1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
4 * Copyright (C) 2004, 2005 Oracle. All rights reserved.
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.
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.
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.
22 #include <linux/kernel.h>
23 #include <linux/sched.h>
24 #include <linux/jiffies.h>
25 #include <linux/module.h>
27 #include <linux/bio.h>
28 #include <linux/blkdev.h>
29 #include <linux/delay.h>
30 #include <linux/file.h>
31 #include <linux/kthread.h>
32 #include <linux/configfs.h>
33 #include <linux/random.h>
34 #include <linux/crc32.h>
35 #include <linux/time.h>
36 #include <linux/debugfs.h>
37 #include <linux/slab.h>
39 #include "heartbeat.h"
41 #include "nodemanager.h"
48 * The first heartbeat pass had one global thread that would serialize all hb
49 * callback calls. This global serializing sem should only be removed once
50 * we've made sure that all callees can deal with being called concurrently
51 * from multiple hb region threads.
53 static DECLARE_RWSEM(o2hb_callback_sem);
56 * multiple hb threads are watching multiple regions. A node is live
57 * whenever any of the threads sees activity from the node in its region.
59 static DEFINE_SPINLOCK(o2hb_live_lock);
60 static struct list_head o2hb_live_slots[O2NM_MAX_NODES];
61 static unsigned long o2hb_live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
62 static LIST_HEAD(o2hb_node_events);
63 static DECLARE_WAIT_QUEUE_HEAD(o2hb_steady_queue);
66 * In global heartbeat, we maintain a series of region bitmaps.
67 * - o2hb_region_bitmap allows us to limit the region number to max region.
68 * - o2hb_live_region_bitmap tracks live regions (seen steady iterations).
69 * - o2hb_quorum_region_bitmap tracks live regions that have seen all nodes
71 * - o2hb_failed_region_bitmap tracks the regions that have seen io timeouts.
73 static unsigned long o2hb_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)];
74 static unsigned long o2hb_live_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)];
75 static unsigned long o2hb_quorum_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)];
76 static unsigned long o2hb_failed_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)];
78 #define O2HB_DB_TYPE_LIVENODES 0
79 #define O2HB_DB_TYPE_LIVEREGIONS 1
80 #define O2HB_DB_TYPE_QUORUMREGIONS 2
81 #define O2HB_DB_TYPE_FAILEDREGIONS 3
82 #define O2HB_DB_TYPE_REGION_LIVENODES 4
83 #define O2HB_DB_TYPE_REGION_NUMBER 5
84 #define O2HB_DB_TYPE_REGION_ELAPSED_TIME 6
85 struct o2hb_debug_buf {
92 static struct o2hb_debug_buf *o2hb_db_livenodes;
93 static struct o2hb_debug_buf *o2hb_db_liveregions;
94 static struct o2hb_debug_buf *o2hb_db_quorumregions;
95 static struct o2hb_debug_buf *o2hb_db_failedregions;
97 #define O2HB_DEBUG_DIR "o2hb"
98 #define O2HB_DEBUG_LIVENODES "livenodes"
99 #define O2HB_DEBUG_LIVEREGIONS "live_regions"
100 #define O2HB_DEBUG_QUORUMREGIONS "quorum_regions"
101 #define O2HB_DEBUG_FAILEDREGIONS "failed_regions"
102 #define O2HB_DEBUG_REGION_NUMBER "num"
103 #define O2HB_DEBUG_REGION_ELAPSED_TIME "elapsed_time_in_ms"
105 static struct dentry *o2hb_debug_dir;
106 static struct dentry *o2hb_debug_livenodes;
107 static struct dentry *o2hb_debug_liveregions;
108 static struct dentry *o2hb_debug_quorumregions;
109 static struct dentry *o2hb_debug_failedregions;
111 static LIST_HEAD(o2hb_all_regions);
113 static struct o2hb_callback {
114 struct list_head list;
115 } o2hb_callbacks[O2HB_NUM_CB];
117 static struct o2hb_callback *hbcall_from_type(enum o2hb_callback_type type);
119 #define O2HB_DEFAULT_BLOCK_BITS 9
121 enum o2hb_heartbeat_modes {
122 O2HB_HEARTBEAT_LOCAL = 0,
123 O2HB_HEARTBEAT_GLOBAL,
124 O2HB_HEARTBEAT_NUM_MODES,
127 char *o2hb_heartbeat_mode_desc[O2HB_HEARTBEAT_NUM_MODES] = {
128 "local", /* O2HB_HEARTBEAT_LOCAL */
129 "global", /* O2HB_HEARTBEAT_GLOBAL */
132 unsigned int o2hb_dead_threshold = O2HB_DEFAULT_DEAD_THRESHOLD;
133 unsigned int o2hb_heartbeat_mode = O2HB_HEARTBEAT_LOCAL;
135 /* Only sets a new threshold if there are no active regions.
137 * No locking or otherwise interesting code is required for reading
138 * o2hb_dead_threshold as it can't change once regions are active and
139 * it's not interesting to anyone until then anyway. */
140 static void o2hb_dead_threshold_set(unsigned int threshold)
142 if (threshold > O2HB_MIN_DEAD_THRESHOLD) {
143 spin_lock(&o2hb_live_lock);
144 if (list_empty(&o2hb_all_regions))
145 o2hb_dead_threshold = threshold;
146 spin_unlock(&o2hb_live_lock);
150 static int o2hb_global_hearbeat_mode_set(unsigned int hb_mode)
154 if (hb_mode < O2HB_HEARTBEAT_NUM_MODES) {
155 spin_lock(&o2hb_live_lock);
156 if (list_empty(&o2hb_all_regions)) {
157 o2hb_heartbeat_mode = hb_mode;
160 spin_unlock(&o2hb_live_lock);
166 struct o2hb_node_event {
167 struct list_head hn_item;
168 enum o2hb_callback_type hn_event_type;
169 struct o2nm_node *hn_node;
173 struct o2hb_disk_slot {
174 struct o2hb_disk_heartbeat_block *ds_raw_block;
177 u64 ds_last_generation;
178 u16 ds_equal_samples;
179 u16 ds_changed_samples;
180 struct list_head ds_live_item;
183 /* each thread owns a region.. when we're asked to tear down the region
184 * we ask the thread to stop, who cleans up the region */
186 struct config_item hr_item;
188 struct list_head hr_all_item;
189 unsigned hr_unclean_stop:1;
191 /* protected by the hr_callback_sem */
192 struct task_struct *hr_task;
194 unsigned int hr_blocks;
195 unsigned long long hr_start_block;
197 unsigned int hr_block_bits;
198 unsigned int hr_block_bytes;
200 unsigned int hr_slots_per_page;
201 unsigned int hr_num_pages;
203 struct page **hr_slot_data;
204 struct block_device *hr_bdev;
205 struct o2hb_disk_slot *hr_slots;
207 /* live node map of this region */
208 unsigned long hr_live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
209 unsigned int hr_region_num;
211 struct dentry *hr_debug_dir;
212 struct dentry *hr_debug_livenodes;
213 struct dentry *hr_debug_regnum;
214 struct dentry *hr_debug_elapsed_time;
215 struct o2hb_debug_buf *hr_db_livenodes;
216 struct o2hb_debug_buf *hr_db_regnum;
217 struct o2hb_debug_buf *hr_db_elapsed_time;
219 /* let the person setting up hb wait for it to return until it
220 * has reached a 'steady' state. This will be fixed when we have
221 * a more complete api that doesn't lead to this sort of fragility. */
222 atomic_t hr_steady_iterations;
224 char hr_dev_name[BDEVNAME_SIZE];
226 unsigned int hr_timeout_ms;
228 /* randomized as the region goes up and down so that a node
229 * recognizes a node going up and down in one iteration */
232 struct delayed_work hr_write_timeout_work;
233 unsigned long hr_last_timeout_start;
235 /* Used during o2hb_check_slot to hold a copy of the block
236 * being checked because we temporarily have to zero out the
238 struct o2hb_disk_heartbeat_block *hr_tmp_block;
241 struct o2hb_bio_wait_ctxt {
242 atomic_t wc_num_reqs;
243 struct completion wc_io_complete;
247 static int o2hb_pop_count(void *map, int count)
251 while ((i = find_next_bit(map, count, i + 1)) < count)
256 static void o2hb_write_timeout(struct work_struct *work)
260 struct o2hb_region *reg =
261 container_of(work, struct o2hb_region,
262 hr_write_timeout_work.work);
264 mlog(ML_ERROR, "Heartbeat write timeout to device %s after %u "
265 "milliseconds\n", reg->hr_dev_name,
266 jiffies_to_msecs(jiffies - reg->hr_last_timeout_start));
268 if (o2hb_global_heartbeat_active()) {
269 spin_lock_irqsave(&o2hb_live_lock, flags);
270 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap))
271 set_bit(reg->hr_region_num, o2hb_failed_region_bitmap);
272 failed = o2hb_pop_count(&o2hb_failed_region_bitmap,
274 quorum = o2hb_pop_count(&o2hb_quorum_region_bitmap,
276 spin_unlock_irqrestore(&o2hb_live_lock, flags);
278 mlog(ML_HEARTBEAT, "Number of regions %d, failed regions %d\n",
282 * Fence if the number of failed regions >= half the number
285 if ((failed << 1) < quorum)
289 o2quo_disk_timeout();
292 static void o2hb_arm_write_timeout(struct o2hb_region *reg)
294 mlog(ML_HEARTBEAT, "Queue write timeout for %u ms\n",
295 O2HB_MAX_WRITE_TIMEOUT_MS);
297 if (o2hb_global_heartbeat_active()) {
298 spin_lock(&o2hb_live_lock);
299 clear_bit(reg->hr_region_num, o2hb_failed_region_bitmap);
300 spin_unlock(&o2hb_live_lock);
302 cancel_delayed_work(®->hr_write_timeout_work);
303 reg->hr_last_timeout_start = jiffies;
304 schedule_delayed_work(®->hr_write_timeout_work,
305 msecs_to_jiffies(O2HB_MAX_WRITE_TIMEOUT_MS));
308 static void o2hb_disarm_write_timeout(struct o2hb_region *reg)
310 cancel_delayed_work_sync(®->hr_write_timeout_work);
313 static inline void o2hb_bio_wait_init(struct o2hb_bio_wait_ctxt *wc)
315 atomic_set(&wc->wc_num_reqs, 1);
316 init_completion(&wc->wc_io_complete);
320 /* Used in error paths too */
321 static inline void o2hb_bio_wait_dec(struct o2hb_bio_wait_ctxt *wc,
324 /* sadly atomic_sub_and_test() isn't available on all platforms. The
325 * good news is that the fast path only completes one at a time */
327 if (atomic_dec_and_test(&wc->wc_num_reqs)) {
329 complete(&wc->wc_io_complete);
334 static void o2hb_wait_on_io(struct o2hb_region *reg,
335 struct o2hb_bio_wait_ctxt *wc)
337 struct address_space *mapping = reg->hr_bdev->bd_inode->i_mapping;
339 blk_run_address_space(mapping);
340 o2hb_bio_wait_dec(wc, 1);
342 wait_for_completion(&wc->wc_io_complete);
345 static void o2hb_bio_end_io(struct bio *bio,
348 struct o2hb_bio_wait_ctxt *wc = bio->bi_private;
351 mlog(ML_ERROR, "IO Error %d\n", error);
352 wc->wc_error = error;
355 o2hb_bio_wait_dec(wc, 1);
359 /* Setup a Bio to cover I/O against num_slots slots starting at
361 static struct bio *o2hb_setup_one_bio(struct o2hb_region *reg,
362 struct o2hb_bio_wait_ctxt *wc,
363 unsigned int *current_slot,
364 unsigned int max_slots)
366 int len, current_page;
367 unsigned int vec_len, vec_start;
368 unsigned int bits = reg->hr_block_bits;
369 unsigned int spp = reg->hr_slots_per_page;
370 unsigned int cs = *current_slot;
374 /* Testing has shown this allocation to take long enough under
375 * GFP_KERNEL that the local node can get fenced. It would be
376 * nicest if we could pre-allocate these bios and avoid this
378 bio = bio_alloc(GFP_ATOMIC, 16);
380 mlog(ML_ERROR, "Could not alloc slots BIO!\n");
381 bio = ERR_PTR(-ENOMEM);
385 /* Must put everything in 512 byte sectors for the bio... */
386 bio->bi_sector = (reg->hr_start_block + cs) << (bits - 9);
387 bio->bi_bdev = reg->hr_bdev;
388 bio->bi_private = wc;
389 bio->bi_end_io = o2hb_bio_end_io;
391 vec_start = (cs << bits) % PAGE_CACHE_SIZE;
392 while(cs < max_slots) {
393 current_page = cs / spp;
394 page = reg->hr_slot_data[current_page];
396 vec_len = min(PAGE_CACHE_SIZE - vec_start,
397 (max_slots-cs) * (PAGE_CACHE_SIZE/spp) );
399 mlog(ML_HB_BIO, "page %d, vec_len = %u, vec_start = %u\n",
400 current_page, vec_len, vec_start);
402 len = bio_add_page(bio, page, vec_len, vec_start);
403 if (len != vec_len) break;
405 cs += vec_len / (PAGE_CACHE_SIZE/spp);
414 static int o2hb_read_slots(struct o2hb_region *reg,
415 unsigned int max_slots)
417 unsigned int current_slot=0;
419 struct o2hb_bio_wait_ctxt wc;
422 o2hb_bio_wait_init(&wc);
424 while(current_slot < max_slots) {
425 bio = o2hb_setup_one_bio(reg, &wc, ¤t_slot, max_slots);
427 status = PTR_ERR(bio);
432 atomic_inc(&wc.wc_num_reqs);
433 submit_bio(READ, bio);
439 o2hb_wait_on_io(reg, &wc);
440 if (wc.wc_error && !status)
441 status = wc.wc_error;
446 static int o2hb_issue_node_write(struct o2hb_region *reg,
447 struct o2hb_bio_wait_ctxt *write_wc)
453 o2hb_bio_wait_init(write_wc);
455 slot = o2nm_this_node();
457 bio = o2hb_setup_one_bio(reg, write_wc, &slot, slot+1);
459 status = PTR_ERR(bio);
464 atomic_inc(&write_wc->wc_num_reqs);
465 submit_bio(WRITE, bio);
472 static u32 o2hb_compute_block_crc_le(struct o2hb_region *reg,
473 struct o2hb_disk_heartbeat_block *hb_block)
478 /* We want to compute the block crc with a 0 value in the
479 * hb_cksum field. Save it off here and replace after the
481 old_cksum = hb_block->hb_cksum;
482 hb_block->hb_cksum = 0;
484 ret = crc32_le(0, (unsigned char *) hb_block, reg->hr_block_bytes);
486 hb_block->hb_cksum = old_cksum;
491 static void o2hb_dump_slot(struct o2hb_disk_heartbeat_block *hb_block)
493 mlog(ML_ERROR, "Dump slot information: seq = 0x%llx, node = %u, "
494 "cksum = 0x%x, generation 0x%llx\n",
495 (long long)le64_to_cpu(hb_block->hb_seq),
496 hb_block->hb_node, le32_to_cpu(hb_block->hb_cksum),
497 (long long)le64_to_cpu(hb_block->hb_generation));
500 static int o2hb_verify_crc(struct o2hb_region *reg,
501 struct o2hb_disk_heartbeat_block *hb_block)
505 read = le32_to_cpu(hb_block->hb_cksum);
506 computed = o2hb_compute_block_crc_le(reg, hb_block);
508 return read == computed;
511 /* We want to make sure that nobody is heartbeating on top of us --
512 * this will help detect an invalid configuration. */
513 static int o2hb_check_last_timestamp(struct o2hb_region *reg)
516 struct o2hb_disk_slot *slot;
517 struct o2hb_disk_heartbeat_block *hb_block;
519 node_num = o2nm_this_node();
522 slot = ®->hr_slots[node_num];
523 /* Don't check on our 1st timestamp */
524 if (slot->ds_last_time) {
525 hb_block = slot->ds_raw_block;
527 if (le64_to_cpu(hb_block->hb_seq) != slot->ds_last_time)
534 static inline void o2hb_prepare_block(struct o2hb_region *reg,
539 struct o2hb_disk_slot *slot;
540 struct o2hb_disk_heartbeat_block *hb_block;
542 node_num = o2nm_this_node();
543 slot = ®->hr_slots[node_num];
545 hb_block = (struct o2hb_disk_heartbeat_block *)slot->ds_raw_block;
546 memset(hb_block, 0, reg->hr_block_bytes);
547 /* TODO: time stuff */
548 cputime = CURRENT_TIME.tv_sec;
552 hb_block->hb_seq = cpu_to_le64(cputime);
553 hb_block->hb_node = node_num;
554 hb_block->hb_generation = cpu_to_le64(generation);
555 hb_block->hb_dead_ms = cpu_to_le32(o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS);
557 /* This step must always happen last! */
558 hb_block->hb_cksum = cpu_to_le32(o2hb_compute_block_crc_le(reg,
561 mlog(ML_HB_BIO, "our node generation = 0x%llx, cksum = 0x%x\n",
562 (long long)generation,
563 le32_to_cpu(hb_block->hb_cksum));
566 static void o2hb_fire_callbacks(struct o2hb_callback *hbcall,
567 struct o2nm_node *node,
570 struct list_head *iter;
571 struct o2hb_callback_func *f;
573 list_for_each(iter, &hbcall->list) {
574 f = list_entry(iter, struct o2hb_callback_func, hc_item);
575 mlog(ML_HEARTBEAT, "calling funcs %p\n", f);
576 (f->hc_func)(node, idx, f->hc_data);
580 /* Will run the list in order until we process the passed event */
581 static void o2hb_run_event_list(struct o2hb_node_event *queued_event)
584 struct o2hb_callback *hbcall;
585 struct o2hb_node_event *event;
587 spin_lock(&o2hb_live_lock);
588 empty = list_empty(&queued_event->hn_item);
589 spin_unlock(&o2hb_live_lock);
593 /* Holding callback sem assures we don't alter the callback
594 * lists when doing this, and serializes ourselves with other
595 * processes wanting callbacks. */
596 down_write(&o2hb_callback_sem);
598 spin_lock(&o2hb_live_lock);
599 while (!list_empty(&o2hb_node_events)
600 && !list_empty(&queued_event->hn_item)) {
601 event = list_entry(o2hb_node_events.next,
602 struct o2hb_node_event,
604 list_del_init(&event->hn_item);
605 spin_unlock(&o2hb_live_lock);
607 mlog(ML_HEARTBEAT, "Node %s event for %d\n",
608 event->hn_event_type == O2HB_NODE_UP_CB ? "UP" : "DOWN",
611 hbcall = hbcall_from_type(event->hn_event_type);
613 /* We should *never* have gotten on to the list with a
614 * bad type... This isn't something that we should try
615 * to recover from. */
616 BUG_ON(IS_ERR(hbcall));
618 o2hb_fire_callbacks(hbcall, event->hn_node, event->hn_node_num);
620 spin_lock(&o2hb_live_lock);
622 spin_unlock(&o2hb_live_lock);
624 up_write(&o2hb_callback_sem);
627 static void o2hb_queue_node_event(struct o2hb_node_event *event,
628 enum o2hb_callback_type type,
629 struct o2nm_node *node,
632 assert_spin_locked(&o2hb_live_lock);
634 BUG_ON((!node) && (type != O2HB_NODE_DOWN_CB));
636 event->hn_event_type = type;
637 event->hn_node = node;
638 event->hn_node_num = node_num;
640 mlog(ML_HEARTBEAT, "Queue node %s event for node %d\n",
641 type == O2HB_NODE_UP_CB ? "UP" : "DOWN", node_num);
643 list_add_tail(&event->hn_item, &o2hb_node_events);
646 static void o2hb_shutdown_slot(struct o2hb_disk_slot *slot)
648 struct o2hb_node_event event =
649 { .hn_item = LIST_HEAD_INIT(event.hn_item), };
650 struct o2nm_node *node;
652 node = o2nm_get_node_by_num(slot->ds_node_num);
656 spin_lock(&o2hb_live_lock);
657 if (!list_empty(&slot->ds_live_item)) {
658 mlog(ML_HEARTBEAT, "Shutdown, node %d leaves region\n",
661 list_del_init(&slot->ds_live_item);
663 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
664 clear_bit(slot->ds_node_num, o2hb_live_node_bitmap);
666 o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB, node,
670 spin_unlock(&o2hb_live_lock);
672 o2hb_run_event_list(&event);
677 static void o2hb_set_quorum_device(struct o2hb_region *reg,
678 struct o2hb_disk_slot *slot)
680 assert_spin_locked(&o2hb_live_lock);
682 if (!o2hb_global_heartbeat_active())
685 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap))
689 * A region can be added to the quorum only when it sees all
690 * live nodes heartbeat on it. In other words, the region has been
691 * added to all nodes.
693 if (memcmp(reg->hr_live_node_bitmap, o2hb_live_node_bitmap,
694 sizeof(o2hb_live_node_bitmap)))
697 if (slot->ds_changed_samples < O2HB_LIVE_THRESHOLD)
700 printk(KERN_NOTICE "o2hb: Region %s is now a quorum device\n",
701 config_item_name(®->hr_item));
703 set_bit(reg->hr_region_num, o2hb_quorum_region_bitmap);
706 static int o2hb_check_slot(struct o2hb_region *reg,
707 struct o2hb_disk_slot *slot)
709 int changed = 0, gen_changed = 0;
710 struct o2hb_node_event event =
711 { .hn_item = LIST_HEAD_INIT(event.hn_item), };
712 struct o2nm_node *node;
713 struct o2hb_disk_heartbeat_block *hb_block = reg->hr_tmp_block;
715 unsigned int dead_ms = o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS;
716 unsigned int slot_dead_ms;
719 memcpy(hb_block, slot->ds_raw_block, reg->hr_block_bytes);
722 * If a node is no longer configured but is still in the livemap, we
723 * may need to clear that bit from the livemap.
725 node = o2nm_get_node_by_num(slot->ds_node_num);
727 spin_lock(&o2hb_live_lock);
728 tmp = test_bit(slot->ds_node_num, o2hb_live_node_bitmap);
729 spin_unlock(&o2hb_live_lock);
734 if (!o2hb_verify_crc(reg, hb_block)) {
735 /* all paths from here will drop o2hb_live_lock for
737 spin_lock(&o2hb_live_lock);
739 /* Don't print an error on the console in this case -
740 * a freshly formatted heartbeat area will not have a
742 if (list_empty(&slot->ds_live_item))
745 /* The node is live but pushed out a bad crc. We
746 * consider it a transient miss but don't populate any
747 * other values as they may be junk. */
748 mlog(ML_ERROR, "Node %d has written a bad crc to %s\n",
749 slot->ds_node_num, reg->hr_dev_name);
750 o2hb_dump_slot(hb_block);
752 slot->ds_equal_samples++;
756 /* we don't care if these wrap.. the state transitions below
757 * clear at the right places */
758 cputime = le64_to_cpu(hb_block->hb_seq);
759 if (slot->ds_last_time != cputime)
760 slot->ds_changed_samples++;
762 slot->ds_equal_samples++;
763 slot->ds_last_time = cputime;
765 /* The node changed heartbeat generations. We assume this to
766 * mean it dropped off but came back before we timed out. We
767 * want to consider it down for the time being but don't want
768 * to lose any changed_samples state we might build up to
769 * considering it live again. */
770 if (slot->ds_last_generation != le64_to_cpu(hb_block->hb_generation)) {
772 slot->ds_equal_samples = 0;
773 mlog(ML_HEARTBEAT, "Node %d changed generation (0x%llx "
774 "to 0x%llx)\n", slot->ds_node_num,
775 (long long)slot->ds_last_generation,
776 (long long)le64_to_cpu(hb_block->hb_generation));
779 slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation);
781 mlog(ML_HEARTBEAT, "Slot %d gen 0x%llx cksum 0x%x "
782 "seq %llu last %llu changed %u equal %u\n",
783 slot->ds_node_num, (long long)slot->ds_last_generation,
784 le32_to_cpu(hb_block->hb_cksum),
785 (unsigned long long)le64_to_cpu(hb_block->hb_seq),
786 (unsigned long long)slot->ds_last_time, slot->ds_changed_samples,
787 slot->ds_equal_samples);
789 spin_lock(&o2hb_live_lock);
792 /* dead nodes only come to life after some number of
793 * changes at any time during their dead time */
794 if (list_empty(&slot->ds_live_item) &&
795 slot->ds_changed_samples >= O2HB_LIVE_THRESHOLD) {
796 mlog(ML_HEARTBEAT, "Node %d (id 0x%llx) joined my region\n",
797 slot->ds_node_num, (long long)slot->ds_last_generation);
799 set_bit(slot->ds_node_num, reg->hr_live_node_bitmap);
801 /* first on the list generates a callback */
802 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
803 mlog(ML_HEARTBEAT, "o2hb: Add node %d to live nodes "
804 "bitmap\n", slot->ds_node_num);
805 set_bit(slot->ds_node_num, o2hb_live_node_bitmap);
807 o2hb_queue_node_event(&event, O2HB_NODE_UP_CB, node,
813 list_add_tail(&slot->ds_live_item,
814 &o2hb_live_slots[slot->ds_node_num]);
816 slot->ds_equal_samples = 0;
818 /* We want to be sure that all nodes agree on the
819 * number of milliseconds before a node will be
820 * considered dead. The self-fencing timeout is
821 * computed from this value, and a discrepancy might
822 * result in heartbeat calling a node dead when it
823 * hasn't self-fenced yet. */
824 slot_dead_ms = le32_to_cpu(hb_block->hb_dead_ms);
825 if (slot_dead_ms && slot_dead_ms != dead_ms) {
826 /* TODO: Perhaps we can fail the region here. */
827 mlog(ML_ERROR, "Node %d on device %s has a dead count "
828 "of %u ms, but our count is %u ms.\n"
829 "Please double check your configuration values "
830 "for 'O2CB_HEARTBEAT_THRESHOLD'\n",
831 slot->ds_node_num, reg->hr_dev_name, slot_dead_ms,
837 /* if the list is dead, we're done.. */
838 if (list_empty(&slot->ds_live_item))
841 /* live nodes only go dead after enough consequtive missed
842 * samples.. reset the missed counter whenever we see
844 if (slot->ds_equal_samples >= o2hb_dead_threshold || gen_changed) {
845 mlog(ML_HEARTBEAT, "Node %d left my region\n",
848 clear_bit(slot->ds_node_num, reg->hr_live_node_bitmap);
850 /* last off the live_slot generates a callback */
851 list_del_init(&slot->ds_live_item);
852 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
853 mlog(ML_HEARTBEAT, "o2hb: Remove node %d from live "
854 "nodes bitmap\n", slot->ds_node_num);
855 clear_bit(slot->ds_node_num, o2hb_live_node_bitmap);
857 /* node can be null */
858 o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB,
859 node, slot->ds_node_num);
864 /* We don't clear this because the node is still
865 * actually writing new blocks. */
867 slot->ds_changed_samples = 0;
870 if (slot->ds_changed_samples) {
871 slot->ds_changed_samples = 0;
872 slot->ds_equal_samples = 0;
875 o2hb_set_quorum_device(reg, slot);
877 spin_unlock(&o2hb_live_lock);
879 o2hb_run_event_list(&event);
886 /* This could be faster if we just implmented a find_last_bit, but I
887 * don't think the circumstances warrant it. */
888 static int o2hb_highest_node(unsigned long *nodes,
895 while ((node = find_next_bit(nodes, numbits, node + 1)) != -1) {
905 static int o2hb_do_disk_heartbeat(struct o2hb_region *reg)
907 int i, ret, highest_node, change = 0;
908 unsigned long configured_nodes[BITS_TO_LONGS(O2NM_MAX_NODES)];
909 unsigned long live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
910 struct o2hb_bio_wait_ctxt write_wc;
912 ret = o2nm_configured_node_map(configured_nodes,
913 sizeof(configured_nodes));
920 * If a node is not configured but is in the livemap, we still need
921 * to read the slot so as to be able to remove it from the livemap.
923 o2hb_fill_node_map(live_node_bitmap, sizeof(live_node_bitmap));
925 while ((i = find_next_bit(live_node_bitmap,
926 O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) {
927 set_bit(i, configured_nodes);
930 highest_node = o2hb_highest_node(configured_nodes, O2NM_MAX_NODES);
931 if (highest_node >= O2NM_MAX_NODES) {
932 mlog(ML_NOTICE, "ocfs2_heartbeat: no configured nodes found!\n");
936 /* No sense in reading the slots of nodes that don't exist
937 * yet. Of course, if the node definitions have holes in them
938 * then we're reading an empty slot anyway... Consider this
940 ret = o2hb_read_slots(reg, highest_node + 1);
946 /* With an up to date view of the slots, we can check that no
947 * other node has been improperly configured to heartbeat in
949 if (!o2hb_check_last_timestamp(reg))
950 mlog(ML_ERROR, "Device \"%s\": another node is heartbeating "
951 "in our slot!\n", reg->hr_dev_name);
953 /* fill in the proper info for our next heartbeat */
954 o2hb_prepare_block(reg, reg->hr_generation);
956 /* And fire off the write. Note that we don't wait on this I/O
958 ret = o2hb_issue_node_write(reg, &write_wc);
965 while((i = find_next_bit(configured_nodes, O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) {
967 change |= o2hb_check_slot(reg, ®->hr_slots[i]);
971 * We have to be sure we've advertised ourselves on disk
972 * before we can go to steady state. This ensures that
973 * people we find in our steady state have seen us.
975 o2hb_wait_on_io(reg, &write_wc);
976 if (write_wc.wc_error) {
977 /* Do not re-arm the write timeout on I/O error - we
978 * can't be sure that the new block ever made it to
980 mlog(ML_ERROR, "Write error %d on device \"%s\"\n",
981 write_wc.wc_error, reg->hr_dev_name);
982 return write_wc.wc_error;
985 o2hb_arm_write_timeout(reg);
987 /* let the person who launched us know when things are steady */
988 if (!change && (atomic_read(®->hr_steady_iterations) != 0)) {
989 if (atomic_dec_and_test(®->hr_steady_iterations))
990 wake_up(&o2hb_steady_queue);
996 /* Subtract b from a, storing the result in a. a *must* have a larger
998 static void o2hb_tv_subtract(struct timeval *a,
1001 /* just return 0 when a is after b */
1002 if (a->tv_sec < b->tv_sec ||
1003 (a->tv_sec == b->tv_sec && a->tv_usec < b->tv_usec)) {
1009 a->tv_sec -= b->tv_sec;
1010 a->tv_usec -= b->tv_usec;
1011 while ( a->tv_usec < 0 ) {
1013 a->tv_usec += 1000000;
1017 static unsigned int o2hb_elapsed_msecs(struct timeval *start,
1018 struct timeval *end)
1020 struct timeval res = *end;
1022 o2hb_tv_subtract(&res, start);
1024 return res.tv_sec * 1000 + res.tv_usec / 1000;
1028 * we ride the region ref that the region dir holds. before the region
1029 * dir is removed and drops it ref it will wait to tear down this
1032 static int o2hb_thread(void *data)
1035 struct o2hb_region *reg = data;
1036 struct o2hb_bio_wait_ctxt write_wc;
1037 struct timeval before_hb, after_hb;
1038 unsigned int elapsed_msec;
1040 mlog(ML_HEARTBEAT|ML_KTHREAD, "hb thread running\n");
1042 set_user_nice(current, -20);
1044 while (!kthread_should_stop() && !reg->hr_unclean_stop) {
1045 /* We track the time spent inside
1046 * o2hb_do_disk_heartbeat so that we avoid more than
1047 * hr_timeout_ms between disk writes. On busy systems
1048 * this should result in a heartbeat which is less
1049 * likely to time itself out. */
1050 do_gettimeofday(&before_hb);
1054 ret = o2hb_do_disk_heartbeat(reg);
1055 } while (ret && ++i < 2);
1057 do_gettimeofday(&after_hb);
1058 elapsed_msec = o2hb_elapsed_msecs(&before_hb, &after_hb);
1061 "start = %lu.%lu, end = %lu.%lu, msec = %u\n",
1062 before_hb.tv_sec, (unsigned long) before_hb.tv_usec,
1063 after_hb.tv_sec, (unsigned long) after_hb.tv_usec,
1066 if (elapsed_msec < reg->hr_timeout_ms) {
1067 /* the kthread api has blocked signals for us so no
1068 * need to record the return value. */
1069 msleep_interruptible(reg->hr_timeout_ms - elapsed_msec);
1073 o2hb_disarm_write_timeout(reg);
1075 /* unclean stop is only used in very bad situation */
1076 for(i = 0; !reg->hr_unclean_stop && i < reg->hr_blocks; i++)
1077 o2hb_shutdown_slot(®->hr_slots[i]);
1079 /* Explicit down notification - avoid forcing the other nodes
1080 * to timeout on this region when we could just as easily
1081 * write a clear generation - thus indicating to them that
1082 * this node has left this region.
1084 * XXX: Should we skip this on unclean_stop? */
1085 o2hb_prepare_block(reg, 0);
1086 ret = o2hb_issue_node_write(reg, &write_wc);
1088 o2hb_wait_on_io(reg, &write_wc);
1093 mlog(ML_HEARTBEAT|ML_KTHREAD, "hb thread exiting\n");
1098 #ifdef CONFIG_DEBUG_FS
1099 static int o2hb_debug_open(struct inode *inode, struct file *file)
1101 struct o2hb_debug_buf *db = inode->i_private;
1102 struct o2hb_region *reg;
1103 unsigned long map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1108 /* max_nodes should be the largest bitmap we pass here */
1109 BUG_ON(sizeof(map) < db->db_size);
1111 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
1115 switch (db->db_type) {
1116 case O2HB_DB_TYPE_LIVENODES:
1117 case O2HB_DB_TYPE_LIVEREGIONS:
1118 case O2HB_DB_TYPE_QUORUMREGIONS:
1119 case O2HB_DB_TYPE_FAILEDREGIONS:
1120 spin_lock(&o2hb_live_lock);
1121 memcpy(map, db->db_data, db->db_size);
1122 spin_unlock(&o2hb_live_lock);
1125 case O2HB_DB_TYPE_REGION_LIVENODES:
1126 spin_lock(&o2hb_live_lock);
1127 reg = (struct o2hb_region *)db->db_data;
1128 memcpy(map, reg->hr_live_node_bitmap, db->db_size);
1129 spin_unlock(&o2hb_live_lock);
1132 case O2HB_DB_TYPE_REGION_NUMBER:
1133 reg = (struct o2hb_region *)db->db_data;
1134 out += snprintf(buf + out, PAGE_SIZE - out, "%d\n",
1135 reg->hr_region_num);
1138 case O2HB_DB_TYPE_REGION_ELAPSED_TIME:
1139 reg = (struct o2hb_region *)db->db_data;
1140 out += snprintf(buf + out, PAGE_SIZE - out, "%u\n",
1141 jiffies_to_msecs(jiffies -
1142 reg->hr_last_timeout_start));
1149 while ((i = find_next_bit(map, db->db_len, i + 1)) < db->db_len)
1150 out += snprintf(buf + out, PAGE_SIZE - out, "%d ", i);
1151 out += snprintf(buf + out, PAGE_SIZE - out, "\n");
1154 i_size_write(inode, out);
1156 file->private_data = buf;
1163 static int o2hb_debug_release(struct inode *inode, struct file *file)
1165 kfree(file->private_data);
1169 static ssize_t o2hb_debug_read(struct file *file, char __user *buf,
1170 size_t nbytes, loff_t *ppos)
1172 return simple_read_from_buffer(buf, nbytes, ppos, file->private_data,
1173 i_size_read(file->f_mapping->host));
1176 static int o2hb_debug_open(struct inode *inode, struct file *file)
1180 static int o2hb_debug_release(struct inode *inode, struct file *file)
1184 static ssize_t o2hb_debug_read(struct file *file, char __user *buf,
1185 size_t nbytes, loff_t *ppos)
1189 #endif /* CONFIG_DEBUG_FS */
1191 static const struct file_operations o2hb_debug_fops = {
1192 .open = o2hb_debug_open,
1193 .release = o2hb_debug_release,
1194 .read = o2hb_debug_read,
1195 .llseek = generic_file_llseek,
1198 void o2hb_exit(void)
1200 kfree(o2hb_db_livenodes);
1201 kfree(o2hb_db_liveregions);
1202 kfree(o2hb_db_quorumregions);
1203 kfree(o2hb_db_failedregions);
1204 debugfs_remove(o2hb_debug_failedregions);
1205 debugfs_remove(o2hb_debug_quorumregions);
1206 debugfs_remove(o2hb_debug_liveregions);
1207 debugfs_remove(o2hb_debug_livenodes);
1208 debugfs_remove(o2hb_debug_dir);
1211 static struct dentry *o2hb_debug_create(const char *name, struct dentry *dir,
1212 struct o2hb_debug_buf **db, int db_len,
1213 int type, int size, int len, void *data)
1215 *db = kmalloc(db_len, GFP_KERNEL);
1219 (*db)->db_type = type;
1220 (*db)->db_size = size;
1221 (*db)->db_len = len;
1222 (*db)->db_data = data;
1224 return debugfs_create_file(name, S_IFREG|S_IRUSR, dir, *db,
1228 static int o2hb_debug_init(void)
1232 o2hb_debug_dir = debugfs_create_dir(O2HB_DEBUG_DIR, NULL);
1233 if (!o2hb_debug_dir) {
1238 o2hb_debug_livenodes = o2hb_debug_create(O2HB_DEBUG_LIVENODES,
1241 sizeof(*o2hb_db_livenodes),
1242 O2HB_DB_TYPE_LIVENODES,
1243 sizeof(o2hb_live_node_bitmap),
1245 o2hb_live_node_bitmap);
1246 if (!o2hb_debug_livenodes) {
1251 o2hb_debug_liveregions = o2hb_debug_create(O2HB_DEBUG_LIVEREGIONS,
1253 &o2hb_db_liveregions,
1254 sizeof(*o2hb_db_liveregions),
1255 O2HB_DB_TYPE_LIVEREGIONS,
1256 sizeof(o2hb_live_region_bitmap),
1258 o2hb_live_region_bitmap);
1259 if (!o2hb_debug_liveregions) {
1264 o2hb_debug_quorumregions =
1265 o2hb_debug_create(O2HB_DEBUG_QUORUMREGIONS,
1267 &o2hb_db_quorumregions,
1268 sizeof(*o2hb_db_quorumregions),
1269 O2HB_DB_TYPE_QUORUMREGIONS,
1270 sizeof(o2hb_quorum_region_bitmap),
1272 o2hb_quorum_region_bitmap);
1273 if (!o2hb_debug_quorumregions) {
1278 o2hb_debug_failedregions =
1279 o2hb_debug_create(O2HB_DEBUG_FAILEDREGIONS,
1281 &o2hb_db_failedregions,
1282 sizeof(*o2hb_db_failedregions),
1283 O2HB_DB_TYPE_FAILEDREGIONS,
1284 sizeof(o2hb_failed_region_bitmap),
1286 o2hb_failed_region_bitmap);
1287 if (!o2hb_debug_failedregions) {
1304 for (i = 0; i < ARRAY_SIZE(o2hb_callbacks); i++)
1305 INIT_LIST_HEAD(&o2hb_callbacks[i].list);
1307 for (i = 0; i < ARRAY_SIZE(o2hb_live_slots); i++)
1308 INIT_LIST_HEAD(&o2hb_live_slots[i]);
1310 INIT_LIST_HEAD(&o2hb_node_events);
1312 memset(o2hb_live_node_bitmap, 0, sizeof(o2hb_live_node_bitmap));
1313 memset(o2hb_region_bitmap, 0, sizeof(o2hb_region_bitmap));
1314 memset(o2hb_live_region_bitmap, 0, sizeof(o2hb_live_region_bitmap));
1315 memset(o2hb_quorum_region_bitmap, 0, sizeof(o2hb_quorum_region_bitmap));
1316 memset(o2hb_failed_region_bitmap, 0, sizeof(o2hb_failed_region_bitmap));
1318 return o2hb_debug_init();
1321 /* if we're already in a callback then we're already serialized by the sem */
1322 static void o2hb_fill_node_map_from_callback(unsigned long *map,
1325 BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1327 memcpy(map, &o2hb_live_node_bitmap, bytes);
1331 * get a map of all nodes that are heartbeating in any regions
1333 void o2hb_fill_node_map(unsigned long *map, unsigned bytes)
1335 /* callers want to serialize this map and callbacks so that they
1336 * can trust that they don't miss nodes coming to the party */
1337 down_read(&o2hb_callback_sem);
1338 spin_lock(&o2hb_live_lock);
1339 o2hb_fill_node_map_from_callback(map, bytes);
1340 spin_unlock(&o2hb_live_lock);
1341 up_read(&o2hb_callback_sem);
1343 EXPORT_SYMBOL_GPL(o2hb_fill_node_map);
1346 * heartbeat configfs bits. The heartbeat set is a default set under
1347 * the cluster set in nodemanager.c.
1350 static struct o2hb_region *to_o2hb_region(struct config_item *item)
1352 return item ? container_of(item, struct o2hb_region, hr_item) : NULL;
1355 /* drop_item only drops its ref after killing the thread, nothing should
1356 * be using the region anymore. this has to clean up any state that
1357 * attributes might have built up. */
1358 static void o2hb_region_release(struct config_item *item)
1362 struct o2hb_region *reg = to_o2hb_region(item);
1364 if (reg->hr_tmp_block)
1365 kfree(reg->hr_tmp_block);
1367 if (reg->hr_slot_data) {
1368 for (i = 0; i < reg->hr_num_pages; i++) {
1369 page = reg->hr_slot_data[i];
1373 kfree(reg->hr_slot_data);
1377 blkdev_put(reg->hr_bdev, FMODE_READ|FMODE_WRITE);
1380 kfree(reg->hr_slots);
1382 kfree(reg->hr_db_regnum);
1383 kfree(reg->hr_db_livenodes);
1384 debugfs_remove(reg->hr_debug_livenodes);
1385 debugfs_remove(reg->hr_debug_regnum);
1386 debugfs_remove(reg->hr_debug_elapsed_time);
1387 debugfs_remove(reg->hr_debug_dir);
1389 spin_lock(&o2hb_live_lock);
1390 list_del(®->hr_all_item);
1391 spin_unlock(&o2hb_live_lock);
1396 static int o2hb_read_block_input(struct o2hb_region *reg,
1399 unsigned long *ret_bytes,
1400 unsigned int *ret_bits)
1402 unsigned long bytes;
1403 char *p = (char *)page;
1405 bytes = simple_strtoul(p, &p, 0);
1406 if (!p || (*p && (*p != '\n')))
1409 /* Heartbeat and fs min / max block sizes are the same. */
1410 if (bytes > 4096 || bytes < 512)
1412 if (hweight16(bytes) != 1)
1418 *ret_bits = ffs(bytes) - 1;
1423 static ssize_t o2hb_region_block_bytes_read(struct o2hb_region *reg,
1426 return sprintf(page, "%u\n", reg->hr_block_bytes);
1429 static ssize_t o2hb_region_block_bytes_write(struct o2hb_region *reg,
1434 unsigned long block_bytes;
1435 unsigned int block_bits;
1440 status = o2hb_read_block_input(reg, page, count,
1441 &block_bytes, &block_bits);
1445 reg->hr_block_bytes = (unsigned int)block_bytes;
1446 reg->hr_block_bits = block_bits;
1451 static ssize_t o2hb_region_start_block_read(struct o2hb_region *reg,
1454 return sprintf(page, "%llu\n", reg->hr_start_block);
1457 static ssize_t o2hb_region_start_block_write(struct o2hb_region *reg,
1461 unsigned long long tmp;
1462 char *p = (char *)page;
1467 tmp = simple_strtoull(p, &p, 0);
1468 if (!p || (*p && (*p != '\n')))
1471 reg->hr_start_block = tmp;
1476 static ssize_t o2hb_region_blocks_read(struct o2hb_region *reg,
1479 return sprintf(page, "%d\n", reg->hr_blocks);
1482 static ssize_t o2hb_region_blocks_write(struct o2hb_region *reg,
1487 char *p = (char *)page;
1492 tmp = simple_strtoul(p, &p, 0);
1493 if (!p || (*p && (*p != '\n')))
1496 if (tmp > O2NM_MAX_NODES || tmp == 0)
1499 reg->hr_blocks = (unsigned int)tmp;
1504 static ssize_t o2hb_region_dev_read(struct o2hb_region *reg,
1507 unsigned int ret = 0;
1510 ret = sprintf(page, "%s\n", reg->hr_dev_name);
1515 static void o2hb_init_region_params(struct o2hb_region *reg)
1517 reg->hr_slots_per_page = PAGE_CACHE_SIZE >> reg->hr_block_bits;
1518 reg->hr_timeout_ms = O2HB_REGION_TIMEOUT_MS;
1520 mlog(ML_HEARTBEAT, "hr_start_block = %llu, hr_blocks = %u\n",
1521 reg->hr_start_block, reg->hr_blocks);
1522 mlog(ML_HEARTBEAT, "hr_block_bytes = %u, hr_block_bits = %u\n",
1523 reg->hr_block_bytes, reg->hr_block_bits);
1524 mlog(ML_HEARTBEAT, "hr_timeout_ms = %u\n", reg->hr_timeout_ms);
1525 mlog(ML_HEARTBEAT, "dead threshold = %u\n", o2hb_dead_threshold);
1528 static int o2hb_map_slot_data(struct o2hb_region *reg)
1531 unsigned int last_slot;
1532 unsigned int spp = reg->hr_slots_per_page;
1535 struct o2hb_disk_slot *slot;
1537 reg->hr_tmp_block = kmalloc(reg->hr_block_bytes, GFP_KERNEL);
1538 if (reg->hr_tmp_block == NULL) {
1539 mlog_errno(-ENOMEM);
1543 reg->hr_slots = kcalloc(reg->hr_blocks,
1544 sizeof(struct o2hb_disk_slot), GFP_KERNEL);
1545 if (reg->hr_slots == NULL) {
1546 mlog_errno(-ENOMEM);
1550 for(i = 0; i < reg->hr_blocks; i++) {
1551 slot = ®->hr_slots[i];
1552 slot->ds_node_num = i;
1553 INIT_LIST_HEAD(&slot->ds_live_item);
1554 slot->ds_raw_block = NULL;
1557 reg->hr_num_pages = (reg->hr_blocks + spp - 1) / spp;
1558 mlog(ML_HEARTBEAT, "Going to require %u pages to cover %u blocks "
1559 "at %u blocks per page\n",
1560 reg->hr_num_pages, reg->hr_blocks, spp);
1562 reg->hr_slot_data = kcalloc(reg->hr_num_pages, sizeof(struct page *),
1564 if (!reg->hr_slot_data) {
1565 mlog_errno(-ENOMEM);
1569 for(i = 0; i < reg->hr_num_pages; i++) {
1570 page = alloc_page(GFP_KERNEL);
1572 mlog_errno(-ENOMEM);
1576 reg->hr_slot_data[i] = page;
1578 last_slot = i * spp;
1579 raw = page_address(page);
1581 (j < spp) && ((j + last_slot) < reg->hr_blocks);
1583 BUG_ON((j + last_slot) >= reg->hr_blocks);
1585 slot = ®->hr_slots[j + last_slot];
1586 slot->ds_raw_block =
1587 (struct o2hb_disk_heartbeat_block *) raw;
1589 raw += reg->hr_block_bytes;
1596 /* Read in all the slots available and populate the tracking
1597 * structures so that we can start with a baseline idea of what's
1599 static int o2hb_populate_slot_data(struct o2hb_region *reg)
1602 struct o2hb_disk_slot *slot;
1603 struct o2hb_disk_heartbeat_block *hb_block;
1607 ret = o2hb_read_slots(reg, reg->hr_blocks);
1613 /* We only want to get an idea of the values initially in each
1614 * slot, so we do no verification - o2hb_check_slot will
1615 * actually determine if each configured slot is valid and
1616 * whether any values have changed. */
1617 for(i = 0; i < reg->hr_blocks; i++) {
1618 slot = ®->hr_slots[i];
1619 hb_block = (struct o2hb_disk_heartbeat_block *) slot->ds_raw_block;
1621 /* Only fill the values that o2hb_check_slot uses to
1622 * determine changing slots */
1623 slot->ds_last_time = le64_to_cpu(hb_block->hb_seq);
1624 slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation);
1632 /* this is acting as commit; we set up all of hr_bdev and hr_task or nothing */
1633 static ssize_t o2hb_region_dev_write(struct o2hb_region *reg,
1637 struct task_struct *hb_task;
1640 char *p = (char *)page;
1641 struct file *filp = NULL;
1642 struct inode *inode = NULL;
1643 ssize_t ret = -EINVAL;
1648 /* We can't heartbeat without having had our node number
1649 * configured yet. */
1650 if (o2nm_this_node() == O2NM_MAX_NODES)
1653 fd = simple_strtol(p, &p, 0);
1654 if (!p || (*p && (*p != '\n')))
1657 if (fd < 0 || fd >= INT_MAX)
1664 if (reg->hr_blocks == 0 || reg->hr_start_block == 0 ||
1665 reg->hr_block_bytes == 0)
1668 inode = igrab(filp->f_mapping->host);
1672 if (!S_ISBLK(inode->i_mode))
1675 reg->hr_bdev = I_BDEV(filp->f_mapping->host);
1676 ret = blkdev_get(reg->hr_bdev, FMODE_WRITE | FMODE_READ);
1678 reg->hr_bdev = NULL;
1683 bdevname(reg->hr_bdev, reg->hr_dev_name);
1685 sectsize = bdev_logical_block_size(reg->hr_bdev);
1686 if (sectsize != reg->hr_block_bytes) {
1688 "blocksize %u incorrect for device, expected %d",
1689 reg->hr_block_bytes, sectsize);
1694 o2hb_init_region_params(reg);
1696 /* Generation of zero is invalid */
1698 get_random_bytes(®->hr_generation,
1699 sizeof(reg->hr_generation));
1700 } while (reg->hr_generation == 0);
1702 ret = o2hb_map_slot_data(reg);
1708 ret = o2hb_populate_slot_data(reg);
1714 INIT_DELAYED_WORK(®->hr_write_timeout_work, o2hb_write_timeout);
1717 * A node is considered live after it has beat LIVE_THRESHOLD
1718 * times. We're not steady until we've given them a chance
1719 * _after_ our first read.
1721 atomic_set(®->hr_steady_iterations, O2HB_LIVE_THRESHOLD + 1);
1723 hb_task = kthread_run(o2hb_thread, reg, "o2hb-%s",
1724 reg->hr_item.ci_name);
1725 if (IS_ERR(hb_task)) {
1726 ret = PTR_ERR(hb_task);
1731 spin_lock(&o2hb_live_lock);
1732 reg->hr_task = hb_task;
1733 spin_unlock(&o2hb_live_lock);
1735 ret = wait_event_interruptible(o2hb_steady_queue,
1736 atomic_read(®->hr_steady_iterations) == 0);
1738 /* We got interrupted (hello ptrace!). Clean up */
1739 spin_lock(&o2hb_live_lock);
1740 hb_task = reg->hr_task;
1741 reg->hr_task = NULL;
1742 spin_unlock(&o2hb_live_lock);
1745 kthread_stop(hb_task);
1749 /* Ok, we were woken. Make sure it wasn't by drop_item() */
1750 spin_lock(&o2hb_live_lock);
1751 hb_task = reg->hr_task;
1752 if (o2hb_global_heartbeat_active())
1753 set_bit(reg->hr_region_num, o2hb_live_region_bitmap);
1754 spin_unlock(&o2hb_live_lock);
1761 if (hb_task && o2hb_global_heartbeat_active())
1762 printk(KERN_NOTICE "o2hb: Heartbeat started on region %s\n",
1763 config_item_name(®->hr_item));
1772 blkdev_put(reg->hr_bdev, FMODE_READ|FMODE_WRITE);
1773 reg->hr_bdev = NULL;
1779 static ssize_t o2hb_region_pid_read(struct o2hb_region *reg,
1784 spin_lock(&o2hb_live_lock);
1786 pid = task_pid_nr(reg->hr_task);
1787 spin_unlock(&o2hb_live_lock);
1792 return sprintf(page, "%u\n", pid);
1795 struct o2hb_region_attribute {
1796 struct configfs_attribute attr;
1797 ssize_t (*show)(struct o2hb_region *, char *);
1798 ssize_t (*store)(struct o2hb_region *, const char *, size_t);
1801 static struct o2hb_region_attribute o2hb_region_attr_block_bytes = {
1802 .attr = { .ca_owner = THIS_MODULE,
1803 .ca_name = "block_bytes",
1804 .ca_mode = S_IRUGO | S_IWUSR },
1805 .show = o2hb_region_block_bytes_read,
1806 .store = o2hb_region_block_bytes_write,
1809 static struct o2hb_region_attribute o2hb_region_attr_start_block = {
1810 .attr = { .ca_owner = THIS_MODULE,
1811 .ca_name = "start_block",
1812 .ca_mode = S_IRUGO | S_IWUSR },
1813 .show = o2hb_region_start_block_read,
1814 .store = o2hb_region_start_block_write,
1817 static struct o2hb_region_attribute o2hb_region_attr_blocks = {
1818 .attr = { .ca_owner = THIS_MODULE,
1819 .ca_name = "blocks",
1820 .ca_mode = S_IRUGO | S_IWUSR },
1821 .show = o2hb_region_blocks_read,
1822 .store = o2hb_region_blocks_write,
1825 static struct o2hb_region_attribute o2hb_region_attr_dev = {
1826 .attr = { .ca_owner = THIS_MODULE,
1828 .ca_mode = S_IRUGO | S_IWUSR },
1829 .show = o2hb_region_dev_read,
1830 .store = o2hb_region_dev_write,
1833 static struct o2hb_region_attribute o2hb_region_attr_pid = {
1834 .attr = { .ca_owner = THIS_MODULE,
1836 .ca_mode = S_IRUGO | S_IRUSR },
1837 .show = o2hb_region_pid_read,
1840 static struct configfs_attribute *o2hb_region_attrs[] = {
1841 &o2hb_region_attr_block_bytes.attr,
1842 &o2hb_region_attr_start_block.attr,
1843 &o2hb_region_attr_blocks.attr,
1844 &o2hb_region_attr_dev.attr,
1845 &o2hb_region_attr_pid.attr,
1849 static ssize_t o2hb_region_show(struct config_item *item,
1850 struct configfs_attribute *attr,
1853 struct o2hb_region *reg = to_o2hb_region(item);
1854 struct o2hb_region_attribute *o2hb_region_attr =
1855 container_of(attr, struct o2hb_region_attribute, attr);
1858 if (o2hb_region_attr->show)
1859 ret = o2hb_region_attr->show(reg, page);
1863 static ssize_t o2hb_region_store(struct config_item *item,
1864 struct configfs_attribute *attr,
1865 const char *page, size_t count)
1867 struct o2hb_region *reg = to_o2hb_region(item);
1868 struct o2hb_region_attribute *o2hb_region_attr =
1869 container_of(attr, struct o2hb_region_attribute, attr);
1870 ssize_t ret = -EINVAL;
1872 if (o2hb_region_attr->store)
1873 ret = o2hb_region_attr->store(reg, page, count);
1877 static struct configfs_item_operations o2hb_region_item_ops = {
1878 .release = o2hb_region_release,
1879 .show_attribute = o2hb_region_show,
1880 .store_attribute = o2hb_region_store,
1883 static struct config_item_type o2hb_region_type = {
1884 .ct_item_ops = &o2hb_region_item_ops,
1885 .ct_attrs = o2hb_region_attrs,
1886 .ct_owner = THIS_MODULE,
1891 struct o2hb_heartbeat_group {
1892 struct config_group hs_group;
1896 static struct o2hb_heartbeat_group *to_o2hb_heartbeat_group(struct config_group *group)
1899 container_of(group, struct o2hb_heartbeat_group, hs_group)
1903 static int o2hb_debug_region_init(struct o2hb_region *reg, struct dentry *dir)
1908 debugfs_create_dir(config_item_name(®->hr_item), dir);
1909 if (!reg->hr_debug_dir) {
1914 reg->hr_debug_livenodes =
1915 o2hb_debug_create(O2HB_DEBUG_LIVENODES,
1917 &(reg->hr_db_livenodes),
1918 sizeof(*(reg->hr_db_livenodes)),
1919 O2HB_DB_TYPE_REGION_LIVENODES,
1920 sizeof(reg->hr_live_node_bitmap),
1921 O2NM_MAX_NODES, reg);
1922 if (!reg->hr_debug_livenodes) {
1927 reg->hr_debug_regnum =
1928 o2hb_debug_create(O2HB_DEBUG_REGION_NUMBER,
1930 &(reg->hr_db_regnum),
1931 sizeof(*(reg->hr_db_regnum)),
1932 O2HB_DB_TYPE_REGION_NUMBER,
1933 0, O2NM_MAX_NODES, reg);
1934 if (!reg->hr_debug_regnum) {
1939 reg->hr_debug_elapsed_time =
1940 o2hb_debug_create(O2HB_DEBUG_REGION_ELAPSED_TIME,
1942 &(reg->hr_db_elapsed_time),
1943 sizeof(*(reg->hr_db_elapsed_time)),
1944 O2HB_DB_TYPE_REGION_ELAPSED_TIME,
1946 if (!reg->hr_debug_elapsed_time) {
1956 static struct config_item *o2hb_heartbeat_group_make_item(struct config_group *group,
1959 struct o2hb_region *reg = NULL;
1962 reg = kzalloc(sizeof(struct o2hb_region), GFP_KERNEL);
1964 return ERR_PTR(-ENOMEM);
1966 if (strlen(name) > O2HB_MAX_REGION_NAME_LEN) {
1967 ret = -ENAMETOOLONG;
1971 spin_lock(&o2hb_live_lock);
1972 reg->hr_region_num = 0;
1973 if (o2hb_global_heartbeat_active()) {
1974 reg->hr_region_num = find_first_zero_bit(o2hb_region_bitmap,
1976 if (reg->hr_region_num >= O2NM_MAX_REGIONS) {
1977 spin_unlock(&o2hb_live_lock);
1981 set_bit(reg->hr_region_num, o2hb_region_bitmap);
1983 list_add_tail(®->hr_all_item, &o2hb_all_regions);
1984 spin_unlock(&o2hb_live_lock);
1986 config_item_init_type_name(®->hr_item, name, &o2hb_region_type);
1988 ret = o2hb_debug_region_init(reg, o2hb_debug_dir);
1990 config_item_put(®->hr_item);
1994 return ®->hr_item;
1997 return ERR_PTR(ret);
2000 static void o2hb_heartbeat_group_drop_item(struct config_group *group,
2001 struct config_item *item)
2003 struct task_struct *hb_task;
2004 struct o2hb_region *reg = to_o2hb_region(item);
2006 /* stop the thread when the user removes the region dir */
2007 spin_lock(&o2hb_live_lock);
2008 if (o2hb_global_heartbeat_active()) {
2009 clear_bit(reg->hr_region_num, o2hb_region_bitmap);
2010 clear_bit(reg->hr_region_num, o2hb_live_region_bitmap);
2012 hb_task = reg->hr_task;
2013 reg->hr_task = NULL;
2014 spin_unlock(&o2hb_live_lock);
2017 kthread_stop(hb_task);
2020 * If we're racing a dev_write(), we need to wake them. They will
2021 * check reg->hr_task
2023 if (atomic_read(®->hr_steady_iterations) != 0) {
2024 atomic_set(®->hr_steady_iterations, 0);
2025 wake_up(&o2hb_steady_queue);
2028 if (o2hb_global_heartbeat_active())
2029 printk(KERN_NOTICE "o2hb: Heartbeat stopped on region %s\n",
2030 config_item_name(®->hr_item));
2031 config_item_put(item);
2034 struct o2hb_heartbeat_group_attribute {
2035 struct configfs_attribute attr;
2036 ssize_t (*show)(struct o2hb_heartbeat_group *, char *);
2037 ssize_t (*store)(struct o2hb_heartbeat_group *, const char *, size_t);
2040 static ssize_t o2hb_heartbeat_group_show(struct config_item *item,
2041 struct configfs_attribute *attr,
2044 struct o2hb_heartbeat_group *reg = to_o2hb_heartbeat_group(to_config_group(item));
2045 struct o2hb_heartbeat_group_attribute *o2hb_heartbeat_group_attr =
2046 container_of(attr, struct o2hb_heartbeat_group_attribute, attr);
2049 if (o2hb_heartbeat_group_attr->show)
2050 ret = o2hb_heartbeat_group_attr->show(reg, page);
2054 static ssize_t o2hb_heartbeat_group_store(struct config_item *item,
2055 struct configfs_attribute *attr,
2056 const char *page, size_t count)
2058 struct o2hb_heartbeat_group *reg = to_o2hb_heartbeat_group(to_config_group(item));
2059 struct o2hb_heartbeat_group_attribute *o2hb_heartbeat_group_attr =
2060 container_of(attr, struct o2hb_heartbeat_group_attribute, attr);
2061 ssize_t ret = -EINVAL;
2063 if (o2hb_heartbeat_group_attr->store)
2064 ret = o2hb_heartbeat_group_attr->store(reg, page, count);
2068 static ssize_t o2hb_heartbeat_group_threshold_show(struct o2hb_heartbeat_group *group,
2071 return sprintf(page, "%u\n", o2hb_dead_threshold);
2074 static ssize_t o2hb_heartbeat_group_threshold_store(struct o2hb_heartbeat_group *group,
2079 char *p = (char *)page;
2081 tmp = simple_strtoul(p, &p, 10);
2082 if (!p || (*p && (*p != '\n')))
2085 /* this will validate ranges for us. */
2086 o2hb_dead_threshold_set((unsigned int) tmp);
2092 ssize_t o2hb_heartbeat_group_mode_show(struct o2hb_heartbeat_group *group,
2095 return sprintf(page, "%s\n",
2096 o2hb_heartbeat_mode_desc[o2hb_heartbeat_mode]);
2100 ssize_t o2hb_heartbeat_group_mode_store(struct o2hb_heartbeat_group *group,
2101 const char *page, size_t count)
2107 len = (page[count - 1] == '\n') ? count - 1 : count;
2111 for (i = 0; i < O2HB_HEARTBEAT_NUM_MODES; ++i) {
2112 if (strnicmp(page, o2hb_heartbeat_mode_desc[i], len))
2115 ret = o2hb_global_hearbeat_mode_set(i);
2117 printk(KERN_NOTICE "o2hb: Heartbeat mode set to %s\n",
2118 o2hb_heartbeat_mode_desc[i]);
2126 static struct o2hb_heartbeat_group_attribute o2hb_heartbeat_group_attr_threshold = {
2127 .attr = { .ca_owner = THIS_MODULE,
2128 .ca_name = "dead_threshold",
2129 .ca_mode = S_IRUGO | S_IWUSR },
2130 .show = o2hb_heartbeat_group_threshold_show,
2131 .store = o2hb_heartbeat_group_threshold_store,
2134 static struct o2hb_heartbeat_group_attribute o2hb_heartbeat_group_attr_mode = {
2135 .attr = { .ca_owner = THIS_MODULE,
2137 .ca_mode = S_IRUGO | S_IWUSR },
2138 .show = o2hb_heartbeat_group_mode_show,
2139 .store = o2hb_heartbeat_group_mode_store,
2142 static struct configfs_attribute *o2hb_heartbeat_group_attrs[] = {
2143 &o2hb_heartbeat_group_attr_threshold.attr,
2144 &o2hb_heartbeat_group_attr_mode.attr,
2148 static struct configfs_item_operations o2hb_hearbeat_group_item_ops = {
2149 .show_attribute = o2hb_heartbeat_group_show,
2150 .store_attribute = o2hb_heartbeat_group_store,
2153 static struct configfs_group_operations o2hb_heartbeat_group_group_ops = {
2154 .make_item = o2hb_heartbeat_group_make_item,
2155 .drop_item = o2hb_heartbeat_group_drop_item,
2158 static struct config_item_type o2hb_heartbeat_group_type = {
2159 .ct_group_ops = &o2hb_heartbeat_group_group_ops,
2160 .ct_item_ops = &o2hb_hearbeat_group_item_ops,
2161 .ct_attrs = o2hb_heartbeat_group_attrs,
2162 .ct_owner = THIS_MODULE,
2165 /* this is just here to avoid touching group in heartbeat.h which the
2166 * entire damn world #includes */
2167 struct config_group *o2hb_alloc_hb_set(void)
2169 struct o2hb_heartbeat_group *hs = NULL;
2170 struct config_group *ret = NULL;
2172 hs = kzalloc(sizeof(struct o2hb_heartbeat_group), GFP_KERNEL);
2176 config_group_init_type_name(&hs->hs_group, "heartbeat",
2177 &o2hb_heartbeat_group_type);
2179 ret = &hs->hs_group;
2186 void o2hb_free_hb_set(struct config_group *group)
2188 struct o2hb_heartbeat_group *hs = to_o2hb_heartbeat_group(group);
2192 /* hb callback registration and issueing */
2194 static struct o2hb_callback *hbcall_from_type(enum o2hb_callback_type type)
2196 if (type == O2HB_NUM_CB)
2197 return ERR_PTR(-EINVAL);
2199 return &o2hb_callbacks[type];
2202 void o2hb_setup_callback(struct o2hb_callback_func *hc,
2203 enum o2hb_callback_type type,
2208 INIT_LIST_HEAD(&hc->hc_item);
2211 hc->hc_priority = priority;
2213 hc->hc_magic = O2HB_CB_MAGIC;
2215 EXPORT_SYMBOL_GPL(o2hb_setup_callback);
2217 static struct o2hb_region *o2hb_find_region(const char *region_uuid)
2219 struct o2hb_region *p, *reg = NULL;
2221 assert_spin_locked(&o2hb_live_lock);
2223 list_for_each_entry(p, &o2hb_all_regions, hr_all_item) {
2224 if (!strcmp(region_uuid, config_item_name(&p->hr_item))) {
2233 static int o2hb_region_get(const char *region_uuid)
2236 struct o2hb_region *reg;
2238 spin_lock(&o2hb_live_lock);
2240 reg = o2hb_find_region(region_uuid);
2243 spin_unlock(&o2hb_live_lock);
2248 ret = o2nm_depend_this_node();
2252 ret = o2nm_depend_item(®->hr_item);
2254 o2nm_undepend_this_node();
2260 static void o2hb_region_put(const char *region_uuid)
2262 struct o2hb_region *reg;
2264 spin_lock(&o2hb_live_lock);
2266 reg = o2hb_find_region(region_uuid);
2268 spin_unlock(&o2hb_live_lock);
2271 o2nm_undepend_item(®->hr_item);
2272 o2nm_undepend_this_node();
2276 int o2hb_register_callback(const char *region_uuid,
2277 struct o2hb_callback_func *hc)
2279 struct o2hb_callback_func *tmp;
2280 struct list_head *iter;
2281 struct o2hb_callback *hbcall;
2284 BUG_ON(hc->hc_magic != O2HB_CB_MAGIC);
2285 BUG_ON(!list_empty(&hc->hc_item));
2287 hbcall = hbcall_from_type(hc->hc_type);
2288 if (IS_ERR(hbcall)) {
2289 ret = PTR_ERR(hbcall);
2294 ret = o2hb_region_get(region_uuid);
2299 down_write(&o2hb_callback_sem);
2301 list_for_each(iter, &hbcall->list) {
2302 tmp = list_entry(iter, struct o2hb_callback_func, hc_item);
2303 if (hc->hc_priority < tmp->hc_priority) {
2304 list_add_tail(&hc->hc_item, iter);
2308 if (list_empty(&hc->hc_item))
2309 list_add_tail(&hc->hc_item, &hbcall->list);
2311 up_write(&o2hb_callback_sem);
2314 mlog(ML_HEARTBEAT, "returning %d on behalf of %p for funcs %p\n",
2315 ret, __builtin_return_address(0), hc);
2318 EXPORT_SYMBOL_GPL(o2hb_register_callback);
2320 void o2hb_unregister_callback(const char *region_uuid,
2321 struct o2hb_callback_func *hc)
2323 BUG_ON(hc->hc_magic != O2HB_CB_MAGIC);
2325 mlog(ML_HEARTBEAT, "on behalf of %p for funcs %p\n",
2326 __builtin_return_address(0), hc);
2328 /* XXX Can this happen _with_ a region reference? */
2329 if (list_empty(&hc->hc_item))
2333 o2hb_region_put(region_uuid);
2335 down_write(&o2hb_callback_sem);
2337 list_del_init(&hc->hc_item);
2339 up_write(&o2hb_callback_sem);
2341 EXPORT_SYMBOL_GPL(o2hb_unregister_callback);
2343 int o2hb_check_node_heartbeating(u8 node_num)
2345 unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
2347 o2hb_fill_node_map(testing_map, sizeof(testing_map));
2348 if (!test_bit(node_num, testing_map)) {
2350 "node (%u) does not have heartbeating enabled.\n",
2357 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating);
2359 int o2hb_check_node_heartbeating_from_callback(u8 node_num)
2361 unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
2363 o2hb_fill_node_map_from_callback(testing_map, sizeof(testing_map));
2364 if (!test_bit(node_num, testing_map)) {
2366 "node (%u) does not have heartbeating enabled.\n",
2373 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating_from_callback);
2375 /* Makes sure our local node is configured with a node number, and is
2377 int o2hb_check_local_node_heartbeating(void)
2381 /* if this node was set then we have networking */
2382 node_num = o2nm_this_node();
2383 if (node_num == O2NM_MAX_NODES) {
2384 mlog(ML_HEARTBEAT, "this node has not been configured.\n");
2388 return o2hb_check_node_heartbeating(node_num);
2390 EXPORT_SYMBOL_GPL(o2hb_check_local_node_heartbeating);
2393 * this is just a hack until we get the plumbing which flips file systems
2394 * read only and drops the hb ref instead of killing the node dead.
2396 void o2hb_stop_all_regions(void)
2398 struct o2hb_region *reg;
2400 mlog(ML_ERROR, "stopping heartbeat on all active regions.\n");
2402 spin_lock(&o2hb_live_lock);
2404 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item)
2405 reg->hr_unclean_stop = 1;
2407 spin_unlock(&o2hb_live_lock);
2409 EXPORT_SYMBOL_GPL(o2hb_stop_all_regions);
2411 int o2hb_get_all_regions(char *region_uuids, u8 max_regions)
2413 struct o2hb_region *reg;
2417 spin_lock(&o2hb_live_lock);
2420 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) {
2421 mlog(0, "Region: %s\n", config_item_name(®->hr_item));
2422 if (numregs < max_regions) {
2423 memcpy(p, config_item_name(®->hr_item),
2424 O2HB_MAX_REGION_NAME_LEN);
2425 p += O2HB_MAX_REGION_NAME_LEN;
2430 spin_unlock(&o2hb_live_lock);
2434 EXPORT_SYMBOL_GPL(o2hb_get_all_regions);
2436 int o2hb_global_heartbeat_active(void)
2438 return (o2hb_heartbeat_mode == O2HB_HEARTBEAT_GLOBAL);
2440 EXPORT_SYMBOL(o2hb_global_heartbeat_active);