lustre/lov/lovsub_lock.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Implementation of cl_lock for LOVSUB layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_LOV
  42
  43 #include "lov_cl_internal.h"
  44
  45 /** \addtogroup lov
  46  *  @{
  47  */
  48
  49 /*****************************************************************************
  50  *
  51  * Lovsub lock operations.
  52  *
  53  */
  54
  55 static void lovsub_lock_fini(const struct lu_env *env,
  56                              struct cl_lock_slice *slice)
  57 {
  58         struct lovsub_lock   *lsl;
  59
  60         ENTRY;
  61         lsl = cl2lovsub_lock(slice);
  62         LASSERT(list_empty(&lsl->lss_parents));
  63         OBD_SLAB_FREE_PTR(lsl, lovsub_lock_kmem);
  64         EXIT;
  65 }
  66
  67 static void lovsub_parent_lock(const struct lu_env *env, struct lov_lock *lov)
  68 {
  69         struct cl_lock *parent;
  70
  71         ENTRY;
  72         parent = lov->lls_cl.cls_lock;
  73         cl_lock_get(parent);
  74         lu_ref_add(&parent->cll_reference, "lovsub-parent", cfs_current());
  75         cl_lock_mutex_get(env, parent);
  76         EXIT;
  77 }
  78
  79 static void lovsub_parent_unlock(const struct lu_env *env, struct lov_lock *lov)
  80 {
  81         struct cl_lock *parent;
  82
  83         ENTRY;
  84         parent = lov->lls_cl.cls_lock;
  85         cl_lock_mutex_put(env, lov->lls_cl.cls_lock);
  86         lu_ref_del(&parent->cll_reference, "lovsub-parent", cfs_current());
  87         cl_lock_put(env, parent);
  88         EXIT;
  89 }
  90
  91 static int lovsub_lock_state_one(const struct lu_env *env,
  92                                  const struct lovsub_lock *lovsub,
  93                                  struct lov_lock *lov)
  94 {
  95         struct cl_lock *parent;
  96         struct cl_lock *child;
  97         int             restart = 0;
  98
  99         ENTRY;
 100         parent = lov->lls_cl.cls_lock;
 101         child  = lovsub->lss_cl.cls_lock;
 102
 103         if (lovsub->lss_active != parent) {
 104                 lovsub_parent_lock(env, lov);
 105                 if (child->cll_error != 0 && parent->cll_error == 0) {
 106                         /*
 107                          * This is a deadlock case:
 108                          * cl_lock_error(for the parent lock)
 109                          *   -> cl_lock_delete
 110                          *     -> lov_lock_delete
 111                          *       -> cl_lock_enclosure
 112                          *         -> cl_lock_mutex_try(for the child lock)
 113                          */
 114                         cl_lock_mutex_put(env, child);
 115                         cl_lock_error(env, parent, child->cll_error);
 116                         restart = 1;
 117                 } else {
 118                         cl_lock_signal(env, parent);
 119                 }
 120                 lovsub_parent_unlock(env, lov);
 121         }
 122         RETURN(restart);
 123 }
 124
 125 /**
 126  * Implements cl_lock_operations::clo_state() method for lovsub layer, which
 127  * method is called whenever sub-lock state changes. Propagates state change
 128  * to the top-locks.
 129  */
 130 static void lovsub_lock_state(const struct lu_env *env,
 131                               const struct cl_lock_slice *slice,
 132                               enum cl_lock_state state)
 133 {
 134         struct lovsub_lock   *sub = cl2lovsub_lock(slice);
 135         struct lov_lock_link *scan;
 136         int                   restart = 0;
 137
 138         LASSERT(cl_lock_is_mutexed(slice->cls_lock));
 139         ENTRY;
 140
 141         do {
 142                 restart = 0;
 143                 list_for_each_entry(scan, &sub->lss_parents, lll_list) {
 144                         restart = lovsub_lock_state_one(env, sub,
 145                                                         scan->lll_super);
 146                         if (restart) {
 147                                 cl_lock_mutex_get(env, slice->cls_lock);
 148                                 break;
 149                         }
 150                 }
 151         } while(restart);
 152         EXIT;
 153 }
 154
 155 /**
 156  * Implementation of cl_lock_operation::clo_weigh() estimating lock weight by
 157  * asking parent lock.
 158  */
 159 static unsigned long lovsub_lock_weigh(const struct lu_env *env,
 160                                        const struct cl_lock_slice *slice)
 161 {
 162         struct lovsub_lock *lock = cl2lovsub_lock(slice);
 163         struct lov_lock    *lov;
 164         unsigned long       dumbbell;
 165
 166         ENTRY;
 167
 168         LASSERT(cl_lock_is_mutexed(slice->cls_lock));
 169
 170         if (!list_empty(&lock->lss_parents)) {
 171                 /*
 172                  * It is not clear whether all parents have to be asked and
 173                  * their estimations summed, or it is enough to ask one. For
 174                  * the current usages, one is always enough.
 175                  */
 176                 lov = container_of(lock->lss_parents.next,
 177                                    struct lov_lock_link, lll_list)->lll_super;
 178
 179                 lovsub_parent_lock(env, lov);
 180                 dumbbell = cl_lock_weigh(env, lov->lls_cl.cls_lock);
 181                 lovsub_parent_unlock(env, lov);
 182         } else
 183                 dumbbell = 0;
 184
 185         RETURN(dumbbell);
 186 }
 187
 188 /**
 189  * Maps start/end offsets within a stripe, to offsets within a file.
 190  */
 191 static void lovsub_lock_descr_map(const struct cl_lock_descr *in,
 192                                   struct lov_object *obj,
 193                                   int stripe, struct cl_lock_descr *out)
 194 {
 195         struct lov_stripe_md *lsm = lov_r0(obj)->lo_lsm;
 196         pgoff_t size; /* stripe size in pages */
 197         pgoff_t skip; /* how many pages in every stripe are occupied by
 198                        * "other" stripes */
 199         pgoff_t start;
 200         pgoff_t end;
 201
 202         ENTRY;
 203         start = in->cld_start;
 204         end   = in->cld_end;
 205
 206         if (lsm->lsm_stripe_count > 1) {
 207                 size = cl_index(lov2cl(obj), lsm->lsm_stripe_size);
 208                 skip = (lsm->lsm_stripe_count - 1) * size;
 209
 210                 /* XXX overflow check here? */
 211                 start += start/size * skip + stripe * size;
 212
 213                 if (end != CL_PAGE_EOF) {
 214                         end += end/size * skip + stripe * size;
 215                         /*
 216                          * And check for overflow...
 217                          */
 218                         if (end < in->cld_end)
 219                                 end = CL_PAGE_EOF;
 220                 }
 221         }
 222         out->cld_start = start;
 223         out->cld_end   = end;
 224         EXIT;
 225 }
 226
 227 /**
 228  * Adjusts parent lock extent when a sub-lock is attached to a parent. This is
 229  * called in two ways:
 230  *
 231  *     - as part of receive call-back, when server returns granted extent to
 232  *       the client, and
 233  *
 234  *     - when top-lock finds existing sub-lock in the cache.
 235  *
 236  * Note, that lock mode is not propagated to the parent: i.e., if CLM_READ
 237  * top-lock matches CLM_WRITE sub-lock, top-lock is still CLM_READ.
 238  */
 239 int lov_sublock_modify(const struct lu_env *env, struct lov_lock *lov,
 240                        struct lovsub_lock *sublock,
 241                        const struct cl_lock_descr *d, int idx)
 242 {
 243         struct cl_lock       *parent;
 244         struct cl_lock       *child;
 245         struct lovsub_object *subobj;
 246         struct cl_lock_descr *pd;
 247         struct cl_lock_descr *parent_descr;
 248         int                   result;
 249
 250         parent       = lov->lls_cl.cls_lock;
 251         parent_descr = &parent->cll_descr;
 252         LASSERT(cl_lock_mode_match(d->cld_mode, parent_descr->cld_mode));
 253
 254         child  = sublock->lss_cl.cls_lock;
 255         subobj = cl2lovsub(sublock->lss_cl.cls_obj);
 256         pd     = &lov_env_info(env)->lti_ldescr;
 257
 258         pd->cld_obj  = parent_descr->cld_obj;
 259         pd->cld_mode = parent_descr->cld_mode;
 260         pd->cld_gid  = parent_descr->cld_gid;
 261         lovsub_lock_descr_map(d, subobj->lso_super, subobj->lso_index, pd);
 262         lov->lls_sub[idx].sub_got = *d;
 263         /*
 264          * Notify top-lock about modification, if lock description changes
 265          * materially.
 266          */
 267         if (!cl_lock_ext_match(parent_descr, pd))
 268                 result = cl_lock_modify(env, parent, pd);
 269         else
 270                 result = 0;
 271         return result;
 272 }
 273
 274 static int lovsub_lock_modify(const struct lu_env *env,
 275                               const struct cl_lock_slice *s,
 276                               const struct cl_lock_descr *d)
 277 {
 278         struct lovsub_lock   *lock   = cl2lovsub_lock(s);
 279         struct lov_lock_link *scan;
 280         struct lov_lock      *lov;
 281         int result                   = 0;
 282
 283         ENTRY;
 284
 285         LASSERT(cl_lock_mode_match(d->cld_mode,
 286                                    s->cls_lock->cll_descr.cld_mode));
 287         list_for_each_entry(scan, &lock->lss_parents, lll_list) {
 288                 int rc;
 289
 290                 lov = scan->lll_super;
 291                 lovsub_parent_lock(env, lov);
 292                 rc = lov_sublock_modify(env, lov, lock, d, scan->lll_idx);
 293                 lovsub_parent_unlock(env, lov);
 294                 result = result ?: rc;
 295         }
 296         RETURN(result);
 297 }
 298
 299 static int lovsub_lock_closure(const struct lu_env *env,
 300                                const struct cl_lock_slice *slice,
 301                                struct cl_lock_closure *closure)
 302 {
 303         struct lovsub_lock   *sub;
 304         struct cl_lock       *parent;
 305         struct lov_lock_link *scan;
 306         int                   result;
 307
 308         LASSERT(cl_lock_is_mutexed(slice->cls_lock));
 309         ENTRY;
 310
 311         sub    = cl2lovsub_lock(slice);
 312         result = 0;
 313
 314         list_for_each_entry(scan, &sub->lss_parents, lll_list) {
 315                 parent = scan->lll_super->lls_cl.cls_lock;
 316                 result = cl_lock_closure_build(env, parent, closure);
 317                 if (result != 0)
 318                         break;
 319         }
 320         RETURN(result);
 321 }
 322
 323 /**
 324  * A helper function for lovsub_lock_delete() that deals with a given parent
 325  * top-lock.
 326  */
 327 static int lovsub_lock_delete_one(const struct lu_env *env,
 328                                   struct cl_lock *child, struct lov_lock *lov)
 329 {
 330         struct cl_lock       *parent;
 331         int             result;
 332         ENTRY;
 333
 334         parent  = lov->lls_cl.cls_lock;
 335         result = 0;
 336
 337         switch (parent->cll_state) {
 338         case CLS_NEW:
 339         case CLS_QUEUING:
 340         case CLS_ENQUEUED:
 341         case CLS_FREEING:
 342                 cl_lock_signal(env, parent);
 343                 break;
 344         case CLS_INTRANSIT:
 345                 /*
 346                  * Here lies a problem: a sub-lock is canceled while top-lock
 347                  * is being unlocked. Top-lock cannot be moved into CLS_NEW
 348                  * state, because unlocking has to succeed eventually by
 349                  * placing lock into CLS_CACHED (or failing it), see
 350                  * cl_unuse_try(). Nor can top-lock be left in CLS_CACHED
 351                  * state, because lov maintains an invariant that all
 352                  * sub-locks exist in CLS_CACHED (this allows cached top-lock
 353                  * to be reused immediately). Nor can we wait for top-lock
 354                  * state to change, because this can be synchronous to the
 355                  * current thread.
 356                  *
 357                  * We know for sure that lov_lock_unuse() will be called at
 358                  * least one more time to finish un-using, so leave a mark on
 359                  * the top-lock, that will be seen by the next call to
 360                  * lov_lock_unuse().
 361                  */
 362                 if (cl_lock_is_intransit(parent))
 363                         lov->lls_cancel_race = 1;
 364                 break;
 365         case CLS_CACHED:
 366                 /*
 367                  * if a sub-lock is canceled move its top-lock into CLS_NEW
 368                  * state to preserve an invariant that a top-lock in
 369                  * CLS_CACHED is immediately ready for re-use (i.e., has all
 370                  * sub-locks), and so that next attempt to re-use the top-lock
 371                  * enqueues missing sub-lock.
 372                  */
 373                 cl_lock_state_set(env, parent, CLS_NEW);
 374                 /*
 375                  * if last sub-lock is canceled, destroy the top-lock (which
 376                  * is now `empty') proactively.
 377                  */
 378                 if (lov->lls_nr_filled == 0) {
 379                         /* ... but unfortunately, this cannot be done easily,
 380                          * as cancellation of a top-lock might acquire mutices
 381                          * of its other sub-locks, violating lock ordering,
 382                          * see cl_lock_{cancel,delete}() preconditions.
 383                          *
 384                          * To work around this, the mutex of this sub-lock is
 385                          * released, top-lock is destroyed, and sub-lock mutex
 386                          * acquired again. The list of parents has to be
 387                          * re-scanned from the beginning after this.
 388                          *
 389                          * Only do this if no mutices other than on @child and
 390                          * @parent are held by the current thread.
 391                          *
 392                          * TODO: The lock modal here is too complex, because
 393                          * the lock may be canceled and deleted by voluntarily:
 394                          *    cl_lock_request
 395                          *      -> osc_lock_enqueue_wait
 396                          *        -> osc_lock_cancel_wait
 397                          *          -> cl_lock_delete
 398                          *            -> lovsub_lock_delete
 399                          *              -> cl_lock_cancel/delete
 400                          *                -> ...
 401                          *
 402                          * The better choice is to spawn a kernel thread for
 403                          * this purpose. -jay
 404                          */
 405                         if (cl_lock_nr_mutexed(env) == 2) {
 406                                 cl_lock_mutex_put(env, child);
 407                                 cl_lock_cancel(env, parent);
 408                                 cl_lock_delete(env, parent);
 409                                 result = 1;
 410                         }
 411                 }
 412                 break;
 413         case CLS_HELD:
 414         default:
 415                 LASSERTF(parent->cll_error != 0, "cll state %d is wrong!\n",
 416                          parent->cll_state);
 417                 break;
 418         }
 419
 420         RETURN(result);
 421 }
 422
 423 /**
 424  * An implementation of cl_lock_operations::clo_delete() method. This is
 425  * invoked in "bottom-to-top" delete, when lock destruction starts from the
 426  * sub-lock (e.g, as a result of ldlm lock LRU policy).
 427  */
 428 static void lovsub_lock_delete(const struct lu_env *env,
 429                                const struct cl_lock_slice *slice)
 430 {
 431         struct cl_lock     *child = slice->cls_lock;
 432         struct lovsub_lock *sub   = cl2lovsub_lock(slice);
 433         int restart;
 434
 435         LASSERT(cl_lock_is_mutexed(child));
 436
 437         ENTRY;
 438         /*
 439          * Destruction of a sub-lock might take multiple iterations, because
 440          * when the last sub-lock of a given top-lock is deleted, top-lock is
 441          * canceled proactively, and this requires to release sub-lock
 442          * mutex. Once sub-lock mutex has been released, list of its parents
 443          * has to be re-scanned from the beginning.
 444          */
 445         do {
 446                 struct lov_lock      *lov;
 447                 struct lov_lock_link *scan;
 448                 struct lov_lock_link *temp;
 449                 struct lov_lock_sub  *subdata;
 450
 451                 restart = 0;
 452                 list_for_each_entry_safe(scan, temp,
 453                                          &sub->lss_parents, lll_list) {
 454                         lov     = scan->lll_super;
 455                         subdata = &lov->lls_sub[scan->lll_idx];
 456                         lovsub_parent_lock(env, lov);
 457                         subdata->sub_got = subdata->sub_descr;
 458                         lov_lock_unlink(env, scan, sub);
 459                         restart = lovsub_lock_delete_one(env, child, lov);
 460                         lovsub_parent_unlock(env, lov);
 461
 462                         if (restart) {
 463                                 cl_lock_mutex_get(env, child);
 464                                 break;
 465                         }
 466                }
 467         } while (restart);
 468         EXIT;
 469 }
 470
 471 static int lovsub_lock_print(const struct lu_env *env, void *cookie,
 472                              lu_printer_t p, const struct cl_lock_slice *slice)
 473 {
 474         struct lovsub_lock   *sub = cl2lovsub_lock(slice);
 475         struct lov_lock      *lov;
 476         struct lov_lock_link *scan;
 477
 478         list_for_each_entry(scan, &sub->lss_parents, lll_list) {
 479                 lov = scan->lll_super;
 480                 (*p)(env, cookie, "[%d %p ", scan->lll_idx, lov);
 481                 if (lov != NULL)
 482                         cl_lock_descr_print(env, cookie, p,
 483                                             &lov->lls_cl.cls_lock->cll_descr);
 484                 (*p)(env, cookie, "] ");
 485         }
 486         return 0;
 487 }
 488
 489 static const struct cl_lock_operations lovsub_lock_ops = {
 490         .clo_fini    = lovsub_lock_fini,
 491         .clo_state   = lovsub_lock_state,
 492         .clo_delete  = lovsub_lock_delete,
 493         .clo_modify  = lovsub_lock_modify,
 494         .clo_closure = lovsub_lock_closure,
 495         .clo_weigh   = lovsub_lock_weigh,
 496         .clo_print   = lovsub_lock_print
 497 };
 498
 499 int lovsub_lock_init(const struct lu_env *env, struct cl_object *obj,
 500                      struct cl_lock *lock, const struct cl_io *io)
 501 {
 502         struct lovsub_lock *lsk;
 503         int result;
 504
 505         ENTRY;
 506         OBD_SLAB_ALLOC_PTR_GFP(lsk, lovsub_lock_kmem, CFS_ALLOC_IO);
 507         if (lsk != NULL) {
 508                 CFS_INIT_LIST_HEAD(&lsk->lss_parents);
 509                 cl_lock_slice_add(lock, &lsk->lss_cl, obj, &lovsub_lock_ops);
 510                 result = 0;
 511         } else
 512                 result = -ENOMEM;
 513         RETURN(result);
 514 }
 515
 516 /** @} lov */