1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* -*- mode: c; c-basic-offset: 8; -*- 3 * vim: noexpandtab sw=8 ts=8 sts=0: 4 * 5 * Copyright (C) 2004, 2005 Oracle. All rights reserved. 6 */ 7 8 #include <linux/kernel.h> 9 #include <linux/sched.h> 10 #include <linux/jiffies.h> 11 #include <linux/module.h> 12 #include <linux/fs.h> 13 #include <linux/bio.h> 14 #include <linux/blkdev.h> 15 #include <linux/delay.h> 16 #include <linux/file.h> 17 #include <linux/kthread.h> 18 #include <linux/configfs.h> 19 #include <linux/random.h> 20 #include <linux/crc32.h> 21 #include <linux/time.h> 22 #include <linux/debugfs.h> 23 #include <linux/slab.h> 24 #include <linux/bitmap.h> 25 #include <linux/ktime.h> 26 #include "heartbeat.h" 27 #include "tcp.h" 28 #include "nodemanager.h" 29 #include "quorum.h" 30 31 #include "masklog.h" 32 33 34 /* 35 * The first heartbeat pass had one global thread that would serialize all hb 36 * callback calls. This global serializing sem should only be removed once 37 * we've made sure that all callees can deal with being called concurrently 38 * from multiple hb region threads. 39 */ 40 static DECLARE_RWSEM(o2hb_callback_sem); 41 42 /* 43 * multiple hb threads are watching multiple regions. A node is live 44 * whenever any of the threads sees activity from the node in its region. 45 */ 46 static DEFINE_SPINLOCK(o2hb_live_lock); 47 static struct list_head o2hb_live_slots[O2NM_MAX_NODES]; 48 static unsigned long o2hb_live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)]; 49 static LIST_HEAD(o2hb_node_events); 50 static DECLARE_WAIT_QUEUE_HEAD(o2hb_steady_queue); 51 52 /* 53 * In global heartbeat, we maintain a series of region bitmaps. 54 * - o2hb_region_bitmap allows us to limit the region number to max region. 55 * - o2hb_live_region_bitmap tracks live regions (seen steady iterations). 56 * - o2hb_quorum_region_bitmap tracks live regions that have seen all nodes 57 * heartbeat on it. 58 * - o2hb_failed_region_bitmap tracks the regions that have seen io timeouts. 59 */ 60 static unsigned long o2hb_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)]; 61 static unsigned long o2hb_live_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)]; 62 static unsigned long o2hb_quorum_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)]; 63 static unsigned long o2hb_failed_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)]; 64 65 #define O2HB_DB_TYPE_LIVENODES 0 66 #define O2HB_DB_TYPE_LIVEREGIONS 1 67 #define O2HB_DB_TYPE_QUORUMREGIONS 2 68 #define O2HB_DB_TYPE_FAILEDREGIONS 3 69 #define O2HB_DB_TYPE_REGION_LIVENODES 4 70 #define O2HB_DB_TYPE_REGION_NUMBER 5 71 #define O2HB_DB_TYPE_REGION_ELAPSED_TIME 6 72 #define O2HB_DB_TYPE_REGION_PINNED 7 73 struct o2hb_debug_buf { 74 int db_type; 75 int db_size; 76 int db_len; 77 void *db_data; 78 }; 79 80 static struct o2hb_debug_buf *o2hb_db_livenodes; 81 static struct o2hb_debug_buf *o2hb_db_liveregions; 82 static struct o2hb_debug_buf *o2hb_db_quorumregions; 83 static struct o2hb_debug_buf *o2hb_db_failedregions; 84 85 #define O2HB_DEBUG_DIR "o2hb" 86 #define O2HB_DEBUG_LIVENODES "livenodes" 87 #define O2HB_DEBUG_LIVEREGIONS "live_regions" 88 #define O2HB_DEBUG_QUORUMREGIONS "quorum_regions" 89 #define O2HB_DEBUG_FAILEDREGIONS "failed_regions" 90 #define O2HB_DEBUG_REGION_NUMBER "num" 91 #define O2HB_DEBUG_REGION_ELAPSED_TIME "elapsed_time_in_ms" 92 #define O2HB_DEBUG_REGION_PINNED "pinned" 93 94 static struct dentry *o2hb_debug_dir; 95 96 static LIST_HEAD(o2hb_all_regions); 97 98 static struct o2hb_callback { 99 struct list_head list; 100 } o2hb_callbacks[O2HB_NUM_CB]; 101 102 static struct o2hb_callback *hbcall_from_type(enum o2hb_callback_type type); 103 104 enum o2hb_heartbeat_modes { 105 O2HB_HEARTBEAT_LOCAL = 0, 106 O2HB_HEARTBEAT_GLOBAL, 107 O2HB_HEARTBEAT_NUM_MODES, 108 }; 109 110 static const char *o2hb_heartbeat_mode_desc[O2HB_HEARTBEAT_NUM_MODES] = { 111 "local", /* O2HB_HEARTBEAT_LOCAL */ 112 "global", /* O2HB_HEARTBEAT_GLOBAL */ 113 }; 114 115 unsigned int o2hb_dead_threshold = O2HB_DEFAULT_DEAD_THRESHOLD; 116 static unsigned int o2hb_heartbeat_mode = O2HB_HEARTBEAT_LOCAL; 117 118 /* 119 * o2hb_dependent_users tracks the number of registered callbacks that depend 120 * on heartbeat. o2net and o2dlm are two entities that register this callback. 121 * However only o2dlm depends on the heartbeat. It does not want the heartbeat 122 * to stop while a dlm domain is still active. 123 */ 124 static unsigned int o2hb_dependent_users; 125 126 /* 127 * In global heartbeat mode, all regions are pinned if there are one or more 128 * dependent users and the quorum region count is <= O2HB_PIN_CUT_OFF. All 129 * regions are unpinned if the region count exceeds the cut off or the number 130 * of dependent users falls to zero. 131 */ 132 #define O2HB_PIN_CUT_OFF 3 133 134 /* 135 * In local heartbeat mode, we assume the dlm domain name to be the same as 136 * region uuid. This is true for domains created for the file system but not 137 * necessarily true for userdlm domains. This is a known limitation. 138 * 139 * In global heartbeat mode, we pin/unpin all o2hb regions. This solution 140 * works for both file system and userdlm domains. 141 */ 142 static int o2hb_region_pin(const char *region_uuid); 143 static void o2hb_region_unpin(const char *region_uuid); 144 145 /* Only sets a new threshold if there are no active regions. 146 * 147 * No locking or otherwise interesting code is required for reading 148 * o2hb_dead_threshold as it can't change once regions are active and 149 * it's not interesting to anyone until then anyway. */ 150 static void o2hb_dead_threshold_set(unsigned int threshold) 151 { 152 if (threshold > O2HB_MIN_DEAD_THRESHOLD) { 153 spin_lock(&o2hb_live_lock); 154 if (list_empty(&o2hb_all_regions)) 155 o2hb_dead_threshold = threshold; 156 spin_unlock(&o2hb_live_lock); 157 } 158 } 159 160 static int o2hb_global_heartbeat_mode_set(unsigned int hb_mode) 161 { 162 int ret = -1; 163 164 if (hb_mode < O2HB_HEARTBEAT_NUM_MODES) { 165 spin_lock(&o2hb_live_lock); 166 if (list_empty(&o2hb_all_regions)) { 167 o2hb_heartbeat_mode = hb_mode; 168 ret = 0; 169 } 170 spin_unlock(&o2hb_live_lock); 171 } 172 173 return ret; 174 } 175 176 struct o2hb_node_event { 177 struct list_head hn_item; 178 enum o2hb_callback_type hn_event_type; 179 struct o2nm_node *hn_node; 180 int hn_node_num; 181 }; 182 183 struct o2hb_disk_slot { 184 struct o2hb_disk_heartbeat_block *ds_raw_block; 185 u8 ds_node_num; 186 u64 ds_last_time; 187 u64 ds_last_generation; 188 u16 ds_equal_samples; 189 u16 ds_changed_samples; 190 struct list_head ds_live_item; 191 }; 192 193 /* each thread owns a region.. when we're asked to tear down the region 194 * we ask the thread to stop, who cleans up the region */ 195 struct o2hb_region { 196 struct config_item hr_item; 197 198 struct list_head hr_all_item; 199 unsigned hr_unclean_stop:1, 200 hr_aborted_start:1, 201 hr_item_pinned:1, 202 hr_item_dropped:1, 203 hr_node_deleted:1; 204 205 /* protected by the hr_callback_sem */ 206 struct task_struct *hr_task; 207 208 unsigned int hr_blocks; 209 unsigned long long hr_start_block; 210 211 unsigned int hr_block_bits; 212 unsigned int hr_block_bytes; 213 214 unsigned int hr_slots_per_page; 215 unsigned int hr_num_pages; 216 217 struct page **hr_slot_data; 218 struct block_device *hr_bdev; 219 struct o2hb_disk_slot *hr_slots; 220 221 /* live node map of this region */ 222 unsigned long hr_live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)]; 223 unsigned int hr_region_num; 224 225 struct dentry *hr_debug_dir; 226 struct o2hb_debug_buf *hr_db_livenodes; 227 struct o2hb_debug_buf *hr_db_regnum; 228 struct o2hb_debug_buf *hr_db_elapsed_time; 229 struct o2hb_debug_buf *hr_db_pinned; 230 231 /* let the person setting up hb wait for it to return until it 232 * has reached a 'steady' state. This will be fixed when we have 233 * a more complete api that doesn't lead to this sort of fragility. */ 234 atomic_t hr_steady_iterations; 235 236 /* terminate o2hb thread if it does not reach steady state 237 * (hr_steady_iterations == 0) within hr_unsteady_iterations */ 238 atomic_t hr_unsteady_iterations; 239 240 char hr_dev_name[BDEVNAME_SIZE]; 241 242 unsigned int hr_timeout_ms; 243 244 /* randomized as the region goes up and down so that a node 245 * recognizes a node going up and down in one iteration */ 246 u64 hr_generation; 247 248 struct delayed_work hr_write_timeout_work; 249 unsigned long hr_last_timeout_start; 250 251 /* negotiate timer, used to negotiate extending hb timeout. */ 252 struct delayed_work hr_nego_timeout_work; 253 unsigned long hr_nego_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)]; 254 255 /* Used during o2hb_check_slot to hold a copy of the block 256 * being checked because we temporarily have to zero out the 257 * crc field. */ 258 struct o2hb_disk_heartbeat_block *hr_tmp_block; 259 260 /* Message key for negotiate timeout message. */ 261 unsigned int hr_key; 262 struct list_head hr_handler_list; 263 264 /* last hb status, 0 for success, other value for error. */ 265 int hr_last_hb_status; 266 }; 267 268 struct o2hb_bio_wait_ctxt { 269 atomic_t wc_num_reqs; 270 struct completion wc_io_complete; 271 int wc_error; 272 }; 273 274 #define O2HB_NEGO_TIMEOUT_MS (O2HB_MAX_WRITE_TIMEOUT_MS/2) 275 276 enum { 277 O2HB_NEGO_TIMEOUT_MSG = 1, 278 O2HB_NEGO_APPROVE_MSG = 2, 279 }; 280 281 struct o2hb_nego_msg { 282 u8 node_num; 283 }; 284 285 static void o2hb_write_timeout(struct work_struct *work) 286 { 287 int failed, quorum; 288 struct o2hb_region *reg = 289 container_of(work, struct o2hb_region, 290 hr_write_timeout_work.work); 291 292 mlog(ML_ERROR, "Heartbeat write timeout to device %s after %u " 293 "milliseconds\n", reg->hr_dev_name, 294 jiffies_to_msecs(jiffies - reg->hr_last_timeout_start)); 295 296 if (o2hb_global_heartbeat_active()) { 297 spin_lock(&o2hb_live_lock); 298 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap)) 299 set_bit(reg->hr_region_num, o2hb_failed_region_bitmap); 300 failed = bitmap_weight(o2hb_failed_region_bitmap, 301 O2NM_MAX_REGIONS); 302 quorum = bitmap_weight(o2hb_quorum_region_bitmap, 303 O2NM_MAX_REGIONS); 304 spin_unlock(&o2hb_live_lock); 305 306 mlog(ML_HEARTBEAT, "Number of regions %d, failed regions %d\n", 307 quorum, failed); 308 309 /* 310 * Fence if the number of failed regions >= half the number 311 * of quorum regions 312 */ 313 if ((failed << 1) < quorum) 314 return; 315 } 316 317 o2quo_disk_timeout(); 318 } 319 320 static void o2hb_arm_timeout(struct o2hb_region *reg) 321 { 322 /* Arm writeout only after thread reaches steady state */ 323 if (atomic_read(®->hr_steady_iterations) != 0) 324 return; 325 326 mlog(ML_HEARTBEAT, "Queue write timeout for %u ms\n", 327 O2HB_MAX_WRITE_TIMEOUT_MS); 328 329 if (o2hb_global_heartbeat_active()) { 330 spin_lock(&o2hb_live_lock); 331 clear_bit(reg->hr_region_num, o2hb_failed_region_bitmap); 332 spin_unlock(&o2hb_live_lock); 333 } 334 cancel_delayed_work(®->hr_write_timeout_work); 335 schedule_delayed_work(®->hr_write_timeout_work, 336 msecs_to_jiffies(O2HB_MAX_WRITE_TIMEOUT_MS)); 337 338 cancel_delayed_work(®->hr_nego_timeout_work); 339 /* negotiate timeout must be less than write timeout. */ 340 schedule_delayed_work(®->hr_nego_timeout_work, 341 msecs_to_jiffies(O2HB_NEGO_TIMEOUT_MS)); 342 memset(reg->hr_nego_node_bitmap, 0, sizeof(reg->hr_nego_node_bitmap)); 343 } 344 345 static void o2hb_disarm_timeout(struct o2hb_region *reg) 346 { 347 cancel_delayed_work_sync(®->hr_write_timeout_work); 348 cancel_delayed_work_sync(®->hr_nego_timeout_work); 349 } 350 351 static int o2hb_send_nego_msg(int key, int type, u8 target) 352 { 353 struct o2hb_nego_msg msg; 354 int status, ret; 355 356 msg.node_num = o2nm_this_node(); 357 again: 358 ret = o2net_send_message(type, key, &msg, sizeof(msg), 359 target, &status); 360 361 if (ret == -EAGAIN || ret == -ENOMEM) { 362 msleep(100); 363 goto again; 364 } 365 366 return ret; 367 } 368 369 static void o2hb_nego_timeout(struct work_struct *work) 370 { 371 unsigned long live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)]; 372 int master_node, i, ret; 373 struct o2hb_region *reg; 374 375 reg = container_of(work, struct o2hb_region, hr_nego_timeout_work.work); 376 /* don't negotiate timeout if last hb failed since it is very 377 * possible io failed. Should let write timeout fence self. 378 */ 379 if (reg->hr_last_hb_status) 380 return; 381 382 o2hb_fill_node_map(live_node_bitmap, sizeof(live_node_bitmap)); 383 /* lowest node as master node to make negotiate decision. */ 384 master_node = find_next_bit(live_node_bitmap, O2NM_MAX_NODES, 0); 385 386 if (master_node == o2nm_this_node()) { 387 if (!test_bit(master_node, reg->hr_nego_node_bitmap)) { 388 printk(KERN_NOTICE "o2hb: node %d hb write hung for %ds on region %s (%s).\n", 389 o2nm_this_node(), O2HB_NEGO_TIMEOUT_MS/1000, 390 config_item_name(®->hr_item), reg->hr_dev_name); 391 set_bit(master_node, reg->hr_nego_node_bitmap); 392 } 393 if (memcmp(reg->hr_nego_node_bitmap, live_node_bitmap, 394 sizeof(reg->hr_nego_node_bitmap))) { 395 /* check negotiate bitmap every second to do timeout 396 * approve decision. 397 */ 398 schedule_delayed_work(®->hr_nego_timeout_work, 399 msecs_to_jiffies(1000)); 400 401 return; 402 } 403 404 printk(KERN_NOTICE "o2hb: all nodes hb write hung, maybe region %s (%s) is down.\n", 405 config_item_name(®->hr_item), reg->hr_dev_name); 406 /* approve negotiate timeout request. */ 407 o2hb_arm_timeout(reg); 408 409 i = -1; 410 while ((i = find_next_bit(live_node_bitmap, 411 O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) { 412 if (i == master_node) 413 continue; 414 415 mlog(ML_HEARTBEAT, "send NEGO_APPROVE msg to node %d\n", i); 416 ret = o2hb_send_nego_msg(reg->hr_key, 417 O2HB_NEGO_APPROVE_MSG, i); 418 if (ret) 419 mlog(ML_ERROR, "send NEGO_APPROVE msg to node %d fail %d\n", 420 i, ret); 421 } 422 } else { 423 /* negotiate timeout with master node. */ 424 printk(KERN_NOTICE "o2hb: node %d hb write hung for %ds on region %s (%s), negotiate timeout with node %d.\n", 425 o2nm_this_node(), O2HB_NEGO_TIMEOUT_MS/1000, config_item_name(®->hr_item), 426 reg->hr_dev_name, master_node); 427 ret = o2hb_send_nego_msg(reg->hr_key, O2HB_NEGO_TIMEOUT_MSG, 428 master_node); 429 if (ret) 430 mlog(ML_ERROR, "send NEGO_TIMEOUT msg to node %d fail %d\n", 431 master_node, ret); 432 } 433 } 434 435 static int o2hb_nego_timeout_handler(struct o2net_msg *msg, u32 len, void *data, 436 void **ret_data) 437 { 438 struct o2hb_region *reg = data; 439 struct o2hb_nego_msg *nego_msg; 440 441 nego_msg = (struct o2hb_nego_msg *)msg->buf; 442 printk(KERN_NOTICE "o2hb: receive negotiate timeout message from node %d on region %s (%s).\n", 443 nego_msg->node_num, config_item_name(®->hr_item), reg->hr_dev_name); 444 if (nego_msg->node_num < O2NM_MAX_NODES) 445 set_bit(nego_msg->node_num, reg->hr_nego_node_bitmap); 446 else 447 mlog(ML_ERROR, "got nego timeout message from bad node.\n"); 448 449 return 0; 450 } 451 452 static int o2hb_nego_approve_handler(struct o2net_msg *msg, u32 len, void *data, 453 void **ret_data) 454 { 455 struct o2hb_region *reg = data; 456 457 printk(KERN_NOTICE "o2hb: negotiate timeout approved by master node on region %s (%s).\n", 458 config_item_name(®->hr_item), reg->hr_dev_name); 459 o2hb_arm_timeout(reg); 460 return 0; 461 } 462 463 static inline void o2hb_bio_wait_init(struct o2hb_bio_wait_ctxt *wc) 464 { 465 atomic_set(&wc->wc_num_reqs, 1); 466 init_completion(&wc->wc_io_complete); 467 wc->wc_error = 0; 468 } 469 470 /* Used in error paths too */ 471 static inline void o2hb_bio_wait_dec(struct o2hb_bio_wait_ctxt *wc, 472 unsigned int num) 473 { 474 /* sadly atomic_sub_and_test() isn't available on all platforms. The 475 * good news is that the fast path only completes one at a time */ 476 while(num--) { 477 if (atomic_dec_and_test(&wc->wc_num_reqs)) { 478 BUG_ON(num > 0); 479 complete(&wc->wc_io_complete); 480 } 481 } 482 } 483 484 static void o2hb_wait_on_io(struct o2hb_bio_wait_ctxt *wc) 485 { 486 o2hb_bio_wait_dec(wc, 1); 487 wait_for_completion(&wc->wc_io_complete); 488 } 489 490 static void o2hb_bio_end_io(struct bio *bio) 491 { 492 struct o2hb_bio_wait_ctxt *wc = bio->bi_private; 493 494 if (bio->bi_status) { 495 mlog(ML_ERROR, "IO Error %d\n", bio->bi_status); 496 wc->wc_error = blk_status_to_errno(bio->bi_status); 497 } 498 499 o2hb_bio_wait_dec(wc, 1); 500 bio_put(bio); 501 } 502 503 /* Setup a Bio to cover I/O against num_slots slots starting at 504 * start_slot. */ 505 static struct bio *o2hb_setup_one_bio(struct o2hb_region *reg, 506 struct o2hb_bio_wait_ctxt *wc, 507 unsigned int *current_slot, 508 unsigned int max_slots, int op, 509 int op_flags) 510 { 511 int len, current_page; 512 unsigned int vec_len, vec_start; 513 unsigned int bits = reg->hr_block_bits; 514 unsigned int spp = reg->hr_slots_per_page; 515 unsigned int cs = *current_slot; 516 struct bio *bio; 517 struct page *page; 518 519 /* Testing has shown this allocation to take long enough under 520 * GFP_KERNEL that the local node can get fenced. It would be 521 * nicest if we could pre-allocate these bios and avoid this 522 * all together. */ 523 bio = bio_alloc(GFP_ATOMIC, 16); 524 if (!bio) { 525 mlog(ML_ERROR, "Could not alloc slots BIO!\n"); 526 bio = ERR_PTR(-ENOMEM); 527 goto bail; 528 } 529 530 /* Must put everything in 512 byte sectors for the bio... */ 531 bio->bi_iter.bi_sector = (reg->hr_start_block + cs) << (bits - 9); 532 bio_set_dev(bio, reg->hr_bdev); 533 bio->bi_private = wc; 534 bio->bi_end_io = o2hb_bio_end_io; 535 bio_set_op_attrs(bio, op, op_flags); 536 537 vec_start = (cs << bits) % PAGE_SIZE; 538 while(cs < max_slots) { 539 current_page = cs / spp; 540 page = reg->hr_slot_data[current_page]; 541 542 vec_len = min(PAGE_SIZE - vec_start, 543 (max_slots-cs) * (PAGE_SIZE/spp) ); 544 545 mlog(ML_HB_BIO, "page %d, vec_len = %u, vec_start = %u\n", 546 current_page, vec_len, vec_start); 547 548 len = bio_add_page(bio, page, vec_len, vec_start); 549 if (len != vec_len) break; 550 551 cs += vec_len / (PAGE_SIZE/spp); 552 vec_start = 0; 553 } 554 555 bail: 556 *current_slot = cs; 557 return bio; 558 } 559 560 static int o2hb_read_slots(struct o2hb_region *reg, 561 unsigned int begin_slot, 562 unsigned int max_slots) 563 { 564 unsigned int current_slot = begin_slot; 565 int status; 566 struct o2hb_bio_wait_ctxt wc; 567 struct bio *bio; 568 569 o2hb_bio_wait_init(&wc); 570 571 while(current_slot < max_slots) { 572 bio = o2hb_setup_one_bio(reg, &wc, ¤t_slot, max_slots, 573 REQ_OP_READ, 0); 574 if (IS_ERR(bio)) { 575 status = PTR_ERR(bio); 576 mlog_errno(status); 577 goto bail_and_wait; 578 } 579 580 atomic_inc(&wc.wc_num_reqs); 581 submit_bio(bio); 582 } 583 584 status = 0; 585 586 bail_and_wait: 587 o2hb_wait_on_io(&wc); 588 if (wc.wc_error && !status) 589 status = wc.wc_error; 590 591 return status; 592 } 593 594 static int o2hb_issue_node_write(struct o2hb_region *reg, 595 struct o2hb_bio_wait_ctxt *write_wc) 596 { 597 int status; 598 unsigned int slot; 599 struct bio *bio; 600 601 o2hb_bio_wait_init(write_wc); 602 603 slot = o2nm_this_node(); 604 605 bio = o2hb_setup_one_bio(reg, write_wc, &slot, slot+1, REQ_OP_WRITE, 606 REQ_SYNC); 607 if (IS_ERR(bio)) { 608 status = PTR_ERR(bio); 609 mlog_errno(status); 610 goto bail; 611 } 612 613 atomic_inc(&write_wc->wc_num_reqs); 614 submit_bio(bio); 615 616 status = 0; 617 bail: 618 return status; 619 } 620 621 static u32 o2hb_compute_block_crc_le(struct o2hb_region *reg, 622 struct o2hb_disk_heartbeat_block *hb_block) 623 { 624 __le32 old_cksum; 625 u32 ret; 626 627 /* We want to compute the block crc with a 0 value in the 628 * hb_cksum field. Save it off here and replace after the 629 * crc. */ 630 old_cksum = hb_block->hb_cksum; 631 hb_block->hb_cksum = 0; 632 633 ret = crc32_le(0, (unsigned char *) hb_block, reg->hr_block_bytes); 634 635 hb_block->hb_cksum = old_cksum; 636 637 return ret; 638 } 639 640 static void o2hb_dump_slot(struct o2hb_disk_heartbeat_block *hb_block) 641 { 642 mlog(ML_ERROR, "Dump slot information: seq = 0x%llx, node = %u, " 643 "cksum = 0x%x, generation 0x%llx\n", 644 (long long)le64_to_cpu(hb_block->hb_seq), 645 hb_block->hb_node, le32_to_cpu(hb_block->hb_cksum), 646 (long long)le64_to_cpu(hb_block->hb_generation)); 647 } 648 649 static int o2hb_verify_crc(struct o2hb_region *reg, 650 struct o2hb_disk_heartbeat_block *hb_block) 651 { 652 u32 read, computed; 653 654 read = le32_to_cpu(hb_block->hb_cksum); 655 computed = o2hb_compute_block_crc_le(reg, hb_block); 656 657 return read == computed; 658 } 659 660 /* 661 * Compare the slot data with what we wrote in the last iteration. 662 * If the match fails, print an appropriate error message. This is to 663 * detect errors like... another node hearting on the same slot, 664 * flaky device that is losing writes, etc. 665 * Returns 1 if check succeeds, 0 otherwise. 666 */ 667 static int o2hb_check_own_slot(struct o2hb_region *reg) 668 { 669 struct o2hb_disk_slot *slot; 670 struct o2hb_disk_heartbeat_block *hb_block; 671 char *errstr; 672 673 slot = ®->hr_slots[o2nm_this_node()]; 674 /* Don't check on our 1st timestamp */ 675 if (!slot->ds_last_time) 676 return 0; 677 678 hb_block = slot->ds_raw_block; 679 if (le64_to_cpu(hb_block->hb_seq) == slot->ds_last_time && 680 le64_to_cpu(hb_block->hb_generation) == slot->ds_last_generation && 681 hb_block->hb_node == slot->ds_node_num) 682 return 1; 683 684 #define ERRSTR1 "Another node is heartbeating on device" 685 #define ERRSTR2 "Heartbeat generation mismatch on device" 686 #define ERRSTR3 "Heartbeat sequence mismatch on device" 687 688 if (hb_block->hb_node != slot->ds_node_num) 689 errstr = ERRSTR1; 690 else if (le64_to_cpu(hb_block->hb_generation) != 691 slot->ds_last_generation) 692 errstr = ERRSTR2; 693 else 694 errstr = ERRSTR3; 695 696 mlog(ML_ERROR, "%s (%s): expected(%u:0x%llx, 0x%llx), " 697 "ondisk(%u:0x%llx, 0x%llx)\n", errstr, reg->hr_dev_name, 698 slot->ds_node_num, (unsigned long long)slot->ds_last_generation, 699 (unsigned long long)slot->ds_last_time, hb_block->hb_node, 700 (unsigned long long)le64_to_cpu(hb_block->hb_generation), 701 (unsigned long long)le64_to_cpu(hb_block->hb_seq)); 702 703 return 0; 704 } 705 706 static inline void o2hb_prepare_block(struct o2hb_region *reg, 707 u64 generation) 708 { 709 int node_num; 710 u64 cputime; 711 struct o2hb_disk_slot *slot; 712 struct o2hb_disk_heartbeat_block *hb_block; 713 714 node_num = o2nm_this_node(); 715 slot = ®->hr_slots[node_num]; 716 717 hb_block = (struct o2hb_disk_heartbeat_block *)slot->ds_raw_block; 718 memset(hb_block, 0, reg->hr_block_bytes); 719 /* TODO: time stuff */ 720 cputime = ktime_get_real_seconds(); 721 if (!cputime) 722 cputime = 1; 723 724 hb_block->hb_seq = cpu_to_le64(cputime); 725 hb_block->hb_node = node_num; 726 hb_block->hb_generation = cpu_to_le64(generation); 727 hb_block->hb_dead_ms = cpu_to_le32(o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS); 728 729 /* This step must always happen last! */ 730 hb_block->hb_cksum = cpu_to_le32(o2hb_compute_block_crc_le(reg, 731 hb_block)); 732 733 mlog(ML_HB_BIO, "our node generation = 0x%llx, cksum = 0x%x\n", 734 (long long)generation, 735 le32_to_cpu(hb_block->hb_cksum)); 736 } 737 738 static void o2hb_fire_callbacks(struct o2hb_callback *hbcall, 739 struct o2nm_node *node, 740 int idx) 741 { 742 struct o2hb_callback_func *f; 743 744 list_for_each_entry(f, &hbcall->list, hc_item) { 745 mlog(ML_HEARTBEAT, "calling funcs %p\n", f); 746 (f->hc_func)(node, idx, f->hc_data); 747 } 748 } 749 750 /* Will run the list in order until we process the passed event */ 751 static void o2hb_run_event_list(struct o2hb_node_event *queued_event) 752 { 753 struct o2hb_callback *hbcall; 754 struct o2hb_node_event *event; 755 756 /* Holding callback sem assures we don't alter the callback 757 * lists when doing this, and serializes ourselves with other 758 * processes wanting callbacks. */ 759 down_write(&o2hb_callback_sem); 760 761 spin_lock(&o2hb_live_lock); 762 while (!list_empty(&o2hb_node_events) 763 && !list_empty(&queued_event->hn_item)) { 764 event = list_entry(o2hb_node_events.next, 765 struct o2hb_node_event, 766 hn_item); 767 list_del_init(&event->hn_item); 768 spin_unlock(&o2hb_live_lock); 769 770 mlog(ML_HEARTBEAT, "Node %s event for %d\n", 771 event->hn_event_type == O2HB_NODE_UP_CB ? "UP" : "DOWN", 772 event->hn_node_num); 773 774 hbcall = hbcall_from_type(event->hn_event_type); 775 776 /* We should *never* have gotten on to the list with a 777 * bad type... This isn't something that we should try 778 * to recover from. */ 779 BUG_ON(IS_ERR(hbcall)); 780 781 o2hb_fire_callbacks(hbcall, event->hn_node, event->hn_node_num); 782 783 spin_lock(&o2hb_live_lock); 784 } 785 spin_unlock(&o2hb_live_lock); 786 787 up_write(&o2hb_callback_sem); 788 } 789 790 static void o2hb_queue_node_event(struct o2hb_node_event *event, 791 enum o2hb_callback_type type, 792 struct o2nm_node *node, 793 int node_num) 794 { 795 assert_spin_locked(&o2hb_live_lock); 796 797 BUG_ON((!node) && (type != O2HB_NODE_DOWN_CB)); 798 799 event->hn_event_type = type; 800 event->hn_node = node; 801 event->hn_node_num = node_num; 802 803 mlog(ML_HEARTBEAT, "Queue node %s event for node %d\n", 804 type == O2HB_NODE_UP_CB ? "UP" : "DOWN", node_num); 805 806 list_add_tail(&event->hn_item, &o2hb_node_events); 807 } 808 809 static void o2hb_shutdown_slot(struct o2hb_disk_slot *slot) 810 { 811 struct o2hb_node_event event = 812 { .hn_item = LIST_HEAD_INIT(event.hn_item), }; 813 struct o2nm_node *node; 814 int queued = 0; 815 816 node = o2nm_get_node_by_num(slot->ds_node_num); 817 if (!node) 818 return; 819 820 spin_lock(&o2hb_live_lock); 821 if (!list_empty(&slot->ds_live_item)) { 822 mlog(ML_HEARTBEAT, "Shutdown, node %d leaves region\n", 823 slot->ds_node_num); 824 825 list_del_init(&slot->ds_live_item); 826 827 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) { 828 clear_bit(slot->ds_node_num, o2hb_live_node_bitmap); 829 830 o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB, node, 831 slot->ds_node_num); 832 queued = 1; 833 } 834 } 835 spin_unlock(&o2hb_live_lock); 836 837 if (queued) 838 o2hb_run_event_list(&event); 839 840 o2nm_node_put(node); 841 } 842 843 static void o2hb_set_quorum_device(struct o2hb_region *reg) 844 { 845 if (!o2hb_global_heartbeat_active()) 846 return; 847 848 /* Prevent race with o2hb_heartbeat_group_drop_item() */ 849 if (kthread_should_stop()) 850 return; 851 852 /* Tag region as quorum only after thread reaches steady state */ 853 if (atomic_read(®->hr_steady_iterations) != 0) 854 return; 855 856 spin_lock(&o2hb_live_lock); 857 858 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap)) 859 goto unlock; 860 861 /* 862 * A region can be added to the quorum only when it sees all 863 * live nodes heartbeat on it. In other words, the region has been 864 * added to all nodes. 865 */ 866 if (memcmp(reg->hr_live_node_bitmap, o2hb_live_node_bitmap, 867 sizeof(o2hb_live_node_bitmap))) 868 goto unlock; 869 870 printk(KERN_NOTICE "o2hb: Region %s (%s) is now a quorum device\n", 871 config_item_name(®->hr_item), reg->hr_dev_name); 872 873 set_bit(reg->hr_region_num, o2hb_quorum_region_bitmap); 874 875 /* 876 * If global heartbeat active, unpin all regions if the 877 * region count > CUT_OFF 878 */ 879 if (bitmap_weight(o2hb_quorum_region_bitmap, 880 O2NM_MAX_REGIONS) > O2HB_PIN_CUT_OFF) 881 o2hb_region_unpin(NULL); 882 unlock: 883 spin_unlock(&o2hb_live_lock); 884 } 885 886 static int o2hb_check_slot(struct o2hb_region *reg, 887 struct o2hb_disk_slot *slot) 888 { 889 int changed = 0, gen_changed = 0; 890 struct o2hb_node_event event = 891 { .hn_item = LIST_HEAD_INIT(event.hn_item), }; 892 struct o2nm_node *node; 893 struct o2hb_disk_heartbeat_block *hb_block = reg->hr_tmp_block; 894 u64 cputime; 895 unsigned int dead_ms = o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS; 896 unsigned int slot_dead_ms; 897 int tmp; 898 int queued = 0; 899 900 memcpy(hb_block, slot->ds_raw_block, reg->hr_block_bytes); 901 902 /* 903 * If a node is no longer configured but is still in the livemap, we 904 * may need to clear that bit from the livemap. 905 */ 906 node = o2nm_get_node_by_num(slot->ds_node_num); 907 if (!node) { 908 spin_lock(&o2hb_live_lock); 909 tmp = test_bit(slot->ds_node_num, o2hb_live_node_bitmap); 910 spin_unlock(&o2hb_live_lock); 911 if (!tmp) 912 return 0; 913 } 914 915 if (!o2hb_verify_crc(reg, hb_block)) { 916 /* all paths from here will drop o2hb_live_lock for 917 * us. */ 918 spin_lock(&o2hb_live_lock); 919 920 /* Don't print an error on the console in this case - 921 * a freshly formatted heartbeat area will not have a 922 * crc set on it. */ 923 if (list_empty(&slot->ds_live_item)) 924 goto out; 925 926 /* The node is live but pushed out a bad crc. We 927 * consider it a transient miss but don't populate any 928 * other values as they may be junk. */ 929 mlog(ML_ERROR, "Node %d has written a bad crc to %s\n", 930 slot->ds_node_num, reg->hr_dev_name); 931 o2hb_dump_slot(hb_block); 932 933 slot->ds_equal_samples++; 934 goto fire_callbacks; 935 } 936 937 /* we don't care if these wrap.. the state transitions below 938 * clear at the right places */ 939 cputime = le64_to_cpu(hb_block->hb_seq); 940 if (slot->ds_last_time != cputime) 941 slot->ds_changed_samples++; 942 else 943 slot->ds_equal_samples++; 944 slot->ds_last_time = cputime; 945 946 /* The node changed heartbeat generations. We assume this to 947 * mean it dropped off but came back before we timed out. We 948 * want to consider it down for the time being but don't want 949 * to lose any changed_samples state we might build up to 950 * considering it live again. */ 951 if (slot->ds_last_generation != le64_to_cpu(hb_block->hb_generation)) { 952 gen_changed = 1; 953 slot->ds_equal_samples = 0; 954 mlog(ML_HEARTBEAT, "Node %d changed generation (0x%llx " 955 "to 0x%llx)\n", slot->ds_node_num, 956 (long long)slot->ds_last_generation, 957 (long long)le64_to_cpu(hb_block->hb_generation)); 958 } 959 960 slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation); 961 962 mlog(ML_HEARTBEAT, "Slot %d gen 0x%llx cksum 0x%x " 963 "seq %llu last %llu changed %u equal %u\n", 964 slot->ds_node_num, (long long)slot->ds_last_generation, 965 le32_to_cpu(hb_block->hb_cksum), 966 (unsigned long long)le64_to_cpu(hb_block->hb_seq), 967 (unsigned long long)slot->ds_last_time, slot->ds_changed_samples, 968 slot->ds_equal_samples); 969 970 spin_lock(&o2hb_live_lock); 971 972 fire_callbacks: 973 /* dead nodes only come to life after some number of 974 * changes at any time during their dead time */ 975 if (list_empty(&slot->ds_live_item) && 976 slot->ds_changed_samples >= O2HB_LIVE_THRESHOLD) { 977 mlog(ML_HEARTBEAT, "Node %d (id 0x%llx) joined my region\n", 978 slot->ds_node_num, (long long)slot->ds_last_generation); 979 980 set_bit(slot->ds_node_num, reg->hr_live_node_bitmap); 981 982 /* first on the list generates a callback */ 983 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) { 984 mlog(ML_HEARTBEAT, "o2hb: Add node %d to live nodes " 985 "bitmap\n", slot->ds_node_num); 986 set_bit(slot->ds_node_num, o2hb_live_node_bitmap); 987 988 o2hb_queue_node_event(&event, O2HB_NODE_UP_CB, node, 989 slot->ds_node_num); 990 991 changed = 1; 992 queued = 1; 993 } 994 995 list_add_tail(&slot->ds_live_item, 996 &o2hb_live_slots[slot->ds_node_num]); 997 998 slot->ds_equal_samples = 0; 999 1000 /* We want to be sure that all nodes agree on the 1001 * number of milliseconds before a node will be 1002 * considered dead. The self-fencing timeout is 1003 * computed from this value, and a discrepancy might 1004 * result in heartbeat calling a node dead when it 1005 * hasn't self-fenced yet. */ 1006 slot_dead_ms = le32_to_cpu(hb_block->hb_dead_ms); 1007 if (slot_dead_ms && slot_dead_ms != dead_ms) { 1008 /* TODO: Perhaps we can fail the region here. */ 1009 mlog(ML_ERROR, "Node %d on device %s has a dead count " 1010 "of %u ms, but our count is %u ms.\n" 1011 "Please double check your configuration values " 1012 "for 'O2CB_HEARTBEAT_THRESHOLD'\n", 1013 slot->ds_node_num, reg->hr_dev_name, slot_dead_ms, 1014 dead_ms); 1015 } 1016 goto out; 1017 } 1018 1019 /* if the list is dead, we're done.. */ 1020 if (list_empty(&slot->ds_live_item)) 1021 goto out; 1022 1023 /* live nodes only go dead after enough consequtive missed 1024 * samples.. reset the missed counter whenever we see 1025 * activity */ 1026 if (slot->ds_equal_samples >= o2hb_dead_threshold || gen_changed) { 1027 mlog(ML_HEARTBEAT, "Node %d left my region\n", 1028 slot->ds_node_num); 1029 1030 clear_bit(slot->ds_node_num, reg->hr_live_node_bitmap); 1031 1032 /* last off the live_slot generates a callback */ 1033 list_del_init(&slot->ds_live_item); 1034 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) { 1035 mlog(ML_HEARTBEAT, "o2hb: Remove node %d from live " 1036 "nodes bitmap\n", slot->ds_node_num); 1037 clear_bit(slot->ds_node_num, o2hb_live_node_bitmap); 1038 1039 /* node can be null */ 1040 o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB, 1041 node, slot->ds_node_num); 1042 1043 changed = 1; 1044 queued = 1; 1045 } 1046 1047 /* We don't clear this because the node is still 1048 * actually writing new blocks. */ 1049 if (!gen_changed) 1050 slot->ds_changed_samples = 0; 1051 goto out; 1052 } 1053 if (slot->ds_changed_samples) { 1054 slot->ds_changed_samples = 0; 1055 slot->ds_equal_samples = 0; 1056 } 1057 out: 1058 spin_unlock(&o2hb_live_lock); 1059 1060 if (queued) 1061 o2hb_run_event_list(&event); 1062 1063 if (node) 1064 o2nm_node_put(node); 1065 return changed; 1066 } 1067 1068 static int o2hb_highest_node(unsigned long *nodes, int numbits) 1069 { 1070 return find_last_bit(nodes, numbits); 1071 } 1072 1073 static int o2hb_lowest_node(unsigned long *nodes, int numbits) 1074 { 1075 return find_first_bit(nodes, numbits); 1076 } 1077 1078 static int o2hb_do_disk_heartbeat(struct o2hb_region *reg) 1079 { 1080 int i, ret, highest_node, lowest_node; 1081 int membership_change = 0, own_slot_ok = 0; 1082 unsigned long configured_nodes[BITS_TO_LONGS(O2NM_MAX_NODES)]; 1083 unsigned long live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)]; 1084 struct o2hb_bio_wait_ctxt write_wc; 1085 1086 ret = o2nm_configured_node_map(configured_nodes, 1087 sizeof(configured_nodes)); 1088 if (ret) { 1089 mlog_errno(ret); 1090 goto bail; 1091 } 1092 1093 /* 1094 * If a node is not configured but is in the livemap, we still need 1095 * to read the slot so as to be able to remove it from the livemap. 1096 */ 1097 o2hb_fill_node_map(live_node_bitmap, sizeof(live_node_bitmap)); 1098 i = -1; 1099 while ((i = find_next_bit(live_node_bitmap, 1100 O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) { 1101 set_bit(i, configured_nodes); 1102 } 1103 1104 highest_node = o2hb_highest_node(configured_nodes, O2NM_MAX_NODES); 1105 lowest_node = o2hb_lowest_node(configured_nodes, O2NM_MAX_NODES); 1106 if (highest_node >= O2NM_MAX_NODES || lowest_node >= O2NM_MAX_NODES) { 1107 mlog(ML_NOTICE, "o2hb: No configured nodes found!\n"); 1108 ret = -EINVAL; 1109 goto bail; 1110 } 1111 1112 /* No sense in reading the slots of nodes that don't exist 1113 * yet. Of course, if the node definitions have holes in them 1114 * then we're reading an empty slot anyway... Consider this 1115 * best-effort. */ 1116 ret = o2hb_read_slots(reg, lowest_node, highest_node + 1); 1117 if (ret < 0) { 1118 mlog_errno(ret); 1119 goto bail; 1120 } 1121 1122 /* With an up to date view of the slots, we can check that no 1123 * other node has been improperly configured to heartbeat in 1124 * our slot. */ 1125 own_slot_ok = o2hb_check_own_slot(reg); 1126 1127 /* fill in the proper info for our next heartbeat */ 1128 o2hb_prepare_block(reg, reg->hr_generation); 1129 1130 ret = o2hb_issue_node_write(reg, &write_wc); 1131 if (ret < 0) { 1132 mlog_errno(ret); 1133 goto bail; 1134 } 1135 1136 i = -1; 1137 while((i = find_next_bit(configured_nodes, 1138 O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) { 1139 membership_change |= o2hb_check_slot(reg, ®->hr_slots[i]); 1140 } 1141 1142 /* 1143 * We have to be sure we've advertised ourselves on disk 1144 * before we can go to steady state. This ensures that 1145 * people we find in our steady state have seen us. 1146 */ 1147 o2hb_wait_on_io(&write_wc); 1148 if (write_wc.wc_error) { 1149 /* Do not re-arm the write timeout on I/O error - we 1150 * can't be sure that the new block ever made it to 1151 * disk */ 1152 mlog(ML_ERROR, "Write error %d on device \"%s\"\n", 1153 write_wc.wc_error, reg->hr_dev_name); 1154 ret = write_wc.wc_error; 1155 goto bail; 1156 } 1157 1158 /* Skip disarming the timeout if own slot has stale/bad data */ 1159 if (own_slot_ok) { 1160 o2hb_set_quorum_device(reg); 1161 o2hb_arm_timeout(reg); 1162 reg->hr_last_timeout_start = jiffies; 1163 } 1164 1165 bail: 1166 /* let the person who launched us know when things are steady */ 1167 if (atomic_read(®->hr_steady_iterations) != 0) { 1168 if (!ret && own_slot_ok && !membership_change) { 1169 if (atomic_dec_and_test(®->hr_steady_iterations)) 1170 wake_up(&o2hb_steady_queue); 1171 } 1172 } 1173 1174 if (atomic_read(®->hr_steady_iterations) != 0) { 1175 if (atomic_dec_and_test(®->hr_unsteady_iterations)) { 1176 printk(KERN_NOTICE "o2hb: Unable to stabilize " 1177 "heartbeat on region %s (%s)\n", 1178 config_item_name(®->hr_item), 1179 reg->hr_dev_name); 1180 atomic_set(®->hr_steady_iterations, 0); 1181 reg->hr_aborted_start = 1; 1182 wake_up(&o2hb_steady_queue); 1183 ret = -EIO; 1184 } 1185 } 1186 1187 return ret; 1188 } 1189 1190 /* 1191 * we ride the region ref that the region dir holds. before the region 1192 * dir is removed and drops it ref it will wait to tear down this 1193 * thread. 1194 */ 1195 static int o2hb_thread(void *data) 1196 { 1197 int i, ret; 1198 struct o2hb_region *reg = data; 1199 struct o2hb_bio_wait_ctxt write_wc; 1200 ktime_t before_hb, after_hb; 1201 unsigned int elapsed_msec; 1202 1203 mlog(ML_HEARTBEAT|ML_KTHREAD, "hb thread running\n"); 1204 1205 set_user_nice(current, MIN_NICE); 1206 1207 /* Pin node */ 1208 ret = o2nm_depend_this_node(); 1209 if (ret) { 1210 mlog(ML_ERROR, "Node has been deleted, ret = %d\n", ret); 1211 reg->hr_node_deleted = 1; 1212 wake_up(&o2hb_steady_queue); 1213 return 0; 1214 } 1215 1216 while (!kthread_should_stop() && 1217 !reg->hr_unclean_stop && !reg->hr_aborted_start) { 1218 /* We track the time spent inside 1219 * o2hb_do_disk_heartbeat so that we avoid more than 1220 * hr_timeout_ms between disk writes. On busy systems 1221 * this should result in a heartbeat which is less 1222 * likely to time itself out. */ 1223 before_hb = ktime_get_real(); 1224 1225 ret = o2hb_do_disk_heartbeat(reg); 1226 reg->hr_last_hb_status = ret; 1227 1228 after_hb = ktime_get_real(); 1229 1230 elapsed_msec = (unsigned int) 1231 ktime_ms_delta(after_hb, before_hb); 1232 1233 mlog(ML_HEARTBEAT, 1234 "start = %lld, end = %lld, msec = %u, ret = %d\n", 1235 before_hb, after_hb, elapsed_msec, ret); 1236 1237 if (!kthread_should_stop() && 1238 elapsed_msec < reg->hr_timeout_ms) { 1239 /* the kthread api has blocked signals for us so no 1240 * need to record the return value. */ 1241 msleep_interruptible(reg->hr_timeout_ms - elapsed_msec); 1242 } 1243 } 1244 1245 o2hb_disarm_timeout(reg); 1246 1247 /* unclean stop is only used in very bad situation */ 1248 for(i = 0; !reg->hr_unclean_stop && i < reg->hr_blocks; i++) 1249 o2hb_shutdown_slot(®->hr_slots[i]); 1250 1251 /* Explicit down notification - avoid forcing the other nodes 1252 * to timeout on this region when we could just as easily 1253 * write a clear generation - thus indicating to them that 1254 * this node has left this region. 1255 */ 1256 if (!reg->hr_unclean_stop && !reg->hr_aborted_start) { 1257 o2hb_prepare_block(reg, 0); 1258 ret = o2hb_issue_node_write(reg, &write_wc); 1259 if (ret == 0) 1260 o2hb_wait_on_io(&write_wc); 1261 else 1262 mlog_errno(ret); 1263 } 1264 1265 /* Unpin node */ 1266 o2nm_undepend_this_node(); 1267 1268 mlog(ML_HEARTBEAT|ML_KTHREAD, "o2hb thread exiting\n"); 1269 1270 return 0; 1271 } 1272 1273 #ifdef CONFIG_DEBUG_FS 1274 static int o2hb_debug_open(struct inode *inode, struct file *file) 1275 { 1276 struct o2hb_debug_buf *db = inode->i_private; 1277 struct o2hb_region *reg; 1278 unsigned long map[BITS_TO_LONGS(O2NM_MAX_NODES)]; 1279 unsigned long lts; 1280 char *buf = NULL; 1281 int i = -1; 1282 int out = 0; 1283 1284 /* max_nodes should be the largest bitmap we pass here */ 1285 BUG_ON(sizeof(map) < db->db_size); 1286 1287 buf = kmalloc(PAGE_SIZE, GFP_KERNEL); 1288 if (!buf) 1289 goto bail; 1290 1291 switch (db->db_type) { 1292 case O2HB_DB_TYPE_LIVENODES: 1293 case O2HB_DB_TYPE_LIVEREGIONS: 1294 case O2HB_DB_TYPE_QUORUMREGIONS: 1295 case O2HB_DB_TYPE_FAILEDREGIONS: 1296 spin_lock(&o2hb_live_lock); 1297 memcpy(map, db->db_data, db->db_size); 1298 spin_unlock(&o2hb_live_lock); 1299 break; 1300 1301 case O2HB_DB_TYPE_REGION_LIVENODES: 1302 spin_lock(&o2hb_live_lock); 1303 reg = (struct o2hb_region *)db->db_data; 1304 memcpy(map, reg->hr_live_node_bitmap, db->db_size); 1305 spin_unlock(&o2hb_live_lock); 1306 break; 1307 1308 case O2HB_DB_TYPE_REGION_NUMBER: 1309 reg = (struct o2hb_region *)db->db_data; 1310 out += scnprintf(buf + out, PAGE_SIZE - out, "%d\n", 1311 reg->hr_region_num); 1312 goto done; 1313 1314 case O2HB_DB_TYPE_REGION_ELAPSED_TIME: 1315 reg = (struct o2hb_region *)db->db_data; 1316 lts = reg->hr_last_timeout_start; 1317 /* If 0, it has never been set before */ 1318 if (lts) 1319 lts = jiffies_to_msecs(jiffies - lts); 1320 out += scnprintf(buf + out, PAGE_SIZE - out, "%lu\n", lts); 1321 goto done; 1322 1323 case O2HB_DB_TYPE_REGION_PINNED: 1324 reg = (struct o2hb_region *)db->db_data; 1325 out += scnprintf(buf + out, PAGE_SIZE - out, "%u\n", 1326 !!reg->hr_item_pinned); 1327 goto done; 1328 1329 default: 1330 goto done; 1331 } 1332 1333 while ((i = find_next_bit(map, db->db_len, i + 1)) < db->db_len) 1334 out += scnprintf(buf + out, PAGE_SIZE - out, "%d ", i); 1335 out += scnprintf(buf + out, PAGE_SIZE - out, "\n"); 1336 1337 done: 1338 i_size_write(inode, out); 1339 1340 file->private_data = buf; 1341 1342 return 0; 1343 bail: 1344 return -ENOMEM; 1345 } 1346 1347 static int o2hb_debug_release(struct inode *inode, struct file *file) 1348 { 1349 kfree(file->private_data); 1350 return 0; 1351 } 1352 1353 static ssize_t o2hb_debug_read(struct file *file, char __user *buf, 1354 size_t nbytes, loff_t *ppos) 1355 { 1356 return simple_read_from_buffer(buf, nbytes, ppos, file->private_data, 1357 i_size_read(file->f_mapping->host)); 1358 } 1359 #else 1360 static int o2hb_debug_open(struct inode *inode, struct file *file) 1361 { 1362 return 0; 1363 } 1364 static int o2hb_debug_release(struct inode *inode, struct file *file) 1365 { 1366 return 0; 1367 } 1368 static ssize_t o2hb_debug_read(struct file *file, char __user *buf, 1369 size_t nbytes, loff_t *ppos) 1370 { 1371 return 0; 1372 } 1373 #endif /* CONFIG_DEBUG_FS */ 1374 1375 static const struct file_operations o2hb_debug_fops = { 1376 .open = o2hb_debug_open, 1377 .release = o2hb_debug_release, 1378 .read = o2hb_debug_read, 1379 .llseek = generic_file_llseek, 1380 }; 1381 1382 void o2hb_exit(void) 1383 { 1384 debugfs_remove_recursive(o2hb_debug_dir); 1385 kfree(o2hb_db_livenodes); 1386 kfree(o2hb_db_liveregions); 1387 kfree(o2hb_db_quorumregions); 1388 kfree(o2hb_db_failedregions); 1389 } 1390 1391 static void o2hb_debug_create(const char *name, struct dentry *dir, 1392 struct o2hb_debug_buf **db, int db_len, int type, 1393 int size, int len, void *data) 1394 { 1395 *db = kmalloc(db_len, GFP_KERNEL); 1396 if (!*db) 1397 return; 1398 1399 (*db)->db_type = type; 1400 (*db)->db_size = size; 1401 (*db)->db_len = len; 1402 (*db)->db_data = data; 1403 1404 debugfs_create_file(name, S_IFREG|S_IRUSR, dir, *db, &o2hb_debug_fops); 1405 } 1406 1407 static void o2hb_debug_init(void) 1408 { 1409 o2hb_debug_dir = debugfs_create_dir(O2HB_DEBUG_DIR, NULL); 1410 1411 o2hb_debug_create(O2HB_DEBUG_LIVENODES, o2hb_debug_dir, 1412 &o2hb_db_livenodes, sizeof(*o2hb_db_livenodes), 1413 O2HB_DB_TYPE_LIVENODES, sizeof(o2hb_live_node_bitmap), 1414 O2NM_MAX_NODES, o2hb_live_node_bitmap); 1415 1416 o2hb_debug_create(O2HB_DEBUG_LIVEREGIONS, o2hb_debug_dir, 1417 &o2hb_db_liveregions, sizeof(*o2hb_db_liveregions), 1418 O2HB_DB_TYPE_LIVEREGIONS, 1419 sizeof(o2hb_live_region_bitmap), O2NM_MAX_REGIONS, 1420 o2hb_live_region_bitmap); 1421 1422 o2hb_debug_create(O2HB_DEBUG_QUORUMREGIONS, o2hb_debug_dir, 1423 &o2hb_db_quorumregions, 1424 sizeof(*o2hb_db_quorumregions), 1425 O2HB_DB_TYPE_QUORUMREGIONS, 1426 sizeof(o2hb_quorum_region_bitmap), O2NM_MAX_REGIONS, 1427 o2hb_quorum_region_bitmap); 1428 1429 o2hb_debug_create(O2HB_DEBUG_FAILEDREGIONS, o2hb_debug_dir, 1430 &o2hb_db_failedregions, 1431 sizeof(*o2hb_db_failedregions), 1432 O2HB_DB_TYPE_FAILEDREGIONS, 1433 sizeof(o2hb_failed_region_bitmap), O2NM_MAX_REGIONS, 1434 o2hb_failed_region_bitmap); 1435 } 1436 1437 void o2hb_init(void) 1438 { 1439 int i; 1440 1441 for (i = 0; i < ARRAY_SIZE(o2hb_callbacks); i++) 1442 INIT_LIST_HEAD(&o2hb_callbacks[i].list); 1443 1444 for (i = 0; i < ARRAY_SIZE(o2hb_live_slots); i++) 1445 INIT_LIST_HEAD(&o2hb_live_slots[i]); 1446 1447 INIT_LIST_HEAD(&o2hb_node_events); 1448 1449 memset(o2hb_live_node_bitmap, 0, sizeof(o2hb_live_node_bitmap)); 1450 memset(o2hb_region_bitmap, 0, sizeof(o2hb_region_bitmap)); 1451 memset(o2hb_live_region_bitmap, 0, sizeof(o2hb_live_region_bitmap)); 1452 memset(o2hb_quorum_region_bitmap, 0, sizeof(o2hb_quorum_region_bitmap)); 1453 memset(o2hb_failed_region_bitmap, 0, sizeof(o2hb_failed_region_bitmap)); 1454 1455 o2hb_dependent_users = 0; 1456 1457 o2hb_debug_init(); 1458 } 1459 1460 /* if we're already in a callback then we're already serialized by the sem */ 1461 static void o2hb_fill_node_map_from_callback(unsigned long *map, 1462 unsigned bytes) 1463 { 1464 BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long))); 1465 1466 memcpy(map, &o2hb_live_node_bitmap, bytes); 1467 } 1468 1469 /* 1470 * get a map of all nodes that are heartbeating in any regions 1471 */ 1472 void o2hb_fill_node_map(unsigned long *map, unsigned bytes) 1473 { 1474 /* callers want to serialize this map and callbacks so that they 1475 * can trust that they don't miss nodes coming to the party */ 1476 down_read(&o2hb_callback_sem); 1477 spin_lock(&o2hb_live_lock); 1478 o2hb_fill_node_map_from_callback(map, bytes); 1479 spin_unlock(&o2hb_live_lock); 1480 up_read(&o2hb_callback_sem); 1481 } 1482 EXPORT_SYMBOL_GPL(o2hb_fill_node_map); 1483 1484 /* 1485 * heartbeat configfs bits. The heartbeat set is a default set under 1486 * the cluster set in nodemanager.c. 1487 */ 1488 1489 static struct o2hb_region *to_o2hb_region(struct config_item *item) 1490 { 1491 return item ? container_of(item, struct o2hb_region, hr_item) : NULL; 1492 } 1493 1494 /* drop_item only drops its ref after killing the thread, nothing should 1495 * be using the region anymore. this has to clean up any state that 1496 * attributes might have built up. */ 1497 static void o2hb_region_release(struct config_item *item) 1498 { 1499 int i; 1500 struct page *page; 1501 struct o2hb_region *reg = to_o2hb_region(item); 1502 1503 mlog(ML_HEARTBEAT, "hb region release (%s)\n", reg->hr_dev_name); 1504 1505 kfree(reg->hr_tmp_block); 1506 1507 if (reg->hr_slot_data) { 1508 for (i = 0; i < reg->hr_num_pages; i++) { 1509 page = reg->hr_slot_data[i]; 1510 if (page) 1511 __free_page(page); 1512 } 1513 kfree(reg->hr_slot_data); 1514 } 1515 1516 if (reg->hr_bdev) 1517 blkdev_put(reg->hr_bdev, FMODE_READ|FMODE_WRITE); 1518 1519 kfree(reg->hr_slots); 1520 1521 debugfs_remove_recursive(reg->hr_debug_dir); 1522 kfree(reg->hr_db_livenodes); 1523 kfree(reg->hr_db_regnum); 1524 kfree(reg->hr_db_elapsed_time); 1525 kfree(reg->hr_db_pinned); 1526 1527 spin_lock(&o2hb_live_lock); 1528 list_del(®->hr_all_item); 1529 spin_unlock(&o2hb_live_lock); 1530 1531 o2net_unregister_handler_list(®->hr_handler_list); 1532 kfree(reg); 1533 } 1534 1535 static int o2hb_read_block_input(struct o2hb_region *reg, 1536 const char *page, 1537 unsigned long *ret_bytes, 1538 unsigned int *ret_bits) 1539 { 1540 unsigned long bytes; 1541 char *p = (char *)page; 1542 1543 bytes = simple_strtoul(p, &p, 0); 1544 if (!p || (*p && (*p != '\n'))) 1545 return -EINVAL; 1546 1547 /* Heartbeat and fs min / max block sizes are the same. */ 1548 if (bytes > 4096 || bytes < 512) 1549 return -ERANGE; 1550 if (hweight16(bytes) != 1) 1551 return -EINVAL; 1552 1553 if (ret_bytes) 1554 *ret_bytes = bytes; 1555 if (ret_bits) 1556 *ret_bits = ffs(bytes) - 1; 1557 1558 return 0; 1559 } 1560 1561 static ssize_t o2hb_region_block_bytes_show(struct config_item *item, 1562 char *page) 1563 { 1564 return sprintf(page, "%u\n", to_o2hb_region(item)->hr_block_bytes); 1565 } 1566 1567 static ssize_t o2hb_region_block_bytes_store(struct config_item *item, 1568 const char *page, 1569 size_t count) 1570 { 1571 struct o2hb_region *reg = to_o2hb_region(item); 1572 int status; 1573 unsigned long block_bytes; 1574 unsigned int block_bits; 1575 1576 if (reg->hr_bdev) 1577 return -EINVAL; 1578 1579 status = o2hb_read_block_input(reg, page, &block_bytes, 1580 &block_bits); 1581 if (status) 1582 return status; 1583 1584 reg->hr_block_bytes = (unsigned int)block_bytes; 1585 reg->hr_block_bits = block_bits; 1586 1587 return count; 1588 } 1589 1590 static ssize_t o2hb_region_start_block_show(struct config_item *item, 1591 char *page) 1592 { 1593 return sprintf(page, "%llu\n", to_o2hb_region(item)->hr_start_block); 1594 } 1595 1596 static ssize_t o2hb_region_start_block_store(struct config_item *item, 1597 const char *page, 1598 size_t count) 1599 { 1600 struct o2hb_region *reg = to_o2hb_region(item); 1601 unsigned long long tmp; 1602 char *p = (char *)page; 1603 1604 if (reg->hr_bdev) 1605 return -EINVAL; 1606 1607 tmp = simple_strtoull(p, &p, 0); 1608 if (!p || (*p && (*p != '\n'))) 1609 return -EINVAL; 1610 1611 reg->hr_start_block = tmp; 1612 1613 return count; 1614 } 1615 1616 static ssize_t o2hb_region_blocks_show(struct config_item *item, char *page) 1617 { 1618 return sprintf(page, "%d\n", to_o2hb_region(item)->hr_blocks); 1619 } 1620 1621 static ssize_t o2hb_region_blocks_store(struct config_item *item, 1622 const char *page, 1623 size_t count) 1624 { 1625 struct o2hb_region *reg = to_o2hb_region(item); 1626 unsigned long tmp; 1627 char *p = (char *)page; 1628 1629 if (reg->hr_bdev) 1630 return -EINVAL; 1631 1632 tmp = simple_strtoul(p, &p, 0); 1633 if (!p || (*p && (*p != '\n'))) 1634 return -EINVAL; 1635 1636 if (tmp > O2NM_MAX_NODES || tmp == 0) 1637 return -ERANGE; 1638 1639 reg->hr_blocks = (unsigned int)tmp; 1640 1641 return count; 1642 } 1643 1644 static ssize_t o2hb_region_dev_show(struct config_item *item, char *page) 1645 { 1646 unsigned int ret = 0; 1647 1648 if (to_o2hb_region(item)->hr_bdev) 1649 ret = sprintf(page, "%s\n", to_o2hb_region(item)->hr_dev_name); 1650 1651 return ret; 1652 } 1653 1654 static void o2hb_init_region_params(struct o2hb_region *reg) 1655 { 1656 reg->hr_slots_per_page = PAGE_SIZE >> reg->hr_block_bits; 1657 reg->hr_timeout_ms = O2HB_REGION_TIMEOUT_MS; 1658 1659 mlog(ML_HEARTBEAT, "hr_start_block = %llu, hr_blocks = %u\n", 1660 reg->hr_start_block, reg->hr_blocks); 1661 mlog(ML_HEARTBEAT, "hr_block_bytes = %u, hr_block_bits = %u\n", 1662 reg->hr_block_bytes, reg->hr_block_bits); 1663 mlog(ML_HEARTBEAT, "hr_timeout_ms = %u\n", reg->hr_timeout_ms); 1664 mlog(ML_HEARTBEAT, "dead threshold = %u\n", o2hb_dead_threshold); 1665 } 1666 1667 static int o2hb_map_slot_data(struct o2hb_region *reg) 1668 { 1669 int i, j; 1670 unsigned int last_slot; 1671 unsigned int spp = reg->hr_slots_per_page; 1672 struct page *page; 1673 char *raw; 1674 struct o2hb_disk_slot *slot; 1675 1676 reg->hr_tmp_block = kmalloc(reg->hr_block_bytes, GFP_KERNEL); 1677 if (reg->hr_tmp_block == NULL) 1678 return -ENOMEM; 1679 1680 reg->hr_slots = kcalloc(reg->hr_blocks, 1681 sizeof(struct o2hb_disk_slot), GFP_KERNEL); 1682 if (reg->hr_slots == NULL) 1683 return -ENOMEM; 1684 1685 for(i = 0; i < reg->hr_blocks; i++) { 1686 slot = ®->hr_slots[i]; 1687 slot->ds_node_num = i; 1688 INIT_LIST_HEAD(&slot->ds_live_item); 1689 slot->ds_raw_block = NULL; 1690 } 1691 1692 reg->hr_num_pages = (reg->hr_blocks + spp - 1) / spp; 1693 mlog(ML_HEARTBEAT, "Going to require %u pages to cover %u blocks " 1694 "at %u blocks per page\n", 1695 reg->hr_num_pages, reg->hr_blocks, spp); 1696 1697 reg->hr_slot_data = kcalloc(reg->hr_num_pages, sizeof(struct page *), 1698 GFP_KERNEL); 1699 if (!reg->hr_slot_data) 1700 return -ENOMEM; 1701 1702 for(i = 0; i < reg->hr_num_pages; i++) { 1703 page = alloc_page(GFP_KERNEL); 1704 if (!page) 1705 return -ENOMEM; 1706 1707 reg->hr_slot_data[i] = page; 1708 1709 last_slot = i * spp; 1710 raw = page_address(page); 1711 for (j = 0; 1712 (j < spp) && ((j + last_slot) < reg->hr_blocks); 1713 j++) { 1714 BUG_ON((j + last_slot) >= reg->hr_blocks); 1715 1716 slot = ®->hr_slots[j + last_slot]; 1717 slot->ds_raw_block = 1718 (struct o2hb_disk_heartbeat_block *) raw; 1719 1720 raw += reg->hr_block_bytes; 1721 } 1722 } 1723 1724 return 0; 1725 } 1726 1727 /* Read in all the slots available and populate the tracking 1728 * structures so that we can start with a baseline idea of what's 1729 * there. */ 1730 static int o2hb_populate_slot_data(struct o2hb_region *reg) 1731 { 1732 int ret, i; 1733 struct o2hb_disk_slot *slot; 1734 struct o2hb_disk_heartbeat_block *hb_block; 1735 1736 ret = o2hb_read_slots(reg, 0, reg->hr_blocks); 1737 if (ret) 1738 goto out; 1739 1740 /* We only want to get an idea of the values initially in each 1741 * slot, so we do no verification - o2hb_check_slot will 1742 * actually determine if each configured slot is valid and 1743 * whether any values have changed. */ 1744 for(i = 0; i < reg->hr_blocks; i++) { 1745 slot = ®->hr_slots[i]; 1746 hb_block = (struct o2hb_disk_heartbeat_block *) slot->ds_raw_block; 1747 1748 /* Only fill the values that o2hb_check_slot uses to 1749 * determine changing slots */ 1750 slot->ds_last_time = le64_to_cpu(hb_block->hb_seq); 1751 slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation); 1752 } 1753 1754 out: 1755 return ret; 1756 } 1757 1758 /* this is acting as commit; we set up all of hr_bdev and hr_task or nothing */ 1759 static ssize_t o2hb_region_dev_store(struct config_item *item, 1760 const char *page, 1761 size_t count) 1762 { 1763 struct o2hb_region *reg = to_o2hb_region(item); 1764 struct task_struct *hb_task; 1765 long fd; 1766 int sectsize; 1767 char *p = (char *)page; 1768 struct fd f; 1769 ssize_t ret = -EINVAL; 1770 int live_threshold; 1771 1772 if (reg->hr_bdev) 1773 goto out; 1774 1775 /* We can't heartbeat without having had our node number 1776 * configured yet. */ 1777 if (o2nm_this_node() == O2NM_MAX_NODES) 1778 goto out; 1779 1780 fd = simple_strtol(p, &p, 0); 1781 if (!p || (*p && (*p != '\n'))) 1782 goto out; 1783 1784 if (fd < 0 || fd >= INT_MAX) 1785 goto out; 1786 1787 f = fdget(fd); 1788 if (f.file == NULL) 1789 goto out; 1790 1791 if (reg->hr_blocks == 0 || reg->hr_start_block == 0 || 1792 reg->hr_block_bytes == 0) 1793 goto out2; 1794 1795 if (!S_ISBLK(f.file->f_mapping->host->i_mode)) 1796 goto out2; 1797 1798 reg->hr_bdev = blkdev_get_by_dev(f.file->f_mapping->host->i_rdev, 1799 FMODE_WRITE | FMODE_READ, NULL); 1800 if (IS_ERR(reg->hr_bdev)) { 1801 ret = PTR_ERR(reg->hr_bdev); 1802 reg->hr_bdev = NULL; 1803 goto out2; 1804 } 1805 1806 bdevname(reg->hr_bdev, reg->hr_dev_name); 1807 1808 sectsize = bdev_logical_block_size(reg->hr_bdev); 1809 if (sectsize != reg->hr_block_bytes) { 1810 mlog(ML_ERROR, 1811 "blocksize %u incorrect for device, expected %d", 1812 reg->hr_block_bytes, sectsize); 1813 ret = -EINVAL; 1814 goto out3; 1815 } 1816 1817 o2hb_init_region_params(reg); 1818 1819 /* Generation of zero is invalid */ 1820 do { 1821 get_random_bytes(®->hr_generation, 1822 sizeof(reg->hr_generation)); 1823 } while (reg->hr_generation == 0); 1824 1825 ret = o2hb_map_slot_data(reg); 1826 if (ret) { 1827 mlog_errno(ret); 1828 goto out3; 1829 } 1830 1831 ret = o2hb_populate_slot_data(reg); 1832 if (ret) { 1833 mlog_errno(ret); 1834 goto out3; 1835 } 1836 1837 INIT_DELAYED_WORK(®->hr_write_timeout_work, o2hb_write_timeout); 1838 INIT_DELAYED_WORK(®->hr_nego_timeout_work, o2hb_nego_timeout); 1839 1840 /* 1841 * A node is considered live after it has beat LIVE_THRESHOLD 1842 * times. We're not steady until we've given them a chance 1843 * _after_ our first read. 1844 * The default threshold is bare minimum so as to limit the delay 1845 * during mounts. For global heartbeat, the threshold doubled for the 1846 * first region. 1847 */ 1848 live_threshold = O2HB_LIVE_THRESHOLD; 1849 if (o2hb_global_heartbeat_active()) { 1850 spin_lock(&o2hb_live_lock); 1851 if (bitmap_weight(o2hb_region_bitmap, O2NM_MAX_REGIONS) == 1) 1852 live_threshold <<= 1; 1853 spin_unlock(&o2hb_live_lock); 1854 } 1855 ++live_threshold; 1856 atomic_set(®->hr_steady_iterations, live_threshold); 1857 /* unsteady_iterations is triple the steady_iterations */ 1858 atomic_set(®->hr_unsteady_iterations, (live_threshold * 3)); 1859 1860 hb_task = kthread_run(o2hb_thread, reg, "o2hb-%s", 1861 reg->hr_item.ci_name); 1862 if (IS_ERR(hb_task)) { 1863 ret = PTR_ERR(hb_task); 1864 mlog_errno(ret); 1865 goto out3; 1866 } 1867 1868 spin_lock(&o2hb_live_lock); 1869 reg->hr_task = hb_task; 1870 spin_unlock(&o2hb_live_lock); 1871 1872 ret = wait_event_interruptible(o2hb_steady_queue, 1873 atomic_read(®->hr_steady_iterations) == 0 || 1874 reg->hr_node_deleted); 1875 if (ret) { 1876 atomic_set(®->hr_steady_iterations, 0); 1877 reg->hr_aborted_start = 1; 1878 } 1879 1880 if (reg->hr_aborted_start) { 1881 ret = -EIO; 1882 goto out3; 1883 } 1884 1885 if (reg->hr_node_deleted) { 1886 ret = -EINVAL; 1887 goto out3; 1888 } 1889 1890 /* Ok, we were woken. Make sure it wasn't by drop_item() */ 1891 spin_lock(&o2hb_live_lock); 1892 hb_task = reg->hr_task; 1893 if (o2hb_global_heartbeat_active()) 1894 set_bit(reg->hr_region_num, o2hb_live_region_bitmap); 1895 spin_unlock(&o2hb_live_lock); 1896 1897 if (hb_task) 1898 ret = count; 1899 else 1900 ret = -EIO; 1901 1902 if (hb_task && o2hb_global_heartbeat_active()) 1903 printk(KERN_NOTICE "o2hb: Heartbeat started on region %s (%s)\n", 1904 config_item_name(®->hr_item), reg->hr_dev_name); 1905 1906 out3: 1907 if (ret < 0) { 1908 blkdev_put(reg->hr_bdev, FMODE_READ | FMODE_WRITE); 1909 reg->hr_bdev = NULL; 1910 } 1911 out2: 1912 fdput(f); 1913 out: 1914 return ret; 1915 } 1916 1917 static ssize_t o2hb_region_pid_show(struct config_item *item, char *page) 1918 { 1919 struct o2hb_region *reg = to_o2hb_region(item); 1920 pid_t pid = 0; 1921 1922 spin_lock(&o2hb_live_lock); 1923 if (reg->hr_task) 1924 pid = task_pid_nr(reg->hr_task); 1925 spin_unlock(&o2hb_live_lock); 1926 1927 if (!pid) 1928 return 0; 1929 1930 return sprintf(page, "%u\n", pid); 1931 } 1932 1933 CONFIGFS_ATTR(o2hb_region_, block_bytes); 1934 CONFIGFS_ATTR(o2hb_region_, start_block); 1935 CONFIGFS_ATTR(o2hb_region_, blocks); 1936 CONFIGFS_ATTR(o2hb_region_, dev); 1937 CONFIGFS_ATTR_RO(o2hb_region_, pid); 1938 1939 static struct configfs_attribute *o2hb_region_attrs[] = { 1940 &o2hb_region_attr_block_bytes, 1941 &o2hb_region_attr_start_block, 1942 &o2hb_region_attr_blocks, 1943 &o2hb_region_attr_dev, 1944 &o2hb_region_attr_pid, 1945 NULL, 1946 }; 1947 1948 static struct configfs_item_operations o2hb_region_item_ops = { 1949 .release = o2hb_region_release, 1950 }; 1951 1952 static const struct config_item_type o2hb_region_type = { 1953 .ct_item_ops = &o2hb_region_item_ops, 1954 .ct_attrs = o2hb_region_attrs, 1955 .ct_owner = THIS_MODULE, 1956 }; 1957 1958 /* heartbeat set */ 1959 1960 struct o2hb_heartbeat_group { 1961 struct config_group hs_group; 1962 /* some stuff? */ 1963 }; 1964 1965 static struct o2hb_heartbeat_group *to_o2hb_heartbeat_group(struct config_group *group) 1966 { 1967 return group ? 1968 container_of(group, struct o2hb_heartbeat_group, hs_group) 1969 : NULL; 1970 } 1971 1972 static void o2hb_debug_region_init(struct o2hb_region *reg, 1973 struct dentry *parent) 1974 { 1975 struct dentry *dir; 1976 1977 dir = debugfs_create_dir(config_item_name(®->hr_item), parent); 1978 reg->hr_debug_dir = dir; 1979 1980 o2hb_debug_create(O2HB_DEBUG_LIVENODES, dir, &(reg->hr_db_livenodes), 1981 sizeof(*(reg->hr_db_livenodes)), 1982 O2HB_DB_TYPE_REGION_LIVENODES, 1983 sizeof(reg->hr_live_node_bitmap), O2NM_MAX_NODES, 1984 reg); 1985 1986 o2hb_debug_create(O2HB_DEBUG_REGION_NUMBER, dir, &(reg->hr_db_regnum), 1987 sizeof(*(reg->hr_db_regnum)), 1988 O2HB_DB_TYPE_REGION_NUMBER, 0, O2NM_MAX_NODES, reg); 1989 1990 o2hb_debug_create(O2HB_DEBUG_REGION_ELAPSED_TIME, dir, 1991 &(reg->hr_db_elapsed_time), 1992 sizeof(*(reg->hr_db_elapsed_time)), 1993 O2HB_DB_TYPE_REGION_ELAPSED_TIME, 0, 0, reg); 1994 1995 o2hb_debug_create(O2HB_DEBUG_REGION_PINNED, dir, &(reg->hr_db_pinned), 1996 sizeof(*(reg->hr_db_pinned)), 1997 O2HB_DB_TYPE_REGION_PINNED, 0, 0, reg); 1998 1999 } 2000 2001 static struct config_item *o2hb_heartbeat_group_make_item(struct config_group *group, 2002 const char *name) 2003 { 2004 struct o2hb_region *reg = NULL; 2005 int ret; 2006 2007 reg = kzalloc(sizeof(struct o2hb_region), GFP_KERNEL); 2008 if (reg == NULL) 2009 return ERR_PTR(-ENOMEM); 2010 2011 if (strlen(name) > O2HB_MAX_REGION_NAME_LEN) { 2012 ret = -ENAMETOOLONG; 2013 goto free; 2014 } 2015 2016 spin_lock(&o2hb_live_lock); 2017 reg->hr_region_num = 0; 2018 if (o2hb_global_heartbeat_active()) { 2019 reg->hr_region_num = find_first_zero_bit(o2hb_region_bitmap, 2020 O2NM_MAX_REGIONS); 2021 if (reg->hr_region_num >= O2NM_MAX_REGIONS) { 2022 spin_unlock(&o2hb_live_lock); 2023 ret = -EFBIG; 2024 goto free; 2025 } 2026 set_bit(reg->hr_region_num, o2hb_region_bitmap); 2027 } 2028 list_add_tail(®->hr_all_item, &o2hb_all_regions); 2029 spin_unlock(&o2hb_live_lock); 2030 2031 config_item_init_type_name(®->hr_item, name, &o2hb_region_type); 2032 2033 /* this is the same way to generate msg key as dlm, for local heartbeat, 2034 * name is also the same, so make initial crc value different to avoid 2035 * message key conflict. 2036 */ 2037 reg->hr_key = crc32_le(reg->hr_region_num + O2NM_MAX_REGIONS, 2038 name, strlen(name)); 2039 INIT_LIST_HEAD(®->hr_handler_list); 2040 ret = o2net_register_handler(O2HB_NEGO_TIMEOUT_MSG, reg->hr_key, 2041 sizeof(struct o2hb_nego_msg), 2042 o2hb_nego_timeout_handler, 2043 reg, NULL, ®->hr_handler_list); 2044 if (ret) 2045 goto free; 2046 2047 ret = o2net_register_handler(O2HB_NEGO_APPROVE_MSG, reg->hr_key, 2048 sizeof(struct o2hb_nego_msg), 2049 o2hb_nego_approve_handler, 2050 reg, NULL, ®->hr_handler_list); 2051 if (ret) 2052 goto unregister_handler; 2053 2054 o2hb_debug_region_init(reg, o2hb_debug_dir); 2055 2056 return ®->hr_item; 2057 2058 unregister_handler: 2059 o2net_unregister_handler_list(®->hr_handler_list); 2060 free: 2061 kfree(reg); 2062 return ERR_PTR(ret); 2063 } 2064 2065 static void o2hb_heartbeat_group_drop_item(struct config_group *group, 2066 struct config_item *item) 2067 { 2068 struct task_struct *hb_task; 2069 struct o2hb_region *reg = to_o2hb_region(item); 2070 int quorum_region = 0; 2071 2072 /* stop the thread when the user removes the region dir */ 2073 spin_lock(&o2hb_live_lock); 2074 hb_task = reg->hr_task; 2075 reg->hr_task = NULL; 2076 reg->hr_item_dropped = 1; 2077 spin_unlock(&o2hb_live_lock); 2078 2079 if (hb_task) 2080 kthread_stop(hb_task); 2081 2082 if (o2hb_global_heartbeat_active()) { 2083 spin_lock(&o2hb_live_lock); 2084 clear_bit(reg->hr_region_num, o2hb_region_bitmap); 2085 clear_bit(reg->hr_region_num, o2hb_live_region_bitmap); 2086 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap)) 2087 quorum_region = 1; 2088 clear_bit(reg->hr_region_num, o2hb_quorum_region_bitmap); 2089 spin_unlock(&o2hb_live_lock); 2090 printk(KERN_NOTICE "o2hb: Heartbeat %s on region %s (%s)\n", 2091 ((atomic_read(®->hr_steady_iterations) == 0) ? 2092 "stopped" : "start aborted"), config_item_name(item), 2093 reg->hr_dev_name); 2094 } 2095 2096 /* 2097 * If we're racing a dev_write(), we need to wake them. They will 2098 * check reg->hr_task 2099 */ 2100 if (atomic_read(®->hr_steady_iterations) != 0) { 2101 reg->hr_aborted_start = 1; 2102 atomic_set(®->hr_steady_iterations, 0); 2103 wake_up(&o2hb_steady_queue); 2104 } 2105 2106 config_item_put(item); 2107 2108 if (!o2hb_global_heartbeat_active() || !quorum_region) 2109 return; 2110 2111 /* 2112 * If global heartbeat active and there are dependent users, 2113 * pin all regions if quorum region count <= CUT_OFF 2114 */ 2115 spin_lock(&o2hb_live_lock); 2116 2117 if (!o2hb_dependent_users) 2118 goto unlock; 2119 2120 if (bitmap_weight(o2hb_quorum_region_bitmap, 2121 O2NM_MAX_REGIONS) <= O2HB_PIN_CUT_OFF) 2122 o2hb_region_pin(NULL); 2123 2124 unlock: 2125 spin_unlock(&o2hb_live_lock); 2126 } 2127 2128 static ssize_t o2hb_heartbeat_group_dead_threshold_show(struct config_item *item, 2129 char *page) 2130 { 2131 return sprintf(page, "%u\n", o2hb_dead_threshold); 2132 } 2133 2134 static ssize_t o2hb_heartbeat_group_dead_threshold_store(struct config_item *item, 2135 const char *page, size_t count) 2136 { 2137 unsigned long tmp; 2138 char *p = (char *)page; 2139 2140 tmp = simple_strtoul(p, &p, 10); 2141 if (!p || (*p && (*p != '\n'))) 2142 return -EINVAL; 2143 2144 /* this will validate ranges for us. */ 2145 o2hb_dead_threshold_set((unsigned int) tmp); 2146 2147 return count; 2148 } 2149 2150 static ssize_t o2hb_heartbeat_group_mode_show(struct config_item *item, 2151 char *page) 2152 { 2153 return sprintf(page, "%s\n", 2154 o2hb_heartbeat_mode_desc[o2hb_heartbeat_mode]); 2155 } 2156 2157 static ssize_t o2hb_heartbeat_group_mode_store(struct config_item *item, 2158 const char *page, size_t count) 2159 { 2160 unsigned int i; 2161 int ret; 2162 size_t len; 2163 2164 len = (page[count - 1] == '\n') ? count - 1 : count; 2165 if (!len) 2166 return -EINVAL; 2167 2168 for (i = 0; i < O2HB_HEARTBEAT_NUM_MODES; ++i) { 2169 if (strncasecmp(page, o2hb_heartbeat_mode_desc[i], len)) 2170 continue; 2171 2172 ret = o2hb_global_heartbeat_mode_set(i); 2173 if (!ret) 2174 printk(KERN_NOTICE "o2hb: Heartbeat mode set to %s\n", 2175 o2hb_heartbeat_mode_desc[i]); 2176 return count; 2177 } 2178 2179 return -EINVAL; 2180 2181 } 2182 2183 CONFIGFS_ATTR(o2hb_heartbeat_group_, dead_threshold); 2184 CONFIGFS_ATTR(o2hb_heartbeat_group_, mode); 2185 2186 static struct configfs_attribute *o2hb_heartbeat_group_attrs[] = { 2187 &o2hb_heartbeat_group_attr_dead_threshold, 2188 &o2hb_heartbeat_group_attr_mode, 2189 NULL, 2190 }; 2191 2192 static struct configfs_group_operations o2hb_heartbeat_group_group_ops = { 2193 .make_item = o2hb_heartbeat_group_make_item, 2194 .drop_item = o2hb_heartbeat_group_drop_item, 2195 }; 2196 2197 static const struct config_item_type o2hb_heartbeat_group_type = { 2198 .ct_group_ops = &o2hb_heartbeat_group_group_ops, 2199 .ct_attrs = o2hb_heartbeat_group_attrs, 2200 .ct_owner = THIS_MODULE, 2201 }; 2202 2203 /* this is just here to avoid touching group in heartbeat.h which the 2204 * entire damn world #includes */ 2205 struct config_group *o2hb_alloc_hb_set(void) 2206 { 2207 struct o2hb_heartbeat_group *hs = NULL; 2208 struct config_group *ret = NULL; 2209 2210 hs = kzalloc(sizeof(struct o2hb_heartbeat_group), GFP_KERNEL); 2211 if (hs == NULL) 2212 goto out; 2213 2214 config_group_init_type_name(&hs->hs_group, "heartbeat", 2215 &o2hb_heartbeat_group_type); 2216 2217 ret = &hs->hs_group; 2218 out: 2219 if (ret == NULL) 2220 kfree(hs); 2221 return ret; 2222 } 2223 2224 void o2hb_free_hb_set(struct config_group *group) 2225 { 2226 struct o2hb_heartbeat_group *hs = to_o2hb_heartbeat_group(group); 2227 kfree(hs); 2228 } 2229 2230 /* hb callback registration and issuing */ 2231 2232 static struct o2hb_callback *hbcall_from_type(enum o2hb_callback_type type) 2233 { 2234 if (type == O2HB_NUM_CB) 2235 return ERR_PTR(-EINVAL); 2236 2237 return &o2hb_callbacks[type]; 2238 } 2239 2240 void o2hb_setup_callback(struct o2hb_callback_func *hc, 2241 enum o2hb_callback_type type, 2242 o2hb_cb_func *func, 2243 void *data, 2244 int priority) 2245 { 2246 INIT_LIST_HEAD(&hc->hc_item); 2247 hc->hc_func = func; 2248 hc->hc_data = data; 2249 hc->hc_priority = priority; 2250 hc->hc_type = type; 2251 hc->hc_magic = O2HB_CB_MAGIC; 2252 } 2253 EXPORT_SYMBOL_GPL(o2hb_setup_callback); 2254 2255 /* 2256 * In local heartbeat mode, region_uuid passed matches the dlm domain name. 2257 * In global heartbeat mode, region_uuid passed is NULL. 2258 * 2259 * In local, we only pin the matching region. In global we pin all the active 2260 * regions. 2261 */ 2262 static int o2hb_region_pin(const char *region_uuid) 2263 { 2264 int ret = 0, found = 0; 2265 struct o2hb_region *reg; 2266 char *uuid; 2267 2268 assert_spin_locked(&o2hb_live_lock); 2269 2270 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) { 2271 if (reg->hr_item_dropped) 2272 continue; 2273 2274 uuid = config_item_name(®->hr_item); 2275 2276 /* local heartbeat */ 2277 if (region_uuid) { 2278 if (strcmp(region_uuid, uuid)) 2279 continue; 2280 found = 1; 2281 } 2282 2283 if (reg->hr_item_pinned || reg->hr_item_dropped) 2284 goto skip_pin; 2285 2286 /* Ignore ENOENT only for local hb (userdlm domain) */ 2287 ret = o2nm_depend_item(®->hr_item); 2288 if (!ret) { 2289 mlog(ML_CLUSTER, "Pin region %s\n", uuid); 2290 reg->hr_item_pinned = 1; 2291 } else { 2292 if (ret == -ENOENT && found) 2293 ret = 0; 2294 else { 2295 mlog(ML_ERROR, "Pin region %s fails with %d\n", 2296 uuid, ret); 2297 break; 2298 } 2299 } 2300 skip_pin: 2301 if (found) 2302 break; 2303 } 2304 2305 return ret; 2306 } 2307 2308 /* 2309 * In local heartbeat mode, region_uuid passed matches the dlm domain name. 2310 * In global heartbeat mode, region_uuid passed is NULL. 2311 * 2312 * In local, we only unpin the matching region. In global we unpin all the 2313 * active regions. 2314 */ 2315 static void o2hb_region_unpin(const char *region_uuid) 2316 { 2317 struct o2hb_region *reg; 2318 char *uuid; 2319 int found = 0; 2320 2321 assert_spin_locked(&o2hb_live_lock); 2322 2323 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) { 2324 if (reg->hr_item_dropped) 2325 continue; 2326 2327 uuid = config_item_name(®->hr_item); 2328 if (region_uuid) { 2329 if (strcmp(region_uuid, uuid)) 2330 continue; 2331 found = 1; 2332 } 2333 2334 if (reg->hr_item_pinned) { 2335 mlog(ML_CLUSTER, "Unpin region %s\n", uuid); 2336 o2nm_undepend_item(®->hr_item); 2337 reg->hr_item_pinned = 0; 2338 } 2339 if (found) 2340 break; 2341 } 2342 } 2343 2344 static int o2hb_region_inc_user(const char *region_uuid) 2345 { 2346 int ret = 0; 2347 2348 spin_lock(&o2hb_live_lock); 2349 2350 /* local heartbeat */ 2351 if (!o2hb_global_heartbeat_active()) { 2352 ret = o2hb_region_pin(region_uuid); 2353 goto unlock; 2354 } 2355 2356 /* 2357 * if global heartbeat active and this is the first dependent user, 2358 * pin all regions if quorum region count <= CUT_OFF 2359 */ 2360 o2hb_dependent_users++; 2361 if (o2hb_dependent_users > 1) 2362 goto unlock; 2363 2364 if (bitmap_weight(o2hb_quorum_region_bitmap, 2365 O2NM_MAX_REGIONS) <= O2HB_PIN_CUT_OFF) 2366 ret = o2hb_region_pin(NULL); 2367 2368 unlock: 2369 spin_unlock(&o2hb_live_lock); 2370 return ret; 2371 } 2372 2373 static void o2hb_region_dec_user(const char *region_uuid) 2374 { 2375 spin_lock(&o2hb_live_lock); 2376 2377 /* local heartbeat */ 2378 if (!o2hb_global_heartbeat_active()) { 2379 o2hb_region_unpin(region_uuid); 2380 goto unlock; 2381 } 2382 2383 /* 2384 * if global heartbeat active and there are no dependent users, 2385 * unpin all quorum regions 2386 */ 2387 o2hb_dependent_users--; 2388 if (!o2hb_dependent_users) 2389 o2hb_region_unpin(NULL); 2390 2391 unlock: 2392 spin_unlock(&o2hb_live_lock); 2393 } 2394 2395 int o2hb_register_callback(const char *region_uuid, 2396 struct o2hb_callback_func *hc) 2397 { 2398 struct o2hb_callback_func *f; 2399 struct o2hb_callback *hbcall; 2400 int ret; 2401 2402 BUG_ON(hc->hc_magic != O2HB_CB_MAGIC); 2403 BUG_ON(!list_empty(&hc->hc_item)); 2404 2405 hbcall = hbcall_from_type(hc->hc_type); 2406 if (IS_ERR(hbcall)) { 2407 ret = PTR_ERR(hbcall); 2408 goto out; 2409 } 2410 2411 if (region_uuid) { 2412 ret = o2hb_region_inc_user(region_uuid); 2413 if (ret) { 2414 mlog_errno(ret); 2415 goto out; 2416 } 2417 } 2418 2419 down_write(&o2hb_callback_sem); 2420 2421 list_for_each_entry(f, &hbcall->list, hc_item) { 2422 if (hc->hc_priority < f->hc_priority) { 2423 list_add_tail(&hc->hc_item, &f->hc_item); 2424 break; 2425 } 2426 } 2427 if (list_empty(&hc->hc_item)) 2428 list_add_tail(&hc->hc_item, &hbcall->list); 2429 2430 up_write(&o2hb_callback_sem); 2431 ret = 0; 2432 out: 2433 mlog(ML_CLUSTER, "returning %d on behalf of %p for funcs %p\n", 2434 ret, __builtin_return_address(0), hc); 2435 return ret; 2436 } 2437 EXPORT_SYMBOL_GPL(o2hb_register_callback); 2438 2439 void o2hb_unregister_callback(const char *region_uuid, 2440 struct o2hb_callback_func *hc) 2441 { 2442 BUG_ON(hc->hc_magic != O2HB_CB_MAGIC); 2443 2444 mlog(ML_CLUSTER, "on behalf of %p for funcs %p\n", 2445 __builtin_return_address(0), hc); 2446 2447 /* XXX Can this happen _with_ a region reference? */ 2448 if (list_empty(&hc->hc_item)) 2449 return; 2450 2451 if (region_uuid) 2452 o2hb_region_dec_user(region_uuid); 2453 2454 down_write(&o2hb_callback_sem); 2455 2456 list_del_init(&hc->hc_item); 2457 2458 up_write(&o2hb_callback_sem); 2459 } 2460 EXPORT_SYMBOL_GPL(o2hb_unregister_callback); 2461 2462 int o2hb_check_node_heartbeating_no_sem(u8 node_num) 2463 { 2464 unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; 2465 2466 spin_lock(&o2hb_live_lock); 2467 o2hb_fill_node_map_from_callback(testing_map, sizeof(testing_map)); 2468 spin_unlock(&o2hb_live_lock); 2469 if (!test_bit(node_num, testing_map)) { 2470 mlog(ML_HEARTBEAT, 2471 "node (%u) does not have heartbeating enabled.\n", 2472 node_num); 2473 return 0; 2474 } 2475 2476 return 1; 2477 } 2478 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating_no_sem); 2479 2480 int o2hb_check_node_heartbeating_from_callback(u8 node_num) 2481 { 2482 unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; 2483 2484 o2hb_fill_node_map_from_callback(testing_map, sizeof(testing_map)); 2485 if (!test_bit(node_num, testing_map)) { 2486 mlog(ML_HEARTBEAT, 2487 "node (%u) does not have heartbeating enabled.\n", 2488 node_num); 2489 return 0; 2490 } 2491 2492 return 1; 2493 } 2494 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating_from_callback); 2495 2496 /* 2497 * this is just a hack until we get the plumbing which flips file systems 2498 * read only and drops the hb ref instead of killing the node dead. 2499 */ 2500 void o2hb_stop_all_regions(void) 2501 { 2502 struct o2hb_region *reg; 2503 2504 mlog(ML_ERROR, "stopping heartbeat on all active regions.\n"); 2505 2506 spin_lock(&o2hb_live_lock); 2507 2508 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) 2509 reg->hr_unclean_stop = 1; 2510 2511 spin_unlock(&o2hb_live_lock); 2512 } 2513 EXPORT_SYMBOL_GPL(o2hb_stop_all_regions); 2514 2515 int o2hb_get_all_regions(char *region_uuids, u8 max_regions) 2516 { 2517 struct o2hb_region *reg; 2518 int numregs = 0; 2519 char *p; 2520 2521 spin_lock(&o2hb_live_lock); 2522 2523 p = region_uuids; 2524 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) { 2525 if (reg->hr_item_dropped) 2526 continue; 2527 2528 mlog(0, "Region: %s\n", config_item_name(®->hr_item)); 2529 if (numregs < max_regions) { 2530 memcpy(p, config_item_name(®->hr_item), 2531 O2HB_MAX_REGION_NAME_LEN); 2532 p += O2HB_MAX_REGION_NAME_LEN; 2533 } 2534 numregs++; 2535 } 2536 2537 spin_unlock(&o2hb_live_lock); 2538 2539 return numregs; 2540 } 2541 EXPORT_SYMBOL_GPL(o2hb_get_all_regions); 2542 2543 int o2hb_global_heartbeat_active(void) 2544 { 2545 return (o2hb_heartbeat_mode == O2HB_HEARTBEAT_GLOBAL); 2546 } 2547 EXPORT_SYMBOL(o2hb_global_heartbeat_active); 2548