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/errno.h>
  25 #include <linux/kernel.h>
  26 #include <linux/sched.h>
  27 #include <linux/smp_lock.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
  33 #define NLMDBG_FACILITY         NLMDBG_SVCLOCK
  34
  35 #ifdef CONFIG_LOCKD_V4
  36 #define nlm_deadlock    nlm4_deadlock
  37 #else
  38 #define nlm_deadlock    nlm_lck_denied
  39 #endif
  40
  41 static void nlmsvc_release_block(struct nlm_block *block);
  42 static void     nlmsvc_insert_block(struct nlm_block *block, unsigned long);
  43 static void     nlmsvc_remove_block(struct nlm_block *block);
  44
  45 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock);
  46 static void nlmsvc_freegrantargs(struct nlm_rqst *call);
  47 static const struct rpc_call_ops nlmsvc_grant_ops;
  48
  49 /*
  50  * The list of blocked locks to retry
  51  */
  52 static LIST_HEAD(nlm_blocked);
  53
  54 /*
  55  * Insert a blocked lock into the global list
  56  */
  57 static void
  58 nlmsvc_insert_block(struct nlm_block *block, unsigned long when)
  59 {
  60         struct nlm_block *b;
  61         struct list_head *pos;
  62
  63         dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block, when);
  64         if (list_empty(&block->b_list)) {
  65                 kref_get(&block->b_count);
  66         } else {
  67                 list_del_init(&block->b_list);
  68         }
  69
  70         pos = &nlm_blocked;
  71         if (when != NLM_NEVER) {
  72                 if ((when += jiffies) == NLM_NEVER)
  73                         when ++;
  74                 list_for_each(pos, &nlm_blocked) {
  75                         b = list_entry(pos, struct nlm_block, b_list);
  76                         if (time_after(b->b_when,when) || b->b_when == NLM_NEVER)
  77                                 break;
  78                 }
  79                 /* On normal exit from the loop, pos == &nlm_blocked,
  80                  * so we will be adding to the end of the list - good
  81                  */
  82         }
  83
  84         list_add_tail(&block->b_list, pos);
  85         block->b_when = when;
  86 }
  87
  88 /*
  89  * Remove a block from the global list
  90  */
  91 static inline void
  92 nlmsvc_remove_block(struct nlm_block *block)
  93 {
  94         if (!list_empty(&block->b_list)) {
  95                 list_del_init(&block->b_list);
  96                 nlmsvc_release_block(block);
  97         }
  98 }
  99
 100 /*
 101  * Find a block for a given lock
 102  */
 103 static struct nlm_block *
 104 nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock)
 105 {
 106         struct nlm_block        *block;
 107         struct file_lock        *fl;
 108
 109         dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
 110                                 file, lock->fl.fl_pid,
 111                                 (long long)lock->fl.fl_start,
 112                                 (long long)lock->fl.fl_end, lock->fl.fl_type);
 113         list_for_each_entry(block, &nlm_blocked, b_list) {
 114                 fl = &block->b_call->a_args.lock.fl;
 115                 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
 116                                 block->b_file, fl->fl_pid,
 117                                 (long long)fl->fl_start,
 118                                 (long long)fl->fl_end, fl->fl_type,
 119                                 nlmdbg_cookie2a(&block->b_call->a_args.cookie));
 120                 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
 121                         kref_get(&block->b_count);
 122                         return block;
 123                 }
 124         }
 125
 126         return NULL;
 127 }
 128
 129 static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
 130 {
 131         if(a->len != b->len)
 132                 return 0;
 133         if(memcmp(a->data,b->data,a->len))
 134                 return 0;
 135         return 1;
 136 }
 137
 138 /*
 139  * Find a block with a given NLM cookie.
 140  */
 141 static inline struct nlm_block *
 142 nlmsvc_find_block(struct nlm_cookie *cookie)
 143 {
 144         struct nlm_block *block;
 145
 146         list_for_each_entry(block, &nlm_blocked, b_list) {
 147                 if (nlm_cookie_match(&block->b_call->a_args.cookie,cookie))
 148                         goto found;
 149         }
 150
 151         return NULL;
 152
 153 found:
 154         dprintk("nlmsvc_find_block(%s): block=%p\n", nlmdbg_cookie2a(cookie), block);
 155         kref_get(&block->b_count);
 156         return block;
 157 }
 158
 159 /*
 160  * Create a block and initialize it.
 161  *
 162  * Note: we explicitly set the cookie of the grant reply to that of
 163  * the blocked lock request. The spec explicitly mentions that the client
 164  * should _not_ rely on the callback containing the same cookie as the
 165  * request, but (as I found out later) that's because some implementations
 166  * do just this. Never mind the standards comittees, they support our
 167  * logging industries.
 168  *
 169  * 10 years later: I hope we can safely ignore these old and broken
 170  * clients by now. Let's fix this so we can uniquely identify an incoming
 171  * GRANTED_RES message by cookie, without having to rely on the client's IP
 172  * address. --okir
 173  */
 174 static inline struct nlm_block *
 175 nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_file *file,
 176                 struct nlm_lock *lock, struct nlm_cookie *cookie)
 177 {
 178         struct nlm_block        *block;
 179         struct nlm_host         *host;
 180         struct nlm_rqst         *call = NULL;
 181
 182         /* Create host handle for callback */
 183         host = nlmsvc_lookup_host(rqstp, lock->caller, lock->len);
 184         if (host == NULL)
 185                 return NULL;
 186
 187         call = nlm_alloc_call(host);
 188         if (call == NULL)
 189                 return NULL;
 190
 191         /* Allocate memory for block, and initialize arguments */
 192         block = kzalloc(sizeof(*block), GFP_KERNEL);
 193         if (block == NULL)
 194                 goto failed;
 195         kref_init(&block->b_count);
 196         INIT_LIST_HEAD(&block->b_list);
 197         INIT_LIST_HEAD(&block->b_flist);
 198
 199         if (!nlmsvc_setgrantargs(call, lock))
 200                 goto failed_free;
 201
 202         /* Set notifier function for VFS, and init args */
 203         call->a_args.lock.fl.fl_flags |= FL_SLEEP;
 204         call->a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
 205         nlmclnt_next_cookie(&call->a_args.cookie);
 206
 207         dprintk("lockd: created block %p...\n", block);
 208
 209         /* Create and initialize the block */
 210         block->b_daemon = rqstp->rq_server;
 211         block->b_host   = host;
 212         block->b_file   = file;
 213         block->b_fl = NULL;
 214         file->f_count++;
 215
 216         /* Add to file's list of blocks */
 217         list_add(&block->b_flist, &file->f_blocks);
 218
 219         /* Set up RPC arguments for callback */
 220         block->b_call = call;
 221         call->a_flags   = RPC_TASK_ASYNC;
 222         call->a_block = block;
 223
 224         return block;
 225
 226 failed_free:
 227         kfree(block);
 228 failed:
 229         nlm_release_call(call);
 230         return NULL;
 231 }
 232
 233 /*
 234  * Delete a block. If the lock was cancelled or the grant callback
 235  * failed, unlock is set to 1.
 236  * It is the caller's responsibility to check whether the file
 237  * can be closed hereafter.
 238  */
 239 static int nlmsvc_unlink_block(struct nlm_block *block)
 240 {
 241         int status;
 242         dprintk("lockd: unlinking block %p...\n", block);
 243
 244         /* Remove block from list */
 245         status = posix_unblock_lock(block->b_file->f_file, &block->b_call->a_args.lock.fl);
 246         nlmsvc_remove_block(block);
 247         return status;
 248 }
 249
 250 static void nlmsvc_free_block(struct kref *kref)
 251 {
 252         struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
 253         struct nlm_file         *file = block->b_file;
 254
 255         dprintk("lockd: freeing block %p...\n", block);
 256
 257         /* Remove block from file's list of blocks */
 258         mutex_lock(&file->f_mutex);
 259         list_del_init(&block->b_flist);
 260         mutex_unlock(&file->f_mutex);
 261
 262         nlmsvc_freegrantargs(block->b_call);
 263         nlm_release_call(block->b_call);
 264         nlm_release_file(block->b_file);
 265         kfree(block->b_fl);
 266         kfree(block);
 267 }
 268
 269 static void nlmsvc_release_block(struct nlm_block *block)
 270 {
 271         if (block != NULL)
 272                 kref_put(&block->b_count, nlmsvc_free_block);
 273 }
 274
 275 /*
 276  * Loop over all blocks and delete blocks held by
 277  * a matching host.
 278  */
 279 void nlmsvc_traverse_blocks(struct nlm_host *host,
 280                         struct nlm_file *file,
 281                         nlm_host_match_fn_t match)
 282 {
 283         struct nlm_block *block, *next;
 284
 285 restart:
 286         mutex_lock(&file->f_mutex);
 287         list_for_each_entry_safe(block, next, &file->f_blocks, b_flist) {
 288                 if (!match(block->b_host, host))
 289                         continue;
 290                 /* Do not destroy blocks that are not on
 291                  * the global retry list - why? */
 292                 if (list_empty(&block->b_list))
 293                         continue;
 294                 kref_get(&block->b_count);
 295                 mutex_unlock(&file->f_mutex);
 296                 nlmsvc_unlink_block(block);
 297                 nlmsvc_release_block(block);
 298                 goto restart;
 299         }
 300         mutex_unlock(&file->f_mutex);
 301 }
 302
 303 /*
 304  * Initialize arguments for GRANTED call. The nlm_rqst structure
 305  * has been cleared already.
 306  */
 307 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
 308 {
 309         locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
 310         memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
 311         call->a_args.lock.caller = utsname()->nodename;
 312         call->a_args.lock.oh.len = lock->oh.len;
 313
 314         /* set default data area */
 315         call->a_args.lock.oh.data = call->a_owner;
 316         call->a_args.lock.svid = lock->fl.fl_pid;
 317
 318         if (lock->oh.len > NLMCLNT_OHSIZE) {
 319                 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
 320                 if (!data)
 321                         return 0;
 322                 call->a_args.lock.oh.data = (u8 *) data;
 323         }
 324
 325         memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
 326         return 1;
 327 }
 328
 329 static void nlmsvc_freegrantargs(struct nlm_rqst *call)
 330 {
 331         if (call->a_args.lock.oh.data != call->a_owner)
 332                 kfree(call->a_args.lock.oh.data);
 333 }
 334
 335 /*
 336  * Deferred lock request handling for non-blocking lock
 337  */
 338 static u32
 339 nlmsvc_defer_lock_rqst(struct svc_rqst *rqstp, struct nlm_block *block)
 340 {
 341         u32 status = nlm_lck_denied_nolocks;
 342
 343         block->b_flags |= B_QUEUED;
 344
 345         nlmsvc_insert_block(block, NLM_TIMEOUT);
 346
 347         block->b_cache_req = &rqstp->rq_chandle;
 348         if (rqstp->rq_chandle.defer) {
 349                 block->b_deferred_req =
 350                         rqstp->rq_chandle.defer(block->b_cache_req);
 351                 if (block->b_deferred_req != NULL)
 352                         status = nlm_drop_reply;
 353         }
 354         dprintk("lockd: nlmsvc_defer_lock_rqst block %p flags %d status %d\n",
 355                 block, block->b_flags, status);
 356
 357         return status;
 358 }
 359
 360 /*
 361  * Attempt to establish a lock, and if it can't be granted, block it
 362  * if required.
 363  */
 364 __be32
 365 nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
 366                         struct nlm_lock *lock, int wait, struct nlm_cookie *cookie)
 367 {
 368         struct nlm_block        *block = NULL;
 369         int                     error;
 370         __be32                  ret;
 371
 372         dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
 373                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
 374                                 file->f_file->f_path.dentry->d_inode->i_ino,
 375                                 lock->fl.fl_type, lock->fl.fl_pid,
 376                                 (long long)lock->fl.fl_start,
 377                                 (long long)lock->fl.fl_end,
 378                                 wait);
 379
 380
 381         /* Lock file against concurrent access */
 382         mutex_lock(&file->f_mutex);
 383         /* Get existing block (in case client is busy-waiting)
 384          * or create new block
 385          */
 386         block = nlmsvc_lookup_block(file, lock);
 387         if (block == NULL) {
 388                 block = nlmsvc_create_block(rqstp, file, lock, cookie);
 389                 ret = nlm_lck_denied_nolocks;
 390                 if (block == NULL)
 391                         goto out;
 392                 lock = &block->b_call->a_args.lock;
 393         } else
 394                 lock->fl.fl_flags &= ~FL_SLEEP;
 395
 396         if (block->b_flags & B_QUEUED) {
 397                 dprintk("lockd: nlmsvc_lock deferred block %p flags %d\n",
 398                                                         block, block->b_flags);
 399                 if (block->b_granted) {
 400                         nlmsvc_unlink_block(block);
 401                         ret = nlm_granted;
 402                         goto out;
 403                 }
 404                 if (block->b_flags & B_TIMED_OUT) {
 405                         nlmsvc_unlink_block(block);
 406                         ret = nlm_lck_denied;
 407                         goto out;
 408                 }
 409                 ret = nlm_drop_reply;
 410                 goto out;
 411         }
 412
 413         if (!wait)
 414                 lock->fl.fl_flags &= ~FL_SLEEP;
 415         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
 416         lock->fl.fl_flags &= ~FL_SLEEP;
 417
 418         dprintk("lockd: vfs_lock_file returned %d\n", error);
 419         switch(error) {
 420                 case 0:
 421                         ret = nlm_granted;
 422                         goto out;
 423                 case -EAGAIN:
 424                         ret = nlm_lck_denied;
 425                         break;
 426                 case -EINPROGRESS:
 427                         if (wait)
 428                                 break;
 429                         /* Filesystem lock operation is in progress
 430                            Add it to the queue waiting for callback */
 431                         ret = nlmsvc_defer_lock_rqst(rqstp, block);
 432                         goto out;
 433                 case -EDEADLK:
 434                         ret = nlm_deadlock;
 435                         goto out;
 436                 default:                        /* includes ENOLCK */
 437                         ret = nlm_lck_denied_nolocks;
 438                         goto out;
 439         }
 440
 441         ret = nlm_lck_denied;
 442         if (!wait)
 443                 goto out;
 444
 445         ret = nlm_lck_blocked;
 446
 447         /* Append to list of blocked */
 448         nlmsvc_insert_block(block, NLM_NEVER);
 449 out:
 450         mutex_unlock(&file->f_mutex);
 451         nlmsvc_release_block(block);
 452         dprintk("lockd: nlmsvc_lock returned %u\n", ret);
 453         return ret;
 454 }
 455
 456 /*
 457  * Test for presence of a conflicting lock.
 458  */
 459 __be32
 460 nlmsvc_testlock(struct svc_rqst *rqstp, struct nlm_file *file,
 461                 struct nlm_lock *lock, struct nlm_lock *conflock,
 462                 struct nlm_cookie *cookie)
 463 {
 464         struct nlm_block        *block = NULL;
 465         int                     error;
 466         __be32                  ret;
 467
 468         dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
 469                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
 470                                 file->f_file->f_path.dentry->d_inode->i_ino,
 471                                 lock->fl.fl_type,
 472                                 (long long)lock->fl.fl_start,
 473                                 (long long)lock->fl.fl_end);
 474
 475         /* Get existing block (in case client is busy-waiting) */
 476         block = nlmsvc_lookup_block(file, lock);
 477
 478         if (block == NULL) {
 479                 struct file_lock *conf = kzalloc(sizeof(*conf), GFP_KERNEL);
 480
 481                 if (conf == NULL)
 482                         return nlm_granted;
 483                 block = nlmsvc_create_block(rqstp, file, lock, cookie);
 484                 if (block == NULL) {
 485                         kfree(conf);
 486                         return nlm_granted;
 487                 }
 488                 block->b_fl = conf;
 489         }
 490         if (block->b_flags & B_QUEUED) {
 491                 dprintk("lockd: nlmsvc_testlock deferred block %p flags %d fl %p\n",
 492                         block, block->b_flags, block->b_fl);
 493                 if (block->b_flags & B_TIMED_OUT) {
 494                         nlmsvc_unlink_block(block);
 495                         return nlm_lck_denied;
 496                 }
 497                 if (block->b_flags & B_GOT_CALLBACK) {
 498                         if (block->b_fl != NULL
 499                                         && block->b_fl->fl_type != F_UNLCK) {
 500                                 lock->fl = *block->b_fl;
 501                                 goto conf_lock;
 502                         }
 503                         else {
 504                                 nlmsvc_unlink_block(block);
 505                                 return nlm_granted;
 506                         }
 507                 }
 508                 return nlm_drop_reply;
 509         }
 510
 511         error = vfs_test_lock(file->f_file, &lock->fl);
 512         if (error == -EINPROGRESS)
 513                 return nlmsvc_defer_lock_rqst(rqstp, block);
 514         if (error) {
 515                 ret = nlm_lck_denied_nolocks;
 516                 goto out;
 517         }
 518         if (lock->fl.fl_type == F_UNLCK) {
 519                 ret = nlm_granted;
 520                 goto out;
 521         }
 522
 523 conf_lock:
 524         dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
 525                 lock->fl.fl_type, (long long)lock->fl.fl_start,
 526                 (long long)lock->fl.fl_end);
 527         conflock->caller = "somehost";  /* FIXME */
 528         conflock->len = strlen(conflock->caller);
 529         conflock->oh.len = 0;           /* don't return OH info */
 530         conflock->svid = lock->fl.fl_pid;
 531         conflock->fl.fl_type = lock->fl.fl_type;
 532         conflock->fl.fl_start = lock->fl.fl_start;
 533         conflock->fl.fl_end = lock->fl.fl_end;
 534         ret = nlm_lck_denied;
 535 out:
 536         if (block)
 537                 nlmsvc_release_block(block);
 538         return ret;
 539 }
 540
 541 /*
 542  * Remove a lock.
 543  * This implies a CANCEL call: We send a GRANT_MSG, the client replies
 544  * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
 545  * afterwards. In this case the block will still be there, and hence
 546  * must be removed.
 547  */
 548 __be32
 549 nlmsvc_unlock(struct nlm_file *file, struct nlm_lock *lock)
 550 {
 551         int     error;
 552
 553         dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
 554                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
 555                                 file->f_file->f_path.dentry->d_inode->i_ino,
 556                                 lock->fl.fl_pid,
 557                                 (long long)lock->fl.fl_start,
 558                                 (long long)lock->fl.fl_end);
 559
 560         /* First, cancel any lock that might be there */
 561         nlmsvc_cancel_blocked(file, lock);
 562
 563         lock->fl.fl_type = F_UNLCK;
 564         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
 565
 566         return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
 567 }
 568
 569 /*
 570  * Cancel a previously blocked request.
 571  *
 572  * A cancel request always overrides any grant that may currently
 573  * be in progress.
 574  * The calling procedure must check whether the file can be closed.
 575  */
 576 __be32
 577 nlmsvc_cancel_blocked(struct nlm_file *file, struct nlm_lock *lock)
 578 {
 579         struct nlm_block        *block;
 580         int status = 0;
 581
 582         dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
 583                                 file->f_file->f_path.dentry->d_inode->i_sb->s_id,
 584                                 file->f_file->f_path.dentry->d_inode->i_ino,
 585                                 lock->fl.fl_pid,
 586                                 (long long)lock->fl.fl_start,
 587                                 (long long)lock->fl.fl_end);
 588
 589         mutex_lock(&file->f_mutex);
 590         block = nlmsvc_lookup_block(file, lock);
 591         mutex_unlock(&file->f_mutex);
 592         if (block != NULL) {
 593                 vfs_cancel_lock(block->b_file->f_file,
 594                                 &block->b_call->a_args.lock.fl);
 595                 status = nlmsvc_unlink_block(block);
 596                 nlmsvc_release_block(block);
 597         }
 598         return status ? nlm_lck_denied : nlm_granted;
 599 }
 600
 601 /*
 602  * This is a callback from the filesystem for VFS file lock requests.
 603  * It will be used if fl_grant is defined and the filesystem can not
 604  * respond to the request immediately.
 605  * For GETLK request it will copy the reply to the nlm_block.
 606  * For SETLK or SETLKW request it will get the local posix lock.
 607  * In all cases it will move the block to the head of nlm_blocked q where
 608  * nlmsvc_retry_blocked() can send back a reply for SETLKW or revisit the
 609  * deferred rpc for GETLK and SETLK.
 610  */
 611 static void
 612 nlmsvc_update_deferred_block(struct nlm_block *block, struct file_lock *conf,
 613                              int result)
 614 {
 615         block->b_flags |= B_GOT_CALLBACK;
 616         if (result == 0)
 617                 block->b_granted = 1;
 618         else
 619                 block->b_flags |= B_TIMED_OUT;
 620         if (conf) {
 621                 if (block->b_fl)
 622                         locks_copy_lock(block->b_fl, conf);
 623         }
 624 }
 625
 626 static int nlmsvc_grant_deferred(struct file_lock *fl, struct file_lock *conf,
 627                                         int result)
 628 {
 629         struct nlm_block *block;
 630         int rc = -ENOENT;
 631
 632         lock_kernel();
 633         list_for_each_entry(block, &nlm_blocked, b_list) {
 634                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
 635                         dprintk("lockd: nlmsvc_notify_blocked block %p flags %d\n",
 636                                                         block, block->b_flags);
 637                         if (block->b_flags & B_QUEUED) {
 638                                 if (block->b_flags & B_TIMED_OUT) {
 639                                         rc = -ENOLCK;
 640                                         break;
 641                                 }
 642                                 nlmsvc_update_deferred_block(block, conf, result);
 643                         } else if (result == 0)
 644                                 block->b_granted = 1;
 645
 646                         nlmsvc_insert_block(block, 0);
 647                         svc_wake_up(block->b_daemon);
 648                         rc = 0;
 649                         break;
 650                 }
 651         }
 652         unlock_kernel();
 653         if (rc == -ENOENT)
 654                 printk(KERN_WARNING "lockd: grant for unknown block\n");
 655         return rc;
 656 }
 657
 658 /*
 659  * Unblock a blocked lock request. This is a callback invoked from the
 660  * VFS layer when a lock on which we blocked is removed.
 661  *
 662  * This function doesn't grant the blocked lock instantly, but rather moves
 663  * the block to the head of nlm_blocked where it can be picked up by lockd.
 664  */
 665 static void
 666 nlmsvc_notify_blocked(struct file_lock *fl)
 667 {
 668         struct nlm_block        *block;
 669
 670         dprintk("lockd: VFS unblock notification for block %p\n", fl);
 671         list_for_each_entry(block, &nlm_blocked, b_list) {
 672                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
 673                         nlmsvc_insert_block(block, 0);
 674                         svc_wake_up(block->b_daemon);
 675                         return;
 676                 }
 677         }
 678
 679         printk(KERN_WARNING "lockd: notification for unknown block!\n");
 680 }
 681
 682 static int nlmsvc_same_owner(struct file_lock *fl1, struct file_lock *fl2)
 683 {
 684         return fl1->fl_owner == fl2->fl_owner && fl1->fl_pid == fl2->fl_pid;
 685 }
 686
 687 struct lock_manager_operations nlmsvc_lock_operations = {
 688         .fl_compare_owner = nlmsvc_same_owner,
 689         .fl_notify = nlmsvc_notify_blocked,
 690         .fl_grant = nlmsvc_grant_deferred,
 691 };
 692
 693 /*
 694  * Try to claim a lock that was previously blocked.
 695  *
 696  * Note that we use both the RPC_GRANTED_MSG call _and_ an async
 697  * RPC thread when notifying the client. This seems like overkill...
 698  * Here's why:
 699  *  -   we don't want to use a synchronous RPC thread, otherwise
 700  *      we might find ourselves hanging on a dead portmapper.
 701  *  -   Some lockd implementations (e.g. HP) don't react to
 702  *      RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
 703  */
 704 static void
 705 nlmsvc_grant_blocked(struct nlm_block *block)
 706 {
 707         struct nlm_file         *file = block->b_file;
 708         struct nlm_lock         *lock = &block->b_call->a_args.lock;
 709         int                     error;
 710
 711         dprintk("lockd: grant blocked lock %p\n", block);
 712
 713         kref_get(&block->b_count);
 714
 715         /* Unlink block request from list */
 716         nlmsvc_unlink_block(block);
 717
 718         /* If b_granted is true this means we've been here before.
 719          * Just retry the grant callback, possibly refreshing the RPC
 720          * binding */
 721         if (block->b_granted) {
 722                 nlm_rebind_host(block->b_host);
 723                 goto callback;
 724         }
 725
 726         /* Try the lock operation again */
 727         lock->fl.fl_flags |= FL_SLEEP;
 728         error = vfs_lock_file(file->f_file, F_SETLK, &lock->fl, NULL);
 729         lock->fl.fl_flags &= ~FL_SLEEP;
 730
 731         switch (error) {
 732         case 0:
 733                 break;
 734         case -EAGAIN:
 735         case -EINPROGRESS:
 736                 dprintk("lockd: lock still blocked error %d\n", error);
 737                 nlmsvc_insert_block(block, NLM_NEVER);
 738                 nlmsvc_release_block(block);
 739                 return;
 740         default:
 741                 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
 742                                 -error, __FUNCTION__);
 743                 nlmsvc_insert_block(block, 10 * HZ);
 744                 nlmsvc_release_block(block);
 745                 return;
 746         }
 747
 748 callback:
 749         /* Lock was granted by VFS. */
 750         dprintk("lockd: GRANTing blocked lock.\n");
 751         block->b_granted = 1;
 752
 753         /* Schedule next grant callback in 30 seconds */
 754         nlmsvc_insert_block(block, 30 * HZ);
 755
 756         /* Call the client */
 757         nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG, &nlmsvc_grant_ops);
 758 }
 759
 760 /*
 761  * This is the callback from the RPC layer when the NLM_GRANTED_MSG
 762  * RPC call has succeeded or timed out.
 763  * Like all RPC callbacks, it is invoked by the rpciod process, so it
 764  * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
 765  * chain once more in order to have it removed by lockd itself (which can
 766  * then sleep on the file semaphore without disrupting e.g. the nfs client).
 767  */
 768 static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
 769 {
 770         struct nlm_rqst         *call = data;
 771         struct nlm_block        *block = call->a_block;
 772         unsigned long           timeout;
 773
 774         dprintk("lockd: GRANT_MSG RPC callback\n");
 775
 776         /* Technically, we should down the file semaphore here. Since we
 777          * move the block towards the head of the queue only, no harm
 778          * can be done, though. */
 779         if (task->tk_status < 0) {
 780                 /* RPC error: Re-insert for retransmission */
 781                 timeout = 10 * HZ;
 782         } else {
 783                 /* Call was successful, now wait for client callback */
 784                 timeout = 60 * HZ;
 785         }
 786         nlmsvc_insert_block(block, timeout);
 787         svc_wake_up(block->b_daemon);
 788 }
 789
 790 static void nlmsvc_grant_release(void *data)
 791 {
 792         struct nlm_rqst         *call = data;
 793
 794         nlmsvc_release_block(call->a_block);
 795 }
 796
 797 static const struct rpc_call_ops nlmsvc_grant_ops = {
 798         .rpc_call_done = nlmsvc_grant_callback,
 799         .rpc_release = nlmsvc_grant_release,
 800 };
 801
 802 /*
 803  * We received a GRANT_RES callback. Try to find the corresponding
 804  * block.
 805  */
 806 void
 807 nlmsvc_grant_reply(struct nlm_cookie *cookie, __be32 status)
 808 {
 809         struct nlm_block        *block;
 810
 811         dprintk("grant_reply: looking for cookie %x, s=%d \n",
 812                 *(unsigned int *)(cookie->data), status);
 813         if (!(block = nlmsvc_find_block(cookie)))
 814                 return;
 815
 816         if (block) {
 817                 if (status == nlm_lck_denied_grace_period) {
 818                         /* Try again in a couple of seconds */
 819                         nlmsvc_insert_block(block, 10 * HZ);
 820                 } else {
 821                         /* Lock is now held by client, or has been rejected.
 822                          * In both cases, the block should be removed. */
 823                         nlmsvc_unlink_block(block);
 824                 }
 825         }
 826         nlmsvc_release_block(block);
 827 }
 828
 829 /* Helper function to handle retry of a deferred block.
 830  * If it is a blocking lock, call grant_blocked.
 831  * For a non-blocking lock or test lock, revisit the request.
 832  */
 833 static void
 834 retry_deferred_block(struct nlm_block *block)
 835 {
 836         if (!(block->b_flags & B_GOT_CALLBACK))
 837                 block->b_flags |= B_TIMED_OUT;
 838         nlmsvc_insert_block(block, NLM_TIMEOUT);
 839         dprintk("revisit block %p flags %d\n",  block, block->b_flags);
 840         if (block->b_deferred_req) {
 841                 block->b_deferred_req->revisit(block->b_deferred_req, 0);
 842                 block->b_deferred_req = NULL;
 843         }
 844 }
 845
 846 /*
 847  * Retry all blocked locks that have been notified. This is where lockd
 848  * picks up locks that can be granted, or grant notifications that must
 849  * be retransmitted.
 850  */
 851 unsigned long
 852 nlmsvc_retry_blocked(void)
 853 {
 854         unsigned long   timeout = MAX_SCHEDULE_TIMEOUT;
 855         struct nlm_block *block;
 856
 857         while (!list_empty(&nlm_blocked)) {
 858                 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
 859
 860                 if (block->b_when == NLM_NEVER)
 861                         break;
 862                 if (time_after(block->b_when,jiffies)) {
 863                         timeout = block->b_when - jiffies;
 864                         break;
 865                 }
 866
 867                 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
 868                         block, block->b_when);
 869                 if (block->b_flags & B_QUEUED) {
 870                         dprintk("nlmsvc_retry_blocked delete block (%p, granted=%d, flags=%d)\n",
 871                                 block, block->b_granted, block->b_flags);
 872                         retry_deferred_block(block);
 873                 } else
 874                         nlmsvc_grant_blocked(block);
 875         }
 876
 877         return timeout;
 878 }