xref: /linux/fs/ocfs2/cluster/heartbeat.c (revision 14b42963f64b98ab61fa9723c03d71aa5ef4f862)
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 	max_sectors = max_pages << (PAGE_SHIFT - 9);
324 
325 	/* Why is fls() 1-based???? */
326 	pow_two_sectors = 1 << (fls(max_sectors) - 1);
327 
328 	return pow_two_sectors;
329 }
330 
331 static inline void o2hb_compute_request_limits(struct o2hb_region *reg,
332 					       unsigned int num_slots,
333 					       unsigned int *num_bios,
334 					       unsigned int *slots_per_bio)
335 {
336 	unsigned int max_sectors, io_sectors;
337 
338 	max_sectors = compute_max_sectors(reg->hr_bdev);
339 
340 	io_sectors = num_slots << (reg->hr_block_bits - 9);
341 
342 	*num_bios = (io_sectors + max_sectors - 1) / max_sectors;
343 	*slots_per_bio = max_sectors >> (reg->hr_block_bits - 9);
344 
345 	mlog(ML_HB_BIO, "My io size is %u sectors for %u slots. This "
346 	     "device can handle %u sectors of I/O\n", io_sectors, num_slots,
347 	     max_sectors);
348 	mlog(ML_HB_BIO, "Will need %u bios holding %u slots each\n",
349 	     *num_bios, *slots_per_bio);
350 }
351 
352 static int o2hb_read_slots(struct o2hb_region *reg,
353 			   unsigned int max_slots)
354 {
355 	unsigned int num_bios, slots_per_bio, start_slot, num_slots;
356 	int i, status;
357 	struct o2hb_bio_wait_ctxt wc;
358 	struct bio **bios;
359 	struct bio *bio;
360 
361 	o2hb_compute_request_limits(reg, max_slots, &num_bios, &slots_per_bio);
362 
363 	bios = kcalloc(num_bios, sizeof(struct bio *), GFP_KERNEL);
364 	if (!bios) {
365 		status = -ENOMEM;
366 		mlog_errno(status);
367 		return status;
368 	}
369 
370 	o2hb_bio_wait_init(&wc, num_bios);
371 
372 	num_slots = slots_per_bio;
373 	for(i = 0; i < num_bios; i++) {
374 		start_slot = i * slots_per_bio;
375 
376 		/* adjust num_slots at last bio */
377 		if (max_slots < (start_slot + num_slots))
378 			num_slots = max_slots - start_slot;
379 
380 		bio = o2hb_setup_one_bio(reg, &wc, start_slot, num_slots);
381 		if (IS_ERR(bio)) {
382 			o2hb_bio_wait_dec(&wc, num_bios - i);
383 
384 			status = PTR_ERR(bio);
385 			mlog_errno(status);
386 			goto bail_and_wait;
387 		}
388 		bios[i] = bio;
389 
390 		submit_bio(READ, bio);
391 	}
392 
393 	status = 0;
394 
395 bail_and_wait:
396 	o2hb_wait_on_io(reg, &wc);
397 	if (wc.wc_error && !status)
398 		status = wc.wc_error;
399 
400 	if (bios) {
401 		for(i = 0; i < num_bios; i++)
402 			if (bios[i])
403 				bio_put(bios[i]);
404 		kfree(bios);
405 	}
406 
407 	return status;
408 }
409 
410 static int o2hb_issue_node_write(struct o2hb_region *reg,
411 				 struct bio **write_bio,
412 				 struct o2hb_bio_wait_ctxt *write_wc)
413 {
414 	int status;
415 	unsigned int slot;
416 	struct bio *bio;
417 
418 	o2hb_bio_wait_init(write_wc, 1);
419 
420 	slot = o2nm_this_node();
421 
422 	bio = o2hb_setup_one_bio(reg, write_wc, slot, 1);
423 	if (IS_ERR(bio)) {
424 		status = PTR_ERR(bio);
425 		mlog_errno(status);
426 		goto bail;
427 	}
428 
429 	submit_bio(WRITE, bio);
430 
431 	*write_bio = bio;
432 	status = 0;
433 bail:
434 	return status;
435 }
436 
437 static u32 o2hb_compute_block_crc_le(struct o2hb_region *reg,
438 				     struct o2hb_disk_heartbeat_block *hb_block)
439 {
440 	__le32 old_cksum;
441 	u32 ret;
442 
443 	/* We want to compute the block crc with a 0 value in the
444 	 * hb_cksum field. Save it off here and replace after the
445 	 * crc. */
446 	old_cksum = hb_block->hb_cksum;
447 	hb_block->hb_cksum = 0;
448 
449 	ret = crc32_le(0, (unsigned char *) hb_block, reg->hr_block_bytes);
450 
451 	hb_block->hb_cksum = old_cksum;
452 
453 	return ret;
454 }
455 
456 static void o2hb_dump_slot(struct o2hb_disk_heartbeat_block *hb_block)
457 {
458 	mlog(ML_ERROR, "Dump slot information: seq = 0x%llx, node = %u, "
459 	     "cksum = 0x%x, generation 0x%llx\n",
460 	     (long long)le64_to_cpu(hb_block->hb_seq),
461 	     hb_block->hb_node, le32_to_cpu(hb_block->hb_cksum),
462 	     (long long)le64_to_cpu(hb_block->hb_generation));
463 }
464 
465 static int o2hb_verify_crc(struct o2hb_region *reg,
466 			   struct o2hb_disk_heartbeat_block *hb_block)
467 {
468 	u32 read, computed;
469 
470 	read = le32_to_cpu(hb_block->hb_cksum);
471 	computed = o2hb_compute_block_crc_le(reg, hb_block);
472 
473 	return read == computed;
474 }
475 
476 /* We want to make sure that nobody is heartbeating on top of us --
477  * this will help detect an invalid configuration. */
478 static int o2hb_check_last_timestamp(struct o2hb_region *reg)
479 {
480 	int node_num, ret;
481 	struct o2hb_disk_slot *slot;
482 	struct o2hb_disk_heartbeat_block *hb_block;
483 
484 	node_num = o2nm_this_node();
485 
486 	ret = 1;
487 	slot = &reg->hr_slots[node_num];
488 	/* Don't check on our 1st timestamp */
489 	if (slot->ds_last_time) {
490 		hb_block = slot->ds_raw_block;
491 
492 		if (le64_to_cpu(hb_block->hb_seq) != slot->ds_last_time)
493 			ret = 0;
494 	}
495 
496 	return ret;
497 }
498 
499 static inline void o2hb_prepare_block(struct o2hb_region *reg,
500 				      u64 generation)
501 {
502 	int node_num;
503 	u64 cputime;
504 	struct o2hb_disk_slot *slot;
505 	struct o2hb_disk_heartbeat_block *hb_block;
506 
507 	node_num = o2nm_this_node();
508 	slot = &reg->hr_slots[node_num];
509 
510 	hb_block = (struct o2hb_disk_heartbeat_block *)slot->ds_raw_block;
511 	memset(hb_block, 0, reg->hr_block_bytes);
512 	/* TODO: time stuff */
513 	cputime = CURRENT_TIME.tv_sec;
514 	if (!cputime)
515 		cputime = 1;
516 
517 	hb_block->hb_seq = cpu_to_le64(cputime);
518 	hb_block->hb_node = node_num;
519 	hb_block->hb_generation = cpu_to_le64(generation);
520 	hb_block->hb_dead_ms = cpu_to_le32(o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS);
521 
522 	/* This step must always happen last! */
523 	hb_block->hb_cksum = cpu_to_le32(o2hb_compute_block_crc_le(reg,
524 								   hb_block));
525 
526 	mlog(ML_HB_BIO, "our node generation = 0x%llx, cksum = 0x%x\n",
527 	     (long long)cpu_to_le64(generation),
528 	     le32_to_cpu(hb_block->hb_cksum));
529 }
530 
531 static void o2hb_fire_callbacks(struct o2hb_callback *hbcall,
532 				struct o2nm_node *node,
533 				int idx)
534 {
535 	struct list_head *iter;
536 	struct o2hb_callback_func *f;
537 
538 	list_for_each(iter, &hbcall->list) {
539 		f = list_entry(iter, struct o2hb_callback_func, hc_item);
540 		mlog(ML_HEARTBEAT, "calling funcs %p\n", f);
541 		(f->hc_func)(node, idx, f->hc_data);
542 	}
543 }
544 
545 /* Will run the list in order until we process the passed event */
546 static void o2hb_run_event_list(struct o2hb_node_event *queued_event)
547 {
548 	int empty;
549 	struct o2hb_callback *hbcall;
550 	struct o2hb_node_event *event;
551 
552 	spin_lock(&o2hb_live_lock);
553 	empty = list_empty(&queued_event->hn_item);
554 	spin_unlock(&o2hb_live_lock);
555 	if (empty)
556 		return;
557 
558 	/* Holding callback sem assures we don't alter the callback
559 	 * lists when doing this, and serializes ourselves with other
560 	 * processes wanting callbacks. */
561 	down_write(&o2hb_callback_sem);
562 
563 	spin_lock(&o2hb_live_lock);
564 	while (!list_empty(&o2hb_node_events)
565 	       && !list_empty(&queued_event->hn_item)) {
566 		event = list_entry(o2hb_node_events.next,
567 				   struct o2hb_node_event,
568 				   hn_item);
569 		list_del_init(&event->hn_item);
570 		spin_unlock(&o2hb_live_lock);
571 
572 		mlog(ML_HEARTBEAT, "Node %s event for %d\n",
573 		     event->hn_event_type == O2HB_NODE_UP_CB ? "UP" : "DOWN",
574 		     event->hn_node_num);
575 
576 		hbcall = hbcall_from_type(event->hn_event_type);
577 
578 		/* We should *never* have gotten on to the list with a
579 		 * bad type... This isn't something that we should try
580 		 * to recover from. */
581 		BUG_ON(IS_ERR(hbcall));
582 
583 		o2hb_fire_callbacks(hbcall, event->hn_node, event->hn_node_num);
584 
585 		spin_lock(&o2hb_live_lock);
586 	}
587 	spin_unlock(&o2hb_live_lock);
588 
589 	up_write(&o2hb_callback_sem);
590 }
591 
592 static void o2hb_queue_node_event(struct o2hb_node_event *event,
593 				  enum o2hb_callback_type type,
594 				  struct o2nm_node *node,
595 				  int node_num)
596 {
597 	assert_spin_locked(&o2hb_live_lock);
598 
599 	event->hn_event_type = type;
600 	event->hn_node = node;
601 	event->hn_node_num = node_num;
602 
603 	mlog(ML_HEARTBEAT, "Queue node %s event for node %d\n",
604 	     type == O2HB_NODE_UP_CB ? "UP" : "DOWN", node_num);
605 
606 	list_add_tail(&event->hn_item, &o2hb_node_events);
607 }
608 
609 static void o2hb_shutdown_slot(struct o2hb_disk_slot *slot)
610 {
611 	struct o2hb_node_event event =
612 		{ .hn_item = LIST_HEAD_INIT(event.hn_item), };
613 	struct o2nm_node *node;
614 
615 	node = o2nm_get_node_by_num(slot->ds_node_num);
616 	if (!node)
617 		return;
618 
619 	spin_lock(&o2hb_live_lock);
620 	if (!list_empty(&slot->ds_live_item)) {
621 		mlog(ML_HEARTBEAT, "Shutdown, node %d leaves region\n",
622 		     slot->ds_node_num);
623 
624 		list_del_init(&slot->ds_live_item);
625 
626 		if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
627 			clear_bit(slot->ds_node_num, o2hb_live_node_bitmap);
628 
629 			o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB, node,
630 					      slot->ds_node_num);
631 		}
632 	}
633 	spin_unlock(&o2hb_live_lock);
634 
635 	o2hb_run_event_list(&event);
636 
637 	o2nm_node_put(node);
638 }
639 
640 static int o2hb_check_slot(struct o2hb_region *reg,
641 			   struct o2hb_disk_slot *slot)
642 {
643 	int changed = 0, gen_changed = 0;
644 	struct o2hb_node_event event =
645 		{ .hn_item = LIST_HEAD_INIT(event.hn_item), };
646 	struct o2nm_node *node;
647 	struct o2hb_disk_heartbeat_block *hb_block = reg->hr_tmp_block;
648 	u64 cputime;
649 	unsigned int dead_ms = o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS;
650 	unsigned int slot_dead_ms;
651 
652 	memcpy(hb_block, slot->ds_raw_block, reg->hr_block_bytes);
653 
654 	/* Is this correct? Do we assume that the node doesn't exist
655 	 * if we're not configured for him? */
656 	node = o2nm_get_node_by_num(slot->ds_node_num);
657 	if (!node)
658 		return 0;
659 
660 	if (!o2hb_verify_crc(reg, hb_block)) {
661 		/* all paths from here will drop o2hb_live_lock for
662 		 * us. */
663 		spin_lock(&o2hb_live_lock);
664 
665 		/* Don't print an error on the console in this case -
666 		 * a freshly formatted heartbeat area will not have a
667 		 * crc set on it. */
668 		if (list_empty(&slot->ds_live_item))
669 			goto out;
670 
671 		/* The node is live but pushed out a bad crc. We
672 		 * consider it a transient miss but don't populate any
673 		 * other values as they may be junk. */
674 		mlog(ML_ERROR, "Node %d has written a bad crc to %s\n",
675 		     slot->ds_node_num, reg->hr_dev_name);
676 		o2hb_dump_slot(hb_block);
677 
678 		slot->ds_equal_samples++;
679 		goto fire_callbacks;
680 	}
681 
682 	/* we don't care if these wrap.. the state transitions below
683 	 * clear at the right places */
684 	cputime = le64_to_cpu(hb_block->hb_seq);
685 	if (slot->ds_last_time != cputime)
686 		slot->ds_changed_samples++;
687 	else
688 		slot->ds_equal_samples++;
689 	slot->ds_last_time = cputime;
690 
691 	/* The node changed heartbeat generations. We assume this to
692 	 * mean it dropped off but came back before we timed out. We
693 	 * want to consider it down for the time being but don't want
694 	 * to lose any changed_samples state we might build up to
695 	 * considering it live again. */
696 	if (slot->ds_last_generation != le64_to_cpu(hb_block->hb_generation)) {
697 		gen_changed = 1;
698 		slot->ds_equal_samples = 0;
699 		mlog(ML_HEARTBEAT, "Node %d changed generation (0x%llx "
700 		     "to 0x%llx)\n", slot->ds_node_num,
701 		     (long long)slot->ds_last_generation,
702 		     (long long)le64_to_cpu(hb_block->hb_generation));
703 	}
704 
705 	slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation);
706 
707 	mlog(ML_HEARTBEAT, "Slot %d gen 0x%llx cksum 0x%x "
708 	     "seq %llu last %llu changed %u equal %u\n",
709 	     slot->ds_node_num, (long long)slot->ds_last_generation,
710 	     le32_to_cpu(hb_block->hb_cksum),
711 	     (unsigned long long)le64_to_cpu(hb_block->hb_seq),
712 	     (unsigned long long)slot->ds_last_time, slot->ds_changed_samples,
713 	     slot->ds_equal_samples);
714 
715 	spin_lock(&o2hb_live_lock);
716 
717 fire_callbacks:
718 	/* dead nodes only come to life after some number of
719 	 * changes at any time during their dead time */
720 	if (list_empty(&slot->ds_live_item) &&
721 	    slot->ds_changed_samples >= O2HB_LIVE_THRESHOLD) {
722 		mlog(ML_HEARTBEAT, "Node %d (id 0x%llx) joined my region\n",
723 		     slot->ds_node_num, (long long)slot->ds_last_generation);
724 
725 		/* first on the list generates a callback */
726 		if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
727 			set_bit(slot->ds_node_num, o2hb_live_node_bitmap);
728 
729 			o2hb_queue_node_event(&event, O2HB_NODE_UP_CB, node,
730 					      slot->ds_node_num);
731 
732 			changed = 1;
733 		}
734 
735 		list_add_tail(&slot->ds_live_item,
736 			      &o2hb_live_slots[slot->ds_node_num]);
737 
738 		slot->ds_equal_samples = 0;
739 
740 		/* We want to be sure that all nodes agree on the
741 		 * number of milliseconds before a node will be
742 		 * considered dead. The self-fencing timeout is
743 		 * computed from this value, and a discrepancy might
744 		 * result in heartbeat calling a node dead when it
745 		 * hasn't self-fenced yet. */
746 		slot_dead_ms = le32_to_cpu(hb_block->hb_dead_ms);
747 		if (slot_dead_ms && slot_dead_ms != dead_ms) {
748 			/* TODO: Perhaps we can fail the region here. */
749 			mlog(ML_ERROR, "Node %d on device %s has a dead count "
750 			     "of %u ms, but our count is %u ms.\n"
751 			     "Please double check your configuration values "
752 			     "for 'O2CB_HEARTBEAT_THRESHOLD'\n",
753 			     slot->ds_node_num, reg->hr_dev_name, slot_dead_ms,
754 			     dead_ms);
755 		}
756 		goto out;
757 	}
758 
759 	/* if the list is dead, we're done.. */
760 	if (list_empty(&slot->ds_live_item))
761 		goto out;
762 
763 	/* live nodes only go dead after enough consequtive missed
764 	 * samples..  reset the missed counter whenever we see
765 	 * activity */
766 	if (slot->ds_equal_samples >= o2hb_dead_threshold || gen_changed) {
767 		mlog(ML_HEARTBEAT, "Node %d left my region\n",
768 		     slot->ds_node_num);
769 
770 		/* last off the live_slot generates a callback */
771 		list_del_init(&slot->ds_live_item);
772 		if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
773 			clear_bit(slot->ds_node_num, o2hb_live_node_bitmap);
774 
775 			o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB, node,
776 					      slot->ds_node_num);
777 
778 			changed = 1;
779 		}
780 
781 		/* We don't clear this because the node is still
782 		 * actually writing new blocks. */
783 		if (!gen_changed)
784 			slot->ds_changed_samples = 0;
785 		goto out;
786 	}
787 	if (slot->ds_changed_samples) {
788 		slot->ds_changed_samples = 0;
789 		slot->ds_equal_samples = 0;
790 	}
791 out:
792 	spin_unlock(&o2hb_live_lock);
793 
794 	o2hb_run_event_list(&event);
795 
796 	o2nm_node_put(node);
797 	return changed;
798 }
799 
800 /* This could be faster if we just implmented a find_last_bit, but I
801  * don't think the circumstances warrant it. */
802 static int o2hb_highest_node(unsigned long *nodes,
803 			     int numbits)
804 {
805 	int highest, node;
806 
807 	highest = numbits;
808 	node = -1;
809 	while ((node = find_next_bit(nodes, numbits, node + 1)) != -1) {
810 		if (node >= numbits)
811 			break;
812 
813 		highest = node;
814 	}
815 
816 	return highest;
817 }
818 
819 static int o2hb_do_disk_heartbeat(struct o2hb_region *reg)
820 {
821 	int i, ret, highest_node, change = 0;
822 	unsigned long configured_nodes[BITS_TO_LONGS(O2NM_MAX_NODES)];
823 	struct bio *write_bio;
824 	struct o2hb_bio_wait_ctxt write_wc;
825 
826 	ret = o2nm_configured_node_map(configured_nodes,
827 				       sizeof(configured_nodes));
828 	if (ret) {
829 		mlog_errno(ret);
830 		return ret;
831 	}
832 
833 	highest_node = o2hb_highest_node(configured_nodes, O2NM_MAX_NODES);
834 	if (highest_node >= O2NM_MAX_NODES) {
835 		mlog(ML_NOTICE, "ocfs2_heartbeat: no configured nodes found!\n");
836 		return -EINVAL;
837 	}
838 
839 	/* No sense in reading the slots of nodes that don't exist
840 	 * yet. Of course, if the node definitions have holes in them
841 	 * then we're reading an empty slot anyway... Consider this
842 	 * best-effort. */
843 	ret = o2hb_read_slots(reg, highest_node + 1);
844 	if (ret < 0) {
845 		mlog_errno(ret);
846 		return ret;
847 	}
848 
849 	/* With an up to date view of the slots, we can check that no
850 	 * other node has been improperly configured to heartbeat in
851 	 * our slot. */
852 	if (!o2hb_check_last_timestamp(reg))
853 		mlog(ML_ERROR, "Device \"%s\": another node is heartbeating "
854 		     "in our slot!\n", reg->hr_dev_name);
855 
856 	/* fill in the proper info for our next heartbeat */
857 	o2hb_prepare_block(reg, reg->hr_generation);
858 
859 	/* And fire off the write. Note that we don't wait on this I/O
860 	 * until later. */
861 	ret = o2hb_issue_node_write(reg, &write_bio, &write_wc);
862 	if (ret < 0) {
863 		mlog_errno(ret);
864 		return ret;
865 	}
866 
867 	i = -1;
868 	while((i = find_next_bit(configured_nodes, O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) {
869 
870 		change |= o2hb_check_slot(reg, &reg->hr_slots[i]);
871 	}
872 
873 	/*
874 	 * We have to be sure we've advertised ourselves on disk
875 	 * before we can go to steady state.  This ensures that
876 	 * people we find in our steady state have seen us.
877 	 */
878 	o2hb_wait_on_io(reg, &write_wc);
879 	bio_put(write_bio);
880 	if (write_wc.wc_error) {
881 		/* Do not re-arm the write timeout on I/O error - we
882 		 * can't be sure that the new block ever made it to
883 		 * disk */
884 		mlog(ML_ERROR, "Write error %d on device \"%s\"\n",
885 		     write_wc.wc_error, reg->hr_dev_name);
886 		return write_wc.wc_error;
887 	}
888 
889 	o2hb_arm_write_timeout(reg);
890 
891 	/* let the person who launched us know when things are steady */
892 	if (!change && (atomic_read(&reg->hr_steady_iterations) != 0)) {
893 		if (atomic_dec_and_test(&reg->hr_steady_iterations))
894 			wake_up(&o2hb_steady_queue);
895 	}
896 
897 	return 0;
898 }
899 
900 /* Subtract b from a, storing the result in a. a *must* have a larger
901  * value than b. */
902 static void o2hb_tv_subtract(struct timeval *a,
903 			     struct timeval *b)
904 {
905 	/* just return 0 when a is after b */
906 	if (a->tv_sec < b->tv_sec ||
907 	    (a->tv_sec == b->tv_sec && a->tv_usec < b->tv_usec)) {
908 		a->tv_sec = 0;
909 		a->tv_usec = 0;
910 		return;
911 	}
912 
913 	a->tv_sec -= b->tv_sec;
914 	a->tv_usec -= b->tv_usec;
915 	while ( a->tv_usec < 0 ) {
916 		a->tv_sec--;
917 		a->tv_usec += 1000000;
918 	}
919 }
920 
921 static unsigned int o2hb_elapsed_msecs(struct timeval *start,
922 				       struct timeval *end)
923 {
924 	struct timeval res = *end;
925 
926 	o2hb_tv_subtract(&res, start);
927 
928 	return res.tv_sec * 1000 + res.tv_usec / 1000;
929 }
930 
931 /*
932  * we ride the region ref that the region dir holds.  before the region
933  * dir is removed and drops it ref it will wait to tear down this
934  * thread.
935  */
936 static int o2hb_thread(void *data)
937 {
938 	int i, ret;
939 	struct o2hb_region *reg = data;
940 	struct bio *write_bio;
941 	struct o2hb_bio_wait_ctxt write_wc;
942 	struct timeval before_hb, after_hb;
943 	unsigned int elapsed_msec;
944 
945 	mlog(ML_HEARTBEAT|ML_KTHREAD, "hb thread running\n");
946 
947 	set_user_nice(current, -20);
948 
949 	while (!kthread_should_stop() && !reg->hr_unclean_stop) {
950 		/* We track the time spent inside
951 		 * o2hb_do_disk_heartbeat so that we avoid more then
952 		 * hr_timeout_ms between disk writes. On busy systems
953 		 * this should result in a heartbeat which is less
954 		 * likely to time itself out. */
955 		do_gettimeofday(&before_hb);
956 
957 		i = 0;
958 		do {
959 			ret = o2hb_do_disk_heartbeat(reg);
960 		} while (ret && ++i < 2);
961 
962 		do_gettimeofday(&after_hb);
963 		elapsed_msec = o2hb_elapsed_msecs(&before_hb, &after_hb);
964 
965 		mlog(0, "start = %lu.%lu, end = %lu.%lu, msec = %u\n",
966 		     before_hb.tv_sec, (unsigned long) before_hb.tv_usec,
967 		     after_hb.tv_sec, (unsigned long) after_hb.tv_usec,
968 		     elapsed_msec);
969 
970 		if (elapsed_msec < reg->hr_timeout_ms) {
971 			/* the kthread api has blocked signals for us so no
972 			 * need to record the return value. */
973 			msleep_interruptible(reg->hr_timeout_ms - elapsed_msec);
974 		}
975 	}
976 
977 	o2hb_disarm_write_timeout(reg);
978 
979 	/* unclean stop is only used in very bad situation */
980 	for(i = 0; !reg->hr_unclean_stop && i < reg->hr_blocks; i++)
981 		o2hb_shutdown_slot(&reg->hr_slots[i]);
982 
983 	/* Explicit down notification - avoid forcing the other nodes
984 	 * to timeout on this region when we could just as easily
985 	 * write a clear generation - thus indicating to them that
986 	 * this node has left this region.
987 	 *
988 	 * XXX: Should we skip this on unclean_stop? */
989 	o2hb_prepare_block(reg, 0);
990 	ret = o2hb_issue_node_write(reg, &write_bio, &write_wc);
991 	if (ret == 0) {
992 		o2hb_wait_on_io(reg, &write_wc);
993 		bio_put(write_bio);
994 	} else {
995 		mlog_errno(ret);
996 	}
997 
998 	mlog(ML_HEARTBEAT|ML_KTHREAD, "hb thread exiting\n");
999 
1000 	return 0;
1001 }
1002 
1003 void o2hb_init(void)
1004 {
1005 	int i;
1006 
1007 	for (i = 0; i < ARRAY_SIZE(o2hb_callbacks); i++)
1008 		INIT_LIST_HEAD(&o2hb_callbacks[i].list);
1009 
1010 	for (i = 0; i < ARRAY_SIZE(o2hb_live_slots); i++)
1011 		INIT_LIST_HEAD(&o2hb_live_slots[i]);
1012 
1013 	INIT_LIST_HEAD(&o2hb_node_events);
1014 
1015 	memset(o2hb_live_node_bitmap, 0, sizeof(o2hb_live_node_bitmap));
1016 }
1017 
1018 /* if we're already in a callback then we're already serialized by the sem */
1019 static void o2hb_fill_node_map_from_callback(unsigned long *map,
1020 					     unsigned bytes)
1021 {
1022 	BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1023 
1024 	memcpy(map, &o2hb_live_node_bitmap, bytes);
1025 }
1026 
1027 /*
1028  * get a map of all nodes that are heartbeating in any regions
1029  */
1030 void o2hb_fill_node_map(unsigned long *map, unsigned bytes)
1031 {
1032 	/* callers want to serialize this map and callbacks so that they
1033 	 * can trust that they don't miss nodes coming to the party */
1034 	down_read(&o2hb_callback_sem);
1035 	spin_lock(&o2hb_live_lock);
1036 	o2hb_fill_node_map_from_callback(map, bytes);
1037 	spin_unlock(&o2hb_live_lock);
1038 	up_read(&o2hb_callback_sem);
1039 }
1040 EXPORT_SYMBOL_GPL(o2hb_fill_node_map);
1041 
1042 /*
1043  * heartbeat configfs bits.  The heartbeat set is a default set under
1044  * the cluster set in nodemanager.c.
1045  */
1046 
1047 static struct o2hb_region *to_o2hb_region(struct config_item *item)
1048 {
1049 	return item ? container_of(item, struct o2hb_region, hr_item) : NULL;
1050 }
1051 
1052 /* drop_item only drops its ref after killing the thread, nothing should
1053  * be using the region anymore.  this has to clean up any state that
1054  * attributes might have built up. */
1055 static void o2hb_region_release(struct config_item *item)
1056 {
1057 	int i;
1058 	struct page *page;
1059 	struct o2hb_region *reg = to_o2hb_region(item);
1060 
1061 	if (reg->hr_tmp_block)
1062 		kfree(reg->hr_tmp_block);
1063 
1064 	if (reg->hr_slot_data) {
1065 		for (i = 0; i < reg->hr_num_pages; i++) {
1066 			page = reg->hr_slot_data[i];
1067 			if (page)
1068 				__free_page(page);
1069 		}
1070 		kfree(reg->hr_slot_data);
1071 	}
1072 
1073 	if (reg->hr_bdev)
1074 		blkdev_put(reg->hr_bdev);
1075 
1076 	if (reg->hr_slots)
1077 		kfree(reg->hr_slots);
1078 
1079 	spin_lock(&o2hb_live_lock);
1080 	list_del(&reg->hr_all_item);
1081 	spin_unlock(&o2hb_live_lock);
1082 
1083 	kfree(reg);
1084 }
1085 
1086 static int o2hb_read_block_input(struct o2hb_region *reg,
1087 				 const char *page,
1088 				 size_t count,
1089 				 unsigned long *ret_bytes,
1090 				 unsigned int *ret_bits)
1091 {
1092 	unsigned long bytes;
1093 	char *p = (char *)page;
1094 
1095 	bytes = simple_strtoul(p, &p, 0);
1096 	if (!p || (*p && (*p != '\n')))
1097 		return -EINVAL;
1098 
1099 	/* Heartbeat and fs min / max block sizes are the same. */
1100 	if (bytes > 4096 || bytes < 512)
1101 		return -ERANGE;
1102 	if (hweight16(bytes) != 1)
1103 		return -EINVAL;
1104 
1105 	if (ret_bytes)
1106 		*ret_bytes = bytes;
1107 	if (ret_bits)
1108 		*ret_bits = ffs(bytes) - 1;
1109 
1110 	return 0;
1111 }
1112 
1113 static ssize_t o2hb_region_block_bytes_read(struct o2hb_region *reg,
1114 					    char *page)
1115 {
1116 	return sprintf(page, "%u\n", reg->hr_block_bytes);
1117 }
1118 
1119 static ssize_t o2hb_region_block_bytes_write(struct o2hb_region *reg,
1120 					     const char *page,
1121 					     size_t count)
1122 {
1123 	int status;
1124 	unsigned long block_bytes;
1125 	unsigned int block_bits;
1126 
1127 	if (reg->hr_bdev)
1128 		return -EINVAL;
1129 
1130 	status = o2hb_read_block_input(reg, page, count,
1131 				       &block_bytes, &block_bits);
1132 	if (status)
1133 		return status;
1134 
1135 	reg->hr_block_bytes = (unsigned int)block_bytes;
1136 	reg->hr_block_bits = block_bits;
1137 
1138 	return count;
1139 }
1140 
1141 static ssize_t o2hb_region_start_block_read(struct o2hb_region *reg,
1142 					    char *page)
1143 {
1144 	return sprintf(page, "%llu\n", reg->hr_start_block);
1145 }
1146 
1147 static ssize_t o2hb_region_start_block_write(struct o2hb_region *reg,
1148 					     const char *page,
1149 					     size_t count)
1150 {
1151 	unsigned long long tmp;
1152 	char *p = (char *)page;
1153 
1154 	if (reg->hr_bdev)
1155 		return -EINVAL;
1156 
1157 	tmp = simple_strtoull(p, &p, 0);
1158 	if (!p || (*p && (*p != '\n')))
1159 		return -EINVAL;
1160 
1161 	reg->hr_start_block = tmp;
1162 
1163 	return count;
1164 }
1165 
1166 static ssize_t o2hb_region_blocks_read(struct o2hb_region *reg,
1167 				       char *page)
1168 {
1169 	return sprintf(page, "%d\n", reg->hr_blocks);
1170 }
1171 
1172 static ssize_t o2hb_region_blocks_write(struct o2hb_region *reg,
1173 					const char *page,
1174 					size_t count)
1175 {
1176 	unsigned long tmp;
1177 	char *p = (char *)page;
1178 
1179 	if (reg->hr_bdev)
1180 		return -EINVAL;
1181 
1182 	tmp = simple_strtoul(p, &p, 0);
1183 	if (!p || (*p && (*p != '\n')))
1184 		return -EINVAL;
1185 
1186 	if (tmp > O2NM_MAX_NODES || tmp == 0)
1187 		return -ERANGE;
1188 
1189 	reg->hr_blocks = (unsigned int)tmp;
1190 
1191 	return count;
1192 }
1193 
1194 static ssize_t o2hb_region_dev_read(struct o2hb_region *reg,
1195 				    char *page)
1196 {
1197 	unsigned int ret = 0;
1198 
1199 	if (reg->hr_bdev)
1200 		ret = sprintf(page, "%s\n", reg->hr_dev_name);
1201 
1202 	return ret;
1203 }
1204 
1205 static void o2hb_init_region_params(struct o2hb_region *reg)
1206 {
1207 	reg->hr_slots_per_page = PAGE_CACHE_SIZE >> reg->hr_block_bits;
1208 	reg->hr_timeout_ms = O2HB_REGION_TIMEOUT_MS;
1209 
1210 	mlog(ML_HEARTBEAT, "hr_start_block = %llu, hr_blocks = %u\n",
1211 	     reg->hr_start_block, reg->hr_blocks);
1212 	mlog(ML_HEARTBEAT, "hr_block_bytes = %u, hr_block_bits = %u\n",
1213 	     reg->hr_block_bytes, reg->hr_block_bits);
1214 	mlog(ML_HEARTBEAT, "hr_timeout_ms = %u\n", reg->hr_timeout_ms);
1215 	mlog(ML_HEARTBEAT, "dead threshold = %u\n", o2hb_dead_threshold);
1216 }
1217 
1218 static int o2hb_map_slot_data(struct o2hb_region *reg)
1219 {
1220 	int i, j;
1221 	unsigned int last_slot;
1222 	unsigned int spp = reg->hr_slots_per_page;
1223 	struct page *page;
1224 	char *raw;
1225 	struct o2hb_disk_slot *slot;
1226 
1227 	reg->hr_tmp_block = kmalloc(reg->hr_block_bytes, GFP_KERNEL);
1228 	if (reg->hr_tmp_block == NULL) {
1229 		mlog_errno(-ENOMEM);
1230 		return -ENOMEM;
1231 	}
1232 
1233 	reg->hr_slots = kcalloc(reg->hr_blocks,
1234 				sizeof(struct o2hb_disk_slot), GFP_KERNEL);
1235 	if (reg->hr_slots == NULL) {
1236 		mlog_errno(-ENOMEM);
1237 		return -ENOMEM;
1238 	}
1239 
1240 	for(i = 0; i < reg->hr_blocks; i++) {
1241 		slot = &reg->hr_slots[i];
1242 		slot->ds_node_num = i;
1243 		INIT_LIST_HEAD(&slot->ds_live_item);
1244 		slot->ds_raw_block = NULL;
1245 	}
1246 
1247 	reg->hr_num_pages = (reg->hr_blocks + spp - 1) / spp;
1248 	mlog(ML_HEARTBEAT, "Going to require %u pages to cover %u blocks "
1249 			   "at %u blocks per page\n",
1250 	     reg->hr_num_pages, reg->hr_blocks, spp);
1251 
1252 	reg->hr_slot_data = kcalloc(reg->hr_num_pages, sizeof(struct page *),
1253 				    GFP_KERNEL);
1254 	if (!reg->hr_slot_data) {
1255 		mlog_errno(-ENOMEM);
1256 		return -ENOMEM;
1257 	}
1258 
1259 	for(i = 0; i < reg->hr_num_pages; i++) {
1260 		page = alloc_page(GFP_KERNEL);
1261 		if (!page) {
1262 			mlog_errno(-ENOMEM);
1263 			return -ENOMEM;
1264 		}
1265 
1266 		reg->hr_slot_data[i] = page;
1267 
1268 		last_slot = i * spp;
1269 		raw = page_address(page);
1270 		for (j = 0;
1271 		     (j < spp) && ((j + last_slot) < reg->hr_blocks);
1272 		     j++) {
1273 			BUG_ON((j + last_slot) >= reg->hr_blocks);
1274 
1275 			slot = &reg->hr_slots[j + last_slot];
1276 			slot->ds_raw_block =
1277 				(struct o2hb_disk_heartbeat_block *) raw;
1278 
1279 			raw += reg->hr_block_bytes;
1280 		}
1281 	}
1282 
1283 	return 0;
1284 }
1285 
1286 /* Read in all the slots available and populate the tracking
1287  * structures so that we can start with a baseline idea of what's
1288  * there. */
1289 static int o2hb_populate_slot_data(struct o2hb_region *reg)
1290 {
1291 	int ret, i;
1292 	struct o2hb_disk_slot *slot;
1293 	struct o2hb_disk_heartbeat_block *hb_block;
1294 
1295 	mlog_entry_void();
1296 
1297 	ret = o2hb_read_slots(reg, reg->hr_blocks);
1298 	if (ret) {
1299 		mlog_errno(ret);
1300 		goto out;
1301 	}
1302 
1303 	/* We only want to get an idea of the values initially in each
1304 	 * slot, so we do no verification - o2hb_check_slot will
1305 	 * actually determine if each configured slot is valid and
1306 	 * whether any values have changed. */
1307 	for(i = 0; i < reg->hr_blocks; i++) {
1308 		slot = &reg->hr_slots[i];
1309 		hb_block = (struct o2hb_disk_heartbeat_block *) slot->ds_raw_block;
1310 
1311 		/* Only fill the values that o2hb_check_slot uses to
1312 		 * determine changing slots */
1313 		slot->ds_last_time = le64_to_cpu(hb_block->hb_seq);
1314 		slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation);
1315 	}
1316 
1317 out:
1318 	mlog_exit(ret);
1319 	return ret;
1320 }
1321 
1322 /* this is acting as commit; we set up all of hr_bdev and hr_task or nothing */
1323 static ssize_t o2hb_region_dev_write(struct o2hb_region *reg,
1324 				     const char *page,
1325 				     size_t count)
1326 {
1327 	long fd;
1328 	int sectsize;
1329 	char *p = (char *)page;
1330 	struct file *filp = NULL;
1331 	struct inode *inode = NULL;
1332 	ssize_t ret = -EINVAL;
1333 
1334 	if (reg->hr_bdev)
1335 		goto out;
1336 
1337 	/* We can't heartbeat without having had our node number
1338 	 * configured yet. */
1339 	if (o2nm_this_node() == O2NM_MAX_NODES)
1340 		goto out;
1341 
1342 	fd = simple_strtol(p, &p, 0);
1343 	if (!p || (*p && (*p != '\n')))
1344 		goto out;
1345 
1346 	if (fd < 0 || fd >= INT_MAX)
1347 		goto out;
1348 
1349 	filp = fget(fd);
1350 	if (filp == NULL)
1351 		goto out;
1352 
1353 	if (reg->hr_blocks == 0 || reg->hr_start_block == 0 ||
1354 	    reg->hr_block_bytes == 0)
1355 		goto out;
1356 
1357 	inode = igrab(filp->f_mapping->host);
1358 	if (inode == NULL)
1359 		goto out;
1360 
1361 	if (!S_ISBLK(inode->i_mode))
1362 		goto out;
1363 
1364 	reg->hr_bdev = I_BDEV(filp->f_mapping->host);
1365 	ret = blkdev_get(reg->hr_bdev, FMODE_WRITE | FMODE_READ, 0);
1366 	if (ret) {
1367 		reg->hr_bdev = NULL;
1368 		goto out;
1369 	}
1370 	inode = NULL;
1371 
1372 	bdevname(reg->hr_bdev, reg->hr_dev_name);
1373 
1374 	sectsize = bdev_hardsect_size(reg->hr_bdev);
1375 	if (sectsize != reg->hr_block_bytes) {
1376 		mlog(ML_ERROR,
1377 		     "blocksize %u incorrect for device, expected %d",
1378 		     reg->hr_block_bytes, sectsize);
1379 		ret = -EINVAL;
1380 		goto out;
1381 	}
1382 
1383 	o2hb_init_region_params(reg);
1384 
1385 	/* Generation of zero is invalid */
1386 	do {
1387 		get_random_bytes(&reg->hr_generation,
1388 				 sizeof(reg->hr_generation));
1389 	} while (reg->hr_generation == 0);
1390 
1391 	ret = o2hb_map_slot_data(reg);
1392 	if (ret) {
1393 		mlog_errno(ret);
1394 		goto out;
1395 	}
1396 
1397 	ret = o2hb_populate_slot_data(reg);
1398 	if (ret) {
1399 		mlog_errno(ret);
1400 		goto out;
1401 	}
1402 
1403 	INIT_WORK(&reg->hr_write_timeout_work, o2hb_write_timeout, reg);
1404 
1405 	/*
1406 	 * A node is considered live after it has beat LIVE_THRESHOLD
1407 	 * times.  We're not steady until we've given them a chance
1408 	 * _after_ our first read.
1409 	 */
1410 	atomic_set(&reg->hr_steady_iterations, O2HB_LIVE_THRESHOLD + 1);
1411 
1412 	reg->hr_task = kthread_run(o2hb_thread, reg, "o2hb-%s",
1413 				   reg->hr_item.ci_name);
1414 	if (IS_ERR(reg->hr_task)) {
1415 		ret = PTR_ERR(reg->hr_task);
1416 		mlog_errno(ret);
1417 		reg->hr_task = NULL;
1418 		goto out;
1419 	}
1420 
1421 	ret = wait_event_interruptible(o2hb_steady_queue,
1422 				atomic_read(&reg->hr_steady_iterations) == 0);
1423 	if (ret) {
1424 		kthread_stop(reg->hr_task);
1425 		reg->hr_task = NULL;
1426 		goto out;
1427 	}
1428 
1429 	ret = count;
1430 out:
1431 	if (filp)
1432 		fput(filp);
1433 	if (inode)
1434 		iput(inode);
1435 	if (ret < 0) {
1436 		if (reg->hr_bdev) {
1437 			blkdev_put(reg->hr_bdev);
1438 			reg->hr_bdev = NULL;
1439 		}
1440 	}
1441 	return ret;
1442 }
1443 
1444 struct o2hb_region_attribute {
1445 	struct configfs_attribute attr;
1446 	ssize_t (*show)(struct o2hb_region *, char *);
1447 	ssize_t (*store)(struct o2hb_region *, const char *, size_t);
1448 };
1449 
1450 static struct o2hb_region_attribute o2hb_region_attr_block_bytes = {
1451 	.attr	= { .ca_owner = THIS_MODULE,
1452 		    .ca_name = "block_bytes",
1453 		    .ca_mode = S_IRUGO | S_IWUSR },
1454 	.show	= o2hb_region_block_bytes_read,
1455 	.store	= o2hb_region_block_bytes_write,
1456 };
1457 
1458 static struct o2hb_region_attribute o2hb_region_attr_start_block = {
1459 	.attr	= { .ca_owner = THIS_MODULE,
1460 		    .ca_name = "start_block",
1461 		    .ca_mode = S_IRUGO | S_IWUSR },
1462 	.show	= o2hb_region_start_block_read,
1463 	.store	= o2hb_region_start_block_write,
1464 };
1465 
1466 static struct o2hb_region_attribute o2hb_region_attr_blocks = {
1467 	.attr	= { .ca_owner = THIS_MODULE,
1468 		    .ca_name = "blocks",
1469 		    .ca_mode = S_IRUGO | S_IWUSR },
1470 	.show	= o2hb_region_blocks_read,
1471 	.store	= o2hb_region_blocks_write,
1472 };
1473 
1474 static struct o2hb_region_attribute o2hb_region_attr_dev = {
1475 	.attr	= { .ca_owner = THIS_MODULE,
1476 		    .ca_name = "dev",
1477 		    .ca_mode = S_IRUGO | S_IWUSR },
1478 	.show	= o2hb_region_dev_read,
1479 	.store	= o2hb_region_dev_write,
1480 };
1481 
1482 static struct configfs_attribute *o2hb_region_attrs[] = {
1483 	&o2hb_region_attr_block_bytes.attr,
1484 	&o2hb_region_attr_start_block.attr,
1485 	&o2hb_region_attr_blocks.attr,
1486 	&o2hb_region_attr_dev.attr,
1487 	NULL,
1488 };
1489 
1490 static ssize_t o2hb_region_show(struct config_item *item,
1491 				struct configfs_attribute *attr,
1492 				char *page)
1493 {
1494 	struct o2hb_region *reg = to_o2hb_region(item);
1495 	struct o2hb_region_attribute *o2hb_region_attr =
1496 		container_of(attr, struct o2hb_region_attribute, attr);
1497 	ssize_t ret = 0;
1498 
1499 	if (o2hb_region_attr->show)
1500 		ret = o2hb_region_attr->show(reg, page);
1501 	return ret;
1502 }
1503 
1504 static ssize_t o2hb_region_store(struct config_item *item,
1505 				 struct configfs_attribute *attr,
1506 				 const char *page, size_t count)
1507 {
1508 	struct o2hb_region *reg = to_o2hb_region(item);
1509 	struct o2hb_region_attribute *o2hb_region_attr =
1510 		container_of(attr, struct o2hb_region_attribute, attr);
1511 	ssize_t ret = -EINVAL;
1512 
1513 	if (o2hb_region_attr->store)
1514 		ret = o2hb_region_attr->store(reg, page, count);
1515 	return ret;
1516 }
1517 
1518 static struct configfs_item_operations o2hb_region_item_ops = {
1519 	.release		= o2hb_region_release,
1520 	.show_attribute		= o2hb_region_show,
1521 	.store_attribute	= o2hb_region_store,
1522 };
1523 
1524 static struct config_item_type o2hb_region_type = {
1525 	.ct_item_ops	= &o2hb_region_item_ops,
1526 	.ct_attrs	= o2hb_region_attrs,
1527 	.ct_owner	= THIS_MODULE,
1528 };
1529 
1530 /* heartbeat set */
1531 
1532 struct o2hb_heartbeat_group {
1533 	struct config_group hs_group;
1534 	/* some stuff? */
1535 };
1536 
1537 static struct o2hb_heartbeat_group *to_o2hb_heartbeat_group(struct config_group *group)
1538 {
1539 	return group ?
1540 		container_of(group, struct o2hb_heartbeat_group, hs_group)
1541 		: NULL;
1542 }
1543 
1544 static struct config_item *o2hb_heartbeat_group_make_item(struct config_group *group,
1545 							  const char *name)
1546 {
1547 	struct o2hb_region *reg = NULL;
1548 	struct config_item *ret = NULL;
1549 
1550 	reg = kcalloc(1, sizeof(struct o2hb_region), GFP_KERNEL);
1551 	if (reg == NULL)
1552 		goto out; /* ENOMEM */
1553 
1554 	config_item_init_type_name(&reg->hr_item, name, &o2hb_region_type);
1555 
1556 	ret = &reg->hr_item;
1557 
1558 	spin_lock(&o2hb_live_lock);
1559 	list_add_tail(&reg->hr_all_item, &o2hb_all_regions);
1560 	spin_unlock(&o2hb_live_lock);
1561 out:
1562 	if (ret == NULL)
1563 		kfree(reg);
1564 
1565 	return ret;
1566 }
1567 
1568 static void o2hb_heartbeat_group_drop_item(struct config_group *group,
1569 					   struct config_item *item)
1570 {
1571 	struct o2hb_region *reg = to_o2hb_region(item);
1572 
1573 	/* stop the thread when the user removes the region dir */
1574 	if (reg->hr_task) {
1575 		kthread_stop(reg->hr_task);
1576 		reg->hr_task = NULL;
1577 	}
1578 
1579 	config_item_put(item);
1580 }
1581 
1582 struct o2hb_heartbeat_group_attribute {
1583 	struct configfs_attribute attr;
1584 	ssize_t (*show)(struct o2hb_heartbeat_group *, char *);
1585 	ssize_t (*store)(struct o2hb_heartbeat_group *, const char *, size_t);
1586 };
1587 
1588 static ssize_t o2hb_heartbeat_group_show(struct config_item *item,
1589 					 struct configfs_attribute *attr,
1590 					 char *page)
1591 {
1592 	struct o2hb_heartbeat_group *reg = to_o2hb_heartbeat_group(to_config_group(item));
1593 	struct o2hb_heartbeat_group_attribute *o2hb_heartbeat_group_attr =
1594 		container_of(attr, struct o2hb_heartbeat_group_attribute, attr);
1595 	ssize_t ret = 0;
1596 
1597 	if (o2hb_heartbeat_group_attr->show)
1598 		ret = o2hb_heartbeat_group_attr->show(reg, page);
1599 	return ret;
1600 }
1601 
1602 static ssize_t o2hb_heartbeat_group_store(struct config_item *item,
1603 					  struct configfs_attribute *attr,
1604 					  const char *page, size_t count)
1605 {
1606 	struct o2hb_heartbeat_group *reg = to_o2hb_heartbeat_group(to_config_group(item));
1607 	struct o2hb_heartbeat_group_attribute *o2hb_heartbeat_group_attr =
1608 		container_of(attr, struct o2hb_heartbeat_group_attribute, attr);
1609 	ssize_t ret = -EINVAL;
1610 
1611 	if (o2hb_heartbeat_group_attr->store)
1612 		ret = o2hb_heartbeat_group_attr->store(reg, page, count);
1613 	return ret;
1614 }
1615 
1616 static ssize_t o2hb_heartbeat_group_threshold_show(struct o2hb_heartbeat_group *group,
1617 						     char *page)
1618 {
1619 	return sprintf(page, "%u\n", o2hb_dead_threshold);
1620 }
1621 
1622 static ssize_t o2hb_heartbeat_group_threshold_store(struct o2hb_heartbeat_group *group,
1623 						    const char *page,
1624 						    size_t count)
1625 {
1626 	unsigned long tmp;
1627 	char *p = (char *)page;
1628 
1629 	tmp = simple_strtoul(p, &p, 10);
1630 	if (!p || (*p && (*p != '\n')))
1631                 return -EINVAL;
1632 
1633 	/* this will validate ranges for us. */
1634 	o2hb_dead_threshold_set((unsigned int) tmp);
1635 
1636 	return count;
1637 }
1638 
1639 static struct o2hb_heartbeat_group_attribute o2hb_heartbeat_group_attr_threshold = {
1640 	.attr	= { .ca_owner = THIS_MODULE,
1641 		    .ca_name = "dead_threshold",
1642 		    .ca_mode = S_IRUGO | S_IWUSR },
1643 	.show	= o2hb_heartbeat_group_threshold_show,
1644 	.store	= o2hb_heartbeat_group_threshold_store,
1645 };
1646 
1647 static struct configfs_attribute *o2hb_heartbeat_group_attrs[] = {
1648 	&o2hb_heartbeat_group_attr_threshold.attr,
1649 	NULL,
1650 };
1651 
1652 static struct configfs_item_operations o2hb_hearbeat_group_item_ops = {
1653 	.show_attribute		= o2hb_heartbeat_group_show,
1654 	.store_attribute	= o2hb_heartbeat_group_store,
1655 };
1656 
1657 static struct configfs_group_operations o2hb_heartbeat_group_group_ops = {
1658 	.make_item	= o2hb_heartbeat_group_make_item,
1659 	.drop_item	= o2hb_heartbeat_group_drop_item,
1660 };
1661 
1662 static struct config_item_type o2hb_heartbeat_group_type = {
1663 	.ct_group_ops	= &o2hb_heartbeat_group_group_ops,
1664 	.ct_item_ops	= &o2hb_hearbeat_group_item_ops,
1665 	.ct_attrs	= o2hb_heartbeat_group_attrs,
1666 	.ct_owner	= THIS_MODULE,
1667 };
1668 
1669 /* this is just here to avoid touching group in heartbeat.h which the
1670  * entire damn world #includes */
1671 struct config_group *o2hb_alloc_hb_set(void)
1672 {
1673 	struct o2hb_heartbeat_group *hs = NULL;
1674 	struct config_group *ret = NULL;
1675 
1676 	hs = kcalloc(1, sizeof(struct o2hb_heartbeat_group), GFP_KERNEL);
1677 	if (hs == NULL)
1678 		goto out;
1679 
1680 	config_group_init_type_name(&hs->hs_group, "heartbeat",
1681 				    &o2hb_heartbeat_group_type);
1682 
1683 	ret = &hs->hs_group;
1684 out:
1685 	if (ret == NULL)
1686 		kfree(hs);
1687 	return ret;
1688 }
1689 
1690 void o2hb_free_hb_set(struct config_group *group)
1691 {
1692 	struct o2hb_heartbeat_group *hs = to_o2hb_heartbeat_group(group);
1693 	kfree(hs);
1694 }
1695 
1696 /* hb callback registration and issueing */
1697 
1698 static struct o2hb_callback *hbcall_from_type(enum o2hb_callback_type type)
1699 {
1700 	if (type == O2HB_NUM_CB)
1701 		return ERR_PTR(-EINVAL);
1702 
1703 	return &o2hb_callbacks[type];
1704 }
1705 
1706 void o2hb_setup_callback(struct o2hb_callback_func *hc,
1707 			 enum o2hb_callback_type type,
1708 			 o2hb_cb_func *func,
1709 			 void *data,
1710 			 int priority)
1711 {
1712 	INIT_LIST_HEAD(&hc->hc_item);
1713 	hc->hc_func = func;
1714 	hc->hc_data = data;
1715 	hc->hc_priority = priority;
1716 	hc->hc_type = type;
1717 	hc->hc_magic = O2HB_CB_MAGIC;
1718 }
1719 EXPORT_SYMBOL_GPL(o2hb_setup_callback);
1720 
1721 int o2hb_register_callback(struct o2hb_callback_func *hc)
1722 {
1723 	struct o2hb_callback_func *tmp;
1724 	struct list_head *iter;
1725 	struct o2hb_callback *hbcall;
1726 	int ret;
1727 
1728 	BUG_ON(hc->hc_magic != O2HB_CB_MAGIC);
1729 	BUG_ON(!list_empty(&hc->hc_item));
1730 
1731 	hbcall = hbcall_from_type(hc->hc_type);
1732 	if (IS_ERR(hbcall)) {
1733 		ret = PTR_ERR(hbcall);
1734 		goto out;
1735 	}
1736 
1737 	down_write(&o2hb_callback_sem);
1738 
1739 	list_for_each(iter, &hbcall->list) {
1740 		tmp = list_entry(iter, struct o2hb_callback_func, hc_item);
1741 		if (hc->hc_priority < tmp->hc_priority) {
1742 			list_add_tail(&hc->hc_item, iter);
1743 			break;
1744 		}
1745 	}
1746 	if (list_empty(&hc->hc_item))
1747 		list_add_tail(&hc->hc_item, &hbcall->list);
1748 
1749 	up_write(&o2hb_callback_sem);
1750 	ret = 0;
1751 out:
1752 	mlog(ML_HEARTBEAT, "returning %d on behalf of %p for funcs %p\n",
1753 	     ret, __builtin_return_address(0), hc);
1754 	return ret;
1755 }
1756 EXPORT_SYMBOL_GPL(o2hb_register_callback);
1757 
1758 int o2hb_unregister_callback(struct o2hb_callback_func *hc)
1759 {
1760 	BUG_ON(hc->hc_magic != O2HB_CB_MAGIC);
1761 
1762 	mlog(ML_HEARTBEAT, "on behalf of %p for funcs %p\n",
1763 	     __builtin_return_address(0), hc);
1764 
1765 	if (list_empty(&hc->hc_item))
1766 		return 0;
1767 
1768 	down_write(&o2hb_callback_sem);
1769 
1770 	list_del_init(&hc->hc_item);
1771 
1772 	up_write(&o2hb_callback_sem);
1773 
1774 	return 0;
1775 }
1776 EXPORT_SYMBOL_GPL(o2hb_unregister_callback);
1777 
1778 int o2hb_check_node_heartbeating(u8 node_num)
1779 {
1780 	unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1781 
1782 	o2hb_fill_node_map(testing_map, sizeof(testing_map));
1783 	if (!test_bit(node_num, testing_map)) {
1784 		mlog(ML_HEARTBEAT,
1785 		     "node (%u) does not have heartbeating enabled.\n",
1786 		     node_num);
1787 		return 0;
1788 	}
1789 
1790 	return 1;
1791 }
1792 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating);
1793 
1794 int o2hb_check_node_heartbeating_from_callback(u8 node_num)
1795 {
1796 	unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1797 
1798 	o2hb_fill_node_map_from_callback(testing_map, sizeof(testing_map));
1799 	if (!test_bit(node_num, testing_map)) {
1800 		mlog(ML_HEARTBEAT,
1801 		     "node (%u) does not have heartbeating enabled.\n",
1802 		     node_num);
1803 		return 0;
1804 	}
1805 
1806 	return 1;
1807 }
1808 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating_from_callback);
1809 
1810 /* Makes sure our local node is configured with a node number, and is
1811  * heartbeating. */
1812 int o2hb_check_local_node_heartbeating(void)
1813 {
1814 	u8 node_num;
1815 
1816 	/* if this node was set then we have networking */
1817 	node_num = o2nm_this_node();
1818 	if (node_num == O2NM_MAX_NODES) {
1819 		mlog(ML_HEARTBEAT, "this node has not been configured.\n");
1820 		return 0;
1821 	}
1822 
1823 	return o2hb_check_node_heartbeating(node_num);
1824 }
1825 EXPORT_SYMBOL_GPL(o2hb_check_local_node_heartbeating);
1826 
1827 /*
1828  * this is just a hack until we get the plumbing which flips file systems
1829  * read only and drops the hb ref instead of killing the node dead.
1830  */
1831 void o2hb_stop_all_regions(void)
1832 {
1833 	struct o2hb_region *reg;
1834 
1835 	mlog(ML_ERROR, "stopping heartbeat on all active regions.\n");
1836 
1837 	spin_lock(&o2hb_live_lock);
1838 
1839 	list_for_each_entry(reg, &o2hb_all_regions, hr_all_item)
1840 		reg->hr_unclean_stop = 1;
1841 
1842 	spin_unlock(&o2hb_live_lock);
1843 }
1844 EXPORT_SYMBOL_GPL(o2hb_stop_all_regions);
1845