1 /* -*- mode: c; c-basic-offset: 8; -*- 2 * vim: noexpandtab sw=8 ts=8 sts=0: 3 * 4 * Copyright (C) 2004, 2005 Oracle. All rights reserved. 5 * 6 * This program is free software; you can redistribute it and/or 7 * modify it under the terms of the GNU General Public 8 * License as published by the Free Software Foundation; either 9 * version 2 of the License, or (at your option) any later version. 10 * 11 * This program is distributed in the hope that it will be useful, 12 * but WITHOUT ANY WARRANTY; without even the implied warranty of 13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 14 * General Public License for more details. 15 * 16 * You should have received a copy of the GNU General Public 17 * License along with this program; if not, write to the 18 * Free Software Foundation, Inc., 59 Temple Place - Suite 330, 19 * Boston, MA 021110-1307, USA. 20 */ 21 22 #include <linux/kernel.h> 23 #include <linux/sched.h> 24 #include <linux/jiffies.h> 25 #include <linux/module.h> 26 #include <linux/fs.h> 27 #include <linux/bio.h> 28 #include <linux/blkdev.h> 29 #include <linux/delay.h> 30 #include <linux/file.h> 31 #include <linux/kthread.h> 32 #include <linux/configfs.h> 33 #include <linux/random.h> 34 #include <linux/crc32.h> 35 #include <linux/time.h> 36 #include <linux/debugfs.h> 37 #include <linux/slab.h> 38 #include <linux/bitmap.h> 39 #include <linux/ktime.h> 40 #include "heartbeat.h" 41 #include "tcp.h" 42 #include "nodemanager.h" 43 #include "quorum.h" 44 45 #include "masklog.h" 46 47 48 /* 49 * The first heartbeat pass had one global thread that would serialize all hb 50 * callback calls. This global serializing sem should only be removed once 51 * we've made sure that all callees can deal with being called concurrently 52 * from multiple hb region threads. 53 */ 54 static DECLARE_RWSEM(o2hb_callback_sem); 55 56 /* 57 * multiple hb threads are watching multiple regions. A node is live 58 * whenever any of the threads sees activity from the node in its region. 59 */ 60 static DEFINE_SPINLOCK(o2hb_live_lock); 61 static struct list_head o2hb_live_slots[O2NM_MAX_NODES]; 62 static unsigned long o2hb_live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)]; 63 static LIST_HEAD(o2hb_node_events); 64 static DECLARE_WAIT_QUEUE_HEAD(o2hb_steady_queue); 65 66 /* 67 * In global heartbeat, we maintain a series of region bitmaps. 68 * - o2hb_region_bitmap allows us to limit the region number to max region. 69 * - o2hb_live_region_bitmap tracks live regions (seen steady iterations). 70 * - o2hb_quorum_region_bitmap tracks live regions that have seen all nodes 71 * heartbeat on it. 72 * - o2hb_failed_region_bitmap tracks the regions that have seen io timeouts. 73 */ 74 static unsigned long o2hb_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)]; 75 static unsigned long o2hb_live_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)]; 76 static unsigned long o2hb_quorum_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)]; 77 static unsigned long o2hb_failed_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)]; 78 79 #define O2HB_DB_TYPE_LIVENODES 0 80 #define O2HB_DB_TYPE_LIVEREGIONS 1 81 #define O2HB_DB_TYPE_QUORUMREGIONS 2 82 #define O2HB_DB_TYPE_FAILEDREGIONS 3 83 #define O2HB_DB_TYPE_REGION_LIVENODES 4 84 #define O2HB_DB_TYPE_REGION_NUMBER 5 85 #define O2HB_DB_TYPE_REGION_ELAPSED_TIME 6 86 #define O2HB_DB_TYPE_REGION_PINNED 7 87 struct o2hb_debug_buf { 88 int db_type; 89 int db_size; 90 int db_len; 91 void *db_data; 92 }; 93 94 static struct o2hb_debug_buf *o2hb_db_livenodes; 95 static struct o2hb_debug_buf *o2hb_db_liveregions; 96 static struct o2hb_debug_buf *o2hb_db_quorumregions; 97 static struct o2hb_debug_buf *o2hb_db_failedregions; 98 99 #define O2HB_DEBUG_DIR "o2hb" 100 #define O2HB_DEBUG_LIVENODES "livenodes" 101 #define O2HB_DEBUG_LIVEREGIONS "live_regions" 102 #define O2HB_DEBUG_QUORUMREGIONS "quorum_regions" 103 #define O2HB_DEBUG_FAILEDREGIONS "failed_regions" 104 #define O2HB_DEBUG_REGION_NUMBER "num" 105 #define O2HB_DEBUG_REGION_ELAPSED_TIME "elapsed_time_in_ms" 106 #define O2HB_DEBUG_REGION_PINNED "pinned" 107 108 static struct dentry *o2hb_debug_dir; 109 static struct dentry *o2hb_debug_livenodes; 110 static struct dentry *o2hb_debug_liveregions; 111 static struct dentry *o2hb_debug_quorumregions; 112 static struct dentry *o2hb_debug_failedregions; 113 114 static LIST_HEAD(o2hb_all_regions); 115 116 static struct o2hb_callback { 117 struct list_head list; 118 } o2hb_callbacks[O2HB_NUM_CB]; 119 120 static struct o2hb_callback *hbcall_from_type(enum o2hb_callback_type type); 121 122 #define O2HB_DEFAULT_BLOCK_BITS 9 123 124 enum o2hb_heartbeat_modes { 125 O2HB_HEARTBEAT_LOCAL = 0, 126 O2HB_HEARTBEAT_GLOBAL, 127 O2HB_HEARTBEAT_NUM_MODES, 128 }; 129 130 char *o2hb_heartbeat_mode_desc[O2HB_HEARTBEAT_NUM_MODES] = { 131 "local", /* O2HB_HEARTBEAT_LOCAL */ 132 "global", /* O2HB_HEARTBEAT_GLOBAL */ 133 }; 134 135 unsigned int o2hb_dead_threshold = O2HB_DEFAULT_DEAD_THRESHOLD; 136 unsigned int o2hb_heartbeat_mode = O2HB_HEARTBEAT_LOCAL; 137 138 /* 139 * o2hb_dependent_users tracks the number of registered callbacks that depend 140 * on heartbeat. o2net and o2dlm are two entities that register this callback. 141 * However only o2dlm depends on the heartbeat. It does not want the heartbeat 142 * to stop while a dlm domain is still active. 143 */ 144 unsigned int o2hb_dependent_users; 145 146 /* 147 * In global heartbeat mode, all regions are pinned if there are one or more 148 * dependent users and the quorum region count is <= O2HB_PIN_CUT_OFF. All 149 * regions are unpinned if the region count exceeds the cut off or the number 150 * of dependent users falls to zero. 151 */ 152 #define O2HB_PIN_CUT_OFF 3 153 154 /* 155 * In local heartbeat mode, we assume the dlm domain name to be the same as 156 * region uuid. This is true for domains created for the file system but not 157 * necessarily true for userdlm domains. This is a known limitation. 158 * 159 * In global heartbeat mode, we pin/unpin all o2hb regions. This solution 160 * works for both file system and userdlm domains. 161 */ 162 static int o2hb_region_pin(const char *region_uuid); 163 static void o2hb_region_unpin(const char *region_uuid); 164 165 /* Only sets a new threshold if there are no active regions. 166 * 167 * No locking or otherwise interesting code is required for reading 168 * o2hb_dead_threshold as it can't change once regions are active and 169 * it's not interesting to anyone until then anyway. */ 170 static void o2hb_dead_threshold_set(unsigned int threshold) 171 { 172 if (threshold > O2HB_MIN_DEAD_THRESHOLD) { 173 spin_lock(&o2hb_live_lock); 174 if (list_empty(&o2hb_all_regions)) 175 o2hb_dead_threshold = threshold; 176 spin_unlock(&o2hb_live_lock); 177 } 178 } 179 180 static int o2hb_global_heartbeat_mode_set(unsigned int hb_mode) 181 { 182 int ret = -1; 183 184 if (hb_mode < O2HB_HEARTBEAT_NUM_MODES) { 185 spin_lock(&o2hb_live_lock); 186 if (list_empty(&o2hb_all_regions)) { 187 o2hb_heartbeat_mode = hb_mode; 188 ret = 0; 189 } 190 spin_unlock(&o2hb_live_lock); 191 } 192 193 return ret; 194 } 195 196 struct o2hb_node_event { 197 struct list_head hn_item; 198 enum o2hb_callback_type hn_event_type; 199 struct o2nm_node *hn_node; 200 int hn_node_num; 201 }; 202 203 struct o2hb_disk_slot { 204 struct o2hb_disk_heartbeat_block *ds_raw_block; 205 u8 ds_node_num; 206 u64 ds_last_time; 207 u64 ds_last_generation; 208 u16 ds_equal_samples; 209 u16 ds_changed_samples; 210 struct list_head ds_live_item; 211 }; 212 213 /* each thread owns a region.. when we're asked to tear down the region 214 * we ask the thread to stop, who cleans up the region */ 215 struct o2hb_region { 216 struct config_item hr_item; 217 218 struct list_head hr_all_item; 219 unsigned hr_unclean_stop:1, 220 hr_aborted_start:1, 221 hr_item_pinned:1, 222 hr_item_dropped:1; 223 224 /* protected by the hr_callback_sem */ 225 struct task_struct *hr_task; 226 227 unsigned int hr_blocks; 228 unsigned long long hr_start_block; 229 230 unsigned int hr_block_bits; 231 unsigned int hr_block_bytes; 232 233 unsigned int hr_slots_per_page; 234 unsigned int hr_num_pages; 235 236 struct page **hr_slot_data; 237 struct block_device *hr_bdev; 238 struct o2hb_disk_slot *hr_slots; 239 240 /* live node map of this region */ 241 unsigned long hr_live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)]; 242 unsigned int hr_region_num; 243 244 struct dentry *hr_debug_dir; 245 struct dentry *hr_debug_livenodes; 246 struct dentry *hr_debug_regnum; 247 struct dentry *hr_debug_elapsed_time; 248 struct dentry *hr_debug_pinned; 249 struct o2hb_debug_buf *hr_db_livenodes; 250 struct o2hb_debug_buf *hr_db_regnum; 251 struct o2hb_debug_buf *hr_db_elapsed_time; 252 struct o2hb_debug_buf *hr_db_pinned; 253 254 /* let the person setting up hb wait for it to return until it 255 * has reached a 'steady' state. This will be fixed when we have 256 * a more complete api that doesn't lead to this sort of fragility. */ 257 atomic_t hr_steady_iterations; 258 259 /* terminate o2hb thread if it does not reach steady state 260 * (hr_steady_iterations == 0) within hr_unsteady_iterations */ 261 atomic_t hr_unsteady_iterations; 262 263 char hr_dev_name[BDEVNAME_SIZE]; 264 265 unsigned int hr_timeout_ms; 266 267 /* randomized as the region goes up and down so that a node 268 * recognizes a node going up and down in one iteration */ 269 u64 hr_generation; 270 271 struct delayed_work hr_write_timeout_work; 272 unsigned long hr_last_timeout_start; 273 274 /* Used during o2hb_check_slot to hold a copy of the block 275 * being checked because we temporarily have to zero out the 276 * crc field. */ 277 struct o2hb_disk_heartbeat_block *hr_tmp_block; 278 }; 279 280 struct o2hb_bio_wait_ctxt { 281 atomic_t wc_num_reqs; 282 struct completion wc_io_complete; 283 int wc_error; 284 }; 285 286 static void o2hb_write_timeout(struct work_struct *work) 287 { 288 int failed, quorum; 289 unsigned long flags; 290 struct o2hb_region *reg = 291 container_of(work, struct o2hb_region, 292 hr_write_timeout_work.work); 293 294 mlog(ML_ERROR, "Heartbeat write timeout to device %s after %u " 295 "milliseconds\n", reg->hr_dev_name, 296 jiffies_to_msecs(jiffies - reg->hr_last_timeout_start)); 297 298 if (o2hb_global_heartbeat_active()) { 299 spin_lock_irqsave(&o2hb_live_lock, flags); 300 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap)) 301 set_bit(reg->hr_region_num, o2hb_failed_region_bitmap); 302 failed = bitmap_weight(o2hb_failed_region_bitmap, 303 O2NM_MAX_REGIONS); 304 quorum = bitmap_weight(o2hb_quorum_region_bitmap, 305 O2NM_MAX_REGIONS); 306 spin_unlock_irqrestore(&o2hb_live_lock, flags); 307 308 mlog(ML_HEARTBEAT, "Number of regions %d, failed regions %d\n", 309 quorum, failed); 310 311 /* 312 * Fence if the number of failed regions >= half the number 313 * of quorum regions 314 */ 315 if ((failed << 1) < quorum) 316 return; 317 } 318 319 o2quo_disk_timeout(); 320 } 321 322 static void o2hb_arm_write_timeout(struct o2hb_region *reg) 323 { 324 /* Arm writeout only after thread reaches steady state */ 325 if (atomic_read(®->hr_steady_iterations) != 0) 326 return; 327 328 mlog(ML_HEARTBEAT, "Queue write timeout for %u ms\n", 329 O2HB_MAX_WRITE_TIMEOUT_MS); 330 331 if (o2hb_global_heartbeat_active()) { 332 spin_lock(&o2hb_live_lock); 333 clear_bit(reg->hr_region_num, o2hb_failed_region_bitmap); 334 spin_unlock(&o2hb_live_lock); 335 } 336 cancel_delayed_work(®->hr_write_timeout_work); 337 reg->hr_last_timeout_start = jiffies; 338 schedule_delayed_work(®->hr_write_timeout_work, 339 msecs_to_jiffies(O2HB_MAX_WRITE_TIMEOUT_MS)); 340 } 341 342 static void o2hb_disarm_write_timeout(struct o2hb_region *reg) 343 { 344 cancel_delayed_work_sync(®->hr_write_timeout_work); 345 } 346 347 static inline void o2hb_bio_wait_init(struct o2hb_bio_wait_ctxt *wc) 348 { 349 atomic_set(&wc->wc_num_reqs, 1); 350 init_completion(&wc->wc_io_complete); 351 wc->wc_error = 0; 352 } 353 354 /* Used in error paths too */ 355 static inline void o2hb_bio_wait_dec(struct o2hb_bio_wait_ctxt *wc, 356 unsigned int num) 357 { 358 /* sadly atomic_sub_and_test() isn't available on all platforms. The 359 * good news is that the fast path only completes one at a time */ 360 while(num--) { 361 if (atomic_dec_and_test(&wc->wc_num_reqs)) { 362 BUG_ON(num > 0); 363 complete(&wc->wc_io_complete); 364 } 365 } 366 } 367 368 static void o2hb_wait_on_io(struct o2hb_region *reg, 369 struct o2hb_bio_wait_ctxt *wc) 370 { 371 o2hb_bio_wait_dec(wc, 1); 372 wait_for_completion(&wc->wc_io_complete); 373 } 374 375 static void o2hb_bio_end_io(struct bio *bio) 376 { 377 struct o2hb_bio_wait_ctxt *wc = bio->bi_private; 378 379 if (bio->bi_error) { 380 mlog(ML_ERROR, "IO Error %d\n", bio->bi_error); 381 wc->wc_error = bio->bi_error; 382 } 383 384 o2hb_bio_wait_dec(wc, 1); 385 bio_put(bio); 386 } 387 388 /* Setup a Bio to cover I/O against num_slots slots starting at 389 * start_slot. */ 390 static struct bio *o2hb_setup_one_bio(struct o2hb_region *reg, 391 struct o2hb_bio_wait_ctxt *wc, 392 unsigned int *current_slot, 393 unsigned int max_slots) 394 { 395 int len, current_page; 396 unsigned int vec_len, vec_start; 397 unsigned int bits = reg->hr_block_bits; 398 unsigned int spp = reg->hr_slots_per_page; 399 unsigned int cs = *current_slot; 400 struct bio *bio; 401 struct page *page; 402 403 /* Testing has shown this allocation to take long enough under 404 * GFP_KERNEL that the local node can get fenced. It would be 405 * nicest if we could pre-allocate these bios and avoid this 406 * all together. */ 407 bio = bio_alloc(GFP_ATOMIC, 16); 408 if (!bio) { 409 mlog(ML_ERROR, "Could not alloc slots BIO!\n"); 410 bio = ERR_PTR(-ENOMEM); 411 goto bail; 412 } 413 414 /* Must put everything in 512 byte sectors for the bio... */ 415 bio->bi_iter.bi_sector = (reg->hr_start_block + cs) << (bits - 9); 416 bio->bi_bdev = reg->hr_bdev; 417 bio->bi_private = wc; 418 bio->bi_end_io = o2hb_bio_end_io; 419 420 vec_start = (cs << bits) % PAGE_CACHE_SIZE; 421 while(cs < max_slots) { 422 current_page = cs / spp; 423 page = reg->hr_slot_data[current_page]; 424 425 vec_len = min(PAGE_CACHE_SIZE - vec_start, 426 (max_slots-cs) * (PAGE_CACHE_SIZE/spp) ); 427 428 mlog(ML_HB_BIO, "page %d, vec_len = %u, vec_start = %u\n", 429 current_page, vec_len, vec_start); 430 431 len = bio_add_page(bio, page, vec_len, vec_start); 432 if (len != vec_len) break; 433 434 cs += vec_len / (PAGE_CACHE_SIZE/spp); 435 vec_start = 0; 436 } 437 438 bail: 439 *current_slot = cs; 440 return bio; 441 } 442 443 static int o2hb_read_slots(struct o2hb_region *reg, 444 unsigned int max_slots) 445 { 446 unsigned int current_slot=0; 447 int status; 448 struct o2hb_bio_wait_ctxt wc; 449 struct bio *bio; 450 451 o2hb_bio_wait_init(&wc); 452 453 while(current_slot < max_slots) { 454 bio = o2hb_setup_one_bio(reg, &wc, ¤t_slot, max_slots); 455 if (IS_ERR(bio)) { 456 status = PTR_ERR(bio); 457 mlog_errno(status); 458 goto bail_and_wait; 459 } 460 461 atomic_inc(&wc.wc_num_reqs); 462 submit_bio(READ, bio); 463 } 464 465 status = 0; 466 467 bail_and_wait: 468 o2hb_wait_on_io(reg, &wc); 469 if (wc.wc_error && !status) 470 status = wc.wc_error; 471 472 return status; 473 } 474 475 static int o2hb_issue_node_write(struct o2hb_region *reg, 476 struct o2hb_bio_wait_ctxt *write_wc) 477 { 478 int status; 479 unsigned int slot; 480 struct bio *bio; 481 482 o2hb_bio_wait_init(write_wc); 483 484 slot = o2nm_this_node(); 485 486 bio = o2hb_setup_one_bio(reg, write_wc, &slot, slot+1); 487 if (IS_ERR(bio)) { 488 status = PTR_ERR(bio); 489 mlog_errno(status); 490 goto bail; 491 } 492 493 atomic_inc(&write_wc->wc_num_reqs); 494 submit_bio(WRITE_SYNC, bio); 495 496 status = 0; 497 bail: 498 return status; 499 } 500 501 static u32 o2hb_compute_block_crc_le(struct o2hb_region *reg, 502 struct o2hb_disk_heartbeat_block *hb_block) 503 { 504 __le32 old_cksum; 505 u32 ret; 506 507 /* We want to compute the block crc with a 0 value in the 508 * hb_cksum field. Save it off here and replace after the 509 * crc. */ 510 old_cksum = hb_block->hb_cksum; 511 hb_block->hb_cksum = 0; 512 513 ret = crc32_le(0, (unsigned char *) hb_block, reg->hr_block_bytes); 514 515 hb_block->hb_cksum = old_cksum; 516 517 return ret; 518 } 519 520 static void o2hb_dump_slot(struct o2hb_disk_heartbeat_block *hb_block) 521 { 522 mlog(ML_ERROR, "Dump slot information: seq = 0x%llx, node = %u, " 523 "cksum = 0x%x, generation 0x%llx\n", 524 (long long)le64_to_cpu(hb_block->hb_seq), 525 hb_block->hb_node, le32_to_cpu(hb_block->hb_cksum), 526 (long long)le64_to_cpu(hb_block->hb_generation)); 527 } 528 529 static int o2hb_verify_crc(struct o2hb_region *reg, 530 struct o2hb_disk_heartbeat_block *hb_block) 531 { 532 u32 read, computed; 533 534 read = le32_to_cpu(hb_block->hb_cksum); 535 computed = o2hb_compute_block_crc_le(reg, hb_block); 536 537 return read == computed; 538 } 539 540 /* 541 * Compare the slot data with what we wrote in the last iteration. 542 * If the match fails, print an appropriate error message. This is to 543 * detect errors like... another node hearting on the same slot, 544 * flaky device that is losing writes, etc. 545 * Returns 1 if check succeeds, 0 otherwise. 546 */ 547 static int o2hb_check_own_slot(struct o2hb_region *reg) 548 { 549 struct o2hb_disk_slot *slot; 550 struct o2hb_disk_heartbeat_block *hb_block; 551 char *errstr; 552 553 slot = ®->hr_slots[o2nm_this_node()]; 554 /* Don't check on our 1st timestamp */ 555 if (!slot->ds_last_time) 556 return 0; 557 558 hb_block = slot->ds_raw_block; 559 if (le64_to_cpu(hb_block->hb_seq) == slot->ds_last_time && 560 le64_to_cpu(hb_block->hb_generation) == slot->ds_last_generation && 561 hb_block->hb_node == slot->ds_node_num) 562 return 1; 563 564 #define ERRSTR1 "Another node is heartbeating on device" 565 #define ERRSTR2 "Heartbeat generation mismatch on device" 566 #define ERRSTR3 "Heartbeat sequence mismatch on device" 567 568 if (hb_block->hb_node != slot->ds_node_num) 569 errstr = ERRSTR1; 570 else if (le64_to_cpu(hb_block->hb_generation) != 571 slot->ds_last_generation) 572 errstr = ERRSTR2; 573 else 574 errstr = ERRSTR3; 575 576 mlog(ML_ERROR, "%s (%s): expected(%u:0x%llx, 0x%llx), " 577 "ondisk(%u:0x%llx, 0x%llx)\n", errstr, reg->hr_dev_name, 578 slot->ds_node_num, (unsigned long long)slot->ds_last_generation, 579 (unsigned long long)slot->ds_last_time, hb_block->hb_node, 580 (unsigned long long)le64_to_cpu(hb_block->hb_generation), 581 (unsigned long long)le64_to_cpu(hb_block->hb_seq)); 582 583 return 0; 584 } 585 586 static inline void o2hb_prepare_block(struct o2hb_region *reg, 587 u64 generation) 588 { 589 int node_num; 590 u64 cputime; 591 struct o2hb_disk_slot *slot; 592 struct o2hb_disk_heartbeat_block *hb_block; 593 594 node_num = o2nm_this_node(); 595 slot = ®->hr_slots[node_num]; 596 597 hb_block = (struct o2hb_disk_heartbeat_block *)slot->ds_raw_block; 598 memset(hb_block, 0, reg->hr_block_bytes); 599 /* TODO: time stuff */ 600 cputime = CURRENT_TIME.tv_sec; 601 if (!cputime) 602 cputime = 1; 603 604 hb_block->hb_seq = cpu_to_le64(cputime); 605 hb_block->hb_node = node_num; 606 hb_block->hb_generation = cpu_to_le64(generation); 607 hb_block->hb_dead_ms = cpu_to_le32(o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS); 608 609 /* This step must always happen last! */ 610 hb_block->hb_cksum = cpu_to_le32(o2hb_compute_block_crc_le(reg, 611 hb_block)); 612 613 mlog(ML_HB_BIO, "our node generation = 0x%llx, cksum = 0x%x\n", 614 (long long)generation, 615 le32_to_cpu(hb_block->hb_cksum)); 616 } 617 618 static void o2hb_fire_callbacks(struct o2hb_callback *hbcall, 619 struct o2nm_node *node, 620 int idx) 621 { 622 struct o2hb_callback_func *f; 623 624 list_for_each_entry(f, &hbcall->list, hc_item) { 625 mlog(ML_HEARTBEAT, "calling funcs %p\n", f); 626 (f->hc_func)(node, idx, f->hc_data); 627 } 628 } 629 630 /* Will run the list in order until we process the passed event */ 631 static void o2hb_run_event_list(struct o2hb_node_event *queued_event) 632 { 633 struct o2hb_callback *hbcall; 634 struct o2hb_node_event *event; 635 636 /* Holding callback sem assures we don't alter the callback 637 * lists when doing this, and serializes ourselves with other 638 * processes wanting callbacks. */ 639 down_write(&o2hb_callback_sem); 640 641 spin_lock(&o2hb_live_lock); 642 while (!list_empty(&o2hb_node_events) 643 && !list_empty(&queued_event->hn_item)) { 644 event = list_entry(o2hb_node_events.next, 645 struct o2hb_node_event, 646 hn_item); 647 list_del_init(&event->hn_item); 648 spin_unlock(&o2hb_live_lock); 649 650 mlog(ML_HEARTBEAT, "Node %s event for %d\n", 651 event->hn_event_type == O2HB_NODE_UP_CB ? "UP" : "DOWN", 652 event->hn_node_num); 653 654 hbcall = hbcall_from_type(event->hn_event_type); 655 656 /* We should *never* have gotten on to the list with a 657 * bad type... This isn't something that we should try 658 * to recover from. */ 659 BUG_ON(IS_ERR(hbcall)); 660 661 o2hb_fire_callbacks(hbcall, event->hn_node, event->hn_node_num); 662 663 spin_lock(&o2hb_live_lock); 664 } 665 spin_unlock(&o2hb_live_lock); 666 667 up_write(&o2hb_callback_sem); 668 } 669 670 static void o2hb_queue_node_event(struct o2hb_node_event *event, 671 enum o2hb_callback_type type, 672 struct o2nm_node *node, 673 int node_num) 674 { 675 assert_spin_locked(&o2hb_live_lock); 676 677 BUG_ON((!node) && (type != O2HB_NODE_DOWN_CB)); 678 679 event->hn_event_type = type; 680 event->hn_node = node; 681 event->hn_node_num = node_num; 682 683 mlog(ML_HEARTBEAT, "Queue node %s event for node %d\n", 684 type == O2HB_NODE_UP_CB ? "UP" : "DOWN", node_num); 685 686 list_add_tail(&event->hn_item, &o2hb_node_events); 687 } 688 689 static void o2hb_shutdown_slot(struct o2hb_disk_slot *slot) 690 { 691 struct o2hb_node_event event = 692 { .hn_item = LIST_HEAD_INIT(event.hn_item), }; 693 struct o2nm_node *node; 694 int queued = 0; 695 696 node = o2nm_get_node_by_num(slot->ds_node_num); 697 if (!node) 698 return; 699 700 spin_lock(&o2hb_live_lock); 701 if (!list_empty(&slot->ds_live_item)) { 702 mlog(ML_HEARTBEAT, "Shutdown, node %d leaves region\n", 703 slot->ds_node_num); 704 705 list_del_init(&slot->ds_live_item); 706 707 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) { 708 clear_bit(slot->ds_node_num, o2hb_live_node_bitmap); 709 710 o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB, node, 711 slot->ds_node_num); 712 queued = 1; 713 } 714 } 715 spin_unlock(&o2hb_live_lock); 716 717 if (queued) 718 o2hb_run_event_list(&event); 719 720 o2nm_node_put(node); 721 } 722 723 static void o2hb_set_quorum_device(struct o2hb_region *reg) 724 { 725 if (!o2hb_global_heartbeat_active()) 726 return; 727 728 /* Prevent race with o2hb_heartbeat_group_drop_item() */ 729 if (kthread_should_stop()) 730 return; 731 732 /* Tag region as quorum only after thread reaches steady state */ 733 if (atomic_read(®->hr_steady_iterations) != 0) 734 return; 735 736 spin_lock(&o2hb_live_lock); 737 738 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap)) 739 goto unlock; 740 741 /* 742 * A region can be added to the quorum only when it sees all 743 * live nodes heartbeat on it. In other words, the region has been 744 * added to all nodes. 745 */ 746 if (memcmp(reg->hr_live_node_bitmap, o2hb_live_node_bitmap, 747 sizeof(o2hb_live_node_bitmap))) 748 goto unlock; 749 750 printk(KERN_NOTICE "o2hb: Region %s (%s) is now a quorum device\n", 751 config_item_name(®->hr_item), reg->hr_dev_name); 752 753 set_bit(reg->hr_region_num, o2hb_quorum_region_bitmap); 754 755 /* 756 * If global heartbeat active, unpin all regions if the 757 * region count > CUT_OFF 758 */ 759 if (bitmap_weight(o2hb_quorum_region_bitmap, 760 O2NM_MAX_REGIONS) > O2HB_PIN_CUT_OFF) 761 o2hb_region_unpin(NULL); 762 unlock: 763 spin_unlock(&o2hb_live_lock); 764 } 765 766 static int o2hb_check_slot(struct o2hb_region *reg, 767 struct o2hb_disk_slot *slot) 768 { 769 int changed = 0, gen_changed = 0; 770 struct o2hb_node_event event = 771 { .hn_item = LIST_HEAD_INIT(event.hn_item), }; 772 struct o2nm_node *node; 773 struct o2hb_disk_heartbeat_block *hb_block = reg->hr_tmp_block; 774 u64 cputime; 775 unsigned int dead_ms = o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS; 776 unsigned int slot_dead_ms; 777 int tmp; 778 int queued = 0; 779 780 memcpy(hb_block, slot->ds_raw_block, reg->hr_block_bytes); 781 782 /* 783 * If a node is no longer configured but is still in the livemap, we 784 * may need to clear that bit from the livemap. 785 */ 786 node = o2nm_get_node_by_num(slot->ds_node_num); 787 if (!node) { 788 spin_lock(&o2hb_live_lock); 789 tmp = test_bit(slot->ds_node_num, o2hb_live_node_bitmap); 790 spin_unlock(&o2hb_live_lock); 791 if (!tmp) 792 return 0; 793 } 794 795 if (!o2hb_verify_crc(reg, hb_block)) { 796 /* all paths from here will drop o2hb_live_lock for 797 * us. */ 798 spin_lock(&o2hb_live_lock); 799 800 /* Don't print an error on the console in this case - 801 * a freshly formatted heartbeat area will not have a 802 * crc set on it. */ 803 if (list_empty(&slot->ds_live_item)) 804 goto out; 805 806 /* The node is live but pushed out a bad crc. We 807 * consider it a transient miss but don't populate any 808 * other values as they may be junk. */ 809 mlog(ML_ERROR, "Node %d has written a bad crc to %s\n", 810 slot->ds_node_num, reg->hr_dev_name); 811 o2hb_dump_slot(hb_block); 812 813 slot->ds_equal_samples++; 814 goto fire_callbacks; 815 } 816 817 /* we don't care if these wrap.. the state transitions below 818 * clear at the right places */ 819 cputime = le64_to_cpu(hb_block->hb_seq); 820 if (slot->ds_last_time != cputime) 821 slot->ds_changed_samples++; 822 else 823 slot->ds_equal_samples++; 824 slot->ds_last_time = cputime; 825 826 /* The node changed heartbeat generations. We assume this to 827 * mean it dropped off but came back before we timed out. We 828 * want to consider it down for the time being but don't want 829 * to lose any changed_samples state we might build up to 830 * considering it live again. */ 831 if (slot->ds_last_generation != le64_to_cpu(hb_block->hb_generation)) { 832 gen_changed = 1; 833 slot->ds_equal_samples = 0; 834 mlog(ML_HEARTBEAT, "Node %d changed generation (0x%llx " 835 "to 0x%llx)\n", slot->ds_node_num, 836 (long long)slot->ds_last_generation, 837 (long long)le64_to_cpu(hb_block->hb_generation)); 838 } 839 840 slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation); 841 842 mlog(ML_HEARTBEAT, "Slot %d gen 0x%llx cksum 0x%x " 843 "seq %llu last %llu changed %u equal %u\n", 844 slot->ds_node_num, (long long)slot->ds_last_generation, 845 le32_to_cpu(hb_block->hb_cksum), 846 (unsigned long long)le64_to_cpu(hb_block->hb_seq), 847 (unsigned long long)slot->ds_last_time, slot->ds_changed_samples, 848 slot->ds_equal_samples); 849 850 spin_lock(&o2hb_live_lock); 851 852 fire_callbacks: 853 /* dead nodes only come to life after some number of 854 * changes at any time during their dead time */ 855 if (list_empty(&slot->ds_live_item) && 856 slot->ds_changed_samples >= O2HB_LIVE_THRESHOLD) { 857 mlog(ML_HEARTBEAT, "Node %d (id 0x%llx) joined my region\n", 858 slot->ds_node_num, (long long)slot->ds_last_generation); 859 860 set_bit(slot->ds_node_num, reg->hr_live_node_bitmap); 861 862 /* first on the list generates a callback */ 863 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) { 864 mlog(ML_HEARTBEAT, "o2hb: Add node %d to live nodes " 865 "bitmap\n", slot->ds_node_num); 866 set_bit(slot->ds_node_num, o2hb_live_node_bitmap); 867 868 o2hb_queue_node_event(&event, O2HB_NODE_UP_CB, node, 869 slot->ds_node_num); 870 871 changed = 1; 872 queued = 1; 873 } 874 875 list_add_tail(&slot->ds_live_item, 876 &o2hb_live_slots[slot->ds_node_num]); 877 878 slot->ds_equal_samples = 0; 879 880 /* We want to be sure that all nodes agree on the 881 * number of milliseconds before a node will be 882 * considered dead. The self-fencing timeout is 883 * computed from this value, and a discrepancy might 884 * result in heartbeat calling a node dead when it 885 * hasn't self-fenced yet. */ 886 slot_dead_ms = le32_to_cpu(hb_block->hb_dead_ms); 887 if (slot_dead_ms && slot_dead_ms != dead_ms) { 888 /* TODO: Perhaps we can fail the region here. */ 889 mlog(ML_ERROR, "Node %d on device %s has a dead count " 890 "of %u ms, but our count is %u ms.\n" 891 "Please double check your configuration values " 892 "for 'O2CB_HEARTBEAT_THRESHOLD'\n", 893 slot->ds_node_num, reg->hr_dev_name, slot_dead_ms, 894 dead_ms); 895 } 896 goto out; 897 } 898 899 /* if the list is dead, we're done.. */ 900 if (list_empty(&slot->ds_live_item)) 901 goto out; 902 903 /* live nodes only go dead after enough consequtive missed 904 * samples.. reset the missed counter whenever we see 905 * activity */ 906 if (slot->ds_equal_samples >= o2hb_dead_threshold || gen_changed) { 907 mlog(ML_HEARTBEAT, "Node %d left my region\n", 908 slot->ds_node_num); 909 910 clear_bit(slot->ds_node_num, reg->hr_live_node_bitmap); 911 912 /* last off the live_slot generates a callback */ 913 list_del_init(&slot->ds_live_item); 914 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) { 915 mlog(ML_HEARTBEAT, "o2hb: Remove node %d from live " 916 "nodes bitmap\n", slot->ds_node_num); 917 clear_bit(slot->ds_node_num, o2hb_live_node_bitmap); 918 919 /* node can be null */ 920 o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB, 921 node, slot->ds_node_num); 922 923 changed = 1; 924 queued = 1; 925 } 926 927 /* We don't clear this because the node is still 928 * actually writing new blocks. */ 929 if (!gen_changed) 930 slot->ds_changed_samples = 0; 931 goto out; 932 } 933 if (slot->ds_changed_samples) { 934 slot->ds_changed_samples = 0; 935 slot->ds_equal_samples = 0; 936 } 937 out: 938 spin_unlock(&o2hb_live_lock); 939 940 if (queued) 941 o2hb_run_event_list(&event); 942 943 if (node) 944 o2nm_node_put(node); 945 return changed; 946 } 947 948 static int o2hb_highest_node(unsigned long *nodes, int numbits) 949 { 950 return find_last_bit(nodes, numbits); 951 } 952 953 static int o2hb_do_disk_heartbeat(struct o2hb_region *reg) 954 { 955 int i, ret, highest_node; 956 int membership_change = 0, own_slot_ok = 0; 957 unsigned long configured_nodes[BITS_TO_LONGS(O2NM_MAX_NODES)]; 958 unsigned long live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)]; 959 struct o2hb_bio_wait_ctxt write_wc; 960 961 ret = o2nm_configured_node_map(configured_nodes, 962 sizeof(configured_nodes)); 963 if (ret) { 964 mlog_errno(ret); 965 goto bail; 966 } 967 968 /* 969 * If a node is not configured but is in the livemap, we still need 970 * to read the slot so as to be able to remove it from the livemap. 971 */ 972 o2hb_fill_node_map(live_node_bitmap, sizeof(live_node_bitmap)); 973 i = -1; 974 while ((i = find_next_bit(live_node_bitmap, 975 O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) { 976 set_bit(i, configured_nodes); 977 } 978 979 highest_node = o2hb_highest_node(configured_nodes, O2NM_MAX_NODES); 980 if (highest_node >= O2NM_MAX_NODES) { 981 mlog(ML_NOTICE, "o2hb: No configured nodes found!\n"); 982 ret = -EINVAL; 983 goto bail; 984 } 985 986 /* No sense in reading the slots of nodes that don't exist 987 * yet. Of course, if the node definitions have holes in them 988 * then we're reading an empty slot anyway... Consider this 989 * best-effort. */ 990 ret = o2hb_read_slots(reg, highest_node + 1); 991 if (ret < 0) { 992 mlog_errno(ret); 993 goto bail; 994 } 995 996 /* With an up to date view of the slots, we can check that no 997 * other node has been improperly configured to heartbeat in 998 * our slot. */ 999 own_slot_ok = o2hb_check_own_slot(reg); 1000 1001 /* fill in the proper info for our next heartbeat */ 1002 o2hb_prepare_block(reg, reg->hr_generation); 1003 1004 ret = o2hb_issue_node_write(reg, &write_wc); 1005 if (ret < 0) { 1006 mlog_errno(ret); 1007 goto bail; 1008 } 1009 1010 i = -1; 1011 while((i = find_next_bit(configured_nodes, 1012 O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) { 1013 membership_change |= o2hb_check_slot(reg, ®->hr_slots[i]); 1014 } 1015 1016 /* 1017 * We have to be sure we've advertised ourselves on disk 1018 * before we can go to steady state. This ensures that 1019 * people we find in our steady state have seen us. 1020 */ 1021 o2hb_wait_on_io(reg, &write_wc); 1022 if (write_wc.wc_error) { 1023 /* Do not re-arm the write timeout on I/O error - we 1024 * can't be sure that the new block ever made it to 1025 * disk */ 1026 mlog(ML_ERROR, "Write error %d on device \"%s\"\n", 1027 write_wc.wc_error, reg->hr_dev_name); 1028 ret = write_wc.wc_error; 1029 goto bail; 1030 } 1031 1032 /* Skip disarming the timeout if own slot has stale/bad data */ 1033 if (own_slot_ok) { 1034 o2hb_set_quorum_device(reg); 1035 o2hb_arm_write_timeout(reg); 1036 } 1037 1038 bail: 1039 /* let the person who launched us know when things are steady */ 1040 if (atomic_read(®->hr_steady_iterations) != 0) { 1041 if (!ret && own_slot_ok && !membership_change) { 1042 if (atomic_dec_and_test(®->hr_steady_iterations)) 1043 wake_up(&o2hb_steady_queue); 1044 } 1045 } 1046 1047 if (atomic_read(®->hr_steady_iterations) != 0) { 1048 if (atomic_dec_and_test(®->hr_unsteady_iterations)) { 1049 printk(KERN_NOTICE "o2hb: Unable to stabilize " 1050 "heartbeart on region %s (%s)\n", 1051 config_item_name(®->hr_item), 1052 reg->hr_dev_name); 1053 atomic_set(®->hr_steady_iterations, 0); 1054 reg->hr_aborted_start = 1; 1055 wake_up(&o2hb_steady_queue); 1056 ret = -EIO; 1057 } 1058 } 1059 1060 return ret; 1061 } 1062 1063 /* 1064 * we ride the region ref that the region dir holds. before the region 1065 * dir is removed and drops it ref it will wait to tear down this 1066 * thread. 1067 */ 1068 static int o2hb_thread(void *data) 1069 { 1070 int i, ret; 1071 struct o2hb_region *reg = data; 1072 struct o2hb_bio_wait_ctxt write_wc; 1073 ktime_t before_hb, after_hb; 1074 unsigned int elapsed_msec; 1075 1076 mlog(ML_HEARTBEAT|ML_KTHREAD, "hb thread running\n"); 1077 1078 set_user_nice(current, MIN_NICE); 1079 1080 /* Pin node */ 1081 o2nm_depend_this_node(); 1082 1083 while (!kthread_should_stop() && 1084 !reg->hr_unclean_stop && !reg->hr_aborted_start) { 1085 /* We track the time spent inside 1086 * o2hb_do_disk_heartbeat so that we avoid more than 1087 * hr_timeout_ms between disk writes. On busy systems 1088 * this should result in a heartbeat which is less 1089 * likely to time itself out. */ 1090 before_hb = ktime_get_real(); 1091 1092 ret = o2hb_do_disk_heartbeat(reg); 1093 1094 after_hb = ktime_get_real(); 1095 1096 elapsed_msec = (unsigned int) 1097 ktime_ms_delta(after_hb, before_hb); 1098 1099 mlog(ML_HEARTBEAT, 1100 "start = %lld, end = %lld, msec = %u, ret = %d\n", 1101 before_hb.tv64, after_hb.tv64, elapsed_msec, ret); 1102 1103 if (!kthread_should_stop() && 1104 elapsed_msec < reg->hr_timeout_ms) { 1105 /* the kthread api has blocked signals for us so no 1106 * need to record the return value. */ 1107 msleep_interruptible(reg->hr_timeout_ms - elapsed_msec); 1108 } 1109 } 1110 1111 o2hb_disarm_write_timeout(reg); 1112 1113 /* unclean stop is only used in very bad situation */ 1114 for(i = 0; !reg->hr_unclean_stop && i < reg->hr_blocks; i++) 1115 o2hb_shutdown_slot(®->hr_slots[i]); 1116 1117 /* Explicit down notification - avoid forcing the other nodes 1118 * to timeout on this region when we could just as easily 1119 * write a clear generation - thus indicating to them that 1120 * this node has left this region. 1121 */ 1122 if (!reg->hr_unclean_stop && !reg->hr_aborted_start) { 1123 o2hb_prepare_block(reg, 0); 1124 ret = o2hb_issue_node_write(reg, &write_wc); 1125 if (ret == 0) 1126 o2hb_wait_on_io(reg, &write_wc); 1127 else 1128 mlog_errno(ret); 1129 } 1130 1131 /* Unpin node */ 1132 o2nm_undepend_this_node(); 1133 1134 mlog(ML_HEARTBEAT|ML_KTHREAD, "o2hb thread exiting\n"); 1135 1136 return 0; 1137 } 1138 1139 #ifdef CONFIG_DEBUG_FS 1140 static int o2hb_debug_open(struct inode *inode, struct file *file) 1141 { 1142 struct o2hb_debug_buf *db = inode->i_private; 1143 struct o2hb_region *reg; 1144 unsigned long map[BITS_TO_LONGS(O2NM_MAX_NODES)]; 1145 unsigned long lts; 1146 char *buf = NULL; 1147 int i = -1; 1148 int out = 0; 1149 1150 /* max_nodes should be the largest bitmap we pass here */ 1151 BUG_ON(sizeof(map) < db->db_size); 1152 1153 buf = kmalloc(PAGE_SIZE, GFP_KERNEL); 1154 if (!buf) 1155 goto bail; 1156 1157 switch (db->db_type) { 1158 case O2HB_DB_TYPE_LIVENODES: 1159 case O2HB_DB_TYPE_LIVEREGIONS: 1160 case O2HB_DB_TYPE_QUORUMREGIONS: 1161 case O2HB_DB_TYPE_FAILEDREGIONS: 1162 spin_lock(&o2hb_live_lock); 1163 memcpy(map, db->db_data, db->db_size); 1164 spin_unlock(&o2hb_live_lock); 1165 break; 1166 1167 case O2HB_DB_TYPE_REGION_LIVENODES: 1168 spin_lock(&o2hb_live_lock); 1169 reg = (struct o2hb_region *)db->db_data; 1170 memcpy(map, reg->hr_live_node_bitmap, db->db_size); 1171 spin_unlock(&o2hb_live_lock); 1172 break; 1173 1174 case O2HB_DB_TYPE_REGION_NUMBER: 1175 reg = (struct o2hb_region *)db->db_data; 1176 out += snprintf(buf + out, PAGE_SIZE - out, "%d\n", 1177 reg->hr_region_num); 1178 goto done; 1179 1180 case O2HB_DB_TYPE_REGION_ELAPSED_TIME: 1181 reg = (struct o2hb_region *)db->db_data; 1182 lts = reg->hr_last_timeout_start; 1183 /* If 0, it has never been set before */ 1184 if (lts) 1185 lts = jiffies_to_msecs(jiffies - lts); 1186 out += snprintf(buf + out, PAGE_SIZE - out, "%lu\n", lts); 1187 goto done; 1188 1189 case O2HB_DB_TYPE_REGION_PINNED: 1190 reg = (struct o2hb_region *)db->db_data; 1191 out += snprintf(buf + out, PAGE_SIZE - out, "%u\n", 1192 !!reg->hr_item_pinned); 1193 goto done; 1194 1195 default: 1196 goto done; 1197 } 1198 1199 while ((i = find_next_bit(map, db->db_len, i + 1)) < db->db_len) 1200 out += snprintf(buf + out, PAGE_SIZE - out, "%d ", i); 1201 out += snprintf(buf + out, PAGE_SIZE - out, "\n"); 1202 1203 done: 1204 i_size_write(inode, out); 1205 1206 file->private_data = buf; 1207 1208 return 0; 1209 bail: 1210 return -ENOMEM; 1211 } 1212 1213 static int o2hb_debug_release(struct inode *inode, struct file *file) 1214 { 1215 kfree(file->private_data); 1216 return 0; 1217 } 1218 1219 static ssize_t o2hb_debug_read(struct file *file, char __user *buf, 1220 size_t nbytes, loff_t *ppos) 1221 { 1222 return simple_read_from_buffer(buf, nbytes, ppos, file->private_data, 1223 i_size_read(file->f_mapping->host)); 1224 } 1225 #else 1226 static int o2hb_debug_open(struct inode *inode, struct file *file) 1227 { 1228 return 0; 1229 } 1230 static int o2hb_debug_release(struct inode *inode, struct file *file) 1231 { 1232 return 0; 1233 } 1234 static ssize_t o2hb_debug_read(struct file *file, char __user *buf, 1235 size_t nbytes, loff_t *ppos) 1236 { 1237 return 0; 1238 } 1239 #endif /* CONFIG_DEBUG_FS */ 1240 1241 static const struct file_operations o2hb_debug_fops = { 1242 .open = o2hb_debug_open, 1243 .release = o2hb_debug_release, 1244 .read = o2hb_debug_read, 1245 .llseek = generic_file_llseek, 1246 }; 1247 1248 void o2hb_exit(void) 1249 { 1250 kfree(o2hb_db_livenodes); 1251 kfree(o2hb_db_liveregions); 1252 kfree(o2hb_db_quorumregions); 1253 kfree(o2hb_db_failedregions); 1254 debugfs_remove(o2hb_debug_failedregions); 1255 debugfs_remove(o2hb_debug_quorumregions); 1256 debugfs_remove(o2hb_debug_liveregions); 1257 debugfs_remove(o2hb_debug_livenodes); 1258 debugfs_remove(o2hb_debug_dir); 1259 } 1260 1261 static struct dentry *o2hb_debug_create(const char *name, struct dentry *dir, 1262 struct o2hb_debug_buf **db, int db_len, 1263 int type, int size, int len, void *data) 1264 { 1265 *db = kmalloc(db_len, GFP_KERNEL); 1266 if (!*db) 1267 return NULL; 1268 1269 (*db)->db_type = type; 1270 (*db)->db_size = size; 1271 (*db)->db_len = len; 1272 (*db)->db_data = data; 1273 1274 return debugfs_create_file(name, S_IFREG|S_IRUSR, dir, *db, 1275 &o2hb_debug_fops); 1276 } 1277 1278 static int o2hb_debug_init(void) 1279 { 1280 int ret = -ENOMEM; 1281 1282 o2hb_debug_dir = debugfs_create_dir(O2HB_DEBUG_DIR, NULL); 1283 if (!o2hb_debug_dir) { 1284 mlog_errno(ret); 1285 goto bail; 1286 } 1287 1288 o2hb_debug_livenodes = o2hb_debug_create(O2HB_DEBUG_LIVENODES, 1289 o2hb_debug_dir, 1290 &o2hb_db_livenodes, 1291 sizeof(*o2hb_db_livenodes), 1292 O2HB_DB_TYPE_LIVENODES, 1293 sizeof(o2hb_live_node_bitmap), 1294 O2NM_MAX_NODES, 1295 o2hb_live_node_bitmap); 1296 if (!o2hb_debug_livenodes) { 1297 mlog_errno(ret); 1298 goto bail; 1299 } 1300 1301 o2hb_debug_liveregions = o2hb_debug_create(O2HB_DEBUG_LIVEREGIONS, 1302 o2hb_debug_dir, 1303 &o2hb_db_liveregions, 1304 sizeof(*o2hb_db_liveregions), 1305 O2HB_DB_TYPE_LIVEREGIONS, 1306 sizeof(o2hb_live_region_bitmap), 1307 O2NM_MAX_REGIONS, 1308 o2hb_live_region_bitmap); 1309 if (!o2hb_debug_liveregions) { 1310 mlog_errno(ret); 1311 goto bail; 1312 } 1313 1314 o2hb_debug_quorumregions = 1315 o2hb_debug_create(O2HB_DEBUG_QUORUMREGIONS, 1316 o2hb_debug_dir, 1317 &o2hb_db_quorumregions, 1318 sizeof(*o2hb_db_quorumregions), 1319 O2HB_DB_TYPE_QUORUMREGIONS, 1320 sizeof(o2hb_quorum_region_bitmap), 1321 O2NM_MAX_REGIONS, 1322 o2hb_quorum_region_bitmap); 1323 if (!o2hb_debug_quorumregions) { 1324 mlog_errno(ret); 1325 goto bail; 1326 } 1327 1328 o2hb_debug_failedregions = 1329 o2hb_debug_create(O2HB_DEBUG_FAILEDREGIONS, 1330 o2hb_debug_dir, 1331 &o2hb_db_failedregions, 1332 sizeof(*o2hb_db_failedregions), 1333 O2HB_DB_TYPE_FAILEDREGIONS, 1334 sizeof(o2hb_failed_region_bitmap), 1335 O2NM_MAX_REGIONS, 1336 o2hb_failed_region_bitmap); 1337 if (!o2hb_debug_failedregions) { 1338 mlog_errno(ret); 1339 goto bail; 1340 } 1341 1342 ret = 0; 1343 bail: 1344 if (ret) 1345 o2hb_exit(); 1346 1347 return ret; 1348 } 1349 1350 int o2hb_init(void) 1351 { 1352 int i; 1353 1354 for (i = 0; i < ARRAY_SIZE(o2hb_callbacks); i++) 1355 INIT_LIST_HEAD(&o2hb_callbacks[i].list); 1356 1357 for (i = 0; i < ARRAY_SIZE(o2hb_live_slots); i++) 1358 INIT_LIST_HEAD(&o2hb_live_slots[i]); 1359 1360 INIT_LIST_HEAD(&o2hb_node_events); 1361 1362 memset(o2hb_live_node_bitmap, 0, sizeof(o2hb_live_node_bitmap)); 1363 memset(o2hb_region_bitmap, 0, sizeof(o2hb_region_bitmap)); 1364 memset(o2hb_live_region_bitmap, 0, sizeof(o2hb_live_region_bitmap)); 1365 memset(o2hb_quorum_region_bitmap, 0, sizeof(o2hb_quorum_region_bitmap)); 1366 memset(o2hb_failed_region_bitmap, 0, sizeof(o2hb_failed_region_bitmap)); 1367 1368 o2hb_dependent_users = 0; 1369 1370 return o2hb_debug_init(); 1371 } 1372 1373 /* if we're already in a callback then we're already serialized by the sem */ 1374 static void o2hb_fill_node_map_from_callback(unsigned long *map, 1375 unsigned bytes) 1376 { 1377 BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long))); 1378 1379 memcpy(map, &o2hb_live_node_bitmap, bytes); 1380 } 1381 1382 /* 1383 * get a map of all nodes that are heartbeating in any regions 1384 */ 1385 void o2hb_fill_node_map(unsigned long *map, unsigned bytes) 1386 { 1387 /* callers want to serialize this map and callbacks so that they 1388 * can trust that they don't miss nodes coming to the party */ 1389 down_read(&o2hb_callback_sem); 1390 spin_lock(&o2hb_live_lock); 1391 o2hb_fill_node_map_from_callback(map, bytes); 1392 spin_unlock(&o2hb_live_lock); 1393 up_read(&o2hb_callback_sem); 1394 } 1395 EXPORT_SYMBOL_GPL(o2hb_fill_node_map); 1396 1397 /* 1398 * heartbeat configfs bits. The heartbeat set is a default set under 1399 * the cluster set in nodemanager.c. 1400 */ 1401 1402 static struct o2hb_region *to_o2hb_region(struct config_item *item) 1403 { 1404 return item ? container_of(item, struct o2hb_region, hr_item) : NULL; 1405 } 1406 1407 /* drop_item only drops its ref after killing the thread, nothing should 1408 * be using the region anymore. this has to clean up any state that 1409 * attributes might have built up. */ 1410 static void o2hb_region_release(struct config_item *item) 1411 { 1412 int i; 1413 struct page *page; 1414 struct o2hb_region *reg = to_o2hb_region(item); 1415 1416 mlog(ML_HEARTBEAT, "hb region release (%s)\n", reg->hr_dev_name); 1417 1418 kfree(reg->hr_tmp_block); 1419 1420 if (reg->hr_slot_data) { 1421 for (i = 0; i < reg->hr_num_pages; i++) { 1422 page = reg->hr_slot_data[i]; 1423 if (page) 1424 __free_page(page); 1425 } 1426 kfree(reg->hr_slot_data); 1427 } 1428 1429 if (reg->hr_bdev) 1430 blkdev_put(reg->hr_bdev, FMODE_READ|FMODE_WRITE); 1431 1432 kfree(reg->hr_slots); 1433 1434 kfree(reg->hr_db_regnum); 1435 kfree(reg->hr_db_livenodes); 1436 debugfs_remove(reg->hr_debug_livenodes); 1437 debugfs_remove(reg->hr_debug_regnum); 1438 debugfs_remove(reg->hr_debug_elapsed_time); 1439 debugfs_remove(reg->hr_debug_pinned); 1440 debugfs_remove(reg->hr_debug_dir); 1441 1442 spin_lock(&o2hb_live_lock); 1443 list_del(®->hr_all_item); 1444 spin_unlock(&o2hb_live_lock); 1445 1446 kfree(reg); 1447 } 1448 1449 static int o2hb_read_block_input(struct o2hb_region *reg, 1450 const char *page, 1451 size_t count, 1452 unsigned long *ret_bytes, 1453 unsigned int *ret_bits) 1454 { 1455 unsigned long bytes; 1456 char *p = (char *)page; 1457 1458 bytes = simple_strtoul(p, &p, 0); 1459 if (!p || (*p && (*p != '\n'))) 1460 return -EINVAL; 1461 1462 /* Heartbeat and fs min / max block sizes are the same. */ 1463 if (bytes > 4096 || bytes < 512) 1464 return -ERANGE; 1465 if (hweight16(bytes) != 1) 1466 return -EINVAL; 1467 1468 if (ret_bytes) 1469 *ret_bytes = bytes; 1470 if (ret_bits) 1471 *ret_bits = ffs(bytes) - 1; 1472 1473 return 0; 1474 } 1475 1476 static ssize_t o2hb_region_block_bytes_show(struct config_item *item, 1477 char *page) 1478 { 1479 return sprintf(page, "%u\n", to_o2hb_region(item)->hr_block_bytes); 1480 } 1481 1482 static ssize_t o2hb_region_block_bytes_store(struct config_item *item, 1483 const char *page, 1484 size_t count) 1485 { 1486 struct o2hb_region *reg = to_o2hb_region(item); 1487 int status; 1488 unsigned long block_bytes; 1489 unsigned int block_bits; 1490 1491 if (reg->hr_bdev) 1492 return -EINVAL; 1493 1494 status = o2hb_read_block_input(reg, page, count, 1495 &block_bytes, &block_bits); 1496 if (status) 1497 return status; 1498 1499 reg->hr_block_bytes = (unsigned int)block_bytes; 1500 reg->hr_block_bits = block_bits; 1501 1502 return count; 1503 } 1504 1505 static ssize_t o2hb_region_start_block_show(struct config_item *item, 1506 char *page) 1507 { 1508 return sprintf(page, "%llu\n", to_o2hb_region(item)->hr_start_block); 1509 } 1510 1511 static ssize_t o2hb_region_start_block_store(struct config_item *item, 1512 const char *page, 1513 size_t count) 1514 { 1515 struct o2hb_region *reg = to_o2hb_region(item); 1516 unsigned long long tmp; 1517 char *p = (char *)page; 1518 1519 if (reg->hr_bdev) 1520 return -EINVAL; 1521 1522 tmp = simple_strtoull(p, &p, 0); 1523 if (!p || (*p && (*p != '\n'))) 1524 return -EINVAL; 1525 1526 reg->hr_start_block = tmp; 1527 1528 return count; 1529 } 1530 1531 static ssize_t o2hb_region_blocks_show(struct config_item *item, char *page) 1532 { 1533 return sprintf(page, "%d\n", to_o2hb_region(item)->hr_blocks); 1534 } 1535 1536 static ssize_t o2hb_region_blocks_store(struct config_item *item, 1537 const char *page, 1538 size_t count) 1539 { 1540 struct o2hb_region *reg = to_o2hb_region(item); 1541 unsigned long tmp; 1542 char *p = (char *)page; 1543 1544 if (reg->hr_bdev) 1545 return -EINVAL; 1546 1547 tmp = simple_strtoul(p, &p, 0); 1548 if (!p || (*p && (*p != '\n'))) 1549 return -EINVAL; 1550 1551 if (tmp > O2NM_MAX_NODES || tmp == 0) 1552 return -ERANGE; 1553 1554 reg->hr_blocks = (unsigned int)tmp; 1555 1556 return count; 1557 } 1558 1559 static ssize_t o2hb_region_dev_show(struct config_item *item, char *page) 1560 { 1561 unsigned int ret = 0; 1562 1563 if (to_o2hb_region(item)->hr_bdev) 1564 ret = sprintf(page, "%s\n", to_o2hb_region(item)->hr_dev_name); 1565 1566 return ret; 1567 } 1568 1569 static void o2hb_init_region_params(struct o2hb_region *reg) 1570 { 1571 reg->hr_slots_per_page = PAGE_CACHE_SIZE >> reg->hr_block_bits; 1572 reg->hr_timeout_ms = O2HB_REGION_TIMEOUT_MS; 1573 1574 mlog(ML_HEARTBEAT, "hr_start_block = %llu, hr_blocks = %u\n", 1575 reg->hr_start_block, reg->hr_blocks); 1576 mlog(ML_HEARTBEAT, "hr_block_bytes = %u, hr_block_bits = %u\n", 1577 reg->hr_block_bytes, reg->hr_block_bits); 1578 mlog(ML_HEARTBEAT, "hr_timeout_ms = %u\n", reg->hr_timeout_ms); 1579 mlog(ML_HEARTBEAT, "dead threshold = %u\n", o2hb_dead_threshold); 1580 } 1581 1582 static int o2hb_map_slot_data(struct o2hb_region *reg) 1583 { 1584 int i, j; 1585 unsigned int last_slot; 1586 unsigned int spp = reg->hr_slots_per_page; 1587 struct page *page; 1588 char *raw; 1589 struct o2hb_disk_slot *slot; 1590 1591 reg->hr_tmp_block = kmalloc(reg->hr_block_bytes, GFP_KERNEL); 1592 if (reg->hr_tmp_block == NULL) 1593 return -ENOMEM; 1594 1595 reg->hr_slots = kcalloc(reg->hr_blocks, 1596 sizeof(struct o2hb_disk_slot), GFP_KERNEL); 1597 if (reg->hr_slots == NULL) 1598 return -ENOMEM; 1599 1600 for(i = 0; i < reg->hr_blocks; i++) { 1601 slot = ®->hr_slots[i]; 1602 slot->ds_node_num = i; 1603 INIT_LIST_HEAD(&slot->ds_live_item); 1604 slot->ds_raw_block = NULL; 1605 } 1606 1607 reg->hr_num_pages = (reg->hr_blocks + spp - 1) / spp; 1608 mlog(ML_HEARTBEAT, "Going to require %u pages to cover %u blocks " 1609 "at %u blocks per page\n", 1610 reg->hr_num_pages, reg->hr_blocks, spp); 1611 1612 reg->hr_slot_data = kcalloc(reg->hr_num_pages, sizeof(struct page *), 1613 GFP_KERNEL); 1614 if (!reg->hr_slot_data) 1615 return -ENOMEM; 1616 1617 for(i = 0; i < reg->hr_num_pages; i++) { 1618 page = alloc_page(GFP_KERNEL); 1619 if (!page) 1620 return -ENOMEM; 1621 1622 reg->hr_slot_data[i] = page; 1623 1624 last_slot = i * spp; 1625 raw = page_address(page); 1626 for (j = 0; 1627 (j < spp) && ((j + last_slot) < reg->hr_blocks); 1628 j++) { 1629 BUG_ON((j + last_slot) >= reg->hr_blocks); 1630 1631 slot = ®->hr_slots[j + last_slot]; 1632 slot->ds_raw_block = 1633 (struct o2hb_disk_heartbeat_block *) raw; 1634 1635 raw += reg->hr_block_bytes; 1636 } 1637 } 1638 1639 return 0; 1640 } 1641 1642 /* Read in all the slots available and populate the tracking 1643 * structures so that we can start with a baseline idea of what's 1644 * there. */ 1645 static int o2hb_populate_slot_data(struct o2hb_region *reg) 1646 { 1647 int ret, i; 1648 struct o2hb_disk_slot *slot; 1649 struct o2hb_disk_heartbeat_block *hb_block; 1650 1651 ret = o2hb_read_slots(reg, reg->hr_blocks); 1652 if (ret) 1653 goto out; 1654 1655 /* We only want to get an idea of the values initially in each 1656 * slot, so we do no verification - o2hb_check_slot will 1657 * actually determine if each configured slot is valid and 1658 * whether any values have changed. */ 1659 for(i = 0; i < reg->hr_blocks; i++) { 1660 slot = ®->hr_slots[i]; 1661 hb_block = (struct o2hb_disk_heartbeat_block *) slot->ds_raw_block; 1662 1663 /* Only fill the values that o2hb_check_slot uses to 1664 * determine changing slots */ 1665 slot->ds_last_time = le64_to_cpu(hb_block->hb_seq); 1666 slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation); 1667 } 1668 1669 out: 1670 return ret; 1671 } 1672 1673 /* this is acting as commit; we set up all of hr_bdev and hr_task or nothing */ 1674 static ssize_t o2hb_region_dev_store(struct config_item *item, 1675 const char *page, 1676 size_t count) 1677 { 1678 struct o2hb_region *reg = to_o2hb_region(item); 1679 struct task_struct *hb_task; 1680 long fd; 1681 int sectsize; 1682 char *p = (char *)page; 1683 struct fd f; 1684 struct inode *inode; 1685 ssize_t ret = -EINVAL; 1686 int live_threshold; 1687 1688 if (reg->hr_bdev) 1689 goto out; 1690 1691 /* We can't heartbeat without having had our node number 1692 * configured yet. */ 1693 if (o2nm_this_node() == O2NM_MAX_NODES) 1694 goto out; 1695 1696 fd = simple_strtol(p, &p, 0); 1697 if (!p || (*p && (*p != '\n'))) 1698 goto out; 1699 1700 if (fd < 0 || fd >= INT_MAX) 1701 goto out; 1702 1703 f = fdget(fd); 1704 if (f.file == NULL) 1705 goto out; 1706 1707 if (reg->hr_blocks == 0 || reg->hr_start_block == 0 || 1708 reg->hr_block_bytes == 0) 1709 goto out2; 1710 1711 inode = igrab(f.file->f_mapping->host); 1712 if (inode == NULL) 1713 goto out2; 1714 1715 if (!S_ISBLK(inode->i_mode)) 1716 goto out3; 1717 1718 reg->hr_bdev = I_BDEV(f.file->f_mapping->host); 1719 ret = blkdev_get(reg->hr_bdev, FMODE_WRITE | FMODE_READ, NULL); 1720 if (ret) { 1721 reg->hr_bdev = NULL; 1722 goto out3; 1723 } 1724 inode = NULL; 1725 1726 bdevname(reg->hr_bdev, reg->hr_dev_name); 1727 1728 sectsize = bdev_logical_block_size(reg->hr_bdev); 1729 if (sectsize != reg->hr_block_bytes) { 1730 mlog(ML_ERROR, 1731 "blocksize %u incorrect for device, expected %d", 1732 reg->hr_block_bytes, sectsize); 1733 ret = -EINVAL; 1734 goto out3; 1735 } 1736 1737 o2hb_init_region_params(reg); 1738 1739 /* Generation of zero is invalid */ 1740 do { 1741 get_random_bytes(®->hr_generation, 1742 sizeof(reg->hr_generation)); 1743 } while (reg->hr_generation == 0); 1744 1745 ret = o2hb_map_slot_data(reg); 1746 if (ret) { 1747 mlog_errno(ret); 1748 goto out3; 1749 } 1750 1751 ret = o2hb_populate_slot_data(reg); 1752 if (ret) { 1753 mlog_errno(ret); 1754 goto out3; 1755 } 1756 1757 INIT_DELAYED_WORK(®->hr_write_timeout_work, o2hb_write_timeout); 1758 1759 /* 1760 * A node is considered live after it has beat LIVE_THRESHOLD 1761 * times. We're not steady until we've given them a chance 1762 * _after_ our first read. 1763 * The default threshold is bare minimum so as to limit the delay 1764 * during mounts. For global heartbeat, the threshold doubled for the 1765 * first region. 1766 */ 1767 live_threshold = O2HB_LIVE_THRESHOLD; 1768 if (o2hb_global_heartbeat_active()) { 1769 spin_lock(&o2hb_live_lock); 1770 if (bitmap_weight(o2hb_region_bitmap, O2NM_MAX_REGIONS) == 1) 1771 live_threshold <<= 1; 1772 spin_unlock(&o2hb_live_lock); 1773 } 1774 ++live_threshold; 1775 atomic_set(®->hr_steady_iterations, live_threshold); 1776 /* unsteady_iterations is double the steady_iterations */ 1777 atomic_set(®->hr_unsteady_iterations, (live_threshold << 1)); 1778 1779 hb_task = kthread_run(o2hb_thread, reg, "o2hb-%s", 1780 reg->hr_item.ci_name); 1781 if (IS_ERR(hb_task)) { 1782 ret = PTR_ERR(hb_task); 1783 mlog_errno(ret); 1784 goto out3; 1785 } 1786 1787 spin_lock(&o2hb_live_lock); 1788 reg->hr_task = hb_task; 1789 spin_unlock(&o2hb_live_lock); 1790 1791 ret = wait_event_interruptible(o2hb_steady_queue, 1792 atomic_read(®->hr_steady_iterations) == 0); 1793 if (ret) { 1794 atomic_set(®->hr_steady_iterations, 0); 1795 reg->hr_aborted_start = 1; 1796 } 1797 1798 if (reg->hr_aborted_start) { 1799 ret = -EIO; 1800 goto out3; 1801 } 1802 1803 /* Ok, we were woken. Make sure it wasn't by drop_item() */ 1804 spin_lock(&o2hb_live_lock); 1805 hb_task = reg->hr_task; 1806 if (o2hb_global_heartbeat_active()) 1807 set_bit(reg->hr_region_num, o2hb_live_region_bitmap); 1808 spin_unlock(&o2hb_live_lock); 1809 1810 if (hb_task) 1811 ret = count; 1812 else 1813 ret = -EIO; 1814 1815 if (hb_task && o2hb_global_heartbeat_active()) 1816 printk(KERN_NOTICE "o2hb: Heartbeat started on region %s (%s)\n", 1817 config_item_name(®->hr_item), reg->hr_dev_name); 1818 1819 out3: 1820 iput(inode); 1821 out2: 1822 fdput(f); 1823 out: 1824 if (ret < 0) { 1825 if (reg->hr_bdev) { 1826 blkdev_put(reg->hr_bdev, FMODE_READ|FMODE_WRITE); 1827 reg->hr_bdev = NULL; 1828 } 1829 } 1830 return ret; 1831 } 1832 1833 static ssize_t o2hb_region_pid_show(struct config_item *item, char *page) 1834 { 1835 struct o2hb_region *reg = to_o2hb_region(item); 1836 pid_t pid = 0; 1837 1838 spin_lock(&o2hb_live_lock); 1839 if (reg->hr_task) 1840 pid = task_pid_nr(reg->hr_task); 1841 spin_unlock(&o2hb_live_lock); 1842 1843 if (!pid) 1844 return 0; 1845 1846 return sprintf(page, "%u\n", pid); 1847 } 1848 1849 CONFIGFS_ATTR(o2hb_region_, block_bytes); 1850 CONFIGFS_ATTR(o2hb_region_, start_block); 1851 CONFIGFS_ATTR(o2hb_region_, blocks); 1852 CONFIGFS_ATTR(o2hb_region_, dev); 1853 CONFIGFS_ATTR_RO(o2hb_region_, pid); 1854 1855 static struct configfs_attribute *o2hb_region_attrs[] = { 1856 &o2hb_region_attr_block_bytes, 1857 &o2hb_region_attr_start_block, 1858 &o2hb_region_attr_blocks, 1859 &o2hb_region_attr_dev, 1860 &o2hb_region_attr_pid, 1861 NULL, 1862 }; 1863 1864 static struct configfs_item_operations o2hb_region_item_ops = { 1865 .release = o2hb_region_release, 1866 }; 1867 1868 static struct config_item_type o2hb_region_type = { 1869 .ct_item_ops = &o2hb_region_item_ops, 1870 .ct_attrs = o2hb_region_attrs, 1871 .ct_owner = THIS_MODULE, 1872 }; 1873 1874 /* heartbeat set */ 1875 1876 struct o2hb_heartbeat_group { 1877 struct config_group hs_group; 1878 /* some stuff? */ 1879 }; 1880 1881 static struct o2hb_heartbeat_group *to_o2hb_heartbeat_group(struct config_group *group) 1882 { 1883 return group ? 1884 container_of(group, struct o2hb_heartbeat_group, hs_group) 1885 : NULL; 1886 } 1887 1888 static int o2hb_debug_region_init(struct o2hb_region *reg, struct dentry *dir) 1889 { 1890 int ret = -ENOMEM; 1891 1892 reg->hr_debug_dir = 1893 debugfs_create_dir(config_item_name(®->hr_item), dir); 1894 if (!reg->hr_debug_dir) { 1895 mlog_errno(ret); 1896 goto bail; 1897 } 1898 1899 reg->hr_debug_livenodes = 1900 o2hb_debug_create(O2HB_DEBUG_LIVENODES, 1901 reg->hr_debug_dir, 1902 &(reg->hr_db_livenodes), 1903 sizeof(*(reg->hr_db_livenodes)), 1904 O2HB_DB_TYPE_REGION_LIVENODES, 1905 sizeof(reg->hr_live_node_bitmap), 1906 O2NM_MAX_NODES, reg); 1907 if (!reg->hr_debug_livenodes) { 1908 mlog_errno(ret); 1909 goto bail; 1910 } 1911 1912 reg->hr_debug_regnum = 1913 o2hb_debug_create(O2HB_DEBUG_REGION_NUMBER, 1914 reg->hr_debug_dir, 1915 &(reg->hr_db_regnum), 1916 sizeof(*(reg->hr_db_regnum)), 1917 O2HB_DB_TYPE_REGION_NUMBER, 1918 0, O2NM_MAX_NODES, reg); 1919 if (!reg->hr_debug_regnum) { 1920 mlog_errno(ret); 1921 goto bail; 1922 } 1923 1924 reg->hr_debug_elapsed_time = 1925 o2hb_debug_create(O2HB_DEBUG_REGION_ELAPSED_TIME, 1926 reg->hr_debug_dir, 1927 &(reg->hr_db_elapsed_time), 1928 sizeof(*(reg->hr_db_elapsed_time)), 1929 O2HB_DB_TYPE_REGION_ELAPSED_TIME, 1930 0, 0, reg); 1931 if (!reg->hr_debug_elapsed_time) { 1932 mlog_errno(ret); 1933 goto bail; 1934 } 1935 1936 reg->hr_debug_pinned = 1937 o2hb_debug_create(O2HB_DEBUG_REGION_PINNED, 1938 reg->hr_debug_dir, 1939 &(reg->hr_db_pinned), 1940 sizeof(*(reg->hr_db_pinned)), 1941 O2HB_DB_TYPE_REGION_PINNED, 1942 0, 0, reg); 1943 if (!reg->hr_debug_pinned) { 1944 mlog_errno(ret); 1945 goto bail; 1946 } 1947 1948 ret = 0; 1949 bail: 1950 return ret; 1951 } 1952 1953 static struct config_item *o2hb_heartbeat_group_make_item(struct config_group *group, 1954 const char *name) 1955 { 1956 struct o2hb_region *reg = NULL; 1957 int ret; 1958 1959 reg = kzalloc(sizeof(struct o2hb_region), GFP_KERNEL); 1960 if (reg == NULL) 1961 return ERR_PTR(-ENOMEM); 1962 1963 if (strlen(name) > O2HB_MAX_REGION_NAME_LEN) { 1964 ret = -ENAMETOOLONG; 1965 goto free; 1966 } 1967 1968 spin_lock(&o2hb_live_lock); 1969 reg->hr_region_num = 0; 1970 if (o2hb_global_heartbeat_active()) { 1971 reg->hr_region_num = find_first_zero_bit(o2hb_region_bitmap, 1972 O2NM_MAX_REGIONS); 1973 if (reg->hr_region_num >= O2NM_MAX_REGIONS) { 1974 spin_unlock(&o2hb_live_lock); 1975 ret = -EFBIG; 1976 goto free; 1977 } 1978 set_bit(reg->hr_region_num, o2hb_region_bitmap); 1979 } 1980 list_add_tail(®->hr_all_item, &o2hb_all_regions); 1981 spin_unlock(&o2hb_live_lock); 1982 1983 config_item_init_type_name(®->hr_item, name, &o2hb_region_type); 1984 1985 ret = o2hb_debug_region_init(reg, o2hb_debug_dir); 1986 if (ret) { 1987 config_item_put(®->hr_item); 1988 goto free; 1989 } 1990 1991 return ®->hr_item; 1992 free: 1993 kfree(reg); 1994 return ERR_PTR(ret); 1995 } 1996 1997 static void o2hb_heartbeat_group_drop_item(struct config_group *group, 1998 struct config_item *item) 1999 { 2000 struct task_struct *hb_task; 2001 struct o2hb_region *reg = to_o2hb_region(item); 2002 int quorum_region = 0; 2003 2004 /* stop the thread when the user removes the region dir */ 2005 spin_lock(&o2hb_live_lock); 2006 hb_task = reg->hr_task; 2007 reg->hr_task = NULL; 2008 reg->hr_item_dropped = 1; 2009 spin_unlock(&o2hb_live_lock); 2010 2011 if (hb_task) 2012 kthread_stop(hb_task); 2013 2014 if (o2hb_global_heartbeat_active()) { 2015 spin_lock(&o2hb_live_lock); 2016 clear_bit(reg->hr_region_num, o2hb_region_bitmap); 2017 clear_bit(reg->hr_region_num, o2hb_live_region_bitmap); 2018 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap)) 2019 quorum_region = 1; 2020 clear_bit(reg->hr_region_num, o2hb_quorum_region_bitmap); 2021 spin_unlock(&o2hb_live_lock); 2022 printk(KERN_NOTICE "o2hb: Heartbeat %s on region %s (%s)\n", 2023 ((atomic_read(®->hr_steady_iterations) == 0) ? 2024 "stopped" : "start aborted"), config_item_name(item), 2025 reg->hr_dev_name); 2026 } 2027 2028 /* 2029 * If we're racing a dev_write(), we need to wake them. They will 2030 * check reg->hr_task 2031 */ 2032 if (atomic_read(®->hr_steady_iterations) != 0) { 2033 reg->hr_aborted_start = 1; 2034 atomic_set(®->hr_steady_iterations, 0); 2035 wake_up(&o2hb_steady_queue); 2036 } 2037 2038 config_item_put(item); 2039 2040 if (!o2hb_global_heartbeat_active() || !quorum_region) 2041 return; 2042 2043 /* 2044 * If global heartbeat active and there are dependent users, 2045 * pin all regions if quorum region count <= CUT_OFF 2046 */ 2047 spin_lock(&o2hb_live_lock); 2048 2049 if (!o2hb_dependent_users) 2050 goto unlock; 2051 2052 if (bitmap_weight(o2hb_quorum_region_bitmap, 2053 O2NM_MAX_REGIONS) <= O2HB_PIN_CUT_OFF) 2054 o2hb_region_pin(NULL); 2055 2056 unlock: 2057 spin_unlock(&o2hb_live_lock); 2058 } 2059 2060 static ssize_t o2hb_heartbeat_group_threshold_show(struct config_item *item, 2061 char *page) 2062 { 2063 return sprintf(page, "%u\n", o2hb_dead_threshold); 2064 } 2065 2066 static ssize_t o2hb_heartbeat_group_threshold_store(struct config_item *item, 2067 const char *page, size_t count) 2068 { 2069 unsigned long tmp; 2070 char *p = (char *)page; 2071 2072 tmp = simple_strtoul(p, &p, 10); 2073 if (!p || (*p && (*p != '\n'))) 2074 return -EINVAL; 2075 2076 /* this will validate ranges for us. */ 2077 o2hb_dead_threshold_set((unsigned int) tmp); 2078 2079 return count; 2080 } 2081 2082 static ssize_t o2hb_heartbeat_group_mode_show(struct config_item *item, 2083 char *page) 2084 { 2085 return sprintf(page, "%s\n", 2086 o2hb_heartbeat_mode_desc[o2hb_heartbeat_mode]); 2087 } 2088 2089 static ssize_t o2hb_heartbeat_group_mode_store(struct config_item *item, 2090 const char *page, size_t count) 2091 { 2092 unsigned int i; 2093 int ret; 2094 size_t len; 2095 2096 len = (page[count - 1] == '\n') ? count - 1 : count; 2097 if (!len) 2098 return -EINVAL; 2099 2100 for (i = 0; i < O2HB_HEARTBEAT_NUM_MODES; ++i) { 2101 if (strncasecmp(page, o2hb_heartbeat_mode_desc[i], len)) 2102 continue; 2103 2104 ret = o2hb_global_heartbeat_mode_set(i); 2105 if (!ret) 2106 printk(KERN_NOTICE "o2hb: Heartbeat mode set to %s\n", 2107 o2hb_heartbeat_mode_desc[i]); 2108 return count; 2109 } 2110 2111 return -EINVAL; 2112 2113 } 2114 2115 CONFIGFS_ATTR(o2hb_heartbeat_group_, threshold); 2116 CONFIGFS_ATTR(o2hb_heartbeat_group_, mode); 2117 2118 static struct configfs_attribute *o2hb_heartbeat_group_attrs[] = { 2119 &o2hb_heartbeat_group_attr_threshold, 2120 &o2hb_heartbeat_group_attr_mode, 2121 NULL, 2122 }; 2123 2124 static struct configfs_group_operations o2hb_heartbeat_group_group_ops = { 2125 .make_item = o2hb_heartbeat_group_make_item, 2126 .drop_item = o2hb_heartbeat_group_drop_item, 2127 }; 2128 2129 static struct config_item_type o2hb_heartbeat_group_type = { 2130 .ct_group_ops = &o2hb_heartbeat_group_group_ops, 2131 .ct_attrs = o2hb_heartbeat_group_attrs, 2132 .ct_owner = THIS_MODULE, 2133 }; 2134 2135 /* this is just here to avoid touching group in heartbeat.h which the 2136 * entire damn world #includes */ 2137 struct config_group *o2hb_alloc_hb_set(void) 2138 { 2139 struct o2hb_heartbeat_group *hs = NULL; 2140 struct config_group *ret = NULL; 2141 2142 hs = kzalloc(sizeof(struct o2hb_heartbeat_group), GFP_KERNEL); 2143 if (hs == NULL) 2144 goto out; 2145 2146 config_group_init_type_name(&hs->hs_group, "heartbeat", 2147 &o2hb_heartbeat_group_type); 2148 2149 ret = &hs->hs_group; 2150 out: 2151 if (ret == NULL) 2152 kfree(hs); 2153 return ret; 2154 } 2155 2156 void o2hb_free_hb_set(struct config_group *group) 2157 { 2158 struct o2hb_heartbeat_group *hs = to_o2hb_heartbeat_group(group); 2159 kfree(hs); 2160 } 2161 2162 /* hb callback registration and issuing */ 2163 2164 static struct o2hb_callback *hbcall_from_type(enum o2hb_callback_type type) 2165 { 2166 if (type == O2HB_NUM_CB) 2167 return ERR_PTR(-EINVAL); 2168 2169 return &o2hb_callbacks[type]; 2170 } 2171 2172 void o2hb_setup_callback(struct o2hb_callback_func *hc, 2173 enum o2hb_callback_type type, 2174 o2hb_cb_func *func, 2175 void *data, 2176 int priority) 2177 { 2178 INIT_LIST_HEAD(&hc->hc_item); 2179 hc->hc_func = func; 2180 hc->hc_data = data; 2181 hc->hc_priority = priority; 2182 hc->hc_type = type; 2183 hc->hc_magic = O2HB_CB_MAGIC; 2184 } 2185 EXPORT_SYMBOL_GPL(o2hb_setup_callback); 2186 2187 /* 2188 * In local heartbeat mode, region_uuid passed matches the dlm domain name. 2189 * In global heartbeat mode, region_uuid passed is NULL. 2190 * 2191 * In local, we only pin the matching region. In global we pin all the active 2192 * regions. 2193 */ 2194 static int o2hb_region_pin(const char *region_uuid) 2195 { 2196 int ret = 0, found = 0; 2197 struct o2hb_region *reg; 2198 char *uuid; 2199 2200 assert_spin_locked(&o2hb_live_lock); 2201 2202 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) { 2203 if (reg->hr_item_dropped) 2204 continue; 2205 2206 uuid = config_item_name(®->hr_item); 2207 2208 /* local heartbeat */ 2209 if (region_uuid) { 2210 if (strcmp(region_uuid, uuid)) 2211 continue; 2212 found = 1; 2213 } 2214 2215 if (reg->hr_item_pinned || reg->hr_item_dropped) 2216 goto skip_pin; 2217 2218 /* Ignore ENOENT only for local hb (userdlm domain) */ 2219 ret = o2nm_depend_item(®->hr_item); 2220 if (!ret) { 2221 mlog(ML_CLUSTER, "Pin region %s\n", uuid); 2222 reg->hr_item_pinned = 1; 2223 } else { 2224 if (ret == -ENOENT && found) 2225 ret = 0; 2226 else { 2227 mlog(ML_ERROR, "Pin region %s fails with %d\n", 2228 uuid, ret); 2229 break; 2230 } 2231 } 2232 skip_pin: 2233 if (found) 2234 break; 2235 } 2236 2237 return ret; 2238 } 2239 2240 /* 2241 * In local heartbeat mode, region_uuid passed matches the dlm domain name. 2242 * In global heartbeat mode, region_uuid passed is NULL. 2243 * 2244 * In local, we only unpin the matching region. In global we unpin all the 2245 * active regions. 2246 */ 2247 static void o2hb_region_unpin(const char *region_uuid) 2248 { 2249 struct o2hb_region *reg; 2250 char *uuid; 2251 int found = 0; 2252 2253 assert_spin_locked(&o2hb_live_lock); 2254 2255 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) { 2256 if (reg->hr_item_dropped) 2257 continue; 2258 2259 uuid = config_item_name(®->hr_item); 2260 if (region_uuid) { 2261 if (strcmp(region_uuid, uuid)) 2262 continue; 2263 found = 1; 2264 } 2265 2266 if (reg->hr_item_pinned) { 2267 mlog(ML_CLUSTER, "Unpin region %s\n", uuid); 2268 o2nm_undepend_item(®->hr_item); 2269 reg->hr_item_pinned = 0; 2270 } 2271 if (found) 2272 break; 2273 } 2274 } 2275 2276 static int o2hb_region_inc_user(const char *region_uuid) 2277 { 2278 int ret = 0; 2279 2280 spin_lock(&o2hb_live_lock); 2281 2282 /* local heartbeat */ 2283 if (!o2hb_global_heartbeat_active()) { 2284 ret = o2hb_region_pin(region_uuid); 2285 goto unlock; 2286 } 2287 2288 /* 2289 * if global heartbeat active and this is the first dependent user, 2290 * pin all regions if quorum region count <= CUT_OFF 2291 */ 2292 o2hb_dependent_users++; 2293 if (o2hb_dependent_users > 1) 2294 goto unlock; 2295 2296 if (bitmap_weight(o2hb_quorum_region_bitmap, 2297 O2NM_MAX_REGIONS) <= O2HB_PIN_CUT_OFF) 2298 ret = o2hb_region_pin(NULL); 2299 2300 unlock: 2301 spin_unlock(&o2hb_live_lock); 2302 return ret; 2303 } 2304 2305 void o2hb_region_dec_user(const char *region_uuid) 2306 { 2307 spin_lock(&o2hb_live_lock); 2308 2309 /* local heartbeat */ 2310 if (!o2hb_global_heartbeat_active()) { 2311 o2hb_region_unpin(region_uuid); 2312 goto unlock; 2313 } 2314 2315 /* 2316 * if global heartbeat active and there are no dependent users, 2317 * unpin all quorum regions 2318 */ 2319 o2hb_dependent_users--; 2320 if (!o2hb_dependent_users) 2321 o2hb_region_unpin(NULL); 2322 2323 unlock: 2324 spin_unlock(&o2hb_live_lock); 2325 } 2326 2327 int o2hb_register_callback(const char *region_uuid, 2328 struct o2hb_callback_func *hc) 2329 { 2330 struct o2hb_callback_func *f; 2331 struct o2hb_callback *hbcall; 2332 int ret; 2333 2334 BUG_ON(hc->hc_magic != O2HB_CB_MAGIC); 2335 BUG_ON(!list_empty(&hc->hc_item)); 2336 2337 hbcall = hbcall_from_type(hc->hc_type); 2338 if (IS_ERR(hbcall)) { 2339 ret = PTR_ERR(hbcall); 2340 goto out; 2341 } 2342 2343 if (region_uuid) { 2344 ret = o2hb_region_inc_user(region_uuid); 2345 if (ret) { 2346 mlog_errno(ret); 2347 goto out; 2348 } 2349 } 2350 2351 down_write(&o2hb_callback_sem); 2352 2353 list_for_each_entry(f, &hbcall->list, hc_item) { 2354 if (hc->hc_priority < f->hc_priority) { 2355 list_add_tail(&hc->hc_item, &f->hc_item); 2356 break; 2357 } 2358 } 2359 if (list_empty(&hc->hc_item)) 2360 list_add_tail(&hc->hc_item, &hbcall->list); 2361 2362 up_write(&o2hb_callback_sem); 2363 ret = 0; 2364 out: 2365 mlog(ML_CLUSTER, "returning %d on behalf of %p for funcs %p\n", 2366 ret, __builtin_return_address(0), hc); 2367 return ret; 2368 } 2369 EXPORT_SYMBOL_GPL(o2hb_register_callback); 2370 2371 void o2hb_unregister_callback(const char *region_uuid, 2372 struct o2hb_callback_func *hc) 2373 { 2374 BUG_ON(hc->hc_magic != O2HB_CB_MAGIC); 2375 2376 mlog(ML_CLUSTER, "on behalf of %p for funcs %p\n", 2377 __builtin_return_address(0), hc); 2378 2379 /* XXX Can this happen _with_ a region reference? */ 2380 if (list_empty(&hc->hc_item)) 2381 return; 2382 2383 if (region_uuid) 2384 o2hb_region_dec_user(region_uuid); 2385 2386 down_write(&o2hb_callback_sem); 2387 2388 list_del_init(&hc->hc_item); 2389 2390 up_write(&o2hb_callback_sem); 2391 } 2392 EXPORT_SYMBOL_GPL(o2hb_unregister_callback); 2393 2394 int o2hb_check_node_heartbeating(u8 node_num) 2395 { 2396 unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; 2397 2398 o2hb_fill_node_map(testing_map, sizeof(testing_map)); 2399 if (!test_bit(node_num, testing_map)) { 2400 mlog(ML_HEARTBEAT, 2401 "node (%u) does not have heartbeating enabled.\n", 2402 node_num); 2403 return 0; 2404 } 2405 2406 return 1; 2407 } 2408 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating); 2409 2410 int o2hb_check_node_heartbeating_no_sem(u8 node_num) 2411 { 2412 unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; 2413 unsigned long flags; 2414 2415 spin_lock_irqsave(&o2hb_live_lock, flags); 2416 o2hb_fill_node_map_from_callback(testing_map, sizeof(testing_map)); 2417 spin_unlock_irqrestore(&o2hb_live_lock, flags); 2418 if (!test_bit(node_num, testing_map)) { 2419 mlog(ML_HEARTBEAT, 2420 "node (%u) does not have heartbeating enabled.\n", 2421 node_num); 2422 return 0; 2423 } 2424 2425 return 1; 2426 } 2427 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating_no_sem); 2428 2429 int o2hb_check_node_heartbeating_from_callback(u8 node_num) 2430 { 2431 unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)]; 2432 2433 o2hb_fill_node_map_from_callback(testing_map, sizeof(testing_map)); 2434 if (!test_bit(node_num, testing_map)) { 2435 mlog(ML_HEARTBEAT, 2436 "node (%u) does not have heartbeating enabled.\n", 2437 node_num); 2438 return 0; 2439 } 2440 2441 return 1; 2442 } 2443 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating_from_callback); 2444 2445 /* Makes sure our local node is configured with a node number, and is 2446 * heartbeating. */ 2447 int o2hb_check_local_node_heartbeating(void) 2448 { 2449 u8 node_num; 2450 2451 /* if this node was set then we have networking */ 2452 node_num = o2nm_this_node(); 2453 if (node_num == O2NM_MAX_NODES) { 2454 mlog(ML_HEARTBEAT, "this node has not been configured.\n"); 2455 return 0; 2456 } 2457 2458 return o2hb_check_node_heartbeating(node_num); 2459 } 2460 EXPORT_SYMBOL_GPL(o2hb_check_local_node_heartbeating); 2461 2462 /* 2463 * this is just a hack until we get the plumbing which flips file systems 2464 * read only and drops the hb ref instead of killing the node dead. 2465 */ 2466 void o2hb_stop_all_regions(void) 2467 { 2468 struct o2hb_region *reg; 2469 2470 mlog(ML_ERROR, "stopping heartbeat on all active regions.\n"); 2471 2472 spin_lock(&o2hb_live_lock); 2473 2474 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) 2475 reg->hr_unclean_stop = 1; 2476 2477 spin_unlock(&o2hb_live_lock); 2478 } 2479 EXPORT_SYMBOL_GPL(o2hb_stop_all_regions); 2480 2481 int o2hb_get_all_regions(char *region_uuids, u8 max_regions) 2482 { 2483 struct o2hb_region *reg; 2484 int numregs = 0; 2485 char *p; 2486 2487 spin_lock(&o2hb_live_lock); 2488 2489 p = region_uuids; 2490 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) { 2491 if (reg->hr_item_dropped) 2492 continue; 2493 2494 mlog(0, "Region: %s\n", config_item_name(®->hr_item)); 2495 if (numregs < max_regions) { 2496 memcpy(p, config_item_name(®->hr_item), 2497 O2HB_MAX_REGION_NAME_LEN); 2498 p += O2HB_MAX_REGION_NAME_LEN; 2499 } 2500 numregs++; 2501 } 2502 2503 spin_unlock(&o2hb_live_lock); 2504 2505 return numregs; 2506 } 2507 EXPORT_SYMBOL_GPL(o2hb_get_all_regions); 2508 2509 int o2hb_global_heartbeat_active(void) 2510 { 2511 return (o2hb_heartbeat_mode == O2HB_HEARTBEAT_GLOBAL); 2512 } 2513 EXPORT_SYMBOL(o2hb_global_heartbeat_active); 2514