xref: /linux/fs/ocfs2/cluster/heartbeat.c (revision f24e9f586b377749dff37554696cf3a105540c94)
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 
37 #include "heartbeat.h"
38 #include "tcp.h"
39 #include "nodemanager.h"
40 #include "quorum.h"
41 
42 #include "masklog.h"
43 
44 
45 /*
46  * The first heartbeat pass had one global thread that would serialize all hb
47  * callback calls.  This global serializing sem should only be removed once
48  * we've made sure that all callees can deal with being called concurrently
49  * from multiple hb region threads.
50  */
51 static DECLARE_RWSEM(o2hb_callback_sem);
52 
53 /*
54  * multiple hb threads are watching multiple regions.  A node is live
55  * whenever any of the threads sees activity from the node in its region.
56  */
57 static DEFINE_SPINLOCK(o2hb_live_lock);
58 static struct list_head o2hb_live_slots[O2NM_MAX_NODES];
59 static unsigned long o2hb_live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
60 static LIST_HEAD(o2hb_node_events);
61 static DECLARE_WAIT_QUEUE_HEAD(o2hb_steady_queue);
62 
63 static LIST_HEAD(o2hb_all_regions);
64 
65 static struct o2hb_callback {
66 	struct list_head list;
67 } o2hb_callbacks[O2HB_NUM_CB];
68 
69 static struct o2hb_callback *hbcall_from_type(enum o2hb_callback_type type);
70 
71 #define O2HB_DEFAULT_BLOCK_BITS       9
72 
73 unsigned int o2hb_dead_threshold = O2HB_DEFAULT_DEAD_THRESHOLD;
74 
75 /* Only sets a new threshold if there are no active regions.
76  *
77  * No locking or otherwise interesting code is required for reading
78  * o2hb_dead_threshold as it can't change once regions are active and
79  * it's not interesting to anyone until then anyway. */
80 static void o2hb_dead_threshold_set(unsigned int threshold)
81 {
82 	if (threshold > O2HB_MIN_DEAD_THRESHOLD) {
83 		spin_lock(&o2hb_live_lock);
84 		if (list_empty(&o2hb_all_regions))
85 			o2hb_dead_threshold = threshold;
86 		spin_unlock(&o2hb_live_lock);
87 	}
88 }
89 
90 struct o2hb_node_event {
91 	struct list_head        hn_item;
92 	enum o2hb_callback_type hn_event_type;
93 	struct o2nm_node        *hn_node;
94 	int                     hn_node_num;
95 };
96 
97 struct o2hb_disk_slot {
98 	struct o2hb_disk_heartbeat_block *ds_raw_block;
99 	u8			ds_node_num;
100 	u64			ds_last_time;
101 	u64			ds_last_generation;
102 	u16			ds_equal_samples;
103 	u16			ds_changed_samples;
104 	struct list_head	ds_live_item;
105 };
106 
107 /* each thread owns a region.. when we're asked to tear down the region
108  * we ask the thread to stop, who cleans up the region */
109 struct o2hb_region {
110 	struct config_item	hr_item;
111 
112 	struct list_head	hr_all_item;
113 	unsigned		hr_unclean_stop:1;
114 
115 	/* protected by the hr_callback_sem */
116 	struct task_struct 	*hr_task;
117 
118 	unsigned int		hr_blocks;
119 	unsigned long long	hr_start_block;
120 
121 	unsigned int		hr_block_bits;
122 	unsigned int		hr_block_bytes;
123 
124 	unsigned int		hr_slots_per_page;
125 	unsigned int		hr_num_pages;
126 
127 	struct page             **hr_slot_data;
128 	struct block_device	*hr_bdev;
129 	struct o2hb_disk_slot	*hr_slots;
130 
131 	/* let the person setting up hb wait for it to return until it
132 	 * has reached a 'steady' state.  This will be fixed when we have
133 	 * a more complete api that doesn't lead to this sort of fragility. */
134 	atomic_t		hr_steady_iterations;
135 
136 	char			hr_dev_name[BDEVNAME_SIZE];
137 
138 	unsigned int		hr_timeout_ms;
139 
140 	/* randomized as the region goes up and down so that a node
141 	 * recognizes a node going up and down in one iteration */
142 	u64			hr_generation;
143 
144 	struct work_struct	hr_write_timeout_work;
145 	unsigned long		hr_last_timeout_start;
146 
147 	/* Used during o2hb_check_slot to hold a copy of the block
148 	 * being checked because we temporarily have to zero out the
149 	 * crc field. */
150 	struct o2hb_disk_heartbeat_block *hr_tmp_block;
151 };
152 
153 struct o2hb_bio_wait_ctxt {
154 	atomic_t          wc_num_reqs;
155 	struct completion wc_io_complete;
156 	int               wc_error;
157 };
158 
159 static void o2hb_write_timeout(void *arg)
160 {
161 	struct o2hb_region *reg = arg;
162 
163 	mlog(ML_ERROR, "Heartbeat write timeout to device %s after %u "
164 	     "milliseconds\n", reg->hr_dev_name,
165 	     jiffies_to_msecs(jiffies - reg->hr_last_timeout_start));
166 	o2quo_disk_timeout();
167 }
168 
169 static void o2hb_arm_write_timeout(struct o2hb_region *reg)
170 {
171 	mlog(0, "Queue write timeout for %u ms\n", O2HB_MAX_WRITE_TIMEOUT_MS);
172 
173 	cancel_delayed_work(&reg->hr_write_timeout_work);
174 	reg->hr_last_timeout_start = jiffies;
175 	schedule_delayed_work(&reg->hr_write_timeout_work,
176 			      msecs_to_jiffies(O2HB_MAX_WRITE_TIMEOUT_MS));
177 }
178 
179 static void o2hb_disarm_write_timeout(struct o2hb_region *reg)
180 {
181 	cancel_delayed_work(&reg->hr_write_timeout_work);
182 	flush_scheduled_work();
183 }
184 
185 static inline void o2hb_bio_wait_init(struct o2hb_bio_wait_ctxt *wc,
186 				      unsigned int num_ios)
187 {
188 	atomic_set(&wc->wc_num_reqs, num_ios);
189 	init_completion(&wc->wc_io_complete);
190 	wc->wc_error = 0;
191 }
192 
193 /* Used in error paths too */
194 static inline void o2hb_bio_wait_dec(struct o2hb_bio_wait_ctxt *wc,
195 				     unsigned int num)
196 {
197 	/* sadly atomic_sub_and_test() isn't available on all platforms.  The
198 	 * good news is that the fast path only completes one at a time */
199 	while(num--) {
200 		if (atomic_dec_and_test(&wc->wc_num_reqs)) {
201 			BUG_ON(num > 0);
202 			complete(&wc->wc_io_complete);
203 		}
204 	}
205 }
206 
207 static void o2hb_wait_on_io(struct o2hb_region *reg,
208 			    struct o2hb_bio_wait_ctxt *wc)
209 {
210 	struct address_space *mapping = reg->hr_bdev->bd_inode->i_mapping;
211 
212 	blk_run_address_space(mapping);
213 
214 	wait_for_completion(&wc->wc_io_complete);
215 }
216 
217 static int o2hb_bio_end_io(struct bio *bio,
218 			   unsigned int bytes_done,
219 			   int error)
220 {
221 	struct o2hb_bio_wait_ctxt *wc = bio->bi_private;
222 
223 	if (error) {
224 		mlog(ML_ERROR, "IO Error %d\n", error);
225 		wc->wc_error = error;
226 	}
227 
228 	if (bio->bi_size)
229 		return 1;
230 
231 	o2hb_bio_wait_dec(wc, 1);
232 	return 0;
233 }
234 
235 /* Setup a Bio to cover I/O against num_slots slots starting at
236  * start_slot. */
237 static struct bio *o2hb_setup_one_bio(struct o2hb_region *reg,
238 				      struct o2hb_bio_wait_ctxt *wc,
239 				      unsigned int start_slot,
240 				      unsigned int num_slots)
241 {
242 	int i, nr_vecs, len, first_page, last_page;
243 	unsigned int vec_len, vec_start;
244 	unsigned int bits = reg->hr_block_bits;
245 	unsigned int spp = reg->hr_slots_per_page;
246 	struct bio *bio;
247 	struct page *page;
248 
249 	nr_vecs = (num_slots + spp - 1) / spp;
250 
251 	/* Testing has shown this allocation to take long enough under
252 	 * GFP_KERNEL that the local node can get fenced. It would be
253 	 * nicest if we could pre-allocate these bios and avoid this
254 	 * all together. */
255 	bio = bio_alloc(GFP_ATOMIC, nr_vecs);
256 	if (!bio) {
257 		mlog(ML_ERROR, "Could not alloc slots BIO!\n");
258 		bio = ERR_PTR(-ENOMEM);
259 		goto bail;
260 	}
261 
262 	/* Must put everything in 512 byte sectors for the bio... */
263 	bio->bi_sector = (reg->hr_start_block + start_slot) << (bits - 9);
264 	bio->bi_bdev = reg->hr_bdev;
265 	bio->bi_private = wc;
266 	bio->bi_end_io = o2hb_bio_end_io;
267 
268 	first_page = start_slot / spp;
269 	last_page = first_page + nr_vecs;
270 	vec_start = (start_slot << bits) % PAGE_CACHE_SIZE;
271 	for(i = first_page; i < last_page; i++) {
272 		page = reg->hr_slot_data[i];
273 
274 		vec_len = PAGE_CACHE_SIZE;
275 		/* last page might be short */
276 		if (((i + 1) * spp) > (start_slot + num_slots))
277 			vec_len = ((num_slots + start_slot) % spp) << bits;
278 		vec_len -=  vec_start;
279 
280 		mlog(ML_HB_BIO, "page %d, vec_len = %u, vec_start = %u\n",
281 		     i, vec_len, vec_start);
282 
283 		len = bio_add_page(bio, page, vec_len, vec_start);
284 		if (len != vec_len) {
285 			bio_put(bio);
286 			bio = ERR_PTR(-EIO);
287 
288 			mlog(ML_ERROR, "Error adding page to bio i = %d, "
289 			     "vec_len = %u, len = %d\n, start = %u\n",
290 			     i, vec_len, len, vec_start);
291 			goto bail;
292 		}
293 
294 		vec_start = 0;
295 	}
296 
297 bail:
298 	return bio;
299 }
300 
301 /*
302  * Compute the maximum number of sectors the bdev can handle in one bio,
303  * as a power of two.
304  *
305  * Stolen from oracleasm, thanks Joel!
306  */
307 static int compute_max_sectors(struct block_device *bdev)
308 {
309 	int max_pages, max_sectors, pow_two_sectors;
310 
311 	struct request_queue *q;
312 
313 	q = bdev_get_queue(bdev);
314 	max_pages = q->max_sectors >> (PAGE_SHIFT - 9);
315 	if (max_pages > BIO_MAX_PAGES)
316 		max_pages = BIO_MAX_PAGES;
317 	if (max_pages > q->max_phys_segments)
318 		max_pages = q->max_phys_segments;
319 	if (max_pages > q->max_hw_segments)
320 		max_pages = q->max_hw_segments;
321 	max_pages--; /* Handle I/Os that straddle a page */
322 
323 	if (max_pages) {
324 		max_sectors = max_pages << (PAGE_SHIFT - 9);
325 	} else {
326 		/* If BIO contains 1 or less than 1 page. */
327 		max_sectors = q->max_sectors;
328 	}
329 	/* Why is fls() 1-based???? */
330 	pow_two_sectors = 1 << (fls(max_sectors) - 1);
331 
332 	return pow_two_sectors;
333 }
334 
335 static inline void o2hb_compute_request_limits(struct o2hb_region *reg,
336 					       unsigned int num_slots,
337 					       unsigned int *num_bios,
338 					       unsigned int *slots_per_bio)
339 {
340 	unsigned int max_sectors, io_sectors;
341 
342 	max_sectors = compute_max_sectors(reg->hr_bdev);
343 
344 	io_sectors = num_slots << (reg->hr_block_bits - 9);
345 
346 	*num_bios = (io_sectors + max_sectors - 1) / max_sectors;
347 	*slots_per_bio = max_sectors >> (reg->hr_block_bits - 9);
348 
349 	mlog(ML_HB_BIO, "My io size is %u sectors for %u slots. This "
350 	     "device can handle %u sectors of I/O\n", io_sectors, num_slots,
351 	     max_sectors);
352 	mlog(ML_HB_BIO, "Will need %u bios holding %u slots each\n",
353 	     *num_bios, *slots_per_bio);
354 }
355 
356 static int o2hb_read_slots(struct o2hb_region *reg,
357 			   unsigned int max_slots)
358 {
359 	unsigned int num_bios, slots_per_bio, start_slot, num_slots;
360 	int i, status;
361 	struct o2hb_bio_wait_ctxt wc;
362 	struct bio **bios;
363 	struct bio *bio;
364 
365 	o2hb_compute_request_limits(reg, max_slots, &num_bios, &slots_per_bio);
366 
367 	bios = kcalloc(num_bios, sizeof(struct bio *), GFP_KERNEL);
368 	if (!bios) {
369 		status = -ENOMEM;
370 		mlog_errno(status);
371 		return status;
372 	}
373 
374 	o2hb_bio_wait_init(&wc, num_bios);
375 
376 	num_slots = slots_per_bio;
377 	for(i = 0; i < num_bios; i++) {
378 		start_slot = i * slots_per_bio;
379 
380 		/* adjust num_slots at last bio */
381 		if (max_slots < (start_slot + num_slots))
382 			num_slots = max_slots - start_slot;
383 
384 		bio = o2hb_setup_one_bio(reg, &wc, start_slot, num_slots);
385 		if (IS_ERR(bio)) {
386 			o2hb_bio_wait_dec(&wc, num_bios - i);
387 
388 			status = PTR_ERR(bio);
389 			mlog_errno(status);
390 			goto bail_and_wait;
391 		}
392 		bios[i] = bio;
393 
394 		submit_bio(READ, bio);
395 	}
396 
397 	status = 0;
398 
399 bail_and_wait:
400 	o2hb_wait_on_io(reg, &wc);
401 	if (wc.wc_error && !status)
402 		status = wc.wc_error;
403 
404 	if (bios) {
405 		for(i = 0; i < num_bios; i++)
406 			if (bios[i])
407 				bio_put(bios[i]);
408 		kfree(bios);
409 	}
410 
411 	return status;
412 }
413 
414 static int o2hb_issue_node_write(struct o2hb_region *reg,
415 				 struct bio **write_bio,
416 				 struct o2hb_bio_wait_ctxt *write_wc)
417 {
418 	int status;
419 	unsigned int slot;
420 	struct bio *bio;
421 
422 	o2hb_bio_wait_init(write_wc, 1);
423 
424 	slot = o2nm_this_node();
425 
426 	bio = o2hb_setup_one_bio(reg, write_wc, slot, 1);
427 	if (IS_ERR(bio)) {
428 		status = PTR_ERR(bio);
429 		mlog_errno(status);
430 		goto bail;
431 	}
432 
433 	submit_bio(WRITE, bio);
434 
435 	*write_bio = bio;
436 	status = 0;
437 bail:
438 	return status;
439 }
440 
441 static u32 o2hb_compute_block_crc_le(struct o2hb_region *reg,
442 				     struct o2hb_disk_heartbeat_block *hb_block)
443 {
444 	__le32 old_cksum;
445 	u32 ret;
446 
447 	/* We want to compute the block crc with a 0 value in the
448 	 * hb_cksum field. Save it off here and replace after the
449 	 * crc. */
450 	old_cksum = hb_block->hb_cksum;
451 	hb_block->hb_cksum = 0;
452 
453 	ret = crc32_le(0, (unsigned char *) hb_block, reg->hr_block_bytes);
454 
455 	hb_block->hb_cksum = old_cksum;
456 
457 	return ret;
458 }
459 
460 static void o2hb_dump_slot(struct o2hb_disk_heartbeat_block *hb_block)
461 {
462 	mlog(ML_ERROR, "Dump slot information: seq = 0x%llx, node = %u, "
463 	     "cksum = 0x%x, generation 0x%llx\n",
464 	     (long long)le64_to_cpu(hb_block->hb_seq),
465 	     hb_block->hb_node, le32_to_cpu(hb_block->hb_cksum),
466 	     (long long)le64_to_cpu(hb_block->hb_generation));
467 }
468 
469 static int o2hb_verify_crc(struct o2hb_region *reg,
470 			   struct o2hb_disk_heartbeat_block *hb_block)
471 {
472 	u32 read, computed;
473 
474 	read = le32_to_cpu(hb_block->hb_cksum);
475 	computed = o2hb_compute_block_crc_le(reg, hb_block);
476 
477 	return read == computed;
478 }
479 
480 /* We want to make sure that nobody is heartbeating on top of us --
481  * this will help detect an invalid configuration. */
482 static int o2hb_check_last_timestamp(struct o2hb_region *reg)
483 {
484 	int node_num, ret;
485 	struct o2hb_disk_slot *slot;
486 	struct o2hb_disk_heartbeat_block *hb_block;
487 
488 	node_num = o2nm_this_node();
489 
490 	ret = 1;
491 	slot = &reg->hr_slots[node_num];
492 	/* Don't check on our 1st timestamp */
493 	if (slot->ds_last_time) {
494 		hb_block = slot->ds_raw_block;
495 
496 		if (le64_to_cpu(hb_block->hb_seq) != slot->ds_last_time)
497 			ret = 0;
498 	}
499 
500 	return ret;
501 }
502 
503 static inline void o2hb_prepare_block(struct o2hb_region *reg,
504 				      u64 generation)
505 {
506 	int node_num;
507 	u64 cputime;
508 	struct o2hb_disk_slot *slot;
509 	struct o2hb_disk_heartbeat_block *hb_block;
510 
511 	node_num = o2nm_this_node();
512 	slot = &reg->hr_slots[node_num];
513 
514 	hb_block = (struct o2hb_disk_heartbeat_block *)slot->ds_raw_block;
515 	memset(hb_block, 0, reg->hr_block_bytes);
516 	/* TODO: time stuff */
517 	cputime = CURRENT_TIME.tv_sec;
518 	if (!cputime)
519 		cputime = 1;
520 
521 	hb_block->hb_seq = cpu_to_le64(cputime);
522 	hb_block->hb_node = node_num;
523 	hb_block->hb_generation = cpu_to_le64(generation);
524 	hb_block->hb_dead_ms = cpu_to_le32(o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS);
525 
526 	/* This step must always happen last! */
527 	hb_block->hb_cksum = cpu_to_le32(o2hb_compute_block_crc_le(reg,
528 								   hb_block));
529 
530 	mlog(ML_HB_BIO, "our node generation = 0x%llx, cksum = 0x%x\n",
531 	     (long long)cpu_to_le64(generation),
532 	     le32_to_cpu(hb_block->hb_cksum));
533 }
534 
535 static void o2hb_fire_callbacks(struct o2hb_callback *hbcall,
536 				struct o2nm_node *node,
537 				int idx)
538 {
539 	struct list_head *iter;
540 	struct o2hb_callback_func *f;
541 
542 	list_for_each(iter, &hbcall->list) {
543 		f = list_entry(iter, struct o2hb_callback_func, hc_item);
544 		mlog(ML_HEARTBEAT, "calling funcs %p\n", f);
545 		(f->hc_func)(node, idx, f->hc_data);
546 	}
547 }
548 
549 /* Will run the list in order until we process the passed event */
550 static void o2hb_run_event_list(struct o2hb_node_event *queued_event)
551 {
552 	int empty;
553 	struct o2hb_callback *hbcall;
554 	struct o2hb_node_event *event;
555 
556 	spin_lock(&o2hb_live_lock);
557 	empty = list_empty(&queued_event->hn_item);
558 	spin_unlock(&o2hb_live_lock);
559 	if (empty)
560 		return;
561 
562 	/* Holding callback sem assures we don't alter the callback
563 	 * lists when doing this, and serializes ourselves with other
564 	 * processes wanting callbacks. */
565 	down_write(&o2hb_callback_sem);
566 
567 	spin_lock(&o2hb_live_lock);
568 	while (!list_empty(&o2hb_node_events)
569 	       && !list_empty(&queued_event->hn_item)) {
570 		event = list_entry(o2hb_node_events.next,
571 				   struct o2hb_node_event,
572 				   hn_item);
573 		list_del_init(&event->hn_item);
574 		spin_unlock(&o2hb_live_lock);
575 
576 		mlog(ML_HEARTBEAT, "Node %s event for %d\n",
577 		     event->hn_event_type == O2HB_NODE_UP_CB ? "UP" : "DOWN",
578 		     event->hn_node_num);
579 
580 		hbcall = hbcall_from_type(event->hn_event_type);
581 
582 		/* We should *never* have gotten on to the list with a
583 		 * bad type... This isn't something that we should try
584 		 * to recover from. */
585 		BUG_ON(IS_ERR(hbcall));
586 
587 		o2hb_fire_callbacks(hbcall, event->hn_node, event->hn_node_num);
588 
589 		spin_lock(&o2hb_live_lock);
590 	}
591 	spin_unlock(&o2hb_live_lock);
592 
593 	up_write(&o2hb_callback_sem);
594 }
595 
596 static void o2hb_queue_node_event(struct o2hb_node_event *event,
597 				  enum o2hb_callback_type type,
598 				  struct o2nm_node *node,
599 				  int node_num)
600 {
601 	assert_spin_locked(&o2hb_live_lock);
602 
603 	event->hn_event_type = type;
604 	event->hn_node = node;
605 	event->hn_node_num = node_num;
606 
607 	mlog(ML_HEARTBEAT, "Queue node %s event for node %d\n",
608 	     type == O2HB_NODE_UP_CB ? "UP" : "DOWN", node_num);
609 
610 	list_add_tail(&event->hn_item, &o2hb_node_events);
611 }
612 
613 static void o2hb_shutdown_slot(struct o2hb_disk_slot *slot)
614 {
615 	struct o2hb_node_event event =
616 		{ .hn_item = LIST_HEAD_INIT(event.hn_item), };
617 	struct o2nm_node *node;
618 
619 	node = o2nm_get_node_by_num(slot->ds_node_num);
620 	if (!node)
621 		return;
622 
623 	spin_lock(&o2hb_live_lock);
624 	if (!list_empty(&slot->ds_live_item)) {
625 		mlog(ML_HEARTBEAT, "Shutdown, node %d leaves region\n",
626 		     slot->ds_node_num);
627 
628 		list_del_init(&slot->ds_live_item);
629 
630 		if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
631 			clear_bit(slot->ds_node_num, o2hb_live_node_bitmap);
632 
633 			o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB, node,
634 					      slot->ds_node_num);
635 		}
636 	}
637 	spin_unlock(&o2hb_live_lock);
638 
639 	o2hb_run_event_list(&event);
640 
641 	o2nm_node_put(node);
642 }
643 
644 static int o2hb_check_slot(struct o2hb_region *reg,
645 			   struct o2hb_disk_slot *slot)
646 {
647 	int changed = 0, gen_changed = 0;
648 	struct o2hb_node_event event =
649 		{ .hn_item = LIST_HEAD_INIT(event.hn_item), };
650 	struct o2nm_node *node;
651 	struct o2hb_disk_heartbeat_block *hb_block = reg->hr_tmp_block;
652 	u64 cputime;
653 	unsigned int dead_ms = o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS;
654 	unsigned int slot_dead_ms;
655 
656 	memcpy(hb_block, slot->ds_raw_block, reg->hr_block_bytes);
657 
658 	/* Is this correct? Do we assume that the node doesn't exist
659 	 * if we're not configured for him? */
660 	node = o2nm_get_node_by_num(slot->ds_node_num);
661 	if (!node)
662 		return 0;
663 
664 	if (!o2hb_verify_crc(reg, hb_block)) {
665 		/* all paths from here will drop o2hb_live_lock for
666 		 * us. */
667 		spin_lock(&o2hb_live_lock);
668 
669 		/* Don't print an error on the console in this case -
670 		 * a freshly formatted heartbeat area will not have a
671 		 * crc set on it. */
672 		if (list_empty(&slot->ds_live_item))
673 			goto out;
674 
675 		/* The node is live but pushed out a bad crc. We
676 		 * consider it a transient miss but don't populate any
677 		 * other values as they may be junk. */
678 		mlog(ML_ERROR, "Node %d has written a bad crc to %s\n",
679 		     slot->ds_node_num, reg->hr_dev_name);
680 		o2hb_dump_slot(hb_block);
681 
682 		slot->ds_equal_samples++;
683 		goto fire_callbacks;
684 	}
685 
686 	/* we don't care if these wrap.. the state transitions below
687 	 * clear at the right places */
688 	cputime = le64_to_cpu(hb_block->hb_seq);
689 	if (slot->ds_last_time != cputime)
690 		slot->ds_changed_samples++;
691 	else
692 		slot->ds_equal_samples++;
693 	slot->ds_last_time = cputime;
694 
695 	/* The node changed heartbeat generations. We assume this to
696 	 * mean it dropped off but came back before we timed out. We
697 	 * want to consider it down for the time being but don't want
698 	 * to lose any changed_samples state we might build up to
699 	 * considering it live again. */
700 	if (slot->ds_last_generation != le64_to_cpu(hb_block->hb_generation)) {
701 		gen_changed = 1;
702 		slot->ds_equal_samples = 0;
703 		mlog(ML_HEARTBEAT, "Node %d changed generation (0x%llx "
704 		     "to 0x%llx)\n", slot->ds_node_num,
705 		     (long long)slot->ds_last_generation,
706 		     (long long)le64_to_cpu(hb_block->hb_generation));
707 	}
708 
709 	slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation);
710 
711 	mlog(ML_HEARTBEAT, "Slot %d gen 0x%llx cksum 0x%x "
712 	     "seq %llu last %llu changed %u equal %u\n",
713 	     slot->ds_node_num, (long long)slot->ds_last_generation,
714 	     le32_to_cpu(hb_block->hb_cksum),
715 	     (unsigned long long)le64_to_cpu(hb_block->hb_seq),
716 	     (unsigned long long)slot->ds_last_time, slot->ds_changed_samples,
717 	     slot->ds_equal_samples);
718 
719 	spin_lock(&o2hb_live_lock);
720 
721 fire_callbacks:
722 	/* dead nodes only come to life after some number of
723 	 * changes at any time during their dead time */
724 	if (list_empty(&slot->ds_live_item) &&
725 	    slot->ds_changed_samples >= O2HB_LIVE_THRESHOLD) {
726 		mlog(ML_HEARTBEAT, "Node %d (id 0x%llx) joined my region\n",
727 		     slot->ds_node_num, (long long)slot->ds_last_generation);
728 
729 		/* first on the list generates a callback */
730 		if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
731 			set_bit(slot->ds_node_num, o2hb_live_node_bitmap);
732 
733 			o2hb_queue_node_event(&event, O2HB_NODE_UP_CB, node,
734 					      slot->ds_node_num);
735 
736 			changed = 1;
737 		}
738 
739 		list_add_tail(&slot->ds_live_item,
740 			      &o2hb_live_slots[slot->ds_node_num]);
741 
742 		slot->ds_equal_samples = 0;
743 
744 		/* We want to be sure that all nodes agree on the
745 		 * number of milliseconds before a node will be
746 		 * considered dead. The self-fencing timeout is
747 		 * computed from this value, and a discrepancy might
748 		 * result in heartbeat calling a node dead when it
749 		 * hasn't self-fenced yet. */
750 		slot_dead_ms = le32_to_cpu(hb_block->hb_dead_ms);
751 		if (slot_dead_ms && slot_dead_ms != dead_ms) {
752 			/* TODO: Perhaps we can fail the region here. */
753 			mlog(ML_ERROR, "Node %d on device %s has a dead count "
754 			     "of %u ms, but our count is %u ms.\n"
755 			     "Please double check your configuration values "
756 			     "for 'O2CB_HEARTBEAT_THRESHOLD'\n",
757 			     slot->ds_node_num, reg->hr_dev_name, slot_dead_ms,
758 			     dead_ms);
759 		}
760 		goto out;
761 	}
762 
763 	/* if the list is dead, we're done.. */
764 	if (list_empty(&slot->ds_live_item))
765 		goto out;
766 
767 	/* live nodes only go dead after enough consequtive missed
768 	 * samples..  reset the missed counter whenever we see
769 	 * activity */
770 	if (slot->ds_equal_samples >= o2hb_dead_threshold || gen_changed) {
771 		mlog(ML_HEARTBEAT, "Node %d left my region\n",
772 		     slot->ds_node_num);
773 
774 		/* last off the live_slot generates a callback */
775 		list_del_init(&slot->ds_live_item);
776 		if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
777 			clear_bit(slot->ds_node_num, o2hb_live_node_bitmap);
778 
779 			o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB, node,
780 					      slot->ds_node_num);
781 
782 			changed = 1;
783 		}
784 
785 		/* We don't clear this because the node is still
786 		 * actually writing new blocks. */
787 		if (!gen_changed)
788 			slot->ds_changed_samples = 0;
789 		goto out;
790 	}
791 	if (slot->ds_changed_samples) {
792 		slot->ds_changed_samples = 0;
793 		slot->ds_equal_samples = 0;
794 	}
795 out:
796 	spin_unlock(&o2hb_live_lock);
797 
798 	o2hb_run_event_list(&event);
799 
800 	o2nm_node_put(node);
801 	return changed;
802 }
803 
804 /* This could be faster if we just implmented a find_last_bit, but I
805  * don't think the circumstances warrant it. */
806 static int o2hb_highest_node(unsigned long *nodes,
807 			     int numbits)
808 {
809 	int highest, node;
810 
811 	highest = numbits;
812 	node = -1;
813 	while ((node = find_next_bit(nodes, numbits, node + 1)) != -1) {
814 		if (node >= numbits)
815 			break;
816 
817 		highest = node;
818 	}
819 
820 	return highest;
821 }
822 
823 static int o2hb_do_disk_heartbeat(struct o2hb_region *reg)
824 {
825 	int i, ret, highest_node, change = 0;
826 	unsigned long configured_nodes[BITS_TO_LONGS(O2NM_MAX_NODES)];
827 	struct bio *write_bio;
828 	struct o2hb_bio_wait_ctxt write_wc;
829 
830 	ret = o2nm_configured_node_map(configured_nodes,
831 				       sizeof(configured_nodes));
832 	if (ret) {
833 		mlog_errno(ret);
834 		return ret;
835 	}
836 
837 	highest_node = o2hb_highest_node(configured_nodes, O2NM_MAX_NODES);
838 	if (highest_node >= O2NM_MAX_NODES) {
839 		mlog(ML_NOTICE, "ocfs2_heartbeat: no configured nodes found!\n");
840 		return -EINVAL;
841 	}
842 
843 	/* No sense in reading the slots of nodes that don't exist
844 	 * yet. Of course, if the node definitions have holes in them
845 	 * then we're reading an empty slot anyway... Consider this
846 	 * best-effort. */
847 	ret = o2hb_read_slots(reg, highest_node + 1);
848 	if (ret < 0) {
849 		mlog_errno(ret);
850 		return ret;
851 	}
852 
853 	/* With an up to date view of the slots, we can check that no
854 	 * other node has been improperly configured to heartbeat in
855 	 * our slot. */
856 	if (!o2hb_check_last_timestamp(reg))
857 		mlog(ML_ERROR, "Device \"%s\": another node is heartbeating "
858 		     "in our slot!\n", reg->hr_dev_name);
859 
860 	/* fill in the proper info for our next heartbeat */
861 	o2hb_prepare_block(reg, reg->hr_generation);
862 
863 	/* And fire off the write. Note that we don't wait on this I/O
864 	 * until later. */
865 	ret = o2hb_issue_node_write(reg, &write_bio, &write_wc);
866 	if (ret < 0) {
867 		mlog_errno(ret);
868 		return ret;
869 	}
870 
871 	i = -1;
872 	while((i = find_next_bit(configured_nodes, O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) {
873 
874 		change |= o2hb_check_slot(reg, &reg->hr_slots[i]);
875 	}
876 
877 	/*
878 	 * We have to be sure we've advertised ourselves on disk
879 	 * before we can go to steady state.  This ensures that
880 	 * people we find in our steady state have seen us.
881 	 */
882 	o2hb_wait_on_io(reg, &write_wc);
883 	bio_put(write_bio);
884 	if (write_wc.wc_error) {
885 		/* Do not re-arm the write timeout on I/O error - we
886 		 * can't be sure that the new block ever made it to
887 		 * disk */
888 		mlog(ML_ERROR, "Write error %d on device \"%s\"\n",
889 		     write_wc.wc_error, reg->hr_dev_name);
890 		return write_wc.wc_error;
891 	}
892 
893 	o2hb_arm_write_timeout(reg);
894 
895 	/* let the person who launched us know when things are steady */
896 	if (!change && (atomic_read(&reg->hr_steady_iterations) != 0)) {
897 		if (atomic_dec_and_test(&reg->hr_steady_iterations))
898 			wake_up(&o2hb_steady_queue);
899 	}
900 
901 	return 0;
902 }
903 
904 /* Subtract b from a, storing the result in a. a *must* have a larger
905  * value than b. */
906 static void o2hb_tv_subtract(struct timeval *a,
907 			     struct timeval *b)
908 {
909 	/* just return 0 when a is after b */
910 	if (a->tv_sec < b->tv_sec ||
911 	    (a->tv_sec == b->tv_sec && a->tv_usec < b->tv_usec)) {
912 		a->tv_sec = 0;
913 		a->tv_usec = 0;
914 		return;
915 	}
916 
917 	a->tv_sec -= b->tv_sec;
918 	a->tv_usec -= b->tv_usec;
919 	while ( a->tv_usec < 0 ) {
920 		a->tv_sec--;
921 		a->tv_usec += 1000000;
922 	}
923 }
924 
925 static unsigned int o2hb_elapsed_msecs(struct timeval *start,
926 				       struct timeval *end)
927 {
928 	struct timeval res = *end;
929 
930 	o2hb_tv_subtract(&res, start);
931 
932 	return res.tv_sec * 1000 + res.tv_usec / 1000;
933 }
934 
935 /*
936  * we ride the region ref that the region dir holds.  before the region
937  * dir is removed and drops it ref it will wait to tear down this
938  * thread.
939  */
940 static int o2hb_thread(void *data)
941 {
942 	int i, ret;
943 	struct o2hb_region *reg = data;
944 	struct bio *write_bio;
945 	struct o2hb_bio_wait_ctxt write_wc;
946 	struct timeval before_hb, after_hb;
947 	unsigned int elapsed_msec;
948 
949 	mlog(ML_HEARTBEAT|ML_KTHREAD, "hb thread running\n");
950 
951 	set_user_nice(current, -20);
952 
953 	while (!kthread_should_stop() && !reg->hr_unclean_stop) {
954 		/* We track the time spent inside
955 		 * o2hb_do_disk_heartbeat so that we avoid more then
956 		 * hr_timeout_ms between disk writes. On busy systems
957 		 * this should result in a heartbeat which is less
958 		 * likely to time itself out. */
959 		do_gettimeofday(&before_hb);
960 
961 		i = 0;
962 		do {
963 			ret = o2hb_do_disk_heartbeat(reg);
964 		} while (ret && ++i < 2);
965 
966 		do_gettimeofday(&after_hb);
967 		elapsed_msec = o2hb_elapsed_msecs(&before_hb, &after_hb);
968 
969 		mlog(0, "start = %lu.%lu, end = %lu.%lu, msec = %u\n",
970 		     before_hb.tv_sec, (unsigned long) before_hb.tv_usec,
971 		     after_hb.tv_sec, (unsigned long) after_hb.tv_usec,
972 		     elapsed_msec);
973 
974 		if (elapsed_msec < reg->hr_timeout_ms) {
975 			/* the kthread api has blocked signals for us so no
976 			 * need to record the return value. */
977 			msleep_interruptible(reg->hr_timeout_ms - elapsed_msec);
978 		}
979 	}
980 
981 	o2hb_disarm_write_timeout(reg);
982 
983 	/* unclean stop is only used in very bad situation */
984 	for(i = 0; !reg->hr_unclean_stop && i < reg->hr_blocks; i++)
985 		o2hb_shutdown_slot(&reg->hr_slots[i]);
986 
987 	/* Explicit down notification - avoid forcing the other nodes
988 	 * to timeout on this region when we could just as easily
989 	 * write a clear generation - thus indicating to them that
990 	 * this node has left this region.
991 	 *
992 	 * XXX: Should we skip this on unclean_stop? */
993 	o2hb_prepare_block(reg, 0);
994 	ret = o2hb_issue_node_write(reg, &write_bio, &write_wc);
995 	if (ret == 0) {
996 		o2hb_wait_on_io(reg, &write_wc);
997 		bio_put(write_bio);
998 	} else {
999 		mlog_errno(ret);
1000 	}
1001 
1002 	mlog(ML_HEARTBEAT|ML_KTHREAD, "hb thread exiting\n");
1003 
1004 	return 0;
1005 }
1006 
1007 void o2hb_init(void)
1008 {
1009 	int i;
1010 
1011 	for (i = 0; i < ARRAY_SIZE(o2hb_callbacks); i++)
1012 		INIT_LIST_HEAD(&o2hb_callbacks[i].list);
1013 
1014 	for (i = 0; i < ARRAY_SIZE(o2hb_live_slots); i++)
1015 		INIT_LIST_HEAD(&o2hb_live_slots[i]);
1016 
1017 	INIT_LIST_HEAD(&o2hb_node_events);
1018 
1019 	memset(o2hb_live_node_bitmap, 0, sizeof(o2hb_live_node_bitmap));
1020 }
1021 
1022 /* if we're already in a callback then we're already serialized by the sem */
1023 static void o2hb_fill_node_map_from_callback(unsigned long *map,
1024 					     unsigned bytes)
1025 {
1026 	BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1027 
1028 	memcpy(map, &o2hb_live_node_bitmap, bytes);
1029 }
1030 
1031 /*
1032  * get a map of all nodes that are heartbeating in any regions
1033  */
1034 void o2hb_fill_node_map(unsigned long *map, unsigned bytes)
1035 {
1036 	/* callers want to serialize this map and callbacks so that they
1037 	 * can trust that they don't miss nodes coming to the party */
1038 	down_read(&o2hb_callback_sem);
1039 	spin_lock(&o2hb_live_lock);
1040 	o2hb_fill_node_map_from_callback(map, bytes);
1041 	spin_unlock(&o2hb_live_lock);
1042 	up_read(&o2hb_callback_sem);
1043 }
1044 EXPORT_SYMBOL_GPL(o2hb_fill_node_map);
1045 
1046 /*
1047  * heartbeat configfs bits.  The heartbeat set is a default set under
1048  * the cluster set in nodemanager.c.
1049  */
1050 
1051 static struct o2hb_region *to_o2hb_region(struct config_item *item)
1052 {
1053 	return item ? container_of(item, struct o2hb_region, hr_item) : NULL;
1054 }
1055 
1056 /* drop_item only drops its ref after killing the thread, nothing should
1057  * be using the region anymore.  this has to clean up any state that
1058  * attributes might have built up. */
1059 static void o2hb_region_release(struct config_item *item)
1060 {
1061 	int i;
1062 	struct page *page;
1063 	struct o2hb_region *reg = to_o2hb_region(item);
1064 
1065 	if (reg->hr_tmp_block)
1066 		kfree(reg->hr_tmp_block);
1067 
1068 	if (reg->hr_slot_data) {
1069 		for (i = 0; i < reg->hr_num_pages; i++) {
1070 			page = reg->hr_slot_data[i];
1071 			if (page)
1072 				__free_page(page);
1073 		}
1074 		kfree(reg->hr_slot_data);
1075 	}
1076 
1077 	if (reg->hr_bdev)
1078 		blkdev_put(reg->hr_bdev);
1079 
1080 	if (reg->hr_slots)
1081 		kfree(reg->hr_slots);
1082 
1083 	spin_lock(&o2hb_live_lock);
1084 	list_del(&reg->hr_all_item);
1085 	spin_unlock(&o2hb_live_lock);
1086 
1087 	kfree(reg);
1088 }
1089 
1090 static int o2hb_read_block_input(struct o2hb_region *reg,
1091 				 const char *page,
1092 				 size_t count,
1093 				 unsigned long *ret_bytes,
1094 				 unsigned int *ret_bits)
1095 {
1096 	unsigned long bytes;
1097 	char *p = (char *)page;
1098 
1099 	bytes = simple_strtoul(p, &p, 0);
1100 	if (!p || (*p && (*p != '\n')))
1101 		return -EINVAL;
1102 
1103 	/* Heartbeat and fs min / max block sizes are the same. */
1104 	if (bytes > 4096 || bytes < 512)
1105 		return -ERANGE;
1106 	if (hweight16(bytes) != 1)
1107 		return -EINVAL;
1108 
1109 	if (ret_bytes)
1110 		*ret_bytes = bytes;
1111 	if (ret_bits)
1112 		*ret_bits = ffs(bytes) - 1;
1113 
1114 	return 0;
1115 }
1116 
1117 static ssize_t o2hb_region_block_bytes_read(struct o2hb_region *reg,
1118 					    char *page)
1119 {
1120 	return sprintf(page, "%u\n", reg->hr_block_bytes);
1121 }
1122 
1123 static ssize_t o2hb_region_block_bytes_write(struct o2hb_region *reg,
1124 					     const char *page,
1125 					     size_t count)
1126 {
1127 	int status;
1128 	unsigned long block_bytes;
1129 	unsigned int block_bits;
1130 
1131 	if (reg->hr_bdev)
1132 		return -EINVAL;
1133 
1134 	status = o2hb_read_block_input(reg, page, count,
1135 				       &block_bytes, &block_bits);
1136 	if (status)
1137 		return status;
1138 
1139 	reg->hr_block_bytes = (unsigned int)block_bytes;
1140 	reg->hr_block_bits = block_bits;
1141 
1142 	return count;
1143 }
1144 
1145 static ssize_t o2hb_region_start_block_read(struct o2hb_region *reg,
1146 					    char *page)
1147 {
1148 	return sprintf(page, "%llu\n", reg->hr_start_block);
1149 }
1150 
1151 static ssize_t o2hb_region_start_block_write(struct o2hb_region *reg,
1152 					     const char *page,
1153 					     size_t count)
1154 {
1155 	unsigned long long tmp;
1156 	char *p = (char *)page;
1157 
1158 	if (reg->hr_bdev)
1159 		return -EINVAL;
1160 
1161 	tmp = simple_strtoull(p, &p, 0);
1162 	if (!p || (*p && (*p != '\n')))
1163 		return -EINVAL;
1164 
1165 	reg->hr_start_block = tmp;
1166 
1167 	return count;
1168 }
1169 
1170 static ssize_t o2hb_region_blocks_read(struct o2hb_region *reg,
1171 				       char *page)
1172 {
1173 	return sprintf(page, "%d\n", reg->hr_blocks);
1174 }
1175 
1176 static ssize_t o2hb_region_blocks_write(struct o2hb_region *reg,
1177 					const char *page,
1178 					size_t count)
1179 {
1180 	unsigned long tmp;
1181 	char *p = (char *)page;
1182 
1183 	if (reg->hr_bdev)
1184 		return -EINVAL;
1185 
1186 	tmp = simple_strtoul(p, &p, 0);
1187 	if (!p || (*p && (*p != '\n')))
1188 		return -EINVAL;
1189 
1190 	if (tmp > O2NM_MAX_NODES || tmp == 0)
1191 		return -ERANGE;
1192 
1193 	reg->hr_blocks = (unsigned int)tmp;
1194 
1195 	return count;
1196 }
1197 
1198 static ssize_t o2hb_region_dev_read(struct o2hb_region *reg,
1199 				    char *page)
1200 {
1201 	unsigned int ret = 0;
1202 
1203 	if (reg->hr_bdev)
1204 		ret = sprintf(page, "%s\n", reg->hr_dev_name);
1205 
1206 	return ret;
1207 }
1208 
1209 static void o2hb_init_region_params(struct o2hb_region *reg)
1210 {
1211 	reg->hr_slots_per_page = PAGE_CACHE_SIZE >> reg->hr_block_bits;
1212 	reg->hr_timeout_ms = O2HB_REGION_TIMEOUT_MS;
1213 
1214 	mlog(ML_HEARTBEAT, "hr_start_block = %llu, hr_blocks = %u\n",
1215 	     reg->hr_start_block, reg->hr_blocks);
1216 	mlog(ML_HEARTBEAT, "hr_block_bytes = %u, hr_block_bits = %u\n",
1217 	     reg->hr_block_bytes, reg->hr_block_bits);
1218 	mlog(ML_HEARTBEAT, "hr_timeout_ms = %u\n", reg->hr_timeout_ms);
1219 	mlog(ML_HEARTBEAT, "dead threshold = %u\n", o2hb_dead_threshold);
1220 }
1221 
1222 static int o2hb_map_slot_data(struct o2hb_region *reg)
1223 {
1224 	int i, j;
1225 	unsigned int last_slot;
1226 	unsigned int spp = reg->hr_slots_per_page;
1227 	struct page *page;
1228 	char *raw;
1229 	struct o2hb_disk_slot *slot;
1230 
1231 	reg->hr_tmp_block = kmalloc(reg->hr_block_bytes, GFP_KERNEL);
1232 	if (reg->hr_tmp_block == NULL) {
1233 		mlog_errno(-ENOMEM);
1234 		return -ENOMEM;
1235 	}
1236 
1237 	reg->hr_slots = kcalloc(reg->hr_blocks,
1238 				sizeof(struct o2hb_disk_slot), GFP_KERNEL);
1239 	if (reg->hr_slots == NULL) {
1240 		mlog_errno(-ENOMEM);
1241 		return -ENOMEM;
1242 	}
1243 
1244 	for(i = 0; i < reg->hr_blocks; i++) {
1245 		slot = &reg->hr_slots[i];
1246 		slot->ds_node_num = i;
1247 		INIT_LIST_HEAD(&slot->ds_live_item);
1248 		slot->ds_raw_block = NULL;
1249 	}
1250 
1251 	reg->hr_num_pages = (reg->hr_blocks + spp - 1) / spp;
1252 	mlog(ML_HEARTBEAT, "Going to require %u pages to cover %u blocks "
1253 			   "at %u blocks per page\n",
1254 	     reg->hr_num_pages, reg->hr_blocks, spp);
1255 
1256 	reg->hr_slot_data = kcalloc(reg->hr_num_pages, sizeof(struct page *),
1257 				    GFP_KERNEL);
1258 	if (!reg->hr_slot_data) {
1259 		mlog_errno(-ENOMEM);
1260 		return -ENOMEM;
1261 	}
1262 
1263 	for(i = 0; i < reg->hr_num_pages; i++) {
1264 		page = alloc_page(GFP_KERNEL);
1265 		if (!page) {
1266 			mlog_errno(-ENOMEM);
1267 			return -ENOMEM;
1268 		}
1269 
1270 		reg->hr_slot_data[i] = page;
1271 
1272 		last_slot = i * spp;
1273 		raw = page_address(page);
1274 		for (j = 0;
1275 		     (j < spp) && ((j + last_slot) < reg->hr_blocks);
1276 		     j++) {
1277 			BUG_ON((j + last_slot) >= reg->hr_blocks);
1278 
1279 			slot = &reg->hr_slots[j + last_slot];
1280 			slot->ds_raw_block =
1281 				(struct o2hb_disk_heartbeat_block *) raw;
1282 
1283 			raw += reg->hr_block_bytes;
1284 		}
1285 	}
1286 
1287 	return 0;
1288 }
1289 
1290 /* Read in all the slots available and populate the tracking
1291  * structures so that we can start with a baseline idea of what's
1292  * there. */
1293 static int o2hb_populate_slot_data(struct o2hb_region *reg)
1294 {
1295 	int ret, i;
1296 	struct o2hb_disk_slot *slot;
1297 	struct o2hb_disk_heartbeat_block *hb_block;
1298 
1299 	mlog_entry_void();
1300 
1301 	ret = o2hb_read_slots(reg, reg->hr_blocks);
1302 	if (ret) {
1303 		mlog_errno(ret);
1304 		goto out;
1305 	}
1306 
1307 	/* We only want to get an idea of the values initially in each
1308 	 * slot, so we do no verification - o2hb_check_slot will
1309 	 * actually determine if each configured slot is valid and
1310 	 * whether any values have changed. */
1311 	for(i = 0; i < reg->hr_blocks; i++) {
1312 		slot = &reg->hr_slots[i];
1313 		hb_block = (struct o2hb_disk_heartbeat_block *) slot->ds_raw_block;
1314 
1315 		/* Only fill the values that o2hb_check_slot uses to
1316 		 * determine changing slots */
1317 		slot->ds_last_time = le64_to_cpu(hb_block->hb_seq);
1318 		slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation);
1319 	}
1320 
1321 out:
1322 	mlog_exit(ret);
1323 	return ret;
1324 }
1325 
1326 /* this is acting as commit; we set up all of hr_bdev and hr_task or nothing */
1327 static ssize_t o2hb_region_dev_write(struct o2hb_region *reg,
1328 				     const char *page,
1329 				     size_t count)
1330 {
1331 	long fd;
1332 	int sectsize;
1333 	char *p = (char *)page;
1334 	struct file *filp = NULL;
1335 	struct inode *inode = NULL;
1336 	ssize_t ret = -EINVAL;
1337 
1338 	if (reg->hr_bdev)
1339 		goto out;
1340 
1341 	/* We can't heartbeat without having had our node number
1342 	 * configured yet. */
1343 	if (o2nm_this_node() == O2NM_MAX_NODES)
1344 		goto out;
1345 
1346 	fd = simple_strtol(p, &p, 0);
1347 	if (!p || (*p && (*p != '\n')))
1348 		goto out;
1349 
1350 	if (fd < 0 || fd >= INT_MAX)
1351 		goto out;
1352 
1353 	filp = fget(fd);
1354 	if (filp == NULL)
1355 		goto out;
1356 
1357 	if (reg->hr_blocks == 0 || reg->hr_start_block == 0 ||
1358 	    reg->hr_block_bytes == 0)
1359 		goto out;
1360 
1361 	inode = igrab(filp->f_mapping->host);
1362 	if (inode == NULL)
1363 		goto out;
1364 
1365 	if (!S_ISBLK(inode->i_mode))
1366 		goto out;
1367 
1368 	reg->hr_bdev = I_BDEV(filp->f_mapping->host);
1369 	ret = blkdev_get(reg->hr_bdev, FMODE_WRITE | FMODE_READ, 0);
1370 	if (ret) {
1371 		reg->hr_bdev = NULL;
1372 		goto out;
1373 	}
1374 	inode = NULL;
1375 
1376 	bdevname(reg->hr_bdev, reg->hr_dev_name);
1377 
1378 	sectsize = bdev_hardsect_size(reg->hr_bdev);
1379 	if (sectsize != reg->hr_block_bytes) {
1380 		mlog(ML_ERROR,
1381 		     "blocksize %u incorrect for device, expected %d",
1382 		     reg->hr_block_bytes, sectsize);
1383 		ret = -EINVAL;
1384 		goto out;
1385 	}
1386 
1387 	o2hb_init_region_params(reg);
1388 
1389 	/* Generation of zero is invalid */
1390 	do {
1391 		get_random_bytes(&reg->hr_generation,
1392 				 sizeof(reg->hr_generation));
1393 	} while (reg->hr_generation == 0);
1394 
1395 	ret = o2hb_map_slot_data(reg);
1396 	if (ret) {
1397 		mlog_errno(ret);
1398 		goto out;
1399 	}
1400 
1401 	ret = o2hb_populate_slot_data(reg);
1402 	if (ret) {
1403 		mlog_errno(ret);
1404 		goto out;
1405 	}
1406 
1407 	INIT_WORK(&reg->hr_write_timeout_work, o2hb_write_timeout, reg);
1408 
1409 	/*
1410 	 * A node is considered live after it has beat LIVE_THRESHOLD
1411 	 * times.  We're not steady until we've given them a chance
1412 	 * _after_ our first read.
1413 	 */
1414 	atomic_set(&reg->hr_steady_iterations, O2HB_LIVE_THRESHOLD + 1);
1415 
1416 	reg->hr_task = kthread_run(o2hb_thread, reg, "o2hb-%s",
1417 				   reg->hr_item.ci_name);
1418 	if (IS_ERR(reg->hr_task)) {
1419 		ret = PTR_ERR(reg->hr_task);
1420 		mlog_errno(ret);
1421 		reg->hr_task = NULL;
1422 		goto out;
1423 	}
1424 
1425 	ret = wait_event_interruptible(o2hb_steady_queue,
1426 				atomic_read(&reg->hr_steady_iterations) == 0);
1427 	if (ret) {
1428 		kthread_stop(reg->hr_task);
1429 		reg->hr_task = NULL;
1430 		goto out;
1431 	}
1432 
1433 	ret = count;
1434 out:
1435 	if (filp)
1436 		fput(filp);
1437 	if (inode)
1438 		iput(inode);
1439 	if (ret < 0) {
1440 		if (reg->hr_bdev) {
1441 			blkdev_put(reg->hr_bdev);
1442 			reg->hr_bdev = NULL;
1443 		}
1444 	}
1445 	return ret;
1446 }
1447 
1448 struct o2hb_region_attribute {
1449 	struct configfs_attribute attr;
1450 	ssize_t (*show)(struct o2hb_region *, char *);
1451 	ssize_t (*store)(struct o2hb_region *, const char *, size_t);
1452 };
1453 
1454 static struct o2hb_region_attribute o2hb_region_attr_block_bytes = {
1455 	.attr	= { .ca_owner = THIS_MODULE,
1456 		    .ca_name = "block_bytes",
1457 		    .ca_mode = S_IRUGO | S_IWUSR },
1458 	.show	= o2hb_region_block_bytes_read,
1459 	.store	= o2hb_region_block_bytes_write,
1460 };
1461 
1462 static struct o2hb_region_attribute o2hb_region_attr_start_block = {
1463 	.attr	= { .ca_owner = THIS_MODULE,
1464 		    .ca_name = "start_block",
1465 		    .ca_mode = S_IRUGO | S_IWUSR },
1466 	.show	= o2hb_region_start_block_read,
1467 	.store	= o2hb_region_start_block_write,
1468 };
1469 
1470 static struct o2hb_region_attribute o2hb_region_attr_blocks = {
1471 	.attr	= { .ca_owner = THIS_MODULE,
1472 		    .ca_name = "blocks",
1473 		    .ca_mode = S_IRUGO | S_IWUSR },
1474 	.show	= o2hb_region_blocks_read,
1475 	.store	= o2hb_region_blocks_write,
1476 };
1477 
1478 static struct o2hb_region_attribute o2hb_region_attr_dev = {
1479 	.attr	= { .ca_owner = THIS_MODULE,
1480 		    .ca_name = "dev",
1481 		    .ca_mode = S_IRUGO | S_IWUSR },
1482 	.show	= o2hb_region_dev_read,
1483 	.store	= o2hb_region_dev_write,
1484 };
1485 
1486 static struct configfs_attribute *o2hb_region_attrs[] = {
1487 	&o2hb_region_attr_block_bytes.attr,
1488 	&o2hb_region_attr_start_block.attr,
1489 	&o2hb_region_attr_blocks.attr,
1490 	&o2hb_region_attr_dev.attr,
1491 	NULL,
1492 };
1493 
1494 static ssize_t o2hb_region_show(struct config_item *item,
1495 				struct configfs_attribute *attr,
1496 				char *page)
1497 {
1498 	struct o2hb_region *reg = to_o2hb_region(item);
1499 	struct o2hb_region_attribute *o2hb_region_attr =
1500 		container_of(attr, struct o2hb_region_attribute, attr);
1501 	ssize_t ret = 0;
1502 
1503 	if (o2hb_region_attr->show)
1504 		ret = o2hb_region_attr->show(reg, page);
1505 	return ret;
1506 }
1507 
1508 static ssize_t o2hb_region_store(struct config_item *item,
1509 				 struct configfs_attribute *attr,
1510 				 const char *page, size_t count)
1511 {
1512 	struct o2hb_region *reg = to_o2hb_region(item);
1513 	struct o2hb_region_attribute *o2hb_region_attr =
1514 		container_of(attr, struct o2hb_region_attribute, attr);
1515 	ssize_t ret = -EINVAL;
1516 
1517 	if (o2hb_region_attr->store)
1518 		ret = o2hb_region_attr->store(reg, page, count);
1519 	return ret;
1520 }
1521 
1522 static struct configfs_item_operations o2hb_region_item_ops = {
1523 	.release		= o2hb_region_release,
1524 	.show_attribute		= o2hb_region_show,
1525 	.store_attribute	= o2hb_region_store,
1526 };
1527 
1528 static struct config_item_type o2hb_region_type = {
1529 	.ct_item_ops	= &o2hb_region_item_ops,
1530 	.ct_attrs	= o2hb_region_attrs,
1531 	.ct_owner	= THIS_MODULE,
1532 };
1533 
1534 /* heartbeat set */
1535 
1536 struct o2hb_heartbeat_group {
1537 	struct config_group hs_group;
1538 	/* some stuff? */
1539 };
1540 
1541 static struct o2hb_heartbeat_group *to_o2hb_heartbeat_group(struct config_group *group)
1542 {
1543 	return group ?
1544 		container_of(group, struct o2hb_heartbeat_group, hs_group)
1545 		: NULL;
1546 }
1547 
1548 static struct config_item *o2hb_heartbeat_group_make_item(struct config_group *group,
1549 							  const char *name)
1550 {
1551 	struct o2hb_region *reg = NULL;
1552 	struct config_item *ret = NULL;
1553 
1554 	reg = kcalloc(1, sizeof(struct o2hb_region), GFP_KERNEL);
1555 	if (reg == NULL)
1556 		goto out; /* ENOMEM */
1557 
1558 	config_item_init_type_name(&reg->hr_item, name, &o2hb_region_type);
1559 
1560 	ret = &reg->hr_item;
1561 
1562 	spin_lock(&o2hb_live_lock);
1563 	list_add_tail(&reg->hr_all_item, &o2hb_all_regions);
1564 	spin_unlock(&o2hb_live_lock);
1565 out:
1566 	if (ret == NULL)
1567 		kfree(reg);
1568 
1569 	return ret;
1570 }
1571 
1572 static void o2hb_heartbeat_group_drop_item(struct config_group *group,
1573 					   struct config_item *item)
1574 {
1575 	struct o2hb_region *reg = to_o2hb_region(item);
1576 
1577 	/* stop the thread when the user removes the region dir */
1578 	if (reg->hr_task) {
1579 		kthread_stop(reg->hr_task);
1580 		reg->hr_task = NULL;
1581 	}
1582 
1583 	config_item_put(item);
1584 }
1585 
1586 struct o2hb_heartbeat_group_attribute {
1587 	struct configfs_attribute attr;
1588 	ssize_t (*show)(struct o2hb_heartbeat_group *, char *);
1589 	ssize_t (*store)(struct o2hb_heartbeat_group *, const char *, size_t);
1590 };
1591 
1592 static ssize_t o2hb_heartbeat_group_show(struct config_item *item,
1593 					 struct configfs_attribute *attr,
1594 					 char *page)
1595 {
1596 	struct o2hb_heartbeat_group *reg = to_o2hb_heartbeat_group(to_config_group(item));
1597 	struct o2hb_heartbeat_group_attribute *o2hb_heartbeat_group_attr =
1598 		container_of(attr, struct o2hb_heartbeat_group_attribute, attr);
1599 	ssize_t ret = 0;
1600 
1601 	if (o2hb_heartbeat_group_attr->show)
1602 		ret = o2hb_heartbeat_group_attr->show(reg, page);
1603 	return ret;
1604 }
1605 
1606 static ssize_t o2hb_heartbeat_group_store(struct config_item *item,
1607 					  struct configfs_attribute *attr,
1608 					  const char *page, size_t count)
1609 {
1610 	struct o2hb_heartbeat_group *reg = to_o2hb_heartbeat_group(to_config_group(item));
1611 	struct o2hb_heartbeat_group_attribute *o2hb_heartbeat_group_attr =
1612 		container_of(attr, struct o2hb_heartbeat_group_attribute, attr);
1613 	ssize_t ret = -EINVAL;
1614 
1615 	if (o2hb_heartbeat_group_attr->store)
1616 		ret = o2hb_heartbeat_group_attr->store(reg, page, count);
1617 	return ret;
1618 }
1619 
1620 static ssize_t o2hb_heartbeat_group_threshold_show(struct o2hb_heartbeat_group *group,
1621 						     char *page)
1622 {
1623 	return sprintf(page, "%u\n", o2hb_dead_threshold);
1624 }
1625 
1626 static ssize_t o2hb_heartbeat_group_threshold_store(struct o2hb_heartbeat_group *group,
1627 						    const char *page,
1628 						    size_t count)
1629 {
1630 	unsigned long tmp;
1631 	char *p = (char *)page;
1632 
1633 	tmp = simple_strtoul(p, &p, 10);
1634 	if (!p || (*p && (*p != '\n')))
1635                 return -EINVAL;
1636 
1637 	/* this will validate ranges for us. */
1638 	o2hb_dead_threshold_set((unsigned int) tmp);
1639 
1640 	return count;
1641 }
1642 
1643 static struct o2hb_heartbeat_group_attribute o2hb_heartbeat_group_attr_threshold = {
1644 	.attr	= { .ca_owner = THIS_MODULE,
1645 		    .ca_name = "dead_threshold",
1646 		    .ca_mode = S_IRUGO | S_IWUSR },
1647 	.show	= o2hb_heartbeat_group_threshold_show,
1648 	.store	= o2hb_heartbeat_group_threshold_store,
1649 };
1650 
1651 static struct configfs_attribute *o2hb_heartbeat_group_attrs[] = {
1652 	&o2hb_heartbeat_group_attr_threshold.attr,
1653 	NULL,
1654 };
1655 
1656 static struct configfs_item_operations o2hb_hearbeat_group_item_ops = {
1657 	.show_attribute		= o2hb_heartbeat_group_show,
1658 	.store_attribute	= o2hb_heartbeat_group_store,
1659 };
1660 
1661 static struct configfs_group_operations o2hb_heartbeat_group_group_ops = {
1662 	.make_item	= o2hb_heartbeat_group_make_item,
1663 	.drop_item	= o2hb_heartbeat_group_drop_item,
1664 };
1665 
1666 static struct config_item_type o2hb_heartbeat_group_type = {
1667 	.ct_group_ops	= &o2hb_heartbeat_group_group_ops,
1668 	.ct_item_ops	= &o2hb_hearbeat_group_item_ops,
1669 	.ct_attrs	= o2hb_heartbeat_group_attrs,
1670 	.ct_owner	= THIS_MODULE,
1671 };
1672 
1673 /* this is just here to avoid touching group in heartbeat.h which the
1674  * entire damn world #includes */
1675 struct config_group *o2hb_alloc_hb_set(void)
1676 {
1677 	struct o2hb_heartbeat_group *hs = NULL;
1678 	struct config_group *ret = NULL;
1679 
1680 	hs = kcalloc(1, sizeof(struct o2hb_heartbeat_group), GFP_KERNEL);
1681 	if (hs == NULL)
1682 		goto out;
1683 
1684 	config_group_init_type_name(&hs->hs_group, "heartbeat",
1685 				    &o2hb_heartbeat_group_type);
1686 
1687 	ret = &hs->hs_group;
1688 out:
1689 	if (ret == NULL)
1690 		kfree(hs);
1691 	return ret;
1692 }
1693 
1694 void o2hb_free_hb_set(struct config_group *group)
1695 {
1696 	struct o2hb_heartbeat_group *hs = to_o2hb_heartbeat_group(group);
1697 	kfree(hs);
1698 }
1699 
1700 /* hb callback registration and issueing */
1701 
1702 static struct o2hb_callback *hbcall_from_type(enum o2hb_callback_type type)
1703 {
1704 	if (type == O2HB_NUM_CB)
1705 		return ERR_PTR(-EINVAL);
1706 
1707 	return &o2hb_callbacks[type];
1708 }
1709 
1710 void o2hb_setup_callback(struct o2hb_callback_func *hc,
1711 			 enum o2hb_callback_type type,
1712 			 o2hb_cb_func *func,
1713 			 void *data,
1714 			 int priority)
1715 {
1716 	INIT_LIST_HEAD(&hc->hc_item);
1717 	hc->hc_func = func;
1718 	hc->hc_data = data;
1719 	hc->hc_priority = priority;
1720 	hc->hc_type = type;
1721 	hc->hc_magic = O2HB_CB_MAGIC;
1722 }
1723 EXPORT_SYMBOL_GPL(o2hb_setup_callback);
1724 
1725 int o2hb_register_callback(struct o2hb_callback_func *hc)
1726 {
1727 	struct o2hb_callback_func *tmp;
1728 	struct list_head *iter;
1729 	struct o2hb_callback *hbcall;
1730 	int ret;
1731 
1732 	BUG_ON(hc->hc_magic != O2HB_CB_MAGIC);
1733 	BUG_ON(!list_empty(&hc->hc_item));
1734 
1735 	hbcall = hbcall_from_type(hc->hc_type);
1736 	if (IS_ERR(hbcall)) {
1737 		ret = PTR_ERR(hbcall);
1738 		goto out;
1739 	}
1740 
1741 	down_write(&o2hb_callback_sem);
1742 
1743 	list_for_each(iter, &hbcall->list) {
1744 		tmp = list_entry(iter, struct o2hb_callback_func, hc_item);
1745 		if (hc->hc_priority < tmp->hc_priority) {
1746 			list_add_tail(&hc->hc_item, iter);
1747 			break;
1748 		}
1749 	}
1750 	if (list_empty(&hc->hc_item))
1751 		list_add_tail(&hc->hc_item, &hbcall->list);
1752 
1753 	up_write(&o2hb_callback_sem);
1754 	ret = 0;
1755 out:
1756 	mlog(ML_HEARTBEAT, "returning %d on behalf of %p for funcs %p\n",
1757 	     ret, __builtin_return_address(0), hc);
1758 	return ret;
1759 }
1760 EXPORT_SYMBOL_GPL(o2hb_register_callback);
1761 
1762 int o2hb_unregister_callback(struct o2hb_callback_func *hc)
1763 {
1764 	BUG_ON(hc->hc_magic != O2HB_CB_MAGIC);
1765 
1766 	mlog(ML_HEARTBEAT, "on behalf of %p for funcs %p\n",
1767 	     __builtin_return_address(0), hc);
1768 
1769 	if (list_empty(&hc->hc_item))
1770 		return 0;
1771 
1772 	down_write(&o2hb_callback_sem);
1773 
1774 	list_del_init(&hc->hc_item);
1775 
1776 	up_write(&o2hb_callback_sem);
1777 
1778 	return 0;
1779 }
1780 EXPORT_SYMBOL_GPL(o2hb_unregister_callback);
1781 
1782 int o2hb_check_node_heartbeating(u8 node_num)
1783 {
1784 	unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1785 
1786 	o2hb_fill_node_map(testing_map, sizeof(testing_map));
1787 	if (!test_bit(node_num, testing_map)) {
1788 		mlog(ML_HEARTBEAT,
1789 		     "node (%u) does not have heartbeating enabled.\n",
1790 		     node_num);
1791 		return 0;
1792 	}
1793 
1794 	return 1;
1795 }
1796 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating);
1797 
1798 int o2hb_check_node_heartbeating_from_callback(u8 node_num)
1799 {
1800 	unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1801 
1802 	o2hb_fill_node_map_from_callback(testing_map, sizeof(testing_map));
1803 	if (!test_bit(node_num, testing_map)) {
1804 		mlog(ML_HEARTBEAT,
1805 		     "node (%u) does not have heartbeating enabled.\n",
1806 		     node_num);
1807 		return 0;
1808 	}
1809 
1810 	return 1;
1811 }
1812 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating_from_callback);
1813 
1814 /* Makes sure our local node is configured with a node number, and is
1815  * heartbeating. */
1816 int o2hb_check_local_node_heartbeating(void)
1817 {
1818 	u8 node_num;
1819 
1820 	/* if this node was set then we have networking */
1821 	node_num = o2nm_this_node();
1822 	if (node_num == O2NM_MAX_NODES) {
1823 		mlog(ML_HEARTBEAT, "this node has not been configured.\n");
1824 		return 0;
1825 	}
1826 
1827 	return o2hb_check_node_heartbeating(node_num);
1828 }
1829 EXPORT_SYMBOL_GPL(o2hb_check_local_node_heartbeating);
1830 
1831 /*
1832  * this is just a hack until we get the plumbing which flips file systems
1833  * read only and drops the hb ref instead of killing the node dead.
1834  */
1835 void o2hb_stop_all_regions(void)
1836 {
1837 	struct o2hb_region *reg;
1838 
1839 	mlog(ML_ERROR, "stopping heartbeat on all active regions.\n");
1840 
1841 	spin_lock(&o2hb_live_lock);
1842 
1843 	list_for_each_entry(reg, &o2hb_all_regions, hr_all_item)
1844 		reg->hr_unclean_stop = 1;
1845 
1846 	spin_unlock(&o2hb_live_lock);
1847 }
1848 EXPORT_SYMBOL_GPL(o2hb_stop_all_regions);
1849