle_timeout(imp->imp_recovery_waitq, + !ptlrpc_import_in_recovery(imp), + timeout_jiffies); + if (rc == 0) + rc = -ETIMEDOUT; + else + rc = 0; 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); + ptlrpc_set_import_discon(imp, 0, false); /* 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); + rc = l_wait_event_abortable( + imp->imp_recovery_waitq, + (atomic_read(&imp->imp_inval_count) == 0)); if (rc) CERROR("Interrupted, inval=%d\n", atomic_read(&imp->imp_inval_count)); @@ -487,7 +496,7 @@ int ptlrpc_reconnect_import(struct obd_import *imp) /* Allow reconnect attempts */ imp->imp_obd->obd_no_recov = 0; /* Remove 'invalid' flag */ - ptlrpc_activate_import(imp); + ptlrpc_activate_import(imp, false); /* Attempt a new connect */ ptlrpc_recover_import(imp, NULL, 0); return 0; @@ -502,108 +511,112 @@ EXPORT_SYMBOL(ptlrpc_reconnect_import); */ static int import_select_connection(struct obd_import *imp) { - struct obd_import_conn *imp_conn = NULL, *conn; - struct obd_export *dlmexp; - char *target_start; - int target_len, tried_all = 1; - ENTRY; + struct obd_import_conn *imp_conn = NULL, *conn; + struct obd_export *dlmexp; + char *target_start; + int target_len, tried_all = 1; + int rc = 0; + ENTRY; spin_lock(&imp->imp_lock); 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); + rc = -EINVAL; + CERROR("%s: no connections available: rc = %d\n", + imp->imp_obd->obd_name, rc); + GOTO(out_unlock, rc); } 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), - conn->oic_last_attempt); - - /* If we have not tried this connection since - the last successful attempt, go with this one */ - if ((conn->oic_last_attempt == 0) || - cfs_time_beforeq_64(conn->oic_last_attempt, - imp->imp_last_success_conn)) { - imp_conn = conn; - tried_all = 0; - break; - } + CDEBUG(D_HA, "%s: connect to NID %s last attempt %lld\n", + imp->imp_obd->obd_name, + libcfs_nid2str(conn->oic_conn->c_peer.nid), + conn->oic_last_attempt); + + /* If we have not tried this connection since + * the last successful attempt, go with this one + */ + if ((conn->oic_last_attempt == 0) || + conn->oic_last_attempt <= imp->imp_last_success_conn) { + imp_conn = conn; + tried_all = 0; + break; + } - /* If all of the connections have already been tried - since the last successful connection; just choose the - least recently used */ - if (!imp_conn) - imp_conn = conn; - else if (cfs_time_before_64(conn->oic_last_attempt, - imp_conn->oic_last_attempt)) - imp_conn = conn; - } + /* If all of the connections have already been tried + * since the last successful connection; just choose the + * least recently used + */ + if (!imp_conn) + imp_conn = conn; + else if (imp_conn->oic_last_attempt > conn->oic_last_attempt) + imp_conn = conn; + } - /* if not found, simply choose the current one */ - if (!imp_conn || imp->imp_force_reconnect) { - LASSERT(imp->imp_conn_current); - imp_conn = imp->imp_conn_current; - tried_all = 0; - } - LASSERT(imp_conn->oic_conn); - - /* If we've tried everything, and we're back to the beginning of the - list, increase our timeout and try again. It will be reset when - 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)) { + /* if not found, simply choose the current one */ + if (!imp_conn || imp->imp_force_reconnect) { + LASSERT(imp->imp_conn_current); + imp_conn = imp->imp_conn_current; + tried_all = 0; + } + LASSERT(imp_conn->oic_conn); + + /* If we've tried everything, and we're back to the beginning of the + * list, increase our timeout and try again. It will be reset when + * 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)) { 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)); + 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(); - - /* switch connection, don't mind if it's same as the current one */ - if (imp->imp_connection) - ptlrpc_connection_put(imp->imp_connection); - imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn); - - dlmexp = class_conn2export(&imp->imp_dlm_handle); - LASSERT(dlmexp != NULL); - if (dlmexp->exp_connection) - ptlrpc_connection_put(dlmexp->exp_connection); - dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn); - class_export_put(dlmexp); - - if (imp->imp_conn_current != imp_conn) { - 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_conn->oic_last_attempt = ktime_get_seconds(); + + /* switch connection, don't mind if it's same as the current one */ + ptlrpc_connection_put(imp->imp_connection); + imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn); + + dlmexp = class_conn2export(&imp->imp_dlm_handle); + if (!dlmexp) + GOTO(out_unlock, rc = -EINVAL); + ptlrpc_connection_put(dlmexp->exp_connection); + dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn); + class_export_put(dlmexp); + + if (imp->imp_conn_current != imp_conn) { + 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; - } + imp->imp_conn_current = imp_conn; + } - CDEBUG(D_HA, "%s: import %p using connection %s/%s\n", - imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid, - libcfs_nid2str(imp_conn->oic_conn->c_peer.nid)); + /* The below message is checked in conf-sanity.sh test_35[ab] */ + CDEBUG(D_HA, "%s: import %p using connection %s/%s\n", + imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid, + libcfs_nid2str(imp_conn->oic_conn->c_peer.nid)); +out_unlock: spin_unlock(&imp->imp_lock); - - RETURN(0); + RETURN(rc); } /* @@ -611,24 +624,24 @@ static int import_select_connection(struct obd_import *imp) */ static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno) { - struct ptlrpc_request *req; - struct list_head *tmp; + struct ptlrpc_request *req; /* 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); + req = list_first_entry(&imp->imp_committed_list, + 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"); + 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); + req = list_first_entry(&imp->imp_committed_list, + 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"); @@ -639,29 +652,41 @@ static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno) return 0; } +int ptlrpc_connect_import(struct obd_import *imp) +{ + spin_lock(&imp->imp_lock); + return ptlrpc_connect_import_locked(imp); +} + /** * Attempt to (re)connect import \a imp. This includes all preparations, * initializing CONNECT RPC request and passing it to ptlrpcd for * actual sending. + * + * Assumes imp->imp_lock is held, and releases it. + * * Returns 0 on success or error code. */ -int ptlrpc_connect_import(struct obd_import *imp) +int ptlrpc_connect_import_locked(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; + struct obd_device *obd = imp->imp_obd; + int initial_connect = 0; + int set_transno = 0; + __u64 committed_before_reconnect = 0; + struct ptlrpc_request *request; + struct obd_connect_data ocd; + char *bufs[] = { NULL, + obd2cli_tgt(imp->imp_obd), + obd->obd_uuid.uuid, + (char *)&imp->imp_dlm_handle, + (char *)&ocd, + NULL }; + struct ptlrpc_connect_async_args *aa; + int rc; + ENTRY; + + assert_spin_locked(&imp->imp_lock); - 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"); @@ -670,97 +695,113 @@ int ptlrpc_connect_import(struct obd_import *imp) spin_unlock(&imp->imp_lock); CERROR("already connected\n"); RETURN(0); - } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) { + } else if (imp->imp_state == LUSTRE_IMP_CONNECTING || + imp->imp_state == LUSTRE_IMP_EVICTED || + imp->imp_connected) { spin_unlock(&imp->imp_lock); CERROR("already connecting\n"); RETURN(-EALREADY); } - IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING); + import_set_state_nolock(imp, LUSTRE_IMP_CONNECTING); - imp->imp_conn_cnt++; - imp->imp_resend_replay = 0; + 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; + 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); + 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 = import_select_connection(imp); + if (rc) + GOTO(out, rc); rc = sptlrpc_import_sec_adapt(imp, NULL, NULL); - if (rc) - GOTO(out, rc); + 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; + /* Reset connect flags to the originally requested flags, in case + * the server is updated on-the-fly we will get the new features. */ + ocd = imp->imp_connect_data; + ocd.ocd_connect_flags = imp->imp_connect_flags_orig; + ocd.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); - } + ocd.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, &ocd, NULL); + if (rc) + GOTO(out, rc); + + request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT); + if (request == NULL) + GOTO(out, rc = -ENOMEM); + + /* get SELinux policy info if any */ + rc = sptlrpc_get_sepol(request); + if (rc < 0) { + ptlrpc_request_free(request); + GOTO(out, rc); + } + + bufs[5] = request->rq_sepol; + + req_capsule_set_size(&request->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT, + strlen(request->rq_sepol) ? + strlen(request->rq_sepol) + 1 : 0); + + 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_timeout(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; + + aa = ptlrpc_req_async_args(aa, request); + memset(aa, 0, sizeof *aa); + + aa->pcaa_peer_committed = committed_before_reconnect; + aa->pcaa_initial_connect = initial_connect; - /* 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) { + 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); - } + 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); + 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); @@ -768,7 +809,7 @@ int ptlrpc_connect_import(struct obd_import *imp) rc = 0; out: if (rc != 0) - IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON); + import_set_state(imp, LUSTRE_IMP_DISCON); RETURN(rc); } @@ -792,35 +833,21 @@ static int ptlrpc_busy_reconnect(int rc) } 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) + 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); + list_move(&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 || @@ -831,7 +858,7 @@ static int ptlrpc_connect_set_flags(struct obd_import *imp, const char *older = "older than client. " "Consider upgrading server"; const char *newer = "newer than client. " - "Consider recompiling application"; + "Consider upgrading client"; LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) " "is much %s (%s)\n", @@ -845,40 +872,19 @@ static int ptlrpc_connect_set_flags(struct obd_import *imp, 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; + obd_cksum_types_supported_client()) == 0) { + LCONSOLE_ERROR("The negotiation of the checksum " + "alogrithm to use with server %s " + "failed (%x/%x)\n", + obd2cli_tgt(imp->imp_obd), + ocd->ocd_cksum_types, + obd_cksum_types_supported_client()); + return -EPROTO; } else { cli->cl_supp_cksum_types = ocd->ocd_cksum_types; } @@ -887,11 +893,13 @@ static int ptlrpc_connect_set_flags(struct obd_import *imp, * 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); + cli->cl_cksum_type = obd_cksum_type_select(imp->imp_obd->obd_name, + cli->cl_supp_cksum_types, + cli->cl_preferred_cksum_type); if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) cli->cl_max_pages_per_rpc = - min(ocd->ocd_brw_size >> PAGE_CACHE_SHIFT, + min(ocd->ocd_brw_size >> PAGE_SHIFT, cli->cl_max_pages_per_rpc); else if (imp->imp_connect_op == MDS_CONNECT || imp->imp_connect_op == MGS_CONNECT) @@ -916,13 +924,17 @@ static int ptlrpc_connect_set_flags(struct obd_import *imp, * this leads to losing user settings done before such as * disable lru_resize, etc. */ if (old_connect_flags != exp_connect_flags(exp) || init_connect) { + struct ldlm_namespace *ns = imp->imp_obd->obd_namespace; + __u64 changed_flags; + + changed_flags = + ns->ns_connect_flags ^ ns->ns_orig_connect_flags; CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server " - "flags: "LPX64"\n", imp->imp_obd->obd_name, + "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; + ns->ns_connect_flags = (ns->ns_connect_flags & changed_flags) | + (ocd->ocd_connect_flags & ~changed_flags); + ns->ns_orig_connect_flags = ocd->ocd_connect_flags; } if (ocd->ocd_connect_flags & OBD_CONNECT_AT) @@ -981,16 +993,17 @@ static void ptlrpc_prepare_replay(struct obd_import *imp) * 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_request *request, + void *data, int rc) { - struct ptlrpc_connect_async_args *aa = data; - struct obd_import *imp = request->rq_import; - struct lustre_handle old_hdl; - __u64 old_connect_flags; - int msg_flags; + struct ptlrpc_connect_async_args *aa = data; + struct obd_import *imp = request->rq_import; + struct lustre_handle old_hdl; + __u64 old_connect_flags; + timeout_t service_timeout; + int msg_flags; struct obd_connect_data *ocd; - struct obd_export *exp; + struct obd_export *exp = NULL; int ret; ENTRY; @@ -1001,19 +1014,39 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, RETURN(0); } + imp->imp_connect_error = rc; if (rc) { + struct ptlrpc_request *free_req; + struct ptlrpc_request *tmp; + + /* abort all delayed requests initiated connection */ + list_for_each_entry_safe(free_req, tmp, &imp->imp_delayed_list, + rq_list) { + spin_lock(&free_req->rq_lock); + if (free_req->rq_no_resend) { + free_req->rq_err = 1; + free_req->rq_status = -EIO; + ptlrpc_client_wake_req(free_req); + } + spin_unlock(&free_req->rq_lock); + } + /* 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); + LASSERT(imp->imp_conn_current); - msg_flags = lustre_msg_get_op_flags(request->rq_repmsg); + msg_flags = lustre_msg_get_op_flags(request->rq_repmsg); ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER); @@ -1030,10 +1063,10 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, spin_lock(&imp->imp_lock); - /* All imports are pingable */ - imp->imp_pingable = 1; - imp->imp_force_reconnect = 0; - imp->imp_force_verify = 0; + /* All imports are pingable */ + imp->imp_pingable = 1; + imp->imp_force_reconnect = 0; + imp->imp_force_verify = 0; imp->imp_connect_data = *ocd; @@ -1043,16 +1076,33 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, 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 granted asked subset of flags: " - "asked="LPX64" grranted="LPX64"\n", - imp->imp_obd->obd_name,imp->imp_connect_flags_orig, + 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) && @@ -1079,27 +1129,23 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, } } - 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); - } 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. */ + * because may reflect other routing, etc. + */ + service_timeout = lustre_msg_get_service_timeout(request->rq_repmsg); at_reinit(&imp->imp_at.iat_net_latency, 0, 0); - ptlrpc_at_adj_net_latency(request, - lustre_msg_get_service_time(request->rq_repmsg)); + ptlrpc_at_adj_net_latency(request, service_timeout); /* 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); + aa->pcaa_initial_connect); class_export_put(exp); + exp = NULL; + if (rc != 0) GOTO(out, rc); @@ -1109,212 +1155,283 @@ static int ptlrpc_connect_interpret(const struct lu_env *env, 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 */ - - imp->imp_remote_handle = - *lustre_msg_get_handle(request->rq_repmsg); - - /* Initial connects are allowed for clients with non-random - * uuids when servers are in recovery. Simply signal the - * servers replay is complete and wait in REPLAY_WAIT. */ - if (msg_flags & MSG_CONNECT_RECOVERING) { - CDEBUG(D_HA, "connect to %s during recovery\n", - 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); - } + /* if applies, adjust the imp->imp_msg_magic here + * according to reply flags + */ - GOTO(finish, rc = 0); - } + imp->imp_remote_handle = + *lustre_msg_get_handle(request->rq_repmsg); - /* Determine what recovery state to move the import to. */ - if (MSG_CONNECT_RECONNECT & msg_flags) { - memset(&old_hdl, 0, sizeof(old_hdl)); - if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg), - sizeof (old_hdl))) { - LCONSOLE_WARN("Reconnect to %s (at @%s) failed due " - "bad handle "LPX64"\n", - obd2cli_tgt(imp->imp_obd), - imp->imp_connection->c_remote_uuid.uuid, - imp->imp_dlm_handle.cookie); - GOTO(out, rc = -ENOTCONN); - } + /* Initial connects are allowed for clients with non-random + * uuids when servers are in recovery. Simply signal the + * servers replay is complete and wait in REPLAY_WAIT. + */ + if (msg_flags & MSG_CONNECT_RECOVERING) { + CDEBUG(D_HA, "connect to %s during recovery\n", + obd2cli_tgt(imp->imp_obd)); + import_set_state_nolock(imp, LUSTRE_IMP_REPLAY_LOCKS); + spin_unlock(&imp->imp_lock); + } else { + spin_unlock(&imp->imp_lock); + ptlrpc_activate_import(imp, true); + } - if (memcmp(&imp->imp_remote_handle, - lustre_msg_get_handle(request->rq_repmsg), - sizeof(imp->imp_remote_handle))) { - int level = msg_flags & MSG_CONNECT_RECOVERING ? - D_HA : D_WARNING; - - /* Bug 16611/14775: if server handle have changed, - * that means some sort of disconnection happened. - * If the server is not in recovery, that also means it - * already erased all of our state because of previous - * eviction. If it is in recovery - we are safe to - * participate since we can reestablish all of our state - * with server again */ - if ((MSG_CONNECT_RECOVERING & msg_flags)) { - CDEBUG(level,"%s@%s changed server handle from " - LPX64" to "LPX64 - " 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 " - LPX64" to "LPX64"\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); - } + GOTO(finish, rc = 0); + } + /* Determine what recovery state to move the import to. */ + if (MSG_CONNECT_RECONNECT & msg_flags) { + memset(&old_hdl, 0, sizeof(old_hdl)); + if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg), + sizeof(old_hdl))) { + 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); + } - imp->imp_remote_handle = - *lustre_msg_get_handle(request->rq_repmsg); + if (memcmp(&imp->imp_remote_handle, + lustre_msg_get_handle(request->rq_repmsg), + sizeof(imp->imp_remote_handle))) { + int level = msg_flags & MSG_CONNECT_RECOVERING ? + D_HA : D_WARNING; + + /* Bug 16611/14775: if server handle have changed, + * that means some sort of disconnection happened. + * If the server is not in recovery, that also means it + * already erased all of our state because of previous + * eviction. If it is in recovery - we are safe to + * participate since we can reestablish all of our state + * with server again + */ + if ((MSG_CONNECT_RECOVERING & msg_flags)) { + CDEBUG_LIMIT(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); + } - if (!(MSG_CONNECT_RECOVERING & msg_flags)) { - IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED); - GOTO(finish, rc = 0); - } + imp->imp_remote_handle = + *lustre_msg_get_handle(request->rq_repmsg); - } else { - CDEBUG(D_HA, "reconnected to %s@%s after partition\n", - obd2cli_tgt(imp->imp_obd), - imp->imp_connection->c_remote_uuid.uuid); - } + if (!(MSG_CONNECT_RECOVERING & msg_flags)) { + import_set_state(imp, LUSTRE_IMP_EVICTED); + GOTO(finish, rc = 0); + } + } else { + CDEBUG(D_HA, "reconnected to %s@%s after partition\n", + obd2cli_tgt(imp->imp_obd), + imp->imp_connection->c_remote_uuid.uuid); + } - if (imp->imp_invalid) { - CDEBUG(D_HA, "%s: reconnected but import is invalid; " - "marking evicted\n", imp->imp_obd->obd_name); - IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED); - } else if (MSG_CONNECT_RECOVERING & msg_flags) { - CDEBUG(D_HA, "%s: reconnected to %s during replay\n", - imp->imp_obd->obd_name, - obd2cli_tgt(imp->imp_obd)); + if (imp->imp_invalid) { + CDEBUG(D_HA, "%s: reconnected but import is invalid; " + "marking evicted\n", imp->imp_obd->obd_name); + import_set_state(imp, LUSTRE_IMP_EVICTED); + } else if (MSG_CONNECT_RECOVERING & msg_flags) { + CDEBUG(D_HA, "%s: reconnected to %s during replay\n", + imp->imp_obd->obd_name, + obd2cli_tgt(imp->imp_obd)); spin_lock(&imp->imp_lock); imp->imp_resend_replay = 1; spin_unlock(&imp->imp_lock); - IMPORT_SET_STATE(imp, imp->imp_replay_state); - } else { - IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER); - } - } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) { - LASSERT(imp->imp_replayable); - imp->imp_remote_handle = - *lustre_msg_get_handle(request->rq_repmsg); - imp->imp_last_replay_transno = 0; - IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY); - } else { - DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags" - " not set: %x)", imp->imp_obd->obd_name, msg_flags); - imp->imp_remote_handle = - *lustre_msg_get_handle(request->rq_repmsg); - IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED); - } + import_set_state(imp, imp->imp_replay_state); + } else { + import_set_state(imp, LUSTRE_IMP_RECOVER); + } + } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) { + LASSERT(imp->imp_replayable); + 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 if ((ocd->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0 && + !imp->imp_invalid) { + + obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE); + /* The below message is checked in recovery-small.sh test_106 */ + DEBUG_REQ(D_HA, request, "%s: lwp recover", + imp->imp_obd->obd_name); + imp->imp_remote_handle = + *lustre_msg_get_handle(request->rq_repmsg); + import_set_state(imp, LUSTRE_IMP_RECOVER); + } else { + DEBUG_REQ(D_HA, request, + "%s: evicting (reconnect/recover flags not set: %x)", + imp->imp_obd->obd_name, msg_flags); + imp->imp_remote_handle = + *lustre_msg_get_handle(request->rq_repmsg); + import_set_state(imp, LUSTRE_IMP_EVICTED); + } - /* Sanity checks for a reconnected import. */ - if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) { - CERROR("imp_replayable flag does not match server " - "after reconnect. We should LBUG right here.\n"); - } + /* Sanity checks for a reconnected import. */ + if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) + CERROR("imp_replayable flag does not match server after reconnect. We should LBUG right here.\n"); + + if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 && + lustre_msg_get_last_committed(request->rq_repmsg) < + aa->pcaa_peer_committed) { + static bool printed; - 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 - ")! See https://bugzilla.lustre.org/show_bug.cgi?" - "id=9646\n", - obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed, + /* The below message is checked in recovery-small.sh test_54 */ + CERROR("%s: went back in time (transno %lld was previously committed, server now claims %lld)!\n", + obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed, lustre_msg_get_last_committed(request->rq_repmsg)); + if (!printed) { + CERROR("For further information, see http://doc.lustre.org/lustre_manual.xhtml#went_back_in_time\n"); + printed = true; + } } finish: 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", + 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: + if (exp != NULL) + class_export_put(exp); + + spin_lock(&imp->imp_lock); + imp->imp_connected = 0; imp->imp_connect_tried = 1; - if (rc != 0) { - IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON); - if (rc == -EACCES) { - /* - * Give up trying to reconnect - * EACCES means client has no permission for connection - */ - imp->imp_obd->obd_no_recov = 1; - ptlrpc_deactivate_import(imp); - } + if (rc != 0) { + bool inact = false; + time64_t now = ktime_get_seconds(); + time64_t next_connect; + + import_set_state_nolock(imp, LUSTRE_IMP_DISCON); + if (rc == -EACCES) { + /* + * Give up trying to reconnect + * EACCES means client has no permission for connection + */ + imp->imp_obd->obd_no_recov = 1; + ptlrpc_deactivate_import_nolock(imp); + inact = true; + } else if (rc == -EPROTO) { + struct obd_connect_data *ocd; + + /* reply message might not be ready */ + if (request->rq_repmsg == NULL) { + spin_unlock(&imp->imp_lock); + RETURN(-EPROTO); + } - if (rc == -EPROTO) { - struct obd_connect_data *ocd; - - /* reply message might not be ready */ - if (request->rq_repmsg == NULL) - RETURN(-EPROTO); - - ocd = req_capsule_server_get(&request->rq_pill, - &RMF_CONNECT_DATA); - if (ocd && - (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) && - (ocd->ocd_version != LUSTRE_VERSION_CODE)) { - /* Actually servers are only supposed to refuse - connection from liblustre clients, so we should - never see this from VFS context */ - LCONSOLE_ERROR_MSG(0x16a, "Server %s version " - "(%d.%d.%d.%d)" - " refused connection from this client " - "with an incompatible version (%s). " - "Client must be recompiled\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), - LUSTRE_VERSION_STRING); - ptlrpc_deactivate_import(imp); - IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED); - } - RETURN(-EPROTO); - } + ocd = req_capsule_server_get(&request->rq_pill, + &RMF_CONNECT_DATA); + /* Servers are not supposed to refuse connections from + * clients based on version, only connection feature + * flags. We should never see this from llite, but it + * may be useful for debugging in the future. */ + if (ocd && + (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) && + (ocd->ocd_version != LUSTRE_VERSION_CODE)) { + LCONSOLE_ERROR_MSG(0x16a, "Server %s version " + "(%d.%d.%d.%d)" + " refused connection from this client " + "with an incompatible version (%s). " + "Client must be recompiled\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), + LUSTRE_VERSION_STRING); + ptlrpc_deactivate_import_nolock(imp); + import_set_state_nolock(imp, LUSTRE_IMP_CLOSED); + inact = true; + } + } else if (rc == -ENODEV || rc == -ETIMEDOUT) { + /* ENODEV means there is no service, force reconnection + * to a pair if attempt happen ptlrpc_next_reconnect + * before now. ETIMEDOUT could be set during network + * error and do not guarantee request deadline happened. + */ + struct obd_import_conn *conn; + time64_t reconnect_time; + + /* Same as ptlrpc_next_reconnect, but in past */ + reconnect_time = now - INITIAL_CONNECT_TIMEOUT; + list_for_each_entry(conn, &imp->imp_conn_list, + oic_item) { + if (conn->oic_last_attempt <= reconnect_time) { + imp->imp_force_verify = 1; + break; + } + } + } + + next_connect = imp->imp_conn_current->oic_last_attempt + + (request->rq_deadline - request->rq_sent); + spin_unlock(&imp->imp_lock); + + if (inact) + obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE); + + if (rc == -EPROTO) + RETURN(rc); + + /* adjust imp_next_ping to request deadline + 1 and reschedule + * a pinger if import lost processing during CONNECTING or far + * away from request deadline. It could happen when connection + * was initiated outside of pinger, like + * ptlrpc_set_import_discon(). + */ + if (!imp->imp_force_verify && (imp->imp_next_ping <= now || + imp->imp_next_ping > next_connect)) { + imp->imp_next_ping = max(now, next_connect) + 1; + ptlrpc_pinger_wake_up(); + } 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); + } else { + spin_unlock(&imp->imp_lock); } - wake_up_all(&imp->imp_recovery_waitq); + wake_up(&imp->imp_recovery_waitq); RETURN(rc); } @@ -1323,13 +1440,12 @@ out: * \see signal_completed_replay */ static int completed_replay_interpret(const struct lu_env *env, - struct ptlrpc_request *req, - void * data, int rc) + struct ptlrpc_request *req, + void *args, int rc) { ENTRY; atomic_dec(&req->rq_import->imp_replay_inflight); - if (req->rq_status == 0 && - !req->rq_import->imp_vbr_failed) { + 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) { @@ -1360,8 +1476,8 @@ static int signal_completed_replay(struct obd_import *imp) 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); + if (!atomic_add_unless(&imp->imp_replay_inflight, 1, 1)) + RETURN(0); req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION, OBD_PING); @@ -1392,9 +1508,6 @@ static int ptlrpc_invalidate_import_thread(void *data) struct obd_import *imp = data; ENTRY; - - unshare_fs_struct(); - CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n", imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd), imp->imp_connection->c_remote_uuid.uuid); @@ -1406,7 +1519,7 @@ static int ptlrpc_invalidate_import_thread(void *data) libcfs_debug_dumplog(); } - IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER); + import_set_state(imp, LUSTRE_IMP_RECOVER); ptlrpc_import_recovery_state_machine(imp); class_import_put(imp); @@ -1442,6 +1555,8 @@ int ptlrpc_import_recovery_state_machine(struct obd_import *imp) ENTRY; if (imp->imp_state == LUSTRE_IMP_EVICTED) { + struct task_struct *task; + deuuidify(obd2cli_tgt(imp->imp_obd), NULL, &target_start, &target_len); /* Don't care about MGC eviction */ @@ -1452,6 +1567,7 @@ int ptlrpc_import_recovery_state_machine(struct obd_import *imp) "using this service will fail.\n", imp->imp_obd->obd_name, target_len, target_start); + LASSERTF(!obd_lbug_on_eviction, "LBUG upon eviction\n"); } CDEBUG(D_HA, "evicted from %s@%s; invalidating\n", obd2cli_tgt(imp->imp_obd), @@ -1461,24 +1577,22 @@ int ptlrpc_import_recovery_state_machine(struct obd_import *imp) 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"); + "ll_imp_inval"); if (IS_ERR(task)) { class_import_put(imp); - CERROR("error starting invalidate thread: %d\n", rc); rc = PTR_ERR(task); + CERROR("%s: can't start invalidate thread: rc = %d\n", + imp->imp_obd->obd_name, rc); } else { rc = 0; } RETURN(rc); - } } if (imp->imp_state == LUSTRE_IMP_REPLAY) { @@ -1487,7 +1601,7 @@ int ptlrpc_import_recovery_state_machine(struct obd_import *imp) rc = ptlrpc_replay_next(imp, &inflight); if (inflight == 0 && atomic_read(&imp->imp_replay_inflight) == 0) { - IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS); + import_set_state(imp, LUSTRE_IMP_REPLAY_LOCKS); rc = ldlm_replay_locks(imp); if (rc) GOTO(out, rc); @@ -1497,7 +1611,7 @@ int ptlrpc_import_recovery_state_machine(struct obd_import *imp) if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) { if (atomic_read(&imp->imp_replay_inflight) == 0) { - IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT); + import_set_state(imp, LUSTRE_IMP_REPLAY_WAIT); rc = signal_completed_replay(imp); if (rc) GOTO(out, rc); @@ -1506,27 +1620,35 @@ int ptlrpc_import_recovery_state_machine(struct obd_import *imp) if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) { if (atomic_read(&imp->imp_replay_inflight) == 0) { - IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER); + import_set_state(imp, LUSTRE_IMP_RECOVER); } } - if (imp->imp_state == LUSTRE_IMP_RECOVER) { + if (imp->imp_state == LUSTRE_IMP_RECOVER) { struct ptlrpc_connection *conn = imp->imp_connection; - rc = ptlrpc_resend(imp); - if (rc) - GOTO(out, rc); - IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL); - ptlrpc_activate_import(imp); - - 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)); - } + rc = ptlrpc_resend(imp); + if (rc) + GOTO(out, rc); + ptlrpc_activate_import(imp, true); + + /* Reverse import are flagged with dlm_fake == 1. + * They do not do recovery and connection are not "restored". + */ + if (!imp->imp_dlm_fake) + CDEBUG_LIMIT(imp->imp_was_idle ? + imp->imp_idle_debug : D_CONSOLE, + "%s: Connection restored to %s (at %s)\n", + imp->imp_obd->obd_name, + obd_uuid2str(&conn->c_remote_uuid), + obd_import_nid2str(imp)); + spin_lock(&imp->imp_lock); + imp->imp_was_idle = 0; + spin_unlock(&imp->imp_lock); + } if (imp->imp_state == LUSTRE_IMP_FULL) { - wake_up_all(&imp->imp_recovery_waitq); + wake_up(&imp->imp_recovery_waitq); ptlrpc_wake_delayed(imp); } @@ -1534,15 +1656,12 @@ out: RETURN(rc); } -int ptlrpc_disconnect_import(struct obd_import *imp, int noclose) +static struct ptlrpc_request *ptlrpc_disconnect_prep_req(struct obd_import *imp) { 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; @@ -1559,64 +1678,91 @@ int ptlrpc_disconnect_import(struct obd_import *imp, int noclose) "(connect_op %d): rc = %d\n", imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd), imp->imp_connect_op, rc); - RETURN(rc); + RETURN(ERR_PTR(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); - else - timeout = cfs_time_seconds(obd_timeout); - } else { - int idx = import_at_get_index(imp, - imp->imp_client->cli_request_portal); - timeout = cfs_time_seconds( - at_get(&imp->imp_at.iat_service_estimate[idx])); - } + req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT, + LUSTRE_OBD_VERSION, rq_opc); + if (req == NULL) + RETURN(ERR_PTR(-ENOMEM)); - lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout), - back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL); - rc = l_wait_event(imp->imp_recovery_waitq, - !ptlrpc_import_in_recovery(imp), &lwi); + /* We are disconnecting, do not retry a failed DISCONNECT rpc if + * it fails. We can get through the above with a down server + * if the client doesn't know the server is gone yet. */ + req->rq_no_resend = 1; - } + /* We want client umounts to happen quickly, no matter the + server state... */ + req->rq_timeout = min_t(timeout_t, req->rq_timeout, + INITIAL_CONNECT_TIMEOUT); + + import_set_state(imp, LUSTRE_IMP_CONNECTING); + req->rq_send_state = LUSTRE_IMP_CONNECTING; + ptlrpc_request_set_replen(req); + + RETURN(req); +} + +int ptlrpc_disconnect_import(struct obd_import *imp, int noclose) +{ + struct ptlrpc_request *req; + int rc = 0; + ENTRY; + + if (imp->imp_obd->obd_force) + GOTO(set_state, rc); + + /* probably the import has been disconnected already being idle */ + spin_lock(&imp->imp_lock); + if (imp->imp_state == LUSTRE_IMP_IDLE) + GOTO(out, rc); + spin_unlock(&imp->imp_lock); + + if (ptlrpc_import_in_recovery(imp)) { + long timeout_jiffies; + time64_t timeout; + + if (AT_OFF) { + if (imp->imp_server_timeout) + timeout = obd_timeout >> 1; + else + timeout = obd_timeout; + } else { + u32 req_portal; + int idx; + + req_portal = imp->imp_client->cli_request_portal; + idx = import_at_get_index(imp, req_portal); + timeout = at_get(&imp->imp_at.iat_service_estimate[idx]); + } + + timeout_jiffies = cfs_time_seconds(timeout); + if (wait_event_idle_timeout(imp->imp_recovery_waitq, + !ptlrpc_import_in_recovery(imp), + timeout_jiffies) == 0 && + l_wait_event_abortable(imp->imp_recovery_waitq, + !ptlrpc_import_in_recovery(imp)) < 0) + rc = -EINTR; + } 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); - if (req) { - /* We are disconnecting, do not retry a failed DISCONNECT rpc if - * it fails. We can get through the above with a down server - * if the client doesn't know the server is gone yet. */ - req->rq_no_resend = 1; - - /* We want client umounts to happen quickly, no matter the - server state... */ - req->rq_timeout = min_t(int, req->rq_timeout, - INITIAL_CONNECT_TIMEOUT); - - IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING); - req->rq_send_state = LUSTRE_IMP_CONNECTING; - ptlrpc_request_set_replen(req); - rc = ptlrpc_queue_wait(req); - ptlrpc_req_finished(req); - } + req = ptlrpc_disconnect_prep_req(imp); + if (IS_ERR(req)) + GOTO(set_state, rc = PTR_ERR(req)); + rc = ptlrpc_queue_wait(req); + ptlrpc_req_finished(req); set_state: spin_lock(&imp->imp_lock); out: if (noclose) - IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON); + import_set_state_nolock(imp, LUSTRE_IMP_DISCON); else - IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED); + import_set_state_nolock(imp, LUSTRE_IMP_CLOSED); memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle)); spin_unlock(&imp->imp_lock); @@ -1626,103 +1772,210 @@ out: } EXPORT_SYMBOL(ptlrpc_disconnect_import); +static void ptlrpc_reset_reqs_generation(struct obd_import *imp) +{ + struct ptlrpc_request *old, *tmp; + + /* tag all resendable requests generated before disconnection + * notice this code is part of disconnect-at-idle path only */ + list_for_each_entry_safe(old, tmp, &imp->imp_delayed_list, + rq_list) { + spin_lock(&old->rq_lock); + if (old->rq_import_generation == imp->imp_generation - 1 && + ((imp->imp_initiated_at == imp->imp_generation) || + !old->rq_no_resend)) + old->rq_import_generation = imp->imp_generation; + spin_unlock(&old->rq_lock); + } +} + +static int ptlrpc_disconnect_idle_interpret(const struct lu_env *env, + struct ptlrpc_request *req, + void *args, int rc) +{ + struct obd_import *imp = req->rq_import; + int connect = 0; + + DEBUG_REQ(D_HA, req, "inflight=%d, refcount=%d: rc = %d", + atomic_read(&imp->imp_inflight), + refcount_read(&imp->imp_refcount), rc); + + spin_lock(&imp->imp_lock); + /* DISCONNECT reply can be late and another connection can just + * be initiated. so we have to abort disconnection. */ + if (req->rq_import_generation == imp->imp_generation && + imp->imp_state != LUSTRE_IMP_CLOSED) { + LASSERTF(imp->imp_state == LUSTRE_IMP_CONNECTING, + "%s\n", ptlrpc_import_state_name(imp->imp_state)); + memset(&imp->imp_remote_handle, 0, + sizeof(imp->imp_remote_handle)); + /* take our DISCONNECT into account */ + if (atomic_read(&imp->imp_reqs) > 1) { + imp->imp_generation++; + imp->imp_initiated_at = imp->imp_generation; + import_set_state_nolock(imp, LUSTRE_IMP_NEW); + ptlrpc_reset_reqs_generation(imp); + connect = 1; + } else { + /* do not expose transient IDLE state */ + import_set_state_nolock(imp, LUSTRE_IMP_IDLE); + } + } + + if (connect) { + rc = ptlrpc_connect_import_locked(imp); + if (rc >= 0) + ptlrpc_pinger_add_import(imp); + } else { + spin_unlock(&imp->imp_lock); + } + + return 0; +} + +int ptlrpc_disconnect_and_idle_import(struct obd_import *imp) +{ + struct ptlrpc_request *req; + ENTRY; + + if (imp->imp_obd->obd_force) + RETURN(0); + + if (ptlrpc_import_in_recovery(imp)) + RETURN(0); + + spin_lock(&imp->imp_lock); + if (imp->imp_state != LUSTRE_IMP_FULL) { + spin_unlock(&imp->imp_lock); + RETURN(0); + } + spin_unlock(&imp->imp_lock); + + req = ptlrpc_disconnect_prep_req(imp); + if (IS_ERR(req)) + RETURN(PTR_ERR(req)); + + CDEBUG_LIMIT(imp->imp_idle_debug, "%s: disconnect after %llus idle\n", + imp->imp_obd->obd_name, + ktime_get_real_seconds() - imp->imp_last_reply_time); + + /* don't make noise at reconnection */ + spin_lock(&imp->imp_lock); + imp->imp_was_idle = 1; + spin_unlock(&imp->imp_lock); + + req->rq_interpret_reply = ptlrpc_disconnect_idle_interpret; + ptlrpcd_add_req(req); + + RETURN(0); +} +EXPORT_SYMBOL(ptlrpc_disconnect_and_idle_import); + void ptlrpc_cleanup_imp(struct obd_import *imp) { ENTRY; spin_lock(&imp->imp_lock); - IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED); + + import_set_state_nolock(imp, LUSTRE_IMP_CLOSED); imp->imp_generation++; - spin_unlock(&imp->imp_lock); ptlrpc_abort_inflight(imp); + spin_unlock(&imp->imp_lock); + EXIT; } /* Adaptive Timeout utils */ -extern unsigned int at_min, at_max, at_history; -/* Update at_current with the specified value (bounded by at_min and at_max), - * as well as the AT history "bins". +/* Update at_current_timeout 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) +timeout_t at_measured(struct adaptive_timeout *at, timeout_t timeout) { - unsigned int old = at->at_current; - time_t now = cfs_time_current_sec(); - time_t binlimit = max_t(time_t, at_history / AT_BINS, 1); + timeout_t old_timeout = at->at_current_timeout; + time64_t now = ktime_get_real_seconds(); + long binlimit = max_t(long, at_history / AT_BINS, 1); LASSERT(at); - CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n", - val, at, now - at->at_binstart, at->at_current, + CDEBUG(D_OTHER, "add %u to %p time=%lld v=%u (%u %u %u %u)\n", + timeout, at, now - at->at_binstart, at->at_current_timeout, at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]); - if (val == 0) - /* 0's don't count, because we never want our timeout to - drop to 0, and because 0 could mean an error */ + if (timeout <= 0) + /* Negative timeouts and 0's don't count, because we never + * want our timeout to drop to 0 or below, and because 0 could + * mean an error + */ return 0; spin_lock(&at->at_lock); if (unlikely(at->at_binstart == 0)) { /* Special case to remove default from history */ - at->at_current = val; - at->at_worst_ever = val; - at->at_worst_time = now; - at->at_hist[0] = val; + at->at_current_timeout = timeout; + at->at_worst_timeout_ever = timeout; + at->at_worst_timestamp = now; + at->at_hist[0] = timeout; at->at_binstart = now; } else if (now - at->at_binstart < binlimit ) { /* in bin 0 */ - at->at_hist[0] = max(val, at->at_hist[0]); - at->at_current = max(val, at->at_current); + at->at_hist[0] = max_t(timeout_t, timeout, at->at_hist[0]); + at->at_current_timeout = max_t(timeout_t, timeout, + at->at_current_timeout); } else { int i, shift; - unsigned int maxv = val; - /* move bins over */ - shift = (now - at->at_binstart) / binlimit; + timeout_t maxv = timeout; + + /* move bins over */ + shift = (u32)(now - at->at_binstart) / binlimit; LASSERT(shift > 0); for(i = AT_BINS - 1; i >= 0; i--) { if (i >= shift) { at->at_hist[i] = at->at_hist[i - shift]; - maxv = max(maxv, at->at_hist[i]); + maxv = max_t(timeout_t, maxv, at->at_hist[i]); } else { at->at_hist[i] = 0; } } - at->at_hist[0] = val; - at->at_current = maxv; + at->at_hist[0] = timeout; + at->at_current_timeout = maxv; at->at_binstart += shift * binlimit; } - if (at->at_current > at->at_worst_ever) { - at->at_worst_ever = at->at_current; - at->at_worst_time = now; - } + if (at->at_current_timeout > at->at_worst_timeout_ever) { + at->at_worst_timeout_ever = at->at_current_timeout; + at->at_worst_timestamp = now; + } - if (at->at_flags & AT_FLG_NOHIST) + if (at->at_flags & AT_FLG_NOHIST) /* Only keep last reported val; keeping the rest of the history - for proc only */ - at->at_current = val; + * for debugfs only + */ + at->at_current_timeout = timeout; if (at_max > 0) - at->at_current = min(at->at_current, at_max); - at->at_current = max(at->at_current, at_min); - - if (at->at_current != old) - CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d " - "(val=%u) hist %u %u %u %u\n", at, - old, at->at_current, at->at_current - old, val, + at->at_current_timeout = min_t(timeout_t, + at->at_current_timeout, at_max); + at->at_current_timeout = max_t(timeout_t, at->at_current_timeout, + at_min); + if (at->at_current_timeout != old_timeout) + CDEBUG(D_OTHER, + "AT %p change: old=%u new=%u delta=%d (val=%d) hist %u %u %u %u\n", + at, old_timeout, at->at_current_timeout, + at->at_current_timeout - old_timeout, timeout, at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]); - /* if we changed, report the old value */ - old = (at->at_current != old) ? old : 0; + /* if we changed, report the old timeout value */ + old_timeout = (at->at_current_timeout != old_timeout) ? old_timeout : 0; spin_unlock(&at->at_lock); - return old; + return old_timeout; } /* Find the imp_at index for a given portal; assign if space available */