1/*
   2 * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3 * Copyright 2004-2011 Red Hat, Inc.
   4 *
   5 * This copyrighted material is made available to anyone wishing to use,
   6 * modify, copy, or redistribute it subject to the terms and conditions
   7 * of the GNU General Public License version 2.
   8 */
   9
  10#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12#include <linux/fs.h>
  13#include <linux/dlm.h>
  14#include <linux/slab.h>
  15#include <linux/types.h>
  16#include <linux/delay.h>
  17#include <linux/gfs2_ondisk.h>
  18#include <linux/sched/signal.h>
  19
  20#include "incore.h"
  21#include "glock.h"
  22#include "util.h"
  23#include "sys.h"
  24#include "trace_gfs2.h"
  25
 
 
  26/**
  27 * gfs2_update_stats - Update time based stats
  28 * @mv: Pointer to mean/variance structure to update
  29 * @sample: New data to include
  30 *
  31 * @delta is the difference between the current rtt sample and the
  32 * running average srtt. We add 1/8 of that to the srtt in order to
  33 * update the current srtt estimate. The variance estimate is a bit
  34 * more complicated. We subtract the abs value of the @delta from
  35 * the current variance estimate and add 1/4 of that to the running
  36 * total.
  37 *
  38 * Note that the index points at the array entry containing the smoothed
  39 * mean value, and the variance is always in the following entry
  40 *
  41 * Reference: TCP/IP Illustrated, vol 2, p. 831,832
  42 * All times are in units of integer nanoseconds. Unlike the TCP/IP case,
  43 * they are not scaled fixed point.
  44 */
  45
  46static inline void gfs2_update_stats(struct gfs2_lkstats *s, unsigned index,
  47				     s64 sample)
  48{
  49	s64 delta = sample - s->stats[index];
  50	s->stats[index] += (delta >> 3);
  51	index++;
  52	s->stats[index] += ((abs(delta) - s->stats[index]) >> 2);
  53}
  54
  55/**
  56 * gfs2_update_reply_times - Update locking statistics
  57 * @gl: The glock to update
  58 *
  59 * This assumes that gl->gl_dstamp has been set earlier.
  60 *
  61 * The rtt (lock round trip time) is an estimate of the time
  62 * taken to perform a dlm lock request. We update it on each
  63 * reply from the dlm.
  64 *
  65 * The blocking flag is set on the glock for all dlm requests
  66 * which may potentially block due to lock requests from other nodes.
  67 * DLM requests where the current lock state is exclusive, the
  68 * requested state is null (or unlocked) or where the TRY or
  69 * TRY_1CB flags are set are classified as non-blocking. All
  70 * other DLM requests are counted as (potentially) blocking.
  71 */
  72static inline void gfs2_update_reply_times(struct gfs2_glock *gl)
  73{
  74	struct gfs2_pcpu_lkstats *lks;
  75	const unsigned gltype = gl->gl_name.ln_type;
  76	unsigned index = test_bit(GLF_BLOCKING, &gl->gl_flags) ?
  77			 GFS2_LKS_SRTTB : GFS2_LKS_SRTT;
  78	s64 rtt;
  79
  80	preempt_disable();
  81	rtt = ktime_to_ns(ktime_sub(ktime_get_real(), gl->gl_dstamp));
  82	lks = this_cpu_ptr(gl->gl_name.ln_sbd->sd_lkstats);
  83	gfs2_update_stats(&gl->gl_stats, index, rtt);		/* Local */
  84	gfs2_update_stats(&lks->lkstats[gltype], index, rtt);	/* Global */
  85	preempt_enable();
  86
  87	trace_gfs2_glock_lock_time(gl, rtt);
  88}
  89
  90/**
  91 * gfs2_update_request_times - Update locking statistics
  92 * @gl: The glock to update
  93 *
  94 * The irt (lock inter-request times) measures the average time
  95 * between requests to the dlm. It is updated immediately before
  96 * each dlm call.
  97 */
  98
  99static inline void gfs2_update_request_times(struct gfs2_glock *gl)
 100{
 101	struct gfs2_pcpu_lkstats *lks;
 102	const unsigned gltype = gl->gl_name.ln_type;
 103	ktime_t dstamp;
 104	s64 irt;
 105
 106	preempt_disable();
 107	dstamp = gl->gl_dstamp;
 108	gl->gl_dstamp = ktime_get_real();
 109	irt = ktime_to_ns(ktime_sub(gl->gl_dstamp, dstamp));
 110	lks = this_cpu_ptr(gl->gl_name.ln_sbd->sd_lkstats);
 111	gfs2_update_stats(&gl->gl_stats, GFS2_LKS_SIRT, irt);		/* Local */
 112	gfs2_update_stats(&lks->lkstats[gltype], GFS2_LKS_SIRT, irt);	/* Global */
 113	preempt_enable();
 114}
 115 
 116static void gdlm_ast(void *arg)
 117{
 118	struct gfs2_glock *gl = arg;
 119	unsigned ret = gl->gl_state;
 120
 121	gfs2_update_reply_times(gl);
 122	BUG_ON(gl->gl_lksb.sb_flags & DLM_SBF_DEMOTED);
 123
 124	if ((gl->gl_lksb.sb_flags & DLM_SBF_VALNOTVALID) && gl->gl_lksb.sb_lvbptr)
 125		memset(gl->gl_lksb.sb_lvbptr, 0, GDLM_LVB_SIZE);
 126
 127	switch (gl->gl_lksb.sb_status) {
 128	case -DLM_EUNLOCK: /* Unlocked, so glock can be freed */
 129		gfs2_glock_free(gl);
 130		return;
 131	case -DLM_ECANCEL: /* Cancel while getting lock */
 132		ret |= LM_OUT_CANCELED;
 133		goto out;
 134	case -EAGAIN: /* Try lock fails */
 135	case -EDEADLK: /* Deadlock detected */
 136		goto out;
 137	case -ETIMEDOUT: /* Canceled due to timeout */
 138		ret |= LM_OUT_ERROR;
 139		goto out;
 140	case 0: /* Success */
 141		break;
 142	default: /* Something unexpected */
 143		BUG();
 144	}
 145
 146	ret = gl->gl_req;
 147	if (gl->gl_lksb.sb_flags & DLM_SBF_ALTMODE) {
 148		if (gl->gl_req == LM_ST_SHARED)
 149			ret = LM_ST_DEFERRED;
 150		else if (gl->gl_req == LM_ST_DEFERRED)
 151			ret = LM_ST_SHARED;
 152		else
 153			BUG();
 154	}
 155
 156	set_bit(GLF_INITIAL, &gl->gl_flags);
 157	gfs2_glock_complete(gl, ret);
 158	return;
 159out:
 160	if (!test_bit(GLF_INITIAL, &gl->gl_flags))
 161		gl->gl_lksb.sb_lkid = 0;
 162	gfs2_glock_complete(gl, ret);
 163}
 164
 165static void gdlm_bast(void *arg, int mode)
 166{
 167	struct gfs2_glock *gl = arg;
 168
 169	switch (mode) {
 170	case DLM_LOCK_EX:
 171		gfs2_glock_cb(gl, LM_ST_UNLOCKED);
 172		break;
 173	case DLM_LOCK_CW:
 174		gfs2_glock_cb(gl, LM_ST_DEFERRED);
 175		break;
 176	case DLM_LOCK_PR:
 177		gfs2_glock_cb(gl, LM_ST_SHARED);
 178		break;
 179	default:
 180		pr_err("unknown bast mode %d\n", mode);
 181		BUG();
 182	}
 183}
 184
 185/* convert gfs lock-state to dlm lock-mode */
 186
 187static int make_mode(const unsigned int lmstate)
 188{
 189	switch (lmstate) {
 190	case LM_ST_UNLOCKED:
 191		return DLM_LOCK_NL;
 192	case LM_ST_EXCLUSIVE:
 193		return DLM_LOCK_EX;
 194	case LM_ST_DEFERRED:
 195		return DLM_LOCK_CW;
 196	case LM_ST_SHARED:
 197		return DLM_LOCK_PR;
 198	}
 199	pr_err("unknown LM state %d\n", lmstate);
 200	BUG();
 201	return -1;
 202}
 203
 204static u32 make_flags(struct gfs2_glock *gl, const unsigned int gfs_flags,
 205		      const int req)
 206{
 207	u32 lkf = 0;
 208
 209	if (gl->gl_lksb.sb_lvbptr)
 210		lkf |= DLM_LKF_VALBLK;
 211
 212	if (gfs_flags & LM_FLAG_TRY)
 213		lkf |= DLM_LKF_NOQUEUE;
 214
 215	if (gfs_flags & LM_FLAG_TRY_1CB) {
 216		lkf |= DLM_LKF_NOQUEUE;
 217		lkf |= DLM_LKF_NOQUEUEBAST;
 218	}
 219
 220	if (gfs_flags & LM_FLAG_PRIORITY) {
 221		lkf |= DLM_LKF_NOORDER;
 222		lkf |= DLM_LKF_HEADQUE;
 223	}
 224
 225	if (gfs_flags & LM_FLAG_ANY) {
 226		if (req == DLM_LOCK_PR)
 227			lkf |= DLM_LKF_ALTCW;
 228		else if (req == DLM_LOCK_CW)
 229			lkf |= DLM_LKF_ALTPR;
 230		else
 231			BUG();
 232	}
 233
 234	if (gl->gl_lksb.sb_lkid != 0) {
 235		lkf |= DLM_LKF_CONVERT;
 236		if (test_bit(GLF_BLOCKING, &gl->gl_flags))
 237			lkf |= DLM_LKF_QUECVT;
 238	}
 239
 240	return lkf;
 241}
 242
 243static void gfs2_reverse_hex(char *c, u64 value)
 244{
 245	*c = '0';
 246	while (value) {
 247		*c-- = hex_asc[value & 0x0f];
 248		value >>= 4;
 249	}
 250}
 251
 252static int gdlm_lock(struct gfs2_glock *gl, unsigned int req_state,
 253		     unsigned int flags)
 254{
 255	struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
 256	int req;
 257	u32 lkf;
 258	char strname[GDLM_STRNAME_BYTES] = "";
 259
 260	req = make_mode(req_state);
 261	lkf = make_flags(gl, flags, req);
 262	gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
 263	gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
 264	if (gl->gl_lksb.sb_lkid) {
 265		gfs2_update_request_times(gl);
 266	} else {
 267		memset(strname, ' ', GDLM_STRNAME_BYTES - 1);
 268		strname[GDLM_STRNAME_BYTES - 1] = '\0';
 269		gfs2_reverse_hex(strname + 7, gl->gl_name.ln_type);
 270		gfs2_reverse_hex(strname + 23, gl->gl_name.ln_number);
 271		gl->gl_dstamp = ktime_get_real();
 272	}
 273	/*
 274	 * Submit the actual lock request.
 275	 */
 276
 277	return dlm_lock(ls->ls_dlm, req, &gl->gl_lksb, lkf, strname,
 278			GDLM_STRNAME_BYTES - 1, 0, gdlm_ast, gl, gdlm_bast);
 279}
 280
 281static void gdlm_put_lock(struct gfs2_glock *gl)
 282{
 283	struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 284	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 285	int lvb_needs_unlock = 0;
 286	int error;
 287
 288	if (gl->gl_lksb.sb_lkid == 0) {
 289		gfs2_glock_free(gl);
 290		return;
 291	}
 292
 293	clear_bit(GLF_BLOCKING, &gl->gl_flags);
 294	gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
 295	gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
 296	gfs2_update_request_times(gl);
 297
 298	/* don't want to skip dlm_unlock writing the lvb when lock is ex */
 299
 300	if (gl->gl_lksb.sb_lvbptr && (gl->gl_state == LM_ST_EXCLUSIVE))
 301		lvb_needs_unlock = 1;
 302
 303	if (test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags) &&
 304	    !lvb_needs_unlock) {
 305		gfs2_glock_free(gl);
 306		return;
 307	}
 308
 309	error = dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_VALBLK,
 310			   NULL, gl);
 311	if (error) {
 312		pr_err("gdlm_unlock %x,%llx err=%d\n",
 313		       gl->gl_name.ln_type,
 314		       (unsigned long long)gl->gl_name.ln_number, error);
 315		return;
 316	}
 317}
 318
 319static void gdlm_cancel(struct gfs2_glock *gl)
 320{
 321	struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
 322	dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_CANCEL, NULL, gl);
 323}
 324
 325/*
 326 * dlm/gfs2 recovery coordination using dlm_recover callbacks
 327 *
 328 *  1. dlm_controld sees lockspace members change
 329 *  2. dlm_controld blocks dlm-kernel locking activity
 330 *  3. dlm_controld within dlm-kernel notifies gfs2 (recover_prep)
 331 *  4. dlm_controld starts and finishes its own user level recovery
 332 *  5. dlm_controld starts dlm-kernel dlm_recoverd to do kernel recovery
 333 *  6. dlm_recoverd notifies gfs2 of failed nodes (recover_slot)
 334 *  7. dlm_recoverd does its own lock recovery
 335 *  8. dlm_recoverd unblocks dlm-kernel locking activity
 336 *  9. dlm_recoverd notifies gfs2 when done (recover_done with new generation)
 337 * 10. gfs2_control updates control_lock lvb with new generation and jid bits
 338 * 11. gfs2_control enqueues journals for gfs2_recover to recover (maybe none)
 339 * 12. gfs2_recover dequeues and recovers journals of failed nodes
 340 * 13. gfs2_recover provides recovery results to gfs2_control (recovery_result)
 341 * 14. gfs2_control updates control_lock lvb jid bits for recovered journals
 342 * 15. gfs2_control unblocks normal locking when all journals are recovered
 343 *
 344 * - failures during recovery
 345 *
 346 * recover_prep() may set BLOCK_LOCKS (step 3) again before gfs2_control
 347 * clears BLOCK_LOCKS (step 15), e.g. another node fails while still
 348 * recovering for a prior failure.  gfs2_control needs a way to detect
 349 * this so it can leave BLOCK_LOCKS set in step 15.  This is managed using
 350 * the recover_block and recover_start values.
 351 *
 352 * recover_done() provides a new lockspace generation number each time it
 353 * is called (step 9).  This generation number is saved as recover_start.
 354 * When recover_prep() is called, it sets BLOCK_LOCKS and sets
 355 * recover_block = recover_start.  So, while recover_block is equal to
 356 * recover_start, BLOCK_LOCKS should remain set.  (recover_spin must
 357 * be held around the BLOCK_LOCKS/recover_block/recover_start logic.)
 358 *
 359 * - more specific gfs2 steps in sequence above
 360 *
 361 *  3. recover_prep sets BLOCK_LOCKS and sets recover_block = recover_start
 362 *  6. recover_slot records any failed jids (maybe none)
 363 *  9. recover_done sets recover_start = new generation number
 364 * 10. gfs2_control sets control_lock lvb = new gen + bits for failed jids
 365 * 12. gfs2_recover does journal recoveries for failed jids identified above
 366 * 14. gfs2_control clears control_lock lvb bits for recovered jids
 367 * 15. gfs2_control checks if recover_block == recover_start (step 3 occured
 368 *     again) then do nothing, otherwise if recover_start > recover_block
 369 *     then clear BLOCK_LOCKS.
 370 *
 371 * - parallel recovery steps across all nodes
 372 *
 373 * All nodes attempt to update the control_lock lvb with the new generation
 374 * number and jid bits, but only the first to get the control_lock EX will
 375 * do so; others will see that it's already done (lvb already contains new
 376 * generation number.)
 377 *
 378 * . All nodes get the same recover_prep/recover_slot/recover_done callbacks
 379 * . All nodes attempt to set control_lock lvb gen + bits for the new gen
 380 * . One node gets control_lock first and writes the lvb, others see it's done
 381 * . All nodes attempt to recover jids for which they see control_lock bits set
 382 * . One node succeeds for a jid, and that one clears the jid bit in the lvb
 383 * . All nodes will eventually see all lvb bits clear and unblock locks
 384 *
 385 * - is there a problem with clearing an lvb bit that should be set
 386 *   and missing a journal recovery?
 387 *
 388 * 1. jid fails
 389 * 2. lvb bit set for step 1
 390 * 3. jid recovered for step 1
 391 * 4. jid taken again (new mount)
 392 * 5. jid fails (for step 4)
 393 * 6. lvb bit set for step 5 (will already be set)
 394 * 7. lvb bit cleared for step 3
 395 *
 396 * This is not a problem because the failure in step 5 does not
 397 * require recovery, because the mount in step 4 could not have
 398 * progressed far enough to unblock locks and access the fs.  The
 399 * control_mount() function waits for all recoveries to be complete
 400 * for the latest lockspace generation before ever unblocking locks
 401 * and returning.  The mount in step 4 waits until the recovery in
 402 * step 1 is done.
 403 *
 404 * - special case of first mounter: first node to mount the fs
 405 *
 406 * The first node to mount a gfs2 fs needs to check all the journals
 407 * and recover any that need recovery before other nodes are allowed
 408 * to mount the fs.  (Others may begin mounting, but they must wait
 409 * for the first mounter to be done before taking locks on the fs
 410 * or accessing the fs.)  This has two parts:
 411 *
 412 * 1. The mounted_lock tells a node it's the first to mount the fs.
 413 * Each node holds the mounted_lock in PR while it's mounted.
 414 * Each node tries to acquire the mounted_lock in EX when it mounts.
 415 * If a node is granted the mounted_lock EX it means there are no
 416 * other mounted nodes (no PR locks exist), and it is the first mounter.
 417 * The mounted_lock is demoted to PR when first recovery is done, so
 418 * others will fail to get an EX lock, but will get a PR lock.
 419 *
 420 * 2. The control_lock blocks others in control_mount() while the first
 421 * mounter is doing first mount recovery of all journals.
 422 * A mounting node needs to acquire control_lock in EX mode before
 423 * it can proceed.  The first mounter holds control_lock in EX while doing
 424 * the first mount recovery, blocking mounts from other nodes, then demotes
 425 * control_lock to NL when it's done (others_may_mount/first_done),
 426 * allowing other nodes to continue mounting.
 427 *
 428 * first mounter:
 429 * control_lock EX/NOQUEUE success
 430 * mounted_lock EX/NOQUEUE success (no other PR, so no other mounters)
 431 * set first=1
 432 * do first mounter recovery
 433 * mounted_lock EX->PR
 434 * control_lock EX->NL, write lvb generation
 435 *
 436 * other mounter:
 437 * control_lock EX/NOQUEUE success (if fail -EAGAIN, retry)
 438 * mounted_lock EX/NOQUEUE fail -EAGAIN (expected due to other mounters PR)
 439 * mounted_lock PR/NOQUEUE success
 440 * read lvb generation
 441 * control_lock EX->NL
 442 * set first=0
 443 *
 444 * - mount during recovery
 445 *
 446 * If a node mounts while others are doing recovery (not first mounter),
 447 * the mounting node will get its initial recover_done() callback without
 448 * having seen any previous failures/callbacks.
 449 *
 450 * It must wait for all recoveries preceding its mount to be finished
 451 * before it unblocks locks.  It does this by repeating the "other mounter"
 452 * steps above until the lvb generation number is >= its mount generation
 453 * number (from initial recover_done) and all lvb bits are clear.
 454 *
 455 * - control_lock lvb format
 456 *
 457 * 4 bytes generation number: the latest dlm lockspace generation number
 458 * from recover_done callback.  Indicates the jid bitmap has been updated
 459 * to reflect all slot failures through that generation.
 460 * 4 bytes unused.
 461 * GDLM_LVB_SIZE-8 bytes of jid bit map. If bit N is set, it indicates
 462 * that jid N needs recovery.
 463 */
 464
 465#define JID_BITMAP_OFFSET 8 /* 4 byte generation number + 4 byte unused */
 466
 467static void control_lvb_read(struct lm_lockstruct *ls, uint32_t *lvb_gen,
 468			     char *lvb_bits)
 469{
 470	__le32 gen;
 471	memcpy(lvb_bits, ls->ls_control_lvb, GDLM_LVB_SIZE);
 472	memcpy(&gen, lvb_bits, sizeof(__le32));
 473	*lvb_gen = le32_to_cpu(gen);
 474}
 475
 476static void control_lvb_write(struct lm_lockstruct *ls, uint32_t lvb_gen,
 477			      char *lvb_bits)
 478{
 479	__le32 gen;
 480	memcpy(ls->ls_control_lvb, lvb_bits, GDLM_LVB_SIZE);
 481	gen = cpu_to_le32(lvb_gen);
 482	memcpy(ls->ls_control_lvb, &gen, sizeof(__le32));
 483}
 484
 485static int all_jid_bits_clear(char *lvb)
 486{
 487	return !memchr_inv(lvb + JID_BITMAP_OFFSET, 0,
 488			GDLM_LVB_SIZE - JID_BITMAP_OFFSET);
 
 
 
 
 489}
 490
 491static void sync_wait_cb(void *arg)
 492{
 493	struct lm_lockstruct *ls = arg;
 494	complete(&ls->ls_sync_wait);
 495}
 496
 497static int sync_unlock(struct gfs2_sbd *sdp, struct dlm_lksb *lksb, char *name)
 498{
 499	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 500	int error;
 501
 502	error = dlm_unlock(ls->ls_dlm, lksb->sb_lkid, 0, lksb, ls);
 503	if (error) {
 504		fs_err(sdp, "%s lkid %x error %d\n",
 505		       name, lksb->sb_lkid, error);
 506		return error;
 507	}
 508
 509	wait_for_completion(&ls->ls_sync_wait);
 510
 511	if (lksb->sb_status != -DLM_EUNLOCK) {
 512		fs_err(sdp, "%s lkid %x status %d\n",
 513		       name, lksb->sb_lkid, lksb->sb_status);
 514		return -1;
 515	}
 516	return 0;
 517}
 518
 519static int sync_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags,
 520		     unsigned int num, struct dlm_lksb *lksb, char *name)
 521{
 522	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 523	char strname[GDLM_STRNAME_BYTES];
 524	int error, status;
 525
 526	memset(strname, 0, GDLM_STRNAME_BYTES);
 527	snprintf(strname, GDLM_STRNAME_BYTES, "%8x%16x", LM_TYPE_NONDISK, num);
 528
 529	error = dlm_lock(ls->ls_dlm, mode, lksb, flags,
 530			 strname, GDLM_STRNAME_BYTES - 1,
 531			 0, sync_wait_cb, ls, NULL);
 532	if (error) {
 533		fs_err(sdp, "%s lkid %x flags %x mode %d error %d\n",
 534		       name, lksb->sb_lkid, flags, mode, error);
 535		return error;
 536	}
 537
 538	wait_for_completion(&ls->ls_sync_wait);
 539
 540	status = lksb->sb_status;
 541
 542	if (status && status != -EAGAIN) {
 543		fs_err(sdp, "%s lkid %x flags %x mode %d status %d\n",
 544		       name, lksb->sb_lkid, flags, mode, status);
 545	}
 546
 547	return status;
 548}
 549
 550static int mounted_unlock(struct gfs2_sbd *sdp)
 551{
 552	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 553	return sync_unlock(sdp, &ls->ls_mounted_lksb, "mounted_lock");
 554}
 555
 556static int mounted_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
 557{
 558	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 559	return sync_lock(sdp, mode, flags, GFS2_MOUNTED_LOCK,
 560			 &ls->ls_mounted_lksb, "mounted_lock");
 561}
 562
 563static int control_unlock(struct gfs2_sbd *sdp)
 564{
 565	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 566	return sync_unlock(sdp, &ls->ls_control_lksb, "control_lock");
 567}
 568
 569static int control_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
 570{
 571	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 572	return sync_lock(sdp, mode, flags, GFS2_CONTROL_LOCK,
 573			 &ls->ls_control_lksb, "control_lock");
 574}
 575
 576static void gfs2_control_func(struct work_struct *work)
 577{
 578	struct gfs2_sbd *sdp = container_of(work, struct gfs2_sbd, sd_control_work.work);
 579	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 
 580	uint32_t block_gen, start_gen, lvb_gen, flags;
 581	int recover_set = 0;
 582	int write_lvb = 0;
 583	int recover_size;
 584	int i, error;
 585
 586	spin_lock(&ls->ls_recover_spin);
 587	/*
 588	 * No MOUNT_DONE means we're still mounting; control_mount()
 589	 * will set this flag, after which this thread will take over
 590	 * all further clearing of BLOCK_LOCKS.
 591	 *
 592	 * FIRST_MOUNT means this node is doing first mounter recovery,
 593	 * for which recovery control is handled by
 594	 * control_mount()/control_first_done(), not this thread.
 595	 */
 596	if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
 597	     test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
 598		spin_unlock(&ls->ls_recover_spin);
 599		return;
 600	}
 601	block_gen = ls->ls_recover_block;
 602	start_gen = ls->ls_recover_start;
 603	spin_unlock(&ls->ls_recover_spin);
 604
 605	/*
 606	 * Equal block_gen and start_gen implies we are between
 607	 * recover_prep and recover_done callbacks, which means
 608	 * dlm recovery is in progress and dlm locking is blocked.
 609	 * There's no point trying to do any work until recover_done.
 610	 */
 611
 612	if (block_gen == start_gen)
 613		return;
 614
 615	/*
 616	 * Propagate recover_submit[] and recover_result[] to lvb:
 617	 * dlm_recoverd adds to recover_submit[] jids needing recovery
 618	 * gfs2_recover adds to recover_result[] journal recovery results
 619	 *
 620	 * set lvb bit for jids in recover_submit[] if the lvb has not
 621	 * yet been updated for the generation of the failure
 622	 *
 623	 * clear lvb bit for jids in recover_result[] if the result of
 624	 * the journal recovery is SUCCESS
 625	 */
 626
 627	error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
 628	if (error) {
 629		fs_err(sdp, "control lock EX error %d\n", error);
 630		return;
 631	}
 632
 633	control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
 634
 635	spin_lock(&ls->ls_recover_spin);
 636	if (block_gen != ls->ls_recover_block ||
 637	    start_gen != ls->ls_recover_start) {
 638		fs_info(sdp, "recover generation %u block1 %u %u\n",
 639			start_gen, block_gen, ls->ls_recover_block);
 640		spin_unlock(&ls->ls_recover_spin);
 641		control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
 642		return;
 643	}
 644
 645	recover_size = ls->ls_recover_size;
 646
 647	if (lvb_gen <= start_gen) {
 648		/*
 649		 * Clear lvb bits for jids we've successfully recovered.
 650		 * Because all nodes attempt to recover failed journals,
 651		 * a journal can be recovered multiple times successfully
 652		 * in succession.  Only the first will really do recovery,
 653		 * the others find it clean, but still report a successful
 654		 * recovery.  So, another node may have already recovered
 655		 * the jid and cleared the lvb bit for it.
 656		 */
 657		for (i = 0; i < recover_size; i++) {
 658			if (ls->ls_recover_result[i] != LM_RD_SUCCESS)
 659				continue;
 660
 661			ls->ls_recover_result[i] = 0;
 662
 663			if (!test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET))
 664				continue;
 665
 666			__clear_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
 667			write_lvb = 1;
 668		}
 669	}
 670
 671	if (lvb_gen == start_gen) {
 672		/*
 673		 * Failed slots before start_gen are already set in lvb.
 674		 */
 675		for (i = 0; i < recover_size; i++) {
 676			if (!ls->ls_recover_submit[i])
 677				continue;
 678			if (ls->ls_recover_submit[i] < lvb_gen)
 679				ls->ls_recover_submit[i] = 0;
 680		}
 681	} else if (lvb_gen < start_gen) {
 682		/*
 683		 * Failed slots before start_gen are not yet set in lvb.
 684		 */
 685		for (i = 0; i < recover_size; i++) {
 686			if (!ls->ls_recover_submit[i])
 687				continue;
 688			if (ls->ls_recover_submit[i] < start_gen) {
 689				ls->ls_recover_submit[i] = 0;
 690				__set_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
 691			}
 692		}
 693		/* even if there are no bits to set, we need to write the
 694		   latest generation to the lvb */
 695		write_lvb = 1;
 696	} else {
 697		/*
 698		 * we should be getting a recover_done() for lvb_gen soon
 699		 */
 700	}
 701	spin_unlock(&ls->ls_recover_spin);
 702
 703	if (write_lvb) {
 704		control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
 705		flags = DLM_LKF_CONVERT | DLM_LKF_VALBLK;
 706	} else {
 707		flags = DLM_LKF_CONVERT;
 708	}
 709
 710	error = control_lock(sdp, DLM_LOCK_NL, flags);
 711	if (error) {
 712		fs_err(sdp, "control lock NL error %d\n", error);
 713		return;
 714	}
 715
 716	/*
 717	 * Everyone will see jid bits set in the lvb, run gfs2_recover_set(),
 718	 * and clear a jid bit in the lvb if the recovery is a success.
 719	 * Eventually all journals will be recovered, all jid bits will
 720	 * be cleared in the lvb, and everyone will clear BLOCK_LOCKS.
 721	 */
 722
 723	for (i = 0; i < recover_size; i++) {
 724		if (test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET)) {
 725			fs_info(sdp, "recover generation %u jid %d\n",
 726				start_gen, i);
 727			gfs2_recover_set(sdp, i);
 728			recover_set++;
 729		}
 730	}
 731	if (recover_set)
 732		return;
 733
 734	/*
 735	 * No more jid bits set in lvb, all recovery is done, unblock locks
 736	 * (unless a new recover_prep callback has occured blocking locks
 737	 * again while working above)
 738	 */
 739
 740	spin_lock(&ls->ls_recover_spin);
 741	if (ls->ls_recover_block == block_gen &&
 742	    ls->ls_recover_start == start_gen) {
 743		clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
 744		spin_unlock(&ls->ls_recover_spin);
 745		fs_info(sdp, "recover generation %u done\n", start_gen);
 746		gfs2_glock_thaw(sdp);
 747	} else {
 748		fs_info(sdp, "recover generation %u block2 %u %u\n",
 749			start_gen, block_gen, ls->ls_recover_block);
 750		spin_unlock(&ls->ls_recover_spin);
 751	}
 752}
 753
 754static int control_mount(struct gfs2_sbd *sdp)
 755{
 756	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 
 757	uint32_t start_gen, block_gen, mount_gen, lvb_gen;
 758	int mounted_mode;
 759	int retries = 0;
 760	int error;
 761
 762	memset(&ls->ls_mounted_lksb, 0, sizeof(struct dlm_lksb));
 763	memset(&ls->ls_control_lksb, 0, sizeof(struct dlm_lksb));
 764	memset(&ls->ls_control_lvb, 0, GDLM_LVB_SIZE);
 765	ls->ls_control_lksb.sb_lvbptr = ls->ls_control_lvb;
 766	init_completion(&ls->ls_sync_wait);
 767
 768	set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
 769
 770	error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_VALBLK);
 771	if (error) {
 772		fs_err(sdp, "control_mount control_lock NL error %d\n", error);
 773		return error;
 774	}
 775
 776	error = mounted_lock(sdp, DLM_LOCK_NL, 0);
 777	if (error) {
 778		fs_err(sdp, "control_mount mounted_lock NL error %d\n", error);
 779		control_unlock(sdp);
 780		return error;
 781	}
 782	mounted_mode = DLM_LOCK_NL;
 783
 784restart:
 785	if (retries++ && signal_pending(current)) {
 786		error = -EINTR;
 787		goto fail;
 788	}
 789
 790	/*
 791	 * We always start with both locks in NL. control_lock is
 792	 * demoted to NL below so we don't need to do it here.
 793	 */
 794
 795	if (mounted_mode != DLM_LOCK_NL) {
 796		error = mounted_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
 797		if (error)
 798			goto fail;
 799		mounted_mode = DLM_LOCK_NL;
 800	}
 801
 802	/*
 803	 * Other nodes need to do some work in dlm recovery and gfs2_control
 804	 * before the recover_done and control_lock will be ready for us below.
 805	 * A delay here is not required but often avoids having to retry.
 806	 */
 807
 808	msleep_interruptible(500);
 809
 810	/*
 811	 * Acquire control_lock in EX and mounted_lock in either EX or PR.
 812	 * control_lock lvb keeps track of any pending journal recoveries.
 813	 * mounted_lock indicates if any other nodes have the fs mounted.
 814	 */
 815
 816	error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE|DLM_LKF_VALBLK);
 817	if (error == -EAGAIN) {
 818		goto restart;
 819	} else if (error) {
 820		fs_err(sdp, "control_mount control_lock EX error %d\n", error);
 821		goto fail;
 822	}
 823
 824	error = mounted_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
 825	if (!error) {
 826		mounted_mode = DLM_LOCK_EX;
 827		goto locks_done;
 828	} else if (error != -EAGAIN) {
 829		fs_err(sdp, "control_mount mounted_lock EX error %d\n", error);
 830		goto fail;
 831	}
 832
 833	error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
 834	if (!error) {
 835		mounted_mode = DLM_LOCK_PR;
 836		goto locks_done;
 837	} else {
 838		/* not even -EAGAIN should happen here */
 839		fs_err(sdp, "control_mount mounted_lock PR error %d\n", error);
 840		goto fail;
 841	}
 842
 843locks_done:
 844	/*
 845	 * If we got both locks above in EX, then we're the first mounter.
 846	 * If not, then we need to wait for the control_lock lvb to be
 847	 * updated by other mounted nodes to reflect our mount generation.
 848	 *
 849	 * In simple first mounter cases, first mounter will see zero lvb_gen,
 850	 * but in cases where all existing nodes leave/fail before mounting
 851	 * nodes finish control_mount, then all nodes will be mounting and
 852	 * lvb_gen will be non-zero.
 853	 */
 854
 855	control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
 856
 857	if (lvb_gen == 0xFFFFFFFF) {
 858		/* special value to force mount attempts to fail */
 859		fs_err(sdp, "control_mount control_lock disabled\n");
 860		error = -EINVAL;
 861		goto fail;
 862	}
 863
 864	if (mounted_mode == DLM_LOCK_EX) {
 865		/* first mounter, keep both EX while doing first recovery */
 866		spin_lock(&ls->ls_recover_spin);
 867		clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
 868		set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
 869		set_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
 870		spin_unlock(&ls->ls_recover_spin);
 871		fs_info(sdp, "first mounter control generation %u\n", lvb_gen);
 872		return 0;
 873	}
 874
 875	error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
 876	if (error)
 877		goto fail;
 878
 879	/*
 880	 * We are not first mounter, now we need to wait for the control_lock
 881	 * lvb generation to be >= the generation from our first recover_done
 882	 * and all lvb bits to be clear (no pending journal recoveries.)
 883	 */
 884
 885	if (!all_jid_bits_clear(ls->ls_lvb_bits)) {
 886		/* journals need recovery, wait until all are clear */
 887		fs_info(sdp, "control_mount wait for journal recovery\n");
 888		goto restart;
 889	}
 890
 891	spin_lock(&ls->ls_recover_spin);
 892	block_gen = ls->ls_recover_block;
 893	start_gen = ls->ls_recover_start;
 894	mount_gen = ls->ls_recover_mount;
 895
 896	if (lvb_gen < mount_gen) {
 897		/* wait for mounted nodes to update control_lock lvb to our
 898		   generation, which might include new recovery bits set */
 899		fs_info(sdp, "control_mount wait1 block %u start %u mount %u "
 900			"lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
 901			lvb_gen, ls->ls_recover_flags);
 902		spin_unlock(&ls->ls_recover_spin);
 903		goto restart;
 904	}
 905
 906	if (lvb_gen != start_gen) {
 907		/* wait for mounted nodes to update control_lock lvb to the
 908		   latest recovery generation */
 909		fs_info(sdp, "control_mount wait2 block %u start %u mount %u "
 910			"lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
 911			lvb_gen, ls->ls_recover_flags);
 912		spin_unlock(&ls->ls_recover_spin);
 913		goto restart;
 914	}
 915
 916	if (block_gen == start_gen) {
 917		/* dlm recovery in progress, wait for it to finish */
 918		fs_info(sdp, "control_mount wait3 block %u start %u mount %u "
 919			"lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
 920			lvb_gen, ls->ls_recover_flags);
 921		spin_unlock(&ls->ls_recover_spin);
 922		goto restart;
 923	}
 924
 925	clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
 926	set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
 927	memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
 928	memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
 929	spin_unlock(&ls->ls_recover_spin);
 930	return 0;
 931
 932fail:
 933	mounted_unlock(sdp);
 934	control_unlock(sdp);
 935	return error;
 936}
 937
 
 
 
 
 
 
 938static int control_first_done(struct gfs2_sbd *sdp)
 939{
 940	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 
 941	uint32_t start_gen, block_gen;
 942	int error;
 943
 944restart:
 945	spin_lock(&ls->ls_recover_spin);
 946	start_gen = ls->ls_recover_start;
 947	block_gen = ls->ls_recover_block;
 948
 949	if (test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags) ||
 950	    !test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
 951	    !test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
 952		/* sanity check, should not happen */
 953		fs_err(sdp, "control_first_done start %u block %u flags %lx\n",
 954		       start_gen, block_gen, ls->ls_recover_flags);
 955		spin_unlock(&ls->ls_recover_spin);
 956		control_unlock(sdp);
 957		return -1;
 958	}
 959
 960	if (start_gen == block_gen) {
 961		/*
 962		 * Wait for the end of a dlm recovery cycle to switch from
 963		 * first mounter recovery.  We can ignore any recover_slot
 964		 * callbacks between the recover_prep and next recover_done
 965		 * because we are still the first mounter and any failed nodes
 966		 * have not fully mounted, so they don't need recovery.
 967		 */
 968		spin_unlock(&ls->ls_recover_spin);
 969		fs_info(sdp, "control_first_done wait gen %u\n", start_gen);
 970
 971		wait_on_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY,
 972			    TASK_UNINTERRUPTIBLE);
 973		goto restart;
 974	}
 975
 976	clear_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
 977	set_bit(DFL_FIRST_MOUNT_DONE, &ls->ls_recover_flags);
 978	memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
 979	memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
 980	spin_unlock(&ls->ls_recover_spin);
 981
 982	memset(ls->ls_lvb_bits, 0, GDLM_LVB_SIZE);
 983	control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
 984
 985	error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT);
 986	if (error)
 987		fs_err(sdp, "control_first_done mounted PR error %d\n", error);
 988
 989	error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
 990	if (error)
 991		fs_err(sdp, "control_first_done control NL error %d\n", error);
 992
 993	return error;
 994}
 995
 996/*
 997 * Expand static jid arrays if necessary (by increments of RECOVER_SIZE_INC)
 998 * to accomodate the largest slot number.  (NB dlm slot numbers start at 1,
 999 * gfs2 jids start at 0, so jid = slot - 1)
1000 */
1001
1002#define RECOVER_SIZE_INC 16
1003
1004static int set_recover_size(struct gfs2_sbd *sdp, struct dlm_slot *slots,
1005			    int num_slots)
1006{
1007	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1008	uint32_t *submit = NULL;
1009	uint32_t *result = NULL;
1010	uint32_t old_size, new_size;
1011	int i, max_jid;
1012
1013	if (!ls->ls_lvb_bits) {
1014		ls->ls_lvb_bits = kzalloc(GDLM_LVB_SIZE, GFP_NOFS);
1015		if (!ls->ls_lvb_bits)
1016			return -ENOMEM;
1017	}
1018
1019	max_jid = 0;
1020	for (i = 0; i < num_slots; i++) {
1021		if (max_jid < slots[i].slot - 1)
1022			max_jid = slots[i].slot - 1;
1023	}
1024
1025	old_size = ls->ls_recover_size;
1026
1027	if (old_size >= max_jid + 1)
1028		return 0;
1029
1030	new_size = old_size + RECOVER_SIZE_INC;
1031
1032	submit = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1033	result = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1034	if (!submit || !result) {
1035		kfree(submit);
1036		kfree(result);
1037		return -ENOMEM;
1038	}
1039
1040	spin_lock(&ls->ls_recover_spin);
1041	memcpy(submit, ls->ls_recover_submit, old_size * sizeof(uint32_t));
1042	memcpy(result, ls->ls_recover_result, old_size * sizeof(uint32_t));
1043	kfree(ls->ls_recover_submit);
1044	kfree(ls->ls_recover_result);
1045	ls->ls_recover_submit = submit;
1046	ls->ls_recover_result = result;
1047	ls->ls_recover_size = new_size;
1048	spin_unlock(&ls->ls_recover_spin);
1049	return 0;
1050}
1051
1052static void free_recover_size(struct lm_lockstruct *ls)
1053{
1054	kfree(ls->ls_lvb_bits);
1055	kfree(ls->ls_recover_submit);
1056	kfree(ls->ls_recover_result);
1057	ls->ls_recover_submit = NULL;
1058	ls->ls_recover_result = NULL;
1059	ls->ls_recover_size = 0;
1060	ls->ls_lvb_bits = NULL;
1061}
1062
1063/* dlm calls before it does lock recovery */
1064
1065static void gdlm_recover_prep(void *arg)
1066{
1067	struct gfs2_sbd *sdp = arg;
1068	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1069
1070	spin_lock(&ls->ls_recover_spin);
1071	ls->ls_recover_block = ls->ls_recover_start;
1072	set_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1073
1074	if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
1075	     test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1076		spin_unlock(&ls->ls_recover_spin);
1077		return;
1078	}
1079	set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
1080	spin_unlock(&ls->ls_recover_spin);
1081}
1082
1083/* dlm calls after recover_prep has been completed on all lockspace members;
1084   identifies slot/jid of failed member */
1085
1086static void gdlm_recover_slot(void *arg, struct dlm_slot *slot)
1087{
1088	struct gfs2_sbd *sdp = arg;
1089	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1090	int jid = slot->slot - 1;
1091
1092	spin_lock(&ls->ls_recover_spin);
1093	if (ls->ls_recover_size < jid + 1) {
1094		fs_err(sdp, "recover_slot jid %d gen %u short size %d\n",
1095		       jid, ls->ls_recover_block, ls->ls_recover_size);
1096		spin_unlock(&ls->ls_recover_spin);
1097		return;
1098	}
1099
1100	if (ls->ls_recover_submit[jid]) {
1101		fs_info(sdp, "recover_slot jid %d gen %u prev %u\n",
1102			jid, ls->ls_recover_block, ls->ls_recover_submit[jid]);
1103	}
1104	ls->ls_recover_submit[jid] = ls->ls_recover_block;
1105	spin_unlock(&ls->ls_recover_spin);
1106}
1107
1108/* dlm calls after recover_slot and after it completes lock recovery */
1109
1110static void gdlm_recover_done(void *arg, struct dlm_slot *slots, int num_slots,
1111			      int our_slot, uint32_t generation)
1112{
1113	struct gfs2_sbd *sdp = arg;
1114	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1115
1116	/* ensure the ls jid arrays are large enough */
1117	set_recover_size(sdp, slots, num_slots);
1118
1119	spin_lock(&ls->ls_recover_spin);
1120	ls->ls_recover_start = generation;
1121
1122	if (!ls->ls_recover_mount) {
1123		ls->ls_recover_mount = generation;
1124		ls->ls_jid = our_slot - 1;
1125	}
1126
1127	if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1128		queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work, 0);
1129
1130	clear_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1131	smp_mb__after_atomic();
1132	wake_up_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY);
1133	spin_unlock(&ls->ls_recover_spin);
1134}
1135
1136/* gfs2_recover thread has a journal recovery result */
1137
1138static void gdlm_recovery_result(struct gfs2_sbd *sdp, unsigned int jid,
1139				 unsigned int result)
1140{
1141	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1142
1143	if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1144		return;
1145
1146	/* don't care about the recovery of own journal during mount */
1147	if (jid == ls->ls_jid)
1148		return;
1149
1150	spin_lock(&ls->ls_recover_spin);
1151	if (test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1152		spin_unlock(&ls->ls_recover_spin);
1153		return;
1154	}
1155	if (ls->ls_recover_size < jid + 1) {
1156		fs_err(sdp, "recovery_result jid %d short size %d\n",
1157		       jid, ls->ls_recover_size);
1158		spin_unlock(&ls->ls_recover_spin);
1159		return;
1160	}
1161
1162	fs_info(sdp, "recover jid %d result %s\n", jid,
1163		result == LM_RD_GAVEUP ? "busy" : "success");
1164
1165	ls->ls_recover_result[jid] = result;
1166
1167	/* GAVEUP means another node is recovering the journal; delay our
1168	   next attempt to recover it, to give the other node a chance to
1169	   finish before trying again */
1170
1171	if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1172		queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work,
1173				   result == LM_RD_GAVEUP ? HZ : 0);
1174	spin_unlock(&ls->ls_recover_spin);
1175}
1176
1177static const struct dlm_lockspace_ops gdlm_lockspace_ops = {
1178	.recover_prep = gdlm_recover_prep,
1179	.recover_slot = gdlm_recover_slot,
1180	.recover_done = gdlm_recover_done,
1181};
1182
1183static int gdlm_mount(struct gfs2_sbd *sdp, const char *table)
1184{
1185	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1186	char cluster[GFS2_LOCKNAME_LEN];
1187	const char *fsname;
1188	uint32_t flags;
1189	int error, ops_result;
1190
1191	/*
1192	 * initialize everything
1193	 */
1194
1195	INIT_DELAYED_WORK(&sdp->sd_control_work, gfs2_control_func);
1196	spin_lock_init(&ls->ls_recover_spin);
1197	ls->ls_recover_flags = 0;
1198	ls->ls_recover_mount = 0;
1199	ls->ls_recover_start = 0;
1200	ls->ls_recover_block = 0;
1201	ls->ls_recover_size = 0;
1202	ls->ls_recover_submit = NULL;
1203	ls->ls_recover_result = NULL;
1204	ls->ls_lvb_bits = NULL;
1205
1206	error = set_recover_size(sdp, NULL, 0);
1207	if (error)
1208		goto fail;
1209
1210	/*
1211	 * prepare dlm_new_lockspace args
1212	 */
1213
1214	fsname = strchr(table, ':');
1215	if (!fsname) {
1216		fs_info(sdp, "no fsname found\n");
1217		error = -EINVAL;
1218		goto fail_free;
1219	}
1220	memset(cluster, 0, sizeof(cluster));
1221	memcpy(cluster, table, strlen(table) - strlen(fsname));
1222	fsname++;
1223
1224	flags = DLM_LSFL_FS | DLM_LSFL_NEWEXCL;
1225
1226	/*
1227	 * create/join lockspace
1228	 */
1229
1230	error = dlm_new_lockspace(fsname, cluster, flags, GDLM_LVB_SIZE,
1231				  &gdlm_lockspace_ops, sdp, &ops_result,
1232				  &ls->ls_dlm);
1233	if (error) {
1234		fs_err(sdp, "dlm_new_lockspace error %d\n", error);
1235		goto fail_free;
1236	}
1237
1238	if (ops_result < 0) {
1239		/*
1240		 * dlm does not support ops callbacks,
1241		 * old dlm_controld/gfs_controld are used, try without ops.
1242		 */
1243		fs_info(sdp, "dlm lockspace ops not used\n");
1244		free_recover_size(ls);
1245		set_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags);
1246		return 0;
1247	}
1248
1249	if (!test_bit(SDF_NOJOURNALID, &sdp->sd_flags)) {
1250		fs_err(sdp, "dlm lockspace ops disallow jid preset\n");
1251		error = -EINVAL;
1252		goto fail_release;
1253	}
1254
1255	/*
1256	 * control_mount() uses control_lock to determine first mounter,
1257	 * and for later mounts, waits for any recoveries to be cleared.
1258	 */
1259
1260	error = control_mount(sdp);
1261	if (error) {
1262		fs_err(sdp, "mount control error %d\n", error);
1263		goto fail_release;
1264	}
1265
1266	ls->ls_first = !!test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
1267	clear_bit(SDF_NOJOURNALID, &sdp->sd_flags);
1268	smp_mb__after_atomic();
1269	wake_up_bit(&sdp->sd_flags, SDF_NOJOURNALID);
1270	return 0;
1271
1272fail_release:
1273	dlm_release_lockspace(ls->ls_dlm, 2);
1274fail_free:
1275	free_recover_size(ls);
1276fail:
1277	return error;
1278}
1279
1280static void gdlm_first_done(struct gfs2_sbd *sdp)
1281{
1282	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1283	int error;
1284
1285	if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1286		return;
1287
1288	error = control_first_done(sdp);
1289	if (error)
1290		fs_err(sdp, "mount first_done error %d\n", error);
1291}
1292
1293static void gdlm_unmount(struct gfs2_sbd *sdp)
1294{
1295	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1296
1297	if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1298		goto release;
1299
1300	/* wait for gfs2_control_wq to be done with this mount */
1301
1302	spin_lock(&ls->ls_recover_spin);
1303	set_bit(DFL_UNMOUNT, &ls->ls_recover_flags);
1304	spin_unlock(&ls->ls_recover_spin);
1305	flush_delayed_work(&sdp->sd_control_work);
1306
1307	/* mounted_lock and control_lock will be purged in dlm recovery */
1308release:
1309	if (ls->ls_dlm) {
1310		dlm_release_lockspace(ls->ls_dlm, 2);
1311		ls->ls_dlm = NULL;
1312	}
1313
1314	free_recover_size(ls);
1315}
1316
1317static const match_table_t dlm_tokens = {
1318	{ Opt_jid, "jid=%d"},
1319	{ Opt_id, "id=%d"},
1320	{ Opt_first, "first=%d"},
1321	{ Opt_nodir, "nodir=%d"},
1322	{ Opt_err, NULL },
1323};
1324
1325const struct lm_lockops gfs2_dlm_ops = {
1326	.lm_proto_name = "lock_dlm",
1327	.lm_mount = gdlm_mount,
1328	.lm_first_done = gdlm_first_done,
1329	.lm_recovery_result = gdlm_recovery_result,
1330	.lm_unmount = gdlm_unmount,
1331	.lm_put_lock = gdlm_put_lock,
1332	.lm_lock = gdlm_lock,
1333	.lm_cancel = gdlm_cancel,
1334	.lm_tokens = &dlm_tokens,
1335};
1336