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 <libcfs/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 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
  43                             void *data, int flag);
  44
  45 /**
  46  * list_for_remaining_safe - iterate over the remaining entries in a list
  47  *              and safeguard against removal of a list entry.
  48  * @pos:        the &struct list_head to use as a loop counter. pos MUST
  49  *              have been initialized prior to using it in this macro.
  50  * @n:          another &struct list_head to use as temporary storage
  51  * @head:       the head for your list.
  52  */
  53 #define list_for_remaining_safe(pos, n, head) \
  54         for (n = pos->next; pos != (head); pos = n, n = pos->next)
  55
  56 static inline int
  57 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
  58 {
  59         return((new->l_policy_data.l_flock.pid ==
  60                 lock->l_policy_data.l_flock.pid) &&
  61                (new->l_export == lock->l_export));
  62 }
  63
  64 static inline int
  65 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
  66 {
  67         return((new->l_policy_data.l_flock.start <=
  68                 lock->l_policy_data.l_flock.end) &&
  69                (new->l_policy_data.l_flock.end >=
  70                 lock->l_policy_data.l_flock.start));
  71 }
  72
  73 static inline void
  74 ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, int flags)
  75 {
  76         ENTRY;
  77
  78         LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: 0x%x)",
  79                    mode, flags);
  80
  81         LASSERT(list_empty(&lock->l_flock_waitq));
  82         list_del_init(&lock->l_res_link);
  83         if (flags == LDLM_FL_WAIT_NOREPROC) {
  84                 /* client side - set a flag to prevent sending a CANCEL */
  85                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
  86                 ldlm_lock_decref_internal(lock, mode);
  87         }
  88
  89         ldlm_lock_destroy(lock);
  90         EXIT;
  91 }
  92
  93 static int
  94 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *blocking_lock)
  95 {
  96         struct obd_export *req_export = req->l_export;
  97         struct obd_export *blocking_export = blocking_lock->l_export;
  98         pid_t req_pid = req->l_policy_data.l_flock.pid;
  99         pid_t blocking_pid = blocking_lock->l_policy_data.l_flock.pid;
 100         struct ldlm_lock *lock;
 101
 102 restart:
 103         list_for_each_entry(lock, &ldlm_flock_waitq, l_flock_waitq) {
 104                 if ((lock->l_policy_data.l_flock.pid != blocking_pid) ||
 105                     (lock->l_export != blocking_export))
 106                         continue;
 107
 108                 blocking_pid = lock->l_policy_data.l_flock.blocking_pid;
 109                 blocking_export = (struct obd_export *)(long)
 110                         lock->l_policy_data.l_flock.blocking_export;
 111                 if (blocking_pid == req_pid && blocking_export == req_export)
 112                         return 1;
 113
 114                 goto restart;
 115         }
 116
 117         return 0;
 118 }
 119
 120 int
 121 ldlm_process_flock_lock(struct ldlm_lock *req, int *flags, int first_enq,
 122                         ldlm_error_t *err, struct list_head *work_list)
 123 {
 124         struct ldlm_resource *res = req->l_resource;
 125         struct ldlm_namespace *ns = res->lr_namespace;
 126         struct list_head *tmp;
 127         struct list_head *ownlocks = NULL;
 128         struct ldlm_lock *lock = NULL;
 129         struct ldlm_lock *new = req;
 130         struct ldlm_lock *new2 = NULL;
 131         ldlm_mode_t mode = req->l_req_mode;
 132         int local = ns->ns_client;
 133         int added = (mode == LCK_NL);
 134         int overlaps = 0;
 135         ENTRY;
 136
 137         CDEBUG(D_DLMTRACE, "flags %#x pid %u mode %u start "LPU64" end "
 138                LPU64"\n", *flags, (unsigned int)new->l_policy_data.l_flock.pid,
 139                mode, req->l_policy_data.l_flock.start,
 140                req->l_policy_data.l_flock.end);
 141
 142         *err = ELDLM_OK;
 143
 144         if (local) {
 145                 /* No blocking ASTs are sent to the clients for
 146                  * Posix file & record locks */
 147                 req->l_blocking_ast = NULL;
 148         } else {
 149                 /* Called on the server for lock cancels. */
 150                 req->l_blocking_ast = ldlm_flock_blocking_ast;
 151         }
 152
 153         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 154                 /* This loop determines where this processes locks start
 155                  * in the resource lr_granted list. */
 156                 list_for_each(tmp, &res->lr_granted) {
 157                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 158                         if (ldlm_same_flock_owner(lock, req)) {
 159                                 ownlocks = tmp;
 160                                 break;
 161                         }
 162                 }
 163         } else {
 164                 lockmode_verify(mode);
 165
 166                 /* This loop determines if there are existing locks
 167                  * that conflict with the new lock request. */
 168                 list_for_each(tmp, &res->lr_granted) {
 169                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 170
 171                         if (ldlm_same_flock_owner(lock, req)) {
 172                                 if (!ownlocks)
 173                                         ownlocks = tmp;
 174                                 continue;
 175                         }
 176
 177                         /* locks are compatible, overlap doesn't matter */
 178                         if (lockmode_compat(lock->l_granted_mode, mode))
 179                                 continue;
 180
 181                         if (!ldlm_flocks_overlap(lock, req))
 182                                 continue;
 183
 184                         if (!first_enq)
 185                                 RETURN(LDLM_ITER_CONTINUE);
 186
 187                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 188                                 ldlm_flock_destroy(req, mode, *flags);
 189                                 *err = -EAGAIN;
 190                                 RETURN(LDLM_ITER_STOP);
 191                         }
 192
 193                         if (*flags & LDLM_FL_TEST_LOCK) {
 194                                 ldlm_flock_destroy(req, mode, *flags);
 195                                 req->l_req_mode = lock->l_granted_mode;
 196                                 req->l_policy_data.l_flock.pid =
 197                                         lock->l_policy_data.l_flock.pid;
 198                                 req->l_policy_data.l_flock.start =
 199                                         lock->l_policy_data.l_flock.start;
 200                                 req->l_policy_data.l_flock.end =
 201                                         lock->l_policy_data.l_flock.end;
 202                                 *flags |= LDLM_FL_LOCK_CHANGED;
 203                                 RETURN(LDLM_ITER_STOP);
 204                         }
 205
 206                         if (ldlm_flock_deadlock(req, lock)) {
 207                                 ldlm_flock_destroy(req, mode, *flags);
 208                                 *err = -EDEADLK;
 209                                 RETURN(LDLM_ITER_STOP);
 210                         }
 211
 212                         req->l_policy_data.l_flock.blocking_pid =
 213                                 lock->l_policy_data.l_flock.pid;
 214                         req->l_policy_data.l_flock.blocking_export =
 215                                 (long)(void *)lock->l_export;
 216
 217                         LASSERT(list_empty(&req->l_flock_waitq));
 218                         list_add_tail(&req->l_flock_waitq, &ldlm_flock_waitq);
 219
 220                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 221                         *flags |= LDLM_FL_BLOCK_GRANTED;
 222                         RETURN(LDLM_ITER_STOP);
 223                 }
 224         }
 225
 226         if (*flags & LDLM_FL_TEST_LOCK) {
 227                 ldlm_flock_destroy(req, mode, *flags);
 228                 req->l_req_mode = LCK_NL;
 229                 *flags |= LDLM_FL_LOCK_CHANGED;
 230                 RETURN(LDLM_ITER_STOP);
 231         }
 232
 233         /* In case we had slept on this lock request take it off of the
 234          * deadlock detection waitq. */
 235         list_del_init(&req->l_flock_waitq);
 236
 237         /* Scan the locks owned by this process that overlap this request.
 238          * We may have to merge or split existing locks. */
 239
 240         if (!ownlocks)
 241                 ownlocks = &res->lr_granted;
 242
 243         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
 244                 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
 245
 246                 if (!ldlm_same_flock_owner(lock, new))
 247                         break;
 248
 249                 if (lock->l_granted_mode == mode) {
 250                         /* If the modes are the same then we need to process
 251                          * locks that overlap OR adjoin the new lock. The extra
 252                          * logic condition is necessary to deal with arithmetic
 253                          * overflow and underflow. */
 254                         if ((new->l_policy_data.l_flock.start >
 255                              (lock->l_policy_data.l_flock.end + 1))
 256                             && (lock->l_policy_data.l_flock.end != ~0))
 257                                 continue;
 258
 259                         if ((new->l_policy_data.l_flock.end <
 260                              (lock->l_policy_data.l_flock.start - 1))
 261                             && (lock->l_policy_data.l_flock.start != 0))
 262                                 break;
 263
 264                         if (new->l_policy_data.l_flock.start <
 265                             lock->l_policy_data.l_flock.start) {
 266                                 lock->l_policy_data.l_flock.start =
 267                                         new->l_policy_data.l_flock.start;
 268                         } else {
 269                                 new->l_policy_data.l_flock.start =
 270                                         lock->l_policy_data.l_flock.start;
 271                         }
 272
 273                         if (new->l_policy_data.l_flock.end >
 274                             lock->l_policy_data.l_flock.end) {
 275                                 lock->l_policy_data.l_flock.end =
 276                                         new->l_policy_data.l_flock.end;
 277                         } else {
 278                                 new->l_policy_data.l_flock.end =
 279                                         lock->l_policy_data.l_flock.end;
 280                         }
 281
 282                         if (added) {
 283                                 ldlm_flock_destroy(lock, mode, *flags);
 284                         } else {
 285                                 new = lock;
 286                                 added = 1;
 287                         }
 288                         continue;
 289                 }
 290
 291                 if (new->l_policy_data.l_flock.start >
 292                     lock->l_policy_data.l_flock.end)
 293                         continue;
 294
 295                 if (new->l_policy_data.l_flock.end <
 296                     lock->l_policy_data.l_flock.start)
 297                         break;
 298
 299                 ++overlaps;
 300
 301                 if (new->l_policy_data.l_flock.start <=
 302                     lock->l_policy_data.l_flock.start) {
 303                         if (new->l_policy_data.l_flock.end <
 304                             lock->l_policy_data.l_flock.end) {
 305                                 lock->l_policy_data.l_flock.start =
 306                                         new->l_policy_data.l_flock.end + 1;
 307                                 break;
 308                         }
 309                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 310                         continue;
 311                 }
 312                 if (new->l_policy_data.l_flock.end >=
 313                     lock->l_policy_data.l_flock.end) {
 314                         lock->l_policy_data.l_flock.end =
 315                                 new->l_policy_data.l_flock.start - 1;
 316                         continue;
 317                 }
 318
 319                 /* split the existing lock into two locks */
 320
 321                 /* if this is an F_UNLCK operation then we could avoid
 322                  * allocating a new lock and use the req lock passed in
 323                  * with the request but this would complicate the reply
 324                  * processing since updates to req get reflected in the
 325                  * reply. The client side replays the lock request so
 326                  * it must see the original lock data in the reply. */
 327
 328                 /* XXX - if ldlm_lock_new() can sleep we should
 329                  * release the ns_lock, allocate the new lock,
 330                  * and restart processing this lock. */
 331                 new2 = ldlm_lock_create(ns, NULL, res->lr_name, LDLM_FLOCK,
 332                                         lock->l_granted_mode, NULL, NULL, NULL,
 333                                         NULL, 0);
 334                 if (!new2) {
 335                         ldlm_flock_destroy(req, lock->l_granted_mode, *flags);
 336                         *err = -ENOLCK;
 337                         RETURN(LDLM_ITER_STOP);
 338                 }
 339
 340                 new2->l_granted_mode = lock->l_granted_mode;
 341                 new2->l_policy_data.l_flock.pid =
 342                         new->l_policy_data.l_flock.pid;
 343                 new2->l_policy_data.l_flock.start =
 344                         lock->l_policy_data.l_flock.start;
 345                 new2->l_policy_data.l_flock.end =
 346                         new->l_policy_data.l_flock.start - 1;
 347                 lock->l_policy_data.l_flock.start =
 348                         new->l_policy_data.l_flock.end + 1;
 349                 new2->l_conn_export = lock->l_conn_export;
 350                 if (lock->l_export != NULL) {
 351                         new2->l_export = class_export_get(lock->l_export);
 352                         list_add(&new2->l_export_chain,
 353                                  &new2->l_export->exp_ldlm_data.led_held_locks);
 354                 }
 355                 if (*flags == LDLM_FL_WAIT_NOREPROC)
 356                         ldlm_lock_addref_internal_nolock(new2, lock->l_granted_mode);
 357
 358                 /* insert new2 at lock */
 359                 ldlm_resource_add_lock(res, ownlocks, new2);
 360                 LDLM_LOCK_PUT(new2);
 361                 break;
 362         }
 363
 364         /* At this point we're granting the lock request. */
 365         req->l_granted_mode = req->l_req_mode;
 366
 367         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 368         if (!added) {
 369                 list_del_init(&req->l_res_link);
 370                 /* insert new lock before ownlocks in list. */
 371                 ldlm_resource_add_lock(res, ownlocks, req);
 372         }
 373
 374         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 375                 if (first_enq) {
 376                         /* If this is an unlock, reprocess the waitq and
 377                          * send completions ASTs for locks that can now be
 378                          * granted. The only problem with doing this
 379                          * reprocessing here is that the completion ASTs for
 380                          * newly granted locks will be sent before the unlock
 381                          * completion is sent. It shouldn't be an issue. Also
 382                          * note that ldlm_process_flock_lock() will recurse,
 383                          * but only once because first_enq will be false from
 384                          * ldlm_reprocess_queue. */
 385                         if ((mode == LCK_NL) && overlaps) {
 386                                 struct list_head rpc_list
 387                                                     = LIST_HEAD_INIT(rpc_list);
 388                                 int rc;
 389 restart:
 390                                 ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
 391                                 unlock_res(res);
 392                                 rc = ldlm_run_cp_ast_work(&rpc_list);
 393                                 lock_res(res);
 394                                 if (rc == -ERESTART)
 395                                         GOTO(restart, -ERESTART);
 396                        }
 397                 } else {
 398                         LASSERT(req->l_completion_ast);
 399                         ldlm_add_ast_work_item(req, NULL, NULL);
 400                 }
 401         }
 402
 403         /* In case we're reprocessing the requested lock we can't destroy
 404          * it until after calling ldlm_ast_work_item() above so that lawi()
 405          * can bump the reference count on req. Otherwise req could be freed
 406          * before the completion AST can be sent.  */
 407         if (added)
 408                 ldlm_flock_destroy(req, mode, *flags);
 409
 410         ldlm_resource_dump(D_OTHER, res);
 411         RETURN(LDLM_ITER_CONTINUE);
 412 }
 413
 414 struct ldlm_flock_wait_data {
 415         struct ldlm_lock *fwd_lock;
 416         int               fwd_generation;
 417 };
 418
 419 static void
 420 ldlm_flock_interrupted_wait(void *data)
 421 {
 422         struct ldlm_lock *lock;
 423         struct lustre_handle lockh;
 424         ENTRY;
 425
 426         lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock;
 427
 428         /* take lock off the deadlock detection waitq. */
 429         list_del_init(&lock->l_flock_waitq);
 430
 431         /* client side - set flag to prevent lock from being put on lru list */
 432         lock->l_flags |= LDLM_FL_CBPENDING;
 433
 434         ldlm_lock_decref_internal(lock, lock->l_req_mode);
 435         ldlm_lock2handle(lock, &lockh);
 436         ldlm_cli_cancel(&lockh);
 437         EXIT;
 438 }
 439
 440 int
 441 ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 442 {
 443         struct ldlm_namespace *ns;
 444         struct file_lock *getlk = lock->l_ast_data;
 445         struct ldlm_flock_wait_data fwd;
 446         unsigned long irqflags;
 447         struct obd_device *obd;
 448         struct obd_import *imp = NULL;
 449         ldlm_error_t err;
 450         int rc = 0;
 451         struct l_wait_info lwi;
 452         ENTRY;
 453
 454         CDEBUG(D_DLMTRACE, "flags: 0x%x data: %p getlk: %p\n",
 455                flags, data, getlk);
 456
 457         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 458
 459         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 460                        LDLM_FL_BLOCK_CONV)))
 461                 goto  granted;
 462
 463         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 464                    "sleeping");
 465
 466         ldlm_lock_dump(D_DLMTRACE, lock, 0);
 467         fwd.fwd_lock = lock;
 468         obd = class_exp2obd(lock->l_conn_export);
 469
 470         /* if this is a local lock, then there is no import */
 471         if (obd != NULL)
 472                 imp = obd->u.cli.cl_import;
 473
 474         if (imp != NULL) {
 475                 spin_lock_irqsave(&imp->imp_lock, irqflags);
 476                 fwd.fwd_generation = imp->imp_generation;
 477                 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
 478         }
 479
 480         lwi = LWI_TIMEOUT_INTR(0, NULL, ldlm_flock_interrupted_wait, &fwd);
 481
 482         /* Go to sleep until the lock is granted. */
 483         rc = l_wait_event(lock->l_waitq,
 484                           ((lock->l_req_mode == lock->l_granted_mode) ||
 485                            lock->l_destroyed), &lwi);
 486
 487         LDLM_DEBUG(lock, "client-side enqueue waking up: rc = %d", rc);
 488         RETURN(rc);
 489
 490 granted:
 491
 492         LDLM_DEBUG(lock, "client-side enqueue granted");
 493         ns = lock->l_resource->lr_namespace;
 494         lock_res(lock->l_resource);
 495
 496         /* take lock off the deadlock detection waitq. */
 497         list_del_init(&lock->l_flock_waitq);
 498
 499         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
 500         list_del_init(&lock->l_res_link);
 501
 502         if (flags & LDLM_FL_TEST_LOCK) {
 503                 /* fcntl(F_GETLK) request */
 504                 /* The old mode was saved in getlk->fl_type so that if the mode
 505                  * in the lock changes we can decref the approprate refcount. */
 506                 ldlm_flock_destroy(lock, getlk->fl_type, LDLM_FL_WAIT_NOREPROC);
 507                 switch (lock->l_granted_mode) {
 508                 case LCK_PR:
 509                         getlk->fl_type = F_RDLCK;
 510                         break;
 511                 case LCK_PW:
 512                         getlk->fl_type = F_WRLCK;
 513                         break;
 514                 default:
 515                         getlk->fl_type = F_UNLCK;
 516                 }
 517                 getlk->fl_pid = lock->l_policy_data.l_flock.pid;
 518                 getlk->fl_start = lock->l_policy_data.l_flock.start;
 519                 getlk->fl_end = lock->l_policy_data.l_flock.end;
 520         } else {
 521                 int noreproc = LDLM_FL_WAIT_NOREPROC;
 522
 523                 /* We need to reprocess the lock to do merges or splits
 524                  * with existing locks owned by this process. */
 525                 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
 526                 if (flags == 0)
 527                         wake_up(&lock->l_waitq);
 528         }
 529         unlock_res(lock->l_resource);
 530         RETURN(0);
 531 }
 532
 533 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 534                             void *data, int flag)
 535 {
 536         ENTRY;
 537
 538         LASSERT(lock);
 539         LASSERT(flag == LDLM_CB_CANCELING);
 540
 541         /* take lock off the deadlock detection waitq. */
 542         lock_res(lock->l_resource);
 543         list_del_init(&lock->l_flock_waitq);
 544         unlock_res(lock->l_resource);
 545         RETURN(0);
 546 }