1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Copyright (C) 2004, 2005 Oracle. All rights reserved.
4 */
5
6 #include <linux/kernel.h>
7 #include <linux/sched.h>
8 #include <linux/jiffies.h>
9 #include <linux/module.h>
10 #include <linux/fs.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/delay.h>
14 #include <linux/file.h>
15 #include <linux/kthread.h>
16 #include <linux/configfs.h>
17 #include <linux/random.h>
18 #include <linux/crc32.h>
19 #include <linux/time.h>
20 #include <linux/debugfs.h>
21 #include <linux/slab.h>
22 #include <linux/bitmap.h>
23 #include <linux/ktime.h>
24 #include "heartbeat.h"
25 #include "tcp.h"
26 #include "nodemanager.h"
27 #include "quorum.h"
28
29 #include "masklog.h"
30
31
32 /*
33 * The first heartbeat pass had one global thread that would serialize all hb
34 * callback calls. This global serializing sem should only be removed once
35 * we've made sure that all callees can deal with being called concurrently
36 * from multiple hb region threads.
37 */
38 static DECLARE_RWSEM(o2hb_callback_sem);
39
40 /*
41 * multiple hb threads are watching multiple regions. A node is live
42 * whenever any of the threads sees activity from the node in its region.
43 */
44 static DEFINE_SPINLOCK(o2hb_live_lock);
45 static struct list_head o2hb_live_slots[O2NM_MAX_NODES];
46 static unsigned long o2hb_live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
47 static LIST_HEAD(o2hb_node_events);
48 static DECLARE_WAIT_QUEUE_HEAD(o2hb_steady_queue);
49
50 /*
51 * In global heartbeat, we maintain a series of region bitmaps.
52 * - o2hb_region_bitmap allows us to limit the region number to max region.
53 * - o2hb_live_region_bitmap tracks live regions (seen steady iterations).
54 * - o2hb_quorum_region_bitmap tracks live regions that have seen all nodes
55 * heartbeat on it.
56 * - o2hb_failed_region_bitmap tracks the regions that have seen io timeouts.
57 */
58 static unsigned long o2hb_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)];
59 static unsigned long o2hb_live_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)];
60 static unsigned long o2hb_quorum_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)];
61 static unsigned long o2hb_failed_region_bitmap[BITS_TO_LONGS(O2NM_MAX_REGIONS)];
62
63 #define O2HB_DB_TYPE_LIVENODES 0
64 #define O2HB_DB_TYPE_LIVEREGIONS 1
65 #define O2HB_DB_TYPE_QUORUMREGIONS 2
66 #define O2HB_DB_TYPE_FAILEDREGIONS 3
67 #define O2HB_DB_TYPE_REGION_LIVENODES 4
68 #define O2HB_DB_TYPE_REGION_NUMBER 5
69 #define O2HB_DB_TYPE_REGION_ELAPSED_TIME 6
70 #define O2HB_DB_TYPE_REGION_PINNED 7
71 struct o2hb_debug_buf {
72 int db_type;
73 int db_size;
74 int db_len;
75 void *db_data;
76 };
77
78 static struct o2hb_debug_buf *o2hb_db_livenodes;
79 static struct o2hb_debug_buf *o2hb_db_liveregions;
80 static struct o2hb_debug_buf *o2hb_db_quorumregions;
81 static struct o2hb_debug_buf *o2hb_db_failedregions;
82
83 #define O2HB_DEBUG_DIR "o2hb"
84 #define O2HB_DEBUG_LIVENODES "livenodes"
85 #define O2HB_DEBUG_LIVEREGIONS "live_regions"
86 #define O2HB_DEBUG_QUORUMREGIONS "quorum_regions"
87 #define O2HB_DEBUG_FAILEDREGIONS "failed_regions"
88 #define O2HB_DEBUG_REGION_NUMBER "num"
89 #define O2HB_DEBUG_REGION_ELAPSED_TIME "elapsed_time_in_ms"
90 #define O2HB_DEBUG_REGION_PINNED "pinned"
91
92 static struct dentry *o2hb_debug_dir;
93
94 static LIST_HEAD(o2hb_all_regions);
95
96 static struct o2hb_callback {
97 struct list_head list;
98 } o2hb_callbacks[O2HB_NUM_CB];
99
100 static struct o2hb_callback *hbcall_from_type(enum o2hb_callback_type type);
101
102 enum o2hb_heartbeat_modes {
103 O2HB_HEARTBEAT_LOCAL = 0,
104 O2HB_HEARTBEAT_GLOBAL,
105 O2HB_HEARTBEAT_NUM_MODES,
106 };
107
108 static const char *o2hb_heartbeat_mode_desc[O2HB_HEARTBEAT_NUM_MODES] = {
109 "local", /* O2HB_HEARTBEAT_LOCAL */
110 "global", /* O2HB_HEARTBEAT_GLOBAL */
111 };
112
113 unsigned int o2hb_dead_threshold = O2HB_DEFAULT_DEAD_THRESHOLD;
114 static unsigned int o2hb_heartbeat_mode = O2HB_HEARTBEAT_LOCAL;
115
116 /*
117 * o2hb_dependent_users tracks the number of registered callbacks that depend
118 * on heartbeat. o2net and o2dlm are two entities that register this callback.
119 * However only o2dlm depends on the heartbeat. It does not want the heartbeat
120 * to stop while a dlm domain is still active.
121 */
122 static unsigned int o2hb_dependent_users;
123
124 /*
125 * In global heartbeat mode, all regions are pinned if there are one or more
126 * dependent users and the quorum region count is <= O2HB_PIN_CUT_OFF. All
127 * regions are unpinned if the region count exceeds the cut off or the number
128 * of dependent users falls to zero.
129 */
130 #define O2HB_PIN_CUT_OFF 3
131
132 /*
133 * In local heartbeat mode, we assume the dlm domain name to be the same as
134 * region uuid. This is true for domains created for the file system but not
135 * necessarily true for userdlm domains. This is a known limitation.
136 *
137 * In global heartbeat mode, we pin/unpin all o2hb regions. This solution
138 * works for both file system and userdlm domains.
139 */
140 static int o2hb_region_pin(const char *region_uuid);
141 static void o2hb_region_unpin(const char *region_uuid);
142
143 /* Only sets a new threshold if there are no active regions.
144 *
145 * No locking or otherwise interesting code is required for reading
146 * o2hb_dead_threshold as it can't change once regions are active and
147 * it's not interesting to anyone until then anyway. */
148 static void o2hb_dead_threshold_set(unsigned int threshold)
149 {
150 if (threshold > O2HB_MIN_DEAD_THRESHOLD) {
151 spin_lock(&o2hb_live_lock);
152 if (list_empty(&o2hb_all_regions))
153 o2hb_dead_threshold = threshold;
154 spin_unlock(&o2hb_live_lock);
155 }
156 }
157
158 static int o2hb_global_heartbeat_mode_set(unsigned int hb_mode)
159 {
160 int ret = -1;
161
162 if (hb_mode < O2HB_HEARTBEAT_NUM_MODES) {
163 spin_lock(&o2hb_live_lock);
164 if (list_empty(&o2hb_all_regions)) {
165 o2hb_heartbeat_mode = hb_mode;
166 ret = 0;
167 }
168 spin_unlock(&o2hb_live_lock);
169 }
170
171 return ret;
172 }
173
174 struct o2hb_node_event {
175 struct list_head hn_item;
176 enum o2hb_callback_type hn_event_type;
177 struct o2nm_node *hn_node;
178 int hn_node_num;
179 };
180
181 struct o2hb_disk_slot {
182 struct o2hb_disk_heartbeat_block *ds_raw_block;
183 u8 ds_node_num;
184 u64 ds_last_time;
185 u64 ds_last_generation;
186 u16 ds_equal_samples;
187 u16 ds_changed_samples;
188 struct list_head ds_live_item;
189 };
190
191 /* each thread owns a region.. when we're asked to tear down the region
192 * we ask the thread to stop, who cleans up the region */
193 struct o2hb_region {
194 struct config_item hr_item;
195
196 struct list_head hr_all_item;
197 unsigned hr_unclean_stop:1,
198 hr_aborted_start:1,
199 hr_item_pinned:1,
200 hr_item_dropped:1,
201 hr_node_deleted:1;
202
203 /* protected by the hr_callback_sem */
204 struct task_struct *hr_task;
205
206 unsigned int hr_blocks;
207 unsigned long long hr_start_block;
208
209 unsigned int hr_block_bits;
210 unsigned int hr_block_bytes;
211
212 unsigned int hr_slots_per_page;
213 unsigned int hr_num_pages;
214
215 struct page **hr_slot_data;
216 struct file *hr_bdev_file;
217 struct o2hb_disk_slot *hr_slots;
218
219 /* live node map of this region */
220 unsigned long hr_live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
221 unsigned int hr_region_num;
222
223 struct dentry *hr_debug_dir;
224 struct o2hb_debug_buf *hr_db_livenodes;
225 struct o2hb_debug_buf *hr_db_regnum;
226 struct o2hb_debug_buf *hr_db_elapsed_time;
227 struct o2hb_debug_buf *hr_db_pinned;
228
229 /* let the person setting up hb wait for it to return until it
230 * has reached a 'steady' state. This will be fixed when we have
231 * a more complete api that doesn't lead to this sort of fragility. */
232 atomic_t hr_steady_iterations;
233
234 /* terminate o2hb thread if it does not reach steady state
235 * (hr_steady_iterations == 0) within hr_unsteady_iterations */
236 atomic_t hr_unsteady_iterations;
237
238 unsigned int hr_timeout_ms;
239
240 /* randomized as the region goes up and down so that a node
241 * recognizes a node going up and down in one iteration */
242 u64 hr_generation;
243
244 struct delayed_work hr_write_timeout_work;
245 unsigned long hr_last_timeout_start;
246
247 /* negotiate timer, used to negotiate extending hb timeout. */
248 struct delayed_work hr_nego_timeout_work;
249 unsigned long hr_nego_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
250
251 /* Used during o2hb_check_slot to hold a copy of the block
252 * being checked because we temporarily have to zero out the
253 * crc field. */
254 struct o2hb_disk_heartbeat_block *hr_tmp_block;
255
256 /* Message key for negotiate timeout message. */
257 unsigned int hr_key;
258 struct list_head hr_handler_list;
259
260 /* last hb status, 0 for success, other value for error. */
261 int hr_last_hb_status;
262 };
263
264 static inline struct block_device *reg_bdev(struct o2hb_region *reg)
265 {
266 return reg->hr_bdev_file ? file_bdev(reg->hr_bdev_file) : NULL;
267 }
268
269 struct o2hb_bio_wait_ctxt {
270 atomic_t wc_num_reqs;
271 struct completion wc_io_complete;
272 int wc_error;
273 };
274
275 #define O2HB_NEGO_TIMEOUT_MS (O2HB_MAX_WRITE_TIMEOUT_MS/2)
276
277 enum {
278 O2HB_NEGO_TIMEOUT_MSG = 1,
279 O2HB_NEGO_APPROVE_MSG = 2,
280 };
281
282 struct o2hb_nego_msg {
283 u8 node_num;
284 };
285
286 static void o2hb_write_timeout(struct work_struct *work)
287 {
288 int failed, quorum;
289 struct o2hb_region *reg =
290 container_of(work, struct o2hb_region,
291 hr_write_timeout_work.work);
292
293 mlog(ML_ERROR, "Heartbeat write timeout to device %pg after %u "
294 "milliseconds\n", reg_bdev(reg),
295 jiffies_to_msecs(jiffies - reg->hr_last_timeout_start));
296
297 if (o2hb_global_heartbeat_active()) {
298 spin_lock(&o2hb_live_lock);
299 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap))
300 set_bit(reg->hr_region_num, o2hb_failed_region_bitmap);
301 failed = bitmap_weight(o2hb_failed_region_bitmap,
302 O2NM_MAX_REGIONS);
303 quorum = bitmap_weight(o2hb_quorum_region_bitmap,
304 O2NM_MAX_REGIONS);
305 spin_unlock(&o2hb_live_lock);
306
307 mlog(ML_HEARTBEAT, "Number of regions %d, failed regions %d\n",
308 quorum, failed);
309
310 /*
311 * Fence if the number of failed regions >= half the number
312 * of quorum regions
313 */
314 if ((failed << 1) < quorum)
315 return;
316 }
317
318 o2quo_disk_timeout();
319 }
320
321 static void o2hb_arm_timeout(struct o2hb_region *reg)
322 {
323 /* Arm writeout only after thread reaches steady state */
324 if (atomic_read(®->hr_steady_iterations) != 0)
325 return;
326
327 mlog(ML_HEARTBEAT, "Queue write timeout for %u ms\n",
328 O2HB_MAX_WRITE_TIMEOUT_MS);
329
330 if (o2hb_global_heartbeat_active()) {
331 spin_lock(&o2hb_live_lock);
332 clear_bit(reg->hr_region_num, o2hb_failed_region_bitmap);
333 spin_unlock(&o2hb_live_lock);
334 }
335 cancel_delayed_work(®->hr_write_timeout_work);
336 schedule_delayed_work(®->hr_write_timeout_work,
337 msecs_to_jiffies(O2HB_MAX_WRITE_TIMEOUT_MS));
338
339 cancel_delayed_work(®->hr_nego_timeout_work);
340 /* negotiate timeout must be less than write timeout. */
341 schedule_delayed_work(®->hr_nego_timeout_work,
342 msecs_to_jiffies(O2HB_NEGO_TIMEOUT_MS));
343 bitmap_zero(reg->hr_nego_node_bitmap, O2NM_MAX_NODES);
344 }
345
346 static void o2hb_disarm_timeout(struct o2hb_region *reg)
347 {
348 cancel_delayed_work_sync(®->hr_write_timeout_work);
349 cancel_delayed_work_sync(®->hr_nego_timeout_work);
350 }
351
352 static int o2hb_send_nego_msg(int key, int type, u8 target)
353 {
354 struct o2hb_nego_msg msg;
355 int status, ret;
356
357 msg.node_num = o2nm_this_node();
358 again:
359 ret = o2net_send_message(type, key, &msg, sizeof(msg),
360 target, &status);
361
362 if (ret == -EAGAIN || ret == -ENOMEM) {
363 msleep(100);
364 goto again;
365 }
366
367 return ret;
368 }
369
370 static void o2hb_nego_timeout(struct work_struct *work)
371 {
372 unsigned long live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
373 int master_node, i, ret;
374 struct o2hb_region *reg;
375
376 reg = container_of(work, struct o2hb_region, hr_nego_timeout_work.work);
377 /* don't negotiate timeout if last hb failed since it is very
378 * possible io failed. Should let write timeout fence self.
379 */
380 if (reg->hr_last_hb_status)
381 return;
382
383 o2hb_fill_node_map(live_node_bitmap, O2NM_MAX_NODES);
384 /* lowest node as master node to make negotiate decision. */
385 master_node = find_first_bit(live_node_bitmap, O2NM_MAX_NODES);
386
387 if (master_node == o2nm_this_node()) {
388 if (!test_bit(master_node, reg->hr_nego_node_bitmap)) {
389 printk(KERN_NOTICE "o2hb: node %d hb write hung for %ds on region %s (%pg).\n",
390 o2nm_this_node(), O2HB_NEGO_TIMEOUT_MS/1000,
391 config_item_name(®->hr_item), reg_bdev(reg));
392 set_bit(master_node, reg->hr_nego_node_bitmap);
393 }
394 if (!bitmap_equal(reg->hr_nego_node_bitmap, live_node_bitmap,
395 O2NM_MAX_NODES)) {
396 /* check negotiate bitmap every second to do timeout
397 * approve decision.
398 */
399 schedule_delayed_work(®->hr_nego_timeout_work,
400 msecs_to_jiffies(1000));
401
402 return;
403 }
404
405 printk(KERN_NOTICE "o2hb: all nodes hb write hung, maybe region %s (%pg) is down.\n",
406 config_item_name(®->hr_item),
407 reg_bdev(reg));
408 /* approve negotiate timeout request. */
409 o2hb_arm_timeout(reg);
410
411 i = -1;
412 while ((i = find_next_bit(live_node_bitmap,
413 O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) {
414 if (i == master_node)
415 continue;
416
417 mlog(ML_HEARTBEAT, "send NEGO_APPROVE msg to node %d\n", i);
418 ret = o2hb_send_nego_msg(reg->hr_key,
419 O2HB_NEGO_APPROVE_MSG, i);
420 if (ret)
421 mlog(ML_ERROR, "send NEGO_APPROVE msg to node %d fail %d\n",
422 i, ret);
423 }
424 } else {
425 /* negotiate timeout with master node. */
426 printk(KERN_NOTICE "o2hb: node %d hb write hung for %ds on region %s (%pg), negotiate timeout with node %d.\n",
427 o2nm_this_node(), O2HB_NEGO_TIMEOUT_MS/1000, config_item_name(®->hr_item),
428 reg_bdev(reg), master_node);
429 ret = o2hb_send_nego_msg(reg->hr_key, O2HB_NEGO_TIMEOUT_MSG,
430 master_node);
431 if (ret)
432 mlog(ML_ERROR, "send NEGO_TIMEOUT msg to node %d fail %d\n",
433 master_node, ret);
434 }
435 }
436
437 static int o2hb_nego_timeout_handler(struct o2net_msg *msg, u32 len, void *data,
438 void **ret_data)
439 {
440 struct o2hb_region *reg = data;
441 struct o2hb_nego_msg *nego_msg;
442
443 nego_msg = (struct o2hb_nego_msg *)msg->buf;
444 printk(KERN_NOTICE "o2hb: receive negotiate timeout message from node %d on region %s (%pg).\n",
445 nego_msg->node_num, config_item_name(®->hr_item),
446 reg_bdev(reg));
447 if (nego_msg->node_num < O2NM_MAX_NODES)
448 set_bit(nego_msg->node_num, reg->hr_nego_node_bitmap);
449 else
450 mlog(ML_ERROR, "got nego timeout message from bad node.\n");
451
452 return 0;
453 }
454
455 static int o2hb_nego_approve_handler(struct o2net_msg *msg, u32 len, void *data,
456 void **ret_data)
457 {
458 struct o2hb_region *reg = data;
459
460 printk(KERN_NOTICE "o2hb: negotiate timeout approved by master node on region %s (%pg).\n",
461 config_item_name(®->hr_item), reg_bdev(reg));
462 o2hb_arm_timeout(reg);
463 return 0;
464 }
465
466 static inline void o2hb_bio_wait_init(struct o2hb_bio_wait_ctxt *wc)
467 {
468 atomic_set(&wc->wc_num_reqs, 1);
469 init_completion(&wc->wc_io_complete);
470 wc->wc_error = 0;
471 }
472
473 /* Used in error paths too */
474 static inline void o2hb_bio_wait_dec(struct o2hb_bio_wait_ctxt *wc,
475 unsigned int num)
476 {
477 /* sadly atomic_sub_and_test() isn't available on all platforms. The
478 * good news is that the fast path only completes one at a time */
479 while(num--) {
480 if (atomic_dec_and_test(&wc->wc_num_reqs)) {
481 BUG_ON(num > 0);
482 complete(&wc->wc_io_complete);
483 }
484 }
485 }
486
487 static void o2hb_wait_on_io(struct o2hb_bio_wait_ctxt *wc)
488 {
489 o2hb_bio_wait_dec(wc, 1);
490 wait_for_completion(&wc->wc_io_complete);
491 }
492
493 static void o2hb_bio_end_io(struct bio *bio)
494 {
495 struct o2hb_bio_wait_ctxt *wc = bio->bi_private;
496
497 if (bio->bi_status) {
498 mlog(ML_ERROR, "IO Error %d\n", bio->bi_status);
499 wc->wc_error = blk_status_to_errno(bio->bi_status);
500 }
501
502 o2hb_bio_wait_dec(wc, 1);
503 bio_put(bio);
504 }
505
506 /* Setup a Bio to cover I/O against num_slots slots starting at
507 * start_slot. */
508 static struct bio *o2hb_setup_one_bio(struct o2hb_region *reg,
509 struct o2hb_bio_wait_ctxt *wc,
510 unsigned int *current_slot,
511 unsigned int max_slots, blk_opf_t opf)
512 {
513 int len, current_page;
514 unsigned int vec_len, vec_start;
515 unsigned int bits = reg->hr_block_bits;
516 unsigned int spp = reg->hr_slots_per_page;
517 unsigned int cs = *current_slot;
518 struct bio *bio;
519 struct page *page;
520
521 /* Testing has shown this allocation to take long enough under
522 * GFP_KERNEL that the local node can get fenced. It would be
523 * nicest if we could pre-allocate these bios and avoid this
524 * all together. */
525 bio = bio_alloc(reg_bdev(reg), 16, opf, GFP_ATOMIC);
526 if (!bio) {
527 mlog(ML_ERROR, "Could not alloc slots BIO!\n");
528 bio = ERR_PTR(-ENOMEM);
529 goto bail;
530 }
531
532 /* Must put everything in 512 byte sectors for the bio... */
533 bio->bi_iter.bi_sector = (reg->hr_start_block + cs) << (bits - 9);
534 bio->bi_private = wc;
535 bio->bi_end_io = o2hb_bio_end_io;
536
537 vec_start = (cs << bits) % PAGE_SIZE;
538 while(cs < max_slots) {
539 current_page = cs / spp;
540 page = reg->hr_slot_data[current_page];
541
542 vec_len = min(PAGE_SIZE - vec_start,
543 (max_slots-cs) * (PAGE_SIZE/spp) );
544
545 mlog(ML_HB_BIO, "page %d, vec_len = %u, vec_start = %u\n",
546 current_page, vec_len, vec_start);
547
548 len = bio_add_page(bio, page, vec_len, vec_start);
549 if (len != vec_len) break;
550
551 cs += vec_len / (PAGE_SIZE/spp);
552 vec_start = 0;
553 }
554
555 bail:
556 *current_slot = cs;
557 return bio;
558 }
559
560 static int o2hb_read_slots(struct o2hb_region *reg,
561 unsigned int begin_slot,
562 unsigned int max_slots)
563 {
564 unsigned int current_slot = begin_slot;
565 int status;
566 struct o2hb_bio_wait_ctxt wc;
567 struct bio *bio;
568
569 o2hb_bio_wait_init(&wc);
570
571 while(current_slot < max_slots) {
572 bio = o2hb_setup_one_bio(reg, &wc, ¤t_slot, max_slots,
573 REQ_OP_READ);
574 if (IS_ERR(bio)) {
575 status = PTR_ERR(bio);
576 mlog_errno(status);
577 goto bail_and_wait;
578 }
579
580 atomic_inc(&wc.wc_num_reqs);
581 submit_bio(bio);
582 }
583
584 status = 0;
585
586 bail_and_wait:
587 o2hb_wait_on_io(&wc);
588 if (wc.wc_error && !status)
589 status = wc.wc_error;
590
591 return status;
592 }
593
594 static int o2hb_issue_node_write(struct o2hb_region *reg,
595 struct o2hb_bio_wait_ctxt *write_wc)
596 {
597 int status;
598 unsigned int slot;
599 struct bio *bio;
600
601 o2hb_bio_wait_init(write_wc);
602
603 slot = o2nm_this_node();
604
605 bio = o2hb_setup_one_bio(reg, write_wc, &slot, slot+1,
606 REQ_OP_WRITE | REQ_SYNC);
607 if (IS_ERR(bio)) {
608 status = PTR_ERR(bio);
609 mlog_errno(status);
610 goto bail;
611 }
612
613 atomic_inc(&write_wc->wc_num_reqs);
614 submit_bio(bio);
615
616 status = 0;
617 bail:
618 return status;
619 }
620
621 static u32 o2hb_compute_block_crc_le(struct o2hb_region *reg,
622 struct o2hb_disk_heartbeat_block *hb_block)
623 {
624 __le32 old_cksum;
625 u32 ret;
626
627 /* We want to compute the block crc with a 0 value in the
628 * hb_cksum field. Save it off here and replace after the
629 * crc. */
630 old_cksum = hb_block->hb_cksum;
631 hb_block->hb_cksum = 0;
632
633 ret = crc32_le(0, (unsigned char *) hb_block, reg->hr_block_bytes);
634
635 hb_block->hb_cksum = old_cksum;
636
637 return ret;
638 }
639
640 static void o2hb_dump_slot(struct o2hb_disk_heartbeat_block *hb_block)
641 {
642 mlog(ML_ERROR, "Dump slot information: seq = 0x%llx, node = %u, "
643 "cksum = 0x%x, generation 0x%llx\n",
644 (long long)le64_to_cpu(hb_block->hb_seq),
645 hb_block->hb_node, le32_to_cpu(hb_block->hb_cksum),
646 (long long)le64_to_cpu(hb_block->hb_generation));
647 }
648
649 static int o2hb_verify_crc(struct o2hb_region *reg,
650 struct o2hb_disk_heartbeat_block *hb_block)
651 {
652 u32 read, computed;
653
654 read = le32_to_cpu(hb_block->hb_cksum);
655 computed = o2hb_compute_block_crc_le(reg, hb_block);
656
657 return read == computed;
658 }
659
660 /*
661 * Compare the slot data with what we wrote in the last iteration.
662 * If the match fails, print an appropriate error message. This is to
663 * detect errors like... another node hearting on the same slot,
664 * flaky device that is losing writes, etc.
665 * Returns 1 if check succeeds, 0 otherwise.
666 */
667 static int o2hb_check_own_slot(struct o2hb_region *reg)
668 {
669 struct o2hb_disk_slot *slot;
670 struct o2hb_disk_heartbeat_block *hb_block;
671 char *errstr;
672
673 slot = ®->hr_slots[o2nm_this_node()];
674 /* Don't check on our 1st timestamp */
675 if (!slot->ds_last_time)
676 return 0;
677
678 hb_block = slot->ds_raw_block;
679 if (le64_to_cpu(hb_block->hb_seq) == slot->ds_last_time &&
680 le64_to_cpu(hb_block->hb_generation) == slot->ds_last_generation &&
681 hb_block->hb_node == slot->ds_node_num)
682 return 1;
683
684 #define ERRSTR1 "Another node is heartbeating on device"
685 #define ERRSTR2 "Heartbeat generation mismatch on device"
686 #define ERRSTR3 "Heartbeat sequence mismatch on device"
687
688 if (hb_block->hb_node != slot->ds_node_num)
689 errstr = ERRSTR1;
690 else if (le64_to_cpu(hb_block->hb_generation) !=
691 slot->ds_last_generation)
692 errstr = ERRSTR2;
693 else
694 errstr = ERRSTR3;
695
696 mlog(ML_ERROR, "%s (%pg): expected(%u:0x%llx, 0x%llx), "
697 "ondisk(%u:0x%llx, 0x%llx)\n", errstr, reg_bdev(reg),
698 slot->ds_node_num, (unsigned long long)slot->ds_last_generation,
699 (unsigned long long)slot->ds_last_time, hb_block->hb_node,
700 (unsigned long long)le64_to_cpu(hb_block->hb_generation),
701 (unsigned long long)le64_to_cpu(hb_block->hb_seq));
702
703 return 0;
704 }
705
706 static inline void o2hb_prepare_block(struct o2hb_region *reg,
707 u64 generation)
708 {
709 int node_num;
710 u64 cputime;
711 struct o2hb_disk_slot *slot;
712 struct o2hb_disk_heartbeat_block *hb_block;
713
714 node_num = o2nm_this_node();
715 slot = ®->hr_slots[node_num];
716
717 hb_block = (struct o2hb_disk_heartbeat_block *)slot->ds_raw_block;
718 memset(hb_block, 0, reg->hr_block_bytes);
719 /* TODO: time stuff */
720 cputime = ktime_get_real_seconds();
721 if (!cputime)
722 cputime = 1;
723
724 hb_block->hb_seq = cpu_to_le64(cputime);
725 hb_block->hb_node = node_num;
726 hb_block->hb_generation = cpu_to_le64(generation);
727 hb_block->hb_dead_ms = cpu_to_le32(o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS);
728
729 /* This step must always happen last! */
730 hb_block->hb_cksum = cpu_to_le32(o2hb_compute_block_crc_le(reg,
731 hb_block));
732
733 mlog(ML_HB_BIO, "our node generation = 0x%llx, cksum = 0x%x\n",
734 (long long)generation,
735 le32_to_cpu(hb_block->hb_cksum));
736 }
737
738 static void o2hb_fire_callbacks(struct o2hb_callback *hbcall,
739 struct o2nm_node *node,
740 int idx)
741 {
742 struct o2hb_callback_func *f;
743
744 list_for_each_entry(f, &hbcall->list, hc_item) {
745 mlog(ML_HEARTBEAT, "calling funcs %p\n", f);
746 (f->hc_func)(node, idx, f->hc_data);
747 }
748 }
749
750 /* Will run the list in order until we process the passed event */
751 static void o2hb_run_event_list(struct o2hb_node_event *queued_event)
752 {
753 struct o2hb_callback *hbcall;
754 struct o2hb_node_event *event;
755
756 /* Holding callback sem assures we don't alter the callback
757 * lists when doing this, and serializes ourselves with other
758 * processes wanting callbacks. */
759 down_write(&o2hb_callback_sem);
760
761 spin_lock(&o2hb_live_lock);
762 while (!list_empty(&o2hb_node_events)
763 && !list_empty(&queued_event->hn_item)) {
764 event = list_entry(o2hb_node_events.next,
765 struct o2hb_node_event,
766 hn_item);
767 list_del_init(&event->hn_item);
768 spin_unlock(&o2hb_live_lock);
769
770 mlog(ML_HEARTBEAT, "Node %s event for %d\n",
771 event->hn_event_type == O2HB_NODE_UP_CB ? "UP" : "DOWN",
772 event->hn_node_num);
773
774 hbcall = hbcall_from_type(event->hn_event_type);
775
776 /* We should *never* have gotten on to the list with a
777 * bad type... This isn't something that we should try
778 * to recover from. */
779 BUG_ON(IS_ERR(hbcall));
780
781 o2hb_fire_callbacks(hbcall, event->hn_node, event->hn_node_num);
782
783 spin_lock(&o2hb_live_lock);
784 }
785 spin_unlock(&o2hb_live_lock);
786
787 up_write(&o2hb_callback_sem);
788 }
789
790 static void o2hb_queue_node_event(struct o2hb_node_event *event,
791 enum o2hb_callback_type type,
792 struct o2nm_node *node,
793 int node_num)
794 {
795 assert_spin_locked(&o2hb_live_lock);
796
797 BUG_ON((!node) && (type != O2HB_NODE_DOWN_CB));
798
799 event->hn_event_type = type;
800 event->hn_node = node;
801 event->hn_node_num = node_num;
802
803 mlog(ML_HEARTBEAT, "Queue node %s event for node %d\n",
804 type == O2HB_NODE_UP_CB ? "UP" : "DOWN", node_num);
805
806 list_add_tail(&event->hn_item, &o2hb_node_events);
807 }
808
809 static void o2hb_shutdown_slot(struct o2hb_disk_slot *slot)
810 {
811 struct o2hb_node_event event =
812 { .hn_item = LIST_HEAD_INIT(event.hn_item), };
813 struct o2nm_node *node;
814 int queued = 0;
815
816 node = o2nm_get_node_by_num(slot->ds_node_num);
817 if (!node)
818 return;
819
820 spin_lock(&o2hb_live_lock);
821 if (!list_empty(&slot->ds_live_item)) {
822 mlog(ML_HEARTBEAT, "Shutdown, node %d leaves region\n",
823 slot->ds_node_num);
824
825 list_del_init(&slot->ds_live_item);
826
827 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
828 clear_bit(slot->ds_node_num, o2hb_live_node_bitmap);
829
830 o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB, node,
831 slot->ds_node_num);
832 queued = 1;
833 }
834 }
835 spin_unlock(&o2hb_live_lock);
836
837 if (queued)
838 o2hb_run_event_list(&event);
839
840 o2nm_node_put(node);
841 }
842
843 static void o2hb_set_quorum_device(struct o2hb_region *reg)
844 {
845 if (!o2hb_global_heartbeat_active())
846 return;
847
848 /* Prevent race with o2hb_heartbeat_group_drop_item() */
849 if (kthread_should_stop())
850 return;
851
852 /* Tag region as quorum only after thread reaches steady state */
853 if (atomic_read(®->hr_steady_iterations) != 0)
854 return;
855
856 spin_lock(&o2hb_live_lock);
857
858 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap))
859 goto unlock;
860
861 /*
862 * A region can be added to the quorum only when it sees all
863 * live nodes heartbeat on it. In other words, the region has been
864 * added to all nodes.
865 */
866 if (!bitmap_equal(reg->hr_live_node_bitmap, o2hb_live_node_bitmap,
867 O2NM_MAX_NODES))
868 goto unlock;
869
870 printk(KERN_NOTICE "o2hb: Region %s (%pg) is now a quorum device\n",
871 config_item_name(®->hr_item), reg_bdev(reg));
872
873 set_bit(reg->hr_region_num, o2hb_quorum_region_bitmap);
874
875 /*
876 * If global heartbeat active, unpin all regions if the
877 * region count > CUT_OFF
878 */
879 if (bitmap_weight(o2hb_quorum_region_bitmap,
880 O2NM_MAX_REGIONS) > O2HB_PIN_CUT_OFF)
881 o2hb_region_unpin(NULL);
882 unlock:
883 spin_unlock(&o2hb_live_lock);
884 }
885
886 static int o2hb_check_slot(struct o2hb_region *reg,
887 struct o2hb_disk_slot *slot)
888 {
889 int changed = 0, gen_changed = 0;
890 struct o2hb_node_event event =
891 { .hn_item = LIST_HEAD_INIT(event.hn_item), };
892 struct o2nm_node *node;
893 struct o2hb_disk_heartbeat_block *hb_block = reg->hr_tmp_block;
894 u64 cputime;
895 unsigned int dead_ms = o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS;
896 unsigned int slot_dead_ms;
897 int tmp;
898 int queued = 0;
899
900 memcpy(hb_block, slot->ds_raw_block, reg->hr_block_bytes);
901
902 /*
903 * If a node is no longer configured but is still in the livemap, we
904 * may need to clear that bit from the livemap.
905 */
906 node = o2nm_get_node_by_num(slot->ds_node_num);
907 if (!node) {
908 spin_lock(&o2hb_live_lock);
909 tmp = test_bit(slot->ds_node_num, o2hb_live_node_bitmap);
910 spin_unlock(&o2hb_live_lock);
911 if (!tmp)
912 return 0;
913 }
914
915 if (!o2hb_verify_crc(reg, hb_block)) {
916 /* all paths from here will drop o2hb_live_lock for
917 * us. */
918 spin_lock(&o2hb_live_lock);
919
920 /* Don't print an error on the console in this case -
921 * a freshly formatted heartbeat area will not have a
922 * crc set on it. */
923 if (list_empty(&slot->ds_live_item))
924 goto out;
925
926 /* The node is live but pushed out a bad crc. We
927 * consider it a transient miss but don't populate any
928 * other values as they may be junk. */
929 mlog(ML_ERROR, "Node %d has written a bad crc to %pg\n",
930 slot->ds_node_num, reg_bdev(reg));
931 o2hb_dump_slot(hb_block);
932
933 slot->ds_equal_samples++;
934 goto fire_callbacks;
935 }
936
937 /* we don't care if these wrap.. the state transitions below
938 * clear at the right places */
939 cputime = le64_to_cpu(hb_block->hb_seq);
940 if (slot->ds_last_time != cputime)
941 slot->ds_changed_samples++;
942 else
943 slot->ds_equal_samples++;
944 slot->ds_last_time = cputime;
945
946 /* The node changed heartbeat generations. We assume this to
947 * mean it dropped off but came back before we timed out. We
948 * want to consider it down for the time being but don't want
949 * to lose any changed_samples state we might build up to
950 * considering it live again. */
951 if (slot->ds_last_generation != le64_to_cpu(hb_block->hb_generation)) {
952 gen_changed = 1;
953 slot->ds_equal_samples = 0;
954 mlog(ML_HEARTBEAT, "Node %d changed generation (0x%llx "
955 "to 0x%llx)\n", slot->ds_node_num,
956 (long long)slot->ds_last_generation,
957 (long long)le64_to_cpu(hb_block->hb_generation));
958 }
959
960 slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation);
961
962 mlog(ML_HEARTBEAT, "Slot %d gen 0x%llx cksum 0x%x "
963 "seq %llu last %llu changed %u equal %u\n",
964 slot->ds_node_num, (long long)slot->ds_last_generation,
965 le32_to_cpu(hb_block->hb_cksum),
966 (unsigned long long)le64_to_cpu(hb_block->hb_seq),
967 (unsigned long long)slot->ds_last_time, slot->ds_changed_samples,
968 slot->ds_equal_samples);
969
970 spin_lock(&o2hb_live_lock);
971
972 fire_callbacks:
973 /* dead nodes only come to life after some number of
974 * changes at any time during their dead time */
975 if (list_empty(&slot->ds_live_item) &&
976 slot->ds_changed_samples >= O2HB_LIVE_THRESHOLD) {
977 mlog(ML_HEARTBEAT, "Node %d (id 0x%llx) joined my region\n",
978 slot->ds_node_num, (long long)slot->ds_last_generation);
979
980 set_bit(slot->ds_node_num, reg->hr_live_node_bitmap);
981
982 /* first on the list generates a callback */
983 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
984 mlog(ML_HEARTBEAT, "o2hb: Add node %d to live nodes "
985 "bitmap\n", slot->ds_node_num);
986 set_bit(slot->ds_node_num, o2hb_live_node_bitmap);
987
988 o2hb_queue_node_event(&event, O2HB_NODE_UP_CB, node,
989 slot->ds_node_num);
990
991 changed = 1;
992 queued = 1;
993 }
994
995 list_add_tail(&slot->ds_live_item,
996 &o2hb_live_slots[slot->ds_node_num]);
997
998 slot->ds_equal_samples = 0;
999
1000 /* We want to be sure that all nodes agree on the
1001 * number of milliseconds before a node will be
1002 * considered dead. The self-fencing timeout is
1003 * computed from this value, and a discrepancy might
1004 * result in heartbeat calling a node dead when it
1005 * hasn't self-fenced yet. */
1006 slot_dead_ms = le32_to_cpu(hb_block->hb_dead_ms);
1007 if (slot_dead_ms && slot_dead_ms != dead_ms) {
1008 /* TODO: Perhaps we can fail the region here. */
1009 mlog(ML_ERROR, "Node %d on device %pg has a dead count "
1010 "of %u ms, but our count is %u ms.\n"
1011 "Please double check your configuration values "
1012 "for 'O2CB_HEARTBEAT_THRESHOLD'\n",
1013 slot->ds_node_num, reg_bdev(reg),
1014 slot_dead_ms, dead_ms);
1015 }
1016 goto out;
1017 }
1018
1019 /* if the list is dead, we're done.. */
1020 if (list_empty(&slot->ds_live_item))
1021 goto out;
1022
1023 /* live nodes only go dead after enough consequtive missed
1024 * samples.. reset the missed counter whenever we see
1025 * activity */
1026 if (slot->ds_equal_samples >= o2hb_dead_threshold || gen_changed) {
1027 mlog(ML_HEARTBEAT, "Node %d left my region\n",
1028 slot->ds_node_num);
1029
1030 clear_bit(slot->ds_node_num, reg->hr_live_node_bitmap);
1031
1032 /* last off the live_slot generates a callback */
1033 list_del_init(&slot->ds_live_item);
1034 if (list_empty(&o2hb_live_slots[slot->ds_node_num])) {
1035 mlog(ML_HEARTBEAT, "o2hb: Remove node %d from live "
1036 "nodes bitmap\n", slot->ds_node_num);
1037 clear_bit(slot->ds_node_num, o2hb_live_node_bitmap);
1038
1039 /* node can be null */
1040 o2hb_queue_node_event(&event, O2HB_NODE_DOWN_CB,
1041 node, slot->ds_node_num);
1042
1043 changed = 1;
1044 queued = 1;
1045 }
1046
1047 /* We don't clear this because the node is still
1048 * actually writing new blocks. */
1049 if (!gen_changed)
1050 slot->ds_changed_samples = 0;
1051 goto out;
1052 }
1053 if (slot->ds_changed_samples) {
1054 slot->ds_changed_samples = 0;
1055 slot->ds_equal_samples = 0;
1056 }
1057 out:
1058 spin_unlock(&o2hb_live_lock);
1059
1060 if (queued)
1061 o2hb_run_event_list(&event);
1062
1063 if (node)
1064 o2nm_node_put(node);
1065 return changed;
1066 }
1067
1068 static int o2hb_highest_node(unsigned long *nodes, int numbits)
1069 {
1070 return find_last_bit(nodes, numbits);
1071 }
1072
1073 static int o2hb_lowest_node(unsigned long *nodes, int numbits)
1074 {
1075 return find_first_bit(nodes, numbits);
1076 }
1077
1078 static int o2hb_do_disk_heartbeat(struct o2hb_region *reg)
1079 {
1080 int i, ret, highest_node, lowest_node;
1081 int membership_change = 0, own_slot_ok = 0;
1082 unsigned long configured_nodes[BITS_TO_LONGS(O2NM_MAX_NODES)];
1083 unsigned long live_node_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
1084 struct o2hb_bio_wait_ctxt write_wc;
1085
1086 ret = o2nm_configured_node_map(configured_nodes,
1087 sizeof(configured_nodes));
1088 if (ret) {
1089 mlog_errno(ret);
1090 goto bail;
1091 }
1092
1093 /*
1094 * If a node is not configured but is in the livemap, we still need
1095 * to read the slot so as to be able to remove it from the livemap.
1096 */
1097 o2hb_fill_node_map(live_node_bitmap, O2NM_MAX_NODES);
1098 i = -1;
1099 while ((i = find_next_bit(live_node_bitmap,
1100 O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) {
1101 set_bit(i, configured_nodes);
1102 }
1103
1104 highest_node = o2hb_highest_node(configured_nodes, O2NM_MAX_NODES);
1105 lowest_node = o2hb_lowest_node(configured_nodes, O2NM_MAX_NODES);
1106 if (highest_node >= O2NM_MAX_NODES || lowest_node >= O2NM_MAX_NODES) {
1107 mlog(ML_NOTICE, "o2hb: No configured nodes found!\n");
1108 ret = -EINVAL;
1109 goto bail;
1110 }
1111
1112 /* No sense in reading the slots of nodes that don't exist
1113 * yet. Of course, if the node definitions have holes in them
1114 * then we're reading an empty slot anyway... Consider this
1115 * best-effort. */
1116 ret = o2hb_read_slots(reg, lowest_node, highest_node + 1);
1117 if (ret < 0) {
1118 mlog_errno(ret);
1119 goto bail;
1120 }
1121
1122 /* With an up to date view of the slots, we can check that no
1123 * other node has been improperly configured to heartbeat in
1124 * our slot. */
1125 own_slot_ok = o2hb_check_own_slot(reg);
1126
1127 /* fill in the proper info for our next heartbeat */
1128 o2hb_prepare_block(reg, reg->hr_generation);
1129
1130 ret = o2hb_issue_node_write(reg, &write_wc);
1131 if (ret < 0) {
1132 mlog_errno(ret);
1133 goto bail;
1134 }
1135
1136 i = -1;
1137 while((i = find_next_bit(configured_nodes,
1138 O2NM_MAX_NODES, i + 1)) < O2NM_MAX_NODES) {
1139 membership_change |= o2hb_check_slot(reg, ®->hr_slots[i]);
1140 }
1141
1142 /*
1143 * We have to be sure we've advertised ourselves on disk
1144 * before we can go to steady state. This ensures that
1145 * people we find in our steady state have seen us.
1146 */
1147 o2hb_wait_on_io(&write_wc);
1148 if (write_wc.wc_error) {
1149 /* Do not re-arm the write timeout on I/O error - we
1150 * can't be sure that the new block ever made it to
1151 * disk */
1152 mlog(ML_ERROR, "Write error %d on device \"%pg\"\n",
1153 write_wc.wc_error, reg_bdev(reg));
1154 ret = write_wc.wc_error;
1155 goto bail;
1156 }
1157
1158 /* Skip disarming the timeout if own slot has stale/bad data */
1159 if (own_slot_ok) {
1160 o2hb_set_quorum_device(reg);
1161 o2hb_arm_timeout(reg);
1162 reg->hr_last_timeout_start = jiffies;
1163 }
1164
1165 bail:
1166 /* let the person who launched us know when things are steady */
1167 if (atomic_read(®->hr_steady_iterations) != 0) {
1168 if (!ret && own_slot_ok && !membership_change) {
1169 if (atomic_dec_and_test(®->hr_steady_iterations))
1170 wake_up(&o2hb_steady_queue);
1171 }
1172 }
1173
1174 if (atomic_read(®->hr_steady_iterations) != 0) {
1175 if (atomic_dec_and_test(®->hr_unsteady_iterations)) {
1176 printk(KERN_NOTICE "o2hb: Unable to stabilize "
1177 "heartbeat on region %s (%pg)\n",
1178 config_item_name(®->hr_item),
1179 reg_bdev(reg));
1180 atomic_set(®->hr_steady_iterations, 0);
1181 reg->hr_aborted_start = 1;
1182 wake_up(&o2hb_steady_queue);
1183 ret = -EIO;
1184 }
1185 }
1186
1187 return ret;
1188 }
1189
1190 /*
1191 * we ride the region ref that the region dir holds. before the region
1192 * dir is removed and drops it ref it will wait to tear down this
1193 * thread.
1194 */
1195 static int o2hb_thread(void *data)
1196 {
1197 int i, ret;
1198 struct o2hb_region *reg = data;
1199 struct o2hb_bio_wait_ctxt write_wc;
1200 ktime_t before_hb, after_hb;
1201 unsigned int elapsed_msec;
1202
1203 mlog(ML_HEARTBEAT|ML_KTHREAD, "hb thread running\n");
1204
1205 set_user_nice(current, MIN_NICE);
1206
1207 /* Pin node */
1208 ret = o2nm_depend_this_node();
1209 if (ret) {
1210 mlog(ML_ERROR, "Node has been deleted, ret = %d\n", ret);
1211 reg->hr_node_deleted = 1;
1212 wake_up(&o2hb_steady_queue);
1213 return 0;
1214 }
1215
1216 while (!kthread_should_stop() &&
1217 !reg->hr_unclean_stop && !reg->hr_aborted_start) {
1218 /* We track the time spent inside
1219 * o2hb_do_disk_heartbeat so that we avoid more than
1220 * hr_timeout_ms between disk writes. On busy systems
1221 * this should result in a heartbeat which is less
1222 * likely to time itself out. */
1223 before_hb = ktime_get_real();
1224
1225 ret = o2hb_do_disk_heartbeat(reg);
1226 reg->hr_last_hb_status = ret;
1227
1228 after_hb = ktime_get_real();
1229
1230 elapsed_msec = (unsigned int)
1231 ktime_ms_delta(after_hb, before_hb);
1232
1233 mlog(ML_HEARTBEAT,
1234 "start = %lld, end = %lld, msec = %u, ret = %d\n",
1235 before_hb, after_hb, elapsed_msec, ret);
1236
1237 if (!kthread_should_stop() &&
1238 elapsed_msec < reg->hr_timeout_ms) {
1239 /* the kthread api has blocked signals for us so no
1240 * need to record the return value. */
1241 msleep_interruptible(reg->hr_timeout_ms - elapsed_msec);
1242 }
1243 }
1244
1245 o2hb_disarm_timeout(reg);
1246
1247 /* unclean stop is only used in very bad situation */
1248 for(i = 0; !reg->hr_unclean_stop && i < reg->hr_blocks; i++)
1249 o2hb_shutdown_slot(®->hr_slots[i]);
1250
1251 /* Explicit down notification - avoid forcing the other nodes
1252 * to timeout on this region when we could just as easily
1253 * write a clear generation - thus indicating to them that
1254 * this node has left this region.
1255 */
1256 if (!reg->hr_unclean_stop && !reg->hr_aborted_start) {
1257 o2hb_prepare_block(reg, 0);
1258 ret = o2hb_issue_node_write(reg, &write_wc);
1259 if (ret == 0)
1260 o2hb_wait_on_io(&write_wc);
1261 else
1262 mlog_errno(ret);
1263 }
1264
1265 /* Unpin node */
1266 o2nm_undepend_this_node();
1267
1268 mlog(ML_HEARTBEAT|ML_KTHREAD, "o2hb thread exiting\n");
1269
1270 return 0;
1271 }
1272
1273 #ifdef CONFIG_DEBUG_FS
1274 static int o2hb_debug_open(struct inode *inode, struct file *file)
1275 {
1276 struct o2hb_debug_buf *db = inode->i_private;
1277 struct o2hb_region *reg;
1278 unsigned long map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1279 unsigned long lts;
1280 char *buf = NULL;
1281 int i = -1;
1282 int out = 0;
1283
1284 /* max_nodes should be the largest bitmap we pass here */
1285 BUG_ON(sizeof(map) < db->db_size);
1286
1287 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
1288 if (!buf)
1289 goto bail;
1290
1291 switch (db->db_type) {
1292 case O2HB_DB_TYPE_LIVENODES:
1293 case O2HB_DB_TYPE_LIVEREGIONS:
1294 case O2HB_DB_TYPE_QUORUMREGIONS:
1295 case O2HB_DB_TYPE_FAILEDREGIONS:
1296 spin_lock(&o2hb_live_lock);
1297 memcpy(map, db->db_data, db->db_size);
1298 spin_unlock(&o2hb_live_lock);
1299 break;
1300
1301 case O2HB_DB_TYPE_REGION_LIVENODES:
1302 spin_lock(&o2hb_live_lock);
1303 reg = (struct o2hb_region *)db->db_data;
1304 memcpy(map, reg->hr_live_node_bitmap, db->db_size);
1305 spin_unlock(&o2hb_live_lock);
1306 break;
1307
1308 case O2HB_DB_TYPE_REGION_NUMBER:
1309 reg = (struct o2hb_region *)db->db_data;
1310 out += scnprintf(buf + out, PAGE_SIZE - out, "%d\n",
1311 reg->hr_region_num);
1312 goto done;
1313
1314 case O2HB_DB_TYPE_REGION_ELAPSED_TIME:
1315 reg = (struct o2hb_region *)db->db_data;
1316 lts = reg->hr_last_timeout_start;
1317 /* If 0, it has never been set before */
1318 if (lts)
1319 lts = jiffies_to_msecs(jiffies - lts);
1320 out += scnprintf(buf + out, PAGE_SIZE - out, "%lu\n", lts);
1321 goto done;
1322
1323 case O2HB_DB_TYPE_REGION_PINNED:
1324 reg = (struct o2hb_region *)db->db_data;
1325 out += scnprintf(buf + out, PAGE_SIZE - out, "%u\n",
1326 !!reg->hr_item_pinned);
1327 goto done;
1328
1329 default:
1330 goto done;
1331 }
1332
1333 while ((i = find_next_bit(map, db->db_len, i + 1)) < db->db_len)
1334 out += scnprintf(buf + out, PAGE_SIZE - out, "%d ", i);
1335 out += scnprintf(buf + out, PAGE_SIZE - out, "\n");
1336
1337 done:
1338 i_size_write(inode, out);
1339
1340 file->private_data = buf;
1341
1342 return 0;
1343 bail:
1344 return -ENOMEM;
1345 }
1346
1347 static int o2hb_debug_release(struct inode *inode, struct file *file)
1348 {
1349 kfree(file->private_data);
1350 return 0;
1351 }
1352
1353 static ssize_t o2hb_debug_read(struct file *file, char __user *buf,
1354 size_t nbytes, loff_t *ppos)
1355 {
1356 return simple_read_from_buffer(buf, nbytes, ppos, file->private_data,
1357 i_size_read(file->f_mapping->host));
1358 }
1359 #else
1360 static int o2hb_debug_open(struct inode *inode, struct file *file)
1361 {
1362 return 0;
1363 }
1364 static int o2hb_debug_release(struct inode *inode, struct file *file)
1365 {
1366 return 0;
1367 }
1368 static ssize_t o2hb_debug_read(struct file *file, char __user *buf,
1369 size_t nbytes, loff_t *ppos)
1370 {
1371 return 0;
1372 }
1373 #endif /* CONFIG_DEBUG_FS */
1374
1375 static const struct file_operations o2hb_debug_fops = {
1376 .open = o2hb_debug_open,
1377 .release = o2hb_debug_release,
1378 .read = o2hb_debug_read,
1379 .llseek = generic_file_llseek,
1380 };
1381
1382 void o2hb_exit(void)
1383 {
1384 debugfs_remove_recursive(o2hb_debug_dir);
1385 kfree(o2hb_db_livenodes);
1386 kfree(o2hb_db_liveregions);
1387 kfree(o2hb_db_quorumregions);
1388 kfree(o2hb_db_failedregions);
1389 }
1390
1391 static void o2hb_debug_create(const char *name, struct dentry *dir,
1392 struct o2hb_debug_buf **db, int db_len, int type,
1393 int size, int len, void *data)
1394 {
1395 *db = kmalloc(db_len, GFP_KERNEL);
1396 if (!*db)
1397 return;
1398
1399 (*db)->db_type = type;
1400 (*db)->db_size = size;
1401 (*db)->db_len = len;
1402 (*db)->db_data = data;
1403
1404 debugfs_create_file(name, S_IFREG|S_IRUSR, dir, *db, &o2hb_debug_fops);
1405 }
1406
1407 static void o2hb_debug_init(void)
1408 {
1409 o2hb_debug_dir = debugfs_create_dir(O2HB_DEBUG_DIR, NULL);
1410
1411 o2hb_debug_create(O2HB_DEBUG_LIVENODES, o2hb_debug_dir,
1412 &o2hb_db_livenodes, sizeof(*o2hb_db_livenodes),
1413 O2HB_DB_TYPE_LIVENODES, sizeof(o2hb_live_node_bitmap),
1414 O2NM_MAX_NODES, o2hb_live_node_bitmap);
1415
1416 o2hb_debug_create(O2HB_DEBUG_LIVEREGIONS, o2hb_debug_dir,
1417 &o2hb_db_liveregions, sizeof(*o2hb_db_liveregions),
1418 O2HB_DB_TYPE_LIVEREGIONS,
1419 sizeof(o2hb_live_region_bitmap), O2NM_MAX_REGIONS,
1420 o2hb_live_region_bitmap);
1421
1422 o2hb_debug_create(O2HB_DEBUG_QUORUMREGIONS, o2hb_debug_dir,
1423 &o2hb_db_quorumregions,
1424 sizeof(*o2hb_db_quorumregions),
1425 O2HB_DB_TYPE_QUORUMREGIONS,
1426 sizeof(o2hb_quorum_region_bitmap), O2NM_MAX_REGIONS,
1427 o2hb_quorum_region_bitmap);
1428
1429 o2hb_debug_create(O2HB_DEBUG_FAILEDREGIONS, o2hb_debug_dir,
1430 &o2hb_db_failedregions,
1431 sizeof(*o2hb_db_failedregions),
1432 O2HB_DB_TYPE_FAILEDREGIONS,
1433 sizeof(o2hb_failed_region_bitmap), O2NM_MAX_REGIONS,
1434 o2hb_failed_region_bitmap);
1435 }
1436
1437 void o2hb_init(void)
1438 {
1439 int i;
1440
1441 for (i = 0; i < ARRAY_SIZE(o2hb_callbacks); i++)
1442 INIT_LIST_HEAD(&o2hb_callbacks[i].list);
1443
1444 for (i = 0; i < ARRAY_SIZE(o2hb_live_slots); i++)
1445 INIT_LIST_HEAD(&o2hb_live_slots[i]);
1446
1447 bitmap_zero(o2hb_live_node_bitmap, O2NM_MAX_NODES);
1448 bitmap_zero(o2hb_region_bitmap, O2NM_MAX_REGIONS);
1449 bitmap_zero(o2hb_live_region_bitmap, O2NM_MAX_REGIONS);
1450 bitmap_zero(o2hb_quorum_region_bitmap, O2NM_MAX_REGIONS);
1451 bitmap_zero(o2hb_failed_region_bitmap, O2NM_MAX_REGIONS);
1452
1453 o2hb_dependent_users = 0;
1454
1455 o2hb_debug_init();
1456 }
1457
1458 /* if we're already in a callback then we're already serialized by the sem */
1459 static void o2hb_fill_node_map_from_callback(unsigned long *map,
1460 unsigned int bits)
1461 {
1462 bitmap_copy(map, o2hb_live_node_bitmap, bits);
1463 }
1464
1465 /*
1466 * get a map of all nodes that are heartbeating in any regions
1467 */
1468 void o2hb_fill_node_map(unsigned long *map, unsigned int bits)
1469 {
1470 /* callers want to serialize this map and callbacks so that they
1471 * can trust that they don't miss nodes coming to the party */
1472 down_read(&o2hb_callback_sem);
1473 spin_lock(&o2hb_live_lock);
1474 o2hb_fill_node_map_from_callback(map, bits);
1475 spin_unlock(&o2hb_live_lock);
1476 up_read(&o2hb_callback_sem);
1477 }
1478 EXPORT_SYMBOL_GPL(o2hb_fill_node_map);
1479
1480 /*
1481 * heartbeat configfs bits. The heartbeat set is a default set under
1482 * the cluster set in nodemanager.c.
1483 */
1484
1485 static struct o2hb_region *to_o2hb_region(struct config_item *item)
1486 {
1487 return item ? container_of(item, struct o2hb_region, hr_item) : NULL;
1488 }
1489
1490 /* drop_item only drops its ref after killing the thread, nothing should
1491 * be using the region anymore. this has to clean up any state that
1492 * attributes might have built up. */
1493 static void o2hb_region_release(struct config_item *item)
1494 {
1495 int i;
1496 struct page *page;
1497 struct o2hb_region *reg = to_o2hb_region(item);
1498
1499 mlog(ML_HEARTBEAT, "hb region release (%pg)\n", reg_bdev(reg));
1500
1501 kfree(reg->hr_tmp_block);
1502
1503 if (reg->hr_slot_data) {
1504 for (i = 0; i < reg->hr_num_pages; i++) {
1505 page = reg->hr_slot_data[i];
1506 if (page)
1507 __free_page(page);
1508 }
1509 kfree(reg->hr_slot_data);
1510 }
1511
1512 if (reg->hr_bdev_file)
1513 fput(reg->hr_bdev_file);
1514
1515 kfree(reg->hr_slots);
1516
1517 debugfs_remove_recursive(reg->hr_debug_dir);
1518 kfree(reg->hr_db_livenodes);
1519 kfree(reg->hr_db_regnum);
1520 kfree(reg->hr_db_elapsed_time);
1521 kfree(reg->hr_db_pinned);
1522
1523 spin_lock(&o2hb_live_lock);
1524 list_del(®->hr_all_item);
1525 spin_unlock(&o2hb_live_lock);
1526
1527 o2net_unregister_handler_list(®->hr_handler_list);
1528 kfree(reg);
1529 }
1530
1531 static int o2hb_read_block_input(struct o2hb_region *reg,
1532 const char *page,
1533 unsigned long *ret_bytes,
1534 unsigned int *ret_bits)
1535 {
1536 unsigned long bytes;
1537 char *p = (char *)page;
1538
1539 bytes = simple_strtoul(p, &p, 0);
1540 if (!p || (*p && (*p != '\n')))
1541 return -EINVAL;
1542
1543 /* Heartbeat and fs min / max block sizes are the same. */
1544 if (bytes > 4096 || bytes < 512)
1545 return -ERANGE;
1546 if (hweight16(bytes) != 1)
1547 return -EINVAL;
1548
1549 if (ret_bytes)
1550 *ret_bytes = bytes;
1551 if (ret_bits)
1552 *ret_bits = ffs(bytes) - 1;
1553
1554 return 0;
1555 }
1556
1557 static ssize_t o2hb_region_block_bytes_show(struct config_item *item,
1558 char *page)
1559 {
1560 return sprintf(page, "%u\n", to_o2hb_region(item)->hr_block_bytes);
1561 }
1562
1563 static ssize_t o2hb_region_block_bytes_store(struct config_item *item,
1564 const char *page,
1565 size_t count)
1566 {
1567 struct o2hb_region *reg = to_o2hb_region(item);
1568 int status;
1569 unsigned long block_bytes;
1570 unsigned int block_bits;
1571
1572 if (reg->hr_bdev_file)
1573 return -EINVAL;
1574
1575 status = o2hb_read_block_input(reg, page, &block_bytes,
1576 &block_bits);
1577 if (status)
1578 return status;
1579
1580 reg->hr_block_bytes = (unsigned int)block_bytes;
1581 reg->hr_block_bits = block_bits;
1582
1583 return count;
1584 }
1585
1586 static ssize_t o2hb_region_start_block_show(struct config_item *item,
1587 char *page)
1588 {
1589 return sprintf(page, "%llu\n", to_o2hb_region(item)->hr_start_block);
1590 }
1591
1592 static ssize_t o2hb_region_start_block_store(struct config_item *item,
1593 const char *page,
1594 size_t count)
1595 {
1596 struct o2hb_region *reg = to_o2hb_region(item);
1597 unsigned long long tmp;
1598 char *p = (char *)page;
1599 ssize_t ret;
1600
1601 if (reg->hr_bdev_file)
1602 return -EINVAL;
1603
1604 ret = kstrtoull(p, 0, &tmp);
1605 if (ret)
1606 return -EINVAL;
1607
1608 reg->hr_start_block = tmp;
1609
1610 return count;
1611 }
1612
1613 static ssize_t o2hb_region_blocks_show(struct config_item *item, char *page)
1614 {
1615 return sprintf(page, "%d\n", to_o2hb_region(item)->hr_blocks);
1616 }
1617
1618 static ssize_t o2hb_region_blocks_store(struct config_item *item,
1619 const char *page,
1620 size_t count)
1621 {
1622 struct o2hb_region *reg = to_o2hb_region(item);
1623 unsigned long tmp;
1624 char *p = (char *)page;
1625
1626 if (reg->hr_bdev_file)
1627 return -EINVAL;
1628
1629 tmp = simple_strtoul(p, &p, 0);
1630 if (!p || (*p && (*p != '\n')))
1631 return -EINVAL;
1632
1633 if (tmp > O2NM_MAX_NODES || tmp == 0)
1634 return -ERANGE;
1635
1636 reg->hr_blocks = (unsigned int)tmp;
1637
1638 return count;
1639 }
1640
1641 static ssize_t o2hb_region_dev_show(struct config_item *item, char *page)
1642 {
1643 unsigned int ret = 0;
1644
1645 if (to_o2hb_region(item)->hr_bdev_file)
1646 ret = sprintf(page, "%pg\n", reg_bdev(to_o2hb_region(item)));
1647
1648 return ret;
1649 }
1650
1651 static void o2hb_init_region_params(struct o2hb_region *reg)
1652 {
1653 reg->hr_slots_per_page = PAGE_SIZE >> reg->hr_block_bits;
1654 reg->hr_timeout_ms = O2HB_REGION_TIMEOUT_MS;
1655
1656 mlog(ML_HEARTBEAT, "hr_start_block = %llu, hr_blocks = %u\n",
1657 reg->hr_start_block, reg->hr_blocks);
1658 mlog(ML_HEARTBEAT, "hr_block_bytes = %u, hr_block_bits = %u\n",
1659 reg->hr_block_bytes, reg->hr_block_bits);
1660 mlog(ML_HEARTBEAT, "hr_timeout_ms = %u\n", reg->hr_timeout_ms);
1661 mlog(ML_HEARTBEAT, "dead threshold = %u\n", o2hb_dead_threshold);
1662 }
1663
1664 static int o2hb_map_slot_data(struct o2hb_region *reg)
1665 {
1666 int i, j;
1667 unsigned int last_slot;
1668 unsigned int spp = reg->hr_slots_per_page;
1669 struct page *page;
1670 char *raw;
1671 struct o2hb_disk_slot *slot;
1672
1673 reg->hr_tmp_block = kmalloc(reg->hr_block_bytes, GFP_KERNEL);
1674 if (reg->hr_tmp_block == NULL)
1675 return -ENOMEM;
1676
1677 reg->hr_slots = kcalloc(reg->hr_blocks,
1678 sizeof(struct o2hb_disk_slot), GFP_KERNEL);
1679 if (reg->hr_slots == NULL)
1680 return -ENOMEM;
1681
1682 for(i = 0; i < reg->hr_blocks; i++) {
1683 slot = ®->hr_slots[i];
1684 slot->ds_node_num = i;
1685 INIT_LIST_HEAD(&slot->ds_live_item);
1686 slot->ds_raw_block = NULL;
1687 }
1688
1689 reg->hr_num_pages = (reg->hr_blocks + spp - 1) / spp;
1690 mlog(ML_HEARTBEAT, "Going to require %u pages to cover %u blocks "
1691 "at %u blocks per page\n",
1692 reg->hr_num_pages, reg->hr_blocks, spp);
1693
1694 reg->hr_slot_data = kcalloc(reg->hr_num_pages, sizeof(struct page *),
1695 GFP_KERNEL);
1696 if (!reg->hr_slot_data)
1697 return -ENOMEM;
1698
1699 for(i = 0; i < reg->hr_num_pages; i++) {
1700 page = alloc_page(GFP_KERNEL);
1701 if (!page)
1702 return -ENOMEM;
1703
1704 reg->hr_slot_data[i] = page;
1705
1706 last_slot = i * spp;
1707 raw = page_address(page);
1708 for (j = 0;
1709 (j < spp) && ((j + last_slot) < reg->hr_blocks);
1710 j++) {
1711 BUG_ON((j + last_slot) >= reg->hr_blocks);
1712
1713 slot = ®->hr_slots[j + last_slot];
1714 slot->ds_raw_block =
1715 (struct o2hb_disk_heartbeat_block *) raw;
1716
1717 raw += reg->hr_block_bytes;
1718 }
1719 }
1720
1721 return 0;
1722 }
1723
1724 /* Read in all the slots available and populate the tracking
1725 * structures so that we can start with a baseline idea of what's
1726 * there. */
1727 static int o2hb_populate_slot_data(struct o2hb_region *reg)
1728 {
1729 int ret, i;
1730 struct o2hb_disk_slot *slot;
1731 struct o2hb_disk_heartbeat_block *hb_block;
1732
1733 ret = o2hb_read_slots(reg, 0, reg->hr_blocks);
1734 if (ret)
1735 goto out;
1736
1737 /* We only want to get an idea of the values initially in each
1738 * slot, so we do no verification - o2hb_check_slot will
1739 * actually determine if each configured slot is valid and
1740 * whether any values have changed. */
1741 for(i = 0; i < reg->hr_blocks; i++) {
1742 slot = ®->hr_slots[i];
1743 hb_block = (struct o2hb_disk_heartbeat_block *) slot->ds_raw_block;
1744
1745 /* Only fill the values that o2hb_check_slot uses to
1746 * determine changing slots */
1747 slot->ds_last_time = le64_to_cpu(hb_block->hb_seq);
1748 slot->ds_last_generation = le64_to_cpu(hb_block->hb_generation);
1749 }
1750
1751 out:
1752 return ret;
1753 }
1754
1755 /*
1756 * this is acting as commit; we set up all of hr_bdev_file and hr_task or
1757 * nothing
1758 */
1759 static ssize_t o2hb_region_dev_store(struct config_item *item,
1760 const char *page,
1761 size_t count)
1762 {
1763 struct o2hb_region *reg = to_o2hb_region(item);
1764 struct task_struct *hb_task;
1765 long fd;
1766 int sectsize;
1767 char *p = (char *)page;
1768 struct fd f;
1769 ssize_t ret = -EINVAL;
1770 int live_threshold;
1771
1772 if (reg->hr_bdev_file)
1773 goto out;
1774
1775 /* We can't heartbeat without having had our node number
1776 * configured yet. */
1777 if (o2nm_this_node() == O2NM_MAX_NODES)
1778 goto out;
1779
1780 fd = simple_strtol(p, &p, 0);
1781 if (!p || (*p && (*p != '\n')))
1782 goto out;
1783
1784 if (fd < 0 || fd >= INT_MAX)
1785 goto out;
1786
1787 f = fdget(fd);
1788 if (f.file == NULL)
1789 goto out;
1790
1791 if (reg->hr_blocks == 0 || reg->hr_start_block == 0 ||
1792 reg->hr_block_bytes == 0)
1793 goto out2;
1794
1795 if (!S_ISBLK(f.file->f_mapping->host->i_mode))
1796 goto out2;
1797
1798 reg->hr_bdev_file = bdev_file_open_by_dev(f.file->f_mapping->host->i_rdev,
1799 BLK_OPEN_WRITE | BLK_OPEN_READ, NULL, NULL);
1800 if (IS_ERR(reg->hr_bdev_file)) {
1801 ret = PTR_ERR(reg->hr_bdev_file);
1802 reg->hr_bdev_file = NULL;
1803 goto out2;
1804 }
1805
1806 sectsize = bdev_logical_block_size(reg_bdev(reg));
1807 if (sectsize != reg->hr_block_bytes) {
1808 mlog(ML_ERROR,
1809 "blocksize %u incorrect for device, expected %d",
1810 reg->hr_block_bytes, sectsize);
1811 ret = -EINVAL;
1812 goto out3;
1813 }
1814
1815 o2hb_init_region_params(reg);
1816
1817 /* Generation of zero is invalid */
1818 do {
1819 get_random_bytes(®->hr_generation,
1820 sizeof(reg->hr_generation));
1821 } while (reg->hr_generation == 0);
1822
1823 ret = o2hb_map_slot_data(reg);
1824 if (ret) {
1825 mlog_errno(ret);
1826 goto out3;
1827 }
1828
1829 ret = o2hb_populate_slot_data(reg);
1830 if (ret) {
1831 mlog_errno(ret);
1832 goto out3;
1833 }
1834
1835 INIT_DELAYED_WORK(®->hr_write_timeout_work, o2hb_write_timeout);
1836 INIT_DELAYED_WORK(®->hr_nego_timeout_work, o2hb_nego_timeout);
1837
1838 /*
1839 * A node is considered live after it has beat LIVE_THRESHOLD
1840 * times. We're not steady until we've given them a chance
1841 * _after_ our first read.
1842 * The default threshold is bare minimum so as to limit the delay
1843 * during mounts. For global heartbeat, the threshold doubled for the
1844 * first region.
1845 */
1846 live_threshold = O2HB_LIVE_THRESHOLD;
1847 if (o2hb_global_heartbeat_active()) {
1848 spin_lock(&o2hb_live_lock);
1849 if (bitmap_weight(o2hb_region_bitmap, O2NM_MAX_REGIONS) == 1)
1850 live_threshold <<= 1;
1851 spin_unlock(&o2hb_live_lock);
1852 }
1853 ++live_threshold;
1854 atomic_set(®->hr_steady_iterations, live_threshold);
1855 /* unsteady_iterations is triple the steady_iterations */
1856 atomic_set(®->hr_unsteady_iterations, (live_threshold * 3));
1857
1858 hb_task = kthread_run(o2hb_thread, reg, "o2hb-%s",
1859 reg->hr_item.ci_name);
1860 if (IS_ERR(hb_task)) {
1861 ret = PTR_ERR(hb_task);
1862 mlog_errno(ret);
1863 goto out3;
1864 }
1865
1866 spin_lock(&o2hb_live_lock);
1867 reg->hr_task = hb_task;
1868 spin_unlock(&o2hb_live_lock);
1869
1870 ret = wait_event_interruptible(o2hb_steady_queue,
1871 atomic_read(®->hr_steady_iterations) == 0 ||
1872 reg->hr_node_deleted);
1873 if (ret) {
1874 atomic_set(®->hr_steady_iterations, 0);
1875 reg->hr_aborted_start = 1;
1876 }
1877
1878 if (reg->hr_aborted_start) {
1879 ret = -EIO;
1880 goto out3;
1881 }
1882
1883 if (reg->hr_node_deleted) {
1884 ret = -EINVAL;
1885 goto out3;
1886 }
1887
1888 /* Ok, we were woken. Make sure it wasn't by drop_item() */
1889 spin_lock(&o2hb_live_lock);
1890 hb_task = reg->hr_task;
1891 if (o2hb_global_heartbeat_active())
1892 set_bit(reg->hr_region_num, o2hb_live_region_bitmap);
1893 spin_unlock(&o2hb_live_lock);
1894
1895 if (hb_task)
1896 ret = count;
1897 else
1898 ret = -EIO;
1899
1900 if (hb_task && o2hb_global_heartbeat_active())
1901 printk(KERN_NOTICE "o2hb: Heartbeat started on region %s (%pg)\n",
1902 config_item_name(®->hr_item), reg_bdev(reg));
1903
1904 out3:
1905 if (ret < 0) {
1906 fput(reg->hr_bdev_file);
1907 reg->hr_bdev_file = NULL;
1908 }
1909 out2:
1910 fdput(f);
1911 out:
1912 return ret;
1913 }
1914
1915 static ssize_t o2hb_region_pid_show(struct config_item *item, char *page)
1916 {
1917 struct o2hb_region *reg = to_o2hb_region(item);
1918 pid_t pid = 0;
1919
1920 spin_lock(&o2hb_live_lock);
1921 if (reg->hr_task)
1922 pid = task_pid_nr(reg->hr_task);
1923 spin_unlock(&o2hb_live_lock);
1924
1925 if (!pid)
1926 return 0;
1927
1928 return sprintf(page, "%u\n", pid);
1929 }
1930
1931 CONFIGFS_ATTR(o2hb_region_, block_bytes);
1932 CONFIGFS_ATTR(o2hb_region_, start_block);
1933 CONFIGFS_ATTR(o2hb_region_, blocks);
1934 CONFIGFS_ATTR(o2hb_region_, dev);
1935 CONFIGFS_ATTR_RO(o2hb_region_, pid);
1936
1937 static struct configfs_attribute *o2hb_region_attrs[] = {
1938 &o2hb_region_attr_block_bytes,
1939 &o2hb_region_attr_start_block,
1940 &o2hb_region_attr_blocks,
1941 &o2hb_region_attr_dev,
1942 &o2hb_region_attr_pid,
1943 NULL,
1944 };
1945
1946 static struct configfs_item_operations o2hb_region_item_ops = {
1947 .release = o2hb_region_release,
1948 };
1949
1950 static const struct config_item_type o2hb_region_type = {
1951 .ct_item_ops = &o2hb_region_item_ops,
1952 .ct_attrs = o2hb_region_attrs,
1953 .ct_owner = THIS_MODULE,
1954 };
1955
1956 /* heartbeat set */
1957
1958 struct o2hb_heartbeat_group {
1959 struct config_group hs_group;
1960 /* some stuff? */
1961 };
1962
1963 static struct o2hb_heartbeat_group *to_o2hb_heartbeat_group(struct config_group *group)
1964 {
1965 return group ?
1966 container_of(group, struct o2hb_heartbeat_group, hs_group)
1967 : NULL;
1968 }
1969
1970 static void o2hb_debug_region_init(struct o2hb_region *reg,
1971 struct dentry *parent)
1972 {
1973 struct dentry *dir;
1974
1975 dir = debugfs_create_dir(config_item_name(®->hr_item), parent);
1976 reg->hr_debug_dir = dir;
1977
1978 o2hb_debug_create(O2HB_DEBUG_LIVENODES, dir, &(reg->hr_db_livenodes),
1979 sizeof(*(reg->hr_db_livenodes)),
1980 O2HB_DB_TYPE_REGION_LIVENODES,
1981 sizeof(reg->hr_live_node_bitmap), O2NM_MAX_NODES,
1982 reg);
1983
1984 o2hb_debug_create(O2HB_DEBUG_REGION_NUMBER, dir, &(reg->hr_db_regnum),
1985 sizeof(*(reg->hr_db_regnum)),
1986 O2HB_DB_TYPE_REGION_NUMBER, 0, O2NM_MAX_NODES, reg);
1987
1988 o2hb_debug_create(O2HB_DEBUG_REGION_ELAPSED_TIME, dir,
1989 &(reg->hr_db_elapsed_time),
1990 sizeof(*(reg->hr_db_elapsed_time)),
1991 O2HB_DB_TYPE_REGION_ELAPSED_TIME, 0, 0, reg);
1992
1993 o2hb_debug_create(O2HB_DEBUG_REGION_PINNED, dir, &(reg->hr_db_pinned),
1994 sizeof(*(reg->hr_db_pinned)),
1995 O2HB_DB_TYPE_REGION_PINNED, 0, 0, reg);
1996
1997 }
1998
1999 static struct config_item *o2hb_heartbeat_group_make_item(struct config_group *group,
2000 const char *name)
2001 {
2002 struct o2hb_region *reg = NULL;
2003 int ret;
2004
2005 reg = kzalloc(sizeof(struct o2hb_region), GFP_KERNEL);
2006 if (reg == NULL)
2007 return ERR_PTR(-ENOMEM);
2008
2009 if (strlen(name) > O2HB_MAX_REGION_NAME_LEN) {
2010 ret = -ENAMETOOLONG;
2011 goto free;
2012 }
2013
2014 spin_lock(&o2hb_live_lock);
2015 reg->hr_region_num = 0;
2016 if (o2hb_global_heartbeat_active()) {
2017 reg->hr_region_num = find_first_zero_bit(o2hb_region_bitmap,
2018 O2NM_MAX_REGIONS);
2019 if (reg->hr_region_num >= O2NM_MAX_REGIONS) {
2020 spin_unlock(&o2hb_live_lock);
2021 ret = -EFBIG;
2022 goto free;
2023 }
2024 set_bit(reg->hr_region_num, o2hb_region_bitmap);
2025 }
2026 list_add_tail(®->hr_all_item, &o2hb_all_regions);
2027 spin_unlock(&o2hb_live_lock);
2028
2029 config_item_init_type_name(®->hr_item, name, &o2hb_region_type);
2030
2031 /* this is the same way to generate msg key as dlm, for local heartbeat,
2032 * name is also the same, so make initial crc value different to avoid
2033 * message key conflict.
2034 */
2035 reg->hr_key = crc32_le(reg->hr_region_num + O2NM_MAX_REGIONS,
2036 name, strlen(name));
2037 INIT_LIST_HEAD(®->hr_handler_list);
2038 ret = o2net_register_handler(O2HB_NEGO_TIMEOUT_MSG, reg->hr_key,
2039 sizeof(struct o2hb_nego_msg),
2040 o2hb_nego_timeout_handler,
2041 reg, NULL, ®->hr_handler_list);
2042 if (ret)
2043 goto remove_item;
2044
2045 ret = o2net_register_handler(O2HB_NEGO_APPROVE_MSG, reg->hr_key,
2046 sizeof(struct o2hb_nego_msg),
2047 o2hb_nego_approve_handler,
2048 reg, NULL, ®->hr_handler_list);
2049 if (ret)
2050 goto unregister_handler;
2051
2052 o2hb_debug_region_init(reg, o2hb_debug_dir);
2053
2054 return ®->hr_item;
2055
2056 unregister_handler:
2057 o2net_unregister_handler_list(®->hr_handler_list);
2058 remove_item:
2059 spin_lock(&o2hb_live_lock);
2060 list_del(®->hr_all_item);
2061 if (o2hb_global_heartbeat_active())
2062 clear_bit(reg->hr_region_num, o2hb_region_bitmap);
2063 spin_unlock(&o2hb_live_lock);
2064 free:
2065 kfree(reg);
2066 return ERR_PTR(ret);
2067 }
2068
2069 static void o2hb_heartbeat_group_drop_item(struct config_group *group,
2070 struct config_item *item)
2071 {
2072 struct task_struct *hb_task;
2073 struct o2hb_region *reg = to_o2hb_region(item);
2074 int quorum_region = 0;
2075
2076 /* stop the thread when the user removes the region dir */
2077 spin_lock(&o2hb_live_lock);
2078 hb_task = reg->hr_task;
2079 reg->hr_task = NULL;
2080 reg->hr_item_dropped = 1;
2081 spin_unlock(&o2hb_live_lock);
2082
2083 if (hb_task)
2084 kthread_stop(hb_task);
2085
2086 if (o2hb_global_heartbeat_active()) {
2087 spin_lock(&o2hb_live_lock);
2088 clear_bit(reg->hr_region_num, o2hb_region_bitmap);
2089 clear_bit(reg->hr_region_num, o2hb_live_region_bitmap);
2090 if (test_bit(reg->hr_region_num, o2hb_quorum_region_bitmap))
2091 quorum_region = 1;
2092 clear_bit(reg->hr_region_num, o2hb_quorum_region_bitmap);
2093 spin_unlock(&o2hb_live_lock);
2094 printk(KERN_NOTICE "o2hb: Heartbeat %s on region %s (%pg)\n",
2095 ((atomic_read(®->hr_steady_iterations) == 0) ?
2096 "stopped" : "start aborted"), config_item_name(item),
2097 reg_bdev(reg));
2098 }
2099
2100 /*
2101 * If we're racing a dev_write(), we need to wake them. They will
2102 * check reg->hr_task
2103 */
2104 if (atomic_read(®->hr_steady_iterations) != 0) {
2105 reg->hr_aborted_start = 1;
2106 atomic_set(®->hr_steady_iterations, 0);
2107 wake_up(&o2hb_steady_queue);
2108 }
2109
2110 config_item_put(item);
2111
2112 if (!o2hb_global_heartbeat_active() || !quorum_region)
2113 return;
2114
2115 /*
2116 * If global heartbeat active and there are dependent users,
2117 * pin all regions if quorum region count <= CUT_OFF
2118 */
2119 spin_lock(&o2hb_live_lock);
2120
2121 if (!o2hb_dependent_users)
2122 goto unlock;
2123
2124 if (bitmap_weight(o2hb_quorum_region_bitmap,
2125 O2NM_MAX_REGIONS) <= O2HB_PIN_CUT_OFF)
2126 o2hb_region_pin(NULL);
2127
2128 unlock:
2129 spin_unlock(&o2hb_live_lock);
2130 }
2131
2132 static ssize_t o2hb_heartbeat_group_dead_threshold_show(struct config_item *item,
2133 char *page)
2134 {
2135 return sprintf(page, "%u\n", o2hb_dead_threshold);
2136 }
2137
2138 static ssize_t o2hb_heartbeat_group_dead_threshold_store(struct config_item *item,
2139 const char *page, size_t count)
2140 {
2141 unsigned long tmp;
2142 char *p = (char *)page;
2143
2144 tmp = simple_strtoul(p, &p, 10);
2145 if (!p || (*p && (*p != '\n')))
2146 return -EINVAL;
2147
2148 /* this will validate ranges for us. */
2149 o2hb_dead_threshold_set((unsigned int) tmp);
2150
2151 return count;
2152 }
2153
2154 static ssize_t o2hb_heartbeat_group_mode_show(struct config_item *item,
2155 char *page)
2156 {
2157 return sprintf(page, "%s\n",
2158 o2hb_heartbeat_mode_desc[o2hb_heartbeat_mode]);
2159 }
2160
2161 static ssize_t o2hb_heartbeat_group_mode_store(struct config_item *item,
2162 const char *page, size_t count)
2163 {
2164 unsigned int i;
2165 int ret;
2166 size_t len;
2167
2168 len = (page[count - 1] == '\n') ? count - 1 : count;
2169 if (!len)
2170 return -EINVAL;
2171
2172 for (i = 0; i < O2HB_HEARTBEAT_NUM_MODES; ++i) {
2173 if (strncasecmp(page, o2hb_heartbeat_mode_desc[i], len))
2174 continue;
2175
2176 ret = o2hb_global_heartbeat_mode_set(i);
2177 if (!ret)
2178 printk(KERN_NOTICE "o2hb: Heartbeat mode set to %s\n",
2179 o2hb_heartbeat_mode_desc[i]);
2180 return count;
2181 }
2182
2183 return -EINVAL;
2184
2185 }
2186
2187 CONFIGFS_ATTR(o2hb_heartbeat_group_, dead_threshold);
2188 CONFIGFS_ATTR(o2hb_heartbeat_group_, mode);
2189
2190 static struct configfs_attribute *o2hb_heartbeat_group_attrs[] = {
2191 &o2hb_heartbeat_group_attr_dead_threshold,
2192 &o2hb_heartbeat_group_attr_mode,
2193 NULL,
2194 };
2195
2196 static struct configfs_group_operations o2hb_heartbeat_group_group_ops = {
2197 .make_item = o2hb_heartbeat_group_make_item,
2198 .drop_item = o2hb_heartbeat_group_drop_item,
2199 };
2200
2201 static const struct config_item_type o2hb_heartbeat_group_type = {
2202 .ct_group_ops = &o2hb_heartbeat_group_group_ops,
2203 .ct_attrs = o2hb_heartbeat_group_attrs,
2204 .ct_owner = THIS_MODULE,
2205 };
2206
2207 /* this is just here to avoid touching group in heartbeat.h which the
2208 * entire damn world #includes */
2209 struct config_group *o2hb_alloc_hb_set(void)
2210 {
2211 struct o2hb_heartbeat_group *hs = NULL;
2212 struct config_group *ret = NULL;
2213
2214 hs = kzalloc(sizeof(struct o2hb_heartbeat_group), GFP_KERNEL);
2215 if (hs == NULL)
2216 goto out;
2217
2218 config_group_init_type_name(&hs->hs_group, "heartbeat",
2219 &o2hb_heartbeat_group_type);
2220
2221 ret = &hs->hs_group;
2222 out:
2223 if (ret == NULL)
2224 kfree(hs);
2225 return ret;
2226 }
2227
2228 void o2hb_free_hb_set(struct config_group *group)
2229 {
2230 struct o2hb_heartbeat_group *hs = to_o2hb_heartbeat_group(group);
2231 kfree(hs);
2232 }
2233
2234 /* hb callback registration and issuing */
2235
2236 static struct o2hb_callback *hbcall_from_type(enum o2hb_callback_type type)
2237 {
2238 if (type == O2HB_NUM_CB)
2239 return ERR_PTR(-EINVAL);
2240
2241 return &o2hb_callbacks[type];
2242 }
2243
2244 void o2hb_setup_callback(struct o2hb_callback_func *hc,
2245 enum o2hb_callback_type type,
2246 o2hb_cb_func *func,
2247 void *data,
2248 int priority)
2249 {
2250 INIT_LIST_HEAD(&hc->hc_item);
2251 hc->hc_func = func;
2252 hc->hc_data = data;
2253 hc->hc_priority = priority;
2254 hc->hc_type = type;
2255 hc->hc_magic = O2HB_CB_MAGIC;
2256 }
2257 EXPORT_SYMBOL_GPL(o2hb_setup_callback);
2258
2259 /*
2260 * In local heartbeat mode, region_uuid passed matches the dlm domain name.
2261 * In global heartbeat mode, region_uuid passed is NULL.
2262 *
2263 * In local, we only pin the matching region. In global we pin all the active
2264 * regions.
2265 */
2266 static int o2hb_region_pin(const char *region_uuid)
2267 {
2268 int ret = 0, found = 0;
2269 struct o2hb_region *reg;
2270 char *uuid;
2271
2272 assert_spin_locked(&o2hb_live_lock);
2273
2274 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) {
2275 if (reg->hr_item_dropped)
2276 continue;
2277
2278 uuid = config_item_name(®->hr_item);
2279
2280 /* local heartbeat */
2281 if (region_uuid) {
2282 if (strcmp(region_uuid, uuid))
2283 continue;
2284 found = 1;
2285 }
2286
2287 if (reg->hr_item_pinned || reg->hr_item_dropped)
2288 goto skip_pin;
2289
2290 /* Ignore ENOENT only for local hb (userdlm domain) */
2291 ret = o2nm_depend_item(®->hr_item);
2292 if (!ret) {
2293 mlog(ML_CLUSTER, "Pin region %s\n", uuid);
2294 reg->hr_item_pinned = 1;
2295 } else {
2296 if (ret == -ENOENT && found)
2297 ret = 0;
2298 else {
2299 mlog(ML_ERROR, "Pin region %s fails with %d\n",
2300 uuid, ret);
2301 break;
2302 }
2303 }
2304 skip_pin:
2305 if (found)
2306 break;
2307 }
2308
2309 return ret;
2310 }
2311
2312 /*
2313 * In local heartbeat mode, region_uuid passed matches the dlm domain name.
2314 * In global heartbeat mode, region_uuid passed is NULL.
2315 *
2316 * In local, we only unpin the matching region. In global we unpin all the
2317 * active regions.
2318 */
2319 static void o2hb_region_unpin(const char *region_uuid)
2320 {
2321 struct o2hb_region *reg;
2322 char *uuid;
2323 int found = 0;
2324
2325 assert_spin_locked(&o2hb_live_lock);
2326
2327 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) {
2328 if (reg->hr_item_dropped)
2329 continue;
2330
2331 uuid = config_item_name(®->hr_item);
2332 if (region_uuid) {
2333 if (strcmp(region_uuid, uuid))
2334 continue;
2335 found = 1;
2336 }
2337
2338 if (reg->hr_item_pinned) {
2339 mlog(ML_CLUSTER, "Unpin region %s\n", uuid);
2340 o2nm_undepend_item(®->hr_item);
2341 reg->hr_item_pinned = 0;
2342 }
2343 if (found)
2344 break;
2345 }
2346 }
2347
2348 static int o2hb_region_inc_user(const char *region_uuid)
2349 {
2350 int ret = 0;
2351
2352 spin_lock(&o2hb_live_lock);
2353
2354 /* local heartbeat */
2355 if (!o2hb_global_heartbeat_active()) {
2356 ret = o2hb_region_pin(region_uuid);
2357 goto unlock;
2358 }
2359
2360 /*
2361 * if global heartbeat active and this is the first dependent user,
2362 * pin all regions if quorum region count <= CUT_OFF
2363 */
2364 o2hb_dependent_users++;
2365 if (o2hb_dependent_users > 1)
2366 goto unlock;
2367
2368 if (bitmap_weight(o2hb_quorum_region_bitmap,
2369 O2NM_MAX_REGIONS) <= O2HB_PIN_CUT_OFF)
2370 ret = o2hb_region_pin(NULL);
2371
2372 unlock:
2373 spin_unlock(&o2hb_live_lock);
2374 return ret;
2375 }
2376
2377 static void o2hb_region_dec_user(const char *region_uuid)
2378 {
2379 spin_lock(&o2hb_live_lock);
2380
2381 /* local heartbeat */
2382 if (!o2hb_global_heartbeat_active()) {
2383 o2hb_region_unpin(region_uuid);
2384 goto unlock;
2385 }
2386
2387 /*
2388 * if global heartbeat active and there are no dependent users,
2389 * unpin all quorum regions
2390 */
2391 o2hb_dependent_users--;
2392 if (!o2hb_dependent_users)
2393 o2hb_region_unpin(NULL);
2394
2395 unlock:
2396 spin_unlock(&o2hb_live_lock);
2397 }
2398
2399 int o2hb_register_callback(const char *region_uuid,
2400 struct o2hb_callback_func *hc)
2401 {
2402 struct o2hb_callback_func *f;
2403 struct o2hb_callback *hbcall;
2404 int ret;
2405
2406 BUG_ON(hc->hc_magic != O2HB_CB_MAGIC);
2407 BUG_ON(!list_empty(&hc->hc_item));
2408
2409 hbcall = hbcall_from_type(hc->hc_type);
2410 if (IS_ERR(hbcall)) {
2411 ret = PTR_ERR(hbcall);
2412 goto out;
2413 }
2414
2415 if (region_uuid) {
2416 ret = o2hb_region_inc_user(region_uuid);
2417 if (ret) {
2418 mlog_errno(ret);
2419 goto out;
2420 }
2421 }
2422
2423 down_write(&o2hb_callback_sem);
2424
2425 list_for_each_entry(f, &hbcall->list, hc_item) {
2426 if (hc->hc_priority < f->hc_priority) {
2427 list_add_tail(&hc->hc_item, &f->hc_item);
2428 break;
2429 }
2430 }
2431 if (list_empty(&hc->hc_item))
2432 list_add_tail(&hc->hc_item, &hbcall->list);
2433
2434 up_write(&o2hb_callback_sem);
2435 ret = 0;
2436 out:
2437 mlog(ML_CLUSTER, "returning %d on behalf of %p for funcs %p\n",
2438 ret, __builtin_return_address(0), hc);
2439 return ret;
2440 }
2441 EXPORT_SYMBOL_GPL(o2hb_register_callback);
2442
2443 void o2hb_unregister_callback(const char *region_uuid,
2444 struct o2hb_callback_func *hc)
2445 {
2446 BUG_ON(hc->hc_magic != O2HB_CB_MAGIC);
2447
2448 mlog(ML_CLUSTER, "on behalf of %p for funcs %p\n",
2449 __builtin_return_address(0), hc);
2450
2451 /* XXX Can this happen _with_ a region reference? */
2452 if (list_empty(&hc->hc_item))
2453 return;
2454
2455 if (region_uuid)
2456 o2hb_region_dec_user(region_uuid);
2457
2458 down_write(&o2hb_callback_sem);
2459
2460 list_del_init(&hc->hc_item);
2461
2462 up_write(&o2hb_callback_sem);
2463 }
2464 EXPORT_SYMBOL_GPL(o2hb_unregister_callback);
2465
2466 int o2hb_check_node_heartbeating_no_sem(u8 node_num)
2467 {
2468 unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
2469
2470 spin_lock(&o2hb_live_lock);
2471 o2hb_fill_node_map_from_callback(testing_map, O2NM_MAX_NODES);
2472 spin_unlock(&o2hb_live_lock);
2473 if (!test_bit(node_num, testing_map)) {
2474 mlog(ML_HEARTBEAT,
2475 "node (%u) does not have heartbeating enabled.\n",
2476 node_num);
2477 return 0;
2478 }
2479
2480 return 1;
2481 }
2482 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating_no_sem);
2483
2484 int o2hb_check_node_heartbeating_from_callback(u8 node_num)
2485 {
2486 unsigned long testing_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
2487
2488 o2hb_fill_node_map_from_callback(testing_map, O2NM_MAX_NODES);
2489 if (!test_bit(node_num, testing_map)) {
2490 mlog(ML_HEARTBEAT,
2491 "node (%u) does not have heartbeating enabled.\n",
2492 node_num);
2493 return 0;
2494 }
2495
2496 return 1;
2497 }
2498 EXPORT_SYMBOL_GPL(o2hb_check_node_heartbeating_from_callback);
2499
2500 /*
2501 * this is just a hack until we get the plumbing which flips file systems
2502 * read only and drops the hb ref instead of killing the node dead.
2503 */
2504 void o2hb_stop_all_regions(void)
2505 {
2506 struct o2hb_region *reg;
2507
2508 mlog(ML_ERROR, "stopping heartbeat on all active regions.\n");
2509
2510 spin_lock(&o2hb_live_lock);
2511
2512 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item)
2513 reg->hr_unclean_stop = 1;
2514
2515 spin_unlock(&o2hb_live_lock);
2516 }
2517 EXPORT_SYMBOL_GPL(o2hb_stop_all_regions);
2518
2519 int o2hb_get_all_regions(char *region_uuids, u8 max_regions)
2520 {
2521 struct o2hb_region *reg;
2522 int numregs = 0;
2523 char *p;
2524
2525 spin_lock(&o2hb_live_lock);
2526
2527 p = region_uuids;
2528 list_for_each_entry(reg, &o2hb_all_regions, hr_all_item) {
2529 if (reg->hr_item_dropped)
2530 continue;
2531
2532 mlog(0, "Region: %s\n", config_item_name(®->hr_item));
2533 if (numregs < max_regions) {
2534 memcpy(p, config_item_name(®->hr_item),
2535 O2HB_MAX_REGION_NAME_LEN);
2536 p += O2HB_MAX_REGION_NAME_LEN;
2537 }
2538 numregs++;
2539 }
2540
2541 spin_unlock(&o2hb_live_lock);
2542
2543 return numregs;
2544 }
2545 EXPORT_SYMBOL_GPL(o2hb_get_all_regions);
2546
2547 int o2hb_global_heartbeat_active(void)
2548 {
2549 return (o2hb_heartbeat_mode == O2HB_HEARTBEAT_GLOBAL);
2550 }
2551 EXPORT_SYMBOL(o2hb_global_heartbeat_active);
2552
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.