blob: 2b13117fb918cbe27775ba61cc68c6f78e5408ff [file] [log] [blame]
Goldwyn Rodrigues8e854e92014-03-07 11:21:15 -06001/*
2 * Copyright (C) 2015, SUSE
3 *
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2, or (at your option)
7 * any later version.
8 *
9 */
10
11
12#include <linux/module.h>
Guoqing Jiang7bcda712016-08-12 13:42:42 +080013#include <linux/kthread.h>
Goldwyn Rodrigues47741b72014-03-07 13:49:26 -060014#include <linux/dlm.h>
15#include <linux/sched.h>
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -050016#include <linux/raid/md_p.h>
Goldwyn Rodrigues47741b72014-03-07 13:49:26 -060017#include "md.h"
Goldwyn Rodriguese94987d2014-06-07 00:45:22 -050018#include "bitmap.h"
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -050019#include "md-cluster.h"
Goldwyn Rodrigues47741b72014-03-07 13:49:26 -060020
21#define LVB_SIZE 64
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -050022#define NEW_DEV_TIMEOUT 5000
Goldwyn Rodrigues47741b72014-03-07 13:49:26 -060023
24struct dlm_lock_resource {
25 dlm_lockspace_t *ls;
26 struct dlm_lksb lksb;
27 char *name; /* lock name. */
28 uint32_t flags; /* flags to pass to dlm_lock() */
Guoqing Jiangfccb60a2016-08-12 13:42:41 +080029 wait_queue_head_t sync_locking; /* wait queue for synchronized locking */
30 bool sync_locking_done;
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -050031 void (*bast)(void *arg, int mode); /* blocking AST function pointer*/
32 struct mddev *mddev; /* pointing back to mddev. */
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -050033 int mode;
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -050034};
35
Goldwyn Rodrigues96ae9232014-06-06 12:35:34 -050036struct suspend_info {
37 int slot;
38 sector_t lo;
39 sector_t hi;
40 struct list_head list;
41};
42
43struct resync_info {
44 __le64 lo;
45 __le64 hi;
46};
47
Goldwyn Rodriguesfa8259d2015-03-02 10:55:49 -060048/* md_cluster_info flags */
49#define MD_CLUSTER_WAITING_FOR_NEWDISK 1
Goldwyn Rodrigues90382ed2015-06-24 09:30:32 -050050#define MD_CLUSTER_SUSPEND_READ_BALANCING 2
Guoqing Jiangeece0752015-07-10 17:01:21 +080051#define MD_CLUSTER_BEGIN_JOIN_CLUSTER 3
Goldwyn Rodriguesfa8259d2015-03-02 10:55:49 -060052
Guoqing Jiang8b9277c2015-12-21 10:51:00 +110053/* Lock the send communication. This is done through
54 * bit manipulation as opposed to a mutex in order to
55 * accomodate lock and hold. See next comment.
56 */
57#define MD_CLUSTER_SEND_LOCK 4
Guoqing Jiange19508f2015-12-21 10:51:01 +110058/* If cluster operations (such as adding a disk) must lock the
59 * communication channel, so as to perform extra operations
60 * (update metadata) and no other operation is allowed on the
61 * MD. Token needs to be locked and held until the operation
62 * completes witha md_update_sb(), which would eventually release
63 * the lock.
Guoqing Jiang8b9277c2015-12-21 10:51:00 +110064 */
65#define MD_CLUSTER_SEND_LOCKED_ALREADY 5
Guoqing Jiang51e453a2016-05-04 02:17:09 -040066/* We should receive message after node joined cluster and
67 * set up all the related infos such as bitmap and personality */
68#define MD_CLUSTER_ALREADY_IN_CLUSTER 6
69#define MD_CLUSTER_PENDING_RECV_EVENT 7
Guoqing Jiang8b9277c2015-12-21 10:51:00 +110070
Goldwyn Rodriguesfa8259d2015-03-02 10:55:49 -060071
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -050072struct md_cluster_info {
73 /* dlm lock space and resources for clustered raid. */
74 dlm_lockspace_t *lockspace;
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -050075 int slot_number;
76 struct completion completion;
Guoqing Jiang8b9277c2015-12-21 10:51:00 +110077 struct mutex recv_mutex;
Goldwyn Rodrigues54519c52014-06-06 12:12:32 -050078 struct dlm_lock_resource *bitmap_lockres;
Guoqing Jiangf6a2dc62015-12-21 10:51:00 +110079 struct dlm_lock_resource **other_bitmap_lockres;
Goldwyn Rodriguesc186b122015-09-30 13:20:35 -050080 struct dlm_lock_resource *resync_lockres;
Goldwyn Rodrigues96ae9232014-06-06 12:35:34 -050081 struct list_head suspend_list;
82 spinlock_t suspend_lock;
Goldwyn Rodriguese94987d2014-06-07 00:45:22 -050083 struct md_thread *recovery_thread;
84 unsigned long recovery_map;
Goldwyn Rodrigues46646802014-06-07 01:08:29 -050085 /* communication loc resources */
86 struct dlm_lock_resource *ack_lockres;
87 struct dlm_lock_resource *message_lockres;
88 struct dlm_lock_resource *token_lockres;
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -050089 struct dlm_lock_resource *no_new_dev_lockres;
Goldwyn Rodrigues46646802014-06-07 01:08:29 -050090 struct md_thread *recv_thread;
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -050091 struct completion newdisk_completion;
Guoqing Jiang8b9277c2015-12-21 10:51:00 +110092 wait_queue_head_t wait;
Goldwyn Rodriguesfa8259d2015-03-02 10:55:49 -060093 unsigned long state;
Guoqing Jiang18c9ff72016-05-02 11:50:12 -040094 /* record the region in RESYNCING message */
95 sector_t sync_low;
96 sector_t sync_hi;
Goldwyn Rodrigues46646802014-06-07 01:08:29 -050097};
98
99enum msg_type {
100 METADATA_UPDATED = 0,
101 RESYNCING,
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500102 NEWDISK,
Goldwyn Rodrigues88bcfef2015-04-14 10:44:44 -0500103 REMOVE,
Goldwyn Rodrigues97f6cd32015-04-14 10:45:42 -0500104 RE_ADD,
Guoqing Jiangdc737d72015-07-10 16:54:04 +0800105 BITMAP_NEEDS_SYNC,
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500106};
107
108struct cluster_msg {
Guoqing Jiangcf97a342015-10-16 15:40:22 +0800109 __le32 type;
110 __le32 slot;
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500111 /* TODO: Unionize this for smaller footprint */
Guoqing Jiangcf97a342015-10-16 15:40:22 +0800112 __le64 low;
113 __le64 high;
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500114 char uuid[16];
Guoqing Jiangcf97a342015-10-16 15:40:22 +0800115 __le32 raid_slot;
Goldwyn Rodrigues47741b72014-03-07 13:49:26 -0600116};
117
118static void sync_ast(void *arg)
119{
120 struct dlm_lock_resource *res;
121
NeilBrown2e2a7cd2015-10-19 15:42:18 +1100122 res = arg;
Guoqing Jiangfccb60a2016-08-12 13:42:41 +0800123 res->sync_locking_done = true;
124 wake_up(&res->sync_locking);
Goldwyn Rodrigues47741b72014-03-07 13:49:26 -0600125}
126
127static int dlm_lock_sync(struct dlm_lock_resource *res, int mode)
128{
129 int ret = 0;
130
Goldwyn Rodrigues47741b72014-03-07 13:49:26 -0600131 ret = dlm_lock(res->ls, mode, &res->lksb,
132 res->flags, res->name, strlen(res->name),
133 0, sync_ast, res, res->bast);
134 if (ret)
135 return ret;
Guoqing Jiangfccb60a2016-08-12 13:42:41 +0800136 wait_event(res->sync_locking, res->sync_locking_done);
137 res->sync_locking_done = false;
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -0500138 if (res->lksb.sb_status == 0)
139 res->mode = mode;
Goldwyn Rodrigues47741b72014-03-07 13:49:26 -0600140 return res->lksb.sb_status;
141}
142
143static int dlm_unlock_sync(struct dlm_lock_resource *res)
144{
145 return dlm_lock_sync(res, DLM_LOCK_NL);
146}
147
Guoqing Jiang7bcda712016-08-12 13:42:42 +0800148/*
149 * An variation of dlm_lock_sync, which make lock request could
150 * be interrupted
151 */
152static int dlm_lock_sync_interruptible(struct dlm_lock_resource *res, int mode,
153 struct mddev *mddev)
154{
155 int ret = 0;
156
157 ret = dlm_lock(res->ls, mode, &res->lksb,
158 res->flags, res->name, strlen(res->name),
159 0, sync_ast, res, res->bast);
160 if (ret)
161 return ret;
162
163 wait_event(res->sync_locking, res->sync_locking_done
Guoqing Jiangd6385db2016-08-12 13:42:43 +0800164 || kthread_should_stop()
165 || test_bit(MD_CLOSING, &mddev->flags));
Guoqing Jiang7bcda712016-08-12 13:42:42 +0800166 if (!res->sync_locking_done) {
167 /*
168 * the convert queue contains the lock request when request is
169 * interrupted, and sync_ast could still be run, so need to
170 * cancel the request and reset completion
171 */
172 ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_CANCEL,
173 &res->lksb, res);
174 res->sync_locking_done = false;
175 if (unlikely(ret != 0))
176 pr_info("failed to cancel previous lock request "
177 "%s return %d\n", res->name, ret);
178 return -EPERM;
179 } else
180 res->sync_locking_done = false;
181 if (res->lksb.sb_status == 0)
182 res->mode = mode;
183 return res->lksb.sb_status;
184}
185
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -0500186static struct dlm_lock_resource *lockres_init(struct mddev *mddev,
Goldwyn Rodrigues47741b72014-03-07 13:49:26 -0600187 char *name, void (*bastfn)(void *arg, int mode), int with_lvb)
188{
189 struct dlm_lock_resource *res = NULL;
190 int ret, namelen;
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -0500191 struct md_cluster_info *cinfo = mddev->cluster_info;
Goldwyn Rodrigues47741b72014-03-07 13:49:26 -0600192
193 res = kzalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
194 if (!res)
195 return NULL;
Guoqing Jiangfccb60a2016-08-12 13:42:41 +0800196 init_waitqueue_head(&res->sync_locking);
197 res->sync_locking_done = false;
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -0500198 res->ls = cinfo->lockspace;
199 res->mddev = mddev;
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -0500200 res->mode = DLM_LOCK_IV;
Goldwyn Rodrigues47741b72014-03-07 13:49:26 -0600201 namelen = strlen(name);
202 res->name = kzalloc(namelen + 1, GFP_KERNEL);
203 if (!res->name) {
204 pr_err("md-cluster: Unable to allocate resource name for resource %s\n", name);
205 goto out_err;
206 }
207 strlcpy(res->name, name, namelen + 1);
208 if (with_lvb) {
209 res->lksb.sb_lvbptr = kzalloc(LVB_SIZE, GFP_KERNEL);
210 if (!res->lksb.sb_lvbptr) {
211 pr_err("md-cluster: Unable to allocate LVB for resource %s\n", name);
212 goto out_err;
213 }
214 res->flags = DLM_LKF_VALBLK;
215 }
216
217 if (bastfn)
218 res->bast = bastfn;
219
220 res->flags |= DLM_LKF_EXPEDITE;
221
222 ret = dlm_lock_sync(res, DLM_LOCK_NL);
223 if (ret) {
224 pr_err("md-cluster: Unable to lock NL on new lock resource %s\n", name);
225 goto out_err;
226 }
227 res->flags &= ~DLM_LKF_EXPEDITE;
228 res->flags |= DLM_LKF_CONVERT;
229
230 return res;
231out_err:
232 kfree(res->lksb.sb_lvbptr);
233 kfree(res->name);
234 kfree(res);
235 return NULL;
236}
237
238static void lockres_free(struct dlm_lock_resource *res)
239{
Guoqing Jiang400cb452016-08-12 13:42:35 +0800240 int ret = 0;
Guoqing Jiangb5ef5672015-07-10 17:01:17 +0800241
Goldwyn Rodrigues47741b72014-03-07 13:49:26 -0600242 if (!res)
243 return;
244
Guoqing Jiang400cb452016-08-12 13:42:35 +0800245 /*
246 * use FORCEUNLOCK flag, so we can unlock even the lock is on the
247 * waiting or convert queue
248 */
249 ret = dlm_unlock(res->ls, res->lksb.sb_lkid, DLM_LKF_FORCEUNLOCK,
250 &res->lksb, res);
251 if (unlikely(ret != 0))
252 pr_err("failed to unlock %s return %d\n", res->name, ret);
253 else
Guoqing Jiangfccb60a2016-08-12 13:42:41 +0800254 wait_event(res->sync_locking, res->sync_locking_done);
Goldwyn Rodrigues47741b72014-03-07 13:49:26 -0600255
256 kfree(res->name);
257 kfree(res->lksb.sb_lvbptr);
258 kfree(res);
259}
Goldwyn Rodrigues8e854e92014-03-07 11:21:15 -0600260
NeilBrown30661b42015-10-19 15:44:00 +1100261static void add_resync_info(struct dlm_lock_resource *lockres,
262 sector_t lo, sector_t hi)
Goldwyn Rodrigues96ae9232014-06-06 12:35:34 -0500263{
264 struct resync_info *ri;
265
266 ri = (struct resync_info *)lockres->lksb.sb_lvbptr;
267 ri->lo = cpu_to_le64(lo);
268 ri->hi = cpu_to_le64(hi);
269}
270
271static struct suspend_info *read_resync_info(struct mddev *mddev, struct dlm_lock_resource *lockres)
272{
273 struct resync_info ri;
274 struct suspend_info *s = NULL;
275 sector_t hi = 0;
276
277 dlm_lock_sync(lockres, DLM_LOCK_CR);
278 memcpy(&ri, lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
279 hi = le64_to_cpu(ri.hi);
Guoqing Jiangcf97a342015-10-16 15:40:22 +0800280 if (hi > 0) {
Goldwyn Rodrigues96ae9232014-06-06 12:35:34 -0500281 s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
282 if (!s)
283 goto out;
284 s->hi = hi;
285 s->lo = le64_to_cpu(ri.lo);
286 }
287 dlm_unlock_sync(lockres);
288out:
289 return s;
290}
291
kbuild test robot6dc69c92015-02-28 07:04:37 +0800292static void recover_bitmaps(struct md_thread *thread)
Goldwyn Rodriguese94987d2014-06-07 00:45:22 -0500293{
294 struct mddev *mddev = thread->mddev;
295 struct md_cluster_info *cinfo = mddev->cluster_info;
296 struct dlm_lock_resource *bm_lockres;
297 char str[64];
298 int slot, ret;
299 struct suspend_info *s, *tmp;
300 sector_t lo, hi;
301
302 while (cinfo->recovery_map) {
303 slot = fls64((u64)cinfo->recovery_map) - 1;
304
305 /* Clear suspend_area associated with the bitmap */
306 spin_lock_irq(&cinfo->suspend_lock);
307 list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
308 if (slot == s->slot) {
309 list_del(&s->list);
310 kfree(s);
311 }
312 spin_unlock_irq(&cinfo->suspend_lock);
313
314 snprintf(str, 64, "bitmap%04d", slot);
315 bm_lockres = lockres_init(mddev, str, NULL, 1);
316 if (!bm_lockres) {
317 pr_err("md-cluster: Cannot initialize bitmaps\n");
318 goto clear_bit;
319 }
320
Guoqing Jiang7bcda712016-08-12 13:42:42 +0800321 ret = dlm_lock_sync_interruptible(bm_lockres, DLM_LOCK_PW, mddev);
Goldwyn Rodriguese94987d2014-06-07 00:45:22 -0500322 if (ret) {
323 pr_err("md-cluster: Could not DLM lock %s: %d\n",
324 str, ret);
325 goto clear_bit;
326 }
Goldwyn Rodrigues97f6cd32015-04-14 10:45:42 -0500327 ret = bitmap_copy_from_slot(mddev, slot, &lo, &hi, true);
Goldwyn Rodrigues4b26a082014-06-07 00:52:29 -0500328 if (ret) {
Goldwyn Rodriguese94987d2014-06-07 00:45:22 -0500329 pr_err("md-cluster: Could not copy data from bitmap %d\n", slot);
Guoqing Jiange3f924d2016-08-12 13:42:36 +0800330 goto clear_bit;
Goldwyn Rodrigues4b26a082014-06-07 00:52:29 -0500331 }
332 if (hi > 0) {
Goldwyn Rodrigues4b26a082014-06-07 00:52:29 -0500333 if (lo < mddev->recovery_cp)
334 mddev->recovery_cp = lo;
Guoqing Jiangeb315cd2016-05-02 11:33:10 -0400335 /* wake up thread to continue resync in case resync
336 * is not finished */
337 if (mddev->recovery_cp != MaxSector) {
338 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
339 md_wakeup_thread(mddev->thread);
340 }
Goldwyn Rodrigues4b26a082014-06-07 00:52:29 -0500341 }
Goldwyn Rodriguese94987d2014-06-07 00:45:22 -0500342clear_bit:
Shaohua Li4ac7a652016-01-22 15:54:42 -0800343 lockres_free(bm_lockres);
Goldwyn Rodriguese94987d2014-06-07 00:45:22 -0500344 clear_bit(slot, &cinfo->recovery_map);
345 }
346}
347
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500348static void recover_prep(void *arg)
349{
Goldwyn Rodrigues90382ed2015-06-24 09:30:32 -0500350 struct mddev *mddev = arg;
351 struct md_cluster_info *cinfo = mddev->cluster_info;
352 set_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500353}
354
Guoqing Jiang05cd0e52015-07-10 16:54:03 +0800355static void __recover_slot(struct mddev *mddev, int slot)
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500356{
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500357 struct md_cluster_info *cinfo = mddev->cluster_info;
358
Guoqing Jiang05cd0e52015-07-10 16:54:03 +0800359 set_bit(slot, &cinfo->recovery_map);
Goldwyn Rodriguese94987d2014-06-07 00:45:22 -0500360 if (!cinfo->recovery_thread) {
361 cinfo->recovery_thread = md_register_thread(recover_bitmaps,
362 mddev, "recover");
363 if (!cinfo->recovery_thread) {
364 pr_warn("md-cluster: Could not create recovery thread\n");
365 return;
366 }
367 }
368 md_wakeup_thread(cinfo->recovery_thread);
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500369}
370
Guoqing Jiang05cd0e52015-07-10 16:54:03 +0800371static void recover_slot(void *arg, struct dlm_slot *slot)
372{
373 struct mddev *mddev = arg;
374 struct md_cluster_info *cinfo = mddev->cluster_info;
375
376 pr_info("md-cluster: %s Node %d/%d down. My slot: %d. Initiating recovery.\n",
377 mddev->bitmap_info.cluster_name,
378 slot->nodeid, slot->slot,
379 cinfo->slot_number);
380 /* deduct one since dlm slot starts from one while the num of
381 * cluster-md begins with 0 */
382 __recover_slot(mddev, slot->slot - 1);
383}
384
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500385static void recover_done(void *arg, struct dlm_slot *slots,
386 int num_slots, int our_slot,
387 uint32_t generation)
388{
389 struct mddev *mddev = arg;
390 struct md_cluster_info *cinfo = mddev->cluster_info;
391
392 cinfo->slot_number = our_slot;
Guoqing Jiangeece0752015-07-10 17:01:21 +0800393 /* completion is only need to be complete when node join cluster,
394 * it doesn't need to run during another node's failure */
395 if (test_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state)) {
396 complete(&cinfo->completion);
397 clear_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
398 }
Goldwyn Rodrigues90382ed2015-06-24 09:30:32 -0500399 clear_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state);
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500400}
401
Guoqing Jiangeece0752015-07-10 17:01:21 +0800402/* the ops is called when node join the cluster, and do lock recovery
403 * if node failure occurs */
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500404static const struct dlm_lockspace_ops md_ls_ops = {
405 .recover_prep = recover_prep,
406 .recover_slot = recover_slot,
407 .recover_done = recover_done,
408};
409
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500410/*
411 * The BAST function for the ack lock resource
412 * This function wakes up the receive thread in
413 * order to receive and process the message.
414 */
415static void ack_bast(void *arg, int mode)
416{
NeilBrown2e2a7cd2015-10-19 15:42:18 +1100417 struct dlm_lock_resource *res = arg;
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500418 struct md_cluster_info *cinfo = res->mddev->cluster_info;
419
Guoqing Jiang51e453a2016-05-04 02:17:09 -0400420 if (mode == DLM_LOCK_EX) {
421 if (test_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state))
422 md_wakeup_thread(cinfo->recv_thread);
423 else
424 set_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state);
425 }
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500426}
427
Goldwyn Rodriguese59721c2014-06-07 02:30:30 -0500428static void __remove_suspend_info(struct md_cluster_info *cinfo, int slot)
429{
430 struct suspend_info *s, *tmp;
431
432 list_for_each_entry_safe(s, tmp, &cinfo->suspend_list, list)
433 if (slot == s->slot) {
Goldwyn Rodriguese59721c2014-06-07 02:30:30 -0500434 list_del(&s->list);
435 kfree(s);
436 break;
437 }
438}
439
Goldwyn Rodriguesb8ca8462015-10-09 11:27:01 -0500440static void remove_suspend_info(struct mddev *mddev, int slot)
Goldwyn Rodriguese59721c2014-06-07 02:30:30 -0500441{
Goldwyn Rodriguesb8ca8462015-10-09 11:27:01 -0500442 struct md_cluster_info *cinfo = mddev->cluster_info;
Goldwyn Rodriguese59721c2014-06-07 02:30:30 -0500443 spin_lock_irq(&cinfo->suspend_lock);
444 __remove_suspend_info(cinfo, slot);
445 spin_unlock_irq(&cinfo->suspend_lock);
Goldwyn Rodriguesb8ca8462015-10-09 11:27:01 -0500446 mddev->pers->quiesce(mddev, 2);
Goldwyn Rodriguese59721c2014-06-07 02:30:30 -0500447}
448
449
Goldwyn Rodrigues9ed38ff52015-08-14 12:19:40 -0500450static void process_suspend_info(struct mddev *mddev,
Goldwyn Rodriguese59721c2014-06-07 02:30:30 -0500451 int slot, sector_t lo, sector_t hi)
452{
Goldwyn Rodrigues9ed38ff52015-08-14 12:19:40 -0500453 struct md_cluster_info *cinfo = mddev->cluster_info;
Goldwyn Rodriguese59721c2014-06-07 02:30:30 -0500454 struct suspend_info *s;
455
456 if (!hi) {
Goldwyn Rodriguesb8ca8462015-10-09 11:27:01 -0500457 remove_suspend_info(mddev, slot);
Goldwyn Rodriguesc186b122015-09-30 13:20:35 -0500458 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
459 md_wakeup_thread(mddev->thread);
Goldwyn Rodriguese59721c2014-06-07 02:30:30 -0500460 return;
461 }
Guoqing Jiang18c9ff72016-05-02 11:50:12 -0400462
463 /*
464 * The bitmaps are not same for different nodes
465 * if RESYNCING is happening in one node, then
466 * the node which received the RESYNCING message
467 * probably will perform resync with the region
468 * [lo, hi] again, so we could reduce resync time
469 * a lot if we can ensure that the bitmaps among
470 * different nodes are match up well.
471 *
472 * sync_low/hi is used to record the region which
473 * arrived in the previous RESYNCING message,
474 *
475 * Call bitmap_sync_with_cluster to clear
476 * NEEDED_MASK and set RESYNC_MASK since
477 * resync thread is running in another node,
478 * so we don't need to do the resync again
479 * with the same section */
480 bitmap_sync_with_cluster(mddev, cinfo->sync_low,
481 cinfo->sync_hi,
482 lo, hi);
483 cinfo->sync_low = lo;
484 cinfo->sync_hi = hi;
485
Goldwyn Rodriguese59721c2014-06-07 02:30:30 -0500486 s = kzalloc(sizeof(struct suspend_info), GFP_KERNEL);
487 if (!s)
488 return;
489 s->slot = slot;
490 s->lo = lo;
491 s->hi = hi;
Goldwyn Rodrigues9ed38ff52015-08-14 12:19:40 -0500492 mddev->pers->quiesce(mddev, 1);
493 mddev->pers->quiesce(mddev, 0);
Goldwyn Rodriguese59721c2014-06-07 02:30:30 -0500494 spin_lock_irq(&cinfo->suspend_lock);
495 /* Remove existing entry (if exists) before adding */
496 __remove_suspend_info(cinfo, slot);
497 list_add(&s->list, &cinfo->suspend_list);
498 spin_unlock_irq(&cinfo->suspend_lock);
Goldwyn Rodriguesb8ca8462015-10-09 11:27:01 -0500499 mddev->pers->quiesce(mddev, 2);
Goldwyn Rodriguese59721c2014-06-07 02:30:30 -0500500}
501
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500502static void process_add_new_disk(struct mddev *mddev, struct cluster_msg *cmsg)
503{
504 char disk_uuid[64];
505 struct md_cluster_info *cinfo = mddev->cluster_info;
506 char event_name[] = "EVENT=ADD_DEVICE";
507 char raid_slot[16];
508 char *envp[] = {event_name, disk_uuid, raid_slot, NULL};
509 int len;
510
511 len = snprintf(disk_uuid, 64, "DEVICE_UUID=");
Guoqing Jiangb89f7042015-07-10 16:54:02 +0800512 sprintf(disk_uuid + len, "%pU", cmsg->uuid);
Guoqing Jiangfaeff832015-10-12 17:21:21 +0800513 snprintf(raid_slot, 16, "RAID_DISK=%d", le32_to_cpu(cmsg->raid_slot));
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500514 pr_info("%s:%d Sending kobject change with %s and %s\n", __func__, __LINE__, disk_uuid, raid_slot);
515 init_completion(&cinfo->newdisk_completion);
Goldwyn Rodriguesfa8259d2015-03-02 10:55:49 -0600516 set_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500517 kobject_uevent_env(&disk_to_dev(mddev->gendisk)->kobj, KOBJ_CHANGE, envp);
518 wait_for_completion_timeout(&cinfo->newdisk_completion,
519 NEW_DEV_TIMEOUT);
Goldwyn Rodriguesfa8259d2015-03-02 10:55:49 -0600520 clear_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state);
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500521}
522
523
524static void process_metadata_update(struct mddev *mddev, struct cluster_msg *msg)
525{
526 struct md_cluster_info *cinfo = mddev->cluster_info;
Guoqing Jiang15858fa2015-12-21 10:51:00 +1100527 mddev->good_device_nr = le32_to_cpu(msg->raid_slot);
528 set_bit(MD_RELOAD_SB, &mddev->flags);
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500529 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
Guoqing Jiang15858fa2015-12-21 10:51:00 +1100530 md_wakeup_thread(mddev->thread);
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500531}
532
Goldwyn Rodrigues88bcfef2015-04-14 10:44:44 -0500533static void process_remove_disk(struct mddev *mddev, struct cluster_msg *msg)
534{
Guoqing Jiang5f0aa212016-08-12 13:42:39 +0800535 struct md_rdev *rdev;
Goldwyn Rodrigues88bcfef2015-04-14 10:44:44 -0500536
Guoqing Jiang5f0aa212016-08-12 13:42:39 +0800537 rcu_read_lock();
538 rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
Guoqing Jiang659b2542015-12-21 10:50:59 +1100539 if (rdev) {
540 set_bit(ClusterRemove, &rdev->flags);
541 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
542 md_wakeup_thread(mddev->thread);
543 }
Goldwyn Rodrigues88bcfef2015-04-14 10:44:44 -0500544 else
Guoqing Jiangfaeff832015-10-12 17:21:21 +0800545 pr_warn("%s: %d Could not find disk(%d) to REMOVE\n",
546 __func__, __LINE__, le32_to_cpu(msg->raid_slot));
Guoqing Jiang5f0aa212016-08-12 13:42:39 +0800547 rcu_read_unlock();
Goldwyn Rodrigues88bcfef2015-04-14 10:44:44 -0500548}
549
Goldwyn Rodrigues97f6cd32015-04-14 10:45:42 -0500550static void process_readd_disk(struct mddev *mddev, struct cluster_msg *msg)
551{
Guoqing Jiang5f0aa212016-08-12 13:42:39 +0800552 struct md_rdev *rdev;
Goldwyn Rodrigues97f6cd32015-04-14 10:45:42 -0500553
Guoqing Jiang5f0aa212016-08-12 13:42:39 +0800554 rcu_read_lock();
555 rdev = md_find_rdev_nr_rcu(mddev, le32_to_cpu(msg->raid_slot));
Goldwyn Rodrigues97f6cd32015-04-14 10:45:42 -0500556 if (rdev && test_bit(Faulty, &rdev->flags))
557 clear_bit(Faulty, &rdev->flags);
558 else
Guoqing Jiangfaeff832015-10-12 17:21:21 +0800559 pr_warn("%s: %d Could not find disk(%d) which is faulty",
560 __func__, __LINE__, le32_to_cpu(msg->raid_slot));
Guoqing Jiang5f0aa212016-08-12 13:42:39 +0800561 rcu_read_unlock();
Goldwyn Rodrigues97f6cd32015-04-14 10:45:42 -0500562}
563
Guoqing Jiang1fa9a1a2016-05-03 22:22:15 -0400564static int process_recvd_msg(struct mddev *mddev, struct cluster_msg *msg)
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500565{
Guoqing Jiang1fa9a1a2016-05-03 22:22:15 -0400566 int ret = 0;
567
Guoqing Jiang256f5b22015-10-12 17:21:23 +0800568 if (WARN(mddev->cluster_info->slot_number - 1 == le32_to_cpu(msg->slot),
569 "node %d received it's own msg\n", le32_to_cpu(msg->slot)))
Guoqing Jiang1fa9a1a2016-05-03 22:22:15 -0400570 return -1;
Guoqing Jiangcf97a342015-10-16 15:40:22 +0800571 switch (le32_to_cpu(msg->type)) {
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500572 case METADATA_UPDATED:
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500573 process_metadata_update(mddev, msg);
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500574 break;
575 case RESYNCING:
Guoqing Jiangcf97a342015-10-16 15:40:22 +0800576 process_suspend_info(mddev, le32_to_cpu(msg->slot),
577 le64_to_cpu(msg->low),
578 le64_to_cpu(msg->high));
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500579 break;
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500580 case NEWDISK:
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500581 process_add_new_disk(mddev, msg);
Goldwyn Rodrigues88bcfef2015-04-14 10:44:44 -0500582 break;
583 case REMOVE:
Goldwyn Rodrigues88bcfef2015-04-14 10:44:44 -0500584 process_remove_disk(mddev, msg);
585 break;
Goldwyn Rodrigues97f6cd32015-04-14 10:45:42 -0500586 case RE_ADD:
Goldwyn Rodrigues97f6cd32015-04-14 10:45:42 -0500587 process_readd_disk(mddev, msg);
588 break;
Guoqing Jiangdc737d72015-07-10 16:54:04 +0800589 case BITMAP_NEEDS_SYNC:
Guoqing Jiangcf97a342015-10-16 15:40:22 +0800590 __recover_slot(mddev, le32_to_cpu(msg->slot));
Guoqing Jiangdc737d72015-07-10 16:54:04 +0800591 break;
Goldwyn Rodrigues88bcfef2015-04-14 10:44:44 -0500592 default:
Guoqing Jiang1fa9a1a2016-05-03 22:22:15 -0400593 ret = -1;
Goldwyn Rodrigues88bcfef2015-04-14 10:44:44 -0500594 pr_warn("%s:%d Received unknown message from %d\n",
595 __func__, __LINE__, msg->slot);
kbuild test robot09dd1af2015-02-28 09:16:08 +0800596 }
Guoqing Jiang1fa9a1a2016-05-03 22:22:15 -0400597 return ret;
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500598}
599
600/*
601 * thread for receiving message
602 */
603static void recv_daemon(struct md_thread *thread)
604{
605 struct md_cluster_info *cinfo = thread->mddev->cluster_info;
606 struct dlm_lock_resource *ack_lockres = cinfo->ack_lockres;
607 struct dlm_lock_resource *message_lockres = cinfo->message_lockres;
608 struct cluster_msg msg;
Guoqing Jiangb5ef5672015-07-10 17:01:17 +0800609 int ret;
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500610
Guoqing Jiang8b9277c2015-12-21 10:51:00 +1100611 mutex_lock(&cinfo->recv_mutex);
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500612 /*get CR on Message*/
613 if (dlm_lock_sync(message_lockres, DLM_LOCK_CR)) {
614 pr_err("md/raid1:failed to get CR on MESSAGE\n");
Guoqing Jiang8b9277c2015-12-21 10:51:00 +1100615 mutex_unlock(&cinfo->recv_mutex);
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500616 return;
617 }
618
619 /* read lvb and wake up thread to process this message_lockres */
620 memcpy(&msg, message_lockres->lksb.sb_lvbptr, sizeof(struct cluster_msg));
Guoqing Jiang1fa9a1a2016-05-03 22:22:15 -0400621 ret = process_recvd_msg(thread->mddev, &msg);
622 if (ret)
623 goto out;
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500624
625 /*release CR on ack_lockres*/
Guoqing Jiangb5ef5672015-07-10 17:01:17 +0800626 ret = dlm_unlock_sync(ack_lockres);
627 if (unlikely(ret != 0))
628 pr_info("unlock ack failed return %d\n", ret);
Guoqing Jiang66099bb2015-07-10 17:01:15 +0800629 /*up-convert to PR on message_lockres*/
Guoqing Jiangb5ef5672015-07-10 17:01:17 +0800630 ret = dlm_lock_sync(message_lockres, DLM_LOCK_PR);
631 if (unlikely(ret != 0))
632 pr_info("lock PR on msg failed return %d\n", ret);
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500633 /*get CR on ack_lockres again*/
Guoqing Jiangb5ef5672015-07-10 17:01:17 +0800634 ret = dlm_lock_sync(ack_lockres, DLM_LOCK_CR);
635 if (unlikely(ret != 0))
636 pr_info("lock CR on ack failed return %d\n", ret);
Guoqing Jiang1fa9a1a2016-05-03 22:22:15 -0400637out:
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500638 /*release CR on message_lockres*/
Guoqing Jiangb5ef5672015-07-10 17:01:17 +0800639 ret = dlm_unlock_sync(message_lockres);
640 if (unlikely(ret != 0))
641 pr_info("unlock msg failed return %d\n", ret);
Guoqing Jiang8b9277c2015-12-21 10:51:00 +1100642 mutex_unlock(&cinfo->recv_mutex);
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500643}
644
Guoqing Jiang8b9277c2015-12-21 10:51:00 +1100645/* lock_token()
Goldwyn Rodrigues601b5152014-06-07 01:28:53 -0500646 * Takes the lock on the TOKEN lock resource so no other
647 * node can communicate while the operation is underway.
648 */
Guoqing Jiang8b9277c2015-12-21 10:51:00 +1100649static int lock_token(struct md_cluster_info *cinfo)
Goldwyn Rodrigues601b5152014-06-07 01:28:53 -0500650{
651 int error;
652
653 error = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
654 if (error)
655 pr_err("md-cluster(%s:%d): failed to get EX on TOKEN (%d)\n",
656 __func__, __LINE__, error);
Guoqing Jiang8b9277c2015-12-21 10:51:00 +1100657
658 /* Lock the receive sequence */
659 mutex_lock(&cinfo->recv_mutex);
Goldwyn Rodrigues601b5152014-06-07 01:28:53 -0500660 return error;
661}
662
Guoqing Jiang8b9277c2015-12-21 10:51:00 +1100663/* lock_comm()
664 * Sets the MD_CLUSTER_SEND_LOCK bit to lock the send channel.
665 */
666static int lock_comm(struct md_cluster_info *cinfo)
667{
668 wait_event(cinfo->wait,
669 !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state));
670
671 return lock_token(cinfo);
672}
673
Goldwyn Rodrigues601b5152014-06-07 01:28:53 -0500674static void unlock_comm(struct md_cluster_info *cinfo)
675{
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -0500676 WARN_ON(cinfo->token_lockres->mode != DLM_LOCK_EX);
Guoqing Jiang8b9277c2015-12-21 10:51:00 +1100677 mutex_unlock(&cinfo->recv_mutex);
Goldwyn Rodrigues601b5152014-06-07 01:28:53 -0500678 dlm_unlock_sync(cinfo->token_lockres);
Guoqing Jiang8b9277c2015-12-21 10:51:00 +1100679 clear_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state);
680 wake_up(&cinfo->wait);
Goldwyn Rodrigues601b5152014-06-07 01:28:53 -0500681}
682
683/* __sendmsg()
684 * This function performs the actual sending of the message. This function is
685 * usually called after performing the encompassing operation
686 * The function:
687 * 1. Grabs the message lockresource in EX mode
688 * 2. Copies the message to the message LVB
Guoqing Jiang66099bb2015-07-10 17:01:15 +0800689 * 3. Downconverts message lockresource to CW
Goldwyn Rodrigues601b5152014-06-07 01:28:53 -0500690 * 4. Upconverts ack lock resource from CR to EX. This forces the BAST on other nodes
691 * and the other nodes read the message. The thread will wait here until all other
692 * nodes have released ack lock resource.
693 * 5. Downconvert ack lockresource to CR
694 */
695static int __sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
696{
697 int error;
698 int slot = cinfo->slot_number - 1;
699
700 cmsg->slot = cpu_to_le32(slot);
701 /*get EX on Message*/
702 error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_EX);
703 if (error) {
704 pr_err("md-cluster: failed to get EX on MESSAGE (%d)\n", error);
705 goto failed_message;
706 }
707
708 memcpy(cinfo->message_lockres->lksb.sb_lvbptr, (void *)cmsg,
709 sizeof(struct cluster_msg));
Guoqing Jiang66099bb2015-07-10 17:01:15 +0800710 /*down-convert EX to CW on Message*/
711 error = dlm_lock_sync(cinfo->message_lockres, DLM_LOCK_CW);
Goldwyn Rodrigues601b5152014-06-07 01:28:53 -0500712 if (error) {
Guoqing Jiang66099bb2015-07-10 17:01:15 +0800713 pr_err("md-cluster: failed to convert EX to CW on MESSAGE(%d)\n",
Goldwyn Rodrigues601b5152014-06-07 01:28:53 -0500714 error);
Guoqing Jiang66099bb2015-07-10 17:01:15 +0800715 goto failed_ack;
Goldwyn Rodrigues601b5152014-06-07 01:28:53 -0500716 }
717
718 /*up-convert CR to EX on Ack*/
719 error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_EX);
720 if (error) {
721 pr_err("md-cluster: failed to convert CR to EX on ACK(%d)\n",
722 error);
723 goto failed_ack;
724 }
725
726 /*down-convert EX to CR on Ack*/
727 error = dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR);
728 if (error) {
729 pr_err("md-cluster: failed to convert EX to CR on ACK(%d)\n",
730 error);
731 goto failed_ack;
732 }
733
734failed_ack:
Guoqing Jiangb5ef5672015-07-10 17:01:17 +0800735 error = dlm_unlock_sync(cinfo->message_lockres);
736 if (unlikely(error != 0)) {
737 pr_err("md-cluster: failed convert to NL on MESSAGE(%d)\n",
738 error);
739 /* in case the message can't be released due to some reason */
740 goto failed_ack;
741 }
Goldwyn Rodrigues601b5152014-06-07 01:28:53 -0500742failed_message:
743 return error;
744}
745
746static int sendmsg(struct md_cluster_info *cinfo, struct cluster_msg *cmsg)
747{
748 int ret;
749
750 lock_comm(cinfo);
751 ret = __sendmsg(cinfo, cmsg);
752 unlock_comm(cinfo);
753 return ret;
754}
755
Goldwyn Rodrigues96ae9232014-06-06 12:35:34 -0500756static int gather_all_resync_info(struct mddev *mddev, int total_slots)
757{
758 struct md_cluster_info *cinfo = mddev->cluster_info;
759 int i, ret = 0;
760 struct dlm_lock_resource *bm_lockres;
761 struct suspend_info *s;
762 char str[64];
Guoqing Jiangabb9b222015-07-10 17:01:22 +0800763 sector_t lo, hi;
Goldwyn Rodrigues96ae9232014-06-06 12:35:34 -0500764
765
766 for (i = 0; i < total_slots; i++) {
767 memset(str, '\0', 64);
768 snprintf(str, 64, "bitmap%04d", i);
769 bm_lockres = lockres_init(mddev, str, NULL, 1);
770 if (!bm_lockres)
771 return -ENOMEM;
Shaohua Li4ac7a652016-01-22 15:54:42 -0800772 if (i == (cinfo->slot_number - 1)) {
773 lockres_free(bm_lockres);
Goldwyn Rodrigues96ae9232014-06-06 12:35:34 -0500774 continue;
Shaohua Li4ac7a652016-01-22 15:54:42 -0800775 }
Goldwyn Rodrigues96ae9232014-06-06 12:35:34 -0500776
777 bm_lockres->flags |= DLM_LKF_NOQUEUE;
778 ret = dlm_lock_sync(bm_lockres, DLM_LOCK_PW);
779 if (ret == -EAGAIN) {
780 memset(bm_lockres->lksb.sb_lvbptr, '\0', LVB_SIZE);
781 s = read_resync_info(mddev, bm_lockres);
782 if (s) {
783 pr_info("%s:%d Resync[%llu..%llu] in progress on %d\n",
784 __func__, __LINE__,
785 (unsigned long long) s->lo,
786 (unsigned long long) s->hi, i);
787 spin_lock_irq(&cinfo->suspend_lock);
788 s->slot = i;
789 list_add(&s->list, &cinfo->suspend_list);
790 spin_unlock_irq(&cinfo->suspend_lock);
791 }
792 ret = 0;
793 lockres_free(bm_lockres);
794 continue;
795 }
Guoqing Jiang6e6d9f22015-07-10 17:01:20 +0800796 if (ret) {
797 lockres_free(bm_lockres);
Goldwyn Rodrigues96ae9232014-06-06 12:35:34 -0500798 goto out;
Guoqing Jiang6e6d9f22015-07-10 17:01:20 +0800799 }
Guoqing Jiangabb9b222015-07-10 17:01:22 +0800800
801 /* Read the disk bitmap sb and check if it needs recovery */
802 ret = bitmap_copy_from_slot(mddev, i, &lo, &hi, false);
803 if (ret) {
804 pr_warn("md-cluster: Could not gather bitmaps from slot %d", i);
805 lockres_free(bm_lockres);
806 continue;
807 }
808 if ((hi > 0) && (lo < mddev->recovery_cp)) {
809 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
810 mddev->recovery_cp = lo;
811 md_check_recovery(mddev);
812 }
813
Goldwyn Rodrigues96ae9232014-06-06 12:35:34 -0500814 lockres_free(bm_lockres);
815 }
816out:
817 return ret;
818}
819
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -0500820static int join(struct mddev *mddev, int nodes)
821{
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -0500822 struct md_cluster_info *cinfo;
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500823 int ret, ops_rv;
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -0500824 char str[64];
825
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -0500826 cinfo = kzalloc(sizeof(struct md_cluster_info), GFP_KERNEL);
827 if (!cinfo)
828 return -ENOMEM;
829
Guoqing Jiang9e3072e2015-07-10 17:01:18 +0800830 INIT_LIST_HEAD(&cinfo->suspend_list);
831 spin_lock_init(&cinfo->suspend_lock);
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500832 init_completion(&cinfo->completion);
Guoqing Jiangeece0752015-07-10 17:01:21 +0800833 set_bit(MD_CLUSTER_BEGIN_JOIN_CLUSTER, &cinfo->state);
Guoqing Jiang8b9277c2015-12-21 10:51:00 +1100834 init_waitqueue_head(&cinfo->wait);
835 mutex_init(&cinfo->recv_mutex);
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500836
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500837 mddev->cluster_info = cinfo;
838
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -0500839 memset(str, 0, 64);
Guoqing Jiangb89f7042015-07-10 16:54:02 +0800840 sprintf(str, "%pU", mddev->uuid);
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500841 ret = dlm_new_lockspace(str, mddev->bitmap_info.cluster_name,
842 DLM_LSFL_FS, LVB_SIZE,
843 &md_ls_ops, mddev, &ops_rv, &cinfo->lockspace);
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -0500844 if (ret)
845 goto err;
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500846 wait_for_completion(&cinfo->completion);
Guoqing Jiang8c58f022015-04-21 11:25:52 -0500847 if (nodes < cinfo->slot_number) {
848 pr_err("md-cluster: Slot allotted(%d) is greater than available slots(%d).",
849 cinfo->slot_number, nodes);
Goldwyn Rodriguesb97e92572014-06-06 11:50:56 -0500850 ret = -ERANGE;
851 goto err;
852 }
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500853 /* Initiate the communication resources */
854 ret = -ENOMEM;
855 cinfo->recv_thread = md_register_thread(recv_daemon, mddev, "cluster_recv");
856 if (!cinfo->recv_thread) {
857 pr_err("md-cluster: cannot allocate memory for recv_thread!\n");
858 goto err;
859 }
860 cinfo->message_lockres = lockres_init(mddev, "message", NULL, 1);
861 if (!cinfo->message_lockres)
862 goto err;
863 cinfo->token_lockres = lockres_init(mddev, "token", NULL, 0);
864 if (!cinfo->token_lockres)
865 goto err;
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500866 cinfo->no_new_dev_lockres = lockres_init(mddev, "no-new-dev", NULL, 0);
867 if (!cinfo->no_new_dev_lockres)
868 goto err;
869
Guoqing Jiang15352122016-05-02 11:33:12 -0400870 ret = dlm_lock_sync(cinfo->token_lockres, DLM_LOCK_EX);
871 if (ret) {
872 ret = -EAGAIN;
873 pr_err("md-cluster: can't join cluster to avoid lock issue\n");
874 goto err;
875 }
876 cinfo->ack_lockres = lockres_init(mddev, "ack", ack_bast, 0);
Wei Yongjun0f6187d2016-08-21 14:42:25 +0000877 if (!cinfo->ack_lockres) {
878 ret = -ENOMEM;
Guoqing Jiang15352122016-05-02 11:33:12 -0400879 goto err;
Wei Yongjun0f6187d2016-08-21 14:42:25 +0000880 }
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500881 /* get sync CR lock on ACK. */
882 if (dlm_lock_sync(cinfo->ack_lockres, DLM_LOCK_CR))
883 pr_err("md-cluster: failed to get a sync CR lock on ACK!(%d)\n",
884 ret);
Guoqing Jiang15352122016-05-02 11:33:12 -0400885 dlm_unlock_sync(cinfo->token_lockres);
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500886 /* get sync CR lock on no-new-dev. */
887 if (dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR))
888 pr_err("md-cluster: failed to get a sync CR lock on no-new-dev!(%d)\n", ret);
889
Goldwyn Rodrigues54519c52014-06-06 12:12:32 -0500890
891 pr_info("md-cluster: Joined cluster %s slot %d\n", str, cinfo->slot_number);
892 snprintf(str, 64, "bitmap%04d", cinfo->slot_number - 1);
893 cinfo->bitmap_lockres = lockres_init(mddev, str, NULL, 1);
Wei Yongjun0f6187d2016-08-21 14:42:25 +0000894 if (!cinfo->bitmap_lockres) {
895 ret = -ENOMEM;
Goldwyn Rodrigues54519c52014-06-06 12:12:32 -0500896 goto err;
Wei Yongjun0f6187d2016-08-21 14:42:25 +0000897 }
Goldwyn Rodrigues54519c52014-06-06 12:12:32 -0500898 if (dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW)) {
899 pr_err("Failed to get bitmap lock\n");
900 ret = -EINVAL;
901 goto err;
902 }
903
Goldwyn Rodriguesc186b122015-09-30 13:20:35 -0500904 cinfo->resync_lockres = lockres_init(mddev, "resync", NULL, 0);
Wei Yongjun0f6187d2016-08-21 14:42:25 +0000905 if (!cinfo->resync_lockres) {
906 ret = -ENOMEM;
Goldwyn Rodriguesc186b122015-09-30 13:20:35 -0500907 goto err;
Wei Yongjun0f6187d2016-08-21 14:42:25 +0000908 }
Goldwyn Rodriguesc186b122015-09-30 13:20:35 -0500909
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -0500910 return 0;
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -0500911err:
Guoqing Jiang5b0fb332016-05-02 11:33:11 -0400912 md_unregister_thread(&cinfo->recovery_thread);
913 md_unregister_thread(&cinfo->recv_thread);
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500914 lockres_free(cinfo->message_lockres);
915 lockres_free(cinfo->token_lockres);
916 lockres_free(cinfo->ack_lockres);
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500917 lockres_free(cinfo->no_new_dev_lockres);
Goldwyn Rodriguesc186b122015-09-30 13:20:35 -0500918 lockres_free(cinfo->resync_lockres);
Goldwyn Rodrigues96ae9232014-06-06 12:35:34 -0500919 lockres_free(cinfo->bitmap_lockres);
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -0500920 if (cinfo->lockspace)
921 dlm_release_lockspace(cinfo->lockspace, 2);
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500922 mddev->cluster_info = NULL;
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -0500923 kfree(cinfo);
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -0500924 return ret;
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -0500925}
926
Guoqing Jiang51e453a2016-05-04 02:17:09 -0400927static void load_bitmaps(struct mddev *mddev, int total_slots)
928{
929 struct md_cluster_info *cinfo = mddev->cluster_info;
930
931 /* load all the node's bitmap info for resync */
932 if (gather_all_resync_info(mddev, total_slots))
933 pr_err("md-cluster: failed to gather all resyn infos\n");
934 set_bit(MD_CLUSTER_ALREADY_IN_CLUSTER, &cinfo->state);
935 /* wake up recv thread in case something need to be handled */
936 if (test_and_clear_bit(MD_CLUSTER_PENDING_RECV_EVENT, &cinfo->state))
937 md_wakeup_thread(cinfo->recv_thread);
938}
939
Guoqing Jiang09995412015-10-01 00:09:18 +0800940static void resync_bitmap(struct mddev *mddev)
941{
942 struct md_cluster_info *cinfo = mddev->cluster_info;
943 struct cluster_msg cmsg = {0};
944 int err;
945
946 cmsg.type = cpu_to_le32(BITMAP_NEEDS_SYNC);
947 err = sendmsg(cinfo, &cmsg);
948 if (err)
949 pr_err("%s:%d: failed to send BITMAP_NEEDS_SYNC message (%d)\n",
950 __func__, __LINE__, err);
951}
952
Guoqing Jiangf6a2dc62015-12-21 10:51:00 +1100953static void unlock_all_bitmaps(struct mddev *mddev);
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -0500954static int leave(struct mddev *mddev)
955{
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -0500956 struct md_cluster_info *cinfo = mddev->cluster_info;
957
958 if (!cinfo)
959 return 0;
Guoqing Jiang09995412015-10-01 00:09:18 +0800960
961 /* BITMAP_NEEDS_SYNC message should be sent when node
962 * is leaving the cluster with dirty bitmap, also we
963 * can only deliver it when dlm connection is available */
964 if (cinfo->slot_number > 0 && mddev->recovery_cp != MaxSector)
965 resync_bitmap(mddev);
966
Goldwyn Rodriguese94987d2014-06-07 00:45:22 -0500967 md_unregister_thread(&cinfo->recovery_thread);
Goldwyn Rodrigues46646802014-06-07 01:08:29 -0500968 md_unregister_thread(&cinfo->recv_thread);
969 lockres_free(cinfo->message_lockres);
970 lockres_free(cinfo->token_lockres);
971 lockres_free(cinfo->ack_lockres);
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -0500972 lockres_free(cinfo->no_new_dev_lockres);
Shaohua Li4ac7a652016-01-22 15:54:42 -0800973 lockres_free(cinfo->resync_lockres);
Goldwyn Rodrigues54519c52014-06-06 12:12:32 -0500974 lockres_free(cinfo->bitmap_lockres);
Guoqing Jiangf6a2dc62015-12-21 10:51:00 +1100975 unlock_all_bitmaps(mddev);
Goldwyn Rodriguesc4ce8672014-03-29 10:20:02 -0500976 dlm_release_lockspace(cinfo->lockspace, 2);
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -0500977 return 0;
978}
979
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -0500980/* slot_number(): Returns the MD slot number to use
981 * DLM starts the slot numbers from 1, wheras cluster-md
982 * wants the number to be from zero, so we deduct one
983 */
984static int slot_number(struct mddev *mddev)
985{
986 struct md_cluster_info *cinfo = mddev->cluster_info;
987
988 return cinfo->slot_number - 1;
989}
990
Guoqing Jiang8b9277c2015-12-21 10:51:00 +1100991/*
992 * Check if the communication is already locked, else lock the communication
993 * channel.
994 * If it is already locked, token is in EX mode, and hence lock_token()
995 * should not be called.
996 */
Goldwyn Rodrigues293467a2014-06-07 01:44:51 -0500997static int metadata_update_start(struct mddev *mddev)
998{
Guoqing Jiang8b9277c2015-12-21 10:51:00 +1100999 struct md_cluster_info *cinfo = mddev->cluster_info;
1000
1001 wait_event(cinfo->wait,
1002 !test_and_set_bit(MD_CLUSTER_SEND_LOCK, &cinfo->state) ||
1003 test_and_clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state));
1004
1005 /* If token is already locked, return 0 */
1006 if (cinfo->token_lockres->mode == DLM_LOCK_EX)
1007 return 0;
1008
1009 return lock_token(cinfo);
Goldwyn Rodrigues293467a2014-06-07 01:44:51 -05001010}
1011
1012static int metadata_update_finish(struct mddev *mddev)
1013{
1014 struct md_cluster_info *cinfo = mddev->cluster_info;
1015 struct cluster_msg cmsg;
Goldwyn Rodrigues70bcecd2015-08-21 10:33:39 -05001016 struct md_rdev *rdev;
1017 int ret = 0;
NeilBrownba2746b2015-10-16 13:48:35 +11001018 int raid_slot = -1;
Goldwyn Rodrigues293467a2014-06-07 01:44:51 -05001019
1020 memset(&cmsg, 0, sizeof(cmsg));
1021 cmsg.type = cpu_to_le32(METADATA_UPDATED);
Goldwyn Rodrigues70bcecd2015-08-21 10:33:39 -05001022 /* Pick up a good active device number to send.
1023 */
1024 rdev_for_each(rdev, mddev)
1025 if (rdev->raid_disk > -1 && !test_bit(Faulty, &rdev->flags)) {
NeilBrownba2746b2015-10-16 13:48:35 +11001026 raid_slot = rdev->desc_nr;
Goldwyn Rodrigues70bcecd2015-08-21 10:33:39 -05001027 break;
1028 }
NeilBrownba2746b2015-10-16 13:48:35 +11001029 if (raid_slot >= 0) {
1030 cmsg.raid_slot = cpu_to_le32(raid_slot);
Goldwyn Rodrigues70bcecd2015-08-21 10:33:39 -05001031 ret = __sendmsg(cinfo, &cmsg);
NeilBrownba2746b2015-10-16 13:48:35 +11001032 } else
Goldwyn Rodrigues70bcecd2015-08-21 10:33:39 -05001033 pr_warn("md-cluster: No good device id found to send\n");
Guoqing Jiang8b9277c2015-12-21 10:51:00 +11001034 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
Goldwyn Rodrigues293467a2014-06-07 01:44:51 -05001035 unlock_comm(cinfo);
1036 return ret;
1037}
1038
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05001039static void metadata_update_cancel(struct mddev *mddev)
Goldwyn Rodrigues293467a2014-06-07 01:44:51 -05001040{
1041 struct md_cluster_info *cinfo = mddev->cluster_info;
Guoqing Jiang8b9277c2015-12-21 10:51:00 +11001042 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05001043 unlock_comm(cinfo);
Goldwyn Rodrigues293467a2014-06-07 01:44:51 -05001044}
1045
Goldwyn Rodriguesc186b122015-09-30 13:20:35 -05001046static int resync_start(struct mddev *mddev)
1047{
1048 struct md_cluster_info *cinfo = mddev->cluster_info;
Guoqing Jiangd6385db2016-08-12 13:42:43 +08001049 return dlm_lock_sync_interruptible(cinfo->resync_lockres, DLM_LOCK_EX, mddev);
Goldwyn Rodriguesc186b122015-09-30 13:20:35 -05001050}
1051
Goldwyn Rodriguesc40f3412015-08-19 08:14:42 +10001052static int resync_info_update(struct mddev *mddev, sector_t lo, sector_t hi)
Goldwyn Rodrigues965400e2014-06-07 02:16:58 -05001053{
1054 struct md_cluster_info *cinfo = mddev->cluster_info;
Goldwyn Rodriguesac277c62015-12-21 10:50:59 +11001055 struct resync_info ri;
Guoqing Jiangaee177a2015-10-12 17:21:24 +08001056 struct cluster_msg cmsg = {0};
Goldwyn Rodrigues965400e2014-06-07 02:16:58 -05001057
Goldwyn Rodriguesac277c62015-12-21 10:50:59 +11001058 /* do not send zero again, if we have sent before */
1059 if (hi == 0) {
1060 memcpy(&ri, cinfo->bitmap_lockres->lksb.sb_lvbptr, sizeof(struct resync_info));
1061 if (le64_to_cpu(ri.hi) == 0)
1062 return 0;
1063 }
1064
NeilBrown30661b42015-10-19 15:44:00 +11001065 add_resync_info(cinfo->bitmap_lockres, lo, hi);
Goldwyn Rodriguesc40f3412015-08-19 08:14:42 +10001066 /* Re-acquire the lock to refresh LVB */
1067 dlm_lock_sync(cinfo->bitmap_lockres, DLM_LOCK_PW);
Goldwyn Rodriguesc40f3412015-08-19 08:14:42 +10001068 cmsg.type = cpu_to_le32(RESYNCING);
Goldwyn Rodrigues965400e2014-06-07 02:16:58 -05001069 cmsg.low = cpu_to_le64(lo);
1070 cmsg.high = cpu_to_le64(hi);
Goldwyn Rodriguesc186b122015-09-30 13:20:35 -05001071
Goldwyn Rodrigues965400e2014-06-07 02:16:58 -05001072 return sendmsg(cinfo, &cmsg);
1073}
1074
Goldwyn Rodriguesc186b122015-09-30 13:20:35 -05001075static int resync_finish(struct mddev *mddev)
1076{
1077 struct md_cluster_info *cinfo = mddev->cluster_info;
Goldwyn Rodriguesc186b122015-09-30 13:20:35 -05001078 dlm_unlock_sync(cinfo->resync_lockres);
1079 return resync_info_update(mddev, 0, 0);
1080}
1081
Goldwyn Rodrigues90382ed2015-06-24 09:30:32 -05001082static int area_resyncing(struct mddev *mddev, int direction,
1083 sector_t lo, sector_t hi)
Goldwyn Rodrigues589a1c42014-06-07 02:39:37 -05001084{
1085 struct md_cluster_info *cinfo = mddev->cluster_info;
1086 int ret = 0;
1087 struct suspend_info *s;
1088
Goldwyn Rodrigues90382ed2015-06-24 09:30:32 -05001089 if ((direction == READ) &&
1090 test_bit(MD_CLUSTER_SUSPEND_READ_BALANCING, &cinfo->state))
1091 return 1;
1092
Goldwyn Rodrigues589a1c42014-06-07 02:39:37 -05001093 spin_lock_irq(&cinfo->suspend_lock);
1094 if (list_empty(&cinfo->suspend_list))
1095 goto out;
1096 list_for_each_entry(s, &cinfo->suspend_list, list)
1097 if (hi > s->lo && lo < s->hi) {
1098 ret = 1;
1099 break;
1100 }
1101out:
1102 spin_unlock_irq(&cinfo->suspend_lock);
1103 return ret;
1104}
1105
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05001106/* add_new_disk() - initiates a disk add
1107 * However, if this fails before writing md_update_sb(),
1108 * add_new_disk_cancel() must be called to release token lock
1109 */
1110static int add_new_disk(struct mddev *mddev, struct md_rdev *rdev)
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05001111{
1112 struct md_cluster_info *cinfo = mddev->cluster_info;
1113 struct cluster_msg cmsg;
1114 int ret = 0;
1115 struct mdp_superblock_1 *sb = page_address(rdev->sb_page);
1116 char *uuid = sb->device_uuid;
1117
1118 memset(&cmsg, 0, sizeof(cmsg));
1119 cmsg.type = cpu_to_le32(NEWDISK);
1120 memcpy(cmsg.uuid, uuid, 16);
Guoqing Jiangfaeff832015-10-12 17:21:21 +08001121 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05001122 lock_comm(cinfo);
1123 ret = __sendmsg(cinfo, &cmsg);
1124 if (ret)
1125 return ret;
1126 cinfo->no_new_dev_lockres->flags |= DLM_LKF_NOQUEUE;
1127 ret = dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_EX);
1128 cinfo->no_new_dev_lockres->flags &= ~DLM_LKF_NOQUEUE;
1129 /* Some node does not "see" the device */
1130 if (ret == -EAGAIN)
1131 ret = -ENOENT;
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05001132 if (ret)
1133 unlock_comm(cinfo);
Guoqing Jiang8b9277c2015-12-21 10:51:00 +11001134 else {
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05001135 dlm_lock_sync(cinfo->no_new_dev_lockres, DLM_LOCK_CR);
Guoqing Jiange19508f2015-12-21 10:51:01 +11001136 /* Since MD_CHANGE_DEVS will be set in add_bound_rdev which
1137 * will run soon after add_new_disk, the below path will be
1138 * invoked:
1139 * md_wakeup_thread(mddev->thread)
1140 * -> conf->thread (raid1d)
1141 * -> md_check_recovery -> md_update_sb
1142 * -> metadata_update_start/finish
1143 * MD_CLUSTER_SEND_LOCKED_ALREADY will be cleared eventually.
1144 *
1145 * For other failure cases, metadata_update_cancel and
1146 * add_new_disk_cancel also clear below bit as well.
1147 * */
Guoqing Jiang8b9277c2015-12-21 10:51:00 +11001148 set_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
1149 wake_up(&cinfo->wait);
1150 }
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05001151 return ret;
1152}
1153
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05001154static void add_new_disk_cancel(struct mddev *mddev)
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05001155{
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05001156 struct md_cluster_info *cinfo = mddev->cluster_info;
Guoqing Jiang8b9277c2015-12-21 10:51:00 +11001157 clear_bit(MD_CLUSTER_SEND_LOCKED_ALREADY, &cinfo->state);
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05001158 unlock_comm(cinfo);
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05001159}
1160
Goldwyn Rodriguesfa8259d2015-03-02 10:55:49 -06001161static int new_disk_ack(struct mddev *mddev, bool ack)
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05001162{
1163 struct md_cluster_info *cinfo = mddev->cluster_info;
1164
Goldwyn Rodriguesfa8259d2015-03-02 10:55:49 -06001165 if (!test_bit(MD_CLUSTER_WAITING_FOR_NEWDISK, &cinfo->state)) {
1166 pr_warn("md-cluster(%s): Spurious cluster confirmation\n", mdname(mddev));
1167 return -EINVAL;
1168 }
1169
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05001170 if (ack)
1171 dlm_unlock_sync(cinfo->no_new_dev_lockres);
1172 complete(&cinfo->newdisk_completion);
Goldwyn Rodriguesfa8259d2015-03-02 10:55:49 -06001173 return 0;
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05001174}
1175
Goldwyn Rodrigues88bcfef2015-04-14 10:44:44 -05001176static int remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1177{
Guoqing Jiangaee177a2015-10-12 17:21:24 +08001178 struct cluster_msg cmsg = {0};
Goldwyn Rodrigues88bcfef2015-04-14 10:44:44 -05001179 struct md_cluster_info *cinfo = mddev->cluster_info;
Guoqing Jiangfaeff832015-10-12 17:21:21 +08001180 cmsg.type = cpu_to_le32(REMOVE);
1181 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
Goldwyn Rodrigues54a88392015-12-21 10:51:00 +11001182 return sendmsg(cinfo, &cmsg);
Goldwyn Rodrigues88bcfef2015-04-14 10:44:44 -05001183}
1184
Guoqing Jiangf6a2dc62015-12-21 10:51:00 +11001185static int lock_all_bitmaps(struct mddev *mddev)
1186{
1187 int slot, my_slot, ret, held = 1, i = 0;
1188 char str[64];
1189 struct md_cluster_info *cinfo = mddev->cluster_info;
1190
1191 cinfo->other_bitmap_lockres = kzalloc((mddev->bitmap_info.nodes - 1) *
1192 sizeof(struct dlm_lock_resource *),
1193 GFP_KERNEL);
1194 if (!cinfo->other_bitmap_lockres) {
1195 pr_err("md: can't alloc mem for other bitmap locks\n");
1196 return 0;
1197 }
1198
1199 my_slot = slot_number(mddev);
1200 for (slot = 0; slot < mddev->bitmap_info.nodes; slot++) {
1201 if (slot == my_slot)
1202 continue;
1203
1204 memset(str, '\0', 64);
1205 snprintf(str, 64, "bitmap%04d", slot);
1206 cinfo->other_bitmap_lockres[i] = lockres_init(mddev, str, NULL, 1);
1207 if (!cinfo->other_bitmap_lockres[i])
1208 return -ENOMEM;
1209
1210 cinfo->other_bitmap_lockres[i]->flags |= DLM_LKF_NOQUEUE;
1211 ret = dlm_lock_sync(cinfo->other_bitmap_lockres[i], DLM_LOCK_PW);
1212 if (ret)
1213 held = -1;
1214 i++;
1215 }
1216
1217 return held;
1218}
1219
1220static void unlock_all_bitmaps(struct mddev *mddev)
1221{
1222 struct md_cluster_info *cinfo = mddev->cluster_info;
1223 int i;
1224
1225 /* release other node's bitmap lock if they are existed */
1226 if (cinfo->other_bitmap_lockres) {
1227 for (i = 0; i < mddev->bitmap_info.nodes - 1; i++) {
1228 if (cinfo->other_bitmap_lockres[i]) {
Guoqing Jiangf6a2dc62015-12-21 10:51:00 +11001229 lockres_free(cinfo->other_bitmap_lockres[i]);
1230 }
1231 }
1232 kfree(cinfo->other_bitmap_lockres);
1233 }
1234}
1235
Goldwyn Rodrigues97f6cd32015-04-14 10:45:42 -05001236static int gather_bitmaps(struct md_rdev *rdev)
1237{
1238 int sn, err;
1239 sector_t lo, hi;
Guoqing Jiangaee177a2015-10-12 17:21:24 +08001240 struct cluster_msg cmsg = {0};
Goldwyn Rodrigues97f6cd32015-04-14 10:45:42 -05001241 struct mddev *mddev = rdev->mddev;
1242 struct md_cluster_info *cinfo = mddev->cluster_info;
1243
Guoqing Jiangfaeff832015-10-12 17:21:21 +08001244 cmsg.type = cpu_to_le32(RE_ADD);
1245 cmsg.raid_slot = cpu_to_le32(rdev->desc_nr);
Goldwyn Rodrigues97f6cd32015-04-14 10:45:42 -05001246 err = sendmsg(cinfo, &cmsg);
1247 if (err)
1248 goto out;
1249
1250 for (sn = 0; sn < mddev->bitmap_info.nodes; sn++) {
1251 if (sn == (cinfo->slot_number - 1))
1252 continue;
1253 err = bitmap_copy_from_slot(mddev, sn, &lo, &hi, false);
1254 if (err) {
1255 pr_warn("md-cluster: Could not gather bitmaps from slot %d", sn);
1256 goto out;
1257 }
1258 if ((hi > 0) && (lo < mddev->recovery_cp))
1259 mddev->recovery_cp = lo;
1260 }
1261out:
1262 return err;
1263}
1264
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -05001265static struct md_cluster_operations cluster_ops = {
1266 .join = join,
1267 .leave = leave,
Goldwyn Rodriguescf921cc2014-03-30 00:42:49 -05001268 .slot_number = slot_number,
Goldwyn Rodriguesc186b122015-09-30 13:20:35 -05001269 .resync_start = resync_start,
1270 .resync_finish = resync_finish,
Goldwyn Rodrigues96ae9232014-06-06 12:35:34 -05001271 .resync_info_update = resync_info_update,
Goldwyn Rodrigues293467a2014-06-07 01:44:51 -05001272 .metadata_update_start = metadata_update_start,
1273 .metadata_update_finish = metadata_update_finish,
1274 .metadata_update_cancel = metadata_update_cancel,
Goldwyn Rodrigues589a1c42014-06-07 02:39:37 -05001275 .area_resyncing = area_resyncing,
Goldwyn Rodriguesdbb64f82015-10-01 13:20:27 -05001276 .add_new_disk = add_new_disk,
1277 .add_new_disk_cancel = add_new_disk_cancel,
Goldwyn Rodrigues1aee41f2014-10-29 18:51:31 -05001278 .new_disk_ack = new_disk_ack,
Goldwyn Rodrigues88bcfef2015-04-14 10:44:44 -05001279 .remove_disk = remove_disk,
Guoqing Jiang51e453a2016-05-04 02:17:09 -04001280 .load_bitmaps = load_bitmaps,
Goldwyn Rodrigues97f6cd32015-04-14 10:45:42 -05001281 .gather_bitmaps = gather_bitmaps,
Guoqing Jiangf6a2dc62015-12-21 10:51:00 +11001282 .lock_all_bitmaps = lock_all_bitmaps,
1283 .unlock_all_bitmaps = unlock_all_bitmaps,
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -05001284};
1285
Goldwyn Rodrigues8e854e92014-03-07 11:21:15 -06001286static int __init cluster_init(void)
1287{
1288 pr_warn("md-cluster: EXPERIMENTAL. Use with caution\n");
1289 pr_info("Registering Cluster MD functions\n");
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -05001290 register_md_cluster_operations(&cluster_ops, THIS_MODULE);
Goldwyn Rodrigues8e854e92014-03-07 11:21:15 -06001291 return 0;
1292}
1293
1294static void cluster_exit(void)
1295{
Goldwyn Rodriguesedb39c92014-03-29 10:01:53 -05001296 unregister_md_cluster_operations();
Goldwyn Rodrigues8e854e92014-03-07 11:21:15 -06001297}
1298
1299module_init(cluster_init);
1300module_exit(cluster_exit);
Guoqing Jiang86b57272015-10-12 17:21:25 +08001301MODULE_AUTHOR("SUSE");
Goldwyn Rodrigues8e854e92014-03-07 11:21:15 -06001302MODULE_LICENSE("GPL");
1303MODULE_DESCRIPTION("Clustering support for MD");