1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Copyright (C) 2004, 2005 Oracle. All rights reserved. 4 */ 5 6 #include <linux/kernel.h> 7 #include <linux/sched.h> 8 #include <linux/jiffies.h> 9 #include <linux/module.h> 10 #include <linux/fs.h> 11 #include <linux/bio.h> 12 #include <linux/blkdev.h> 13 #include <linux/delay.h> 14 #include <linux/file.h> 15 #include <linux/kthread.h> 16 #include <linux/configfs.h> 17 #include <linux/random.h> 18 #include <linux/crc32.h> 19 #include <linux/time.h> 20 #include <linux/debugfs.h> 21 #include <linux/slab.h> 22 #include <linux/bitmap.h> 23 #include <linux/ktime.h> 24 #include "heartbeat.h" 25 #include "tcp.h" 26 #include "nodemanager.h" 27 #include "quorum.h" 28 29 #include "masklog.h" 30 31 32 /* 33 * The first heartbeat pass had one global thread that would serialize all hb 34 * callback calls. This global serializing sem should only be removed once 35 * we've made sure that all callees can deal with being called concurrently 36 * from multiple hb region threads. 37 */ 38 static DECLARE_RWSEM(o2hb_callback_sem); 39 40 /* 41 * multiple hb threads are watching multiple regions. A node is live 42 * whenever any of the threads sees activity from the node in its region. 43 */ 44 static DEFINE_SPINLOCK(o2hb_live_lock); 45 static struct list_head o2hb_live_slots[O2NM_MAX_NODES]; 46 static unsigned long o2hb_live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)]; 47 static LIST_HEAD(o2hb_node_events); 48 static DECLARE_WAIT_QUEUE_HEAD(o2hb_steady_queue); 49 50 /* 51 * In global heartbeat, we maintain a series of region bitmaps. 52 * - o2hb_region_bitmap allows us to limit the region number to max region. 53 * - o2hb_live_region_bitmap tracks live regions (seen steady iterations). 54 * - o2hb_quorum_region_bitmap tracks live regions that have seen all nodes 55 * heartbeat on it. 56 * - o2hb_failed_region_bitmap tracks the regions that have seen io timeouts. 57 */ 58 static unsigned long o2hb_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)]; 59 static unsigned long o2hb_live_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)]; 60 static unsigned long o2hb_quorum_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)]; 61 static unsigned long o2hb_failed_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)]; 62 63 #define O2HB_DB_TYPE_LIVENODES 0 64 #define O2HB_DB_TYPE_LIVEREGIONS 1 65 #define O2HB_DB_TYPE_QUORUMREGIONS 2 66 #define O2HB_DB_TYPE_FAILEDREGIONS 3 67 #define O2HB_DB_TYPE_REGION_LIVENODES 4 68 #define O2HB_DB_TYPE_REGION_NUMBER 5 69 #define O2HB_DB_TYPE_REGION_ELAPSED_TIME 6 70 #define O2HB_DB_TYPE_REGION_PINNED 7 71 struct o2hb_debug_buf { 72 int db_type; 73 int db_size; 74 int db_len; 75 void *db_data; 76 }; 77 78 static struct o2hb_debug_buf *o2hb_db_livenodes; 79 static struct o2hb_debug_buf *o2hb_db_liveregions; 80 static struct o2hb_debug_buf *o2hb_db_quorumregions; 81 static struct o2hb_debug_buf *o2hb_db_failedregions; 82 83 #define O2HB_DEBUG_DIR "o2hb" 84 #define O2HB_DEBUG_LIVENODES "livenodes" 85 #define O2HB_DEBUG_LIVEREGIONS "live_regions" 86 #define O2HB_DEBUG_QUORUMREGIONS "quorum_regions" 87 #define O2HB_DEBUG_FAILEDREGIONS "failed_regions" 88 #define O2HB_DEBUG_REGION_NUMBER "num" 89 #define O2HB_DEBUG_REGION_ELAPSED_TIME "elapsed_time_in_ms" 90 #define O2HB_DEBUG_REGION_PINNED "pinned" 91 92 static struct dentry *o2hb_debug_dir; 93 94 static LIST_HEAD(o2hb_all_regions); 95 96 static struct o2hb_callback { 97 struct list_head list; 98 } o2hb_callbacks[O2HB_NUM_CB]; 99 100 static struct o2hb_callback *hbcall_from_type(enum o2hb_callback_type type); 101 102 enum o2hb_heartbeat_modes { 103 O2HB_HEARTBEAT_LOCAL = 0, 104 O2HB_HEARTBEAT_GLOBAL, 105 O2HB_HEARTBEAT_NUM_MODES, 106 }; 107 108 static const char *o2hb_heartbeat_mode_desc[O2HB_HEARTBEAT_NUM_MODES] = { 109 "local", /* O2HB_HEARTBEAT_LOCAL */ 110 "global", /* O2HB_HEARTBEAT_GLOBAL */ 111 }; 112 113 unsigned int o2hb_dead_threshold = O2HB_DEFAULT_DEAD_THRESHOLD; 114 static unsigned int o2hb_heartbeat_mode = O2HB_HEARTBEAT_LOCAL; 115 116 /* 117 * o2hb_dependent_users tracks the number of registered callbacks that depend 118 * on heartbeat. o2net and o2dlm are two entities that register this callback. 119 * However only o2dlm depends on the heartbeat. It does not want the heartbeat 120 * to stop while a dlm domain is still active. 121 */ 122 static unsigned int o2hb_dependent_users; 123 124 /* 125 * In global heartbeat mode, all regions are pinned if there are one or more 126 * dependent users and the quorum region count is <= O2HB_PIN_CUT_OFF. All 127 * regions are unpinned if the region count exceeds the cut off or the number 128 * of dependent users falls to zero. 129 */ 130 #define O2HB_PIN_CUT_OFF 3 131 132 /* 133 * In local heartbeat mode, we assume the dlm domain name to be the same as 134 * region uuid. This is true for domains created for the file system but not 135 * necessarily true for userdlm domains. This is a known limitation. 136 * 137 * In global heartbeat mode, we pin/unpin all o2hb regions. This solution 138 * works for both file system and userdlm domains. 139 */ 140 static int o2hb_region_pin(const char *region_uuid); 141 static void o2hb_region_unpin(const char *region_uuid); 142 143 /* Only sets a new threshold if there are no active regions. 144 * 145 * No locking or otherwise interesting code is required for reading 146 * o2hb_dead_threshold as it can't change once regions are active and 147 * it's not interesting to anyone until then anyway. */ 148 static void o2hb_dead_threshold_set(unsigned int threshold) 149 { 150 if (threshold > O2HB_MIN_DEAD_THRESHOLD) { 151 spin_lock(&o2hb_live_lock); 152 if (list_empty(&o2hb_all_regions)) 153 o2hb_dead_threshold = threshold; 154 spin_unlock(&o2hb_live_lock); 155 } 156 } 157 158 static int o2hb_global_heartbeat_mode_set(unsigned int hb_mode) 159 { 160 int ret = -1; 161 162 if (hb_mode < O2HB_HEARTBEAT_NUM_MODES) { 163 spin_lock(&o2hb_live_lock); 164 if (list_empty(&o2hb_all_regions)) { 165 o2hb_heartbeat_mode = hb_mode; 166 ret = 0; 167 } 168 spin_unlock(&o2hb_live_lock); 169 } 170 171 return ret; 172 } 173 174 struct o2hb_node_event { 175 struct list_head hn_item; 176 enum o2hb_callback_type hn_event_type; 177 struct o2nm_node *hn_node; 178 int hn_node_num; 179 }; 180 181 struct o2hb_disk_slot { 182 struct o2hb_disk_heartbeat_block *ds_raw_block; 183 u8 ds_node_num; 184 u64 ds_last_time; 185 u64 ds_last_generation; 186 u16 ds_equal_samples; 187 u16 ds_changed_samples; 188 struct list_head ds_live_item; 189 }; 190 191 /* each thread owns a region.. when we're asked to tear down the region 192 * we ask the thread to stop, who cleans up the region */ 193 struct o2hb_region { 194 struct config_item hr_item; 195 196 struct list_head hr_all_item; 197 unsigned hr_unclean_stop:1, 198 hr_aborted_start:1, 199 hr_item_pinned:1, 200 hr_item_dropped:1, 201 hr_node_deleted:1; 202 203 /* protected by the hr_callback_sem */ 204 struct task_struct *hr_task; 205 206 unsigned int hr_blocks; 207 unsigned long long hr_start_block; 208 209 unsigned int hr_block_bits; 210 unsigned int hr_block_bytes; 211 212 unsigned int hr_slots_per_page; 213 unsigned int hr_num_pages; 214 215 struct page **hr_slot_data; 216 struct block_device *hr_bdev; 217 struct o2hb_disk_slot *hr_slots; 218 219 /* live node map of this region */ 220 unsigned long hr_live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)]; 221 unsigned int hr_region_num; 222 223 struct dentry *hr_debug_dir; 224 struct o2hb_debug_buf *hr_db_livenodes; 225 struct o2hb_debug_buf *hr_db_regnum; 226 struct o2hb_debug_buf *hr_db_elapsed_time; 227 struct o2hb_debug_buf *hr_db_pinned; 228 229 /* let the person setting up hb wait for it to return until it 230 * has reached a 'steady' state. This will be fixed when we have 231 * a more complete api that doesn't lead to this sort of fragility. */ 232 atomic_t hr_steady_iterations; 233 234 /* terminate o2hb thread if it does not reach steady state 235 * (hr_steady_iterations == 0) within hr_unsteady_iterations */ 236 atomic_t hr_unsteady_iterations; 237 238 char hr_dev_name[BDEVNAME_SIZE]; 239 240 unsigned int hr_timeout_ms; 241 242 /* randomized as the region goes up and down so that a node 243 * recognizes a node going up and down in one iteration */ 244 u64 hr_generation; 245 246 struct delayed_work hr_write_timeout_work; 247 unsigned long hr_last_timeout_start; 248 249 /* negotiate timer, used to negotiate extending hb timeout. */ 250 struct delayed_work hr_nego_timeout_work; 251 unsigned long hr_nego_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)]; 252 253 /* Used during o2hb_check_slot to hold a copy of the block 254 * being checked because we temporarily have to zero out the 255 * crc field. */ 256 struct o2hb_disk_heartbeat_block *hr_tmp_block; 257 258 /* Message key for negotiate timeout message. */ 259 unsigned int hr_key; 260 struct list_head hr_handler_list; 261 262 /* last hb status, 0 for success, other value for error. */ 263 int hr_last_hb_status; 264 }; 265 266 struct o2hb_bio_wait_ctxt { 267 atomic_t wc_num_reqs; 268 struct completion wc_io_complete; 269 int wc_error; 270 }; 271 272 #define O2HB_NEGO_TIMEOUT_MS (O2HB_MAX_WRITE_TIMEOUT_MS/2) 273 274 enum { 275 O2HB_NEGO_TIMEOUT_MSG = 1, 276 O2HB_NEGO_APPROVE_MSG = 2, 277 }; 278 279 struct o2hb_nego_msg { 280 u8 node_num; 281 }; 282 283 static void o2hb_write_timeout(struct work_struct *work) 284 { 285 int failed, quorum; 286 struct o2hb_region *reg = 287 container_of(work, struct o2hb_region, 288 hr_write_timeout_work.work); 289 290 mlog(ML_ERROR, "Heartbeat write timeout to device %s after %u " 291 "milliseconds\n", reg->hr_dev_name, 292 jiffies_to_msecs(jiffies - reg->hr_last_timeout_start)); 293 294 if (o2hb_global_heartbeat_active()) { 295 spin_lock(&o2hb_live_lock); 296 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap)) 297 set_bit(reg->hr_region_num, o2hb_failed_region_bitmap); 298 failed = bitmap_weight(o2hb_failed_region_bitmap, 299 O2NM_MAX_REGIONS); 300 quorum = bitmap_weight(o2hb_quorum_region_bitmap, 301 O2NM_MAX_REGIONS); 302 spin_unlock(&o2hb_live_lock); 303 304 mlog(ML_HEARTBEAT, "Number of regions %d, failed regions %d\n", 305 quorum, failed); 306 307 /* 308 * Fence if the number of failed regions >= half the number 309 * of quorum regions 310 */ 311 if ((failed << 1) < quorum) 312 return; 313 } 314 315 o2quo_disk_timeout(); 316 } 317 318 static void o2hb_arm_timeout(struct o2hb_region *reg) 319 { 320 /* Arm writeout only after thread reaches steady state */ 321 if (atomic_read(®->hr_steady_iterations) != 0) 322 return; 323 324 mlog(ML_HEARTBEAT, "Queue write timeout for %u ms\n", 325 O2HB_MAX_WRITE_TIMEOUT_MS); 326 327 if (o2hb_global_heartbeat_active()) { 328 spin_lock(&o2hb_live_lock); 329 clear_bit(reg->hr_region_num, o2hb_failed_region_bitmap); 330 spin_unlock(&o2hb_live_lock); 331 } 332 cancel_delayed_work(®->hr_write_timeout_work); 333 schedule_delayed_work(®->hr_write_timeout_work, 334 msecs_to_jiffies(O2HB_MAX_WRITE_TIMEOUT_MS)); 335 336 cancel_delayed_work(®->hr_nego_timeout_work); 337 /* negotiate timeout must be less than write timeout. */ 338 schedule_delayed_work(®->hr_nego_timeout_work, 339 msecs_to_jiffies(O2HB_NEGO_TIMEOUT_MS)); 340 memset(reg->hr_nego_node_bitmap, 0, sizeof(reg->hr_nego_node_bitmap)); 341 } 342 343 static void o2hb_disarm_timeout(struct o2hb_region *reg) 344 { 345 cancel_delayed_work_sync(®->hr_write_timeout_work); 346 cancel_delayed_work_sync(®->hr_nego_timeout_work); 347 } 348 349 static int o2hb_send_nego_msg(int key, int type, u8 target) 350 { 351 struct o2hb_nego_msg msg; 352 int status, ret; 353 354 msg.node_num = o2nm_this_node(); 355 again: 356 ret = o2net_send_message(type, key, &msg, sizeof(msg), 357 target, &status); 358 359 if (ret == -EAGAIN || ret == -ENOMEM) { 360 msleep(100); 361 goto again; 362 } 363 364 return ret; 365 } 366 367 static void o2hb_nego_timeout(struct work_struct *work) 368 { 369 unsigned long live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)]; 370 int master_node, i, ret; 371 struct o2hb_region *reg; 372 373 reg = container_of(work, struct o2hb_region, hr_nego_timeout_work.work); 374 /* don't negotiate timeout if last hb failed since it is very 375 * possible io failed. Should let write timeout fence self. 376 */ 377 if (reg->hr_last_hb_status) 378 return; 379 380 o2hb_fill_node_map(live_node_bitmap, sizeof(live_node_bitmap)); 381 /* lowest node as master node to make negotiate decision. */ 382 master_node = find_next_bit(live_node_bitmap, O2NM_MAX_NODES, 0); 383 384 if (master_node == o2nm_this_node()) { 385 if (!test_bit(master_node, reg->hr_nego_node_bitmap)) { 386 printk(KERN_NOTICE "o2hb: node %d hb write hung for %ds on region %s (%s).\n", 387 o2nm_this_node(), O2HB_NEGO_TIMEOUT_MS/1000, 388 config_item_name(®->hr_item), reg->hr_dev_name); 389 set_bit(master_node, reg->hr_nego_node_bitmap); 390 } 391 if (memcmp(reg->hr_nego_node_bitmap, live_node_bitmap, 392 sizeof(reg->hr_nego_node_bitmap))) { 393 /* check negotiate bitmap every second to do timeout 394 * approve decision. 395 */ 396 schedule_delayed_work(®->hr_nego_timeout_work, 397 msecs_to_jiffies(1000)); 398 399 return; 400 } 401 402 printk(KERN_NOTICE "o2hb: all nodes hb write hung, maybe region %s (%s) is down.\n", 403 config_item_name(®->hr_item), reg->hr_dev_name); 404 /* approve negotiate timeout request. */ 405 o2hb_arm_timeout(reg); 406 407 i = -1; 408 while ((i = find_next_bit(live_node_bitmap, 409 O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) { 410 if (i == master_node) 411 continue; 412 413 mlog(ML_HEARTBEAT, "send NEGO_APPROVE msg to node %d\n", i); 414 ret = o2hb_send_nego_msg(reg->hr_key, 415 O2HB_NEGO_APPROVE_MSG, i); 416 if (ret) 417 mlog(ML_ERROR, "send NEGO_APPROVE msg to node %d fail %d\n", 418 i, ret); 419 } 420 } else { 421 /* negotiate timeout with master node. */ 422 printk(KERN_NOTICE "o2hb: node %d hb write hung for %ds on region %s (%s), negotiate timeout with node %d.\n", 423 o2nm_this_node(), O2HB_NEGO_TIMEOUT_MS/1000, config_item_name(®->hr_item), 424 reg->hr_dev_name, master_node); 425 ret = o2hb_send_nego_msg(reg->hr_key, O2HB_NEGO_TIMEOUT_MSG, 426 master_node); 427 if (ret) 428 mlog(ML_ERROR, "send NEGO_TIMEOUT msg to node %d fail %d\n", 429 master_node, ret); 430 } 431 } 432 433 static int o2hb_nego_timeout_handler(struct o2net_msg *msg, u32 len, void *data, 434 void **ret_data) 435 { 436 struct o2hb_region *reg = data; 437 struct o2hb_nego_msg *nego_msg; 438 439 nego_msg = (struct o2hb_nego_msg *)msg->buf; 440 printk(KERN_NOTICE "o2hb: receive negotiate timeout message from node %d on region %s (%s).\n", 441 nego_msg->node_num, config_item_name(®->hr_item), reg->hr_dev_name); 442 if (nego_msg->node_num < O2NM_MAX_NODES) 443 set_bit(nego_msg->node_num, reg->hr_nego_node_bitmap); 444 else 445 mlog(ML_ERROR, "got nego timeout message from bad node.\n"); 446 447 return 0; 448 } 449 450 static int o2hb_nego_approve_handler(struct o2net_msg *msg, u32 len, void *data, 451 void **ret_data) 452 { 453 struct o2hb_region *reg = data; 454 455 printk(KERN_NOTICE "o2hb: negotiate timeout approved by master node on region %s (%s).\n", 456 config_item_name(®->hr_item), reg->hr_dev_name); 457 o2hb_arm_timeout(reg); 458 return 0; 459 } 460 461 static inline void o2hb_bio_wait_init(struct o2hb_bio_wait_ctxt *wc) 462 { 463 atomic_set(&wc->wc_num_reqs, 1); 464 init_completion(&wc->wc_io_complete); 465 wc->wc_error = 0; 466 } 467 468 /* Used in error paths too */ 469 static inline void o2hb_bio_wait_dec(struct o2hb_bio_wait_ctxt *wc, 470 unsigned int num) 471 { 472 /* sadly atomic_sub_and_test() isn't available on all platforms. The 473 * good news is that the fast path only completes one at a time */ 474 while(num--) { 475 if (atomic_dec_and_test(&wc->wc_num_reqs)) { 476 BUG_ON(num > 0); 477 complete(&wc->wc_io_complete); 478 } 479 } 480 } 481 482 static void o2hb_wait_on_io(struct o2hb_bio_wait_ctxt *wc) 483 { 484 o2hb_bio_wait_dec(wc, 1); 485 wait_for_completion(&wc->wc_io_complete); 486 } 487 488 static void o2hb_bio_end_io(struct bio *bio) 489 { 490 struct o2hb_bio_wait_ctxt *wc = bio->bi_private; 491 492 if (bio->bi_status) { 493 mlog(ML_ERROR, "IO Error %d\n", bio->bi_status); 494 wc->wc_error = blk_status_to_errno(bio->bi_status); 495 } 496 497 o2hb_bio_wait_dec(wc, 1); 498 bio_put(bio); 499 } 500 501 /* Setup a Bio to cover I/O against num_slots slots starting at 502 * start_slot. */ 503 static struct bio *o2hb_setup_one_bio(struct o2hb_region *reg, 504 struct o2hb_bio_wait_ctxt *wc, 505 unsigned int *current_slot, 506 unsigned int max_slots, int op, 507 int op_flags) 508 { 509 int len, current_page; 510 unsigned int vec_len, vec_start; 511 unsigned int bits = reg->hr_block_bits; 512 unsigned int spp = reg->hr_slots_per_page; 513 unsigned int cs = *current_slot; 514 struct bio *bio; 515 struct page *page; 516 517 /* Testing has shown this allocation to take long enough under 518 * GFP_KERNEL that the local node can get fenced. It would be 519 * nicest if we could pre-allocate these bios and avoid this 520 * all together. */ 521 bio = bio_alloc(GFP_ATOMIC, 16); 522 if (!bio) { 523 mlog(ML_ERROR, "Could not alloc slots BIO!\n"); 524 bio = ERR_PTR(-ENOMEM); 525 goto bail; 526 } 527 528 /* Must put everything in 512 byte sectors for the bio... */ 529 bio->bi_iter.bi_sector = (reg->hr_start_block + cs) << (bits - 9); 530 bio_set_dev(bio, reg->hr_bdev); 531 bio->bi_private = wc; 532 bio->bi_end_io = o2hb_bio_end_io; 533 bio_set_op_attrs(bio, op, op_flags); 534 535 vec_start = (cs << bits) % PAGE_SIZE; 536 while(cs < max_slots) { 537 current_page = cs / spp; 538 page = reg->hr_slot_data[current_page]; 539 540 vec_len = min(PAGE_SIZE - vec_start, 541 (max_slots-cs) * (PAGE_SIZE/spp) ); 542 543 mlog(ML_HB_BIO, "page %d, vec_len = %u, vec_start = %u\n", 544 current_page, vec_len, vec_start); 545 546 len = bio_add_page(bio, page, vec_len, vec_start); 547 if (len != vec_len) break; 548 549 cs += vec_len / (PAGE_SIZE/spp); 550 vec_start = 0; 551 } 552 553 bail: 554 *current_slot = cs; 555 return bio; 556 } 557 558 static int o2hb_read_slots(struct o2hb_region *reg, 559 unsigned int begin_slot, 560 unsigned int max_slots) 561 { 562 unsigned int current_slot = begin_slot; 563 int status; 564 struct o2hb_bio_wait_ctxt wc; 565 struct bio *bio; 566 567 o2hb_bio_wait_init(&wc); 568 569 while(current_slot < max_slots) { 570 bio = o2hb_setup_one_bio(reg, &wc, ¤t_slot, max_slots, 571 REQ_OP_READ, 0); 572 if (IS_ERR(bio)) { 573 status = PTR_ERR(bio); 574 mlog_errno(status); 575 goto bail_and_wait; 576 } 577 578 atomic_inc(&wc.wc_num_reqs); 579 submit_bio(bio); 580 } 581 582 status = 0; 583 584 bail_and_wait: 585 o2hb_wait_on_io(&wc); 586 if (wc.wc_error && !status) 587 status = wc.wc_error; 588 589 return status; 590 } 591 592 static int o2hb_issue_node_write(struct o2hb_region *reg, 593 struct o2hb_bio_wait_ctxt *write_wc) 594 { 595 int status; 596 unsigned int slot; 597 struct bio *bio; 598 599 o2hb_bio_wait_init(write_wc); 600 601 slot = o2nm_this_node(); 602 603 bio = o2hb_setup_one_bio(reg, write_wc, &slot, slot+1, REQ_OP_WRITE, 604 REQ_SYNC); 605 if (IS_ERR(bio)) { 606 status = PTR_ERR(bio); 607 mlog_errno(status); 608 goto bail; 609 } 610 611 atomic_inc(&write_wc->wc_num_reqs); 612 submit_bio(bio); 613 614 status = 0; 615 bail: 616 return status; 617 } 618 619 static u32 o2hb_compute_block_crc_le(struct o2hb_region *reg, 620 struct o2hb_disk_heartbeat_block *hb_block) 621 { 622 __le32 old_cksum; 623 u32 ret; 624 625 /* We want to compute the block crc with a 0 value in the 626 * hb_cksum field. Save it off here and replace after the 627 * crc. */ 628 old_cksum = hb_block->hb_cksum; 629 hb_block->hb_cksum = 0; 630 631 ret = crc32_le(0, (unsigned char *) hb_block, reg->hr_block_bytes); 632 633 hb_block->hb_cksum = old_cksum; 634 635 return ret; 636 } 637 638 static void o2hb_dump_slot(struct o2hb_disk_heartbeat_block *hb_block) 639 { 640 mlog(ML_ERROR, "Dump slot information: seq = 0x%llx, node = %u, " 641 "cksum = 0x%x, generation 0x%llx\n", 642 (long long)le64_to_cpu(hb_block->hb_seq), 643 hb_block->hb_node, le32_to_cpu(hb_block->hb_cksum), 644 (long long)le64_to_cpu(hb_block->hb_generation)); 645 } 646 647 static int o2hb_verify_crc(struct o2hb_region *reg, 648 struct o2hb_disk_heartbeat_block *hb_block) 649 { 650 u32 read, computed; 651 652 read = le32_to_cpu(hb_block->hb_cksum); 653 computed = o2hb_compute_block_crc_le(reg, hb_block); 654 655 return read == computed; 656 } 657 658 /* 659 * Compare the slot data with what we wrote in the last iteration. 660 * If the match fails, print an appropriate error message. This is to 661 * detect errors like... another node hearting on the same slot, 662 * flaky device that is losing writes, etc. 663 * Returns 1 if check succeeds, 0 otherwise. 664 */ 665 static int o2hb_check_own_slot(struct o2hb_region *reg) 666 { 667 struct o2hb_disk_slot *slot; 668 struct o2hb_disk_heartbeat_block *hb_block; 669 char *errstr; 670 671 slot = ®->hr_slots[o2nm_this_node()]; 672 /* Don't check on our 1st timestamp */ 673 if (!slot->ds_last_time) 674 return 0; 675 676 hb_block = slot->ds_raw_block; 677 if (le64_to_cpu(hb_block->hb_seq) == slot->ds_last_time && 678 le64_to_cpu(hb_block->hb_generation) == slot->ds_last_generation && 679 hb_block->hb_node == slot->ds_node_num) 680 return 1; 681 682 #define ERRSTR1 "Another node is heartbeating on device" 683 #define ERRSTR2 "Heartbeat generation mismatch on device" 684 #define ERRSTR3 "Heartbeat sequence mismatch on device" 685 686 if (hb_block->hb_node != slot->ds_node_num) 687 errstr = ERRSTR1; 688 else if (le64_to_cpu(hb_block->hb_generation) != 689 slot->ds_last_generation) 690 errstr = ERRSTR2; 691 else 692 errstr = ERRSTR3; 693 694 mlog(ML_ERROR, "%s (%s): expected(%u:0x%llx, 0x%llx), " 695 "ondisk(%u:0x%llx, 0x%llx)\n", errstr, reg->hr_dev_name, 696 slot->ds_node_num, (unsigned long long)slot->ds_last_generation, 697 (unsigned long long)slot->ds_last_time, hb_block->hb_node, 698 (unsigned long long)le64_to_cpu(hb_block->hb_generation), 699 (unsigned long long)le64_to_cpu(hb_block->hb_seq)); 700 701 return 0; 702 } 703 704 static inline void o2hb_prepare_block(struct o2hb_region *reg, 705 u64 generation) 706 { 707 int node_num; 708 u64 cputime; 709 struct o2hb_disk_slot *slot; 710 struct o2hb_disk_heartbeat_block *hb_block; 711 712 node_num = o2nm_this_node(); 713 slot = ®->hr_slots[node_num]; 714 715 hb_block = (struct o2hb_disk_heartbeat_block *)slot->ds_raw_block; 716 memset(hb_block, 0, reg->hr_block_bytes); 717 /* TODO: time stuff */ 718 cputime = ktime_get_real_seconds(); 719 if (!cputime) 720 cputime = 1; 721 722 hb_block->hb_seq = cpu_to_le64(cputime); 723 hb_block->hb_node = node_num; 724 hb_block->hb_generation = cpu_to_le64(generation); 725 hb_block->hb_dead_ms = cpu_to_le32(o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS); 726 727 /* This step must always happen last! */ 728 hb_block->hb_cksum = cpu_to_le32(o2hb_compute_block_crc_le(reg, 729 hb_block)); 730 731 mlog(ML_HB_BIO, "our node generation = 0x%llx, cksum = 0x%x\n", 732 (long long)generation, 733 le32_to_cpu(hb_block->hb_cksum)); 734 } 735 736 static void o2hb_fire_callbacks(struct o2hb_callback *hbcall, 737 struct o2nm_node *node, 738 int idx) 739 { 740 struct o2hb_callback_func *f; 741 742 list_for_each_entry(f, &hbcall->list, hc_item) { 743 mlog(ML_HEARTBEAT, "calling funcs %p\n", f); 744 (f->hc_func)(node, idx, f->hc_data); 745 } 746 } 747 748 /* Will run the list in order until we process the passed event */ 749 static void o2hb_run_event_list(struct o2hb_node_event *queued_event) 750 { 751 struct o2hb_callback *hbcall; 752 struct o2hb_node_event *event; 753 754 /* Holding callback sem assures we don't alter the callback 755 * lists when doing this, and serializes ourselves with other 756 * processes wanting callbacks. */ 757 down_write(&o2hb_callback_sem); 758 759 spin_lock(&o2hb_live_lock); 760 while (!list_empty(&o2hb_node_events) 761 && !list_empty(&queued_event->hn_item)) { 762 event = list_entry(o2hb_node_events.next, 763 struct o2hb_node_event, 764 hn_item); 765 list_del_init(&event->hn_item); 766 spin_unlock(&o2hb_live_lock); 767 768 mlog(ML_HEARTBEAT, "Node %s event for %d\n", 769 event->hn_event_type == O2HB_NODE_UP_CB ? "UP" : "DOWN", 770 event->hn_node_num); 771 772 hbcall = hbcall_from_type(event->hn_event_type); 773 774 /* We should *never* have gotten on to the list with a 775 * bad type... This isn't something that we should try 776 * to recover from. */ 777 BUG_ON(IS_ERR(hbcall)); 778 779 o2hb_fire_callbacks(hbcall, event->hn_node, event->hn_node_num); 780 781 spin_lock(&o2hb_live_lock); 782 } 783 spin_unlock(&o2hb_live_lock); 784 785 up_write(&o2hb_callback_sem); 786 } 787 788 static void o2hb_queue_node_event(struct o2hb_node_event *event, 789 enum o2hb_callback_type type, 790 struct o2nm_node *node, 791 int node_num) 792 { 793 assert_spin_locked(&o2hb_live_lock); 794 795 BUG_ON((!node) && (type != O2HB_NODE_DOWN_CB)); 796 797 event->hn_event_type = type; 798 event->hn_node = node; 799 event->hn_node_num = node_num; 800 801 mlog(ML_HEARTBEAT, "Queue node %s event for node %d\n", 802 type == O2HB_NODE_UP_CB ? "UP" : "DOWN", node_num); 803 804 list_add_tail(&event->hn_item, &o2hb_node_events); 805 } 806 807 static void o2hb_shutdown_slot(struct o2hb_disk_slot *slot) 808 { 809 struct o2hb_node_event event = 810 { .hn_item = LIST_HEAD_INIT(event.hn_item), }; 811 struct o2nm_node *node; 812 int queued = 0; 813 814 node = o2nm_get_node_by_num(slot->ds_node_num); 815 if (!node) 816 return; 817 818 spin_lock(&o2hb_live_lock); 819 if (!list_empty(&slot->ds_live_item)) { 820 mlog(ML_HEARTBEAT, "Shutdown, node %d leaves region\n", 821 slot->ds_node_num); 822 823 list_del_init(&slot->ds_live_item); 824 825 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) { 826 clear_bit(slot->ds_node_num, o2hb_live_node_bitmap); 827 828 o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB, node, 829 slot->ds_node_num); 830 queued = 1; 831 } 832 } 833 spin_unlock(&o2hb_live_lock); 834 835 if (queued) 836 o2hb_run_event_list(&event); 837 838 o2nm_node_put(node); 839 } 840 841 static void o2hb_set_quorum_device(struct o2hb_region *reg) 842 { 843 if (!o2hb_global_heartbeat_active()) 844 return; 845 846 /* Prevent race with o2hb_heartbeat_group_drop_item() */ 847 if (kthread_should_stop()) 848 return; 849 850 /* Tag region as quorum only after thread reaches steady state */ 851 if (atomic_read(®->hr_steady_iterations) != 0) 852 return; 853 854 spin_lock(&o2hb_live_lock); 855 856 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap)) 857 goto unlock; 858 859 /* 860 * A region can be added to the quorum only when it sees all 861 * live nodes heartbeat on it. In other words, the region has been 862 * added to all nodes. 863 */ 864 if (memcmp(reg->hr_live_node_bitmap, o2hb_live_node_bitmap, 865 sizeof(o2hb_live_node_bitmap))) 866 goto unlock; 867 868 printk(KERN_NOTICE "o2hb: Region %s (%s) is now a quorum device\n", 869 config_item_name(®->hr_item), reg->hr_dev_name); 870 871 set_bit(reg->hr_region_num, o2hb_quorum_region_bitmap); 872 873 /* 874 * If global heartbeat active, unpin all regions if the 875 * region count > CUT_OFF 876 */ 877 if (bitmap_weight(o2hb_quorum_region_bitmap, 878 O2NM_MAX_REGIONS) > O2HB_PIN_CUT_OFF) 879 o2hb_region_unpin(NULL); 880 unlock: 881 spin_unlock(&o2hb_live_lock); 882 } 883 884 static int o2hb_check_slot(struct o2hb_region *reg, 885 struct o2hb_disk_slot *slot) 886 { 887 int changed = 0, gen_changed = 0; 888 struct o2hb_node_event event = 889 { .hn_item = LIST_HEAD_INIT(event.hn_item), }; 890 struct o2nm_node *node; 891 struct o2hb_disk_heartbeat_block *hb_block = reg->hr_tmp_block; 892 u64 cputime; 893 unsigned int dead_ms = o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS; 894 unsigned int slot_dead_ms; 895 int tmp; 896 int queued = 0; 897 898 memcpy(hb_block, slot->ds_raw_block, reg->hr_block_bytes); 899 900 /* 901 * If a node is no longer configured but is still in the livemap, we 902 * may need to clear that bit from the livemap. 903 */ 904 node = o2nm_get_node_by_num(slot->ds_node_num); 905 if (!node) { 906 spin_lock(&o2hb_live_lock); 907 tmp = test_bit(slot->ds_node_num, o2hb_live_node_bitmap); 908 spin_unlock(&o2hb_live_lock); 909 if (!tmp) 910 return 0; 911 } 912 913 if (!o2hb_verify_crc(reg, hb_block)) { 914 /* all paths from here will drop o2hb_live_lock for 915 * us. */ 916 spin_lock(&o2hb_live_lock); 917 918 /* Don't print an error on the console in this case - 919 * a freshly formatted heartbeat area will not have a 920 * crc set on it. */ 921 if (list_empty(&slot->ds_live_item)) 922 goto out; 923 924 /* The node is live but pushed out a bad crc. We 925 * consider it a transient miss but don't populate any 926 * other values as they may be junk. */ 927 mlog(ML_ERROR, "Node %d has written a bad crc to %s\n", 928 slot->ds_node_num, reg->hr_dev_name); 929 o2hb_dump_slot(hb_block); 930 931 slot->ds_equal_samples++; 932 goto fire_callbacks; 933 } 934 935 /* we don't care if these wrap.. the state transitions below 936 * clear at the right places */ 937 cputime = le64_to_cpu(hb_block->hb_seq); 938 if (slot->ds_last_time != cputime) 939 slot->ds_changed_samples++; 940 else 941 slot->ds_equal_samples++; 942 slot->ds_last_time = cputime; 943 944 /* The node changed heartbeat generations. We assume this to 945 * mean it dropped off but came back before we timed out. We 946 * want to consider it down for the time being but don't want 947 * to lose any changed_samples state we might build up to 948 * considering it live again. */ 949 if (slot->ds_last_generation != le64_to_cpu(hb_block->hb_generation)) { 950 gen_changed = 1; 951 slot->ds_equal_samples = 0; 952 mlog(ML_HEARTBEAT, "Node %d changed generation (0x%llx " 953 "to 0x%llx)\n", slot->ds_node_num, 954 (long long)slot->ds_last_generation, 955 (long long)le64_to_cpu(hb_block->hb_generation)); 956 } 957 958 slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation); 959 960 mlog(ML_HEARTBEAT, "Slot %d gen 0x%llx cksum 0x%x " 961 "seq %llu last %llu changed %u equal %u\n", 962 slot->ds_node_num, (long long)slot->ds_last_generation, 963 le32_to_cpu(hb_block->hb_cksum), 964 (unsigned long long)le64_to_cpu(hb_block->hb_seq), 965 (unsigned long long)slot->ds_last_time, slot->ds_changed_samples, 966 slot->ds_equal_samples); 967 968 spin_lock(&o2hb_live_lock); 969 970 fire_callbacks: 971 /* dead nodes only come to life after some number of 972 * changes at any time during their dead time */ 973 if (list_empty(&slot->ds_live_item) && 974 slot->ds_changed_samples >= O2HB_LIVE_THRESHOLD) { 975 mlog(ML_HEARTBEAT, "Node %d (id 0x%llx) joined my region\n", 976 slot->ds_node_num, (long long)slot->ds_last_generation); 977 978 set_bit(slot->ds_node_num, reg->hr_live_node_bitmap); 979 980 /* first on the list generates a callback */ 981 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) { 982 mlog(ML_HEARTBEAT, "o2hb: Add node %d to live nodes " 983 "bitmap\n", slot->ds_node_num); 984 set_bit(slot->ds_node_num, o2hb_live_node_bitmap); 985 986 o2hb_queue_node_event(&event, O2HB_NODE_UP_CB, node, 987 slot->ds_node_num); 988 989 changed = 1; 990 queued = 1; 991 } 992 993 list_add_tail(&slot->ds_live_item, 994 &o2hb_live_slots[slot->ds_node_num]); 995 996 slot->ds_equal_samples = 0; 997 998 /* We want to be sure that all nodes agree on the 999 * number of milliseconds before a node will be 1000 * considered dead. The self-fencing timeout is 1001 * computed from this value, and a discrepancy might 1002 * result in heartbeat calling a node dead when it 1003 * hasn't self-fenced yet. */ 1004 slot_dead_ms = le32_to_cpu(hb_block->hb_dead_ms); 1005 if (slot_dead_ms && slot_dead_ms != dead_ms) { 1006 /* TODO: Perhaps we can fail the region here. */ 1007 mlog(ML_ERROR, "Node %d on device %s has a dead count " 1008 "of %u ms, but our count is %u ms.\n" 1009 "Please double check your configuration values " 1010 "for 'O2CB_HEARTBEAT_THRESHOLD'\n", 1011 slot->ds_node_num, reg->hr_dev_name, slot_dead_ms, 1012 dead_ms); 1013 } 1014 goto out; 1015 } 1016 1017 /* if the list is dead, we're done.. */ 1018 if (list_empty(&slot->ds_live_item)) 1019 goto out; 1020 1021 /* live nodes only go dead after enough consequtive missed 1022 * samples.. reset the missed counter whenever we see 1023 * activity */ 1024 if (slot->ds_equal_samples >= o2hb_dead_threshold || gen_changed) { 1025 mlog(ML_HEARTBEAT, "Node %d left my region\n", 1026 slot->ds_node_num); 1027 1028 clear_bit(slot->ds_node_num, reg->hr_live_node_bitmap); 1029 1030 /* last off the live_slot generates a callback */ 1031 list_del_init(&slot->ds_live_item); 1032 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) { 1033 mlog(ML_HEARTBEAT, "o2hb: Remove node %d from live " 1034 "nodes bitmap\n", slot->ds_node_num); 1035 clear_bit(slot->ds_node_num, o2hb_live_node_bitmap); 1036 1037 /* node can be null */ 1038 o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB, 1039 node, slot->ds_node_num); 1040 1041 changed = 1; 1042 queued = 1; 1043 } 1044 1045 /* We don't clear this because the node is still 1046 * actually writing new blocks. */ 1047 if (!gen_changed) 1048 slot->ds_changed_samples = 0; 1049 goto out; 1050 } 1051 if (slot->ds_changed_samples) { 1052 slot->ds_changed_samples = 0; 1053 slot->ds_equal_samples = 0; 1054 } 1055 out: 1056 spin_unlock(&o2hb_live_lock); 1057 1058 if (queued) 1059 o2hb_run_event_list(&event); 1060 1061 if (node) 1062 o2nm_node_put(node); 1063 return changed; 1064 } 1065 1066 static int o2hb_highest_node(unsigned long *nodes, int numbits) 1067 { 1068 return find_last_bit(nodes, numbits); 1069 } 1070 1071 static int o2hb_lowest_node(unsigned long *nodes, int numbits) 1072 { 1073 return find_first_bit(nodes, numbits); 1074 } 1075 1076 static int o2hb_do_disk_heartbeat(struct o2hb_region *reg) 1077 { 1078 int i, ret, highest_node, lowest_node; 1079 int membership_change = 0, own_slot_ok = 0; 1080 unsigned long configured_nodes[BITS_TO_LONGS(O2NM_MAX_NODES)]; 1081 unsigned long live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)]; 1082 struct o2hb_bio_wait_ctxt write_wc; 1083 1084 ret = o2nm_configured_node_map(configured_nodes, 1085 sizeof(configured_nodes)); 1086 if (ret) { 1087 mlog_errno(ret); 1088 goto bail; 1089 } 1090 1091 /* 1092 * If a node is not configured but is in the livemap, we still need 1093 * to read the slot so as to be able to remove it from the livemap. 1094 */ 1095 o2hb_fill_node_map(live_node_bitmap, sizeof(live_node_bitmap)); 1096 i = -1; 1097 while ((i = find_next_bit(live_node_bitmap, 1098 O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) { 1099 set_bit(i, configured_nodes); 1100 } 1101 1102 highest_node = o2hb_highest_node(configured_nodes, O2NM_MAX_NODES); 1103 lowest_node = o2hb_lowest_node(configured_nodes, O2NM_MAX_NODES); 1104 if (highest_node >= O2NM_MAX_NODES || lowest_node >= O2NM_MAX_NODES) { 1105 mlog(ML_NOTICE, "o2hb: No configured nodes found!\n"); 1106 ret = -EINVAL; 1107 goto bail; 1108 } 1109 1110 /* No sense in reading the slots of nodes that don't exist 1111 * yet. Of course, if the node definitions have holes in them 1112 * then we're reading an empty slot anyway... Consider this 1113 * best-effort. */ 1114 ret = o2hb_read_slots(reg, lowest_node, highest_node + 1); 1115 if (ret < 0) { 1116 mlog_errno(ret); 1117 goto bail; 1118 } 1119 1120 /* With an up to date view of the slots, we can check that no 1121 * other node has been improperly configured to heartbeat in 1122 * our slot. */ 1123 own_slot_ok = o2hb_check_own_slot(reg); 1124 1125 /* fill in the proper info for our next heartbeat */ 1126 o2hb_prepare_block(reg, reg->hr_generation); 1127 1128 ret = o2hb_issue_node_write(reg, &write_wc); 1129 if (ret < 0) { 1130 mlog_errno(ret); 1131 goto bail; 1132 } 1133 1134 i = -1; 1135 while((i = find_next_bit(configured_nodes, 1136 O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) { 1137 membership_change |= o2hb_check_slot(reg, ®->hr_slots[i]); 1138 } 1139 1140 /* 1141 * We have to be sure we've advertised ourselves on disk 1142 * before we can go to steady state. This ensures that 1143 * people we find in our steady state have seen us. 1144 */ 1145 o2hb_wait_on_io(&write_wc); 1146 if (write_wc.wc_error) { 1147 /* Do not re-arm the write timeout on I/O error - we 1148 * can't be sure that the new block ever made it to 1149 * disk */ 1150 mlog(ML_ERROR, "Write error %d on device \"%s\"\n", 1151 write_wc.wc_error, reg->hr_dev_name); 1152 ret = write_wc.wc_error; 1153 goto bail; 1154 } 1155 1156 /* Skip disarming the timeout if own slot has stale/bad data */ 1157 if (own_slot_ok) { 1158 o2hb_set_quorum_device(reg); 1159 o2hb_arm_timeout(reg); 1160 reg->hr_last_timeout_start = jiffies; 1161 } 1162 1163 bail: 1164 /* let the person who launched us know when things are steady */ 1165 if (atomic_read(®->hr_steady_iterations) != 0) { 1166 if (!ret && own_slot_ok && !membership_change) { 1167 if (atomic_dec_and_test(®->hr_steady_iterations)) 1168 wake_up(&o2hb_steady_queue); 1169 } 1170 } 1171 1172 if (atomic_read(®->hr_steady_iterations) != 0) { 1173 if (atomic_dec_and_test(®->hr_unsteady_iterations)) { 1174 printk(KERN_NOTICE "o2hb: Unable to stabilize " 1175 "heartbeat on region %s (%s)\n", 1176 config_item_name(®->hr_item), 1177 reg->hr_dev_name); 1178 atomic_set(®->hr_steady_iterations, 0); 1179 reg->hr_aborted_start = 1; 1180 wake_up(&o2hb_steady_queue); 1181 ret = -EIO; 1182 } 1183 } 1184 1185 return ret; 1186 } 1187 1188 /* 1189 * we ride the region ref that the region dir holds. before the region 1190 * dir is removed and drops it ref it will wait to tear down this 1191 * thread. 1192 */ 1193 static int o2hb_thread(void *data) 1194 { 1195 int i, ret; 1196 struct o2hb_region *reg = data; 1197 struct o2hb_bio_wait_ctxt write_wc; 1198 ktime_t before_hb, after_hb; 1199 unsigned int elapsed_msec; 1200 1201 mlog(ML_HEARTBEAT|ML_KTHREAD, "hb thread running\n"); 1202 1203 set_user_nice(current, MIN_NICE); 1204 1205 /* Pin node */ 1206 ret = o2nm_depend_this_node(); 1207 if (ret) { 1208 mlog(ML_ERROR, "Node has been deleted, ret = %d\n", ret); 1209 reg->hr_node_deleted = 1; 1210 wake_up(&o2hb_steady_queue); 1211 return 0; 1212 } 1213 1214 while (!kthread_should_stop() && 1215 !reg->hr_unclean_stop && !reg->hr_aborted_start) { 1216 /* We track the time spent inside 1217 * o2hb_do_disk_heartbeat so that we avoid more than 1218 * hr_timeout_ms between disk writes. On busy systems 1219 * this should result in a heartbeat which is less 1220 * likely to time itself out. */ 1221 before_hb = ktime_get_real(); 1222 1223 ret = o2hb_do_disk_heartbeat(reg); 1224 reg->hr_last_hb_status = ret; 1225 1226 after_hb = ktime_get_real(); 1227 1228 elapsed_msec = (unsigned int) 1229 ktime_ms_delta(after_hb, before_hb); 1230 1231 mlog(ML_HEARTBEAT, 1232 "start = %lld, end = %lld, msec = %u, ret = %d\n", 1233 before_hb, after_hb, elapsed_msec, ret); 1234 1235 if (!kthread_should_stop() && 1236 elapsed_msec < reg->hr_timeout_ms) { 1237 /* the kthread api has blocked signals for us so no 1238 * need to record the return value. */ 1239 msleep_interruptible(reg->hr_timeout_ms - elapsed_msec); 1240 } 1241 } 1242 1243 o2hb_disarm_timeout(reg); 1244 1245 /* unclean stop is only used in very bad situation */ 1246 for(i = 0; !reg->hr_unclean_stop && i < reg->hr_blocks; i++) 1247 o2hb_shutdown_slot(®->hr_slots[i]); 1248 1249 /* Explicit down notification - avoid forcing the other nodes 1250 * to timeout on this region when we could just as easily 1251 * write a clear generation - thus indicating to them that 1252 * this node has left this region. 1253 */ 1254 if (!reg->hr_unclean_stop && !reg->hr_aborted_start) { 1255 o2hb_prepare_block(reg, 0); 1256 ret = o2hb_issue_node_write(reg, &write_wc); 1257 if (ret == 0) 1258 o2hb_wait_on_io(&write_wc); 1259 else 1260 mlog_errno(ret); 1261 } 1262 1263 /* Unpin node */ 1264 o2nm_undepend_this_node(); 1265 1266 mlog(ML_HEARTBEAT|ML_KTHREAD, "o2hb thread exiting\n"); 1267 1268 return 0; 1269 } 1270 1271 #ifdef CONFIG_DEBUG_FS 1272 static int o2hb_debug_open(struct inode *inode, struct file *file) 1273 { 1274 struct o2hb_debug_buf *db = inode->i_private; 1275 struct o2hb_region *reg; 1276 unsigned long map[BITS_TO_LONGS(O2NM_MAX_NODES)]; 1277 unsigned long lts; 1278 char *buf = NULL; 1279 int i = -1; 1280 int out = 0; 1281 1282 /* max_nodes should be the largest bitmap we pass here */ 1283 BUG_ON(sizeof(map) < db->db_size); 1284 1285 buf = kmalloc(PAGE_SIZE, GFP_KERNEL); 1286 if (!buf) 1287 goto bail; 1288 1289 switch (db->db_type) { 1290 case O2HB_DB_TYPE_LIVENODES: 1291 case O2HB_DB_TYPE_LIVEREGIONS: 1292 case O2HB_DB_TYPE_QUORUMREGIONS: 1293 case O2HB_DB_TYPE_FAILEDREGIONS: 1294 spin_lock(&o2hb_live_lock); 1295 memcpy(map, db->db_data, db->db_size); 1296 spin_unlock(&o2hb_live_lock); 1297 break; 1298 1299 case O2HB_DB_TYPE_REGION_LIVENODES: 1300 spin_lock(&o2hb_live_lock); 1301 reg = (struct o2hb_region *)db->db_data; 1302 memcpy(map, reg->hr_live_node_bitmap, db->db_size); 1303 spin_unlock(&o2hb_live_lock); 1304 break; 1305 1306 case O2HB_DB_TYPE_REGION_NUMBER: 1307 reg = (struct o2hb_region *)db->db_data; 1308 out += scnprintf(buf + out, PAGE_SIZE - out, "%d\n", 1309 reg->hr_region_num); 1310 goto done; 1311 1312 case O2HB_DB_TYPE_REGION_ELAPSED_TIME: 1313 reg = (struct o2hb_region *)db->db_data; 1314 lts = reg->hr_last_timeout_start; 1315 /* If 0, it has never been set before */ 1316 if (lts) 1317 lts = jiffies_to_msecs(jiffies - lts); 1318 out += scnprintf(buf + out, PAGE_SIZE - out, "%lu\n", lts); 1319 goto done; 1320 1321 case O2HB_DB_TYPE_REGION_PINNED: 1322 reg = (struct o2hb_region *)db->db_data; 1323 out += scnprintf(buf + out, PAGE_SIZE - out, "%u\n", 1324 !!reg->hr_item_pinned); 1325 goto done; 1326 1327 default: 1328 goto done; 1329 } 1330 1331 while ((i = find_next_bit(map, db->db_len, i + 1)) < db->db_len) 1332 out += scnprintf(buf + out, PAGE_SIZE - out, "%d ", i); 1333 out += scnprintf(buf + out, PAGE_SIZE - out, "\n"); 1334 1335 done: 1336 i_size_write(inode, out); 1337 1338 file->private_data = buf; 1339 1340 return 0; 1341 bail: 1342 return -ENOMEM; 1343 } 1344 1345 static int o2hb_debug_release(struct inode *inode, struct file *file) 1346 { 1347 kfree(file->private_data); 1348 return 0; 1349 } 1350 1351 static ssize_t o2hb_debug_read(struct file *file, char __user *buf, 1352 size_t nbytes, loff_t *ppos) 1353 { 1354 return simple_read_from_buffer(buf, nbytes, ppos, file->private_data, 1355 i_size_read(file->f_mapping->host)); 1356 } 1357 #else 1358 static int o2hb_debug_open(struct inode *inode, struct file *file) 1359 { 1360 return 0; 1361 } 1362 static int o2hb_debug_release(struct inode *inode, struct file *file) 1363 { 1364 return 0; 1365 } 1366 static ssize_t o2hb_debug_read(struct file *file, char __user *buf, 1367 size_t nbytes, loff_t *ppos) 1368 { 1369 return 0; 1370 } 1371 #endif /* CONFIG_DEBUG_FS */ 1372 1373 static const struct file_operations o2hb_debug_fops = { 1374 .open = o2hb_debug_open, 1375 .release = o2hb_debug_release, 1376 .read = o2hb_debug_read, 1377 .llseek = generic_file_llseek, 1378 }; 1379 1380 void o2hb_exit(void) 1381 { 1382 debugfs_remove_recursive(o2hb_debug_dir); 1383 kfree(o2hb_db_livenodes); 1384 kfree(o2hb_db_liveregions); 1385 kfree(o2hb_db_quorumregions); 1386 kfree(o2hb_db_failedregions); 1387 } 1388 1389 static void o2hb_debug_create(const char *name, struct dentry *dir, 1390 struct o2hb_debug_buf **db, int db_len, int type, 1391 int size, int len, void *data) 1392 { 1393 *db = kmalloc(db_len, GFP_KERNEL); 1394 if (!*db) 1395 return; 1396 1397 (*db)->db_type = type; 1398 (*db)->db_size = size; 1399 (*db)->db_len = len; 1400 (*db)->db_data = data; 1401 1402 debugfs_create_file(name, S_IFREG|S_IRUSR, dir, *db, &o2hb_debug_fops); 1403 } 1404 1405 static void o2hb_debug_init(void) 1406 { 1407 o2hb_debug_dir = debugfs_create_dir(O2HB_DEBUG_DIR, NULL); 1408 1409 o2hb_debug_create(O2HB_DEBUG_LIVENODES, o2hb_debug_dir, 1410 &o2hb_db_livenodes, sizeof(*o2hb_db_livenodes), 1411 O2HB_DB_TYPE_LIVENODES, sizeof(o2hb_live_node_bitmap), 1412 O2NM_MAX_NODES, o2hb_live_node_bitmap); 1413 1414 o2hb_debug_create(O2HB_DEBUG_LIVEREGIONS, o2hb_debug_dir, 1415 &o2hb_db_liveregions, sizeof(*o2hb_db_liveregions), 1416 O2HB_DB_TYPE_LIVEREGIONS, 1417 sizeof(o2hb_live_region_bitmap), O2NM_MAX_REGIONS, 1418 o2hb_live_region_bitmap); 1419 1420 o2hb_debug_create(O2HB_DEBUG_QUORUMREGIONS, o2hb_debug_dir, 1421 &o2hb_db_quorumregions, 1422 sizeof(*o2hb_db_quorumregions), 1423 O2HB_DB_TYPE_QUORUMREGIONS, 1424 sizeof(o2hb_quorum_region_bitmap), O2NM_MAX_REGIONS, 1425 o2hb_quorum_region_bitmap); 1426 1427 o2hb_debug_create(O2HB_DEBUG_FAILEDREGIONS, o2hb_debug_dir, 1428 &o2hb_db_failedregions, 1429 sizeof(*o2hb_db_failedregions), 1430 O2HB_DB_TYPE_FAILEDREGIONS, 1431 sizeof(o2hb_failed_region_bitmap), O2NM_MAX_REGIONS, 1432 o2hb_failed_region_bitmap); 1433 } 1434 1435 void o2hb_init(void) 1436 { 1437 int i; 1438 1439 for (i = 0; i < ARRAY_SIZE(o2hb_callbacks); i++) 1440 INIT_LIST_HEAD(&o2hb_callbacks[i].list); 1441 1442 for (i = 0; i < ARRAY_SIZE(o2hb_live_slots); i++) 1443 INIT_LIST_HEAD(&o2hb_live_slots[i]); 1444 1445 memset(o2hb_live_node_bitmap, 0, sizeof(o2hb_live_node_bitmap)); 1446 memset(o2hb_region_bitmap, 0, sizeof(o2hb_region_bitmap)); 1447 memset(o2hb_live_region_bitmap, 0, sizeof(o2hb_live_region_bitmap)); 1448 memset(o2hb_quorum_region_bitmap, 0, sizeof(o2hb_quorum_region_bitmap)); 1449 memset(o2hb_failed_region_bitmap, 0, sizeof(o2hb_failed_region_bitmap)); 1450 1451 o2hb_dependent_users = 0; 1452 1453 o2hb_debug_init(); 1454 } 1455 1456 /* if we're already in a callback then we're already serialized by the sem */ 1457 static void o2hb_fill_node_map_from_callback(unsigned long *map, 1458 unsigned bytes) 1459 { 1460 BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long))); 1461 1462 memcpy(map, &o2hb_live_node_bitmap, bytes); 1463 } 1464 1465 /* 1466 * get a map of all nodes that are heartbeating in any regions 1467 */ 1468 void o2hb_fill_node_map(unsigned long *map, unsigned bytes) 1469 { 1470 /* callers want to serialize this map and callbacks so that they 1471 * can trust that they don't miss nodes coming to the party */ 1472 down_read(&o2hb_callback_sem); 1473 spin_lock(&o2hb_live_lock); 1474 o2hb_fill_node_map_from_callback(map, bytes); 1475 spin_unlock(&o2hb_live_lock); 1476 up_read(&o2hb_callback_sem); 1477 } 1478 EXPORT_SYMBOL_GPL(o2hb_fill_node_map); 1479 1480 /* 1481 * heartbeat configfs bits. The heartbeat set is a default set under 1482 * the cluster set in nodemanager.c. 1483 */ 1484 1485 static struct o2hb_region *to_o2hb_region(struct config_item *item) 1486 { 1487 return item ? container_of(item, struct o2hb_region, hr_item) : NULL; 1488 } 1489 1490 /* drop_item only drops its ref after killing the thread, nothing should 1491 * be using the region anymore. this has to clean up any state that 1492 * attributes might have built up. */ 1493 static void o2hb_region_release(struct config_item *item) 1494 { 1495 int i; 1496 struct page *page; 1497 struct o2hb_region *reg = to_o2hb_region(item); 1498 1499 mlog(ML_HEARTBEAT, "hb region release (%s)\n", reg->hr_dev_name); 1500 1501 kfree(reg->hr_tmp_block); 1502 1503 if (reg->hr_slot_data) { 1504 for (i = 0; i < reg->hr_num_pages; i++) { 1505 page = reg->hr_slot_data[i]; 1506 if (page) 1507 __free_page(page); 1508 } 1509 kfree(reg->hr_slot_data); 1510 } 1511 1512 if (reg->hr_bdev) 1513 blkdev_put(reg->hr_bdev, FMODE_READ|FMODE_WRITE); 1514 1515 kfree(reg->hr_slots); 1516 1517 debugfs_remove_recursive(reg->hr_debug_dir); 1518 kfree(reg->hr_db_livenodes); 1519 kfree(reg->hr_db_regnum); 1520 kfree(reg->hr_db_elapsed_time); 1521 kfree(reg->hr_db_pinned); 1522 1523 spin_lock(&o2hb_live_lock); 1524 list_del(®->hr_all_item); 1525 spin_unlock(&o2hb_live_lock); 1526 1527 o2net_unregister_handler_list(®->hr_handler_list); 1528 kfree(reg); 1529 } 1530 1531 static int o2hb_read_block_input(struct o2hb_region *reg, 1532 const char *page, 1533 unsigned long *ret_bytes, 1534 unsigned int *ret_bits) 1535 { 1536 unsigned long bytes; 1537 char *p = (char *)page; 1538 1539 bytes = simple_strtoul(p, &p, 0); 1540 if (!p || (*p && (*p != '\n'))) 1541 return -EINVAL; 1542 1543 /* Heartbeat and fs min / max block sizes are the same. */ 1544 if (bytes > 4096 || bytes < 512) 1545 return -ERANGE; 1546 if (hweight16(bytes) != 1) 1547 return -EINVAL; 1548 1549 if (ret_bytes) 1550 *ret_bytes = bytes; 1551 if (ret_bits) 1552 *ret_bits = ffs(bytes) - 1; 1553 1554 return 0; 1555 } 1556 1557 static ssize_t o2hb_region_block_bytes_show(struct config_item *item, 1558 char *page) 1559 { 1560 return sprintf(page, "%u\n", to_o2hb_region(item)->hr_block_bytes); 1561 } 1562 1563 static ssize_t o2hb_region_block_bytes_store(struct config_item *item, 1564 const char *page, 1565 size_t count) 1566 { 1567 struct o2hb_region *reg = to_o2hb_region(item); 1568 int status; 1569 unsigned long block_bytes; 1570 unsigned int block_bits; 1571 1572 if (reg->hr_bdev) 1573 return -EINVAL; 1574 1575 status = o2hb_read_block_input(reg, page, &block_bytes, 1576 &block_bits); 1577 if (status) 1578 return status; 1579 1580 reg->hr_block_bytes = (unsigned int)block_bytes; 1581 reg->hr_block_bits = block_bits; 1582 1583 return count; 1584 } 1585 1586 static ssize_t o2hb_region_start_block_show(struct config_item *item, 1587 char *page) 1588 { 1589 return sprintf(page, "%llu\n", to_o2hb_region(item)->hr_start_block); 1590 } 1591 1592 static ssize_t o2hb_region_start_block_store(struct config_item *item, 1593 const char *page, 1594 size_t count) 1595 { 1596 struct o2hb_region *reg = to_o2hb_region(item); 1597 unsigned long long tmp; 1598 char *p = (char *)page; 1599 ssize_t ret; 1600 1601 if (reg->hr_bdev) 1602 return -EINVAL; 1603 1604 ret = kstrtoull(p, 0, &tmp); 1605 if (ret) 1606 return -EINVAL; 1607 1608 reg->hr_start_block = tmp; 1609 1610 return count; 1611 } 1612 1613 static ssize_t o2hb_region_blocks_show(struct config_item *item, char *page) 1614 { 1615 return sprintf(page, "%d\n", to_o2hb_region(item)->hr_blocks); 1616 } 1617 1618 static ssize_t o2hb_region_blocks_store(struct config_item *item, 1619 const char *page, 1620 size_t count) 1621 { 1622 struct o2hb_region *reg = to_o2hb_region(item); 1623 unsigned long tmp; 1624 char *p = (char *)page; 1625 1626 if (reg->hr_bdev) 1627 return -EINVAL; 1628 1629 tmp = simple_strtoul(p, &p, 0); 1630 if (!p || (*p && (*p != '\n'))) 1631 return -EINVAL; 1632 1633 if (tmp > O2NM_MAX_NODES || tmp == 0) 1634 return -ERANGE; 1635 1636 reg->hr_blocks = (unsigned int)tmp; 1637 1638 return count; 1639 } 1640 1641 static ssize_t o2hb_region_dev_show(struct config_item *item, char *page) 1642 { 1643 unsigned int ret = 0; 1644 1645 if (to_o2hb_region(item)->hr_bdev) 1646 ret = sprintf(page, "%s\n", to_o2hb_region(item)->hr_dev_name); 1647 1648 return ret; 1649 } 1650 1651 static void o2hb_init_region_params(struct o2hb_region *reg) 1652 { 1653 reg->hr_slots_per_page = PAGE_SIZE >> reg->hr_block_bits; 1654 reg->hr_timeout_ms = O2HB_REGION_TIMEOUT_MS; 1655 1656 mlog(ML_HEARTBEAT, "hr_start_block = %llu, hr_blocks = %u\n", 1657 reg->hr_start_block, reg->hr_blocks); 1658 mlog(ML_HEARTBEAT, "hr_block_bytes = %u, hr_block_bits = %u\n", 1659 reg->hr_block_bytes, reg->hr_block_bits); 1660 mlog(ML_HEARTBEAT, "hr_timeout_ms = %u\n", reg->hr_timeout_ms); 1661 mlog(ML_HEARTBEAT, "dead threshold = %u\n", o2hb_dead_threshold); 1662 } 1663 1664 static int o2hb_map_slot_data(struct o2hb_region *reg) 1665 { 1666 int i, j; 1667 unsigned int last_slot; 1668 unsigned int spp = reg->hr_slots_per_page; 1669 struct page *page; 1670 char *raw; 1671 struct o2hb_disk_slot *slot; 1672 1673 reg->hr_tmp_block = kmalloc(reg->hr_block_bytes, GFP_KERNEL); 1674 if (reg->hr_tmp_block == NULL) 1675 return -ENOMEM; 1676 1677 reg->hr_slots = kcalloc(reg->hr_blocks, 1678 sizeof(struct o2hb_disk_slot), GFP_KERNEL); 1679 if (reg->hr_slots == NULL) 1680 return -ENOMEM; 1681 1682 for(i = 0; i < reg->hr_blocks; i++) { 1683 slot = ®->hr_slots[i]; 1684 slot->ds_node_num = i; 1685 INIT_LIST_HEAD(&slot->ds_live_item); 1686 slot->ds_raw_block = NULL; 1687 } 1688 1689 reg->hr_num_pages = (reg->hr_blocks + spp - 1) / spp; 1690 mlog(ML_HEARTBEAT, "Going to require %u pages to cover %u blocks " 1691 "at %u blocks per page\n", 1692 reg->hr_num_pages, reg->hr_blocks, spp); 1693 1694 reg->hr_slot_data = kcalloc(reg->hr_num_pages, sizeof(struct page *), 1695 GFP_KERNEL); 1696 if (!reg->hr_slot_data) 1697 return -ENOMEM; 1698 1699 for(i = 0; i < reg->hr_num_pages; i++) { 1700 page = alloc_page(GFP_KERNEL); 1701 if (!page) 1702 return -ENOMEM; 1703 1704 reg->hr_slot_data[i] = page; 1705 1706 last_slot = i * spp; 1707 raw = page_address(page); 1708 for (j = 0; 1709 (j < spp) && ((j + last_slot) < reg->hr_blocks); 1710 j++) { 1711 BUG_ON((j + last_slot) >= reg->hr_blocks); 1712 1713 slot = ®->hr_slots[j + last_slot]; 1714 slot->ds_raw_block = 1715 (struct o2hb_disk_heartbeat_block *) raw; 1716 1717 raw += reg->hr_block_bytes; 1718 } 1719 } 1720 1721 return 0; 1722 } 1723 1724 /* Read in all the slots available and populate the tracking 1725 * structures so that we can start with a baseline idea of what's 1726 * there. */ 1727 static int o2hb_populate_slot_data(struct o2hb_region *reg) 1728 { 1729 int ret, i; 1730 struct o2hb_disk_slot *slot; 1731 struct o2hb_disk_heartbeat_block *hb_block; 1732 1733 ret = o2hb_read_slots(reg, 0, reg->hr_blocks); 1734 if (ret) 1735 goto out; 1736 1737 /* We only want to get an idea of the values initially in each 1738 * slot, so we do no verification - o2hb_check_slot will 1739 * actually determine if each configured slot is valid and 1740 * whether any values have changed. */ 1741 for(i = 0; i < reg->hr_blocks; i++) { 1742 slot = ®->hr_slots[i]; 1743 hb_block = (struct o2hb_disk_heartbeat_block *) slot->ds_raw_block; 1744 1745 /* Only fill the values that o2hb_check_slot uses to 1746 * determine changing slots */ 1747 slot->ds_last_time = le64_to_cpu(hb_block->hb_seq); 1748 slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation); 1749 } 1750 1751 out: 1752 return ret; 1753 } 1754 1755 /* this is acting as commit; we set up all of hr_bdev and hr_task or nothing */ 1756 static ssize_t o2hb_region_dev_store(struct config_item *item, 1757 const char *page, 1758 size_t count) 1759 { 1760 struct o2hb_region *reg = to_o2hb_region(item); 1761 struct task_struct *hb_task; 1762 long fd; 1763 int sectsize; 1764 char *p = (char *)page; 1765 struct fd f; 1766 ssize_t ret = -EINVAL; 1767 int live_threshold; 1768 1769 if (reg->hr_bdev) 1770 goto out; 1771 1772 /* We can't heartbeat without having had our node number 1773 * configured yet. */ 1774 if (o2nm_this_node() == O2NM_MAX_NODES) 1775 goto out; 1776 1777 fd = simple_strtol(p, &p, 0); 1778 if (!p || (*p && (*p != '\n'))) 1779 goto out; 1780 1781 if (fd < 0 || fd >= INT_MAX) 1782 goto out; 1783 1784 f = fdget(fd); 1785 if (f.file == NULL) 1786 goto out; 1787 1788 if (reg->hr_blocks == 0 || reg->hr_start_block == 0 || 1789 reg->hr_block_bytes == 0) 1790 goto out2; 1791 1792 if (!S_ISBLK(f.file->f_mapping->host->i_mode)) 1793 goto out2; 1794 1795 reg->hr_bdev = blkdev_get_by_dev(f.file->f_mapping->host->i_rdev, 1796 FMODE_WRITE | FMODE_READ, NULL); 1797 if (IS_ERR(reg->hr_bdev)) { 1798 ret = PTR_ERR(reg->hr_bdev); 1799 reg->hr_bdev = NULL; 1800 goto out2; 1801 } 1802 1803 bdevname(reg->hr_bdev, reg->hr_dev_name); 1804 1805 sectsize = bdev_logical_block_size(reg->hr_bdev); 1806 if (sectsize != reg->hr_block_bytes) { 1807 mlog(ML_ERROR, 1808 "blocksize %u incorrect for device, expected %d", 1809 reg->hr_block_bytes, sectsize); 1810 ret = -EINVAL; 1811 goto out3; 1812 } 1813 1814 o2hb_init_region_params(reg); 1815 1816 /* Generation of zero is invalid */ 1817 do { 1818 get_random_bytes(®->hr_generation, 1819 sizeof(reg->hr_generation)); 1820 } while (reg->hr_generation == 0); 1821 1822 ret = o2hb_map_slot_data(reg); 1823 if (ret) { 1824 mlog_errno(ret); 1825 goto out3; 1826 } 1827 1828 ret = o2hb_populate_slot_data(reg); 1829 if (ret) { 1830 mlog_errno(ret); 1831 goto out3; 1832 } 1833 1834 INIT_DELAYED_WORK(®->hr_write_timeout_work, o2hb_write_timeout); 1835 INIT_DELAYED_WORK(®->hr_nego_timeout_work, o2hb_nego_timeout); 1836 1837 /* 1838 * A node is considered live after it has beat LIVE_THRESHOLD 1839 * times. We're not steady until we've given them a chance 1840 * _after_ our first read. 1841 * The default threshold is bare minimum so as to limit the delay 1842 * during mounts. For global heartbeat, the threshold doubled for the 1843 * first region. 1844 */ 1845 live_threshold = O2HB_LIVE_THRESHOLD; 1846 if (o2hb_global_heartbeat_active()) { 1847 spin_lock(&o2hb_live_lock); 1848 if (bitmap_weight(o2hb_region_bitmap, O2NM_MAX_REGIONS) == 1) 1849 live_threshold <<= 1; 1850 spin_unlock(&o2hb_live_lock); 1851 } 1852 ++live_threshold; 1853 atomic_set(®->hr_steady_iterations, live_threshold); 1854 /* unsteady_iterations is triple the steady_iterations */ 1855 atomic_set(®->hr_unsteady_iterations, (live_threshold * 3)); 1856 1857 hb_task = kthread_run(o2hb_thread, reg, "o2hb-%s", 1858 reg->hr_item.ci_name); 1859 if (IS_ERR(hb_task)) { 1860 ret = PTR_ERR(hb_task); 1861 mlog_errno(ret); 1862 goto out3; 1863 } 1864 1865 spin_lock(&o2hb_live_lock); 1866 reg->hr_task = hb_task; 1867 spin_unlock(&o2hb_live_lock); 1868 1869 ret = wait_event_interruptible(o2hb_steady_queue, 1870 atomic_read(®->hr_steady_iterations) == 0 || 1871 reg->hr_node_deleted); 1872 if (ret) { 1873 atomic_set(®->hr_steady_iterations, 0); 1874 reg->hr_aborted_start = 1; 1875 } 1876 1877 if (reg->hr_aborted_start) { 1878 ret = -EIO; 1879 goto out3; 1880 } 1881 1882 if (reg->hr_node_deleted) { 1883 ret = -EINVAL; 1884 goto out3; 1885 } 1886 1887 /* Ok, we were woken. Make sure it wasn't by drop_item() */ 1888 spin_lock(&o2hb_live_lock); 1889 hb_task = reg->hr_task; 1890 if (o2hb_global_heartbeat_active()) 1891 set_bit(reg->hr_region_num, o2hb_live_region_bitmap); 1892 spin_unlock(&o2hb_live_lock); 1893 1894 if (hb_task) 1895 ret = count; 1896 else 1897 ret = -EIO; 1898 1899 if (hb_task && o2hb_global_heartbeat_active()) 1900 printk(KERN_NOTICE "o2hb: Heartbeat started on region %s (%s)\n", 1901 config_item_name(®->hr_item), reg->hr_dev_name); 1902 1903 out3: 1904 if (ret < 0) { 1905 blkdev_put(reg->hr_bdev, FMODE_READ | FMODE_WRITE); 1906 reg->hr_bdev = NULL; 1907 } 1908 out2: 1909 fdput(f); 1910 out: 1911 return ret; 1912 } 1913 1914 static ssize_t o2hb_region_pid_show(struct config_item *item, char *page) 1915 { 1916 struct o2hb_region *reg = to_o2hb_region(item); 1917 pid_t pid = 0; 1918 1919 spin_lock(&o2hb_live_lock); 1920 if (reg->hr_task) 1921 pid = task_pid_nr(reg->hr_task); 1922 spin_unlock(&o2hb_live_lock); 1923 1924 if (!pid) 1925 return 0; 1926 1927 return sprintf(page, "%u\n", pid); 1928 } 1929 1930 CONFIGFS_ATTR(o2hb_region_, block_bytes); 1931 CONFIGFS_ATTR(o2hb_region_, start_block); 1932 CONFIGFS_ATTR(o2hb_region_, blocks); 1933 CONFIGFS_ATTR(o2hb_region_, dev); 1934 CONFIGFS_ATTR_RO(o2hb_region_, pid); 1935 1936 static struct configfs_attribute *o2hb_region_attrs[] = { 1937 &o2hb_region_attr_block_bytes, 1938 &o2hb_region_attr_start_block, 1939 &o2hb_region_attr_blocks, 1940 &o2hb_region_attr_dev, 1941 &o2hb_region_attr_pid, 1942 NULL, 1943 }; 1944 1945 static struct configfs_item_operations o2hb_region_item_ops = { 1946 .release = o2hb_region_release, 1947 }; 1948 1949 static const struct config_item_type o2hb_region_type = { 1950 .ct_item_ops = &o2hb_region_item_ops, 1951 .ct_attrs = o2hb_region_attrs, 1952 .ct_owner = THIS_MODULE, 1953 }; 1954 1955 /* heartbeat set */ 1956 1957 struct o2hb_heartbeat_group { 1958 struct config_group hs_group; 1959 /* some stuff? */ 1960 }; 1961 1962 static struct o2hb_heartbeat_group *to_o2hb_heartbeat_group(struct config_group *group) 1963 { 1964 return group ? 1965 container_of(group, struct o2hb_heartbeat_group, hs_group) 1966 : NULL; 1967 } 1968 1969 static void o2hb_debug_region_init(struct o2hb_region *reg, 1970 struct dentry *parent) 1971 { 1972 struct dentry *dir; 1973 1974 dir = debugfs_create_dir(config_item_name(®->hr_item), parent); 1975 reg->hr_debug_dir = dir; 1976 1977 o2hb_debug_create(O2HB_DEBUG_LIVENODES, dir, &(reg->hr_db_livenodes), 1978 sizeof(*(reg->hr_db_livenodes)), 1979 O2HB_DB_TYPE_REGION_LIVENODES, 1980 sizeof(reg->hr_live_node_bitmap), O2NM_MAX_NODES, 1981 reg); 1982 1983 o2hb_debug_create(O2HB_DEBUG_REGION_NUMBER, dir, &(reg->hr_db_regnum), 1984 sizeof(*(reg->hr_db_regnum)), 1985 O2HB_DB_TYPE_REGION_NUMBER, 0, O2NM_MAX_NODES, reg); 1986 1987 o2hb_debug_create(O2HB_DEBUG_REGION_ELAPSED_TIME, dir, 1988 &(reg->hr_db_elapsed_time), 1989 sizeof(*(reg->hr_db_elapsed_time)), 1990 O2HB_DB_TYPE_REGION_ELAPSED_TIME, 0, 0, reg); 1991 1992 o2hb_debug_create(O2HB_DEBUG_REGION_PINNED, dir, &(reg->hr_db_pinned), 1993 sizeof(*(reg->hr_db_pinned)), 1994 O2HB_DB_TYPE_REGION_PINNED, 0, 0, reg); 1995 1996 } 1997 1998 static struct config_item *o2hb_heartbeat_group_make_item(struct config_group *group, 1999 const char *name) 2000 { 2001 struct o2hb_region *reg = NULL; 2002 int ret; 2003 2004 reg = kzalloc(sizeof(struct o2hb_region), GFP_KERNEL); 2005 if (reg == NULL) 2006 return ERR_PTR(-ENOMEM); 2007 2008 if (strlen(name) > O2HB_MAX_REGION_NAME_LEN) { 2009 ret = -ENAMETOOLONG; 2010 goto free; 2011 } 2012 2013 spin_lock(&o2hb_live_lock); 2014 reg->hr_region_num = 0; 2015 if (o2hb_global_heartbeat_active()) { 2016 reg->hr_region_num = find_first_zero_bit(o2hb_region_bitmap, 2017 O2NM_MAX_REGIONS); 2018 if (reg->hr_region_num >= O2NM_MAX_REGIONS) { 2019 spin_unlock(&o2hb_live_lock); 2020 ret = -EFBIG; 2021 goto free; 2022 } 2023 set_bit(reg->hr_region_num, o2hb_region_bitmap); 2024 } 2025 list_add_tail(®->hr_all_item, &o2hb_all_regions); 2026 spin_unlock(&o2hb_live_lock); 2027 2028 config_item_init_type_name(®->hr_item, name, &o2hb_region_type); 2029 2030 /* this is the same way to generate msg key as dlm, for local heartbeat, 2031 * name is also the same, so make initial crc value different to avoid 2032 * message key conflict. 2033 */ 2034 reg->hr_key = crc32_le(reg->hr_region_num + O2NM_MAX_REGIONS, 2035 name, strlen(name)); 2036 INIT_LIST_HEAD(®->hr_handler_list); 2037 ret = o2net_register_handler(O2HB_NEGO_TIMEOUT_MSG, reg->hr_key, 2038 sizeof(struct o2hb_nego_msg), 2039 o2hb_nego_timeout_handler, 2040 reg, NULL, ®->hr_handler_list); 2041 if (ret) 2042 goto remove_item; 2043 2044 ret = o2net_register_handler(O2HB_NEGO_APPROVE_MSG, reg->hr_key, 2045 sizeof(struct o2hb_nego_msg), 2046 o2hb_nego_approve_handler, 2047 reg, NULL, ®->hr_handler_list); 2048 if (ret) 2049 goto unregister_handler; 2050 2051 o2hb_debug_region_init(reg, o2hb_debug_dir); 2052 2053 return ®->hr_item; 2054 2055 unregister_handler: 2056 o2net_unregister_handler_list(®->hr_handler_list); 2057 remove_item: 2058 spin_lock(&o2hb_live_lock); 2059 list_del(®->hr_all_item); 2060 if (o2hb_global_heartbeat_active()) 2061 clear_bit(reg->hr_region_num, o2hb_region_bitmap); 2062 spin_unlock(&o2hb_live_lock); 2063 free: 2064 kfree(reg); 2065 return ERR_PTR(ret); 2066 } 2067 2068 static void o2hb_heartbeat_group_drop_item(struct config_group *group, 2069 struct config_item *item) 2070 { 2071 struct task_struct *hb_task; 2072 struct o2hb_region *reg = to_o2hb_region(item); 2073 int quorum_region = 0; 2074 2075 /* stop the thread when the user removes the region dir */ 2076 spin_lock(&o2hb_live_lock); 2077 hb_task = reg->hr_task; 2078 reg->hr_task = NULL; 2079 reg->hr_item_dropped = 1; 2080 spin_unlock(&o2hb_live_lock); 2081 2082 if (hb_task) 2083 kthread_stop(hb_task); 2084 2085 if (o2hb_global_heartbeat_active()) { 2086 spin_lock(&o2hb_live_lock); 2087 clear_bit(reg->hr_region_num, o2hb_region_bitmap); 2088 clear_bit(reg->hr_region_num, o2hb_live_region_bitmap); 2089 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap)) 2090 quorum_region = 1; 2091 clear_bit(reg->hr_region_num, o2hb_quorum_region_bitmap); 2092 spin_unlock(&o2hb_live_lock); 2093 printk(KERN_NOTICE "o2hb: Heartbeat %s on region %s (%s)\n", 2094 ((atomic_read(®->hr_steady_iterations) == 0) ? 2095 "stopped" : "start aborted"), config_item_name(item), 2096 reg->hr_dev_name); 2097 } 2098 2099 /* 2100 * If we're racing a dev_write(), we need to wake them. They will 2101 * check reg->hr_task 2102 */ 2103 if (atomic_read(®->hr_steady_iterations) != 0) { 2104 reg->hr_aborted_start = 1; 2105 atomic_set(®->hr_steady_iterations, 0); 2106 wake_up(&o2hb_steady_queue); 2107 } 2108 2109 config_item_put(item); 2110 2111 if (!o2hb_global_heartbeat_active() || !quorum_region) 2112 return; 2113 2114 /* 2115 * If global heartbeat active and there are dependent users, 2116 * pin all regions if quorum region count <= CUT_OFF 2117 */ 2118 spin_lock(&o2hb_live_lock); 2119 2120 if (!o2hb_dependent_users) 2121 goto unlock; 2122 2123 if (bitmap_weight(o2hb_quorum_region_bitmap, 2124 O2NM_MAX_REGIONS) <= O2HB_PIN_CUT_OFF) 2125 o2hb_region_pin(NULL); 2126 2127 unlock: 2128 spin_unlock(&o2hb_live_lock); 2129 } 2130 2131 static ssize_t o2hb_heartbeat_group_dead_threshold_show(struct config_item *item, 2132 char *page) 2133 { 2134 return sprintf(page, "%u\n", o2hb_dead_threshold); 2135 } 2136 2137 static ssize_t o2hb_heartbeat_group_dead_threshold_store(struct config_item *item, 2138 const char *page, size_t count) 2139 { 2140 unsigned long tmp; 2141 char *p = (char *)page; 2142 2143 tmp = simple_strtoul(p, &p, 10); 2144 if (!p || (*p && (*p != '\n'))) 2145 return -EINVAL; 2146 2147 /* this will validate ranges for us. */ 2148 o2hb_dead_threshold_set((unsigned int) tmp); 2149 2150 return count; 2151 } 2152 2153 static ssize_t o2hb_heartbeat_group_mode_show(struct config_item *item, 2154 char *page) 2155 { 2156 return sprintf(page, "%s\n", 2157 o2hb_heartbeat_mode_desc[o2hb_heartbeat_mode]); 2158 } 2159 2160 static ssize_t o2hb_heartbeat_group_mode_store(struct config_item *item, 2161 const char *page, size_t count) 2162 { 2163 unsigned int i; 2164 int ret; 2165 size_t len; 2166 2167 len = (page[count - 1] == '\n') ? count - 1 : count; 2168 if (!len) 2169 return -EINVAL; 2170 2171 for (i = 0; i < O2HB_HEARTBEAT_NUM_MODES; ++i) { 2172 if (strncasecmp(page, o2hb_heartbeat_mode_desc[i], len)) 2173 continue; 2174 2175 ret = o2hb_global_heartbeat_mode_set(i); 2176 if (!ret) 2177 printk(KERN_NOTICE "o2hb: Heartbeat mode set to %s\n", 2178 o2hb_heartbeat_mode_desc[i]); 2179 return count; 2180 } 2181 2182 return -EINVAL; 2183 2184 } 2185 2186 CONFIGFS_ATTR(o2hb_heartbeat_group_, dead_threshold); 2187 CONFIGFS_ATTR(o2hb_heartbeat_group_, mode); 2188 2189 static struct configfs_attribute *o2hb_heartbeat_group_attrs[] = { 2190 &o2hb_heartbeat_group_attr_dead_threshold, 2191 &o2hb_heartbeat_group_attr_mode, 2192 NULL, 2193 }; 2194 2195 static struct configfs_group_operations o2hb_heartbeat_group_group_ops = { 2196 .make_item = o2hb_heartbeat_group_make_item, 2197 .drop_item = o2hb_heartbeat_group_drop_item, 2198 }; 2199 2200 static const struct config_item_type o2hb_heartbeat_group_type = { 2201 .ct_group_ops = &o2hb_heartbeat_group_group_ops, 2202 .ct_attrs = o2hb_heartbeat_group_attrs, 2203 .ct_owner = THIS_MODULE, 2204 }; 2205 2206 /* this is just here to avoid touching group in heartbeat.h which the 2207 * entire damn world #includes */ 2208 struct config_group *o2hb_alloc_hb_set(void) 2209 { 2210 struct o2hb_heartbeat_group *hs = NULL; 2211 struct config_group *ret = NULL; 2212 2213 hs = kzalloc(sizeof(struct o2hb_heartbeat_group), GFP_KERNEL); 2214 if (hs == NULL) 2215 goto out; 2216 2217 config_group_init_type_name(&hs->hs_group, "heartbeat", 2218 &o2hb_heartbeat_group_type); 2219 2220 ret = &hs->hs_group; 2221 out: 2222 if (ret == NULL) 2223 kfree(hs); 2224 return ret; 2225 } 2226 2227 void o2hb_free_hb_set(struct config_group *group) 2228 { 2229 struct o2hb_heartbeat_group *hs = to_o2hb_heartbeat_group(group); 2230 kfree(hs); 2231 } 2232 2233 /* hb callback registration and issuing */ 2234 2235 static struct o2hb_callback *hbcall_from_type(enum o2hb_callback_type type) 2236 { 2237 if (type == O2HB_NUM_CB) 2238 return ERR_PTR(-EINVAL); 2239 2240 return &o2hb_callbacks[type]; 2241 } 2242 2243 void o2hb_setup_callback(struct o2hb_callback_func *hc, 2244 enum o2hb_callback_type type, 2245 o2hb_cb_func *func, 2246 void *data, 2247 int priority) 2248 { 2249 INIT_LIST_HEAD(&hc->hc_item); 2250 hc->hc_func = func; 2251 hc->hc_data = data; 2252 hc->hc_priority = priority; 2253 hc->hc_type = type; 2254 hc->hc_magic = O2HB_CB_MAGIC; 2255 } 2256 EXPORT_SYMBOL_GPL(o2hb_setup_callback); 2257 2258 /* 2259 * In local heartbeat mode, region_uuid passed matches the dlm domain name. 2260 * In global heartbeat mode, region_uuid passed is NULL. 2261 * 2262 * In local, we only pin the matching region. In global we pin all the active 2263 * regions. 2264 */ 2265 static int o2hb_region_pin(const char *region_uuid) 2266 { 2267 int ret = 0, found = 0; 2268 struct o2hb_region *reg; 2269 char *uuid; 2270 2271 assert_spin_locked(&o2hb_live_lock); 2272 2273 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) { 2274 if (reg->hr_item_dropped) 2275 continue; 2276 2277 uuid = config_item_name(®->hr_item); 2278 2279 /* local heartbeat */ 2280 if (region_uuid) { 2281 if (strcmp(region_uuid, uuid)) 2282 continue; 2283 found = 1; 2284 } 2285 2286 if (reg->hr_item_pinned || reg->hr_item_dropped) 2287 goto skip_pin; 2288 2289 /* Ignore ENOENT only for local hb (userdlm domain) */ 2290 ret = o2nm_depend_item(®->hr_item); 2291 if (!ret) { 2292 mlog(ML_CLUSTER, "Pin region %s\n", uuid); 2293 reg->hr_item_pinned = 1; 2294 } else { 2295 if (ret == -ENOENT && found) 2296 ret = 0; 2297 else { 2298 mlog(ML_ERROR, "Pin region %s fails with %d\n", 2299 uuid, ret); 2300 break; 2301 } 2302 } 2303 skip_pin: 2304 if (found) 2305 break; 2306 } 2307 2308 return ret; 2309 } 2310 2311 /* 2312 * In local heartbeat mode, region_uuid passed matches the dlm domain name. 2313 * In global heartbeat mode, region_uuid passed is NULL. 2314 * 2315 * In local, we only unpin the matching region. In global we unpin all the 2316 * active regions. 2317 */ 2318 static void o2hb_region_unpin(const char *region_uuid) 2319 { 2320 struct o2hb_region *reg; 2321 char *uuid; 2322 int found = 0; 2323 2324 assert_spin_locked(&o2hb_live_lock); 2325 2326 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) { 2327 if (reg->hr_item_dropped) 2328 continue; 2329 2330 uuid = config_item_name(®->hr_item); 2331 if (region_uuid) { 2332 if (strcmp(region_uuid, uuid)) 2333 continue; 2334 found = 1; 2335 } 2336 2337 if (reg->hr_item_pinned) { 2338 mlog(ML_CLUSTER, "Unpin region %s\n", uuid); 2339 o2nm_undepend_item(®->hr_item); 2340 reg->hr_item_pinned = 0; 2341 } 2342 if (found) 2343 break; 2344 } 2345 } 2346 2347 static int o2hb_region_inc_user(const char *region_uuid) 2348 { 2349 int ret = 0; 2350 2351 spin_lock(&o2hb_live_lock); 2352 2353 /* local heartbeat */ 2354 if (!o2hb_global_heartbeat_active()) { 2355 ret = o2hb_region_pin(region_uuid); 2356 goto unlock; 2357 } 2358 2359 /* 2360 * if global heartbeat active and this is the first dependent user, 2361 * pin all regions if quorum region count <= CUT_OFF 2362 */ 2363 o2hb_dependent_users++; 2364 if (o2hb_dependent_users > 1) 2365 goto unlock; 2366 2367 if (bitmap_weight(o2hb_quorum_region_bitmap, 2368 O2NM_MAX_REGIONS) <= O2HB_PIN_CUT_OFF) 2369 ret = o2hb_region_pin(NULL); 2370 2371 unlock: 2372 spin_unlock(&o2hb_live_lock); 2373 return ret; 2374 } 2375 2376 static void o2hb_region_dec_user(const char *region_uuid) 2377 { 2378 spin_lock(&o2hb_live_lock); 2379 2380 /* local heartbeat */ 2381 if (!o2hb_global_heartbeat_active()) { 2382 o2hb_region_unpin(region_uuid); 2383 goto unlock; 2384 } 2385 2386 /* 2387 * if global heartbeat active and there are no dependent users, 2388 * unpin all quorum regions 2389 */ 2390 o2hb_dependent_users--; 2391 if (!o2hb_dependent_users) 2392 o2hb_region_unpin(NULL); 2393 2394 unlock: 2395 spin_unlock(&o2hb_live_lock); 2396 } 2397 2398 int o2hb_register_callback(const char *region_uuid, 2399 struct o2hb_callback_func *hc) 2400 { 2401 struct o2hb_callback_func *f; 2402 struct o2hb_callback *hbcall; 2403 int ret; 2404 2405 BUG_ON(hc->hc_magic != O2HB_CB_MAGIC); 2406 BUG_ON(!list_empty(&hc->hc_item)); 2407 2408 hbcall = hbcall_from_type(hc->hc_type); 2409 if (IS_ERR(hbcall)) { 2410 ret = PTR_ERR(hbcall); 2411 goto out; 2412 } 2413 2414 if (region_uuid) { 2415 ret = o2hb_region_inc_user(region_uuid); 2416 if (ret) { 2417 mlog_errno(ret); 2418 goto out; 2419 } 2420 } 2421 2422 down_write(&o2hb_callback_sem); 2423 2424 list_for_each_entry(f, &hbcall->list, hc_item) { 2425 if (hc->hc_priority < f->hc_priority) { 2426 list_add_tail(&hc->hc_item, &f->hc_item); 2427 break; 2428 } 2429 } 2430 if (list_empty(&hc->hc_item)) 2431 list_add_tail(&hc->hc_item, &hbcall->list); 2432 2433 up_write(&o2hb_callback_sem); 2434 ret = 0; 2435 out: 2436 mlog(ML_CLUSTER, "returning %d on behalf of %p for funcs %p\n", 2437 ret, __builtin_return_address(0), hc); 2438 return ret; 2439 } 2440 EXPORT_SYMBOL_GPL(o2hb_register_callback); 2441 2442 void o2hb_unregister_callback(const char *region_uuid, 2443 struct o2hb_callback_func *hc) 2444 { 2445 BUG_ON(hc->hc_magic != O2HB_CB_MAGIC); 2446 2447 mlog(ML_CLUSTER, "on behalf of %p for funcs %p\n", 2448 __builtin_return_address(0), hc); 2449 2450 /* XXX Can this happen _with_ a region reference? */ 2451 if (list_empty(&hc->hc_item)) 2452 return; 2453 2454 if (region_uuid) 2455 o2hb_region_dec_user(region_uuid); 2456 2457 down_write(&o2hb_callback_sem); 2458 2459 list_del_init(&hc->hc_item); 2460 2461 up_write(&o2hb_callback_sem); 2462 } 2463 EXPORT_SYMBOL_GPL(o2hb_unregister_callback); 2464 2465 int o2hb_check_node_heartbeating_no_sem(u8 node_num) 2466 { 2467 unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; 2468 2469 spin_lock(&o2hb_live_lock); 2470 o2hb_fill_node_map_from_callback(testing_map, sizeof(testing_map)); 2471 spin_unlock(&o2hb_live_lock); 2472 if (!test_bit(node_num, testing_map)) { 2473 mlog(ML_HEARTBEAT, 2474 "node (%u) does not have heartbeating enabled.\n", 2475 node_num); 2476 return 0; 2477 } 2478 2479 return 1; 2480 } 2481 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating_no_sem); 2482 2483 int o2hb_check_node_heartbeating_from_callback(u8 node_num) 2484 { 2485 unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; 2486 2487 o2hb_fill_node_map_from_callback(testing_map, sizeof(testing_map)); 2488 if (!test_bit(node_num, testing_map)) { 2489 mlog(ML_HEARTBEAT, 2490 "node (%u) does not have heartbeating enabled.\n", 2491 node_num); 2492 return 0; 2493 } 2494 2495 return 1; 2496 } 2497 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating_from_callback); 2498 2499 /* 2500 * this is just a hack until we get the plumbing which flips file systems 2501 * read only and drops the hb ref instead of killing the node dead. 2502 */ 2503 void o2hb_stop_all_regions(void) 2504 { 2505 struct o2hb_region *reg; 2506 2507 mlog(ML_ERROR, "stopping heartbeat on all active regions.\n"); 2508 2509 spin_lock(&o2hb_live_lock); 2510 2511 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) 2512 reg->hr_unclean_stop = 1; 2513 2514 spin_unlock(&o2hb_live_lock); 2515 } 2516 EXPORT_SYMBOL_GPL(o2hb_stop_all_regions); 2517 2518 int o2hb_get_all_regions(char *region_uuids, u8 max_regions) 2519 { 2520 struct o2hb_region *reg; 2521 int numregs = 0; 2522 char *p; 2523 2524 spin_lock(&o2hb_live_lock); 2525 2526 p = region_uuids; 2527 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) { 2528 if (reg->hr_item_dropped) 2529 continue; 2530 2531 mlog(0, "Region: %s\n", config_item_name(®->hr_item)); 2532 if (numregs < max_regions) { 2533 memcpy(p, config_item_name(®->hr_item), 2534 O2HB_MAX_REGION_NAME_LEN); 2535 p += O2HB_MAX_REGION_NAME_LEN; 2536 } 2537 numregs++; 2538 } 2539 2540 spin_unlock(&o2hb_live_lock); 2541 2542 return numregs; 2543 } 2544 EXPORT_SYMBOL_GPL(o2hb_get_all_regions); 2545 2546 int o2hb_global_heartbeat_active(void) 2547 { 2548 return (o2hb_heartbeat_mode == O2HB_HEARTBEAT_GLOBAL); 2549 } 2550 EXPORT_SYMBOL(o2hb_global_heartbeat_active); 2551