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_xprt.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         call = nlm_alloc_call(host);
 223         if (call == NULL)
 224                 return NULL;
 225
 226         /* Allocate memory for block, and initialize arguments */
 227         block = kzalloc(sizeof(*block), GFP_KERNEL);
 228         if (block == NULL)
 229                 goto failed;
 230         kref_init(&block->b_count);
 231         INIT_LIST_HEAD(&block->b_list);
 232         INIT_LIST_HEAD(&block->b_flist);
 233
 234         if (!nlmsvc_setgrantargs(call, lock))
 235                 goto failed_free;
 236
 237         /* Set notifier function for VFS, and init args */
 238         call->a_args.lock.fl.fl_flags |= FL_SLEEP;
 239         call->a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
 240         nlmclnt_next_cookie(&call->a_args.cookie);
 241
 242         dprintk("lockd: created block %p...\n", block);
 243
 244         /* Create and initialize the block */
 245         block->b_daemon = rqstp->rq_server;
 246         block->b_host   = host;
 247         block->b_file   = file;
 248         block->b_fl = NULL;
 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 = posix_unblock_lock(block->b_file->f_file, &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         mutex_lock(&file->f_mutex);
 293         list_del_init(&block->b_flist);
 294         mutex_unlock(&file->f_mutex);
 295
 296         nlmsvc_freegrantargs(block->b_call);
 297         nlmsvc_release_call(block->b_call);
 298         nlm_release_file(block->b_file);
 299         kfree(block->b_fl);
 300         kfree(block);
 301 }
 302
 303 static void nlmsvc_release_block(struct nlm_block *block)
 304 {
 305         if (block != NULL)
 306                 kref_put(&block->b_count, nlmsvc_free_block);
 307 }
 308
 309 /*
 310  * Loop over all blocks and delete blocks held by
 311  * a matching host.
 312  */
 313 void nlmsvc_traverse_blocks(struct nlm_host *host,
 314                         struct nlm_file *file,
 315                         nlm_host_match_fn_t match)
 316 {
 317         struct nlm_block *block, *next;
 318
 319 restart:
 320         mutex_lock(&file->f_mutex);
 321         list_for_each_entry_safe(block, next, &file->f_blocks, b_flist) {
 322                 if (!match(block->b_host, host))
 323                         continue;
 324                 /* Do not destroy blocks that are not on
 325                  * the global retry list - why? */
 326                 if (list_empty(&block->b_list))
 327                         continue;
 328                 kref_get(&block->b_count);
 329                 mutex_unlock(&file->f_mutex);
 330                 nlmsvc_unlink_block(block);
 331                 nlmsvc_release_block(block);
 332                 goto restart;
 333         }
 334         mutex_unlock(&file->f_mutex);
 335 }
 336
 337 /*
 338  * Initialize arguments for GRANTED call. The nlm_rqst structure
 339  * has been cleared already.
 340  */
 341 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
 342 {
 343         locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
 344         memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
 345         call->a_args.lock.caller = utsname()->nodename;
 346         call->a_args.lock.oh.len = lock->oh.len;
 347
 348         /* set default data area */
 349         call->a_args.lock.oh.data = call->a_owner;
 350         call->a_args.lock.svid = lock->fl.fl_pid;
 351
 352         if (lock->oh.len > NLMCLNT_OHSIZE) {
 353                 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
 354                 if (!data)
 355                         return 0;
 356                 call->a_args.lock.oh.data = (u8 *) data;
 357         }
 358
 359         memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
 360         return 1;
 361 }
 362
 363 static void nlmsvc_freegrantargs(struct nlm_rqst *call)
 364 {
 365         if (call->a_args.lock.oh.data != call->a_owner)
 366                 kfree(call->a_args.lock.oh.data);
 367
 368         locks_release_private(&call->a_args.lock.fl);
 369 }
 370
 371 /*
 372  * Deferred lock request handling for non-blocking lock
 373  */
 374 static __be32
 375 nlmsvc_defer_lock_rqst(struct svc_rqst *rqstp, struct nlm_block *block)
 376 {
 377         __be32 status = nlm_lck_denied_nolocks;
 378
 379         block->b_flags |= B_QUEUED;
 380
 381         nlmsvc_insert_block(block, NLM_TIMEOUT);
 382
 383         block->b_cache_req = &rqstp->rq_chandle;
 384         if (rqstp->rq_chandle.defer) {
 385                 block->b_deferred_req =
 386                         rqstp->rq_chandle.defer(block->b_cache_req);
 387                 if (block->b_deferred_req != NULL)
 388                         status = nlm_drop_reply;
 389         }
 390         dprintk("lockd: nlmsvc_defer_lock_rqst block %p flags %d status %d\n",
 391                 block, block->b_flags, ntohl(status));
 392
 393         return status;
 394 }
 395
 396 /*
 397  * Attempt to establish a lock, and if it can't be granted, block it
 398  * if required.
 399  */
 400 __be32
 401 nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
 402             struct nlm_host *host, struct nlm_lock *lock, int wait,
 403             struct nlm_cookie *cookie, int reclaim)
 404 {
 405         struct nlm_block        *block = NULL;
 406         int                     error;
 407         __be32                  ret;
 408
 409         dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
 410                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
 411                                 file->f_file->f_path.dentry->d_inode->i_ino,
 412                                 lock->fl.fl_type, lock->fl.fl_pid,
 413                                 (long long)lock->fl.fl_start,
 414                                 (long long)lock->fl.fl_end,
 415                                 wait);
 416
 417         /* Lock file against concurrent access */
 418         mutex_lock(&file->f_mutex);
 419         /* Get existing block (in case client is busy-waiting)
 420          * or create new block
 421          */
 422         block = nlmsvc_lookup_block(file, lock);
 423         if (block == NULL) {
 424                 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
 425                 ret = nlm_lck_denied_nolocks;
 426                 if (block == NULL)
 427                         goto out;
 428                 lock = &block->b_call->a_args.lock;
 429         } else
 430                 lock->fl.fl_flags &= ~FL_SLEEP;
 431
 432         if (block->b_flags & B_QUEUED) {
 433                 dprintk("lockd: nlmsvc_lock deferred block %p flags %d\n",
 434                                                         block, block->b_flags);
 435                 if (block->b_granted) {
 436                         nlmsvc_unlink_block(block);
 437                         ret = nlm_granted;
 438                         goto out;
 439                 }
 440                 if (block->b_flags & B_TIMED_OUT) {
 441                         nlmsvc_unlink_block(block);
 442                         ret = nlm_lck_denied;
 443                         goto out;
 444                 }
 445                 ret = nlm_drop_reply;
 446                 goto out;
 447         }
 448
 449         if (locks_in_grace(SVC_NET(rqstp)) && !reclaim) {
 450                 ret = nlm_lck_denied_grace_period;
 451                 goto out;
 452         }
 453         if (reclaim && !locks_in_grace(SVC_NET(rqstp))) {
 454                 ret = nlm_lck_denied_grace_period;
 455                 goto out;
 456         }
 457
 458         if (!wait)
 459                 lock->fl.fl_flags &= ~FL_SLEEP;
 460         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
 461         lock->fl.fl_flags &= ~FL_SLEEP;
 462
 463         dprintk("lockd: vfs_lock_file returned %d\n", error);
 464         switch (error) {
 465                 case 0:
 466                         ret = nlm_granted;
 467                         goto out;
 468                 case -EAGAIN:
 469                         /*
 470                          * If this is a blocking request for an
 471                          * already pending lock request then we need
 472                          * to put it back on lockd's block list
 473                          */
 474                         if (wait)
 475                                 break;
 476                         ret = nlm_lck_denied;
 477                         goto out;
 478                 case FILE_LOCK_DEFERRED:
 479                         if (wait)
 480                                 break;
 481                         /* Filesystem lock operation is in progress
 482                            Add it to the queue waiting for callback */
 483                         ret = nlmsvc_defer_lock_rqst(rqstp, block);
 484                         goto out;
 485                 case -EDEADLK:
 486                         ret = nlm_deadlock;
 487                         goto out;
 488                 default:                        /* includes ENOLCK */
 489                         ret = nlm_lck_denied_nolocks;
 490                         goto out;
 491         }
 492
 493         ret = nlm_lck_blocked;
 494
 495         /* Append to list of blocked */
 496         nlmsvc_insert_block(block, NLM_NEVER);
 497 out:
 498         mutex_unlock(&file->f_mutex);
 499         nlmsvc_release_block(block);
 500         dprintk("lockd: nlmsvc_lock returned %u\n", ret);
 501         return ret;
 502 }
 503
 504 /*
 505  * Test for presence of a conflicting lock.
 506  */
 507 __be32
 508 nlmsvc_testlock(struct svc_rqst *rqstp, struct nlm_file *file,
 509                 struct nlm_host *host, struct nlm_lock *lock,
 510                 struct nlm_lock *conflock, struct nlm_cookie *cookie)
 511 {
 512         struct nlm_block        *block = NULL;
 513         int                     error;
 514         __be32                  ret;
 515
 516         dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
 517                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
 518                                 file->f_file->f_path.dentry->d_inode->i_ino,
 519                                 lock->fl.fl_type,
 520                                 (long long)lock->fl.fl_start,
 521                                 (long long)lock->fl.fl_end);
 522
 523         /* Get existing block (in case client is busy-waiting) */
 524         block = nlmsvc_lookup_block(file, lock);
 525
 526         if (block == NULL) {
 527                 struct file_lock *conf = kzalloc(sizeof(*conf), GFP_KERNEL);
 528
 529                 if (conf == NULL)
 530                         return nlm_granted;
 531                 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
 532                 if (block == NULL) {
 533                         kfree(conf);
 534                         return nlm_granted;
 535                 }
 536                 block->b_fl = conf;
 537         }
 538         if (block->b_flags & B_QUEUED) {
 539                 dprintk("lockd: nlmsvc_testlock deferred block %p flags %d fl %p\n",
 540                         block, block->b_flags, block->b_fl);
 541                 if (block->b_flags & B_TIMED_OUT) {
 542                         nlmsvc_unlink_block(block);
 543                         ret = nlm_lck_denied;
 544                         goto out;
 545                 }
 546                 if (block->b_flags & B_GOT_CALLBACK) {
 547                         nlmsvc_unlink_block(block);
 548                         if (block->b_fl != NULL
 549                                         && block->b_fl->fl_type != F_UNLCK) {
 550                                 lock->fl = *block->b_fl;
 551                                 goto conf_lock;
 552                         } else {
 553                                 ret = nlm_granted;
 554                                 goto out;
 555                         }
 556                 }
 557                 ret = nlm_drop_reply;
 558                 goto out;
 559         }
 560
 561         if (locks_in_grace(SVC_NET(rqstp))) {
 562                 ret = nlm_lck_denied_grace_period;
 563                 goto out;
 564         }
 565         error = vfs_test_lock(file->f_file, &lock->fl);
 566         if (error == FILE_LOCK_DEFERRED) {
 567                 ret = nlmsvc_defer_lock_rqst(rqstp, block);
 568                 goto out;
 569         }
 570         if (error) {
 571                 ret = nlm_lck_denied_nolocks;
 572                 goto out;
 573         }
 574         if (lock->fl.fl_type == F_UNLCK) {
 575                 ret = nlm_granted;
 576                 goto out;
 577         }
 578
 579 conf_lock:
 580         dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
 581                 lock->fl.fl_type, (long long)lock->fl.fl_start,
 582                 (long long)lock->fl.fl_end);
 583         conflock->caller = "somehost";  /* FIXME */
 584         conflock->len = strlen(conflock->caller);
 585         conflock->oh.len = 0;           /* don't return OH info */
 586         conflock->svid = lock->fl.fl_pid;
 587         conflock->fl.fl_type = lock->fl.fl_type;
 588         conflock->fl.fl_start = lock->fl.fl_start;
 589         conflock->fl.fl_end = lock->fl.fl_end;
 590         ret = nlm_lck_denied;
 591 out:
 592         if (block)
 593                 nlmsvc_release_block(block);
 594         return ret;
 595 }
 596
 597 /*
 598  * Remove a lock.
 599  * This implies a CANCEL call: We send a GRANT_MSG, the client replies
 600  * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
 601  * afterwards. In this case the block will still be there, and hence
 602  * must be removed.
 603  */
 604 __be32
 605 nlmsvc_unlock(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
 606 {
 607         int     error;
 608
 609         dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
 610                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
 611                                 file->f_file->f_path.dentry->d_inode->i_ino,
 612                                 lock->fl.fl_pid,
 613                                 (long long)lock->fl.fl_start,
 614                                 (long long)lock->fl.fl_end);
 615
 616         /* First, cancel any lock that might be there */
 617         nlmsvc_cancel_blocked(net, file, lock);
 618
 619         lock->fl.fl_type = F_UNLCK;
 620         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
 621
 622         return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
 623 }
 624
 625 /*
 626  * Cancel a previously blocked request.
 627  *
 628  * A cancel request always overrides any grant that may currently
 629  * be in progress.
 630  * The calling procedure must check whether the file can be closed.
 631  */
 632 __be32
 633 nlmsvc_cancel_blocked(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
 634 {
 635         struct nlm_block        *block;
 636         int status = 0;
 637
 638         dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
 639                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
 640                                 file->f_file->f_path.dentry->d_inode->i_ino,
 641                                 lock->fl.fl_pid,
 642                                 (long long)lock->fl.fl_start,
 643                                 (long long)lock->fl.fl_end);
 644
 645         if (locks_in_grace(net))
 646                 return nlm_lck_denied_grace_period;
 647
 648         mutex_lock(&file->f_mutex);
 649         block = nlmsvc_lookup_block(file, lock);
 650         mutex_unlock(&file->f_mutex);
 651         if (block != NULL) {
 652                 vfs_cancel_lock(block->b_file->f_file,
 653                                 &block->b_call->a_args.lock.fl);
 654                 status = nlmsvc_unlink_block(block);
 655                 nlmsvc_release_block(block);
 656         }
 657         return status ? nlm_lck_denied : nlm_granted;
 658 }
 659
 660 /*
 661  * This is a callback from the filesystem for VFS file lock requests.
 662  * It will be used if lm_grant is defined and the filesystem can not
 663  * respond to the request immediately.
 664  * For GETLK request it will copy the reply to the nlm_block.
 665  * For SETLK or SETLKW request it will get the local posix lock.
 666  * In all cases it will move the block to the head of nlm_blocked q where
 667  * nlmsvc_retry_blocked() can send back a reply for SETLKW or revisit the
 668  * deferred rpc for GETLK and SETLK.
 669  */
 670 static void
 671 nlmsvc_update_deferred_block(struct nlm_block *block, struct file_lock *conf,
 672                              int result)
 673 {
 674         block->b_flags |= B_GOT_CALLBACK;
 675         if (result == 0)
 676                 block->b_granted = 1;
 677         else
 678                 block->b_flags |= B_TIMED_OUT;
 679         if (conf) {
 680                 if (block->b_fl)
 681                         __locks_copy_lock(block->b_fl, conf);
 682         }
 683 }
 684
 685 static int nlmsvc_grant_deferred(struct file_lock *fl, struct file_lock *conf,
 686                                         int result)
 687 {
 688         struct nlm_block *block;
 689         int rc = -ENOENT;
 690
 691         spin_lock(&nlm_blocked_lock);
 692         list_for_each_entry(block, &nlm_blocked, b_list) {
 693                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
 694                         dprintk("lockd: nlmsvc_notify_blocked block %p flags %d\n",
 695                                                         block, block->b_flags);
 696                         if (block->b_flags & B_QUEUED) {
 697                                 if (block->b_flags & B_TIMED_OUT) {
 698                                         rc = -ENOLCK;
 699                                         break;
 700                                 }
 701                                 nlmsvc_update_deferred_block(block, conf, result);
 702                         } else if (result == 0)
 703                                 block->b_granted = 1;
 704
 705                         nlmsvc_insert_block_locked(block, 0);
 706                         svc_wake_up(block->b_daemon);
 707                         rc = 0;
 708                         break;
 709                 }
 710         }
 711         spin_unlock(&nlm_blocked_lock);
 712         if (rc == -ENOENT)
 713                 printk(KERN_WARNING "lockd: grant for unknown block\n");
 714         return rc;
 715 }
 716
 717 /*
 718  * Unblock a blocked lock request. This is a callback invoked from the
 719  * VFS layer when a lock on which we blocked is removed.
 720  *
 721  * This function doesn't grant the blocked lock instantly, but rather moves
 722  * the block to the head of nlm_blocked where it can be picked up by lockd.
 723  */
 724 static void
 725 nlmsvc_notify_blocked(struct file_lock *fl)
 726 {
 727         struct nlm_block        *block;
 728
 729         dprintk("lockd: VFS unblock notification for block %p\n", fl);
 730         spin_lock(&nlm_blocked_lock);
 731         list_for_each_entry(block, &nlm_blocked, b_list) {
 732                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
 733                         nlmsvc_insert_block_locked(block, 0);
 734                         spin_unlock(&nlm_blocked_lock);
 735                         svc_wake_up(block->b_daemon);
 736                         return;
 737                 }
 738         }
 739         spin_unlock(&nlm_blocked_lock);
 740         printk(KERN_WARNING "lockd: notification for unknown block!\n");
 741 }
 742
 743 static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
 744 {
 745         return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
 746 }
 747
 748 const struct lock_manager_operations nlmsvc_lock_operations = {
 749         .lm_compare_owner = nlmsvc_same_owner,
 750         .lm_notify = nlmsvc_notify_blocked,
 751         .lm_grant = nlmsvc_grant_deferred,
 752 };
 753
 754 /*
 755  * Try to claim a lock that was previously blocked.
 756  *
 757  * Note that we use both the RPC_GRANTED_MSG call _and_ an async
 758  * RPC thread when notifying the client. This seems like overkill...
 759  * Here's why:
 760  *  -   we don't want to use a synchronous RPC thread, otherwise
 761  *      we might find ourselves hanging on a dead portmapper.
 762  *  -   Some lockd implementations (e.g. HP) don't react to
 763  *      RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
 764  */
 765 static void
 766 nlmsvc_grant_blocked(struct nlm_block *block)
 767 {
 768         struct nlm_file         *file = block->b_file;
 769         struct nlm_lock         *lock = &block->b_call->a_args.lock;
 770         int                     error;
 771
 772         dprintk("lockd: grant blocked lock %p\n", block);
 773
 774         kref_get(&block->b_count);
 775
 776         /* Unlink block request from list */
 777         nlmsvc_unlink_block(block);
 778
 779         /* If b_granted is true this means we've been here before.
 780          * Just retry the grant callback, possibly refreshing the RPC
 781          * binding */
 782         if (block->b_granted) {
 783                 nlm_rebind_host(block->b_host);
 784                 goto callback;
 785         }
 786
 787         /* Try the lock operation again */
 788         lock->fl.fl_flags |= FL_SLEEP;
 789         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
 790         lock->fl.fl_flags &= ~FL_SLEEP;
 791
 792         switch (error) {
 793         case 0:
 794                 break;
 795         case FILE_LOCK_DEFERRED:
 796                 dprintk("lockd: lock still blocked error %d\n", error);
 797                 nlmsvc_insert_block(block, NLM_NEVER);
 798                 nlmsvc_release_block(block);
 799                 return;
 800         default:
 801                 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
 802                                 -error, __func__);
 803                 nlmsvc_insert_block(block, 10 * HZ);
 804                 nlmsvc_release_block(block);
 805                 return;
 806         }
 807
 808 callback:
 809         /* Lock was granted by VFS. */
 810         dprintk("lockd: GRANTing blocked lock.\n");
 811         block->b_granted = 1;
 812
 813         /* keep block on the list, but don't reattempt until the RPC
 814          * completes or the submission fails
 815          */
 816         nlmsvc_insert_block(block, NLM_NEVER);
 817
 818         /* Call the client -- use a soft RPC task since nlmsvc_retry_blocked
 819          * will queue up a new one if this one times out
 820          */
 821         error = nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
 822                                 &nlmsvc_grant_ops);
 823
 824         /* RPC submission failed, wait a bit and retry */
 825         if (error < 0)
 826                 nlmsvc_insert_block(block, 10 * HZ);
 827 }
 828
 829 /*
 830  * This is the callback from the RPC layer when the NLM_GRANTED_MSG
 831  * RPC call has succeeded or timed out.
 832  * Like all RPC callbacks, it is invoked by the rpciod process, so it
 833  * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
 834  * chain once more in order to have it removed by lockd itself (which can
 835  * then sleep on the file semaphore without disrupting e.g. the nfs client).
 836  */
 837 static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
 838 {
 839         struct nlm_rqst         *call = data;
 840         struct nlm_block        *block = call->a_block;
 841         unsigned long           timeout;
 842
 843         dprintk("lockd: GRANT_MSG RPC callback\n");
 844
 845         spin_lock(&nlm_blocked_lock);
 846         /* if the block is not on a list at this point then it has
 847          * been invalidated. Don't try to requeue it.
 848          *
 849          * FIXME: it's possible that the block is removed from the list
 850          * after this check but before the nlmsvc_insert_block. In that
 851          * case it will be added back. Perhaps we need better locking
 852          * for nlm_blocked?
 853          */
 854         if (list_empty(&block->b_list))
 855                 goto out;
 856
 857         /* Technically, we should down the file semaphore here. Since we
 858          * move the block towards the head of the queue only, no harm
 859          * can be done, though. */
 860         if (task->tk_status < 0) {
 861                 /* RPC error: Re-insert for retransmission */
 862                 timeout = 10 * HZ;
 863         } else {
 864                 /* Call was successful, now wait for client callback */
 865                 timeout = 60 * HZ;
 866         }
 867         nlmsvc_insert_block_locked(block, timeout);
 868         svc_wake_up(block->b_daemon);
 869 out:
 870         spin_unlock(&nlm_blocked_lock);
 871 }
 872
 873 /*
 874  * FIXME: nlmsvc_release_block() grabs a mutex.  This is not allowed for an
 875  * .rpc_release rpc_call_op
 876  */
 877 static void nlmsvc_grant_release(void *data)
 878 {
 879         struct nlm_rqst         *call = data;
 880         nlmsvc_release_block(call->a_block);
 881 }
 882
 883 static const struct rpc_call_ops nlmsvc_grant_ops = {
 884         .rpc_call_done = nlmsvc_grant_callback,
 885         .rpc_release = nlmsvc_grant_release,
 886 };
 887
 888 /*
 889  * We received a GRANT_RES callback. Try to find the corresponding
 890  * block.
 891  */
 892 void
 893 nlmsvc_grant_reply(struct nlm_cookie *cookie, __be32 status)
 894 {
 895         struct nlm_block        *block;
 896
 897         dprintk("grant_reply: looking for cookie %x, s=%d \n",
 898                 *(unsigned int *)(cookie->data), status);
 899         if (!(block = nlmsvc_find_block(cookie)))
 900                 return;
 901
 902         if (block) {
 903                 if (status == nlm_lck_denied_grace_period) {
 904                         /* Try again in a couple of seconds */
 905                         nlmsvc_insert_block(block, 10 * HZ);
 906                 } else {
 907                         /* Lock is now held by client, or has been rejected.
 908                          * In both cases, the block should be removed. */
 909                         nlmsvc_unlink_block(block);
 910                 }
 911         }
 912         nlmsvc_release_block(block);
 913 }
 914
 915 /* Helper function to handle retry of a deferred block.
 916  * If it is a blocking lock, call grant_blocked.
 917  * For a non-blocking lock or test lock, revisit the request.
 918  */
 919 static void
 920 retry_deferred_block(struct nlm_block *block)
 921 {
 922         if (!(block->b_flags & B_GOT_CALLBACK))
 923                 block->b_flags |= B_TIMED_OUT;
 924         nlmsvc_insert_block(block, NLM_TIMEOUT);
 925         dprintk("revisit block %p flags %d\n",  block, block->b_flags);
 926         if (block->b_deferred_req) {
 927                 block->b_deferred_req->revisit(block->b_deferred_req, 0);
 928                 block->b_deferred_req = NULL;
 929         }
 930 }
 931
 932 /*
 933  * Retry all blocked locks that have been notified. This is where lockd
 934  * picks up locks that can be granted, or grant notifications that must
 935  * be retransmitted.
 936  */
 937 unsigned long
 938 nlmsvc_retry_blocked(void)
 939 {
 940         unsigned long   timeout = MAX_SCHEDULE_TIMEOUT;
 941         struct nlm_block *block;
 942
 943         while (!list_empty(&nlm_blocked) && !kthread_should_stop()) {
 944                 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
 945
 946                 if (block->b_when == NLM_NEVER)
 947                         break;
 948                 if (time_after(block->b_when, jiffies)) {
 949                         timeout = block->b_when - jiffies;
 950                         break;
 951                 }
 952
 953                 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
 954                         block, block->b_when);
 955                 if (block->b_flags & B_QUEUED) {
 956                         dprintk("nlmsvc_retry_blocked delete block (%p, granted=%d, flags=%d)\n",
 957                                 block, block->b_granted, block->b_flags);
 958                         retry_deferred_block(block);
 959                 } else
 960                         nlmsvc_grant_blocked(block);
 961         }
 962
 963         return timeout;
 964 }