Linux Audio

Check our new training course

Buildroot integration, development and maintenance

Need a Buildroot system for your embedded project?
Loading...
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-only
   2/******************************************************************************
   3*******************************************************************************
   4**
   5**  Copyright (C) 2005-2010 Red Hat, Inc.  All rights reserved.
   6**
   7**
   8*******************************************************************************
   9******************************************************************************/
  10
  11/* Central locking logic has four stages:
  12
  13   dlm_lock()
  14   dlm_unlock()
  15
  16   request_lock(ls, lkb)
  17   convert_lock(ls, lkb)
  18   unlock_lock(ls, lkb)
  19   cancel_lock(ls, lkb)
  20
  21   _request_lock(r, lkb)
  22   _convert_lock(r, lkb)
  23   _unlock_lock(r, lkb)
  24   _cancel_lock(r, lkb)
  25
  26   do_request(r, lkb)
  27   do_convert(r, lkb)
  28   do_unlock(r, lkb)
  29   do_cancel(r, lkb)
  30
  31   Stage 1 (lock, unlock) is mainly about checking input args and
  32   splitting into one of the four main operations:
  33
  34       dlm_lock          = request_lock
  35       dlm_lock+CONVERT  = convert_lock
  36       dlm_unlock        = unlock_lock
  37       dlm_unlock+CANCEL = cancel_lock
  38
  39   Stage 2, xxxx_lock(), just finds and locks the relevant rsb which is
  40   provided to the next stage.
  41
  42   Stage 3, _xxxx_lock(), determines if the operation is local or remote.
  43   When remote, it calls send_xxxx(), when local it calls do_xxxx().
  44
  45   Stage 4, do_xxxx(), is the guts of the operation.  It manipulates the
  46   given rsb and lkb and queues callbacks.
  47
  48   For remote operations, send_xxxx() results in the corresponding do_xxxx()
  49   function being executed on the remote node.  The connecting send/receive
  50   calls on local (L) and remote (R) nodes:
  51
  52   L: send_xxxx()              ->  R: receive_xxxx()
  53                                   R: do_xxxx()
  54   L: receive_xxxx_reply()     <-  R: send_xxxx_reply()
  55*/
 
 
  56#include <linux/types.h>
  57#include <linux/rbtree.h>
  58#include <linux/slab.h>
  59#include "dlm_internal.h"
  60#include <linux/dlm_device.h>
  61#include "memory.h"
  62#include "midcomms.h"
  63#include "requestqueue.h"
  64#include "util.h"
  65#include "dir.h"
  66#include "member.h"
  67#include "lockspace.h"
  68#include "ast.h"
  69#include "lock.h"
  70#include "rcom.h"
  71#include "recover.h"
  72#include "lvb_table.h"
  73#include "user.h"
  74#include "config.h"
  75
  76static int send_request(struct dlm_rsb *r, struct dlm_lkb *lkb);
  77static int send_convert(struct dlm_rsb *r, struct dlm_lkb *lkb);
  78static int send_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb);
  79static int send_cancel(struct dlm_rsb *r, struct dlm_lkb *lkb);
  80static int send_grant(struct dlm_rsb *r, struct dlm_lkb *lkb);
  81static int send_bast(struct dlm_rsb *r, struct dlm_lkb *lkb, int mode);
  82static int send_lookup(struct dlm_rsb *r, struct dlm_lkb *lkb);
  83static int send_remove(struct dlm_rsb *r);
  84static int _request_lock(struct dlm_rsb *r, struct dlm_lkb *lkb);
  85static int _cancel_lock(struct dlm_rsb *r, struct dlm_lkb *lkb);
  86static void __receive_convert_reply(struct dlm_rsb *r, struct dlm_lkb *lkb,
  87				    struct dlm_message *ms);
  88static int receive_extralen(struct dlm_message *ms);
  89static void do_purge(struct dlm_ls *ls, int nodeid, int pid);
  90static void del_timeout(struct dlm_lkb *lkb);
  91static void toss_rsb(struct kref *kref);
  92
  93/*
  94 * Lock compatibilty matrix - thanks Steve
  95 * UN = Unlocked state. Not really a state, used as a flag
  96 * PD = Padding. Used to make the matrix a nice power of two in size
  97 * Other states are the same as the VMS DLM.
  98 * Usage: matrix[grmode+1][rqmode+1]  (although m[rq+1][gr+1] is the same)
  99 */
 100
 101static const int __dlm_compat_matrix[8][8] = {
 102      /* UN NL CR CW PR PW EX PD */
 103        {1, 1, 1, 1, 1, 1, 1, 0},       /* UN */
 104        {1, 1, 1, 1, 1, 1, 1, 0},       /* NL */
 105        {1, 1, 1, 1, 1, 1, 0, 0},       /* CR */
 106        {1, 1, 1, 1, 0, 0, 0, 0},       /* CW */
 107        {1, 1, 1, 0, 1, 0, 0, 0},       /* PR */
 108        {1, 1, 1, 0, 0, 0, 0, 0},       /* PW */
 109        {1, 1, 0, 0, 0, 0, 0, 0},       /* EX */
 110        {0, 0, 0, 0, 0, 0, 0, 0}        /* PD */
 111};
 112
 113/*
 114 * This defines the direction of transfer of LVB data.
 115 * Granted mode is the row; requested mode is the column.
 116 * Usage: matrix[grmode+1][rqmode+1]
 117 * 1 = LVB is returned to the caller
 118 * 0 = LVB is written to the resource
 119 * -1 = nothing happens to the LVB
 120 */
 121
 122const int dlm_lvb_operations[8][8] = {
 123        /* UN   NL  CR  CW  PR  PW  EX  PD*/
 124        {  -1,  1,  1,  1,  1,  1,  1, -1 }, /* UN */
 125        {  -1,  1,  1,  1,  1,  1,  1,  0 }, /* NL */
 126        {  -1, -1,  1,  1,  1,  1,  1,  0 }, /* CR */
 127        {  -1, -1, -1,  1,  1,  1,  1,  0 }, /* CW */
 128        {  -1, -1, -1, -1,  1,  1,  1,  0 }, /* PR */
 129        {  -1,  0,  0,  0,  0,  0,  1,  0 }, /* PW */
 130        {  -1,  0,  0,  0,  0,  0,  0,  0 }, /* EX */
 131        {  -1,  0,  0,  0,  0,  0,  0,  0 }  /* PD */
 132};
 133
 134#define modes_compat(gr, rq) \
 135	__dlm_compat_matrix[(gr)->lkb_grmode + 1][(rq)->lkb_rqmode + 1]
 136
 137int dlm_modes_compat(int mode1, int mode2)
 138{
 139	return __dlm_compat_matrix[mode1 + 1][mode2 + 1];
 140}
 141
 142/*
 143 * Compatibility matrix for conversions with QUECVT set.
 144 * Granted mode is the row; requested mode is the column.
 145 * Usage: matrix[grmode+1][rqmode+1]
 146 */
 147
 148static const int __quecvt_compat_matrix[8][8] = {
 149      /* UN NL CR CW PR PW EX PD */
 150        {0, 0, 0, 0, 0, 0, 0, 0},       /* UN */
 151        {0, 0, 1, 1, 1, 1, 1, 0},       /* NL */
 152        {0, 0, 0, 1, 1, 1, 1, 0},       /* CR */
 153        {0, 0, 0, 0, 1, 1, 1, 0},       /* CW */
 154        {0, 0, 0, 1, 0, 1, 1, 0},       /* PR */
 155        {0, 0, 0, 0, 0, 0, 1, 0},       /* PW */
 156        {0, 0, 0, 0, 0, 0, 0, 0},       /* EX */
 157        {0, 0, 0, 0, 0, 0, 0, 0}        /* PD */
 158};
 159
 160void dlm_print_lkb(struct dlm_lkb *lkb)
 161{
 162	printk(KERN_ERR "lkb: nodeid %d id %x remid %x exflags %x flags %x "
 163	       "sts %d rq %d gr %d wait_type %d wait_nodeid %d seq %llu\n",
 164	       lkb->lkb_nodeid, lkb->lkb_id, lkb->lkb_remid, lkb->lkb_exflags,
 165	       lkb->lkb_flags, lkb->lkb_status, lkb->lkb_rqmode,
 166	       lkb->lkb_grmode, lkb->lkb_wait_type, lkb->lkb_wait_nodeid,
 167	       (unsigned long long)lkb->lkb_recover_seq);
 168}
 169
 170static void dlm_print_rsb(struct dlm_rsb *r)
 171{
 172	printk(KERN_ERR "rsb: nodeid %d master %d dir %d flags %lx first %x "
 173	       "rlc %d name %s\n",
 174	       r->res_nodeid, r->res_master_nodeid, r->res_dir_nodeid,
 175	       r->res_flags, r->res_first_lkid, r->res_recover_locks_count,
 176	       r->res_name);
 177}
 178
 179void dlm_dump_rsb(struct dlm_rsb *r)
 180{
 181	struct dlm_lkb *lkb;
 182
 183	dlm_print_rsb(r);
 184
 185	printk(KERN_ERR "rsb: root_list empty %d recover_list empty %d\n",
 186	       list_empty(&r->res_root_list), list_empty(&r->res_recover_list));
 187	printk(KERN_ERR "rsb lookup list\n");
 188	list_for_each_entry(lkb, &r->res_lookup, lkb_rsb_lookup)
 189		dlm_print_lkb(lkb);
 190	printk(KERN_ERR "rsb grant queue:\n");
 191	list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue)
 192		dlm_print_lkb(lkb);
 193	printk(KERN_ERR "rsb convert queue:\n");
 194	list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue)
 195		dlm_print_lkb(lkb);
 196	printk(KERN_ERR "rsb wait queue:\n");
 197	list_for_each_entry(lkb, &r->res_waitqueue, lkb_statequeue)
 198		dlm_print_lkb(lkb);
 199}
 200
 201/* Threads cannot use the lockspace while it's being recovered */
 202
 203static inline void dlm_lock_recovery(struct dlm_ls *ls)
 204{
 205	down_read(&ls->ls_in_recovery);
 206}
 207
 208void dlm_unlock_recovery(struct dlm_ls *ls)
 209{
 210	up_read(&ls->ls_in_recovery);
 211}
 212
 213int dlm_lock_recovery_try(struct dlm_ls *ls)
 214{
 215	return down_read_trylock(&ls->ls_in_recovery);
 216}
 217
 218static inline int can_be_queued(struct dlm_lkb *lkb)
 219{
 220	return !(lkb->lkb_exflags & DLM_LKF_NOQUEUE);
 221}
 222
 223static inline int force_blocking_asts(struct dlm_lkb *lkb)
 224{
 225	return (lkb->lkb_exflags & DLM_LKF_NOQUEUEBAST);
 226}
 227
 228static inline int is_demoted(struct dlm_lkb *lkb)
 229{
 230	return (lkb->lkb_sbflags & DLM_SBF_DEMOTED);
 231}
 232
 233static inline int is_altmode(struct dlm_lkb *lkb)
 234{
 235	return (lkb->lkb_sbflags & DLM_SBF_ALTMODE);
 236}
 237
 238static inline int is_granted(struct dlm_lkb *lkb)
 239{
 240	return (lkb->lkb_status == DLM_LKSTS_GRANTED);
 241}
 242
 243static inline int is_remote(struct dlm_rsb *r)
 244{
 245	DLM_ASSERT(r->res_nodeid >= 0, dlm_print_rsb(r););
 246	return !!r->res_nodeid;
 247}
 248
 249static inline int is_process_copy(struct dlm_lkb *lkb)
 250{
 251	return (lkb->lkb_nodeid && !(lkb->lkb_flags & DLM_IFL_MSTCPY));
 
 252}
 253
 254static inline int is_master_copy(struct dlm_lkb *lkb)
 255{
 256	return (lkb->lkb_flags & DLM_IFL_MSTCPY) ? 1 : 0;
 257}
 258
 259static inline int middle_conversion(struct dlm_lkb *lkb)
 260{
 261	if ((lkb->lkb_grmode==DLM_LOCK_PR && lkb->lkb_rqmode==DLM_LOCK_CW) ||
 262	    (lkb->lkb_rqmode==DLM_LOCK_PR && lkb->lkb_grmode==DLM_LOCK_CW))
 263		return 1;
 264	return 0;
 265}
 266
 267static inline int down_conversion(struct dlm_lkb *lkb)
 268{
 269	return (!middle_conversion(lkb) && lkb->lkb_rqmode < lkb->lkb_grmode);
 270}
 271
 272static inline int is_overlap_unlock(struct dlm_lkb *lkb)
 273{
 274	return lkb->lkb_flags & DLM_IFL_OVERLAP_UNLOCK;
 275}
 276
 277static inline int is_overlap_cancel(struct dlm_lkb *lkb)
 278{
 279	return lkb->lkb_flags & DLM_IFL_OVERLAP_CANCEL;
 280}
 281
 282static inline int is_overlap(struct dlm_lkb *lkb)
 283{
 284	return (lkb->lkb_flags & (DLM_IFL_OVERLAP_UNLOCK |
 285				  DLM_IFL_OVERLAP_CANCEL));
 286}
 287
 288static void queue_cast(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
 289{
 290	if (is_master_copy(lkb))
 291		return;
 292
 293	del_timeout(lkb);
 294
 295	DLM_ASSERT(lkb->lkb_lksb, dlm_print_lkb(lkb););
 296
 297	/* if the operation was a cancel, then return -DLM_ECANCEL, if a
 298	   timeout caused the cancel then return -ETIMEDOUT */
 299	if (rv == -DLM_ECANCEL && (lkb->lkb_flags & DLM_IFL_TIMEOUT_CANCEL)) {
 300		lkb->lkb_flags &= ~DLM_IFL_TIMEOUT_CANCEL;
 301		rv = -ETIMEDOUT;
 302	}
 303
 304	if (rv == -DLM_ECANCEL && (lkb->lkb_flags & DLM_IFL_DEADLOCK_CANCEL)) {
 305		lkb->lkb_flags &= ~DLM_IFL_DEADLOCK_CANCEL;
 306		rv = -EDEADLK;
 307	}
 308
 309	dlm_add_cb(lkb, DLM_CB_CAST, lkb->lkb_grmode, rv, lkb->lkb_sbflags);
 310}
 311
 312static inline void queue_cast_overlap(struct dlm_rsb *r, struct dlm_lkb *lkb)
 313{
 314	queue_cast(r, lkb,
 315		   is_overlap_unlock(lkb) ? -DLM_EUNLOCK : -DLM_ECANCEL);
 316}
 317
 318static void queue_bast(struct dlm_rsb *r, struct dlm_lkb *lkb, int rqmode)
 319{
 320	if (is_master_copy(lkb)) {
 321		send_bast(r, lkb, rqmode);
 322	} else {
 323		dlm_add_cb(lkb, DLM_CB_BAST, rqmode, 0, 0);
 324	}
 325}
 326
 327/*
 328 * Basic operations on rsb's and lkb's
 329 */
 330
 331/* This is only called to add a reference when the code already holds
 332   a valid reference to the rsb, so there's no need for locking. */
 333
 334static inline void hold_rsb(struct dlm_rsb *r)
 335{
 336	kref_get(&r->res_ref);
 337}
 338
 339void dlm_hold_rsb(struct dlm_rsb *r)
 340{
 341	hold_rsb(r);
 342}
 343
 344/* When all references to the rsb are gone it's transferred to
 345   the tossed list for later disposal. */
 346
 347static void put_rsb(struct dlm_rsb *r)
 348{
 349	struct dlm_ls *ls = r->res_ls;
 350	uint32_t bucket = r->res_bucket;
 
 351
 352	spin_lock(&ls->ls_rsbtbl[bucket].lock);
 353	kref_put(&r->res_ref, toss_rsb);
 354	spin_unlock(&ls->ls_rsbtbl[bucket].lock);
 
 355}
 356
 357void dlm_put_rsb(struct dlm_rsb *r)
 358{
 359	put_rsb(r);
 360}
 361
 362static int pre_rsb_struct(struct dlm_ls *ls)
 363{
 364	struct dlm_rsb *r1, *r2;
 365	int count = 0;
 366
 367	spin_lock(&ls->ls_new_rsb_spin);
 368	if (ls->ls_new_rsb_count > dlm_config.ci_new_rsb_count / 2) {
 369		spin_unlock(&ls->ls_new_rsb_spin);
 370		return 0;
 371	}
 372	spin_unlock(&ls->ls_new_rsb_spin);
 373
 374	r1 = dlm_allocate_rsb(ls);
 375	r2 = dlm_allocate_rsb(ls);
 376
 377	spin_lock(&ls->ls_new_rsb_spin);
 378	if (r1) {
 379		list_add(&r1->res_hashchain, &ls->ls_new_rsb);
 380		ls->ls_new_rsb_count++;
 381	}
 382	if (r2) {
 383		list_add(&r2->res_hashchain, &ls->ls_new_rsb);
 384		ls->ls_new_rsb_count++;
 385	}
 386	count = ls->ls_new_rsb_count;
 387	spin_unlock(&ls->ls_new_rsb_spin);
 388
 389	if (!count)
 390		return -ENOMEM;
 391	return 0;
 392}
 393
 394/* If ls->ls_new_rsb is empty, return -EAGAIN, so the caller can
 395   unlock any spinlocks, go back and call pre_rsb_struct again.
 396   Otherwise, take an rsb off the list and return it. */
 397
 398static int get_rsb_struct(struct dlm_ls *ls, char *name, int len,
 399			  struct dlm_rsb **r_ret)
 400{
 401	struct dlm_rsb *r;
 402	int count;
 403
 404	spin_lock(&ls->ls_new_rsb_spin);
 405	if (list_empty(&ls->ls_new_rsb)) {
 406		count = ls->ls_new_rsb_count;
 407		spin_unlock(&ls->ls_new_rsb_spin);
 408		log_debug(ls, "find_rsb retry %d %d %s",
 409			  count, dlm_config.ci_new_rsb_count, name);
 
 410		return -EAGAIN;
 411	}
 412
 413	r = list_first_entry(&ls->ls_new_rsb, struct dlm_rsb, res_hashchain);
 414	list_del(&r->res_hashchain);
 415	/* Convert the empty list_head to a NULL rb_node for tree usage: */
 416	memset(&r->res_hashnode, 0, sizeof(struct rb_node));
 417	ls->ls_new_rsb_count--;
 418	spin_unlock(&ls->ls_new_rsb_spin);
 419
 420	r->res_ls = ls;
 421	r->res_length = len;
 422	memcpy(r->res_name, name, len);
 423	mutex_init(&r->res_mutex);
 424
 425	INIT_LIST_HEAD(&r->res_lookup);
 426	INIT_LIST_HEAD(&r->res_grantqueue);
 427	INIT_LIST_HEAD(&r->res_convertqueue);
 428	INIT_LIST_HEAD(&r->res_waitqueue);
 429	INIT_LIST_HEAD(&r->res_root_list);
 430	INIT_LIST_HEAD(&r->res_recover_list);
 431
 432	*r_ret = r;
 433	return 0;
 434}
 435
 436static int rsb_cmp(struct dlm_rsb *r, const char *name, int nlen)
 437{
 438	char maxname[DLM_RESNAME_MAXLEN];
 439
 440	memset(maxname, 0, DLM_RESNAME_MAXLEN);
 441	memcpy(maxname, name, nlen);
 442	return memcmp(r->res_name, maxname, DLM_RESNAME_MAXLEN);
 443}
 444
 445int dlm_search_rsb_tree(struct rb_root *tree, char *name, int len,
 446			struct dlm_rsb **r_ret)
 447{
 448	struct rb_node *node = tree->rb_node;
 449	struct dlm_rsb *r;
 450	int rc;
 451
 452	while (node) {
 453		r = rb_entry(node, struct dlm_rsb, res_hashnode);
 454		rc = rsb_cmp(r, name, len);
 455		if (rc < 0)
 456			node = node->rb_left;
 457		else if (rc > 0)
 458			node = node->rb_right;
 459		else
 460			goto found;
 461	}
 462	*r_ret = NULL;
 463	return -EBADR;
 464
 465 found:
 466	*r_ret = r;
 467	return 0;
 468}
 469
 470static int rsb_insert(struct dlm_rsb *rsb, struct rb_root *tree)
 471{
 472	struct rb_node **newn = &tree->rb_node;
 473	struct rb_node *parent = NULL;
 474	int rc;
 475
 476	while (*newn) {
 477		struct dlm_rsb *cur = rb_entry(*newn, struct dlm_rsb,
 478					       res_hashnode);
 479
 480		parent = *newn;
 481		rc = rsb_cmp(cur, rsb->res_name, rsb->res_length);
 482		if (rc < 0)
 483			newn = &parent->rb_left;
 484		else if (rc > 0)
 485			newn = &parent->rb_right;
 486		else {
 487			log_print("rsb_insert match");
 488			dlm_dump_rsb(rsb);
 489			dlm_dump_rsb(cur);
 490			return -EEXIST;
 491		}
 492	}
 493
 494	rb_link_node(&rsb->res_hashnode, parent, newn);
 495	rb_insert_color(&rsb->res_hashnode, tree);
 496	return 0;
 497}
 498
 499/*
 500 * Find rsb in rsbtbl and potentially create/add one
 501 *
 502 * Delaying the release of rsb's has a similar benefit to applications keeping
 503 * NL locks on an rsb, but without the guarantee that the cached master value
 504 * will still be valid when the rsb is reused.  Apps aren't always smart enough
 505 * to keep NL locks on an rsb that they may lock again shortly; this can lead
 506 * to excessive master lookups and removals if we don't delay the release.
 507 *
 508 * Searching for an rsb means looking through both the normal list and toss
 509 * list.  When found on the toss list the rsb is moved to the normal list with
 510 * ref count of 1; when found on normal list the ref count is incremented.
 511 *
 512 * rsb's on the keep list are being used locally and refcounted.
 513 * rsb's on the toss list are not being used locally, and are not refcounted.
 514 *
 515 * The toss list rsb's were either
 516 * - previously used locally but not any more (were on keep list, then
 517 *   moved to toss list when last refcount dropped)
 518 * - created and put on toss list as a directory record for a lookup
 519 *   (we are the dir node for the res, but are not using the res right now,
 520 *   but some other node is)
 521 *
 522 * The purpose of find_rsb() is to return a refcounted rsb for local use.
 523 * So, if the given rsb is on the toss list, it is moved to the keep list
 524 * before being returned.
 525 *
 526 * toss_rsb() happens when all local usage of the rsb is done, i.e. no
 527 * more refcounts exist, so the rsb is moved from the keep list to the
 528 * toss list.
 529 *
 530 * rsb's on both keep and toss lists are used for doing a name to master
 531 * lookups.  rsb's that are in use locally (and being refcounted) are on
 532 * the keep list, rsb's that are not in use locally (not refcounted) and
 533 * only exist for name/master lookups are on the toss list.
 534 *
 535 * rsb's on the toss list who's dir_nodeid is not local can have stale
 536 * name/master mappings.  So, remote requests on such rsb's can potentially
 537 * return with an error, which means the mapping is stale and needs to
 538 * be updated with a new lookup.  (The idea behind MASTER UNCERTAIN and
 539 * first_lkid is to keep only a single outstanding request on an rsb
 540 * while that rsb has a potentially stale master.)
 541 */
 542
 543static int find_rsb_dir(struct dlm_ls *ls, char *name, int len,
 544			uint32_t hash, uint32_t b,
 545			int dir_nodeid, int from_nodeid,
 546			unsigned int flags, struct dlm_rsb **r_ret)
 547{
 548	struct dlm_rsb *r = NULL;
 549	int our_nodeid = dlm_our_nodeid();
 550	int from_local = 0;
 551	int from_other = 0;
 552	int from_dir = 0;
 553	int create = 0;
 554	int error;
 555
 556	if (flags & R_RECEIVE_REQUEST) {
 557		if (from_nodeid == dir_nodeid)
 558			from_dir = 1;
 559		else
 560			from_other = 1;
 561	} else if (flags & R_REQUEST) {
 562		from_local = 1;
 563	}
 564
 565	/*
 566	 * flags & R_RECEIVE_RECOVER is from dlm_recover_master_copy, so
 567	 * from_nodeid has sent us a lock in dlm_recover_locks, believing
 568	 * we're the new master.  Our local recovery may not have set
 569	 * res_master_nodeid to our_nodeid yet, so allow either.  Don't
 570	 * create the rsb; dlm_recover_process_copy() will handle EBADR
 571	 * by resending.
 572	 *
 573	 * If someone sends us a request, we are the dir node, and we do
 574	 * not find the rsb anywhere, then recreate it.  This happens if
 575	 * someone sends us a request after we have removed/freed an rsb
 576	 * from our toss list.  (They sent a request instead of lookup
 577	 * because they are using an rsb from their toss list.)
 578	 */
 579
 580	if (from_local || from_dir ||
 581	    (from_other && (dir_nodeid == our_nodeid))) {
 582		create = 1;
 583	}
 584
 585 retry:
 586	if (create) {
 587		error = pre_rsb_struct(ls);
 588		if (error < 0)
 589			goto out;
 590	}
 591
 592	spin_lock(&ls->ls_rsbtbl[b].lock);
 593
 594	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
 595	if (error)
 596		goto do_toss;
 597	
 598	/*
 599	 * rsb is active, so we can't check master_nodeid without lock_rsb.
 600	 */
 601
 602	kref_get(&r->res_ref);
 603	error = 0;
 604	goto out_unlock;
 605
 606
 607 do_toss:
 608	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
 609	if (error)
 610		goto do_new;
 611
 612	/*
 613	 * rsb found inactive (master_nodeid may be out of date unless
 614	 * we are the dir_nodeid or were the master)  No other thread
 615	 * is using this rsb because it's on the toss list, so we can
 616	 * look at or update res_master_nodeid without lock_rsb.
 617	 */
 618
 619	if ((r->res_master_nodeid != our_nodeid) && from_other) {
 620		/* our rsb was not master, and another node (not the dir node)
 621		   has sent us a request */
 622		log_debug(ls, "find_rsb toss from_other %d master %d dir %d %s",
 623			  from_nodeid, r->res_master_nodeid, dir_nodeid,
 624			  r->res_name);
 625		error = -ENOTBLK;
 626		goto out_unlock;
 627	}
 628
 629	if ((r->res_master_nodeid != our_nodeid) && from_dir) {
 630		/* don't think this should ever happen */
 631		log_error(ls, "find_rsb toss from_dir %d master %d",
 632			  from_nodeid, r->res_master_nodeid);
 633		dlm_print_rsb(r);
 634		/* fix it and go on */
 635		r->res_master_nodeid = our_nodeid;
 636		r->res_nodeid = 0;
 637		rsb_clear_flag(r, RSB_MASTER_UNCERTAIN);
 638		r->res_first_lkid = 0;
 639	}
 640
 641	if (from_local && (r->res_master_nodeid != our_nodeid)) {
 642		/* Because we have held no locks on this rsb,
 643		   res_master_nodeid could have become stale. */
 644		rsb_set_flag(r, RSB_MASTER_UNCERTAIN);
 645		r->res_first_lkid = 0;
 646	}
 647
 648	rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
 649	error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
 650	goto out_unlock;
 651
 652
 653 do_new:
 654	/*
 655	 * rsb not found
 656	 */
 657
 658	if (error == -EBADR && !create)
 659		goto out_unlock;
 660
 661	error = get_rsb_struct(ls, name, len, &r);
 662	if (error == -EAGAIN) {
 663		spin_unlock(&ls->ls_rsbtbl[b].lock);
 664		goto retry;
 665	}
 666	if (error)
 667		goto out_unlock;
 668
 669	r->res_hash = hash;
 670	r->res_bucket = b;
 671	r->res_dir_nodeid = dir_nodeid;
 672	kref_init(&r->res_ref);
 673
 674	if (from_dir) {
 675		/* want to see how often this happens */
 676		log_debug(ls, "find_rsb new from_dir %d recreate %s",
 677			  from_nodeid, r->res_name);
 678		r->res_master_nodeid = our_nodeid;
 679		r->res_nodeid = 0;
 680		goto out_add;
 681	}
 682
 683	if (from_other && (dir_nodeid != our_nodeid)) {
 684		/* should never happen */
 685		log_error(ls, "find_rsb new from_other %d dir %d our %d %s",
 686			  from_nodeid, dir_nodeid, our_nodeid, r->res_name);
 687		dlm_free_rsb(r);
 688		r = NULL;
 689		error = -ENOTBLK;
 690		goto out_unlock;
 691	}
 692
 693	if (from_other) {
 694		log_debug(ls, "find_rsb new from_other %d dir %d %s",
 695			  from_nodeid, dir_nodeid, r->res_name);
 696	}
 697
 698	if (dir_nodeid == our_nodeid) {
 699		/* When we are the dir nodeid, we can set the master
 700		   node immediately */
 701		r->res_master_nodeid = our_nodeid;
 702		r->res_nodeid = 0;
 703	} else {
 704		/* set_master will send_lookup to dir_nodeid */
 705		r->res_master_nodeid = 0;
 706		r->res_nodeid = -1;
 707	}
 708
 709 out_add:
 710	error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
 711 out_unlock:
 712	spin_unlock(&ls->ls_rsbtbl[b].lock);
 713 out:
 714	*r_ret = r;
 715	return error;
 716}
 717
 718/* During recovery, other nodes can send us new MSTCPY locks (from
 719   dlm_recover_locks) before we've made ourself master (in
 720   dlm_recover_masters). */
 721
 722static int find_rsb_nodir(struct dlm_ls *ls, char *name, int len,
 723			  uint32_t hash, uint32_t b,
 724			  int dir_nodeid, int from_nodeid,
 725			  unsigned int flags, struct dlm_rsb **r_ret)
 726{
 727	struct dlm_rsb *r = NULL;
 728	int our_nodeid = dlm_our_nodeid();
 729	int recover = (flags & R_RECEIVE_RECOVER);
 730	int error;
 731
 732 retry:
 733	error = pre_rsb_struct(ls);
 734	if (error < 0)
 735		goto out;
 736
 737	spin_lock(&ls->ls_rsbtbl[b].lock);
 738
 739	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
 740	if (error)
 741		goto do_toss;
 742
 743	/*
 744	 * rsb is active, so we can't check master_nodeid without lock_rsb.
 745	 */
 746
 747	kref_get(&r->res_ref);
 748	goto out_unlock;
 749
 750
 751 do_toss:
 752	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
 753	if (error)
 754		goto do_new;
 755
 756	/*
 757	 * rsb found inactive. No other thread is using this rsb because
 758	 * it's on the toss list, so we can look at or update
 759	 * res_master_nodeid without lock_rsb.
 760	 */
 761
 762	if (!recover && (r->res_master_nodeid != our_nodeid) && from_nodeid) {
 763		/* our rsb is not master, and another node has sent us a
 764		   request; this should never happen */
 765		log_error(ls, "find_rsb toss from_nodeid %d master %d dir %d",
 766			  from_nodeid, r->res_master_nodeid, dir_nodeid);
 767		dlm_print_rsb(r);
 768		error = -ENOTBLK;
 769		goto out_unlock;
 770	}
 771
 772	if (!recover && (r->res_master_nodeid != our_nodeid) &&
 773	    (dir_nodeid == our_nodeid)) {
 774		/* our rsb is not master, and we are dir; may as well fix it;
 775		   this should never happen */
 776		log_error(ls, "find_rsb toss our %d master %d dir %d",
 777			  our_nodeid, r->res_master_nodeid, dir_nodeid);
 778		dlm_print_rsb(r);
 779		r->res_master_nodeid = our_nodeid;
 780		r->res_nodeid = 0;
 781	}
 782
 783	rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
 784	error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
 785	goto out_unlock;
 786
 787
 788 do_new:
 789	/*
 790	 * rsb not found
 791	 */
 792
 793	error = get_rsb_struct(ls, name, len, &r);
 794	if (error == -EAGAIN) {
 795		spin_unlock(&ls->ls_rsbtbl[b].lock);
 796		goto retry;
 797	}
 798	if (error)
 799		goto out_unlock;
 800
 801	r->res_hash = hash;
 802	r->res_bucket = b;
 803	r->res_dir_nodeid = dir_nodeid;
 804	r->res_master_nodeid = dir_nodeid;
 805	r->res_nodeid = (dir_nodeid == our_nodeid) ? 0 : dir_nodeid;
 806	kref_init(&r->res_ref);
 807
 808	error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
 809 out_unlock:
 810	spin_unlock(&ls->ls_rsbtbl[b].lock);
 811 out:
 812	*r_ret = r;
 813	return error;
 814}
 815
 816static int find_rsb(struct dlm_ls *ls, char *name, int len, int from_nodeid,
 817		    unsigned int flags, struct dlm_rsb **r_ret)
 
 818{
 819	uint32_t hash, b;
 820	int dir_nodeid;
 821
 822	if (len > DLM_RESNAME_MAXLEN)
 823		return -EINVAL;
 824
 825	hash = jhash(name, len, 0);
 826	b = hash & (ls->ls_rsbtbl_size - 1);
 827
 828	dir_nodeid = dlm_hash2nodeid(ls, hash);
 829
 830	if (dlm_no_directory(ls))
 831		return find_rsb_nodir(ls, name, len, hash, b, dir_nodeid,
 832				      from_nodeid, flags, r_ret);
 833	else
 834		return find_rsb_dir(ls, name, len, hash, b, dir_nodeid,
 835				      from_nodeid, flags, r_ret);
 836}
 837
 838/* we have received a request and found that res_master_nodeid != our_nodeid,
 839   so we need to return an error or make ourself the master */
 840
 841static int validate_master_nodeid(struct dlm_ls *ls, struct dlm_rsb *r,
 842				  int from_nodeid)
 843{
 844	if (dlm_no_directory(ls)) {
 845		log_error(ls, "find_rsb keep from_nodeid %d master %d dir %d",
 846			  from_nodeid, r->res_master_nodeid,
 847			  r->res_dir_nodeid);
 848		dlm_print_rsb(r);
 849		return -ENOTBLK;
 850	}
 851
 852	if (from_nodeid != r->res_dir_nodeid) {
 853		/* our rsb is not master, and another node (not the dir node)
 854	   	   has sent us a request.  this is much more common when our
 855	   	   master_nodeid is zero, so limit debug to non-zero.  */
 856
 857		if (r->res_master_nodeid) {
 858			log_debug(ls, "validate master from_other %d master %d "
 859				  "dir %d first %x %s", from_nodeid,
 860				  r->res_master_nodeid, r->res_dir_nodeid,
 861				  r->res_first_lkid, r->res_name);
 862		}
 863		return -ENOTBLK;
 864	} else {
 865		/* our rsb is not master, but the dir nodeid has sent us a
 866	   	   request; this could happen with master 0 / res_nodeid -1 */
 867
 868		if (r->res_master_nodeid) {
 869			log_error(ls, "validate master from_dir %d master %d "
 870				  "first %x %s",
 871				  from_nodeid, r->res_master_nodeid,
 872				  r->res_first_lkid, r->res_name);
 873		}
 874
 875		r->res_master_nodeid = dlm_our_nodeid();
 876		r->res_nodeid = 0;
 877		return 0;
 878	}
 879}
 880
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 881/*
 882 * We're the dir node for this res and another node wants to know the
 883 * master nodeid.  During normal operation (non recovery) this is only
 884 * called from receive_lookup(); master lookups when the local node is
 885 * the dir node are done by find_rsb().
 886 *
 887 * normal operation, we are the dir node for a resource
 888 * . _request_lock
 889 * . set_master
 890 * . send_lookup
 891 * . receive_lookup
 892 * . dlm_master_lookup flags 0
 893 *
 894 * recover directory, we are rebuilding dir for all resources
 895 * . dlm_recover_directory
 896 * . dlm_rcom_names
 897 *   remote node sends back the rsb names it is master of and we are dir of
 898 * . dlm_master_lookup RECOVER_DIR (fix_master 0, from_master 1)
 899 *   we either create new rsb setting remote node as master, or find existing
 900 *   rsb and set master to be the remote node.
 901 *
 902 * recover masters, we are finding the new master for resources
 903 * . dlm_recover_masters
 904 * . recover_master
 905 * . dlm_send_rcom_lookup
 906 * . receive_rcom_lookup
 907 * . dlm_master_lookup RECOVER_MASTER (fix_master 1, from_master 0)
 908 */
 909
 910int dlm_master_lookup(struct dlm_ls *ls, int from_nodeid, char *name, int len,
 911		      unsigned int flags, int *r_nodeid, int *result)
 912{
 913	struct dlm_rsb *r = NULL;
 914	uint32_t hash, b;
 915	int from_master = (flags & DLM_LU_RECOVER_DIR);
 916	int fix_master = (flags & DLM_LU_RECOVER_MASTER);
 917	int our_nodeid = dlm_our_nodeid();
 918	int dir_nodeid, error, toss_list = 0;
 919
 920	if (len > DLM_RESNAME_MAXLEN)
 921		return -EINVAL;
 922
 923	if (from_nodeid == our_nodeid) {
 924		log_error(ls, "dlm_master_lookup from our_nodeid %d flags %x",
 925			  our_nodeid, flags);
 926		return -EINVAL;
 927	}
 928
 929	hash = jhash(name, len, 0);
 930	b = hash & (ls->ls_rsbtbl_size - 1);
 931
 932	dir_nodeid = dlm_hash2nodeid(ls, hash);
 933	if (dir_nodeid != our_nodeid) {
 934		log_error(ls, "dlm_master_lookup from %d dir %d our %d h %x %d",
 935			  from_nodeid, dir_nodeid, our_nodeid, hash,
 936			  ls->ls_num_nodes);
 937		*r_nodeid = -1;
 938		return -EINVAL;
 939	}
 940
 941 retry:
 942	error = pre_rsb_struct(ls);
 943	if (error < 0)
 944		return error;
 945
 946	spin_lock(&ls->ls_rsbtbl[b].lock);
 947	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
 948	if (!error) {
 949		/* because the rsb is active, we need to lock_rsb before
 950		   checking/changing re_master_nodeid */
 
 951
 952		hold_rsb(r);
 953		spin_unlock(&ls->ls_rsbtbl[b].lock);
 954		lock_rsb(r);
 955		goto found;
 
 
 
 
 
 
 
 
 956	}
 957
 958	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
 959	if (error)
 960		goto not_found;
 961
 962	/* because the rsb is inactive (on toss list), it's not refcounted
 963	   and lock_rsb is not used, but is protected by the rsbtbl lock */
 964
 965	toss_list = 1;
 966 found:
 967	if (r->res_dir_nodeid != our_nodeid) {
 968		/* should not happen, but may as well fix it and carry on */
 969		log_error(ls, "dlm_master_lookup res_dir %d our %d %s",
 970			  r->res_dir_nodeid, our_nodeid, r->res_name);
 971		r->res_dir_nodeid = our_nodeid;
 972	}
 973
 974	if (fix_master && dlm_is_removed(ls, r->res_master_nodeid)) {
 975		/* Recovery uses this function to set a new master when
 976		   the previous master failed.  Setting NEW_MASTER will
 977		   force dlm_recover_masters to call recover_master on this
 978		   rsb even though the res_nodeid is no longer removed. */
 979
 980		r->res_master_nodeid = from_nodeid;
 981		r->res_nodeid = from_nodeid;
 982		rsb_set_flag(r, RSB_NEW_MASTER);
 983
 984		if (toss_list) {
 985			/* I don't think we should ever find it on toss list. */
 986			log_error(ls, "dlm_master_lookup fix_master on toss");
 987			dlm_dump_rsb(r);
 988		}
 989	}
 990
 991	if (from_master && (r->res_master_nodeid != from_nodeid)) {
 992		/* this will happen if from_nodeid became master during
 993		   a previous recovery cycle, and we aborted the previous
 994		   cycle before recovering this master value */
 995
 996		log_limit(ls, "dlm_master_lookup from_master %d "
 997			  "master_nodeid %d res_nodeid %d first %x %s",
 998			  from_nodeid, r->res_master_nodeid, r->res_nodeid,
 999			  r->res_first_lkid, r->res_name);
1000
1001		if (r->res_master_nodeid == our_nodeid) {
1002			log_error(ls, "from_master %d our_master", from_nodeid);
1003			dlm_dump_rsb(r);
1004			goto out_found;
1005		}
1006
1007		r->res_master_nodeid = from_nodeid;
1008		r->res_nodeid = from_nodeid;
1009		rsb_set_flag(r, RSB_NEW_MASTER);
1010	}
1011
1012	if (!r->res_master_nodeid) {
1013		/* this will happen if recovery happens while we're looking
1014		   up the master for this rsb */
1015
1016		log_debug(ls, "dlm_master_lookup master 0 to %d first %x %s",
1017			  from_nodeid, r->res_first_lkid, r->res_name);
1018		r->res_master_nodeid = from_nodeid;
1019		r->res_nodeid = from_nodeid;
1020	}
1021
1022	if (!from_master && !fix_master &&
1023	    (r->res_master_nodeid == from_nodeid)) {
1024		/* this can happen when the master sends remove, the dir node
1025		   finds the rsb on the keep list and ignores the remove,
1026		   and the former master sends a lookup */
1027
1028		log_limit(ls, "dlm_master_lookup from master %d flags %x "
1029			  "first %x %s", from_nodeid, flags,
1030			  r->res_first_lkid, r->res_name);
1031	}
1032
1033 out_found:
1034	*r_nodeid = r->res_master_nodeid;
1035	if (result)
1036		*result = DLM_LU_MATCH;
1037
1038	if (toss_list) {
1039		r->res_toss_time = jiffies;
1040		/* the rsb was inactive (on toss list) */
1041		spin_unlock(&ls->ls_rsbtbl[b].lock);
1042	} else {
1043		/* the rsb was active */
1044		unlock_rsb(r);
1045		put_rsb(r);
1046	}
1047	return 0;
1048
1049 not_found:
1050	error = get_rsb_struct(ls, name, len, &r);
1051	if (error == -EAGAIN) {
1052		spin_unlock(&ls->ls_rsbtbl[b].lock);
1053		goto retry;
1054	}
1055	if (error)
1056		goto out_unlock;
1057
1058	r->res_hash = hash;
1059	r->res_bucket = b;
1060	r->res_dir_nodeid = our_nodeid;
1061	r->res_master_nodeid = from_nodeid;
1062	r->res_nodeid = from_nodeid;
1063	kref_init(&r->res_ref);
1064	r->res_toss_time = jiffies;
1065
1066	error = rsb_insert(r, &ls->ls_rsbtbl[b].toss);
1067	if (error) {
1068		/* should never happen */
1069		dlm_free_rsb(r);
1070		spin_unlock(&ls->ls_rsbtbl[b].lock);
1071		goto retry;
1072	}
1073
1074	if (result)
1075		*result = DLM_LU_ADD;
1076	*r_nodeid = from_nodeid;
1077	error = 0;
1078 out_unlock:
1079	spin_unlock(&ls->ls_rsbtbl[b].lock);
1080	return error;
1081}
1082
1083static void dlm_dump_rsb_hash(struct dlm_ls *ls, uint32_t hash)
1084{
1085	struct rb_node *n;
1086	struct dlm_rsb *r;
1087	int i;
1088
1089	for (i = 0; i < ls->ls_rsbtbl_size; i++) {
1090		spin_lock(&ls->ls_rsbtbl[i].lock);
1091		for (n = rb_first(&ls->ls_rsbtbl[i].keep); n; n = rb_next(n)) {
1092			r = rb_entry(n, struct dlm_rsb, res_hashnode);
1093			if (r->res_hash == hash)
1094				dlm_dump_rsb(r);
1095		}
1096		spin_unlock(&ls->ls_rsbtbl[i].lock);
1097	}
1098}
1099
1100void dlm_dump_rsb_name(struct dlm_ls *ls, char *name, int len)
1101{
1102	struct dlm_rsb *r = NULL;
1103	uint32_t hash, b;
1104	int error;
1105
1106	hash = jhash(name, len, 0);
1107	b = hash & (ls->ls_rsbtbl_size - 1);
1108
1109	spin_lock(&ls->ls_rsbtbl[b].lock);
1110	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
1111	if (!error)
1112		goto out_dump;
1113
1114	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
1115	if (error)
1116		goto out;
1117 out_dump:
1118	dlm_dump_rsb(r);
1119 out:
1120	spin_unlock(&ls->ls_rsbtbl[b].lock);
1121}
1122
1123static void toss_rsb(struct kref *kref)
1124{
1125	struct dlm_rsb *r = container_of(kref, struct dlm_rsb, res_ref);
1126	struct dlm_ls *ls = r->res_ls;
1127
1128	DLM_ASSERT(list_empty(&r->res_root_list), dlm_print_rsb(r););
1129	kref_init(&r->res_ref);
1130	rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[r->res_bucket].keep);
1131	rsb_insert(r, &ls->ls_rsbtbl[r->res_bucket].toss);
1132	r->res_toss_time = jiffies;
1133	ls->ls_rsbtbl[r->res_bucket].flags |= DLM_RTF_SHRINK;
1134	if (r->res_lvbptr) {
1135		dlm_free_lvb(r->res_lvbptr);
1136		r->res_lvbptr = NULL;
1137	}
1138}
1139
1140/* See comment for unhold_lkb */
1141
1142static void unhold_rsb(struct dlm_rsb *r)
1143{
1144	int rv;
1145	rv = kref_put(&r->res_ref, toss_rsb);
1146	DLM_ASSERT(!rv, dlm_dump_rsb(r););
1147}
1148
1149static void kill_rsb(struct kref *kref)
1150{
1151	struct dlm_rsb *r = container_of(kref, struct dlm_rsb, res_ref);
1152
1153	/* All work is done after the return from kref_put() so we
1154	   can release the write_lock before the remove and free. */
1155
1156	DLM_ASSERT(list_empty(&r->res_lookup), dlm_dump_rsb(r););
1157	DLM_ASSERT(list_empty(&r->res_grantqueue), dlm_dump_rsb(r););
1158	DLM_ASSERT(list_empty(&r->res_convertqueue), dlm_dump_rsb(r););
1159	DLM_ASSERT(list_empty(&r->res_waitqueue), dlm_dump_rsb(r););
1160	DLM_ASSERT(list_empty(&r->res_root_list), dlm_dump_rsb(r););
1161	DLM_ASSERT(list_empty(&r->res_recover_list), dlm_dump_rsb(r););
1162}
1163
1164/* Attaching/detaching lkb's from rsb's is for rsb reference counting.
1165   The rsb must exist as long as any lkb's for it do. */
1166
1167static void attach_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb)
1168{
1169	hold_rsb(r);
1170	lkb->lkb_resource = r;
1171}
1172
1173static void detach_lkb(struct dlm_lkb *lkb)
1174{
1175	if (lkb->lkb_resource) {
1176		put_rsb(lkb->lkb_resource);
1177		lkb->lkb_resource = NULL;
1178	}
1179}
1180
1181static int create_lkb(struct dlm_ls *ls, struct dlm_lkb **lkb_ret)
 
1182{
1183	struct dlm_lkb *lkb;
1184	int rv;
1185
1186	lkb = dlm_allocate_lkb(ls);
1187	if (!lkb)
1188		return -ENOMEM;
1189
 
1190	lkb->lkb_nodeid = -1;
1191	lkb->lkb_grmode = DLM_LOCK_IV;
1192	kref_init(&lkb->lkb_ref);
1193	INIT_LIST_HEAD(&lkb->lkb_ownqueue);
1194	INIT_LIST_HEAD(&lkb->lkb_rsb_lookup);
1195	INIT_LIST_HEAD(&lkb->lkb_time_list);
1196	INIT_LIST_HEAD(&lkb->lkb_cb_list);
1197	mutex_init(&lkb->lkb_cb_mutex);
 
1198	INIT_WORK(&lkb->lkb_cb_work, dlm_callback_work);
1199
1200	idr_preload(GFP_NOFS);
1201	spin_lock(&ls->ls_lkbidr_spin);
1202	rv = idr_alloc(&ls->ls_lkbidr, lkb, 1, 0, GFP_NOWAIT);
1203	if (rv >= 0)
1204		lkb->lkb_id = rv;
1205	spin_unlock(&ls->ls_lkbidr_spin);
1206	idr_preload_end();
1207
1208	if (rv < 0) {
1209		log_error(ls, "create_lkb idr error %d", rv);
1210		dlm_free_lkb(lkb);
1211		return rv;
1212	}
1213
1214	*lkb_ret = lkb;
1215	return 0;
1216}
1217
 
 
 
 
 
1218static int find_lkb(struct dlm_ls *ls, uint32_t lkid, struct dlm_lkb **lkb_ret)
1219{
1220	struct dlm_lkb *lkb;
1221
1222	spin_lock(&ls->ls_lkbidr_spin);
1223	lkb = idr_find(&ls->ls_lkbidr, lkid);
1224	if (lkb)
1225		kref_get(&lkb->lkb_ref);
1226	spin_unlock(&ls->ls_lkbidr_spin);
1227
1228	*lkb_ret = lkb;
1229	return lkb ? 0 : -ENOENT;
1230}
1231
1232static void kill_lkb(struct kref *kref)
1233{
1234	struct dlm_lkb *lkb = container_of(kref, struct dlm_lkb, lkb_ref);
1235
1236	/* All work is done after the return from kref_put() so we
1237	   can release the write_lock before the detach_lkb */
1238
1239	DLM_ASSERT(!lkb->lkb_status, dlm_print_lkb(lkb););
1240}
1241
1242/* __put_lkb() is used when an lkb may not have an rsb attached to
1243   it so we need to provide the lockspace explicitly */
1244
1245static int __put_lkb(struct dlm_ls *ls, struct dlm_lkb *lkb)
1246{
1247	uint32_t lkid = lkb->lkb_id;
 
1248
1249	spin_lock(&ls->ls_lkbidr_spin);
1250	if (kref_put(&lkb->lkb_ref, kill_lkb)) {
 
1251		idr_remove(&ls->ls_lkbidr, lkid);
1252		spin_unlock(&ls->ls_lkbidr_spin);
1253
1254		detach_lkb(lkb);
1255
1256		/* for local/process lkbs, lvbptr points to caller's lksb */
1257		if (lkb->lkb_lvbptr && is_master_copy(lkb))
1258			dlm_free_lvb(lkb->lkb_lvbptr);
1259		dlm_free_lkb(lkb);
1260		return 1;
1261	} else {
1262		spin_unlock(&ls->ls_lkbidr_spin);
1263		return 0;
1264	}
 
 
1265}
1266
1267int dlm_put_lkb(struct dlm_lkb *lkb)
1268{
1269	struct dlm_ls *ls;
1270
1271	DLM_ASSERT(lkb->lkb_resource, dlm_print_lkb(lkb););
1272	DLM_ASSERT(lkb->lkb_resource->res_ls, dlm_print_lkb(lkb););
1273
1274	ls = lkb->lkb_resource->res_ls;
1275	return __put_lkb(ls, lkb);
1276}
1277
1278/* This is only called to add a reference when the code already holds
1279   a valid reference to the lkb, so there's no need for locking. */
1280
1281static inline void hold_lkb(struct dlm_lkb *lkb)
1282{
1283	kref_get(&lkb->lkb_ref);
1284}
1285
 
 
 
 
 
 
 
1286/* This is called when we need to remove a reference and are certain
1287   it's not the last ref.  e.g. del_lkb is always called between a
1288   find_lkb/put_lkb and is always the inverse of a previous add_lkb.
1289   put_lkb would work fine, but would involve unnecessary locking */
1290
1291static inline void unhold_lkb(struct dlm_lkb *lkb)
1292{
1293	int rv;
1294	rv = kref_put(&lkb->lkb_ref, kill_lkb);
1295	DLM_ASSERT(!rv, dlm_print_lkb(lkb););
1296}
1297
1298static void lkb_add_ordered(struct list_head *new, struct list_head *head,
1299			    int mode)
1300{
1301	struct dlm_lkb *lkb = NULL;
1302
1303	list_for_each_entry(lkb, head, lkb_statequeue)
1304		if (lkb->lkb_rqmode < mode)
 
 
1305			break;
 
1306
1307	__list_add(new, lkb->lkb_statequeue.prev, &lkb->lkb_statequeue);
 
1308}
1309
1310/* add/remove lkb to rsb's grant/convert/wait queue */
1311
1312static void add_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb, int status)
1313{
1314	kref_get(&lkb->lkb_ref);
1315
1316	DLM_ASSERT(!lkb->lkb_status, dlm_print_lkb(lkb););
1317
1318	lkb->lkb_timestamp = ktime_get();
1319
1320	lkb->lkb_status = status;
1321
1322	switch (status) {
1323	case DLM_LKSTS_WAITING:
1324		if (lkb->lkb_exflags & DLM_LKF_HEADQUE)
1325			list_add(&lkb->lkb_statequeue, &r->res_waitqueue);
1326		else
1327			list_add_tail(&lkb->lkb_statequeue, &r->res_waitqueue);
1328		break;
1329	case DLM_LKSTS_GRANTED:
1330		/* convention says granted locks kept in order of grmode */
1331		lkb_add_ordered(&lkb->lkb_statequeue, &r->res_grantqueue,
1332				lkb->lkb_grmode);
1333		break;
1334	case DLM_LKSTS_CONVERT:
1335		if (lkb->lkb_exflags & DLM_LKF_HEADQUE)
1336			list_add(&lkb->lkb_statequeue, &r->res_convertqueue);
1337		else
1338			list_add_tail(&lkb->lkb_statequeue,
1339				      &r->res_convertqueue);
1340		break;
1341	default:
1342		DLM_ASSERT(0, dlm_print_lkb(lkb); printk("sts=%d\n", status););
1343	}
1344}
1345
1346static void del_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb)
1347{
1348	lkb->lkb_status = 0;
1349	list_del(&lkb->lkb_statequeue);
1350	unhold_lkb(lkb);
1351}
1352
1353static void move_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb, int sts)
1354{
1355	hold_lkb(lkb);
1356	del_lkb(r, lkb);
1357	add_lkb(r, lkb, sts);
1358	unhold_lkb(lkb);
1359}
1360
1361static int msg_reply_type(int mstype)
1362{
1363	switch (mstype) {
1364	case DLM_MSG_REQUEST:
1365		return DLM_MSG_REQUEST_REPLY;
1366	case DLM_MSG_CONVERT:
1367		return DLM_MSG_CONVERT_REPLY;
1368	case DLM_MSG_UNLOCK:
1369		return DLM_MSG_UNLOCK_REPLY;
1370	case DLM_MSG_CANCEL:
1371		return DLM_MSG_CANCEL_REPLY;
1372	case DLM_MSG_LOOKUP:
1373		return DLM_MSG_LOOKUP_REPLY;
1374	}
1375	return -1;
1376}
1377
1378static int nodeid_warned(int nodeid, int num_nodes, int *warned)
1379{
1380	int i;
1381
1382	for (i = 0; i < num_nodes; i++) {
1383		if (!warned[i]) {
1384			warned[i] = nodeid;
1385			return 0;
1386		}
1387		if (warned[i] == nodeid)
1388			return 1;
1389	}
1390	return 0;
1391}
1392
1393void dlm_scan_waiters(struct dlm_ls *ls)
1394{
1395	struct dlm_lkb *lkb;
1396	s64 us;
1397	s64 debug_maxus = 0;
1398	u32 debug_scanned = 0;
1399	u32 debug_expired = 0;
1400	int num_nodes = 0;
1401	int *warned = NULL;
1402
1403	if (!dlm_config.ci_waitwarn_us)
1404		return;
1405
1406	mutex_lock(&ls->ls_waiters_mutex);
1407
1408	list_for_each_entry(lkb, &ls->ls_waiters, lkb_wait_reply) {
1409		if (!lkb->lkb_wait_time)
1410			continue;
1411
1412		debug_scanned++;
1413
1414		us = ktime_to_us(ktime_sub(ktime_get(), lkb->lkb_wait_time));
1415
1416		if (us < dlm_config.ci_waitwarn_us)
1417			continue;
1418
1419		lkb->lkb_wait_time = 0;
1420
1421		debug_expired++;
1422		if (us > debug_maxus)
1423			debug_maxus = us;
1424
1425		if (!num_nodes) {
1426			num_nodes = ls->ls_num_nodes;
1427			warned = kcalloc(num_nodes, sizeof(int), GFP_KERNEL);
1428		}
1429		if (!warned)
1430			continue;
1431		if (nodeid_warned(lkb->lkb_wait_nodeid, num_nodes, warned))
1432			continue;
1433
1434		log_error(ls, "waitwarn %x %lld %d us check connection to "
1435			  "node %d", lkb->lkb_id, (long long)us,
1436			  dlm_config.ci_waitwarn_us, lkb->lkb_wait_nodeid);
1437	}
1438	mutex_unlock(&ls->ls_waiters_mutex);
1439	kfree(warned);
1440
1441	if (debug_expired)
1442		log_debug(ls, "scan_waiters %u warn %u over %d us max %lld us",
1443			  debug_scanned, debug_expired,
1444			  dlm_config.ci_waitwarn_us, (long long)debug_maxus);
1445}
1446
1447/* add/remove lkb from global waiters list of lkb's waiting for
1448   a reply from a remote node */
1449
1450static int add_to_waiters(struct dlm_lkb *lkb, int mstype, int to_nodeid)
1451{
1452	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1453	int error = 0;
 
1454
1455	mutex_lock(&ls->ls_waiters_mutex);
1456
1457	if (is_overlap_unlock(lkb) ||
1458	    (is_overlap_cancel(lkb) && (mstype == DLM_MSG_CANCEL))) {
1459		error = -EINVAL;
1460		goto out;
1461	}
1462
1463	if (lkb->lkb_wait_type || is_overlap_cancel(lkb)) {
1464		switch (mstype) {
1465		case DLM_MSG_UNLOCK:
1466			lkb->lkb_flags |= DLM_IFL_OVERLAP_UNLOCK;
1467			break;
1468		case DLM_MSG_CANCEL:
1469			lkb->lkb_flags |= DLM_IFL_OVERLAP_CANCEL;
1470			break;
1471		default:
1472			error = -EBUSY;
1473			goto out;
1474		}
1475		lkb->lkb_wait_count++;
1476		hold_lkb(lkb);
1477
1478		log_debug(ls, "addwait %x cur %d overlap %d count %d f %x",
1479			  lkb->lkb_id, lkb->lkb_wait_type, mstype,
1480			  lkb->lkb_wait_count, lkb->lkb_flags);
1481		goto out;
1482	}
1483
1484	DLM_ASSERT(!lkb->lkb_wait_count,
1485		   dlm_print_lkb(lkb);
1486		   printk("wait_count %d\n", lkb->lkb_wait_count););
1487
1488	lkb->lkb_wait_count++;
1489	lkb->lkb_wait_type = mstype;
1490	lkb->lkb_wait_time = ktime_get();
1491	lkb->lkb_wait_nodeid = to_nodeid; /* for debugging */
1492	hold_lkb(lkb);
1493	list_add(&lkb->lkb_wait_reply, &ls->ls_waiters);
1494 out:
1495	if (error)
1496		log_error(ls, "addwait error %x %d flags %x %d %d %s",
1497			  lkb->lkb_id, error, lkb->lkb_flags, mstype,
1498			  lkb->lkb_wait_type, lkb->lkb_resource->res_name);
1499	mutex_unlock(&ls->ls_waiters_mutex);
1500	return error;
1501}
1502
1503/* We clear the RESEND flag because we might be taking an lkb off the waiters
1504   list as part of process_requestqueue (e.g. a lookup that has an optimized
1505   request reply on the requestqueue) between dlm_recover_waiters_pre() which
1506   set RESEND and dlm_recover_waiters_post() */
1507
1508static int _remove_from_waiters(struct dlm_lkb *lkb, int mstype,
1509				struct dlm_message *ms)
1510{
1511	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1512	int overlap_done = 0;
1513
1514	if (is_overlap_unlock(lkb) && (mstype == DLM_MSG_UNLOCK_REPLY)) {
 
1515		log_debug(ls, "remwait %x unlock_reply overlap", lkb->lkb_id);
1516		lkb->lkb_flags &= ~DLM_IFL_OVERLAP_UNLOCK;
1517		overlap_done = 1;
1518		goto out_del;
1519	}
1520
1521	if (is_overlap_cancel(lkb) && (mstype == DLM_MSG_CANCEL_REPLY)) {
 
1522		log_debug(ls, "remwait %x cancel_reply overlap", lkb->lkb_id);
1523		lkb->lkb_flags &= ~DLM_IFL_OVERLAP_CANCEL;
1524		overlap_done = 1;
1525		goto out_del;
1526	}
1527
1528	/* Cancel state was preemptively cleared by a successful convert,
1529	   see next comment, nothing to do. */
1530
1531	if ((mstype == DLM_MSG_CANCEL_REPLY) &&
1532	    (lkb->lkb_wait_type != DLM_MSG_CANCEL)) {
1533		log_debug(ls, "remwait %x cancel_reply wait_type %d",
1534			  lkb->lkb_id, lkb->lkb_wait_type);
1535		return -1;
1536	}
1537
1538	/* Remove for the convert reply, and premptively remove for the
1539	   cancel reply.  A convert has been granted while there's still
1540	   an outstanding cancel on it (the cancel is moot and the result
1541	   in the cancel reply should be 0).  We preempt the cancel reply
1542	   because the app gets the convert result and then can follow up
1543	   with another op, like convert.  This subsequent op would see the
1544	   lingering state of the cancel and fail with -EBUSY. */
1545
1546	if ((mstype == DLM_MSG_CONVERT_REPLY) &&
1547	    (lkb->lkb_wait_type == DLM_MSG_CONVERT) &&
1548	    is_overlap_cancel(lkb) && ms && !ms->m_result) {
1549		log_debug(ls, "remwait %x convert_reply zap overlap_cancel",
1550			  lkb->lkb_id);
1551		lkb->lkb_wait_type = 0;
1552		lkb->lkb_flags &= ~DLM_IFL_OVERLAP_CANCEL;
1553		lkb->lkb_wait_count--;
1554		goto out_del;
1555	}
1556
1557	/* N.B. type of reply may not always correspond to type of original
1558	   msg due to lookup->request optimization, verify others? */
1559
1560	if (lkb->lkb_wait_type) {
1561		lkb->lkb_wait_type = 0;
1562		goto out_del;
1563	}
1564
1565	log_error(ls, "remwait error %x remote %d %x msg %d flags %x no wait",
1566		  lkb->lkb_id, ms ? ms->m_header.h_nodeid : 0, lkb->lkb_remid,
1567		  mstype, lkb->lkb_flags);
1568	return -1;
1569
1570 out_del:
1571	/* the force-unlock/cancel has completed and we haven't recvd a reply
1572	   to the op that was in progress prior to the unlock/cancel; we
1573	   give up on any reply to the earlier op.  FIXME: not sure when/how
1574	   this would happen */
1575
1576	if (overlap_done && lkb->lkb_wait_type) {
1577		log_error(ls, "remwait error %x reply %d wait_type %d overlap",
1578			  lkb->lkb_id, mstype, lkb->lkb_wait_type);
1579		lkb->lkb_wait_count--;
 
1580		lkb->lkb_wait_type = 0;
1581	}
1582
1583	DLM_ASSERT(lkb->lkb_wait_count, dlm_print_lkb(lkb););
1584
1585	lkb->lkb_flags &= ~DLM_IFL_RESEND;
1586	lkb->lkb_wait_count--;
1587	if (!lkb->lkb_wait_count)
1588		list_del_init(&lkb->lkb_wait_reply);
1589	unhold_lkb(lkb);
1590	return 0;
1591}
1592
1593static int remove_from_waiters(struct dlm_lkb *lkb, int mstype)
1594{
1595	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1596	int error;
1597
1598	mutex_lock(&ls->ls_waiters_mutex);
1599	error = _remove_from_waiters(lkb, mstype, NULL);
1600	mutex_unlock(&ls->ls_waiters_mutex);
1601	return error;
1602}
1603
1604/* Handles situations where we might be processing a "fake" or "stub" reply in
1605   which we can't try to take waiters_mutex again. */
1606
1607static int remove_from_waiters_ms(struct dlm_lkb *lkb, struct dlm_message *ms)
 
1608{
1609	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1610	int error;
1611
1612	if (ms->m_flags != DLM_IFL_STUB_MS)
1613		mutex_lock(&ls->ls_waiters_mutex);
1614	error = _remove_from_waiters(lkb, ms->m_type, ms);
1615	if (ms->m_flags != DLM_IFL_STUB_MS)
1616		mutex_unlock(&ls->ls_waiters_mutex);
1617	return error;
1618}
1619
1620/* If there's an rsb for the same resource being removed, ensure
1621   that the remove message is sent before the new lookup message.
1622   It should be rare to need a delay here, but if not, then it may
1623   be worthwhile to add a proper wait mechanism rather than a delay. */
1624
1625static void wait_pending_remove(struct dlm_rsb *r)
1626{
1627	struct dlm_ls *ls = r->res_ls;
1628 restart:
1629	spin_lock(&ls->ls_remove_spin);
1630	if (ls->ls_remove_len &&
1631	    !rsb_cmp(r, ls->ls_remove_name, ls->ls_remove_len)) {
1632		log_debug(ls, "delay lookup for remove dir %d %s",
1633		  	  r->res_dir_nodeid, r->res_name);
1634		spin_unlock(&ls->ls_remove_spin);
1635		msleep(1);
1636		goto restart;
1637	}
1638	spin_unlock(&ls->ls_remove_spin);
1639}
1640
1641/*
1642 * ls_remove_spin protects ls_remove_name and ls_remove_len which are
1643 * read by other threads in wait_pending_remove.  ls_remove_names
1644 * and ls_remove_lens are only used by the scan thread, so they do
1645 * not need protection.
1646 */
1647
1648static void shrink_bucket(struct dlm_ls *ls, int b)
1649{
1650	struct rb_node *n, *next;
1651	struct dlm_rsb *r;
1652	char *name;
1653	int our_nodeid = dlm_our_nodeid();
1654	int remote_count = 0;
1655	int need_shrink = 0;
1656	int i, len, rv;
1657
1658	memset(&ls->ls_remove_lens, 0, sizeof(int) * DLM_REMOVE_NAMES_MAX);
1659
1660	spin_lock(&ls->ls_rsbtbl[b].lock);
1661
1662	if (!(ls->ls_rsbtbl[b].flags & DLM_RTF_SHRINK)) {
1663		spin_unlock(&ls->ls_rsbtbl[b].lock);
1664		return;
1665	}
1666
1667	for (n = rb_first(&ls->ls_rsbtbl[b].toss); n; n = next) {
1668		next = rb_next(n);
1669		r = rb_entry(n, struct dlm_rsb, res_hashnode);
1670
1671		/* If we're the directory record for this rsb, and
1672		   we're not the master of it, then we need to wait
1673		   for the master node to send us a dir remove for
1674		   before removing the dir record. */
1675
1676		if (!dlm_no_directory(ls) &&
1677		    (r->res_master_nodeid != our_nodeid) &&
1678		    (dlm_dir_nodeid(r) == our_nodeid)) {
1679			continue;
1680		}
1681
1682		need_shrink = 1;
1683
1684		if (!time_after_eq(jiffies, r->res_toss_time +
1685				   dlm_config.ci_toss_secs * HZ)) {
1686			continue;
1687		}
1688
1689		if (!dlm_no_directory(ls) &&
1690		    (r->res_master_nodeid == our_nodeid) &&
1691		    (dlm_dir_nodeid(r) != our_nodeid)) {
1692
1693			/* We're the master of this rsb but we're not
1694			   the directory record, so we need to tell the
1695			   dir node to remove the dir record. */
1696
1697			ls->ls_remove_lens[remote_count] = r->res_length;
1698			memcpy(ls->ls_remove_names[remote_count], r->res_name,
1699			       DLM_RESNAME_MAXLEN);
1700			remote_count++;
1701
1702			if (remote_count >= DLM_REMOVE_NAMES_MAX)
1703				break;
1704			continue;
1705		}
1706
1707		if (!kref_put(&r->res_ref, kill_rsb)) {
1708			log_error(ls, "tossed rsb in use %s", r->res_name);
1709			continue;
1710		}
1711
1712		rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
1713		dlm_free_rsb(r);
1714	}
1715
1716	if (need_shrink)
1717		ls->ls_rsbtbl[b].flags |= DLM_RTF_SHRINK;
1718	else
1719		ls->ls_rsbtbl[b].flags &= ~DLM_RTF_SHRINK;
1720	spin_unlock(&ls->ls_rsbtbl[b].lock);
1721
1722	/*
1723	 * While searching for rsb's to free, we found some that require
1724	 * remote removal.  We leave them in place and find them again here
1725	 * so there is a very small gap between removing them from the toss
1726	 * list and sending the removal.  Keeping this gap small is
1727	 * important to keep us (the master node) from being out of sync
1728	 * with the remote dir node for very long.
1729	 *
1730	 * From the time the rsb is removed from toss until just after
1731	 * send_remove, the rsb name is saved in ls_remove_name.  A new
1732	 * lookup checks this to ensure that a new lookup message for the
1733	 * same resource name is not sent just before the remove message.
1734	 */
1735
1736	for (i = 0; i < remote_count; i++) {
1737		name = ls->ls_remove_names[i];
1738		len = ls->ls_remove_lens[i];
1739
1740		spin_lock(&ls->ls_rsbtbl[b].lock);
1741		rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
1742		if (rv) {
1743			spin_unlock(&ls->ls_rsbtbl[b].lock);
1744			log_debug(ls, "remove_name not toss %s", name);
1745			continue;
1746		}
1747
1748		if (r->res_master_nodeid != our_nodeid) {
1749			spin_unlock(&ls->ls_rsbtbl[b].lock);
1750			log_debug(ls, "remove_name master %d dir %d our %d %s",
1751				  r->res_master_nodeid, r->res_dir_nodeid,
1752				  our_nodeid, name);
1753			continue;
1754		}
1755
1756		if (r->res_dir_nodeid == our_nodeid) {
1757			/* should never happen */
1758			spin_unlock(&ls->ls_rsbtbl[b].lock);
1759			log_error(ls, "remove_name dir %d master %d our %d %s",
1760				  r->res_dir_nodeid, r->res_master_nodeid,
1761				  our_nodeid, name);
1762			continue;
1763		}
1764
1765		if (!time_after_eq(jiffies, r->res_toss_time +
1766				   dlm_config.ci_toss_secs * HZ)) {
1767			spin_unlock(&ls->ls_rsbtbl[b].lock);
1768			log_debug(ls, "remove_name toss_time %lu now %lu %s",
1769				  r->res_toss_time, jiffies, name);
1770			continue;
1771		}
1772
1773		if (!kref_put(&r->res_ref, kill_rsb)) {
1774			spin_unlock(&ls->ls_rsbtbl[b].lock);
1775			log_error(ls, "remove_name in use %s", name);
1776			continue;
1777		}
1778
1779		rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
1780
1781		/* block lookup of same name until we've sent remove */
1782		spin_lock(&ls->ls_remove_spin);
1783		ls->ls_remove_len = len;
1784		memcpy(ls->ls_remove_name, name, DLM_RESNAME_MAXLEN);
1785		spin_unlock(&ls->ls_remove_spin);
1786		spin_unlock(&ls->ls_rsbtbl[b].lock);
1787
1788		send_remove(r);
1789
1790		/* allow lookup of name again */
1791		spin_lock(&ls->ls_remove_spin);
1792		ls->ls_remove_len = 0;
1793		memset(ls->ls_remove_name, 0, DLM_RESNAME_MAXLEN);
1794		spin_unlock(&ls->ls_remove_spin);
1795
1796		dlm_free_rsb(r);
1797	}
1798}
1799
1800void dlm_scan_rsbs(struct dlm_ls *ls)
1801{
1802	int i;
1803
1804	for (i = 0; i < ls->ls_rsbtbl_size; i++) {
1805		shrink_bucket(ls, i);
1806		if (dlm_locking_stopped(ls))
1807			break;
1808		cond_resched();
1809	}
1810}
1811
1812static void add_timeout(struct dlm_lkb *lkb)
1813{
1814	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1815
1816	if (is_master_copy(lkb))
1817		return;
1818
1819	if (test_bit(LSFL_TIMEWARN, &ls->ls_flags) &&
1820	    !(lkb->lkb_exflags & DLM_LKF_NODLCKWT)) {
1821		lkb->lkb_flags |= DLM_IFL_WATCH_TIMEWARN;
1822		goto add_it;
1823	}
1824	if (lkb->lkb_exflags & DLM_LKF_TIMEOUT)
1825		goto add_it;
1826	return;
1827
1828 add_it:
1829	DLM_ASSERT(list_empty(&lkb->lkb_time_list), dlm_print_lkb(lkb););
1830	mutex_lock(&ls->ls_timeout_mutex);
1831	hold_lkb(lkb);
1832	list_add_tail(&lkb->lkb_time_list, &ls->ls_timeout);
1833	mutex_unlock(&ls->ls_timeout_mutex);
1834}
1835
1836static void del_timeout(struct dlm_lkb *lkb)
1837{
1838	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1839
1840	mutex_lock(&ls->ls_timeout_mutex);
1841	if (!list_empty(&lkb->lkb_time_list)) {
1842		list_del_init(&lkb->lkb_time_list);
1843		unhold_lkb(lkb);
1844	}
1845	mutex_unlock(&ls->ls_timeout_mutex);
1846}
1847
1848/* FIXME: is it safe to look at lkb_exflags, lkb_flags, lkb_timestamp, and
1849   lkb_lksb_timeout without lock_rsb?  Note: we can't lock timeout_mutex
1850   and then lock rsb because of lock ordering in add_timeout.  We may need
1851   to specify some special timeout-related bits in the lkb that are just to
1852   be accessed under the timeout_mutex. */
1853
1854void dlm_scan_timeout(struct dlm_ls *ls)
1855{
1856	struct dlm_rsb *r;
1857	struct dlm_lkb *lkb;
1858	int do_cancel, do_warn;
1859	s64 wait_us;
1860
1861	for (;;) {
1862		if (dlm_locking_stopped(ls))
1863			break;
1864
1865		do_cancel = 0;
1866		do_warn = 0;
1867		mutex_lock(&ls->ls_timeout_mutex);
1868		list_for_each_entry(lkb, &ls->ls_timeout, lkb_time_list) {
1869
1870			wait_us = ktime_to_us(ktime_sub(ktime_get(),
1871					      		lkb->lkb_timestamp));
1872
1873			if ((lkb->lkb_exflags & DLM_LKF_TIMEOUT) &&
1874			    wait_us >= (lkb->lkb_timeout_cs * 10000))
1875				do_cancel = 1;
1876
1877			if ((lkb->lkb_flags & DLM_IFL_WATCH_TIMEWARN) &&
1878			    wait_us >= dlm_config.ci_timewarn_cs * 10000)
1879				do_warn = 1;
1880
1881			if (!do_cancel && !do_warn)
1882				continue;
1883			hold_lkb(lkb);
1884			break;
1885		}
1886		mutex_unlock(&ls->ls_timeout_mutex);
1887
1888		if (!do_cancel && !do_warn)
1889			break;
1890
1891		r = lkb->lkb_resource;
1892		hold_rsb(r);
1893		lock_rsb(r);
1894
1895		if (do_warn) {
1896			/* clear flag so we only warn once */
1897			lkb->lkb_flags &= ~DLM_IFL_WATCH_TIMEWARN;
1898			if (!(lkb->lkb_exflags & DLM_LKF_TIMEOUT))
1899				del_timeout(lkb);
1900			dlm_timeout_warn(lkb);
1901		}
1902
1903		if (do_cancel) {
1904			log_debug(ls, "timeout cancel %x node %d %s",
1905				  lkb->lkb_id, lkb->lkb_nodeid, r->res_name);
1906			lkb->lkb_flags &= ~DLM_IFL_WATCH_TIMEWARN;
1907			lkb->lkb_flags |= DLM_IFL_TIMEOUT_CANCEL;
1908			del_timeout(lkb);
1909			_cancel_lock(r, lkb);
1910		}
1911
1912		unlock_rsb(r);
1913		unhold_rsb(r);
1914		dlm_put_lkb(lkb);
1915	}
1916}
1917
1918/* This is only called by dlm_recoverd, and we rely on dlm_ls_stop() stopping
1919   dlm_recoverd before checking/setting ls_recover_begin. */
1920
1921void dlm_adjust_timeouts(struct dlm_ls *ls)
1922{
1923	struct dlm_lkb *lkb;
1924	u64 adj_us = jiffies_to_usecs(jiffies - ls->ls_recover_begin);
1925
1926	ls->ls_recover_begin = 0;
1927	mutex_lock(&ls->ls_timeout_mutex);
1928	list_for_each_entry(lkb, &ls->ls_timeout, lkb_time_list)
1929		lkb->lkb_timestamp = ktime_add_us(lkb->lkb_timestamp, adj_us);
1930	mutex_unlock(&ls->ls_timeout_mutex);
1931
1932	if (!dlm_config.ci_waitwarn_us)
1933		return;
1934
1935	mutex_lock(&ls->ls_waiters_mutex);
1936	list_for_each_entry(lkb, &ls->ls_waiters, lkb_wait_reply) {
1937		if (ktime_to_us(lkb->lkb_wait_time))
1938			lkb->lkb_wait_time = ktime_get();
1939	}
1940	mutex_unlock(&ls->ls_waiters_mutex);
1941}
1942
1943/* lkb is master or local copy */
1944
1945static void set_lvb_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1946{
1947	int b, len = r->res_ls->ls_lvblen;
1948
1949	/* b=1 lvb returned to caller
1950	   b=0 lvb written to rsb or invalidated
1951	   b=-1 do nothing */
1952
1953	b =  dlm_lvb_operations[lkb->lkb_grmode + 1][lkb->lkb_rqmode + 1];
1954
1955	if (b == 1) {
1956		if (!lkb->lkb_lvbptr)
1957			return;
1958
1959		if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
1960			return;
1961
1962		if (!r->res_lvbptr)
1963			return;
1964
1965		memcpy(lkb->lkb_lvbptr, r->res_lvbptr, len);
1966		lkb->lkb_lvbseq = r->res_lvbseq;
1967
1968	} else if (b == 0) {
1969		if (lkb->lkb_exflags & DLM_LKF_IVVALBLK) {
1970			rsb_set_flag(r, RSB_VALNOTVALID);
1971			return;
1972		}
1973
1974		if (!lkb->lkb_lvbptr)
1975			return;
1976
1977		if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
1978			return;
1979
1980		if (!r->res_lvbptr)
1981			r->res_lvbptr = dlm_allocate_lvb(r->res_ls);
1982
1983		if (!r->res_lvbptr)
1984			return;
1985
1986		memcpy(r->res_lvbptr, lkb->lkb_lvbptr, len);
1987		r->res_lvbseq++;
1988		lkb->lkb_lvbseq = r->res_lvbseq;
1989		rsb_clear_flag(r, RSB_VALNOTVALID);
1990	}
1991
1992	if (rsb_flag(r, RSB_VALNOTVALID))
1993		lkb->lkb_sbflags |= DLM_SBF_VALNOTVALID;
1994}
1995
1996static void set_lvb_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1997{
1998	if (lkb->lkb_grmode < DLM_LOCK_PW)
1999		return;
2000
2001	if (lkb->lkb_exflags & DLM_LKF_IVVALBLK) {
2002		rsb_set_flag(r, RSB_VALNOTVALID);
2003		return;
2004	}
2005
2006	if (!lkb->lkb_lvbptr)
2007		return;
2008
2009	if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
2010		return;
2011
2012	if (!r->res_lvbptr)
2013		r->res_lvbptr = dlm_allocate_lvb(r->res_ls);
2014
2015	if (!r->res_lvbptr)
2016		return;
2017
2018	memcpy(r->res_lvbptr, lkb->lkb_lvbptr, r->res_ls->ls_lvblen);
2019	r->res_lvbseq++;
2020	rsb_clear_flag(r, RSB_VALNOTVALID);
2021}
2022
2023/* lkb is process copy (pc) */
2024
2025static void set_lvb_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb,
2026			    struct dlm_message *ms)
2027{
2028	int b;
2029
2030	if (!lkb->lkb_lvbptr)
2031		return;
2032
2033	if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
2034		return;
2035
2036	b = dlm_lvb_operations[lkb->lkb_grmode + 1][lkb->lkb_rqmode + 1];
2037	if (b == 1) {
2038		int len = receive_extralen(ms);
2039		if (len > r->res_ls->ls_lvblen)
2040			len = r->res_ls->ls_lvblen;
2041		memcpy(lkb->lkb_lvbptr, ms->m_extra, len);
2042		lkb->lkb_lvbseq = ms->m_lvbseq;
2043	}
2044}
2045
2046/* Manipulate lkb's on rsb's convert/granted/waiting queues
2047   remove_lock -- used for unlock, removes lkb from granted
2048   revert_lock -- used for cancel, moves lkb from convert to granted
2049   grant_lock  -- used for request and convert, adds lkb to granted or
2050                  moves lkb from convert or waiting to granted
2051
2052   Each of these is used for master or local copy lkb's.  There is
2053   also a _pc() variation used to make the corresponding change on
2054   a process copy (pc) lkb. */
2055
2056static void _remove_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
2057{
2058	del_lkb(r, lkb);
2059	lkb->lkb_grmode = DLM_LOCK_IV;
2060	/* this unhold undoes the original ref from create_lkb()
2061	   so this leads to the lkb being freed */
2062	unhold_lkb(lkb);
2063}
2064
2065static void remove_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
2066{
2067	set_lvb_unlock(r, lkb);
2068	_remove_lock(r, lkb);
2069}
2070
2071static void remove_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb)
2072{
2073	_remove_lock(r, lkb);
2074}
2075
2076/* returns: 0 did nothing
2077	    1 moved lock to granted
2078	   -1 removed lock */
2079
2080static int revert_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
2081{
2082	int rv = 0;
2083
2084	lkb->lkb_rqmode = DLM_LOCK_IV;
2085
2086	switch (lkb->lkb_status) {
2087	case DLM_LKSTS_GRANTED:
2088		break;
2089	case DLM_LKSTS_CONVERT:
2090		move_lkb(r, lkb, DLM_LKSTS_GRANTED);
2091		rv = 1;
2092		break;
2093	case DLM_LKSTS_WAITING:
2094		del_lkb(r, lkb);
2095		lkb->lkb_grmode = DLM_LOCK_IV;
2096		/* this unhold undoes the original ref from create_lkb()
2097		   so this leads to the lkb being freed */
2098		unhold_lkb(lkb);
2099		rv = -1;
2100		break;
2101	default:
2102		log_print("invalid status for revert %d", lkb->lkb_status);
2103	}
2104	return rv;
2105}
2106
2107static int revert_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb)
2108{
2109	return revert_lock(r, lkb);
2110}
2111
2112static void _grant_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
2113{
2114	if (lkb->lkb_grmode != lkb->lkb_rqmode) {
2115		lkb->lkb_grmode = lkb->lkb_rqmode;
2116		if (lkb->lkb_status)
2117			move_lkb(r, lkb, DLM_LKSTS_GRANTED);
2118		else
2119			add_lkb(r, lkb, DLM_LKSTS_GRANTED);
2120	}
2121
2122	lkb->lkb_rqmode = DLM_LOCK_IV;
2123	lkb->lkb_highbast = 0;
2124}
2125
2126static void grant_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
2127{
2128	set_lvb_lock(r, lkb);
2129	_grant_lock(r, lkb);
2130}
2131
2132static void grant_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb,
2133			  struct dlm_message *ms)
2134{
2135	set_lvb_lock_pc(r, lkb, ms);
2136	_grant_lock(r, lkb);
2137}
2138
2139/* called by grant_pending_locks() which means an async grant message must
2140   be sent to the requesting node in addition to granting the lock if the
2141   lkb belongs to a remote node. */
2142
2143static void grant_lock_pending(struct dlm_rsb *r, struct dlm_lkb *lkb)
2144{
2145	grant_lock(r, lkb);
2146	if (is_master_copy(lkb))
2147		send_grant(r, lkb);
2148	else
2149		queue_cast(r, lkb, 0);
2150}
2151
2152/* The special CONVDEADLK, ALTPR and ALTCW flags allow the master to
2153   change the granted/requested modes.  We're munging things accordingly in
2154   the process copy.
2155   CONVDEADLK: our grmode may have been forced down to NL to resolve a
2156   conversion deadlock
2157   ALTPR/ALTCW: our rqmode may have been changed to PR or CW to become
2158   compatible with other granted locks */
2159
2160static void munge_demoted(struct dlm_lkb *lkb)
2161{
2162	if (lkb->lkb_rqmode == DLM_LOCK_IV || lkb->lkb_grmode == DLM_LOCK_IV) {
2163		log_print("munge_demoted %x invalid modes gr %d rq %d",
2164			  lkb->lkb_id, lkb->lkb_grmode, lkb->lkb_rqmode);
2165		return;
2166	}
2167
2168	lkb->lkb_grmode = DLM_LOCK_NL;
2169}
2170
2171static void munge_altmode(struct dlm_lkb *lkb, struct dlm_message *ms)
2172{
2173	if (ms->m_type != DLM_MSG_REQUEST_REPLY &&
2174	    ms->m_type != DLM_MSG_GRANT) {
2175		log_print("munge_altmode %x invalid reply type %d",
2176			  lkb->lkb_id, ms->m_type);
2177		return;
2178	}
2179
2180	if (lkb->lkb_exflags & DLM_LKF_ALTPR)
2181		lkb->lkb_rqmode = DLM_LOCK_PR;
2182	else if (lkb->lkb_exflags & DLM_LKF_ALTCW)
2183		lkb->lkb_rqmode = DLM_LOCK_CW;
2184	else {
2185		log_print("munge_altmode invalid exflags %x", lkb->lkb_exflags);
2186		dlm_print_lkb(lkb);
2187	}
2188}
2189
2190static inline int first_in_list(struct dlm_lkb *lkb, struct list_head *head)
2191{
2192	struct dlm_lkb *first = list_entry(head->next, struct dlm_lkb,
2193					   lkb_statequeue);
2194	if (lkb->lkb_id == first->lkb_id)
2195		return 1;
2196
2197	return 0;
2198}
2199
2200/* Check if the given lkb conflicts with another lkb on the queue. */
2201
2202static int queue_conflict(struct list_head *head, struct dlm_lkb *lkb)
2203{
2204	struct dlm_lkb *this;
2205
2206	list_for_each_entry(this, head, lkb_statequeue) {
2207		if (this == lkb)
2208			continue;
2209		if (!modes_compat(this, lkb))
2210			return 1;
2211	}
2212	return 0;
2213}
2214
2215/*
2216 * "A conversion deadlock arises with a pair of lock requests in the converting
2217 * queue for one resource.  The granted mode of each lock blocks the requested
2218 * mode of the other lock."
2219 *
2220 * Part 2: if the granted mode of lkb is preventing an earlier lkb in the
2221 * convert queue from being granted, then deadlk/demote lkb.
2222 *
2223 * Example:
2224 * Granted Queue: empty
2225 * Convert Queue: NL->EX (first lock)
2226 *                PR->EX (second lock)
2227 *
2228 * The first lock can't be granted because of the granted mode of the second
2229 * lock and the second lock can't be granted because it's not first in the
2230 * list.  We either cancel lkb's conversion (PR->EX) and return EDEADLK, or we
2231 * demote the granted mode of lkb (from PR to NL) if it has the CONVDEADLK
2232 * flag set and return DEMOTED in the lksb flags.
2233 *
2234 * Originally, this function detected conv-deadlk in a more limited scope:
2235 * - if !modes_compat(lkb1, lkb2) && !modes_compat(lkb2, lkb1), or
2236 * - if lkb1 was the first entry in the queue (not just earlier), and was
2237 *   blocked by the granted mode of lkb2, and there was nothing on the
2238 *   granted queue preventing lkb1 from being granted immediately, i.e.
2239 *   lkb2 was the only thing preventing lkb1 from being granted.
2240 *
2241 * That second condition meant we'd only say there was conv-deadlk if
2242 * resolving it (by demotion) would lead to the first lock on the convert
2243 * queue being granted right away.  It allowed conversion deadlocks to exist
2244 * between locks on the convert queue while they couldn't be granted anyway.
2245 *
2246 * Now, we detect and take action on conversion deadlocks immediately when
2247 * they're created, even if they may not be immediately consequential.  If
2248 * lkb1 exists anywhere in the convert queue and lkb2 comes in with a granted
2249 * mode that would prevent lkb1's conversion from being granted, we do a
2250 * deadlk/demote on lkb2 right away and don't let it onto the convert queue.
2251 * I think this means that the lkb_is_ahead condition below should always
2252 * be zero, i.e. there will never be conv-deadlk between two locks that are
2253 * both already on the convert queue.
2254 */
2255
2256static int conversion_deadlock_detect(struct dlm_rsb *r, struct dlm_lkb *lkb2)
2257{
2258	struct dlm_lkb *lkb1;
2259	int lkb_is_ahead = 0;
2260
2261	list_for_each_entry(lkb1, &r->res_convertqueue, lkb_statequeue) {
2262		if (lkb1 == lkb2) {
2263			lkb_is_ahead = 1;
2264			continue;
2265		}
2266
2267		if (!lkb_is_ahead) {
2268			if (!modes_compat(lkb2, lkb1))
2269				return 1;
2270		} else {
2271			if (!modes_compat(lkb2, lkb1) &&
2272			    !modes_compat(lkb1, lkb2))
2273				return 1;
2274		}
2275	}
2276	return 0;
2277}
2278
2279/*
2280 * Return 1 if the lock can be granted, 0 otherwise.
2281 * Also detect and resolve conversion deadlocks.
2282 *
2283 * lkb is the lock to be granted
2284 *
2285 * now is 1 if the function is being called in the context of the
2286 * immediate request, it is 0 if called later, after the lock has been
2287 * queued.
2288 *
2289 * recover is 1 if dlm_recover_grant() is trying to grant conversions
2290 * after recovery.
2291 *
2292 * References are from chapter 6 of "VAXcluster Principles" by Roy Davis
2293 */
2294
2295static int _can_be_granted(struct dlm_rsb *r, struct dlm_lkb *lkb, int now,
2296			   int recover)
2297{
2298	int8_t conv = (lkb->lkb_grmode != DLM_LOCK_IV);
2299
2300	/*
2301	 * 6-10: Version 5.4 introduced an option to address the phenomenon of
2302	 * a new request for a NL mode lock being blocked.
2303	 *
2304	 * 6-11: If the optional EXPEDITE flag is used with the new NL mode
2305	 * request, then it would be granted.  In essence, the use of this flag
2306	 * tells the Lock Manager to expedite theis request by not considering
2307	 * what may be in the CONVERTING or WAITING queues...  As of this
2308	 * writing, the EXPEDITE flag can be used only with new requests for NL
2309	 * mode locks.  This flag is not valid for conversion requests.
2310	 *
2311	 * A shortcut.  Earlier checks return an error if EXPEDITE is used in a
2312	 * conversion or used with a non-NL requested mode.  We also know an
2313	 * EXPEDITE request is always granted immediately, so now must always
2314	 * be 1.  The full condition to grant an expedite request: (now &&
2315	 * !conv && lkb->rqmode == DLM_LOCK_NL && (flags & EXPEDITE)) can
2316	 * therefore be shortened to just checking the flag.
2317	 */
2318
2319	if (lkb->lkb_exflags & DLM_LKF_EXPEDITE)
2320		return 1;
2321
2322	/*
2323	 * A shortcut. Without this, !queue_conflict(grantqueue, lkb) would be
2324	 * added to the remaining conditions.
2325	 */
2326
2327	if (queue_conflict(&r->res_grantqueue, lkb))
2328		return 0;
2329
2330	/*
2331	 * 6-3: By default, a conversion request is immediately granted if the
2332	 * requested mode is compatible with the modes of all other granted
2333	 * locks
2334	 */
2335
2336	if (queue_conflict(&r->res_convertqueue, lkb))
2337		return 0;
2338
2339	/*
2340	 * The RECOVER_GRANT flag means dlm_recover_grant() is granting
2341	 * locks for a recovered rsb, on which lkb's have been rebuilt.
2342	 * The lkb's may have been rebuilt on the queues in a different
2343	 * order than they were in on the previous master.  So, granting
2344	 * queued conversions in order after recovery doesn't make sense
2345	 * since the order hasn't been preserved anyway.  The new order
2346	 * could also have created a new "in place" conversion deadlock.
2347	 * (e.g. old, failed master held granted EX, with PR->EX, NL->EX.
2348	 * After recovery, there would be no granted locks, and possibly
2349	 * NL->EX, PR->EX, an in-place conversion deadlock.)  So, after
2350	 * recovery, grant conversions without considering order.
2351	 */
2352
2353	if (conv && recover)
2354		return 1;
2355
2356	/*
2357	 * 6-5: But the default algorithm for deciding whether to grant or
2358	 * queue conversion requests does not by itself guarantee that such
2359	 * requests are serviced on a "first come first serve" basis.  This, in
2360	 * turn, can lead to a phenomenon known as "indefinate postponement".
2361	 *
2362	 * 6-7: This issue is dealt with by using the optional QUECVT flag with
2363	 * the system service employed to request a lock conversion.  This flag
2364	 * forces certain conversion requests to be queued, even if they are
2365	 * compatible with the granted modes of other locks on the same
2366	 * resource.  Thus, the use of this flag results in conversion requests
2367	 * being ordered on a "first come first servce" basis.
2368	 *
2369	 * DCT: This condition is all about new conversions being able to occur
2370	 * "in place" while the lock remains on the granted queue (assuming
2371	 * nothing else conflicts.)  IOW if QUECVT isn't set, a conversion
2372	 * doesn't _have_ to go onto the convert queue where it's processed in
2373	 * order.  The "now" variable is necessary to distinguish converts
2374	 * being received and processed for the first time now, because once a
2375	 * convert is moved to the conversion queue the condition below applies
2376	 * requiring fifo granting.
2377	 */
2378
2379	if (now && conv && !(lkb->lkb_exflags & DLM_LKF_QUECVT))
2380		return 1;
2381
2382	/*
2383	 * Even if the convert is compat with all granted locks,
2384	 * QUECVT forces it behind other locks on the convert queue.
2385	 */
2386
2387	if (now && conv && (lkb->lkb_exflags & DLM_LKF_QUECVT)) {
2388		if (list_empty(&r->res_convertqueue))
2389			return 1;
2390		else
2391			return 0;
2392	}
2393
2394	/*
2395	 * The NOORDER flag is set to avoid the standard vms rules on grant
2396	 * order.
2397	 */
2398
2399	if (lkb->lkb_exflags & DLM_LKF_NOORDER)
2400		return 1;
2401
2402	/*
2403	 * 6-3: Once in that queue [CONVERTING], a conversion request cannot be
2404	 * granted until all other conversion requests ahead of it are granted
2405	 * and/or canceled.
2406	 */
2407
2408	if (!now && conv && first_in_list(lkb, &r->res_convertqueue))
2409		return 1;
2410
2411	/*
2412	 * 6-4: By default, a new request is immediately granted only if all
2413	 * three of the following conditions are satisfied when the request is
2414	 * issued:
2415	 * - The queue of ungranted conversion requests for the resource is
2416	 *   empty.
2417	 * - The queue of ungranted new requests for the resource is empty.
2418	 * - The mode of the new request is compatible with the most
2419	 *   restrictive mode of all granted locks on the resource.
2420	 */
2421
2422	if (now && !conv && list_empty(&r->res_convertqueue) &&
2423	    list_empty(&r->res_waitqueue))
2424		return 1;
2425
2426	/*
2427	 * 6-4: Once a lock request is in the queue of ungranted new requests,
2428	 * it cannot be granted until the queue of ungranted conversion
2429	 * requests is empty, all ungranted new requests ahead of it are
2430	 * granted and/or canceled, and it is compatible with the granted mode
2431	 * of the most restrictive lock granted on the resource.
2432	 */
2433
2434	if (!now && !conv && list_empty(&r->res_convertqueue) &&
2435	    first_in_list(lkb, &r->res_waitqueue))
2436		return 1;
2437
2438	return 0;
2439}
2440
2441static int can_be_granted(struct dlm_rsb *r, struct dlm_lkb *lkb, int now,
2442			  int recover, int *err)
2443{
2444	int rv;
2445	int8_t alt = 0, rqmode = lkb->lkb_rqmode;
2446	int8_t is_convert = (lkb->lkb_grmode != DLM_LOCK_IV);
2447
2448	if (err)
2449		*err = 0;
2450
2451	rv = _can_be_granted(r, lkb, now, recover);
2452	if (rv)
2453		goto out;
2454
2455	/*
2456	 * The CONVDEADLK flag is non-standard and tells the dlm to resolve
2457	 * conversion deadlocks by demoting grmode to NL, otherwise the dlm
2458	 * cancels one of the locks.
2459	 */
2460
2461	if (is_convert && can_be_queued(lkb) &&
2462	    conversion_deadlock_detect(r, lkb)) {
2463		if (lkb->lkb_exflags & DLM_LKF_CONVDEADLK) {
2464			lkb->lkb_grmode = DLM_LOCK_NL;
2465			lkb->lkb_sbflags |= DLM_SBF_DEMOTED;
2466		} else if (err) {
2467			*err = -EDEADLK;
2468		} else {
2469			log_print("can_be_granted deadlock %x now %d",
2470				  lkb->lkb_id, now);
2471			dlm_dump_rsb(r);
2472		}
2473		goto out;
2474	}
2475
2476	/*
2477	 * The ALTPR and ALTCW flags are non-standard and tell the dlm to try
2478	 * to grant a request in a mode other than the normal rqmode.  It's a
2479	 * simple way to provide a big optimization to applications that can
2480	 * use them.
2481	 */
2482
2483	if (rqmode != DLM_LOCK_PR && (lkb->lkb_exflags & DLM_LKF_ALTPR))
2484		alt = DLM_LOCK_PR;
2485	else if (rqmode != DLM_LOCK_CW && (lkb->lkb_exflags & DLM_LKF_ALTCW))
2486		alt = DLM_LOCK_CW;
2487
2488	if (alt) {
2489		lkb->lkb_rqmode = alt;
2490		rv = _can_be_granted(r, lkb, now, 0);
2491		if (rv)
2492			lkb->lkb_sbflags |= DLM_SBF_ALTMODE;
2493		else
2494			lkb->lkb_rqmode = rqmode;
2495	}
2496 out:
2497	return rv;
2498}
2499
2500/* Returns the highest requested mode of all blocked conversions; sets
2501   cw if there's a blocked conversion to DLM_LOCK_CW. */
2502
2503static int grant_pending_convert(struct dlm_rsb *r, int high, int *cw,
2504				 unsigned int *count)
2505{
2506	struct dlm_lkb *lkb, *s;
2507	int recover = rsb_flag(r, RSB_RECOVER_GRANT);
2508	int hi, demoted, quit, grant_restart, demote_restart;
2509	int deadlk;
2510
2511	quit = 0;
2512 restart:
2513	grant_restart = 0;
2514	demote_restart = 0;
2515	hi = DLM_LOCK_IV;
2516
2517	list_for_each_entry_safe(lkb, s, &r->res_convertqueue, lkb_statequeue) {
2518		demoted = is_demoted(lkb);
2519		deadlk = 0;
2520
2521		if (can_be_granted(r, lkb, 0, recover, &deadlk)) {
2522			grant_lock_pending(r, lkb);
2523			grant_restart = 1;
2524			if (count)
2525				(*count)++;
2526			continue;
2527		}
2528
2529		if (!demoted && is_demoted(lkb)) {
2530			log_print("WARN: pending demoted %x node %d %s",
2531				  lkb->lkb_id, lkb->lkb_nodeid, r->res_name);
2532			demote_restart = 1;
2533			continue;
2534		}
2535
2536		if (deadlk) {
2537			/*
2538			 * If DLM_LKB_NODLKWT flag is set and conversion
2539			 * deadlock is detected, we request blocking AST and
2540			 * down (or cancel) conversion.
2541			 */
2542			if (lkb->lkb_exflags & DLM_LKF_NODLCKWT) {
2543				if (lkb->lkb_highbast < lkb->lkb_rqmode) {
2544					queue_bast(r, lkb, lkb->lkb_rqmode);
2545					lkb->lkb_highbast = lkb->lkb_rqmode;
2546				}
2547			} else {
2548				log_print("WARN: pending deadlock %x node %d %s",
2549					  lkb->lkb_id, lkb->lkb_nodeid,
2550					  r->res_name);
2551				dlm_dump_rsb(r);
2552			}
2553			continue;
2554		}
2555
2556		hi = max_t(int, lkb->lkb_rqmode, hi);
2557
2558		if (cw && lkb->lkb_rqmode == DLM_LOCK_CW)
2559			*cw = 1;
2560	}
2561
2562	if (grant_restart)
2563		goto restart;
2564	if (demote_restart && !quit) {
2565		quit = 1;
2566		goto restart;
2567	}
2568
2569	return max_t(int, high, hi);
2570}
2571
2572static int grant_pending_wait(struct dlm_rsb *r, int high, int *cw,
2573			      unsigned int *count)
2574{
2575	struct dlm_lkb *lkb, *s;
2576
2577	list_for_each_entry_safe(lkb, s, &r->res_waitqueue, lkb_statequeue) {
2578		if (can_be_granted(r, lkb, 0, 0, NULL)) {
2579			grant_lock_pending(r, lkb);
2580			if (count)
2581				(*count)++;
2582		} else {
2583			high = max_t(int, lkb->lkb_rqmode, high);
2584			if (lkb->lkb_rqmode == DLM_LOCK_CW)
2585				*cw = 1;
2586		}
2587	}
2588
2589	return high;
2590}
2591
2592/* cw of 1 means there's a lock with a rqmode of DLM_LOCK_CW that's blocked
2593   on either the convert or waiting queue.
2594   high is the largest rqmode of all locks blocked on the convert or
2595   waiting queue. */
2596
2597static int lock_requires_bast(struct dlm_lkb *gr, int high, int cw)
2598{
2599	if (gr->lkb_grmode == DLM_LOCK_PR && cw) {
2600		if (gr->lkb_highbast < DLM_LOCK_EX)
2601			return 1;
2602		return 0;
2603	}
2604
2605	if (gr->lkb_highbast < high &&
2606	    !__dlm_compat_matrix[gr->lkb_grmode+1][high+1])
2607		return 1;
2608	return 0;
2609}
2610
2611static void grant_pending_locks(struct dlm_rsb *r, unsigned int *count)
2612{
2613	struct dlm_lkb *lkb, *s;
2614	int high = DLM_LOCK_IV;
2615	int cw = 0;
2616
2617	if (!is_master(r)) {
2618		log_print("grant_pending_locks r nodeid %d", r->res_nodeid);
2619		dlm_dump_rsb(r);
2620		return;
2621	}
2622
2623	high = grant_pending_convert(r, high, &cw, count);
2624	high = grant_pending_wait(r, high, &cw, count);
2625
2626	if (high == DLM_LOCK_IV)
2627		return;
2628
2629	/*
2630	 * If there are locks left on the wait/convert queue then send blocking
2631	 * ASTs to granted locks based on the largest requested mode (high)
2632	 * found above.
2633	 */
2634
2635	list_for_each_entry_safe(lkb, s, &r->res_grantqueue, lkb_statequeue) {
2636		if (lkb->lkb_bastfn && lock_requires_bast(lkb, high, cw)) {
2637			if (cw && high == DLM_LOCK_PR &&
2638			    lkb->lkb_grmode == DLM_LOCK_PR)
2639				queue_bast(r, lkb, DLM_LOCK_CW);
2640			else
2641				queue_bast(r, lkb, high);
2642			lkb->lkb_highbast = high;
2643		}
2644	}
2645}
2646
2647static int modes_require_bast(struct dlm_lkb *gr, struct dlm_lkb *rq)
2648{
2649	if ((gr->lkb_grmode == DLM_LOCK_PR && rq->lkb_rqmode == DLM_LOCK_CW) ||
2650	    (gr->lkb_grmode == DLM_LOCK_CW && rq->lkb_rqmode == DLM_LOCK_PR)) {
2651		if (gr->lkb_highbast < DLM_LOCK_EX)
2652			return 1;
2653		return 0;
2654	}
2655
2656	if (gr->lkb_highbast < rq->lkb_rqmode && !modes_compat(gr, rq))
2657		return 1;
2658	return 0;
2659}
2660
2661static void send_bast_queue(struct dlm_rsb *r, struct list_head *head,
2662			    struct dlm_lkb *lkb)
2663{
2664	struct dlm_lkb *gr;
2665
2666	list_for_each_entry(gr, head, lkb_statequeue) {
2667		/* skip self when sending basts to convertqueue */
2668		if (gr == lkb)
2669			continue;
2670		if (gr->lkb_bastfn && modes_require_bast(gr, lkb)) {
2671			queue_bast(r, gr, lkb->lkb_rqmode);
2672			gr->lkb_highbast = lkb->lkb_rqmode;
2673		}
2674	}
2675}
2676
2677static void send_blocking_asts(struct dlm_rsb *r, struct dlm_lkb *lkb)
2678{
2679	send_bast_queue(r, &r->res_grantqueue, lkb);
2680}
2681
2682static void send_blocking_asts_all(struct dlm_rsb *r, struct dlm_lkb *lkb)
2683{
2684	send_bast_queue(r, &r->res_grantqueue, lkb);
2685	send_bast_queue(r, &r->res_convertqueue, lkb);
2686}
2687
2688/* set_master(r, lkb) -- set the master nodeid of a resource
2689
2690   The purpose of this function is to set the nodeid field in the given
2691   lkb using the nodeid field in the given rsb.  If the rsb's nodeid is
2692   known, it can just be copied to the lkb and the function will return
2693   0.  If the rsb's nodeid is _not_ known, it needs to be looked up
2694   before it can be copied to the lkb.
2695
2696   When the rsb nodeid is being looked up remotely, the initial lkb
2697   causing the lookup is kept on the ls_waiters list waiting for the
2698   lookup reply.  Other lkb's waiting for the same rsb lookup are kept
2699   on the rsb's res_lookup list until the master is verified.
2700
2701   Return values:
2702   0: nodeid is set in rsb/lkb and the caller should go ahead and use it
2703   1: the rsb master is not available and the lkb has been placed on
2704      a wait queue
2705*/
2706
2707static int set_master(struct dlm_rsb *r, struct dlm_lkb *lkb)
2708{
2709	int our_nodeid = dlm_our_nodeid();
2710
2711	if (rsb_flag(r, RSB_MASTER_UNCERTAIN)) {
2712		rsb_clear_flag(r, RSB_MASTER_UNCERTAIN);
2713		r->res_first_lkid = lkb->lkb_id;
2714		lkb->lkb_nodeid = r->res_nodeid;
2715		return 0;
2716	}
2717
2718	if (r->res_first_lkid && r->res_first_lkid != lkb->lkb_id) {
2719		list_add_tail(&lkb->lkb_rsb_lookup, &r->res_lookup);
2720		return 1;
2721	}
2722
2723	if (r->res_master_nodeid == our_nodeid) {
2724		lkb->lkb_nodeid = 0;
2725		return 0;
2726	}
2727
2728	if (r->res_master_nodeid) {
2729		lkb->lkb_nodeid = r->res_master_nodeid;
2730		return 0;
2731	}
2732
2733	if (dlm_dir_nodeid(r) == our_nodeid) {
2734		/* This is a somewhat unusual case; find_rsb will usually
2735		   have set res_master_nodeid when dir nodeid is local, but
2736		   there are cases where we become the dir node after we've
2737		   past find_rsb and go through _request_lock again.
2738		   confirm_master() or process_lookup_list() needs to be
2739		   called after this. */
2740		log_debug(r->res_ls, "set_master %x self master %d dir %d %s",
2741			  lkb->lkb_id, r->res_master_nodeid, r->res_dir_nodeid,
2742			  r->res_name);
2743		r->res_master_nodeid = our_nodeid;
2744		r->res_nodeid = 0;
2745		lkb->lkb_nodeid = 0;
2746		return 0;
2747	}
2748
2749	wait_pending_remove(r);
2750
2751	r->res_first_lkid = lkb->lkb_id;
2752	send_lookup(r, lkb);
2753	return 1;
2754}
2755
2756static void process_lookup_list(struct dlm_rsb *r)
2757{
2758	struct dlm_lkb *lkb, *safe;
2759
2760	list_for_each_entry_safe(lkb, safe, &r->res_lookup, lkb_rsb_lookup) {
2761		list_del_init(&lkb->lkb_rsb_lookup);
2762		_request_lock(r, lkb);
2763		schedule();
2764	}
2765}
2766
2767/* confirm_master -- confirm (or deny) an rsb's master nodeid */
2768
2769static void confirm_master(struct dlm_rsb *r, int error)
2770{
2771	struct dlm_lkb *lkb;
2772
2773	if (!r->res_first_lkid)
2774		return;
2775
2776	switch (error) {
2777	case 0:
2778	case -EINPROGRESS:
2779		r->res_first_lkid = 0;
2780		process_lookup_list(r);
2781		break;
2782
2783	case -EAGAIN:
2784	case -EBADR:
2785	case -ENOTBLK:
2786		/* the remote request failed and won't be retried (it was
2787		   a NOQUEUE, or has been canceled/unlocked); make a waiting
2788		   lkb the first_lkid */
2789
2790		r->res_first_lkid = 0;
2791
2792		if (!list_empty(&r->res_lookup)) {
2793			lkb = list_entry(r->res_lookup.next, struct dlm_lkb,
2794					 lkb_rsb_lookup);
2795			list_del_init(&lkb->lkb_rsb_lookup);
2796			r->res_first_lkid = lkb->lkb_id;
2797			_request_lock(r, lkb);
2798		}
2799		break;
2800
2801	default:
2802		log_error(r->res_ls, "confirm_master unknown error %d", error);
2803	}
2804}
2805
2806static int set_lock_args(int mode, struct dlm_lksb *lksb, uint32_t flags,
2807			 int namelen, unsigned long timeout_cs,
2808			 void (*ast) (void *astparam),
2809			 void *astparam,
2810			 void (*bast) (void *astparam, int mode),
2811			 struct dlm_args *args)
2812{
2813	int rv = -EINVAL;
2814
2815	/* check for invalid arg usage */
2816
2817	if (mode < 0 || mode > DLM_LOCK_EX)
2818		goto out;
2819
2820	if (!(flags & DLM_LKF_CONVERT) && (namelen > DLM_RESNAME_MAXLEN))
2821		goto out;
2822
2823	if (flags & DLM_LKF_CANCEL)
2824		goto out;
2825
2826	if (flags & DLM_LKF_QUECVT && !(flags & DLM_LKF_CONVERT))
2827		goto out;
2828
2829	if (flags & DLM_LKF_CONVDEADLK && !(flags & DLM_LKF_CONVERT))
2830		goto out;
2831
2832	if (flags & DLM_LKF_CONVDEADLK && flags & DLM_LKF_NOQUEUE)
2833		goto out;
2834
2835	if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_CONVERT)
2836		goto out;
2837
2838	if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_QUECVT)
2839		goto out;
2840
2841	if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_NOQUEUE)
2842		goto out;
2843
2844	if (flags & DLM_LKF_EXPEDITE && mode != DLM_LOCK_NL)
2845		goto out;
2846
2847	if (!ast || !lksb)
2848		goto out;
2849
2850	if (flags & DLM_LKF_VALBLK && !lksb->sb_lvbptr)
2851		goto out;
2852
2853	if (flags & DLM_LKF_CONVERT && !lksb->sb_lkid)
2854		goto out;
2855
2856	/* these args will be copied to the lkb in validate_lock_args,
2857	   it cannot be done now because when converting locks, fields in
2858	   an active lkb cannot be modified before locking the rsb */
2859
2860	args->flags = flags;
2861	args->astfn = ast;
2862	args->astparam = astparam;
2863	args->bastfn = bast;
2864	args->timeout = timeout_cs;
2865	args->mode = mode;
2866	args->lksb = lksb;
2867	rv = 0;
2868 out:
2869	return rv;
2870}
2871
2872static int set_unlock_args(uint32_t flags, void *astarg, struct dlm_args *args)
2873{
2874	if (flags & ~(DLM_LKF_CANCEL | DLM_LKF_VALBLK | DLM_LKF_IVVALBLK |
2875 		      DLM_LKF_FORCEUNLOCK))
2876		return -EINVAL;
2877
2878	if (flags & DLM_LKF_CANCEL && flags & DLM_LKF_FORCEUNLOCK)
2879		return -EINVAL;
2880
2881	args->flags = flags;
2882	args->astparam = astarg;
2883	return 0;
2884}
2885
2886static int validate_lock_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
2887			      struct dlm_args *args)
2888{
2889	int rv = -EINVAL;
2890
2891	if (args->flags & DLM_LKF_CONVERT) {
2892		if (lkb->lkb_flags & DLM_IFL_MSTCPY)
2893			goto out;
2894
2895		if (args->flags & DLM_LKF_QUECVT &&
2896		    !__quecvt_compat_matrix[lkb->lkb_grmode+1][args->mode+1])
2897			goto out;
2898
2899		rv = -EBUSY;
2900		if (lkb->lkb_status != DLM_LKSTS_GRANTED)
2901			goto out;
2902
2903		if (lkb->lkb_wait_type)
 
2904			goto out;
2905
2906		if (is_overlap(lkb))
 
2907			goto out;
2908	}
2909
2910	lkb->lkb_exflags = args->flags;
2911	lkb->lkb_sbflags = 0;
2912	lkb->lkb_astfn = args->astfn;
2913	lkb->lkb_astparam = args->astparam;
2914	lkb->lkb_bastfn = args->bastfn;
2915	lkb->lkb_rqmode = args->mode;
2916	lkb->lkb_lksb = args->lksb;
2917	lkb->lkb_lvbptr = args->lksb->sb_lvbptr;
2918	lkb->lkb_ownpid = (int) current->pid;
2919	lkb->lkb_timeout_cs = args->timeout;
2920	rv = 0;
2921 out:
2922	if (rv)
2923		log_debug(ls, "validate_lock_args %d %x %x %x %d %d %s",
2924			  rv, lkb->lkb_id, lkb->lkb_flags, args->flags,
 
 
 
 
 
 
 
 
 
 
 
2925			  lkb->lkb_status, lkb->lkb_wait_type,
2926			  lkb->lkb_resource->res_name);
 
 
 
2927	return rv;
2928}
2929
2930/* when dlm_unlock() sees -EBUSY with CANCEL/FORCEUNLOCK it returns 0
2931   for success */
2932
2933/* note: it's valid for lkb_nodeid/res_nodeid to be -1 when we get here
2934   because there may be a lookup in progress and it's valid to do
2935   cancel/unlockf on it */
2936
2937static int validate_unlock_args(struct dlm_lkb *lkb, struct dlm_args *args)
2938{
2939	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
2940	int rv = -EINVAL;
2941
2942	if (lkb->lkb_flags & DLM_IFL_MSTCPY) {
2943		log_error(ls, "unlock on MSTCPY %x", lkb->lkb_id);
2944		dlm_print_lkb(lkb);
2945		goto out;
2946	}
2947
2948	/* an lkb may still exist even though the lock is EOL'ed due to a
2949	   cancel, unlock or failed noqueue request; an app can't use these
2950	   locks; return same error as if the lkid had not been found at all */
2951
2952	if (lkb->lkb_flags & DLM_IFL_ENDOFLIFE) {
2953		log_debug(ls, "unlock on ENDOFLIFE %x", lkb->lkb_id);
2954		rv = -ENOENT;
2955		goto out;
2956	}
2957
2958	/* an lkb may be waiting for an rsb lookup to complete where the
2959	   lookup was initiated by another lock */
2960
2961	if (!list_empty(&lkb->lkb_rsb_lookup)) {
2962		if (args->flags & (DLM_LKF_CANCEL | DLM_LKF_FORCEUNLOCK)) {
2963			log_debug(ls, "unlock on rsb_lookup %x", lkb->lkb_id);
2964			list_del_init(&lkb->lkb_rsb_lookup);
2965			queue_cast(lkb->lkb_resource, lkb,
2966				   args->flags & DLM_LKF_CANCEL ?
2967				   -DLM_ECANCEL : -DLM_EUNLOCK);
2968			unhold_lkb(lkb); /* undoes create_lkb() */
2969		}
2970		/* caller changes -EBUSY to 0 for CANCEL and FORCEUNLOCK */
2971		rv = -EBUSY;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2972		goto out;
2973	}
2974
2975	/* cancel not allowed with another cancel/unlock in progress */
2976
2977	if (args->flags & DLM_LKF_CANCEL) {
2978		if (lkb->lkb_exflags & DLM_LKF_CANCEL)
2979			goto out;
2980
2981		if (is_overlap(lkb))
2982			goto out;
2983
2984		/* don't let scand try to do a cancel */
2985		del_timeout(lkb);
2986
2987		if (lkb->lkb_flags & DLM_IFL_RESEND) {
2988			lkb->lkb_flags |= DLM_IFL_OVERLAP_CANCEL;
2989			rv = -EBUSY;
2990			goto out;
2991		}
2992
2993		/* there's nothing to cancel */
2994		if (lkb->lkb_status == DLM_LKSTS_GRANTED &&
2995		    !lkb->lkb_wait_type) {
2996			rv = -EBUSY;
2997			goto out;
2998		}
2999
3000		switch (lkb->lkb_wait_type) {
3001		case DLM_MSG_LOOKUP:
3002		case DLM_MSG_REQUEST:
3003			lkb->lkb_flags |= DLM_IFL_OVERLAP_CANCEL;
3004			rv = -EBUSY;
3005			goto out;
3006		case DLM_MSG_UNLOCK:
3007		case DLM_MSG_CANCEL:
3008			goto out;
3009		}
3010		/* add_to_waiters() will set OVERLAP_CANCEL */
3011		goto out_ok;
3012	}
3013
3014	/* do we need to allow a force-unlock if there's a normal unlock
3015	   already in progress?  in what conditions could the normal unlock
3016	   fail such that we'd want to send a force-unlock to be sure? */
3017
3018	if (args->flags & DLM_LKF_FORCEUNLOCK) {
3019		if (lkb->lkb_exflags & DLM_LKF_FORCEUNLOCK)
3020			goto out;
3021
3022		if (is_overlap_unlock(lkb))
3023			goto out;
3024
3025		/* don't let scand try to do a cancel */
3026		del_timeout(lkb);
3027
3028		if (lkb->lkb_flags & DLM_IFL_RESEND) {
3029			lkb->lkb_flags |= DLM_IFL_OVERLAP_UNLOCK;
3030			rv = -EBUSY;
3031			goto out;
3032		}
3033
3034		switch (lkb->lkb_wait_type) {
3035		case DLM_MSG_LOOKUP:
3036		case DLM_MSG_REQUEST:
3037			lkb->lkb_flags |= DLM_IFL_OVERLAP_UNLOCK;
3038			rv = -EBUSY;
3039			goto out;
3040		case DLM_MSG_UNLOCK:
3041			goto out;
3042		}
3043		/* add_to_waiters() will set OVERLAP_UNLOCK */
3044		goto out_ok;
3045	}
3046
3047	/* normal unlock not allowed if there's any op in progress */
3048	rv = -EBUSY;
3049	if (lkb->lkb_wait_type || lkb->lkb_wait_count)
3050		goto out;
3051
3052 out_ok:
3053	/* an overlapping op shouldn't blow away exflags from other op */
3054	lkb->lkb_exflags |= args->flags;
3055	lkb->lkb_sbflags = 0;
3056	lkb->lkb_astparam = args->astparam;
3057	rv = 0;
3058 out:
3059	if (rv)
3060		log_debug(ls, "validate_unlock_args %d %x %x %x %x %d %s", rv,
3061			  lkb->lkb_id, lkb->lkb_flags, lkb->lkb_exflags,
 
 
 
 
 
 
 
 
 
 
 
3062			  args->flags, lkb->lkb_wait_type,
3063			  lkb->lkb_resource->res_name);
 
 
 
3064	return rv;
3065}
3066
3067/*
3068 * Four stage 4 varieties:
3069 * do_request(), do_convert(), do_unlock(), do_cancel()
3070 * These are called on the master node for the given lock and
3071 * from the central locking logic.
3072 */
3073
3074static int do_request(struct dlm_rsb *r, struct dlm_lkb *lkb)
3075{
3076	int error = 0;
3077
3078	if (can_be_granted(r, lkb, 1, 0, NULL)) {
3079		grant_lock(r, lkb);
3080		queue_cast(r, lkb, 0);
3081		goto out;
3082	}
3083
3084	if (can_be_queued(lkb)) {
3085		error = -EINPROGRESS;
3086		add_lkb(r, lkb, DLM_LKSTS_WAITING);
3087		add_timeout(lkb);
3088		goto out;
3089	}
3090
3091	error = -EAGAIN;
3092	queue_cast(r, lkb, -EAGAIN);
3093 out:
3094	return error;
3095}
3096
3097static void do_request_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
3098			       int error)
3099{
3100	switch (error) {
3101	case -EAGAIN:
3102		if (force_blocking_asts(lkb))
3103			send_blocking_asts_all(r, lkb);
3104		break;
3105	case -EINPROGRESS:
3106		send_blocking_asts(r, lkb);
3107		break;
3108	}
3109}
3110
3111static int do_convert(struct dlm_rsb *r, struct dlm_lkb *lkb)
3112{
3113	int error = 0;
3114	int deadlk = 0;
3115
3116	/* changing an existing lock may allow others to be granted */
3117
3118	if (can_be_granted(r, lkb, 1, 0, &deadlk)) {
3119		grant_lock(r, lkb);
3120		queue_cast(r, lkb, 0);
3121		goto out;
3122	}
3123
3124	/* can_be_granted() detected that this lock would block in a conversion
3125	   deadlock, so we leave it on the granted queue and return EDEADLK in
3126	   the ast for the convert. */
3127
3128	if (deadlk && !(lkb->lkb_exflags & DLM_LKF_NODLCKWT)) {
3129		/* it's left on the granted queue */
3130		revert_lock(r, lkb);
3131		queue_cast(r, lkb, -EDEADLK);
3132		error = -EDEADLK;
3133		goto out;
3134	}
3135
3136	/* is_demoted() means the can_be_granted() above set the grmode
3137	   to NL, and left us on the granted queue.  This auto-demotion
3138	   (due to CONVDEADLK) might mean other locks, and/or this lock, are
3139	   now grantable.  We have to try to grant other converting locks
3140	   before we try again to grant this one. */
3141
3142	if (is_demoted(lkb)) {
3143		grant_pending_convert(r, DLM_LOCK_IV, NULL, NULL);
3144		if (_can_be_granted(r, lkb, 1, 0)) {
3145			grant_lock(r, lkb);
3146			queue_cast(r, lkb, 0);
3147			goto out;
3148		}
3149		/* else fall through and move to convert queue */
3150	}
3151
3152	if (can_be_queued(lkb)) {
3153		error = -EINPROGRESS;
3154		del_lkb(r, lkb);
3155		add_lkb(r, lkb, DLM_LKSTS_CONVERT);
3156		add_timeout(lkb);
3157		goto out;
3158	}
3159
3160	error = -EAGAIN;
3161	queue_cast(r, lkb, -EAGAIN);
3162 out:
3163	return error;
3164}
3165
3166static void do_convert_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
3167			       int error)
3168{
3169	switch (error) {
3170	case 0:
3171		grant_pending_locks(r, NULL);
3172		/* grant_pending_locks also sends basts */
3173		break;
3174	case -EAGAIN:
3175		if (force_blocking_asts(lkb))
3176			send_blocking_asts_all(r, lkb);
3177		break;
3178	case -EINPROGRESS:
3179		send_blocking_asts(r, lkb);
3180		break;
3181	}
3182}
3183
3184static int do_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3185{
3186	remove_lock(r, lkb);
3187	queue_cast(r, lkb, -DLM_EUNLOCK);
3188	return -DLM_EUNLOCK;
3189}
3190
3191static void do_unlock_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
3192			      int error)
3193{
3194	grant_pending_locks(r, NULL);
3195}
3196
3197/* returns: 0 did nothing, -DLM_ECANCEL canceled lock */
3198
3199static int do_cancel(struct dlm_rsb *r, struct dlm_lkb *lkb)
3200{
3201	int error;
3202
3203	error = revert_lock(r, lkb);
3204	if (error) {
3205		queue_cast(r, lkb, -DLM_ECANCEL);
3206		return -DLM_ECANCEL;
3207	}
3208	return 0;
3209}
3210
3211static void do_cancel_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
3212			      int error)
3213{
3214	if (error)
3215		grant_pending_locks(r, NULL);
3216}
3217
3218/*
3219 * Four stage 3 varieties:
3220 * _request_lock(), _convert_lock(), _unlock_lock(), _cancel_lock()
3221 */
3222
3223/* add a new lkb to a possibly new rsb, called by requesting process */
3224
3225static int _request_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3226{
3227	int error;
3228
3229	/* set_master: sets lkb nodeid from r */
3230
3231	error = set_master(r, lkb);
3232	if (error < 0)
3233		goto out;
3234	if (error) {
3235		error = 0;
3236		goto out;
3237	}
3238
3239	if (is_remote(r)) {
3240		/* receive_request() calls do_request() on remote node */
3241		error = send_request(r, lkb);
3242	} else {
3243		error = do_request(r, lkb);
3244		/* for remote locks the request_reply is sent
3245		   between do_request and do_request_effects */
3246		do_request_effects(r, lkb, error);
3247	}
3248 out:
3249	return error;
3250}
3251
3252/* change some property of an existing lkb, e.g. mode */
3253
3254static int _convert_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3255{
3256	int error;
3257
3258	if (is_remote(r)) {
3259		/* receive_convert() calls do_convert() on remote node */
3260		error = send_convert(r, lkb);
3261	} else {
3262		error = do_convert(r, lkb);
3263		/* for remote locks the convert_reply is sent
3264		   between do_convert and do_convert_effects */
3265		do_convert_effects(r, lkb, error);
3266	}
3267
3268	return error;
3269}
3270
3271/* remove an existing lkb from the granted queue */
3272
3273static int _unlock_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3274{
3275	int error;
3276
3277	if (is_remote(r)) {
3278		/* receive_unlock() calls do_unlock() on remote node */
3279		error = send_unlock(r, lkb);
3280	} else {
3281		error = do_unlock(r, lkb);
3282		/* for remote locks the unlock_reply is sent
3283		   between do_unlock and do_unlock_effects */
3284		do_unlock_effects(r, lkb, error);
3285	}
3286
3287	return error;
3288}
3289
3290/* remove an existing lkb from the convert or wait queue */
3291
3292static int _cancel_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3293{
3294	int error;
3295
3296	if (is_remote(r)) {
3297		/* receive_cancel() calls do_cancel() on remote node */
3298		error = send_cancel(r, lkb);
3299	} else {
3300		error = do_cancel(r, lkb);
3301		/* for remote locks the cancel_reply is sent
3302		   between do_cancel and do_cancel_effects */
3303		do_cancel_effects(r, lkb, error);
3304	}
3305
3306	return error;
3307}
3308
3309/*
3310 * Four stage 2 varieties:
3311 * request_lock(), convert_lock(), unlock_lock(), cancel_lock()
3312 */
3313
3314static int request_lock(struct dlm_ls *ls, struct dlm_lkb *lkb, char *name,
3315			int len, struct dlm_args *args)
 
3316{
3317	struct dlm_rsb *r;
3318	int error;
3319
3320	error = validate_lock_args(ls, lkb, args);
3321	if (error)
3322		return error;
3323
3324	error = find_rsb(ls, name, len, 0, R_REQUEST, &r);
3325	if (error)
3326		return error;
3327
3328	lock_rsb(r);
3329
3330	attach_lkb(r, lkb);
3331	lkb->lkb_lksb->sb_lkid = lkb->lkb_id;
3332
3333	error = _request_lock(r, lkb);
3334
3335	unlock_rsb(r);
3336	put_rsb(r);
3337	return error;
3338}
3339
3340static int convert_lock(struct dlm_ls *ls, struct dlm_lkb *lkb,
3341			struct dlm_args *args)
3342{
3343	struct dlm_rsb *r;
3344	int error;
3345
3346	r = lkb->lkb_resource;
3347
3348	hold_rsb(r);
3349	lock_rsb(r);
3350
3351	error = validate_lock_args(ls, lkb, args);
3352	if (error)
3353		goto out;
3354
3355	error = _convert_lock(r, lkb);
3356 out:
3357	unlock_rsb(r);
3358	put_rsb(r);
3359	return error;
3360}
3361
3362static int unlock_lock(struct dlm_ls *ls, struct dlm_lkb *lkb,
3363		       struct dlm_args *args)
3364{
3365	struct dlm_rsb *r;
3366	int error;
3367
3368	r = lkb->lkb_resource;
3369
3370	hold_rsb(r);
3371	lock_rsb(r);
3372
3373	error = validate_unlock_args(lkb, args);
3374	if (error)
3375		goto out;
3376
3377	error = _unlock_lock(r, lkb);
3378 out:
3379	unlock_rsb(r);
3380	put_rsb(r);
3381	return error;
3382}
3383
3384static int cancel_lock(struct dlm_ls *ls, struct dlm_lkb *lkb,
3385		       struct dlm_args *args)
3386{
3387	struct dlm_rsb *r;
3388	int error;
3389
3390	r = lkb->lkb_resource;
3391
3392	hold_rsb(r);
3393	lock_rsb(r);
3394
3395	error = validate_unlock_args(lkb, args);
3396	if (error)
3397		goto out;
3398
3399	error = _cancel_lock(r, lkb);
3400 out:
3401	unlock_rsb(r);
3402	put_rsb(r);
3403	return error;
3404}
3405
3406/*
3407 * Two stage 1 varieties:  dlm_lock() and dlm_unlock()
3408 */
3409
3410int dlm_lock(dlm_lockspace_t *lockspace,
3411	     int mode,
3412	     struct dlm_lksb *lksb,
3413	     uint32_t flags,
3414	     void *name,
3415	     unsigned int namelen,
3416	     uint32_t parent_lkid,
3417	     void (*ast) (void *astarg),
3418	     void *astarg,
3419	     void (*bast) (void *astarg, int mode))
3420{
3421	struct dlm_ls *ls;
3422	struct dlm_lkb *lkb;
3423	struct dlm_args args;
3424	int error, convert = flags & DLM_LKF_CONVERT;
3425
3426	ls = dlm_find_lockspace_local(lockspace);
3427	if (!ls)
3428		return -EINVAL;
3429
3430	dlm_lock_recovery(ls);
3431
3432	if (convert)
3433		error = find_lkb(ls, lksb->sb_lkid, &lkb);
3434	else
3435		error = create_lkb(ls, &lkb);
3436
3437	if (error)
3438		goto out;
3439
3440	error = set_lock_args(mode, lksb, flags, namelen, 0, ast,
3441			      astarg, bast, &args);
 
 
3442	if (error)
3443		goto out_put;
3444
3445	if (convert)
3446		error = convert_lock(ls, lkb, &args);
3447	else
3448		error = request_lock(ls, lkb, name, namelen, &args);
3449
3450	if (error == -EINPROGRESS)
3451		error = 0;
3452 out_put:
 
 
3453	if (convert || error)
3454		__put_lkb(ls, lkb);
3455	if (error == -EAGAIN || error == -EDEADLK)
3456		error = 0;
3457 out:
3458	dlm_unlock_recovery(ls);
3459	dlm_put_lockspace(ls);
3460	return error;
3461}
3462
3463int dlm_unlock(dlm_lockspace_t *lockspace,
3464	       uint32_t lkid,
3465	       uint32_t flags,
3466	       struct dlm_lksb *lksb,
3467	       void *astarg)
3468{
3469	struct dlm_ls *ls;
3470	struct dlm_lkb *lkb;
3471	struct dlm_args args;
3472	int error;
3473
3474	ls = dlm_find_lockspace_local(lockspace);
3475	if (!ls)
3476		return -EINVAL;
3477
3478	dlm_lock_recovery(ls);
3479
3480	error = find_lkb(ls, lkid, &lkb);
3481	if (error)
3482		goto out;
3483
 
 
3484	error = set_unlock_args(flags, astarg, &args);
3485	if (error)
3486		goto out_put;
3487
3488	if (flags & DLM_LKF_CANCEL)
3489		error = cancel_lock(ls, lkb, &args);
3490	else
3491		error = unlock_lock(ls, lkb, &args);
3492
3493	if (error == -DLM_EUNLOCK || error == -DLM_ECANCEL)
3494		error = 0;
3495	if (error == -EBUSY && (flags & (DLM_LKF_CANCEL | DLM_LKF_FORCEUNLOCK)))
3496		error = 0;
3497 out_put:
 
 
3498	dlm_put_lkb(lkb);
3499 out:
3500	dlm_unlock_recovery(ls);
3501	dlm_put_lockspace(ls);
3502	return error;
3503}
3504
3505/*
3506 * send/receive routines for remote operations and replies
3507 *
3508 * send_args
3509 * send_common
3510 * send_request			receive_request
3511 * send_convert			receive_convert
3512 * send_unlock			receive_unlock
3513 * send_cancel			receive_cancel
3514 * send_grant			receive_grant
3515 * send_bast			receive_bast
3516 * send_lookup			receive_lookup
3517 * send_remove			receive_remove
3518 *
3519 * 				send_common_reply
3520 * receive_request_reply	send_request_reply
3521 * receive_convert_reply	send_convert_reply
3522 * receive_unlock_reply		send_unlock_reply
3523 * receive_cancel_reply		send_cancel_reply
3524 * receive_lookup_reply		send_lookup_reply
3525 */
3526
3527static int _create_message(struct dlm_ls *ls, int mb_len,
3528			   int to_nodeid, int mstype,
3529			   struct dlm_message **ms_ret,
3530			   struct dlm_mhandle **mh_ret)
 
3531{
3532	struct dlm_message *ms;
3533	struct dlm_mhandle *mh;
3534	char *mb;
3535
3536	/* get_buffer gives us a message handle (mh) that we need to
3537	   pass into midcomms_commit and a message buffer (mb) that we
3538	   write our data into */
3539
3540	mh = dlm_midcomms_get_mhandle(to_nodeid, mb_len, GFP_NOFS, &mb);
3541	if (!mh)
3542		return -ENOBUFS;
3543
3544	ms = (struct dlm_message *) mb;
3545
3546	ms->m_header.h_version = (DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
3547	ms->m_header.u.h_lockspace = ls->ls_global_id;
3548	ms->m_header.h_nodeid = dlm_our_nodeid();
3549	ms->m_header.h_length = mb_len;
3550	ms->m_header.h_cmd = DLM_MSG;
3551
3552	ms->m_type = mstype;
3553
3554	*mh_ret = mh;
3555	*ms_ret = ms;
3556	return 0;
3557}
3558
3559static int create_message(struct dlm_rsb *r, struct dlm_lkb *lkb,
3560			  int to_nodeid, int mstype,
3561			  struct dlm_message **ms_ret,
3562			  struct dlm_mhandle **mh_ret)
 
3563{
3564	int mb_len = sizeof(struct dlm_message);
3565
3566	switch (mstype) {
3567	case DLM_MSG_REQUEST:
3568	case DLM_MSG_LOOKUP:
3569	case DLM_MSG_REMOVE:
3570		mb_len += r->res_length;
3571		break;
3572	case DLM_MSG_CONVERT:
3573	case DLM_MSG_UNLOCK:
3574	case DLM_MSG_REQUEST_REPLY:
3575	case DLM_MSG_CONVERT_REPLY:
3576	case DLM_MSG_GRANT:
3577		if (lkb && lkb->lkb_lvbptr)
3578			mb_len += r->res_ls->ls_lvblen;
3579		break;
3580	}
3581
3582	return _create_message(r->res_ls, mb_len, to_nodeid, mstype,
3583			       ms_ret, mh_ret);
3584}
3585
3586/* further lowcomms enhancements or alternate implementations may make
3587   the return value from this function useful at some point */
3588
3589static int send_message(struct dlm_mhandle *mh, struct dlm_message *ms)
 
3590{
3591	dlm_message_out(ms);
3592	dlm_midcomms_commit_mhandle(mh);
3593	return 0;
3594}
3595
3596static void send_args(struct dlm_rsb *r, struct dlm_lkb *lkb,
3597		      struct dlm_message *ms)
3598{
3599	ms->m_nodeid   = lkb->lkb_nodeid;
3600	ms->m_pid      = lkb->lkb_ownpid;
3601	ms->m_lkid     = lkb->lkb_id;
3602	ms->m_remid    = lkb->lkb_remid;
3603	ms->m_exflags  = lkb->lkb_exflags;
3604	ms->m_sbflags  = lkb->lkb_sbflags;
3605	ms->m_flags    = lkb->lkb_flags;
3606	ms->m_lvbseq   = lkb->lkb_lvbseq;
3607	ms->m_status   = lkb->lkb_status;
3608	ms->m_grmode   = lkb->lkb_grmode;
3609	ms->m_rqmode   = lkb->lkb_rqmode;
3610	ms->m_hash     = r->res_hash;
3611
3612	/* m_result and m_bastmode are set from function args,
3613	   not from lkb fields */
3614
3615	if (lkb->lkb_bastfn)
3616		ms->m_asts |= DLM_CB_BAST;
3617	if (lkb->lkb_astfn)
3618		ms->m_asts |= DLM_CB_CAST;
3619
3620	/* compare with switch in create_message; send_remove() doesn't
3621	   use send_args() */
3622
3623	switch (ms->m_type) {
3624	case DLM_MSG_REQUEST:
3625	case DLM_MSG_LOOKUP:
3626		memcpy(ms->m_extra, r->res_name, r->res_length);
3627		break;
3628	case DLM_MSG_CONVERT:
3629	case DLM_MSG_UNLOCK:
3630	case DLM_MSG_REQUEST_REPLY:
3631	case DLM_MSG_CONVERT_REPLY:
3632	case DLM_MSG_GRANT:
3633		if (!lkb->lkb_lvbptr)
3634			break;
3635		memcpy(ms->m_extra, lkb->lkb_lvbptr, r->res_ls->ls_lvblen);
3636		break;
3637	}
3638}
3639
3640static int send_common(struct dlm_rsb *r, struct dlm_lkb *lkb, int mstype)
3641{
3642	struct dlm_message *ms;
3643	struct dlm_mhandle *mh;
3644	int to_nodeid, error;
3645
3646	to_nodeid = r->res_nodeid;
3647
3648	error = add_to_waiters(lkb, mstype, to_nodeid);
3649	if (error)
3650		return error;
3651
3652	error = create_message(r, lkb, to_nodeid, mstype, &ms, &mh);
3653	if (error)
3654		goto fail;
3655
3656	send_args(r, lkb, ms);
3657
3658	error = send_message(mh, ms);
3659	if (error)
3660		goto fail;
3661	return 0;
3662
3663 fail:
3664	remove_from_waiters(lkb, msg_reply_type(mstype));
3665	return error;
3666}
3667
3668static int send_request(struct dlm_rsb *r, struct dlm_lkb *lkb)
3669{
3670	return send_common(r, lkb, DLM_MSG_REQUEST);
3671}
3672
3673static int send_convert(struct dlm_rsb *r, struct dlm_lkb *lkb)
3674{
3675	int error;
3676
3677	error = send_common(r, lkb, DLM_MSG_CONVERT);
3678
3679	/* down conversions go without a reply from the master */
3680	if (!error && down_conversion(lkb)) {
3681		remove_from_waiters(lkb, DLM_MSG_CONVERT_REPLY);
3682		r->res_ls->ls_stub_ms.m_flags = DLM_IFL_STUB_MS;
3683		r->res_ls->ls_stub_ms.m_type = DLM_MSG_CONVERT_REPLY;
3684		r->res_ls->ls_stub_ms.m_result = 0;
3685		__receive_convert_reply(r, lkb, &r->res_ls->ls_stub_ms);
3686	}
3687
3688	return error;
3689}
3690
3691/* FIXME: if this lkb is the only lock we hold on the rsb, then set
3692   MASTER_UNCERTAIN to force the next request on the rsb to confirm
3693   that the master is still correct. */
3694
3695static int send_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3696{
3697	return send_common(r, lkb, DLM_MSG_UNLOCK);
3698}
3699
3700static int send_cancel(struct dlm_rsb *r, struct dlm_lkb *lkb)
3701{
3702	return send_common(r, lkb, DLM_MSG_CANCEL);
3703}
3704
3705static int send_grant(struct dlm_rsb *r, struct dlm_lkb *lkb)
3706{
3707	struct dlm_message *ms;
3708	struct dlm_mhandle *mh;
3709	int to_nodeid, error;
3710
3711	to_nodeid = lkb->lkb_nodeid;
3712
3713	error = create_message(r, lkb, to_nodeid, DLM_MSG_GRANT, &ms, &mh);
 
3714	if (error)
3715		goto out;
3716
3717	send_args(r, lkb, ms);
3718
3719	ms->m_result = 0;
3720
3721	error = send_message(mh, ms);
3722 out:
3723	return error;
3724}
3725
3726static int send_bast(struct dlm_rsb *r, struct dlm_lkb *lkb, int mode)
3727{
3728	struct dlm_message *ms;
3729	struct dlm_mhandle *mh;
3730	int to_nodeid, error;
3731
3732	to_nodeid = lkb->lkb_nodeid;
3733
3734	error = create_message(r, NULL, to_nodeid, DLM_MSG_BAST, &ms, &mh);
 
3735	if (error)
3736		goto out;
3737
3738	send_args(r, lkb, ms);
3739
3740	ms->m_bastmode = mode;
3741
3742	error = send_message(mh, ms);
3743 out:
3744	return error;
3745}
3746
3747static int send_lookup(struct dlm_rsb *r, struct dlm_lkb *lkb)
3748{
3749	struct dlm_message *ms;
3750	struct dlm_mhandle *mh;
3751	int to_nodeid, error;
3752
3753	to_nodeid = dlm_dir_nodeid(r);
3754
3755	error = add_to_waiters(lkb, DLM_MSG_LOOKUP, to_nodeid);
3756	if (error)
3757		return error;
3758
3759	error = create_message(r, NULL, to_nodeid, DLM_MSG_LOOKUP, &ms, &mh);
 
3760	if (error)
3761		goto fail;
3762
3763	send_args(r, lkb, ms);
3764
3765	error = send_message(mh, ms);
3766	if (error)
3767		goto fail;
3768	return 0;
3769
3770 fail:
3771	remove_from_waiters(lkb, DLM_MSG_LOOKUP_REPLY);
3772	return error;
3773}
3774
3775static int send_remove(struct dlm_rsb *r)
3776{
3777	struct dlm_message *ms;
3778	struct dlm_mhandle *mh;
3779	int to_nodeid, error;
3780
3781	to_nodeid = dlm_dir_nodeid(r);
3782
3783	error = create_message(r, NULL, to_nodeid, DLM_MSG_REMOVE, &ms, &mh);
 
3784	if (error)
3785		goto out;
3786
3787	memcpy(ms->m_extra, r->res_name, r->res_length);
3788	ms->m_hash = r->res_hash;
3789
3790	error = send_message(mh, ms);
3791 out:
3792	return error;
3793}
3794
3795static int send_common_reply(struct dlm_rsb *r, struct dlm_lkb *lkb,
3796			     int mstype, int rv)
3797{
3798	struct dlm_message *ms;
3799	struct dlm_mhandle *mh;
3800	int to_nodeid, error;
3801
3802	to_nodeid = lkb->lkb_nodeid;
3803
3804	error = create_message(r, lkb, to_nodeid, mstype, &ms, &mh);
3805	if (error)
3806		goto out;
3807
3808	send_args(r, lkb, ms);
3809
3810	ms->m_result = rv;
3811
3812	error = send_message(mh, ms);
3813 out:
3814	return error;
3815}
3816
3817static int send_request_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3818{
3819	return send_common_reply(r, lkb, DLM_MSG_REQUEST_REPLY, rv);
3820}
3821
3822static int send_convert_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3823{
3824	return send_common_reply(r, lkb, DLM_MSG_CONVERT_REPLY, rv);
3825}
3826
3827static int send_unlock_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3828{
3829	return send_common_reply(r, lkb, DLM_MSG_UNLOCK_REPLY, rv);
3830}
3831
3832static int send_cancel_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3833{
3834	return send_common_reply(r, lkb, DLM_MSG_CANCEL_REPLY, rv);
3835}
3836
3837static int send_lookup_reply(struct dlm_ls *ls, struct dlm_message *ms_in,
3838			     int ret_nodeid, int rv)
 
3839{
3840	struct dlm_rsb *r = &ls->ls_stub_rsb;
3841	struct dlm_message *ms;
3842	struct dlm_mhandle *mh;
3843	int error, nodeid = ms_in->m_header.h_nodeid;
3844
3845	error = create_message(r, NULL, nodeid, DLM_MSG_LOOKUP_REPLY, &ms, &mh);
 
3846	if (error)
3847		goto out;
3848
3849	ms->m_lkid = ms_in->m_lkid;
3850	ms->m_result = rv;
3851	ms->m_nodeid = ret_nodeid;
3852
3853	error = send_message(mh, ms);
3854 out:
3855	return error;
3856}
3857
3858/* which args we save from a received message depends heavily on the type
3859   of message, unlike the send side where we can safely send everything about
3860   the lkb for any type of message */
3861
3862static void receive_flags(struct dlm_lkb *lkb, struct dlm_message *ms)
3863{
3864	lkb->lkb_exflags = ms->m_exflags;
3865	lkb->lkb_sbflags = ms->m_sbflags;
3866	lkb->lkb_flags = (lkb->lkb_flags & 0xFFFF0000) |
3867		         (ms->m_flags & 0x0000FFFF);
3868}
3869
3870static void receive_flags_reply(struct dlm_lkb *lkb, struct dlm_message *ms)
 
 
3871{
3872	if (ms->m_flags == DLM_IFL_STUB_MS)
3873		return;
3874
3875	lkb->lkb_sbflags = ms->m_sbflags;
3876	lkb->lkb_flags = (lkb->lkb_flags & 0xFFFF0000) |
3877		         (ms->m_flags & 0x0000FFFF);
3878}
3879
3880static int receive_extralen(struct dlm_message *ms)
3881{
3882	return (ms->m_header.h_length - sizeof(struct dlm_message));
 
3883}
3884
3885static int receive_lvb(struct dlm_ls *ls, struct dlm_lkb *lkb,
3886		       struct dlm_message *ms)
3887{
3888	int len;
3889
3890	if (lkb->lkb_exflags & DLM_LKF_VALBLK) {
3891		if (!lkb->lkb_lvbptr)
3892			lkb->lkb_lvbptr = dlm_allocate_lvb(ls);
3893		if (!lkb->lkb_lvbptr)
3894			return -ENOMEM;
3895		len = receive_extralen(ms);
3896		if (len > ls->ls_lvblen)
3897			len = ls->ls_lvblen;
3898		memcpy(lkb->lkb_lvbptr, ms->m_extra, len);
3899	}
3900	return 0;
3901}
3902
3903static void fake_bastfn(void *astparam, int mode)
3904{
3905	log_print("fake_bastfn should not be called");
3906}
3907
3908static void fake_astfn(void *astparam)
3909{
3910	log_print("fake_astfn should not be called");
3911}
3912
3913static int receive_request_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
3914				struct dlm_message *ms)
3915{
3916	lkb->lkb_nodeid = ms->m_header.h_nodeid;
3917	lkb->lkb_ownpid = ms->m_pid;
3918	lkb->lkb_remid = ms->m_lkid;
3919	lkb->lkb_grmode = DLM_LOCK_IV;
3920	lkb->lkb_rqmode = ms->m_rqmode;
3921
3922	lkb->lkb_bastfn = (ms->m_asts & DLM_CB_BAST) ? &fake_bastfn : NULL;
3923	lkb->lkb_astfn = (ms->m_asts & DLM_CB_CAST) ? &fake_astfn : NULL;
3924
3925	if (lkb->lkb_exflags & DLM_LKF_VALBLK) {
3926		/* lkb was just created so there won't be an lvb yet */
3927		lkb->lkb_lvbptr = dlm_allocate_lvb(ls);
3928		if (!lkb->lkb_lvbptr)
3929			return -ENOMEM;
3930	}
3931
3932	return 0;
3933}
3934
3935static int receive_convert_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
3936				struct dlm_message *ms)
3937{
3938	if (lkb->lkb_status != DLM_LKSTS_GRANTED)
3939		return -EBUSY;
3940
3941	if (receive_lvb(ls, lkb, ms))
3942		return -ENOMEM;
3943
3944	lkb->lkb_rqmode = ms->m_rqmode;
3945	lkb->lkb_lvbseq = ms->m_lvbseq;
3946
3947	return 0;
3948}
3949
3950static int receive_unlock_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
3951			       struct dlm_message *ms)
3952{
3953	if (receive_lvb(ls, lkb, ms))
3954		return -ENOMEM;
3955	return 0;
3956}
3957
3958/* We fill in the stub-lkb fields with the info that send_xxxx_reply()
3959   uses to send a reply and that the remote end uses to process the reply. */
3960
3961static void setup_stub_lkb(struct dlm_ls *ls, struct dlm_message *ms)
3962{
3963	struct dlm_lkb *lkb = &ls->ls_stub_lkb;
3964	lkb->lkb_nodeid = ms->m_header.h_nodeid;
3965	lkb->lkb_remid = ms->m_lkid;
3966}
3967
3968/* This is called after the rsb is locked so that we can safely inspect
3969   fields in the lkb. */
3970
3971static int validate_message(struct dlm_lkb *lkb, struct dlm_message *ms)
3972{
3973	int from = ms->m_header.h_nodeid;
3974	int error = 0;
3975
 
 
 
 
 
 
 
 
 
3976	switch (ms->m_type) {
3977	case DLM_MSG_CONVERT:
3978	case DLM_MSG_UNLOCK:
3979	case DLM_MSG_CANCEL:
3980		if (!is_master_copy(lkb) || lkb->lkb_nodeid != from)
3981			error = -EINVAL;
3982		break;
3983
3984	case DLM_MSG_CONVERT_REPLY:
3985	case DLM_MSG_UNLOCK_REPLY:
3986	case DLM_MSG_CANCEL_REPLY:
3987	case DLM_MSG_GRANT:
3988	case DLM_MSG_BAST:
3989		if (!is_process_copy(lkb) || lkb->lkb_nodeid != from)
3990			error = -EINVAL;
3991		break;
3992
3993	case DLM_MSG_REQUEST_REPLY:
3994		if (!is_process_copy(lkb))
3995			error = -EINVAL;
3996		else if (lkb->lkb_nodeid != -1 && lkb->lkb_nodeid != from)
3997			error = -EINVAL;
3998		break;
3999
4000	default:
4001		error = -EINVAL;
4002	}
4003
 
4004	if (error)
4005		log_error(lkb->lkb_resource->res_ls,
4006			  "ignore invalid message %d from %d %x %x %x %d",
4007			  ms->m_type, from, lkb->lkb_id, lkb->lkb_remid,
4008			  lkb->lkb_flags, lkb->lkb_nodeid);
 
4009	return error;
4010}
4011
4012static void send_repeat_remove(struct dlm_ls *ls, char *ms_name, int len)
4013{
4014	char name[DLM_RESNAME_MAXLEN + 1];
4015	struct dlm_message *ms;
4016	struct dlm_mhandle *mh;
4017	struct dlm_rsb *r;
4018	uint32_t hash, b;
4019	int rv, dir_nodeid;
4020
4021	memset(name, 0, sizeof(name));
4022	memcpy(name, ms_name, len);
4023
4024	hash = jhash(name, len, 0);
4025	b = hash & (ls->ls_rsbtbl_size - 1);
4026
4027	dir_nodeid = dlm_hash2nodeid(ls, hash);
4028
4029	log_error(ls, "send_repeat_remove dir %d %s", dir_nodeid, name);
4030
4031	spin_lock(&ls->ls_rsbtbl[b].lock);
4032	rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
4033	if (!rv) {
4034		spin_unlock(&ls->ls_rsbtbl[b].lock);
4035		log_error(ls, "repeat_remove on keep %s", name);
4036		return;
4037	}
4038
4039	rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
4040	if (!rv) {
4041		spin_unlock(&ls->ls_rsbtbl[b].lock);
4042		log_error(ls, "repeat_remove on toss %s", name);
4043		return;
4044	}
4045
4046	/* use ls->remove_name2 to avoid conflict with shrink? */
4047
4048	spin_lock(&ls->ls_remove_spin);
4049	ls->ls_remove_len = len;
4050	memcpy(ls->ls_remove_name, name, DLM_RESNAME_MAXLEN);
4051	spin_unlock(&ls->ls_remove_spin);
4052	spin_unlock(&ls->ls_rsbtbl[b].lock);
4053
4054	rv = _create_message(ls, sizeof(struct dlm_message) + len,
4055			     dir_nodeid, DLM_MSG_REMOVE, &ms, &mh);
4056	if (rv)
4057		return;
4058
4059	memcpy(ms->m_extra, name, len);
4060	ms->m_hash = hash;
4061
4062	send_message(mh, ms);
4063
4064	spin_lock(&ls->ls_remove_spin);
4065	ls->ls_remove_len = 0;
4066	memset(ls->ls_remove_name, 0, DLM_RESNAME_MAXLEN);
4067	spin_unlock(&ls->ls_remove_spin);
4068}
4069
4070static int receive_request(struct dlm_ls *ls, struct dlm_message *ms)
4071{
4072	struct dlm_lkb *lkb;
4073	struct dlm_rsb *r;
4074	int from_nodeid;
4075	int error, namelen = 0;
4076
4077	from_nodeid = ms->m_header.h_nodeid;
4078
4079	error = create_lkb(ls, &lkb);
4080	if (error)
4081		goto fail;
4082
4083	receive_flags(lkb, ms);
4084	lkb->lkb_flags |= DLM_IFL_MSTCPY;
4085	error = receive_request_args(ls, lkb, ms);
4086	if (error) {
4087		__put_lkb(ls, lkb);
4088		goto fail;
4089	}
4090
4091	/* The dir node is the authority on whether we are the master
4092	   for this rsb or not, so if the master sends us a request, we should
4093	   recreate the rsb if we've destroyed it.   This race happens when we
4094	   send a remove message to the dir node at the same time that the dir
4095	   node sends us a request for the rsb. */
4096
4097	namelen = receive_extralen(ms);
4098
4099	error = find_rsb(ls, ms->m_extra, namelen, from_nodeid,
4100			 R_RECEIVE_REQUEST, &r);
4101	if (error) {
4102		__put_lkb(ls, lkb);
4103		goto fail;
4104	}
4105
4106	lock_rsb(r);
4107
4108	if (r->res_master_nodeid != dlm_our_nodeid()) {
4109		error = validate_master_nodeid(ls, r, from_nodeid);
4110		if (error) {
4111			unlock_rsb(r);
4112			put_rsb(r);
4113			__put_lkb(ls, lkb);
4114			goto fail;
4115		}
4116	}
4117
4118	attach_lkb(r, lkb);
4119	error = do_request(r, lkb);
4120	send_request_reply(r, lkb, error);
4121	do_request_effects(r, lkb, error);
4122
4123	unlock_rsb(r);
4124	put_rsb(r);
4125
4126	if (error == -EINPROGRESS)
4127		error = 0;
4128	if (error)
4129		dlm_put_lkb(lkb);
4130	return 0;
4131
4132 fail:
4133	/* TODO: instead of returning ENOTBLK, add the lkb to res_lookup
4134	   and do this receive_request again from process_lookup_list once
4135	   we get the lookup reply.  This would avoid a many repeated
4136	   ENOTBLK request failures when the lookup reply designating us
4137	   as master is delayed. */
4138
4139	/* We could repeatedly return -EBADR here if our send_remove() is
4140	   delayed in being sent/arriving/being processed on the dir node.
4141	   Another node would repeatedly lookup up the master, and the dir
4142	   node would continue returning our nodeid until our send_remove
4143	   took effect.
4144
4145	   We send another remove message in case our previous send_remove
4146	   was lost/ignored/missed somehow. */
4147
4148	if (error != -ENOTBLK) {
4149		log_limit(ls, "receive_request %x from %d %d",
4150			  ms->m_lkid, from_nodeid, error);
4151	}
4152
4153	if (namelen && error == -EBADR) {
4154		send_repeat_remove(ls, ms->m_extra, namelen);
4155		msleep(1000);
4156	}
4157
4158	setup_stub_lkb(ls, ms);
4159	send_request_reply(&ls->ls_stub_rsb, &ls->ls_stub_lkb, error);
4160	return error;
4161}
4162
4163static int receive_convert(struct dlm_ls *ls, struct dlm_message *ms)
4164{
4165	struct dlm_lkb *lkb;
4166	struct dlm_rsb *r;
4167	int error, reply = 1;
4168
4169	error = find_lkb(ls, ms->m_remid, &lkb);
4170	if (error)
4171		goto fail;
4172
4173	if (lkb->lkb_remid != ms->m_lkid) {
4174		log_error(ls, "receive_convert %x remid %x recover_seq %llu "
4175			  "remote %d %x", lkb->lkb_id, lkb->lkb_remid,
4176			  (unsigned long long)lkb->lkb_recover_seq,
4177			  ms->m_header.h_nodeid, ms->m_lkid);
 
4178		error = -ENOENT;
4179		dlm_put_lkb(lkb);
4180		goto fail;
4181	}
4182
4183	r = lkb->lkb_resource;
4184
4185	hold_rsb(r);
4186	lock_rsb(r);
4187
4188	error = validate_message(lkb, ms);
4189	if (error)
4190		goto out;
4191
4192	receive_flags(lkb, ms);
4193
4194	error = receive_convert_args(ls, lkb, ms);
4195	if (error) {
4196		send_convert_reply(r, lkb, error);
4197		goto out;
4198	}
4199
4200	reply = !down_conversion(lkb);
4201
4202	error = do_convert(r, lkb);
4203	if (reply)
4204		send_convert_reply(r, lkb, error);
4205	do_convert_effects(r, lkb, error);
4206 out:
4207	unlock_rsb(r);
4208	put_rsb(r);
4209	dlm_put_lkb(lkb);
4210	return 0;
4211
4212 fail:
4213	setup_stub_lkb(ls, ms);
4214	send_convert_reply(&ls->ls_stub_rsb, &ls->ls_stub_lkb, error);
4215	return error;
4216}
4217
4218static int receive_unlock(struct dlm_ls *ls, struct dlm_message *ms)
4219{
4220	struct dlm_lkb *lkb;
4221	struct dlm_rsb *r;
4222	int error;
4223
4224	error = find_lkb(ls, ms->m_remid, &lkb);
4225	if (error)
4226		goto fail;
4227
4228	if (lkb->lkb_remid != ms->m_lkid) {
4229		log_error(ls, "receive_unlock %x remid %x remote %d %x",
4230			  lkb->lkb_id, lkb->lkb_remid,
4231			  ms->m_header.h_nodeid, ms->m_lkid);
 
4232		error = -ENOENT;
4233		dlm_put_lkb(lkb);
4234		goto fail;
4235	}
4236
4237	r = lkb->lkb_resource;
4238
4239	hold_rsb(r);
4240	lock_rsb(r);
4241
4242	error = validate_message(lkb, ms);
4243	if (error)
4244		goto out;
4245
4246	receive_flags(lkb, ms);
4247
4248	error = receive_unlock_args(ls, lkb, ms);
4249	if (error) {
4250		send_unlock_reply(r, lkb, error);
4251		goto out;
4252	}
4253
4254	error = do_unlock(r, lkb);
4255	send_unlock_reply(r, lkb, error);
4256	do_unlock_effects(r, lkb, error);
4257 out:
4258	unlock_rsb(r);
4259	put_rsb(r);
4260	dlm_put_lkb(lkb);
4261	return 0;
4262
4263 fail:
4264	setup_stub_lkb(ls, ms);
4265	send_unlock_reply(&ls->ls_stub_rsb, &ls->ls_stub_lkb, error);
4266	return error;
4267}
4268
4269static int receive_cancel(struct dlm_ls *ls, struct dlm_message *ms)
4270{
4271	struct dlm_lkb *lkb;
4272	struct dlm_rsb *r;
4273	int error;
4274
4275	error = find_lkb(ls, ms->m_remid, &lkb);
4276	if (error)
4277		goto fail;
4278
4279	receive_flags(lkb, ms);
4280
4281	r = lkb->lkb_resource;
4282
4283	hold_rsb(r);
4284	lock_rsb(r);
4285
4286	error = validate_message(lkb, ms);
4287	if (error)
4288		goto out;
4289
4290	error = do_cancel(r, lkb);
4291	send_cancel_reply(r, lkb, error);
4292	do_cancel_effects(r, lkb, error);
4293 out:
4294	unlock_rsb(r);
4295	put_rsb(r);
4296	dlm_put_lkb(lkb);
4297	return 0;
4298
4299 fail:
4300	setup_stub_lkb(ls, ms);
4301	send_cancel_reply(&ls->ls_stub_rsb, &ls->ls_stub_lkb, error);
4302	return error;
4303}
4304
4305static int receive_grant(struct dlm_ls *ls, struct dlm_message *ms)
4306{
4307	struct dlm_lkb *lkb;
4308	struct dlm_rsb *r;
4309	int error;
4310
4311	error = find_lkb(ls, ms->m_remid, &lkb);
4312	if (error)
4313		return error;
4314
4315	r = lkb->lkb_resource;
4316
4317	hold_rsb(r);
4318	lock_rsb(r);
4319
4320	error = validate_message(lkb, ms);
4321	if (error)
4322		goto out;
4323
4324	receive_flags_reply(lkb, ms);
4325	if (is_altmode(lkb))
4326		munge_altmode(lkb, ms);
4327	grant_lock_pc(r, lkb, ms);
4328	queue_cast(r, lkb, 0);
4329 out:
4330	unlock_rsb(r);
4331	put_rsb(r);
4332	dlm_put_lkb(lkb);
4333	return 0;
4334}
4335
4336static int receive_bast(struct dlm_ls *ls, struct dlm_message *ms)
4337{
4338	struct dlm_lkb *lkb;
4339	struct dlm_rsb *r;
4340	int error;
4341
4342	error = find_lkb(ls, ms->m_remid, &lkb);
4343	if (error)
4344		return error;
4345
4346	r = lkb->lkb_resource;
4347
4348	hold_rsb(r);
4349	lock_rsb(r);
4350
4351	error = validate_message(lkb, ms);
4352	if (error)
4353		goto out;
4354
4355	queue_bast(r, lkb, ms->m_bastmode);
4356	lkb->lkb_highbast = ms->m_bastmode;
4357 out:
4358	unlock_rsb(r);
4359	put_rsb(r);
4360	dlm_put_lkb(lkb);
4361	return 0;
4362}
4363
4364static void receive_lookup(struct dlm_ls *ls, struct dlm_message *ms)
4365{
4366	int len, error, ret_nodeid, from_nodeid, our_nodeid;
4367
4368	from_nodeid = ms->m_header.h_nodeid;
4369	our_nodeid = dlm_our_nodeid();
4370
4371	len = receive_extralen(ms);
4372
4373	error = dlm_master_lookup(ls, from_nodeid, ms->m_extra, len, 0,
4374				  &ret_nodeid, NULL);
4375
4376	/* Optimization: we're master so treat lookup as a request */
4377	if (!error && ret_nodeid == our_nodeid) {
4378		receive_request(ls, ms);
4379		return;
4380	}
4381	send_lookup_reply(ls, ms, ret_nodeid, error);
4382}
4383
4384static void receive_remove(struct dlm_ls *ls, struct dlm_message *ms)
4385{
4386	char name[DLM_RESNAME_MAXLEN+1];
4387	struct dlm_rsb *r;
4388	uint32_t hash, b;
4389	int rv, len, dir_nodeid, from_nodeid;
4390
4391	from_nodeid = ms->m_header.h_nodeid;
4392
4393	len = receive_extralen(ms);
4394
4395	if (len > DLM_RESNAME_MAXLEN) {
4396		log_error(ls, "receive_remove from %d bad len %d",
4397			  from_nodeid, len);
4398		return;
4399	}
4400
4401	dir_nodeid = dlm_hash2nodeid(ls, ms->m_hash);
4402	if (dir_nodeid != dlm_our_nodeid()) {
4403		log_error(ls, "receive_remove from %d bad nodeid %d",
4404			  from_nodeid, dir_nodeid);
4405		return;
4406	}
4407
4408	/* Look for name on rsbtbl.toss, if it's there, kill it.
4409	   If it's on rsbtbl.keep, it's being used, and we should ignore this
4410	   message.  This is an expected race between the dir node sending a
4411	   request to the master node at the same time as the master node sends
4412	   a remove to the dir node.  The resolution to that race is for the
4413	   dir node to ignore the remove message, and the master node to
4414	   recreate the master rsb when it gets a request from the dir node for
4415	   an rsb it doesn't have. */
4416
4417	memset(name, 0, sizeof(name));
4418	memcpy(name, ms->m_extra, len);
4419
4420	hash = jhash(name, len, 0);
4421	b = hash & (ls->ls_rsbtbl_size - 1);
4422
4423	spin_lock(&ls->ls_rsbtbl[b].lock);
4424
4425	rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
4426	if (rv) {
4427		/* verify the rsb is on keep list per comment above */
4428		rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
4429		if (rv) {
4430			/* should not happen */
4431			log_error(ls, "receive_remove from %d not found %s",
4432				  from_nodeid, name);
4433			spin_unlock(&ls->ls_rsbtbl[b].lock);
4434			return;
4435		}
4436		if (r->res_master_nodeid != from_nodeid) {
4437			/* should not happen */
4438			log_error(ls, "receive_remove keep from %d master %d",
4439				  from_nodeid, r->res_master_nodeid);
4440			dlm_print_rsb(r);
4441			spin_unlock(&ls->ls_rsbtbl[b].lock);
4442			return;
4443		}
4444
4445		log_debug(ls, "receive_remove from %d master %d first %x %s",
4446			  from_nodeid, r->res_master_nodeid, r->res_first_lkid,
4447			  name);
4448		spin_unlock(&ls->ls_rsbtbl[b].lock);
4449		return;
4450	}
4451
4452	if (r->res_master_nodeid != from_nodeid) {
4453		log_error(ls, "receive_remove toss from %d master %d",
4454			  from_nodeid, r->res_master_nodeid);
4455		dlm_print_rsb(r);
4456		spin_unlock(&ls->ls_rsbtbl[b].lock);
4457		return;
4458	}
4459
4460	if (kref_put(&r->res_ref, kill_rsb)) {
4461		rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
4462		spin_unlock(&ls->ls_rsbtbl[b].lock);
4463		dlm_free_rsb(r);
4464	} else {
4465		log_error(ls, "receive_remove from %d rsb ref error",
4466			  from_nodeid);
4467		dlm_print_rsb(r);
4468		spin_unlock(&ls->ls_rsbtbl[b].lock);
4469	}
4470}
4471
4472static void receive_purge(struct dlm_ls *ls, struct dlm_message *ms)
4473{
4474	do_purge(ls, ms->m_nodeid, ms->m_pid);
4475}
4476
4477static int receive_request_reply(struct dlm_ls *ls, struct dlm_message *ms)
 
4478{
4479	struct dlm_lkb *lkb;
4480	struct dlm_rsb *r;
4481	int error, mstype, result;
4482	int from_nodeid = ms->m_header.h_nodeid;
4483
4484	error = find_lkb(ls, ms->m_remid, &lkb);
4485	if (error)
4486		return error;
4487
4488	r = lkb->lkb_resource;
4489	hold_rsb(r);
4490	lock_rsb(r);
4491
4492	error = validate_message(lkb, ms);
4493	if (error)
4494		goto out;
4495
4496	mstype = lkb->lkb_wait_type;
4497	error = remove_from_waiters(lkb, DLM_MSG_REQUEST_REPLY);
4498	if (error) {
4499		log_error(ls, "receive_request_reply %x remote %d %x result %d",
4500			  lkb->lkb_id, from_nodeid, ms->m_lkid, ms->m_result);
 
4501		dlm_dump_rsb(r);
4502		goto out;
4503	}
4504
4505	/* Optimization: the dir node was also the master, so it took our
4506	   lookup as a request and sent request reply instead of lookup reply */
4507	if (mstype == DLM_MSG_LOOKUP) {
4508		r->res_master_nodeid = from_nodeid;
4509		r->res_nodeid = from_nodeid;
4510		lkb->lkb_nodeid = from_nodeid;
4511	}
4512
4513	/* this is the value returned from do_request() on the master */
4514	result = ms->m_result;
4515
4516	switch (result) {
4517	case -EAGAIN:
4518		/* request would block (be queued) on remote master */
4519		queue_cast(r, lkb, -EAGAIN);
4520		confirm_master(r, -EAGAIN);
4521		unhold_lkb(lkb); /* undoes create_lkb() */
4522		break;
4523
4524	case -EINPROGRESS:
4525	case 0:
4526		/* request was queued or granted on remote master */
4527		receive_flags_reply(lkb, ms);
4528		lkb->lkb_remid = ms->m_lkid;
4529		if (is_altmode(lkb))
4530			munge_altmode(lkb, ms);
4531		if (result) {
4532			add_lkb(r, lkb, DLM_LKSTS_WAITING);
4533			add_timeout(lkb);
4534		} else {
4535			grant_lock_pc(r, lkb, ms);
4536			queue_cast(r, lkb, 0);
4537		}
4538		confirm_master(r, result);
4539		break;
4540
4541	case -EBADR:
4542	case -ENOTBLK:
4543		/* find_rsb failed to find rsb or rsb wasn't master */
4544		log_limit(ls, "receive_request_reply %x from %d %d "
4545			  "master %d dir %d first %x %s", lkb->lkb_id,
4546			  from_nodeid, result, r->res_master_nodeid,
4547			  r->res_dir_nodeid, r->res_first_lkid, r->res_name);
4548
4549		if (r->res_dir_nodeid != dlm_our_nodeid() &&
4550		    r->res_master_nodeid != dlm_our_nodeid()) {
4551			/* cause _request_lock->set_master->send_lookup */
4552			r->res_master_nodeid = 0;
4553			r->res_nodeid = -1;
4554			lkb->lkb_nodeid = -1;
4555		}
4556
4557		if (is_overlap(lkb)) {
4558			/* we'll ignore error in cancel/unlock reply */
4559			queue_cast_overlap(r, lkb);
4560			confirm_master(r, result);
4561			unhold_lkb(lkb); /* undoes create_lkb() */
4562		} else {
4563			_request_lock(r, lkb);
4564
4565			if (r->res_master_nodeid == dlm_our_nodeid())
4566				confirm_master(r, 0);
4567		}
4568		break;
4569
4570	default:
4571		log_error(ls, "receive_request_reply %x error %d",
4572			  lkb->lkb_id, result);
4573	}
4574
4575	if (is_overlap_unlock(lkb) && (result == 0 || result == -EINPROGRESS)) {
 
4576		log_debug(ls, "receive_request_reply %x result %d unlock",
4577			  lkb->lkb_id, result);
4578		lkb->lkb_flags &= ~DLM_IFL_OVERLAP_UNLOCK;
4579		lkb->lkb_flags &= ~DLM_IFL_OVERLAP_CANCEL;
4580		send_unlock(r, lkb);
4581	} else if (is_overlap_cancel(lkb) && (result == -EINPROGRESS)) {
 
 
4582		log_debug(ls, "receive_request_reply %x cancel", lkb->lkb_id);
4583		lkb->lkb_flags &= ~DLM_IFL_OVERLAP_UNLOCK;
4584		lkb->lkb_flags &= ~DLM_IFL_OVERLAP_CANCEL;
4585		send_cancel(r, lkb);
4586	} else {
4587		lkb->lkb_flags &= ~DLM_IFL_OVERLAP_CANCEL;
4588		lkb->lkb_flags &= ~DLM_IFL_OVERLAP_UNLOCK;
4589	}
4590 out:
4591	unlock_rsb(r);
4592	put_rsb(r);
4593	dlm_put_lkb(lkb);
4594	return 0;
4595}
4596
4597static void __receive_convert_reply(struct dlm_rsb *r, struct dlm_lkb *lkb,
4598				    struct dlm_message *ms)
4599{
4600	/* this is the value returned from do_convert() on the master */
4601	switch (ms->m_result) {
4602	case -EAGAIN:
4603		/* convert would block (be queued) on remote master */
4604		queue_cast(r, lkb, -EAGAIN);
4605		break;
4606
4607	case -EDEADLK:
4608		receive_flags_reply(lkb, ms);
4609		revert_lock_pc(r, lkb);
4610		queue_cast(r, lkb, -EDEADLK);
4611		break;
4612
4613	case -EINPROGRESS:
4614		/* convert was queued on remote master */
4615		receive_flags_reply(lkb, ms);
4616		if (is_demoted(lkb))
4617			munge_demoted(lkb);
4618		del_lkb(r, lkb);
4619		add_lkb(r, lkb, DLM_LKSTS_CONVERT);
4620		add_timeout(lkb);
4621		break;
4622
4623	case 0:
4624		/* convert was granted on remote master */
4625		receive_flags_reply(lkb, ms);
4626		if (is_demoted(lkb))
4627			munge_demoted(lkb);
4628		grant_lock_pc(r, lkb, ms);
4629		queue_cast(r, lkb, 0);
4630		break;
4631
4632	default:
4633		log_error(r->res_ls, "receive_convert_reply %x remote %d %x %d",
4634			  lkb->lkb_id, ms->m_header.h_nodeid, ms->m_lkid,
4635			  ms->m_result);
 
4636		dlm_print_rsb(r);
4637		dlm_print_lkb(lkb);
4638	}
4639}
4640
4641static void _receive_convert_reply(struct dlm_lkb *lkb, struct dlm_message *ms)
 
4642{
4643	struct dlm_rsb *r = lkb->lkb_resource;
4644	int error;
4645
4646	hold_rsb(r);
4647	lock_rsb(r);
4648
4649	error = validate_message(lkb, ms);
4650	if (error)
4651		goto out;
4652
4653	/* stub reply can happen with waiters_mutex held */
4654	error = remove_from_waiters_ms(lkb, ms);
4655	if (error)
4656		goto out;
4657
4658	__receive_convert_reply(r, lkb, ms);
4659 out:
4660	unlock_rsb(r);
4661	put_rsb(r);
4662}
4663
4664static int receive_convert_reply(struct dlm_ls *ls, struct dlm_message *ms)
 
4665{
4666	struct dlm_lkb *lkb;
4667	int error;
4668
4669	error = find_lkb(ls, ms->m_remid, &lkb);
4670	if (error)
4671		return error;
4672
4673	_receive_convert_reply(lkb, ms);
4674	dlm_put_lkb(lkb);
4675	return 0;
4676}
4677
4678static void _receive_unlock_reply(struct dlm_lkb *lkb, struct dlm_message *ms)
 
4679{
4680	struct dlm_rsb *r = lkb->lkb_resource;
4681	int error;
4682
4683	hold_rsb(r);
4684	lock_rsb(r);
4685
4686	error = validate_message(lkb, ms);
4687	if (error)
4688		goto out;
4689
4690	/* stub reply can happen with waiters_mutex held */
4691	error = remove_from_waiters_ms(lkb, ms);
4692	if (error)
4693		goto out;
4694
4695	/* this is the value returned from do_unlock() on the master */
4696
4697	switch (ms->m_result) {
4698	case -DLM_EUNLOCK:
4699		receive_flags_reply(lkb, ms);
4700		remove_lock_pc(r, lkb);
4701		queue_cast(r, lkb, -DLM_EUNLOCK);
4702		break;
4703	case -ENOENT:
4704		break;
4705	default:
4706		log_error(r->res_ls, "receive_unlock_reply %x error %d",
4707			  lkb->lkb_id, ms->m_result);
4708	}
4709 out:
4710	unlock_rsb(r);
4711	put_rsb(r);
4712}
4713
4714static int receive_unlock_reply(struct dlm_ls *ls, struct dlm_message *ms)
 
4715{
4716	struct dlm_lkb *lkb;
4717	int error;
4718
4719	error = find_lkb(ls, ms->m_remid, &lkb);
4720	if (error)
4721		return error;
4722
4723	_receive_unlock_reply(lkb, ms);
4724	dlm_put_lkb(lkb);
4725	return 0;
4726}
4727
4728static void _receive_cancel_reply(struct dlm_lkb *lkb, struct dlm_message *ms)
 
4729{
4730	struct dlm_rsb *r = lkb->lkb_resource;
4731	int error;
4732
4733	hold_rsb(r);
4734	lock_rsb(r);
4735
4736	error = validate_message(lkb, ms);
4737	if (error)
4738		goto out;
4739
4740	/* stub reply can happen with waiters_mutex held */
4741	error = remove_from_waiters_ms(lkb, ms);
4742	if (error)
4743		goto out;
4744
4745	/* this is the value returned from do_cancel() on the master */
4746
4747	switch (ms->m_result) {
4748	case -DLM_ECANCEL:
4749		receive_flags_reply(lkb, ms);
4750		revert_lock_pc(r, lkb);
4751		queue_cast(r, lkb, -DLM_ECANCEL);
4752		break;
4753	case 0:
4754		break;
4755	default:
4756		log_error(r->res_ls, "receive_cancel_reply %x error %d",
4757			  lkb->lkb_id, ms->m_result);
 
4758	}
4759 out:
4760	unlock_rsb(r);
4761	put_rsb(r);
4762}
4763
4764static int receive_cancel_reply(struct dlm_ls *ls, struct dlm_message *ms)
 
4765{
4766	struct dlm_lkb *lkb;
4767	int error;
4768
4769	error = find_lkb(ls, ms->m_remid, &lkb);
4770	if (error)
4771		return error;
4772
4773	_receive_cancel_reply(lkb, ms);
4774	dlm_put_lkb(lkb);
4775	return 0;
4776}
4777
4778static void receive_lookup_reply(struct dlm_ls *ls, struct dlm_message *ms)
 
4779{
4780	struct dlm_lkb *lkb;
4781	struct dlm_rsb *r;
4782	int error, ret_nodeid;
4783	int do_lookup_list = 0;
4784
4785	error = find_lkb(ls, ms->m_lkid, &lkb);
4786	if (error) {
4787		log_error(ls, "receive_lookup_reply no lkid %x", ms->m_lkid);
 
4788		return;
4789	}
4790
4791	/* ms->m_result is the value returned by dlm_master_lookup on dir node
4792	   FIXME: will a non-zero error ever be returned? */
4793
4794	r = lkb->lkb_resource;
4795	hold_rsb(r);
4796	lock_rsb(r);
4797
4798	error = remove_from_waiters(lkb, DLM_MSG_LOOKUP_REPLY);
4799	if (error)
4800		goto out;
4801
4802	ret_nodeid = ms->m_nodeid;
4803
4804	/* We sometimes receive a request from the dir node for this
4805	   rsb before we've received the dir node's loookup_reply for it.
4806	   The request from the dir node implies we're the master, so we set
4807	   ourself as master in receive_request_reply, and verify here that
4808	   we are indeed the master. */
4809
4810	if (r->res_master_nodeid && (r->res_master_nodeid != ret_nodeid)) {
4811		/* This should never happen */
4812		log_error(ls, "receive_lookup_reply %x from %d ret %d "
4813			  "master %d dir %d our %d first %x %s",
4814			  lkb->lkb_id, ms->m_header.h_nodeid, ret_nodeid,
4815			  r->res_master_nodeid, r->res_dir_nodeid,
4816			  dlm_our_nodeid(), r->res_first_lkid, r->res_name);
4817	}
4818
4819	if (ret_nodeid == dlm_our_nodeid()) {
4820		r->res_master_nodeid = ret_nodeid;
4821		r->res_nodeid = 0;
4822		do_lookup_list = 1;
4823		r->res_first_lkid = 0;
4824	} else if (ret_nodeid == -1) {
4825		/* the remote node doesn't believe it's the dir node */
4826		log_error(ls, "receive_lookup_reply %x from %d bad ret_nodeid",
4827			  lkb->lkb_id, ms->m_header.h_nodeid);
4828		r->res_master_nodeid = 0;
4829		r->res_nodeid = -1;
4830		lkb->lkb_nodeid = -1;
4831	} else {
4832		/* set_master() will set lkb_nodeid from r */
4833		r->res_master_nodeid = ret_nodeid;
4834		r->res_nodeid = ret_nodeid;
4835	}
4836
4837	if (is_overlap(lkb)) {
4838		log_debug(ls, "receive_lookup_reply %x unlock %x",
4839			  lkb->lkb_id, lkb->lkb_flags);
4840		queue_cast_overlap(r, lkb);
4841		unhold_lkb(lkb); /* undoes create_lkb() */
4842		goto out_list;
4843	}
4844
4845	_request_lock(r, lkb);
4846
4847 out_list:
4848	if (do_lookup_list)
4849		process_lookup_list(r);
4850 out:
4851	unlock_rsb(r);
4852	put_rsb(r);
4853	dlm_put_lkb(lkb);
4854}
4855
4856static void _receive_message(struct dlm_ls *ls, struct dlm_message *ms,
4857			     uint32_t saved_seq)
4858{
4859	int error = 0, noent = 0;
4860
4861	if (!dlm_is_member(ls, ms->m_header.h_nodeid)) {
4862		log_limit(ls, "receive %d from non-member %d %x %x %d",
4863			  ms->m_type, ms->m_header.h_nodeid, ms->m_lkid,
4864			  ms->m_remid, ms->m_result);
 
 
4865		return;
4866	}
4867
4868	switch (ms->m_type) {
4869
4870	/* messages sent to a master node */
4871
4872	case DLM_MSG_REQUEST:
4873		error = receive_request(ls, ms);
4874		break;
4875
4876	case DLM_MSG_CONVERT:
4877		error = receive_convert(ls, ms);
4878		break;
4879
4880	case DLM_MSG_UNLOCK:
4881		error = receive_unlock(ls, ms);
4882		break;
4883
4884	case DLM_MSG_CANCEL:
4885		noent = 1;
4886		error = receive_cancel(ls, ms);
4887		break;
4888
4889	/* messages sent from a master node (replies to above) */
4890
4891	case DLM_MSG_REQUEST_REPLY:
4892		error = receive_request_reply(ls, ms);
4893		break;
4894
4895	case DLM_MSG_CONVERT_REPLY:
4896		error = receive_convert_reply(ls, ms);
4897		break;
4898
4899	case DLM_MSG_UNLOCK_REPLY:
4900		error = receive_unlock_reply(ls, ms);
4901		break;
4902
4903	case DLM_MSG_CANCEL_REPLY:
4904		error = receive_cancel_reply(ls, ms);
4905		break;
4906
4907	/* messages sent from a master node (only two types of async msg) */
4908
4909	case DLM_MSG_GRANT:
4910		noent = 1;
4911		error = receive_grant(ls, ms);
4912		break;
4913
4914	case DLM_MSG_BAST:
4915		noent = 1;
4916		error = receive_bast(ls, ms);
4917		break;
4918
4919	/* messages sent to a dir node */
4920
4921	case DLM_MSG_LOOKUP:
4922		receive_lookup(ls, ms);
4923		break;
4924
4925	case DLM_MSG_REMOVE:
4926		receive_remove(ls, ms);
4927		break;
4928
4929	/* messages sent from a dir node (remove has no reply) */
4930
4931	case DLM_MSG_LOOKUP_REPLY:
4932		receive_lookup_reply(ls, ms);
4933		break;
4934
4935	/* other messages */
4936
4937	case DLM_MSG_PURGE:
4938		receive_purge(ls, ms);
4939		break;
4940
4941	default:
4942		log_error(ls, "unknown message type %d", ms->m_type);
 
4943	}
4944
4945	/*
4946	 * When checking for ENOENT, we're checking the result of
4947	 * find_lkb(m_remid):
4948	 *
4949	 * The lock id referenced in the message wasn't found.  This may
4950	 * happen in normal usage for the async messages and cancel, so
4951	 * only use log_debug for them.
4952	 *
4953	 * Some errors are expected and normal.
4954	 */
4955
4956	if (error == -ENOENT && noent) {
4957		log_debug(ls, "receive %d no %x remote %d %x saved_seq %u",
4958			  ms->m_type, ms->m_remid, ms->m_header.h_nodeid,
4959			  ms->m_lkid, saved_seq);
 
4960	} else if (error == -ENOENT) {
4961		log_error(ls, "receive %d no %x remote %d %x saved_seq %u",
4962			  ms->m_type, ms->m_remid, ms->m_header.h_nodeid,
4963			  ms->m_lkid, saved_seq);
 
4964
4965		if (ms->m_type == DLM_MSG_CONVERT)
4966			dlm_dump_rsb_hash(ls, ms->m_hash);
4967	}
4968
4969	if (error == -EINVAL) {
4970		log_error(ls, "receive %d inval from %d lkid %x remid %x "
4971			  "saved_seq %u",
4972			  ms->m_type, ms->m_header.h_nodeid,
4973			  ms->m_lkid, ms->m_remid, saved_seq);
 
 
4974	}
4975}
4976
4977/* If the lockspace is in recovery mode (locking stopped), then normal
4978   messages are saved on the requestqueue for processing after recovery is
4979   done.  When not in recovery mode, we wait for dlm_recoverd to drain saved
4980   messages off the requestqueue before we process new ones. This occurs right
4981   after recovery completes when we transition from saving all messages on
4982   requestqueue, to processing all the saved messages, to processing new
4983   messages as they arrive. */
4984
4985static void dlm_receive_message(struct dlm_ls *ls, struct dlm_message *ms,
4986				int nodeid)
4987{
4988	if (dlm_locking_stopped(ls)) {
4989		/* If we were a member of this lockspace, left, and rejoined,
4990		   other nodes may still be sending us messages from the
4991		   lockspace generation before we left. */
4992		if (!ls->ls_generation) {
4993			log_limit(ls, "receive %d from %d ignore old gen",
4994				  ms->m_type, nodeid);
4995			return;
4996		}
4997
4998		dlm_add_requestqueue(ls, nodeid, ms);
4999	} else {
5000		dlm_wait_requestqueue(ls);
5001		_receive_message(ls, ms, 0);
5002	}
5003}
5004
5005/* This is called by dlm_recoverd to process messages that were saved on
5006   the requestqueue. */
5007
5008void dlm_receive_message_saved(struct dlm_ls *ls, struct dlm_message *ms,
5009			       uint32_t saved_seq)
5010{
5011	_receive_message(ls, ms, saved_seq);
5012}
5013
5014/* This is called by the midcomms layer when something is received for
5015   the lockspace.  It could be either a MSG (normal message sent as part of
5016   standard locking activity) or an RCOM (recovery message sent as part of
5017   lockspace recovery). */
5018
5019void dlm_receive_buffer(union dlm_packet *p, int nodeid)
5020{
5021	struct dlm_header *hd = &p->header;
5022	struct dlm_ls *ls;
5023	int type = 0;
5024
5025	switch (hd->h_cmd) {
5026	case DLM_MSG:
5027		dlm_message_in(&p->message);
5028		type = p->message.m_type;
5029		break;
5030	case DLM_RCOM:
5031		dlm_rcom_in(&p->rcom);
5032		type = p->rcom.rc_type;
5033		break;
5034	default:
5035		log_print("invalid h_cmd %d from %u", hd->h_cmd, nodeid);
5036		return;
5037	}
5038
5039	if (hd->h_nodeid != nodeid) {
5040		log_print("invalid h_nodeid %d from %d lockspace %x",
5041			  hd->h_nodeid, nodeid, hd->u.h_lockspace);
 
5042		return;
5043	}
5044
5045	ls = dlm_find_lockspace_global(hd->u.h_lockspace);
5046	if (!ls) {
5047		if (dlm_config.ci_log_debug) {
5048			printk_ratelimited(KERN_DEBUG "dlm: invalid lockspace "
5049				"%u from %d cmd %d type %d\n",
5050				hd->u.h_lockspace, nodeid, hd->h_cmd, type);
 
5051		}
5052
5053		if (hd->h_cmd == DLM_RCOM && type == DLM_RCOM_STATUS)
5054			dlm_send_ls_not_ready(nodeid, &p->rcom);
5055		return;
5056	}
5057
5058	/* this rwsem allows dlm_ls_stop() to wait for all dlm_recv threads to
5059	   be inactive (in this ls) before transitioning to recovery mode */
5060
5061	down_read(&ls->ls_recv_active);
5062	if (hd->h_cmd == DLM_MSG)
5063		dlm_receive_message(ls, &p->message, nodeid);
5064	else
5065		dlm_receive_rcom(ls, &p->rcom, nodeid);
 
 
 
5066	up_read(&ls->ls_recv_active);
5067
5068	dlm_put_lockspace(ls);
5069}
5070
5071static void recover_convert_waiter(struct dlm_ls *ls, struct dlm_lkb *lkb,
5072				   struct dlm_message *ms_stub)
5073{
5074	if (middle_conversion(lkb)) {
5075		hold_lkb(lkb);
5076		memset(ms_stub, 0, sizeof(struct dlm_message));
5077		ms_stub->m_flags = DLM_IFL_STUB_MS;
5078		ms_stub->m_type = DLM_MSG_CONVERT_REPLY;
5079		ms_stub->m_result = -EINPROGRESS;
5080		ms_stub->m_header.h_nodeid = lkb->lkb_nodeid;
5081		_receive_convert_reply(lkb, ms_stub);
5082
5083		/* Same special case as in receive_rcom_lock_args() */
5084		lkb->lkb_grmode = DLM_LOCK_IV;
5085		rsb_set_flag(lkb->lkb_resource, RSB_RECOVER_CONVERT);
5086		unhold_lkb(lkb);
5087
5088	} else if (lkb->lkb_rqmode >= lkb->lkb_grmode) {
5089		lkb->lkb_flags |= DLM_IFL_RESEND;
5090	}
5091
5092	/* lkb->lkb_rqmode < lkb->lkb_grmode shouldn't happen since down
5093	   conversions are async; there's no reply from the remote master */
5094}
5095
5096/* A waiting lkb needs recovery if the master node has failed, or
5097   the master node is changing (only when no directory is used) */
5098
5099static int waiter_needs_recovery(struct dlm_ls *ls, struct dlm_lkb *lkb,
5100				 int dir_nodeid)
5101{
5102	if (dlm_no_directory(ls))
5103		return 1;
5104
5105	if (dlm_is_removed(ls, lkb->lkb_wait_nodeid))
5106		return 1;
5107
5108	return 0;
5109}
5110
5111/* Recovery for locks that are waiting for replies from nodes that are now
5112   gone.  We can just complete unlocks and cancels by faking a reply from the
5113   dead node.  Requests and up-conversions we flag to be resent after
5114   recovery.  Down-conversions can just be completed with a fake reply like
5115   unlocks.  Conversions between PR and CW need special attention. */
5116
5117void dlm_recover_waiters_pre(struct dlm_ls *ls)
5118{
5119	struct dlm_lkb *lkb, *safe;
5120	struct dlm_message *ms_stub;
5121	int wait_type, stub_unlock_result, stub_cancel_result;
5122	int dir_nodeid;
5123
5124	ms_stub = kmalloc(sizeof(*ms_stub), GFP_KERNEL);
5125	if (!ms_stub)
5126		return;
5127
5128	mutex_lock(&ls->ls_waiters_mutex);
5129
5130	list_for_each_entry_safe(lkb, safe, &ls->ls_waiters, lkb_wait_reply) {
5131
5132		dir_nodeid = dlm_dir_nodeid(lkb->lkb_resource);
5133
5134		/* exclude debug messages about unlocks because there can be so
5135		   many and they aren't very interesting */
5136
5137		if (lkb->lkb_wait_type != DLM_MSG_UNLOCK) {
5138			log_debug(ls, "waiter %x remote %x msg %d r_nodeid %d "
5139				  "lkb_nodeid %d wait_nodeid %d dir_nodeid %d",
5140				  lkb->lkb_id,
5141				  lkb->lkb_remid,
5142				  lkb->lkb_wait_type,
5143				  lkb->lkb_resource->res_nodeid,
5144				  lkb->lkb_nodeid,
5145				  lkb->lkb_wait_nodeid,
5146				  dir_nodeid);
5147		}
5148
5149		/* all outstanding lookups, regardless of destination  will be
5150		   resent after recovery is done */
5151
5152		if (lkb->lkb_wait_type == DLM_MSG_LOOKUP) {
5153			lkb->lkb_flags |= DLM_IFL_RESEND;
5154			continue;
5155		}
5156
5157		if (!waiter_needs_recovery(ls, lkb, dir_nodeid))
5158			continue;
5159
5160		wait_type = lkb->lkb_wait_type;
5161		stub_unlock_result = -DLM_EUNLOCK;
5162		stub_cancel_result = -DLM_ECANCEL;
5163
5164		/* Main reply may have been received leaving a zero wait_type,
5165		   but a reply for the overlapping op may not have been
5166		   received.  In that case we need to fake the appropriate
5167		   reply for the overlap op. */
5168
5169		if (!wait_type) {
5170			if (is_overlap_cancel(lkb)) {
5171				wait_type = DLM_MSG_CANCEL;
5172				if (lkb->lkb_grmode == DLM_LOCK_IV)
5173					stub_cancel_result = 0;
5174			}
5175			if (is_overlap_unlock(lkb)) {
5176				wait_type = DLM_MSG_UNLOCK;
5177				if (lkb->lkb_grmode == DLM_LOCK_IV)
5178					stub_unlock_result = -ENOENT;
5179			}
5180
5181			log_debug(ls, "rwpre overlap %x %x %d %d %d",
5182				  lkb->lkb_id, lkb->lkb_flags, wait_type,
5183				  stub_cancel_result, stub_unlock_result);
5184		}
5185
5186		switch (wait_type) {
5187
5188		case DLM_MSG_REQUEST:
5189			lkb->lkb_flags |= DLM_IFL_RESEND;
5190			break;
5191
5192		case DLM_MSG_CONVERT:
5193			recover_convert_waiter(ls, lkb, ms_stub);
5194			break;
5195
5196		case DLM_MSG_UNLOCK:
5197			hold_lkb(lkb);
5198			memset(ms_stub, 0, sizeof(struct dlm_message));
5199			ms_stub->m_flags = DLM_IFL_STUB_MS;
5200			ms_stub->m_type = DLM_MSG_UNLOCK_REPLY;
5201			ms_stub->m_result = stub_unlock_result;
5202			ms_stub->m_header.h_nodeid = lkb->lkb_nodeid;
5203			_receive_unlock_reply(lkb, ms_stub);
5204			dlm_put_lkb(lkb);
5205			break;
5206
5207		case DLM_MSG_CANCEL:
5208			hold_lkb(lkb);
5209			memset(ms_stub, 0, sizeof(struct dlm_message));
5210			ms_stub->m_flags = DLM_IFL_STUB_MS;
5211			ms_stub->m_type = DLM_MSG_CANCEL_REPLY;
5212			ms_stub->m_result = stub_cancel_result;
5213			ms_stub->m_header.h_nodeid = lkb->lkb_nodeid;
5214			_receive_cancel_reply(lkb, ms_stub);
5215			dlm_put_lkb(lkb);
5216			break;
5217
5218		default:
5219			log_error(ls, "invalid lkb wait_type %d %d",
5220				  lkb->lkb_wait_type, wait_type);
5221		}
5222		schedule();
5223	}
5224	mutex_unlock(&ls->ls_waiters_mutex);
5225	kfree(ms_stub);
5226}
5227
5228static struct dlm_lkb *find_resend_waiter(struct dlm_ls *ls)
5229{
5230	struct dlm_lkb *lkb;
5231	int found = 0;
5232
5233	mutex_lock(&ls->ls_waiters_mutex);
5234	list_for_each_entry(lkb, &ls->ls_waiters, lkb_wait_reply) {
5235		if (lkb->lkb_flags & DLM_IFL_RESEND) {
5236			hold_lkb(lkb);
5237			found = 1;
5238			break;
5239		}
5240	}
5241	mutex_unlock(&ls->ls_waiters_mutex);
5242
5243	if (!found)
5244		lkb = NULL;
5245	return lkb;
5246}
5247
5248/* Deal with lookups and lkb's marked RESEND from _pre.  We may now be the
5249   master or dir-node for r.  Processing the lkb may result in it being placed
5250   back on waiters. */
5251
5252/* We do this after normal locking has been enabled and any saved messages
5253   (in requestqueue) have been processed.  We should be confident that at
5254   this point we won't get or process a reply to any of these waiting
5255   operations.  But, new ops may be coming in on the rsbs/locks here from
5256   userspace or remotely. */
5257
5258/* there may have been an overlap unlock/cancel prior to recovery or after
5259   recovery.  if before, the lkb may still have a pos wait_count; if after, the
5260   overlap flag would just have been set and nothing new sent.  we can be
5261   confident here than any replies to either the initial op or overlap ops
5262   prior to recovery have been received. */
5263
5264int dlm_recover_waiters_post(struct dlm_ls *ls)
5265{
5266	struct dlm_lkb *lkb;
5267	struct dlm_rsb *r;
5268	int error = 0, mstype, err, oc, ou;
5269
5270	while (1) {
5271		if (dlm_locking_stopped(ls)) {
5272			log_debug(ls, "recover_waiters_post aborted");
5273			error = -EINTR;
5274			break;
5275		}
5276
5277		lkb = find_resend_waiter(ls);
5278		if (!lkb)
5279			break;
5280
5281		r = lkb->lkb_resource;
5282		hold_rsb(r);
5283		lock_rsb(r);
5284
5285		mstype = lkb->lkb_wait_type;
5286		oc = is_overlap_cancel(lkb);
5287		ou = is_overlap_unlock(lkb);
 
 
5288		err = 0;
5289
5290		log_debug(ls, "waiter %x remote %x msg %d r_nodeid %d "
5291			  "lkb_nodeid %d wait_nodeid %d dir_nodeid %d "
5292			  "overlap %d %d", lkb->lkb_id, lkb->lkb_remid, mstype,
5293			  r->res_nodeid, lkb->lkb_nodeid, lkb->lkb_wait_nodeid,
5294			  dlm_dir_nodeid(r), oc, ou);
5295
5296		/* At this point we assume that we won't get a reply to any
5297		   previous op or overlap op on this lock.  First, do a big
5298		   remove_from_waiters() for all previous ops. */
5299
5300		lkb->lkb_flags &= ~DLM_IFL_RESEND;
5301		lkb->lkb_flags &= ~DLM_IFL_OVERLAP_UNLOCK;
5302		lkb->lkb_flags &= ~DLM_IFL_OVERLAP_CANCEL;
5303		lkb->lkb_wait_type = 0;
5304		lkb->lkb_wait_count = 0;
 
 
 
 
 
5305		mutex_lock(&ls->ls_waiters_mutex);
5306		list_del_init(&lkb->lkb_wait_reply);
5307		mutex_unlock(&ls->ls_waiters_mutex);
5308		unhold_lkb(lkb); /* for waiters list */
5309
5310		if (oc || ou) {
5311			/* do an unlock or cancel instead of resending */
5312			switch (mstype) {
5313			case DLM_MSG_LOOKUP:
5314			case DLM_MSG_REQUEST:
5315				queue_cast(r, lkb, ou ? -DLM_EUNLOCK :
5316							-DLM_ECANCEL);
5317				unhold_lkb(lkb); /* undoes create_lkb() */
5318				break;
5319			case DLM_MSG_CONVERT:
5320				if (oc) {
5321					queue_cast(r, lkb, -DLM_ECANCEL);
5322				} else {
5323					lkb->lkb_exflags |= DLM_LKF_FORCEUNLOCK;
5324					_unlock_lock(r, lkb);
5325				}
5326				break;
5327			default:
5328				err = 1;
5329			}
5330		} else {
5331			switch (mstype) {
5332			case DLM_MSG_LOOKUP:
5333			case DLM_MSG_REQUEST:
5334				_request_lock(r, lkb);
5335				if (is_master(r))
5336					confirm_master(r, 0);
5337				break;
5338			case DLM_MSG_CONVERT:
5339				_convert_lock(r, lkb);
5340				break;
5341			default:
5342				err = 1;
5343			}
5344		}
5345
5346		if (err) {
5347			log_error(ls, "waiter %x msg %d r_nodeid %d "
5348				  "dir_nodeid %d overlap %d %d",
5349				  lkb->lkb_id, mstype, r->res_nodeid,
5350				  dlm_dir_nodeid(r), oc, ou);
5351		}
5352		unlock_rsb(r);
5353		put_rsb(r);
5354		dlm_put_lkb(lkb);
5355	}
5356
5357	return error;
5358}
5359
5360static void purge_mstcpy_list(struct dlm_ls *ls, struct dlm_rsb *r,
5361			      struct list_head *list)
5362{
5363	struct dlm_lkb *lkb, *safe;
5364
5365	list_for_each_entry_safe(lkb, safe, list, lkb_statequeue) {
5366		if (!is_master_copy(lkb))
5367			continue;
5368
5369		/* don't purge lkbs we've added in recover_master_copy for
5370		   the current recovery seq */
5371
5372		if (lkb->lkb_recover_seq == ls->ls_recover_seq)
5373			continue;
5374
5375		del_lkb(r, lkb);
5376
5377		/* this put should free the lkb */
5378		if (!dlm_put_lkb(lkb))
5379			log_error(ls, "purged mstcpy lkb not released");
5380	}
5381}
5382
5383void dlm_purge_mstcpy_locks(struct dlm_rsb *r)
5384{
5385	struct dlm_ls *ls = r->res_ls;
5386
5387	purge_mstcpy_list(ls, r, &r->res_grantqueue);
5388	purge_mstcpy_list(ls, r, &r->res_convertqueue);
5389	purge_mstcpy_list(ls, r, &r->res_waitqueue);
5390}
5391
5392static void purge_dead_list(struct dlm_ls *ls, struct dlm_rsb *r,
5393			    struct list_head *list,
5394			    int nodeid_gone, unsigned int *count)
5395{
5396	struct dlm_lkb *lkb, *safe;
5397
5398	list_for_each_entry_safe(lkb, safe, list, lkb_statequeue) {
5399		if (!is_master_copy(lkb))
5400			continue;
5401
5402		if ((lkb->lkb_nodeid == nodeid_gone) ||
5403		    dlm_is_removed(ls, lkb->lkb_nodeid)) {
5404
5405			/* tell recover_lvb to invalidate the lvb
5406			   because a node holding EX/PW failed */
5407			if ((lkb->lkb_exflags & DLM_LKF_VALBLK) &&
5408			    (lkb->lkb_grmode >= DLM_LOCK_PW)) {
5409				rsb_set_flag(r, RSB_RECOVER_LVB_INVAL);
5410			}
5411
5412			del_lkb(r, lkb);
5413
5414			/* this put should free the lkb */
5415			if (!dlm_put_lkb(lkb))
5416				log_error(ls, "purged dead lkb not released");
5417
5418			rsb_set_flag(r, RSB_RECOVER_GRANT);
5419
5420			(*count)++;
5421		}
5422	}
5423}
5424
5425/* Get rid of locks held by nodes that are gone. */
5426
5427void dlm_recover_purge(struct dlm_ls *ls)
5428{
5429	struct dlm_rsb *r;
5430	struct dlm_member *memb;
5431	int nodes_count = 0;
5432	int nodeid_gone = 0;
5433	unsigned int lkb_count = 0;
5434
5435	/* cache one removed nodeid to optimize the common
5436	   case of a single node removed */
5437
5438	list_for_each_entry(memb, &ls->ls_nodes_gone, list) {
5439		nodes_count++;
5440		nodeid_gone = memb->nodeid;
5441	}
5442
5443	if (!nodes_count)
5444		return;
5445
5446	down_write(&ls->ls_root_sem);
5447	list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
5448		hold_rsb(r);
5449		lock_rsb(r);
5450		if (is_master(r)) {
5451			purge_dead_list(ls, r, &r->res_grantqueue,
5452					nodeid_gone, &lkb_count);
5453			purge_dead_list(ls, r, &r->res_convertqueue,
5454					nodeid_gone, &lkb_count);
5455			purge_dead_list(ls, r, &r->res_waitqueue,
5456					nodeid_gone, &lkb_count);
5457		}
5458		unlock_rsb(r);
5459		unhold_rsb(r);
5460		cond_resched();
5461	}
5462	up_write(&ls->ls_root_sem);
5463
5464	if (lkb_count)
5465		log_rinfo(ls, "dlm_recover_purge %u locks for %u nodes",
5466			  lkb_count, nodes_count);
5467}
5468
5469static struct dlm_rsb *find_grant_rsb(struct dlm_ls *ls, int bucket)
5470{
5471	struct rb_node *n;
5472	struct dlm_rsb *r;
5473
5474	spin_lock(&ls->ls_rsbtbl[bucket].lock);
5475	for (n = rb_first(&ls->ls_rsbtbl[bucket].keep); n; n = rb_next(n)) {
5476		r = rb_entry(n, struct dlm_rsb, res_hashnode);
5477
5478		if (!rsb_flag(r, RSB_RECOVER_GRANT))
5479			continue;
5480		if (!is_master(r)) {
5481			rsb_clear_flag(r, RSB_RECOVER_GRANT);
5482			continue;
5483		}
5484		hold_rsb(r);
5485		spin_unlock(&ls->ls_rsbtbl[bucket].lock);
5486		return r;
5487	}
5488	spin_unlock(&ls->ls_rsbtbl[bucket].lock);
5489	return NULL;
5490}
5491
5492/*
5493 * Attempt to grant locks on resources that we are the master of.
5494 * Locks may have become grantable during recovery because locks
5495 * from departed nodes have been purged (or not rebuilt), allowing
5496 * previously blocked locks to now be granted.  The subset of rsb's
5497 * we are interested in are those with lkb's on either the convert or
5498 * waiting queues.
5499 *
5500 * Simplest would be to go through each master rsb and check for non-empty
5501 * convert or waiting queues, and attempt to grant on those rsbs.
5502 * Checking the queues requires lock_rsb, though, for which we'd need
5503 * to release the rsbtbl lock.  This would make iterating through all
5504 * rsb's very inefficient.  So, we rely on earlier recovery routines
5505 * to set RECOVER_GRANT on any rsb's that we should attempt to grant
5506 * locks for.
5507 */
5508
5509void dlm_recover_grant(struct dlm_ls *ls)
5510{
5511	struct dlm_rsb *r;
5512	int bucket = 0;
5513	unsigned int count = 0;
5514	unsigned int rsb_count = 0;
5515	unsigned int lkb_count = 0;
5516
5517	while (1) {
5518		r = find_grant_rsb(ls, bucket);
5519		if (!r) {
5520			if (bucket == ls->ls_rsbtbl_size - 1)
5521				break;
5522			bucket++;
5523			continue;
5524		}
5525		rsb_count++;
5526		count = 0;
5527		lock_rsb(r);
5528		/* the RECOVER_GRANT flag is checked in the grant path */
5529		grant_pending_locks(r, &count);
5530		rsb_clear_flag(r, RSB_RECOVER_GRANT);
5531		lkb_count += count;
5532		confirm_master(r, 0);
5533		unlock_rsb(r);
5534		put_rsb(r);
5535		cond_resched();
5536	}
5537
5538	if (lkb_count)
5539		log_rinfo(ls, "dlm_recover_grant %u locks on %u resources",
5540			  lkb_count, rsb_count);
5541}
5542
5543static struct dlm_lkb *search_remid_list(struct list_head *head, int nodeid,
5544					 uint32_t remid)
5545{
5546	struct dlm_lkb *lkb;
5547
5548	list_for_each_entry(lkb, head, lkb_statequeue) {
5549		if (lkb->lkb_nodeid == nodeid && lkb->lkb_remid == remid)
5550			return lkb;
5551	}
5552	return NULL;
5553}
5554
5555static struct dlm_lkb *search_remid(struct dlm_rsb *r, int nodeid,
5556				    uint32_t remid)
5557{
5558	struct dlm_lkb *lkb;
5559
5560	lkb = search_remid_list(&r->res_grantqueue, nodeid, remid);
5561	if (lkb)
5562		return lkb;
5563	lkb = search_remid_list(&r->res_convertqueue, nodeid, remid);
5564	if (lkb)
5565		return lkb;
5566	lkb = search_remid_list(&r->res_waitqueue, nodeid, remid);
5567	if (lkb)
5568		return lkb;
5569	return NULL;
5570}
5571
5572/* needs at least dlm_rcom + rcom_lock */
5573static int receive_rcom_lock_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
5574				  struct dlm_rsb *r, struct dlm_rcom *rc)
5575{
5576	struct rcom_lock *rl = (struct rcom_lock *) rc->rc_buf;
5577
5578	lkb->lkb_nodeid = rc->rc_header.h_nodeid;
5579	lkb->lkb_ownpid = le32_to_cpu(rl->rl_ownpid);
5580	lkb->lkb_remid = le32_to_cpu(rl->rl_lkid);
5581	lkb->lkb_exflags = le32_to_cpu(rl->rl_exflags);
5582	lkb->lkb_flags = le32_to_cpu(rl->rl_flags) & 0x0000FFFF;
5583	lkb->lkb_flags |= DLM_IFL_MSTCPY;
5584	lkb->lkb_lvbseq = le32_to_cpu(rl->rl_lvbseq);
5585	lkb->lkb_rqmode = rl->rl_rqmode;
5586	lkb->lkb_grmode = rl->rl_grmode;
5587	/* don't set lkb_status because add_lkb wants to itself */
5588
5589	lkb->lkb_bastfn = (rl->rl_asts & DLM_CB_BAST) ? &fake_bastfn : NULL;
5590	lkb->lkb_astfn = (rl->rl_asts & DLM_CB_CAST) ? &fake_astfn : NULL;
5591
5592	if (lkb->lkb_exflags & DLM_LKF_VALBLK) {
5593		int lvblen = rc->rc_header.h_length - sizeof(struct dlm_rcom) -
5594			 sizeof(struct rcom_lock);
5595		if (lvblen > ls->ls_lvblen)
5596			return -EINVAL;
5597		lkb->lkb_lvbptr = dlm_allocate_lvb(ls);
5598		if (!lkb->lkb_lvbptr)
5599			return -ENOMEM;
5600		memcpy(lkb->lkb_lvbptr, rl->rl_lvb, lvblen);
5601	}
5602
5603	/* Conversions between PR and CW (middle modes) need special handling.
5604	   The real granted mode of these converting locks cannot be determined
5605	   until all locks have been rebuilt on the rsb (recover_conversion) */
5606
5607	if (rl->rl_wait_type == cpu_to_le16(DLM_MSG_CONVERT) &&
5608	    middle_conversion(lkb)) {
5609		rl->rl_status = DLM_LKSTS_CONVERT;
5610		lkb->lkb_grmode = DLM_LOCK_IV;
5611		rsb_set_flag(r, RSB_RECOVER_CONVERT);
5612	}
5613
5614	return 0;
5615}
5616
5617/* This lkb may have been recovered in a previous aborted recovery so we need
5618   to check if the rsb already has an lkb with the given remote nodeid/lkid.
5619   If so we just send back a standard reply.  If not, we create a new lkb with
5620   the given values and send back our lkid.  We send back our lkid by sending
5621   back the rcom_lock struct we got but with the remid field filled in. */
5622
5623/* needs at least dlm_rcom + rcom_lock */
5624int dlm_recover_master_copy(struct dlm_ls *ls, struct dlm_rcom *rc)
 
5625{
5626	struct rcom_lock *rl = (struct rcom_lock *) rc->rc_buf;
5627	struct dlm_rsb *r;
5628	struct dlm_lkb *lkb;
5629	uint32_t remid = 0;
5630	int from_nodeid = rc->rc_header.h_nodeid;
5631	int error;
5632
 
 
 
5633	if (rl->rl_parent_lkid) {
5634		error = -EOPNOTSUPP;
5635		goto out;
5636	}
5637
5638	remid = le32_to_cpu(rl->rl_lkid);
5639
5640	/* In general we expect the rsb returned to be R_MASTER, but we don't
5641	   have to require it.  Recovery of masters on one node can overlap
5642	   recovery of locks on another node, so one node can send us MSTCPY
5643	   locks before we've made ourselves master of this rsb.  We can still
5644	   add new MSTCPY locks that we receive here without any harm; when
5645	   we make ourselves master, dlm_recover_masters() won't touch the
5646	   MSTCPY locks we've received early. */
5647
5648	error = find_rsb(ls, rl->rl_name, le16_to_cpu(rl->rl_namelen),
5649			 from_nodeid, R_RECEIVE_RECOVER, &r);
5650	if (error)
5651		goto out;
5652
5653	lock_rsb(r);
5654
5655	if (dlm_no_directory(ls) && (dlm_dir_nodeid(r) != dlm_our_nodeid())) {
5656		log_error(ls, "dlm_recover_master_copy remote %d %x not dir",
5657			  from_nodeid, remid);
5658		error = -EBADR;
5659		goto out_unlock;
5660	}
5661
5662	lkb = search_remid(r, from_nodeid, remid);
5663	if (lkb) {
5664		error = -EEXIST;
5665		goto out_remid;
5666	}
5667
5668	error = create_lkb(ls, &lkb);
5669	if (error)
5670		goto out_unlock;
5671
5672	error = receive_rcom_lock_args(ls, lkb, r, rc);
5673	if (error) {
5674		__put_lkb(ls, lkb);
5675		goto out_unlock;
5676	}
5677
5678	attach_lkb(r, lkb);
5679	add_lkb(r, lkb, rl->rl_status);
5680	error = 0;
5681	ls->ls_recover_locks_in++;
5682
5683	if (!list_empty(&r->res_waitqueue) || !list_empty(&r->res_convertqueue))
5684		rsb_set_flag(r, RSB_RECOVER_GRANT);
5685
5686 out_remid:
5687	/* this is the new value returned to the lock holder for
5688	   saving in its process-copy lkb */
5689	rl->rl_remid = cpu_to_le32(lkb->lkb_id);
5690
5691	lkb->lkb_recover_seq = ls->ls_recover_seq;
5692
5693 out_unlock:
5694	unlock_rsb(r);
5695	put_rsb(r);
5696 out:
5697	if (error && error != -EEXIST)
5698		log_rinfo(ls, "dlm_recover_master_copy remote %d %x error %d",
5699			  from_nodeid, remid, error);
5700	rl->rl_result = cpu_to_le32(error);
5701	return error;
5702}
5703
5704/* needs at least dlm_rcom + rcom_lock */
5705int dlm_recover_process_copy(struct dlm_ls *ls, struct dlm_rcom *rc)
 
5706{
5707	struct rcom_lock *rl = (struct rcom_lock *) rc->rc_buf;
5708	struct dlm_rsb *r;
5709	struct dlm_lkb *lkb;
5710	uint32_t lkid, remid;
5711	int error, result;
5712
5713	lkid = le32_to_cpu(rl->rl_lkid);
5714	remid = le32_to_cpu(rl->rl_remid);
5715	result = le32_to_cpu(rl->rl_result);
5716
5717	error = find_lkb(ls, lkid, &lkb);
5718	if (error) {
5719		log_error(ls, "dlm_recover_process_copy no %x remote %d %x %d",
5720			  lkid, rc->rc_header.h_nodeid, remid, result);
 
5721		return error;
5722	}
5723
5724	r = lkb->lkb_resource;
5725	hold_rsb(r);
5726	lock_rsb(r);
5727
5728	if (!is_process_copy(lkb)) {
5729		log_error(ls, "dlm_recover_process_copy bad %x remote %d %x %d",
5730			  lkid, rc->rc_header.h_nodeid, remid, result);
 
5731		dlm_dump_rsb(r);
5732		unlock_rsb(r);
5733		put_rsb(r);
5734		dlm_put_lkb(lkb);
5735		return -EINVAL;
5736	}
5737
5738	switch (result) {
5739	case -EBADR:
5740		/* There's a chance the new master received our lock before
5741		   dlm_recover_master_reply(), this wouldn't happen if we did
5742		   a barrier between recover_masters and recover_locks. */
5743
5744		log_debug(ls, "dlm_recover_process_copy %x remote %d %x %d",
5745			  lkid, rc->rc_header.h_nodeid, remid, result);
 
5746	
5747		dlm_send_rcom_lock(r, lkb);
5748		goto out;
5749	case -EEXIST:
5750	case 0:
5751		lkb->lkb_remid = remid;
5752		break;
5753	default:
5754		log_error(ls, "dlm_recover_process_copy %x remote %d %x %d unk",
5755			  lkid, rc->rc_header.h_nodeid, remid, result);
 
5756	}
5757
5758	/* an ack for dlm_recover_locks() which waits for replies from
5759	   all the locks it sends to new masters */
5760	dlm_recovered_lock(r);
5761 out:
5762	unlock_rsb(r);
5763	put_rsb(r);
5764	dlm_put_lkb(lkb);
5765
5766	return 0;
5767}
5768
5769int dlm_user_request(struct dlm_ls *ls, struct dlm_user_args *ua,
5770		     int mode, uint32_t flags, void *name, unsigned int namelen,
5771		     unsigned long timeout_cs)
5772{
5773	struct dlm_lkb *lkb;
5774	struct dlm_args args;
 
5775	int error;
5776
5777	dlm_lock_recovery(ls);
5778
5779	error = create_lkb(ls, &lkb);
5780	if (error) {
5781		kfree(ua);
5782		goto out;
5783	}
5784
 
 
5785	if (flags & DLM_LKF_VALBLK) {
5786		ua->lksb.sb_lvbptr = kzalloc(DLM_USER_LVB_LEN, GFP_NOFS);
5787		if (!ua->lksb.sb_lvbptr) {
5788			kfree(ua);
5789			__put_lkb(ls, lkb);
5790			error = -ENOMEM;
5791			goto out;
5792		}
5793	}
5794	error = set_lock_args(mode, &ua->lksb, flags, namelen, timeout_cs,
5795			      fake_astfn, ua, fake_bastfn, &args);
5796	if (error) {
5797		kfree(ua->lksb.sb_lvbptr);
5798		ua->lksb.sb_lvbptr = NULL;
5799		kfree(ua);
5800		__put_lkb(ls, lkb);
5801		goto out;
5802	}
5803
5804	/* After ua is attached to lkb it will be freed by dlm_free_lkb().
5805	   When DLM_IFL_USER is set, the dlm knows that this is a userspace
5806	   lock and that lkb_astparam is the dlm_user_args structure. */
5807	lkb->lkb_flags |= DLM_IFL_USER;
5808	error = request_lock(ls, lkb, name, namelen, &args);
5809
5810	switch (error) {
5811	case 0:
5812		break;
5813	case -EINPROGRESS:
5814		error = 0;
5815		break;
5816	case -EAGAIN:
5817		error = 0;
5818		fallthrough;
5819	default:
5820		__put_lkb(ls, lkb);
5821		goto out;
5822	}
5823
5824	/* add this new lkb to the per-process list of locks */
5825	spin_lock(&ua->proc->locks_spin);
5826	hold_lkb(lkb);
5827	list_add_tail(&lkb->lkb_ownqueue, &ua->proc->locks);
5828	spin_unlock(&ua->proc->locks_spin);
 
 
 
 
 
5829 out:
5830	dlm_unlock_recovery(ls);
5831	return error;
5832}
5833
5834int dlm_user_convert(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
5835		     int mode, uint32_t flags, uint32_t lkid, char *lvb_in,
5836		     unsigned long timeout_cs)
5837{
5838	struct dlm_lkb *lkb;
5839	struct dlm_args args;
5840	struct dlm_user_args *ua;
5841	int error;
5842
5843	dlm_lock_recovery(ls);
5844
5845	error = find_lkb(ls, lkid, &lkb);
5846	if (error)
5847		goto out;
5848
 
 
5849	/* user can change the params on its lock when it converts it, or
5850	   add an lvb that didn't exist before */
5851
5852	ua = lkb->lkb_ua;
5853
5854	if (flags & DLM_LKF_VALBLK && !ua->lksb.sb_lvbptr) {
5855		ua->lksb.sb_lvbptr = kzalloc(DLM_USER_LVB_LEN, GFP_NOFS);
5856		if (!ua->lksb.sb_lvbptr) {
5857			error = -ENOMEM;
5858			goto out_put;
5859		}
5860	}
5861	if (lvb_in && ua->lksb.sb_lvbptr)
5862		memcpy(ua->lksb.sb_lvbptr, lvb_in, DLM_USER_LVB_LEN);
5863
5864	ua->xid = ua_tmp->xid;
5865	ua->castparam = ua_tmp->castparam;
5866	ua->castaddr = ua_tmp->castaddr;
5867	ua->bastparam = ua_tmp->bastparam;
5868	ua->bastaddr = ua_tmp->bastaddr;
5869	ua->user_lksb = ua_tmp->user_lksb;
5870
5871	error = set_lock_args(mode, &ua->lksb, flags, 0, timeout_cs,
5872			      fake_astfn, ua, fake_bastfn, &args);
5873	if (error)
5874		goto out_put;
5875
5876	error = convert_lock(ls, lkb, &args);
5877
5878	if (error == -EINPROGRESS || error == -EAGAIN || error == -EDEADLK)
5879		error = 0;
5880 out_put:
 
5881	dlm_put_lkb(lkb);
5882 out:
5883	dlm_unlock_recovery(ls);
5884	kfree(ua_tmp);
5885	return error;
5886}
5887
5888/*
5889 * The caller asks for an orphan lock on a given resource with a given mode.
5890 * If a matching lock exists, it's moved to the owner's list of locks and
5891 * the lkid is returned.
5892 */
5893
5894int dlm_user_adopt_orphan(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
5895		     int mode, uint32_t flags, void *name, unsigned int namelen,
5896		     unsigned long timeout_cs, uint32_t *lkid)
5897{
5898	struct dlm_lkb *lkb;
5899	struct dlm_user_args *ua;
5900	int found_other_mode = 0;
5901	int found = 0;
5902	int rv = 0;
5903
5904	mutex_lock(&ls->ls_orphans_mutex);
5905	list_for_each_entry(lkb, &ls->ls_orphans, lkb_ownqueue) {
5906		if (lkb->lkb_resource->res_length != namelen)
5907			continue;
5908		if (memcmp(lkb->lkb_resource->res_name, name, namelen))
5909			continue;
5910		if (lkb->lkb_grmode != mode) {
5911			found_other_mode = 1;
5912			continue;
5913		}
5914
5915		found = 1;
5916		list_del_init(&lkb->lkb_ownqueue);
5917		lkb->lkb_flags &= ~DLM_IFL_ORPHAN;
5918		*lkid = lkb->lkb_id;
5919		break;
5920	}
5921	mutex_unlock(&ls->ls_orphans_mutex);
5922
5923	if (!found && found_other_mode) {
5924		rv = -EAGAIN;
5925		goto out;
5926	}
5927
5928	if (!found) {
5929		rv = -ENOENT;
5930		goto out;
5931	}
5932
5933	lkb->lkb_exflags = flags;
5934	lkb->lkb_ownpid = (int) current->pid;
5935
5936	ua = lkb->lkb_ua;
5937
5938	ua->proc = ua_tmp->proc;
5939	ua->xid = ua_tmp->xid;
5940	ua->castparam = ua_tmp->castparam;
5941	ua->castaddr = ua_tmp->castaddr;
5942	ua->bastparam = ua_tmp->bastparam;
5943	ua->bastaddr = ua_tmp->bastaddr;
5944	ua->user_lksb = ua_tmp->user_lksb;
5945
5946	/*
5947	 * The lkb reference from the ls_orphans list was not
5948	 * removed above, and is now considered the reference
5949	 * for the proc locks list.
5950	 */
5951
5952	spin_lock(&ua->proc->locks_spin);
5953	list_add_tail(&lkb->lkb_ownqueue, &ua->proc->locks);
5954	spin_unlock(&ua->proc->locks_spin);
5955 out:
5956	kfree(ua_tmp);
5957	return rv;
5958}
5959
5960int dlm_user_unlock(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
5961		    uint32_t flags, uint32_t lkid, char *lvb_in)
5962{
5963	struct dlm_lkb *lkb;
5964	struct dlm_args args;
5965	struct dlm_user_args *ua;
5966	int error;
5967
5968	dlm_lock_recovery(ls);
5969
5970	error = find_lkb(ls, lkid, &lkb);
5971	if (error)
5972		goto out;
5973
 
 
5974	ua = lkb->lkb_ua;
5975
5976	if (lvb_in && ua->lksb.sb_lvbptr)
5977		memcpy(ua->lksb.sb_lvbptr, lvb_in, DLM_USER_LVB_LEN);
5978	if (ua_tmp->castparam)
5979		ua->castparam = ua_tmp->castparam;
5980	ua->user_lksb = ua_tmp->user_lksb;
5981
5982	error = set_unlock_args(flags, ua, &args);
5983	if (error)
5984		goto out_put;
5985
5986	error = unlock_lock(ls, lkb, &args);
5987
5988	if (error == -DLM_EUNLOCK)
5989		error = 0;
5990	/* from validate_unlock_args() */
5991	if (error == -EBUSY && (flags & DLM_LKF_FORCEUNLOCK))
5992		error = 0;
5993	if (error)
5994		goto out_put;
5995
5996	spin_lock(&ua->proc->locks_spin);
5997	/* dlm_user_add_cb() may have already taken lkb off the proc list */
5998	if (!list_empty(&lkb->lkb_ownqueue))
5999		list_move(&lkb->lkb_ownqueue, &ua->proc->unlocking);
6000	spin_unlock(&ua->proc->locks_spin);
6001 out_put:
 
6002	dlm_put_lkb(lkb);
6003 out:
6004	dlm_unlock_recovery(ls);
6005	kfree(ua_tmp);
6006	return error;
6007}
6008
6009int dlm_user_cancel(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
6010		    uint32_t flags, uint32_t lkid)
6011{
6012	struct dlm_lkb *lkb;
6013	struct dlm_args args;
6014	struct dlm_user_args *ua;
6015	int error;
6016
6017	dlm_lock_recovery(ls);
6018
6019	error = find_lkb(ls, lkid, &lkb);
6020	if (error)
6021		goto out;
6022
 
 
6023	ua = lkb->lkb_ua;
6024	if (ua_tmp->castparam)
6025		ua->castparam = ua_tmp->castparam;
6026	ua->user_lksb = ua_tmp->user_lksb;
6027
6028	error = set_unlock_args(flags, ua, &args);
6029	if (error)
6030		goto out_put;
6031
6032	error = cancel_lock(ls, lkb, &args);
6033
6034	if (error == -DLM_ECANCEL)
6035		error = 0;
6036	/* from validate_unlock_args() */
6037	if (error == -EBUSY)
6038		error = 0;
6039 out_put:
 
6040	dlm_put_lkb(lkb);
6041 out:
6042	dlm_unlock_recovery(ls);
6043	kfree(ua_tmp);
6044	return error;
6045}
6046
6047int dlm_user_deadlock(struct dlm_ls *ls, uint32_t flags, uint32_t lkid)
6048{
6049	struct dlm_lkb *lkb;
6050	struct dlm_args args;
6051	struct dlm_user_args *ua;
6052	struct dlm_rsb *r;
6053	int error;
6054
6055	dlm_lock_recovery(ls);
6056
6057	error = find_lkb(ls, lkid, &lkb);
6058	if (error)
6059		goto out;
6060
 
 
6061	ua = lkb->lkb_ua;
6062
6063	error = set_unlock_args(flags, ua, &args);
6064	if (error)
6065		goto out_put;
6066
6067	/* same as cancel_lock(), but set DEADLOCK_CANCEL after lock_rsb */
6068
6069	r = lkb->lkb_resource;
6070	hold_rsb(r);
6071	lock_rsb(r);
6072
6073	error = validate_unlock_args(lkb, &args);
6074	if (error)
6075		goto out_r;
6076	lkb->lkb_flags |= DLM_IFL_DEADLOCK_CANCEL;
6077
6078	error = _cancel_lock(r, lkb);
6079 out_r:
6080	unlock_rsb(r);
6081	put_rsb(r);
6082
6083	if (error == -DLM_ECANCEL)
6084		error = 0;
6085	/* from validate_unlock_args() */
6086	if (error == -EBUSY)
6087		error = 0;
6088 out_put:
 
6089	dlm_put_lkb(lkb);
6090 out:
6091	dlm_unlock_recovery(ls);
6092	return error;
6093}
6094
6095/* lkb's that are removed from the waiters list by revert are just left on the
6096   orphans list with the granted orphan locks, to be freed by purge */
6097
6098static int orphan_proc_lock(struct dlm_ls *ls, struct dlm_lkb *lkb)
6099{
6100	struct dlm_args args;
6101	int error;
6102
6103	hold_lkb(lkb); /* reference for the ls_orphans list */
6104	mutex_lock(&ls->ls_orphans_mutex);
6105	list_add_tail(&lkb->lkb_ownqueue, &ls->ls_orphans);
6106	mutex_unlock(&ls->ls_orphans_mutex);
6107
6108	set_unlock_args(0, lkb->lkb_ua, &args);
6109
6110	error = cancel_lock(ls, lkb, &args);
6111	if (error == -DLM_ECANCEL)
6112		error = 0;
6113	return error;
6114}
6115
6116/* The FORCEUNLOCK flag allows the unlock to go ahead even if the lkb isn't
6117   granted.  Regardless of what rsb queue the lock is on, it's removed and
6118   freed.  The IVVALBLK flag causes the lvb on the resource to be invalidated
6119   if our lock is PW/EX (it's ignored if our granted mode is smaller.) */
6120
6121static int unlock_proc_lock(struct dlm_ls *ls, struct dlm_lkb *lkb)
6122{
6123	struct dlm_args args;
6124	int error;
6125
6126	set_unlock_args(DLM_LKF_FORCEUNLOCK | DLM_LKF_IVVALBLK,
6127			lkb->lkb_ua, &args);
6128
6129	error = unlock_lock(ls, lkb, &args);
6130	if (error == -DLM_EUNLOCK)
6131		error = 0;
6132	return error;
6133}
6134
6135/* We have to release clear_proc_locks mutex before calling unlock_proc_lock()
6136   (which does lock_rsb) due to deadlock with receiving a message that does
6137   lock_rsb followed by dlm_user_add_cb() */
6138
6139static struct dlm_lkb *del_proc_lock(struct dlm_ls *ls,
6140				     struct dlm_user_proc *proc)
6141{
6142	struct dlm_lkb *lkb = NULL;
6143
6144	mutex_lock(&ls->ls_clear_proc_locks);
6145	if (list_empty(&proc->locks))
6146		goto out;
6147
6148	lkb = list_entry(proc->locks.next, struct dlm_lkb, lkb_ownqueue);
6149	list_del_init(&lkb->lkb_ownqueue);
6150
6151	if (lkb->lkb_exflags & DLM_LKF_PERSISTENT)
6152		lkb->lkb_flags |= DLM_IFL_ORPHAN;
6153	else
6154		lkb->lkb_flags |= DLM_IFL_DEAD;
6155 out:
6156	mutex_unlock(&ls->ls_clear_proc_locks);
6157	return lkb;
6158}
6159
6160/* The ls_clear_proc_locks mutex protects against dlm_user_add_cb() which
6161   1) references lkb->ua which we free here and 2) adds lkbs to proc->asts,
6162   which we clear here. */
6163
6164/* proc CLOSING flag is set so no more device_reads should look at proc->asts
6165   list, and no more device_writes should add lkb's to proc->locks list; so we
6166   shouldn't need to take asts_spin or locks_spin here.  this assumes that
6167   device reads/writes/closes are serialized -- FIXME: we may need to serialize
6168   them ourself. */
6169
6170void dlm_clear_proc_locks(struct dlm_ls *ls, struct dlm_user_proc *proc)
6171{
6172	struct dlm_lkb *lkb, *safe;
6173
6174	dlm_lock_recovery(ls);
6175
6176	while (1) {
6177		lkb = del_proc_lock(ls, proc);
6178		if (!lkb)
6179			break;
6180		del_timeout(lkb);
6181		if (lkb->lkb_exflags & DLM_LKF_PERSISTENT)
6182			orphan_proc_lock(ls, lkb);
6183		else
6184			unlock_proc_lock(ls, lkb);
6185
6186		/* this removes the reference for the proc->locks list
6187		   added by dlm_user_request, it may result in the lkb
6188		   being freed */
6189
6190		dlm_put_lkb(lkb);
6191	}
6192
6193	mutex_lock(&ls->ls_clear_proc_locks);
6194
6195	/* in-progress unlocks */
6196	list_for_each_entry_safe(lkb, safe, &proc->unlocking, lkb_ownqueue) {
6197		list_del_init(&lkb->lkb_ownqueue);
6198		lkb->lkb_flags |= DLM_IFL_DEAD;
6199		dlm_put_lkb(lkb);
6200	}
6201
6202	list_for_each_entry_safe(lkb, safe, &proc->asts, lkb_cb_list) {
6203		memset(&lkb->lkb_callbacks, 0,
6204		       sizeof(struct dlm_callback) * DLM_CALLBACKS_SIZE);
6205		list_del_init(&lkb->lkb_cb_list);
6206		dlm_put_lkb(lkb);
6207	}
6208
6209	mutex_unlock(&ls->ls_clear_proc_locks);
6210	dlm_unlock_recovery(ls);
6211}
6212
6213static void purge_proc_locks(struct dlm_ls *ls, struct dlm_user_proc *proc)
6214{
6215	struct dlm_lkb *lkb, *safe;
6216
6217	while (1) {
6218		lkb = NULL;
6219		spin_lock(&proc->locks_spin);
6220		if (!list_empty(&proc->locks)) {
6221			lkb = list_entry(proc->locks.next, struct dlm_lkb,
6222					 lkb_ownqueue);
6223			list_del_init(&lkb->lkb_ownqueue);
6224		}
6225		spin_unlock(&proc->locks_spin);
6226
6227		if (!lkb)
6228			break;
6229
6230		lkb->lkb_flags |= DLM_IFL_DEAD;
6231		unlock_proc_lock(ls, lkb);
6232		dlm_put_lkb(lkb); /* ref from proc->locks list */
6233	}
6234
6235	spin_lock(&proc->locks_spin);
6236	list_for_each_entry_safe(lkb, safe, &proc->unlocking, lkb_ownqueue) {
6237		list_del_init(&lkb->lkb_ownqueue);
6238		lkb->lkb_flags |= DLM_IFL_DEAD;
6239		dlm_put_lkb(lkb);
6240	}
6241	spin_unlock(&proc->locks_spin);
6242
6243	spin_lock(&proc->asts_spin);
6244	list_for_each_entry_safe(lkb, safe, &proc->asts, lkb_cb_list) {
6245		memset(&lkb->lkb_callbacks, 0,
6246		       sizeof(struct dlm_callback) * DLM_CALLBACKS_SIZE);
6247		list_del_init(&lkb->lkb_cb_list);
6248		dlm_put_lkb(lkb);
6249	}
6250	spin_unlock(&proc->asts_spin);
6251}
6252
6253/* pid of 0 means purge all orphans */
6254
6255static void do_purge(struct dlm_ls *ls, int nodeid, int pid)
6256{
6257	struct dlm_lkb *lkb, *safe;
6258
6259	mutex_lock(&ls->ls_orphans_mutex);
6260	list_for_each_entry_safe(lkb, safe, &ls->ls_orphans, lkb_ownqueue) {
6261		if (pid && lkb->lkb_ownpid != pid)
6262			continue;
6263		unlock_proc_lock(ls, lkb);
6264		list_del_init(&lkb->lkb_ownqueue);
6265		dlm_put_lkb(lkb);
6266	}
6267	mutex_unlock(&ls->ls_orphans_mutex);
6268}
6269
6270static int send_purge(struct dlm_ls *ls, int nodeid, int pid)
6271{
6272	struct dlm_message *ms;
6273	struct dlm_mhandle *mh;
6274	int error;
6275
6276	error = _create_message(ls, sizeof(struct dlm_message), nodeid,
6277				DLM_MSG_PURGE, &ms, &mh);
6278	if (error)
6279		return error;
6280	ms->m_nodeid = nodeid;
6281	ms->m_pid = pid;
6282
6283	return send_message(mh, ms);
6284}
6285
6286int dlm_user_purge(struct dlm_ls *ls, struct dlm_user_proc *proc,
6287		   int nodeid, int pid)
6288{
6289	int error = 0;
6290
6291	if (nodeid && (nodeid != dlm_our_nodeid())) {
6292		error = send_purge(ls, nodeid, pid);
6293	} else {
6294		dlm_lock_recovery(ls);
6295		if (pid == current->pid)
6296			purge_proc_locks(ls, proc);
6297		else
6298			do_purge(ls, nodeid, pid);
6299		dlm_unlock_recovery(ls);
6300	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6301	return error;
6302}
6303
v6.8
   1// SPDX-License-Identifier: GPL-2.0-only
   2/******************************************************************************
   3*******************************************************************************
   4**
   5**  Copyright (C) 2005-2010 Red Hat, Inc.  All rights reserved.
   6**
   7**
   8*******************************************************************************
   9******************************************************************************/
  10
  11/* Central locking logic has four stages:
  12
  13   dlm_lock()
  14   dlm_unlock()
  15
  16   request_lock(ls, lkb)
  17   convert_lock(ls, lkb)
  18   unlock_lock(ls, lkb)
  19   cancel_lock(ls, lkb)
  20
  21   _request_lock(r, lkb)
  22   _convert_lock(r, lkb)
  23   _unlock_lock(r, lkb)
  24   _cancel_lock(r, lkb)
  25
  26   do_request(r, lkb)
  27   do_convert(r, lkb)
  28   do_unlock(r, lkb)
  29   do_cancel(r, lkb)
  30
  31   Stage 1 (lock, unlock) is mainly about checking input args and
  32   splitting into one of the four main operations:
  33
  34       dlm_lock          = request_lock
  35       dlm_lock+CONVERT  = convert_lock
  36       dlm_unlock        = unlock_lock
  37       dlm_unlock+CANCEL = cancel_lock
  38
  39   Stage 2, xxxx_lock(), just finds and locks the relevant rsb which is
  40   provided to the next stage.
  41
  42   Stage 3, _xxxx_lock(), determines if the operation is local or remote.
  43   When remote, it calls send_xxxx(), when local it calls do_xxxx().
  44
  45   Stage 4, do_xxxx(), is the guts of the operation.  It manipulates the
  46   given rsb and lkb and queues callbacks.
  47
  48   For remote operations, send_xxxx() results in the corresponding do_xxxx()
  49   function being executed on the remote node.  The connecting send/receive
  50   calls on local (L) and remote (R) nodes:
  51
  52   L: send_xxxx()              ->  R: receive_xxxx()
  53                                   R: do_xxxx()
  54   L: receive_xxxx_reply()     <-  R: send_xxxx_reply()
  55*/
  56#include <trace/events/dlm.h>
  57
  58#include <linux/types.h>
  59#include <linux/rbtree.h>
  60#include <linux/slab.h>
  61#include "dlm_internal.h"
  62#include <linux/dlm_device.h>
  63#include "memory.h"
  64#include "midcomms.h"
  65#include "requestqueue.h"
  66#include "util.h"
  67#include "dir.h"
  68#include "member.h"
  69#include "lockspace.h"
  70#include "ast.h"
  71#include "lock.h"
  72#include "rcom.h"
  73#include "recover.h"
  74#include "lvb_table.h"
  75#include "user.h"
  76#include "config.h"
  77
  78static int send_request(struct dlm_rsb *r, struct dlm_lkb *lkb);
  79static int send_convert(struct dlm_rsb *r, struct dlm_lkb *lkb);
  80static int send_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb);
  81static int send_cancel(struct dlm_rsb *r, struct dlm_lkb *lkb);
  82static int send_grant(struct dlm_rsb *r, struct dlm_lkb *lkb);
  83static int send_bast(struct dlm_rsb *r, struct dlm_lkb *lkb, int mode);
  84static int send_lookup(struct dlm_rsb *r, struct dlm_lkb *lkb);
  85static int send_remove(struct dlm_rsb *r);
  86static int _request_lock(struct dlm_rsb *r, struct dlm_lkb *lkb);
  87static int _cancel_lock(struct dlm_rsb *r, struct dlm_lkb *lkb);
  88static void __receive_convert_reply(struct dlm_rsb *r, struct dlm_lkb *lkb,
  89				    const struct dlm_message *ms, bool local);
  90static int receive_extralen(const struct dlm_message *ms);
  91static void do_purge(struct dlm_ls *ls, int nodeid, int pid);
 
  92static void toss_rsb(struct kref *kref);
  93
  94/*
  95 * Lock compatibilty matrix - thanks Steve
  96 * UN = Unlocked state. Not really a state, used as a flag
  97 * PD = Padding. Used to make the matrix a nice power of two in size
  98 * Other states are the same as the VMS DLM.
  99 * Usage: matrix[grmode+1][rqmode+1]  (although m[rq+1][gr+1] is the same)
 100 */
 101
 102static const int __dlm_compat_matrix[8][8] = {
 103      /* UN NL CR CW PR PW EX PD */
 104        {1, 1, 1, 1, 1, 1, 1, 0},       /* UN */
 105        {1, 1, 1, 1, 1, 1, 1, 0},       /* NL */
 106        {1, 1, 1, 1, 1, 1, 0, 0},       /* CR */
 107        {1, 1, 1, 1, 0, 0, 0, 0},       /* CW */
 108        {1, 1, 1, 0, 1, 0, 0, 0},       /* PR */
 109        {1, 1, 1, 0, 0, 0, 0, 0},       /* PW */
 110        {1, 1, 0, 0, 0, 0, 0, 0},       /* EX */
 111        {0, 0, 0, 0, 0, 0, 0, 0}        /* PD */
 112};
 113
 114/*
 115 * This defines the direction of transfer of LVB data.
 116 * Granted mode is the row; requested mode is the column.
 117 * Usage: matrix[grmode+1][rqmode+1]
 118 * 1 = LVB is returned to the caller
 119 * 0 = LVB is written to the resource
 120 * -1 = nothing happens to the LVB
 121 */
 122
 123const int dlm_lvb_operations[8][8] = {
 124        /* UN   NL  CR  CW  PR  PW  EX  PD*/
 125        {  -1,  1,  1,  1,  1,  1,  1, -1 }, /* UN */
 126        {  -1,  1,  1,  1,  1,  1,  1,  0 }, /* NL */
 127        {  -1, -1,  1,  1,  1,  1,  1,  0 }, /* CR */
 128        {  -1, -1, -1,  1,  1,  1,  1,  0 }, /* CW */
 129        {  -1, -1, -1, -1,  1,  1,  1,  0 }, /* PR */
 130        {  -1,  0,  0,  0,  0,  0,  1,  0 }, /* PW */
 131        {  -1,  0,  0,  0,  0,  0,  0,  0 }, /* EX */
 132        {  -1,  0,  0,  0,  0,  0,  0,  0 }  /* PD */
 133};
 134
 135#define modes_compat(gr, rq) \
 136	__dlm_compat_matrix[(gr)->lkb_grmode + 1][(rq)->lkb_rqmode + 1]
 137
 138int dlm_modes_compat(int mode1, int mode2)
 139{
 140	return __dlm_compat_matrix[mode1 + 1][mode2 + 1];
 141}
 142
 143/*
 144 * Compatibility matrix for conversions with QUECVT set.
 145 * Granted mode is the row; requested mode is the column.
 146 * Usage: matrix[grmode+1][rqmode+1]
 147 */
 148
 149static const int __quecvt_compat_matrix[8][8] = {
 150      /* UN NL CR CW PR PW EX PD */
 151        {0, 0, 0, 0, 0, 0, 0, 0},       /* UN */
 152        {0, 0, 1, 1, 1, 1, 1, 0},       /* NL */
 153        {0, 0, 0, 1, 1, 1, 1, 0},       /* CR */
 154        {0, 0, 0, 0, 1, 1, 1, 0},       /* CW */
 155        {0, 0, 0, 1, 0, 1, 1, 0},       /* PR */
 156        {0, 0, 0, 0, 0, 0, 1, 0},       /* PW */
 157        {0, 0, 0, 0, 0, 0, 0, 0},       /* EX */
 158        {0, 0, 0, 0, 0, 0, 0, 0}        /* PD */
 159};
 160
 161void dlm_print_lkb(struct dlm_lkb *lkb)
 162{
 163	printk(KERN_ERR "lkb: nodeid %d id %x remid %x exflags %x flags %x "
 164	       "sts %d rq %d gr %d wait_type %d wait_nodeid %d seq %llu\n",
 165	       lkb->lkb_nodeid, lkb->lkb_id, lkb->lkb_remid, lkb->lkb_exflags,
 166	       dlm_iflags_val(lkb), lkb->lkb_status, lkb->lkb_rqmode,
 167	       lkb->lkb_grmode, lkb->lkb_wait_type, lkb->lkb_wait_nodeid,
 168	       (unsigned long long)lkb->lkb_recover_seq);
 169}
 170
 171static void dlm_print_rsb(struct dlm_rsb *r)
 172{
 173	printk(KERN_ERR "rsb: nodeid %d master %d dir %d flags %lx first %x "
 174	       "rlc %d name %s\n",
 175	       r->res_nodeid, r->res_master_nodeid, r->res_dir_nodeid,
 176	       r->res_flags, r->res_first_lkid, r->res_recover_locks_count,
 177	       r->res_name);
 178}
 179
 180void dlm_dump_rsb(struct dlm_rsb *r)
 181{
 182	struct dlm_lkb *lkb;
 183
 184	dlm_print_rsb(r);
 185
 186	printk(KERN_ERR "rsb: root_list empty %d recover_list empty %d\n",
 187	       list_empty(&r->res_root_list), list_empty(&r->res_recover_list));
 188	printk(KERN_ERR "rsb lookup list\n");
 189	list_for_each_entry(lkb, &r->res_lookup, lkb_rsb_lookup)
 190		dlm_print_lkb(lkb);
 191	printk(KERN_ERR "rsb grant queue:\n");
 192	list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue)
 193		dlm_print_lkb(lkb);
 194	printk(KERN_ERR "rsb convert queue:\n");
 195	list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue)
 196		dlm_print_lkb(lkb);
 197	printk(KERN_ERR "rsb wait queue:\n");
 198	list_for_each_entry(lkb, &r->res_waitqueue, lkb_statequeue)
 199		dlm_print_lkb(lkb);
 200}
 201
 202/* Threads cannot use the lockspace while it's being recovered */
 203
 204static inline void dlm_lock_recovery(struct dlm_ls *ls)
 205{
 206	down_read(&ls->ls_in_recovery);
 207}
 208
 209void dlm_unlock_recovery(struct dlm_ls *ls)
 210{
 211	up_read(&ls->ls_in_recovery);
 212}
 213
 214int dlm_lock_recovery_try(struct dlm_ls *ls)
 215{
 216	return down_read_trylock(&ls->ls_in_recovery);
 217}
 218
 219static inline int can_be_queued(struct dlm_lkb *lkb)
 220{
 221	return !(lkb->lkb_exflags & DLM_LKF_NOQUEUE);
 222}
 223
 224static inline int force_blocking_asts(struct dlm_lkb *lkb)
 225{
 226	return (lkb->lkb_exflags & DLM_LKF_NOQUEUEBAST);
 227}
 228
 229static inline int is_demoted(struct dlm_lkb *lkb)
 230{
 231	return test_bit(DLM_SBF_DEMOTED_BIT, &lkb->lkb_sbflags);
 232}
 233
 234static inline int is_altmode(struct dlm_lkb *lkb)
 235{
 236	return test_bit(DLM_SBF_ALTMODE_BIT, &lkb->lkb_sbflags);
 237}
 238
 239static inline int is_granted(struct dlm_lkb *lkb)
 240{
 241	return (lkb->lkb_status == DLM_LKSTS_GRANTED);
 242}
 243
 244static inline int is_remote(struct dlm_rsb *r)
 245{
 246	DLM_ASSERT(r->res_nodeid >= 0, dlm_print_rsb(r););
 247	return !!r->res_nodeid;
 248}
 249
 250static inline int is_process_copy(struct dlm_lkb *lkb)
 251{
 252	return lkb->lkb_nodeid &&
 253	       !test_bit(DLM_IFL_MSTCPY_BIT, &lkb->lkb_iflags);
 254}
 255
 256static inline int is_master_copy(struct dlm_lkb *lkb)
 257{
 258	return test_bit(DLM_IFL_MSTCPY_BIT, &lkb->lkb_iflags);
 259}
 260
 261static inline int middle_conversion(struct dlm_lkb *lkb)
 262{
 263	if ((lkb->lkb_grmode==DLM_LOCK_PR && lkb->lkb_rqmode==DLM_LOCK_CW) ||
 264	    (lkb->lkb_rqmode==DLM_LOCK_PR && lkb->lkb_grmode==DLM_LOCK_CW))
 265		return 1;
 266	return 0;
 267}
 268
 269static inline int down_conversion(struct dlm_lkb *lkb)
 270{
 271	return (!middle_conversion(lkb) && lkb->lkb_rqmode < lkb->lkb_grmode);
 272}
 273
 274static inline int is_overlap_unlock(struct dlm_lkb *lkb)
 275{
 276	return test_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags);
 277}
 278
 279static inline int is_overlap_cancel(struct dlm_lkb *lkb)
 280{
 281	return test_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags);
 282}
 283
 284static inline int is_overlap(struct dlm_lkb *lkb)
 285{
 286	return test_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags) ||
 287	       test_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags);
 288}
 289
 290static void queue_cast(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
 291{
 292	if (is_master_copy(lkb))
 293		return;
 294
 
 
 295	DLM_ASSERT(lkb->lkb_lksb, dlm_print_lkb(lkb););
 296
 297	if (rv == -DLM_ECANCEL &&
 298	    test_and_clear_bit(DLM_IFL_DEADLOCK_CANCEL_BIT, &lkb->lkb_iflags))
 
 
 
 
 
 
 
 299		rv = -EDEADLK;
 
 300
 301	dlm_add_cb(lkb, DLM_CB_CAST, lkb->lkb_grmode, rv, dlm_sbflags_val(lkb));
 302}
 303
 304static inline void queue_cast_overlap(struct dlm_rsb *r, struct dlm_lkb *lkb)
 305{
 306	queue_cast(r, lkb,
 307		   is_overlap_unlock(lkb) ? -DLM_EUNLOCK : -DLM_ECANCEL);
 308}
 309
 310static void queue_bast(struct dlm_rsb *r, struct dlm_lkb *lkb, int rqmode)
 311{
 312	if (is_master_copy(lkb)) {
 313		send_bast(r, lkb, rqmode);
 314	} else {
 315		dlm_add_cb(lkb, DLM_CB_BAST, rqmode, 0, 0);
 316	}
 317}
 318
 319/*
 320 * Basic operations on rsb's and lkb's
 321 */
 322
 323/* This is only called to add a reference when the code already holds
 324   a valid reference to the rsb, so there's no need for locking. */
 325
 326static inline void hold_rsb(struct dlm_rsb *r)
 327{
 328	kref_get(&r->res_ref);
 329}
 330
 331void dlm_hold_rsb(struct dlm_rsb *r)
 332{
 333	hold_rsb(r);
 334}
 335
 336/* When all references to the rsb are gone it's transferred to
 337   the tossed list for later disposal. */
 338
 339static void put_rsb(struct dlm_rsb *r)
 340{
 341	struct dlm_ls *ls = r->res_ls;
 342	uint32_t bucket = r->res_bucket;
 343	int rv;
 344
 345	rv = kref_put_lock(&r->res_ref, toss_rsb,
 346			   &ls->ls_rsbtbl[bucket].lock);
 347	if (rv)
 348		spin_unlock(&ls->ls_rsbtbl[bucket].lock);
 349}
 350
 351void dlm_put_rsb(struct dlm_rsb *r)
 352{
 353	put_rsb(r);
 354}
 355
 356static int pre_rsb_struct(struct dlm_ls *ls)
 357{
 358	struct dlm_rsb *r1, *r2;
 359	int count = 0;
 360
 361	spin_lock(&ls->ls_new_rsb_spin);
 362	if (ls->ls_new_rsb_count > dlm_config.ci_new_rsb_count / 2) {
 363		spin_unlock(&ls->ls_new_rsb_spin);
 364		return 0;
 365	}
 366	spin_unlock(&ls->ls_new_rsb_spin);
 367
 368	r1 = dlm_allocate_rsb(ls);
 369	r2 = dlm_allocate_rsb(ls);
 370
 371	spin_lock(&ls->ls_new_rsb_spin);
 372	if (r1) {
 373		list_add(&r1->res_hashchain, &ls->ls_new_rsb);
 374		ls->ls_new_rsb_count++;
 375	}
 376	if (r2) {
 377		list_add(&r2->res_hashchain, &ls->ls_new_rsb);
 378		ls->ls_new_rsb_count++;
 379	}
 380	count = ls->ls_new_rsb_count;
 381	spin_unlock(&ls->ls_new_rsb_spin);
 382
 383	if (!count)
 384		return -ENOMEM;
 385	return 0;
 386}
 387
 388/* If ls->ls_new_rsb is empty, return -EAGAIN, so the caller can
 389   unlock any spinlocks, go back and call pre_rsb_struct again.
 390   Otherwise, take an rsb off the list and return it. */
 391
 392static int get_rsb_struct(struct dlm_ls *ls, const void *name, int len,
 393			  struct dlm_rsb **r_ret)
 394{
 395	struct dlm_rsb *r;
 396	int count;
 397
 398	spin_lock(&ls->ls_new_rsb_spin);
 399	if (list_empty(&ls->ls_new_rsb)) {
 400		count = ls->ls_new_rsb_count;
 401		spin_unlock(&ls->ls_new_rsb_spin);
 402		log_debug(ls, "find_rsb retry %d %d %s",
 403			  count, dlm_config.ci_new_rsb_count,
 404			  (const char *)name);
 405		return -EAGAIN;
 406	}
 407
 408	r = list_first_entry(&ls->ls_new_rsb, struct dlm_rsb, res_hashchain);
 409	list_del(&r->res_hashchain);
 410	/* Convert the empty list_head to a NULL rb_node for tree usage: */
 411	memset(&r->res_hashnode, 0, sizeof(struct rb_node));
 412	ls->ls_new_rsb_count--;
 413	spin_unlock(&ls->ls_new_rsb_spin);
 414
 415	r->res_ls = ls;
 416	r->res_length = len;
 417	memcpy(r->res_name, name, len);
 418	mutex_init(&r->res_mutex);
 419
 420	INIT_LIST_HEAD(&r->res_lookup);
 421	INIT_LIST_HEAD(&r->res_grantqueue);
 422	INIT_LIST_HEAD(&r->res_convertqueue);
 423	INIT_LIST_HEAD(&r->res_waitqueue);
 424	INIT_LIST_HEAD(&r->res_root_list);
 425	INIT_LIST_HEAD(&r->res_recover_list);
 426
 427	*r_ret = r;
 428	return 0;
 429}
 430
 431static int rsb_cmp(struct dlm_rsb *r, const char *name, int nlen)
 432{
 433	char maxname[DLM_RESNAME_MAXLEN];
 434
 435	memset(maxname, 0, DLM_RESNAME_MAXLEN);
 436	memcpy(maxname, name, nlen);
 437	return memcmp(r->res_name, maxname, DLM_RESNAME_MAXLEN);
 438}
 439
 440int dlm_search_rsb_tree(struct rb_root *tree, const void *name, int len,
 441			struct dlm_rsb **r_ret)
 442{
 443	struct rb_node *node = tree->rb_node;
 444	struct dlm_rsb *r;
 445	int rc;
 446
 447	while (node) {
 448		r = rb_entry(node, struct dlm_rsb, res_hashnode);
 449		rc = rsb_cmp(r, name, len);
 450		if (rc < 0)
 451			node = node->rb_left;
 452		else if (rc > 0)
 453			node = node->rb_right;
 454		else
 455			goto found;
 456	}
 457	*r_ret = NULL;
 458	return -EBADR;
 459
 460 found:
 461	*r_ret = r;
 462	return 0;
 463}
 464
 465static int rsb_insert(struct dlm_rsb *rsb, struct rb_root *tree)
 466{
 467	struct rb_node **newn = &tree->rb_node;
 468	struct rb_node *parent = NULL;
 469	int rc;
 470
 471	while (*newn) {
 472		struct dlm_rsb *cur = rb_entry(*newn, struct dlm_rsb,
 473					       res_hashnode);
 474
 475		parent = *newn;
 476		rc = rsb_cmp(cur, rsb->res_name, rsb->res_length);
 477		if (rc < 0)
 478			newn = &parent->rb_left;
 479		else if (rc > 0)
 480			newn = &parent->rb_right;
 481		else {
 482			log_print("rsb_insert match");
 483			dlm_dump_rsb(rsb);
 484			dlm_dump_rsb(cur);
 485			return -EEXIST;
 486		}
 487	}
 488
 489	rb_link_node(&rsb->res_hashnode, parent, newn);
 490	rb_insert_color(&rsb->res_hashnode, tree);
 491	return 0;
 492}
 493
 494/*
 495 * Find rsb in rsbtbl and potentially create/add one
 496 *
 497 * Delaying the release of rsb's has a similar benefit to applications keeping
 498 * NL locks on an rsb, but without the guarantee that the cached master value
 499 * will still be valid when the rsb is reused.  Apps aren't always smart enough
 500 * to keep NL locks on an rsb that they may lock again shortly; this can lead
 501 * to excessive master lookups and removals if we don't delay the release.
 502 *
 503 * Searching for an rsb means looking through both the normal list and toss
 504 * list.  When found on the toss list the rsb is moved to the normal list with
 505 * ref count of 1; when found on normal list the ref count is incremented.
 506 *
 507 * rsb's on the keep list are being used locally and refcounted.
 508 * rsb's on the toss list are not being used locally, and are not refcounted.
 509 *
 510 * The toss list rsb's were either
 511 * - previously used locally but not any more (were on keep list, then
 512 *   moved to toss list when last refcount dropped)
 513 * - created and put on toss list as a directory record for a lookup
 514 *   (we are the dir node for the res, but are not using the res right now,
 515 *   but some other node is)
 516 *
 517 * The purpose of find_rsb() is to return a refcounted rsb for local use.
 518 * So, if the given rsb is on the toss list, it is moved to the keep list
 519 * before being returned.
 520 *
 521 * toss_rsb() happens when all local usage of the rsb is done, i.e. no
 522 * more refcounts exist, so the rsb is moved from the keep list to the
 523 * toss list.
 524 *
 525 * rsb's on both keep and toss lists are used for doing a name to master
 526 * lookups.  rsb's that are in use locally (and being refcounted) are on
 527 * the keep list, rsb's that are not in use locally (not refcounted) and
 528 * only exist for name/master lookups are on the toss list.
 529 *
 530 * rsb's on the toss list who's dir_nodeid is not local can have stale
 531 * name/master mappings.  So, remote requests on such rsb's can potentially
 532 * return with an error, which means the mapping is stale and needs to
 533 * be updated with a new lookup.  (The idea behind MASTER UNCERTAIN and
 534 * first_lkid is to keep only a single outstanding request on an rsb
 535 * while that rsb has a potentially stale master.)
 536 */
 537
 538static int find_rsb_dir(struct dlm_ls *ls, const void *name, int len,
 539			uint32_t hash, uint32_t b,
 540			int dir_nodeid, int from_nodeid,
 541			unsigned int flags, struct dlm_rsb **r_ret)
 542{
 543	struct dlm_rsb *r = NULL;
 544	int our_nodeid = dlm_our_nodeid();
 545	int from_local = 0;
 546	int from_other = 0;
 547	int from_dir = 0;
 548	int create = 0;
 549	int error;
 550
 551	if (flags & R_RECEIVE_REQUEST) {
 552		if (from_nodeid == dir_nodeid)
 553			from_dir = 1;
 554		else
 555			from_other = 1;
 556	} else if (flags & R_REQUEST) {
 557		from_local = 1;
 558	}
 559
 560	/*
 561	 * flags & R_RECEIVE_RECOVER is from dlm_recover_master_copy, so
 562	 * from_nodeid has sent us a lock in dlm_recover_locks, believing
 563	 * we're the new master.  Our local recovery may not have set
 564	 * res_master_nodeid to our_nodeid yet, so allow either.  Don't
 565	 * create the rsb; dlm_recover_process_copy() will handle EBADR
 566	 * by resending.
 567	 *
 568	 * If someone sends us a request, we are the dir node, and we do
 569	 * not find the rsb anywhere, then recreate it.  This happens if
 570	 * someone sends us a request after we have removed/freed an rsb
 571	 * from our toss list.  (They sent a request instead of lookup
 572	 * because they are using an rsb from their toss list.)
 573	 */
 574
 575	if (from_local || from_dir ||
 576	    (from_other && (dir_nodeid == our_nodeid))) {
 577		create = 1;
 578	}
 579
 580 retry:
 581	if (create) {
 582		error = pre_rsb_struct(ls);
 583		if (error < 0)
 584			goto out;
 585	}
 586
 587	spin_lock(&ls->ls_rsbtbl[b].lock);
 588
 589	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
 590	if (error)
 591		goto do_toss;
 592	
 593	/*
 594	 * rsb is active, so we can't check master_nodeid without lock_rsb.
 595	 */
 596
 597	kref_get(&r->res_ref);
 
 598	goto out_unlock;
 599
 600
 601 do_toss:
 602	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
 603	if (error)
 604		goto do_new;
 605
 606	/*
 607	 * rsb found inactive (master_nodeid may be out of date unless
 608	 * we are the dir_nodeid or were the master)  No other thread
 609	 * is using this rsb because it's on the toss list, so we can
 610	 * look at or update res_master_nodeid without lock_rsb.
 611	 */
 612
 613	if ((r->res_master_nodeid != our_nodeid) && from_other) {
 614		/* our rsb was not master, and another node (not the dir node)
 615		   has sent us a request */
 616		log_debug(ls, "find_rsb toss from_other %d master %d dir %d %s",
 617			  from_nodeid, r->res_master_nodeid, dir_nodeid,
 618			  r->res_name);
 619		error = -ENOTBLK;
 620		goto out_unlock;
 621	}
 622
 623	if ((r->res_master_nodeid != our_nodeid) && from_dir) {
 624		/* don't think this should ever happen */
 625		log_error(ls, "find_rsb toss from_dir %d master %d",
 626			  from_nodeid, r->res_master_nodeid);
 627		dlm_print_rsb(r);
 628		/* fix it and go on */
 629		r->res_master_nodeid = our_nodeid;
 630		r->res_nodeid = 0;
 631		rsb_clear_flag(r, RSB_MASTER_UNCERTAIN);
 632		r->res_first_lkid = 0;
 633	}
 634
 635	if (from_local && (r->res_master_nodeid != our_nodeid)) {
 636		/* Because we have held no locks on this rsb,
 637		   res_master_nodeid could have become stale. */
 638		rsb_set_flag(r, RSB_MASTER_UNCERTAIN);
 639		r->res_first_lkid = 0;
 640	}
 641
 642	rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
 643	error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
 644	goto out_unlock;
 645
 646
 647 do_new:
 648	/*
 649	 * rsb not found
 650	 */
 651
 652	if (error == -EBADR && !create)
 653		goto out_unlock;
 654
 655	error = get_rsb_struct(ls, name, len, &r);
 656	if (error == -EAGAIN) {
 657		spin_unlock(&ls->ls_rsbtbl[b].lock);
 658		goto retry;
 659	}
 660	if (error)
 661		goto out_unlock;
 662
 663	r->res_hash = hash;
 664	r->res_bucket = b;
 665	r->res_dir_nodeid = dir_nodeid;
 666	kref_init(&r->res_ref);
 667
 668	if (from_dir) {
 669		/* want to see how often this happens */
 670		log_debug(ls, "find_rsb new from_dir %d recreate %s",
 671			  from_nodeid, r->res_name);
 672		r->res_master_nodeid = our_nodeid;
 673		r->res_nodeid = 0;
 674		goto out_add;
 675	}
 676
 677	if (from_other && (dir_nodeid != our_nodeid)) {
 678		/* should never happen */
 679		log_error(ls, "find_rsb new from_other %d dir %d our %d %s",
 680			  from_nodeid, dir_nodeid, our_nodeid, r->res_name);
 681		dlm_free_rsb(r);
 682		r = NULL;
 683		error = -ENOTBLK;
 684		goto out_unlock;
 685	}
 686
 687	if (from_other) {
 688		log_debug(ls, "find_rsb new from_other %d dir %d %s",
 689			  from_nodeid, dir_nodeid, r->res_name);
 690	}
 691
 692	if (dir_nodeid == our_nodeid) {
 693		/* When we are the dir nodeid, we can set the master
 694		   node immediately */
 695		r->res_master_nodeid = our_nodeid;
 696		r->res_nodeid = 0;
 697	} else {
 698		/* set_master will send_lookup to dir_nodeid */
 699		r->res_master_nodeid = 0;
 700		r->res_nodeid = -1;
 701	}
 702
 703 out_add:
 704	error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
 705 out_unlock:
 706	spin_unlock(&ls->ls_rsbtbl[b].lock);
 707 out:
 708	*r_ret = r;
 709	return error;
 710}
 711
 712/* During recovery, other nodes can send us new MSTCPY locks (from
 713   dlm_recover_locks) before we've made ourself master (in
 714   dlm_recover_masters). */
 715
 716static int find_rsb_nodir(struct dlm_ls *ls, const void *name, int len,
 717			  uint32_t hash, uint32_t b,
 718			  int dir_nodeid, int from_nodeid,
 719			  unsigned int flags, struct dlm_rsb **r_ret)
 720{
 721	struct dlm_rsb *r = NULL;
 722	int our_nodeid = dlm_our_nodeid();
 723	int recover = (flags & R_RECEIVE_RECOVER);
 724	int error;
 725
 726 retry:
 727	error = pre_rsb_struct(ls);
 728	if (error < 0)
 729		goto out;
 730
 731	spin_lock(&ls->ls_rsbtbl[b].lock);
 732
 733	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
 734	if (error)
 735		goto do_toss;
 736
 737	/*
 738	 * rsb is active, so we can't check master_nodeid without lock_rsb.
 739	 */
 740
 741	kref_get(&r->res_ref);
 742	goto out_unlock;
 743
 744
 745 do_toss:
 746	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
 747	if (error)
 748		goto do_new;
 749
 750	/*
 751	 * rsb found inactive. No other thread is using this rsb because
 752	 * it's on the toss list, so we can look at or update
 753	 * res_master_nodeid without lock_rsb.
 754	 */
 755
 756	if (!recover && (r->res_master_nodeid != our_nodeid) && from_nodeid) {
 757		/* our rsb is not master, and another node has sent us a
 758		   request; this should never happen */
 759		log_error(ls, "find_rsb toss from_nodeid %d master %d dir %d",
 760			  from_nodeid, r->res_master_nodeid, dir_nodeid);
 761		dlm_print_rsb(r);
 762		error = -ENOTBLK;
 763		goto out_unlock;
 764	}
 765
 766	if (!recover && (r->res_master_nodeid != our_nodeid) &&
 767	    (dir_nodeid == our_nodeid)) {
 768		/* our rsb is not master, and we are dir; may as well fix it;
 769		   this should never happen */
 770		log_error(ls, "find_rsb toss our %d master %d dir %d",
 771			  our_nodeid, r->res_master_nodeid, dir_nodeid);
 772		dlm_print_rsb(r);
 773		r->res_master_nodeid = our_nodeid;
 774		r->res_nodeid = 0;
 775	}
 776
 777	rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
 778	error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
 779	goto out_unlock;
 780
 781
 782 do_new:
 783	/*
 784	 * rsb not found
 785	 */
 786
 787	error = get_rsb_struct(ls, name, len, &r);
 788	if (error == -EAGAIN) {
 789		spin_unlock(&ls->ls_rsbtbl[b].lock);
 790		goto retry;
 791	}
 792	if (error)
 793		goto out_unlock;
 794
 795	r->res_hash = hash;
 796	r->res_bucket = b;
 797	r->res_dir_nodeid = dir_nodeid;
 798	r->res_master_nodeid = dir_nodeid;
 799	r->res_nodeid = (dir_nodeid == our_nodeid) ? 0 : dir_nodeid;
 800	kref_init(&r->res_ref);
 801
 802	error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
 803 out_unlock:
 804	spin_unlock(&ls->ls_rsbtbl[b].lock);
 805 out:
 806	*r_ret = r;
 807	return error;
 808}
 809
 810static int find_rsb(struct dlm_ls *ls, const void *name, int len,
 811		    int from_nodeid, unsigned int flags,
 812		    struct dlm_rsb **r_ret)
 813{
 814	uint32_t hash, b;
 815	int dir_nodeid;
 816
 817	if (len > DLM_RESNAME_MAXLEN)
 818		return -EINVAL;
 819
 820	hash = jhash(name, len, 0);
 821	b = hash & (ls->ls_rsbtbl_size - 1);
 822
 823	dir_nodeid = dlm_hash2nodeid(ls, hash);
 824
 825	if (dlm_no_directory(ls))
 826		return find_rsb_nodir(ls, name, len, hash, b, dir_nodeid,
 827				      from_nodeid, flags, r_ret);
 828	else
 829		return find_rsb_dir(ls, name, len, hash, b, dir_nodeid,
 830				      from_nodeid, flags, r_ret);
 831}
 832
 833/* we have received a request and found that res_master_nodeid != our_nodeid,
 834   so we need to return an error or make ourself the master */
 835
 836static int validate_master_nodeid(struct dlm_ls *ls, struct dlm_rsb *r,
 837				  int from_nodeid)
 838{
 839	if (dlm_no_directory(ls)) {
 840		log_error(ls, "find_rsb keep from_nodeid %d master %d dir %d",
 841			  from_nodeid, r->res_master_nodeid,
 842			  r->res_dir_nodeid);
 843		dlm_print_rsb(r);
 844		return -ENOTBLK;
 845	}
 846
 847	if (from_nodeid != r->res_dir_nodeid) {
 848		/* our rsb is not master, and another node (not the dir node)
 849	   	   has sent us a request.  this is much more common when our
 850	   	   master_nodeid is zero, so limit debug to non-zero.  */
 851
 852		if (r->res_master_nodeid) {
 853			log_debug(ls, "validate master from_other %d master %d "
 854				  "dir %d first %x %s", from_nodeid,
 855				  r->res_master_nodeid, r->res_dir_nodeid,
 856				  r->res_first_lkid, r->res_name);
 857		}
 858		return -ENOTBLK;
 859	} else {
 860		/* our rsb is not master, but the dir nodeid has sent us a
 861	   	   request; this could happen with master 0 / res_nodeid -1 */
 862
 863		if (r->res_master_nodeid) {
 864			log_error(ls, "validate master from_dir %d master %d "
 865				  "first %x %s",
 866				  from_nodeid, r->res_master_nodeid,
 867				  r->res_first_lkid, r->res_name);
 868		}
 869
 870		r->res_master_nodeid = dlm_our_nodeid();
 871		r->res_nodeid = 0;
 872		return 0;
 873	}
 874}
 875
 876static void __dlm_master_lookup(struct dlm_ls *ls, struct dlm_rsb *r, int our_nodeid,
 877				int from_nodeid, bool toss_list, unsigned int flags,
 878				int *r_nodeid, int *result)
 879{
 880	int fix_master = (flags & DLM_LU_RECOVER_MASTER);
 881	int from_master = (flags & DLM_LU_RECOVER_DIR);
 882
 883	if (r->res_dir_nodeid != our_nodeid) {
 884		/* should not happen, but may as well fix it and carry on */
 885		log_error(ls, "%s res_dir %d our %d %s", __func__,
 886			  r->res_dir_nodeid, our_nodeid, r->res_name);
 887		r->res_dir_nodeid = our_nodeid;
 888	}
 889
 890	if (fix_master && dlm_is_removed(ls, r->res_master_nodeid)) {
 891		/* Recovery uses this function to set a new master when
 892		 * the previous master failed.  Setting NEW_MASTER will
 893		 * force dlm_recover_masters to call recover_master on this
 894		 * rsb even though the res_nodeid is no longer removed.
 895		 */
 896
 897		r->res_master_nodeid = from_nodeid;
 898		r->res_nodeid = from_nodeid;
 899		rsb_set_flag(r, RSB_NEW_MASTER);
 900
 901		if (toss_list) {
 902			/* I don't think we should ever find it on toss list. */
 903			log_error(ls, "%s fix_master on toss", __func__);
 904			dlm_dump_rsb(r);
 905		}
 906	}
 907
 908	if (from_master && (r->res_master_nodeid != from_nodeid)) {
 909		/* this will happen if from_nodeid became master during
 910		 * a previous recovery cycle, and we aborted the previous
 911		 * cycle before recovering this master value
 912		 */
 913
 914		log_limit(ls, "%s from_master %d master_nodeid %d res_nodeid %d first %x %s",
 915			  __func__, from_nodeid, r->res_master_nodeid,
 916			  r->res_nodeid, r->res_first_lkid, r->res_name);
 917
 918		if (r->res_master_nodeid == our_nodeid) {
 919			log_error(ls, "from_master %d our_master", from_nodeid);
 920			dlm_dump_rsb(r);
 921			goto ret_assign;
 922		}
 923
 924		r->res_master_nodeid = from_nodeid;
 925		r->res_nodeid = from_nodeid;
 926		rsb_set_flag(r, RSB_NEW_MASTER);
 927	}
 928
 929	if (!r->res_master_nodeid) {
 930		/* this will happen if recovery happens while we're looking
 931		 * up the master for this rsb
 932		 */
 933
 934		log_debug(ls, "%s master 0 to %d first %x %s", __func__,
 935			  from_nodeid, r->res_first_lkid, r->res_name);
 936		r->res_master_nodeid = from_nodeid;
 937		r->res_nodeid = from_nodeid;
 938	}
 939
 940	if (!from_master && !fix_master &&
 941	    (r->res_master_nodeid == from_nodeid)) {
 942		/* this can happen when the master sends remove, the dir node
 943		 * finds the rsb on the keep list and ignores the remove,
 944		 * and the former master sends a lookup
 945		 */
 946
 947		log_limit(ls, "%s from master %d flags %x first %x %s",
 948			  __func__, from_nodeid, flags, r->res_first_lkid,
 949			  r->res_name);
 950	}
 951
 952 ret_assign:
 953	*r_nodeid = r->res_master_nodeid;
 954	if (result)
 955		*result = DLM_LU_MATCH;
 956}
 957
 958/*
 959 * We're the dir node for this res and another node wants to know the
 960 * master nodeid.  During normal operation (non recovery) this is only
 961 * called from receive_lookup(); master lookups when the local node is
 962 * the dir node are done by find_rsb().
 963 *
 964 * normal operation, we are the dir node for a resource
 965 * . _request_lock
 966 * . set_master
 967 * . send_lookup
 968 * . receive_lookup
 969 * . dlm_master_lookup flags 0
 970 *
 971 * recover directory, we are rebuilding dir for all resources
 972 * . dlm_recover_directory
 973 * . dlm_rcom_names
 974 *   remote node sends back the rsb names it is master of and we are dir of
 975 * . dlm_master_lookup RECOVER_DIR (fix_master 0, from_master 1)
 976 *   we either create new rsb setting remote node as master, or find existing
 977 *   rsb and set master to be the remote node.
 978 *
 979 * recover masters, we are finding the new master for resources
 980 * . dlm_recover_masters
 981 * . recover_master
 982 * . dlm_send_rcom_lookup
 983 * . receive_rcom_lookup
 984 * . dlm_master_lookup RECOVER_MASTER (fix_master 1, from_master 0)
 985 */
 986
 987int dlm_master_lookup(struct dlm_ls *ls, int from_nodeid, const char *name,
 988		      int len, unsigned int flags, int *r_nodeid, int *result)
 989{
 990	struct dlm_rsb *r = NULL;
 991	uint32_t hash, b;
 
 
 992	int our_nodeid = dlm_our_nodeid();
 993	int dir_nodeid, error;
 994
 995	if (len > DLM_RESNAME_MAXLEN)
 996		return -EINVAL;
 997
 998	if (from_nodeid == our_nodeid) {
 999		log_error(ls, "dlm_master_lookup from our_nodeid %d flags %x",
1000			  our_nodeid, flags);
1001		return -EINVAL;
1002	}
1003
1004	hash = jhash(name, len, 0);
1005	b = hash & (ls->ls_rsbtbl_size - 1);
1006
1007	dir_nodeid = dlm_hash2nodeid(ls, hash);
1008	if (dir_nodeid != our_nodeid) {
1009		log_error(ls, "dlm_master_lookup from %d dir %d our %d h %x %d",
1010			  from_nodeid, dir_nodeid, our_nodeid, hash,
1011			  ls->ls_num_nodes);
1012		*r_nodeid = -1;
1013		return -EINVAL;
1014	}
1015
1016 retry:
1017	error = pre_rsb_struct(ls);
1018	if (error < 0)
1019		return error;
1020
1021	spin_lock(&ls->ls_rsbtbl[b].lock);
1022	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
1023	if (!error) {
1024		/* because the rsb is active, we need to lock_rsb before
1025		 * checking/changing re_master_nodeid
1026		 */
1027
1028		hold_rsb(r);
1029		spin_unlock(&ls->ls_rsbtbl[b].lock);
1030		lock_rsb(r);
1031
1032		__dlm_master_lookup(ls, r, our_nodeid, from_nodeid, false,
1033				    flags, r_nodeid, result);
1034
1035		/* the rsb was active */
1036		unlock_rsb(r);
1037		put_rsb(r);
1038
1039		return 0;
1040	}
1041
1042	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
1043	if (error)
1044		goto not_found;
1045
1046	/* because the rsb is inactive (on toss list), it's not refcounted
1047	 * and lock_rsb is not used, but is protected by the rsbtbl lock
1048	 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1049
1050	__dlm_master_lookup(ls, r, our_nodeid, from_nodeid, true, flags,
1051			    r_nodeid, result);
 
 
1052
1053	r->res_toss_time = jiffies;
1054	/* the rsb was inactive (on toss list) */
1055	spin_unlock(&ls->ls_rsbtbl[b].lock);
 
1056
 
 
 
 
 
 
 
 
 
1057	return 0;
1058
1059 not_found:
1060	error = get_rsb_struct(ls, name, len, &r);
1061	if (error == -EAGAIN) {
1062		spin_unlock(&ls->ls_rsbtbl[b].lock);
1063		goto retry;
1064	}
1065	if (error)
1066		goto out_unlock;
1067
1068	r->res_hash = hash;
1069	r->res_bucket = b;
1070	r->res_dir_nodeid = our_nodeid;
1071	r->res_master_nodeid = from_nodeid;
1072	r->res_nodeid = from_nodeid;
1073	kref_init(&r->res_ref);
1074	r->res_toss_time = jiffies;
1075
1076	error = rsb_insert(r, &ls->ls_rsbtbl[b].toss);
1077	if (error) {
1078		/* should never happen */
1079		dlm_free_rsb(r);
1080		spin_unlock(&ls->ls_rsbtbl[b].lock);
1081		goto retry;
1082	}
1083
1084	if (result)
1085		*result = DLM_LU_ADD;
1086	*r_nodeid = from_nodeid;
 
1087 out_unlock:
1088	spin_unlock(&ls->ls_rsbtbl[b].lock);
1089	return error;
1090}
1091
1092static void dlm_dump_rsb_hash(struct dlm_ls *ls, uint32_t hash)
1093{
1094	struct rb_node *n;
1095	struct dlm_rsb *r;
1096	int i;
1097
1098	for (i = 0; i < ls->ls_rsbtbl_size; i++) {
1099		spin_lock(&ls->ls_rsbtbl[i].lock);
1100		for (n = rb_first(&ls->ls_rsbtbl[i].keep); n; n = rb_next(n)) {
1101			r = rb_entry(n, struct dlm_rsb, res_hashnode);
1102			if (r->res_hash == hash)
1103				dlm_dump_rsb(r);
1104		}
1105		spin_unlock(&ls->ls_rsbtbl[i].lock);
1106	}
1107}
1108
1109void dlm_dump_rsb_name(struct dlm_ls *ls, const char *name, int len)
1110{
1111	struct dlm_rsb *r = NULL;
1112	uint32_t hash, b;
1113	int error;
1114
1115	hash = jhash(name, len, 0);
1116	b = hash & (ls->ls_rsbtbl_size - 1);
1117
1118	spin_lock(&ls->ls_rsbtbl[b].lock);
1119	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
1120	if (!error)
1121		goto out_dump;
1122
1123	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
1124	if (error)
1125		goto out;
1126 out_dump:
1127	dlm_dump_rsb(r);
1128 out:
1129	spin_unlock(&ls->ls_rsbtbl[b].lock);
1130}
1131
1132static void toss_rsb(struct kref *kref)
1133{
1134	struct dlm_rsb *r = container_of(kref, struct dlm_rsb, res_ref);
1135	struct dlm_ls *ls = r->res_ls;
1136
1137	DLM_ASSERT(list_empty(&r->res_root_list), dlm_print_rsb(r););
1138	kref_init(&r->res_ref);
1139	rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[r->res_bucket].keep);
1140	rsb_insert(r, &ls->ls_rsbtbl[r->res_bucket].toss);
1141	r->res_toss_time = jiffies;
1142	set_bit(DLM_RTF_SHRINK_BIT, &ls->ls_rsbtbl[r->res_bucket].flags);
1143	if (r->res_lvbptr) {
1144		dlm_free_lvb(r->res_lvbptr);
1145		r->res_lvbptr = NULL;
1146	}
1147}
1148
1149/* See comment for unhold_lkb */
1150
1151static void unhold_rsb(struct dlm_rsb *r)
1152{
1153	int rv;
1154	rv = kref_put(&r->res_ref, toss_rsb);
1155	DLM_ASSERT(!rv, dlm_dump_rsb(r););
1156}
1157
1158static void kill_rsb(struct kref *kref)
1159{
1160	struct dlm_rsb *r = container_of(kref, struct dlm_rsb, res_ref);
1161
1162	/* All work is done after the return from kref_put() so we
1163	   can release the write_lock before the remove and free. */
1164
1165	DLM_ASSERT(list_empty(&r->res_lookup), dlm_dump_rsb(r););
1166	DLM_ASSERT(list_empty(&r->res_grantqueue), dlm_dump_rsb(r););
1167	DLM_ASSERT(list_empty(&r->res_convertqueue), dlm_dump_rsb(r););
1168	DLM_ASSERT(list_empty(&r->res_waitqueue), dlm_dump_rsb(r););
1169	DLM_ASSERT(list_empty(&r->res_root_list), dlm_dump_rsb(r););
1170	DLM_ASSERT(list_empty(&r->res_recover_list), dlm_dump_rsb(r););
1171}
1172
1173/* Attaching/detaching lkb's from rsb's is for rsb reference counting.
1174   The rsb must exist as long as any lkb's for it do. */
1175
1176static void attach_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb)
1177{
1178	hold_rsb(r);
1179	lkb->lkb_resource = r;
1180}
1181
1182static void detach_lkb(struct dlm_lkb *lkb)
1183{
1184	if (lkb->lkb_resource) {
1185		put_rsb(lkb->lkb_resource);
1186		lkb->lkb_resource = NULL;
1187	}
1188}
1189
1190static int _create_lkb(struct dlm_ls *ls, struct dlm_lkb **lkb_ret,
1191		       int start, int end)
1192{
1193	struct dlm_lkb *lkb;
1194	int rv;
1195
1196	lkb = dlm_allocate_lkb(ls);
1197	if (!lkb)
1198		return -ENOMEM;
1199
1200	lkb->lkb_last_bast_mode = -1;
1201	lkb->lkb_nodeid = -1;
1202	lkb->lkb_grmode = DLM_LOCK_IV;
1203	kref_init(&lkb->lkb_ref);
1204	INIT_LIST_HEAD(&lkb->lkb_ownqueue);
1205	INIT_LIST_HEAD(&lkb->lkb_rsb_lookup);
 
1206	INIT_LIST_HEAD(&lkb->lkb_cb_list);
1207	INIT_LIST_HEAD(&lkb->lkb_callbacks);
1208	spin_lock_init(&lkb->lkb_cb_lock);
1209	INIT_WORK(&lkb->lkb_cb_work, dlm_callback_work);
1210
1211	idr_preload(GFP_NOFS);
1212	spin_lock(&ls->ls_lkbidr_spin);
1213	rv = idr_alloc(&ls->ls_lkbidr, lkb, start, end, GFP_NOWAIT);
1214	if (rv >= 0)
1215		lkb->lkb_id = rv;
1216	spin_unlock(&ls->ls_lkbidr_spin);
1217	idr_preload_end();
1218
1219	if (rv < 0) {
1220		log_error(ls, "create_lkb idr error %d", rv);
1221		dlm_free_lkb(lkb);
1222		return rv;
1223	}
1224
1225	*lkb_ret = lkb;
1226	return 0;
1227}
1228
1229static int create_lkb(struct dlm_ls *ls, struct dlm_lkb **lkb_ret)
1230{
1231	return _create_lkb(ls, lkb_ret, 1, 0);
1232}
1233
1234static int find_lkb(struct dlm_ls *ls, uint32_t lkid, struct dlm_lkb **lkb_ret)
1235{
1236	struct dlm_lkb *lkb;
1237
1238	spin_lock(&ls->ls_lkbidr_spin);
1239	lkb = idr_find(&ls->ls_lkbidr, lkid);
1240	if (lkb)
1241		kref_get(&lkb->lkb_ref);
1242	spin_unlock(&ls->ls_lkbidr_spin);
1243
1244	*lkb_ret = lkb;
1245	return lkb ? 0 : -ENOENT;
1246}
1247
1248static void kill_lkb(struct kref *kref)
1249{
1250	struct dlm_lkb *lkb = container_of(kref, struct dlm_lkb, lkb_ref);
1251
1252	/* All work is done after the return from kref_put() so we
1253	   can release the write_lock before the detach_lkb */
1254
1255	DLM_ASSERT(!lkb->lkb_status, dlm_print_lkb(lkb););
1256}
1257
1258/* __put_lkb() is used when an lkb may not have an rsb attached to
1259   it so we need to provide the lockspace explicitly */
1260
1261static int __put_lkb(struct dlm_ls *ls, struct dlm_lkb *lkb)
1262{
1263	uint32_t lkid = lkb->lkb_id;
1264	int rv;
1265
1266	rv = kref_put_lock(&lkb->lkb_ref, kill_lkb,
1267			   &ls->ls_lkbidr_spin);
1268	if (rv) {
1269		idr_remove(&ls->ls_lkbidr, lkid);
1270		spin_unlock(&ls->ls_lkbidr_spin);
1271
1272		detach_lkb(lkb);
1273
1274		/* for local/process lkbs, lvbptr points to caller's lksb */
1275		if (lkb->lkb_lvbptr && is_master_copy(lkb))
1276			dlm_free_lvb(lkb->lkb_lvbptr);
1277		dlm_free_lkb(lkb);
 
 
 
 
1278	}
1279
1280	return rv;
1281}
1282
1283int dlm_put_lkb(struct dlm_lkb *lkb)
1284{
1285	struct dlm_ls *ls;
1286
1287	DLM_ASSERT(lkb->lkb_resource, dlm_print_lkb(lkb););
1288	DLM_ASSERT(lkb->lkb_resource->res_ls, dlm_print_lkb(lkb););
1289
1290	ls = lkb->lkb_resource->res_ls;
1291	return __put_lkb(ls, lkb);
1292}
1293
1294/* This is only called to add a reference when the code already holds
1295   a valid reference to the lkb, so there's no need for locking. */
1296
1297static inline void hold_lkb(struct dlm_lkb *lkb)
1298{
1299	kref_get(&lkb->lkb_ref);
1300}
1301
1302static void unhold_lkb_assert(struct kref *kref)
1303{
1304	struct dlm_lkb *lkb = container_of(kref, struct dlm_lkb, lkb_ref);
1305
1306	DLM_ASSERT(false, dlm_print_lkb(lkb););
1307}
1308
1309/* This is called when we need to remove a reference and are certain
1310   it's not the last ref.  e.g. del_lkb is always called between a
1311   find_lkb/put_lkb and is always the inverse of a previous add_lkb.
1312   put_lkb would work fine, but would involve unnecessary locking */
1313
1314static inline void unhold_lkb(struct dlm_lkb *lkb)
1315{
1316	kref_put(&lkb->lkb_ref, unhold_lkb_assert);
 
 
1317}
1318
1319static void lkb_add_ordered(struct list_head *new, struct list_head *head,
1320			    int mode)
1321{
1322	struct dlm_lkb *lkb = NULL, *iter;
1323
1324	list_for_each_entry(iter, head, lkb_statequeue)
1325		if (iter->lkb_rqmode < mode) {
1326			lkb = iter;
1327			list_add_tail(new, &iter->lkb_statequeue);
1328			break;
1329		}
1330
1331	if (!lkb)
1332		list_add_tail(new, head);
1333}
1334
1335/* add/remove lkb to rsb's grant/convert/wait queue */
1336
1337static void add_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb, int status)
1338{
1339	kref_get(&lkb->lkb_ref);
1340
1341	DLM_ASSERT(!lkb->lkb_status, dlm_print_lkb(lkb););
1342
1343	lkb->lkb_timestamp = ktime_get();
1344
1345	lkb->lkb_status = status;
1346
1347	switch (status) {
1348	case DLM_LKSTS_WAITING:
1349		if (lkb->lkb_exflags & DLM_LKF_HEADQUE)
1350			list_add(&lkb->lkb_statequeue, &r->res_waitqueue);
1351		else
1352			list_add_tail(&lkb->lkb_statequeue, &r->res_waitqueue);
1353		break;
1354	case DLM_LKSTS_GRANTED:
1355		/* convention says granted locks kept in order of grmode */
1356		lkb_add_ordered(&lkb->lkb_statequeue, &r->res_grantqueue,
1357				lkb->lkb_grmode);
1358		break;
1359	case DLM_LKSTS_CONVERT:
1360		if (lkb->lkb_exflags & DLM_LKF_HEADQUE)
1361			list_add(&lkb->lkb_statequeue, &r->res_convertqueue);
1362		else
1363			list_add_tail(&lkb->lkb_statequeue,
1364				      &r->res_convertqueue);
1365		break;
1366	default:
1367		DLM_ASSERT(0, dlm_print_lkb(lkb); printk("sts=%d\n", status););
1368	}
1369}
1370
1371static void del_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb)
1372{
1373	lkb->lkb_status = 0;
1374	list_del(&lkb->lkb_statequeue);
1375	unhold_lkb(lkb);
1376}
1377
1378static void move_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb, int sts)
1379{
1380	hold_lkb(lkb);
1381	del_lkb(r, lkb);
1382	add_lkb(r, lkb, sts);
1383	unhold_lkb(lkb);
1384}
1385
1386static int msg_reply_type(int mstype)
1387{
1388	switch (mstype) {
1389	case DLM_MSG_REQUEST:
1390		return DLM_MSG_REQUEST_REPLY;
1391	case DLM_MSG_CONVERT:
1392		return DLM_MSG_CONVERT_REPLY;
1393	case DLM_MSG_UNLOCK:
1394		return DLM_MSG_UNLOCK_REPLY;
1395	case DLM_MSG_CANCEL:
1396		return DLM_MSG_CANCEL_REPLY;
1397	case DLM_MSG_LOOKUP:
1398		return DLM_MSG_LOOKUP_REPLY;
1399	}
1400	return -1;
1401}
1402
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1403/* add/remove lkb from global waiters list of lkb's waiting for
1404   a reply from a remote node */
1405
1406static int add_to_waiters(struct dlm_lkb *lkb, int mstype, int to_nodeid)
1407{
1408	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1409	int error = 0;
1410	int wc;
1411
1412	mutex_lock(&ls->ls_waiters_mutex);
1413
1414	if (is_overlap_unlock(lkb) ||
1415	    (is_overlap_cancel(lkb) && (mstype == DLM_MSG_CANCEL))) {
1416		error = -EINVAL;
1417		goto out;
1418	}
1419
1420	if (lkb->lkb_wait_type || is_overlap_cancel(lkb)) {
1421		switch (mstype) {
1422		case DLM_MSG_UNLOCK:
1423			set_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags);
1424			break;
1425		case DLM_MSG_CANCEL:
1426			set_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags);
1427			break;
1428		default:
1429			error = -EBUSY;
1430			goto out;
1431		}
1432		wc = atomic_inc_return(&lkb->lkb_wait_count);
1433		hold_lkb(lkb);
1434
1435		log_debug(ls, "addwait %x cur %d overlap %d count %d f %x",
1436			  lkb->lkb_id, lkb->lkb_wait_type, mstype, wc,
1437			  dlm_iflags_val(lkb));
1438		goto out;
1439	}
1440
1441	wc = atomic_fetch_inc(&lkb->lkb_wait_count);
1442	DLM_ASSERT(!wc, dlm_print_lkb(lkb); printk("wait_count %d\n", wc););
 
 
 
1443	lkb->lkb_wait_type = mstype;
 
1444	lkb->lkb_wait_nodeid = to_nodeid; /* for debugging */
1445	hold_lkb(lkb);
1446	list_add(&lkb->lkb_wait_reply, &ls->ls_waiters);
1447 out:
1448	if (error)
1449		log_error(ls, "addwait error %x %d flags %x %d %d %s",
1450			  lkb->lkb_id, error, dlm_iflags_val(lkb), mstype,
1451			  lkb->lkb_wait_type, lkb->lkb_resource->res_name);
1452	mutex_unlock(&ls->ls_waiters_mutex);
1453	return error;
1454}
1455
1456/* We clear the RESEND flag because we might be taking an lkb off the waiters
1457   list as part of process_requestqueue (e.g. a lookup that has an optimized
1458   request reply on the requestqueue) between dlm_recover_waiters_pre() which
1459   set RESEND and dlm_recover_waiters_post() */
1460
1461static int _remove_from_waiters(struct dlm_lkb *lkb, int mstype,
1462				const struct dlm_message *ms)
1463{
1464	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1465	int overlap_done = 0;
1466
1467	if (mstype == DLM_MSG_UNLOCK_REPLY &&
1468	    test_and_clear_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags)) {
1469		log_debug(ls, "remwait %x unlock_reply overlap", lkb->lkb_id);
 
1470		overlap_done = 1;
1471		goto out_del;
1472	}
1473
1474	if (mstype == DLM_MSG_CANCEL_REPLY &&
1475	    test_and_clear_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags)) {
1476		log_debug(ls, "remwait %x cancel_reply overlap", lkb->lkb_id);
 
1477		overlap_done = 1;
1478		goto out_del;
1479	}
1480
1481	/* Cancel state was preemptively cleared by a successful convert,
1482	   see next comment, nothing to do. */
1483
1484	if ((mstype == DLM_MSG_CANCEL_REPLY) &&
1485	    (lkb->lkb_wait_type != DLM_MSG_CANCEL)) {
1486		log_debug(ls, "remwait %x cancel_reply wait_type %d",
1487			  lkb->lkb_id, lkb->lkb_wait_type);
1488		return -1;
1489	}
1490
1491	/* Remove for the convert reply, and premptively remove for the
1492	   cancel reply.  A convert has been granted while there's still
1493	   an outstanding cancel on it (the cancel is moot and the result
1494	   in the cancel reply should be 0).  We preempt the cancel reply
1495	   because the app gets the convert result and then can follow up
1496	   with another op, like convert.  This subsequent op would see the
1497	   lingering state of the cancel and fail with -EBUSY. */
1498
1499	if ((mstype == DLM_MSG_CONVERT_REPLY) &&
1500	    (lkb->lkb_wait_type == DLM_MSG_CONVERT) && ms && !ms->m_result &&
1501	    test_and_clear_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags)) {
1502		log_debug(ls, "remwait %x convert_reply zap overlap_cancel",
1503			  lkb->lkb_id);
1504		lkb->lkb_wait_type = 0;
1505		atomic_dec(&lkb->lkb_wait_count);
1506		unhold_lkb(lkb);
1507		goto out_del;
1508	}
1509
1510	/* N.B. type of reply may not always correspond to type of original
1511	   msg due to lookup->request optimization, verify others? */
1512
1513	if (lkb->lkb_wait_type) {
1514		lkb->lkb_wait_type = 0;
1515		goto out_del;
1516	}
1517
1518	log_error(ls, "remwait error %x remote %d %x msg %d flags %x no wait",
1519		  lkb->lkb_id, ms ? le32_to_cpu(ms->m_header.h_nodeid) : 0,
1520		  lkb->lkb_remid, mstype, dlm_iflags_val(lkb));
1521	return -1;
1522
1523 out_del:
1524	/* the force-unlock/cancel has completed and we haven't recvd a reply
1525	   to the op that was in progress prior to the unlock/cancel; we
1526	   give up on any reply to the earlier op.  FIXME: not sure when/how
1527	   this would happen */
1528
1529	if (overlap_done && lkb->lkb_wait_type) {
1530		log_error(ls, "remwait error %x reply %d wait_type %d overlap",
1531			  lkb->lkb_id, mstype, lkb->lkb_wait_type);
1532		atomic_dec(&lkb->lkb_wait_count);
1533		unhold_lkb(lkb);
1534		lkb->lkb_wait_type = 0;
1535	}
1536
1537	DLM_ASSERT(atomic_read(&lkb->lkb_wait_count), dlm_print_lkb(lkb););
1538
1539	clear_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags);
1540	if (atomic_dec_and_test(&lkb->lkb_wait_count))
 
1541		list_del_init(&lkb->lkb_wait_reply);
1542	unhold_lkb(lkb);
1543	return 0;
1544}
1545
1546static int remove_from_waiters(struct dlm_lkb *lkb, int mstype)
1547{
1548	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1549	int error;
1550
1551	mutex_lock(&ls->ls_waiters_mutex);
1552	error = _remove_from_waiters(lkb, mstype, NULL);
1553	mutex_unlock(&ls->ls_waiters_mutex);
1554	return error;
1555}
1556
1557/* Handles situations where we might be processing a "fake" or "local" reply in
1558   which we can't try to take waiters_mutex again. */
1559
1560static int remove_from_waiters_ms(struct dlm_lkb *lkb,
1561				  const struct dlm_message *ms, bool local)
1562{
1563	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
1564	int error;
1565
1566	if (!local)
1567		mutex_lock(&ls->ls_waiters_mutex);
1568	error = _remove_from_waiters(lkb, le32_to_cpu(ms->m_type), ms);
1569	if (!local)
1570		mutex_unlock(&ls->ls_waiters_mutex);
1571	return error;
1572}
1573
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1574static void shrink_bucket(struct dlm_ls *ls, int b)
1575{
1576	struct rb_node *n, *next;
1577	struct dlm_rsb *r;
1578	char *name;
1579	int our_nodeid = dlm_our_nodeid();
1580	int remote_count = 0;
1581	int need_shrink = 0;
1582	int i, len, rv;
1583
1584	memset(&ls->ls_remove_lens, 0, sizeof(int) * DLM_REMOVE_NAMES_MAX);
1585
1586	spin_lock(&ls->ls_rsbtbl[b].lock);
1587
1588	if (!test_bit(DLM_RTF_SHRINK_BIT, &ls->ls_rsbtbl[b].flags)) {
1589		spin_unlock(&ls->ls_rsbtbl[b].lock);
1590		return;
1591	}
1592
1593	for (n = rb_first(&ls->ls_rsbtbl[b].toss); n; n = next) {
1594		next = rb_next(n);
1595		r = rb_entry(n, struct dlm_rsb, res_hashnode);
1596
1597		/* If we're the directory record for this rsb, and
1598		   we're not the master of it, then we need to wait
1599		   for the master node to send us a dir remove for
1600		   before removing the dir record. */
1601
1602		if (!dlm_no_directory(ls) &&
1603		    (r->res_master_nodeid != our_nodeid) &&
1604		    (dlm_dir_nodeid(r) == our_nodeid)) {
1605			continue;
1606		}
1607
1608		need_shrink = 1;
1609
1610		if (!time_after_eq(jiffies, r->res_toss_time +
1611				   dlm_config.ci_toss_secs * HZ)) {
1612			continue;
1613		}
1614
1615		if (!dlm_no_directory(ls) &&
1616		    (r->res_master_nodeid == our_nodeid) &&
1617		    (dlm_dir_nodeid(r) != our_nodeid)) {
1618
1619			/* We're the master of this rsb but we're not
1620			   the directory record, so we need to tell the
1621			   dir node to remove the dir record. */
1622
1623			ls->ls_remove_lens[remote_count] = r->res_length;
1624			memcpy(ls->ls_remove_names[remote_count], r->res_name,
1625			       DLM_RESNAME_MAXLEN);
1626			remote_count++;
1627
1628			if (remote_count >= DLM_REMOVE_NAMES_MAX)
1629				break;
1630			continue;
1631		}
1632
1633		if (!kref_put(&r->res_ref, kill_rsb)) {
1634			log_error(ls, "tossed rsb in use %s", r->res_name);
1635			continue;
1636		}
1637
1638		rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
1639		dlm_free_rsb(r);
1640	}
1641
1642	if (need_shrink)
1643		set_bit(DLM_RTF_SHRINK_BIT, &ls->ls_rsbtbl[b].flags);
1644	else
1645		clear_bit(DLM_RTF_SHRINK_BIT, &ls->ls_rsbtbl[b].flags);
1646	spin_unlock(&ls->ls_rsbtbl[b].lock);
1647
1648	/*
1649	 * While searching for rsb's to free, we found some that require
1650	 * remote removal.  We leave them in place and find them again here
1651	 * so there is a very small gap between removing them from the toss
1652	 * list and sending the removal.  Keeping this gap small is
1653	 * important to keep us (the master node) from being out of sync
1654	 * with the remote dir node for very long.
 
 
 
 
 
1655	 */
1656
1657	for (i = 0; i < remote_count; i++) {
1658		name = ls->ls_remove_names[i];
1659		len = ls->ls_remove_lens[i];
1660
1661		spin_lock(&ls->ls_rsbtbl[b].lock);
1662		rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
1663		if (rv) {
1664			spin_unlock(&ls->ls_rsbtbl[b].lock);
1665			log_debug(ls, "remove_name not toss %s", name);
1666			continue;
1667		}
1668
1669		if (r->res_master_nodeid != our_nodeid) {
1670			spin_unlock(&ls->ls_rsbtbl[b].lock);
1671			log_debug(ls, "remove_name master %d dir %d our %d %s",
1672				  r->res_master_nodeid, r->res_dir_nodeid,
1673				  our_nodeid, name);
1674			continue;
1675		}
1676
1677		if (r->res_dir_nodeid == our_nodeid) {
1678			/* should never happen */
1679			spin_unlock(&ls->ls_rsbtbl[b].lock);
1680			log_error(ls, "remove_name dir %d master %d our %d %s",
1681				  r->res_dir_nodeid, r->res_master_nodeid,
1682				  our_nodeid, name);
1683			continue;
1684		}
1685
1686		if (!time_after_eq(jiffies, r->res_toss_time +
1687				   dlm_config.ci_toss_secs * HZ)) {
1688			spin_unlock(&ls->ls_rsbtbl[b].lock);
1689			log_debug(ls, "remove_name toss_time %lu now %lu %s",
1690				  r->res_toss_time, jiffies, name);
1691			continue;
1692		}
1693
1694		if (!kref_put(&r->res_ref, kill_rsb)) {
1695			spin_unlock(&ls->ls_rsbtbl[b].lock);
1696			log_error(ls, "remove_name in use %s", name);
1697			continue;
1698		}
1699
1700		rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
 
 
 
 
 
 
 
 
1701		send_remove(r);
1702		spin_unlock(&ls->ls_rsbtbl[b].lock);
 
 
 
 
 
1703
1704		dlm_free_rsb(r);
1705	}
1706}
1707
1708void dlm_scan_rsbs(struct dlm_ls *ls)
1709{
1710	int i;
1711
1712	for (i = 0; i < ls->ls_rsbtbl_size; i++) {
1713		shrink_bucket(ls, i);
1714		if (dlm_locking_stopped(ls))
1715			break;
1716		cond_resched();
1717	}
1718}
1719
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1720/* lkb is master or local copy */
1721
1722static void set_lvb_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1723{
1724	int b, len = r->res_ls->ls_lvblen;
1725
1726	/* b=1 lvb returned to caller
1727	   b=0 lvb written to rsb or invalidated
1728	   b=-1 do nothing */
1729
1730	b =  dlm_lvb_operations[lkb->lkb_grmode + 1][lkb->lkb_rqmode + 1];
1731
1732	if (b == 1) {
1733		if (!lkb->lkb_lvbptr)
1734			return;
1735
1736		if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
1737			return;
1738
1739		if (!r->res_lvbptr)
1740			return;
1741
1742		memcpy(lkb->lkb_lvbptr, r->res_lvbptr, len);
1743		lkb->lkb_lvbseq = r->res_lvbseq;
1744
1745	} else if (b == 0) {
1746		if (lkb->lkb_exflags & DLM_LKF_IVVALBLK) {
1747			rsb_set_flag(r, RSB_VALNOTVALID);
1748			return;
1749		}
1750
1751		if (!lkb->lkb_lvbptr)
1752			return;
1753
1754		if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
1755			return;
1756
1757		if (!r->res_lvbptr)
1758			r->res_lvbptr = dlm_allocate_lvb(r->res_ls);
1759
1760		if (!r->res_lvbptr)
1761			return;
1762
1763		memcpy(r->res_lvbptr, lkb->lkb_lvbptr, len);
1764		r->res_lvbseq++;
1765		lkb->lkb_lvbseq = r->res_lvbseq;
1766		rsb_clear_flag(r, RSB_VALNOTVALID);
1767	}
1768
1769	if (rsb_flag(r, RSB_VALNOTVALID))
1770		set_bit(DLM_SBF_VALNOTVALID_BIT, &lkb->lkb_sbflags);
1771}
1772
1773static void set_lvb_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1774{
1775	if (lkb->lkb_grmode < DLM_LOCK_PW)
1776		return;
1777
1778	if (lkb->lkb_exflags & DLM_LKF_IVVALBLK) {
1779		rsb_set_flag(r, RSB_VALNOTVALID);
1780		return;
1781	}
1782
1783	if (!lkb->lkb_lvbptr)
1784		return;
1785
1786	if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
1787		return;
1788
1789	if (!r->res_lvbptr)
1790		r->res_lvbptr = dlm_allocate_lvb(r->res_ls);
1791
1792	if (!r->res_lvbptr)
1793		return;
1794
1795	memcpy(r->res_lvbptr, lkb->lkb_lvbptr, r->res_ls->ls_lvblen);
1796	r->res_lvbseq++;
1797	rsb_clear_flag(r, RSB_VALNOTVALID);
1798}
1799
1800/* lkb is process copy (pc) */
1801
1802static void set_lvb_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb,
1803			    const struct dlm_message *ms)
1804{
1805	int b;
1806
1807	if (!lkb->lkb_lvbptr)
1808		return;
1809
1810	if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
1811		return;
1812
1813	b = dlm_lvb_operations[lkb->lkb_grmode + 1][lkb->lkb_rqmode + 1];
1814	if (b == 1) {
1815		int len = receive_extralen(ms);
1816		if (len > r->res_ls->ls_lvblen)
1817			len = r->res_ls->ls_lvblen;
1818		memcpy(lkb->lkb_lvbptr, ms->m_extra, len);
1819		lkb->lkb_lvbseq = le32_to_cpu(ms->m_lvbseq);
1820	}
1821}
1822
1823/* Manipulate lkb's on rsb's convert/granted/waiting queues
1824   remove_lock -- used for unlock, removes lkb from granted
1825   revert_lock -- used for cancel, moves lkb from convert to granted
1826   grant_lock  -- used for request and convert, adds lkb to granted or
1827                  moves lkb from convert or waiting to granted
1828
1829   Each of these is used for master or local copy lkb's.  There is
1830   also a _pc() variation used to make the corresponding change on
1831   a process copy (pc) lkb. */
1832
1833static void _remove_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1834{
1835	del_lkb(r, lkb);
1836	lkb->lkb_grmode = DLM_LOCK_IV;
1837	/* this unhold undoes the original ref from create_lkb()
1838	   so this leads to the lkb being freed */
1839	unhold_lkb(lkb);
1840}
1841
1842static void remove_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1843{
1844	set_lvb_unlock(r, lkb);
1845	_remove_lock(r, lkb);
1846}
1847
1848static void remove_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb)
1849{
1850	_remove_lock(r, lkb);
1851}
1852
1853/* returns: 0 did nothing
1854	    1 moved lock to granted
1855	   -1 removed lock */
1856
1857static int revert_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1858{
1859	int rv = 0;
1860
1861	lkb->lkb_rqmode = DLM_LOCK_IV;
1862
1863	switch (lkb->lkb_status) {
1864	case DLM_LKSTS_GRANTED:
1865		break;
1866	case DLM_LKSTS_CONVERT:
1867		move_lkb(r, lkb, DLM_LKSTS_GRANTED);
1868		rv = 1;
1869		break;
1870	case DLM_LKSTS_WAITING:
1871		del_lkb(r, lkb);
1872		lkb->lkb_grmode = DLM_LOCK_IV;
1873		/* this unhold undoes the original ref from create_lkb()
1874		   so this leads to the lkb being freed */
1875		unhold_lkb(lkb);
1876		rv = -1;
1877		break;
1878	default:
1879		log_print("invalid status for revert %d", lkb->lkb_status);
1880	}
1881	return rv;
1882}
1883
1884static int revert_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb)
1885{
1886	return revert_lock(r, lkb);
1887}
1888
1889static void _grant_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1890{
1891	if (lkb->lkb_grmode != lkb->lkb_rqmode) {
1892		lkb->lkb_grmode = lkb->lkb_rqmode;
1893		if (lkb->lkb_status)
1894			move_lkb(r, lkb, DLM_LKSTS_GRANTED);
1895		else
1896			add_lkb(r, lkb, DLM_LKSTS_GRANTED);
1897	}
1898
1899	lkb->lkb_rqmode = DLM_LOCK_IV;
1900	lkb->lkb_highbast = 0;
1901}
1902
1903static void grant_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
1904{
1905	set_lvb_lock(r, lkb);
1906	_grant_lock(r, lkb);
1907}
1908
1909static void grant_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb,
1910			  const struct dlm_message *ms)
1911{
1912	set_lvb_lock_pc(r, lkb, ms);
1913	_grant_lock(r, lkb);
1914}
1915
1916/* called by grant_pending_locks() which means an async grant message must
1917   be sent to the requesting node in addition to granting the lock if the
1918   lkb belongs to a remote node. */
1919
1920static void grant_lock_pending(struct dlm_rsb *r, struct dlm_lkb *lkb)
1921{
1922	grant_lock(r, lkb);
1923	if (is_master_copy(lkb))
1924		send_grant(r, lkb);
1925	else
1926		queue_cast(r, lkb, 0);
1927}
1928
1929/* The special CONVDEADLK, ALTPR and ALTCW flags allow the master to
1930   change the granted/requested modes.  We're munging things accordingly in
1931   the process copy.
1932   CONVDEADLK: our grmode may have been forced down to NL to resolve a
1933   conversion deadlock
1934   ALTPR/ALTCW: our rqmode may have been changed to PR or CW to become
1935   compatible with other granted locks */
1936
1937static void munge_demoted(struct dlm_lkb *lkb)
1938{
1939	if (lkb->lkb_rqmode == DLM_LOCK_IV || lkb->lkb_grmode == DLM_LOCK_IV) {
1940		log_print("munge_demoted %x invalid modes gr %d rq %d",
1941			  lkb->lkb_id, lkb->lkb_grmode, lkb->lkb_rqmode);
1942		return;
1943	}
1944
1945	lkb->lkb_grmode = DLM_LOCK_NL;
1946}
1947
1948static void munge_altmode(struct dlm_lkb *lkb, const struct dlm_message *ms)
1949{
1950	if (ms->m_type != cpu_to_le32(DLM_MSG_REQUEST_REPLY) &&
1951	    ms->m_type != cpu_to_le32(DLM_MSG_GRANT)) {
1952		log_print("munge_altmode %x invalid reply type %d",
1953			  lkb->lkb_id, le32_to_cpu(ms->m_type));
1954		return;
1955	}
1956
1957	if (lkb->lkb_exflags & DLM_LKF_ALTPR)
1958		lkb->lkb_rqmode = DLM_LOCK_PR;
1959	else if (lkb->lkb_exflags & DLM_LKF_ALTCW)
1960		lkb->lkb_rqmode = DLM_LOCK_CW;
1961	else {
1962		log_print("munge_altmode invalid exflags %x", lkb->lkb_exflags);
1963		dlm_print_lkb(lkb);
1964	}
1965}
1966
1967static inline int first_in_list(struct dlm_lkb *lkb, struct list_head *head)
1968{
1969	struct dlm_lkb *first = list_entry(head->next, struct dlm_lkb,
1970					   lkb_statequeue);
1971	if (lkb->lkb_id == first->lkb_id)
1972		return 1;
1973
1974	return 0;
1975}
1976
1977/* Check if the given lkb conflicts with another lkb on the queue. */
1978
1979static int queue_conflict(struct list_head *head, struct dlm_lkb *lkb)
1980{
1981	struct dlm_lkb *this;
1982
1983	list_for_each_entry(this, head, lkb_statequeue) {
1984		if (this == lkb)
1985			continue;
1986		if (!modes_compat(this, lkb))
1987			return 1;
1988	}
1989	return 0;
1990}
1991
1992/*
1993 * "A conversion deadlock arises with a pair of lock requests in the converting
1994 * queue for one resource.  The granted mode of each lock blocks the requested
1995 * mode of the other lock."
1996 *
1997 * Part 2: if the granted mode of lkb is preventing an earlier lkb in the
1998 * convert queue from being granted, then deadlk/demote lkb.
1999 *
2000 * Example:
2001 * Granted Queue: empty
2002 * Convert Queue: NL->EX (first lock)
2003 *                PR->EX (second lock)
2004 *
2005 * The first lock can't be granted because of the granted mode of the second
2006 * lock and the second lock can't be granted because it's not first in the
2007 * list.  We either cancel lkb's conversion (PR->EX) and return EDEADLK, or we
2008 * demote the granted mode of lkb (from PR to NL) if it has the CONVDEADLK
2009 * flag set and return DEMOTED in the lksb flags.
2010 *
2011 * Originally, this function detected conv-deadlk in a more limited scope:
2012 * - if !modes_compat(lkb1, lkb2) && !modes_compat(lkb2, lkb1), or
2013 * - if lkb1 was the first entry in the queue (not just earlier), and was
2014 *   blocked by the granted mode of lkb2, and there was nothing on the
2015 *   granted queue preventing lkb1 from being granted immediately, i.e.
2016 *   lkb2 was the only thing preventing lkb1 from being granted.
2017 *
2018 * That second condition meant we'd only say there was conv-deadlk if
2019 * resolving it (by demotion) would lead to the first lock on the convert
2020 * queue being granted right away.  It allowed conversion deadlocks to exist
2021 * between locks on the convert queue while they couldn't be granted anyway.
2022 *
2023 * Now, we detect and take action on conversion deadlocks immediately when
2024 * they're created, even if they may not be immediately consequential.  If
2025 * lkb1 exists anywhere in the convert queue and lkb2 comes in with a granted
2026 * mode that would prevent lkb1's conversion from being granted, we do a
2027 * deadlk/demote on lkb2 right away and don't let it onto the convert queue.
2028 * I think this means that the lkb_is_ahead condition below should always
2029 * be zero, i.e. there will never be conv-deadlk between two locks that are
2030 * both already on the convert queue.
2031 */
2032
2033static int conversion_deadlock_detect(struct dlm_rsb *r, struct dlm_lkb *lkb2)
2034{
2035	struct dlm_lkb *lkb1;
2036	int lkb_is_ahead = 0;
2037
2038	list_for_each_entry(lkb1, &r->res_convertqueue, lkb_statequeue) {
2039		if (lkb1 == lkb2) {
2040			lkb_is_ahead = 1;
2041			continue;
2042		}
2043
2044		if (!lkb_is_ahead) {
2045			if (!modes_compat(lkb2, lkb1))
2046				return 1;
2047		} else {
2048			if (!modes_compat(lkb2, lkb1) &&
2049			    !modes_compat(lkb1, lkb2))
2050				return 1;
2051		}
2052	}
2053	return 0;
2054}
2055
2056/*
2057 * Return 1 if the lock can be granted, 0 otherwise.
2058 * Also detect and resolve conversion deadlocks.
2059 *
2060 * lkb is the lock to be granted
2061 *
2062 * now is 1 if the function is being called in the context of the
2063 * immediate request, it is 0 if called later, after the lock has been
2064 * queued.
2065 *
2066 * recover is 1 if dlm_recover_grant() is trying to grant conversions
2067 * after recovery.
2068 *
2069 * References are from chapter 6 of "VAXcluster Principles" by Roy Davis
2070 */
2071
2072static int _can_be_granted(struct dlm_rsb *r, struct dlm_lkb *lkb, int now,
2073			   int recover)
2074{
2075	int8_t conv = (lkb->lkb_grmode != DLM_LOCK_IV);
2076
2077	/*
2078	 * 6-10: Version 5.4 introduced an option to address the phenomenon of
2079	 * a new request for a NL mode lock being blocked.
2080	 *
2081	 * 6-11: If the optional EXPEDITE flag is used with the new NL mode
2082	 * request, then it would be granted.  In essence, the use of this flag
2083	 * tells the Lock Manager to expedite theis request by not considering
2084	 * what may be in the CONVERTING or WAITING queues...  As of this
2085	 * writing, the EXPEDITE flag can be used only with new requests for NL
2086	 * mode locks.  This flag is not valid for conversion requests.
2087	 *
2088	 * A shortcut.  Earlier checks return an error if EXPEDITE is used in a
2089	 * conversion or used with a non-NL requested mode.  We also know an
2090	 * EXPEDITE request is always granted immediately, so now must always
2091	 * be 1.  The full condition to grant an expedite request: (now &&
2092	 * !conv && lkb->rqmode == DLM_LOCK_NL && (flags & EXPEDITE)) can
2093	 * therefore be shortened to just checking the flag.
2094	 */
2095
2096	if (lkb->lkb_exflags & DLM_LKF_EXPEDITE)
2097		return 1;
2098
2099	/*
2100	 * A shortcut. Without this, !queue_conflict(grantqueue, lkb) would be
2101	 * added to the remaining conditions.
2102	 */
2103
2104	if (queue_conflict(&r->res_grantqueue, lkb))
2105		return 0;
2106
2107	/*
2108	 * 6-3: By default, a conversion request is immediately granted if the
2109	 * requested mode is compatible with the modes of all other granted
2110	 * locks
2111	 */
2112
2113	if (queue_conflict(&r->res_convertqueue, lkb))
2114		return 0;
2115
2116	/*
2117	 * The RECOVER_GRANT flag means dlm_recover_grant() is granting
2118	 * locks for a recovered rsb, on which lkb's have been rebuilt.
2119	 * The lkb's may have been rebuilt on the queues in a different
2120	 * order than they were in on the previous master.  So, granting
2121	 * queued conversions in order after recovery doesn't make sense
2122	 * since the order hasn't been preserved anyway.  The new order
2123	 * could also have created a new "in place" conversion deadlock.
2124	 * (e.g. old, failed master held granted EX, with PR->EX, NL->EX.
2125	 * After recovery, there would be no granted locks, and possibly
2126	 * NL->EX, PR->EX, an in-place conversion deadlock.)  So, after
2127	 * recovery, grant conversions without considering order.
2128	 */
2129
2130	if (conv && recover)
2131		return 1;
2132
2133	/*
2134	 * 6-5: But the default algorithm for deciding whether to grant or
2135	 * queue conversion requests does not by itself guarantee that such
2136	 * requests are serviced on a "first come first serve" basis.  This, in
2137	 * turn, can lead to a phenomenon known as "indefinate postponement".
2138	 *
2139	 * 6-7: This issue is dealt with by using the optional QUECVT flag with
2140	 * the system service employed to request a lock conversion.  This flag
2141	 * forces certain conversion requests to be queued, even if they are
2142	 * compatible with the granted modes of other locks on the same
2143	 * resource.  Thus, the use of this flag results in conversion requests
2144	 * being ordered on a "first come first servce" basis.
2145	 *
2146	 * DCT: This condition is all about new conversions being able to occur
2147	 * "in place" while the lock remains on the granted queue (assuming
2148	 * nothing else conflicts.)  IOW if QUECVT isn't set, a conversion
2149	 * doesn't _have_ to go onto the convert queue where it's processed in
2150	 * order.  The "now" variable is necessary to distinguish converts
2151	 * being received and processed for the first time now, because once a
2152	 * convert is moved to the conversion queue the condition below applies
2153	 * requiring fifo granting.
2154	 */
2155
2156	if (now && conv && !(lkb->lkb_exflags & DLM_LKF_QUECVT))
2157		return 1;
2158
2159	/*
2160	 * Even if the convert is compat with all granted locks,
2161	 * QUECVT forces it behind other locks on the convert queue.
2162	 */
2163
2164	if (now && conv && (lkb->lkb_exflags & DLM_LKF_QUECVT)) {
2165		if (list_empty(&r->res_convertqueue))
2166			return 1;
2167		else
2168			return 0;
2169	}
2170
2171	/*
2172	 * The NOORDER flag is set to avoid the standard vms rules on grant
2173	 * order.
2174	 */
2175
2176	if (lkb->lkb_exflags & DLM_LKF_NOORDER)
2177		return 1;
2178
2179	/*
2180	 * 6-3: Once in that queue [CONVERTING], a conversion request cannot be
2181	 * granted until all other conversion requests ahead of it are granted
2182	 * and/or canceled.
2183	 */
2184
2185	if (!now && conv && first_in_list(lkb, &r->res_convertqueue))
2186		return 1;
2187
2188	/*
2189	 * 6-4: By default, a new request is immediately granted only if all
2190	 * three of the following conditions are satisfied when the request is
2191	 * issued:
2192	 * - The queue of ungranted conversion requests for the resource is
2193	 *   empty.
2194	 * - The queue of ungranted new requests for the resource is empty.
2195	 * - The mode of the new request is compatible with the most
2196	 *   restrictive mode of all granted locks on the resource.
2197	 */
2198
2199	if (now && !conv && list_empty(&r->res_convertqueue) &&
2200	    list_empty(&r->res_waitqueue))
2201		return 1;
2202
2203	/*
2204	 * 6-4: Once a lock request is in the queue of ungranted new requests,
2205	 * it cannot be granted until the queue of ungranted conversion
2206	 * requests is empty, all ungranted new requests ahead of it are
2207	 * granted and/or canceled, and it is compatible with the granted mode
2208	 * of the most restrictive lock granted on the resource.
2209	 */
2210
2211	if (!now && !conv && list_empty(&r->res_convertqueue) &&
2212	    first_in_list(lkb, &r->res_waitqueue))
2213		return 1;
2214
2215	return 0;
2216}
2217
2218static int can_be_granted(struct dlm_rsb *r, struct dlm_lkb *lkb, int now,
2219			  int recover, int *err)
2220{
2221	int rv;
2222	int8_t alt = 0, rqmode = lkb->lkb_rqmode;
2223	int8_t is_convert = (lkb->lkb_grmode != DLM_LOCK_IV);
2224
2225	if (err)
2226		*err = 0;
2227
2228	rv = _can_be_granted(r, lkb, now, recover);
2229	if (rv)
2230		goto out;
2231
2232	/*
2233	 * The CONVDEADLK flag is non-standard and tells the dlm to resolve
2234	 * conversion deadlocks by demoting grmode to NL, otherwise the dlm
2235	 * cancels one of the locks.
2236	 */
2237
2238	if (is_convert && can_be_queued(lkb) &&
2239	    conversion_deadlock_detect(r, lkb)) {
2240		if (lkb->lkb_exflags & DLM_LKF_CONVDEADLK) {
2241			lkb->lkb_grmode = DLM_LOCK_NL;
2242			set_bit(DLM_SBF_DEMOTED_BIT, &lkb->lkb_sbflags);
2243		} else if (err) {
2244			*err = -EDEADLK;
2245		} else {
2246			log_print("can_be_granted deadlock %x now %d",
2247				  lkb->lkb_id, now);
2248			dlm_dump_rsb(r);
2249		}
2250		goto out;
2251	}
2252
2253	/*
2254	 * The ALTPR and ALTCW flags are non-standard and tell the dlm to try
2255	 * to grant a request in a mode other than the normal rqmode.  It's a
2256	 * simple way to provide a big optimization to applications that can
2257	 * use them.
2258	 */
2259
2260	if (rqmode != DLM_LOCK_PR && (lkb->lkb_exflags & DLM_LKF_ALTPR))
2261		alt = DLM_LOCK_PR;
2262	else if (rqmode != DLM_LOCK_CW && (lkb->lkb_exflags & DLM_LKF_ALTCW))
2263		alt = DLM_LOCK_CW;
2264
2265	if (alt) {
2266		lkb->lkb_rqmode = alt;
2267		rv = _can_be_granted(r, lkb, now, 0);
2268		if (rv)
2269			set_bit(DLM_SBF_ALTMODE_BIT, &lkb->lkb_sbflags);
2270		else
2271			lkb->lkb_rqmode = rqmode;
2272	}
2273 out:
2274	return rv;
2275}
2276
2277/* Returns the highest requested mode of all blocked conversions; sets
2278   cw if there's a blocked conversion to DLM_LOCK_CW. */
2279
2280static int grant_pending_convert(struct dlm_rsb *r, int high, int *cw,
2281				 unsigned int *count)
2282{
2283	struct dlm_lkb *lkb, *s;
2284	int recover = rsb_flag(r, RSB_RECOVER_GRANT);
2285	int hi, demoted, quit, grant_restart, demote_restart;
2286	int deadlk;
2287
2288	quit = 0;
2289 restart:
2290	grant_restart = 0;
2291	demote_restart = 0;
2292	hi = DLM_LOCK_IV;
2293
2294	list_for_each_entry_safe(lkb, s, &r->res_convertqueue, lkb_statequeue) {
2295		demoted = is_demoted(lkb);
2296		deadlk = 0;
2297
2298		if (can_be_granted(r, lkb, 0, recover, &deadlk)) {
2299			grant_lock_pending(r, lkb);
2300			grant_restart = 1;
2301			if (count)
2302				(*count)++;
2303			continue;
2304		}
2305
2306		if (!demoted && is_demoted(lkb)) {
2307			log_print("WARN: pending demoted %x node %d %s",
2308				  lkb->lkb_id, lkb->lkb_nodeid, r->res_name);
2309			demote_restart = 1;
2310			continue;
2311		}
2312
2313		if (deadlk) {
2314			/*
2315			 * If DLM_LKB_NODLKWT flag is set and conversion
2316			 * deadlock is detected, we request blocking AST and
2317			 * down (or cancel) conversion.
2318			 */
2319			if (lkb->lkb_exflags & DLM_LKF_NODLCKWT) {
2320				if (lkb->lkb_highbast < lkb->lkb_rqmode) {
2321					queue_bast(r, lkb, lkb->lkb_rqmode);
2322					lkb->lkb_highbast = lkb->lkb_rqmode;
2323				}
2324			} else {
2325				log_print("WARN: pending deadlock %x node %d %s",
2326					  lkb->lkb_id, lkb->lkb_nodeid,
2327					  r->res_name);
2328				dlm_dump_rsb(r);
2329			}
2330			continue;
2331		}
2332
2333		hi = max_t(int, lkb->lkb_rqmode, hi);
2334
2335		if (cw && lkb->lkb_rqmode == DLM_LOCK_CW)
2336			*cw = 1;
2337	}
2338
2339	if (grant_restart)
2340		goto restart;
2341	if (demote_restart && !quit) {
2342		quit = 1;
2343		goto restart;
2344	}
2345
2346	return max_t(int, high, hi);
2347}
2348
2349static int grant_pending_wait(struct dlm_rsb *r, int high, int *cw,
2350			      unsigned int *count)
2351{
2352	struct dlm_lkb *lkb, *s;
2353
2354	list_for_each_entry_safe(lkb, s, &r->res_waitqueue, lkb_statequeue) {
2355		if (can_be_granted(r, lkb, 0, 0, NULL)) {
2356			grant_lock_pending(r, lkb);
2357			if (count)
2358				(*count)++;
2359		} else {
2360			high = max_t(int, lkb->lkb_rqmode, high);
2361			if (lkb->lkb_rqmode == DLM_LOCK_CW)
2362				*cw = 1;
2363		}
2364	}
2365
2366	return high;
2367}
2368
2369/* cw of 1 means there's a lock with a rqmode of DLM_LOCK_CW that's blocked
2370   on either the convert or waiting queue.
2371   high is the largest rqmode of all locks blocked on the convert or
2372   waiting queue. */
2373
2374static int lock_requires_bast(struct dlm_lkb *gr, int high, int cw)
2375{
2376	if (gr->lkb_grmode == DLM_LOCK_PR && cw) {
2377		if (gr->lkb_highbast < DLM_LOCK_EX)
2378			return 1;
2379		return 0;
2380	}
2381
2382	if (gr->lkb_highbast < high &&
2383	    !__dlm_compat_matrix[gr->lkb_grmode+1][high+1])
2384		return 1;
2385	return 0;
2386}
2387
2388static void grant_pending_locks(struct dlm_rsb *r, unsigned int *count)
2389{
2390	struct dlm_lkb *lkb, *s;
2391	int high = DLM_LOCK_IV;
2392	int cw = 0;
2393
2394	if (!is_master(r)) {
2395		log_print("grant_pending_locks r nodeid %d", r->res_nodeid);
2396		dlm_dump_rsb(r);
2397		return;
2398	}
2399
2400	high = grant_pending_convert(r, high, &cw, count);
2401	high = grant_pending_wait(r, high, &cw, count);
2402
2403	if (high == DLM_LOCK_IV)
2404		return;
2405
2406	/*
2407	 * If there are locks left on the wait/convert queue then send blocking
2408	 * ASTs to granted locks based on the largest requested mode (high)
2409	 * found above.
2410	 */
2411
2412	list_for_each_entry_safe(lkb, s, &r->res_grantqueue, lkb_statequeue) {
2413		if (lkb->lkb_bastfn && lock_requires_bast(lkb, high, cw)) {
2414			if (cw && high == DLM_LOCK_PR &&
2415			    lkb->lkb_grmode == DLM_LOCK_PR)
2416				queue_bast(r, lkb, DLM_LOCK_CW);
2417			else
2418				queue_bast(r, lkb, high);
2419			lkb->lkb_highbast = high;
2420		}
2421	}
2422}
2423
2424static int modes_require_bast(struct dlm_lkb *gr, struct dlm_lkb *rq)
2425{
2426	if ((gr->lkb_grmode == DLM_LOCK_PR && rq->lkb_rqmode == DLM_LOCK_CW) ||
2427	    (gr->lkb_grmode == DLM_LOCK_CW && rq->lkb_rqmode == DLM_LOCK_PR)) {
2428		if (gr->lkb_highbast < DLM_LOCK_EX)
2429			return 1;
2430		return 0;
2431	}
2432
2433	if (gr->lkb_highbast < rq->lkb_rqmode && !modes_compat(gr, rq))
2434		return 1;
2435	return 0;
2436}
2437
2438static void send_bast_queue(struct dlm_rsb *r, struct list_head *head,
2439			    struct dlm_lkb *lkb)
2440{
2441	struct dlm_lkb *gr;
2442
2443	list_for_each_entry(gr, head, lkb_statequeue) {
2444		/* skip self when sending basts to convertqueue */
2445		if (gr == lkb)
2446			continue;
2447		if (gr->lkb_bastfn && modes_require_bast(gr, lkb)) {
2448			queue_bast(r, gr, lkb->lkb_rqmode);
2449			gr->lkb_highbast = lkb->lkb_rqmode;
2450		}
2451	}
2452}
2453
2454static void send_blocking_asts(struct dlm_rsb *r, struct dlm_lkb *lkb)
2455{
2456	send_bast_queue(r, &r->res_grantqueue, lkb);
2457}
2458
2459static void send_blocking_asts_all(struct dlm_rsb *r, struct dlm_lkb *lkb)
2460{
2461	send_bast_queue(r, &r->res_grantqueue, lkb);
2462	send_bast_queue(r, &r->res_convertqueue, lkb);
2463}
2464
2465/* set_master(r, lkb) -- set the master nodeid of a resource
2466
2467   The purpose of this function is to set the nodeid field in the given
2468   lkb using the nodeid field in the given rsb.  If the rsb's nodeid is
2469   known, it can just be copied to the lkb and the function will return
2470   0.  If the rsb's nodeid is _not_ known, it needs to be looked up
2471   before it can be copied to the lkb.
2472
2473   When the rsb nodeid is being looked up remotely, the initial lkb
2474   causing the lookup is kept on the ls_waiters list waiting for the
2475   lookup reply.  Other lkb's waiting for the same rsb lookup are kept
2476   on the rsb's res_lookup list until the master is verified.
2477
2478   Return values:
2479   0: nodeid is set in rsb/lkb and the caller should go ahead and use it
2480   1: the rsb master is not available and the lkb has been placed on
2481      a wait queue
2482*/
2483
2484static int set_master(struct dlm_rsb *r, struct dlm_lkb *lkb)
2485{
2486	int our_nodeid = dlm_our_nodeid();
2487
2488	if (rsb_flag(r, RSB_MASTER_UNCERTAIN)) {
2489		rsb_clear_flag(r, RSB_MASTER_UNCERTAIN);
2490		r->res_first_lkid = lkb->lkb_id;
2491		lkb->lkb_nodeid = r->res_nodeid;
2492		return 0;
2493	}
2494
2495	if (r->res_first_lkid && r->res_first_lkid != lkb->lkb_id) {
2496		list_add_tail(&lkb->lkb_rsb_lookup, &r->res_lookup);
2497		return 1;
2498	}
2499
2500	if (r->res_master_nodeid == our_nodeid) {
2501		lkb->lkb_nodeid = 0;
2502		return 0;
2503	}
2504
2505	if (r->res_master_nodeid) {
2506		lkb->lkb_nodeid = r->res_master_nodeid;
2507		return 0;
2508	}
2509
2510	if (dlm_dir_nodeid(r) == our_nodeid) {
2511		/* This is a somewhat unusual case; find_rsb will usually
2512		   have set res_master_nodeid when dir nodeid is local, but
2513		   there are cases where we become the dir node after we've
2514		   past find_rsb and go through _request_lock again.
2515		   confirm_master() or process_lookup_list() needs to be
2516		   called after this. */
2517		log_debug(r->res_ls, "set_master %x self master %d dir %d %s",
2518			  lkb->lkb_id, r->res_master_nodeid, r->res_dir_nodeid,
2519			  r->res_name);
2520		r->res_master_nodeid = our_nodeid;
2521		r->res_nodeid = 0;
2522		lkb->lkb_nodeid = 0;
2523		return 0;
2524	}
2525
 
 
2526	r->res_first_lkid = lkb->lkb_id;
2527	send_lookup(r, lkb);
2528	return 1;
2529}
2530
2531static void process_lookup_list(struct dlm_rsb *r)
2532{
2533	struct dlm_lkb *lkb, *safe;
2534
2535	list_for_each_entry_safe(lkb, safe, &r->res_lookup, lkb_rsb_lookup) {
2536		list_del_init(&lkb->lkb_rsb_lookup);
2537		_request_lock(r, lkb);
2538		schedule();
2539	}
2540}
2541
2542/* confirm_master -- confirm (or deny) an rsb's master nodeid */
2543
2544static void confirm_master(struct dlm_rsb *r, int error)
2545{
2546	struct dlm_lkb *lkb;
2547
2548	if (!r->res_first_lkid)
2549		return;
2550
2551	switch (error) {
2552	case 0:
2553	case -EINPROGRESS:
2554		r->res_first_lkid = 0;
2555		process_lookup_list(r);
2556		break;
2557
2558	case -EAGAIN:
2559	case -EBADR:
2560	case -ENOTBLK:
2561		/* the remote request failed and won't be retried (it was
2562		   a NOQUEUE, or has been canceled/unlocked); make a waiting
2563		   lkb the first_lkid */
2564
2565		r->res_first_lkid = 0;
2566
2567		if (!list_empty(&r->res_lookup)) {
2568			lkb = list_entry(r->res_lookup.next, struct dlm_lkb,
2569					 lkb_rsb_lookup);
2570			list_del_init(&lkb->lkb_rsb_lookup);
2571			r->res_first_lkid = lkb->lkb_id;
2572			_request_lock(r, lkb);
2573		}
2574		break;
2575
2576	default:
2577		log_error(r->res_ls, "confirm_master unknown error %d", error);
2578	}
2579}
2580
2581static int set_lock_args(int mode, struct dlm_lksb *lksb, uint32_t flags,
2582			 int namelen, void (*ast)(void *astparam),
 
2583			 void *astparam,
2584			 void (*bast)(void *astparam, int mode),
2585			 struct dlm_args *args)
2586{
2587	int rv = -EINVAL;
2588
2589	/* check for invalid arg usage */
2590
2591	if (mode < 0 || mode > DLM_LOCK_EX)
2592		goto out;
2593
2594	if (!(flags & DLM_LKF_CONVERT) && (namelen > DLM_RESNAME_MAXLEN))
2595		goto out;
2596
2597	if (flags & DLM_LKF_CANCEL)
2598		goto out;
2599
2600	if (flags & DLM_LKF_QUECVT && !(flags & DLM_LKF_CONVERT))
2601		goto out;
2602
2603	if (flags & DLM_LKF_CONVDEADLK && !(flags & DLM_LKF_CONVERT))
2604		goto out;
2605
2606	if (flags & DLM_LKF_CONVDEADLK && flags & DLM_LKF_NOQUEUE)
2607		goto out;
2608
2609	if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_CONVERT)
2610		goto out;
2611
2612	if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_QUECVT)
2613		goto out;
2614
2615	if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_NOQUEUE)
2616		goto out;
2617
2618	if (flags & DLM_LKF_EXPEDITE && mode != DLM_LOCK_NL)
2619		goto out;
2620
2621	if (!ast || !lksb)
2622		goto out;
2623
2624	if (flags & DLM_LKF_VALBLK && !lksb->sb_lvbptr)
2625		goto out;
2626
2627	if (flags & DLM_LKF_CONVERT && !lksb->sb_lkid)
2628		goto out;
2629
2630	/* these args will be copied to the lkb in validate_lock_args,
2631	   it cannot be done now because when converting locks, fields in
2632	   an active lkb cannot be modified before locking the rsb */
2633
2634	args->flags = flags;
2635	args->astfn = ast;
2636	args->astparam = astparam;
2637	args->bastfn = bast;
 
2638	args->mode = mode;
2639	args->lksb = lksb;
2640	rv = 0;
2641 out:
2642	return rv;
2643}
2644
2645static int set_unlock_args(uint32_t flags, void *astarg, struct dlm_args *args)
2646{
2647	if (flags & ~(DLM_LKF_CANCEL | DLM_LKF_VALBLK | DLM_LKF_IVVALBLK |
2648 		      DLM_LKF_FORCEUNLOCK))
2649		return -EINVAL;
2650
2651	if (flags & DLM_LKF_CANCEL && flags & DLM_LKF_FORCEUNLOCK)
2652		return -EINVAL;
2653
2654	args->flags = flags;
2655	args->astparam = astarg;
2656	return 0;
2657}
2658
2659static int validate_lock_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
2660			      struct dlm_args *args)
2661{
2662	int rv = -EBUSY;
2663
2664	if (args->flags & DLM_LKF_CONVERT) {
2665		if (lkb->lkb_status != DLM_LKSTS_GRANTED)
2666			goto out;
2667
2668		/* lock not allowed if there's any op in progress */
2669		if (lkb->lkb_wait_type || atomic_read(&lkb->lkb_wait_count))
2670			goto out;
2671
2672		if (is_overlap(lkb))
 
2673			goto out;
2674
2675		rv = -EINVAL;
2676		if (test_bit(DLM_IFL_MSTCPY_BIT, &lkb->lkb_iflags))
2677			goto out;
2678
2679		if (args->flags & DLM_LKF_QUECVT &&
2680		    !__quecvt_compat_matrix[lkb->lkb_grmode+1][args->mode+1])
2681			goto out;
2682	}
2683
2684	lkb->lkb_exflags = args->flags;
2685	dlm_set_sbflags_val(lkb, 0);
2686	lkb->lkb_astfn = args->astfn;
2687	lkb->lkb_astparam = args->astparam;
2688	lkb->lkb_bastfn = args->bastfn;
2689	lkb->lkb_rqmode = args->mode;
2690	lkb->lkb_lksb = args->lksb;
2691	lkb->lkb_lvbptr = args->lksb->sb_lvbptr;
2692	lkb->lkb_ownpid = (int) current->pid;
 
2693	rv = 0;
2694 out:
2695	switch (rv) {
2696	case 0:
2697		break;
2698	case -EINVAL:
2699		/* annoy the user because dlm usage is wrong */
2700		WARN_ON(1);
2701		log_error(ls, "%s %d %x %x %x %d %d %s", __func__,
2702			  rv, lkb->lkb_id, dlm_iflags_val(lkb), args->flags,
2703			  lkb->lkb_status, lkb->lkb_wait_type,
2704			  lkb->lkb_resource->res_name);
2705		break;
2706	default:
2707		log_debug(ls, "%s %d %x %x %x %d %d %s", __func__,
2708			  rv, lkb->lkb_id, dlm_iflags_val(lkb), args->flags,
2709			  lkb->lkb_status, lkb->lkb_wait_type,
2710			  lkb->lkb_resource->res_name);
2711		break;
2712	}
2713
2714	return rv;
2715}
2716
2717/* when dlm_unlock() sees -EBUSY with CANCEL/FORCEUNLOCK it returns 0
2718   for success */
2719
2720/* note: it's valid for lkb_nodeid/res_nodeid to be -1 when we get here
2721   because there may be a lookup in progress and it's valid to do
2722   cancel/unlockf on it */
2723
2724static int validate_unlock_args(struct dlm_lkb *lkb, struct dlm_args *args)
2725{
2726	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
2727	int rv = -EBUSY;
2728
2729	/* normal unlock not allowed if there's any op in progress */
2730	if (!(args->flags & (DLM_LKF_CANCEL | DLM_LKF_FORCEUNLOCK)) &&
2731	    (lkb->lkb_wait_type || atomic_read(&lkb->lkb_wait_count)))
2732		goto out;
 
 
 
 
 
 
 
 
 
 
 
2733
2734	/* an lkb may be waiting for an rsb lookup to complete where the
2735	   lookup was initiated by another lock */
2736
2737	if (!list_empty(&lkb->lkb_rsb_lookup)) {
2738		if (args->flags & (DLM_LKF_CANCEL | DLM_LKF_FORCEUNLOCK)) {
2739			log_debug(ls, "unlock on rsb_lookup %x", lkb->lkb_id);
2740			list_del_init(&lkb->lkb_rsb_lookup);
2741			queue_cast(lkb->lkb_resource, lkb,
2742				   args->flags & DLM_LKF_CANCEL ?
2743				   -DLM_ECANCEL : -DLM_EUNLOCK);
2744			unhold_lkb(lkb); /* undoes create_lkb() */
2745		}
2746		/* caller changes -EBUSY to 0 for CANCEL and FORCEUNLOCK */
2747		goto out;
2748	}
2749
2750	rv = -EINVAL;
2751	if (test_bit(DLM_IFL_MSTCPY_BIT, &lkb->lkb_iflags)) {
2752		log_error(ls, "unlock on MSTCPY %x", lkb->lkb_id);
2753		dlm_print_lkb(lkb);
2754		goto out;
2755	}
2756
2757	/* an lkb may still exist even though the lock is EOL'ed due to a
2758	 * cancel, unlock or failed noqueue request; an app can't use these
2759	 * locks; return same error as if the lkid had not been found at all
2760	 */
2761
2762	if (test_bit(DLM_IFL_ENDOFLIFE_BIT, &lkb->lkb_iflags)) {
2763		log_debug(ls, "unlock on ENDOFLIFE %x", lkb->lkb_id);
2764		rv = -ENOENT;
2765		goto out;
2766	}
2767
2768	/* cancel not allowed with another cancel/unlock in progress */
2769
2770	if (args->flags & DLM_LKF_CANCEL) {
2771		if (lkb->lkb_exflags & DLM_LKF_CANCEL)
2772			goto out;
2773
2774		if (is_overlap(lkb))
2775			goto out;
2776
2777		if (test_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags)) {
2778			set_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags);
 
 
 
2779			rv = -EBUSY;
2780			goto out;
2781		}
2782
2783		/* there's nothing to cancel */
2784		if (lkb->lkb_status == DLM_LKSTS_GRANTED &&
2785		    !lkb->lkb_wait_type) {
2786			rv = -EBUSY;
2787			goto out;
2788		}
2789
2790		switch (lkb->lkb_wait_type) {
2791		case DLM_MSG_LOOKUP:
2792		case DLM_MSG_REQUEST:
2793			set_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags);
2794			rv = -EBUSY;
2795			goto out;
2796		case DLM_MSG_UNLOCK:
2797		case DLM_MSG_CANCEL:
2798			goto out;
2799		}
2800		/* add_to_waiters() will set OVERLAP_CANCEL */
2801		goto out_ok;
2802	}
2803
2804	/* do we need to allow a force-unlock if there's a normal unlock
2805	   already in progress?  in what conditions could the normal unlock
2806	   fail such that we'd want to send a force-unlock to be sure? */
2807
2808	if (args->flags & DLM_LKF_FORCEUNLOCK) {
2809		if (lkb->lkb_exflags & DLM_LKF_FORCEUNLOCK)
2810			goto out;
2811
2812		if (is_overlap_unlock(lkb))
2813			goto out;
2814
2815		if (test_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags)) {
2816			set_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags);
 
 
 
2817			rv = -EBUSY;
2818			goto out;
2819		}
2820
2821		switch (lkb->lkb_wait_type) {
2822		case DLM_MSG_LOOKUP:
2823		case DLM_MSG_REQUEST:
2824			set_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags);
2825			rv = -EBUSY;
2826			goto out;
2827		case DLM_MSG_UNLOCK:
2828			goto out;
2829		}
2830		/* add_to_waiters() will set OVERLAP_UNLOCK */
 
2831	}
2832
 
 
 
 
 
2833 out_ok:
2834	/* an overlapping op shouldn't blow away exflags from other op */
2835	lkb->lkb_exflags |= args->flags;
2836	dlm_set_sbflags_val(lkb, 0);
2837	lkb->lkb_astparam = args->astparam;
2838	rv = 0;
2839 out:
2840	switch (rv) {
2841	case 0:
2842		break;
2843	case -EINVAL:
2844		/* annoy the user because dlm usage is wrong */
2845		WARN_ON(1);
2846		log_error(ls, "%s %d %x %x %x %x %d %s", __func__, rv,
2847			  lkb->lkb_id, dlm_iflags_val(lkb), lkb->lkb_exflags,
2848			  args->flags, lkb->lkb_wait_type,
2849			  lkb->lkb_resource->res_name);
2850		break;
2851	default:
2852		log_debug(ls, "%s %d %x %x %x %x %d %s", __func__, rv,
2853			  lkb->lkb_id, dlm_iflags_val(lkb), lkb->lkb_exflags,
2854			  args->flags, lkb->lkb_wait_type,
2855			  lkb->lkb_resource->res_name);
2856		break;
2857	}
2858
2859	return rv;
2860}
2861
2862/*
2863 * Four stage 4 varieties:
2864 * do_request(), do_convert(), do_unlock(), do_cancel()
2865 * These are called on the master node for the given lock and
2866 * from the central locking logic.
2867 */
2868
2869static int do_request(struct dlm_rsb *r, struct dlm_lkb *lkb)
2870{
2871	int error = 0;
2872
2873	if (can_be_granted(r, lkb, 1, 0, NULL)) {
2874		grant_lock(r, lkb);
2875		queue_cast(r, lkb, 0);
2876		goto out;
2877	}
2878
2879	if (can_be_queued(lkb)) {
2880		error = -EINPROGRESS;
2881		add_lkb(r, lkb, DLM_LKSTS_WAITING);
 
2882		goto out;
2883	}
2884
2885	error = -EAGAIN;
2886	queue_cast(r, lkb, -EAGAIN);
2887 out:
2888	return error;
2889}
2890
2891static void do_request_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
2892			       int error)
2893{
2894	switch (error) {
2895	case -EAGAIN:
2896		if (force_blocking_asts(lkb))
2897			send_blocking_asts_all(r, lkb);
2898		break;
2899	case -EINPROGRESS:
2900		send_blocking_asts(r, lkb);
2901		break;
2902	}
2903}
2904
2905static int do_convert(struct dlm_rsb *r, struct dlm_lkb *lkb)
2906{
2907	int error = 0;
2908	int deadlk = 0;
2909
2910	/* changing an existing lock may allow others to be granted */
2911
2912	if (can_be_granted(r, lkb, 1, 0, &deadlk)) {
2913		grant_lock(r, lkb);
2914		queue_cast(r, lkb, 0);
2915		goto out;
2916	}
2917
2918	/* can_be_granted() detected that this lock would block in a conversion
2919	   deadlock, so we leave it on the granted queue and return EDEADLK in
2920	   the ast for the convert. */
2921
2922	if (deadlk && !(lkb->lkb_exflags & DLM_LKF_NODLCKWT)) {
2923		/* it's left on the granted queue */
2924		revert_lock(r, lkb);
2925		queue_cast(r, lkb, -EDEADLK);
2926		error = -EDEADLK;
2927		goto out;
2928	}
2929
2930	/* is_demoted() means the can_be_granted() above set the grmode
2931	   to NL, and left us on the granted queue.  This auto-demotion
2932	   (due to CONVDEADLK) might mean other locks, and/or this lock, are
2933	   now grantable.  We have to try to grant other converting locks
2934	   before we try again to grant this one. */
2935
2936	if (is_demoted(lkb)) {
2937		grant_pending_convert(r, DLM_LOCK_IV, NULL, NULL);
2938		if (_can_be_granted(r, lkb, 1, 0)) {
2939			grant_lock(r, lkb);
2940			queue_cast(r, lkb, 0);
2941			goto out;
2942		}
2943		/* else fall through and move to convert queue */
2944	}
2945
2946	if (can_be_queued(lkb)) {
2947		error = -EINPROGRESS;
2948		del_lkb(r, lkb);
2949		add_lkb(r, lkb, DLM_LKSTS_CONVERT);
 
2950		goto out;
2951	}
2952
2953	error = -EAGAIN;
2954	queue_cast(r, lkb, -EAGAIN);
2955 out:
2956	return error;
2957}
2958
2959static void do_convert_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
2960			       int error)
2961{
2962	switch (error) {
2963	case 0:
2964		grant_pending_locks(r, NULL);
2965		/* grant_pending_locks also sends basts */
2966		break;
2967	case -EAGAIN:
2968		if (force_blocking_asts(lkb))
2969			send_blocking_asts_all(r, lkb);
2970		break;
2971	case -EINPROGRESS:
2972		send_blocking_asts(r, lkb);
2973		break;
2974	}
2975}
2976
2977static int do_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb)
2978{
2979	remove_lock(r, lkb);
2980	queue_cast(r, lkb, -DLM_EUNLOCK);
2981	return -DLM_EUNLOCK;
2982}
2983
2984static void do_unlock_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
2985			      int error)
2986{
2987	grant_pending_locks(r, NULL);
2988}
2989
2990/* returns: 0 did nothing, -DLM_ECANCEL canceled lock */
2991
2992static int do_cancel(struct dlm_rsb *r, struct dlm_lkb *lkb)
2993{
2994	int error;
2995
2996	error = revert_lock(r, lkb);
2997	if (error) {
2998		queue_cast(r, lkb, -DLM_ECANCEL);
2999		return -DLM_ECANCEL;
3000	}
3001	return 0;
3002}
3003
3004static void do_cancel_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
3005			      int error)
3006{
3007	if (error)
3008		grant_pending_locks(r, NULL);
3009}
3010
3011/*
3012 * Four stage 3 varieties:
3013 * _request_lock(), _convert_lock(), _unlock_lock(), _cancel_lock()
3014 */
3015
3016/* add a new lkb to a possibly new rsb, called by requesting process */
3017
3018static int _request_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3019{
3020	int error;
3021
3022	/* set_master: sets lkb nodeid from r */
3023
3024	error = set_master(r, lkb);
3025	if (error < 0)
3026		goto out;
3027	if (error) {
3028		error = 0;
3029		goto out;
3030	}
3031
3032	if (is_remote(r)) {
3033		/* receive_request() calls do_request() on remote node */
3034		error = send_request(r, lkb);
3035	} else {
3036		error = do_request(r, lkb);
3037		/* for remote locks the request_reply is sent
3038		   between do_request and do_request_effects */
3039		do_request_effects(r, lkb, error);
3040	}
3041 out:
3042	return error;
3043}
3044
3045/* change some property of an existing lkb, e.g. mode */
3046
3047static int _convert_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3048{
3049	int error;
3050
3051	if (is_remote(r)) {
3052		/* receive_convert() calls do_convert() on remote node */
3053		error = send_convert(r, lkb);
3054	} else {
3055		error = do_convert(r, lkb);
3056		/* for remote locks the convert_reply is sent
3057		   between do_convert and do_convert_effects */
3058		do_convert_effects(r, lkb, error);
3059	}
3060
3061	return error;
3062}
3063
3064/* remove an existing lkb from the granted queue */
3065
3066static int _unlock_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3067{
3068	int error;
3069
3070	if (is_remote(r)) {
3071		/* receive_unlock() calls do_unlock() on remote node */
3072		error = send_unlock(r, lkb);
3073	} else {
3074		error = do_unlock(r, lkb);
3075		/* for remote locks the unlock_reply is sent
3076		   between do_unlock and do_unlock_effects */
3077		do_unlock_effects(r, lkb, error);
3078	}
3079
3080	return error;
3081}
3082
3083/* remove an existing lkb from the convert or wait queue */
3084
3085static int _cancel_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3086{
3087	int error;
3088
3089	if (is_remote(r)) {
3090		/* receive_cancel() calls do_cancel() on remote node */
3091		error = send_cancel(r, lkb);
3092	} else {
3093		error = do_cancel(r, lkb);
3094		/* for remote locks the cancel_reply is sent
3095		   between do_cancel and do_cancel_effects */
3096		do_cancel_effects(r, lkb, error);
3097	}
3098
3099	return error;
3100}
3101
3102/*
3103 * Four stage 2 varieties:
3104 * request_lock(), convert_lock(), unlock_lock(), cancel_lock()
3105 */
3106
3107static int request_lock(struct dlm_ls *ls, struct dlm_lkb *lkb,
3108			const void *name, int len,
3109			struct dlm_args *args)
3110{
3111	struct dlm_rsb *r;
3112	int error;
3113
3114	error = validate_lock_args(ls, lkb, args);
3115	if (error)
3116		return error;
3117
3118	error = find_rsb(ls, name, len, 0, R_REQUEST, &r);
3119	if (error)
3120		return error;
3121
3122	lock_rsb(r);
3123
3124	attach_lkb(r, lkb);
3125	lkb->lkb_lksb->sb_lkid = lkb->lkb_id;
3126
3127	error = _request_lock(r, lkb);
3128
3129	unlock_rsb(r);
3130	put_rsb(r);
3131	return error;
3132}
3133
3134static int convert_lock(struct dlm_ls *ls, struct dlm_lkb *lkb,
3135			struct dlm_args *args)
3136{
3137	struct dlm_rsb *r;
3138	int error;
3139
3140	r = lkb->lkb_resource;
3141
3142	hold_rsb(r);
3143	lock_rsb(r);
3144
3145	error = validate_lock_args(ls, lkb, args);
3146	if (error)
3147		goto out;
3148
3149	error = _convert_lock(r, lkb);
3150 out:
3151	unlock_rsb(r);
3152	put_rsb(r);
3153	return error;
3154}
3155
3156static int unlock_lock(struct dlm_ls *ls, struct dlm_lkb *lkb,
3157		       struct dlm_args *args)
3158{
3159	struct dlm_rsb *r;
3160	int error;
3161
3162	r = lkb->lkb_resource;
3163
3164	hold_rsb(r);
3165	lock_rsb(r);
3166
3167	error = validate_unlock_args(lkb, args);
3168	if (error)
3169		goto out;
3170
3171	error = _unlock_lock(r, lkb);
3172 out:
3173	unlock_rsb(r);
3174	put_rsb(r);
3175	return error;
3176}
3177
3178static int cancel_lock(struct dlm_ls *ls, struct dlm_lkb *lkb,
3179		       struct dlm_args *args)
3180{
3181	struct dlm_rsb *r;
3182	int error;
3183
3184	r = lkb->lkb_resource;
3185
3186	hold_rsb(r);
3187	lock_rsb(r);
3188
3189	error = validate_unlock_args(lkb, args);
3190	if (error)
3191		goto out;
3192
3193	error = _cancel_lock(r, lkb);
3194 out:
3195	unlock_rsb(r);
3196	put_rsb(r);
3197	return error;
3198}
3199
3200/*
3201 * Two stage 1 varieties:  dlm_lock() and dlm_unlock()
3202 */
3203
3204int dlm_lock(dlm_lockspace_t *lockspace,
3205	     int mode,
3206	     struct dlm_lksb *lksb,
3207	     uint32_t flags,
3208	     const void *name,
3209	     unsigned int namelen,
3210	     uint32_t parent_lkid,
3211	     void (*ast) (void *astarg),
3212	     void *astarg,
3213	     void (*bast) (void *astarg, int mode))
3214{
3215	struct dlm_ls *ls;
3216	struct dlm_lkb *lkb;
3217	struct dlm_args args;
3218	int error, convert = flags & DLM_LKF_CONVERT;
3219
3220	ls = dlm_find_lockspace_local(lockspace);
3221	if (!ls)
3222		return -EINVAL;
3223
3224	dlm_lock_recovery(ls);
3225
3226	if (convert)
3227		error = find_lkb(ls, lksb->sb_lkid, &lkb);
3228	else
3229		error = create_lkb(ls, &lkb);
3230
3231	if (error)
3232		goto out;
3233
3234	trace_dlm_lock_start(ls, lkb, name, namelen, mode, flags);
3235
3236	error = set_lock_args(mode, lksb, flags, namelen, ast, astarg, bast,
3237			      &args);
3238	if (error)
3239		goto out_put;
3240
3241	if (convert)
3242		error = convert_lock(ls, lkb, &args);
3243	else
3244		error = request_lock(ls, lkb, name, namelen, &args);
3245
3246	if (error == -EINPROGRESS)
3247		error = 0;
3248 out_put:
3249	trace_dlm_lock_end(ls, lkb, name, namelen, mode, flags, error, true);
3250
3251	if (convert || error)
3252		__put_lkb(ls, lkb);
3253	if (error == -EAGAIN || error == -EDEADLK)
3254		error = 0;
3255 out:
3256	dlm_unlock_recovery(ls);
3257	dlm_put_lockspace(ls);
3258	return error;
3259}
3260
3261int dlm_unlock(dlm_lockspace_t *lockspace,
3262	       uint32_t lkid,
3263	       uint32_t flags,
3264	       struct dlm_lksb *lksb,
3265	       void *astarg)
3266{
3267	struct dlm_ls *ls;
3268	struct dlm_lkb *lkb;
3269	struct dlm_args args;
3270	int error;
3271
3272	ls = dlm_find_lockspace_local(lockspace);
3273	if (!ls)
3274		return -EINVAL;
3275
3276	dlm_lock_recovery(ls);
3277
3278	error = find_lkb(ls, lkid, &lkb);
3279	if (error)
3280		goto out;
3281
3282	trace_dlm_unlock_start(ls, lkb, flags);
3283
3284	error = set_unlock_args(flags, astarg, &args);
3285	if (error)
3286		goto out_put;
3287
3288	if (flags & DLM_LKF_CANCEL)
3289		error = cancel_lock(ls, lkb, &args);
3290	else
3291		error = unlock_lock(ls, lkb, &args);
3292
3293	if (error == -DLM_EUNLOCK || error == -DLM_ECANCEL)
3294		error = 0;
3295	if (error == -EBUSY && (flags & (DLM_LKF_CANCEL | DLM_LKF_FORCEUNLOCK)))
3296		error = 0;
3297 out_put:
3298	trace_dlm_unlock_end(ls, lkb, flags, error);
3299
3300	dlm_put_lkb(lkb);
3301 out:
3302	dlm_unlock_recovery(ls);
3303	dlm_put_lockspace(ls);
3304	return error;
3305}
3306
3307/*
3308 * send/receive routines for remote operations and replies
3309 *
3310 * send_args
3311 * send_common
3312 * send_request			receive_request
3313 * send_convert			receive_convert
3314 * send_unlock			receive_unlock
3315 * send_cancel			receive_cancel
3316 * send_grant			receive_grant
3317 * send_bast			receive_bast
3318 * send_lookup			receive_lookup
3319 * send_remove			receive_remove
3320 *
3321 * 				send_common_reply
3322 * receive_request_reply	send_request_reply
3323 * receive_convert_reply	send_convert_reply
3324 * receive_unlock_reply		send_unlock_reply
3325 * receive_cancel_reply		send_cancel_reply
3326 * receive_lookup_reply		send_lookup_reply
3327 */
3328
3329static int _create_message(struct dlm_ls *ls, int mb_len,
3330			   int to_nodeid, int mstype,
3331			   struct dlm_message **ms_ret,
3332			   struct dlm_mhandle **mh_ret,
3333			   gfp_t allocation)
3334{
3335	struct dlm_message *ms;
3336	struct dlm_mhandle *mh;
3337	char *mb;
3338
3339	/* get_buffer gives us a message handle (mh) that we need to
3340	   pass into midcomms_commit and a message buffer (mb) that we
3341	   write our data into */
3342
3343	mh = dlm_midcomms_get_mhandle(to_nodeid, mb_len, allocation, &mb);
3344	if (!mh)
3345		return -ENOBUFS;
3346
3347	ms = (struct dlm_message *) mb;
3348
3349	ms->m_header.h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
3350	ms->m_header.u.h_lockspace = cpu_to_le32(ls->ls_global_id);
3351	ms->m_header.h_nodeid = cpu_to_le32(dlm_our_nodeid());
3352	ms->m_header.h_length = cpu_to_le16(mb_len);
3353	ms->m_header.h_cmd = DLM_MSG;
3354
3355	ms->m_type = cpu_to_le32(mstype);
3356
3357	*mh_ret = mh;
3358	*ms_ret = ms;
3359	return 0;
3360}
3361
3362static int create_message(struct dlm_rsb *r, struct dlm_lkb *lkb,
3363			  int to_nodeid, int mstype,
3364			  struct dlm_message **ms_ret,
3365			  struct dlm_mhandle **mh_ret,
3366			  gfp_t allocation)
3367{
3368	int mb_len = sizeof(struct dlm_message);
3369
3370	switch (mstype) {
3371	case DLM_MSG_REQUEST:
3372	case DLM_MSG_LOOKUP:
3373	case DLM_MSG_REMOVE:
3374		mb_len += r->res_length;
3375		break;
3376	case DLM_MSG_CONVERT:
3377	case DLM_MSG_UNLOCK:
3378	case DLM_MSG_REQUEST_REPLY:
3379	case DLM_MSG_CONVERT_REPLY:
3380	case DLM_MSG_GRANT:
3381		if (lkb && lkb->lkb_lvbptr && (lkb->lkb_exflags & DLM_LKF_VALBLK))
3382			mb_len += r->res_ls->ls_lvblen;
3383		break;
3384	}
3385
3386	return _create_message(r->res_ls, mb_len, to_nodeid, mstype,
3387			       ms_ret, mh_ret, allocation);
3388}
3389
3390/* further lowcomms enhancements or alternate implementations may make
3391   the return value from this function useful at some point */
3392
3393static int send_message(struct dlm_mhandle *mh, struct dlm_message *ms,
3394			const void *name, int namelen)
3395{
3396	dlm_midcomms_commit_mhandle(mh, name, namelen);
 
3397	return 0;
3398}
3399
3400static void send_args(struct dlm_rsb *r, struct dlm_lkb *lkb,
3401		      struct dlm_message *ms)
3402{
3403	ms->m_nodeid   = cpu_to_le32(lkb->lkb_nodeid);
3404	ms->m_pid      = cpu_to_le32(lkb->lkb_ownpid);
3405	ms->m_lkid     = cpu_to_le32(lkb->lkb_id);
3406	ms->m_remid    = cpu_to_le32(lkb->lkb_remid);
3407	ms->m_exflags  = cpu_to_le32(lkb->lkb_exflags);
3408	ms->m_sbflags  = cpu_to_le32(dlm_sbflags_val(lkb));
3409	ms->m_flags    = cpu_to_le32(dlm_dflags_val(lkb));
3410	ms->m_lvbseq   = cpu_to_le32(lkb->lkb_lvbseq);
3411	ms->m_status   = cpu_to_le32(lkb->lkb_status);
3412	ms->m_grmode   = cpu_to_le32(lkb->lkb_grmode);
3413	ms->m_rqmode   = cpu_to_le32(lkb->lkb_rqmode);
3414	ms->m_hash     = cpu_to_le32(r->res_hash);
3415
3416	/* m_result and m_bastmode are set from function args,
3417	   not from lkb fields */
3418
3419	if (lkb->lkb_bastfn)
3420		ms->m_asts |= cpu_to_le32(DLM_CB_BAST);
3421	if (lkb->lkb_astfn)
3422		ms->m_asts |= cpu_to_le32(DLM_CB_CAST);
3423
3424	/* compare with switch in create_message; send_remove() doesn't
3425	   use send_args() */
3426
3427	switch (ms->m_type) {
3428	case cpu_to_le32(DLM_MSG_REQUEST):
3429	case cpu_to_le32(DLM_MSG_LOOKUP):
3430		memcpy(ms->m_extra, r->res_name, r->res_length);
3431		break;
3432	case cpu_to_le32(DLM_MSG_CONVERT):
3433	case cpu_to_le32(DLM_MSG_UNLOCK):
3434	case cpu_to_le32(DLM_MSG_REQUEST_REPLY):
3435	case cpu_to_le32(DLM_MSG_CONVERT_REPLY):
3436	case cpu_to_le32(DLM_MSG_GRANT):
3437		if (!lkb->lkb_lvbptr || !(lkb->lkb_exflags & DLM_LKF_VALBLK))
3438			break;
3439		memcpy(ms->m_extra, lkb->lkb_lvbptr, r->res_ls->ls_lvblen);
3440		break;
3441	}
3442}
3443
3444static int send_common(struct dlm_rsb *r, struct dlm_lkb *lkb, int mstype)
3445{
3446	struct dlm_message *ms;
3447	struct dlm_mhandle *mh;
3448	int to_nodeid, error;
3449
3450	to_nodeid = r->res_nodeid;
3451
3452	error = add_to_waiters(lkb, mstype, to_nodeid);
3453	if (error)
3454		return error;
3455
3456	error = create_message(r, lkb, to_nodeid, mstype, &ms, &mh, GFP_NOFS);
3457	if (error)
3458		goto fail;
3459
3460	send_args(r, lkb, ms);
3461
3462	error = send_message(mh, ms, r->res_name, r->res_length);
3463	if (error)
3464		goto fail;
3465	return 0;
3466
3467 fail:
3468	remove_from_waiters(lkb, msg_reply_type(mstype));
3469	return error;
3470}
3471
3472static int send_request(struct dlm_rsb *r, struct dlm_lkb *lkb)
3473{
3474	return send_common(r, lkb, DLM_MSG_REQUEST);
3475}
3476
3477static int send_convert(struct dlm_rsb *r, struct dlm_lkb *lkb)
3478{
3479	int error;
3480
3481	error = send_common(r, lkb, DLM_MSG_CONVERT);
3482
3483	/* down conversions go without a reply from the master */
3484	if (!error && down_conversion(lkb)) {
3485		remove_from_waiters(lkb, DLM_MSG_CONVERT_REPLY);
3486		r->res_ls->ls_local_ms.m_type = cpu_to_le32(DLM_MSG_CONVERT_REPLY);
3487		r->res_ls->ls_local_ms.m_result = 0;
3488		__receive_convert_reply(r, lkb, &r->res_ls->ls_local_ms, true);
 
3489	}
3490
3491	return error;
3492}
3493
3494/* FIXME: if this lkb is the only lock we hold on the rsb, then set
3495   MASTER_UNCERTAIN to force the next request on the rsb to confirm
3496   that the master is still correct. */
3497
3498static int send_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb)
3499{
3500	return send_common(r, lkb, DLM_MSG_UNLOCK);
3501}
3502
3503static int send_cancel(struct dlm_rsb *r, struct dlm_lkb *lkb)
3504{
3505	return send_common(r, lkb, DLM_MSG_CANCEL);
3506}
3507
3508static int send_grant(struct dlm_rsb *r, struct dlm_lkb *lkb)
3509{
3510	struct dlm_message *ms;
3511	struct dlm_mhandle *mh;
3512	int to_nodeid, error;
3513
3514	to_nodeid = lkb->lkb_nodeid;
3515
3516	error = create_message(r, lkb, to_nodeid, DLM_MSG_GRANT, &ms, &mh,
3517			       GFP_NOFS);
3518	if (error)
3519		goto out;
3520
3521	send_args(r, lkb, ms);
3522
3523	ms->m_result = 0;
3524
3525	error = send_message(mh, ms, r->res_name, r->res_length);
3526 out:
3527	return error;
3528}
3529
3530static int send_bast(struct dlm_rsb *r, struct dlm_lkb *lkb, int mode)
3531{
3532	struct dlm_message *ms;
3533	struct dlm_mhandle *mh;
3534	int to_nodeid, error;
3535
3536	to_nodeid = lkb->lkb_nodeid;
3537
3538	error = create_message(r, NULL, to_nodeid, DLM_MSG_BAST, &ms, &mh,
3539			       GFP_NOFS);
3540	if (error)
3541		goto out;
3542
3543	send_args(r, lkb, ms);
3544
3545	ms->m_bastmode = cpu_to_le32(mode);
3546
3547	error = send_message(mh, ms, r->res_name, r->res_length);
3548 out:
3549	return error;
3550}
3551
3552static int send_lookup(struct dlm_rsb *r, struct dlm_lkb *lkb)
3553{
3554	struct dlm_message *ms;
3555	struct dlm_mhandle *mh;
3556	int to_nodeid, error;
3557
3558	to_nodeid = dlm_dir_nodeid(r);
3559
3560	error = add_to_waiters(lkb, DLM_MSG_LOOKUP, to_nodeid);
3561	if (error)
3562		return error;
3563
3564	error = create_message(r, NULL, to_nodeid, DLM_MSG_LOOKUP, &ms, &mh,
3565			       GFP_NOFS);
3566	if (error)
3567		goto fail;
3568
3569	send_args(r, lkb, ms);
3570
3571	error = send_message(mh, ms, r->res_name, r->res_length);
3572	if (error)
3573		goto fail;
3574	return 0;
3575
3576 fail:
3577	remove_from_waiters(lkb, DLM_MSG_LOOKUP_REPLY);
3578	return error;
3579}
3580
3581static int send_remove(struct dlm_rsb *r)
3582{
3583	struct dlm_message *ms;
3584	struct dlm_mhandle *mh;
3585	int to_nodeid, error;
3586
3587	to_nodeid = dlm_dir_nodeid(r);
3588
3589	error = create_message(r, NULL, to_nodeid, DLM_MSG_REMOVE, &ms, &mh,
3590			       GFP_ATOMIC);
3591	if (error)
3592		goto out;
3593
3594	memcpy(ms->m_extra, r->res_name, r->res_length);
3595	ms->m_hash = cpu_to_le32(r->res_hash);
3596
3597	error = send_message(mh, ms, r->res_name, r->res_length);
3598 out:
3599	return error;
3600}
3601
3602static int send_common_reply(struct dlm_rsb *r, struct dlm_lkb *lkb,
3603			     int mstype, int rv)
3604{
3605	struct dlm_message *ms;
3606	struct dlm_mhandle *mh;
3607	int to_nodeid, error;
3608
3609	to_nodeid = lkb->lkb_nodeid;
3610
3611	error = create_message(r, lkb, to_nodeid, mstype, &ms, &mh, GFP_NOFS);
3612	if (error)
3613		goto out;
3614
3615	send_args(r, lkb, ms);
3616
3617	ms->m_result = cpu_to_le32(to_dlm_errno(rv));
3618
3619	error = send_message(mh, ms, r->res_name, r->res_length);
3620 out:
3621	return error;
3622}
3623
3624static int send_request_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3625{
3626	return send_common_reply(r, lkb, DLM_MSG_REQUEST_REPLY, rv);
3627}
3628
3629static int send_convert_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3630{
3631	return send_common_reply(r, lkb, DLM_MSG_CONVERT_REPLY, rv);
3632}
3633
3634static int send_unlock_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3635{
3636	return send_common_reply(r, lkb, DLM_MSG_UNLOCK_REPLY, rv);
3637}
3638
3639static int send_cancel_reply(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
3640{
3641	return send_common_reply(r, lkb, DLM_MSG_CANCEL_REPLY, rv);
3642}
3643
3644static int send_lookup_reply(struct dlm_ls *ls,
3645			     const struct dlm_message *ms_in, int ret_nodeid,
3646			     int rv)
3647{
3648	struct dlm_rsb *r = &ls->ls_local_rsb;
3649	struct dlm_message *ms;
3650	struct dlm_mhandle *mh;
3651	int error, nodeid = le32_to_cpu(ms_in->m_header.h_nodeid);
3652
3653	error = create_message(r, NULL, nodeid, DLM_MSG_LOOKUP_REPLY, &ms, &mh,
3654			       GFP_NOFS);
3655	if (error)
3656		goto out;
3657
3658	ms->m_lkid = ms_in->m_lkid;
3659	ms->m_result = cpu_to_le32(to_dlm_errno(rv));
3660	ms->m_nodeid = cpu_to_le32(ret_nodeid);
3661
3662	error = send_message(mh, ms, ms_in->m_extra, receive_extralen(ms_in));
3663 out:
3664	return error;
3665}
3666
3667/* which args we save from a received message depends heavily on the type
3668   of message, unlike the send side where we can safely send everything about
3669   the lkb for any type of message */
3670
3671static void receive_flags(struct dlm_lkb *lkb, const struct dlm_message *ms)
3672{
3673	lkb->lkb_exflags = le32_to_cpu(ms->m_exflags);
3674	dlm_set_sbflags_val(lkb, le32_to_cpu(ms->m_sbflags));
3675	dlm_set_dflags_val(lkb, le32_to_cpu(ms->m_flags));
 
3676}
3677
3678static void receive_flags_reply(struct dlm_lkb *lkb,
3679				const struct dlm_message *ms,
3680				bool local)
3681{
3682	if (local)
3683		return;
3684
3685	dlm_set_sbflags_val(lkb, le32_to_cpu(ms->m_sbflags));
3686	dlm_set_dflags_val(lkb, le32_to_cpu(ms->m_flags));
 
3687}
3688
3689static int receive_extralen(const struct dlm_message *ms)
3690{
3691	return (le16_to_cpu(ms->m_header.h_length) -
3692		sizeof(struct dlm_message));
3693}
3694
3695static int receive_lvb(struct dlm_ls *ls, struct dlm_lkb *lkb,
3696		       const struct dlm_message *ms)
3697{
3698	int len;
3699
3700	if (lkb->lkb_exflags & DLM_LKF_VALBLK) {
3701		if (!lkb->lkb_lvbptr)
3702			lkb->lkb_lvbptr = dlm_allocate_lvb(ls);
3703		if (!lkb->lkb_lvbptr)
3704			return -ENOMEM;
3705		len = receive_extralen(ms);
3706		if (len > ls->ls_lvblen)
3707			len = ls->ls_lvblen;
3708		memcpy(lkb->lkb_lvbptr, ms->m_extra, len);
3709	}
3710	return 0;
3711}
3712
3713static void fake_bastfn(void *astparam, int mode)
3714{
3715	log_print("fake_bastfn should not be called");
3716}
3717
3718static void fake_astfn(void *astparam)
3719{
3720	log_print("fake_astfn should not be called");
3721}
3722
3723static int receive_request_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
3724				const struct dlm_message *ms)
3725{
3726	lkb->lkb_nodeid = le32_to_cpu(ms->m_header.h_nodeid);
3727	lkb->lkb_ownpid = le32_to_cpu(ms->m_pid);
3728	lkb->lkb_remid = le32_to_cpu(ms->m_lkid);
3729	lkb->lkb_grmode = DLM_LOCK_IV;
3730	lkb->lkb_rqmode = le32_to_cpu(ms->m_rqmode);
3731
3732	lkb->lkb_bastfn = (ms->m_asts & cpu_to_le32(DLM_CB_BAST)) ? &fake_bastfn : NULL;
3733	lkb->lkb_astfn = (ms->m_asts & cpu_to_le32(DLM_CB_CAST)) ? &fake_astfn : NULL;
3734
3735	if (lkb->lkb_exflags & DLM_LKF_VALBLK) {
3736		/* lkb was just created so there won't be an lvb yet */
3737		lkb->lkb_lvbptr = dlm_allocate_lvb(ls);
3738		if (!lkb->lkb_lvbptr)
3739			return -ENOMEM;
3740	}
3741
3742	return 0;
3743}
3744
3745static int receive_convert_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
3746				const struct dlm_message *ms)
3747{
3748	if (lkb->lkb_status != DLM_LKSTS_GRANTED)
3749		return -EBUSY;
3750
3751	if (receive_lvb(ls, lkb, ms))
3752		return -ENOMEM;
3753
3754	lkb->lkb_rqmode = le32_to_cpu(ms->m_rqmode);
3755	lkb->lkb_lvbseq = le32_to_cpu(ms->m_lvbseq);
3756
3757	return 0;
3758}
3759
3760static int receive_unlock_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
3761			       const struct dlm_message *ms)
3762{
3763	if (receive_lvb(ls, lkb, ms))
3764		return -ENOMEM;
3765	return 0;
3766}
3767
3768/* We fill in the local-lkb fields with the info that send_xxxx_reply()
3769   uses to send a reply and that the remote end uses to process the reply. */
3770
3771static void setup_local_lkb(struct dlm_ls *ls, const struct dlm_message *ms)
3772{
3773	struct dlm_lkb *lkb = &ls->ls_local_lkb;
3774	lkb->lkb_nodeid = le32_to_cpu(ms->m_header.h_nodeid);
3775	lkb->lkb_remid = le32_to_cpu(ms->m_lkid);
3776}
3777
3778/* This is called after the rsb is locked so that we can safely inspect
3779   fields in the lkb. */
3780
3781static int validate_message(struct dlm_lkb *lkb, const struct dlm_message *ms)
3782{
3783	int from = le32_to_cpu(ms->m_header.h_nodeid);
3784	int error = 0;
3785
3786	/* currently mixing of user/kernel locks are not supported */
3787	if (ms->m_flags & cpu_to_le32(BIT(DLM_DFL_USER_BIT)) &&
3788	    !test_bit(DLM_DFL_USER_BIT, &lkb->lkb_dflags)) {
3789		log_error(lkb->lkb_resource->res_ls,
3790			  "got user dlm message for a kernel lock");
3791		error = -EINVAL;
3792		goto out;
3793	}
3794
3795	switch (ms->m_type) {
3796	case cpu_to_le32(DLM_MSG_CONVERT):
3797	case cpu_to_le32(DLM_MSG_UNLOCK):
3798	case cpu_to_le32(DLM_MSG_CANCEL):
3799		if (!is_master_copy(lkb) || lkb->lkb_nodeid != from)
3800			error = -EINVAL;
3801		break;
3802
3803	case cpu_to_le32(DLM_MSG_CONVERT_REPLY):
3804	case cpu_to_le32(DLM_MSG_UNLOCK_REPLY):
3805	case cpu_to_le32(DLM_MSG_CANCEL_REPLY):
3806	case cpu_to_le32(DLM_MSG_GRANT):
3807	case cpu_to_le32(DLM_MSG_BAST):
3808		if (!is_process_copy(lkb) || lkb->lkb_nodeid != from)
3809			error = -EINVAL;
3810		break;
3811
3812	case cpu_to_le32(DLM_MSG_REQUEST_REPLY):
3813		if (!is_process_copy(lkb))
3814			error = -EINVAL;
3815		else if (lkb->lkb_nodeid != -1 && lkb->lkb_nodeid != from)
3816			error = -EINVAL;
3817		break;
3818
3819	default:
3820		error = -EINVAL;
3821	}
3822
3823out:
3824	if (error)
3825		log_error(lkb->lkb_resource->res_ls,
3826			  "ignore invalid message %d from %d %x %x %x %d",
3827			  le32_to_cpu(ms->m_type), from, lkb->lkb_id,
3828			  lkb->lkb_remid, dlm_iflags_val(lkb),
3829			  lkb->lkb_nodeid);
3830	return error;
3831}
3832
3833static int receive_request(struct dlm_ls *ls, const struct dlm_message *ms)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3834{
3835	struct dlm_lkb *lkb;
3836	struct dlm_rsb *r;
3837	int from_nodeid;
3838	int error, namelen = 0;
3839
3840	from_nodeid = le32_to_cpu(ms->m_header.h_nodeid);
3841
3842	error = create_lkb(ls, &lkb);
3843	if (error)
3844		goto fail;
3845
3846	receive_flags(lkb, ms);
3847	set_bit(DLM_IFL_MSTCPY_BIT, &lkb->lkb_iflags);
3848	error = receive_request_args(ls, lkb, ms);
3849	if (error) {
3850		__put_lkb(ls, lkb);
3851		goto fail;
3852	}
3853
3854	/* The dir node is the authority on whether we are the master
3855	   for this rsb or not, so if the master sends us a request, we should
3856	   recreate the rsb if we've destroyed it.   This race happens when we
3857	   send a remove message to the dir node at the same time that the dir
3858	   node sends us a request for the rsb. */
3859
3860	namelen = receive_extralen(ms);
3861
3862	error = find_rsb(ls, ms->m_extra, namelen, from_nodeid,
3863			 R_RECEIVE_REQUEST, &r);
3864	if (error) {
3865		__put_lkb(ls, lkb);
3866		goto fail;
3867	}
3868
3869	lock_rsb(r);
3870
3871	if (r->res_master_nodeid != dlm_our_nodeid()) {
3872		error = validate_master_nodeid(ls, r, from_nodeid);
3873		if (error) {
3874			unlock_rsb(r);
3875			put_rsb(r);
3876			__put_lkb(ls, lkb);
3877			goto fail;
3878		}
3879	}
3880
3881	attach_lkb(r, lkb);
3882	error = do_request(r, lkb);
3883	send_request_reply(r, lkb, error);
3884	do_request_effects(r, lkb, error);
3885
3886	unlock_rsb(r);
3887	put_rsb(r);
3888
3889	if (error == -EINPROGRESS)
3890		error = 0;
3891	if (error)
3892		dlm_put_lkb(lkb);
3893	return 0;
3894
3895 fail:
3896	/* TODO: instead of returning ENOTBLK, add the lkb to res_lookup
3897	   and do this receive_request again from process_lookup_list once
3898	   we get the lookup reply.  This would avoid a many repeated
3899	   ENOTBLK request failures when the lookup reply designating us
3900	   as master is delayed. */
3901
 
 
 
 
 
 
 
 
 
3902	if (error != -ENOTBLK) {
3903		log_limit(ls, "receive_request %x from %d %d",
3904			  le32_to_cpu(ms->m_lkid), from_nodeid, error);
3905	}
3906
3907	setup_local_lkb(ls, ms);
3908	send_request_reply(&ls->ls_local_rsb, &ls->ls_local_lkb, error);
 
 
 
 
 
3909	return error;
3910}
3911
3912static int receive_convert(struct dlm_ls *ls, const struct dlm_message *ms)
3913{
3914	struct dlm_lkb *lkb;
3915	struct dlm_rsb *r;
3916	int error, reply = 1;
3917
3918	error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
3919	if (error)
3920		goto fail;
3921
3922	if (lkb->lkb_remid != le32_to_cpu(ms->m_lkid)) {
3923		log_error(ls, "receive_convert %x remid %x recover_seq %llu "
3924			  "remote %d %x", lkb->lkb_id, lkb->lkb_remid,
3925			  (unsigned long long)lkb->lkb_recover_seq,
3926			  le32_to_cpu(ms->m_header.h_nodeid),
3927			  le32_to_cpu(ms->m_lkid));
3928		error = -ENOENT;
3929		dlm_put_lkb(lkb);
3930		goto fail;
3931	}
3932
3933	r = lkb->lkb_resource;
3934
3935	hold_rsb(r);
3936	lock_rsb(r);
3937
3938	error = validate_message(lkb, ms);
3939	if (error)
3940		goto out;
3941
3942	receive_flags(lkb, ms);
3943
3944	error = receive_convert_args(ls, lkb, ms);
3945	if (error) {
3946		send_convert_reply(r, lkb, error);
3947		goto out;
3948	}
3949
3950	reply = !down_conversion(lkb);
3951
3952	error = do_convert(r, lkb);
3953	if (reply)
3954		send_convert_reply(r, lkb, error);
3955	do_convert_effects(r, lkb, error);
3956 out:
3957	unlock_rsb(r);
3958	put_rsb(r);
3959	dlm_put_lkb(lkb);
3960	return 0;
3961
3962 fail:
3963	setup_local_lkb(ls, ms);
3964	send_convert_reply(&ls->ls_local_rsb, &ls->ls_local_lkb, error);
3965	return error;
3966}
3967
3968static int receive_unlock(struct dlm_ls *ls, const struct dlm_message *ms)
3969{
3970	struct dlm_lkb *lkb;
3971	struct dlm_rsb *r;
3972	int error;
3973
3974	error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
3975	if (error)
3976		goto fail;
3977
3978	if (lkb->lkb_remid != le32_to_cpu(ms->m_lkid)) {
3979		log_error(ls, "receive_unlock %x remid %x remote %d %x",
3980			  lkb->lkb_id, lkb->lkb_remid,
3981			  le32_to_cpu(ms->m_header.h_nodeid),
3982			  le32_to_cpu(ms->m_lkid));
3983		error = -ENOENT;
3984		dlm_put_lkb(lkb);
3985		goto fail;
3986	}
3987
3988	r = lkb->lkb_resource;
3989
3990	hold_rsb(r);
3991	lock_rsb(r);
3992
3993	error = validate_message(lkb, ms);
3994	if (error)
3995		goto out;
3996
3997	receive_flags(lkb, ms);
3998
3999	error = receive_unlock_args(ls, lkb, ms);
4000	if (error) {
4001		send_unlock_reply(r, lkb, error);
4002		goto out;
4003	}
4004
4005	error = do_unlock(r, lkb);
4006	send_unlock_reply(r, lkb, error);
4007	do_unlock_effects(r, lkb, error);
4008 out:
4009	unlock_rsb(r);
4010	put_rsb(r);
4011	dlm_put_lkb(lkb);
4012	return 0;
4013
4014 fail:
4015	setup_local_lkb(ls, ms);
4016	send_unlock_reply(&ls->ls_local_rsb, &ls->ls_local_lkb, error);
4017	return error;
4018}
4019
4020static int receive_cancel(struct dlm_ls *ls, const struct dlm_message *ms)
4021{
4022	struct dlm_lkb *lkb;
4023	struct dlm_rsb *r;
4024	int error;
4025
4026	error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
4027	if (error)
4028		goto fail;
4029
4030	receive_flags(lkb, ms);
4031
4032	r = lkb->lkb_resource;
4033
4034	hold_rsb(r);
4035	lock_rsb(r);
4036
4037	error = validate_message(lkb, ms);
4038	if (error)
4039		goto out;
4040
4041	error = do_cancel(r, lkb);
4042	send_cancel_reply(r, lkb, error);
4043	do_cancel_effects(r, lkb, error);
4044 out:
4045	unlock_rsb(r);
4046	put_rsb(r);
4047	dlm_put_lkb(lkb);
4048	return 0;
4049
4050 fail:
4051	setup_local_lkb(ls, ms);
4052	send_cancel_reply(&ls->ls_local_rsb, &ls->ls_local_lkb, error);
4053	return error;
4054}
4055
4056static int receive_grant(struct dlm_ls *ls, const struct dlm_message *ms)
4057{
4058	struct dlm_lkb *lkb;
4059	struct dlm_rsb *r;
4060	int error;
4061
4062	error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
4063	if (error)
4064		return error;
4065
4066	r = lkb->lkb_resource;
4067
4068	hold_rsb(r);
4069	lock_rsb(r);
4070
4071	error = validate_message(lkb, ms);
4072	if (error)
4073		goto out;
4074
4075	receive_flags_reply(lkb, ms, false);
4076	if (is_altmode(lkb))
4077		munge_altmode(lkb, ms);
4078	grant_lock_pc(r, lkb, ms);
4079	queue_cast(r, lkb, 0);
4080 out:
4081	unlock_rsb(r);
4082	put_rsb(r);
4083	dlm_put_lkb(lkb);
4084	return 0;
4085}
4086
4087static int receive_bast(struct dlm_ls *ls, const struct dlm_message *ms)
4088{
4089	struct dlm_lkb *lkb;
4090	struct dlm_rsb *r;
4091	int error;
4092
4093	error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
4094	if (error)
4095		return error;
4096
4097	r = lkb->lkb_resource;
4098
4099	hold_rsb(r);
4100	lock_rsb(r);
4101
4102	error = validate_message(lkb, ms);
4103	if (error)
4104		goto out;
4105
4106	queue_bast(r, lkb, le32_to_cpu(ms->m_bastmode));
4107	lkb->lkb_highbast = le32_to_cpu(ms->m_bastmode);
4108 out:
4109	unlock_rsb(r);
4110	put_rsb(r);
4111	dlm_put_lkb(lkb);
4112	return 0;
4113}
4114
4115static void receive_lookup(struct dlm_ls *ls, const struct dlm_message *ms)
4116{
4117	int len, error, ret_nodeid, from_nodeid, our_nodeid;
4118
4119	from_nodeid = le32_to_cpu(ms->m_header.h_nodeid);
4120	our_nodeid = dlm_our_nodeid();
4121
4122	len = receive_extralen(ms);
4123
4124	error = dlm_master_lookup(ls, from_nodeid, ms->m_extra, len, 0,
4125				  &ret_nodeid, NULL);
4126
4127	/* Optimization: we're master so treat lookup as a request */
4128	if (!error && ret_nodeid == our_nodeid) {
4129		receive_request(ls, ms);
4130		return;
4131	}
4132	send_lookup_reply(ls, ms, ret_nodeid, error);
4133}
4134
4135static void receive_remove(struct dlm_ls *ls, const struct dlm_message *ms)
4136{
4137	char name[DLM_RESNAME_MAXLEN+1];
4138	struct dlm_rsb *r;
4139	uint32_t hash, b;
4140	int rv, len, dir_nodeid, from_nodeid;
4141
4142	from_nodeid = le32_to_cpu(ms->m_header.h_nodeid);
4143
4144	len = receive_extralen(ms);
4145
4146	if (len > DLM_RESNAME_MAXLEN) {
4147		log_error(ls, "receive_remove from %d bad len %d",
4148			  from_nodeid, len);
4149		return;
4150	}
4151
4152	dir_nodeid = dlm_hash2nodeid(ls, le32_to_cpu(ms->m_hash));
4153	if (dir_nodeid != dlm_our_nodeid()) {
4154		log_error(ls, "receive_remove from %d bad nodeid %d",
4155			  from_nodeid, dir_nodeid);
4156		return;
4157	}
4158
4159	/* Look for name on rsbtbl.toss, if it's there, kill it.
4160	   If it's on rsbtbl.keep, it's being used, and we should ignore this
4161	   message.  This is an expected race between the dir node sending a
4162	   request to the master node at the same time as the master node sends
4163	   a remove to the dir node.  The resolution to that race is for the
4164	   dir node to ignore the remove message, and the master node to
4165	   recreate the master rsb when it gets a request from the dir node for
4166	   an rsb it doesn't have. */
4167
4168	memset(name, 0, sizeof(name));
4169	memcpy(name, ms->m_extra, len);
4170
4171	hash = jhash(name, len, 0);
4172	b = hash & (ls->ls_rsbtbl_size - 1);
4173
4174	spin_lock(&ls->ls_rsbtbl[b].lock);
4175
4176	rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
4177	if (rv) {
4178		/* verify the rsb is on keep list per comment above */
4179		rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
4180		if (rv) {
4181			/* should not happen */
4182			log_error(ls, "receive_remove from %d not found %s",
4183				  from_nodeid, name);
4184			spin_unlock(&ls->ls_rsbtbl[b].lock);
4185			return;
4186		}
4187		if (r->res_master_nodeid != from_nodeid) {
4188			/* should not happen */
4189			log_error(ls, "receive_remove keep from %d master %d",
4190				  from_nodeid, r->res_master_nodeid);
4191			dlm_print_rsb(r);
4192			spin_unlock(&ls->ls_rsbtbl[b].lock);
4193			return;
4194		}
4195
4196		log_debug(ls, "receive_remove from %d master %d first %x %s",
4197			  from_nodeid, r->res_master_nodeid, r->res_first_lkid,
4198			  name);
4199		spin_unlock(&ls->ls_rsbtbl[b].lock);
4200		return;
4201	}
4202
4203	if (r->res_master_nodeid != from_nodeid) {
4204		log_error(ls, "receive_remove toss from %d master %d",
4205			  from_nodeid, r->res_master_nodeid);
4206		dlm_print_rsb(r);
4207		spin_unlock(&ls->ls_rsbtbl[b].lock);
4208		return;
4209	}
4210
4211	if (kref_put(&r->res_ref, kill_rsb)) {
4212		rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
4213		spin_unlock(&ls->ls_rsbtbl[b].lock);
4214		dlm_free_rsb(r);
4215	} else {
4216		log_error(ls, "receive_remove from %d rsb ref error",
4217			  from_nodeid);
4218		dlm_print_rsb(r);
4219		spin_unlock(&ls->ls_rsbtbl[b].lock);
4220	}
4221}
4222
4223static void receive_purge(struct dlm_ls *ls, const struct dlm_message *ms)
4224{
4225	do_purge(ls, le32_to_cpu(ms->m_nodeid), le32_to_cpu(ms->m_pid));
4226}
4227
4228static int receive_request_reply(struct dlm_ls *ls,
4229				 const struct dlm_message *ms)
4230{
4231	struct dlm_lkb *lkb;
4232	struct dlm_rsb *r;
4233	int error, mstype, result;
4234	int from_nodeid = le32_to_cpu(ms->m_header.h_nodeid);
4235
4236	error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
4237	if (error)
4238		return error;
4239
4240	r = lkb->lkb_resource;
4241	hold_rsb(r);
4242	lock_rsb(r);
4243
4244	error = validate_message(lkb, ms);
4245	if (error)
4246		goto out;
4247
4248	mstype = lkb->lkb_wait_type;
4249	error = remove_from_waiters(lkb, DLM_MSG_REQUEST_REPLY);
4250	if (error) {
4251		log_error(ls, "receive_request_reply %x remote %d %x result %d",
4252			  lkb->lkb_id, from_nodeid, le32_to_cpu(ms->m_lkid),
4253			  from_dlm_errno(le32_to_cpu(ms->m_result)));
4254		dlm_dump_rsb(r);
4255		goto out;
4256	}
4257
4258	/* Optimization: the dir node was also the master, so it took our
4259	   lookup as a request and sent request reply instead of lookup reply */
4260	if (mstype == DLM_MSG_LOOKUP) {
4261		r->res_master_nodeid = from_nodeid;
4262		r->res_nodeid = from_nodeid;
4263		lkb->lkb_nodeid = from_nodeid;
4264	}
4265
4266	/* this is the value returned from do_request() on the master */
4267	result = from_dlm_errno(le32_to_cpu(ms->m_result));
4268
4269	switch (result) {
4270	case -EAGAIN:
4271		/* request would block (be queued) on remote master */
4272		queue_cast(r, lkb, -EAGAIN);
4273		confirm_master(r, -EAGAIN);
4274		unhold_lkb(lkb); /* undoes create_lkb() */
4275		break;
4276
4277	case -EINPROGRESS:
4278	case 0:
4279		/* request was queued or granted on remote master */
4280		receive_flags_reply(lkb, ms, false);
4281		lkb->lkb_remid = le32_to_cpu(ms->m_lkid);
4282		if (is_altmode(lkb))
4283			munge_altmode(lkb, ms);
4284		if (result) {
4285			add_lkb(r, lkb, DLM_LKSTS_WAITING);
 
4286		} else {
4287			grant_lock_pc(r, lkb, ms);
4288			queue_cast(r, lkb, 0);
4289		}
4290		confirm_master(r, result);
4291		break;
4292
4293	case -EBADR:
4294	case -ENOTBLK:
4295		/* find_rsb failed to find rsb or rsb wasn't master */
4296		log_limit(ls, "receive_request_reply %x from %d %d "
4297			  "master %d dir %d first %x %s", lkb->lkb_id,
4298			  from_nodeid, result, r->res_master_nodeid,
4299			  r->res_dir_nodeid, r->res_first_lkid, r->res_name);
4300
4301		if (r->res_dir_nodeid != dlm_our_nodeid() &&
4302		    r->res_master_nodeid != dlm_our_nodeid()) {
4303			/* cause _request_lock->set_master->send_lookup */
4304			r->res_master_nodeid = 0;
4305			r->res_nodeid = -1;
4306			lkb->lkb_nodeid = -1;
4307		}
4308
4309		if (is_overlap(lkb)) {
4310			/* we'll ignore error in cancel/unlock reply */
4311			queue_cast_overlap(r, lkb);
4312			confirm_master(r, result);
4313			unhold_lkb(lkb); /* undoes create_lkb() */
4314		} else {
4315			_request_lock(r, lkb);
4316
4317			if (r->res_master_nodeid == dlm_our_nodeid())
4318				confirm_master(r, 0);
4319		}
4320		break;
4321
4322	default:
4323		log_error(ls, "receive_request_reply %x error %d",
4324			  lkb->lkb_id, result);
4325	}
4326
4327	if ((result == 0 || result == -EINPROGRESS) &&
4328	    test_and_clear_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags)) {
4329		log_debug(ls, "receive_request_reply %x result %d unlock",
4330			  lkb->lkb_id, result);
4331		clear_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags);
 
4332		send_unlock(r, lkb);
4333	} else if ((result == -EINPROGRESS) &&
4334		   test_and_clear_bit(DLM_IFL_OVERLAP_CANCEL_BIT,
4335				      &lkb->lkb_iflags)) {
4336		log_debug(ls, "receive_request_reply %x cancel", lkb->lkb_id);
4337		clear_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags);
 
4338		send_cancel(r, lkb);
4339	} else {
4340		clear_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags);
4341		clear_bit(DLM_IFL_OVERLAP_UNLOCK_BIT, &lkb->lkb_iflags);
4342	}
4343 out:
4344	unlock_rsb(r);
4345	put_rsb(r);
4346	dlm_put_lkb(lkb);
4347	return 0;
4348}
4349
4350static void __receive_convert_reply(struct dlm_rsb *r, struct dlm_lkb *lkb,
4351				    const struct dlm_message *ms, bool local)
4352{
4353	/* this is the value returned from do_convert() on the master */
4354	switch (from_dlm_errno(le32_to_cpu(ms->m_result))) {
4355	case -EAGAIN:
4356		/* convert would block (be queued) on remote master */
4357		queue_cast(r, lkb, -EAGAIN);
4358		break;
4359
4360	case -EDEADLK:
4361		receive_flags_reply(lkb, ms, local);
4362		revert_lock_pc(r, lkb);
4363		queue_cast(r, lkb, -EDEADLK);
4364		break;
4365
4366	case -EINPROGRESS:
4367		/* convert was queued on remote master */
4368		receive_flags_reply(lkb, ms, local);
4369		if (is_demoted(lkb))
4370			munge_demoted(lkb);
4371		del_lkb(r, lkb);
4372		add_lkb(r, lkb, DLM_LKSTS_CONVERT);
 
4373		break;
4374
4375	case 0:
4376		/* convert was granted on remote master */
4377		receive_flags_reply(lkb, ms, local);
4378		if (is_demoted(lkb))
4379			munge_demoted(lkb);
4380		grant_lock_pc(r, lkb, ms);
4381		queue_cast(r, lkb, 0);
4382		break;
4383
4384	default:
4385		log_error(r->res_ls, "receive_convert_reply %x remote %d %x %d",
4386			  lkb->lkb_id, le32_to_cpu(ms->m_header.h_nodeid),
4387			  le32_to_cpu(ms->m_lkid),
4388			  from_dlm_errno(le32_to_cpu(ms->m_result)));
4389		dlm_print_rsb(r);
4390		dlm_print_lkb(lkb);
4391	}
4392}
4393
4394static void _receive_convert_reply(struct dlm_lkb *lkb,
4395				   const struct dlm_message *ms, bool local)
4396{
4397	struct dlm_rsb *r = lkb->lkb_resource;
4398	int error;
4399
4400	hold_rsb(r);
4401	lock_rsb(r);
4402
4403	error = validate_message(lkb, ms);
4404	if (error)
4405		goto out;
4406
4407	/* local reply can happen with waiters_mutex held */
4408	error = remove_from_waiters_ms(lkb, ms, local);
4409	if (error)
4410		goto out;
4411
4412	__receive_convert_reply(r, lkb, ms, local);
4413 out:
4414	unlock_rsb(r);
4415	put_rsb(r);
4416}
4417
4418static int receive_convert_reply(struct dlm_ls *ls,
4419				 const struct dlm_message *ms)
4420{
4421	struct dlm_lkb *lkb;
4422	int error;
4423
4424	error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
4425	if (error)
4426		return error;
4427
4428	_receive_convert_reply(lkb, ms, false);
4429	dlm_put_lkb(lkb);
4430	return 0;
4431}
4432
4433static void _receive_unlock_reply(struct dlm_lkb *lkb,
4434				  const struct dlm_message *ms, bool local)
4435{
4436	struct dlm_rsb *r = lkb->lkb_resource;
4437	int error;
4438
4439	hold_rsb(r);
4440	lock_rsb(r);
4441
4442	error = validate_message(lkb, ms);
4443	if (error)
4444		goto out;
4445
4446	/* local reply can happen with waiters_mutex held */
4447	error = remove_from_waiters_ms(lkb, ms, local);
4448	if (error)
4449		goto out;
4450
4451	/* this is the value returned from do_unlock() on the master */
4452
4453	switch (from_dlm_errno(le32_to_cpu(ms->m_result))) {
4454	case -DLM_EUNLOCK:
4455		receive_flags_reply(lkb, ms, local);
4456		remove_lock_pc(r, lkb);
4457		queue_cast(r, lkb, -DLM_EUNLOCK);
4458		break;
4459	case -ENOENT:
4460		break;
4461	default:
4462		log_error(r->res_ls, "receive_unlock_reply %x error %d",
4463			  lkb->lkb_id, from_dlm_errno(le32_to_cpu(ms->m_result)));
4464	}
4465 out:
4466	unlock_rsb(r);
4467	put_rsb(r);
4468}
4469
4470static int receive_unlock_reply(struct dlm_ls *ls,
4471				const struct dlm_message *ms)
4472{
4473	struct dlm_lkb *lkb;
4474	int error;
4475
4476	error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
4477	if (error)
4478		return error;
4479
4480	_receive_unlock_reply(lkb, ms, false);
4481	dlm_put_lkb(lkb);
4482	return 0;
4483}
4484
4485static void _receive_cancel_reply(struct dlm_lkb *lkb,
4486				  const struct dlm_message *ms, bool local)
4487{
4488	struct dlm_rsb *r = lkb->lkb_resource;
4489	int error;
4490
4491	hold_rsb(r);
4492	lock_rsb(r);
4493
4494	error = validate_message(lkb, ms);
4495	if (error)
4496		goto out;
4497
4498	/* local reply can happen with waiters_mutex held */
4499	error = remove_from_waiters_ms(lkb, ms, local);
4500	if (error)
4501		goto out;
4502
4503	/* this is the value returned from do_cancel() on the master */
4504
4505	switch (from_dlm_errno(le32_to_cpu(ms->m_result))) {
4506	case -DLM_ECANCEL:
4507		receive_flags_reply(lkb, ms, local);
4508		revert_lock_pc(r, lkb);
4509		queue_cast(r, lkb, -DLM_ECANCEL);
4510		break;
4511	case 0:
4512		break;
4513	default:
4514		log_error(r->res_ls, "receive_cancel_reply %x error %d",
4515			  lkb->lkb_id,
4516			  from_dlm_errno(le32_to_cpu(ms->m_result)));
4517	}
4518 out:
4519	unlock_rsb(r);
4520	put_rsb(r);
4521}
4522
4523static int receive_cancel_reply(struct dlm_ls *ls,
4524				const struct dlm_message *ms)
4525{
4526	struct dlm_lkb *lkb;
4527	int error;
4528
4529	error = find_lkb(ls, le32_to_cpu(ms->m_remid), &lkb);
4530	if (error)
4531		return error;
4532
4533	_receive_cancel_reply(lkb, ms, false);
4534	dlm_put_lkb(lkb);
4535	return 0;
4536}
4537
4538static void receive_lookup_reply(struct dlm_ls *ls,
4539				 const struct dlm_message *ms)
4540{
4541	struct dlm_lkb *lkb;
4542	struct dlm_rsb *r;
4543	int error, ret_nodeid;
4544	int do_lookup_list = 0;
4545
4546	error = find_lkb(ls, le32_to_cpu(ms->m_lkid), &lkb);
4547	if (error) {
4548		log_error(ls, "%s no lkid %x", __func__,
4549			  le32_to_cpu(ms->m_lkid));
4550		return;
4551	}
4552
4553	/* ms->m_result is the value returned by dlm_master_lookup on dir node
4554	   FIXME: will a non-zero error ever be returned? */
4555
4556	r = lkb->lkb_resource;
4557	hold_rsb(r);
4558	lock_rsb(r);
4559
4560	error = remove_from_waiters(lkb, DLM_MSG_LOOKUP_REPLY);
4561	if (error)
4562		goto out;
4563
4564	ret_nodeid = le32_to_cpu(ms->m_nodeid);
4565
4566	/* We sometimes receive a request from the dir node for this
4567	   rsb before we've received the dir node's loookup_reply for it.
4568	   The request from the dir node implies we're the master, so we set
4569	   ourself as master in receive_request_reply, and verify here that
4570	   we are indeed the master. */
4571
4572	if (r->res_master_nodeid && (r->res_master_nodeid != ret_nodeid)) {
4573		/* This should never happen */
4574		log_error(ls, "receive_lookup_reply %x from %d ret %d "
4575			  "master %d dir %d our %d first %x %s",
4576			  lkb->lkb_id, le32_to_cpu(ms->m_header.h_nodeid),
4577			  ret_nodeid, r->res_master_nodeid, r->res_dir_nodeid,
4578			  dlm_our_nodeid(), r->res_first_lkid, r->res_name);
4579	}
4580
4581	if (ret_nodeid == dlm_our_nodeid()) {
4582		r->res_master_nodeid = ret_nodeid;
4583		r->res_nodeid = 0;
4584		do_lookup_list = 1;
4585		r->res_first_lkid = 0;
4586	} else if (ret_nodeid == -1) {
4587		/* the remote node doesn't believe it's the dir node */
4588		log_error(ls, "receive_lookup_reply %x from %d bad ret_nodeid",
4589			  lkb->lkb_id, le32_to_cpu(ms->m_header.h_nodeid));
4590		r->res_master_nodeid = 0;
4591		r->res_nodeid = -1;
4592		lkb->lkb_nodeid = -1;
4593	} else {
4594		/* set_master() will set lkb_nodeid from r */
4595		r->res_master_nodeid = ret_nodeid;
4596		r->res_nodeid = ret_nodeid;
4597	}
4598
4599	if (is_overlap(lkb)) {
4600		log_debug(ls, "receive_lookup_reply %x unlock %x",
4601			  lkb->lkb_id, dlm_iflags_val(lkb));
4602		queue_cast_overlap(r, lkb);
4603		unhold_lkb(lkb); /* undoes create_lkb() */
4604		goto out_list;
4605	}
4606
4607	_request_lock(r, lkb);
4608
4609 out_list:
4610	if (do_lookup_list)
4611		process_lookup_list(r);
4612 out:
4613	unlock_rsb(r);
4614	put_rsb(r);
4615	dlm_put_lkb(lkb);
4616}
4617
4618static void _receive_message(struct dlm_ls *ls, const struct dlm_message *ms,
4619			     uint32_t saved_seq)
4620{
4621	int error = 0, noent = 0;
4622
4623	if (WARN_ON_ONCE(!dlm_is_member(ls, le32_to_cpu(ms->m_header.h_nodeid)))) {
4624		log_limit(ls, "receive %d from non-member %d %x %x %d",
4625			  le32_to_cpu(ms->m_type),
4626			  le32_to_cpu(ms->m_header.h_nodeid),
4627			  le32_to_cpu(ms->m_lkid), le32_to_cpu(ms->m_remid),
4628			  from_dlm_errno(le32_to_cpu(ms->m_result)));
4629		return;
4630	}
4631
4632	switch (ms->m_type) {
4633
4634	/* messages sent to a master node */
4635
4636	case cpu_to_le32(DLM_MSG_REQUEST):
4637		error = receive_request(ls, ms);
4638		break;
4639
4640	case cpu_to_le32(DLM_MSG_CONVERT):
4641		error = receive_convert(ls, ms);
4642		break;
4643
4644	case cpu_to_le32(DLM_MSG_UNLOCK):
4645		error = receive_unlock(ls, ms);
4646		break;
4647
4648	case cpu_to_le32(DLM_MSG_CANCEL):
4649		noent = 1;
4650		error = receive_cancel(ls, ms);
4651		break;
4652
4653	/* messages sent from a master node (replies to above) */
4654
4655	case cpu_to_le32(DLM_MSG_REQUEST_REPLY):
4656		error = receive_request_reply(ls, ms);
4657		break;
4658
4659	case cpu_to_le32(DLM_MSG_CONVERT_REPLY):
4660		error = receive_convert_reply(ls, ms);
4661		break;
4662
4663	case cpu_to_le32(DLM_MSG_UNLOCK_REPLY):
4664		error = receive_unlock_reply(ls, ms);
4665		break;
4666
4667	case cpu_to_le32(DLM_MSG_CANCEL_REPLY):
4668		error = receive_cancel_reply(ls, ms);
4669		break;
4670
4671	/* messages sent from a master node (only two types of async msg) */
4672
4673	case cpu_to_le32(DLM_MSG_GRANT):
4674		noent = 1;
4675		error = receive_grant(ls, ms);
4676		break;
4677
4678	case cpu_to_le32(DLM_MSG_BAST):
4679		noent = 1;
4680		error = receive_bast(ls, ms);
4681		break;
4682
4683	/* messages sent to a dir node */
4684
4685	case cpu_to_le32(DLM_MSG_LOOKUP):
4686		receive_lookup(ls, ms);
4687		break;
4688
4689	case cpu_to_le32(DLM_MSG_REMOVE):
4690		receive_remove(ls, ms);
4691		break;
4692
4693	/* messages sent from a dir node (remove has no reply) */
4694
4695	case cpu_to_le32(DLM_MSG_LOOKUP_REPLY):
4696		receive_lookup_reply(ls, ms);
4697		break;
4698
4699	/* other messages */
4700
4701	case cpu_to_le32(DLM_MSG_PURGE):
4702		receive_purge(ls, ms);
4703		break;
4704
4705	default:
4706		log_error(ls, "unknown message type %d",
4707			  le32_to_cpu(ms->m_type));
4708	}
4709
4710	/*
4711	 * When checking for ENOENT, we're checking the result of
4712	 * find_lkb(m_remid):
4713	 *
4714	 * The lock id referenced in the message wasn't found.  This may
4715	 * happen in normal usage for the async messages and cancel, so
4716	 * only use log_debug for them.
4717	 *
4718	 * Some errors are expected and normal.
4719	 */
4720
4721	if (error == -ENOENT && noent) {
4722		log_debug(ls, "receive %d no %x remote %d %x saved_seq %u",
4723			  le32_to_cpu(ms->m_type), le32_to_cpu(ms->m_remid),
4724			  le32_to_cpu(ms->m_header.h_nodeid),
4725			  le32_to_cpu(ms->m_lkid), saved_seq);
4726	} else if (error == -ENOENT) {
4727		log_error(ls, "receive %d no %x remote %d %x saved_seq %u",
4728			  le32_to_cpu(ms->m_type), le32_to_cpu(ms->m_remid),
4729			  le32_to_cpu(ms->m_header.h_nodeid),
4730			  le32_to_cpu(ms->m_lkid), saved_seq);
4731
4732		if (ms->m_type == cpu_to_le32(DLM_MSG_CONVERT))
4733			dlm_dump_rsb_hash(ls, le32_to_cpu(ms->m_hash));
4734	}
4735
4736	if (error == -EINVAL) {
4737		log_error(ls, "receive %d inval from %d lkid %x remid %x "
4738			  "saved_seq %u",
4739			  le32_to_cpu(ms->m_type),
4740			  le32_to_cpu(ms->m_header.h_nodeid),
4741			  le32_to_cpu(ms->m_lkid), le32_to_cpu(ms->m_remid),
4742			  saved_seq);
4743	}
4744}
4745
4746/* If the lockspace is in recovery mode (locking stopped), then normal
4747   messages are saved on the requestqueue for processing after recovery is
4748   done.  When not in recovery mode, we wait for dlm_recoverd to drain saved
4749   messages off the requestqueue before we process new ones. This occurs right
4750   after recovery completes when we transition from saving all messages on
4751   requestqueue, to processing all the saved messages, to processing new
4752   messages as they arrive. */
4753
4754static void dlm_receive_message(struct dlm_ls *ls, const struct dlm_message *ms,
4755				int nodeid)
4756{
4757	if (dlm_locking_stopped(ls)) {
4758		/* If we were a member of this lockspace, left, and rejoined,
4759		   other nodes may still be sending us messages from the
4760		   lockspace generation before we left. */
4761		if (WARN_ON_ONCE(!ls->ls_generation)) {
4762			log_limit(ls, "receive %d from %d ignore old gen",
4763				  le32_to_cpu(ms->m_type), nodeid);
4764			return;
4765		}
4766
4767		dlm_add_requestqueue(ls, nodeid, ms);
4768	} else {
4769		dlm_wait_requestqueue(ls);
4770		_receive_message(ls, ms, 0);
4771	}
4772}
4773
4774/* This is called by dlm_recoverd to process messages that were saved on
4775   the requestqueue. */
4776
4777void dlm_receive_message_saved(struct dlm_ls *ls, const struct dlm_message *ms,
4778			       uint32_t saved_seq)
4779{
4780	_receive_message(ls, ms, saved_seq);
4781}
4782
4783/* This is called by the midcomms layer when something is received for
4784   the lockspace.  It could be either a MSG (normal message sent as part of
4785   standard locking activity) or an RCOM (recovery message sent as part of
4786   lockspace recovery). */
4787
4788void dlm_receive_buffer(const union dlm_packet *p, int nodeid)
4789{
4790	const struct dlm_header *hd = &p->header;
4791	struct dlm_ls *ls;
4792	int type = 0;
4793
4794	switch (hd->h_cmd) {
4795	case DLM_MSG:
4796		type = le32_to_cpu(p->message.m_type);
 
4797		break;
4798	case DLM_RCOM:
4799		type = le32_to_cpu(p->rcom.rc_type);
 
4800		break;
4801	default:
4802		log_print("invalid h_cmd %d from %u", hd->h_cmd, nodeid);
4803		return;
4804	}
4805
4806	if (le32_to_cpu(hd->h_nodeid) != nodeid) {
4807		log_print("invalid h_nodeid %d from %d lockspace %x",
4808			  le32_to_cpu(hd->h_nodeid), nodeid,
4809			  le32_to_cpu(hd->u.h_lockspace));
4810		return;
4811	}
4812
4813	ls = dlm_find_lockspace_global(le32_to_cpu(hd->u.h_lockspace));
4814	if (!ls) {
4815		if (dlm_config.ci_log_debug) {
4816			printk_ratelimited(KERN_DEBUG "dlm: invalid lockspace "
4817				"%u from %d cmd %d type %d\n",
4818				le32_to_cpu(hd->u.h_lockspace), nodeid,
4819				hd->h_cmd, type);
4820		}
4821
4822		if (hd->h_cmd == DLM_RCOM && type == DLM_RCOM_STATUS)
4823			dlm_send_ls_not_ready(nodeid, &p->rcom);
4824		return;
4825	}
4826
4827	/* this rwsem allows dlm_ls_stop() to wait for all dlm_recv threads to
4828	   be inactive (in this ls) before transitioning to recovery mode */
4829
4830	down_read(&ls->ls_recv_active);
4831	if (hd->h_cmd == DLM_MSG)
4832		dlm_receive_message(ls, &p->message, nodeid);
4833	else if (hd->h_cmd == DLM_RCOM)
4834		dlm_receive_rcom(ls, &p->rcom, nodeid);
4835	else
4836		log_error(ls, "invalid h_cmd %d from %d lockspace %x",
4837			  hd->h_cmd, nodeid, le32_to_cpu(hd->u.h_lockspace));
4838	up_read(&ls->ls_recv_active);
4839
4840	dlm_put_lockspace(ls);
4841}
4842
4843static void recover_convert_waiter(struct dlm_ls *ls, struct dlm_lkb *lkb,
4844				   struct dlm_message *ms_local)
4845{
4846	if (middle_conversion(lkb)) {
4847		hold_lkb(lkb);
4848		memset(ms_local, 0, sizeof(struct dlm_message));
4849		ms_local->m_type = cpu_to_le32(DLM_MSG_CONVERT_REPLY);
4850		ms_local->m_result = cpu_to_le32(to_dlm_errno(-EINPROGRESS));
4851		ms_local->m_header.h_nodeid = cpu_to_le32(lkb->lkb_nodeid);
4852		_receive_convert_reply(lkb, ms_local, true);
 
4853
4854		/* Same special case as in receive_rcom_lock_args() */
4855		lkb->lkb_grmode = DLM_LOCK_IV;
4856		rsb_set_flag(lkb->lkb_resource, RSB_RECOVER_CONVERT);
4857		unhold_lkb(lkb);
4858
4859	} else if (lkb->lkb_rqmode >= lkb->lkb_grmode) {
4860		set_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags);
4861	}
4862
4863	/* lkb->lkb_rqmode < lkb->lkb_grmode shouldn't happen since down
4864	   conversions are async; there's no reply from the remote master */
4865}
4866
4867/* A waiting lkb needs recovery if the master node has failed, or
4868   the master node is changing (only when no directory is used) */
4869
4870static int waiter_needs_recovery(struct dlm_ls *ls, struct dlm_lkb *lkb,
4871				 int dir_nodeid)
4872{
4873	if (dlm_no_directory(ls))
4874		return 1;
4875
4876	if (dlm_is_removed(ls, lkb->lkb_wait_nodeid))
4877		return 1;
4878
4879	return 0;
4880}
4881
4882/* Recovery for locks that are waiting for replies from nodes that are now
4883   gone.  We can just complete unlocks and cancels by faking a reply from the
4884   dead node.  Requests and up-conversions we flag to be resent after
4885   recovery.  Down-conversions can just be completed with a fake reply like
4886   unlocks.  Conversions between PR and CW need special attention. */
4887
4888void dlm_recover_waiters_pre(struct dlm_ls *ls)
4889{
4890	struct dlm_lkb *lkb, *safe;
4891	struct dlm_message *ms_local;
4892	int wait_type, local_unlock_result, local_cancel_result;
4893	int dir_nodeid;
4894
4895	ms_local = kmalloc(sizeof(*ms_local), GFP_KERNEL);
4896	if (!ms_local)
4897		return;
4898
4899	mutex_lock(&ls->ls_waiters_mutex);
4900
4901	list_for_each_entry_safe(lkb, safe, &ls->ls_waiters, lkb_wait_reply) {
4902
4903		dir_nodeid = dlm_dir_nodeid(lkb->lkb_resource);
4904
4905		/* exclude debug messages about unlocks because there can be so
4906		   many and they aren't very interesting */
4907
4908		if (lkb->lkb_wait_type != DLM_MSG_UNLOCK) {
4909			log_debug(ls, "waiter %x remote %x msg %d r_nodeid %d "
4910				  "lkb_nodeid %d wait_nodeid %d dir_nodeid %d",
4911				  lkb->lkb_id,
4912				  lkb->lkb_remid,
4913				  lkb->lkb_wait_type,
4914				  lkb->lkb_resource->res_nodeid,
4915				  lkb->lkb_nodeid,
4916				  lkb->lkb_wait_nodeid,
4917				  dir_nodeid);
4918		}
4919
4920		/* all outstanding lookups, regardless of destination  will be
4921		   resent after recovery is done */
4922
4923		if (lkb->lkb_wait_type == DLM_MSG_LOOKUP) {
4924			set_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags);
4925			continue;
4926		}
4927
4928		if (!waiter_needs_recovery(ls, lkb, dir_nodeid))
4929			continue;
4930
4931		wait_type = lkb->lkb_wait_type;
4932		local_unlock_result = -DLM_EUNLOCK;
4933		local_cancel_result = -DLM_ECANCEL;
4934
4935		/* Main reply may have been received leaving a zero wait_type,
4936		   but a reply for the overlapping op may not have been
4937		   received.  In that case we need to fake the appropriate
4938		   reply for the overlap op. */
4939
4940		if (!wait_type) {
4941			if (is_overlap_cancel(lkb)) {
4942				wait_type = DLM_MSG_CANCEL;
4943				if (lkb->lkb_grmode == DLM_LOCK_IV)
4944					local_cancel_result = 0;
4945			}
4946			if (is_overlap_unlock(lkb)) {
4947				wait_type = DLM_MSG_UNLOCK;
4948				if (lkb->lkb_grmode == DLM_LOCK_IV)
4949					local_unlock_result = -ENOENT;
4950			}
4951
4952			log_debug(ls, "rwpre overlap %x %x %d %d %d",
4953				  lkb->lkb_id, dlm_iflags_val(lkb), wait_type,
4954				  local_cancel_result, local_unlock_result);
4955		}
4956
4957		switch (wait_type) {
4958
4959		case DLM_MSG_REQUEST:
4960			set_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags);
4961			break;
4962
4963		case DLM_MSG_CONVERT:
4964			recover_convert_waiter(ls, lkb, ms_local);
4965			break;
4966
4967		case DLM_MSG_UNLOCK:
4968			hold_lkb(lkb);
4969			memset(ms_local, 0, sizeof(struct dlm_message));
4970			ms_local->m_type = cpu_to_le32(DLM_MSG_UNLOCK_REPLY);
4971			ms_local->m_result = cpu_to_le32(to_dlm_errno(local_unlock_result));
4972			ms_local->m_header.h_nodeid = cpu_to_le32(lkb->lkb_nodeid);
4973			_receive_unlock_reply(lkb, ms_local, true);
 
4974			dlm_put_lkb(lkb);
4975			break;
4976
4977		case DLM_MSG_CANCEL:
4978			hold_lkb(lkb);
4979			memset(ms_local, 0, sizeof(struct dlm_message));
4980			ms_local->m_type = cpu_to_le32(DLM_MSG_CANCEL_REPLY);
4981			ms_local->m_result = cpu_to_le32(to_dlm_errno(local_cancel_result));
4982			ms_local->m_header.h_nodeid = cpu_to_le32(lkb->lkb_nodeid);
4983			_receive_cancel_reply(lkb, ms_local, true);
 
4984			dlm_put_lkb(lkb);
4985			break;
4986
4987		default:
4988			log_error(ls, "invalid lkb wait_type %d %d",
4989				  lkb->lkb_wait_type, wait_type);
4990		}
4991		schedule();
4992	}
4993	mutex_unlock(&ls->ls_waiters_mutex);
4994	kfree(ms_local);
4995}
4996
4997static struct dlm_lkb *find_resend_waiter(struct dlm_ls *ls)
4998{
4999	struct dlm_lkb *lkb = NULL, *iter;
 
5000
5001	mutex_lock(&ls->ls_waiters_mutex);
5002	list_for_each_entry(iter, &ls->ls_waiters, lkb_wait_reply) {
5003		if (test_bit(DLM_IFL_RESEND_BIT, &iter->lkb_iflags)) {
5004			hold_lkb(iter);
5005			lkb = iter;
5006			break;
5007		}
5008	}
5009	mutex_unlock(&ls->ls_waiters_mutex);
5010
 
 
5011	return lkb;
5012}
5013
5014/* Deal with lookups and lkb's marked RESEND from _pre.  We may now be the
5015   master or dir-node for r.  Processing the lkb may result in it being placed
5016   back on waiters. */
5017
5018/* We do this after normal locking has been enabled and any saved messages
5019   (in requestqueue) have been processed.  We should be confident that at
5020   this point we won't get or process a reply to any of these waiting
5021   operations.  But, new ops may be coming in on the rsbs/locks here from
5022   userspace or remotely. */
5023
5024/* there may have been an overlap unlock/cancel prior to recovery or after
5025   recovery.  if before, the lkb may still have a pos wait_count; if after, the
5026   overlap flag would just have been set and nothing new sent.  we can be
5027   confident here than any replies to either the initial op or overlap ops
5028   prior to recovery have been received. */
5029
5030int dlm_recover_waiters_post(struct dlm_ls *ls)
5031{
5032	struct dlm_lkb *lkb;
5033	struct dlm_rsb *r;
5034	int error = 0, mstype, err, oc, ou;
5035
5036	while (1) {
5037		if (dlm_locking_stopped(ls)) {
5038			log_debug(ls, "recover_waiters_post aborted");
5039			error = -EINTR;
5040			break;
5041		}
5042
5043		lkb = find_resend_waiter(ls);
5044		if (!lkb)
5045			break;
5046
5047		r = lkb->lkb_resource;
5048		hold_rsb(r);
5049		lock_rsb(r);
5050
5051		mstype = lkb->lkb_wait_type;
5052		oc = test_and_clear_bit(DLM_IFL_OVERLAP_CANCEL_BIT,
5053					&lkb->lkb_iflags);
5054		ou = test_and_clear_bit(DLM_IFL_OVERLAP_UNLOCK_BIT,
5055					&lkb->lkb_iflags);
5056		err = 0;
5057
5058		log_debug(ls, "waiter %x remote %x msg %d r_nodeid %d "
5059			  "lkb_nodeid %d wait_nodeid %d dir_nodeid %d "
5060			  "overlap %d %d", lkb->lkb_id, lkb->lkb_remid, mstype,
5061			  r->res_nodeid, lkb->lkb_nodeid, lkb->lkb_wait_nodeid,
5062			  dlm_dir_nodeid(r), oc, ou);
5063
5064		/* At this point we assume that we won't get a reply to any
5065		   previous op or overlap op on this lock.  First, do a big
5066		   remove_from_waiters() for all previous ops. */
5067
5068		clear_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags);
 
 
5069		lkb->lkb_wait_type = 0;
5070		/* drop all wait_count references we still
5071		 * hold a reference for this iteration.
5072		 */
5073		while (!atomic_dec_and_test(&lkb->lkb_wait_count))
5074			unhold_lkb(lkb);
5075
5076		mutex_lock(&ls->ls_waiters_mutex);
5077		list_del_init(&lkb->lkb_wait_reply);
5078		mutex_unlock(&ls->ls_waiters_mutex);
 
5079
5080		if (oc || ou) {
5081			/* do an unlock or cancel instead of resending */
5082			switch (mstype) {
5083			case DLM_MSG_LOOKUP:
5084			case DLM_MSG_REQUEST:
5085				queue_cast(r, lkb, ou ? -DLM_EUNLOCK :
5086							-DLM_ECANCEL);
5087				unhold_lkb(lkb); /* undoes create_lkb() */
5088				break;
5089			case DLM_MSG_CONVERT:
5090				if (oc) {
5091					queue_cast(r, lkb, -DLM_ECANCEL);
5092				} else {
5093					lkb->lkb_exflags |= DLM_LKF_FORCEUNLOCK;
5094					_unlock_lock(r, lkb);
5095				}
5096				break;
5097			default:
5098				err = 1;
5099			}
5100		} else {
5101			switch (mstype) {
5102			case DLM_MSG_LOOKUP:
5103			case DLM_MSG_REQUEST:
5104				_request_lock(r, lkb);
5105				if (is_master(r))
5106					confirm_master(r, 0);
5107				break;
5108			case DLM_MSG_CONVERT:
5109				_convert_lock(r, lkb);
5110				break;
5111			default:
5112				err = 1;
5113			}
5114		}
5115
5116		if (err) {
5117			log_error(ls, "waiter %x msg %d r_nodeid %d "
5118				  "dir_nodeid %d overlap %d %d",
5119				  lkb->lkb_id, mstype, r->res_nodeid,
5120				  dlm_dir_nodeid(r), oc, ou);
5121		}
5122		unlock_rsb(r);
5123		put_rsb(r);
5124		dlm_put_lkb(lkb);
5125	}
5126
5127	return error;
5128}
5129
5130static void purge_mstcpy_list(struct dlm_ls *ls, struct dlm_rsb *r,
5131			      struct list_head *list)
5132{
5133	struct dlm_lkb *lkb, *safe;
5134
5135	list_for_each_entry_safe(lkb, safe, list, lkb_statequeue) {
5136		if (!is_master_copy(lkb))
5137			continue;
5138
5139		/* don't purge lkbs we've added in recover_master_copy for
5140		   the current recovery seq */
5141
5142		if (lkb->lkb_recover_seq == ls->ls_recover_seq)
5143			continue;
5144
5145		del_lkb(r, lkb);
5146
5147		/* this put should free the lkb */
5148		if (!dlm_put_lkb(lkb))
5149			log_error(ls, "purged mstcpy lkb not released");
5150	}
5151}
5152
5153void dlm_purge_mstcpy_locks(struct dlm_rsb *r)
5154{
5155	struct dlm_ls *ls = r->res_ls;
5156
5157	purge_mstcpy_list(ls, r, &r->res_grantqueue);
5158	purge_mstcpy_list(ls, r, &r->res_convertqueue);
5159	purge_mstcpy_list(ls, r, &r->res_waitqueue);
5160}
5161
5162static void purge_dead_list(struct dlm_ls *ls, struct dlm_rsb *r,
5163			    struct list_head *list,
5164			    int nodeid_gone, unsigned int *count)
5165{
5166	struct dlm_lkb *lkb, *safe;
5167
5168	list_for_each_entry_safe(lkb, safe, list, lkb_statequeue) {
5169		if (!is_master_copy(lkb))
5170			continue;
5171
5172		if ((lkb->lkb_nodeid == nodeid_gone) ||
5173		    dlm_is_removed(ls, lkb->lkb_nodeid)) {
5174
5175			/* tell recover_lvb to invalidate the lvb
5176			   because a node holding EX/PW failed */
5177			if ((lkb->lkb_exflags & DLM_LKF_VALBLK) &&
5178			    (lkb->lkb_grmode >= DLM_LOCK_PW)) {
5179				rsb_set_flag(r, RSB_RECOVER_LVB_INVAL);
5180			}
5181
5182			del_lkb(r, lkb);
5183
5184			/* this put should free the lkb */
5185			if (!dlm_put_lkb(lkb))
5186				log_error(ls, "purged dead lkb not released");
5187
5188			rsb_set_flag(r, RSB_RECOVER_GRANT);
5189
5190			(*count)++;
5191		}
5192	}
5193}
5194
5195/* Get rid of locks held by nodes that are gone. */
5196
5197void dlm_recover_purge(struct dlm_ls *ls)
5198{
5199	struct dlm_rsb *r;
5200	struct dlm_member *memb;
5201	int nodes_count = 0;
5202	int nodeid_gone = 0;
5203	unsigned int lkb_count = 0;
5204
5205	/* cache one removed nodeid to optimize the common
5206	   case of a single node removed */
5207
5208	list_for_each_entry(memb, &ls->ls_nodes_gone, list) {
5209		nodes_count++;
5210		nodeid_gone = memb->nodeid;
5211	}
5212
5213	if (!nodes_count)
5214		return;
5215
5216	down_write(&ls->ls_root_sem);
5217	list_for_each_entry(r, &ls->ls_root_list, res_root_list) {
5218		hold_rsb(r);
5219		lock_rsb(r);
5220		if (is_master(r)) {
5221			purge_dead_list(ls, r, &r->res_grantqueue,
5222					nodeid_gone, &lkb_count);
5223			purge_dead_list(ls, r, &r->res_convertqueue,
5224					nodeid_gone, &lkb_count);
5225			purge_dead_list(ls, r, &r->res_waitqueue,
5226					nodeid_gone, &lkb_count);
5227		}
5228		unlock_rsb(r);
5229		unhold_rsb(r);
5230		cond_resched();
5231	}
5232	up_write(&ls->ls_root_sem);
5233
5234	if (lkb_count)
5235		log_rinfo(ls, "dlm_recover_purge %u locks for %u nodes",
5236			  lkb_count, nodes_count);
5237}
5238
5239static struct dlm_rsb *find_grant_rsb(struct dlm_ls *ls, int bucket)
5240{
5241	struct rb_node *n;
5242	struct dlm_rsb *r;
5243
5244	spin_lock(&ls->ls_rsbtbl[bucket].lock);
5245	for (n = rb_first(&ls->ls_rsbtbl[bucket].keep); n; n = rb_next(n)) {
5246		r = rb_entry(n, struct dlm_rsb, res_hashnode);
5247
5248		if (!rsb_flag(r, RSB_RECOVER_GRANT))
5249			continue;
5250		if (!is_master(r)) {
5251			rsb_clear_flag(r, RSB_RECOVER_GRANT);
5252			continue;
5253		}
5254		hold_rsb(r);
5255		spin_unlock(&ls->ls_rsbtbl[bucket].lock);
5256		return r;
5257	}
5258	spin_unlock(&ls->ls_rsbtbl[bucket].lock);
5259	return NULL;
5260}
5261
5262/*
5263 * Attempt to grant locks on resources that we are the master of.
5264 * Locks may have become grantable during recovery because locks
5265 * from departed nodes have been purged (or not rebuilt), allowing
5266 * previously blocked locks to now be granted.  The subset of rsb's
5267 * we are interested in are those with lkb's on either the convert or
5268 * waiting queues.
5269 *
5270 * Simplest would be to go through each master rsb and check for non-empty
5271 * convert or waiting queues, and attempt to grant on those rsbs.
5272 * Checking the queues requires lock_rsb, though, for which we'd need
5273 * to release the rsbtbl lock.  This would make iterating through all
5274 * rsb's very inefficient.  So, we rely on earlier recovery routines
5275 * to set RECOVER_GRANT on any rsb's that we should attempt to grant
5276 * locks for.
5277 */
5278
5279void dlm_recover_grant(struct dlm_ls *ls)
5280{
5281	struct dlm_rsb *r;
5282	int bucket = 0;
5283	unsigned int count = 0;
5284	unsigned int rsb_count = 0;
5285	unsigned int lkb_count = 0;
5286
5287	while (1) {
5288		r = find_grant_rsb(ls, bucket);
5289		if (!r) {
5290			if (bucket == ls->ls_rsbtbl_size - 1)
5291				break;
5292			bucket++;
5293			continue;
5294		}
5295		rsb_count++;
5296		count = 0;
5297		lock_rsb(r);
5298		/* the RECOVER_GRANT flag is checked in the grant path */
5299		grant_pending_locks(r, &count);
5300		rsb_clear_flag(r, RSB_RECOVER_GRANT);
5301		lkb_count += count;
5302		confirm_master(r, 0);
5303		unlock_rsb(r);
5304		put_rsb(r);
5305		cond_resched();
5306	}
5307
5308	if (lkb_count)
5309		log_rinfo(ls, "dlm_recover_grant %u locks on %u resources",
5310			  lkb_count, rsb_count);
5311}
5312
5313static struct dlm_lkb *search_remid_list(struct list_head *head, int nodeid,
5314					 uint32_t remid)
5315{
5316	struct dlm_lkb *lkb;
5317
5318	list_for_each_entry(lkb, head, lkb_statequeue) {
5319		if (lkb->lkb_nodeid == nodeid && lkb->lkb_remid == remid)
5320			return lkb;
5321	}
5322	return NULL;
5323}
5324
5325static struct dlm_lkb *search_remid(struct dlm_rsb *r, int nodeid,
5326				    uint32_t remid)
5327{
5328	struct dlm_lkb *lkb;
5329
5330	lkb = search_remid_list(&r->res_grantqueue, nodeid, remid);
5331	if (lkb)
5332		return lkb;
5333	lkb = search_remid_list(&r->res_convertqueue, nodeid, remid);
5334	if (lkb)
5335		return lkb;
5336	lkb = search_remid_list(&r->res_waitqueue, nodeid, remid);
5337	if (lkb)
5338		return lkb;
5339	return NULL;
5340}
5341
5342/* needs at least dlm_rcom + rcom_lock */
5343static int receive_rcom_lock_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
5344				  struct dlm_rsb *r, const struct dlm_rcom *rc)
5345{
5346	struct rcom_lock *rl = (struct rcom_lock *) rc->rc_buf;
5347
5348	lkb->lkb_nodeid = le32_to_cpu(rc->rc_header.h_nodeid);
5349	lkb->lkb_ownpid = le32_to_cpu(rl->rl_ownpid);
5350	lkb->lkb_remid = le32_to_cpu(rl->rl_lkid);
5351	lkb->lkb_exflags = le32_to_cpu(rl->rl_exflags);
5352	dlm_set_dflags_val(lkb, le32_to_cpu(rl->rl_flags));
5353	set_bit(DLM_IFL_MSTCPY_BIT, &lkb->lkb_iflags);
5354	lkb->lkb_lvbseq = le32_to_cpu(rl->rl_lvbseq);
5355	lkb->lkb_rqmode = rl->rl_rqmode;
5356	lkb->lkb_grmode = rl->rl_grmode;
5357	/* don't set lkb_status because add_lkb wants to itself */
5358
5359	lkb->lkb_bastfn = (rl->rl_asts & DLM_CB_BAST) ? &fake_bastfn : NULL;
5360	lkb->lkb_astfn = (rl->rl_asts & DLM_CB_CAST) ? &fake_astfn : NULL;
5361
5362	if (lkb->lkb_exflags & DLM_LKF_VALBLK) {
5363		int lvblen = le16_to_cpu(rc->rc_header.h_length) -
5364			sizeof(struct dlm_rcom) - sizeof(struct rcom_lock);
5365		if (lvblen > ls->ls_lvblen)
5366			return -EINVAL;
5367		lkb->lkb_lvbptr = dlm_allocate_lvb(ls);
5368		if (!lkb->lkb_lvbptr)
5369			return -ENOMEM;
5370		memcpy(lkb->lkb_lvbptr, rl->rl_lvb, lvblen);
5371	}
5372
5373	/* Conversions between PR and CW (middle modes) need special handling.
5374	   The real granted mode of these converting locks cannot be determined
5375	   until all locks have been rebuilt on the rsb (recover_conversion) */
5376
5377	if (rl->rl_wait_type == cpu_to_le16(DLM_MSG_CONVERT) &&
5378	    middle_conversion(lkb)) {
5379		rl->rl_status = DLM_LKSTS_CONVERT;
5380		lkb->lkb_grmode = DLM_LOCK_IV;
5381		rsb_set_flag(r, RSB_RECOVER_CONVERT);
5382	}
5383
5384	return 0;
5385}
5386
5387/* This lkb may have been recovered in a previous aborted recovery so we need
5388   to check if the rsb already has an lkb with the given remote nodeid/lkid.
5389   If so we just send back a standard reply.  If not, we create a new lkb with
5390   the given values and send back our lkid.  We send back our lkid by sending
5391   back the rcom_lock struct we got but with the remid field filled in. */
5392
5393/* needs at least dlm_rcom + rcom_lock */
5394int dlm_recover_master_copy(struct dlm_ls *ls, const struct dlm_rcom *rc,
5395			    __le32 *rl_remid, __le32 *rl_result)
5396{
5397	struct rcom_lock *rl = (struct rcom_lock *) rc->rc_buf;
5398	struct dlm_rsb *r;
5399	struct dlm_lkb *lkb;
5400	uint32_t remid = 0;
5401	int from_nodeid = le32_to_cpu(rc->rc_header.h_nodeid);
5402	int error;
5403
5404	/* init rl_remid with rcom lock rl_remid */
5405	*rl_remid = rl->rl_remid;
5406
5407	if (rl->rl_parent_lkid) {
5408		error = -EOPNOTSUPP;
5409		goto out;
5410	}
5411
5412	remid = le32_to_cpu(rl->rl_lkid);
5413
5414	/* In general we expect the rsb returned to be R_MASTER, but we don't
5415	   have to require it.  Recovery of masters on one node can overlap
5416	   recovery of locks on another node, so one node can send us MSTCPY
5417	   locks before we've made ourselves master of this rsb.  We can still
5418	   add new MSTCPY locks that we receive here without any harm; when
5419	   we make ourselves master, dlm_recover_masters() won't touch the
5420	   MSTCPY locks we've received early. */
5421
5422	error = find_rsb(ls, rl->rl_name, le16_to_cpu(rl->rl_namelen),
5423			 from_nodeid, R_RECEIVE_RECOVER, &r);
5424	if (error)
5425		goto out;
5426
5427	lock_rsb(r);
5428
5429	if (dlm_no_directory(ls) && (dlm_dir_nodeid(r) != dlm_our_nodeid())) {
5430		log_error(ls, "dlm_recover_master_copy remote %d %x not dir",
5431			  from_nodeid, remid);
5432		error = -EBADR;
5433		goto out_unlock;
5434	}
5435
5436	lkb = search_remid(r, from_nodeid, remid);
5437	if (lkb) {
5438		error = -EEXIST;
5439		goto out_remid;
5440	}
5441
5442	error = create_lkb(ls, &lkb);
5443	if (error)
5444		goto out_unlock;
5445
5446	error = receive_rcom_lock_args(ls, lkb, r, rc);
5447	if (error) {
5448		__put_lkb(ls, lkb);
5449		goto out_unlock;
5450	}
5451
5452	attach_lkb(r, lkb);
5453	add_lkb(r, lkb, rl->rl_status);
 
5454	ls->ls_recover_locks_in++;
5455
5456	if (!list_empty(&r->res_waitqueue) || !list_empty(&r->res_convertqueue))
5457		rsb_set_flag(r, RSB_RECOVER_GRANT);
5458
5459 out_remid:
5460	/* this is the new value returned to the lock holder for
5461	   saving in its process-copy lkb */
5462	*rl_remid = cpu_to_le32(lkb->lkb_id);
5463
5464	lkb->lkb_recover_seq = ls->ls_recover_seq;
5465
5466 out_unlock:
5467	unlock_rsb(r);
5468	put_rsb(r);
5469 out:
5470	if (error && error != -EEXIST)
5471		log_rinfo(ls, "dlm_recover_master_copy remote %d %x error %d",
5472			  from_nodeid, remid, error);
5473	*rl_result = cpu_to_le32(error);
5474	return error;
5475}
5476
5477/* needs at least dlm_rcom + rcom_lock */
5478int dlm_recover_process_copy(struct dlm_ls *ls, const struct dlm_rcom *rc,
5479			     uint64_t seq)
5480{
5481	struct rcom_lock *rl = (struct rcom_lock *) rc->rc_buf;
5482	struct dlm_rsb *r;
5483	struct dlm_lkb *lkb;
5484	uint32_t lkid, remid;
5485	int error, result;
5486
5487	lkid = le32_to_cpu(rl->rl_lkid);
5488	remid = le32_to_cpu(rl->rl_remid);
5489	result = le32_to_cpu(rl->rl_result);
5490
5491	error = find_lkb(ls, lkid, &lkb);
5492	if (error) {
5493		log_error(ls, "dlm_recover_process_copy no %x remote %d %x %d",
5494			  lkid, le32_to_cpu(rc->rc_header.h_nodeid), remid,
5495			  result);
5496		return error;
5497	}
5498
5499	r = lkb->lkb_resource;
5500	hold_rsb(r);
5501	lock_rsb(r);
5502
5503	if (!is_process_copy(lkb)) {
5504		log_error(ls, "dlm_recover_process_copy bad %x remote %d %x %d",
5505			  lkid, le32_to_cpu(rc->rc_header.h_nodeid), remid,
5506			  result);
5507		dlm_dump_rsb(r);
5508		unlock_rsb(r);
5509		put_rsb(r);
5510		dlm_put_lkb(lkb);
5511		return -EINVAL;
5512	}
5513
5514	switch (result) {
5515	case -EBADR:
5516		/* There's a chance the new master received our lock before
5517		   dlm_recover_master_reply(), this wouldn't happen if we did
5518		   a barrier between recover_masters and recover_locks. */
5519
5520		log_debug(ls, "dlm_recover_process_copy %x remote %d %x %d",
5521			  lkid, le32_to_cpu(rc->rc_header.h_nodeid), remid,
5522			  result);
5523	
5524		dlm_send_rcom_lock(r, lkb, seq);
5525		goto out;
5526	case -EEXIST:
5527	case 0:
5528		lkb->lkb_remid = remid;
5529		break;
5530	default:
5531		log_error(ls, "dlm_recover_process_copy %x remote %d %x %d unk",
5532			  lkid, le32_to_cpu(rc->rc_header.h_nodeid), remid,
5533			  result);
5534	}
5535
5536	/* an ack for dlm_recover_locks() which waits for replies from
5537	   all the locks it sends to new masters */
5538	dlm_recovered_lock(r);
5539 out:
5540	unlock_rsb(r);
5541	put_rsb(r);
5542	dlm_put_lkb(lkb);
5543
5544	return 0;
5545}
5546
5547int dlm_user_request(struct dlm_ls *ls, struct dlm_user_args *ua,
5548		     int mode, uint32_t flags, void *name, unsigned int namelen)
 
5549{
5550	struct dlm_lkb *lkb;
5551	struct dlm_args args;
5552	bool do_put = true;
5553	int error;
5554
5555	dlm_lock_recovery(ls);
5556
5557	error = create_lkb(ls, &lkb);
5558	if (error) {
5559		kfree(ua);
5560		goto out;
5561	}
5562
5563	trace_dlm_lock_start(ls, lkb, name, namelen, mode, flags);
5564
5565	if (flags & DLM_LKF_VALBLK) {
5566		ua->lksb.sb_lvbptr = kzalloc(DLM_USER_LVB_LEN, GFP_NOFS);
5567		if (!ua->lksb.sb_lvbptr) {
5568			kfree(ua);
 
5569			error = -ENOMEM;
5570			goto out_put;
5571		}
5572	}
5573	error = set_lock_args(mode, &ua->lksb, flags, namelen, fake_astfn, ua,
5574			      fake_bastfn, &args);
5575	if (error) {
5576		kfree(ua->lksb.sb_lvbptr);
5577		ua->lksb.sb_lvbptr = NULL;
5578		kfree(ua);
5579		goto out_put;
 
5580	}
5581
5582	/* After ua is attached to lkb it will be freed by dlm_free_lkb().
5583	   When DLM_DFL_USER_BIT is set, the dlm knows that this is a userspace
5584	   lock and that lkb_astparam is the dlm_user_args structure. */
5585	set_bit(DLM_DFL_USER_BIT, &lkb->lkb_dflags);
5586	error = request_lock(ls, lkb, name, namelen, &args);
5587
5588	switch (error) {
5589	case 0:
5590		break;
5591	case -EINPROGRESS:
5592		error = 0;
5593		break;
5594	case -EAGAIN:
5595		error = 0;
5596		fallthrough;
5597	default:
5598		goto out_put;
 
5599	}
5600
5601	/* add this new lkb to the per-process list of locks */
5602	spin_lock(&ua->proc->locks_spin);
5603	hold_lkb(lkb);
5604	list_add_tail(&lkb->lkb_ownqueue, &ua->proc->locks);
5605	spin_unlock(&ua->proc->locks_spin);
5606	do_put = false;
5607 out_put:
5608	trace_dlm_lock_end(ls, lkb, name, namelen, mode, flags, error, false);
5609	if (do_put)
5610		__put_lkb(ls, lkb);
5611 out:
5612	dlm_unlock_recovery(ls);
5613	return error;
5614}
5615
5616int dlm_user_convert(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
5617		     int mode, uint32_t flags, uint32_t lkid, char *lvb_in)
 
5618{
5619	struct dlm_lkb *lkb;
5620	struct dlm_args args;
5621	struct dlm_user_args *ua;
5622	int error;
5623
5624	dlm_lock_recovery(ls);
5625
5626	error = find_lkb(ls, lkid, &lkb);
5627	if (error)
5628		goto out;
5629
5630	trace_dlm_lock_start(ls, lkb, NULL, 0, mode, flags);
5631
5632	/* user can change the params on its lock when it converts it, or
5633	   add an lvb that didn't exist before */
5634
5635	ua = lkb->lkb_ua;
5636
5637	if (flags & DLM_LKF_VALBLK && !ua->lksb.sb_lvbptr) {
5638		ua->lksb.sb_lvbptr = kzalloc(DLM_USER_LVB_LEN, GFP_NOFS);
5639		if (!ua->lksb.sb_lvbptr) {
5640			error = -ENOMEM;
5641			goto out_put;
5642		}
5643	}
5644	if (lvb_in && ua->lksb.sb_lvbptr)
5645		memcpy(ua->lksb.sb_lvbptr, lvb_in, DLM_USER_LVB_LEN);
5646
5647	ua->xid = ua_tmp->xid;
5648	ua->castparam = ua_tmp->castparam;
5649	ua->castaddr = ua_tmp->castaddr;
5650	ua->bastparam = ua_tmp->bastparam;
5651	ua->bastaddr = ua_tmp->bastaddr;
5652	ua->user_lksb = ua_tmp->user_lksb;
5653
5654	error = set_lock_args(mode, &ua->lksb, flags, 0, fake_astfn, ua,
5655			      fake_bastfn, &args);
5656	if (error)
5657		goto out_put;
5658
5659	error = convert_lock(ls, lkb, &args);
5660
5661	if (error == -EINPROGRESS || error == -EAGAIN || error == -EDEADLK)
5662		error = 0;
5663 out_put:
5664	trace_dlm_lock_end(ls, lkb, NULL, 0, mode, flags, error, false);
5665	dlm_put_lkb(lkb);
5666 out:
5667	dlm_unlock_recovery(ls);
5668	kfree(ua_tmp);
5669	return error;
5670}
5671
5672/*
5673 * The caller asks for an orphan lock on a given resource with a given mode.
5674 * If a matching lock exists, it's moved to the owner's list of locks and
5675 * the lkid is returned.
5676 */
5677
5678int dlm_user_adopt_orphan(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
5679		     int mode, uint32_t flags, void *name, unsigned int namelen,
5680		     uint32_t *lkid)
5681{
5682	struct dlm_lkb *lkb = NULL, *iter;
5683	struct dlm_user_args *ua;
5684	int found_other_mode = 0;
 
5685	int rv = 0;
5686
5687	mutex_lock(&ls->ls_orphans_mutex);
5688	list_for_each_entry(iter, &ls->ls_orphans, lkb_ownqueue) {
5689		if (iter->lkb_resource->res_length != namelen)
5690			continue;
5691		if (memcmp(iter->lkb_resource->res_name, name, namelen))
5692			continue;
5693		if (iter->lkb_grmode != mode) {
5694			found_other_mode = 1;
5695			continue;
5696		}
5697
5698		lkb = iter;
5699		list_del_init(&iter->lkb_ownqueue);
5700		clear_bit(DLM_DFL_ORPHAN_BIT, &iter->lkb_dflags);
5701		*lkid = iter->lkb_id;
5702		break;
5703	}
5704	mutex_unlock(&ls->ls_orphans_mutex);
5705
5706	if (!lkb && found_other_mode) {
5707		rv = -EAGAIN;
5708		goto out;
5709	}
5710
5711	if (!lkb) {
5712		rv = -ENOENT;
5713		goto out;
5714	}
5715
5716	lkb->lkb_exflags = flags;
5717	lkb->lkb_ownpid = (int) current->pid;
5718
5719	ua = lkb->lkb_ua;
5720
5721	ua->proc = ua_tmp->proc;
5722	ua->xid = ua_tmp->xid;
5723	ua->castparam = ua_tmp->castparam;
5724	ua->castaddr = ua_tmp->castaddr;
5725	ua->bastparam = ua_tmp->bastparam;
5726	ua->bastaddr = ua_tmp->bastaddr;
5727	ua->user_lksb = ua_tmp->user_lksb;
5728
5729	/*
5730	 * The lkb reference from the ls_orphans list was not
5731	 * removed above, and is now considered the reference
5732	 * for the proc locks list.
5733	 */
5734
5735	spin_lock(&ua->proc->locks_spin);
5736	list_add_tail(&lkb->lkb_ownqueue, &ua->proc->locks);
5737	spin_unlock(&ua->proc->locks_spin);
5738 out:
5739	kfree(ua_tmp);
5740	return rv;
5741}
5742
5743int dlm_user_unlock(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
5744		    uint32_t flags, uint32_t lkid, char *lvb_in)
5745{
5746	struct dlm_lkb *lkb;
5747	struct dlm_args args;
5748	struct dlm_user_args *ua;
5749	int error;
5750
5751	dlm_lock_recovery(ls);
5752
5753	error = find_lkb(ls, lkid, &lkb);
5754	if (error)
5755		goto out;
5756
5757	trace_dlm_unlock_start(ls, lkb, flags);
5758
5759	ua = lkb->lkb_ua;
5760
5761	if (lvb_in && ua->lksb.sb_lvbptr)
5762		memcpy(ua->lksb.sb_lvbptr, lvb_in, DLM_USER_LVB_LEN);
5763	if (ua_tmp->castparam)
5764		ua->castparam = ua_tmp->castparam;
5765	ua->user_lksb = ua_tmp->user_lksb;
5766
5767	error = set_unlock_args(flags, ua, &args);
5768	if (error)
5769		goto out_put;
5770
5771	error = unlock_lock(ls, lkb, &args);
5772
5773	if (error == -DLM_EUNLOCK)
5774		error = 0;
5775	/* from validate_unlock_args() */
5776	if (error == -EBUSY && (flags & DLM_LKF_FORCEUNLOCK))
5777		error = 0;
5778	if (error)
5779		goto out_put;
5780
5781	spin_lock(&ua->proc->locks_spin);
5782	/* dlm_user_add_cb() may have already taken lkb off the proc list */
5783	if (!list_empty(&lkb->lkb_ownqueue))
5784		list_move(&lkb->lkb_ownqueue, &ua->proc->unlocking);
5785	spin_unlock(&ua->proc->locks_spin);
5786 out_put:
5787	trace_dlm_unlock_end(ls, lkb, flags, error);
5788	dlm_put_lkb(lkb);
5789 out:
5790	dlm_unlock_recovery(ls);
5791	kfree(ua_tmp);
5792	return error;
5793}
5794
5795int dlm_user_cancel(struct dlm_ls *ls, struct dlm_user_args *ua_tmp,
5796		    uint32_t flags, uint32_t lkid)
5797{
5798	struct dlm_lkb *lkb;
5799	struct dlm_args args;
5800	struct dlm_user_args *ua;
5801	int error;
5802
5803	dlm_lock_recovery(ls);
5804
5805	error = find_lkb(ls, lkid, &lkb);
5806	if (error)
5807		goto out;
5808
5809	trace_dlm_unlock_start(ls, lkb, flags);
5810
5811	ua = lkb->lkb_ua;
5812	if (ua_tmp->castparam)
5813		ua->castparam = ua_tmp->castparam;
5814	ua->user_lksb = ua_tmp->user_lksb;
5815
5816	error = set_unlock_args(flags, ua, &args);
5817	if (error)
5818		goto out_put;
5819
5820	error = cancel_lock(ls, lkb, &args);
5821
5822	if (error == -DLM_ECANCEL)
5823		error = 0;
5824	/* from validate_unlock_args() */
5825	if (error == -EBUSY)
5826		error = 0;
5827 out_put:
5828	trace_dlm_unlock_end(ls, lkb, flags, error);
5829	dlm_put_lkb(lkb);
5830 out:
5831	dlm_unlock_recovery(ls);
5832	kfree(ua_tmp);
5833	return error;
5834}
5835
5836int dlm_user_deadlock(struct dlm_ls *ls, uint32_t flags, uint32_t lkid)
5837{
5838	struct dlm_lkb *lkb;
5839	struct dlm_args args;
5840	struct dlm_user_args *ua;
5841	struct dlm_rsb *r;
5842	int error;
5843
5844	dlm_lock_recovery(ls);
5845
5846	error = find_lkb(ls, lkid, &lkb);
5847	if (error)
5848		goto out;
5849
5850	trace_dlm_unlock_start(ls, lkb, flags);
5851
5852	ua = lkb->lkb_ua;
5853
5854	error = set_unlock_args(flags, ua, &args);
5855	if (error)
5856		goto out_put;
5857
5858	/* same as cancel_lock(), but set DEADLOCK_CANCEL after lock_rsb */
5859
5860	r = lkb->lkb_resource;
5861	hold_rsb(r);
5862	lock_rsb(r);
5863
5864	error = validate_unlock_args(lkb, &args);
5865	if (error)
5866		goto out_r;
5867	set_bit(DLM_IFL_DEADLOCK_CANCEL_BIT, &lkb->lkb_iflags);
5868
5869	error = _cancel_lock(r, lkb);
5870 out_r:
5871	unlock_rsb(r);
5872	put_rsb(r);
5873
5874	if (error == -DLM_ECANCEL)
5875		error = 0;
5876	/* from validate_unlock_args() */
5877	if (error == -EBUSY)
5878		error = 0;
5879 out_put:
5880	trace_dlm_unlock_end(ls, lkb, flags, error);
5881	dlm_put_lkb(lkb);
5882 out:
5883	dlm_unlock_recovery(ls);
5884	return error;
5885}
5886
5887/* lkb's that are removed from the waiters list by revert are just left on the
5888   orphans list with the granted orphan locks, to be freed by purge */
5889
5890static int orphan_proc_lock(struct dlm_ls *ls, struct dlm_lkb *lkb)
5891{
5892	struct dlm_args args;
5893	int error;
5894
5895	hold_lkb(lkb); /* reference for the ls_orphans list */
5896	mutex_lock(&ls->ls_orphans_mutex);
5897	list_add_tail(&lkb->lkb_ownqueue, &ls->ls_orphans);
5898	mutex_unlock(&ls->ls_orphans_mutex);
5899
5900	set_unlock_args(0, lkb->lkb_ua, &args);
5901
5902	error = cancel_lock(ls, lkb, &args);
5903	if (error == -DLM_ECANCEL)
5904		error = 0;
5905	return error;
5906}
5907
5908/* The FORCEUNLOCK flag allows the unlock to go ahead even if the lkb isn't
5909   granted.  Regardless of what rsb queue the lock is on, it's removed and
5910   freed.  The IVVALBLK flag causes the lvb on the resource to be invalidated
5911   if our lock is PW/EX (it's ignored if our granted mode is smaller.) */
5912
5913static int unlock_proc_lock(struct dlm_ls *ls, struct dlm_lkb *lkb)
5914{
5915	struct dlm_args args;
5916	int error;
5917
5918	set_unlock_args(DLM_LKF_FORCEUNLOCK | DLM_LKF_IVVALBLK,
5919			lkb->lkb_ua, &args);
5920
5921	error = unlock_lock(ls, lkb, &args);
5922	if (error == -DLM_EUNLOCK)
5923		error = 0;
5924	return error;
5925}
5926
5927/* We have to release clear_proc_locks mutex before calling unlock_proc_lock()
5928   (which does lock_rsb) due to deadlock with receiving a message that does
5929   lock_rsb followed by dlm_user_add_cb() */
5930
5931static struct dlm_lkb *del_proc_lock(struct dlm_ls *ls,
5932				     struct dlm_user_proc *proc)
5933{
5934	struct dlm_lkb *lkb = NULL;
5935
5936	spin_lock(&ls->ls_clear_proc_locks);
5937	if (list_empty(&proc->locks))
5938		goto out;
5939
5940	lkb = list_entry(proc->locks.next, struct dlm_lkb, lkb_ownqueue);
5941	list_del_init(&lkb->lkb_ownqueue);
5942
5943	if (lkb->lkb_exflags & DLM_LKF_PERSISTENT)
5944		set_bit(DLM_DFL_ORPHAN_BIT, &lkb->lkb_dflags);
5945	else
5946		set_bit(DLM_IFL_DEAD_BIT, &lkb->lkb_iflags);
5947 out:
5948	spin_unlock(&ls->ls_clear_proc_locks);
5949	return lkb;
5950}
5951
5952/* The ls_clear_proc_locks mutex protects against dlm_user_add_cb() which
5953   1) references lkb->ua which we free here and 2) adds lkbs to proc->asts,
5954   which we clear here. */
5955
5956/* proc CLOSING flag is set so no more device_reads should look at proc->asts
5957   list, and no more device_writes should add lkb's to proc->locks list; so we
5958   shouldn't need to take asts_spin or locks_spin here.  this assumes that
5959   device reads/writes/closes are serialized -- FIXME: we may need to serialize
5960   them ourself. */
5961
5962void dlm_clear_proc_locks(struct dlm_ls *ls, struct dlm_user_proc *proc)
5963{
5964	struct dlm_lkb *lkb, *safe;
5965
5966	dlm_lock_recovery(ls);
5967
5968	while (1) {
5969		lkb = del_proc_lock(ls, proc);
5970		if (!lkb)
5971			break;
 
5972		if (lkb->lkb_exflags & DLM_LKF_PERSISTENT)
5973			orphan_proc_lock(ls, lkb);
5974		else
5975			unlock_proc_lock(ls, lkb);
5976
5977		/* this removes the reference for the proc->locks list
5978		   added by dlm_user_request, it may result in the lkb
5979		   being freed */
5980
5981		dlm_put_lkb(lkb);
5982	}
5983
5984	spin_lock(&ls->ls_clear_proc_locks);
5985
5986	/* in-progress unlocks */
5987	list_for_each_entry_safe(lkb, safe, &proc->unlocking, lkb_ownqueue) {
5988		list_del_init(&lkb->lkb_ownqueue);
5989		set_bit(DLM_IFL_DEAD_BIT, &lkb->lkb_iflags);
5990		dlm_put_lkb(lkb);
5991	}
5992
5993	list_for_each_entry_safe(lkb, safe, &proc->asts, lkb_cb_list) {
5994		dlm_purge_lkb_callbacks(lkb);
 
5995		list_del_init(&lkb->lkb_cb_list);
5996		dlm_put_lkb(lkb);
5997	}
5998
5999	spin_unlock(&ls->ls_clear_proc_locks);
6000	dlm_unlock_recovery(ls);
6001}
6002
6003static void purge_proc_locks(struct dlm_ls *ls, struct dlm_user_proc *proc)
6004{
6005	struct dlm_lkb *lkb, *safe;
6006
6007	while (1) {
6008		lkb = NULL;
6009		spin_lock(&proc->locks_spin);
6010		if (!list_empty(&proc->locks)) {
6011			lkb = list_entry(proc->locks.next, struct dlm_lkb,
6012					 lkb_ownqueue);
6013			list_del_init(&lkb->lkb_ownqueue);
6014		}
6015		spin_unlock(&proc->locks_spin);
6016
6017		if (!lkb)
6018			break;
6019
6020		set_bit(DLM_IFL_DEAD_BIT, &lkb->lkb_iflags);
6021		unlock_proc_lock(ls, lkb);
6022		dlm_put_lkb(lkb); /* ref from proc->locks list */
6023	}
6024
6025	spin_lock(&proc->locks_spin);
6026	list_for_each_entry_safe(lkb, safe, &proc->unlocking, lkb_ownqueue) {
6027		list_del_init(&lkb->lkb_ownqueue);
6028		set_bit(DLM_IFL_DEAD_BIT, &lkb->lkb_iflags);
6029		dlm_put_lkb(lkb);
6030	}
6031	spin_unlock(&proc->locks_spin);
6032
6033	spin_lock(&proc->asts_spin);
6034	list_for_each_entry_safe(lkb, safe, &proc->asts, lkb_cb_list) {
6035		dlm_purge_lkb_callbacks(lkb);
 
6036		list_del_init(&lkb->lkb_cb_list);
6037		dlm_put_lkb(lkb);
6038	}
6039	spin_unlock(&proc->asts_spin);
6040}
6041
6042/* pid of 0 means purge all orphans */
6043
6044static void do_purge(struct dlm_ls *ls, int nodeid, int pid)
6045{
6046	struct dlm_lkb *lkb, *safe;
6047
6048	mutex_lock(&ls->ls_orphans_mutex);
6049	list_for_each_entry_safe(lkb, safe, &ls->ls_orphans, lkb_ownqueue) {
6050		if (pid && lkb->lkb_ownpid != pid)
6051			continue;
6052		unlock_proc_lock(ls, lkb);
6053		list_del_init(&lkb->lkb_ownqueue);
6054		dlm_put_lkb(lkb);
6055	}
6056	mutex_unlock(&ls->ls_orphans_mutex);
6057}
6058
6059static int send_purge(struct dlm_ls *ls, int nodeid, int pid)
6060{
6061	struct dlm_message *ms;
6062	struct dlm_mhandle *mh;
6063	int error;
6064
6065	error = _create_message(ls, sizeof(struct dlm_message), nodeid,
6066				DLM_MSG_PURGE, &ms, &mh, GFP_NOFS);
6067	if (error)
6068		return error;
6069	ms->m_nodeid = cpu_to_le32(nodeid);
6070	ms->m_pid = cpu_to_le32(pid);
6071
6072	return send_message(mh, ms, NULL, 0);
6073}
6074
6075int dlm_user_purge(struct dlm_ls *ls, struct dlm_user_proc *proc,
6076		   int nodeid, int pid)
6077{
6078	int error = 0;
6079
6080	if (nodeid && (nodeid != dlm_our_nodeid())) {
6081		error = send_purge(ls, nodeid, pid);
6082	} else {
6083		dlm_lock_recovery(ls);
6084		if (pid == current->pid)
6085			purge_proc_locks(ls, proc);
6086		else
6087			do_purge(ls, nodeid, pid);
6088		dlm_unlock_recovery(ls);
6089	}
6090	return error;
6091}
6092
6093/* debug functionality */
6094int dlm_debug_add_lkb(struct dlm_ls *ls, uint32_t lkb_id, char *name, int len,
6095		      int lkb_nodeid, unsigned int lkb_dflags, int lkb_status)
6096{
6097	struct dlm_lksb *lksb;
6098	struct dlm_lkb *lkb;
6099	struct dlm_rsb *r;
6100	int error;
6101
6102	/* we currently can't set a valid user lock */
6103	if (lkb_dflags & BIT(DLM_DFL_USER_BIT))
6104		return -EOPNOTSUPP;
6105
6106	lksb = kzalloc(sizeof(*lksb), GFP_NOFS);
6107	if (!lksb)
6108		return -ENOMEM;
6109
6110	error = _create_lkb(ls, &lkb, lkb_id, lkb_id + 1);
6111	if (error) {
6112		kfree(lksb);
6113		return error;
6114	}
6115
6116	dlm_set_dflags_val(lkb, lkb_dflags);
6117	lkb->lkb_nodeid = lkb_nodeid;
6118	lkb->lkb_lksb = lksb;
6119	/* user specific pointer, just don't have it NULL for kernel locks */
6120	if (~lkb_dflags & BIT(DLM_DFL_USER_BIT))
6121		lkb->lkb_astparam = (void *)0xDEADBEEF;
6122
6123	error = find_rsb(ls, name, len, 0, R_REQUEST, &r);
6124	if (error) {
6125		kfree(lksb);
6126		__put_lkb(ls, lkb);
6127		return error;
6128	}
6129
6130	lock_rsb(r);
6131	attach_lkb(r, lkb);
6132	add_lkb(r, lkb, lkb_status);
6133	unlock_rsb(r);
6134	put_rsb(r);
6135
6136	return 0;
6137}
6138
6139int dlm_debug_add_lkb_to_waiters(struct dlm_ls *ls, uint32_t lkb_id,
6140				 int mstype, int to_nodeid)
6141{
6142	struct dlm_lkb *lkb;
6143	int error;
6144
6145	error = find_lkb(ls, lkb_id, &lkb);
6146	if (error)
6147		return error;
6148
6149	error = add_to_waiters(lkb, mstype, to_nodeid);
6150	dlm_put_lkb(lkb);
6151	return error;
6152}
6153