Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1 | /* -*- mode: c; c-basic-offset: 8; -*- |
| 2 | * vim: noexpandtab sw=8 ts=8 sts=0: |
| 3 | * |
| 4 | * dlmrecovery.c |
| 5 | * |
| 6 | * recovery stuff |
| 7 | * |
| 8 | * Copyright (C) 2004 Oracle. All rights reserved. |
| 9 | * |
| 10 | * This program is free software; you can redistribute it and/or |
| 11 | * modify it under the terms of the GNU General Public |
| 12 | * License as published by the Free Software Foundation; either |
| 13 | * version 2 of the License, or (at your option) any later version. |
| 14 | * |
| 15 | * This program is distributed in the hope that it will be useful, |
| 16 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 17 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 18 | * General Public License for more details. |
| 19 | * |
| 20 | * You should have received a copy of the GNU General Public |
| 21 | * License along with this program; if not, write to the |
| 22 | * Free Software Foundation, Inc., 59 Temple Place - Suite 330, |
| 23 | * Boston, MA 021110-1307, USA. |
| 24 | * |
| 25 | */ |
| 26 | |
| 27 | |
| 28 | #include <linux/module.h> |
| 29 | #include <linux/fs.h> |
| 30 | #include <linux/types.h> |
| 31 | #include <linux/slab.h> |
| 32 | #include <linux/highmem.h> |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 33 | #include <linux/init.h> |
| 34 | #include <linux/sysctl.h> |
| 35 | #include <linux/random.h> |
| 36 | #include <linux/blkdev.h> |
| 37 | #include <linux/socket.h> |
| 38 | #include <linux/inet.h> |
| 39 | #include <linux/timer.h> |
| 40 | #include <linux/kthread.h> |
Adrian Bunk | b4c7f53 | 2006-01-14 20:55:10 +0100 | [diff] [blame] | 41 | #include <linux/delay.h> |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 42 | |
| 43 | |
| 44 | #include "cluster/heartbeat.h" |
| 45 | #include "cluster/nodemanager.h" |
| 46 | #include "cluster/tcp.h" |
| 47 | |
| 48 | #include "dlmapi.h" |
| 49 | #include "dlmcommon.h" |
| 50 | #include "dlmdomain.h" |
| 51 | |
| 52 | #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY) |
| 53 | #include "cluster/masklog.h" |
| 54 | |
| 55 | static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node); |
| 56 | |
| 57 | static int dlm_recovery_thread(void *data); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 58 | static int dlm_do_recovery(struct dlm_ctxt *dlm); |
| 59 | |
| 60 | static int dlm_pick_recovery_master(struct dlm_ctxt *dlm); |
| 61 | static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node); |
| 62 | static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node); |
| 63 | static int dlm_request_all_locks(struct dlm_ctxt *dlm, |
| 64 | u8 request_from, u8 dead_node); |
| 65 | static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node); |
| 66 | |
| 67 | static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res); |
| 68 | static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres, |
| 69 | const char *lockname, int namelen, |
| 70 | int total_locks, u64 cookie, |
| 71 | u8 flags, u8 master); |
| 72 | static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm, |
| 73 | struct dlm_migratable_lockres *mres, |
| 74 | u8 send_to, |
| 75 | struct dlm_lock_resource *res, |
| 76 | int total_locks); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 77 | static int dlm_process_recovery_data(struct dlm_ctxt *dlm, |
| 78 | struct dlm_lock_resource *res, |
| 79 | struct dlm_migratable_lockres *mres); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 80 | static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm); |
| 81 | static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, |
| 82 | u8 dead_node, u8 send_to); |
| 83 | static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node); |
| 84 | static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm, |
| 85 | struct list_head *list, u8 dead_node); |
| 86 | static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm, |
| 87 | u8 dead_node, u8 new_master); |
| 88 | static void dlm_reco_ast(void *astdata); |
| 89 | static void dlm_reco_bast(void *astdata, int blocked_type); |
| 90 | static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st); |
| 91 | static void dlm_request_all_locks_worker(struct dlm_work_item *item, |
| 92 | void *data); |
| 93 | static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data); |
Adrian Bunk | 8169cae | 2006-03-31 16:53:55 +0200 | [diff] [blame] | 94 | static int dlm_lockres_master_requery(struct dlm_ctxt *dlm, |
| 95 | struct dlm_lock_resource *res, |
| 96 | u8 *real_master); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 97 | |
| 98 | static u64 dlm_get_next_mig_cookie(void); |
| 99 | |
Ingo Molnar | 34af946 | 2006-06-27 02:53:55 -0700 | [diff] [blame] | 100 | static DEFINE_SPINLOCK(dlm_reco_state_lock); |
| 101 | static DEFINE_SPINLOCK(dlm_mig_cookie_lock); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 102 | static u64 dlm_mig_cookie = 1; |
| 103 | |
| 104 | static u64 dlm_get_next_mig_cookie(void) |
| 105 | { |
| 106 | u64 c; |
| 107 | spin_lock(&dlm_mig_cookie_lock); |
| 108 | c = dlm_mig_cookie; |
| 109 | if (dlm_mig_cookie == (~0ULL)) |
| 110 | dlm_mig_cookie = 1; |
| 111 | else |
| 112 | dlm_mig_cookie++; |
| 113 | spin_unlock(&dlm_mig_cookie_lock); |
| 114 | return c; |
| 115 | } |
| 116 | |
Kurt Hackel | ab27eb6 | 2006-04-27 18:03:49 -0700 | [diff] [blame] | 117 | static inline void dlm_set_reco_dead_node(struct dlm_ctxt *dlm, |
| 118 | u8 dead_node) |
| 119 | { |
| 120 | assert_spin_locked(&dlm->spinlock); |
| 121 | if (dlm->reco.dead_node != dead_node) |
| 122 | mlog(0, "%s: changing dead_node from %u to %u\n", |
| 123 | dlm->name, dlm->reco.dead_node, dead_node); |
| 124 | dlm->reco.dead_node = dead_node; |
| 125 | } |
| 126 | |
| 127 | static inline void dlm_set_reco_master(struct dlm_ctxt *dlm, |
| 128 | u8 master) |
| 129 | { |
| 130 | assert_spin_locked(&dlm->spinlock); |
| 131 | mlog(0, "%s: changing new_master from %u to %u\n", |
| 132 | dlm->name, dlm->reco.new_master, master); |
| 133 | dlm->reco.new_master = master; |
| 134 | } |
| 135 | |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 136 | static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm) |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 137 | { |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 138 | assert_spin_locked(&dlm->spinlock); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 139 | clear_bit(dlm->reco.dead_node, dlm->recovery_map); |
Kurt Hackel | ab27eb6 | 2006-04-27 18:03:49 -0700 | [diff] [blame] | 140 | dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); |
| 141 | dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM); |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 142 | } |
| 143 | |
| 144 | static inline void dlm_reset_recovery(struct dlm_ctxt *dlm) |
| 145 | { |
| 146 | spin_lock(&dlm->spinlock); |
| 147 | __dlm_reset_recovery(dlm); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 148 | spin_unlock(&dlm->spinlock); |
| 149 | } |
| 150 | |
| 151 | /* Worker function used during recovery. */ |
David Howells | c402895 | 2006-11-22 14:57:56 +0000 | [diff] [blame] | 152 | void dlm_dispatch_work(struct work_struct *work) |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 153 | { |
David Howells | c402895 | 2006-11-22 14:57:56 +0000 | [diff] [blame] | 154 | struct dlm_ctxt *dlm = |
| 155 | container_of(work, struct dlm_ctxt, dispatched_work); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 156 | LIST_HEAD(tmp_list); |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 157 | struct dlm_work_item *item, *next; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 158 | dlm_workfunc_t *workfunc; |
Kurt Hackel | 3156d26 | 2006-05-01 14:39:29 -0700 | [diff] [blame] | 159 | int tot=0; |
| 160 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 161 | spin_lock(&dlm->work_lock); |
| 162 | list_splice_init(&dlm->work_list, &tmp_list); |
| 163 | spin_unlock(&dlm->work_lock); |
| 164 | |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 165 | list_for_each_entry(item, &tmp_list, list) { |
Kurt Hackel | 3156d26 | 2006-05-01 14:39:29 -0700 | [diff] [blame] | 166 | tot++; |
| 167 | } |
| 168 | mlog(0, "%s: work thread has %d work items\n", dlm->name, tot); |
| 169 | |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 170 | list_for_each_entry_safe(item, next, &tmp_list, list) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 171 | workfunc = item->func; |
| 172 | list_del_init(&item->list); |
| 173 | |
| 174 | /* already have ref on dlm to avoid having |
| 175 | * it disappear. just double-check. */ |
| 176 | BUG_ON(item->dlm != dlm); |
| 177 | |
| 178 | /* this is allowed to sleep and |
| 179 | * call network stuff */ |
| 180 | workfunc(item, item->data); |
| 181 | |
| 182 | dlm_put(dlm); |
| 183 | kfree(item); |
| 184 | } |
| 185 | } |
| 186 | |
| 187 | /* |
| 188 | * RECOVERY THREAD |
| 189 | */ |
| 190 | |
Kurt Hackel | c03872f | 2006-03-06 14:08:49 -0800 | [diff] [blame] | 191 | void dlm_kick_recovery_thread(struct dlm_ctxt *dlm) |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 192 | { |
| 193 | /* wake the recovery thread |
| 194 | * this will wake the reco thread in one of three places |
| 195 | * 1) sleeping with no recovery happening |
| 196 | * 2) sleeping with recovery mastered elsewhere |
| 197 | * 3) recovery mastered here, waiting on reco data */ |
| 198 | |
| 199 | wake_up(&dlm->dlm_reco_thread_wq); |
| 200 | } |
| 201 | |
| 202 | /* Launch the recovery thread */ |
| 203 | int dlm_launch_recovery_thread(struct dlm_ctxt *dlm) |
| 204 | { |
| 205 | mlog(0, "starting dlm recovery thread...\n"); |
| 206 | |
| 207 | dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm, |
| 208 | "dlm_reco_thread"); |
| 209 | if (IS_ERR(dlm->dlm_reco_thread_task)) { |
| 210 | mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task)); |
| 211 | dlm->dlm_reco_thread_task = NULL; |
| 212 | return -EINVAL; |
| 213 | } |
| 214 | |
| 215 | return 0; |
| 216 | } |
| 217 | |
| 218 | void dlm_complete_recovery_thread(struct dlm_ctxt *dlm) |
| 219 | { |
| 220 | if (dlm->dlm_reco_thread_task) { |
| 221 | mlog(0, "waiting for dlm recovery thread to exit\n"); |
| 222 | kthread_stop(dlm->dlm_reco_thread_task); |
| 223 | dlm->dlm_reco_thread_task = NULL; |
| 224 | } |
| 225 | } |
| 226 | |
| 227 | |
| 228 | |
| 229 | /* |
| 230 | * this is lame, but here's how recovery works... |
| 231 | * 1) all recovery threads cluster wide will work on recovering |
| 232 | * ONE node at a time |
| 233 | * 2) negotiate who will take over all the locks for the dead node. |
| 234 | * thats right... ALL the locks. |
| 235 | * 3) once a new master is chosen, everyone scans all locks |
| 236 | * and moves aside those mastered by the dead guy |
| 237 | * 4) each of these locks should be locked until recovery is done |
| 238 | * 5) the new master collects up all of secondary lock queue info |
| 239 | * one lock at a time, forcing each node to communicate back |
| 240 | * before continuing |
| 241 | * 6) each secondary lock queue responds with the full known lock info |
| 242 | * 7) once the new master has run all its locks, it sends a ALLDONE! |
| 243 | * message to everyone |
| 244 | * 8) upon receiving this message, the secondary queue node unlocks |
| 245 | * and responds to the ALLDONE |
| 246 | * 9) once the new master gets responses from everyone, he unlocks |
| 247 | * everything and recovery for this dead node is done |
| 248 | *10) go back to 2) while there are still dead nodes |
| 249 | * |
| 250 | */ |
| 251 | |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 252 | static void dlm_print_reco_node_status(struct dlm_ctxt *dlm) |
| 253 | { |
| 254 | struct dlm_reco_node_data *ndata; |
| 255 | struct dlm_lock_resource *res; |
| 256 | |
| 257 | mlog(ML_NOTICE, "%s(%d): recovery info, state=%s, dead=%u, master=%u\n", |
Pavel Emelyanov | ba25f9d | 2007-10-18 23:40:40 -0700 | [diff] [blame] | 258 | dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 259 | dlm->reco.state & DLM_RECO_STATE_ACTIVE ? "ACTIVE" : "inactive", |
| 260 | dlm->reco.dead_node, dlm->reco.new_master); |
| 261 | |
| 262 | list_for_each_entry(ndata, &dlm->reco.node_data, list) { |
| 263 | char *st = "unknown"; |
| 264 | switch (ndata->state) { |
| 265 | case DLM_RECO_NODE_DATA_INIT: |
| 266 | st = "init"; |
| 267 | break; |
| 268 | case DLM_RECO_NODE_DATA_REQUESTING: |
| 269 | st = "requesting"; |
| 270 | break; |
| 271 | case DLM_RECO_NODE_DATA_DEAD: |
| 272 | st = "dead"; |
| 273 | break; |
| 274 | case DLM_RECO_NODE_DATA_RECEIVING: |
| 275 | st = "receiving"; |
| 276 | break; |
| 277 | case DLM_RECO_NODE_DATA_REQUESTED: |
| 278 | st = "requested"; |
| 279 | break; |
| 280 | case DLM_RECO_NODE_DATA_DONE: |
| 281 | st = "done"; |
| 282 | break; |
| 283 | case DLM_RECO_NODE_DATA_FINALIZE_SENT: |
| 284 | st = "finalize-sent"; |
| 285 | break; |
| 286 | default: |
| 287 | st = "bad"; |
| 288 | break; |
| 289 | } |
| 290 | mlog(ML_NOTICE, "%s: reco state, node %u, state=%s\n", |
| 291 | dlm->name, ndata->node_num, st); |
| 292 | } |
| 293 | list_for_each_entry(res, &dlm->reco.resources, recovering) { |
| 294 | mlog(ML_NOTICE, "%s: lockres %.*s on recovering list\n", |
| 295 | dlm->name, res->lockname.len, res->lockname.name); |
| 296 | } |
| 297 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 298 | |
| 299 | #define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000) |
| 300 | |
| 301 | static int dlm_recovery_thread(void *data) |
| 302 | { |
| 303 | int status; |
| 304 | struct dlm_ctxt *dlm = data; |
| 305 | unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS); |
| 306 | |
| 307 | mlog(0, "dlm thread running for %s...\n", dlm->name); |
| 308 | |
| 309 | while (!kthread_should_stop()) { |
Srinivas Eeda | bc9838c | 2010-02-26 12:53:51 -0800 | [diff] [blame] | 310 | if (dlm_domain_fully_joined(dlm)) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 311 | status = dlm_do_recovery(dlm); |
| 312 | if (status == -EAGAIN) { |
| 313 | /* do not sleep, recheck immediately. */ |
| 314 | continue; |
| 315 | } |
| 316 | if (status < 0) |
| 317 | mlog_errno(status); |
| 318 | } |
| 319 | |
| 320 | wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq, |
| 321 | kthread_should_stop(), |
| 322 | timeout); |
| 323 | } |
| 324 | |
| 325 | mlog(0, "quitting DLM recovery thread\n"); |
| 326 | return 0; |
| 327 | } |
| 328 | |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 329 | /* returns true when the recovery master has contacted us */ |
| 330 | static int dlm_reco_master_ready(struct dlm_ctxt *dlm) |
| 331 | { |
| 332 | int ready; |
| 333 | spin_lock(&dlm->spinlock); |
| 334 | ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM); |
| 335 | spin_unlock(&dlm->spinlock); |
| 336 | return ready; |
| 337 | } |
| 338 | |
| 339 | /* returns true if node is no longer in the domain |
| 340 | * could be dead or just not joined */ |
| 341 | int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node) |
| 342 | { |
| 343 | int dead; |
| 344 | spin_lock(&dlm->spinlock); |
Kurt Hackel | aba9aac | 2006-04-27 18:00:21 -0700 | [diff] [blame] | 345 | dead = !test_bit(node, dlm->domain_map); |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 346 | spin_unlock(&dlm->spinlock); |
| 347 | return dead; |
| 348 | } |
| 349 | |
Kurt Hackel | b7084ab | 2006-05-01 13:54:07 -0700 | [diff] [blame] | 350 | /* returns true if node is no longer in the domain |
| 351 | * could be dead or just not joined */ |
Adrian Bunk | 3fb5a98 | 2006-05-16 17:26:41 +0200 | [diff] [blame] | 352 | static int dlm_is_node_recovered(struct dlm_ctxt *dlm, u8 node) |
Kurt Hackel | b7084ab | 2006-05-01 13:54:07 -0700 | [diff] [blame] | 353 | { |
| 354 | int recovered; |
| 355 | spin_lock(&dlm->spinlock); |
| 356 | recovered = !test_bit(node, dlm->recovery_map); |
| 357 | spin_unlock(&dlm->spinlock); |
| 358 | return recovered; |
| 359 | } |
| 360 | |
| 361 | |
Sunil Mushran | ed8625c | 2011-07-24 10:28:54 -0700 | [diff] [blame] | 362 | void dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout) |
Kurt Hackel | 44465a7 | 2006-01-18 17:05:38 -0800 | [diff] [blame] | 363 | { |
Sunil Mushran | ed8625c | 2011-07-24 10:28:54 -0700 | [diff] [blame] | 364 | if (dlm_is_node_dead(dlm, node)) |
| 365 | return; |
| 366 | |
| 367 | printk(KERN_NOTICE "o2dlm: Waiting on the death of node %u in " |
| 368 | "domain %s\n", node, dlm->name); |
| 369 | |
| 370 | if (timeout) |
Kurt Hackel | 44465a7 | 2006-01-18 17:05:38 -0800 | [diff] [blame] | 371 | wait_event_timeout(dlm->dlm_reco_thread_wq, |
Sunil Mushran | ed8625c | 2011-07-24 10:28:54 -0700 | [diff] [blame] | 372 | dlm_is_node_dead(dlm, node), |
| 373 | msecs_to_jiffies(timeout)); |
| 374 | else |
Kurt Hackel | 44465a7 | 2006-01-18 17:05:38 -0800 | [diff] [blame] | 375 | wait_event(dlm->dlm_reco_thread_wq, |
| 376 | dlm_is_node_dead(dlm, node)); |
Kurt Hackel | 44465a7 | 2006-01-18 17:05:38 -0800 | [diff] [blame] | 377 | } |
| 378 | |
Sunil Mushran | ed8625c | 2011-07-24 10:28:54 -0700 | [diff] [blame] | 379 | void dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout) |
Kurt Hackel | b7084ab | 2006-05-01 13:54:07 -0700 | [diff] [blame] | 380 | { |
Sunil Mushran | ed8625c | 2011-07-24 10:28:54 -0700 | [diff] [blame] | 381 | if (dlm_is_node_recovered(dlm, node)) |
| 382 | return; |
| 383 | |
| 384 | printk(KERN_NOTICE "o2dlm: Waiting on the recovery of node %u in " |
| 385 | "domain %s\n", node, dlm->name); |
| 386 | |
| 387 | if (timeout) |
Kurt Hackel | b7084ab | 2006-05-01 13:54:07 -0700 | [diff] [blame] | 388 | wait_event_timeout(dlm->dlm_reco_thread_wq, |
Sunil Mushran | ed8625c | 2011-07-24 10:28:54 -0700 | [diff] [blame] | 389 | dlm_is_node_recovered(dlm, node), |
| 390 | msecs_to_jiffies(timeout)); |
| 391 | else |
Kurt Hackel | b7084ab | 2006-05-01 13:54:07 -0700 | [diff] [blame] | 392 | wait_event(dlm->dlm_reco_thread_wq, |
| 393 | dlm_is_node_recovered(dlm, node)); |
Kurt Hackel | b7084ab | 2006-05-01 13:54:07 -0700 | [diff] [blame] | 394 | } |
| 395 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 396 | /* callers of the top-level api calls (dlmlock/dlmunlock) should |
| 397 | * block on the dlm->reco.event when recovery is in progress. |
| 398 | * the dlm recovery thread will set this state when it begins |
| 399 | * recovering a dead node (as the new master or not) and clear |
| 400 | * the state and wake as soon as all affected lock resources have |
| 401 | * been marked with the RECOVERY flag */ |
| 402 | static int dlm_in_recovery(struct dlm_ctxt *dlm) |
| 403 | { |
| 404 | int in_recovery; |
| 405 | spin_lock(&dlm->spinlock); |
| 406 | in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE); |
| 407 | spin_unlock(&dlm->spinlock); |
| 408 | return in_recovery; |
| 409 | } |
| 410 | |
| 411 | |
| 412 | void dlm_wait_for_recovery(struct dlm_ctxt *dlm) |
| 413 | { |
Kurt Hackel | 56a7c10 | 2006-05-01 14:30:39 -0700 | [diff] [blame] | 414 | if (dlm_in_recovery(dlm)) { |
Kurt Hackel | 3b3b84a | 2006-05-01 14:31:37 -0700 | [diff] [blame] | 415 | mlog(0, "%s: reco thread %d in recovery: " |
Kurt Hackel | 56a7c10 | 2006-05-01 14:30:39 -0700 | [diff] [blame] | 416 | "state=%d, master=%u, dead=%u\n", |
Pavel Emelyanov | ba25f9d | 2007-10-18 23:40:40 -0700 | [diff] [blame] | 417 | dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), |
Kurt Hackel | 56a7c10 | 2006-05-01 14:30:39 -0700 | [diff] [blame] | 418 | dlm->reco.state, dlm->reco.new_master, |
| 419 | dlm->reco.dead_node); |
| 420 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 421 | wait_event(dlm->reco.event, !dlm_in_recovery(dlm)); |
| 422 | } |
| 423 | |
| 424 | static void dlm_begin_recovery(struct dlm_ctxt *dlm) |
| 425 | { |
| 426 | spin_lock(&dlm->spinlock); |
| 427 | BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE); |
Sunil Mushran | 8decab3 | 2011-07-24 10:23:54 -0700 | [diff] [blame] | 428 | printk(KERN_NOTICE "o2dlm: Begin recovery on domain %s for node %u\n", |
| 429 | dlm->name, dlm->reco.dead_node); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 430 | dlm->reco.state |= DLM_RECO_STATE_ACTIVE; |
| 431 | spin_unlock(&dlm->spinlock); |
| 432 | } |
| 433 | |
| 434 | static void dlm_end_recovery(struct dlm_ctxt *dlm) |
| 435 | { |
| 436 | spin_lock(&dlm->spinlock); |
| 437 | BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE)); |
| 438 | dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE; |
| 439 | spin_unlock(&dlm->spinlock); |
Sunil Mushran | 8decab3 | 2011-07-24 10:23:54 -0700 | [diff] [blame] | 440 | printk(KERN_NOTICE "o2dlm: End recovery on domain %s\n", dlm->name); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 441 | wake_up(&dlm->reco.event); |
| 442 | } |
| 443 | |
Sunil Mushran | 8decab3 | 2011-07-24 10:23:54 -0700 | [diff] [blame] | 444 | static void dlm_print_recovery_master(struct dlm_ctxt *dlm) |
| 445 | { |
| 446 | printk(KERN_NOTICE "o2dlm: Node %u (%s) is the Recovery Master for the " |
| 447 | "dead node %u in domain %s\n", dlm->reco.new_master, |
| 448 | (dlm->node_num == dlm->reco.new_master ? "me" : "he"), |
| 449 | dlm->reco.dead_node, dlm->name); |
| 450 | } |
| 451 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 452 | static int dlm_do_recovery(struct dlm_ctxt *dlm) |
| 453 | { |
| 454 | int status = 0; |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 455 | int ret; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 456 | |
| 457 | spin_lock(&dlm->spinlock); |
| 458 | |
| 459 | /* check to see if the new master has died */ |
| 460 | if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM && |
| 461 | test_bit(dlm->reco.new_master, dlm->recovery_map)) { |
| 462 | mlog(0, "new master %u died while recovering %u!\n", |
| 463 | dlm->reco.new_master, dlm->reco.dead_node); |
| 464 | /* unset the new_master, leave dead_node */ |
Kurt Hackel | ab27eb6 | 2006-04-27 18:03:49 -0700 | [diff] [blame] | 465 | dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 466 | } |
| 467 | |
| 468 | /* select a target to recover */ |
| 469 | if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { |
| 470 | int bit; |
| 471 | |
Wengang Wang | f471c9d | 2010-06-30 20:23:30 +0800 | [diff] [blame] | 472 | bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES, 0); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 473 | if (bit >= O2NM_MAX_NODES || bit < 0) |
Kurt Hackel | ab27eb6 | 2006-04-27 18:03:49 -0700 | [diff] [blame] | 474 | dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 475 | else |
Kurt Hackel | ab27eb6 | 2006-04-27 18:03:49 -0700 | [diff] [blame] | 476 | dlm_set_reco_dead_node(dlm, bit); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 477 | } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) { |
| 478 | /* BUG? */ |
| 479 | mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n", |
| 480 | dlm->reco.dead_node); |
Kurt Hackel | ab27eb6 | 2006-04-27 18:03:49 -0700 | [diff] [blame] | 481 | dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 482 | } |
| 483 | |
| 484 | if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { |
| 485 | // mlog(0, "nothing to recover! sleeping now!\n"); |
| 486 | spin_unlock(&dlm->spinlock); |
| 487 | /* return to main thread loop and sleep. */ |
| 488 | return 0; |
| 489 | } |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 490 | mlog(0, "%s(%d):recovery thread found node %u in the recovery map!\n", |
Pavel Emelyanov | ba25f9d | 2007-10-18 23:40:40 -0700 | [diff] [blame] | 491 | dlm->name, task_pid_nr(dlm->dlm_reco_thread_task), |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 492 | dlm->reco.dead_node); |
| 493 | spin_unlock(&dlm->spinlock); |
| 494 | |
| 495 | /* take write barrier */ |
| 496 | /* (stops the list reshuffling thread, proxy ast handling) */ |
| 497 | dlm_begin_recovery(dlm); |
| 498 | |
| 499 | if (dlm->reco.new_master == dlm->node_num) |
| 500 | goto master_here; |
| 501 | |
| 502 | if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) { |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 503 | /* choose a new master, returns 0 if this node |
| 504 | * is the master, -EEXIST if it's another node. |
| 505 | * this does not return until a new master is chosen |
| 506 | * or recovery completes entirely. */ |
| 507 | ret = dlm_pick_recovery_master(dlm); |
| 508 | if (!ret) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 509 | /* already notified everyone. go. */ |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 510 | goto master_here; |
| 511 | } |
| 512 | mlog(0, "another node will master this recovery session.\n"); |
| 513 | } |
Sunil Mushran | 8decab3 | 2011-07-24 10:23:54 -0700 | [diff] [blame] | 514 | |
| 515 | dlm_print_recovery_master(dlm); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 516 | |
| 517 | /* it is safe to start everything back up here |
| 518 | * because all of the dead node's lock resources |
| 519 | * have been marked as in-recovery */ |
| 520 | dlm_end_recovery(dlm); |
| 521 | |
| 522 | /* sleep out in main dlm_recovery_thread loop. */ |
| 523 | return 0; |
| 524 | |
| 525 | master_here: |
Sunil Mushran | 8decab3 | 2011-07-24 10:23:54 -0700 | [diff] [blame] | 526 | dlm_print_recovery_master(dlm); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 527 | |
| 528 | status = dlm_remaster_locks(dlm, dlm->reco.dead_node); |
| 529 | if (status < 0) { |
Kurt Hackel | 6a41321 | 2006-05-01 13:49:20 -0700 | [diff] [blame] | 530 | /* we should never hit this anymore */ |
Sunil Mushran | 8decab3 | 2011-07-24 10:23:54 -0700 | [diff] [blame] | 531 | mlog(ML_ERROR, "%s: Error %d remastering locks for node %u, " |
| 532 | "retrying.\n", dlm->name, status, dlm->reco.dead_node); |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 533 | /* yield a bit to allow any final network messages |
| 534 | * to get handled on remaining nodes */ |
| 535 | msleep(100); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 536 | } else { |
| 537 | /* success! see if any other nodes need recovery */ |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 538 | mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n", |
| 539 | dlm->name, dlm->reco.dead_node, dlm->node_num); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 540 | dlm_reset_recovery(dlm); |
| 541 | } |
| 542 | dlm_end_recovery(dlm); |
| 543 | |
| 544 | /* continue and look for another dead node */ |
| 545 | return -EAGAIN; |
| 546 | } |
| 547 | |
| 548 | static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node) |
| 549 | { |
| 550 | int status = 0; |
| 551 | struct dlm_reco_node_data *ndata; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 552 | int all_nodes_done; |
| 553 | int destroy = 0; |
| 554 | int pass = 0; |
| 555 | |
Kurt Hackel | 6a41321 | 2006-05-01 13:49:20 -0700 | [diff] [blame] | 556 | do { |
| 557 | /* we have become recovery master. there is no escaping |
| 558 | * this, so just keep trying until we get it. */ |
| 559 | status = dlm_init_recovery_area(dlm, dead_node); |
| 560 | if (status < 0) { |
| 561 | mlog(ML_ERROR, "%s: failed to alloc recovery area, " |
| 562 | "retrying\n", dlm->name); |
| 563 | msleep(1000); |
| 564 | } |
| 565 | } while (status != 0); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 566 | |
| 567 | /* safe to access the node data list without a lock, since this |
| 568 | * process is the only one to change the list */ |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 569 | list_for_each_entry(ndata, &dlm->reco.node_data, list) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 570 | BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT); |
| 571 | ndata->state = DLM_RECO_NODE_DATA_REQUESTING; |
| 572 | |
Sunil Mushran | 8decab3 | 2011-07-24 10:23:54 -0700 | [diff] [blame] | 573 | mlog(0, "%s: Requesting lock info from node %u\n", dlm->name, |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 574 | ndata->node_num); |
| 575 | |
| 576 | if (ndata->node_num == dlm->node_num) { |
| 577 | ndata->state = DLM_RECO_NODE_DATA_DONE; |
| 578 | continue; |
| 579 | } |
| 580 | |
Kurt Hackel | 6a41321 | 2006-05-01 13:49:20 -0700 | [diff] [blame] | 581 | do { |
| 582 | status = dlm_request_all_locks(dlm, ndata->node_num, |
| 583 | dead_node); |
| 584 | if (status < 0) { |
| 585 | mlog_errno(status); |
| 586 | if (dlm_is_host_down(status)) { |
| 587 | /* node died, ignore it for recovery */ |
| 588 | status = 0; |
| 589 | ndata->state = DLM_RECO_NODE_DATA_DEAD; |
| 590 | /* wait for the domain map to catch up |
| 591 | * with the network state. */ |
| 592 | wait_event_timeout(dlm->dlm_reco_thread_wq, |
| 593 | dlm_is_node_dead(dlm, |
| 594 | ndata->node_num), |
| 595 | msecs_to_jiffies(1000)); |
| 596 | mlog(0, "waited 1 sec for %u, " |
| 597 | "dead? %s\n", ndata->node_num, |
| 598 | dlm_is_node_dead(dlm, ndata->node_num) ? |
| 599 | "yes" : "no"); |
| 600 | } else { |
| 601 | /* -ENOMEM on the other node */ |
| 602 | mlog(0, "%s: node %u returned " |
| 603 | "%d during recovery, retrying " |
| 604 | "after a short wait\n", |
| 605 | dlm->name, ndata->node_num, |
| 606 | status); |
| 607 | msleep(100); |
| 608 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 609 | } |
Kurt Hackel | 6a41321 | 2006-05-01 13:49:20 -0700 | [diff] [blame] | 610 | } while (status != 0); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 611 | |
Srinivas Eeda | 756a150 | 2007-04-17 13:26:33 -0700 | [diff] [blame] | 612 | spin_lock(&dlm_reco_state_lock); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 613 | switch (ndata->state) { |
| 614 | case DLM_RECO_NODE_DATA_INIT: |
| 615 | case DLM_RECO_NODE_DATA_FINALIZE_SENT: |
| 616 | case DLM_RECO_NODE_DATA_REQUESTED: |
| 617 | BUG(); |
| 618 | break; |
| 619 | case DLM_RECO_NODE_DATA_DEAD: |
| 620 | mlog(0, "node %u died after requesting " |
| 621 | "recovery info for node %u\n", |
| 622 | ndata->node_num, dead_node); |
Kurt Hackel | 6a41321 | 2006-05-01 13:49:20 -0700 | [diff] [blame] | 623 | /* fine. don't need this node's info. |
| 624 | * continue without it. */ |
| 625 | break; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 626 | case DLM_RECO_NODE_DATA_REQUESTING: |
| 627 | ndata->state = DLM_RECO_NODE_DATA_REQUESTED; |
| 628 | mlog(0, "now receiving recovery data from " |
| 629 | "node %u for dead node %u\n", |
| 630 | ndata->node_num, dead_node); |
| 631 | break; |
| 632 | case DLM_RECO_NODE_DATA_RECEIVING: |
| 633 | mlog(0, "already receiving recovery data from " |
| 634 | "node %u for dead node %u\n", |
| 635 | ndata->node_num, dead_node); |
| 636 | break; |
| 637 | case DLM_RECO_NODE_DATA_DONE: |
| 638 | mlog(0, "already DONE receiving recovery data " |
| 639 | "from node %u for dead node %u\n", |
| 640 | ndata->node_num, dead_node); |
| 641 | break; |
| 642 | } |
Srinivas Eeda | 756a150 | 2007-04-17 13:26:33 -0700 | [diff] [blame] | 643 | spin_unlock(&dlm_reco_state_lock); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 644 | } |
| 645 | |
Sunil Mushran | 8decab3 | 2011-07-24 10:23:54 -0700 | [diff] [blame] | 646 | mlog(0, "%s: Done requesting all lock info\n", dlm->name); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 647 | |
| 648 | /* nodes should be sending reco data now |
| 649 | * just need to wait */ |
| 650 | |
| 651 | while (1) { |
| 652 | /* check all the nodes now to see if we are |
| 653 | * done, or if anyone died */ |
| 654 | all_nodes_done = 1; |
| 655 | spin_lock(&dlm_reco_state_lock); |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 656 | list_for_each_entry(ndata, &dlm->reco.node_data, list) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 657 | mlog(0, "checking recovery state of node %u\n", |
| 658 | ndata->node_num); |
| 659 | switch (ndata->state) { |
| 660 | case DLM_RECO_NODE_DATA_INIT: |
| 661 | case DLM_RECO_NODE_DATA_REQUESTING: |
| 662 | mlog(ML_ERROR, "bad ndata state for " |
| 663 | "node %u: state=%d\n", |
| 664 | ndata->node_num, ndata->state); |
| 665 | BUG(); |
| 666 | break; |
| 667 | case DLM_RECO_NODE_DATA_DEAD: |
Kurt Hackel | 6a41321 | 2006-05-01 13:49:20 -0700 | [diff] [blame] | 668 | mlog(0, "node %u died after " |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 669 | "requesting recovery info for " |
| 670 | "node %u\n", ndata->node_num, |
| 671 | dead_node); |
Kurt Hackel | 6a41321 | 2006-05-01 13:49:20 -0700 | [diff] [blame] | 672 | break; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 673 | case DLM_RECO_NODE_DATA_RECEIVING: |
| 674 | case DLM_RECO_NODE_DATA_REQUESTED: |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 675 | mlog(0, "%s: node %u still in state %s\n", |
| 676 | dlm->name, ndata->node_num, |
| 677 | ndata->state==DLM_RECO_NODE_DATA_RECEIVING ? |
| 678 | "receiving" : "requested"); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 679 | all_nodes_done = 0; |
| 680 | break; |
| 681 | case DLM_RECO_NODE_DATA_DONE: |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 682 | mlog(0, "%s: node %u state is done\n", |
| 683 | dlm->name, ndata->node_num); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 684 | break; |
| 685 | case DLM_RECO_NODE_DATA_FINALIZE_SENT: |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 686 | mlog(0, "%s: node %u state is finalize\n", |
| 687 | dlm->name, ndata->node_num); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 688 | break; |
| 689 | } |
| 690 | } |
| 691 | spin_unlock(&dlm_reco_state_lock); |
| 692 | |
| 693 | mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass, |
| 694 | all_nodes_done?"yes":"no"); |
| 695 | if (all_nodes_done) { |
| 696 | int ret; |
| 697 | |
| 698 | /* all nodes are now in DLM_RECO_NODE_DATA_DONE state |
| 699 | * just send a finalize message to everyone and |
| 700 | * clean up */ |
| 701 | mlog(0, "all nodes are done! send finalize\n"); |
| 702 | ret = dlm_send_finalize_reco_message(dlm); |
| 703 | if (ret < 0) |
| 704 | mlog_errno(ret); |
| 705 | |
| 706 | spin_lock(&dlm->spinlock); |
| 707 | dlm_finish_local_lockres_recovery(dlm, dead_node, |
| 708 | dlm->node_num); |
| 709 | spin_unlock(&dlm->spinlock); |
| 710 | mlog(0, "should be done with recovery!\n"); |
| 711 | |
| 712 | mlog(0, "finishing recovery of %s at %lu, " |
| 713 | "dead=%u, this=%u, new=%u\n", dlm->name, |
| 714 | jiffies, dlm->reco.dead_node, |
| 715 | dlm->node_num, dlm->reco.new_master); |
| 716 | destroy = 1; |
Kurt Hackel | 6a41321 | 2006-05-01 13:49:20 -0700 | [diff] [blame] | 717 | status = 0; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 718 | /* rescan everything marked dirty along the way */ |
| 719 | dlm_kick_thread(dlm, NULL); |
| 720 | break; |
| 721 | } |
| 722 | /* wait to be signalled, with periodic timeout |
| 723 | * to check for node death */ |
| 724 | wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq, |
| 725 | kthread_should_stop(), |
| 726 | msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS)); |
| 727 | |
| 728 | } |
| 729 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 730 | if (destroy) |
| 731 | dlm_destroy_recovery_area(dlm, dead_node); |
| 732 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 733 | return status; |
| 734 | } |
| 735 | |
| 736 | static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node) |
| 737 | { |
| 738 | int num=0; |
| 739 | struct dlm_reco_node_data *ndata; |
| 740 | |
| 741 | spin_lock(&dlm->spinlock); |
| 742 | memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map)); |
| 743 | /* nodes can only be removed (by dying) after dropping |
| 744 | * this lock, and death will be trapped later, so this should do */ |
| 745 | spin_unlock(&dlm->spinlock); |
| 746 | |
| 747 | while (1) { |
| 748 | num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num); |
| 749 | if (num >= O2NM_MAX_NODES) { |
| 750 | break; |
| 751 | } |
| 752 | BUG_ON(num == dead_node); |
| 753 | |
Robert P. J. Day | cd86128 | 2006-12-13 00:34:52 -0800 | [diff] [blame] | 754 | ndata = kzalloc(sizeof(*ndata), GFP_NOFS); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 755 | if (!ndata) { |
| 756 | dlm_destroy_recovery_area(dlm, dead_node); |
| 757 | return -ENOMEM; |
| 758 | } |
| 759 | ndata->node_num = num; |
| 760 | ndata->state = DLM_RECO_NODE_DATA_INIT; |
| 761 | spin_lock(&dlm_reco_state_lock); |
| 762 | list_add_tail(&ndata->list, &dlm->reco.node_data); |
| 763 | spin_unlock(&dlm_reco_state_lock); |
| 764 | num++; |
| 765 | } |
| 766 | |
| 767 | return 0; |
| 768 | } |
| 769 | |
| 770 | static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node) |
| 771 | { |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 772 | struct dlm_reco_node_data *ndata, *next; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 773 | LIST_HEAD(tmplist); |
| 774 | |
| 775 | spin_lock(&dlm_reco_state_lock); |
| 776 | list_splice_init(&dlm->reco.node_data, &tmplist); |
| 777 | spin_unlock(&dlm_reco_state_lock); |
| 778 | |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 779 | list_for_each_entry_safe(ndata, next, &tmplist, list) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 780 | list_del_init(&ndata->list); |
| 781 | kfree(ndata); |
| 782 | } |
| 783 | } |
| 784 | |
| 785 | static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from, |
| 786 | u8 dead_node) |
| 787 | { |
| 788 | struct dlm_lock_request lr; |
Joseph Qi | 22ab901 | 2013-07-03 15:00:47 -0700 | [diff] [blame] | 789 | int ret; |
Xue jiufei | 98ac912 | 2013-09-11 14:19:46 -0700 | [diff] [blame] | 790 | int status; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 791 | |
| 792 | mlog(0, "\n"); |
| 793 | |
| 794 | |
| 795 | mlog(0, "dlm_request_all_locks: dead node is %u, sending request " |
| 796 | "to %u\n", dead_node, request_from); |
| 797 | |
| 798 | memset(&lr, 0, sizeof(lr)); |
| 799 | lr.node_idx = dlm->node_num; |
| 800 | lr.dead_node = dead_node; |
| 801 | |
| 802 | // send message |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 803 | ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key, |
Xue jiufei | 98ac912 | 2013-09-11 14:19:46 -0700 | [diff] [blame] | 804 | &lr, sizeof(lr), request_from, &status); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 805 | |
| 806 | /* negative status is handled by caller */ |
| 807 | if (ret < 0) |
Sunil Mushran | 8decab3 | 2011-07-24 10:23:54 -0700 | [diff] [blame] | 808 | mlog(ML_ERROR, "%s: Error %d send LOCK_REQUEST to node %u " |
| 809 | "to recover dead node %u\n", dlm->name, ret, |
| 810 | request_from, dead_node); |
Xue jiufei | 98ac912 | 2013-09-11 14:19:46 -0700 | [diff] [blame] | 811 | else |
| 812 | ret = status; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 813 | // return from here, then |
| 814 | // sleep until all received or error |
| 815 | return ret; |
| 816 | |
| 817 | } |
| 818 | |
Kurt Hackel | d74c980 | 2007-01-17 17:04:25 -0800 | [diff] [blame] | 819 | int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data, |
| 820 | void **ret_data) |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 821 | { |
| 822 | struct dlm_ctxt *dlm = data; |
| 823 | struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf; |
| 824 | char *buf = NULL; |
| 825 | struct dlm_work_item *item = NULL; |
| 826 | |
| 827 | if (!dlm_grab(dlm)) |
| 828 | return -EINVAL; |
| 829 | |
Kurt Hackel | c3187ce | 2006-04-27 18:05:41 -0700 | [diff] [blame] | 830 | if (lr->dead_node != dlm->reco.dead_node) { |
| 831 | mlog(ML_ERROR, "%s: node %u sent dead_node=%u, but local " |
| 832 | "dead_node is %u\n", dlm->name, lr->node_idx, |
| 833 | lr->dead_node, dlm->reco.dead_node); |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 834 | dlm_print_reco_node_status(dlm); |
Kurt Hackel | c3187ce | 2006-04-27 18:05:41 -0700 | [diff] [blame] | 835 | /* this is a hack */ |
| 836 | dlm_put(dlm); |
| 837 | return -ENOMEM; |
| 838 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 839 | BUG_ON(lr->dead_node != dlm->reco.dead_node); |
| 840 | |
Robert P. J. Day | cd86128 | 2006-12-13 00:34:52 -0800 | [diff] [blame] | 841 | item = kzalloc(sizeof(*item), GFP_NOFS); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 842 | if (!item) { |
| 843 | dlm_put(dlm); |
| 844 | return -ENOMEM; |
| 845 | } |
| 846 | |
| 847 | /* this will get freed by dlm_request_all_locks_worker */ |
Kurt Hackel | ad8100e | 2006-05-01 14:25:21 -0700 | [diff] [blame] | 848 | buf = (char *) __get_free_page(GFP_NOFS); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 849 | if (!buf) { |
| 850 | kfree(item); |
| 851 | dlm_put(dlm); |
| 852 | return -ENOMEM; |
| 853 | } |
| 854 | |
| 855 | /* queue up work for dlm_request_all_locks_worker */ |
| 856 | dlm_grab(dlm); /* get an extra ref for the work item */ |
| 857 | dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf); |
| 858 | item->u.ral.reco_master = lr->node_idx; |
| 859 | item->u.ral.dead_node = lr->dead_node; |
| 860 | spin_lock(&dlm->work_lock); |
| 861 | list_add_tail(&item->list, &dlm->work_list); |
| 862 | spin_unlock(&dlm->work_lock); |
Kurt Hackel | 3156d26 | 2006-05-01 14:39:29 -0700 | [diff] [blame] | 863 | queue_work(dlm->dlm_worker, &dlm->dispatched_work); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 864 | |
| 865 | dlm_put(dlm); |
| 866 | return 0; |
| 867 | } |
| 868 | |
| 869 | static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data) |
| 870 | { |
| 871 | struct dlm_migratable_lockres *mres; |
| 872 | struct dlm_lock_resource *res; |
| 873 | struct dlm_ctxt *dlm; |
| 874 | LIST_HEAD(resources); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 875 | int ret; |
| 876 | u8 dead_node, reco_master; |
Kurt Hackel | 29c0fa0 | 2006-04-27 18:06:58 -0700 | [diff] [blame] | 877 | int skip_all_done = 0; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 878 | |
| 879 | dlm = item->dlm; |
| 880 | dead_node = item->u.ral.dead_node; |
| 881 | reco_master = item->u.ral.reco_master; |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 882 | mres = (struct dlm_migratable_lockres *)data; |
| 883 | |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 884 | mlog(0, "%s: recovery worker started, dead=%u, master=%u\n", |
| 885 | dlm->name, dead_node, reco_master); |
| 886 | |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 887 | if (dead_node != dlm->reco.dead_node || |
| 888 | reco_master != dlm->reco.new_master) { |
Kurt Hackel | 6a41321 | 2006-05-01 13:49:20 -0700 | [diff] [blame] | 889 | /* worker could have been created before the recovery master |
| 890 | * died. if so, do not continue, but do not error. */ |
| 891 | if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) { |
| 892 | mlog(ML_NOTICE, "%s: will not send recovery state, " |
| 893 | "recovery master %u died, thread=(dead=%u,mas=%u)" |
| 894 | " current=(dead=%u,mas=%u)\n", dlm->name, |
| 895 | reco_master, dead_node, reco_master, |
| 896 | dlm->reco.dead_node, dlm->reco.new_master); |
| 897 | } else { |
| 898 | mlog(ML_NOTICE, "%s: reco state invalid: reco(dead=%u, " |
| 899 | "master=%u), request(dead=%u, master=%u)\n", |
| 900 | dlm->name, dlm->reco.dead_node, |
| 901 | dlm->reco.new_master, dead_node, reco_master); |
| 902 | } |
| 903 | goto leave; |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 904 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 905 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 906 | /* lock resources should have already been moved to the |
| 907 | * dlm->reco.resources list. now move items from that list |
| 908 | * to a temp list if the dead owner matches. note that the |
| 909 | * whole cluster recovers only one node at a time, so we |
| 910 | * can safely move UNKNOWN lock resources for each recovery |
| 911 | * session. */ |
| 912 | dlm_move_reco_locks_to_list(dlm, &resources, dead_node); |
| 913 | |
| 914 | /* now we can begin blasting lockreses without the dlm lock */ |
Kurt Hackel | 29c0fa0 | 2006-04-27 18:06:58 -0700 | [diff] [blame] | 915 | |
| 916 | /* any errors returned will be due to the new_master dying, |
| 917 | * the dlm_reco_thread should detect this */ |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 918 | list_for_each_entry(res, &resources, recovering) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 919 | ret = dlm_send_one_lockres(dlm, res, mres, reco_master, |
| 920 | DLM_MRES_RECOVERY); |
Kurt Hackel | 29c0fa0 | 2006-04-27 18:06:58 -0700 | [diff] [blame] | 921 | if (ret < 0) { |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 922 | mlog(ML_ERROR, "%s: node %u went down while sending " |
| 923 | "recovery state for dead node %u, ret=%d\n", dlm->name, |
| 924 | reco_master, dead_node, ret); |
Kurt Hackel | 29c0fa0 | 2006-04-27 18:06:58 -0700 | [diff] [blame] | 925 | skip_all_done = 1; |
| 926 | break; |
| 927 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 928 | } |
| 929 | |
| 930 | /* move the resources back to the list */ |
| 931 | spin_lock(&dlm->spinlock); |
| 932 | list_splice_init(&resources, &dlm->reco.resources); |
| 933 | spin_unlock(&dlm->spinlock); |
| 934 | |
Kurt Hackel | 29c0fa0 | 2006-04-27 18:06:58 -0700 | [diff] [blame] | 935 | if (!skip_all_done) { |
| 936 | ret = dlm_send_all_done_msg(dlm, dead_node, reco_master); |
| 937 | if (ret < 0) { |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 938 | mlog(ML_ERROR, "%s: node %u went down while sending " |
| 939 | "recovery all-done for dead node %u, ret=%d\n", |
| 940 | dlm->name, reco_master, dead_node, ret); |
Kurt Hackel | 29c0fa0 | 2006-04-27 18:06:58 -0700 | [diff] [blame] | 941 | } |
| 942 | } |
Kurt Hackel | 6a41321 | 2006-05-01 13:49:20 -0700 | [diff] [blame] | 943 | leave: |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 944 | free_page((unsigned long)data); |
| 945 | } |
| 946 | |
| 947 | |
| 948 | static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to) |
| 949 | { |
| 950 | int ret, tmpret; |
| 951 | struct dlm_reco_data_done done_msg; |
| 952 | |
| 953 | memset(&done_msg, 0, sizeof(done_msg)); |
| 954 | done_msg.node_idx = dlm->node_num; |
| 955 | done_msg.dead_node = dead_node; |
| 956 | mlog(0, "sending DATA DONE message to %u, " |
| 957 | "my node=%u, dead node=%u\n", send_to, done_msg.node_idx, |
| 958 | done_msg.dead_node); |
| 959 | |
| 960 | ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg, |
| 961 | sizeof(done_msg), send_to, &tmpret); |
Kurt Hackel | 29c0fa0 | 2006-04-27 18:06:58 -0700 | [diff] [blame] | 962 | if (ret < 0) { |
Sunil Mushran | 8decab3 | 2011-07-24 10:23:54 -0700 | [diff] [blame] | 963 | mlog(ML_ERROR, "%s: Error %d send RECO_DATA_DONE to node %u " |
| 964 | "to recover dead node %u\n", dlm->name, ret, send_to, |
| 965 | dead_node); |
Kurt Hackel | 29c0fa0 | 2006-04-27 18:06:58 -0700 | [diff] [blame] | 966 | if (!dlm_is_host_down(ret)) { |
Kurt Hackel | 29c0fa0 | 2006-04-27 18:06:58 -0700 | [diff] [blame] | 967 | BUG(); |
| 968 | } |
| 969 | } else |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 970 | ret = tmpret; |
| 971 | return ret; |
| 972 | } |
| 973 | |
| 974 | |
Kurt Hackel | d74c980 | 2007-01-17 17:04:25 -0800 | [diff] [blame] | 975 | int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data, |
| 976 | void **ret_data) |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 977 | { |
| 978 | struct dlm_ctxt *dlm = data; |
| 979 | struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 980 | struct dlm_reco_node_data *ndata = NULL; |
| 981 | int ret = -EINVAL; |
| 982 | |
| 983 | if (!dlm_grab(dlm)) |
| 984 | return -EINVAL; |
| 985 | |
| 986 | mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, " |
| 987 | "node_idx=%u, this node=%u\n", done->dead_node, |
| 988 | dlm->reco.dead_node, done->node_idx, dlm->node_num); |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 989 | |
| 990 | mlog_bug_on_msg((done->dead_node != dlm->reco.dead_node), |
| 991 | "Got DATA DONE: dead_node=%u, reco.dead_node=%u, " |
| 992 | "node_idx=%u, this node=%u\n", done->dead_node, |
| 993 | dlm->reco.dead_node, done->node_idx, dlm->node_num); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 994 | |
| 995 | spin_lock(&dlm_reco_state_lock); |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 996 | list_for_each_entry(ndata, &dlm->reco.node_data, list) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 997 | if (ndata->node_num != done->node_idx) |
| 998 | continue; |
| 999 | |
| 1000 | switch (ndata->state) { |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 1001 | /* should have moved beyond INIT but not to FINALIZE yet */ |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1002 | case DLM_RECO_NODE_DATA_INIT: |
| 1003 | case DLM_RECO_NODE_DATA_DEAD: |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1004 | case DLM_RECO_NODE_DATA_FINALIZE_SENT: |
| 1005 | mlog(ML_ERROR, "bad ndata state for node %u:" |
| 1006 | " state=%d\n", ndata->node_num, |
| 1007 | ndata->state); |
| 1008 | BUG(); |
| 1009 | break; |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 1010 | /* these states are possible at this point, anywhere along |
| 1011 | * the line of recovery */ |
| 1012 | case DLM_RECO_NODE_DATA_DONE: |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1013 | case DLM_RECO_NODE_DATA_RECEIVING: |
| 1014 | case DLM_RECO_NODE_DATA_REQUESTED: |
| 1015 | case DLM_RECO_NODE_DATA_REQUESTING: |
| 1016 | mlog(0, "node %u is DONE sending " |
| 1017 | "recovery data!\n", |
| 1018 | ndata->node_num); |
| 1019 | |
| 1020 | ndata->state = DLM_RECO_NODE_DATA_DONE; |
| 1021 | ret = 0; |
| 1022 | break; |
| 1023 | } |
| 1024 | } |
| 1025 | spin_unlock(&dlm_reco_state_lock); |
| 1026 | |
| 1027 | /* wake the recovery thread, some node is done */ |
| 1028 | if (!ret) |
| 1029 | dlm_kick_recovery_thread(dlm); |
| 1030 | |
| 1031 | if (ret < 0) |
| 1032 | mlog(ML_ERROR, "failed to find recovery node data for node " |
| 1033 | "%u\n", done->node_idx); |
| 1034 | dlm_put(dlm); |
| 1035 | |
| 1036 | mlog(0, "leaving reco data done handler, ret=%d\n", ret); |
| 1037 | return ret; |
| 1038 | } |
| 1039 | |
| 1040 | static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm, |
| 1041 | struct list_head *list, |
| 1042 | u8 dead_node) |
| 1043 | { |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 1044 | struct dlm_lock_resource *res, *next; |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 1045 | struct dlm_lock *lock; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1046 | |
| 1047 | spin_lock(&dlm->spinlock); |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 1048 | list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) { |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 1049 | /* always prune any $RECOVERY entries for dead nodes, |
| 1050 | * otherwise hangs can occur during later recovery */ |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1051 | if (dlm_is_recovery_lock(res->lockname.name, |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 1052 | res->lockname.len)) { |
| 1053 | spin_lock(&res->spinlock); |
| 1054 | list_for_each_entry(lock, &res->granted, list) { |
| 1055 | if (lock->ml.node == dead_node) { |
| 1056 | mlog(0, "AHA! there was " |
| 1057 | "a $RECOVERY lock for dead " |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 1058 | "node %u (%s)!\n", |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 1059 | dead_node, dlm->name); |
| 1060 | list_del_init(&lock->list); |
| 1061 | dlm_lock_put(lock); |
| 1062 | break; |
| 1063 | } |
| 1064 | } |
| 1065 | spin_unlock(&res->spinlock); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1066 | continue; |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 1067 | } |
| 1068 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1069 | if (res->owner == dead_node) { |
| 1070 | mlog(0, "found lockres owned by dead node while " |
| 1071 | "doing recovery for node %u. sending it.\n", |
| 1072 | dead_node); |
Akinobu Mita | f116629 | 2006-06-26 00:24:46 -0700 | [diff] [blame] | 1073 | list_move_tail(&res->recovering, list); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1074 | } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) { |
| 1075 | mlog(0, "found UNKNOWN owner while doing recovery " |
| 1076 | "for node %u. sending it.\n", dead_node); |
Akinobu Mita | f116629 | 2006-06-26 00:24:46 -0700 | [diff] [blame] | 1077 | list_move_tail(&res->recovering, list); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1078 | } |
| 1079 | } |
| 1080 | spin_unlock(&dlm->spinlock); |
| 1081 | } |
| 1082 | |
| 1083 | static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res) |
| 1084 | { |
| 1085 | int total_locks = 0; |
| 1086 | struct list_head *iter, *queue = &res->granted; |
| 1087 | int i; |
| 1088 | |
| 1089 | for (i=0; i<3; i++) { |
| 1090 | list_for_each(iter, queue) |
| 1091 | total_locks++; |
| 1092 | queue++; |
| 1093 | } |
| 1094 | return total_locks; |
| 1095 | } |
| 1096 | |
| 1097 | |
| 1098 | static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm, |
| 1099 | struct dlm_migratable_lockres *mres, |
| 1100 | u8 send_to, |
| 1101 | struct dlm_lock_resource *res, |
| 1102 | int total_locks) |
| 1103 | { |
| 1104 | u64 mig_cookie = be64_to_cpu(mres->mig_cookie); |
| 1105 | int mres_total_locks = be32_to_cpu(mres->total_locks); |
| 1106 | int sz, ret = 0, status = 0; |
| 1107 | u8 orig_flags = mres->flags, |
| 1108 | orig_master = mres->master; |
| 1109 | |
| 1110 | BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS); |
| 1111 | if (!mres->num_locks) |
| 1112 | return 0; |
| 1113 | |
| 1114 | sz = sizeof(struct dlm_migratable_lockres) + |
| 1115 | (mres->num_locks * sizeof(struct dlm_migratable_lock)); |
| 1116 | |
| 1117 | /* add an all-done flag if we reached the last lock */ |
| 1118 | orig_flags = mres->flags; |
| 1119 | BUG_ON(total_locks > mres_total_locks); |
| 1120 | if (total_locks == mres_total_locks) |
| 1121 | mres->flags |= DLM_MRES_ALL_DONE; |
| 1122 | |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1123 | mlog(0, "%s:%.*s: sending mig lockres (%s) to %u\n", |
| 1124 | dlm->name, res->lockname.len, res->lockname.name, |
Jeff Liu | 17ae26b | 2009-07-07 15:51:40 +0800 | [diff] [blame] | 1125 | orig_flags & DLM_MRES_MIGRATION ? "migration" : "recovery", |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1126 | send_to); |
| 1127 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1128 | /* send it */ |
| 1129 | ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres, |
| 1130 | sz, send_to, &status); |
| 1131 | if (ret < 0) { |
| 1132 | /* XXX: negative status is not handled. |
| 1133 | * this will end up killing this node. */ |
Sunil Mushran | 8decab3 | 2011-07-24 10:23:54 -0700 | [diff] [blame] | 1134 | mlog(ML_ERROR, "%s: res %.*s, Error %d send MIG_LOCKRES to " |
| 1135 | "node %u (%s)\n", dlm->name, mres->lockname_len, |
| 1136 | mres->lockname, ret, send_to, |
| 1137 | (orig_flags & DLM_MRES_MIGRATION ? |
| 1138 | "migration" : "recovery")); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1139 | } else { |
| 1140 | /* might get an -ENOMEM back here */ |
| 1141 | ret = status; |
| 1142 | if (ret < 0) { |
| 1143 | mlog_errno(ret); |
| 1144 | |
| 1145 | if (ret == -EFAULT) { |
| 1146 | mlog(ML_ERROR, "node %u told me to kill " |
| 1147 | "myself!\n", send_to); |
| 1148 | BUG(); |
| 1149 | } |
| 1150 | } |
| 1151 | } |
| 1152 | |
| 1153 | /* zero and reinit the message buffer */ |
| 1154 | dlm_init_migratable_lockres(mres, res->lockname.name, |
| 1155 | res->lockname.len, mres_total_locks, |
| 1156 | mig_cookie, orig_flags, orig_master); |
| 1157 | return ret; |
| 1158 | } |
| 1159 | |
| 1160 | static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres, |
| 1161 | const char *lockname, int namelen, |
| 1162 | int total_locks, u64 cookie, |
| 1163 | u8 flags, u8 master) |
| 1164 | { |
| 1165 | /* mres here is one full page */ |
Shani Moideen | 5fb0f7f0 | 2007-06-11 09:38:19 +0530 | [diff] [blame] | 1166 | clear_page(mres); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1167 | mres->lockname_len = namelen; |
| 1168 | memcpy(mres->lockname, lockname, namelen); |
| 1169 | mres->num_locks = 0; |
| 1170 | mres->total_locks = cpu_to_be32(total_locks); |
| 1171 | mres->mig_cookie = cpu_to_be64(cookie); |
| 1172 | mres->flags = flags; |
| 1173 | mres->master = master; |
| 1174 | } |
| 1175 | |
Sunil Mushran | 71656fa | 2010-01-25 16:57:39 -0800 | [diff] [blame] | 1176 | static void dlm_prepare_lvb_for_migration(struct dlm_lock *lock, |
| 1177 | struct dlm_migratable_lockres *mres, |
| 1178 | int queue) |
| 1179 | { |
| 1180 | if (!lock->lksb) |
| 1181 | return; |
| 1182 | |
| 1183 | /* Ignore lvb in all locks in the blocked list */ |
| 1184 | if (queue == DLM_BLOCKED_LIST) |
| 1185 | return; |
| 1186 | |
| 1187 | /* Only consider lvbs in locks with granted EX or PR lock levels */ |
| 1188 | if (lock->ml.type != LKM_EXMODE && lock->ml.type != LKM_PRMODE) |
| 1189 | return; |
| 1190 | |
| 1191 | if (dlm_lvb_is_empty(mres->lvb)) { |
| 1192 | memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN); |
| 1193 | return; |
| 1194 | } |
| 1195 | |
| 1196 | /* Ensure the lvb copied for migration matches in other valid locks */ |
| 1197 | if (!memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN)) |
| 1198 | return; |
| 1199 | |
| 1200 | mlog(ML_ERROR, "Mismatched lvb in lock cookie=%u:%llu, name=%.*s, " |
| 1201 | "node=%u\n", |
| 1202 | dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)), |
| 1203 | dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)), |
| 1204 | lock->lockres->lockname.len, lock->lockres->lockname.name, |
| 1205 | lock->ml.node); |
| 1206 | dlm_print_one_lock_resource(lock->lockres); |
| 1207 | BUG(); |
| 1208 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1209 | |
| 1210 | /* returns 1 if this lock fills the network structure, |
| 1211 | * 0 otherwise */ |
| 1212 | static int dlm_add_lock_to_array(struct dlm_lock *lock, |
| 1213 | struct dlm_migratable_lockres *mres, int queue) |
| 1214 | { |
| 1215 | struct dlm_migratable_lock *ml; |
| 1216 | int lock_num = mres->num_locks; |
| 1217 | |
| 1218 | ml = &(mres->ml[lock_num]); |
| 1219 | ml->cookie = lock->ml.cookie; |
| 1220 | ml->type = lock->ml.type; |
| 1221 | ml->convert_type = lock->ml.convert_type; |
| 1222 | ml->highest_blocked = lock->ml.highest_blocked; |
| 1223 | ml->list = queue; |
| 1224 | if (lock->lksb) { |
| 1225 | ml->flags = lock->lksb->flags; |
Sunil Mushran | 71656fa | 2010-01-25 16:57:39 -0800 | [diff] [blame] | 1226 | dlm_prepare_lvb_for_migration(lock, mres, queue); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1227 | } |
| 1228 | ml->node = lock->ml.node; |
| 1229 | mres->num_locks++; |
| 1230 | /* we reached the max, send this network message */ |
| 1231 | if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS) |
| 1232 | return 1; |
| 1233 | return 0; |
| 1234 | } |
| 1235 | |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1236 | static void dlm_add_dummy_lock(struct dlm_ctxt *dlm, |
| 1237 | struct dlm_migratable_lockres *mres) |
| 1238 | { |
| 1239 | struct dlm_lock dummy; |
| 1240 | memset(&dummy, 0, sizeof(dummy)); |
| 1241 | dummy.ml.cookie = 0; |
| 1242 | dummy.ml.type = LKM_IVMODE; |
| 1243 | dummy.ml.convert_type = LKM_IVMODE; |
| 1244 | dummy.ml.highest_blocked = LKM_IVMODE; |
| 1245 | dummy.lksb = NULL; |
| 1246 | dummy.ml.node = dlm->node_num; |
| 1247 | dlm_add_lock_to_array(&dummy, mres, DLM_BLOCKED_LIST); |
| 1248 | } |
| 1249 | |
| 1250 | static inline int dlm_is_dummy_lock(struct dlm_ctxt *dlm, |
| 1251 | struct dlm_migratable_lock *ml, |
| 1252 | u8 *nodenum) |
| 1253 | { |
| 1254 | if (unlikely(ml->cookie == 0 && |
| 1255 | ml->type == LKM_IVMODE && |
| 1256 | ml->convert_type == LKM_IVMODE && |
| 1257 | ml->highest_blocked == LKM_IVMODE && |
| 1258 | ml->list == DLM_BLOCKED_LIST)) { |
| 1259 | *nodenum = ml->node; |
| 1260 | return 1; |
| 1261 | } |
| 1262 | return 0; |
| 1263 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1264 | |
| 1265 | int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, |
| 1266 | struct dlm_migratable_lockres *mres, |
| 1267 | u8 send_to, u8 flags) |
| 1268 | { |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 1269 | struct list_head *queue; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1270 | int total_locks, i; |
| 1271 | u64 mig_cookie = 0; |
| 1272 | struct dlm_lock *lock; |
| 1273 | int ret = 0; |
| 1274 | |
| 1275 | BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION))); |
| 1276 | |
| 1277 | mlog(0, "sending to %u\n", send_to); |
| 1278 | |
| 1279 | total_locks = dlm_num_locks_in_lockres(res); |
| 1280 | if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) { |
| 1281 | /* rare, but possible */ |
| 1282 | mlog(0, "argh. lockres has %d locks. this will " |
| 1283 | "require more than one network packet to " |
| 1284 | "migrate\n", total_locks); |
| 1285 | mig_cookie = dlm_get_next_mig_cookie(); |
| 1286 | } |
| 1287 | |
| 1288 | dlm_init_migratable_lockres(mres, res->lockname.name, |
| 1289 | res->lockname.len, total_locks, |
| 1290 | mig_cookie, flags, res->owner); |
| 1291 | |
| 1292 | total_locks = 0; |
| 1293 | for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) { |
| 1294 | queue = dlm_list_idx_to_ptr(res, i); |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 1295 | list_for_each_entry(lock, queue, list) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1296 | /* add another lock. */ |
| 1297 | total_locks++; |
| 1298 | if (!dlm_add_lock_to_array(lock, mres, i)) |
| 1299 | continue; |
| 1300 | |
| 1301 | /* this filled the lock message, |
| 1302 | * we must send it immediately. */ |
| 1303 | ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, |
| 1304 | res, total_locks); |
Kurt Hackel | 29c0fa0 | 2006-04-27 18:06:58 -0700 | [diff] [blame] | 1305 | if (ret < 0) |
| 1306 | goto error; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1307 | } |
| 1308 | } |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1309 | if (total_locks == 0) { |
| 1310 | /* send a dummy lock to indicate a mastery reference only */ |
| 1311 | mlog(0, "%s:%.*s: sending dummy lock to %u, %s\n", |
| 1312 | dlm->name, res->lockname.len, res->lockname.name, |
| 1313 | send_to, flags & DLM_MRES_RECOVERY ? "recovery" : |
| 1314 | "migration"); |
| 1315 | dlm_add_dummy_lock(dlm, mres); |
| 1316 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1317 | /* flush any remaining locks */ |
| 1318 | ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks); |
Kurt Hackel | 29c0fa0 | 2006-04-27 18:06:58 -0700 | [diff] [blame] | 1319 | if (ret < 0) |
| 1320 | goto error; |
| 1321 | return ret; |
| 1322 | |
| 1323 | error: |
| 1324 | mlog(ML_ERROR, "%s: dlm_send_mig_lockres_msg returned %d\n", |
| 1325 | dlm->name, ret); |
| 1326 | if (!dlm_is_host_down(ret)) |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1327 | BUG(); |
Kurt Hackel | 29c0fa0 | 2006-04-27 18:06:58 -0700 | [diff] [blame] | 1328 | mlog(0, "%s: node %u went down while sending %s " |
| 1329 | "lockres %.*s\n", dlm->name, send_to, |
| 1330 | flags & DLM_MRES_RECOVERY ? "recovery" : "migration", |
| 1331 | res->lockname.len, res->lockname.name); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1332 | return ret; |
| 1333 | } |
| 1334 | |
| 1335 | |
| 1336 | |
| 1337 | /* |
| 1338 | * this message will contain no more than one page worth of |
| 1339 | * recovery data, and it will work on only one lockres. |
| 1340 | * there may be many locks in this page, and we may need to wait |
| 1341 | * for additional packets to complete all the locks (rare, but |
| 1342 | * possible). |
| 1343 | */ |
| 1344 | /* |
| 1345 | * NOTE: the allocation error cases here are scary |
| 1346 | * we really cannot afford to fail an alloc in recovery |
| 1347 | * do we spin? returning an error only delays the problem really |
| 1348 | */ |
| 1349 | |
Kurt Hackel | d74c980 | 2007-01-17 17:04:25 -0800 | [diff] [blame] | 1350 | int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data, |
| 1351 | void **ret_data) |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1352 | { |
| 1353 | struct dlm_ctxt *dlm = data; |
| 1354 | struct dlm_migratable_lockres *mres = |
| 1355 | (struct dlm_migratable_lockres *)msg->buf; |
| 1356 | int ret = 0; |
| 1357 | u8 real_master; |
Sunil Mushran | 52987e2 | 2008-03-01 14:04:21 -0800 | [diff] [blame] | 1358 | u8 extra_refs = 0; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1359 | char *buf = NULL; |
| 1360 | struct dlm_work_item *item = NULL; |
| 1361 | struct dlm_lock_resource *res = NULL; |
| 1362 | |
| 1363 | if (!dlm_grab(dlm)) |
| 1364 | return -EINVAL; |
| 1365 | |
| 1366 | BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION))); |
| 1367 | |
| 1368 | real_master = mres->master; |
| 1369 | if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { |
| 1370 | /* cannot migrate a lockres with no master */ |
| 1371 | BUG_ON(!(mres->flags & DLM_MRES_RECOVERY)); |
| 1372 | } |
| 1373 | |
| 1374 | mlog(0, "%s message received from node %u\n", |
| 1375 | (mres->flags & DLM_MRES_RECOVERY) ? |
| 1376 | "recovery" : "migration", mres->master); |
| 1377 | if (mres->flags & DLM_MRES_ALL_DONE) |
| 1378 | mlog(0, "all done flag. all lockres data received!\n"); |
| 1379 | |
| 1380 | ret = -ENOMEM; |
Kurt Hackel | ad8100e | 2006-05-01 14:25:21 -0700 | [diff] [blame] | 1381 | buf = kmalloc(be16_to_cpu(msg->data_len), GFP_NOFS); |
Robert P. J. Day | cd86128 | 2006-12-13 00:34:52 -0800 | [diff] [blame] | 1382 | item = kzalloc(sizeof(*item), GFP_NOFS); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1383 | if (!buf || !item) |
| 1384 | goto leave; |
| 1385 | |
| 1386 | /* lookup the lock to see if we have a secondary queue for this |
| 1387 | * already... just add the locks in and this will have its owner |
| 1388 | * and RECOVERY flag changed when it completes. */ |
| 1389 | res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len); |
| 1390 | if (res) { |
| 1391 | /* this will get a ref on res */ |
| 1392 | /* mark it as recovering/migrating and hash it */ |
| 1393 | spin_lock(&res->spinlock); |
| 1394 | if (mres->flags & DLM_MRES_RECOVERY) { |
| 1395 | res->state |= DLM_LOCK_RES_RECOVERING; |
| 1396 | } else { |
| 1397 | if (res->state & DLM_LOCK_RES_MIGRATING) { |
| 1398 | /* this is at least the second |
| 1399 | * lockres message */ |
| 1400 | mlog(0, "lock %.*s is already migrating\n", |
| 1401 | mres->lockname_len, |
| 1402 | mres->lockname); |
| 1403 | } else if (res->state & DLM_LOCK_RES_RECOVERING) { |
| 1404 | /* caller should BUG */ |
| 1405 | mlog(ML_ERROR, "node is attempting to migrate " |
| 1406 | "lock %.*s, but marked as recovering!\n", |
| 1407 | mres->lockname_len, mres->lockname); |
| 1408 | ret = -EFAULT; |
| 1409 | spin_unlock(&res->spinlock); |
Xue jiufei | 27749f2 | 2013-06-12 14:05:03 -0700 | [diff] [blame] | 1410 | dlm_lockres_put(res); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1411 | goto leave; |
| 1412 | } |
| 1413 | res->state |= DLM_LOCK_RES_MIGRATING; |
| 1414 | } |
| 1415 | spin_unlock(&res->spinlock); |
| 1416 | } else { |
| 1417 | /* need to allocate, just like if it was |
| 1418 | * mastered here normally */ |
| 1419 | res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len); |
| 1420 | if (!res) |
| 1421 | goto leave; |
| 1422 | |
| 1423 | /* to match the ref that we would have gotten if |
| 1424 | * dlm_lookup_lockres had succeeded */ |
| 1425 | dlm_lockres_get(res); |
| 1426 | |
| 1427 | /* mark it as recovering/migrating and hash it */ |
| 1428 | if (mres->flags & DLM_MRES_RECOVERY) |
| 1429 | res->state |= DLM_LOCK_RES_RECOVERING; |
| 1430 | else |
| 1431 | res->state |= DLM_LOCK_RES_MIGRATING; |
| 1432 | |
| 1433 | spin_lock(&dlm->spinlock); |
| 1434 | __dlm_insert_lockres(dlm, res); |
| 1435 | spin_unlock(&dlm->spinlock); |
| 1436 | |
Sunil Mushran | 52987e2 | 2008-03-01 14:04:21 -0800 | [diff] [blame] | 1437 | /* Add an extra ref for this lock-less lockres lest the |
| 1438 | * dlm_thread purges it before we get the chance to add |
| 1439 | * locks to it */ |
| 1440 | dlm_lockres_get(res); |
| 1441 | |
| 1442 | /* There are three refs that need to be put. |
| 1443 | * 1. Taken above. |
| 1444 | * 2. kref_init in dlm_new_lockres()->dlm_init_lockres(). |
| 1445 | * 3. dlm_lookup_lockres() |
| 1446 | * The first one is handled at the end of this function. The |
| 1447 | * other two are handled in the worker thread after locks have |
| 1448 | * been attached. Yes, we don't wait for purge time to match |
| 1449 | * kref_init. The lockres will still have atleast one ref |
| 1450 | * added because it is in the hash __dlm_insert_lockres() */ |
| 1451 | extra_refs++; |
| 1452 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1453 | /* now that the new lockres is inserted, |
| 1454 | * make it usable by other processes */ |
| 1455 | spin_lock(&res->spinlock); |
| 1456 | res->state &= ~DLM_LOCK_RES_IN_PROGRESS; |
| 1457 | spin_unlock(&res->spinlock); |
Kurt Hackel | a6fa364 | 2007-01-17 14:59:12 -0800 | [diff] [blame] | 1458 | wake_up(&res->wq); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1459 | } |
| 1460 | |
| 1461 | /* at this point we have allocated everything we need, |
| 1462 | * and we have a hashed lockres with an extra ref and |
| 1463 | * the proper res->state flags. */ |
| 1464 | ret = 0; |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1465 | spin_lock(&res->spinlock); |
| 1466 | /* drop this either when master requery finds a different master |
| 1467 | * or when a lock is added by the recovery worker */ |
| 1468 | dlm_lockres_grab_inflight_ref(dlm, res); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1469 | if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) { |
| 1470 | /* migration cannot have an unknown master */ |
| 1471 | BUG_ON(!(mres->flags & DLM_MRES_RECOVERY)); |
| 1472 | mlog(0, "recovery has passed me a lockres with an " |
| 1473 | "unknown owner.. will need to requery: " |
| 1474 | "%.*s\n", mres->lockname_len, mres->lockname); |
| 1475 | } else { |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1476 | /* take a reference now to pin the lockres, drop it |
| 1477 | * when locks are added in the worker */ |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1478 | dlm_change_lockres_owner(dlm, res, dlm->node_num); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1479 | } |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1480 | spin_unlock(&res->spinlock); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1481 | |
| 1482 | /* queue up work for dlm_mig_lockres_worker */ |
| 1483 | dlm_grab(dlm); /* get an extra ref for the work item */ |
| 1484 | memcpy(buf, msg->buf, be16_to_cpu(msg->data_len)); /* copy the whole message */ |
| 1485 | dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf); |
| 1486 | item->u.ml.lockres = res; /* already have a ref */ |
| 1487 | item->u.ml.real_master = real_master; |
Sunil Mushran | 52987e2 | 2008-03-01 14:04:21 -0800 | [diff] [blame] | 1488 | item->u.ml.extra_ref = extra_refs; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1489 | spin_lock(&dlm->work_lock); |
| 1490 | list_add_tail(&item->list, &dlm->work_list); |
| 1491 | spin_unlock(&dlm->work_lock); |
Kurt Hackel | 3156d26 | 2006-05-01 14:39:29 -0700 | [diff] [blame] | 1492 | queue_work(dlm->dlm_worker, &dlm->dispatched_work); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1493 | |
| 1494 | leave: |
Sunil Mushran | 52987e2 | 2008-03-01 14:04:21 -0800 | [diff] [blame] | 1495 | /* One extra ref taken needs to be put here */ |
| 1496 | if (extra_refs) |
| 1497 | dlm_lockres_put(res); |
| 1498 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1499 | dlm_put(dlm); |
| 1500 | if (ret < 0) { |
Sachin Kamat | 7cfa74d | 2013-04-29 15:06:00 -0700 | [diff] [blame] | 1501 | kfree(buf); |
| 1502 | kfree(item); |
Tao Ma | c1e8d35 | 2011-03-07 16:43:21 +0800 | [diff] [blame] | 1503 | mlog_errno(ret); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1504 | } |
| 1505 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1506 | return ret; |
| 1507 | } |
| 1508 | |
| 1509 | |
| 1510 | static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data) |
| 1511 | { |
Sunil Mushran | 52987e2 | 2008-03-01 14:04:21 -0800 | [diff] [blame] | 1512 | struct dlm_ctxt *dlm; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1513 | struct dlm_migratable_lockres *mres; |
| 1514 | int ret = 0; |
| 1515 | struct dlm_lock_resource *res; |
| 1516 | u8 real_master; |
Sunil Mushran | 52987e2 | 2008-03-01 14:04:21 -0800 | [diff] [blame] | 1517 | u8 extra_ref; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1518 | |
| 1519 | dlm = item->dlm; |
| 1520 | mres = (struct dlm_migratable_lockres *)data; |
| 1521 | |
| 1522 | res = item->u.ml.lockres; |
| 1523 | real_master = item->u.ml.real_master; |
Sunil Mushran | 52987e2 | 2008-03-01 14:04:21 -0800 | [diff] [blame] | 1524 | extra_ref = item->u.ml.extra_ref; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1525 | |
| 1526 | if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { |
| 1527 | /* this case is super-rare. only occurs if |
| 1528 | * node death happens during migration. */ |
| 1529 | again: |
| 1530 | ret = dlm_lockres_master_requery(dlm, res, &real_master); |
| 1531 | if (ret < 0) { |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 1532 | mlog(0, "dlm_lockres_master_requery ret=%d\n", |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1533 | ret); |
| 1534 | goto again; |
| 1535 | } |
| 1536 | if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) { |
| 1537 | mlog(0, "lockres %.*s not claimed. " |
| 1538 | "this node will take it.\n", |
| 1539 | res->lockname.len, res->lockname.name); |
| 1540 | } else { |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1541 | spin_lock(&res->spinlock); |
| 1542 | dlm_lockres_drop_inflight_ref(dlm, res); |
| 1543 | spin_unlock(&res->spinlock); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1544 | mlog(0, "master needs to respond to sender " |
| 1545 | "that node %u still owns %.*s\n", |
| 1546 | real_master, res->lockname.len, |
| 1547 | res->lockname.name); |
| 1548 | /* cannot touch this lockres */ |
| 1549 | goto leave; |
| 1550 | } |
| 1551 | } |
| 1552 | |
| 1553 | ret = dlm_process_recovery_data(dlm, res, mres); |
| 1554 | if (ret < 0) |
| 1555 | mlog(0, "dlm_process_recovery_data returned %d\n", ret); |
| 1556 | else |
| 1557 | mlog(0, "dlm_process_recovery_data succeeded\n"); |
| 1558 | |
| 1559 | if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) == |
| 1560 | (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) { |
| 1561 | ret = dlm_finish_migration(dlm, res, mres->master); |
| 1562 | if (ret < 0) |
| 1563 | mlog_errno(ret); |
| 1564 | } |
| 1565 | |
| 1566 | leave: |
Sunil Mushran | 52987e2 | 2008-03-01 14:04:21 -0800 | [diff] [blame] | 1567 | /* See comment in dlm_mig_lockres_handler() */ |
| 1568 | if (res) { |
| 1569 | if (extra_ref) |
| 1570 | dlm_lockres_put(res); |
| 1571 | dlm_lockres_put(res); |
| 1572 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1573 | kfree(data); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1574 | } |
| 1575 | |
| 1576 | |
| 1577 | |
Adrian Bunk | 8169cae | 2006-03-31 16:53:55 +0200 | [diff] [blame] | 1578 | static int dlm_lockres_master_requery(struct dlm_ctxt *dlm, |
| 1579 | struct dlm_lock_resource *res, |
| 1580 | u8 *real_master) |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1581 | { |
| 1582 | struct dlm_node_iter iter; |
| 1583 | int nodenum; |
| 1584 | int ret = 0; |
| 1585 | |
| 1586 | *real_master = DLM_LOCK_RES_OWNER_UNKNOWN; |
| 1587 | |
| 1588 | /* we only reach here if one of the two nodes in a |
| 1589 | * migration died while the migration was in progress. |
| 1590 | * at this point we need to requery the master. we |
| 1591 | * know that the new_master got as far as creating |
| 1592 | * an mle on at least one node, but we do not know |
| 1593 | * if any nodes had actually cleared the mle and set |
| 1594 | * the master to the new_master. the old master |
| 1595 | * is supposed to set the owner to UNKNOWN in the |
| 1596 | * event of a new_master death, so the only possible |
| 1597 | * responses that we can get from nodes here are |
| 1598 | * that the master is new_master, or that the master |
| 1599 | * is UNKNOWN. |
| 1600 | * if all nodes come back with UNKNOWN then we know |
| 1601 | * the lock needs remastering here. |
| 1602 | * if any node comes back with a valid master, check |
| 1603 | * to see if that master is the one that we are |
| 1604 | * recovering. if so, then the new_master died and |
| 1605 | * we need to remaster this lock. if not, then the |
| 1606 | * new_master survived and that node will respond to |
| 1607 | * other nodes about the owner. |
| 1608 | * if there is an owner, this node needs to dump this |
| 1609 | * lockres and alert the sender that this lockres |
| 1610 | * was rejected. */ |
| 1611 | spin_lock(&dlm->spinlock); |
| 1612 | dlm_node_iter_init(dlm->domain_map, &iter); |
| 1613 | spin_unlock(&dlm->spinlock); |
| 1614 | |
| 1615 | while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { |
| 1616 | /* do not send to self */ |
| 1617 | if (nodenum == dlm->node_num) |
| 1618 | continue; |
| 1619 | ret = dlm_do_master_requery(dlm, res, nodenum, real_master); |
| 1620 | if (ret < 0) { |
| 1621 | mlog_errno(ret); |
Kurt Hackel | c03872f | 2006-03-06 14:08:49 -0800 | [diff] [blame] | 1622 | if (!dlm_is_host_down(ret)) |
| 1623 | BUG(); |
| 1624 | /* host is down, so answer for that node would be |
| 1625 | * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */ |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1626 | } |
| 1627 | if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) { |
| 1628 | mlog(0, "lock master is %u\n", *real_master); |
| 1629 | break; |
| 1630 | } |
| 1631 | } |
| 1632 | return ret; |
| 1633 | } |
| 1634 | |
| 1635 | |
Kurt Hackel | c03872f | 2006-03-06 14:08:49 -0800 | [diff] [blame] | 1636 | int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res, |
| 1637 | u8 nodenum, u8 *real_master) |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1638 | { |
| 1639 | int ret = -EINVAL; |
| 1640 | struct dlm_master_requery req; |
| 1641 | int status = DLM_LOCK_RES_OWNER_UNKNOWN; |
| 1642 | |
| 1643 | memset(&req, 0, sizeof(req)); |
| 1644 | req.node_idx = dlm->node_num; |
| 1645 | req.namelen = res->lockname.len; |
| 1646 | memcpy(req.name, res->lockname.name, res->lockname.len); |
| 1647 | |
| 1648 | ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key, |
| 1649 | &req, sizeof(req), nodenum, &status); |
| 1650 | /* XXX: negative status not handled properly here. */ |
| 1651 | if (ret < 0) |
Wengang Wang | a5196ec | 2010-03-30 12:09:22 +0800 | [diff] [blame] | 1652 | mlog(ML_ERROR, "Error %d when sending message %u (key " |
| 1653 | "0x%x) to node %u\n", ret, DLM_MASTER_REQUERY_MSG, |
| 1654 | dlm->key, nodenum); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1655 | else { |
| 1656 | BUG_ON(status < 0); |
| 1657 | BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN); |
| 1658 | *real_master = (u8) (status & 0xff); |
| 1659 | mlog(0, "node %u responded to master requery with %u\n", |
| 1660 | nodenum, *real_master); |
| 1661 | ret = 0; |
| 1662 | } |
| 1663 | return ret; |
| 1664 | } |
| 1665 | |
| 1666 | |
| 1667 | /* this function cannot error, so unless the sending |
| 1668 | * or receiving of the message failed, the owner can |
| 1669 | * be trusted */ |
Kurt Hackel | d74c980 | 2007-01-17 17:04:25 -0800 | [diff] [blame] | 1670 | int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data, |
| 1671 | void **ret_data) |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1672 | { |
| 1673 | struct dlm_ctxt *dlm = data; |
| 1674 | struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf; |
| 1675 | struct dlm_lock_resource *res = NULL; |
Mark Fasheh | a3d3329 | 2006-03-09 17:55:56 -0800 | [diff] [blame] | 1676 | unsigned int hash; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1677 | int master = DLM_LOCK_RES_OWNER_UNKNOWN; |
| 1678 | u32 flags = DLM_ASSERT_MASTER_REQUERY; |
| 1679 | |
| 1680 | if (!dlm_grab(dlm)) { |
| 1681 | /* since the domain has gone away on this |
| 1682 | * node, the proper response is UNKNOWN */ |
| 1683 | return master; |
| 1684 | } |
| 1685 | |
Mark Fasheh | a3d3329 | 2006-03-09 17:55:56 -0800 | [diff] [blame] | 1686 | hash = dlm_lockid_hash(req->name, req->namelen); |
| 1687 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1688 | spin_lock(&dlm->spinlock); |
Mark Fasheh | a3d3329 | 2006-03-09 17:55:56 -0800 | [diff] [blame] | 1689 | res = __dlm_lookup_lockres(dlm, req->name, req->namelen, hash); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1690 | if (res) { |
| 1691 | spin_lock(&res->spinlock); |
| 1692 | master = res->owner; |
| 1693 | if (master == dlm->node_num) { |
| 1694 | int ret = dlm_dispatch_assert_master(dlm, res, |
| 1695 | 0, 0, flags); |
| 1696 | if (ret < 0) { |
| 1697 | mlog_errno(-ENOMEM); |
| 1698 | /* retry!? */ |
| 1699 | BUG(); |
| 1700 | } |
Sunil Mushran | 52987e2 | 2008-03-01 14:04:21 -0800 | [diff] [blame] | 1701 | } else /* put.. incase we are not the master */ |
| 1702 | dlm_lockres_put(res); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1703 | spin_unlock(&res->spinlock); |
| 1704 | } |
| 1705 | spin_unlock(&dlm->spinlock); |
| 1706 | |
| 1707 | dlm_put(dlm); |
| 1708 | return master; |
| 1709 | } |
| 1710 | |
| 1711 | static inline struct list_head * |
| 1712 | dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num) |
| 1713 | { |
| 1714 | struct list_head *ret; |
| 1715 | BUG_ON(list_num < 0); |
| 1716 | BUG_ON(list_num > 2); |
| 1717 | ret = &(res->granted); |
| 1718 | ret += list_num; |
| 1719 | return ret; |
| 1720 | } |
| 1721 | /* TODO: do ast flush business |
| 1722 | * TODO: do MIGRATING and RECOVERING spinning |
| 1723 | */ |
| 1724 | |
| 1725 | /* |
| 1726 | * NOTE about in-flight requests during migration: |
| 1727 | * |
| 1728 | * Before attempting the migrate, the master has marked the lockres as |
| 1729 | * MIGRATING and then flushed all of its pending ASTS. So any in-flight |
| 1730 | * requests either got queued before the MIGRATING flag got set, in which |
| 1731 | * case the lock data will reflect the change and a return message is on |
| 1732 | * the way, or the request failed to get in before MIGRATING got set. In |
| 1733 | * this case, the caller will be told to spin and wait for the MIGRATING |
| 1734 | * flag to be dropped, then recheck the master. |
| 1735 | * This holds true for the convert, cancel and unlock cases, and since lvb |
| 1736 | * updates are tied to these same messages, it applies to lvb updates as |
| 1737 | * well. For the lock case, there is no way a lock can be on the master |
| 1738 | * queue and not be on the secondary queue since the lock is always added |
| 1739 | * locally first. This means that the new target node will never be sent |
| 1740 | * a lock that he doesn't already have on the list. |
| 1741 | * In total, this means that the local lock is correct and should not be |
| 1742 | * updated to match the one sent by the master. Any messages sent back |
| 1743 | * from the master before the MIGRATING flag will bring the lock properly |
| 1744 | * up-to-date, and the change will be ordered properly for the waiter. |
| 1745 | * We will *not* attempt to modify the lock underneath the waiter. |
| 1746 | */ |
| 1747 | |
| 1748 | static int dlm_process_recovery_data(struct dlm_ctxt *dlm, |
| 1749 | struct dlm_lock_resource *res, |
| 1750 | struct dlm_migratable_lockres *mres) |
| 1751 | { |
| 1752 | struct dlm_migratable_lock *ml; |
Junxiao Bi | 34aa8da | 2014-04-03 14:46:49 -0700 | [diff] [blame^] | 1753 | struct list_head *queue, *iter; |
Kurt Hackel | e17e75e | 2007-01-05 15:04:49 -0800 | [diff] [blame] | 1754 | struct list_head *tmpq = NULL; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1755 | struct dlm_lock *newlock = NULL; |
| 1756 | struct dlm_lockstatus *lksb = NULL; |
| 1757 | int ret = 0; |
Kurt Hackel | e17e75e | 2007-01-05 15:04:49 -0800 | [diff] [blame] | 1758 | int i, j, bad; |
Junxiao Bi | 34aa8da | 2014-04-03 14:46:49 -0700 | [diff] [blame^] | 1759 | struct dlm_lock *lock; |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1760 | u8 from = O2NM_MAX_NODES; |
| 1761 | unsigned int added = 0; |
Sunil Mushran | 26636bf | 2010-01-25 16:57:40 -0800 | [diff] [blame] | 1762 | __be64 c; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1763 | |
| 1764 | mlog(0, "running %d locks for this lockres\n", mres->num_locks); |
| 1765 | for (i=0; i<mres->num_locks; i++) { |
| 1766 | ml = &(mres->ml[i]); |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1767 | |
| 1768 | if (dlm_is_dummy_lock(dlm, ml, &from)) { |
| 1769 | /* placeholder, just need to set the refmap bit */ |
| 1770 | BUG_ON(mres->num_locks != 1); |
| 1771 | mlog(0, "%s:%.*s: dummy lock for %u\n", |
| 1772 | dlm->name, mres->lockname_len, mres->lockname, |
| 1773 | from); |
| 1774 | spin_lock(&res->spinlock); |
Sunil Mushran | 8d400b8 | 2011-07-24 10:26:54 -0700 | [diff] [blame] | 1775 | dlm_lockres_set_refmap_bit(dlm, res, from); |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1776 | spin_unlock(&res->spinlock); |
| 1777 | added++; |
| 1778 | break; |
| 1779 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1780 | BUG_ON(ml->highest_blocked != LKM_IVMODE); |
| 1781 | newlock = NULL; |
| 1782 | lksb = NULL; |
| 1783 | |
| 1784 | queue = dlm_list_num_to_pointer(res, ml->list); |
Kurt Hackel | e17e75e | 2007-01-05 15:04:49 -0800 | [diff] [blame] | 1785 | tmpq = NULL; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1786 | |
| 1787 | /* if the lock is for the local node it needs to |
| 1788 | * be moved to the proper location within the queue. |
| 1789 | * do not allocate a new lock structure. */ |
| 1790 | if (ml->node == dlm->node_num) { |
| 1791 | /* MIGRATION ONLY! */ |
| 1792 | BUG_ON(!(mres->flags & DLM_MRES_MIGRATION)); |
| 1793 | |
Junxiao Bi | 34aa8da | 2014-04-03 14:46:49 -0700 | [diff] [blame^] | 1794 | lock = NULL; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1795 | spin_lock(&res->spinlock); |
Kurt Hackel | e17e75e | 2007-01-05 15:04:49 -0800 | [diff] [blame] | 1796 | for (j = DLM_GRANTED_LIST; j <= DLM_BLOCKED_LIST; j++) { |
| 1797 | tmpq = dlm_list_idx_to_ptr(res, j); |
Junxiao Bi | 34aa8da | 2014-04-03 14:46:49 -0700 | [diff] [blame^] | 1798 | list_for_each(iter, tmpq) { |
| 1799 | lock = list_entry(iter, |
| 1800 | struct dlm_lock, list); |
| 1801 | if (lock->ml.cookie == ml->cookie) |
Kurt Hackel | e17e75e | 2007-01-05 15:04:49 -0800 | [diff] [blame] | 1802 | break; |
Junxiao Bi | 34aa8da | 2014-04-03 14:46:49 -0700 | [diff] [blame^] | 1803 | lock = NULL; |
Kurt Hackel | e17e75e | 2007-01-05 15:04:49 -0800 | [diff] [blame] | 1804 | } |
| 1805 | if (lock) |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1806 | break; |
| 1807 | } |
| 1808 | |
| 1809 | /* lock is always created locally first, and |
| 1810 | * destroyed locally last. it must be on the list */ |
| 1811 | if (!lock) { |
Sunil Mushran | 26636bf | 2010-01-25 16:57:40 -0800 | [diff] [blame] | 1812 | c = ml->cookie; |
| 1813 | mlog(ML_ERROR, "Could not find local lock " |
| 1814 | "with cookie %u:%llu, node %u, " |
| 1815 | "list %u, flags 0x%x, type %d, " |
| 1816 | "conv %d, highest blocked %d\n", |
Kurt Hackel | 74aa258 | 2007-01-17 15:11:36 -0800 | [diff] [blame] | 1817 | dlm_get_lock_cookie_node(be64_to_cpu(c)), |
Sunil Mushran | 26636bf | 2010-01-25 16:57:40 -0800 | [diff] [blame] | 1818 | dlm_get_lock_cookie_seq(be64_to_cpu(c)), |
| 1819 | ml->node, ml->list, ml->flags, ml->type, |
| 1820 | ml->convert_type, ml->highest_blocked); |
Kurt Hackel | 71ac106 | 2007-01-05 15:02:30 -0800 | [diff] [blame] | 1821 | __dlm_print_one_lock_resource(res); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1822 | BUG(); |
| 1823 | } |
Sunil Mushran | 26636bf | 2010-01-25 16:57:40 -0800 | [diff] [blame] | 1824 | |
| 1825 | if (lock->ml.node != ml->node) { |
| 1826 | c = lock->ml.cookie; |
| 1827 | mlog(ML_ERROR, "Mismatched node# in lock " |
| 1828 | "cookie %u:%llu, name %.*s, node %u\n", |
| 1829 | dlm_get_lock_cookie_node(be64_to_cpu(c)), |
| 1830 | dlm_get_lock_cookie_seq(be64_to_cpu(c)), |
| 1831 | res->lockname.len, res->lockname.name, |
| 1832 | lock->ml.node); |
| 1833 | c = ml->cookie; |
| 1834 | mlog(ML_ERROR, "Migrate lock cookie %u:%llu, " |
| 1835 | "node %u, list %u, flags 0x%x, type %d, " |
| 1836 | "conv %d, highest blocked %d\n", |
| 1837 | dlm_get_lock_cookie_node(be64_to_cpu(c)), |
| 1838 | dlm_get_lock_cookie_seq(be64_to_cpu(c)), |
| 1839 | ml->node, ml->list, ml->flags, ml->type, |
| 1840 | ml->convert_type, ml->highest_blocked); |
| 1841 | __dlm_print_one_lock_resource(res); |
| 1842 | BUG(); |
| 1843 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1844 | |
Kurt Hackel | e17e75e | 2007-01-05 15:04:49 -0800 | [diff] [blame] | 1845 | if (tmpq != queue) { |
Sunil Mushran | 26636bf | 2010-01-25 16:57:40 -0800 | [diff] [blame] | 1846 | c = ml->cookie; |
| 1847 | mlog(0, "Lock cookie %u:%llu was on list %u " |
| 1848 | "instead of list %u for %.*s\n", |
| 1849 | dlm_get_lock_cookie_node(be64_to_cpu(c)), |
| 1850 | dlm_get_lock_cookie_seq(be64_to_cpu(c)), |
| 1851 | j, ml->list, res->lockname.len, |
| 1852 | res->lockname.name); |
| 1853 | __dlm_print_one_lock_resource(res); |
Kurt Hackel | e17e75e | 2007-01-05 15:04:49 -0800 | [diff] [blame] | 1854 | spin_unlock(&res->spinlock); |
| 1855 | continue; |
| 1856 | } |
| 1857 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1858 | /* see NOTE above about why we do not update |
| 1859 | * to match the master here */ |
| 1860 | |
| 1861 | /* move the lock to its proper place */ |
| 1862 | /* do not alter lock refcount. switching lists. */ |
Akinobu Mita | f116629 | 2006-06-26 00:24:46 -0700 | [diff] [blame] | 1863 | list_move_tail(&lock->list, queue); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1864 | spin_unlock(&res->spinlock); |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1865 | added++; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1866 | |
| 1867 | mlog(0, "just reordered a local lock!\n"); |
| 1868 | continue; |
| 1869 | } |
| 1870 | |
| 1871 | /* lock is for another node. */ |
| 1872 | newlock = dlm_new_lock(ml->type, ml->node, |
| 1873 | be64_to_cpu(ml->cookie), NULL); |
| 1874 | if (!newlock) { |
| 1875 | ret = -ENOMEM; |
| 1876 | goto leave; |
| 1877 | } |
| 1878 | lksb = newlock->lksb; |
| 1879 | dlm_lock_attach_lockres(newlock, res); |
| 1880 | |
| 1881 | if (ml->convert_type != LKM_IVMODE) { |
| 1882 | BUG_ON(queue != &res->converting); |
| 1883 | newlock->ml.convert_type = ml->convert_type; |
| 1884 | } |
| 1885 | lksb->flags |= (ml->flags & |
| 1886 | (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB)); |
Kurt Hackel | ccd8b1f | 2006-05-01 11:32:14 -0700 | [diff] [blame] | 1887 | |
| 1888 | if (ml->type == LKM_NLMODE) |
| 1889 | goto skip_lvb; |
| 1890 | |
Xue jiufei | 750e3c6 | 2013-11-12 15:06:59 -0800 | [diff] [blame] | 1891 | /* |
| 1892 | * If the lock is in the blocked list it can't have a valid lvb, |
| 1893 | * so skip it |
| 1894 | */ |
| 1895 | if (ml->list == DLM_BLOCKED_LIST) |
| 1896 | goto skip_lvb; |
| 1897 | |
Kurt Hackel | 8bc674c | 2006-04-27 18:02:10 -0700 | [diff] [blame] | 1898 | if (!dlm_lvb_is_empty(mres->lvb)) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1899 | if (lksb->flags & DLM_LKSB_PUT_LVB) { |
| 1900 | /* other node was trying to update |
| 1901 | * lvb when node died. recreate the |
| 1902 | * lksb with the updated lvb. */ |
| 1903 | memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN); |
Kurt Hackel | ccd8b1f | 2006-05-01 11:32:14 -0700 | [diff] [blame] | 1904 | /* the lock resource lvb update must happen |
| 1905 | * NOW, before the spinlock is dropped. |
| 1906 | * we no longer wait for the AST to update |
| 1907 | * the lvb. */ |
| 1908 | memcpy(res->lvb, mres->lvb, DLM_LVB_LEN); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1909 | } else { |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 1910 | /* otherwise, the node is sending its |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1911 | * most recent valid lvb info */ |
| 1912 | BUG_ON(ml->type != LKM_EXMODE && |
| 1913 | ml->type != LKM_PRMODE); |
Kurt Hackel | 8bc674c | 2006-04-27 18:02:10 -0700 | [diff] [blame] | 1914 | if (!dlm_lvb_is_empty(res->lvb) && |
Kurt Hackel | ccd8b1f | 2006-05-01 11:32:14 -0700 | [diff] [blame] | 1915 | (ml->type == LKM_EXMODE || |
| 1916 | memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) { |
| 1917 | int i; |
| 1918 | mlog(ML_ERROR, "%s:%.*s: received bad " |
| 1919 | "lvb! type=%d\n", dlm->name, |
| 1920 | res->lockname.len, |
| 1921 | res->lockname.name, ml->type); |
| 1922 | printk("lockres lvb=["); |
| 1923 | for (i=0; i<DLM_LVB_LEN; i++) |
| 1924 | printk("%02x", res->lvb[i]); |
| 1925 | printk("]\nmigrated lvb=["); |
| 1926 | for (i=0; i<DLM_LVB_LEN; i++) |
| 1927 | printk("%02x", mres->lvb[i]); |
| 1928 | printk("]\n"); |
| 1929 | dlm_print_one_lock_resource(res); |
| 1930 | BUG(); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1931 | } |
| 1932 | memcpy(res->lvb, mres->lvb, DLM_LVB_LEN); |
| 1933 | } |
| 1934 | } |
Kurt Hackel | ccd8b1f | 2006-05-01 11:32:14 -0700 | [diff] [blame] | 1935 | skip_lvb: |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1936 | |
| 1937 | /* NOTE: |
| 1938 | * wrt lock queue ordering and recovery: |
| 1939 | * 1. order of locks on granted queue is |
| 1940 | * meaningless. |
| 1941 | * 2. order of locks on converting queue is |
| 1942 | * LOST with the node death. sorry charlie. |
| 1943 | * 3. order of locks on the blocked queue is |
| 1944 | * also LOST. |
| 1945 | * order of locks does not affect integrity, it |
| 1946 | * just means that a lock request may get pushed |
| 1947 | * back in line as a result of the node death. |
| 1948 | * also note that for a given node the lock order |
| 1949 | * for its secondary queue locks is preserved |
| 1950 | * relative to each other, but clearly *not* |
| 1951 | * preserved relative to locks from other nodes. |
| 1952 | */ |
Kurt Hackel | c3187ce | 2006-04-27 18:05:41 -0700 | [diff] [blame] | 1953 | bad = 0; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1954 | spin_lock(&res->spinlock); |
Kurt Hackel | c3187ce | 2006-04-27 18:05:41 -0700 | [diff] [blame] | 1955 | list_for_each_entry(lock, queue, list) { |
| 1956 | if (lock->ml.cookie == ml->cookie) { |
Sunil Mushran | 26636bf | 2010-01-25 16:57:40 -0800 | [diff] [blame] | 1957 | c = lock->ml.cookie; |
Kurt Hackel | c3187ce | 2006-04-27 18:05:41 -0700 | [diff] [blame] | 1958 | mlog(ML_ERROR, "%s:%.*s: %u:%llu: lock already " |
| 1959 | "exists on this lockres!\n", dlm->name, |
| 1960 | res->lockname.len, res->lockname.name, |
Kurt Hackel | 74aa258 | 2007-01-17 15:11:36 -0800 | [diff] [blame] | 1961 | dlm_get_lock_cookie_node(be64_to_cpu(c)), |
| 1962 | dlm_get_lock_cookie_seq(be64_to_cpu(c))); |
Kurt Hackel | c3187ce | 2006-04-27 18:05:41 -0700 | [diff] [blame] | 1963 | |
| 1964 | mlog(ML_NOTICE, "sent lock: type=%d, conv=%d, " |
| 1965 | "node=%u, cookie=%u:%llu, queue=%d\n", |
| 1966 | ml->type, ml->convert_type, ml->node, |
Kurt Hackel | 74aa258 | 2007-01-17 15:11:36 -0800 | [diff] [blame] | 1967 | dlm_get_lock_cookie_node(be64_to_cpu(ml->cookie)), |
| 1968 | dlm_get_lock_cookie_seq(be64_to_cpu(ml->cookie)), |
Kurt Hackel | c3187ce | 2006-04-27 18:05:41 -0700 | [diff] [blame] | 1969 | ml->list); |
| 1970 | |
| 1971 | __dlm_print_one_lock_resource(res); |
| 1972 | bad = 1; |
| 1973 | break; |
| 1974 | } |
| 1975 | } |
| 1976 | if (!bad) { |
| 1977 | dlm_lock_get(newlock); |
| 1978 | list_add_tail(&newlock->list, queue); |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1979 | mlog(0, "%s:%.*s: added lock for node %u, " |
| 1980 | "setting refmap bit\n", dlm->name, |
| 1981 | res->lockname.len, res->lockname.name, ml->node); |
Sunil Mushran | 8d400b8 | 2011-07-24 10:26:54 -0700 | [diff] [blame] | 1982 | dlm_lockres_set_refmap_bit(dlm, res, ml->node); |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1983 | added++; |
Kurt Hackel | c3187ce | 2006-04-27 18:05:41 -0700 | [diff] [blame] | 1984 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1985 | spin_unlock(&res->spinlock); |
| 1986 | } |
| 1987 | mlog(0, "done running all the locks\n"); |
| 1988 | |
| 1989 | leave: |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1990 | /* balance the ref taken when the work was queued */ |
Kurt Hackel | 50635f1 | 2007-01-17 14:54:39 -0800 | [diff] [blame] | 1991 | spin_lock(&res->spinlock); |
| 1992 | dlm_lockres_drop_inflight_ref(dlm, res); |
| 1993 | spin_unlock(&res->spinlock); |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 1994 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 1995 | if (ret < 0) { |
| 1996 | mlog_errno(ret); |
| 1997 | if (newlock) |
| 1998 | dlm_lock_put(newlock); |
| 1999 | } |
| 2000 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2001 | return ret; |
| 2002 | } |
| 2003 | |
| 2004 | void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm, |
| 2005 | struct dlm_lock_resource *res) |
| 2006 | { |
| 2007 | int i; |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 2008 | struct list_head *queue; |
| 2009 | struct dlm_lock *lock, *next; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2010 | |
Wengang Wang | a524812 | 2010-07-30 16:14:44 +0800 | [diff] [blame] | 2011 | assert_spin_locked(&dlm->spinlock); |
| 2012 | assert_spin_locked(&res->spinlock); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2013 | res->state |= DLM_LOCK_RES_RECOVERING; |
Kurt Hackel | 69d72b0 | 2006-05-01 10:57:51 -0700 | [diff] [blame] | 2014 | if (!list_empty(&res->recovering)) { |
| 2015 | mlog(0, |
| 2016 | "Recovering res %s:%.*s, is already on recovery list!\n", |
| 2017 | dlm->name, res->lockname.len, res->lockname.name); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2018 | list_del_init(&res->recovering); |
Sunil Mushran | 52987e2 | 2008-03-01 14:04:21 -0800 | [diff] [blame] | 2019 | dlm_lockres_put(res); |
Kurt Hackel | 69d72b0 | 2006-05-01 10:57:51 -0700 | [diff] [blame] | 2020 | } |
| 2021 | /* We need to hold a reference while on the recovery list */ |
| 2022 | dlm_lockres_get(res); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2023 | list_add_tail(&res->recovering, &dlm->reco.resources); |
| 2024 | |
| 2025 | /* find any pending locks and put them back on proper list */ |
| 2026 | for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) { |
| 2027 | queue = dlm_list_idx_to_ptr(res, i); |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 2028 | list_for_each_entry_safe(lock, next, queue, list) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2029 | dlm_lock_get(lock); |
| 2030 | if (lock->convert_pending) { |
| 2031 | /* move converting lock back to granted */ |
| 2032 | BUG_ON(i != DLM_CONVERTING_LIST); |
| 2033 | mlog(0, "node died with convert pending " |
| 2034 | "on %.*s. move back to granted list.\n", |
| 2035 | res->lockname.len, res->lockname.name); |
| 2036 | dlm_revert_pending_convert(res, lock); |
| 2037 | lock->convert_pending = 0; |
| 2038 | } else if (lock->lock_pending) { |
| 2039 | /* remove pending lock requests completely */ |
| 2040 | BUG_ON(i != DLM_BLOCKED_LIST); |
| 2041 | mlog(0, "node died with lock pending " |
| 2042 | "on %.*s. remove from blocked list and skip.\n", |
| 2043 | res->lockname.len, res->lockname.name); |
| 2044 | /* lock will be floating until ref in |
| 2045 | * dlmlock_remote is freed after the network |
| 2046 | * call returns. ok for it to not be on any |
| 2047 | * list since no ast can be called |
| 2048 | * (the master is dead). */ |
| 2049 | dlm_revert_pending_lock(res, lock); |
| 2050 | lock->lock_pending = 0; |
| 2051 | } else if (lock->unlock_pending) { |
| 2052 | /* if an unlock was in progress, treat as |
| 2053 | * if this had completed successfully |
| 2054 | * before sending this lock state to the |
| 2055 | * new master. note that the dlm_unlock |
| 2056 | * call is still responsible for calling |
| 2057 | * the unlockast. that will happen after |
| 2058 | * the network call times out. for now, |
| 2059 | * just move lists to prepare the new |
| 2060 | * recovery master. */ |
| 2061 | BUG_ON(i != DLM_GRANTED_LIST); |
| 2062 | mlog(0, "node died with unlock pending " |
| 2063 | "on %.*s. remove from blocked list and skip.\n", |
| 2064 | res->lockname.len, res->lockname.name); |
| 2065 | dlm_commit_pending_unlock(res, lock); |
| 2066 | lock->unlock_pending = 0; |
| 2067 | } else if (lock->cancel_pending) { |
| 2068 | /* if a cancel was in progress, treat as |
| 2069 | * if this had completed successfully |
| 2070 | * before sending this lock state to the |
| 2071 | * new master */ |
| 2072 | BUG_ON(i != DLM_CONVERTING_LIST); |
| 2073 | mlog(0, "node died with cancel pending " |
| 2074 | "on %.*s. move back to granted list.\n", |
| 2075 | res->lockname.len, res->lockname.name); |
| 2076 | dlm_commit_pending_cancel(res, lock); |
| 2077 | lock->cancel_pending = 0; |
| 2078 | } |
| 2079 | dlm_lock_put(lock); |
| 2080 | } |
| 2081 | } |
| 2082 | } |
| 2083 | |
| 2084 | |
| 2085 | |
| 2086 | /* removes all recovered locks from the recovery list. |
| 2087 | * sets the res->owner to the new master. |
| 2088 | * unsets the RECOVERY flag and wakes waiters. */ |
| 2089 | static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm, |
| 2090 | u8 dead_node, u8 new_master) |
| 2091 | { |
| 2092 | int i; |
Mark Fasheh | 81f2094 | 2006-02-28 17:31:22 -0800 | [diff] [blame] | 2093 | struct hlist_head *bucket; |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 2094 | struct dlm_lock_resource *res, *next; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2095 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2096 | assert_spin_locked(&dlm->spinlock); |
| 2097 | |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 2098 | list_for_each_entry_safe(res, next, &dlm->reco.resources, recovering) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2099 | if (res->owner == dead_node) { |
Sunil Mushran | 0afbba1 | 2011-07-24 10:25:54 -0700 | [diff] [blame] | 2100 | mlog(0, "%s: res %.*s, Changing owner from %u to %u\n", |
| 2101 | dlm->name, res->lockname.len, res->lockname.name, |
| 2102 | res->owner, new_master); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2103 | list_del_init(&res->recovering); |
| 2104 | spin_lock(&res->spinlock); |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 2105 | /* new_master has our reference from |
| 2106 | * the lock state sent during recovery */ |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2107 | dlm_change_lockres_owner(dlm, res, new_master); |
| 2108 | res->state &= ~DLM_LOCK_RES_RECOVERING; |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 2109 | if (__dlm_lockres_has_locks(res)) |
Kurt Hackel | 69d72b0 | 2006-05-01 10:57:51 -0700 | [diff] [blame] | 2110 | __dlm_dirty_lockres(dlm, res); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2111 | spin_unlock(&res->spinlock); |
| 2112 | wake_up(&res->wq); |
Kurt Hackel | 69d72b0 | 2006-05-01 10:57:51 -0700 | [diff] [blame] | 2113 | dlm_lockres_put(res); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2114 | } |
| 2115 | } |
| 2116 | |
| 2117 | /* this will become unnecessary eventually, but |
| 2118 | * for now we need to run the whole hash, clear |
| 2119 | * the RECOVERING state and set the owner |
| 2120 | * if necessary */ |
Mark Fasheh | 81f2094 | 2006-02-28 17:31:22 -0800 | [diff] [blame] | 2121 | for (i = 0; i < DLM_HASH_BUCKETS; i++) { |
Daniel Phillips | 03d864c | 2006-03-10 18:08:16 -0800 | [diff] [blame] | 2122 | bucket = dlm_lockres_hash(dlm, i); |
Sasha Levin | b67bfe0 | 2013-02-27 17:06:00 -0800 | [diff] [blame] | 2123 | hlist_for_each_entry(res, bucket, hash_node) { |
Sunil Mushran | 0afbba1 | 2011-07-24 10:25:54 -0700 | [diff] [blame] | 2124 | if (!(res->state & DLM_LOCK_RES_RECOVERING)) |
| 2125 | continue; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2126 | |
Sunil Mushran | 0afbba1 | 2011-07-24 10:25:54 -0700 | [diff] [blame] | 2127 | if (res->owner != dead_node && |
| 2128 | res->owner != dlm->node_num) |
| 2129 | continue; |
| 2130 | |
| 2131 | if (!list_empty(&res->recovering)) { |
| 2132 | list_del_init(&res->recovering); |
| 2133 | dlm_lockres_put(res); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2134 | } |
Sunil Mushran | 0afbba1 | 2011-07-24 10:25:54 -0700 | [diff] [blame] | 2135 | |
| 2136 | /* new_master has our reference from |
| 2137 | * the lock state sent during recovery */ |
| 2138 | mlog(0, "%s: res %.*s, Changing owner from %u to %u\n", |
| 2139 | dlm->name, res->lockname.len, res->lockname.name, |
| 2140 | res->owner, new_master); |
| 2141 | spin_lock(&res->spinlock); |
| 2142 | dlm_change_lockres_owner(dlm, res, new_master); |
| 2143 | res->state &= ~DLM_LOCK_RES_RECOVERING; |
| 2144 | if (__dlm_lockres_has_locks(res)) |
| 2145 | __dlm_dirty_lockres(dlm, res); |
| 2146 | spin_unlock(&res->spinlock); |
| 2147 | wake_up(&res->wq); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2148 | } |
| 2149 | } |
| 2150 | } |
| 2151 | |
| 2152 | static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local) |
| 2153 | { |
| 2154 | if (local) { |
| 2155 | if (lock->ml.type != LKM_EXMODE && |
| 2156 | lock->ml.type != LKM_PRMODE) |
| 2157 | return 1; |
| 2158 | } else if (lock->ml.type == LKM_EXMODE) |
| 2159 | return 1; |
| 2160 | return 0; |
| 2161 | } |
| 2162 | |
| 2163 | static void dlm_revalidate_lvb(struct dlm_ctxt *dlm, |
| 2164 | struct dlm_lock_resource *res, u8 dead_node) |
| 2165 | { |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 2166 | struct list_head *queue; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2167 | struct dlm_lock *lock; |
| 2168 | int blank_lvb = 0, local = 0; |
| 2169 | int i; |
| 2170 | u8 search_node; |
| 2171 | |
| 2172 | assert_spin_locked(&dlm->spinlock); |
| 2173 | assert_spin_locked(&res->spinlock); |
| 2174 | |
| 2175 | if (res->owner == dlm->node_num) |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2176 | /* if this node owned the lockres, and if the dead node |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2177 | * had an EX when he died, blank out the lvb */ |
| 2178 | search_node = dead_node; |
| 2179 | else { |
| 2180 | /* if this is a secondary lockres, and we had no EX or PR |
| 2181 | * locks granted, we can no longer trust the lvb */ |
| 2182 | search_node = dlm->node_num; |
| 2183 | local = 1; /* check local state for valid lvb */ |
| 2184 | } |
| 2185 | |
| 2186 | for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) { |
| 2187 | queue = dlm_list_idx_to_ptr(res, i); |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 2188 | list_for_each_entry(lock, queue, list) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2189 | if (lock->ml.node == search_node) { |
| 2190 | if (dlm_lvb_needs_invalidation(lock, local)) { |
| 2191 | /* zero the lksb lvb and lockres lvb */ |
| 2192 | blank_lvb = 1; |
| 2193 | memset(lock->lksb->lvb, 0, DLM_LVB_LEN); |
| 2194 | } |
| 2195 | } |
| 2196 | } |
| 2197 | } |
| 2198 | |
| 2199 | if (blank_lvb) { |
| 2200 | mlog(0, "clearing %.*s lvb, dead node %u had EX\n", |
| 2201 | res->lockname.len, res->lockname.name, dead_node); |
| 2202 | memset(res->lvb, 0, DLM_LVB_LEN); |
| 2203 | } |
| 2204 | } |
| 2205 | |
| 2206 | static void dlm_free_dead_locks(struct dlm_ctxt *dlm, |
| 2207 | struct dlm_lock_resource *res, u8 dead_node) |
| 2208 | { |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 2209 | struct dlm_lock *lock, *next; |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 2210 | unsigned int freed = 0; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2211 | |
| 2212 | /* this node is the lockres master: |
| 2213 | * 1) remove any stale locks for the dead node |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2214 | * 2) if the dead node had an EX when he died, blank out the lvb |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2215 | */ |
| 2216 | assert_spin_locked(&dlm->spinlock); |
| 2217 | assert_spin_locked(&res->spinlock); |
| 2218 | |
Sunil Mushran | 2c5c54a | 2008-03-01 14:04:20 -0800 | [diff] [blame] | 2219 | /* We do two dlm_lock_put(). One for removing from list and the other is |
| 2220 | * to force the DLM_UNLOCK_FREE_LOCK action so as to free the locks */ |
| 2221 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2222 | /* TODO: check pending_asts, pending_basts here */ |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 2223 | list_for_each_entry_safe(lock, next, &res->granted, list) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2224 | if (lock->ml.node == dead_node) { |
| 2225 | list_del_init(&lock->list); |
| 2226 | dlm_lock_put(lock); |
Sunil Mushran | 2c5c54a | 2008-03-01 14:04:20 -0800 | [diff] [blame] | 2227 | /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ |
| 2228 | dlm_lock_put(lock); |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 2229 | freed++; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2230 | } |
| 2231 | } |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 2232 | list_for_each_entry_safe(lock, next, &res->converting, list) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2233 | if (lock->ml.node == dead_node) { |
| 2234 | list_del_init(&lock->list); |
| 2235 | dlm_lock_put(lock); |
Sunil Mushran | 2c5c54a | 2008-03-01 14:04:20 -0800 | [diff] [blame] | 2236 | /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ |
| 2237 | dlm_lock_put(lock); |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 2238 | freed++; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2239 | } |
| 2240 | } |
Christoph Hellwig | 800deef | 2007-05-17 16:03:13 +0200 | [diff] [blame] | 2241 | list_for_each_entry_safe(lock, next, &res->blocked, list) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2242 | if (lock->ml.node == dead_node) { |
| 2243 | list_del_init(&lock->list); |
| 2244 | dlm_lock_put(lock); |
Sunil Mushran | 2c5c54a | 2008-03-01 14:04:20 -0800 | [diff] [blame] | 2245 | /* Can't schedule DLM_UNLOCK_FREE_LOCK - do manually */ |
| 2246 | dlm_lock_put(lock); |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 2247 | freed++; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2248 | } |
| 2249 | } |
| 2250 | |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 2251 | if (freed) { |
| 2252 | mlog(0, "%s:%.*s: freed %u locks for dead node %u, " |
| 2253 | "dropping ref from lockres\n", dlm->name, |
| 2254 | res->lockname.len, res->lockname.name, freed, dead_node); |
Sunil Mushran | cda70ba | 2010-02-01 17:34:58 -0800 | [diff] [blame] | 2255 | if(!test_bit(dead_node, res->refmap)) { |
| 2256 | mlog(ML_ERROR, "%s:%.*s: freed %u locks for dead node %u, " |
| 2257 | "but ref was not set\n", dlm->name, |
| 2258 | res->lockname.len, res->lockname.name, freed, dead_node); |
| 2259 | __dlm_print_one_lock_resource(res); |
| 2260 | } |
Sunil Mushran | 8d400b8 | 2011-07-24 10:26:54 -0700 | [diff] [blame] | 2261 | dlm_lockres_clear_refmap_bit(dlm, res, dead_node); |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 2262 | } else if (test_bit(dead_node, res->refmap)) { |
| 2263 | mlog(0, "%s:%.*s: dead node %u had a ref, but had " |
| 2264 | "no locks and had not purged before dying\n", dlm->name, |
| 2265 | res->lockname.len, res->lockname.name, dead_node); |
Sunil Mushran | 8d400b8 | 2011-07-24 10:26:54 -0700 | [diff] [blame] | 2266 | dlm_lockres_clear_refmap_bit(dlm, res, dead_node); |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 2267 | } |
| 2268 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2269 | /* do not kick thread yet */ |
| 2270 | __dlm_dirty_lockres(dlm, res); |
| 2271 | } |
| 2272 | |
| 2273 | /* if this node is the recovery master, and there are no |
| 2274 | * locks for a given lockres owned by this node that are in |
| 2275 | * either PR or EX mode, zero out the lvb before requesting. |
| 2276 | * |
| 2277 | */ |
| 2278 | |
| 2279 | |
| 2280 | static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node) |
| 2281 | { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2282 | struct dlm_lock_resource *res; |
| 2283 | int i; |
Mark Fasheh | 81f2094 | 2006-02-28 17:31:22 -0800 | [diff] [blame] | 2284 | struct hlist_head *bucket; |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2285 | struct dlm_lock *lock; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2286 | |
| 2287 | |
| 2288 | /* purge any stale mles */ |
| 2289 | dlm_clean_master_list(dlm, dead_node); |
| 2290 | |
| 2291 | /* |
| 2292 | * now clean up all lock resources. there are two rules: |
| 2293 | * |
| 2294 | * 1) if the dead node was the master, move the lockres |
| 2295 | * to the recovering list. set the RECOVERING flag. |
| 2296 | * this lockres needs to be cleaned up before it can |
| 2297 | * be used further. |
| 2298 | * |
| 2299 | * 2) if this node was the master, remove all locks from |
| 2300 | * each of the lockres queues that were owned by the |
| 2301 | * dead node. once recovery finishes, the dlm thread |
| 2302 | * can be kicked again to see if any ASTs or BASTs |
| 2303 | * need to be fired as a result. |
| 2304 | */ |
Mark Fasheh | 81f2094 | 2006-02-28 17:31:22 -0800 | [diff] [blame] | 2305 | for (i = 0; i < DLM_HASH_BUCKETS; i++) { |
Daniel Phillips | 03d864c | 2006-03-10 18:08:16 -0800 | [diff] [blame] | 2306 | bucket = dlm_lockres_hash(dlm, i); |
Sasha Levin | b67bfe0 | 2013-02-27 17:06:00 -0800 | [diff] [blame] | 2307 | hlist_for_each_entry(res, bucket, hash_node) { |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2308 | /* always prune any $RECOVERY entries for dead nodes, |
| 2309 | * otherwise hangs can occur during later recovery */ |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2310 | if (dlm_is_recovery_lock(res->lockname.name, |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2311 | res->lockname.len)) { |
| 2312 | spin_lock(&res->spinlock); |
| 2313 | list_for_each_entry(lock, &res->granted, list) { |
| 2314 | if (lock->ml.node == dead_node) { |
| 2315 | mlog(0, "AHA! there was " |
| 2316 | "a $RECOVERY lock for dead " |
| 2317 | "node %u (%s)!\n", |
| 2318 | dead_node, dlm->name); |
| 2319 | list_del_init(&lock->list); |
| 2320 | dlm_lock_put(lock); |
| 2321 | break; |
| 2322 | } |
| 2323 | } |
| 2324 | spin_unlock(&res->spinlock); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2325 | continue; |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2326 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2327 | spin_lock(&res->spinlock); |
| 2328 | /* zero the lvb if necessary */ |
| 2329 | dlm_revalidate_lvb(dlm, res, dead_node); |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 2330 | if (res->owner == dead_node) { |
Wengang Wang | a524812 | 2010-07-30 16:14:44 +0800 | [diff] [blame] | 2331 | if (res->state & DLM_LOCK_RES_DROPPING_REF) { |
Sunil Mushran | 8decab3 | 2011-07-24 10:23:54 -0700 | [diff] [blame] | 2332 | mlog(ML_NOTICE, "%s: res %.*s, Skip " |
Wengang Wang | a524812 | 2010-07-30 16:14:44 +0800 | [diff] [blame] | 2333 | "recovery as it is being freed\n", |
Sunil Mushran | 8decab3 | 2011-07-24 10:23:54 -0700 | [diff] [blame] | 2334 | dlm->name, res->lockname.len, |
Wengang Wang | a524812 | 2010-07-30 16:14:44 +0800 | [diff] [blame] | 2335 | res->lockname.name); |
| 2336 | } else |
| 2337 | dlm_move_lockres_to_recovery_list(dlm, |
| 2338 | res); |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 2339 | |
Kurt Hackel | ba2bf21 | 2006-12-01 14:47:20 -0800 | [diff] [blame] | 2340 | } else if (res->owner == dlm->node_num) { |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2341 | dlm_free_dead_locks(dlm, res, dead_node); |
| 2342 | __dlm_lockres_calc_usage(dlm, res); |
Xue jiufei | 69b2bd16 | 2013-09-11 14:19:57 -0700 | [diff] [blame] | 2343 | } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) { |
| 2344 | if (test_bit(dead_node, res->refmap)) { |
| 2345 | mlog(0, "%s:%.*s: dead node %u had a ref, but had " |
| 2346 | "no locks and had not purged before dying\n", |
| 2347 | dlm->name, res->lockname.len, |
| 2348 | res->lockname.name, dead_node); |
| 2349 | dlm_lockres_clear_refmap_bit(dlm, res, dead_node); |
| 2350 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2351 | } |
| 2352 | spin_unlock(&res->spinlock); |
| 2353 | } |
| 2354 | } |
| 2355 | |
| 2356 | } |
| 2357 | |
| 2358 | static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx) |
| 2359 | { |
| 2360 | assert_spin_locked(&dlm->spinlock); |
| 2361 | |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2362 | if (dlm->reco.new_master == idx) { |
| 2363 | mlog(0, "%s: recovery master %d just died\n", |
| 2364 | dlm->name, idx); |
| 2365 | if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { |
| 2366 | /* finalize1 was reached, so it is safe to clear |
| 2367 | * the new_master and dead_node. that recovery |
| 2368 | * is complete. */ |
| 2369 | mlog(0, "%s: dead master %d had reached " |
| 2370 | "finalize1 state, clearing\n", dlm->name, idx); |
| 2371 | dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE; |
| 2372 | __dlm_reset_recovery(dlm); |
| 2373 | } |
| 2374 | } |
| 2375 | |
Tao Ma | 2d4b1cb | 2008-01-10 15:20:55 +0800 | [diff] [blame] | 2376 | /* Clean up join state on node death. */ |
| 2377 | if (dlm->joining_node == idx) { |
| 2378 | mlog(0, "Clearing join state for node %u\n", idx); |
| 2379 | __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN); |
| 2380 | } |
| 2381 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2382 | /* check to see if the node is already considered dead */ |
| 2383 | if (!test_bit(idx, dlm->live_nodes_map)) { |
| 2384 | mlog(0, "for domain %s, node %d is already dead. " |
| 2385 | "another node likely did recovery already.\n", |
| 2386 | dlm->name, idx); |
| 2387 | return; |
| 2388 | } |
| 2389 | |
| 2390 | /* check to see if we do not care about this node */ |
| 2391 | if (!test_bit(idx, dlm->domain_map)) { |
| 2392 | /* This also catches the case that we get a node down |
| 2393 | * but haven't joined the domain yet. */ |
| 2394 | mlog(0, "node %u already removed from domain!\n", idx); |
| 2395 | return; |
| 2396 | } |
| 2397 | |
| 2398 | clear_bit(idx, dlm->live_nodes_map); |
| 2399 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2400 | /* make sure local cleanup occurs before the heartbeat events */ |
| 2401 | if (!test_bit(idx, dlm->recovery_map)) |
| 2402 | dlm_do_local_recovery_cleanup(dlm, idx); |
| 2403 | |
| 2404 | /* notify anything attached to the heartbeat events */ |
| 2405 | dlm_hb_event_notify_attached(dlm, idx, 0); |
| 2406 | |
| 2407 | mlog(0, "node %u being removed from domain map!\n", idx); |
| 2408 | clear_bit(idx, dlm->domain_map); |
Sunil Mushran | bddefde | 2011-05-19 14:34:11 -0700 | [diff] [blame] | 2409 | clear_bit(idx, dlm->exit_domain_map); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2410 | /* wake up migration waiters if a node goes down. |
| 2411 | * perhaps later we can genericize this for other waiters. */ |
| 2412 | wake_up(&dlm->migration_wq); |
| 2413 | |
| 2414 | if (test_bit(idx, dlm->recovery_map)) |
| 2415 | mlog(0, "domain %s, node %u already added " |
| 2416 | "to recovery map!\n", dlm->name, idx); |
| 2417 | else |
| 2418 | set_bit(idx, dlm->recovery_map); |
| 2419 | } |
| 2420 | |
| 2421 | void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data) |
| 2422 | { |
| 2423 | struct dlm_ctxt *dlm = data; |
| 2424 | |
| 2425 | if (!dlm_grab(dlm)) |
| 2426 | return; |
| 2427 | |
Mark Fasheh | 6561168 | 2007-09-07 11:11:10 -0700 | [diff] [blame] | 2428 | /* |
| 2429 | * This will notify any dlm users that a node in our domain |
| 2430 | * went away without notifying us first. |
| 2431 | */ |
| 2432 | if (test_bit(idx, dlm->domain_map)) |
| 2433 | dlm_fire_domain_eviction_callbacks(dlm, idx); |
| 2434 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2435 | spin_lock(&dlm->spinlock); |
| 2436 | __dlm_hb_node_down(dlm, idx); |
| 2437 | spin_unlock(&dlm->spinlock); |
| 2438 | |
| 2439 | dlm_put(dlm); |
| 2440 | } |
| 2441 | |
| 2442 | void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data) |
| 2443 | { |
| 2444 | struct dlm_ctxt *dlm = data; |
| 2445 | |
| 2446 | if (!dlm_grab(dlm)) |
| 2447 | return; |
| 2448 | |
| 2449 | spin_lock(&dlm->spinlock); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2450 | set_bit(idx, dlm->live_nodes_map); |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2451 | /* do NOT notify mle attached to the heartbeat events. |
| 2452 | * new nodes are not interesting in mastery until joined. */ |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2453 | spin_unlock(&dlm->spinlock); |
| 2454 | |
| 2455 | dlm_put(dlm); |
| 2456 | } |
| 2457 | |
| 2458 | static void dlm_reco_ast(void *astdata) |
| 2459 | { |
| 2460 | struct dlm_ctxt *dlm = astdata; |
| 2461 | mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n", |
| 2462 | dlm->node_num, dlm->name); |
| 2463 | } |
| 2464 | static void dlm_reco_bast(void *astdata, int blocked_type) |
| 2465 | { |
| 2466 | struct dlm_ctxt *dlm = astdata; |
| 2467 | mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n", |
| 2468 | dlm->node_num, dlm->name); |
| 2469 | } |
| 2470 | static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st) |
| 2471 | { |
| 2472 | mlog(0, "unlockast for recovery lock fired!\n"); |
| 2473 | } |
| 2474 | |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2475 | /* |
| 2476 | * dlm_pick_recovery_master will continually attempt to use |
| 2477 | * dlmlock() on the special "$RECOVERY" lockres with the |
| 2478 | * LKM_NOQUEUE flag to get an EX. every thread that enters |
| 2479 | * this function on each node racing to become the recovery |
| 2480 | * master will not stop attempting this until either: |
| 2481 | * a) this node gets the EX (and becomes the recovery master), |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2482 | * or b) dlm->reco.new_master gets set to some nodenum |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2483 | * != O2NM_INVALID_NODE_NUM (another node will do the reco). |
| 2484 | * so each time a recovery master is needed, the entire cluster |
| 2485 | * will sync at this point. if the new master dies, that will |
| 2486 | * be detected in dlm_do_recovery */ |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2487 | static int dlm_pick_recovery_master(struct dlm_ctxt *dlm) |
| 2488 | { |
| 2489 | enum dlm_status ret; |
| 2490 | struct dlm_lockstatus lksb; |
| 2491 | int status = -EINVAL; |
| 2492 | |
| 2493 | mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n", |
| 2494 | dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num); |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2495 | again: |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2496 | memset(&lksb, 0, sizeof(lksb)); |
| 2497 | |
| 2498 | ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY, |
Mark Fasheh | 3384f3d | 2006-09-08 11:38:29 -0700 | [diff] [blame] | 2499 | DLM_RECOVERY_LOCK_NAME, DLM_RECOVERY_LOCK_NAME_LEN, |
| 2500 | dlm_reco_ast, dlm, dlm_reco_bast); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2501 | |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2502 | mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n", |
| 2503 | dlm->name, ret, lksb.status); |
| 2504 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2505 | if (ret == DLM_NORMAL) { |
| 2506 | mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n", |
| 2507 | dlm->name, dlm->node_num); |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2508 | |
| 2509 | /* got the EX lock. check to see if another node |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2510 | * just became the reco master */ |
| 2511 | if (dlm_reco_master_ready(dlm)) { |
| 2512 | mlog(0, "%s: got reco EX lock, but %u will " |
| 2513 | "do the recovery\n", dlm->name, |
| 2514 | dlm->reco.new_master); |
| 2515 | status = -EEXIST; |
| 2516 | } else { |
Kurt Hackel | 898effa | 2006-01-18 17:01:25 -0800 | [diff] [blame] | 2517 | status = 0; |
| 2518 | |
| 2519 | /* see if recovery was already finished elsewhere */ |
| 2520 | spin_lock(&dlm->spinlock); |
| 2521 | if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) { |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2522 | status = -EINVAL; |
Kurt Hackel | 898effa | 2006-01-18 17:01:25 -0800 | [diff] [blame] | 2523 | mlog(0, "%s: got reco EX lock, but " |
| 2524 | "node got recovered already\n", dlm->name); |
| 2525 | if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) { |
| 2526 | mlog(ML_ERROR, "%s: new master is %u " |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2527 | "but no dead node!\n", |
Kurt Hackel | 898effa | 2006-01-18 17:01:25 -0800 | [diff] [blame] | 2528 | dlm->name, dlm->reco.new_master); |
| 2529 | BUG(); |
| 2530 | } |
| 2531 | } |
| 2532 | spin_unlock(&dlm->spinlock); |
| 2533 | } |
| 2534 | |
| 2535 | /* if this node has actually become the recovery master, |
| 2536 | * set the master and send the messages to begin recovery */ |
| 2537 | if (!status) { |
| 2538 | mlog(0, "%s: dead=%u, this=%u, sending " |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2539 | "begin_reco now\n", dlm->name, |
Kurt Hackel | 898effa | 2006-01-18 17:01:25 -0800 | [diff] [blame] | 2540 | dlm->reco.dead_node, dlm->node_num); |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2541 | status = dlm_send_begin_reco_message(dlm, |
| 2542 | dlm->reco.dead_node); |
| 2543 | /* this always succeeds */ |
| 2544 | BUG_ON(status); |
| 2545 | |
| 2546 | /* set the new_master to this node */ |
| 2547 | spin_lock(&dlm->spinlock); |
Kurt Hackel | ab27eb6 | 2006-04-27 18:03:49 -0700 | [diff] [blame] | 2548 | dlm_set_reco_master(dlm, dlm->node_num); |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2549 | spin_unlock(&dlm->spinlock); |
| 2550 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2551 | |
| 2552 | /* recovery lock is a special case. ast will not get fired, |
| 2553 | * so just go ahead and unlock it. */ |
| 2554 | ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm); |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2555 | if (ret == DLM_DENIED) { |
| 2556 | mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n"); |
| 2557 | ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm); |
| 2558 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2559 | if (ret != DLM_NORMAL) { |
| 2560 | /* this would really suck. this could only happen |
| 2561 | * if there was a network error during the unlock |
| 2562 | * because of node death. this means the unlock |
| 2563 | * is actually "done" and the lock structure is |
| 2564 | * even freed. we can continue, but only |
| 2565 | * because this specific lock name is special. */ |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2566 | mlog(ML_ERROR, "dlmunlock returned %d\n", ret); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2567 | } |
| 2568 | } else if (ret == DLM_NOTQUEUED) { |
| 2569 | mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n", |
| 2570 | dlm->name, dlm->node_num); |
| 2571 | /* another node is master. wait on |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2572 | * reco.new_master != O2NM_INVALID_NODE_NUM |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2573 | * for at most one second */ |
| 2574 | wait_event_timeout(dlm->dlm_reco_thread_wq, |
| 2575 | dlm_reco_master_ready(dlm), |
| 2576 | msecs_to_jiffies(1000)); |
| 2577 | if (!dlm_reco_master_ready(dlm)) { |
| 2578 | mlog(0, "%s: reco master taking awhile\n", |
| 2579 | dlm->name); |
| 2580 | goto again; |
| 2581 | } |
| 2582 | /* another node has informed this one that it is reco master */ |
| 2583 | mlog(0, "%s: reco master %u is ready to recover %u\n", |
| 2584 | dlm->name, dlm->reco.new_master, dlm->reco.dead_node); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2585 | status = -EEXIST; |
Kurt Hackel | c8df412 | 2006-05-01 13:47:50 -0700 | [diff] [blame] | 2586 | } else if (ret == DLM_RECOVERING) { |
| 2587 | mlog(0, "dlm=%s dlmlock says master node died (this=%u)\n", |
| 2588 | dlm->name, dlm->node_num); |
| 2589 | goto again; |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2590 | } else { |
| 2591 | struct dlm_lock_resource *res; |
| 2592 | |
| 2593 | /* dlmlock returned something other than NOTQUEUED or NORMAL */ |
| 2594 | mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), " |
| 2595 | "lksb.status=%s\n", dlm->name, dlm_errname(ret), |
| 2596 | dlm_errname(lksb.status)); |
| 2597 | res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME, |
| 2598 | DLM_RECOVERY_LOCK_NAME_LEN); |
| 2599 | if (res) { |
| 2600 | dlm_print_one_lock_resource(res); |
| 2601 | dlm_lockres_put(res); |
| 2602 | } else { |
| 2603 | mlog(ML_ERROR, "recovery lock not found\n"); |
| 2604 | } |
| 2605 | BUG(); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2606 | } |
| 2607 | |
| 2608 | return status; |
| 2609 | } |
| 2610 | |
| 2611 | static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node) |
| 2612 | { |
| 2613 | struct dlm_begin_reco br; |
| 2614 | int ret = 0; |
| 2615 | struct dlm_node_iter iter; |
| 2616 | int nodenum; |
| 2617 | int status; |
| 2618 | |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 2619 | mlog(0, "%s: dead node is %u\n", dlm->name, dead_node); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2620 | |
| 2621 | spin_lock(&dlm->spinlock); |
| 2622 | dlm_node_iter_init(dlm->domain_map, &iter); |
| 2623 | spin_unlock(&dlm->spinlock); |
| 2624 | |
| 2625 | clear_bit(dead_node, iter.node_map); |
| 2626 | |
| 2627 | memset(&br, 0, sizeof(br)); |
| 2628 | br.node_idx = dlm->node_num; |
| 2629 | br.dead_node = dead_node; |
| 2630 | |
| 2631 | while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { |
| 2632 | ret = 0; |
| 2633 | if (nodenum == dead_node) { |
| 2634 | mlog(0, "not sending begin reco to dead node " |
| 2635 | "%u\n", dead_node); |
| 2636 | continue; |
| 2637 | } |
| 2638 | if (nodenum == dlm->node_num) { |
| 2639 | mlog(0, "not sending begin reco to self\n"); |
| 2640 | continue; |
| 2641 | } |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2642 | retry: |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2643 | ret = -EINVAL; |
| 2644 | mlog(0, "attempting to send begin reco msg to %d\n", |
| 2645 | nodenum); |
| 2646 | ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key, |
| 2647 | &br, sizeof(br), nodenum, &status); |
| 2648 | /* negative status is handled ok by caller here */ |
| 2649 | if (ret >= 0) |
| 2650 | ret = status; |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2651 | if (dlm_is_host_down(ret)) { |
| 2652 | /* node is down. not involved in recovery |
| 2653 | * so just keep going */ |
Wengang Wang | a5196ec | 2010-03-30 12:09:22 +0800 | [diff] [blame] | 2654 | mlog(ML_NOTICE, "%s: node %u was down when sending " |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2655 | "begin reco msg (%d)\n", dlm->name, nodenum, ret); |
| 2656 | ret = 0; |
| 2657 | } |
Sunil Mushran | cd34edd | 2010-01-25 17:58:30 -0800 | [diff] [blame] | 2658 | |
| 2659 | /* |
| 2660 | * Prior to commit aad1b15310b9bcd59fa81ab8f2b1513b59553ea8, |
| 2661 | * dlm_begin_reco_handler() returned EAGAIN and not -EAGAIN. |
| 2662 | * We are handling both for compatibility reasons. |
| 2663 | */ |
| 2664 | if (ret == -EAGAIN || ret == EAGAIN) { |
Tiger Yang | aad1b15 | 2009-11-19 10:17:46 +0800 | [diff] [blame] | 2665 | mlog(0, "%s: trying to start recovery of node " |
| 2666 | "%u, but node %u is waiting for last recovery " |
| 2667 | "to complete, backoff for a bit\n", dlm->name, |
| 2668 | dead_node, nodenum); |
| 2669 | msleep(100); |
| 2670 | goto retry; |
| 2671 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2672 | if (ret < 0) { |
| 2673 | struct dlm_lock_resource *res; |
Wengang Wang | a5196ec | 2010-03-30 12:09:22 +0800 | [diff] [blame] | 2674 | |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2675 | /* this is now a serious problem, possibly ENOMEM |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2676 | * in the network stack. must retry */ |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2677 | mlog_errno(ret); |
| 2678 | mlog(ML_ERROR, "begin reco of dlm %s to node %u " |
Wengang Wang | a5196ec | 2010-03-30 12:09:22 +0800 | [diff] [blame] | 2679 | "returned %d\n", dlm->name, nodenum, ret); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2680 | res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME, |
| 2681 | DLM_RECOVERY_LOCK_NAME_LEN); |
| 2682 | if (res) { |
| 2683 | dlm_print_one_lock_resource(res); |
| 2684 | dlm_lockres_put(res); |
| 2685 | } else { |
| 2686 | mlog(ML_ERROR, "recovery lock not found\n"); |
| 2687 | } |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2688 | /* sleep for a bit in hopes that we can avoid |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2689 | * another ENOMEM */ |
| 2690 | msleep(100); |
| 2691 | goto retry; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2692 | } |
| 2693 | } |
| 2694 | |
| 2695 | return ret; |
| 2696 | } |
| 2697 | |
Kurt Hackel | d74c980 | 2007-01-17 17:04:25 -0800 | [diff] [blame] | 2698 | int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data, |
| 2699 | void **ret_data) |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2700 | { |
| 2701 | struct dlm_ctxt *dlm = data; |
| 2702 | struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf; |
| 2703 | |
| 2704 | /* ok to return 0, domain has gone away */ |
| 2705 | if (!dlm_grab(dlm)) |
| 2706 | return 0; |
| 2707 | |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2708 | spin_lock(&dlm->spinlock); |
| 2709 | if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { |
| 2710 | mlog(0, "%s: node %u wants to recover node %u (%u:%u) " |
| 2711 | "but this node is in finalize state, waiting on finalize2\n", |
| 2712 | dlm->name, br->node_idx, br->dead_node, |
| 2713 | dlm->reco.dead_node, dlm->reco.new_master); |
| 2714 | spin_unlock(&dlm->spinlock); |
Xue jiufei | 40c7f2e | 2013-07-03 15:00:50 -0700 | [diff] [blame] | 2715 | dlm_put(dlm); |
Tiger Yang | aad1b15 | 2009-11-19 10:17:46 +0800 | [diff] [blame] | 2716 | return -EAGAIN; |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2717 | } |
| 2718 | spin_unlock(&dlm->spinlock); |
| 2719 | |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 2720 | mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n", |
| 2721 | dlm->name, br->node_idx, br->dead_node, |
| 2722 | dlm->reco.dead_node, dlm->reco.new_master); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2723 | |
| 2724 | dlm_fire_domain_eviction_callbacks(dlm, br->dead_node); |
| 2725 | |
| 2726 | spin_lock(&dlm->spinlock); |
| 2727 | if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) { |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2728 | if (test_bit(dlm->reco.new_master, dlm->recovery_map)) { |
| 2729 | mlog(0, "%s: new_master %u died, changing " |
| 2730 | "to %u\n", dlm->name, dlm->reco.new_master, |
| 2731 | br->node_idx); |
| 2732 | } else { |
| 2733 | mlog(0, "%s: new_master %u NOT DEAD, changing " |
| 2734 | "to %u\n", dlm->name, dlm->reco.new_master, |
| 2735 | br->node_idx); |
| 2736 | /* may not have seen the new master as dead yet */ |
| 2737 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2738 | } |
| 2739 | if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) { |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2740 | mlog(ML_NOTICE, "%s: dead_node previously set to %u, " |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2741 | "node %u changing it to %u\n", dlm->name, |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2742 | dlm->reco.dead_node, br->node_idx, br->dead_node); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2743 | } |
Kurt Hackel | ab27eb6 | 2006-04-27 18:03:49 -0700 | [diff] [blame] | 2744 | dlm_set_reco_master(dlm, br->node_idx); |
| 2745 | dlm_set_reco_dead_node(dlm, br->dead_node); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2746 | if (!test_bit(br->dead_node, dlm->recovery_map)) { |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2747 | mlog(0, "recovery master %u sees %u as dead, but this " |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2748 | "node has not yet. marking %u as dead\n", |
| 2749 | br->node_idx, br->dead_node, br->dead_node); |
Kurt Hackel | e2faea4 | 2006-01-12 14:24:55 -0800 | [diff] [blame] | 2750 | if (!test_bit(br->dead_node, dlm->domain_map) || |
| 2751 | !test_bit(br->dead_node, dlm->live_nodes_map)) |
| 2752 | mlog(0, "%u not in domain/live_nodes map " |
| 2753 | "so setting it in reco map manually\n", |
| 2754 | br->dead_node); |
Kurt Hackel | c03872f | 2006-03-06 14:08:49 -0800 | [diff] [blame] | 2755 | /* force the recovery cleanup in __dlm_hb_node_down |
| 2756 | * both of these will be cleared in a moment */ |
| 2757 | set_bit(br->dead_node, dlm->domain_map); |
| 2758 | set_bit(br->dead_node, dlm->live_nodes_map); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2759 | __dlm_hb_node_down(dlm, br->dead_node); |
| 2760 | } |
| 2761 | spin_unlock(&dlm->spinlock); |
| 2762 | |
| 2763 | dlm_kick_recovery_thread(dlm); |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 2764 | |
| 2765 | mlog(0, "%s: recovery started by node %u, for %u (%u:%u)\n", |
| 2766 | dlm->name, br->node_idx, br->dead_node, |
| 2767 | dlm->reco.dead_node, dlm->reco.new_master); |
| 2768 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2769 | dlm_put(dlm); |
| 2770 | return 0; |
| 2771 | } |
| 2772 | |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2773 | #define DLM_FINALIZE_STAGE2 0x01 |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2774 | static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm) |
| 2775 | { |
| 2776 | int ret = 0; |
| 2777 | struct dlm_finalize_reco fr; |
| 2778 | struct dlm_node_iter iter; |
| 2779 | int nodenum; |
| 2780 | int status; |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2781 | int stage = 1; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2782 | |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2783 | mlog(0, "finishing recovery for node %s:%u, " |
| 2784 | "stage %d\n", dlm->name, dlm->reco.dead_node, stage); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2785 | |
| 2786 | spin_lock(&dlm->spinlock); |
| 2787 | dlm_node_iter_init(dlm->domain_map, &iter); |
| 2788 | spin_unlock(&dlm->spinlock); |
| 2789 | |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2790 | stage2: |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2791 | memset(&fr, 0, sizeof(fr)); |
| 2792 | fr.node_idx = dlm->node_num; |
| 2793 | fr.dead_node = dlm->reco.dead_node; |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2794 | if (stage == 2) |
| 2795 | fr.flags |= DLM_FINALIZE_STAGE2; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2796 | |
| 2797 | while ((nodenum = dlm_node_iter_next(&iter)) >= 0) { |
| 2798 | if (nodenum == dlm->node_num) |
| 2799 | continue; |
| 2800 | ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key, |
| 2801 | &fr, sizeof(fr), nodenum, &status); |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2802 | if (ret >= 0) |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2803 | ret = status; |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2804 | if (ret < 0) { |
Wengang Wang | a5196ec | 2010-03-30 12:09:22 +0800 | [diff] [blame] | 2805 | mlog(ML_ERROR, "Error %d when sending message %u (key " |
| 2806 | "0x%x) to node %u\n", ret, DLM_FINALIZE_RECO_MSG, |
| 2807 | dlm->key, nodenum); |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2808 | if (dlm_is_host_down(ret)) { |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2809 | /* this has no effect on this recovery |
| 2810 | * session, so set the status to zero to |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2811 | * finish out the last recovery */ |
| 2812 | mlog(ML_ERROR, "node %u went down after this " |
| 2813 | "node finished recovery.\n", nodenum); |
| 2814 | ret = 0; |
Kurt Hackel | c27069e | 2006-05-01 13:51:49 -0700 | [diff] [blame] | 2815 | continue; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2816 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2817 | break; |
| 2818 | } |
| 2819 | } |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2820 | if (stage == 1) { |
| 2821 | /* reset the node_iter back to the top and send finalize2 */ |
| 2822 | iter.curnode = -1; |
| 2823 | stage = 2; |
| 2824 | goto stage2; |
| 2825 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2826 | |
| 2827 | return ret; |
| 2828 | } |
| 2829 | |
Kurt Hackel | d74c980 | 2007-01-17 17:04:25 -0800 | [diff] [blame] | 2830 | int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data, |
| 2831 | void **ret_data) |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2832 | { |
| 2833 | struct dlm_ctxt *dlm = data; |
| 2834 | struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf; |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2835 | int stage = 1; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2836 | |
| 2837 | /* ok to return 0, domain has gone away */ |
| 2838 | if (!dlm_grab(dlm)) |
| 2839 | return 0; |
| 2840 | |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2841 | if (fr->flags & DLM_FINALIZE_STAGE2) |
| 2842 | stage = 2; |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2843 | |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2844 | mlog(0, "%s: node %u finalizing recovery stage%d of " |
| 2845 | "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage, |
| 2846 | fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master); |
Sunil Mushran | 2bd6321 | 2010-01-25 16:57:38 -0800 | [diff] [blame] | 2847 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2848 | spin_lock(&dlm->spinlock); |
| 2849 | |
| 2850 | if (dlm->reco.new_master != fr->node_idx) { |
| 2851 | mlog(ML_ERROR, "node %u sent recovery finalize msg, but node " |
| 2852 | "%u is supposed to be the new master, dead=%u\n", |
| 2853 | fr->node_idx, dlm->reco.new_master, fr->dead_node); |
| 2854 | BUG(); |
| 2855 | } |
| 2856 | if (dlm->reco.dead_node != fr->dead_node) { |
| 2857 | mlog(ML_ERROR, "node %u sent recovery finalize msg for dead " |
| 2858 | "node %u, but node %u is supposed to be dead\n", |
| 2859 | fr->node_idx, fr->dead_node, dlm->reco.dead_node); |
| 2860 | BUG(); |
| 2861 | } |
| 2862 | |
Kurt Hackel | 466d1a4 | 2006-05-01 11:11:13 -0700 | [diff] [blame] | 2863 | switch (stage) { |
| 2864 | case 1: |
| 2865 | dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx); |
| 2866 | if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) { |
| 2867 | mlog(ML_ERROR, "%s: received finalize1 from " |
| 2868 | "new master %u for dead node %u, but " |
| 2869 | "this node has already received it!\n", |
| 2870 | dlm->name, fr->node_idx, fr->dead_node); |
| 2871 | dlm_print_reco_node_status(dlm); |
| 2872 | BUG(); |
| 2873 | } |
| 2874 | dlm->reco.state |= DLM_RECO_STATE_FINALIZE; |
| 2875 | spin_unlock(&dlm->spinlock); |
| 2876 | break; |
| 2877 | case 2: |
| 2878 | if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) { |
| 2879 | mlog(ML_ERROR, "%s: received finalize2 from " |
| 2880 | "new master %u for dead node %u, but " |
| 2881 | "this node did not have finalize1!\n", |
| 2882 | dlm->name, fr->node_idx, fr->dead_node); |
| 2883 | dlm_print_reco_node_status(dlm); |
| 2884 | BUG(); |
| 2885 | } |
| 2886 | dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE; |
| 2887 | spin_unlock(&dlm->spinlock); |
| 2888 | dlm_reset_recovery(dlm); |
| 2889 | dlm_kick_recovery_thread(dlm); |
| 2890 | break; |
| 2891 | default: |
| 2892 | BUG(); |
| 2893 | } |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2894 | |
Kurt Hackel | d6dea6e | 2006-04-27 18:08:51 -0700 | [diff] [blame] | 2895 | mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n", |
| 2896 | dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master); |
| 2897 | |
Kurt Hackel | 6714d8e | 2005-12-15 14:31:23 -0800 | [diff] [blame] | 2898 | dlm_put(dlm); |
| 2899 | return 0; |
| 2900 | } |