fs/lockd/svclock.c

   1 /*
   2  * linux/fs/lockd/svclock.c
   3  *
   4  * Handling of server-side locks, mostly of the blocked variety.
   5  * This is the ugliest part of lockd because we tread on very thin ice.
   6  * GRANT and CANCEL calls may get stuck, meet in mid-flight, etc.
   7  * IMNSHO introducing the grant callback into the NLM protocol was one
   8  * of the worst ideas Sun ever had. Except maybe for the idea of doing
   9  * NFS file locking at all.
  10  *
  11  * I'm trying hard to avoid race conditions by protecting most accesses
  12  * to a file's list of blocked locks through a semaphore. The global
  13  * list of blocked locks is not protected in this fashion however.
  14  * Therefore, some functions (such as the RPC callback for the async grant
  15  * call) move blocked locks towards the head of the list *while some other
  16  * process might be traversing it*. This should not be a problem in
  17  * practice, because this will only cause functions traversing the list
  18  * to visit some blocks twice.
  19  *
  20  * Copyright (C) 1996, Olaf Kirch <okir@monad.swb.de>
  21  */
  22
  23 #include <linux/types.h>
  24 #include <linux/slab.h>
  25 #include <linux/errno.h>
  26 #include <linux/kernel.h>
  27 #include <linux/sched.h>
  28 #include <linux/sunrpc/clnt.h>
  29 #include <linux/sunrpc/svc.h>
  30 #include <linux/lockd/nlm.h>
  31 #include <linux/lockd/lockd.h>
  32 #include <linux/kthread.h>
  33
  34 #define NLMDBG_FACILITY         NLMDBG_SVCLOCK
  35
  36 #ifdef CONFIG_LOCKD_V4
  37 #define nlm_deadlock    nlm4_deadlock
  38 #else
  39 #define nlm_deadlock    nlm_lck_denied
  40 #endif
  41
  42 static void nlmsvc_release_block(struct nlm_block *block);
  43 static void     nlmsvc_insert_block(struct nlm_block *block, unsigned long);
  44 static void     nlmsvc_remove_block(struct nlm_block *block);
  45
  46 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock);
  47 static void nlmsvc_freegrantargs(struct nlm_rqst *call);
  48 static const struct rpc_call_ops nlmsvc_grant_ops;
  49
  50 /*
  51  * The list of blocked locks to retry
  52  */
  53 static LIST_HEAD(nlm_blocked);
  54 static DEFINE_SPINLOCK(nlm_blocked_lock);
  55
  56 #ifdef LOCKD_DEBUG
  57 static const char *nlmdbg_cookie2a(const struct nlm_cookie *cookie)
  58 {
  59         /*
  60          * We can get away with a static buffer because we're only
  61          * called with BKL held.
  62          */
  63         static char buf[2*NLM_MAXCOOKIELEN+1];
  64         unsigned int i, len = sizeof(buf);
  65         char *p = buf;
  66
  67         len--;  /* allow for trailing \0 */
  68         if (len < 3)
  69                 return "???";
  70         for (i = 0 ; i < cookie->len ; i++) {
  71                 if (len < 2) {
  72                         strcpy(p-3, "...");
  73                         break;
  74                 }
  75                 sprintf(p, "%02x", cookie->data[i]);
  76                 p += 2;
  77                 len -= 2;
  78         }
  79         *p = '\0';
  80
  81         return buf;
  82 }
  83 #endif
  84
  85 /*
  86  * Insert a blocked lock into the global list
  87  */
  88 static void
  89 nlmsvc_insert_block_locked(struct nlm_block *block, unsigned long when)
  90 {
  91         struct nlm_block *b;
  92         struct list_head *pos;
  93
  94         dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block, when);
  95         if (list_empty(&block->b_list)) {
  96                 kref_get(&block->b_count);
  97         } else {
  98                 list_del_init(&block->b_list);
  99         }
 100
 101         pos = &nlm_blocked;
 102         if (when != NLM_NEVER) {
 103                 if ((when += jiffies) == NLM_NEVER)
 104                         when ++;
 105                 list_for_each(pos, &nlm_blocked) {
 106                         b = list_entry(pos, struct nlm_block, b_list);
 107                         if (time_after(b->b_when,when) || b->b_when == NLM_NEVER)
 108                                 break;
 109                 }
 110                 /* On normal exit from the loop, pos == &nlm_blocked,
 111                  * so we will be adding to the end of the list - good
 112                  */
 113         }
 114
 115         list_add_tail(&block->b_list, pos);
 116         block->b_when = when;
 117 }
 118
 119 static void nlmsvc_insert_block(struct nlm_block *block, unsigned long when)
 120 {
 121         spin_lock(&nlm_blocked_lock);
 122         nlmsvc_insert_block_locked(block, when);
 123         spin_unlock(&nlm_blocked_lock);
 124 }
 125
 126 /*
 127  * Remove a block from the global list
 128  */
 129 static inline void
 130 nlmsvc_remove_block(struct nlm_block *block)
 131 {
 132         if (!list_empty(&block->b_list)) {
 133                 spin_lock(&nlm_blocked_lock);
 134                 list_del_init(&block->b_list);
 135                 spin_unlock(&nlm_blocked_lock);
 136                 nlmsvc_release_block(block);
 137         }
 138 }
 139
 140 /*
 141  * Find a block for a given lock
 142  */
 143 static struct nlm_block *
 144 nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock)
 145 {
 146         struct nlm_block        *block;
 147         struct file_lock        *fl;
 148
 149         dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
 150                                 file, lock->fl.fl_pid,
 151                                 (long long)lock->fl.fl_start,
 152                                 (long long)lock->fl.fl_end, lock->fl.fl_type);
 153         list_for_each_entry(block, &nlm_blocked, b_list) {
 154                 fl = &block->b_call->a_args.lock.fl;
 155                 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
 156                                 block->b_file, fl->fl_pid,
 157                                 (long long)fl->fl_start,
 158                                 (long long)fl->fl_end, fl->fl_type,
 159                                 nlmdbg_cookie2a(&block->b_call->a_args.cookie));
 160                 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
 161                         kref_get(&block->b_count);
 162                         return block;
 163                 }
 164         }
 165
 166         return NULL;
 167 }
 168
 169 static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
 170 {
 171         if (a->len != b->len)
 172                 return 0;
 173         if (memcmp(a->data, b->data, a->len))
 174                 return 0;
 175         return 1;
 176 }
 177
 178 /*
 179  * Find a block with a given NLM cookie.
 180  */
 181 static inline struct nlm_block *
 182 nlmsvc_find_block(struct nlm_cookie *cookie)
 183 {
 184         struct nlm_block *block;
 185
 186         list_for_each_entry(block, &nlm_blocked, b_list) {
 187                 if (nlm_cookie_match(&block->b_call->a_args.cookie,cookie))
 188                         goto found;
 189         }
 190
 191         return NULL;
 192
 193 found:
 194         dprintk("nlmsvc_find_block(%s): block=%p\n", nlmdbg_cookie2a(cookie), block);
 195         kref_get(&block->b_count);
 196         return block;
 197 }
 198
 199 /*
 200  * Create a block and initialize it.
 201  *
 202  * Note: we explicitly set the cookie of the grant reply to that of
 203  * the blocked lock request. The spec explicitly mentions that the client
 204  * should _not_ rely on the callback containing the same cookie as the
 205  * request, but (as I found out later) that's because some implementations
 206  * do just this. Never mind the standards comittees, they support our
 207  * logging industries.
 208  *
 209  * 10 years later: I hope we can safely ignore these old and broken
 210  * clients by now. Let's fix this so we can uniquely identify an incoming
 211  * GRANTED_RES message by cookie, without having to rely on the client's IP
 212  * address. --okir
 213  */
 214 static struct nlm_block *
 215 nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_host *host,
 216                     struct nlm_file *file, struct nlm_lock *lock,
 217                     struct nlm_cookie *cookie)
 218 {
 219         struct nlm_block        *block;
 220         struct nlm_rqst         *call = NULL;
 221
 222         nlm_get_host(host);
 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         block->b_fl = NULL;
 250         file->f_count++;
 251
 252         /* Add to file's list of blocks */
 253         list_add(&block->b_flist, &file->f_blocks);
 254
 255         /* Set up RPC arguments for callback */
 256         block->b_call = call;
 257         call->a_flags   = RPC_TASK_ASYNC;
 258         call->a_block = block;
 259
 260         return block;
 261
 262 failed_free:
 263         kfree(block);
 264 failed:
 265         nlmsvc_release_call(call);
 266         return NULL;
 267 }
 268
 269 /*
 270  * Delete a block.
 271  * It is the caller's responsibility to check whether the file
 272  * can be closed hereafter.
 273  */
 274 static int nlmsvc_unlink_block(struct nlm_block *block)
 275 {
 276         int status;
 277         dprintk("lockd: unlinking block %p...\n", block);
 278
 279         /* Remove block from list */
 280         status = posix_unblock_lock(block->b_file->f_file, &block->b_call->a_args.lock.fl);
 281         nlmsvc_remove_block(block);
 282         return status;
 283 }
 284
 285 static void nlmsvc_free_block(struct kref *kref)
 286 {
 287         struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
 288         struct nlm_file         *file = block->b_file;
 289
 290         dprintk("lockd: freeing block %p...\n", block);
 291
 292         /* Remove block from file's list of blocks */
 293         mutex_lock(&file->f_mutex);
 294         list_del_init(&block->b_flist);
 295         mutex_unlock(&file->f_mutex);
 296
 297         nlmsvc_freegrantargs(block->b_call);
 298         nlmsvc_release_call(block->b_call);
 299         nlm_release_file(block->b_file);
 300         kfree(block->b_fl);
 301         kfree(block);
 302 }
 303
 304 static void nlmsvc_release_block(struct nlm_block *block)
 305 {
 306         if (block != NULL)
 307                 kref_put(&block->b_count, nlmsvc_free_block);
 308 }
 309
 310 /*
 311  * Loop over all blocks and delete blocks held by
 312  * a matching host.
 313  */
 314 void nlmsvc_traverse_blocks(struct nlm_host *host,
 315                         struct nlm_file *file,
 316                         nlm_host_match_fn_t match)
 317 {
 318         struct nlm_block *block, *next;
 319
 320 restart:
 321         mutex_lock(&file->f_mutex);
 322         list_for_each_entry_safe(block, next, &file->f_blocks, b_flist) {
 323                 if (!match(block->b_host, host))
 324                         continue;
 325                 /* Do not destroy blocks that are not on
 326                  * the global retry list - why? */
 327                 if (list_empty(&block->b_list))
 328                         continue;
 329                 kref_get(&block->b_count);
 330                 mutex_unlock(&file->f_mutex);
 331                 nlmsvc_unlink_block(block);
 332                 nlmsvc_release_block(block);
 333                 goto restart;
 334         }
 335         mutex_unlock(&file->f_mutex);
 336 }
 337
 338 /*
 339  * Initialize arguments for GRANTED call. The nlm_rqst structure
 340  * has been cleared already.
 341  */
 342 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
 343 {
 344         locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
 345         memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
 346         call->a_args.lock.caller = utsname()->nodename;
 347         call->a_args.lock.oh.len = lock->oh.len;
 348
 349         /* set default data area */
 350         call->a_args.lock.oh.data = call->a_owner;
 351         call->a_args.lock.svid = lock->fl.fl_pid;
 352
 353         if (lock->oh.len > NLMCLNT_OHSIZE) {
 354                 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
 355                 if (!data)
 356                         return 0;
 357                 call->a_args.lock.oh.data = (u8 *) data;
 358         }
 359
 360         memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
 361         return 1;
 362 }
 363
 364 static void nlmsvc_freegrantargs(struct nlm_rqst *call)
 365 {
 366         if (call->a_args.lock.oh.data != call->a_owner)
 367                 kfree(call->a_args.lock.oh.data);
 368
 369         locks_release_private(&call->a_args.lock.fl);
 370 }
 371
 372 /*
 373  * Deferred lock request handling for non-blocking lock
 374  */
 375 static __be32
 376 nlmsvc_defer_lock_rqst(struct svc_rqst *rqstp, struct nlm_block *block)
 377 {
 378         __be32 status = nlm_lck_denied_nolocks;
 379
 380         block->b_flags |= B_QUEUED;
 381
 382         nlmsvc_insert_block(block, NLM_TIMEOUT);
 383
 384         block->b_cache_req = &rqstp->rq_chandle;
 385         if (rqstp->rq_chandle.defer) {
 386                 block->b_deferred_req =
 387                         rqstp->rq_chandle.defer(block->b_cache_req);
 388                 if (block->b_deferred_req != NULL)
 389                         status = nlm_drop_reply;
 390         }
 391         dprintk("lockd: nlmsvc_defer_lock_rqst block %p flags %d status %d\n",
 392                 block, block->b_flags, ntohl(status));
 393
 394         return status;
 395 }
 396
 397 /*
 398  * Attempt to establish a lock, and if it can't be granted, block it
 399  * if required.
 400  */
 401 __be32
 402 nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
 403             struct nlm_host *host, struct nlm_lock *lock, int wait,
 404             struct nlm_cookie *cookie, int reclaim)
 405 {
 406         struct nlm_block        *block = NULL;
 407         int                     error;
 408         __be32                  ret;
 409
 410         dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
 411                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
 412                                 file->f_file->f_path.dentry->d_inode->i_ino,
 413                                 lock->fl.fl_type, lock->fl.fl_pid,
 414                                 (long long)lock->fl.fl_start,
 415                                 (long long)lock->fl.fl_end,
 416                                 wait);
 417
 418         /* Lock file against concurrent access */
 419         mutex_lock(&file->f_mutex);
 420         /* Get existing block (in case client is busy-waiting)
 421          * or create new block
 422          */
 423         block = nlmsvc_lookup_block(file, lock);
 424         if (block == NULL) {
 425                 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
 426                 ret = nlm_lck_denied_nolocks;
 427                 if (block == NULL)
 428                         goto out;
 429                 lock = &block->b_call->a_args.lock;
 430         } else
 431                 lock->fl.fl_flags &= ~FL_SLEEP;
 432
 433         if (block->b_flags & B_QUEUED) {
 434                 dprintk("lockd: nlmsvc_lock deferred block %p flags %d\n",
 435                                                         block, block->b_flags);
 436                 if (block->b_granted) {
 437                         nlmsvc_unlink_block(block);
 438                         ret = nlm_granted;
 439                         goto out;
 440                 }
 441                 if (block->b_flags & B_TIMED_OUT) {
 442                         nlmsvc_unlink_block(block);
 443                         ret = nlm_lck_denied;
 444                         goto out;
 445                 }
 446                 ret = nlm_drop_reply;
 447                 goto out;
 448         }
 449
 450         if (locks_in_grace() && !reclaim) {
 451                 ret = nlm_lck_denied_grace_period;
 452                 goto out;
 453         }
 454         if (reclaim && !locks_in_grace()) {
 455                 ret = nlm_lck_denied_grace_period;
 456                 goto out;
 457         }
 458
 459         if (!wait)
 460                 lock->fl.fl_flags &= ~FL_SLEEP;
 461         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
 462         lock->fl.fl_flags &= ~FL_SLEEP;
 463
 464         dprintk("lockd: vfs_lock_file returned %d\n", error);
 465         switch (error) {
 466                 case 0:
 467                         ret = nlm_granted;
 468                         goto out;
 469                 case -EAGAIN:
 470                         /*
 471                          * If this is a blocking request for an
 472                          * already pending lock request then we need
 473                          * to put it back on lockd's block list
 474                          */
 475                         if (wait)
 476                                 break;
 477                         ret = nlm_lck_denied;
 478                         goto out;
 479                 case FILE_LOCK_DEFERRED:
 480                         if (wait)
 481                                 break;
 482                         /* Filesystem lock operation is in progress
 483                            Add it to the queue waiting for callback */
 484                         ret = nlmsvc_defer_lock_rqst(rqstp, block);
 485                         goto out;
 486                 case -EDEADLK:
 487                         ret = nlm_deadlock;
 488                         goto out;
 489                 default:                        /* includes ENOLCK */
 490                         ret = nlm_lck_denied_nolocks;
 491                         goto out;
 492         }
 493
 494         ret = nlm_lck_blocked;
 495
 496         /* Append to list of blocked */
 497         nlmsvc_insert_block(block, NLM_NEVER);
 498 out:
 499         mutex_unlock(&file->f_mutex);
 500         nlmsvc_release_block(block);
 501         dprintk("lockd: nlmsvc_lock returned %u\n", ret);
 502         return ret;
 503 }
 504
 505 /*
 506  * Test for presence of a conflicting lock.
 507  */
 508 __be32
 509 nlmsvc_testlock(struct svc_rqst *rqstp, struct nlm_file *file,
 510                 struct nlm_host *host, struct nlm_lock *lock,
 511                 struct nlm_lock *conflock, struct nlm_cookie *cookie)
 512 {
 513         struct nlm_block        *block = NULL;
 514         int                     error;
 515         __be32                  ret;
 516
 517         dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
 518                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
 519                                 file->f_file->f_path.dentry->d_inode->i_ino,
 520                                 lock->fl.fl_type,
 521                                 (long long)lock->fl.fl_start,
 522                                 (long long)lock->fl.fl_end);
 523
 524         /* Get existing block (in case client is busy-waiting) */
 525         block = nlmsvc_lookup_block(file, lock);
 526
 527         if (block == NULL) {
 528                 struct file_lock *conf = kzalloc(sizeof(*conf), GFP_KERNEL);
 529
 530                 if (conf == NULL)
 531                         return nlm_granted;
 532                 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
 533                 if (block == NULL) {
 534                         kfree(conf);
 535                         return nlm_granted;
 536                 }
 537                 block->b_fl = conf;
 538         }
 539         if (block->b_flags & B_QUEUED) {
 540                 dprintk("lockd: nlmsvc_testlock deferred block %p flags %d fl %p\n",
 541                         block, block->b_flags, block->b_fl);
 542                 if (block->b_flags & B_TIMED_OUT) {
 543                         nlmsvc_unlink_block(block);
 544                         ret = nlm_lck_denied;
 545                         goto out;
 546                 }
 547                 if (block->b_flags & B_GOT_CALLBACK) {
 548                         nlmsvc_unlink_block(block);
 549                         if (block->b_fl != NULL
 550                                         && block->b_fl->fl_type != F_UNLCK) {
 551                                 lock->fl = *block->b_fl;
 552                                 goto conf_lock;
 553                         } else {
 554                                 ret = nlm_granted;
 555                                 goto out;
 556                         }
 557                 }
 558                 ret = nlm_drop_reply;
 559                 goto out;
 560         }
 561
 562         if (locks_in_grace()) {
 563                 ret = nlm_lck_denied_grace_period;
 564                 goto out;
 565         }
 566         error = vfs_test_lock(file->f_file, &lock->fl);
 567         if (error == FILE_LOCK_DEFERRED) {
 568                 ret = nlmsvc_defer_lock_rqst(rqstp, block);
 569                 goto out;
 570         }
 571         if (error) {
 572                 ret = nlm_lck_denied_nolocks;
 573                 goto out;
 574         }
 575         if (lock->fl.fl_type == F_UNLCK) {
 576                 ret = nlm_granted;
 577                 goto out;
 578         }
 579
 580 conf_lock:
 581         dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
 582                 lock->fl.fl_type, (long long)lock->fl.fl_start,
 583                 (long long)lock->fl.fl_end);
 584         conflock->caller = "somehost";  /* FIXME */
 585         conflock->len = strlen(conflock->caller);
 586         conflock->oh.len = 0;           /* don't return OH info */
 587         conflock->svid = lock->fl.fl_pid;
 588         conflock->fl.fl_type = lock->fl.fl_type;
 589         conflock->fl.fl_start = lock->fl.fl_start;
 590         conflock->fl.fl_end = lock->fl.fl_end;
 591         ret = nlm_lck_denied;
 592 out:
 593         if (block)
 594                 nlmsvc_release_block(block);
 595         return ret;
 596 }
 597
 598 /*
 599  * Remove a lock.
 600  * This implies a CANCEL call: We send a GRANT_MSG, the client replies
 601  * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
 602  * afterwards. In this case the block will still be there, and hence
 603  * must be removed.
 604  */
 605 __be32
 606 nlmsvc_unlock(struct nlm_file *file, struct nlm_lock *lock)
 607 {
 608         int     error;
 609
 610         dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
 611                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
 612                                 file->f_file->f_path.dentry->d_inode->i_ino,
 613                                 lock->fl.fl_pid,
 614                                 (long long)lock->fl.fl_start,
 615                                 (long long)lock->fl.fl_end);
 616
 617         /* First, cancel any lock that might be there */
 618         nlmsvc_cancel_blocked(file, lock);
 619
 620         lock->fl.fl_type = F_UNLCK;
 621         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
 622
 623         return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
 624 }
 625
 626 /*
 627  * Cancel a previously blocked request.
 628  *
 629  * A cancel request always overrides any grant that may currently
 630  * be in progress.
 631  * The calling procedure must check whether the file can be closed.
 632  */
 633 __be32
 634 nlmsvc_cancel_blocked(struct nlm_file *file, struct nlm_lock *lock)
 635 {
 636         struct nlm_block        *block;
 637         int status = 0;
 638
 639         dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
 640                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
 641                                 file->f_file->f_path.dentry->d_inode->i_ino,
 642                                 lock->fl.fl_pid,
 643                                 (long long)lock->fl.fl_start,
 644                                 (long long)lock->fl.fl_end);
 645
 646         if (locks_in_grace())
 647                 return nlm_lck_denied_grace_period;
 648
 649         mutex_lock(&file->f_mutex);
 650         block = nlmsvc_lookup_block(file, lock);
 651         mutex_unlock(&file->f_mutex);
 652         if (block != NULL) {
 653                 vfs_cancel_lock(block->b_file->f_file,
 654                                 &block->b_call->a_args.lock.fl);
 655                 status = nlmsvc_unlink_block(block);
 656                 nlmsvc_release_block(block);
 657         }
 658         return status ? nlm_lck_denied : nlm_granted;
 659 }
 660
 661 /*
 662  * This is a callback from the filesystem for VFS file lock requests.
 663  * It will be used if lm_grant is defined and the filesystem can not
 664  * respond to the request immediately.
 665  * For GETLK request it will copy the reply to the nlm_block.
 666  * For SETLK or SETLKW request it will get the local posix lock.
 667  * In all cases it will move the block to the head of nlm_blocked q where
 668  * nlmsvc_retry_blocked() can send back a reply for SETLKW or revisit the
 669  * deferred rpc for GETLK and SETLK.
 670  */
 671 static void
 672 nlmsvc_update_deferred_block(struct nlm_block *block, struct file_lock *conf,
 673                              int result)
 674 {
 675         block->b_flags |= B_GOT_CALLBACK;
 676         if (result == 0)
 677                 block->b_granted = 1;
 678         else
 679                 block->b_flags |= B_TIMED_OUT;
 680         if (conf) {
 681                 if (block->b_fl)
 682                         __locks_copy_lock(block->b_fl, conf);
 683         }
 684 }
 685
 686 static int nlmsvc_grant_deferred(struct file_lock *fl, struct file_lock *conf,
 687                                         int result)
 688 {
 689         struct nlm_block *block;
 690         int rc = -ENOENT;
 691
 692         spin_lock(&nlm_blocked_lock);
 693         list_for_each_entry(block, &nlm_blocked, b_list) {
 694                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
 695                         dprintk("lockd: nlmsvc_notify_blocked block %p flags %d\n",
 696                                                         block, block->b_flags);
 697                         if (block->b_flags & B_QUEUED) {
 698                                 if (block->b_flags & B_TIMED_OUT) {
 699                                         rc = -ENOLCK;
 700                                         break;
 701                                 }
 702                                 nlmsvc_update_deferred_block(block, conf, result);
 703                         } else if (result == 0)
 704                                 block->b_granted = 1;
 705
 706                         nlmsvc_insert_block_locked(block, 0);
 707                         svc_wake_up(block->b_daemon);
 708                         rc = 0;
 709                         break;
 710                 }
 711         }
 712         spin_unlock(&nlm_blocked_lock);
 713         if (rc == -ENOENT)
 714                 printk(KERN_WARNING "lockd: grant for unknown block\n");
 715         return rc;
 716 }
 717
 718 /*
 719  * Unblock a blocked lock request. This is a callback invoked from the
 720  * VFS layer when a lock on which we blocked is removed.
 721  *
 722  * This function doesn't grant the blocked lock instantly, but rather moves
 723  * the block to the head of nlm_blocked where it can be picked up by lockd.
 724  */
 725 static void
 726 nlmsvc_notify_blocked(struct file_lock *fl)
 727 {
 728         struct nlm_block        *block;
 729
 730         dprintk("lockd: VFS unblock notification for block %p\n", fl);
 731         spin_lock(&nlm_blocked_lock);
 732         list_for_each_entry(block, &nlm_blocked, b_list) {
 733                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
 734                         nlmsvc_insert_block_locked(block, 0);
 735                         spin_unlock(&nlm_blocked_lock);
 736                         svc_wake_up(block->b_daemon);
 737                         return;
 738                 }
 739         }
 740         spin_unlock(&nlm_blocked_lock);
 741         printk(KERN_WARNING "lockd: notification for unknown block!\n");
 742 }
 743
 744 static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
 745 {
 746         return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
 747 }
 748
 749 const struct lock_manager_operations nlmsvc_lock_operations = {
 750         .lm_compare_owner = nlmsvc_same_owner,
 751         .lm_notify = nlmsvc_notify_blocked,
 752         .lm_grant = nlmsvc_grant_deferred,
 753 };
 754
 755 /*
 756  * Try to claim a lock that was previously blocked.
 757  *
 758  * Note that we use both the RPC_GRANTED_MSG call _and_ an async
 759  * RPC thread when notifying the client. This seems like overkill...
 760  * Here's why:
 761  *  -   we don't want to use a synchronous RPC thread, otherwise
 762  *      we might find ourselves hanging on a dead portmapper.
 763  *  -   Some lockd implementations (e.g. HP) don't react to
 764  *      RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
 765  */
 766 static void
 767 nlmsvc_grant_blocked(struct nlm_block *block)
 768 {
 769         struct nlm_file         *file = block->b_file;
 770         struct nlm_lock         *lock = &block->b_call->a_args.lock;
 771         int                     error;
 772
 773         dprintk("lockd: grant blocked lock %p\n", block);
 774
 775         kref_get(&block->b_count);
 776
 777         /* Unlink block request from list */
 778         nlmsvc_unlink_block(block);
 779
 780         /* If b_granted is true this means we've been here before.
 781          * Just retry the grant callback, possibly refreshing the RPC
 782          * binding */
 783         if (block->b_granted) {
 784                 nlm_rebind_host(block->b_host);
 785                 goto callback;
 786         }
 787
 788         /* Try the lock operation again */
 789         lock->fl.fl_flags |= FL_SLEEP;
 790         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
 791         lock->fl.fl_flags &= ~FL_SLEEP;
 792
 793         switch (error) {
 794         case 0:
 795                 break;
 796         case FILE_LOCK_DEFERRED:
 797                 dprintk("lockd: lock still blocked error %d\n", error);
 798                 nlmsvc_insert_block(block, NLM_NEVER);
 799                 nlmsvc_release_block(block);
 800                 return;
 801         default:
 802                 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
 803                                 -error, __func__);
 804                 nlmsvc_insert_block(block, 10 * HZ);
 805                 nlmsvc_release_block(block);
 806                 return;
 807         }
 808
 809 callback:
 810         /* Lock was granted by VFS. */
 811         dprintk("lockd: GRANTing blocked lock.\n");
 812         block->b_granted = 1;
 813
 814         /* keep block on the list, but don't reattempt until the RPC
 815          * completes or the submission fails
 816          */
 817         nlmsvc_insert_block(block, NLM_NEVER);
 818
 819         /* Call the client -- use a soft RPC task since nlmsvc_retry_blocked
 820          * will queue up a new one if this one times out
 821          */
 822         error = nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
 823                                 &nlmsvc_grant_ops);
 824
 825         /* RPC submission failed, wait a bit and retry */
 826         if (error < 0)
 827                 nlmsvc_insert_block(block, 10 * HZ);
 828 }
 829
 830 /*
 831  * This is the callback from the RPC layer when the NLM_GRANTED_MSG
 832  * RPC call has succeeded or timed out.
 833  * Like all RPC callbacks, it is invoked by the rpciod process, so it
 834  * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
 835  * chain once more in order to have it removed by lockd itself (which can
 836  * then sleep on the file semaphore without disrupting e.g. the nfs client).
 837  */
 838 static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
 839 {
 840         struct nlm_rqst         *call = data;
 841         struct nlm_block        *block = call->a_block;
 842         unsigned long           timeout;
 843
 844         dprintk("lockd: GRANT_MSG RPC callback\n");
 845
 846         spin_lock(&nlm_blocked_lock);
 847         /* if the block is not on a list at this point then it has
 848          * been invalidated. Don't try to requeue it.
 849          *
 850          * FIXME: it's possible that the block is removed from the list
 851          * after this check but before the nlmsvc_insert_block. In that
 852          * case it will be added back. Perhaps we need better locking
 853          * for nlm_blocked?
 854          */
 855         if (list_empty(&block->b_list))
 856                 goto out;
 857
 858         /* Technically, we should down the file semaphore here. Since we
 859          * move the block towards the head of the queue only, no harm
 860          * can be done, though. */
 861         if (task->tk_status < 0) {
 862                 /* RPC error: Re-insert for retransmission */
 863                 timeout = 10 * HZ;
 864         } else {
 865                 /* Call was successful, now wait for client callback */
 866                 timeout = 60 * HZ;
 867         }
 868         nlmsvc_insert_block_locked(block, timeout);
 869         svc_wake_up(block->b_daemon);
 870 out:
 871         spin_unlock(&nlm_blocked_lock);
 872 }
 873
 874 /*
 875  * FIXME: nlmsvc_release_block() grabs a mutex.  This is not allowed for an
 876  * .rpc_release rpc_call_op
 877  */
 878 static void nlmsvc_grant_release(void *data)
 879 {
 880         struct nlm_rqst         *call = data;
 881         nlmsvc_release_block(call->a_block);
 882 }
 883
 884 static const struct rpc_call_ops nlmsvc_grant_ops = {
 885         .rpc_call_done = nlmsvc_grant_callback,
 886         .rpc_release = nlmsvc_grant_release,
 887 };
 888
 889 /*
 890  * We received a GRANT_RES callback. Try to find the corresponding
 891  * block.
 892  */
 893 void
 894 nlmsvc_grant_reply(struct nlm_cookie *cookie, __be32 status)
 895 {
 896         struct nlm_block        *block;
 897
 898         dprintk("grant_reply: looking for cookie %x, s=%d \n",
 899                 *(unsigned int *)(cookie->data), status);
 900         if (!(block = nlmsvc_find_block(cookie)))
 901                 return;
 902
 903         if (block) {
 904                 if (status == nlm_lck_denied_grace_period) {
 905                         /* Try again in a couple of seconds */
 906                         nlmsvc_insert_block(block, 10 * HZ);
 907                 } else {
 908                         /* Lock is now held by client, or has been rejected.
 909                          * In both cases, the block should be removed. */
 910                         nlmsvc_unlink_block(block);
 911                 }
 912         }
 913         nlmsvc_release_block(block);
 914 }
 915
 916 /* Helper function to handle retry of a deferred block.
 917  * If it is a blocking lock, call grant_blocked.
 918  * For a non-blocking lock or test lock, revisit the request.
 919  */
 920 static void
 921 retry_deferred_block(struct nlm_block *block)
 922 {
 923         if (!(block->b_flags & B_GOT_CALLBACK))
 924                 block->b_flags |= B_TIMED_OUT;
 925         nlmsvc_insert_block(block, NLM_TIMEOUT);
 926         dprintk("revisit block %p flags %d\n",  block, block->b_flags);
 927         if (block->b_deferred_req) {
 928                 block->b_deferred_req->revisit(block->b_deferred_req, 0);
 929                 block->b_deferred_req = NULL;
 930         }
 931 }
 932
 933 /*
 934  * Retry all blocked locks that have been notified. This is where lockd
 935  * picks up locks that can be granted, or grant notifications that must
 936  * be retransmitted.
 937  */
 938 unsigned long
 939 nlmsvc_retry_blocked(void)
 940 {
 941         unsigned long   timeout = MAX_SCHEDULE_TIMEOUT;
 942         struct nlm_block *block;
 943
 944         while (!list_empty(&nlm_blocked) && !kthread_should_stop()) {
 945                 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
 946
 947                 if (block->b_when == NLM_NEVER)
 948                         break;
 949                 if (time_after(block->b_when, jiffies)) {
 950                         timeout = block->b_when - jiffies;
 951                         break;
 952                 }
 953
 954                 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
 955                         block, block->b_when);
 956                 if (block->b_flags & B_QUEUED) {
 957                         dprintk("nlmsvc_retry_blocked delete block (%p, granted=%d, flags=%d)\n",
 958                                 block, block->b_granted, block->b_flags);
 959                         retry_deferred_block(block);
 960                 } else
 961                         nlmsvc_grant_blocked(block);
 962         }
 963
 964         return timeout;
 965 }