lustre/ldlm/ldlm_flock.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #define DEBUG_SUBSYSTEM S_LDLM
  25
  26 #ifdef __KERNEL__
  27 #include <linux/lustre_dlm.h>
  28 #include <linux/obd_support.h>
  29 #include <linux/obd_class.h>
  30 #include <linux/lustre_lib.h>
  31 #include <portals/list.h>
  32 #else
  33 #include <liblustre.h>
  34 #endif
  35
  36 #include "ldlm_internal.h"
  37
  38 #define l_flock_waitq   l_lru
  39
  40 static struct list_head ldlm_flock_waitq = LIST_HEAD_INIT(ldlm_flock_waitq);
  41
  42 /**
  43  * list_for_remaining_safe - iterate over the remaining entries in a list
  44  *              and safeguard against removal of a list entry.
  45  * @pos:        the &struct list_head to use as a loop counter. pos MUST
  46  *              have been initialized prior to using it in this macro.
  47  * @n:          another &struct list_head to use as temporary storage
  48  * @head:       the head for your list.
  49  */
  50 #define list_for_remaining_safe(pos, n, head) \
  51         for (n = pos->next; pos != (head); pos = n, n = pos->next)
  52
  53 static inline int
  54 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
  55 {
  56         return((new->l_policy_data.l_flock.pid ==
  57                 lock->l_policy_data.l_flock.pid) &&
  58                (new->l_export == lock->l_export));
  59 }
  60
  61 static inline int
  62 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
  63 {
  64         return((new->l_policy_data.l_flock.start <=
  65                 lock->l_policy_data.l_flock.end) &&
  66                (new->l_policy_data.l_flock.end >=
  67                 lock->l_policy_data.l_flock.start));
  68 }
  69
  70 static inline void
  71 ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, int flags)
  72 {
  73         ENTRY;
  74
  75         LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: 0x%x)",
  76                    mode, flags);
  77
  78         list_del_init(&lock->l_res_link);
  79         if (flags == LDLM_FL_WAIT_NOREPROC) {
  80                 /* client side - set a flag to prevent sending a CANCEL */
  81                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
  82                 ldlm_lock_decref_internal(lock, mode);
  83         }
  84
  85         ldlm_lock_destroy(lock);
  86         EXIT;
  87 }
  88
  89 static int
  90 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *blocking_lock)
  91 {
  92         struct obd_export *req_export = req->l_export;
  93         struct obd_export *blocking_export = blocking_lock->l_export;
  94         pid_t req_pid = req->l_policy_data.l_flock.pid;
  95         pid_t blocking_pid = blocking_lock->l_policy_data.l_flock.pid;
  96         struct ldlm_lock *lock;
  97
  98 restart:
  99         list_for_each_entry(lock, &ldlm_flock_waitq, l_flock_waitq) {
 100                 if ((lock->l_policy_data.l_flock.pid != blocking_pid) ||
 101                     (lock->l_export != blocking_export))
 102                         continue;
 103
 104                 blocking_pid = lock->l_policy_data.l_flock.blocking_pid;
 105                 blocking_export = (struct obd_export *)(long)
 106                         lock->l_policy_data.l_flock.blocking_export;
 107                 if (blocking_pid == req_pid && blocking_export == req_export)
 108                         return 1;
 109
 110                 goto restart;
 111         }
 112
 113         return 0;
 114 }
 115
 116 int
 117 ldlm_process_flock_lock(struct ldlm_lock *req, int *flags, int first_enq,
 118                         ldlm_error_t *err)
 119 {
 120         struct ldlm_resource *res = req->l_resource;
 121         struct ldlm_namespace *ns = res->lr_namespace;
 122         struct list_head *tmp;
 123         struct list_head *ownlocks = NULL;
 124         struct ldlm_lock *lock = NULL;
 125         struct ldlm_lock *new = req;
 126         struct ldlm_lock *new2 = NULL;
 127         ldlm_mode_t mode = req->l_req_mode;
 128         int added = (mode == LCK_NL);
 129         int overlaps = 0;
 130         ENTRY;
 131
 132         CDEBUG(D_DLMTRACE, "flags %#x pid %u mode %u start "LPU64" end "LPU64
 133                "\n", *flags, new->l_policy_data.l_flock.pid, mode,
 134                req->l_policy_data.l_flock.start,
 135                req->l_policy_data.l_flock.end);
 136
 137         *err = ELDLM_OK;
 138
 139         /* No blocking ASTs are sent for Posix file & record locks */
 140         req->l_blocking_ast = NULL;
 141
 142         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 143                 /* This loop determines where this processes locks start
 144                  * in the resource lr_granted list. */
 145                 list_for_each(tmp, &res->lr_granted) {
 146                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 147                         if (ldlm_same_flock_owner(lock, req)) {
 148                                 ownlocks = tmp;
 149                                 break;
 150                         }
 151                 }
 152         } else {
 153                 /* This loop determines if there are existing locks
 154                  * that conflict with the new lock request. */
 155                 list_for_each(tmp, &res->lr_granted) {
 156                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 157
 158                         if (ldlm_same_flock_owner(lock, req)) {
 159                                 if (!ownlocks)
 160                                         ownlocks = tmp;
 161                                 continue;
 162                         }
 163
 164                         /* locks are compatible, overlap doesn't matter */
 165                         if (lockmode_compat(lock->l_granted_mode, mode))
 166                                 continue;
 167
 168                         if (!ldlm_flocks_overlap(lock, req))
 169                                 continue;
 170
 171                         if (!first_enq)
 172                                 RETURN(LDLM_ITER_CONTINUE);
 173
 174                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 175                                 ldlm_flock_destroy(req, mode, *flags);
 176                                 *err = -EAGAIN;
 177                                 RETURN(LDLM_ITER_STOP);
 178                         }
 179
 180                         if (*flags & LDLM_FL_TEST_LOCK) {
 181                                 ldlm_flock_destroy(req, mode, *flags);
 182                                 req->l_req_mode = lock->l_granted_mode;
 183                                 req->l_policy_data.l_flock.pid =
 184                                         lock->l_policy_data.l_flock.pid;
 185                                 req->l_policy_data.l_flock.start =
 186                                         lock->l_policy_data.l_flock.start;
 187                                 req->l_policy_data.l_flock.end =
 188                                         lock->l_policy_data.l_flock.end;
 189                                 *flags |= LDLM_FL_LOCK_CHANGED;
 190                                 RETURN(LDLM_ITER_STOP);
 191                         }
 192
 193                         if (ldlm_flock_deadlock(req, lock)) {
 194                                 ldlm_flock_destroy(req, mode, *flags);
 195                                 *err = -EDEADLK;
 196                                 RETURN(LDLM_ITER_STOP);
 197                         }
 198
 199                         req->l_policy_data.l_flock.blocking_pid =
 200                                 lock->l_policy_data.l_flock.pid;
 201                         req->l_policy_data.l_flock.blocking_export =
 202                                 (long)(void *)lock->l_export;
 203
 204                         LASSERT(list_empty(&req->l_flock_waitq));
 205                         list_add_tail(&req->l_flock_waitq, &ldlm_flock_waitq);
 206
 207                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 208                         *flags |= LDLM_FL_BLOCK_GRANTED;
 209                         RETURN(LDLM_ITER_STOP);
 210                 }
 211         }
 212
 213         if (*flags & LDLM_FL_TEST_LOCK) {
 214                 ldlm_flock_destroy(req, mode, *flags);
 215                 req->l_req_mode = LCK_NL;
 216                 *flags |= LDLM_FL_LOCK_CHANGED;
 217                 RETURN(LDLM_ITER_STOP);
 218         }
 219
 220         /* Scan the locks owned by this process that overlap this request.
 221          * We may have to merge or split existing locks. */
 222
 223         if (!ownlocks)
 224                 ownlocks = &res->lr_granted;
 225
 226         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
 227                 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
 228
 229                 if (!ldlm_same_flock_owner(lock, new))
 230                         break;
 231
 232                 if (lock->l_granted_mode == mode) {
 233                         /* If the modes are the same then we need to process
 234                          * locks that overlap OR adjoin the new lock. The extra
 235                          * logic condition is necessary to deal with arithmetic
 236                          * overflow and underflow. */
 237                         if ((new->l_policy_data.l_flock.start >
 238                              (lock->l_policy_data.l_flock.end + 1))
 239                             && (lock->l_policy_data.l_flock.end != ~0))
 240                                 continue;
 241
 242                         if ((new->l_policy_data.l_flock.end <
 243                              (lock->l_policy_data.l_flock.start - 1))
 244                             && (lock->l_policy_data.l_flock.start != 0))
 245                                 break;
 246
 247                         if (new->l_policy_data.l_flock.start <
 248                             lock->l_policy_data.l_flock.start) {
 249                                 lock->l_policy_data.l_flock.start =
 250                                         new->l_policy_data.l_flock.start;
 251                         } else {
 252                                 new->l_policy_data.l_flock.start =
 253                                         lock->l_policy_data.l_flock.start;
 254                         }
 255
 256                         if (new->l_policy_data.l_flock.end >
 257                             lock->l_policy_data.l_flock.end) {
 258                                 lock->l_policy_data.l_flock.end =
 259                                         new->l_policy_data.l_flock.end;
 260                         } else {
 261                                 new->l_policy_data.l_flock.end =
 262                                         lock->l_policy_data.l_flock.end;
 263                         }
 264
 265                         if (added) {
 266                                 ldlm_flock_destroy(lock, mode, *flags);
 267                         } else {
 268                                 new = lock;
 269                                 added = 1;
 270                         }
 271                         continue;
 272                 }
 273
 274                 if (new->l_policy_data.l_flock.start >
 275                     lock->l_policy_data.l_flock.end)
 276                         continue;
 277
 278                 if (new->l_policy_data.l_flock.end <
 279                     lock->l_policy_data.l_flock.start)
 280                         break;
 281
 282                 ++overlaps;
 283
 284                 if (new->l_policy_data.l_flock.start <=
 285                     lock->l_policy_data.l_flock.start) {
 286                         if (new->l_policy_data.l_flock.end <
 287                             lock->l_policy_data.l_flock.end) {
 288                                 lock->l_policy_data.l_flock.start =
 289                                         new->l_policy_data.l_flock.end + 1;
 290                                 break;
 291                         }
 292                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 293                         continue;
 294                 }
 295                 if (new->l_policy_data.l_flock.end >=
 296                     lock->l_policy_data.l_flock.end) {
 297                         lock->l_policy_data.l_flock.end =
 298                                 new->l_policy_data.l_flock.start - 1;
 299                         continue;
 300                 }
 301
 302                 /* split the existing lock into two locks */
 303
 304                 /* if this is an F_UNLCK operation then we could avoid
 305                  * allocating a new lock and use the req lock passed in
 306                  * with the request but this would complicate the reply
 307                  * processing since updates to req get reflected in the
 308                  * reply. The client side replays the lock request so
 309                  * it must see the original lock data in the reply. */
 310
 311                 /* XXX - if ldlm_lock_new() can sleep we should
 312                  * release the ns_lock, allocate the new lock,
 313                  * and restart processing this lock. */
 314                 new2 = ldlm_lock_create(ns, NULL, res->lr_name, LDLM_FLOCK,
 315                                         lock->l_granted_mode, NULL, NULL, NULL);
 316                 if (!new2) {
 317                         ldlm_flock_destroy(req, lock->l_granted_mode, *flags);
 318                         *err = -ENOLCK;
 319                         RETURN(LDLM_ITER_STOP);
 320                 }
 321
 322                 new2->l_granted_mode = lock->l_granted_mode;
 323                 new2->l_policy_data.l_flock.pid =
 324                         new->l_policy_data.l_flock.pid;
 325                 new2->l_policy_data.l_flock.start =
 326                         lock->l_policy_data.l_flock.start;
 327                 new2->l_policy_data.l_flock.end =
 328                         new->l_policy_data.l_flock.start - 1;
 329                 lock->l_policy_data.l_flock.start =
 330                         new->l_policy_data.l_flock.end + 1;
 331                 new2->l_conn_export = lock->l_conn_export;
 332                 if (lock->l_export != NULL) {
 333                         new2->l_export = class_export_get(lock->l_export);
 334                         list_add(&new2->l_export_chain,
 335                                  &new2->l_export->exp_ldlm_data.led_held_locks);
 336                 }
 337                 if (*flags == LDLM_FL_WAIT_NOREPROC)
 338                         ldlm_lock_addref_internal(new2, lock->l_granted_mode);
 339
 340                 /* insert new2 at lock */
 341                 ldlm_resource_add_lock(res, ownlocks, new2);
 342                 LDLM_LOCK_PUT(new2);
 343                 break;
 344         }
 345
 346         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 347         if (!added) {
 348                 req->l_granted_mode = req->l_req_mode;
 349                 list_del_init(&req->l_res_link);
 350                 /* insert new lock before ownlocks in list. */
 351                 ldlm_resource_add_lock(res, ownlocks, req);
 352         }
 353
 354         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 355                 if (first_enq) {
 356                         /* If this is an unlock, reprocess the waitq and
 357                          * send completions ASTs for locks that can now be
 358                          * granted. The only problem with doing this
 359                          * reprocessing here is that the completion ASTs for
 360                          * newly granted locks will be sent before the unlock
 361                          * completion is sent. It shouldn't be an issue. Also
 362                          * note that ldlm_process_flock_lock() will recurse,
 363                          * but only once because first_enq will be false from
 364                          * ldlm_reprocess_queue. */
 365                         if ((mode == LCK_NL) && overlaps) {
 366                                 struct list_head rpc_list
 367                                                     = LIST_HEAD_INIT(rpc_list);
 368                                 int rc;
 369 restart:
 370                                 res->lr_tmp = &rpc_list;
 371                                 ldlm_reprocess_queue(res, &res->lr_waiting);
 372                                 res->lr_tmp = NULL;
 373
 374                                 l_unlock(&ns->ns_lock);
 375                                 rc = ldlm_run_ast_work(res->lr_namespace,
 376                                                        &rpc_list);
 377                                 l_lock(&ns->ns_lock);
 378                                 if (rc == -ERESTART)
 379                                         GOTO(restart, -ERESTART);
 380                        }
 381                 } else {
 382                         LASSERT(req->l_completion_ast);
 383                         ldlm_add_ast_work_item(req, NULL, NULL, 0);
 384                 }
 385         }
 386
 387         /* In case we're reprocessing the requested lock we can't destroy
 388          * it until after calling ldlm_ast_work_item() above so that lawi()
 389          * can bump the reference count on req. Otherwise req could be freed
 390          * before the completion AST can be sent.  */
 391         if (added)
 392                 ldlm_flock_destroy(req, mode, *flags);
 393
 394         ldlm_resource_dump(res);
 395         RETURN(LDLM_ITER_CONTINUE);
 396 }
 397
 398 struct ldlm_flock_wait_data {
 399         struct ldlm_lock *fwd_lock;
 400         int               fwd_generation;
 401 };
 402
 403 static void
 404 ldlm_flock_interrupted_wait(void *data)
 405 {
 406         struct ldlm_lock *lock;
 407         struct lustre_handle lockh;
 408         int rc;
 409         ENTRY;
 410
 411         lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock;
 412         ldlm_lock_decref_internal(lock, lock->l_req_mode);
 413         ldlm_lock2handle(lock, &lockh);
 414         rc = ldlm_cli_cancel(&lockh);
 415         CDEBUG(D_DLMTRACE, "ldlm_cli_cancel: %d\n", rc);
 416         EXIT;
 417 }
 418
 419 int
 420 ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 421 {
 422         struct ldlm_namespace *ns;
 423         struct file_lock *getlk = lock->l_ast_data;
 424         struct ldlm_flock_wait_data fwd;
 425         unsigned long irqflags;
 426         struct obd_device *obd;
 427         struct obd_import *imp = NULL;
 428         ldlm_error_t err;
 429         int rc = 0;
 430         struct l_wait_info lwi;
 431         ENTRY;
 432
 433         CDEBUG(D_DLMTRACE, "flags: 0x%x data: %p getlk: %p\n",
 434                flags, data, getlk);
 435
 436         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 437
 438         if (flags == 0) {
 439                 wake_up(&lock->l_waitq);
 440                 RETURN(0);
 441         }
 442
 443         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 444                        LDLM_FL_BLOCK_CONV)))
 445                 goto  granted;
 446
 447         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 448                    "sleeping");
 449
 450         ldlm_lock_dump(D_OTHER, lock, 0);
 451
 452         fwd.fwd_lock = lock;
 453         obd = class_exp2obd(lock->l_conn_export);
 454
 455         /* if this is a local lock, then there is no import */
 456         if (obd != NULL)
 457                 imp = obd->u.cli.cl_import;
 458
 459         if (imp != NULL) {
 460                 spin_lock_irqsave(&imp->imp_lock, irqflags);
 461                 fwd.fwd_generation = imp->imp_generation;
 462                 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
 463         }
 464
 465         lwi = LWI_TIMEOUT_INTR(0,NULL,ldlm_flock_interrupted_wait,&fwd);
 466
 467         /* Go to sleep until the lock is granted. */
 468         rc = l_wait_event(lock->l_waitq,
 469                           ((lock->l_req_mode == lock->l_granted_mode) ||
 470                            lock->l_destroyed), &lwi);
 471
 472         if (rc) {
 473                 LDLM_DEBUG(lock, "client-side enqueue waking up: failed (%d)",
 474                            rc);
 475                 RETURN(rc);
 476         }
 477
 478         LASSERT(!(lock->l_destroyed));
 479
 480 granted:
 481
 482         LDLM_DEBUG(lock, "client-side enqueue waking up");
 483         ns = lock->l_resource->lr_namespace;
 484         l_lock(&ns->ns_lock);
 485
 486         /* take data off of deadlock detection waitq. */
 487         list_del_init(&lock->l_flock_waitq);
 488
 489         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
 490         list_del_init(&lock->l_res_link);
 491
 492         if (flags & LDLM_FL_TEST_LOCK) {
 493                 /* fcntl(F_GETLK) request */
 494                 /* The old mode was saved in getlk->fl_type so that if the mode
 495                  * in the lock changes we can decref the approprate refcount. */
 496                 ldlm_flock_destroy(lock, getlk->fl_type, LDLM_FL_WAIT_NOREPROC);
 497                 switch (lock->l_granted_mode) {
 498                 case LCK_PR:
 499                         getlk->fl_type = F_RDLCK;
 500                         break;
 501                 case LCK_PW:
 502                         getlk->fl_type = F_WRLCK;
 503                         break;
 504                 default:
 505                         getlk->fl_type = F_UNLCK;
 506                 }
 507                 getlk->fl_pid = lock->l_policy_data.l_flock.pid;
 508                 getlk->fl_start = lock->l_policy_data.l_flock.start;
 509                 getlk->fl_end = lock->l_policy_data.l_flock.end;
 510         } else {
 511                 /* We need to reprocess the lock to do merges or splits
 512                  * with existing locks owne by this process. */
 513                 flags = LDLM_FL_WAIT_NOREPROC;
 514                 ldlm_process_flock_lock(lock, &flags, 1, &err);
 515         }
 516         l_unlock(&ns->ns_lock);
 517         RETURN(0);
 518 }