xref: /linux/fs/ocfs2/dlm/dlmconvert.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * dlmconvert.c
5  *
6  * underlying calls for lock conversion
7  *
8  * Copyright (C) 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  *
25  */
26 
27 
28 #include <linux/module.h>
29 #include <linux/fs.h>
30 #include <linux/types.h>
31 #include <linux/highmem.h>
32 #include <linux/init.h>
33 #include <linux/sysctl.h>
34 #include <linux/random.h>
35 #include <linux/blkdev.h>
36 #include <linux/socket.h>
37 #include <linux/inet.h>
38 #include <linux/spinlock.h>
39 
40 
41 #include "cluster/heartbeat.h"
42 #include "cluster/nodemanager.h"
43 #include "cluster/tcp.h"
44 
45 #include "dlmapi.h"
46 #include "dlmcommon.h"
47 
48 #include "dlmconvert.h"
49 
50 #define MLOG_MASK_PREFIX ML_DLM
51 #include "cluster/masklog.h"
52 
53 /* NOTE: __dlmconvert_master is the only function in here that
54  * needs a spinlock held on entry (res->spinlock) and it is the
55  * only one that holds a lock on exit (res->spinlock).
56  * All other functions in here need no locks and drop all of
57  * the locks that they acquire. */
58 static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm,
59 					   struct dlm_lock_resource *res,
60 					   struct dlm_lock *lock, int flags,
61 					   int type, int *call_ast,
62 					   int *kick_thread);
63 static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm,
64 					   struct dlm_lock_resource *res,
65 					   struct dlm_lock *lock, int flags, int type);
66 
67 /*
68  * this is only called directly by dlmlock(), and only when the
69  * local node is the owner of the lockres
70  * locking:
71  *   caller needs:  none
72  *   taken:         takes and drops res->spinlock
73  *   held on exit:  none
74  * returns: see __dlmconvert_master
75  */
76 enum dlm_status dlmconvert_master(struct dlm_ctxt *dlm,
77 				  struct dlm_lock_resource *res,
78 				  struct dlm_lock *lock, int flags, int type)
79 {
80 	int call_ast = 0, kick_thread = 0;
81 	enum dlm_status status;
82 
83 	spin_lock(&res->spinlock);
84 	/* we are not in a network handler, this is fine */
85 	__dlm_wait_on_lockres(res);
86 	__dlm_lockres_reserve_ast(res);
87 	res->state |= DLM_LOCK_RES_IN_PROGRESS;
88 
89 	status = __dlmconvert_master(dlm, res, lock, flags, type,
90 				     &call_ast, &kick_thread);
91 
92 	res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
93 	spin_unlock(&res->spinlock);
94 	wake_up(&res->wq);
95 	if (status != DLM_NORMAL && status != DLM_NOTQUEUED)
96 		dlm_error(status);
97 
98 	/* either queue the ast or release it */
99 	if (call_ast)
100 		dlm_queue_ast(dlm, lock);
101 	else
102 		dlm_lockres_release_ast(dlm, res);
103 
104 	if (kick_thread)
105 		dlm_kick_thread(dlm, res);
106 
107 	return status;
108 }
109 
110 /* performs lock conversion at the lockres master site
111  * locking:
112  *   caller needs:  res->spinlock
113  *   taken:         takes and drops lock->spinlock
114  *   held on exit:  res->spinlock
115  * returns: DLM_NORMAL, DLM_NOTQUEUED, DLM_DENIED
116  *   call_ast: whether ast should be called for this lock
117  *   kick_thread: whether dlm_kick_thread should be called
118  */
119 static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm,
120 					   struct dlm_lock_resource *res,
121 					   struct dlm_lock *lock, int flags,
122 					   int type, int *call_ast,
123 					   int *kick_thread)
124 {
125 	enum dlm_status status = DLM_NORMAL;
126 	struct dlm_lock *tmplock=NULL;
127 
128 	assert_spin_locked(&res->spinlock);
129 
130 	mlog(0, "type=%d, convert_type=%d, new convert_type=%d\n",
131 	     lock->ml.type, lock->ml.convert_type, type);
132 
133 	spin_lock(&lock->spinlock);
134 
135 	/* already converting? */
136 	if (lock->ml.convert_type != LKM_IVMODE) {
137 		mlog(ML_ERROR, "attempted to convert a lock with a lock "
138 		     "conversion pending\n");
139 		status = DLM_DENIED;
140 		goto unlock_exit;
141 	}
142 
143 	/* must be on grant queue to convert */
144 	if (!dlm_lock_on_list(&res->granted, lock)) {
145 		mlog(ML_ERROR, "attempted to convert a lock not on grant "
146 		     "queue\n");
147 		status = DLM_DENIED;
148 		goto unlock_exit;
149 	}
150 
151 	if (flags & LKM_VALBLK) {
152 		switch (lock->ml.type) {
153 			case LKM_EXMODE:
154 				/* EX + LKM_VALBLK + convert == set lvb */
155 				mlog(0, "will set lvb: converting %s->%s\n",
156 				     dlm_lock_mode_name(lock->ml.type),
157 				     dlm_lock_mode_name(type));
158 				lock->lksb->flags |= DLM_LKSB_PUT_LVB;
159 				break;
160 			case LKM_PRMODE:
161 			case LKM_NLMODE:
162 				/* refetch if new level is not NL */
163 				if (type > LKM_NLMODE) {
164 					mlog(0, "will fetch new value into "
165 					     "lvb: converting %s->%s\n",
166 					     dlm_lock_mode_name(lock->ml.type),
167 					     dlm_lock_mode_name(type));
168 					lock->lksb->flags |= DLM_LKSB_GET_LVB;
169 				} else {
170 					mlog(0, "will NOT fetch new value "
171 					     "into lvb: converting %s->%s\n",
172 					     dlm_lock_mode_name(lock->ml.type),
173 					     dlm_lock_mode_name(type));
174 					flags &= ~(LKM_VALBLK);
175 				}
176 				break;
177 		}
178 	}
179 
180 
181 	/* in-place downconvert? */
182 	if (type <= lock->ml.type)
183 		goto grant;
184 
185 	/* upconvert from here on */
186 	status = DLM_NORMAL;
187 	list_for_each_entry(tmplock, &res->granted, list) {
188 		if (tmplock == lock)
189 			continue;
190 		if (!dlm_lock_compatible(tmplock->ml.type, type))
191 			goto switch_queues;
192 	}
193 
194 	list_for_each_entry(tmplock, &res->converting, list) {
195 		if (!dlm_lock_compatible(tmplock->ml.type, type))
196 			goto switch_queues;
197 		/* existing conversion requests take precedence */
198 		if (!dlm_lock_compatible(tmplock->ml.convert_type, type))
199 			goto switch_queues;
200 	}
201 
202 	/* fall thru to grant */
203 
204 grant:
205 	mlog(0, "res %.*s, granting %s lock\n", res->lockname.len,
206 	     res->lockname.name, dlm_lock_mode_name(type));
207 	/* immediately grant the new lock type */
208 	lock->lksb->status = DLM_NORMAL;
209 	if (lock->ml.node == dlm->node_num)
210 		mlog(0, "doing in-place convert for nonlocal lock\n");
211 	lock->ml.type = type;
212 	if (lock->lksb->flags & DLM_LKSB_PUT_LVB)
213 		memcpy(res->lvb, lock->lksb->lvb, DLM_LVB_LEN);
214 
215 	status = DLM_NORMAL;
216 	*call_ast = 1;
217 	goto unlock_exit;
218 
219 switch_queues:
220 	if (flags & LKM_NOQUEUE) {
221 		mlog(0, "failed to convert NOQUEUE lock %.*s from "
222 		     "%d to %d...\n", res->lockname.len, res->lockname.name,
223 		     lock->ml.type, type);
224 		status = DLM_NOTQUEUED;
225 		goto unlock_exit;
226 	}
227 	mlog(0, "res %.*s, queueing...\n", res->lockname.len,
228 	     res->lockname.name);
229 
230 	lock->ml.convert_type = type;
231 	/* do not alter lock refcount.  switching lists. */
232 	list_move_tail(&lock->list, &res->converting);
233 
234 unlock_exit:
235 	spin_unlock(&lock->spinlock);
236 	if (status == DLM_DENIED) {
237 		__dlm_print_one_lock_resource(res);
238 	}
239 	if (status == DLM_NORMAL)
240 		*kick_thread = 1;
241 	return status;
242 }
243 
244 void dlm_revert_pending_convert(struct dlm_lock_resource *res,
245 				struct dlm_lock *lock)
246 {
247 	/* do not alter lock refcount.  switching lists. */
248 	list_move_tail(&lock->list, &res->granted);
249 	lock->ml.convert_type = LKM_IVMODE;
250 	lock->lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB);
251 }
252 
253 /* messages the master site to do lock conversion
254  * locking:
255  *   caller needs:  none
256  *   taken:         takes and drops res->spinlock, uses DLM_LOCK_RES_IN_PROGRESS
257  *   held on exit:  none
258  * returns: DLM_NORMAL, DLM_RECOVERING, status from remote node
259  */
260 enum dlm_status dlmconvert_remote(struct dlm_ctxt *dlm,
261 				  struct dlm_lock_resource *res,
262 				  struct dlm_lock *lock, int flags, int type)
263 {
264 	enum dlm_status status;
265 
266 	mlog(0, "type=%d, convert_type=%d, busy=%d\n", lock->ml.type,
267 	     lock->ml.convert_type, res->state & DLM_LOCK_RES_IN_PROGRESS);
268 
269 	spin_lock(&res->spinlock);
270 	if (res->state & DLM_LOCK_RES_RECOVERING) {
271 		mlog(0, "bailing out early since res is RECOVERING "
272 		     "on secondary queue\n");
273 		/* __dlm_print_one_lock_resource(res); */
274 		status = DLM_RECOVERING;
275 		goto bail;
276 	}
277 	/* will exit this call with spinlock held */
278 	__dlm_wait_on_lockres(res);
279 
280 	if (lock->ml.convert_type != LKM_IVMODE) {
281 		__dlm_print_one_lock_resource(res);
282 		mlog(ML_ERROR, "converting a remote lock that is already "
283 		     "converting! (cookie=%u:%llu, conv=%d)\n",
284 		     dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
285 		     dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
286 		     lock->ml.convert_type);
287 		status = DLM_DENIED;
288 		goto bail;
289 	}
290 	res->state |= DLM_LOCK_RES_IN_PROGRESS;
291 	/* move lock to local convert queue */
292 	/* do not alter lock refcount.  switching lists. */
293 	list_move_tail(&lock->list, &res->converting);
294 	lock->convert_pending = 1;
295 	lock->ml.convert_type = type;
296 
297 	if (flags & LKM_VALBLK) {
298 		if (lock->ml.type == LKM_EXMODE) {
299 			flags |= LKM_PUT_LVB;
300 			lock->lksb->flags |= DLM_LKSB_PUT_LVB;
301 		} else {
302 			if (lock->ml.convert_type == LKM_NLMODE)
303 				flags &= ~LKM_VALBLK;
304 			else {
305 				flags |= LKM_GET_LVB;
306 				lock->lksb->flags |= DLM_LKSB_GET_LVB;
307 			}
308 		}
309 	}
310 	spin_unlock(&res->spinlock);
311 
312 	/* no locks held here.
313 	 * need to wait for a reply as to whether it got queued or not. */
314 	status = dlm_send_remote_convert_request(dlm, res, lock, flags, type);
315 
316 	spin_lock(&res->spinlock);
317 	res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
318 	lock->convert_pending = 0;
319 	/* if it failed, move it back to granted queue */
320 	if (status != DLM_NORMAL) {
321 		if (status != DLM_NOTQUEUED)
322 			dlm_error(status);
323 		dlm_revert_pending_convert(res, lock);
324 	}
325 bail:
326 	spin_unlock(&res->spinlock);
327 
328 	/* TODO: should this be a wake_one? */
329 	/* wake up any IN_PROGRESS waiters */
330 	wake_up(&res->wq);
331 
332 	return status;
333 }
334 
335 /* sends DLM_CONVERT_LOCK_MSG to master site
336  * locking:
337  *   caller needs:  none
338  *   taken:         none
339  *   held on exit:  none
340  * returns: DLM_NOLOCKMGR, status from remote node
341  */
342 static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm,
343 					   struct dlm_lock_resource *res,
344 					   struct dlm_lock *lock, int flags, int type)
345 {
346 	struct dlm_convert_lock convert;
347 	int tmpret;
348 	enum dlm_status ret;
349 	int status = 0;
350 	struct kvec vec[2];
351 	size_t veclen = 1;
352 
353 	mlog(0, "%.*s\n", res->lockname.len, res->lockname.name);
354 
355 	memset(&convert, 0, sizeof(struct dlm_convert_lock));
356 	convert.node_idx = dlm->node_num;
357 	convert.requested_type = type;
358 	convert.cookie = lock->ml.cookie;
359 	convert.namelen = res->lockname.len;
360 	convert.flags = cpu_to_be32(flags);
361 	memcpy(convert.name, res->lockname.name, convert.namelen);
362 
363 	vec[0].iov_len = sizeof(struct dlm_convert_lock);
364 	vec[0].iov_base = &convert;
365 
366 	if (flags & LKM_PUT_LVB) {
367 		/* extra data to send if we are updating lvb */
368 		vec[1].iov_len = DLM_LVB_LEN;
369 		vec[1].iov_base = lock->lksb->lvb;
370 		veclen++;
371 	}
372 
373 	tmpret = o2net_send_message_vec(DLM_CONVERT_LOCK_MSG, dlm->key,
374 					vec, veclen, res->owner, &status);
375 	if (tmpret >= 0) {
376 		// successfully sent and received
377 		ret = status;  // this is already a dlm_status
378 		if (ret == DLM_RECOVERING) {
379 			mlog(0, "node %u returned DLM_RECOVERING from convert "
380 			     "message!\n", res->owner);
381 		} else if (ret == DLM_MIGRATING) {
382 			mlog(0, "node %u returned DLM_MIGRATING from convert "
383 			     "message!\n", res->owner);
384 		} else if (ret == DLM_FORWARD) {
385 			mlog(0, "node %u returned DLM_FORWARD from convert "
386 			     "message!\n", res->owner);
387 		} else if (ret != DLM_NORMAL && ret != DLM_NOTQUEUED)
388 			dlm_error(ret);
389 	} else {
390 		mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
391 		     "node %u\n", tmpret, DLM_CONVERT_LOCK_MSG, dlm->key,
392 		     res->owner);
393 		if (dlm_is_host_down(tmpret)) {
394 			/* instead of logging the same network error over
395 			 * and over, sleep here and wait for the heartbeat
396 			 * to notice the node is dead.  times out after 5s. */
397 			dlm_wait_for_node_death(dlm, res->owner,
398 						DLM_NODE_DEATH_WAIT_MAX);
399 			ret = DLM_RECOVERING;
400 			mlog(0, "node %u died so returning DLM_RECOVERING "
401 			     "from convert message!\n", res->owner);
402 		} else {
403 			ret = dlm_err_to_dlm_status(tmpret);
404 		}
405 	}
406 
407 	return ret;
408 }
409 
410 /* handler for DLM_CONVERT_LOCK_MSG on master site
411  * locking:
412  *   caller needs:  none
413  *   taken:         takes and drop res->spinlock
414  *   held on exit:  none
415  * returns: DLM_NORMAL, DLM_IVLOCKID, DLM_BADARGS,
416  *          status from __dlmconvert_master
417  */
418 int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data,
419 			     void **ret_data)
420 {
421 	struct dlm_ctxt *dlm = data;
422 	struct dlm_convert_lock *cnv = (struct dlm_convert_lock *)msg->buf;
423 	struct dlm_lock_resource *res = NULL;
424 	struct dlm_lock *lock = NULL;
425 	struct dlm_lock *tmp_lock;
426 	struct dlm_lockstatus *lksb;
427 	enum dlm_status status = DLM_NORMAL;
428 	u32 flags;
429 	int call_ast = 0, kick_thread = 0, ast_reserved = 0, wake = 0;
430 
431 	if (!dlm_grab(dlm)) {
432 		dlm_error(DLM_REJECTED);
433 		return DLM_REJECTED;
434 	}
435 
436 	mlog_bug_on_msg(!dlm_domain_fully_joined(dlm),
437 			"Domain %s not fully joined!\n", dlm->name);
438 
439 	if (cnv->namelen > DLM_LOCKID_NAME_MAX) {
440 		status = DLM_IVBUFLEN;
441 		dlm_error(status);
442 		goto leave;
443 	}
444 
445 	flags = be32_to_cpu(cnv->flags);
446 
447 	if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) ==
448 	     (LKM_PUT_LVB|LKM_GET_LVB)) {
449 		mlog(ML_ERROR, "both PUT and GET lvb specified\n");
450 		status = DLM_BADARGS;
451 		goto leave;
452 	}
453 
454 	mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" :
455 	     (flags & LKM_GET_LVB ? "get lvb" : "none"));
456 
457 	status = DLM_IVLOCKID;
458 	res = dlm_lookup_lockres(dlm, cnv->name, cnv->namelen);
459 	if (!res) {
460 		dlm_error(status);
461 		goto leave;
462 	}
463 
464 	spin_lock(&res->spinlock);
465 	status = __dlm_lockres_state_to_status(res);
466 	if (status != DLM_NORMAL) {
467 		spin_unlock(&res->spinlock);
468 		dlm_error(status);
469 		goto leave;
470 	}
471 	list_for_each_entry(tmp_lock, &res->granted, list) {
472 		if (tmp_lock->ml.cookie == cnv->cookie &&
473 		    tmp_lock->ml.node == cnv->node_idx) {
474 			lock = tmp_lock;
475 			dlm_lock_get(lock);
476 			break;
477 		}
478 	}
479 	spin_unlock(&res->spinlock);
480 	if (!lock) {
481 		status = DLM_IVLOCKID;
482 		mlog(ML_ERROR, "did not find lock to convert on grant queue! "
483 			       "cookie=%u:%llu\n",
484 		     dlm_get_lock_cookie_node(be64_to_cpu(cnv->cookie)),
485 		     dlm_get_lock_cookie_seq(be64_to_cpu(cnv->cookie)));
486 		dlm_print_one_lock_resource(res);
487 		goto leave;
488 	}
489 
490 	/* found the lock */
491 	lksb = lock->lksb;
492 
493 	/* see if caller needed to get/put lvb */
494 	if (flags & LKM_PUT_LVB) {
495 		BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
496 		lksb->flags |= DLM_LKSB_PUT_LVB;
497 		memcpy(&lksb->lvb[0], &cnv->lvb[0], DLM_LVB_LEN);
498 	} else if (flags & LKM_GET_LVB) {
499 		BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
500 		lksb->flags |= DLM_LKSB_GET_LVB;
501 	}
502 
503 	spin_lock(&res->spinlock);
504 	status = __dlm_lockres_state_to_status(res);
505 	if (status == DLM_NORMAL) {
506 		__dlm_lockres_reserve_ast(res);
507 		ast_reserved = 1;
508 		res->state |= DLM_LOCK_RES_IN_PROGRESS;
509 		status = __dlmconvert_master(dlm, res, lock, flags,
510 					     cnv->requested_type,
511 					     &call_ast, &kick_thread);
512 		res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
513 		wake = 1;
514 	}
515 	spin_unlock(&res->spinlock);
516 	if (wake)
517 		wake_up(&res->wq);
518 
519 	if (status != DLM_NORMAL) {
520 		if (status != DLM_NOTQUEUED)
521 			dlm_error(status);
522 		lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB);
523 	}
524 
525 leave:
526 	if (lock)
527 		dlm_lock_put(lock);
528 
529 	/* either queue the ast or release it, if reserved */
530 	if (call_ast)
531 		dlm_queue_ast(dlm, lock);
532 	else if (ast_reserved)
533 		dlm_lockres_release_ast(dlm, res);
534 
535 	if (kick_thread)
536 		dlm_kick_thread(dlm, res);
537 
538 	if (res)
539 		dlm_lockres_put(res);
540 
541 	dlm_put(dlm);
542 
543 	return status;
544 }
545