lock_dlm.c

   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   4 * Copyright 2004-2011 Red Hat, Inc.
   5 */
   6
   7#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   8
   9#include <linux/fs.h>
  10#include <linux/dlm.h>
  11#include <linux/slab.h>
  12#include <linux/types.h>
  13#include <linux/delay.h>
  14#include <linux/gfs2_ondisk.h>
  15#include <linux/sched/signal.h>
  16
  17#include "incore.h"
  18#include "util.h"
  19#include "sys.h"
  20#include "trace_gfs2.h"
  21
  22/**
  23 * gfs2_update_stats - Update time based stats
  24 * @s: The stats to update (local or global)
  25 * @index: The index inside @s
  26 * @sample: New data to include
  27 */
  28static inline void gfs2_update_stats(struct gfs2_lkstats *s, unsigned index,
  29				     s64 sample)
  30{
  31	/*
  32	 * @delta is the difference between the current rtt sample and the
  33	 * running average srtt. We add 1/8 of that to the srtt in order to
  34	 * update the current srtt estimate. The variance estimate is a bit
  35	 * more complicated. We subtract the current variance estimate from
  36	 * the abs value of the @delta and add 1/4 of that to the running
  37	 * total.  That's equivalent to 3/4 of the current variance
  38	 * estimate plus 1/4 of the abs of @delta.
  39	 *
  40	 * Note that the index points at the array entry containing the
  41	 * smoothed mean value, and the variance is always in the following
  42	 * entry
  43	 *
  44	 * Reference: TCP/IP Illustrated, vol 2, p. 831,832
  45	 * All times are in units of integer nanoseconds. Unlike the TCP/IP
  46	 * case, they are not scaled fixed point.
  47	 */
  48
  49	s64 delta = sample - s->stats[index];
  50	s->stats[index] += (delta >> 3);
  51	index++;
  52	s->stats[index] += (s64)(abs(delta) - s->stats[index]) >> 2;
  53}
  54
  55/**
  56 * gfs2_update_reply_times - Update locking statistics
  57 * @gl: The glock to update
  58 * @blocking: The operation may have been blocking
  59 *
  60 * This assumes that gl->gl_dstamp has been set earlier.
  61 *
  62 * The rtt (lock round trip time) is an estimate of the time
  63 * taken to perform a dlm lock request. We update it on each
  64 * reply from the dlm.
  65 *
  66 * The blocking flag is set on the glock for all dlm requests
  67 * which may potentially block due to lock requests from other nodes.
  68 * DLM requests where the current lock state is exclusive, the
  69 * requested state is null (or unlocked) or where the TRY or
  70 * TRY_1CB flags are set are classified as non-blocking. All
  71 * other DLM requests are counted as (potentially) blocking.
  72 */
  73static inline void gfs2_update_reply_times(struct gfs2_glock *gl,
  74					   bool blocking)
  75{
  76	struct gfs2_pcpu_lkstats *lks;
  77	const unsigned gltype = gl->gl_name.ln_type;
  78	unsigned index = blocking ? GFS2_LKS_SRTTB : GFS2_LKS_SRTT;
  79	s64 rtt;
  80
  81	preempt_disable();
  82	rtt = ktime_to_ns(ktime_sub(ktime_get_real(), gl->gl_dstamp));
  83	lks = this_cpu_ptr(gl->gl_name.ln_sbd->sd_lkstats);
  84	gfs2_update_stats(&gl->gl_stats, index, rtt);		/* Local */
  85	gfs2_update_stats(&lks->lkstats[gltype], index, rtt);	/* Global */
  86	preempt_enable();
  87
  88	trace_gfs2_glock_lock_time(gl, rtt);
  89}
  90
  91/**
  92 * gfs2_update_request_times - Update locking statistics
  93 * @gl: The glock to update
  94 *
  95 * The irt (lock inter-request times) measures the average time
  96 * between requests to the dlm. It is updated immediately before
  97 * each dlm call.
  98 */
  99
 100static inline void gfs2_update_request_times(struct gfs2_glock *gl)
 101{
 102	struct gfs2_pcpu_lkstats *lks;
 103	const unsigned gltype = gl->gl_name.ln_type;
 104	ktime_t dstamp;
 105	s64 irt;
 106
 107	preempt_disable();
 108	dstamp = gl->gl_dstamp;
 109	gl->gl_dstamp = ktime_get_real();
 110	irt = ktime_to_ns(ktime_sub(gl->gl_dstamp, dstamp));
 111	lks = this_cpu_ptr(gl->gl_name.ln_sbd->sd_lkstats);
 112	gfs2_update_stats(&gl->gl_stats, GFS2_LKS_SIRT, irt);		/* Local */
 113	gfs2_update_stats(&lks->lkstats[gltype], GFS2_LKS_SIRT, irt);	/* Global */
 114	preempt_enable();
 115}
 116 
 117static void gdlm_ast(void *arg)
 118{
 119	struct gfs2_glock *gl = arg;
 120	bool blocking;
 121	unsigned ret;
 122
 123	blocking = test_bit(GLF_BLOCKING, &gl->gl_flags);
 124	gfs2_update_reply_times(gl, blocking);
 125	clear_bit(GLF_BLOCKING, &gl->gl_flags);
 126
 127	/* If the glock is dead, we only react to a dlm_unlock() reply. */
 128	if (__lockref_is_dead(&gl->gl_lockref) &&
 129	    gl->gl_lksb.sb_status != -DLM_EUNLOCK)
 130		return;
 131
 132	BUG_ON(gl->gl_lksb.sb_flags & DLM_SBF_DEMOTED);
 133
 134	if ((gl->gl_lksb.sb_flags & DLM_SBF_VALNOTVALID) && gl->gl_lksb.sb_lvbptr)
 135		memset(gl->gl_lksb.sb_lvbptr, 0, GDLM_LVB_SIZE);
 136
 137	switch (gl->gl_lksb.sb_status) {
 138	case -DLM_EUNLOCK: /* Unlocked, so glock can be freed */
 139		gfs2_glock_free(gl);
 140		return;
 141	case -DLM_ECANCEL: /* Cancel while getting lock */
 142		ret = LM_OUT_CANCELED;
 143		goto out;
 144	case -EAGAIN: /* Try lock fails */
 145		ret = LM_OUT_TRY_AGAIN;
 146		goto out;
 147	case -EDEADLK: /* Deadlock detected */
 148		ret = LM_OUT_DEADLOCK;
 149		goto out;
 150	case -ETIMEDOUT: /* Canceled due to timeout */
 151		ret = LM_OUT_ERROR;
 152		goto out;
 153	case 0: /* Success */
 154		break;
 155	default: /* Something unexpected */
 156		BUG();
 157	}
 158
 159	ret = gl->gl_req;
 160
 161	/*
 162	 * The GLF_INITIAL flag is initially set for new glocks.  Upon the
 163	 * first successful new (non-conversion) request, we clear this flag to
 164	 * indicate that a DLM lock exists and that gl->gl_lksb.sb_lkid is the
 165	 * identifier to use for identifying it.
 166	 *
 167	 * Any failed initial requests do not create a DLM lock, so we ignore
 168	 * the gl->gl_lksb.sb_lkid values that come with such requests.
 169	 */
 170
 171	clear_bit(GLF_INITIAL, &gl->gl_flags);
 172	gfs2_glock_complete(gl, ret);
 173	return;
 174out:
 175	if (test_bit(GLF_INITIAL, &gl->gl_flags))
 176		gl->gl_lksb.sb_lkid = 0;
 177	gfs2_glock_complete(gl, ret);
 178}
 179
 180static void gdlm_bast(void *arg, int mode)
 181{
 182	struct gfs2_glock *gl = arg;
 183
 184	if (__lockref_is_dead(&gl->gl_lockref))
 185		return;
 186
 187	switch (mode) {
 188	case DLM_LOCK_EX:
 189		gfs2_glock_cb(gl, LM_ST_UNLOCKED);
 190		break;
 191	case DLM_LOCK_CW:
 192		gfs2_glock_cb(gl, LM_ST_DEFERRED);
 193		break;
 194	case DLM_LOCK_PR:
 195		gfs2_glock_cb(gl, LM_ST_SHARED);
 196		break;
 197	default:
 198		fs_err(gl->gl_name.ln_sbd, "unknown bast mode %d\n", mode);
 199		BUG();
 200	}
 201}
 202
 203/* convert gfs lock-state to dlm lock-mode */
 204
 205static int make_mode(struct gfs2_sbd *sdp, const unsigned int lmstate)
 206{
 207	switch (lmstate) {
 208	case LM_ST_UNLOCKED:
 209		return DLM_LOCK_NL;
 210	case LM_ST_EXCLUSIVE:
 211		return DLM_LOCK_EX;
 212	case LM_ST_DEFERRED:
 213		return DLM_LOCK_CW;
 214	case LM_ST_SHARED:
 215		return DLM_LOCK_PR;
 216	}
 217	fs_err(sdp, "unknown LM state %d\n", lmstate);
 218	BUG();
 219	return -1;
 220}
 221
 222/* Taken from fs/dlm/lock.c. */
 223
 224static bool middle_conversion(int cur, int req)
 225{
 226	return (cur == DLM_LOCK_PR && req == DLM_LOCK_CW) ||
 227	       (cur == DLM_LOCK_CW && req == DLM_LOCK_PR);
 228}
 229
 230static bool down_conversion(int cur, int req)
 231{
 232	return !middle_conversion(cur, req) && req < cur;
 233}
 234
 235static u32 make_flags(struct gfs2_glock *gl, const unsigned int gfs_flags,
 236		      const int req, bool blocking)
 237{
 238	u32 lkf = 0;
 239
 240	if (gl->gl_lksb.sb_lvbptr)
 241		lkf |= DLM_LKF_VALBLK;
 242
 243	if (gfs_flags & LM_FLAG_TRY)
 244		lkf |= DLM_LKF_NOQUEUE;
 245
 246	if (gfs_flags & LM_FLAG_TRY_1CB) {
 247		lkf |= DLM_LKF_NOQUEUE;
 248		lkf |= DLM_LKF_NOQUEUEBAST;
 249	}
 250
 251	if (!test_bit(GLF_INITIAL, &gl->gl_flags)) {
 252		lkf |= DLM_LKF_CONVERT;
 253
 254		/*
 255		 * The DLM_LKF_QUECVT flag needs to be set for "first come,
 256		 * first served" semantics, but it must only be set for
 257		 * "upward" lock conversions or else DLM will reject the
 258		 * request as invalid.
 259		 */
 260		if (blocking)
 261			lkf |= DLM_LKF_QUECVT;
 262	}
 263
 264	return lkf;
 265}
 266
 267static void gfs2_reverse_hex(char *c, u64 value)
 268{
 269	*c = '0';
 270	while (value) {
 271		*c-- = hex_asc[value & 0x0f];
 272		value >>= 4;
 273	}
 274}
 275
 276static int gdlm_lock(struct gfs2_glock *gl, unsigned int req_state,
 277		     unsigned int flags)
 278{
 279	struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
 280	bool blocking;
 281	int cur, req;
 282	u32 lkf;
 283	char strname[GDLM_STRNAME_BYTES] = "";
 284	int error;
 285
 286	gl->gl_req = req_state;
 287	cur = make_mode(gl->gl_name.ln_sbd, gl->gl_state);
 288	req = make_mode(gl->gl_name.ln_sbd, req_state);
 289	blocking = !down_conversion(cur, req) &&
 290		   !(flags & (LM_FLAG_TRY|LM_FLAG_TRY_1CB));
 291	lkf = make_flags(gl, flags, req, blocking);
 292	if (blocking)
 293		set_bit(GLF_BLOCKING, &gl->gl_flags);
 294	gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
 295	gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
 296	if (test_bit(GLF_INITIAL, &gl->gl_flags)) {
 297		memset(strname, ' ', GDLM_STRNAME_BYTES - 1);
 298		strname[GDLM_STRNAME_BYTES - 1] = '\0';
 299		gfs2_reverse_hex(strname + 7, gl->gl_name.ln_type);
 300		gfs2_reverse_hex(strname + 23, gl->gl_name.ln_number);
 301		gl->gl_dstamp = ktime_get_real();
 302	} else {
 303		gfs2_update_request_times(gl);
 304	}
 305	/*
 306	 * Submit the actual lock request.
 307	 */
 308
 309again:
 310	down_read(&ls->ls_sem);
 311	error = -ENODEV;
 312	if (likely(ls->ls_dlm != NULL)) {
 313		error = dlm_lock(ls->ls_dlm, req, &gl->gl_lksb, lkf, strname,
 314				GDLM_STRNAME_BYTES - 1, 0, gdlm_ast, gl, gdlm_bast);
 315	}
 316	up_read(&ls->ls_sem);
 317	if (error == -EBUSY) {
 318		msleep(20);
 319		goto again;
 320	}
 321	return error;
 322}
 323
 324static void gdlm_put_lock(struct gfs2_glock *gl)
 325{
 326	struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 327	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 328	uint32_t flags = 0;
 329	int error;
 330
 331	BUG_ON(!__lockref_is_dead(&gl->gl_lockref));
 332
 333	if (test_bit(GLF_INITIAL, &gl->gl_flags)) {
 334		gfs2_glock_free(gl);
 335		return;
 336	}
 337
 338	gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
 339	gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
 340	gfs2_update_request_times(gl);
 341
 342	/*
 343	 * When the lockspace is released, all remaining glocks will be
 344	 * unlocked automatically.  This is more efficient than unlocking them
 345	 * individually, but when the lock is held in DLM_LOCK_EX or
 346	 * DLM_LOCK_PW mode, the lock value block (LVB) would be lost.
 347	 */
 348
 349	if (test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags) &&
 350	    (!gl->gl_lksb.sb_lvbptr || gl->gl_state != LM_ST_EXCLUSIVE)) {
 351		gfs2_glock_free_later(gl);
 352		return;
 353	}
 354
 355	if (gl->gl_lksb.sb_lvbptr)
 356		flags |= DLM_LKF_VALBLK;
 357
 358again:
 359	down_read(&ls->ls_sem);
 360	error = -ENODEV;
 361	if (likely(ls->ls_dlm != NULL)) {
 362		error = dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, flags,
 363				   NULL, gl);
 364	}
 365	up_read(&ls->ls_sem);
 366	if (error == -EBUSY) {
 367		msleep(20);
 368		goto again;
 369	}
 370
 371	if (error == -ENODEV) {
 372		gfs2_glock_free(gl);
 373		return;
 374	}
 375
 376	if (error) {
 377		fs_err(sdp, "gdlm_unlock %x,%llx err=%d\n",
 378		       gl->gl_name.ln_type,
 379		       (unsigned long long)gl->gl_name.ln_number, error);
 380	}
 381}
 382
 383static void gdlm_cancel(struct gfs2_glock *gl)
 384{
 385	struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
 386
 387	down_read(&ls->ls_sem);
 388	if (likely(ls->ls_dlm != NULL)) {
 389		dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_CANCEL, NULL, gl);
 390	}
 391	up_read(&ls->ls_sem);
 392}
 393
 394/*
 395 * dlm/gfs2 recovery coordination using dlm_recover callbacks
 396 *
 397 *  1. dlm_controld sees lockspace members change
 398 *  2. dlm_controld blocks dlm-kernel locking activity
 399 *  3. dlm_controld within dlm-kernel notifies gfs2 (recover_prep)
 400 *  4. dlm_controld starts and finishes its own user level recovery
 401 *  5. dlm_controld starts dlm-kernel dlm_recoverd to do kernel recovery
 402 *  6. dlm_recoverd notifies gfs2 of failed nodes (recover_slot)
 403 *  7. dlm_recoverd does its own lock recovery
 404 *  8. dlm_recoverd unblocks dlm-kernel locking activity
 405 *  9. dlm_recoverd notifies gfs2 when done (recover_done with new generation)
 406 * 10. gfs2_control updates control_lock lvb with new generation and jid bits
 407 * 11. gfs2_control enqueues journals for gfs2_recover to recover (maybe none)
 408 * 12. gfs2_recover dequeues and recovers journals of failed nodes
 409 * 13. gfs2_recover provides recovery results to gfs2_control (recovery_result)
 410 * 14. gfs2_control updates control_lock lvb jid bits for recovered journals
 411 * 15. gfs2_control unblocks normal locking when all journals are recovered
 412 *
 413 * - failures during recovery
 414 *
 415 * recover_prep() may set BLOCK_LOCKS (step 3) again before gfs2_control
 416 * clears BLOCK_LOCKS (step 15), e.g. another node fails while still
 417 * recovering for a prior failure.  gfs2_control needs a way to detect
 418 * this so it can leave BLOCK_LOCKS set in step 15.  This is managed using
 419 * the recover_block and recover_start values.
 420 *
 421 * recover_done() provides a new lockspace generation number each time it
 422 * is called (step 9).  This generation number is saved as recover_start.
 423 * When recover_prep() is called, it sets BLOCK_LOCKS and sets
 424 * recover_block = recover_start.  So, while recover_block is equal to
 425 * recover_start, BLOCK_LOCKS should remain set.  (recover_spin must
 426 * be held around the BLOCK_LOCKS/recover_block/recover_start logic.)
 427 *
 428 * - more specific gfs2 steps in sequence above
 429 *
 430 *  3. recover_prep sets BLOCK_LOCKS and sets recover_block = recover_start
 431 *  6. recover_slot records any failed jids (maybe none)
 432 *  9. recover_done sets recover_start = new generation number
 433 * 10. gfs2_control sets control_lock lvb = new gen + bits for failed jids
 434 * 12. gfs2_recover does journal recoveries for failed jids identified above
 435 * 14. gfs2_control clears control_lock lvb bits for recovered jids
 436 * 15. gfs2_control checks if recover_block == recover_start (step 3 occured
 437 *     again) then do nothing, otherwise if recover_start > recover_block
 438 *     then clear BLOCK_LOCKS.
 439 *
 440 * - parallel recovery steps across all nodes
 441 *
 442 * All nodes attempt to update the control_lock lvb with the new generation
 443 * number and jid bits, but only the first to get the control_lock EX will
 444 * do so; others will see that it's already done (lvb already contains new
 445 * generation number.)
 446 *
 447 * . All nodes get the same recover_prep/recover_slot/recover_done callbacks
 448 * . All nodes attempt to set control_lock lvb gen + bits for the new gen
 449 * . One node gets control_lock first and writes the lvb, others see it's done
 450 * . All nodes attempt to recover jids for which they see control_lock bits set
 451 * . One node succeeds for a jid, and that one clears the jid bit in the lvb
 452 * . All nodes will eventually see all lvb bits clear and unblock locks
 453 *
 454 * - is there a problem with clearing an lvb bit that should be set
 455 *   and missing a journal recovery?
 456 *
 457 * 1. jid fails
 458 * 2. lvb bit set for step 1
 459 * 3. jid recovered for step 1
 460 * 4. jid taken again (new mount)
 461 * 5. jid fails (for step 4)
 462 * 6. lvb bit set for step 5 (will already be set)
 463 * 7. lvb bit cleared for step 3
 464 *
 465 * This is not a problem because the failure in step 5 does not
 466 * require recovery, because the mount in step 4 could not have
 467 * progressed far enough to unblock locks and access the fs.  The
 468 * control_mount() function waits for all recoveries to be complete
 469 * for the latest lockspace generation before ever unblocking locks
 470 * and returning.  The mount in step 4 waits until the recovery in
 471 * step 1 is done.
 472 *
 473 * - special case of first mounter: first node to mount the fs
 474 *
 475 * The first node to mount a gfs2 fs needs to check all the journals
 476 * and recover any that need recovery before other nodes are allowed
 477 * to mount the fs.  (Others may begin mounting, but they must wait
 478 * for the first mounter to be done before taking locks on the fs
 479 * or accessing the fs.)  This has two parts:
 480 *
 481 * 1. The mounted_lock tells a node it's the first to mount the fs.
 482 * Each node holds the mounted_lock in PR while it's mounted.
 483 * Each node tries to acquire the mounted_lock in EX when it mounts.
 484 * If a node is granted the mounted_lock EX it means there are no
 485 * other mounted nodes (no PR locks exist), and it is the first mounter.
 486 * The mounted_lock is demoted to PR when first recovery is done, so
 487 * others will fail to get an EX lock, but will get a PR lock.
 488 *
 489 * 2. The control_lock blocks others in control_mount() while the first
 490 * mounter is doing first mount recovery of all journals.
 491 * A mounting node needs to acquire control_lock in EX mode before
 492 * it can proceed.  The first mounter holds control_lock in EX while doing
 493 * the first mount recovery, blocking mounts from other nodes, then demotes
 494 * control_lock to NL when it's done (others_may_mount/first_done),
 495 * allowing other nodes to continue mounting.
 496 *
 497 * first mounter:
 498 * control_lock EX/NOQUEUE success
 499 * mounted_lock EX/NOQUEUE success (no other PR, so no other mounters)
 500 * set first=1
 501 * do first mounter recovery
 502 * mounted_lock EX->PR
 503 * control_lock EX->NL, write lvb generation
 504 *
 505 * other mounter:
 506 * control_lock EX/NOQUEUE success (if fail -EAGAIN, retry)
 507 * mounted_lock EX/NOQUEUE fail -EAGAIN (expected due to other mounters PR)
 508 * mounted_lock PR/NOQUEUE success
 509 * read lvb generation
 510 * control_lock EX->NL
 511 * set first=0
 512 *
 513 * - mount during recovery
 514 *
 515 * If a node mounts while others are doing recovery (not first mounter),
 516 * the mounting node will get its initial recover_done() callback without
 517 * having seen any previous failures/callbacks.
 518 *
 519 * It must wait for all recoveries preceding its mount to be finished
 520 * before it unblocks locks.  It does this by repeating the "other mounter"
 521 * steps above until the lvb generation number is >= its mount generation
 522 * number (from initial recover_done) and all lvb bits are clear.
 523 *
 524 * - control_lock lvb format
 525 *
 526 * 4 bytes generation number: the latest dlm lockspace generation number
 527 * from recover_done callback.  Indicates the jid bitmap has been updated
 528 * to reflect all slot failures through that generation.
 529 * 4 bytes unused.
 530 * GDLM_LVB_SIZE-8 bytes of jid bit map. If bit N is set, it indicates
 531 * that jid N needs recovery.
 532 */
 533
 534#define JID_BITMAP_OFFSET 8 /* 4 byte generation number + 4 byte unused */
 535
 536static void control_lvb_read(struct lm_lockstruct *ls, uint32_t *lvb_gen,
 537			     char *lvb_bits)
 538{
 539	__le32 gen;
 540	memcpy(lvb_bits, ls->ls_control_lvb, GDLM_LVB_SIZE);
 541	memcpy(&gen, lvb_bits, sizeof(__le32));
 542	*lvb_gen = le32_to_cpu(gen);
 543}
 544
 545static void control_lvb_write(struct lm_lockstruct *ls, uint32_t lvb_gen,
 546			      char *lvb_bits)
 547{
 548	__le32 gen;
 549	memcpy(ls->ls_control_lvb, lvb_bits, GDLM_LVB_SIZE);
 550	gen = cpu_to_le32(lvb_gen);
 551	memcpy(ls->ls_control_lvb, &gen, sizeof(__le32));
 552}
 553
 554static int all_jid_bits_clear(char *lvb)
 555{
 556	return !memchr_inv(lvb + JID_BITMAP_OFFSET, 0,
 557			GDLM_LVB_SIZE - JID_BITMAP_OFFSET);
 558}
 559
 560static void sync_wait_cb(void *arg)
 561{
 562	struct lm_lockstruct *ls = arg;
 563	complete(&ls->ls_sync_wait);
 564}
 565
 566static int sync_unlock(struct gfs2_sbd *sdp, struct dlm_lksb *lksb, char *name)
 567{
 568	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 569	int error;
 570
 571	down_read(&ls->ls_sem);
 572	error = -ENODEV;
 573	if (likely(ls->ls_dlm != NULL))
 574		error = dlm_unlock(ls->ls_dlm, lksb->sb_lkid, 0, lksb, ls);
 575	up_read(&ls->ls_sem);
 576	if (error) {
 577		fs_err(sdp, "%s lkid %x error %d\n",
 578		       name, lksb->sb_lkid, error);
 579		return error;
 580	}
 581
 582	wait_for_completion(&ls->ls_sync_wait);
 583
 584	if (lksb->sb_status != -DLM_EUNLOCK) {
 585		fs_err(sdp, "%s lkid %x status %d\n",
 586		       name, lksb->sb_lkid, lksb->sb_status);
 587		return -1;
 588	}
 589	return 0;
 590}
 591
 592static int sync_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags,
 593		     unsigned int num, struct dlm_lksb *lksb, char *name)
 594{
 595	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 596	char strname[GDLM_STRNAME_BYTES];
 597	int error, status;
 598
 599	memset(strname, 0, GDLM_STRNAME_BYTES);
 600	snprintf(strname, GDLM_STRNAME_BYTES, "%8x%16x", LM_TYPE_NONDISK, num);
 601
 602	down_read(&ls->ls_sem);
 603	error = -ENODEV;
 604	if (likely(ls->ls_dlm != NULL)) {
 605		error = dlm_lock(ls->ls_dlm, mode, lksb, flags,
 606				 strname, GDLM_STRNAME_BYTES - 1,
 607				 0, sync_wait_cb, ls, NULL);
 608	}
 609	up_read(&ls->ls_sem);
 610	if (error) {
 611		fs_err(sdp, "%s lkid %x flags %x mode %d error %d\n",
 612		       name, lksb->sb_lkid, flags, mode, error);
 613		return error;
 614	}
 615
 616	wait_for_completion(&ls->ls_sync_wait);
 617
 618	status = lksb->sb_status;
 619
 620	if (status && status != -EAGAIN) {
 621		fs_err(sdp, "%s lkid %x flags %x mode %d status %d\n",
 622		       name, lksb->sb_lkid, flags, mode, status);
 623	}
 624
 625	return status;
 626}
 627
 628static int mounted_unlock(struct gfs2_sbd *sdp)
 629{
 630	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 631	return sync_unlock(sdp, &ls->ls_mounted_lksb, "mounted_lock");
 632}
 633
 634static int mounted_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
 635{
 636	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 637	return sync_lock(sdp, mode, flags, GFS2_MOUNTED_LOCK,
 638			 &ls->ls_mounted_lksb, "mounted_lock");
 639}
 640
 641static int control_unlock(struct gfs2_sbd *sdp)
 642{
 643	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 644	return sync_unlock(sdp, &ls->ls_control_lksb, "control_lock");
 645}
 646
 647static int control_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
 648{
 649	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 650	return sync_lock(sdp, mode, flags, GFS2_CONTROL_LOCK,
 651			 &ls->ls_control_lksb, "control_lock");
 652}
 653
 654static void gfs2_control_func(struct work_struct *work)
 655{
 656	struct gfs2_sbd *sdp = container_of(work, struct gfs2_sbd, sd_control_work.work);
 657	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 658	uint32_t block_gen, start_gen, lvb_gen, flags;
 659	int recover_set = 0;
 660	int write_lvb = 0;
 661	int recover_size;
 662	int i, error;
 663
 664	spin_lock(&ls->ls_recover_spin);
 665	/*
 666	 * No MOUNT_DONE means we're still mounting; control_mount()
 667	 * will set this flag, after which this thread will take over
 668	 * all further clearing of BLOCK_LOCKS.
 669	 *
 670	 * FIRST_MOUNT means this node is doing first mounter recovery,
 671	 * for which recovery control is handled by
 672	 * control_mount()/control_first_done(), not this thread.
 673	 */
 674	if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
 675	     test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
 676		spin_unlock(&ls->ls_recover_spin);
 677		return;
 678	}
 679	block_gen = ls->ls_recover_block;
 680	start_gen = ls->ls_recover_start;
 681	spin_unlock(&ls->ls_recover_spin);
 682
 683	/*
 684	 * Equal block_gen and start_gen implies we are between
 685	 * recover_prep and recover_done callbacks, which means
 686	 * dlm recovery is in progress and dlm locking is blocked.
 687	 * There's no point trying to do any work until recover_done.
 688	 */
 689
 690	if (block_gen == start_gen)
 691		return;
 692
 693	/*
 694	 * Propagate recover_submit[] and recover_result[] to lvb:
 695	 * dlm_recoverd adds to recover_submit[] jids needing recovery
 696	 * gfs2_recover adds to recover_result[] journal recovery results
 697	 *
 698	 * set lvb bit for jids in recover_submit[] if the lvb has not
 699	 * yet been updated for the generation of the failure
 700	 *
 701	 * clear lvb bit for jids in recover_result[] if the result of
 702	 * the journal recovery is SUCCESS
 703	 */
 704
 705	error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
 706	if (error) {
 707		fs_err(sdp, "control lock EX error %d\n", error);
 708		return;
 709	}
 710
 711	control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
 712
 713	spin_lock(&ls->ls_recover_spin);
 714	if (block_gen != ls->ls_recover_block ||
 715	    start_gen != ls->ls_recover_start) {
 716		fs_info(sdp, "recover generation %u block1 %u %u\n",
 717			start_gen, block_gen, ls->ls_recover_block);
 718		spin_unlock(&ls->ls_recover_spin);
 719		control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
 720		return;
 721	}
 722
 723	recover_size = ls->ls_recover_size;
 724
 725	if (lvb_gen <= start_gen) {
 726		/*
 727		 * Clear lvb bits for jids we've successfully recovered.
 728		 * Because all nodes attempt to recover failed journals,
 729		 * a journal can be recovered multiple times successfully
 730		 * in succession.  Only the first will really do recovery,
 731		 * the others find it clean, but still report a successful
 732		 * recovery.  So, another node may have already recovered
 733		 * the jid and cleared the lvb bit for it.
 734		 */
 735		for (i = 0; i < recover_size; i++) {
 736			if (ls->ls_recover_result[i] != LM_RD_SUCCESS)
 737				continue;
 738
 739			ls->ls_recover_result[i] = 0;
 740
 741			if (!test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET))
 742				continue;
 743
 744			__clear_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
 745			write_lvb = 1;
 746		}
 747	}
 748
 749	if (lvb_gen == start_gen) {
 750		/*
 751		 * Failed slots before start_gen are already set in lvb.
 752		 */
 753		for (i = 0; i < recover_size; i++) {
 754			if (!ls->ls_recover_submit[i])
 755				continue;
 756			if (ls->ls_recover_submit[i] < lvb_gen)
 757				ls->ls_recover_submit[i] = 0;
 758		}
 759	} else if (lvb_gen < start_gen) {
 760		/*
 761		 * Failed slots before start_gen are not yet set in lvb.
 762		 */
 763		for (i = 0; i < recover_size; i++) {
 764			if (!ls->ls_recover_submit[i])
 765				continue;
 766			if (ls->ls_recover_submit[i] < start_gen) {
 767				ls->ls_recover_submit[i] = 0;
 768				__set_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
 769			}
 770		}
 771		/* even if there are no bits to set, we need to write the
 772		   latest generation to the lvb */
 773		write_lvb = 1;
 774	} else {
 775		/*
 776		 * we should be getting a recover_done() for lvb_gen soon
 777		 */
 778	}
 779	spin_unlock(&ls->ls_recover_spin);
 780
 781	if (write_lvb) {
 782		control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
 783		flags = DLM_LKF_CONVERT | DLM_LKF_VALBLK;
 784	} else {
 785		flags = DLM_LKF_CONVERT;
 786	}
 787
 788	error = control_lock(sdp, DLM_LOCK_NL, flags);
 789	if (error) {
 790		fs_err(sdp, "control lock NL error %d\n", error);
 791		return;
 792	}
 793
 794	/*
 795	 * Everyone will see jid bits set in the lvb, run gfs2_recover_set(),
 796	 * and clear a jid bit in the lvb if the recovery is a success.
 797	 * Eventually all journals will be recovered, all jid bits will
 798	 * be cleared in the lvb, and everyone will clear BLOCK_LOCKS.
 799	 */
 800
 801	for (i = 0; i < recover_size; i++) {
 802		if (test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET)) {
 803			fs_info(sdp, "recover generation %u jid %d\n",
 804				start_gen, i);
 805			gfs2_recover_set(sdp, i);
 806			recover_set++;
 807		}
 808	}
 809	if (recover_set)
 810		return;
 811
 812	/*
 813	 * No more jid bits set in lvb, all recovery is done, unblock locks
 814	 * (unless a new recover_prep callback has occured blocking locks
 815	 * again while working above)
 816	 */
 817
 818	spin_lock(&ls->ls_recover_spin);
 819	if (ls->ls_recover_block == block_gen &&
 820	    ls->ls_recover_start == start_gen) {
 821		clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
 822		spin_unlock(&ls->ls_recover_spin);
 823		fs_info(sdp, "recover generation %u done\n", start_gen);
 824		gfs2_glock_thaw(sdp);
 825	} else {
 826		fs_info(sdp, "recover generation %u block2 %u %u\n",
 827			start_gen, block_gen, ls->ls_recover_block);
 828		spin_unlock(&ls->ls_recover_spin);
 829	}
 830}
 831
 832static int control_mount(struct gfs2_sbd *sdp)
 833{
 834	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
 835	uint32_t start_gen, block_gen, mount_gen, lvb_gen;
 836	int mounted_mode;
 837	int retries = 0;
 838	int error;
 839
 840	memset(&ls->ls_mounted_lksb, 0, sizeof(struct dlm_lksb));
 841	memset(&ls->ls_control_lksb, 0, sizeof(struct dlm_lksb));
 842	memset(&ls->ls_control_lvb, 0, GDLM_LVB_SIZE);
 843	ls->ls_control_lksb.sb_lvbptr = ls->ls_control_lvb;
 844	init_completion(&ls->ls_sync_wait);
 845
 846	set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
 847
 848	error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_VALBLK);
 849	if (error) {
 850		fs_err(sdp, "control_mount control_lock NL error %d\n", error);
 851		return error;
 852	}
 853
 854	error = mounted_lock(sdp, DLM_LOCK_NL, 0);
 855	if (error) {
 856		fs_err(sdp, "control_mount mounted_lock NL error %d\n", error);
 857		control_unlock(sdp);
 858		return error;
 859	}
 860	mounted_mode = DLM_LOCK_NL;
 861
 862restart:
 863	if (retries++ && signal_pending(current)) {
 864		error = -EINTR;
 865		goto fail;
 866	}
 867
 868	/*
 869	 * We always start with both locks in NL. control_lock is
 870	 * demoted to NL below so we don't need to do it here.
 871	 */
 872
 873	if (mounted_mode != DLM_LOCK_NL) {
 874		error = mounted_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
 875		if (error)
 876			goto fail;
 877		mounted_mode = DLM_LOCK_NL;
 878	}
 879
 880	/*
 881	 * Other nodes need to do some work in dlm recovery and gfs2_control
 882	 * before the recover_done and control_lock will be ready for us below.
 883	 * A delay here is not required but often avoids having to retry.
 884	 */
 885
 886	msleep_interruptible(500);
 887
 888	/*
 889	 * Acquire control_lock in EX and mounted_lock in either EX or PR.
 890	 * control_lock lvb keeps track of any pending journal recoveries.
 891	 * mounted_lock indicates if any other nodes have the fs mounted.
 892	 */
 893
 894	error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE|DLM_LKF_VALBLK);
 895	if (error == -EAGAIN) {
 896		goto restart;
 897	} else if (error) {
 898		fs_err(sdp, "control_mount control_lock EX error %d\n", error);
 899		goto fail;
 900	}
 901
 902	/**
 903	 * If we're a spectator, we don't want to take the lock in EX because
 904	 * we cannot do the first-mount responsibility it implies: recovery.
 905	 */
 906	if (sdp->sd_args.ar_spectator)
 907		goto locks_done;
 908
 909	error = mounted_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
 910	if (!error) {
 911		mounted_mode = DLM_LOCK_EX;
 912		goto locks_done;
 913	} else if (error != -EAGAIN) {
 914		fs_err(sdp, "control_mount mounted_lock EX error %d\n", error);
 915		goto fail;
 916	}
 917
 918	error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
 919	if (!error) {
 920		mounted_mode = DLM_LOCK_PR;
 921		goto locks_done;
 922	} else {
 923		/* not even -EAGAIN should happen here */
 924		fs_err(sdp, "control_mount mounted_lock PR error %d\n", error);
 925		goto fail;
 926	}
 927
 928locks_done:
 929	/*
 930	 * If we got both locks above in EX, then we're the first mounter.
 931	 * If not, then we need to wait for the control_lock lvb to be
 932	 * updated by other mounted nodes to reflect our mount generation.
 933	 *
 934	 * In simple first mounter cases, first mounter will see zero lvb_gen,
 935	 * but in cases where all existing nodes leave/fail before mounting
 936	 * nodes finish control_mount, then all nodes will be mounting and
 937	 * lvb_gen will be non-zero.
 938	 */
 939
 940	control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
 941
 942	if (lvb_gen == 0xFFFFFFFF) {
 943		/* special value to force mount attempts to fail */
 944		fs_err(sdp, "control_mount control_lock disabled\n");
 945		error = -EINVAL;
 946		goto fail;
 947	}
 948
 949	if (mounted_mode == DLM_LOCK_EX) {
 950		/* first mounter, keep both EX while doing first recovery */
 951		spin_lock(&ls->ls_recover_spin);
 952		clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
 953		set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
 954		set_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
 955		spin_unlock(&ls->ls_recover_spin);
 956		fs_info(sdp, "first mounter control generation %u\n", lvb_gen);
 957		return 0;
 958	}
 959
 960	error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
 961	if (error)
 962		goto fail;
 963
 964	/*
 965	 * We are not first mounter, now we need to wait for the control_lock
 966	 * lvb generation to be >= the generation from our first recover_done
 967	 * and all lvb bits to be clear (no pending journal recoveries.)
 968	 */
 969
 970	if (!all_jid_bits_clear(ls->ls_lvb_bits)) {
 971		/* journals need recovery, wait until all are clear */
 972		fs_info(sdp, "control_mount wait for journal recovery\n");
 973		goto restart;
 974	}
 975
 976	spin_lock(&ls->ls_recover_spin);
 977	block_gen = ls->ls_recover_block;
 978	start_gen = ls->ls_recover_start;
 979	mount_gen = ls->ls_recover_mount;
 980
 981	if (lvb_gen < mount_gen) {
 982		/* wait for mounted nodes to update control_lock lvb to our
 983		   generation, which might include new recovery bits set */
 984		if (sdp->sd_args.ar_spectator) {
 985			fs_info(sdp, "Recovery is required. Waiting for a "
 986				"non-spectator to mount.\n");
 987			spin_unlock(&ls->ls_recover_spin);
 988			msleep_interruptible(1000);
 989		} else {
 990			fs_info(sdp, "control_mount wait1 block %u start %u "
 991				"mount %u lvb %u flags %lx\n", block_gen,
 992				start_gen, mount_gen, lvb_gen,
 993				ls->ls_recover_flags);
 994			spin_unlock(&ls->ls_recover_spin);
 995		}
 996		goto restart;
 997	}
 998
 999	if (lvb_gen != start_gen) {
1000		/* wait for mounted nodes to update control_lock lvb to the
1001		   latest recovery generation */
1002		fs_info(sdp, "control_mount wait2 block %u start %u mount %u "
1003			"lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
1004			lvb_gen, ls->ls_recover_flags);
1005		spin_unlock(&ls->ls_recover_spin);
1006		goto restart;
1007	}
1008
1009	if (block_gen == start_gen) {
1010		/* dlm recovery in progress, wait for it to finish */
1011		fs_info(sdp, "control_mount wait3 block %u start %u mount %u "
1012			"lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
1013			lvb_gen, ls->ls_recover_flags);
1014		spin_unlock(&ls->ls_recover_spin);
1015		goto restart;
1016	}
1017
1018	clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
1019	set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
1020	memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
1021	memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
1022	spin_unlock(&ls->ls_recover_spin);
1023	return 0;
1024
1025fail:
1026	mounted_unlock(sdp);
1027	control_unlock(sdp);
1028	return error;
1029}
1030
1031static int control_first_done(struct gfs2_sbd *sdp)
1032{
1033	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1034	uint32_t start_gen, block_gen;
1035	int error;
1036
1037restart:
1038	spin_lock(&ls->ls_recover_spin);
1039	start_gen = ls->ls_recover_start;
1040	block_gen = ls->ls_recover_block;
1041
1042	if (test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags) ||
1043	    !test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
1044	    !test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1045		/* sanity check, should not happen */
1046		fs_err(sdp, "control_first_done start %u block %u flags %lx\n",
1047		       start_gen, block_gen, ls->ls_recover_flags);
1048		spin_unlock(&ls->ls_recover_spin);
1049		control_unlock(sdp);
1050		return -1;
1051	}
1052
1053	if (start_gen == block_gen) {
1054		/*
1055		 * Wait for the end of a dlm recovery cycle to switch from
1056		 * first mounter recovery.  We can ignore any recover_slot
1057		 * callbacks between the recover_prep and next recover_done
1058		 * because we are still the first mounter and any failed nodes
1059		 * have not fully mounted, so they don't need recovery.
1060		 */
1061		spin_unlock(&ls->ls_recover_spin);
1062		fs_info(sdp, "control_first_done wait gen %u\n", start_gen);
1063
1064		wait_on_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY,
1065			    TASK_UNINTERRUPTIBLE);
1066		goto restart;
1067	}
1068
1069	clear_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
1070	set_bit(DFL_FIRST_MOUNT_DONE, &ls->ls_recover_flags);
1071	memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
1072	memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
1073	spin_unlock(&ls->ls_recover_spin);
1074
1075	memset(ls->ls_lvb_bits, 0, GDLM_LVB_SIZE);
1076	control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
1077
1078	error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT);
1079	if (error)
1080		fs_err(sdp, "control_first_done mounted PR error %d\n", error);
1081
1082	error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
1083	if (error)
1084		fs_err(sdp, "control_first_done control NL error %d\n", error);
1085
1086	return error;
1087}
1088
1089/*
1090 * Expand static jid arrays if necessary (by increments of RECOVER_SIZE_INC)
1091 * to accommodate the largest slot number.  (NB dlm slot numbers start at 1,
1092 * gfs2 jids start at 0, so jid = slot - 1)
1093 */
1094
1095#define RECOVER_SIZE_INC 16
1096
1097static int set_recover_size(struct gfs2_sbd *sdp, struct dlm_slot *slots,
1098			    int num_slots)
1099{
1100	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1101	uint32_t *submit = NULL;
1102	uint32_t *result = NULL;
1103	uint32_t old_size, new_size;
1104	int i, max_jid;
1105
1106	if (!ls->ls_lvb_bits) {
1107		ls->ls_lvb_bits = kzalloc(GDLM_LVB_SIZE, GFP_NOFS);
1108		if (!ls->ls_lvb_bits)
1109			return -ENOMEM;
1110	}
1111
1112	max_jid = 0;
1113	for (i = 0; i < num_slots; i++) {
1114		if (max_jid < slots[i].slot - 1)
1115			max_jid = slots[i].slot - 1;
1116	}
1117
1118	old_size = ls->ls_recover_size;
1119	new_size = old_size;
1120	while (new_size < max_jid + 1)
1121		new_size += RECOVER_SIZE_INC;
1122	if (new_size == old_size)
1123		return 0;
1124
1125	submit = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1126	result = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1127	if (!submit || !result) {
1128		kfree(submit);
1129		kfree(result);
1130		return -ENOMEM;
1131	}
1132
1133	spin_lock(&ls->ls_recover_spin);
1134	memcpy(submit, ls->ls_recover_submit, old_size * sizeof(uint32_t));
1135	memcpy(result, ls->ls_recover_result, old_size * sizeof(uint32_t));
1136	kfree(ls->ls_recover_submit);
1137	kfree(ls->ls_recover_result);
1138	ls->ls_recover_submit = submit;
1139	ls->ls_recover_result = result;
1140	ls->ls_recover_size = new_size;
1141	spin_unlock(&ls->ls_recover_spin);
1142	return 0;
1143}
1144
1145static void free_recover_size(struct lm_lockstruct *ls)
1146{
1147	kfree(ls->ls_lvb_bits);
1148	kfree(ls->ls_recover_submit);
1149	kfree(ls->ls_recover_result);
1150	ls->ls_recover_submit = NULL;
1151	ls->ls_recover_result = NULL;
1152	ls->ls_recover_size = 0;
1153	ls->ls_lvb_bits = NULL;
1154}
1155
1156/* dlm calls before it does lock recovery */
1157
1158static void gdlm_recover_prep(void *arg)
1159{
1160	struct gfs2_sbd *sdp = arg;
1161	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1162
1163	if (gfs2_withdrawn(sdp)) {
1164		fs_err(sdp, "recover_prep ignored due to withdraw.\n");
1165		return;
1166	}
1167	spin_lock(&ls->ls_recover_spin);
1168	ls->ls_recover_block = ls->ls_recover_start;
1169	set_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1170
1171	if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
1172	     test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1173		spin_unlock(&ls->ls_recover_spin);
1174		return;
1175	}
1176	set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
1177	spin_unlock(&ls->ls_recover_spin);
1178}
1179
1180/* dlm calls after recover_prep has been completed on all lockspace members;
1181   identifies slot/jid of failed member */
1182
1183static void gdlm_recover_slot(void *arg, struct dlm_slot *slot)
1184{
1185	struct gfs2_sbd *sdp = arg;
1186	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1187	int jid = slot->slot - 1;
1188
1189	if (gfs2_withdrawn(sdp)) {
1190		fs_err(sdp, "recover_slot jid %d ignored due to withdraw.\n",
1191		       jid);
1192		return;
1193	}
1194	spin_lock(&ls->ls_recover_spin);
1195	if (ls->ls_recover_size < jid + 1) {
1196		fs_err(sdp, "recover_slot jid %d gen %u short size %d\n",
1197		       jid, ls->ls_recover_block, ls->ls_recover_size);
1198		spin_unlock(&ls->ls_recover_spin);
1199		return;
1200	}
1201
1202	if (ls->ls_recover_submit[jid]) {
1203		fs_info(sdp, "recover_slot jid %d gen %u prev %u\n",
1204			jid, ls->ls_recover_block, ls->ls_recover_submit[jid]);
1205	}
1206	ls->ls_recover_submit[jid] = ls->ls_recover_block;
1207	spin_unlock(&ls->ls_recover_spin);
1208}
1209
1210/* dlm calls after recover_slot and after it completes lock recovery */
1211
1212static void gdlm_recover_done(void *arg, struct dlm_slot *slots, int num_slots,
1213			      int our_slot, uint32_t generation)
1214{
1215	struct gfs2_sbd *sdp = arg;
1216	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1217
1218	if (gfs2_withdrawn(sdp)) {
1219		fs_err(sdp, "recover_done ignored due to withdraw.\n");
1220		return;
1221	}
1222	/* ensure the ls jid arrays are large enough */
1223	set_recover_size(sdp, slots, num_slots);
1224
1225	spin_lock(&ls->ls_recover_spin);
1226	ls->ls_recover_start = generation;
1227
1228	if (!ls->ls_recover_mount) {
1229		ls->ls_recover_mount = generation;
1230		ls->ls_jid = our_slot - 1;
1231	}
1232
1233	if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1234		queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work, 0);
1235
1236	clear_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1237	smp_mb__after_atomic();
1238	wake_up_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY);
1239	spin_unlock(&ls->ls_recover_spin);
1240}
1241
1242/* gfs2_recover thread has a journal recovery result */
1243
1244static void gdlm_recovery_result(struct gfs2_sbd *sdp, unsigned int jid,
1245				 unsigned int result)
1246{
1247	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1248
1249	if (gfs2_withdrawn(sdp)) {
1250		fs_err(sdp, "recovery_result jid %d ignored due to withdraw.\n",
1251		       jid);
1252		return;
1253	}
1254	if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1255		return;
1256
1257	/* don't care about the recovery of own journal during mount */
1258	if (jid == ls->ls_jid)
1259		return;
1260
1261	spin_lock(&ls->ls_recover_spin);
1262	if (test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1263		spin_unlock(&ls->ls_recover_spin);
1264		return;
1265	}
1266	if (ls->ls_recover_size < jid + 1) {
1267		fs_err(sdp, "recovery_result jid %d short size %d\n",
1268		       jid, ls->ls_recover_size);
1269		spin_unlock(&ls->ls_recover_spin);
1270		return;
1271	}
1272
1273	fs_info(sdp, "recover jid %d result %s\n", jid,
1274		result == LM_RD_GAVEUP ? "busy" : "success");
1275
1276	ls->ls_recover_result[jid] = result;
1277
1278	/* GAVEUP means another node is recovering the journal; delay our
1279	   next attempt to recover it, to give the other node a chance to
1280	   finish before trying again */
1281
1282	if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1283		queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work,
1284				   result == LM_RD_GAVEUP ? HZ : 0);
1285	spin_unlock(&ls->ls_recover_spin);
1286}
1287
1288static const struct dlm_lockspace_ops gdlm_lockspace_ops = {
1289	.recover_prep = gdlm_recover_prep,
1290	.recover_slot = gdlm_recover_slot,
1291	.recover_done = gdlm_recover_done,
1292};
1293
1294static int gdlm_mount(struct gfs2_sbd *sdp, const char *table)
1295{
1296	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1297	char cluster[GFS2_LOCKNAME_LEN];
1298	const char *fsname;
1299	uint32_t flags;
1300	int error, ops_result;
1301
1302	/*
1303	 * initialize everything
1304	 */
1305
1306	INIT_DELAYED_WORK(&sdp->sd_control_work, gfs2_control_func);
1307	ls->ls_dlm = NULL;
1308	spin_lock_init(&ls->ls_recover_spin);
1309	ls->ls_recover_flags = 0;
1310	ls->ls_recover_mount = 0;
1311	ls->ls_recover_start = 0;
1312	ls->ls_recover_block = 0;
1313	ls->ls_recover_size = 0;
1314	ls->ls_recover_submit = NULL;
1315	ls->ls_recover_result = NULL;
1316	ls->ls_lvb_bits = NULL;
1317
1318	error = set_recover_size(sdp, NULL, 0);
1319	if (error)
1320		goto fail;
1321
1322	/*
1323	 * prepare dlm_new_lockspace args
1324	 */
1325
1326	fsname = strchr(table, ':');
1327	if (!fsname) {
1328		fs_info(sdp, "no fsname found\n");
1329		error = -EINVAL;
1330		goto fail_free;
1331	}
1332	memset(cluster, 0, sizeof(cluster));
1333	memcpy(cluster, table, strlen(table) - strlen(fsname));
1334	fsname++;
1335
1336	flags = DLM_LSFL_NEWEXCL;
1337
1338	/*
1339	 * create/join lockspace
1340	 */
1341
1342	init_rwsem(&ls->ls_sem);
1343	error = dlm_new_lockspace(fsname, cluster, flags, GDLM_LVB_SIZE,
1344				  &gdlm_lockspace_ops, sdp, &ops_result,
1345				  &ls->ls_dlm);
1346	if (error) {
1347		fs_err(sdp, "dlm_new_lockspace error %d\n", error);
1348		goto fail_free;
1349	}
1350
1351	if (ops_result < 0) {
1352		/*
1353		 * dlm does not support ops callbacks,
1354		 * old dlm_controld/gfs_controld are used, try without ops.
1355		 */
1356		fs_info(sdp, "dlm lockspace ops not used\n");
1357		free_recover_size(ls);
1358		set_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags);
1359		return 0;
1360	}
1361
1362	if (!test_bit(SDF_NOJOURNALID, &sdp->sd_flags)) {
1363		fs_err(sdp, "dlm lockspace ops disallow jid preset\n");
1364		error = -EINVAL;
1365		goto fail_release;
1366	}
1367
1368	/*
1369	 * control_mount() uses control_lock to determine first mounter,
1370	 * and for later mounts, waits for any recoveries to be cleared.
1371	 */
1372
1373	error = control_mount(sdp);
1374	if (error) {
1375		fs_err(sdp, "mount control error %d\n", error);
1376		goto fail_release;
1377	}
1378
1379	ls->ls_first = !!test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
1380	clear_bit(SDF_NOJOURNALID, &sdp->sd_flags);
1381	smp_mb__after_atomic();
1382	wake_up_bit(&sdp->sd_flags, SDF_NOJOURNALID);
1383	return 0;
1384
1385fail_release:
1386	dlm_release_lockspace(ls->ls_dlm, DLM_RELEASE_NORMAL);
1387fail_free:
1388	free_recover_size(ls);
1389fail:
1390	return error;
1391}
1392
1393static void gdlm_first_done(struct gfs2_sbd *sdp)
1394{
1395	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1396	int error;
1397
1398	if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1399		return;
1400
1401	error = control_first_done(sdp);
1402	if (error)
1403		fs_err(sdp, "mount first_done error %d\n", error);
1404}
1405
1406/*
1407 * gdlm_unmount - release our lockspace
1408 * @sdp: the superblock
1409 * @clean: Indicates whether or not the remaining nodes in the cluster should
1410 *	   perform recovery.  Recovery is necessary when a node withdraws and
1411 *	   its journal remains dirty.  Recovery isn't necessary when a node
1412 *	   cleanly unmounts a filesystem.
1413 */
1414static void gdlm_unmount(struct gfs2_sbd *sdp, bool clean)
1415{
1416	struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1417
1418	if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1419		goto release;
1420
1421	/* wait for gfs2_control_wq to be done with this mount */
1422
1423	spin_lock(&ls->ls_recover_spin);
1424	set_bit(DFL_UNMOUNT, &ls->ls_recover_flags);
1425	spin_unlock(&ls->ls_recover_spin);
1426	flush_delayed_work(&sdp->sd_control_work);
1427
1428	/* mounted_lock and control_lock will be purged in dlm recovery */
1429release:
1430	down_write(&ls->ls_sem);
1431	if (ls->ls_dlm) {
1432		dlm_release_lockspace(ls->ls_dlm,
1433				      clean ? DLM_RELEASE_NORMAL :
1434					      DLM_RELEASE_RECOVER);
1435		ls->ls_dlm = NULL;
1436	}
1437	up_write(&ls->ls_sem);
1438
1439	free_recover_size(ls);
1440}
1441
1442static const match_table_t dlm_tokens = {
1443	{ Opt_jid, "jid=%d"},
1444	{ Opt_id, "id=%d"},
1445	{ Opt_first, "first=%d"},
1446	{ Opt_nodir, "nodir=%d"},
1447	{ Opt_err, NULL },
1448};
1449
1450const struct lm_lockops gfs2_dlm_ops = {
1451	.lm_proto_name = "lock_dlm",
1452	.lm_mount = gdlm_mount,
1453	.lm_first_done = gdlm_first_done,
1454	.lm_recovery_result = gdlm_recovery_result,
1455	.lm_unmount = gdlm_unmount,
1456	.lm_put_lock = gdlm_put_lock,
1457	.lm_lock = gdlm_lock,
1458	.lm_cancel = gdlm_cancel,
1459	.lm_tokens = &dlm_tokens,
1460};
1461