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(®->hr_write_timeout_work);
311 flush_scheduled_work();
314 static inline void o2hb_bio_wait_init(struct o2hb_bio_wait_ctxt *wc)
316 atomic_set(&wc->wc_num_reqs, 1);
317 init_completion(&wc->wc_io_complete);
321 /* Used in error paths too */
322 static inline void o2hb_bio_wait_dec(struct o2hb_bio_wait_ctxt *wc,
325 /* sadly atomic_sub_and_test() isn't available on all platforms. The
326 * good news is that the fast path only completes one at a time */
328 if (atomic_dec_and_test(&wc->wc_num_reqs)) {
330 complete(&wc->wc_io_complete);
335 static void o2hb_wait_on_io(struct o2hb_region *reg,
336 struct o2hb_bio_wait_ctxt *wc)
338 struct address_space *mapping = reg->hr_bdev->bd_inode->i_mapping;
340 blk_run_address_space(mapping);
341 o2hb_bio_wait_dec(wc, 1);
343 wait_for_completion(&wc->wc_io_complete);
346 static void o2hb_bio_end_io(struct bio *bio,
349 struct o2hb_bio_wait_ctxt *wc = bio->bi_private;
352 mlog(ML_ERROR, "IO Error %d\n", error);
353 wc->wc_error = error;
356 o2hb_bio_wait_dec(wc, 1);
360 /* Setup a Bio to cover I/O against num_slots slots starting at
362 static struct bio *o2hb_setup_one_bio(struct o2hb_region *reg,
363 struct o2hb_bio_wait_ctxt *wc,
364 unsigned int *current_slot,
365 unsigned int max_slots)
367 int len, current_page;
368 unsigned int vec_len, vec_start;
369 unsigned int bits = reg->hr_block_bits;
370 unsigned int spp = reg->hr_slots_per_page;
371 unsigned int cs = *current_slot;
375 /* Testing has shown this allocation to take long enough under
376 * GFP_KERNEL that the local node can get fenced. It would be
377 * nicest if we could pre-allocate these bios and avoid this
379 bio = bio_alloc(GFP_ATOMIC, 16);
381 mlog(ML_ERROR, "Could not alloc slots BIO!\n");
382 bio = ERR_PTR(-ENOMEM);
386 /* Must put everything in 512 byte sectors for the bio... */
387 bio->bi_sector = (reg->hr_start_block + cs) << (bits - 9);
388 bio->bi_bdev = reg->hr_bdev;
389 bio->bi_private = wc;
390 bio->bi_end_io = o2hb_bio_end_io;
392 vec_start = (cs << bits) % PAGE_CACHE_SIZE;
393 while(cs < max_slots) {
394 current_page = cs / spp;
395 page = reg->hr_slot_data[current_page];
397 vec_len = min(PAGE_CACHE_SIZE - vec_start,
398 (max_slots-cs) * (PAGE_CACHE_SIZE/spp) );
400 mlog(ML_HB_BIO, "page %d, vec_len = %u, vec_start = %u\n",
401 current_page, vec_len, vec_start);
403 len = bio_add_page(bio, page, vec_len, vec_start);
404 if (len != vec_len) break;
406 cs += vec_len / (PAGE_CACHE_SIZE/spp);
415 static int o2hb_read_slots(struct o2hb_region *reg,
416 unsigned int max_slots)
418 unsigned int current_slot=0;
420 struct o2hb_bio_wait_ctxt wc;
423 o2hb_bio_wait_init(&wc);
425 while(current_slot < max_slots) {
426 bio = o2hb_setup_one_bio(reg, &wc, ¤t_slot, max_slots);
428 status = PTR_ERR(bio);
433 atomic_inc(&wc.wc_num_reqs);
434 submit_bio(READ, bio);
440 o2hb_wait_on_io(reg, &wc);
441 if (wc.wc_error && !status)
442 status = wc.wc_error;
447 static int o2hb_issue_node_write(struct o2hb_region *reg,
448 struct o2hb_bio_wait_ctxt *write_wc)
454 o2hb_bio_wait_init(write_wc);
456 slot = o2nm_this_node();
458 bio = o2hb_setup_one_bio(reg, write_wc, &slot, slot+1);
460 status = PTR_ERR(bio);
465 atomic_inc(&write_wc->wc_num_reqs);
466 submit_bio(WRITE, bio);
473 static u32 o2hb_compute_block_crc_le(struct o2hb_region *reg,
474 struct o2hb_disk_heartbeat_block *hb_block)
479 /* We want to compute the block crc with a 0 value in the
480 * hb_cksum field. Save it off here and replace after the
482 old_cksum = hb_block->hb_cksum;
483 hb_block->hb_cksum = 0;
485 ret = crc32_le(0, (unsigned char *) hb_block, reg->hr_block_bytes);
487 hb_block->hb_cksum = old_cksum;
492 static void o2hb_dump_slot(struct o2hb_disk_heartbeat_block *hb_block)
494 mlog(ML_ERROR, "Dump slot information: seq = 0x%llx, node = %u, "
495 "cksum = 0x%x, generation 0x%llx\n",
496 (long long)le64_to_cpu(hb_block->hb_seq),
497 hb_block->hb_node, le32_to_cpu(hb_block->hb_cksum),
498 (long long)le64_to_cpu(hb_block->hb_generation));
501 static int o2hb_verify_crc(struct o2hb_region *reg,
502 struct o2hb_disk_heartbeat_block *hb_block)
506 read = le32_to_cpu(hb_block->hb_cksum);
507 computed = o2hb_compute_block_crc_le(reg, hb_block);
509 return read == computed;
512 /* We want to make sure that nobody is heartbeating on top of us --
513 * this will help detect an invalid configuration. */
514 static int o2hb_check_last_timestamp(struct o2hb_region *reg)
517 struct o2hb_disk_slot *slot;
518 struct o2hb_disk_heartbeat_block *hb_block;
520 node_num = o2nm_this_node();
523 slot = ®->hr_slots[node_num];
524 /* Don't check on our 1st timestamp */
525 if (slot->ds_last_time) {
526 hb_block = slot->ds_raw_block;
528 if (le64_to_cpu(hb_block->hb_seq) != slot->ds_last_time)
535 static inline void o2hb_prepare_block(struct o2hb_region *reg,
540 struct o2hb_disk_slot *slot;
541 struct o2hb_disk_heartbeat_block *hb_block;
543 node_num = o2nm_this_node();
544 slot = ®->hr_slots[node_num];
546 hb_block = (struct o2hb_disk_heartbeat_block *)slot->ds_raw_block;
547 memset(hb_block, 0, reg->hr_block_bytes);
548 /* TODO: time stuff */
549 cputime = CURRENT_TIME.tv_sec;
553 hb_block->hb_seq = cpu_to_le64(cputime);
554 hb_block->hb_node = node_num;
555 hb_block->hb_generation = cpu_to_le64(generation);
556 hb_block->hb_dead_ms = cpu_to_le32(o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS);
558 /* This step must always happen last! */
559 hb_block->hb_cksum = cpu_to_le32(o2hb_compute_block_crc_le(reg,
562 mlog(ML_HB_BIO, "our node generation = 0x%llx, cksum = 0x%x\n",
563 (long long)generation,
564 le32_to_cpu(hb_block->hb_cksum));
567 static void o2hb_fire_callbacks(struct o2hb_callback *hbcall,
568 struct o2nm_node *node,
571 struct list_head *iter;
572 struct o2hb_callback_func *f;
574 list_for_each(iter, &hbcall->list) {
575 f = list_entry(iter, struct o2hb_callback_func, hc_item);
576 mlog(ML_HEARTBEAT, "calling funcs %p\n", f);
577 (f->hc_func)(node, idx, f->hc_data);
581 /* Will run the list in order until we process the passed event */
582 static void o2hb_run_event_list(struct o2hb_node_event *queued_event)
585 struct o2hb_callback *hbcall;
586 struct o2hb_node_event *event;
588 spin_lock(&o2hb_live_lock);
589 empty = list_empty(&queued_event->hn_item);
590 spin_unlock(&o2hb_live_lock);
594 /* Holding callback sem assures we don't alter the callback
595 * lists when doing this, and serializes ourselves with other
596 * processes wanting callbacks. */
597 down_write(&o2hb_callback_sem);
599 spin_lock(&o2hb_live_lock);
600 while (!list_empty(&o2hb_node_events)
601 && !list_empty(&queued_event->hn_item)) {
602 event = list_entry(o2hb_node_events.next,
603 struct o2hb_node_event,
605 list_del_init(&event->hn_item);
606 spin_unlock(&o2hb_live_lock);
608 mlog(ML_HEARTBEAT, "Node %s event for %d\n",
609 event->hn_event_type == O2HB_NODE_UP_CB ? "UP" : "DOWN",
612 hbcall = hbcall_from_type(event->hn_event_type);
614 /* We should *never* have gotten on to the list with a
615 * bad type... This isn't something that we should try
616 * to recover from. */
617 BUG_ON(IS_ERR(hbcall));
619 o2hb_fire_callbacks(hbcall, event->hn_node, event->hn_node_num);
621 spin_lock(&o2hb_live_lock);
623 spin_unlock(&o2hb_live_lock);
625 up_write(&o2hb_callback_sem);
628 static void o2hb_queue_node_event(struct o2hb_node_event *event,
629 enum o2hb_callback_type type,
630 struct o2nm_node *node,
633 assert_spin_locked(&o2hb_live_lock);
635 BUG_ON((!node) && (type != O2HB_NODE_DOWN_CB));
637 event->hn_event_type = type;
638 event->hn_node = node;
639 event->hn_node_num = node_num;
641 mlog(ML_HEARTBEAT, "Queue node %s event for node %d\n",
642 type == O2HB_NODE_UP_CB ? "UP" : "DOWN", node_num);
644 list_add_tail(&event->hn_item, &o2hb_node_events);
647 static void o2hb_shutdown_slot(struct o2hb_disk_slot *slot)
649 struct o2hb_node_event event =
650 { .hn_item = LIST_HEAD_INIT(event.hn_item), };
651 struct o2nm_node *node;
653 node = o2nm_get_node_by_num(slot->ds_node_num);
657 spin_lock(&o2hb_live_lock);
658 if (!list_empty(&slot->ds_live_item)) {
659 mlog(ML_HEARTBEAT, "Shutdown, node %d leaves region\n",
662 list_del_init(&slot->ds_live_item);
664 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
665 clear_bit(slot->ds_node_num, o2hb_live_node_bitmap);
667 o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB, node,
671 spin_unlock(&o2hb_live_lock);
673 o2hb_run_event_list(&event);
678 static void o2hb_set_quorum_device(struct o2hb_region *reg,
679 struct o2hb_disk_slot *slot)
681 assert_spin_locked(&o2hb_live_lock);
683 if (!o2hb_global_heartbeat_active())
686 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap))
690 * A region can be added to the quorum only when it sees all
691 * live nodes heartbeat on it. In other words, the region has been
692 * added to all nodes.
694 if (memcmp(reg->hr_live_node_bitmap, o2hb_live_node_bitmap,
695 sizeof(o2hb_live_node_bitmap)))
698 if (slot->ds_changed_samples < O2HB_LIVE_THRESHOLD)
701 printk(KERN_NOTICE "o2hb: Region %s is now a quorum device\n",
702 config_item_name(®->hr_item));
704 set_bit(reg->hr_region_num, o2hb_quorum_region_bitmap);
707 static int o2hb_check_slot(struct o2hb_region *reg,
708 struct o2hb_disk_slot *slot)
710 int changed = 0, gen_changed = 0;
711 struct o2hb_node_event event =
712 { .hn_item = LIST_HEAD_INIT(event.hn_item), };
713 struct o2nm_node *node;
714 struct o2hb_disk_heartbeat_block *hb_block = reg->hr_tmp_block;
716 unsigned int dead_ms = o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS;
717 unsigned int slot_dead_ms;
720 memcpy(hb_block, slot->ds_raw_block, reg->hr_block_bytes);
723 * If a node is no longer configured but is still in the livemap, we
724 * may need to clear that bit from the livemap.
726 node = o2nm_get_node_by_num(slot->ds_node_num);
728 spin_lock(&o2hb_live_lock);
729 tmp = test_bit(slot->ds_node_num, o2hb_live_node_bitmap);
730 spin_unlock(&o2hb_live_lock);
735 if (!o2hb_verify_crc(reg, hb_block)) {
736 /* all paths from here will drop o2hb_live_lock for
738 spin_lock(&o2hb_live_lock);
740 /* Don't print an error on the console in this case -
741 * a freshly formatted heartbeat area will not have a
743 if (list_empty(&slot->ds_live_item))
746 /* The node is live but pushed out a bad crc. We
747 * consider it a transient miss but don't populate any
748 * other values as they may be junk. */
749 mlog(ML_ERROR, "Node %d has written a bad crc to %s\n",
750 slot->ds_node_num, reg->hr_dev_name);
751 o2hb_dump_slot(hb_block);
753 slot->ds_equal_samples++;
757 /* we don't care if these wrap.. the state transitions below
758 * clear at the right places */
759 cputime = le64_to_cpu(hb_block->hb_seq);
760 if (slot->ds_last_time != cputime)
761 slot->ds_changed_samples++;
763 slot->ds_equal_samples++;
764 slot->ds_last_time = cputime;
766 /* The node changed heartbeat generations. We assume this to
767 * mean it dropped off but came back before we timed out. We
768 * want to consider it down for the time being but don't want
769 * to lose any changed_samples state we might build up to
770 * considering it live again. */
771 if (slot->ds_last_generation != le64_to_cpu(hb_block->hb_generation)) {
773 slot->ds_equal_samples = 0;
774 mlog(ML_HEARTBEAT, "Node %d changed generation (0x%llx "
775 "to 0x%llx)\n", slot->ds_node_num,
776 (long long)slot->ds_last_generation,
777 (long long)le64_to_cpu(hb_block->hb_generation));
780 slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation);
782 mlog(ML_HEARTBEAT, "Slot %d gen 0x%llx cksum 0x%x "
783 "seq %llu last %llu changed %u equal %u\n",
784 slot->ds_node_num, (long long)slot->ds_last_generation,
785 le32_to_cpu(hb_block->hb_cksum),
786 (unsigned long long)le64_to_cpu(hb_block->hb_seq),
787 (unsigned long long)slot->ds_last_time, slot->ds_changed_samples,
788 slot->ds_equal_samples);
790 spin_lock(&o2hb_live_lock);
793 /* dead nodes only come to life after some number of
794 * changes at any time during their dead time */
795 if (list_empty(&slot->ds_live_item) &&
796 slot->ds_changed_samples >= O2HB_LIVE_THRESHOLD) {
797 mlog(ML_HEARTBEAT, "Node %d (id 0x%llx) joined my region\n",
798 slot->ds_node_num, (long long)slot->ds_last_generation);
800 set_bit(slot->ds_node_num, reg->hr_live_node_bitmap);
802 /* first on the list generates a callback */
803 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
804 mlog(ML_HEARTBEAT, "o2hb: Add node %d to live nodes "
805 "bitmap\n", slot->ds_node_num);
806 set_bit(slot->ds_node_num, o2hb_live_node_bitmap);
808 o2hb_queue_node_event(&event, O2HB_NODE_UP_CB, node,
814 list_add_tail(&slot->ds_live_item,
815 &o2hb_live_slots[slot->ds_node_num]);
817 slot->ds_equal_samples = 0;
819 /* We want to be sure that all nodes agree on the
820 * number of milliseconds before a node will be
821 * considered dead. The self-fencing timeout is
822 * computed from this value, and a discrepancy might
823 * result in heartbeat calling a node dead when it
824 * hasn't self-fenced yet. */
825 slot_dead_ms = le32_to_cpu(hb_block->hb_dead_ms);
826 if (slot_dead_ms && slot_dead_ms != dead_ms) {
827 /* TODO: Perhaps we can fail the region here. */
828 mlog(ML_ERROR, "Node %d on device %s has a dead count "
829 "of %u ms, but our count is %u ms.\n"
830 "Please double check your configuration values "
831 "for 'O2CB_HEARTBEAT_THRESHOLD'\n",
832 slot->ds_node_num, reg->hr_dev_name, slot_dead_ms,
838 /* if the list is dead, we're done.. */
839 if (list_empty(&slot->ds_live_item))
842 /* live nodes only go dead after enough consequtive missed
843 * samples.. reset the missed counter whenever we see
845 if (slot->ds_equal_samples >= o2hb_dead_threshold || gen_changed) {
846 mlog(ML_HEARTBEAT, "Node %d left my region\n",
849 clear_bit(slot->ds_node_num, reg->hr_live_node_bitmap);
851 /* last off the live_slot generates a callback */
852 list_del_init(&slot->ds_live_item);
853 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
854 mlog(ML_HEARTBEAT, "o2hb: Remove node %d from live "
855 "nodes bitmap\n", slot->ds_node_num);
856 clear_bit(slot->ds_node_num, o2hb_live_node_bitmap);
858 /* node can be null */
859 o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB,
860 node, slot->ds_node_num);
865 /* We don't clear this because the node is still
866 * actually writing new blocks. */
868 slot->ds_changed_samples = 0;
871 if (slot->ds_changed_samples) {
872 slot->ds_changed_samples = 0;
873 slot->ds_equal_samples = 0;
876 o2hb_set_quorum_device(reg, slot);
878 spin_unlock(&o2hb_live_lock);
880 o2hb_run_event_list(&event);
887 /* This could be faster if we just implmented a find_last_bit, but I
888 * don't think the circumstances warrant it. */
889 static int o2hb_highest_node(unsigned long *nodes,
896 while ((node = find_next_bit(nodes, numbits, node + 1)) != -1) {
906 static int o2hb_do_disk_heartbeat(struct o2hb_region *reg)
908 int i, ret, highest_node, change = 0;
909 unsigned long configured_nodes[BITS_TO_LONGS(O2NM_MAX_NODES)];
910 unsigned long live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
911 struct o2hb_bio_wait_ctxt write_wc;
913 ret = o2nm_configured_node_map(configured_nodes,
914 sizeof(configured_nodes));
921 * If a node is not configured but is in the livemap, we still need
922 * to read the slot so as to be able to remove it from the livemap.
924 o2hb_fill_node_map(live_node_bitmap, sizeof(live_node_bitmap));
926 while ((i = find_next_bit(live_node_bitmap,
927 O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) {
928 set_bit(i, configured_nodes);
931 highest_node = o2hb_highest_node(configured_nodes, O2NM_MAX_NODES);
932 if (highest_node >= O2NM_MAX_NODES) {
933 mlog(ML_NOTICE, "ocfs2_heartbeat: no configured nodes found!\n");
937 /* No sense in reading the slots of nodes that don't exist
938 * yet. Of course, if the node definitions have holes in them
939 * then we're reading an empty slot anyway... Consider this
941 ret = o2hb_read_slots(reg, highest_node + 1);
947 /* With an up to date view of the slots, we can check that no
948 * other node has been improperly configured to heartbeat in
950 if (!o2hb_check_last_timestamp(reg))
951 mlog(ML_ERROR, "Device \"%s\": another node is heartbeating "
952 "in our slot!\n", reg->hr_dev_name);
954 /* fill in the proper info for our next heartbeat */
955 o2hb_prepare_block(reg, reg->hr_generation);
957 /* And fire off the write. Note that we don't wait on this I/O
959 ret = o2hb_issue_node_write(reg, &write_wc);
966 while((i = find_next_bit(configured_nodes, O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) {
968 change |= o2hb_check_slot(reg, ®->hr_slots[i]);
972 * We have to be sure we've advertised ourselves on disk
973 * before we can go to steady state. This ensures that
974 * people we find in our steady state have seen us.
976 o2hb_wait_on_io(reg, &write_wc);
977 if (write_wc.wc_error) {
978 /* Do not re-arm the write timeout on I/O error - we
979 * can't be sure that the new block ever made it to
981 mlog(ML_ERROR, "Write error %d on device \"%s\"\n",
982 write_wc.wc_error, reg->hr_dev_name);
983 return write_wc.wc_error;
986 o2hb_arm_write_timeout(reg);
988 /* let the person who launched us know when things are steady */
989 if (!change && (atomic_read(®->hr_steady_iterations) != 0)) {
990 if (atomic_dec_and_test(®->hr_steady_iterations))
991 wake_up(&o2hb_steady_queue);
997 /* Subtract b from a, storing the result in a. a *must* have a larger
999 static void o2hb_tv_subtract(struct timeval *a,
1002 /* just return 0 when a is after b */
1003 if (a->tv_sec < b->tv_sec ||
1004 (a->tv_sec == b->tv_sec && a->tv_usec < b->tv_usec)) {
1010 a->tv_sec -= b->tv_sec;
1011 a->tv_usec -= b->tv_usec;
1012 while ( a->tv_usec < 0 ) {
1014 a->tv_usec += 1000000;
1018 static unsigned int o2hb_elapsed_msecs(struct timeval *start,
1019 struct timeval *end)
1021 struct timeval res = *end;
1023 o2hb_tv_subtract(&res, start);
1025 return res.tv_sec * 1000 + res.tv_usec / 1000;
1029 * we ride the region ref that the region dir holds. before the region
1030 * dir is removed and drops it ref it will wait to tear down this
1033 static int o2hb_thread(void *data)
1036 struct o2hb_region *reg = data;
1037 struct o2hb_bio_wait_ctxt write_wc;
1038 struct timeval before_hb, after_hb;
1039 unsigned int elapsed_msec;
1041 mlog(ML_HEARTBEAT|ML_KTHREAD, "hb thread running\n");
1043 set_user_nice(current, -20);
1045 while (!kthread_should_stop() && !reg->hr_unclean_stop) {
1046 /* We track the time spent inside
1047 * o2hb_do_disk_heartbeat so that we avoid more than
1048 * hr_timeout_ms between disk writes. On busy systems
1049 * this should result in a heartbeat which is less
1050 * likely to time itself out. */
1051 do_gettimeofday(&before_hb);
1055 ret = o2hb_do_disk_heartbeat(reg);
1056 } while (ret && ++i < 2);
1058 do_gettimeofday(&after_hb);
1059 elapsed_msec = o2hb_elapsed_msecs(&before_hb, &after_hb);
1062 "start = %lu.%lu, end = %lu.%lu, msec = %u\n",
1063 before_hb.tv_sec, (unsigned long) before_hb.tv_usec,
1064 after_hb.tv_sec, (unsigned long) after_hb.tv_usec,
1067 if (elapsed_msec < reg->hr_timeout_ms) {
1068 /* the kthread api has blocked signals for us so no
1069 * need to record the return value. */
1070 msleep_interruptible(reg->hr_timeout_ms - elapsed_msec);
1074 o2hb_disarm_write_timeout(reg);
1076 /* unclean stop is only used in very bad situation */
1077 for(i = 0; !reg->hr_unclean_stop && i < reg->hr_blocks; i++)
1078 o2hb_shutdown_slot(®->hr_slots[i]);
1080 /* Explicit down notification - avoid forcing the other nodes
1081 * to timeout on this region when we could just as easily
1082 * write a clear generation - thus indicating to them that
1083 * this node has left this region.
1085 * XXX: Should we skip this on unclean_stop? */
1086 o2hb_prepare_block(reg, 0);
1087 ret = o2hb_issue_node_write(reg, &write_wc);
1089 o2hb_wait_on_io(reg, &write_wc);
1094 mlog(ML_HEARTBEAT|ML_KTHREAD, "hb thread exiting\n");
1099 #ifdef CONFIG_DEBUG_FS
1100 static int o2hb_debug_open(struct inode *inode, struct file *file)
1102 struct o2hb_debug_buf *db = inode->i_private;
1103 struct o2hb_region *reg;
1104 unsigned long map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1109 /* max_nodes should be the largest bitmap we pass here */
1110 BUG_ON(sizeof(map) < db->db_size);
1112 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
1116 switch (db->db_type) {
1117 case O2HB_DB_TYPE_LIVENODES:
1118 case O2HB_DB_TYPE_LIVEREGIONS:
1119 case O2HB_DB_TYPE_QUORUMREGIONS:
1120 case O2HB_DB_TYPE_FAILEDREGIONS:
1121 spin_lock(&o2hb_live_lock);
1122 memcpy(map, db->db_data, db->db_size);
1123 spin_unlock(&o2hb_live_lock);
1126 case O2HB_DB_TYPE_REGION_LIVENODES:
1127 spin_lock(&o2hb_live_lock);
1128 reg = (struct o2hb_region *)db->db_data;
1129 memcpy(map, reg->hr_live_node_bitmap, db->db_size);
1130 spin_unlock(&o2hb_live_lock);
1133 case O2HB_DB_TYPE_REGION_NUMBER:
1134 reg = (struct o2hb_region *)db->db_data;
1135 out += snprintf(buf + out, PAGE_SIZE - out, "%d\n",
1136 reg->hr_region_num);
1139 case O2HB_DB_TYPE_REGION_ELAPSED_TIME:
1140 reg = (struct o2hb_region *)db->db_data;
1141 out += snprintf(buf + out, PAGE_SIZE - out, "%u\n",
1142 jiffies_to_msecs(jiffies -
1143 reg->hr_last_timeout_start));
1150 while ((i = find_next_bit(map, db->db_len, i + 1)) < db->db_len)
1151 out += snprintf(buf + out, PAGE_SIZE - out, "%d ", i);
1152 out += snprintf(buf + out, PAGE_SIZE - out, "\n");
1155 i_size_write(inode, out);
1157 file->private_data = buf;
1164 static int o2hb_debug_release(struct inode *inode, struct file *file)
1166 kfree(file->private_data);
1170 static ssize_t o2hb_debug_read(struct file *file, char __user *buf,
1171 size_t nbytes, loff_t *ppos)
1173 return simple_read_from_buffer(buf, nbytes, ppos, file->private_data,
1174 i_size_read(file->f_mapping->host));
1177 static int o2hb_debug_open(struct inode *inode, struct file *file)
1181 static int o2hb_debug_release(struct inode *inode, struct file *file)
1185 static ssize_t o2hb_debug_read(struct file *file, char __user *buf,
1186 size_t nbytes, loff_t *ppos)
1190 #endif /* CONFIG_DEBUG_FS */
1192 static const struct file_operations o2hb_debug_fops = {
1193 .open = o2hb_debug_open,
1194 .release = o2hb_debug_release,
1195 .read = o2hb_debug_read,
1196 .llseek = generic_file_llseek,
1199 void o2hb_exit(void)
1201 kfree(o2hb_db_livenodes);
1202 kfree(o2hb_db_liveregions);
1203 kfree(o2hb_db_quorumregions);
1204 kfree(o2hb_db_failedregions);
1205 debugfs_remove(o2hb_debug_failedregions);
1206 debugfs_remove(o2hb_debug_quorumregions);
1207 debugfs_remove(o2hb_debug_liveregions);
1208 debugfs_remove(o2hb_debug_livenodes);
1209 debugfs_remove(o2hb_debug_dir);
1212 static struct dentry *o2hb_debug_create(const char *name, struct dentry *dir,
1213 struct o2hb_debug_buf **db, int db_len,
1214 int type, int size, int len, void *data)
1216 *db = kmalloc(db_len, GFP_KERNEL);
1220 (*db)->db_type = type;
1221 (*db)->db_size = size;
1222 (*db)->db_len = len;
1223 (*db)->db_data = data;
1225 return debugfs_create_file(name, S_IFREG|S_IRUSR, dir, *db,
1229 static int o2hb_debug_init(void)
1233 o2hb_debug_dir = debugfs_create_dir(O2HB_DEBUG_DIR, NULL);
1234 if (!o2hb_debug_dir) {
1239 o2hb_debug_livenodes = o2hb_debug_create(O2HB_DEBUG_LIVENODES,
1242 sizeof(*o2hb_db_livenodes),
1243 O2HB_DB_TYPE_LIVENODES,
1244 sizeof(o2hb_live_node_bitmap),
1246 o2hb_live_node_bitmap);
1247 if (!o2hb_debug_livenodes) {
1252 o2hb_debug_liveregions = o2hb_debug_create(O2HB_DEBUG_LIVEREGIONS,
1254 &o2hb_db_liveregions,
1255 sizeof(*o2hb_db_liveregions),
1256 O2HB_DB_TYPE_LIVEREGIONS,
1257 sizeof(o2hb_live_region_bitmap),
1259 o2hb_live_region_bitmap);
1260 if (!o2hb_debug_liveregions) {
1265 o2hb_debug_quorumregions =
1266 o2hb_debug_create(O2HB_DEBUG_QUORUMREGIONS,
1268 &o2hb_db_quorumregions,
1269 sizeof(*o2hb_db_quorumregions),
1270 O2HB_DB_TYPE_QUORUMREGIONS,
1271 sizeof(o2hb_quorum_region_bitmap),
1273 o2hb_quorum_region_bitmap);
1274 if (!o2hb_debug_quorumregions) {
1279 o2hb_debug_failedregions =
1280 o2hb_debug_create(O2HB_DEBUG_FAILEDREGIONS,
1282 &o2hb_db_failedregions,
1283 sizeof(*o2hb_db_failedregions),
1284 O2HB_DB_TYPE_FAILEDREGIONS,
1285 sizeof(o2hb_failed_region_bitmap),
1287 o2hb_failed_region_bitmap);
1288 if (!o2hb_debug_failedregions) {
1305 for (i = 0; i < ARRAY_SIZE(o2hb_callbacks); i++)
1306 INIT_LIST_HEAD(&o2hb_callbacks[i].list);
1308 for (i = 0; i < ARRAY_SIZE(o2hb_live_slots); i++)
1309 INIT_LIST_HEAD(&o2hb_live_slots[i]);
1311 INIT_LIST_HEAD(&o2hb_node_events);
1313 memset(o2hb_live_node_bitmap, 0, sizeof(o2hb_live_node_bitmap));
1314 memset(o2hb_region_bitmap, 0, sizeof(o2hb_region_bitmap));
1315 memset(o2hb_live_region_bitmap, 0, sizeof(o2hb_live_region_bitmap));
1316 memset(o2hb_quorum_region_bitmap, 0, sizeof(o2hb_quorum_region_bitmap));
1317 memset(o2hb_failed_region_bitmap, 0, sizeof(o2hb_failed_region_bitmap));
1319 return o2hb_debug_init();
1322 /* if we're already in a callback then we're already serialized by the sem */
1323 static void o2hb_fill_node_map_from_callback(unsigned long *map,
1326 BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1328 memcpy(map, &o2hb_live_node_bitmap, bytes);
1332 * get a map of all nodes that are heartbeating in any regions
1334 void o2hb_fill_node_map(unsigned long *map, unsigned bytes)
1336 /* callers want to serialize this map and callbacks so that they
1337 * can trust that they don't miss nodes coming to the party */
1338 down_read(&o2hb_callback_sem);
1339 spin_lock(&o2hb_live_lock);
1340 o2hb_fill_node_map_from_callback(map, bytes);
1341 spin_unlock(&o2hb_live_lock);
1342 up_read(&o2hb_callback_sem);
1344 EXPORT_SYMBOL_GPL(o2hb_fill_node_map);
1347 * heartbeat configfs bits. The heartbeat set is a default set under
1348 * the cluster set in nodemanager.c.
1351 static struct o2hb_region *to_o2hb_region(struct config_item *item)
1353 return item ? container_of(item, struct o2hb_region, hr_item) : NULL;
1356 /* drop_item only drops its ref after killing the thread, nothing should
1357 * be using the region anymore. this has to clean up any state that
1358 * attributes might have built up. */
1359 static void o2hb_region_release(struct config_item *item)
1363 struct o2hb_region *reg = to_o2hb_region(item);
1365 if (reg->hr_tmp_block)
1366 kfree(reg->hr_tmp_block);
1368 if (reg->hr_slot_data) {
1369 for (i = 0; i < reg->hr_num_pages; i++) {
1370 page = reg->hr_slot_data[i];
1374 kfree(reg->hr_slot_data);
1378 blkdev_put(reg->hr_bdev, FMODE_READ|FMODE_WRITE);
1381 kfree(reg->hr_slots);
1383 kfree(reg->hr_db_regnum);
1384 kfree(reg->hr_db_livenodes);
1385 debugfs_remove(reg->hr_debug_livenodes);
1386 debugfs_remove(reg->hr_debug_regnum);
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 return ERR_PTR(-ENAMETOOLONG);
1969 spin_lock(&o2hb_live_lock);
1970 reg->hr_region_num = 0;
1971 if (o2hb_global_heartbeat_active()) {
1972 reg->hr_region_num = find_first_zero_bit(o2hb_region_bitmap,
1974 if (reg->hr_region_num >= O2NM_MAX_REGIONS) {
1975 spin_unlock(&o2hb_live_lock);
1976 return ERR_PTR(-EFBIG);
1978 set_bit(reg->hr_region_num, o2hb_region_bitmap);
1980 list_add_tail(®->hr_all_item, &o2hb_all_regions);
1981 spin_unlock(&o2hb_live_lock);
1983 config_item_init_type_name(®->hr_item, name, &o2hb_region_type);
1985 ret = o2hb_debug_region_init(reg, o2hb_debug_dir);
1987 config_item_put(®->hr_item);
1988 return ERR_PTR(ret);
1991 return ®->hr_item;
1994 static void o2hb_heartbeat_group_drop_item(struct config_group *group,
1995 struct config_item *item)
1997 struct task_struct *hb_task;
1998 struct o2hb_region *reg = to_o2hb_region(item);
2000 /* stop the thread when the user removes the region dir */
2001 spin_lock(&o2hb_live_lock);
2002 if (o2hb_global_heartbeat_active()) {
2003 clear_bit(reg->hr_region_num, o2hb_region_bitmap);
2004 clear_bit(reg->hr_region_num, o2hb_live_region_bitmap);
2006 hb_task = reg->hr_task;
2007 reg->hr_task = NULL;
2008 spin_unlock(&o2hb_live_lock);
2011 kthread_stop(hb_task);
2014 * If we're racing a dev_write(), we need to wake them. They will
2015 * check reg->hr_task
2017 if (atomic_read(®->hr_steady_iterations) != 0) {
2018 atomic_set(®->hr_steady_iterations, 0);
2019 wake_up(&o2hb_steady_queue);
2022 if (o2hb_global_heartbeat_active())
2023 printk(KERN_NOTICE "o2hb: Heartbeat stopped on region %s\n",
2024 config_item_name(®->hr_item));
2025 config_item_put(item);
2028 struct o2hb_heartbeat_group_attribute {
2029 struct configfs_attribute attr;
2030 ssize_t (*show)(struct o2hb_heartbeat_group *, char *);
2031 ssize_t (*store)(struct o2hb_heartbeat_group *, const char *, size_t);
2034 static ssize_t o2hb_heartbeat_group_show(struct config_item *item,
2035 struct configfs_attribute *attr,
2038 struct o2hb_heartbeat_group *reg = to_o2hb_heartbeat_group(to_config_group(item));
2039 struct o2hb_heartbeat_group_attribute *o2hb_heartbeat_group_attr =
2040 container_of(attr, struct o2hb_heartbeat_group_attribute, attr);
2043 if (o2hb_heartbeat_group_attr->show)
2044 ret = o2hb_heartbeat_group_attr->show(reg, page);
2048 static ssize_t o2hb_heartbeat_group_store(struct config_item *item,
2049 struct configfs_attribute *attr,
2050 const char *page, size_t count)
2052 struct o2hb_heartbeat_group *reg = to_o2hb_heartbeat_group(to_config_group(item));
2053 struct o2hb_heartbeat_group_attribute *o2hb_heartbeat_group_attr =
2054 container_of(attr, struct o2hb_heartbeat_group_attribute, attr);
2055 ssize_t ret = -EINVAL;
2057 if (o2hb_heartbeat_group_attr->store)
2058 ret = o2hb_heartbeat_group_attr->store(reg, page, count);
2062 static ssize_t o2hb_heartbeat_group_threshold_show(struct o2hb_heartbeat_group *group,
2065 return sprintf(page, "%u\n", o2hb_dead_threshold);
2068 static ssize_t o2hb_heartbeat_group_threshold_store(struct o2hb_heartbeat_group *group,
2073 char *p = (char *)page;
2075 tmp = simple_strtoul(p, &p, 10);
2076 if (!p || (*p && (*p != '\n')))
2079 /* this will validate ranges for us. */
2080 o2hb_dead_threshold_set((unsigned int) tmp);
2086 ssize_t o2hb_heartbeat_group_mode_show(struct o2hb_heartbeat_group *group,
2089 return sprintf(page, "%s\n",
2090 o2hb_heartbeat_mode_desc[o2hb_heartbeat_mode]);
2094 ssize_t o2hb_heartbeat_group_mode_store(struct o2hb_heartbeat_group *group,
2095 const char *page, size_t count)
2101 len = (page[count - 1] == '\n') ? count - 1 : count;
2105 for (i = 0; i < O2HB_HEARTBEAT_NUM_MODES; ++i) {
2106 if (strnicmp(page, o2hb_heartbeat_mode_desc[i], len))
2109 ret = o2hb_global_hearbeat_mode_set(i);
2111 printk(KERN_NOTICE "o2hb: Heartbeat mode set to %s\n",
2112 o2hb_heartbeat_mode_desc[i]);
2120 static struct o2hb_heartbeat_group_attribute o2hb_heartbeat_group_attr_threshold = {
2121 .attr = { .ca_owner = THIS_MODULE,
2122 .ca_name = "dead_threshold",
2123 .ca_mode = S_IRUGO | S_IWUSR },
2124 .show = o2hb_heartbeat_group_threshold_show,
2125 .store = o2hb_heartbeat_group_threshold_store,
2128 static struct o2hb_heartbeat_group_attribute o2hb_heartbeat_group_attr_mode = {
2129 .attr = { .ca_owner = THIS_MODULE,
2131 .ca_mode = S_IRUGO | S_IWUSR },
2132 .show = o2hb_heartbeat_group_mode_show,
2133 .store = o2hb_heartbeat_group_mode_store,
2136 static struct configfs_attribute *o2hb_heartbeat_group_attrs[] = {
2137 &o2hb_heartbeat_group_attr_threshold.attr,
2138 &o2hb_heartbeat_group_attr_mode.attr,
2142 static struct configfs_item_operations o2hb_hearbeat_group_item_ops = {
2143 .show_attribute = o2hb_heartbeat_group_show,
2144 .store_attribute = o2hb_heartbeat_group_store,
2147 static struct configfs_group_operations o2hb_heartbeat_group_group_ops = {
2148 .make_item = o2hb_heartbeat_group_make_item,
2149 .drop_item = o2hb_heartbeat_group_drop_item,
2152 static struct config_item_type o2hb_heartbeat_group_type = {
2153 .ct_group_ops = &o2hb_heartbeat_group_group_ops,
2154 .ct_item_ops = &o2hb_hearbeat_group_item_ops,
2155 .ct_attrs = o2hb_heartbeat_group_attrs,
2156 .ct_owner = THIS_MODULE,
2159 /* this is just here to avoid touching group in heartbeat.h which the
2160 * entire damn world #includes */
2161 struct config_group *o2hb_alloc_hb_set(void)
2163 struct o2hb_heartbeat_group *hs = NULL;
2164 struct config_group *ret = NULL;
2166 hs = kzalloc(sizeof(struct o2hb_heartbeat_group), GFP_KERNEL);
2170 config_group_init_type_name(&hs->hs_group, "heartbeat",
2171 &o2hb_heartbeat_group_type);
2173 ret = &hs->hs_group;
2180 void o2hb_free_hb_set(struct config_group *group)
2182 struct o2hb_heartbeat_group *hs = to_o2hb_heartbeat_group(group);
2186 /* hb callback registration and issueing */
2188 static struct o2hb_callback *hbcall_from_type(enum o2hb_callback_type type)
2190 if (type == O2HB_NUM_CB)
2191 return ERR_PTR(-EINVAL);
2193 return &o2hb_callbacks[type];
2196 void o2hb_setup_callback(struct o2hb_callback_func *hc,
2197 enum o2hb_callback_type type,
2202 INIT_LIST_HEAD(&hc->hc_item);
2205 hc->hc_priority = priority;
2207 hc->hc_magic = O2HB_CB_MAGIC;
2209 EXPORT_SYMBOL_GPL(o2hb_setup_callback);
2211 static struct o2hb_region *o2hb_find_region(const char *region_uuid)
2213 struct o2hb_region *p, *reg = NULL;
2215 assert_spin_locked(&o2hb_live_lock);
2217 list_for_each_entry(p, &o2hb_all_regions, hr_all_item) {
2218 if (!strcmp(region_uuid, config_item_name(&p->hr_item))) {
2227 static int o2hb_region_get(const char *region_uuid)
2230 struct o2hb_region *reg;
2232 spin_lock(&o2hb_live_lock);
2234 reg = o2hb_find_region(region_uuid);
2237 spin_unlock(&o2hb_live_lock);
2242 ret = o2nm_depend_this_node();
2246 ret = o2nm_depend_item(®->hr_item);
2248 o2nm_undepend_this_node();
2254 static void o2hb_region_put(const char *region_uuid)
2256 struct o2hb_region *reg;
2258 spin_lock(&o2hb_live_lock);
2260 reg = o2hb_find_region(region_uuid);
2262 spin_unlock(&o2hb_live_lock);
2265 o2nm_undepend_item(®->hr_item);
2266 o2nm_undepend_this_node();
2270 int o2hb_register_callback(const char *region_uuid,
2271 struct o2hb_callback_func *hc)
2273 struct o2hb_callback_func *tmp;
2274 struct list_head *iter;
2275 struct o2hb_callback *hbcall;
2278 BUG_ON(hc->hc_magic != O2HB_CB_MAGIC);
2279 BUG_ON(!list_empty(&hc->hc_item));
2281 hbcall = hbcall_from_type(hc->hc_type);
2282 if (IS_ERR(hbcall)) {
2283 ret = PTR_ERR(hbcall);
2288 ret = o2hb_region_get(region_uuid);
2293 down_write(&o2hb_callback_sem);
2295 list_for_each(iter, &hbcall->list) {
2296 tmp = list_entry(iter, struct o2hb_callback_func, hc_item);
2297 if (hc->hc_priority < tmp->hc_priority) {
2298 list_add_tail(&hc->hc_item, iter);
2302 if (list_empty(&hc->hc_item))
2303 list_add_tail(&hc->hc_item, &hbcall->list);
2305 up_write(&o2hb_callback_sem);
2308 mlog(ML_HEARTBEAT, "returning %d on behalf of %p for funcs %p\n",
2309 ret, __builtin_return_address(0), hc);
2312 EXPORT_SYMBOL_GPL(o2hb_register_callback);
2314 void o2hb_unregister_callback(const char *region_uuid,
2315 struct o2hb_callback_func *hc)
2317 BUG_ON(hc->hc_magic != O2HB_CB_MAGIC);
2319 mlog(ML_HEARTBEAT, "on behalf of %p for funcs %p\n",
2320 __builtin_return_address(0), hc);
2322 /* XXX Can this happen _with_ a region reference? */
2323 if (list_empty(&hc->hc_item))
2327 o2hb_region_put(region_uuid);
2329 down_write(&o2hb_callback_sem);
2331 list_del_init(&hc->hc_item);
2333 up_write(&o2hb_callback_sem);
2335 EXPORT_SYMBOL_GPL(o2hb_unregister_callback);
2337 int o2hb_check_node_heartbeating(u8 node_num)
2339 unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
2341 o2hb_fill_node_map(testing_map, sizeof(testing_map));
2342 if (!test_bit(node_num, testing_map)) {
2344 "node (%u) does not have heartbeating enabled.\n",
2351 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating);
2353 int o2hb_check_node_heartbeating_from_callback(u8 node_num)
2355 unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
2357 o2hb_fill_node_map_from_callback(testing_map, sizeof(testing_map));
2358 if (!test_bit(node_num, testing_map)) {
2360 "node (%u) does not have heartbeating enabled.\n",
2367 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating_from_callback);
2369 /* Makes sure our local node is configured with a node number, and is
2371 int o2hb_check_local_node_heartbeating(void)
2375 /* if this node was set then we have networking */
2376 node_num = o2nm_this_node();
2377 if (node_num == O2NM_MAX_NODES) {
2378 mlog(ML_HEARTBEAT, "this node has not been configured.\n");
2382 return o2hb_check_node_heartbeating(node_num);
2384 EXPORT_SYMBOL_GPL(o2hb_check_local_node_heartbeating);
2387 * this is just a hack until we get the plumbing which flips file systems
2388 * read only and drops the hb ref instead of killing the node dead.
2390 void o2hb_stop_all_regions(void)
2392 struct o2hb_region *reg;
2394 mlog(ML_ERROR, "stopping heartbeat on all active regions.\n");
2396 spin_lock(&o2hb_live_lock);
2398 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item)
2399 reg->hr_unclean_stop = 1;
2401 spin_unlock(&o2hb_live_lock);
2403 EXPORT_SYMBOL_GPL(o2hb_stop_all_regions);
2405 int o2hb_get_all_regions(char *region_uuids, u8 max_regions)
2407 struct o2hb_region *reg;
2411 spin_lock(&o2hb_live_lock);
2414 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) {
2415 mlog(0, "Region: %s\n", config_item_name(®->hr_item));
2416 if (numregs < max_regions) {
2417 memcpy(p, config_item_name(®->hr_item),
2418 O2HB_MAX_REGION_NAME_LEN);
2419 p += O2HB_MAX_REGION_NAME_LEN;
2424 spin_unlock(&o2hb_live_lock);
2428 EXPORT_SYMBOL_GPL(o2hb_get_all_regions);
2430 int o2hb_global_heartbeat_active(void)
2432 return (o2hb_heartbeat_mode == O2HB_HEARTBEAT_GLOBAL);
2434 EXPORT_SYMBOL(o2hb_global_heartbeat_active);