fs/lockd/svclock.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * linux/fs/lockd/svclock.c
   4  *
   5  * Handling of server-side locks, mostly of the blocked variety.
   6  * This is the ugliest part of lockd because we tread on very thin ice.
   7  * GRANT and CANCEL calls may get stuck, meet in mid-flight, etc.
   8  * IMNSHO introducing the grant callback into the NLM protocol was one
   9  * of the worst ideas Sun ever had. Except maybe for the idea of doing
  10  * NFS file locking at all.
  11  *
  12  * I'm trying hard to avoid race conditions by protecting most accesses
  13  * to a file's list of blocked locks through a semaphore. The global
  14  * list of blocked locks is not protected in this fashion however.
  15  * Therefore, some functions (such as the RPC callback for the async grant
  16  * call) move blocked locks towards the head of the list *while some other
  17  * process might be traversing it*. This should not be a problem in
  18  * practice, because this will only cause functions traversing the list
  19  * to visit some blocks twice.
  20  *
  21  * Copyright (C) 1996, Olaf Kirch <okir@monad.swb.de>
  22  */
  23
  24 #include <linux/types.h>
  25 #include <linux/slab.h>
  26 #include <linux/errno.h>
  27 #include <linux/kernel.h>
  28 #include <linux/sched.h>
  29 #include <linux/sunrpc/clnt.h>
  30 #include <linux/sunrpc/svc_xprt.h>
  31 #include <linux/lockd/nlm.h>
  32 #include <linux/lockd/lockd.h>
  33 #include <linux/kthread.h>
  34 #include <linux/exportfs.h>
  35
  36 #define NLMDBG_FACILITY         NLMDBG_SVCLOCK
  37
  38 #ifdef CONFIG_LOCKD_V4
  39 #define nlm_deadlock    nlm4_deadlock
  40 #else
  41 #define nlm_deadlock    nlm_lck_denied
  42 #endif
  43
  44 static void nlmsvc_release_block(struct nlm_block *block);
  45 static void     nlmsvc_insert_block(struct nlm_block *block, unsigned long);
  46 static void     nlmsvc_remove_block(struct nlm_block *block);
  47
  48 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock);
  49 static void nlmsvc_freegrantargs(struct nlm_rqst *call);
  50 static const struct rpc_call_ops nlmsvc_grant_ops;
  51
  52 /*
  53  * The list of blocked locks to retry
  54  */
  55 static LIST_HEAD(nlm_blocked);
  56 static DEFINE_SPINLOCK(nlm_blocked_lock);
  57
  58 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  59 static const char *nlmdbg_cookie2a(const struct nlm_cookie *cookie)
  60 {
  61         /*
  62          * We can get away with a static buffer because this is only called
  63          * from lockd, which is single-threaded.
  64          */
  65         static char buf[2*NLM_MAXCOOKIELEN+1];
  66         unsigned int i, len = sizeof(buf);
  67         char *p = buf;
  68
  69         len--;  /* allow for trailing \0 */
  70         if (len < 3)
  71                 return "???";
  72         for (i = 0 ; i < cookie->len ; i++) {
  73                 if (len < 2) {
  74                         strcpy(p-3, "...");
  75                         break;
  76                 }
  77                 sprintf(p, "%02x", cookie->data[i]);
  78                 p += 2;
  79                 len -= 2;
  80         }
  81         *p = '\0';
  82
  83         return buf;
  84 }
  85 #endif
  86
  87 /*
  88  * Insert a blocked lock into the global list
  89  */
  90 static void
  91 nlmsvc_insert_block_locked(struct nlm_block *block, unsigned long when)
  92 {
  93         struct nlm_block *b;
  94         struct list_head *pos;
  95
  96         dprintk("lockd: nlmsvc_insert_block(%p, %ld)\n", block, when);
  97         if (list_empty(&block->b_list)) {
  98                 kref_get(&block->b_count);
  99         } else {
 100                 list_del_init(&block->b_list);
 101         }
 102
 103         pos = &nlm_blocked;
 104         if (when != NLM_NEVER) {
 105                 if ((when += jiffies) == NLM_NEVER)
 106                         when ++;
 107                 list_for_each(pos, &nlm_blocked) {
 108                         b = list_entry(pos, struct nlm_block, b_list);
 109                         if (time_after(b->b_when,when) || b->b_when == NLM_NEVER)
 110                                 break;
 111                 }
 112                 /* On normal exit from the loop, pos == &nlm_blocked,
 113                  * so we will be adding to the end of the list - good
 114                  */
 115         }
 116
 117         list_add_tail(&block->b_list, pos);
 118         block->b_when = when;
 119 }
 120
 121 static void nlmsvc_insert_block(struct nlm_block *block, unsigned long when)
 122 {
 123         spin_lock(&nlm_blocked_lock);
 124         nlmsvc_insert_block_locked(block, when);
 125         spin_unlock(&nlm_blocked_lock);
 126 }
 127
 128 /*
 129  * Remove a block from the global list
 130  */
 131 static inline void
 132 nlmsvc_remove_block(struct nlm_block *block)
 133 {
 134         spin_lock(&nlm_blocked_lock);
 135         if (!list_empty(&block->b_list)) {
 136                 list_del_init(&block->b_list);
 137                 spin_unlock(&nlm_blocked_lock);
 138                 nlmsvc_release_block(block);
 139                 return;
 140         }
 141         spin_unlock(&nlm_blocked_lock);
 142 }
 143
 144 /*
 145  * Find a block for a given lock
 146  */
 147 static struct nlm_block *
 148 nlmsvc_lookup_block(struct nlm_file *file, struct nlm_lock *lock)
 149 {
 150         struct nlm_block        *block;
 151         struct file_lock        *fl;
 152
 153         dprintk("lockd: nlmsvc_lookup_block f=%p pd=%d %Ld-%Ld ty=%d\n",
 154                                 file, lock->fl.fl_pid,
 155                                 (long long)lock->fl.fl_start,
 156                                 (long long)lock->fl.fl_end, lock->fl.fl_type);
 157         spin_lock(&nlm_blocked_lock);
 158         list_for_each_entry(block, &nlm_blocked, b_list) {
 159                 fl = &block->b_call->a_args.lock.fl;
 160                 dprintk("lockd: check f=%p pd=%d %Ld-%Ld ty=%d cookie=%s\n",
 161                                 block->b_file, fl->fl_pid,
 162                                 (long long)fl->fl_start,
 163                                 (long long)fl->fl_end, fl->fl_type,
 164                                 nlmdbg_cookie2a(&block->b_call->a_args.cookie));
 165                 if (block->b_file == file && nlm_compare_locks(fl, &lock->fl)) {
 166                         kref_get(&block->b_count);
 167                         spin_unlock(&nlm_blocked_lock);
 168                         return block;
 169                 }
 170         }
 171         spin_unlock(&nlm_blocked_lock);
 172
 173         return NULL;
 174 }
 175
 176 static inline int nlm_cookie_match(struct nlm_cookie *a, struct nlm_cookie *b)
 177 {
 178         if (a->len != b->len)
 179                 return 0;
 180         if (memcmp(a->data, b->data, a->len))
 181                 return 0;
 182         return 1;
 183 }
 184
 185 /*
 186  * Find a block with a given NLM cookie.
 187  */
 188 static inline struct nlm_block *
 189 nlmsvc_find_block(struct nlm_cookie *cookie)
 190 {
 191         struct nlm_block *block;
 192
 193         spin_lock(&nlm_blocked_lock);
 194         list_for_each_entry(block, &nlm_blocked, b_list) {
 195                 if (nlm_cookie_match(&block->b_call->a_args.cookie,cookie))
 196                         goto found;
 197         }
 198         spin_unlock(&nlm_blocked_lock);
 199
 200         return NULL;
 201
 202 found:
 203         dprintk("nlmsvc_find_block(%s): block=%p\n", nlmdbg_cookie2a(cookie), block);
 204         kref_get(&block->b_count);
 205         spin_unlock(&nlm_blocked_lock);
 206         return block;
 207 }
 208
 209 /*
 210  * Create a block and initialize it.
 211  *
 212  * Note: we explicitly set the cookie of the grant reply to that of
 213  * the blocked lock request. The spec explicitly mentions that the client
 214  * should _not_ rely on the callback containing the same cookie as the
 215  * request, but (as I found out later) that's because some implementations
 216  * do just this. Never mind the standards comittees, they support our
 217  * logging industries.
 218  *
 219  * 10 years later: I hope we can safely ignore these old and broken
 220  * clients by now. Let's fix this so we can uniquely identify an incoming
 221  * GRANTED_RES message by cookie, without having to rely on the client's IP
 222  * address. --okir
 223  */
 224 static struct nlm_block *
 225 nlmsvc_create_block(struct svc_rqst *rqstp, struct nlm_host *host,
 226                     struct nlm_file *file, struct nlm_lock *lock,
 227                     struct nlm_cookie *cookie)
 228 {
 229         struct nlm_block        *block;
 230         struct nlm_rqst         *call = NULL;
 231
 232         call = nlm_alloc_call(host);
 233         if (call == NULL)
 234                 return NULL;
 235
 236         /* Allocate memory for block, and initialize arguments */
 237         block = kzalloc(sizeof(*block), GFP_KERNEL);
 238         if (block == NULL)
 239                 goto failed;
 240         kref_init(&block->b_count);
 241         INIT_LIST_HEAD(&block->b_list);
 242         INIT_LIST_HEAD(&block->b_flist);
 243
 244         if (!nlmsvc_setgrantargs(call, lock))
 245                 goto failed_free;
 246
 247         /* Set notifier function for VFS, and init args */
 248         call->a_args.lock.fl.fl_flags |= FL_SLEEP;
 249         call->a_args.lock.fl.fl_lmops = &nlmsvc_lock_operations;
 250         nlmclnt_next_cookie(&call->a_args.cookie);
 251
 252         dprintk("lockd: created block %p...\n", block);
 253
 254         /* Create and initialize the block */
 255         block->b_daemon = rqstp->rq_server;
 256         block->b_host   = host;
 257         block->b_file   = file;
 258         file->f_count++;
 259
 260         /* Add to file's list of blocks */
 261         list_add(&block->b_flist, &file->f_blocks);
 262
 263         /* Set up RPC arguments for callback */
 264         block->b_call = call;
 265         call->a_flags   = RPC_TASK_ASYNC;
 266         call->a_block = block;
 267
 268         return block;
 269
 270 failed_free:
 271         kfree(block);
 272 failed:
 273         nlmsvc_release_call(call);
 274         return NULL;
 275 }
 276
 277 /*
 278  * Delete a block.
 279  * It is the caller's responsibility to check whether the file
 280  * can be closed hereafter.
 281  */
 282 static int nlmsvc_unlink_block(struct nlm_block *block)
 283 {
 284         int status;
 285         dprintk("lockd: unlinking block %p...\n", block);
 286
 287         /* Remove block from list */
 288         status = locks_delete_block(&block->b_call->a_args.lock.fl);
 289         nlmsvc_remove_block(block);
 290         return status;
 291 }
 292
 293 static void nlmsvc_free_block(struct kref *kref)
 294 {
 295         struct nlm_block *block = container_of(kref, struct nlm_block, b_count);
 296         struct nlm_file         *file = block->b_file;
 297
 298         dprintk("lockd: freeing block %p...\n", block);
 299
 300         /* Remove block from file's list of blocks */
 301         list_del_init(&block->b_flist);
 302         mutex_unlock(&file->f_mutex);
 303
 304         nlmsvc_freegrantargs(block->b_call);
 305         nlmsvc_release_call(block->b_call);
 306         nlm_release_file(block->b_file);
 307         kfree(block);
 308 }
 309
 310 static void nlmsvc_release_block(struct nlm_block *block)
 311 {
 312         if (block != NULL)
 313                 kref_put_mutex(&block->b_count, nlmsvc_free_block, &block->b_file->f_mutex);
 314 }
 315
 316 /*
 317  * Loop over all blocks and delete blocks held by
 318  * a matching host.
 319  */
 320 void nlmsvc_traverse_blocks(struct nlm_host *host,
 321                         struct nlm_file *file,
 322                         nlm_host_match_fn_t match)
 323 {
 324         struct nlm_block *block, *next;
 325
 326 restart:
 327         mutex_lock(&file->f_mutex);
 328         spin_lock(&nlm_blocked_lock);
 329         list_for_each_entry_safe(block, next, &file->f_blocks, b_flist) {
 330                 if (!match(block->b_host, host))
 331                         continue;
 332                 /* Do not destroy blocks that are not on
 333                  * the global retry list - why? */
 334                 if (list_empty(&block->b_list))
 335                         continue;
 336                 kref_get(&block->b_count);
 337                 spin_unlock(&nlm_blocked_lock);
 338                 mutex_unlock(&file->f_mutex);
 339                 nlmsvc_unlink_block(block);
 340                 nlmsvc_release_block(block);
 341                 goto restart;
 342         }
 343         spin_unlock(&nlm_blocked_lock);
 344         mutex_unlock(&file->f_mutex);
 345 }
 346
 347 static struct nlm_lockowner *
 348 nlmsvc_get_lockowner(struct nlm_lockowner *lockowner)
 349 {
 350         refcount_inc(&lockowner->count);
 351         return lockowner;
 352 }
 353
 354 void nlmsvc_put_lockowner(struct nlm_lockowner *lockowner)
 355 {
 356         if (!refcount_dec_and_lock(&lockowner->count, &lockowner->host->h_lock))
 357                 return;
 358         list_del(&lockowner->list);
 359         spin_unlock(&lockowner->host->h_lock);
 360         nlmsvc_release_host(lockowner->host);
 361         kfree(lockowner);
 362 }
 363
 364 static struct nlm_lockowner *__nlmsvc_find_lockowner(struct nlm_host *host, pid_t pid)
 365 {
 366         struct nlm_lockowner *lockowner;
 367         list_for_each_entry(lockowner, &host->h_lockowners, list) {
 368                 if (lockowner->pid != pid)
 369                         continue;
 370                 return nlmsvc_get_lockowner(lockowner);
 371         }
 372         return NULL;
 373 }
 374
 375 static struct nlm_lockowner *nlmsvc_find_lockowner(struct nlm_host *host, pid_t pid)
 376 {
 377         struct nlm_lockowner *res, *new = NULL;
 378
 379         spin_lock(&host->h_lock);
 380         res = __nlmsvc_find_lockowner(host, pid);
 381
 382         if (res == NULL) {
 383                 spin_unlock(&host->h_lock);
 384                 new = kmalloc(sizeof(*res), GFP_KERNEL);
 385                 spin_lock(&host->h_lock);
 386                 res = __nlmsvc_find_lockowner(host, pid);
 387                 if (res == NULL && new != NULL) {
 388                         res = new;
 389                         /* fs/locks.c will manage the refcount through lock_ops */
 390                         refcount_set(&new->count, 1);
 391                         new->pid = pid;
 392                         new->host = nlm_get_host(host);
 393                         list_add(&new->list, &host->h_lockowners);
 394                         new = NULL;
 395                 }
 396         }
 397
 398         spin_unlock(&host->h_lock);
 399         kfree(new);
 400         return res;
 401 }
 402
 403 void
 404 nlmsvc_release_lockowner(struct nlm_lock *lock)
 405 {
 406         if (lock->fl.fl_owner)
 407                 nlmsvc_put_lockowner(lock->fl.fl_owner);
 408 }
 409
 410 void nlmsvc_locks_init_private(struct file_lock *fl, struct nlm_host *host,
 411                                                 pid_t pid)
 412 {
 413         fl->fl_owner = nlmsvc_find_lockowner(host, pid);
 414 }
 415
 416 /*
 417  * Initialize arguments for GRANTED call. The nlm_rqst structure
 418  * has been cleared already.
 419  */
 420 static int nlmsvc_setgrantargs(struct nlm_rqst *call, struct nlm_lock *lock)
 421 {
 422         locks_copy_lock(&call->a_args.lock.fl, &lock->fl);
 423         memcpy(&call->a_args.lock.fh, &lock->fh, sizeof(call->a_args.lock.fh));
 424         call->a_args.lock.caller = utsname()->nodename;
 425         call->a_args.lock.oh.len = lock->oh.len;
 426
 427         /* set default data area */
 428         call->a_args.lock.oh.data = call->a_owner;
 429         call->a_args.lock.svid = ((struct nlm_lockowner *)lock->fl.fl_owner)->pid;
 430
 431         if (lock->oh.len > NLMCLNT_OHSIZE) {
 432                 void *data = kmalloc(lock->oh.len, GFP_KERNEL);
 433                 if (!data)
 434                         return 0;
 435                 call->a_args.lock.oh.data = (u8 *) data;
 436         }
 437
 438         memcpy(call->a_args.lock.oh.data, lock->oh.data, lock->oh.len);
 439         return 1;
 440 }
 441
 442 static void nlmsvc_freegrantargs(struct nlm_rqst *call)
 443 {
 444         if (call->a_args.lock.oh.data != call->a_owner)
 445                 kfree(call->a_args.lock.oh.data);
 446
 447         locks_release_private(&call->a_args.lock.fl);
 448 }
 449
 450 /*
 451  * Deferred lock request handling for non-blocking lock
 452  */
 453 static __be32
 454 nlmsvc_defer_lock_rqst(struct svc_rqst *rqstp, struct nlm_block *block)
 455 {
 456         __be32 status = nlm_lck_denied_nolocks;
 457
 458         block->b_flags |= B_QUEUED;
 459
 460         nlmsvc_insert_block(block, NLM_TIMEOUT);
 461
 462         block->b_cache_req = &rqstp->rq_chandle;
 463         if (rqstp->rq_chandle.defer) {
 464                 block->b_deferred_req =
 465                         rqstp->rq_chandle.defer(block->b_cache_req);
 466                 if (block->b_deferred_req != NULL)
 467                         status = nlm_drop_reply;
 468         }
 469         dprintk("lockd: nlmsvc_defer_lock_rqst block %p flags %d status %d\n",
 470                 block, block->b_flags, ntohl(status));
 471
 472         return status;
 473 }
 474
 475 /*
 476  * Attempt to establish a lock, and if it can't be granted, block it
 477  * if required.
 478  */
 479 __be32
 480 nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
 481             struct nlm_host *host, struct nlm_lock *lock, int wait,
 482             struct nlm_cookie *cookie, int reclaim)
 483 {
 484 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 485         struct inode            *inode = nlmsvc_file_inode(file);
 486 #endif
 487         struct nlm_block        *block = NULL;
 488         int                     error;
 489         int                     mode;
 490         int                     async_block = 0;
 491         __be32                  ret;
 492
 493         dprintk("lockd: nlmsvc_lock(%s/%ld, ty=%d, pi=%d, %Ld-%Ld, bl=%d)\n",
 494                                 inode->i_sb->s_id, inode->i_ino,
 495                                 lock->fl.fl_type, lock->fl.fl_pid,
 496                                 (long long)lock->fl.fl_start,
 497                                 (long long)lock->fl.fl_end,
 498                                 wait);
 499
 500         if (nlmsvc_file_file(file)->f_op->lock) {
 501                 async_block = wait;
 502                 wait = 0;
 503         }
 504
 505         /* Lock file against concurrent access */
 506         mutex_lock(&file->f_mutex);
 507         /* Get existing block (in case client is busy-waiting)
 508          * or create new block
 509          */
 510         block = nlmsvc_lookup_block(file, lock);
 511         if (block == NULL) {
 512                 block = nlmsvc_create_block(rqstp, host, file, lock, cookie);
 513                 ret = nlm_lck_denied_nolocks;
 514                 if (block == NULL)
 515                         goto out;
 516                 lock = &block->b_call->a_args.lock;
 517         } else
 518                 lock->fl.fl_flags &= ~FL_SLEEP;
 519
 520         if (block->b_flags & B_QUEUED) {
 521                 dprintk("lockd: nlmsvc_lock deferred block %p flags %d\n",
 522                                                         block, block->b_flags);
 523                 if (block->b_granted) {
 524                         nlmsvc_unlink_block(block);
 525                         ret = nlm_granted;
 526                         goto out;
 527                 }
 528                 if (block->b_flags & B_TIMED_OUT) {
 529                         nlmsvc_unlink_block(block);
 530                         ret = nlm_lck_denied;
 531                         goto out;
 532                 }
 533                 ret = nlm_drop_reply;
 534                 goto out;
 535         }
 536
 537         if (locks_in_grace(SVC_NET(rqstp)) && !reclaim) {
 538                 ret = nlm_lck_denied_grace_period;
 539                 goto out;
 540         }
 541         if (reclaim && !locks_in_grace(SVC_NET(rqstp))) {
 542                 ret = nlm_lck_denied_grace_period;
 543                 goto out;
 544         }
 545
 546         if (!wait)
 547                 lock->fl.fl_flags &= ~FL_SLEEP;
 548         mode = lock_to_openmode(&lock->fl);
 549         error = vfs_lock_file(file->f_file[mode], F_SETLK, &lock->fl, NULL);
 550         lock->fl.fl_flags &= ~FL_SLEEP;
 551
 552         dprintk("lockd: vfs_lock_file returned %d\n", error);
 553         switch (error) {
 554                 case 0:
 555                         ret = nlm_granted;
 556                         goto out;
 557                 case -EAGAIN:
 558                         /*
 559                          * If this is a blocking request for an
 560                          * already pending lock request then we need
 561                          * to put it back on lockd's block list
 562                          */
 563                         if (wait)
 564                                 break;
 565                         ret = async_block ? nlm_lck_blocked : nlm_lck_denied;
 566                         goto out;
 567                 case FILE_LOCK_DEFERRED:
 568                         if (wait)
 569                                 break;
 570                         /* Filesystem lock operation is in progress
 571                            Add it to the queue waiting for callback */
 572                         ret = nlmsvc_defer_lock_rqst(rqstp, block);
 573                         goto out;
 574                 case -EDEADLK:
 575                         ret = nlm_deadlock;
 576                         goto out;
 577                 default:                        /* includes ENOLCK */
 578                         ret = nlm_lck_denied_nolocks;
 579                         goto out;
 580         }
 581
 582         ret = nlm_lck_blocked;
 583
 584         /* Append to list of blocked */
 585         nlmsvc_insert_block(block, NLM_NEVER);
 586 out:
 587         mutex_unlock(&file->f_mutex);
 588         nlmsvc_release_block(block);
 589         dprintk("lockd: nlmsvc_lock returned %u\n", ret);
 590         return ret;
 591 }
 592
 593 /*
 594  * Test for presence of a conflicting lock.
 595  */
 596 __be32
 597 nlmsvc_testlock(struct svc_rqst *rqstp, struct nlm_file *file,
 598                 struct nlm_host *host, struct nlm_lock *lock,
 599                 struct nlm_lock *conflock, struct nlm_cookie *cookie)
 600 {
 601         int                     error;
 602         int                     mode;
 603         __be32                  ret;
 604
 605         dprintk("lockd: nlmsvc_testlock(%s/%ld, ty=%d, %Ld-%Ld)\n",
 606                                 nlmsvc_file_inode(file)->i_sb->s_id,
 607                                 nlmsvc_file_inode(file)->i_ino,
 608                                 lock->fl.fl_type,
 609                                 (long long)lock->fl.fl_start,
 610                                 (long long)lock->fl.fl_end);
 611
 612         if (locks_in_grace(SVC_NET(rqstp))) {
 613                 ret = nlm_lck_denied_grace_period;
 614                 goto out;
 615         }
 616
 617         mode = lock_to_openmode(&lock->fl);
 618         error = vfs_test_lock(file->f_file[mode], &lock->fl);
 619         if (error) {
 620                 /* We can't currently deal with deferred test requests */
 621                 if (error == FILE_LOCK_DEFERRED)
 622                         WARN_ON_ONCE(1);
 623
 624                 ret = nlm_lck_denied_nolocks;
 625                 goto out;
 626         }
 627
 628         if (lock->fl.fl_type == F_UNLCK) {
 629                 ret = nlm_granted;
 630                 goto out;
 631         }
 632
 633         dprintk("lockd: conflicting lock(ty=%d, %Ld-%Ld)\n",
 634                 lock->fl.fl_type, (long long)lock->fl.fl_start,
 635                 (long long)lock->fl.fl_end);
 636         conflock->caller = "somehost";  /* FIXME */
 637         conflock->len = strlen(conflock->caller);
 638         conflock->oh.len = 0;           /* don't return OH info */
 639         conflock->svid = lock->fl.fl_pid;
 640         conflock->fl.fl_type = lock->fl.fl_type;
 641         conflock->fl.fl_start = lock->fl.fl_start;
 642         conflock->fl.fl_end = lock->fl.fl_end;
 643         locks_release_private(&lock->fl);
 644
 645         ret = nlm_lck_denied;
 646 out:
 647         return ret;
 648 }
 649
 650 /*
 651  * Remove a lock.
 652  * This implies a CANCEL call: We send a GRANT_MSG, the client replies
 653  * with a GRANT_RES call which gets lost, and calls UNLOCK immediately
 654  * afterwards. In this case the block will still be there, and hence
 655  * must be removed.
 656  */
 657 __be32
 658 nlmsvc_unlock(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
 659 {
 660         int     error = 0;
 661
 662         dprintk("lockd: nlmsvc_unlock(%s/%ld, pi=%d, %Ld-%Ld)\n",
 663                                 nlmsvc_file_inode(file)->i_sb->s_id,
 664                                 nlmsvc_file_inode(file)->i_ino,
 665                                 lock->fl.fl_pid,
 666                                 (long long)lock->fl.fl_start,
 667                                 (long long)lock->fl.fl_end);
 668
 669         /* First, cancel any lock that might be there */
 670         nlmsvc_cancel_blocked(net, file, lock);
 671
 672         lock->fl.fl_type = F_UNLCK;
 673         lock->fl.fl_file = file->f_file[O_RDONLY];
 674         if (lock->fl.fl_file)
 675                 error = vfs_lock_file(lock->fl.fl_file, F_SETLK,
 676                                         &lock->fl, NULL);
 677         lock->fl.fl_file = file->f_file[O_WRONLY];
 678         if (lock->fl.fl_file)
 679                 error |= vfs_lock_file(lock->fl.fl_file, F_SETLK,
 680                                         &lock->fl, NULL);
 681
 682         return (error < 0)? nlm_lck_denied_nolocks : nlm_granted;
 683 }
 684
 685 /*
 686  * Cancel a previously blocked request.
 687  *
 688  * A cancel request always overrides any grant that may currently
 689  * be in progress.
 690  * The calling procedure must check whether the file can be closed.
 691  */
 692 __be32
 693 nlmsvc_cancel_blocked(struct net *net, struct nlm_file *file, struct nlm_lock *lock)
 694 {
 695         struct nlm_block        *block;
 696         int status = 0;
 697         int mode;
 698
 699         dprintk("lockd: nlmsvc_cancel(%s/%ld, pi=%d, %Ld-%Ld)\n",
 700                                 nlmsvc_file_inode(file)->i_sb->s_id,
 701                                 nlmsvc_file_inode(file)->i_ino,
 702                                 lock->fl.fl_pid,
 703                                 (long long)lock->fl.fl_start,
 704                                 (long long)lock->fl.fl_end);
 705
 706         if (locks_in_grace(net))
 707                 return nlm_lck_denied_grace_period;
 708
 709         mutex_lock(&file->f_mutex);
 710         block = nlmsvc_lookup_block(file, lock);
 711         mutex_unlock(&file->f_mutex);
 712         if (block != NULL) {
 713                 struct file_lock *fl = &block->b_call->a_args.lock.fl;
 714
 715                 mode = lock_to_openmode(fl);
 716                 vfs_cancel_lock(block->b_file->f_file[mode], fl);
 717                 status = nlmsvc_unlink_block(block);
 718                 nlmsvc_release_block(block);
 719         }
 720         return status ? nlm_lck_denied : nlm_granted;
 721 }
 722
 723 /*
 724  * This is a callback from the filesystem for VFS file lock requests.
 725  * It will be used if lm_grant is defined and the filesystem can not
 726  * respond to the request immediately.
 727  * For SETLK or SETLKW request it will get the local posix lock.
 728  * In all cases it will move the block to the head of nlm_blocked q where
 729  * nlmsvc_retry_blocked() can send back a reply for SETLKW or revisit the
 730  * deferred rpc for GETLK and SETLK.
 731  */
 732 static void
 733 nlmsvc_update_deferred_block(struct nlm_block *block, int result)
 734 {
 735         block->b_flags |= B_GOT_CALLBACK;
 736         if (result == 0)
 737                 block->b_granted = 1;
 738         else
 739                 block->b_flags |= B_TIMED_OUT;
 740 }
 741
 742 static int nlmsvc_grant_deferred(struct file_lock *fl, int result)
 743 {
 744         struct nlm_block *block;
 745         int rc = -ENOENT;
 746
 747         spin_lock(&nlm_blocked_lock);
 748         list_for_each_entry(block, &nlm_blocked, b_list) {
 749                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
 750                         dprintk("lockd: nlmsvc_notify_blocked block %p flags %d\n",
 751                                                         block, block->b_flags);
 752                         if (block->b_flags & B_QUEUED) {
 753                                 if (block->b_flags & B_TIMED_OUT) {
 754                                         rc = -ENOLCK;
 755                                         break;
 756                                 }
 757                                 nlmsvc_update_deferred_block(block, result);
 758                         } else if (result == 0)
 759                                 block->b_granted = 1;
 760
 761                         nlmsvc_insert_block_locked(block, 0);
 762                         svc_wake_up(block->b_daemon);
 763                         rc = 0;
 764                         break;
 765                 }
 766         }
 767         spin_unlock(&nlm_blocked_lock);
 768         if (rc == -ENOENT)
 769                 printk(KERN_WARNING "lockd: grant for unknown block\n");
 770         return rc;
 771 }
 772
 773 /*
 774  * Unblock a blocked lock request. This is a callback invoked from the
 775  * VFS layer when a lock on which we blocked is removed.
 776  *
 777  * This function doesn't grant the blocked lock instantly, but rather moves
 778  * the block to the head of nlm_blocked where it can be picked up by lockd.
 779  */
 780 static void
 781 nlmsvc_notify_blocked(struct file_lock *fl)
 782 {
 783         struct nlm_block        *block;
 784
 785         dprintk("lockd: VFS unblock notification for block %p\n", fl);
 786         spin_lock(&nlm_blocked_lock);
 787         list_for_each_entry(block, &nlm_blocked, b_list) {
 788                 if (nlm_compare_locks(&block->b_call->a_args.lock.fl, fl)) {
 789                         nlmsvc_insert_block_locked(block, 0);
 790                         spin_unlock(&nlm_blocked_lock);
 791                         svc_wake_up(block->b_daemon);
 792                         return;
 793                 }
 794         }
 795         spin_unlock(&nlm_blocked_lock);
 796         printk(KERN_WARNING "lockd: notification for unknown block!\n");
 797 }
 798
 799 static fl_owner_t nlmsvc_get_owner(fl_owner_t owner)
 800 {
 801         return nlmsvc_get_lockowner(owner);
 802 }
 803
 804 static void nlmsvc_put_owner(fl_owner_t owner)
 805 {
 806         nlmsvc_put_lockowner(owner);
 807 }
 808
 809 const struct lock_manager_operations nlmsvc_lock_operations = {
 810         .lm_notify = nlmsvc_notify_blocked,
 811         .lm_grant = nlmsvc_grant_deferred,
 812         .lm_get_owner = nlmsvc_get_owner,
 813         .lm_put_owner = nlmsvc_put_owner,
 814 };
 815
 816 /*
 817  * Try to claim a lock that was previously blocked.
 818  *
 819  * Note that we use both the RPC_GRANTED_MSG call _and_ an async
 820  * RPC thread when notifying the client. This seems like overkill...
 821  * Here's why:
 822  *  -   we don't want to use a synchronous RPC thread, otherwise
 823  *      we might find ourselves hanging on a dead portmapper.
 824  *  -   Some lockd implementations (e.g. HP) don't react to
 825  *      RPC_GRANTED calls; they seem to insist on RPC_GRANTED_MSG calls.
 826  */
 827 static void
 828 nlmsvc_grant_blocked(struct nlm_block *block)
 829 {
 830         struct nlm_file         *file = block->b_file;
 831         struct nlm_lock         *lock = &block->b_call->a_args.lock;
 832         int                     mode;
 833         int                     error;
 834         loff_t                  fl_start, fl_end;
 835
 836         dprintk("lockd: grant blocked lock %p\n", block);
 837
 838         kref_get(&block->b_count);
 839
 840         /* Unlink block request from list */
 841         nlmsvc_unlink_block(block);
 842
 843         /* If b_granted is true this means we've been here before.
 844          * Just retry the grant callback, possibly refreshing the RPC
 845          * binding */
 846         if (block->b_granted) {
 847                 nlm_rebind_host(block->b_host);
 848                 goto callback;
 849         }
 850
 851         /* Try the lock operation again */
 852         /* vfs_lock_file() can mangle fl_start and fl_end, but we need
 853          * them unchanged for the GRANT_MSG
 854          */
 855         lock->fl.fl_flags |= FL_SLEEP;
 856         fl_start = lock->fl.fl_start;
 857         fl_end = lock->fl.fl_end;
 858         mode = lock_to_openmode(&lock->fl);
 859         error = vfs_lock_file(file->f_file[mode], F_SETLK, &lock->fl, NULL);
 860         lock->fl.fl_flags &= ~FL_SLEEP;
 861         lock->fl.fl_start = fl_start;
 862         lock->fl.fl_end = fl_end;
 863
 864         switch (error) {
 865         case 0:
 866                 break;
 867         case FILE_LOCK_DEFERRED:
 868                 dprintk("lockd: lock still blocked error %d\n", error);
 869                 nlmsvc_insert_block(block, NLM_NEVER);
 870                 nlmsvc_release_block(block);
 871                 return;
 872         default:
 873                 printk(KERN_WARNING "lockd: unexpected error %d in %s!\n",
 874                                 -error, __func__);
 875                 nlmsvc_insert_block(block, 10 * HZ);
 876                 nlmsvc_release_block(block);
 877                 return;
 878         }
 879
 880 callback:
 881         /* Lock was granted by VFS. */
 882         dprintk("lockd: GRANTing blocked lock.\n");
 883         block->b_granted = 1;
 884
 885         /* keep block on the list, but don't reattempt until the RPC
 886          * completes or the submission fails
 887          */
 888         nlmsvc_insert_block(block, NLM_NEVER);
 889
 890         /* Call the client -- use a soft RPC task since nlmsvc_retry_blocked
 891          * will queue up a new one if this one times out
 892          */
 893         error = nlm_async_call(block->b_call, NLMPROC_GRANTED_MSG,
 894                                 &nlmsvc_grant_ops);
 895
 896         /* RPC submission failed, wait a bit and retry */
 897         if (error < 0)
 898                 nlmsvc_insert_block(block, 10 * HZ);
 899 }
 900
 901 /*
 902  * This is the callback from the RPC layer when the NLM_GRANTED_MSG
 903  * RPC call has succeeded or timed out.
 904  * Like all RPC callbacks, it is invoked by the rpciod process, so it
 905  * better not sleep. Therefore, we put the blocked lock on the nlm_blocked
 906  * chain once more in order to have it removed by lockd itself (which can
 907  * then sleep on the file semaphore without disrupting e.g. the nfs client).
 908  */
 909 static void nlmsvc_grant_callback(struct rpc_task *task, void *data)
 910 {
 911         struct nlm_rqst         *call = data;
 912         struct nlm_block        *block = call->a_block;
 913         unsigned long           timeout;
 914
 915         dprintk("lockd: GRANT_MSG RPC callback\n");
 916
 917         spin_lock(&nlm_blocked_lock);
 918         /* if the block is not on a list at this point then it has
 919          * been invalidated. Don't try to requeue it.
 920          *
 921          * FIXME: it's possible that the block is removed from the list
 922          * after this check but before the nlmsvc_insert_block. In that
 923          * case it will be added back. Perhaps we need better locking
 924          * for nlm_blocked?
 925          */
 926         if (list_empty(&block->b_list))
 927                 goto out;
 928
 929         /* Technically, we should down the file semaphore here. Since we
 930          * move the block towards the head of the queue only, no harm
 931          * can be done, though. */
 932         if (task->tk_status < 0) {
 933                 /* RPC error: Re-insert for retransmission */
 934                 timeout = 10 * HZ;
 935         } else {
 936                 /* Call was successful, now wait for client callback */
 937                 timeout = 60 * HZ;
 938         }
 939         nlmsvc_insert_block_locked(block, timeout);
 940         svc_wake_up(block->b_daemon);
 941 out:
 942         spin_unlock(&nlm_blocked_lock);
 943 }
 944
 945 /*
 946  * FIXME: nlmsvc_release_block() grabs a mutex.  This is not allowed for an
 947  * .rpc_release rpc_call_op
 948  */
 949 static void nlmsvc_grant_release(void *data)
 950 {
 951         struct nlm_rqst         *call = data;
 952         nlmsvc_release_block(call->a_block);
 953 }
 954
 955 static const struct rpc_call_ops nlmsvc_grant_ops = {
 956         .rpc_call_done = nlmsvc_grant_callback,
 957         .rpc_release = nlmsvc_grant_release,
 958 };
 959
 960 /*
 961  * We received a GRANT_RES callback. Try to find the corresponding
 962  * block.
 963  */
 964 void
 965 nlmsvc_grant_reply(struct nlm_cookie *cookie, __be32 status)
 966 {
 967         struct nlm_block        *block;
 968         struct file_lock        *fl;
 969         int                     error;
 970
 971         dprintk("grant_reply: looking for cookie %x, s=%d \n",
 972                 *(unsigned int *)(cookie->data), status);
 973         if (!(block = nlmsvc_find_block(cookie)))
 974                 return;
 975
 976         switch (status) {
 977         case nlm_lck_denied_grace_period:
 978                 /* Try again in a couple of seconds */
 979                 nlmsvc_insert_block(block, 10 * HZ);
 980                 break;
 981         case nlm_lck_denied:
 982                 /* Client doesn't want it, just unlock it */
 983                 nlmsvc_unlink_block(block);
 984                 fl = &block->b_call->a_args.lock.fl;
 985                 fl->fl_type = F_UNLCK;
 986                 error = vfs_lock_file(fl->fl_file, F_SETLK, fl, NULL);
 987                 if (error)
 988                         pr_warn("lockd: unable to unlock lock rejected by client!\n");
 989                 break;
 990         default:
 991                 /*
 992                  * Either it was accepted or the status makes no sense
 993                  * just unlink it either way.
 994                  */
 995                 nlmsvc_unlink_block(block);
 996         }
 997         nlmsvc_release_block(block);
 998 }
 999
1000 /* Helper function to handle retry of a deferred block.
1001  * If it is a blocking lock, call grant_blocked.
1002  * For a non-blocking lock or test lock, revisit the request.
1003  */
1004 static void
1005 retry_deferred_block(struct nlm_block *block)
1006 {
1007         if (!(block->b_flags & B_GOT_CALLBACK))
1008                 block->b_flags |= B_TIMED_OUT;
1009         nlmsvc_insert_block(block, NLM_TIMEOUT);
1010         dprintk("revisit block %p flags %d\n",  block, block->b_flags);
1011         if (block->b_deferred_req) {
1012                 block->b_deferred_req->revisit(block->b_deferred_req, 0);
1013                 block->b_deferred_req = NULL;
1014         }
1015 }
1016
1017 /*
1018  * Retry all blocked locks that have been notified. This is where lockd
1019  * picks up locks that can be granted, or grant notifications that must
1020  * be retransmitted.
1021  */
1022 void
1023 nlmsvc_retry_blocked(void)
1024 {
1025         unsigned long   timeout = MAX_SCHEDULE_TIMEOUT;
1026         struct nlm_block *block;
1027
1028         spin_lock(&nlm_blocked_lock);
1029         while (!list_empty(&nlm_blocked) && !kthread_should_stop()) {
1030                 block = list_entry(nlm_blocked.next, struct nlm_block, b_list);
1031
1032                 if (block->b_when == NLM_NEVER)
1033                         break;
1034                 if (time_after(block->b_when, jiffies)) {
1035                         timeout = block->b_when - jiffies;
1036                         break;
1037                 }
1038                 spin_unlock(&nlm_blocked_lock);
1039
1040                 dprintk("nlmsvc_retry_blocked(%p, when=%ld)\n",
1041                         block, block->b_when);
1042                 if (block->b_flags & B_QUEUED) {
1043                         dprintk("nlmsvc_retry_blocked delete block (%p, granted=%d, flags=%d)\n",
1044                                 block, block->b_granted, block->b_flags);
1045                         retry_deferred_block(block);
1046                 } else
1047                         nlmsvc_grant_blocked(block);
1048                 spin_lock(&nlm_blocked_lock);
1049         }
1050         spin_unlock(&nlm_blocked_lock);
1051
1052         if (timeout < MAX_SCHEDULE_TIMEOUT)
1053                 mod_timer(&nlmsvc_retry, jiffies + timeout);
1054 }