fs/lockd/svclock.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * linux/fs/lockd/svclock.c
   4  *
   5  * Handling of server-side locks, mostly of the blocked variety.
   6  * This is the ugliest part of lockd because we tread on very thin ice.
   7  * GRANT and CANCEL calls may get stuck, meet in mid-flight, etc.
   8  * IMNSHO introducing the grant callback into the NLM protocol was one
   9  * of the worst ideas Sun ever had. Except maybe for the idea of doing
  10  * NFS file locking at all.
  11  *
  12  * I'm trying hard to avoid race conditions by protecting most accesses
  13  * to a file's list of blocked locks through a semaphore. The global
  14  * list of blocked locks is not protected in this fashion however.
  15  * Therefore, some functions (such as the RPC callback for the async grant
  16  * call) move blocked locks towards the head of the list *while some other
  17  * process might be traversing it*. This should not be a problem in
  18  * practice, because this will only cause functions traversing the list
  19  * to visit some blocks twice.
  20  *
  21  * Copyright (C) 1996, Olaf Kirch <okir@monad.swb.de>
  22  */
  23
  24 #include <linux/types.h>
  25 #include <linux/slab.h>
  26 #include <linux/errno.h>
  27 #include <linux/kernel.h>
  28 #include <linux/sched.h>
  29 #include <linux/sunrpc/clnt.h>
  30 #include <linux/sunrpc/svc_xprt.h>
  31 #include <linux/lockd/nlm.h>
  32 #include <linux/lockd/lockd.h>
  33 #include <linux/kthread.h>
  34
  35 #define NLMDBG_FACILITY         NLMDBG_SVCLOCK
  36
  37 #ifdef CONFIG_LOCKD_V4
  38 #define nlm_deadlock    nlm4_deadlock
  39 #else
  40 #define nlm_deadlock    nlm_lck_denied
  41 #endif
  42
  43 static void nlmsvc_release_block(struct nlm_block *block);
  44 static void     nlmsvc_insert_block(struct nlm_block *block, unsigned long);
  45 static void     nlmsvc_remove_block(struct nlm_block *block);
  46
  47 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock);
  48 static void nlmsvc_freegrantargs(struct nlm_rqst *call);
  49 static const struct rpc_call_ops nlmsvc_grant_ops;
  50
  51 /*
  52  * The list of blocked locks to retry
  53  */
  54 static LIST_HEAD(nlm_blocked);
  55 static DEFINE_SPINLOCK(nlm_blocked_lock);
  56
  57 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  58 static const char *nlmdbg_cookie2a(const struct nlm_cookie *cookie)
  59 {
  60         /*
  61          * We can get away with a static buffer because this is only called
  62          * from lockd, which is single-threaded.
  63          */
  64         static char buf[2*NLM_MAXCOOKIELEN+1];
  65         unsigned int i, len = sizeof(buf);
  66         char *p = buf;
  67
  68         len--;  /* allow for trailing \0 */
  69         if (len < 3)
  70                 return "???";
  71         for (i = 0 ; i < cookie->len ; i++) {
  72                 if (len < 2) {
  73                         strcpy(p-3, "...");
  74                         break;
  75                 }
  76                 sprintf(p, "%02x", cookie->data[i]);
  77                 p += 2;
  78                 len -= 2;
  79         }
  80         *p = '\0';
  81
  82         return buf;
  83 }
  84 #endif
  85
  86 /*
  87  * Insert a blocked lock into the global list
  88  */
  89 static void
  90 nlmsvc_insert_block_locked(struct nlm_block *block, unsigned long when)
  91 {
  92         struct nlm_block *b;
  93         struct list_head *pos;
  94
  95         dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block, when);
  96         if (list_empty(&block->b_list)) {
  97                 kref_get(&block->b_count);
  98         } else {
  99                 list_del_init(&block->b_list);
 100         }
 101
 102         pos = &nlm_blocked;
 103         if (when != NLM_NEVER) {
 104                 if ((when += jiffies) == NLM_NEVER)
 105                         when ++;
 106                 list_for_each(pos, &nlm_blocked) {
 107                         b = list_entry(pos, struct nlm_block, b_list);
 108                         if (time_after(b->b_when,when) || b->b_when == NLM_NEVER)
 109                                 break;
 110                 }
 111                 /* On normal exit from the loop, pos == &nlm_blocked,
 112                  * so we will be adding to the end of the list - good
 113                  */
 114         }
 115
 116         list_add_tail(&block->b_list, pos);
 117         block->b_when = when;
 118 }
 119
 120 static void nlmsvc_insert_block(struct nlm_block *block, unsigned long when)
 121 {
 122         spin_lock(&nlm_blocked_lock);
 123         nlmsvc_insert_block_locked(block, when);
 124         spin_unlock(&nlm_blocked_lock);
 125 }
 126
 127 /*
 128  * Remove a block from the global list
 129  */
 130 static inline void
 131 nlmsvc_remove_block(struct nlm_block *block)
 132 {
 133         if (!list_empty(&block->b_list)) {
 134                 spin_lock(&nlm_blocked_lock);
 135                 list_del_init(&block->b_list);
 136                 spin_unlock(&nlm_blocked_lock);
 137                 nlmsvc_release_block(block);
 138         }
 139 }
 140
 141 /*
 142  * Find a block for a given lock
 143  */
 144 static struct nlm_block *
 145 nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock)
 146 {
 147         struct nlm_block        *block;
 148         struct file_lock        *fl;
 149
 150         dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
 151                                 file, lock->fl.fl_pid,
 152                                 (long long)lock->fl.fl_start,
 153                                 (long long)lock->fl.fl_end, lock->fl.fl_type);
 154         list_for_each_entry(block, &nlm_blocked, b_list) {
 155                 fl = &block->b_call->a_args.lock.fl;
 156                 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
 157                                 block->b_file, fl->fl_pid,
 158                                 (long long)fl->fl_start,
 159                                 (long long)fl->fl_end, fl->fl_type,
 160                                 nlmdbg_cookie2a(&block->b_call->a_args.cookie));
 161                 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
 162                         kref_get(&block->b_count);
 163                         return block;
 164                 }
 165         }
 166
 167         return NULL;
 168 }
 169
 170 static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
 171 {
 172         if (a->len != b->len)
 173                 return 0;
 174         if (memcmp(a->data, b->data, a->len))
 175                 return 0;
 176         return 1;
 177 }
 178
 179 /*
 180  * Find a block with a given NLM cookie.
 181  */
 182 static inline struct nlm_block *
 183 nlmsvc_find_block(struct nlm_cookie *cookie)
 184 {
 185         struct nlm_block *block;
 186
 187         list_for_each_entry(block, &nlm_blocked, b_list) {
 188                 if (nlm_cookie_match(&block->b_call->a_args.cookie,cookie))
 189                         goto found;
 190         }
 191
 192         return NULL;
 193
 194 found:
 195         dprintk("nlmsvc_find_block(%s): block=%p\n", nlmdbg_cookie2a(cookie), block);
 196         kref_get(&block->b_count);
 197         return block;
 198 }
 199
 200 /*
 201  * Create a block and initialize it.
 202  *
 203  * Note: we explicitly set the cookie of the grant reply to that of
 204  * the blocked lock request. The spec explicitly mentions that the client
 205  * should _not_ rely on the callback containing the same cookie as the
 206  * request, but (as I found out later) that's because some implementations
 207  * do just this. Never mind the standards comittees, they support our
 208  * logging industries.
 209  *
 210  * 10 years later: I hope we can safely ignore these old and broken
 211  * clients by now. Let's fix this so we can uniquely identify an incoming
 212  * GRANTED_RES message by cookie, without having to rely on the client's IP
 213  * address. --okir
 214  */
 215 static struct nlm_block *
 216 nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_host *host,
 217                     struct nlm_file *file, struct nlm_lock *lock,
 218                     struct nlm_cookie *cookie)
 219 {
 220         struct nlm_block        *block;
 221         struct nlm_rqst         *call = NULL;
 222
 223         call = nlm_alloc_call(host);
 224         if (call == NULL)
 225                 return NULL;
 226
 227         /* Allocate memory for block, and initialize arguments */
 228         block = kzalloc(sizeof(*block), GFP_KERNEL);
 229         if (block == NULL)
 230                 goto failed;
 231         kref_init(&block->b_count);
 232         INIT_LIST_HEAD(&block->b_list);
 233         INIT_LIST_HEAD(&block->b_flist);
 234
 235         if (!nlmsvc_setgrantargs(call, lock))
 236                 goto failed_free;
 237
 238         /* Set notifier function for VFS, and init args */
 239         call->a_args.lock.fl.fl_flags |= FL_SLEEP;
 240         call->a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
 241         nlmclnt_next_cookie(&call->a_args.cookie);
 242
 243         dprintk("lockd: created block %p...\n", block);
 244
 245         /* Create and initialize the block */
 246         block->b_daemon = rqstp->rq_server;
 247         block->b_host   = host;
 248         block->b_file   = file;
 249         file->f_count++;
 250
 251         /* Add to file's list of blocks */
 252         list_add(&block->b_flist, &file->f_blocks);
 253
 254         /* Set up RPC arguments for callback */
 255         block->b_call = call;
 256         call->a_flags   = RPC_TASK_ASYNC;
 257         call->a_block = block;
 258
 259         return block;
 260
 261 failed_free:
 262         kfree(block);
 263 failed:
 264         nlmsvc_release_call(call);
 265         return NULL;
 266 }
 267
 268 /*
 269  * Delete a block.
 270  * It is the caller's responsibility to check whether the file
 271  * can be closed hereafter.
 272  */
 273 static int nlmsvc_unlink_block(struct nlm_block *block)
 274 {
 275         int status;
 276         dprintk("lockd: unlinking block %p...\n", block);
 277
 278         /* Remove block from list */
 279         status = locks_delete_block(&block->b_call->a_args.lock.fl);
 280         nlmsvc_remove_block(block);
 281         return status;
 282 }
 283
 284 static void nlmsvc_free_block(struct kref *kref)
 285 {
 286         struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
 287         struct nlm_file         *file = block->b_file;
 288
 289         dprintk("lockd: freeing block %p...\n", block);
 290
 291         /* Remove block from file's list of blocks */
 292         list_del_init(&block->b_flist);
 293         mutex_unlock(&file->f_mutex);
 294
 295         nlmsvc_freegrantargs(block->b_call);
 296         nlmsvc_release_call(block->b_call);
 297         nlm_release_file(block->b_file);
 298         kfree(block);
 299 }
 300
 301 static void nlmsvc_release_block(struct nlm_block *block)
 302 {
 303         if (block != NULL)
 304                 kref_put_mutex(&block->b_count, nlmsvc_free_block, &block->b_file->f_mutex);
 305 }
 306
 307 /*
 308  * Loop over all blocks and delete blocks held by
 309  * a matching host.
 310  */
 311 void nlmsvc_traverse_blocks(struct nlm_host *host,
 312                         struct nlm_file *file,
 313                         nlm_host_match_fn_t match)
 314 {
 315         struct nlm_block *block, *next;
 316
 317 restart:
 318         mutex_lock(&file->f_mutex);
 319         list_for_each_entry_safe(block, next, &file->f_blocks, b_flist) {
 320                 if (!match(block->b_host, host))
 321                         continue;
 322                 /* Do not destroy blocks that are not on
 323                  * the global retry list - why? */
 324                 if (list_empty(&block->b_list))
 325                         continue;
 326                 kref_get(&block->b_count);
 327                 mutex_unlock(&file->f_mutex);
 328                 nlmsvc_unlink_block(block);
 329                 nlmsvc_release_block(block);
 330                 goto restart;
 331         }
 332         mutex_unlock(&file->f_mutex);
 333 }
 334
 335 static struct nlm_lockowner *
 336 nlmsvc_get_lockowner(struct nlm_lockowner *lockowner)
 337 {
 338         refcount_inc(&lockowner->count);
 339         return lockowner;
 340 }
 341
 342 static void nlmsvc_put_lockowner(struct nlm_lockowner *lockowner)
 343 {
 344         if (!refcount_dec_and_lock(&lockowner->count, &lockowner->host->h_lock))
 345                 return;
 346         list_del(&lockowner->list);
 347         spin_unlock(&lockowner->host->h_lock);
 348         nlmsvc_release_host(lockowner->host);
 349         kfree(lockowner);
 350 }
 351
 352 static struct nlm_lockowner *__nlmsvc_find_lockowner(struct nlm_host *host, pid_t pid)
 353 {
 354         struct nlm_lockowner *lockowner;
 355         list_for_each_entry(lockowner, &host->h_lockowners, list) {
 356                 if (lockowner->pid != pid)
 357                         continue;
 358                 return nlmsvc_get_lockowner(lockowner);
 359         }
 360         return NULL;
 361 }
 362
 363 static struct nlm_lockowner *nlmsvc_find_lockowner(struct nlm_host *host, pid_t pid)
 364 {
 365         struct nlm_lockowner *res, *new = NULL;
 366
 367         spin_lock(&host->h_lock);
 368         res = __nlmsvc_find_lockowner(host, pid);
 369
 370         if (res == NULL) {
 371                 spin_unlock(&host->h_lock);
 372                 new = kmalloc(sizeof(*res), GFP_KERNEL);
 373                 spin_lock(&host->h_lock);
 374                 res = __nlmsvc_find_lockowner(host, pid);
 375                 if (res == NULL && new != NULL) {
 376                         res = new;
 377                         /* fs/locks.c will manage the refcount through lock_ops */
 378                         refcount_set(&new->count, 1);
 379                         new->pid = pid;
 380                         new->host = nlm_get_host(host);
 381                         list_add(&new->list, &host->h_lockowners);
 382                         new = NULL;
 383                 }
 384         }
 385
 386         spin_unlock(&host->h_lock);
 387         kfree(new);
 388         return res;
 389 }
 390
 391 void
 392 nlmsvc_release_lockowner(struct nlm_lock *lock)
 393 {
 394         if (lock->fl.fl_owner)
 395                 nlmsvc_put_lockowner(lock->fl.fl_owner);
 396 }
 397
 398 void nlmsvc_locks_init_private(struct file_lock *fl, struct nlm_host *host,
 399                                                 pid_t pid)
 400 {
 401         fl->fl_owner = nlmsvc_find_lockowner(host, pid);
 402 }
 403
 404 /*
 405  * Initialize arguments for GRANTED call. The nlm_rqst structure
 406  * has been cleared already.
 407  */
 408 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
 409 {
 410         locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
 411         memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
 412         call->a_args.lock.caller = utsname()->nodename;
 413         call->a_args.lock.oh.len = lock->oh.len;
 414
 415         /* set default data area */
 416         call->a_args.lock.oh.data = call->a_owner;
 417         call->a_args.lock.svid = ((struct nlm_lockowner *)lock->fl.fl_owner)->pid;
 418
 419         if (lock->oh.len > NLMCLNT_OHSIZE) {
 420                 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
 421                 if (!data)
 422                         return 0;
 423                 call->a_args.lock.oh.data = (u8 *) data;
 424         }
 425
 426         memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
 427         return 1;
 428 }
 429
 430 static void nlmsvc_freegrantargs(struct nlm_rqst *call)
 431 {
 432         if (call->a_args.lock.oh.data != call->a_owner)
 433                 kfree(call->a_args.lock.oh.data);
 434
 435         locks_release_private(&call->a_args.lock.fl);
 436 }
 437
 438 /*
 439  * Deferred lock request handling for non-blocking lock
 440  */
 441 static __be32
 442 nlmsvc_defer_lock_rqst(struct svc_rqst *rqstp, struct nlm_block *block)
 443 {
 444         __be32 status = nlm_lck_denied_nolocks;
 445
 446         block->b_flags |= B_QUEUED;
 447
 448         nlmsvc_insert_block(block, NLM_TIMEOUT);
 449
 450         block->b_cache_req = &rqstp->rq_chandle;
 451         if (rqstp->rq_chandle.defer) {
 452                 block->b_deferred_req =
 453                         rqstp->rq_chandle.defer(block->b_cache_req);
 454                 if (block->b_deferred_req != NULL)
 455                         status = nlm_drop_reply;
 456         }
 457         dprintk("lockd: nlmsvc_defer_lock_rqst block %p flags %d status %d\n",
 458                 block, block->b_flags, ntohl(status));
 459
 460         return status;
 461 }
 462
 463 /*
 464  * Attempt to establish a lock, and if it can't be granted, block it
 465  * if required.
 466  */
 467 __be32
 468 nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
 469             struct nlm_host *host, struct nlm_lock *lock, int wait,
 470             struct nlm_cookie *cookie, int reclaim)
 471 {
 472         struct nlm_block        *block = NULL;
 473         int                     error;
 474         __be32                  ret;
 475
 476         dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
 477                                 nlmsvc_file_inode(file)->i_sb->s_id,
 478                                 nlmsvc_file_inode(file)->i_ino,
 479                                 lock->fl.fl_type, lock->fl.fl_pid,
 480                                 (long long)lock->fl.fl_start,
 481                                 (long long)lock->fl.fl_end,
 482                                 wait);
 483
 484         /* Lock file against concurrent access */
 485         mutex_lock(&file->f_mutex);
 486         /* Get existing block (in case client is busy-waiting)
 487          * or create new block
 488          */
 489         block = nlmsvc_lookup_block(file, lock);
 490         if (block == NULL) {
 491                 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
 492                 ret = nlm_lck_denied_nolocks;
 493                 if (block == NULL)
 494                         goto out;
 495                 lock = &block->b_call->a_args.lock;
 496         } else
 497                 lock->fl.fl_flags &= ~FL_SLEEP;
 498
 499         if (block->b_flags & B_QUEUED) {
 500                 dprintk("lockd: nlmsvc_lock deferred block %p flags %d\n",
 501                                                         block, block->b_flags);
 502                 if (block->b_granted) {
 503                         nlmsvc_unlink_block(block);
 504                         ret = nlm_granted;
 505                         goto out;
 506                 }
 507                 if (block->b_flags & B_TIMED_OUT) {
 508                         nlmsvc_unlink_block(block);
 509                         ret = nlm_lck_denied;
 510                         goto out;
 511                 }
 512                 ret = nlm_drop_reply;
 513                 goto out;
 514         }
 515
 516         if (locks_in_grace(SVC_NET(rqstp)) && !reclaim) {
 517                 ret = nlm_lck_denied_grace_period;
 518                 goto out;
 519         }
 520         if (reclaim && !locks_in_grace(SVC_NET(rqstp))) {
 521                 ret = nlm_lck_denied_grace_period;
 522                 goto out;
 523         }
 524
 525         if (!wait)
 526                 lock->fl.fl_flags &= ~FL_SLEEP;
 527         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
 528         lock->fl.fl_flags &= ~FL_SLEEP;
 529
 530         dprintk("lockd: vfs_lock_file returned %d\n", error);
 531         switch (error) {
 532                 case 0:
 533                         ret = nlm_granted;
 534                         goto out;
 535                 case -EAGAIN:
 536                         /*
 537                          * If this is a blocking request for an
 538                          * already pending lock request then we need
 539                          * to put it back on lockd's block list
 540                          */
 541                         if (wait)
 542                                 break;
 543                         ret = nlm_lck_denied;
 544                         goto out;
 545                 case FILE_LOCK_DEFERRED:
 546                         if (wait)
 547                                 break;
 548                         /* Filesystem lock operation is in progress
 549                            Add it to the queue waiting for callback */
 550                         ret = nlmsvc_defer_lock_rqst(rqstp, block);
 551                         goto out;
 552                 case -EDEADLK:
 553                         ret = nlm_deadlock;
 554                         goto out;
 555                 default:                        /* includes ENOLCK */
 556                         ret = nlm_lck_denied_nolocks;
 557                         goto out;
 558         }
 559
 560         ret = nlm_lck_blocked;
 561
 562         /* Append to list of blocked */
 563         nlmsvc_insert_block(block, NLM_NEVER);
 564 out:
 565         mutex_unlock(&file->f_mutex);
 566         nlmsvc_release_block(block);
 567         dprintk("lockd: nlmsvc_lock returned %u\n", ret);
 568         return ret;
 569 }
 570
 571 /*
 572  * Test for presence of a conflicting lock.
 573  */
 574 __be32
 575 nlmsvc_testlock(struct svc_rqst *rqstp, struct nlm_file *file,
 576                 struct nlm_host *host, struct nlm_lock *lock,
 577                 struct nlm_lock *conflock, struct nlm_cookie *cookie)
 578 {
 579         int                     error;
 580         __be32                  ret;
 581         struct nlm_lockowner    *test_owner;
 582
 583         dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
 584                                 nlmsvc_file_inode(file)->i_sb->s_id,
 585                                 nlmsvc_file_inode(file)->i_ino,
 586                                 lock->fl.fl_type,
 587                                 (long long)lock->fl.fl_start,
 588                                 (long long)lock->fl.fl_end);
 589
 590         if (locks_in_grace(SVC_NET(rqstp))) {
 591                 ret = nlm_lck_denied_grace_period;
 592                 goto out;
 593         }
 594
 595         /* If there's a conflicting lock, remember to clean up the test lock */
 596         test_owner = (struct nlm_lockowner *)lock->fl.fl_owner;
 597
 598         error = vfs_test_lock(file->f_file, &lock->fl);
 599         if (error) {
 600                 /* We can't currently deal with deferred test requests */
 601                 if (error == FILE_LOCK_DEFERRED)
 602                         WARN_ON_ONCE(1);
 603
 604                 ret = nlm_lck_denied_nolocks;
 605                 goto out;
 606         }
 607
 608         if (lock->fl.fl_type == F_UNLCK) {
 609                 ret = nlm_granted;
 610                 goto out;
 611         }
 612
 613         dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
 614                 lock->fl.fl_type, (long long)lock->fl.fl_start,
 615                 (long long)lock->fl.fl_end);
 616         conflock->caller = "somehost";  /* FIXME */
 617         conflock->len = strlen(conflock->caller);
 618         conflock->oh.len = 0;           /* don't return OH info */
 619         conflock->svid = lock->fl.fl_pid;
 620         conflock->fl.fl_type = lock->fl.fl_type;
 621         conflock->fl.fl_start = lock->fl.fl_start;
 622         conflock->fl.fl_end = lock->fl.fl_end;
 623         locks_release_private(&lock->fl);
 624
 625         /* Clean up the test lock */
 626         lock->fl.fl_owner = NULL;
 627         nlmsvc_put_lockowner(test_owner);
 628
 629         ret = nlm_lck_denied;
 630 out:
 631         return ret;
 632 }
 633
 634 /*
 635  * Remove a lock.
 636  * This implies a CANCEL call: We send a GRANT_MSG, the client replies
 637  * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
 638  * afterwards. In this case the block will still be there, and hence
 639  * must be removed.
 640  */
 641 __be32
 642 nlmsvc_unlock(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
 643 {
 644         int     error;
 645
 646         dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
 647                                 nlmsvc_file_inode(file)->i_sb->s_id,
 648                                 nlmsvc_file_inode(file)->i_ino,
 649                                 lock->fl.fl_pid,
 650                                 (long long)lock->fl.fl_start,
 651                                 (long long)lock->fl.fl_end);
 652
 653         /* First, cancel any lock that might be there */
 654         nlmsvc_cancel_blocked(net, file, lock);
 655
 656         lock->fl.fl_type = F_UNLCK;
 657         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
 658
 659         return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
 660 }
 661
 662 /*
 663  * Cancel a previously blocked request.
 664  *
 665  * A cancel request always overrides any grant that may currently
 666  * be in progress.
 667  * The calling procedure must check whether the file can be closed.
 668  */
 669 __be32
 670 nlmsvc_cancel_blocked(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
 671 {
 672         struct nlm_block        *block;
 673         int status = 0;
 674
 675         dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
 676                                 nlmsvc_file_inode(file)->i_sb->s_id,
 677                                 nlmsvc_file_inode(file)->i_ino,
 678                                 lock->fl.fl_pid,
 679                                 (long long)lock->fl.fl_start,
 680                                 (long long)lock->fl.fl_end);
 681
 682         if (locks_in_grace(net))
 683                 return nlm_lck_denied_grace_period;
 684
 685         mutex_lock(&file->f_mutex);
 686         block = nlmsvc_lookup_block(file, lock);
 687         mutex_unlock(&file->f_mutex);
 688         if (block != NULL) {
 689                 vfs_cancel_lock(block->b_file->f_file,
 690                                 &block->b_call->a_args.lock.fl);
 691                 status = nlmsvc_unlink_block(block);
 692                 nlmsvc_release_block(block);
 693         }
 694         return status ? nlm_lck_denied : nlm_granted;
 695 }
 696
 697 /*
 698  * This is a callback from the filesystem for VFS file lock requests.
 699  * It will be used if lm_grant is defined and the filesystem can not
 700  * respond to the request immediately.
 701  * For SETLK or SETLKW request it will get the local posix lock.
 702  * In all cases it will move the block to the head of nlm_blocked q where
 703  * nlmsvc_retry_blocked() can send back a reply for SETLKW or revisit the
 704  * deferred rpc for GETLK and SETLK.
 705  */
 706 static void
 707 nlmsvc_update_deferred_block(struct nlm_block *block, int result)
 708 {
 709         block->b_flags |= B_GOT_CALLBACK;
 710         if (result == 0)
 711                 block->b_granted = 1;
 712         else
 713                 block->b_flags |= B_TIMED_OUT;
 714 }
 715
 716 static int nlmsvc_grant_deferred(struct file_lock *fl, int result)
 717 {
 718         struct nlm_block *block;
 719         int rc = -ENOENT;
 720
 721         spin_lock(&nlm_blocked_lock);
 722         list_for_each_entry(block, &nlm_blocked, b_list) {
 723                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
 724                         dprintk("lockd: nlmsvc_notify_blocked block %p flags %d\n",
 725                                                         block, block->b_flags);
 726                         if (block->b_flags & B_QUEUED) {
 727                                 if (block->b_flags & B_TIMED_OUT) {
 728                                         rc = -ENOLCK;
 729                                         break;
 730                                 }
 731                                 nlmsvc_update_deferred_block(block, result);
 732                         } else if (result == 0)
 733                                 block->b_granted = 1;
 734
 735                         nlmsvc_insert_block_locked(block, 0);
 736                         svc_wake_up(block->b_daemon);
 737                         rc = 0;
 738                         break;
 739                 }
 740         }
 741         spin_unlock(&nlm_blocked_lock);
 742         if (rc == -ENOENT)
 743                 printk(KERN_WARNING "lockd: grant for unknown block\n");
 744         return rc;
 745 }
 746
 747 /*
 748  * Unblock a blocked lock request. This is a callback invoked from the
 749  * VFS layer when a lock on which we blocked is removed.
 750  *
 751  * This function doesn't grant the blocked lock instantly, but rather moves
 752  * the block to the head of nlm_blocked where it can be picked up by lockd.
 753  */
 754 static void
 755 nlmsvc_notify_blocked(struct file_lock *fl)
 756 {
 757         struct nlm_block        *block;
 758
 759         dprintk("lockd: VFS unblock notification for block %p\n", fl);
 760         spin_lock(&nlm_blocked_lock);
 761         list_for_each_entry(block, &nlm_blocked, b_list) {
 762                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
 763                         nlmsvc_insert_block_locked(block, 0);
 764                         spin_unlock(&nlm_blocked_lock);
 765                         svc_wake_up(block->b_daemon);
 766                         return;
 767                 }
 768         }
 769         spin_unlock(&nlm_blocked_lock);
 770         printk(KERN_WARNING "lockd: notification for unknown block!\n");
 771 }
 772
 773 static fl_owner_t nlmsvc_get_owner(fl_owner_t owner)
 774 {
 775         return nlmsvc_get_lockowner(owner);
 776 }
 777
 778 static void nlmsvc_put_owner(fl_owner_t owner)
 779 {
 780         nlmsvc_put_lockowner(owner);
 781 }
 782
 783 const struct lock_manager_operations nlmsvc_lock_operations = {
 784         .lm_notify = nlmsvc_notify_blocked,
 785         .lm_grant = nlmsvc_grant_deferred,
 786         .lm_get_owner = nlmsvc_get_owner,
 787         .lm_put_owner = nlmsvc_put_owner,
 788 };
 789
 790 /*
 791  * Try to claim a lock that was previously blocked.
 792  *
 793  * Note that we use both the RPC_GRANTED_MSG call _and_ an async
 794  * RPC thread when notifying the client. This seems like overkill...
 795  * Here's why:
 796  *  -   we don't want to use a synchronous RPC thread, otherwise
 797  *      we might find ourselves hanging on a dead portmapper.
 798  *  -   Some lockd implementations (e.g. HP) don't react to
 799  *      RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
 800  */
 801 static void
 802 nlmsvc_grant_blocked(struct nlm_block *block)
 803 {
 804         struct nlm_file         *file = block->b_file;
 805         struct nlm_lock         *lock = &block->b_call->a_args.lock;
 806         int                     error;
 807         loff_t                  fl_start, fl_end;
 808
 809         dprintk("lockd: grant blocked lock %p\n", block);
 810
 811         kref_get(&block->b_count);
 812
 813         /* Unlink block request from list */
 814         nlmsvc_unlink_block(block);
 815
 816         /* If b_granted is true this means we've been here before.
 817          * Just retry the grant callback, possibly refreshing the RPC
 818          * binding */
 819         if (block->b_granted) {
 820                 nlm_rebind_host(block->b_host);
 821                 goto callback;
 822         }
 823
 824         /* Try the lock operation again */
 825         /* vfs_lock_file() can mangle fl_start and fl_end, but we need
 826          * them unchanged for the GRANT_MSG
 827          */
 828         lock->fl.fl_flags |= FL_SLEEP;
 829         fl_start = lock->fl.fl_start;
 830         fl_end = lock->fl.fl_end;
 831         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
 832         lock->fl.fl_flags &= ~FL_SLEEP;
 833         lock->fl.fl_start = fl_start;
 834         lock->fl.fl_end = fl_end;
 835
 836         switch (error) {
 837         case 0:
 838                 break;
 839         case FILE_LOCK_DEFERRED:
 840                 dprintk("lockd: lock still blocked error %d\n", error);
 841                 nlmsvc_insert_block(block, NLM_NEVER);
 842                 nlmsvc_release_block(block);
 843                 return;
 844         default:
 845                 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
 846                                 -error, __func__);
 847                 nlmsvc_insert_block(block, 10 * HZ);
 848                 nlmsvc_release_block(block);
 849                 return;
 850         }
 851
 852 callback:
 853         /* Lock was granted by VFS. */
 854         dprintk("lockd: GRANTing blocked lock.\n");
 855         block->b_granted = 1;
 856
 857         /* keep block on the list, but don't reattempt until the RPC
 858          * completes or the submission fails
 859          */
 860         nlmsvc_insert_block(block, NLM_NEVER);
 861
 862         /* Call the client -- use a soft RPC task since nlmsvc_retry_blocked
 863          * will queue up a new one if this one times out
 864          */
 865         error = nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
 866                                 &nlmsvc_grant_ops);
 867
 868         /* RPC submission failed, wait a bit and retry */
 869         if (error < 0)
 870                 nlmsvc_insert_block(block, 10 * HZ);
 871 }
 872
 873 /*
 874  * This is the callback from the RPC layer when the NLM_GRANTED_MSG
 875  * RPC call has succeeded or timed out.
 876  * Like all RPC callbacks, it is invoked by the rpciod process, so it
 877  * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
 878  * chain once more in order to have it removed by lockd itself (which can
 879  * then sleep on the file semaphore without disrupting e.g. the nfs client).
 880  */
 881 static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
 882 {
 883         struct nlm_rqst         *call = data;
 884         struct nlm_block        *block = call->a_block;
 885         unsigned long           timeout;
 886
 887         dprintk("lockd: GRANT_MSG RPC callback\n");
 888
 889         spin_lock(&nlm_blocked_lock);
 890         /* if the block is not on a list at this point then it has
 891          * been invalidated. Don't try to requeue it.
 892          *
 893          * FIXME: it's possible that the block is removed from the list
 894          * after this check but before the nlmsvc_insert_block. In that
 895          * case it will be added back. Perhaps we need better locking
 896          * for nlm_blocked?
 897          */
 898         if (list_empty(&block->b_list))
 899                 goto out;
 900
 901         /* Technically, we should down the file semaphore here. Since we
 902          * move the block towards the head of the queue only, no harm
 903          * can be done, though. */
 904         if (task->tk_status < 0) {
 905                 /* RPC error: Re-insert for retransmission */
 906                 timeout = 10 * HZ;
 907         } else {
 908                 /* Call was successful, now wait for client callback */
 909                 timeout = 60 * HZ;
 910         }
 911         nlmsvc_insert_block_locked(block, timeout);
 912         svc_wake_up(block->b_daemon);
 913 out:
 914         spin_unlock(&nlm_blocked_lock);
 915 }
 916
 917 /*
 918  * FIXME: nlmsvc_release_block() grabs a mutex.  This is not allowed for an
 919  * .rpc_release rpc_call_op
 920  */
 921 static void nlmsvc_grant_release(void *data)
 922 {
 923         struct nlm_rqst         *call = data;
 924         nlmsvc_release_block(call->a_block);
 925 }
 926
 927 static const struct rpc_call_ops nlmsvc_grant_ops = {
 928         .rpc_call_done = nlmsvc_grant_callback,
 929         .rpc_release = nlmsvc_grant_release,
 930 };
 931
 932 /*
 933  * We received a GRANT_RES callback. Try to find the corresponding
 934  * block.
 935  */
 936 void
 937 nlmsvc_grant_reply(struct nlm_cookie *cookie, __be32 status)
 938 {
 939         struct nlm_block        *block;
 940
 941         dprintk("grant_reply: looking for cookie %x, s=%d \n",
 942                 *(unsigned int *)(cookie->data), status);
 943         if (!(block = nlmsvc_find_block(cookie)))
 944                 return;
 945
 946         if (status == nlm_lck_denied_grace_period) {
 947                 /* Try again in a couple of seconds */
 948                 nlmsvc_insert_block(block, 10 * HZ);
 949         } else {
 950                 /*
 951                  * Lock is now held by client, or has been rejected.
 952                  * In both cases, the block should be removed.
 953                  */
 954                 nlmsvc_unlink_block(block);
 955         }
 956         nlmsvc_release_block(block);
 957 }
 958
 959 /* Helper function to handle retry of a deferred block.
 960  * If it is a blocking lock, call grant_blocked.
 961  * For a non-blocking lock or test lock, revisit the request.
 962  */
 963 static void
 964 retry_deferred_block(struct nlm_block *block)
 965 {
 966         if (!(block->b_flags & B_GOT_CALLBACK))
 967                 block->b_flags |= B_TIMED_OUT;
 968         nlmsvc_insert_block(block, NLM_TIMEOUT);
 969         dprintk("revisit block %p flags %d\n",  block, block->b_flags);
 970         if (block->b_deferred_req) {
 971                 block->b_deferred_req->revisit(block->b_deferred_req, 0);
 972                 block->b_deferred_req = NULL;
 973         }
 974 }
 975
 976 /*
 977  * Retry all blocked locks that have been notified. This is where lockd
 978  * picks up locks that can be granted, or grant notifications that must
 979  * be retransmitted.
 980  */
 981 unsigned long
 982 nlmsvc_retry_blocked(void)
 983 {
 984         unsigned long   timeout = MAX_SCHEDULE_TIMEOUT;
 985         struct nlm_block *block;
 986
 987         spin_lock(&nlm_blocked_lock);
 988         while (!list_empty(&nlm_blocked) && !kthread_should_stop()) {
 989                 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
 990
 991                 if (block->b_when == NLM_NEVER)
 992                         break;
 993                 if (time_after(block->b_when, jiffies)) {
 994                         timeout = block->b_when - jiffies;
 995                         break;
 996                 }
 997                 spin_unlock(&nlm_blocked_lock);
 998
 999                 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
1000                         block, block->b_when);
1001                 if (block->b_flags & B_QUEUED) {
1002                         dprintk("nlmsvc_retry_blocked delete block (%p, granted=%d, flags=%d)\n",
1003                                 block, block->b_granted, block->b_flags);
1004                         retry_deferred_block(block);
1005                 } else
1006                         nlmsvc_grant_blocked(block);
1007                 spin_lock(&nlm_blocked_lock);
1008         }
1009         spin_unlock(&nlm_blocked_lock);
1010
1011         return timeout;
1012 }