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) 2003 Hewlett-Packard Development Company LP.
   5  *   Developed under the sponsorship of the US Government under
   6  *   Subcontract No. B514193
   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 #include <linux/obd_class.h>
  35 #endif
  36
  37 #include "ldlm_internal.h"
  38
  39 #define l_flock_waitq   l_lru
  40
  41 static struct list_head ldlm_flock_waitq = LIST_HEAD_INIT(ldlm_flock_waitq);
  42
  43 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
  44                             void *data, int flag);
  45
  46 /**
  47  * list_for_remaining_safe - iterate over the remaining entries in a list
  48  *              and safeguard against removal of a list entry.
  49  * @pos:        the &struct list_head to use as a loop counter. pos MUST
  50  *              have been initialized prior to using it in this macro.
  51  * @n:          another &struct list_head to use as temporary storage
  52  * @head:       the head for your list.
  53  */
  54 #define list_for_remaining_safe(pos, n, head) \
  55         for (n = pos->next; pos != (head); pos = n, n = pos->next)
  56
  57 static inline int
  58 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
  59 {
  60         return((new->l_policy_data.l_flock.pid ==
  61                 lock->l_policy_data.l_flock.pid) &&
  62                (new->l_export == lock->l_export));
  63 }
  64
  65 static inline int
  66 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
  67 {
  68         return((new->l_policy_data.l_flock.start <=
  69                 lock->l_policy_data.l_flock.end) &&
  70                (new->l_policy_data.l_flock.end >=
  71                 lock->l_policy_data.l_flock.start));
  72 }
  73
  74 static inline void
  75 ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, int flags)
  76 {
  77         ENTRY;
  78
  79         LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: 0x%x)",
  80                    mode, flags);
  81
  82         LASSERT(list_empty(&lock->l_flock_waitq));
  83
  84         list_del_init(&lock->l_res_link);
  85         if (flags == LDLM_FL_WAIT_NOREPROC) {
  86                 /* client side - set a flag to prevent sending a CANCEL */
  87                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
  88                 ldlm_lock_decref_internal(lock, mode);
  89         }
  90
  91         ldlm_lock_destroy(lock);
  92         EXIT;
  93 }
  94
  95 static int
  96 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *blocking_lock)
  97 {
  98         struct obd_export *req_export = req->l_export;
  99         struct obd_export *blocking_export = blocking_lock->l_export;
 100         pid_t req_pid = req->l_policy_data.l_flock.pid;
 101         pid_t blocking_pid = blocking_lock->l_policy_data.l_flock.pid;
 102         struct ldlm_lock *lock;
 103
 104 restart:
 105         list_for_each_entry(lock, &ldlm_flock_waitq, l_flock_waitq) {
 106                 if ((lock->l_policy_data.l_flock.pid != blocking_pid) ||
 107                     (lock->l_export != blocking_export))
 108                         continue;
 109
 110                 blocking_pid = lock->l_policy_data.l_flock.blocking_pid;
 111                 blocking_export = (struct obd_export *)(long)
 112                         lock->l_policy_data.l_flock.blocking_export;
 113                 if (blocking_pid == req_pid && blocking_export == req_export)
 114                         return 1;
 115
 116                 goto restart;
 117         }
 118
 119         return 0;
 120 }
 121
 122 int
 123 ldlm_process_flock_lock(struct ldlm_lock *req, int *flags, int first_enq,
 124                         ldlm_error_t *err)
 125 {
 126         struct ldlm_resource *res = req->l_resource;
 127         struct ldlm_namespace *ns = res->lr_namespace;
 128         struct list_head *tmp;
 129         struct list_head *ownlocks = NULL;
 130         struct ldlm_lock *lock = NULL;
 131         struct ldlm_lock *new = req;
 132         struct ldlm_lock *new2 = NULL;
 133         ldlm_mode_t mode = req->l_req_mode;
 134         int local = ns->ns_client;
 135         int added = (mode == LCK_NL);
 136         int overlaps = 0;
 137         ENTRY;
 138
 139         CDEBUG(D_DLMTRACE, "flags %#x pid %u mode %u start "LPU64" end "LPU64
 140                "\n", *flags, new->l_policy_data.l_flock.pid, mode,
 141                req->l_policy_data.l_flock.start,
 142                req->l_policy_data.l_flock.end);
 143
 144         *err = ELDLM_OK;
 145
 146         if (local) {
 147                 /* No blocking ASTs are sent to the clients for
 148                  * Posix file & record locks */
 149                 req->l_blocking_ast = NULL;
 150         } else {
 151                 /* Called on the server for lock cancels. */
 152                 req->l_blocking_ast = ldlm_flock_blocking_ast;
 153         }
 154
 155         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 156                 /* This loop determines where this processes locks start
 157                  * in the resource lr_granted list. */
 158                 list_for_each(tmp, &res->lr_granted) {
 159                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 160                         if (ldlm_same_flock_owner(lock, req)) {
 161                                 ownlocks = tmp;
 162                                 break;
 163                         }
 164                 }
 165         } else {
 166                 lockmode_verify(mode);
 167
 168                 /* This loop determines if there are existing locks
 169                  * that conflict with the new lock request. */
 170                 list_for_each(tmp, &res->lr_granted) {
 171                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 172
 173                         if (ldlm_same_flock_owner(lock, req)) {
 174                                 if (!ownlocks)
 175                                         ownlocks = tmp;
 176                                 continue;
 177                         }
 178
 179                         /* locks are compatible, overlap doesn't matter */
 180                         if (lockmode_compat(lock->l_granted_mode, mode))
 181                                 continue;
 182
 183                         if (!ldlm_flocks_overlap(lock, req))
 184                                 continue;
 185
 186                         if (!first_enq)
 187                                 RETURN(LDLM_ITER_CONTINUE);
 188
 189                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 190                                 ldlm_flock_destroy(req, mode, *flags);
 191                                 *err = -EAGAIN;
 192                                 RETURN(LDLM_ITER_STOP);
 193                         }
 194
 195                         if (*flags & LDLM_FL_TEST_LOCK) {
 196                                 ldlm_flock_destroy(req, mode, *flags);
 197                                 req->l_req_mode = lock->l_granted_mode;
 198                                 req->l_policy_data.l_flock.pid =
 199                                         lock->l_policy_data.l_flock.pid;
 200                                 req->l_policy_data.l_flock.start =
 201                                         lock->l_policy_data.l_flock.start;
 202                                 req->l_policy_data.l_flock.end =
 203                                         lock->l_policy_data.l_flock.end;
 204                                 *flags |= LDLM_FL_LOCK_CHANGED;
 205                                 RETURN(LDLM_ITER_STOP);
 206                         }
 207
 208                         if (ldlm_flock_deadlock(req, lock)) {
 209                                 ldlm_flock_destroy(req, mode, *flags);
 210                                 *err = -EDEADLK;
 211                                 RETURN(LDLM_ITER_STOP);
 212                         }
 213
 214                         req->l_policy_data.l_flock.blocking_pid =
 215                                 lock->l_policy_data.l_flock.pid;
 216                         req->l_policy_data.l_flock.blocking_export =
 217                                 (long)(void *)lock->l_export;
 218
 219                         LASSERT(list_empty(&req->l_flock_waitq));
 220                         list_add_tail(&req->l_flock_waitq, &ldlm_flock_waitq);
 221
 222                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 223                         *flags |= LDLM_FL_BLOCK_GRANTED;
 224                         RETURN(LDLM_ITER_STOP);
 225                 }
 226         }
 227
 228         if (*flags & LDLM_FL_TEST_LOCK) {
 229                 ldlm_flock_destroy(req, mode, *flags);
 230                 req->l_req_mode = LCK_NL;
 231                 *flags |= LDLM_FL_LOCK_CHANGED;
 232                 RETURN(LDLM_ITER_STOP);
 233         }
 234
 235         /* In case we had slept on this lock request take it off of the
 236          * deadlock detection waitq. */
 237         list_del_init(&req->l_flock_waitq);
 238
 239         /* Scan the locks owned by this process that overlap this request.
 240          * We may have to merge or split existing locks. */
 241
 242         if (!ownlocks)
 243                 ownlocks = &res->lr_granted;
 244
 245         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
 246                 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
 247
 248                 if (!ldlm_same_flock_owner(lock, new))
 249                         break;
 250
 251                 if (lock->l_granted_mode == mode) {
 252                         /* If the modes are the same then we need to process
 253                          * locks that overlap OR adjoin the new lock. The extra
 254                          * logic condition is necessary to deal with arithmetic
 255                          * overflow and underflow. */
 256                         if ((new->l_policy_data.l_flock.start >
 257                              (lock->l_policy_data.l_flock.end + 1))
 258                             && (lock->l_policy_data.l_flock.end != ~0))
 259                                 continue;
 260
 261                         if ((new->l_policy_data.l_flock.end <
 262                              (lock->l_policy_data.l_flock.start - 1))
 263                             && (lock->l_policy_data.l_flock.start != 0))
 264                                 break;
 265
 266                         if (new->l_policy_data.l_flock.start <
 267                             lock->l_policy_data.l_flock.start) {
 268                                 lock->l_policy_data.l_flock.start =
 269                                         new->l_policy_data.l_flock.start;
 270                         } else {
 271                                 new->l_policy_data.l_flock.start =
 272                                         lock->l_policy_data.l_flock.start;
 273                         }
 274
 275                         if (new->l_policy_data.l_flock.end >
 276                             lock->l_policy_data.l_flock.end) {
 277                                 lock->l_policy_data.l_flock.end =
 278                                         new->l_policy_data.l_flock.end;
 279                         } else {
 280                                 new->l_policy_data.l_flock.end =
 281                                         lock->l_policy_data.l_flock.end;
 282                         }
 283
 284                         if (added) {
 285                                 ldlm_flock_destroy(lock, mode, *flags);
 286                         } else {
 287                                 new = lock;
 288                                 added = 1;
 289                         }
 290                         continue;
 291                 }
 292
 293                 if (new->l_policy_data.l_flock.start >
 294                     lock->l_policy_data.l_flock.end)
 295                         continue;
 296
 297                 if (new->l_policy_data.l_flock.end <
 298                     lock->l_policy_data.l_flock.start)
 299                         break;
 300
 301                 ++overlaps;
 302
 303                 if (new->l_policy_data.l_flock.start <=
 304                     lock->l_policy_data.l_flock.start) {
 305                         if (new->l_policy_data.l_flock.end <
 306                             lock->l_policy_data.l_flock.end) {
 307                                 lock->l_policy_data.l_flock.start =
 308                                         new->l_policy_data.l_flock.end + 1;
 309                                 break;
 310                         }
 311                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 312                         continue;
 313                 }
 314                 if (new->l_policy_data.l_flock.end >=
 315                     lock->l_policy_data.l_flock.end) {
 316                         lock->l_policy_data.l_flock.end =
 317                                 new->l_policy_data.l_flock.start - 1;
 318                         continue;
 319                 }
 320
 321                 /* split the existing lock into two locks */
 322
 323                 /* if this is an F_UNLCK operation then we could avoid
 324                  * allocating a new lock and use the req lock passed in
 325                  * with the request but this would complicate the reply
 326                  * processing since updates to req get reflected in the
 327                  * reply. The client side replays the lock request so
 328                  * it must see the original lock data in the reply. */
 329
 330                 /* XXX - if ldlm_lock_new() can sleep we should
 331                  * release the ns_lock, allocate the new lock,
 332                  * and restart processing this lock. */
 333                 new2 = ldlm_lock_create(ns, NULL, res->lr_name, LDLM_FLOCK,
 334                                         lock->l_granted_mode, NULL, NULL, NULL,
 335                                         NULL, 0);
 336                 if (!new2) {
 337                         ldlm_flock_destroy(req, lock->l_granted_mode, *flags);
 338                         *err = -ENOLCK;
 339                         RETURN(LDLM_ITER_STOP);
 340                 }
 341
 342                 new2->l_granted_mode = lock->l_granted_mode;
 343                 new2->l_policy_data.l_flock.pid =
 344                         new->l_policy_data.l_flock.pid;
 345                 new2->l_policy_data.l_flock.start =
 346                         lock->l_policy_data.l_flock.start;
 347                 new2->l_policy_data.l_flock.end =
 348                         new->l_policy_data.l_flock.start - 1;
 349                 lock->l_policy_data.l_flock.start =
 350                         new->l_policy_data.l_flock.end + 1;
 351                 new2->l_conn_export = lock->l_conn_export;
 352                 if (lock->l_export != NULL) {
 353                         new2->l_export = class_export_get(lock->l_export);
 354                         list_add(&new2->l_export_chain,
 355                                  &new2->l_export->exp_ldlm_data.led_held_locks);
 356                 }
 357                 if (*flags == LDLM_FL_WAIT_NOREPROC)
 358                         ldlm_lock_addref_internal(new2, lock->l_granted_mode);
 359
 360                 /* insert new2 at lock */
 361                 ldlm_resource_add_lock(res, ownlocks, new2);
 362                 LDLM_LOCK_PUT(new2);
 363                 break;
 364         }
 365
 366         /* At this point we're granting the lock request. */
 367         req->l_granted_mode = req->l_req_mode;
 368
 369         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 370         if (!added) {
 371                 list_del_init(&req->l_res_link);
 372                 /* insert new lock before ownlocks in list. */
 373                 ldlm_resource_add_lock(res, ownlocks, req);
 374         }
 375
 376         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 377                 if (first_enq) {
 378                         /* If this is an unlock, reprocess the waitq and
 379                          * send completions ASTs for locks that can now be
 380                          * granted. The only problem with doing this
 381                          * reprocessing here is that the completion ASTs for
 382                          * newly granted locks will be sent before the unlock
 383                          * completion is sent. It shouldn't be an issue. Also
 384                          * note that ldlm_process_flock_lock() will recurse,
 385                          * but only once because first_enq will be false from
 386                          * ldlm_reprocess_queue. */
 387                         if ((mode == LCK_NL) && overlaps) {
 388                                 struct list_head rpc_list
 389                                                     = LIST_HEAD_INIT(rpc_list);
 390                                 int rc;
 391 restart:
 392                                 res->lr_tmp = &rpc_list;
 393                                 ldlm_reprocess_queue(res, &res->lr_waiting);
 394                                 res->lr_tmp = NULL;
 395
 396                                 l_unlock(&ns->ns_lock);
 397                                 rc = ldlm_run_ast_work(res->lr_namespace,
 398                                                        &rpc_list);
 399                                 l_lock(&ns->ns_lock);
 400                                 if (rc == -ERESTART)
 401                                         GOTO(restart, -ERESTART);
 402                        }
 403                 } else {
 404                         LASSERT(req->l_completion_ast);
 405                         ldlm_add_ast_work_item(req, NULL, NULL, 0);
 406                 }
 407         }
 408
 409         /* In case we're reprocessing the requested lock we can't destroy
 410          * it until after calling ldlm_ast_work_item() above so that lawi()
 411          * can bump the reference count on req. Otherwise req could be freed
 412          * before the completion AST can be sent.  */
 413         if (added)
 414                 ldlm_flock_destroy(req, mode, *flags);
 415
 416         ldlm_resource_dump(D_OTHER, res);
 417         RETURN(LDLM_ITER_CONTINUE);
 418 }
 419
 420 struct ldlm_flock_wait_data {
 421         struct ldlm_lock *fwd_lock;
 422         int               fwd_generation;
 423 };
 424
 425 static void
 426 ldlm_flock_interrupted_wait(void *data)
 427 {
 428         struct ldlm_lock *lock;
 429         struct lustre_handle lockh;
 430         ENTRY;
 431
 432         lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock;
 433
 434         /* take lock off the deadlock detection waitq. */
 435         list_del_init(&lock->l_flock_waitq);
 436
 437         /* client side - set flag to prevent lock from being put on lru list */
 438         lock->l_flags |= LDLM_FL_CBPENDING;
 439
 440         ldlm_lock_decref_internal(lock, lock->l_req_mode);
 441         ldlm_lock2handle(lock, &lockh);
 442         ldlm_cli_cancel(&lockh);
 443         EXIT;
 444 }
 445
 446 int
 447 ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 448 {
 449         struct ldlm_namespace *ns;
 450         struct file_lock *getlk = lock->l_ast_data;
 451         struct ldlm_flock_wait_data fwd;
 452         unsigned long irqflags;
 453         struct obd_device *obd;
 454         struct obd_import *imp = NULL;
 455         ldlm_error_t err;
 456         int rc = 0;
 457         struct l_wait_info lwi;
 458         ENTRY;
 459
 460         CDEBUG(D_DLMTRACE, "flags: 0x%x data: %p getlk: %p\n",
 461                flags, data, getlk);
 462
 463         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 464
 465         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 466                        LDLM_FL_BLOCK_CONV)))
 467                 goto  granted;
 468
 469         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 470                    "sleeping");
 471
 472         fwd.fwd_lock = lock;
 473         obd = class_exp2obd(lock->l_conn_export);
 474
 475         /* if this is a local lock, then there is no import */
 476         if (obd != NULL)
 477                 imp = obd->u.cli.cl_import;
 478
 479         if (imp != NULL) {
 480                 spin_lock_irqsave(&imp->imp_lock, irqflags);
 481                 fwd.fwd_generation = imp->imp_generation;
 482                 spin_unlock_irqrestore(&imp->imp_lock, irqflags);
 483         }
 484
 485         lwi = LWI_TIMEOUT_INTR(0, NULL, ldlm_flock_interrupted_wait, &fwd);
 486
 487         /* Go to sleep until the lock is granted. */
 488         rc = l_wait_event(lock->l_waitq,
 489                           ((lock->l_req_mode == lock->l_granted_mode) ||
 490                            lock->l_destroyed), &lwi);
 491
 492         LDLM_DEBUG(lock, "client-side enqueue waking up: rc = %d", rc);
 493         RETURN(rc);
 494
 495 granted:
 496
 497         LDLM_DEBUG(lock, "client-side enqueue granted");
 498         ns = lock->l_resource->lr_namespace;
 499         l_lock(&ns->ns_lock);
 500
 501         /* take lock off the deadlock detection waitq. */
 502         list_del_init(&lock->l_flock_waitq);
 503
 504         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
 505         list_del_init(&lock->l_res_link);
 506
 507         if (flags & LDLM_FL_TEST_LOCK) {
 508                 /* fcntl(F_GETLK) request */
 509                 /* The old mode was saved in getlk->fl_type so that if the mode
 510                  * in the lock changes we can decref the approprate refcount. */
 511                 ldlm_flock_destroy(lock, getlk->fl_type, LDLM_FL_WAIT_NOREPROC);
 512                 switch (lock->l_granted_mode) {
 513                 case LCK_PR:
 514                         getlk->fl_type = F_RDLCK;
 515                         break;
 516                 case LCK_PW:
 517                         getlk->fl_type = F_WRLCK;
 518                         break;
 519                 default:
 520                         getlk->fl_type = F_UNLCK;
 521                 }
 522                 getlk->fl_pid = lock->l_policy_data.l_flock.pid;
 523                 getlk->fl_start = lock->l_policy_data.l_flock.start;
 524                 getlk->fl_end = lock->l_policy_data.l_flock.end;
 525         } else {
 526                 int noreproc = LDLM_FL_WAIT_NOREPROC;
 527
 528                 /* We need to reprocess the lock to do merges or splits
 529                  * with existing locks owned by this process. */
 530                 ldlm_process_flock_lock(lock, &noreproc, 1, &err);
 531                 if (flags == 0)
 532                         wake_up(&lock->l_waitq);
 533         }
 534         l_unlock(&ns->ns_lock);
 535         RETURN(0);
 536 }
 537 EXPORT_SYMBOL(ldlm_flock_completion_ast);
 538
 539 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 540                             void *data, int flag)
 541 {
 542         struct ldlm_namespace *ns;
 543         ENTRY;
 544
 545         LASSERT(lock);
 546         LASSERT(flag == LDLM_CB_CANCELING);
 547
 548         ns = lock->l_resource->lr_namespace;
 549
 550         /* take lock off the deadlock detection waitq. */
 551         l_lock(&ns->ns_lock);
 552         list_del_init(&lock->l_flock_waitq);
 553         l_unlock(&ns->ns_lock);
 554         RETURN(0);
 555 }