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 (c) 2008, 2010, Oracle and/or its affiliates. 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(cfs_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 /**
  92  * Implements cl_lock_operations::clo_state() method for lovsub layer, which
  93  * method is called whenever sub-lock state changes. Propagates state change
  94  * to the top-locks.
  95  */
  96 static void lovsub_lock_state(const struct lu_env *env,
  97                               const struct cl_lock_slice *slice,
  98                               enum cl_lock_state state)
  99 {
 100         struct lovsub_lock   *sub = cl2lovsub_lock(slice);
 101         struct lov_lock_link *scan;
 102
 103         LASSERT(cl_lock_is_mutexed(slice->cls_lock));
 104         ENTRY;
 105
 106         cfs_list_for_each_entry(scan, &sub->lss_parents, lll_list) {
 107                 struct lov_lock *lov    = scan->lll_super;
 108                 struct cl_lock  *parent = lov->lls_cl.cls_lock;
 109
 110                 if (sub->lss_active != parent) {
 111                         lovsub_parent_lock(env, lov);
 112                         cl_lock_signal(env, parent);
 113                         lovsub_parent_unlock(env, lov);
 114                 }
 115         }
 116         EXIT;
 117 }
 118
 119 /**
 120  * Implementation of cl_lock_operation::clo_weigh() estimating lock weight by
 121  * asking parent lock.
 122  */
 123 static unsigned long lovsub_lock_weigh(const struct lu_env *env,
 124                                        const struct cl_lock_slice *slice)
 125 {
 126         struct lovsub_lock *lock = cl2lovsub_lock(slice);
 127         struct lov_lock    *lov;
 128         unsigned long       dumbbell;
 129
 130         ENTRY;
 131
 132         LASSERT(cl_lock_is_mutexed(slice->cls_lock));
 133
 134         if (!cfs_list_empty(&lock->lss_parents)) {
 135                 /*
 136                  * It is not clear whether all parents have to be asked and
 137                  * their estimations summed, or it is enough to ask one. For
 138                  * the current usages, one is always enough.
 139                  */
 140                 lov = container_of(lock->lss_parents.next,
 141                                    struct lov_lock_link, lll_list)->lll_super;
 142
 143                 lovsub_parent_lock(env, lov);
 144                 dumbbell = cl_lock_weigh(env, lov->lls_cl.cls_lock);
 145                 lovsub_parent_unlock(env, lov);
 146         } else
 147                 dumbbell = 0;
 148
 149         RETURN(dumbbell);
 150 }
 151
 152 /**
 153  * Maps start/end offsets within a stripe, to offsets within a file.
 154  */
 155 static void lovsub_lock_descr_map(const struct cl_lock_descr *in,
 156                                   struct lov_object *obj,
 157                                   int stripe, struct cl_lock_descr *out)
 158 {
 159         struct lov_stripe_md *lsm = lov_r0(obj)->lo_lsm;
 160         pgoff_t size; /* stripe size in pages */
 161         pgoff_t skip; /* how many pages in every stripe are occupied by
 162                        * "other" stripes */
 163         pgoff_t start;
 164         pgoff_t end;
 165
 166         ENTRY;
 167         start = in->cld_start;
 168         end   = in->cld_end;
 169
 170         if (lsm->lsm_stripe_count > 1) {
 171                 size = cl_index(lov2cl(obj), lsm->lsm_stripe_size);
 172                 skip = (lsm->lsm_stripe_count - 1) * size;
 173
 174                 /* XXX overflow check here? */
 175                 start += start/size * skip + stripe * size;
 176
 177                 if (end != CL_PAGE_EOF) {
 178                         end += end/size * skip + stripe * size;
 179                         /*
 180                          * And check for overflow...
 181                          */
 182                         if (end < in->cld_end)
 183                                 end = CL_PAGE_EOF;
 184                 }
 185         }
 186         out->cld_start = start;
 187         out->cld_end   = end;
 188         EXIT;
 189 }
 190
 191 /**
 192  * Adjusts parent lock extent when a sub-lock is attached to a parent. This is
 193  * called in two ways:
 194  *
 195  *     - as part of receive call-back, when server returns granted extent to
 196  *       the client, and
 197  *
 198  *     - when top-lock finds existing sub-lock in the cache.
 199  *
 200  * Note, that lock mode is not propagated to the parent: i.e., if CLM_READ
 201  * top-lock matches CLM_WRITE sub-lock, top-lock is still CLM_READ.
 202  */
 203 int lov_sublock_modify(const struct lu_env *env, struct lov_lock *lov,
 204                        struct lovsub_lock *sublock,
 205                        const struct cl_lock_descr *d, int idx)
 206 {
 207         struct cl_lock       *parent;
 208         struct cl_lock       *child;
 209         struct lovsub_object *subobj;
 210         struct cl_lock_descr *pd;
 211         struct cl_lock_descr *parent_descr;
 212         int                   result;
 213
 214         parent       = lov->lls_cl.cls_lock;
 215         parent_descr = &parent->cll_descr;
 216         LASSERT(cl_lock_mode_match(d->cld_mode, parent_descr->cld_mode));
 217
 218         child  = sublock->lss_cl.cls_lock;
 219         subobj = cl2lovsub(sublock->lss_cl.cls_obj);
 220         pd     = &lov_env_info(env)->lti_ldescr;
 221
 222         pd->cld_obj  = parent_descr->cld_obj;
 223         pd->cld_mode = parent_descr->cld_mode;
 224         pd->cld_gid  = parent_descr->cld_gid;
 225         lovsub_lock_descr_map(d, subobj->lso_super, subobj->lso_index, pd);
 226         lov->lls_sub[idx].sub_got = *d;
 227         /*
 228          * Notify top-lock about modification, if lock description changes
 229          * materially.
 230          */
 231         if (!cl_lock_ext_match(parent_descr, pd))
 232                 result = cl_lock_modify(env, parent, pd);
 233         else
 234                 result = 0;
 235         return result;
 236 }
 237
 238 static int lovsub_lock_modify(const struct lu_env *env,
 239                               const struct cl_lock_slice *s,
 240                               const struct cl_lock_descr *d)
 241 {
 242         struct lovsub_lock   *lock   = cl2lovsub_lock(s);
 243         struct lov_lock_link *scan;
 244         struct lov_lock      *lov;
 245         int result                   = 0;
 246
 247         ENTRY;
 248
 249         LASSERT(cl_lock_mode_match(d->cld_mode,
 250                                    s->cls_lock->cll_descr.cld_mode));
 251         cfs_list_for_each_entry(scan, &lock->lss_parents, lll_list) {
 252                 int rc;
 253
 254                 lov = scan->lll_super;
 255                 lovsub_parent_lock(env, lov);
 256                 rc = lov_sublock_modify(env, lov, lock, d, scan->lll_idx);
 257                 lovsub_parent_unlock(env, lov);
 258                 result = result ?: rc;
 259         }
 260         RETURN(result);
 261 }
 262
 263 static int lovsub_lock_closure(const struct lu_env *env,
 264                                const struct cl_lock_slice *slice,
 265                                struct cl_lock_closure *closure)
 266 {
 267         struct lovsub_lock   *sub;
 268         struct cl_lock       *parent;
 269         struct lov_lock_link *scan;
 270         int                   result;
 271
 272         LASSERT(cl_lock_is_mutexed(slice->cls_lock));
 273         ENTRY;
 274
 275         sub    = cl2lovsub_lock(slice);
 276         result = 0;
 277
 278         cfs_list_for_each_entry(scan, &sub->lss_parents, lll_list) {
 279                 parent = scan->lll_super->lls_cl.cls_lock;
 280                 result = cl_lock_closure_build(env, parent, closure);
 281                 if (result != 0)
 282                         break;
 283         }
 284         RETURN(result);
 285 }
 286
 287 /**
 288  * A helper function for lovsub_lock_delete() that deals with a given parent
 289  * top-lock.
 290  */
 291 static int lovsub_lock_delete_one(const struct lu_env *env,
 292                                   struct cl_lock *child, struct lov_lock *lov)
 293 {
 294         struct cl_lock *parent;
 295         int             result;
 296         ENTRY;
 297
 298         parent = lov->lls_cl.cls_lock;
 299         if (parent->cll_error)
 300                 RETURN(0);
 301
 302         result = 0;
 303         switch (parent->cll_state) {
 304         case CLS_ENQUEUED:
 305                 /* See LU-1355 for the case that a glimpse lock is
 306                  * interrupted by signal */
 307                 LASSERT(parent->cll_flags & CLF_CANCELLED);
 308                 break;
 309         case CLS_QUEUING:
 310         case CLS_FREEING:
 311                 cl_lock_signal(env, parent);
 312                 break;
 313         case CLS_INTRANSIT:
 314                 /*
 315                  * Here lies a problem: a sub-lock is canceled while top-lock
 316                  * is being unlocked. Top-lock cannot be moved into CLS_NEW
 317                  * state, because unlocking has to succeed eventually by
 318                  * placing lock into CLS_CACHED (or failing it), see
 319                  * cl_unuse_try(). Nor can top-lock be left in CLS_CACHED
 320                  * state, because lov maintains an invariant that all
 321                  * sub-locks exist in CLS_CACHED (this allows cached top-lock
 322                  * to be reused immediately). Nor can we wait for top-lock
 323                  * state to change, because this can be synchronous to the
 324                  * current thread.
 325                  *
 326                  * We know for sure that lov_lock_unuse() will be called at
 327                  * least one more time to finish un-using, so leave a mark on
 328                  * the top-lock, that will be seen by the next call to
 329                  * lov_lock_unuse().
 330                  */
 331                 if (cl_lock_is_intransit(parent))
 332                         lov->lls_cancel_race = 1;
 333                 break;
 334         case CLS_CACHED:
 335                 /*
 336                  * if a sub-lock is canceled move its top-lock into CLS_NEW
 337                  * state to preserve an invariant that a top-lock in
 338                  * CLS_CACHED is immediately ready for re-use (i.e., has all
 339                  * sub-locks), and so that next attempt to re-use the top-lock
 340                  * enqueues missing sub-lock.
 341                  */
 342                 cl_lock_state_set(env, parent, CLS_NEW);
 343                 /* fall through */
 344         case CLS_NEW:
 345                 /*
 346                  * if last sub-lock is canceled, destroy the top-lock (which
 347                  * is now `empty') proactively.
 348                  */
 349                 if (lov->lls_nr_filled == 0) {
 350                         /* ... but unfortunately, this cannot be done easily,
 351                          * as cancellation of a top-lock might acquire mutices
 352                          * of its other sub-locks, violating lock ordering,
 353                          * see cl_lock_{cancel,delete}() preconditions.
 354                          *
 355                          * To work around this, the mutex of this sub-lock is
 356                          * released, top-lock is destroyed, and sub-lock mutex
 357                          * acquired again. The list of parents has to be
 358                          * re-scanned from the beginning after this.
 359                          *
 360                          * Only do this if no mutices other than on @child and
 361                          * @parent are held by the current thread.
 362                          *
 363                          * TODO: The lock modal here is too complex, because
 364                          * the lock may be canceled and deleted by voluntarily:
 365                          *    cl_lock_request
 366                          *      -> osc_lock_enqueue_wait
 367                          *        -> osc_lock_cancel_wait
 368                          *          -> cl_lock_delete
 369                          *            -> lovsub_lock_delete
 370                          *              -> cl_lock_cancel/delete
 371                          *                -> ...
 372                          *
 373                          * The better choice is to spawn a kernel thread for
 374                          * this purpose. -jay
 375                          */
 376                         if (cl_lock_nr_mutexed(env) == 2) {
 377                                 cl_lock_mutex_put(env, child);
 378                                 cl_lock_cancel(env, parent);
 379                                 cl_lock_delete(env, parent);
 380                                 result = 1;
 381                         }
 382                 }
 383                 break;
 384         case CLS_HELD:
 385                 CL_LOCK_DEBUG(D_ERROR, env, parent, "Delete CLS_HELD lock\n");
 386         default:
 387                 CERROR("Impossible state: %d\n", parent->cll_state);
 388                 LBUG();
 389                 break;
 390         }
 391
 392         RETURN(result);
 393 }
 394
 395 /**
 396  * An implementation of cl_lock_operations::clo_delete() method. This is
 397  * invoked in "bottom-to-top" delete, when lock destruction starts from the
 398  * sub-lock (e.g, as a result of ldlm lock LRU policy).
 399  */
 400 static void lovsub_lock_delete(const struct lu_env *env,
 401                                const struct cl_lock_slice *slice)
 402 {
 403         struct cl_lock     *child = slice->cls_lock;
 404         struct lovsub_lock *sub   = cl2lovsub_lock(slice);
 405         int restart;
 406
 407         LASSERT(cl_lock_is_mutexed(child));
 408
 409         ENTRY;
 410         /*
 411          * Destruction of a sub-lock might take multiple iterations, because
 412          * when the last sub-lock of a given top-lock is deleted, top-lock is
 413          * canceled proactively, and this requires to release sub-lock
 414          * mutex. Once sub-lock mutex has been released, list of its parents
 415          * has to be re-scanned from the beginning.
 416          */
 417         do {
 418                 struct lov_lock      *lov;
 419                 struct lov_lock_link *scan;
 420                 struct lov_lock_link *temp;
 421                 struct lov_lock_sub  *subdata;
 422
 423                 restart = 0;
 424                 cfs_list_for_each_entry_safe(scan, temp,
 425                                              &sub->lss_parents, lll_list) {
 426                         lov     = scan->lll_super;
 427                         subdata = &lov->lls_sub[scan->lll_idx];
 428                         lovsub_parent_lock(env, lov);
 429                         subdata->sub_got = subdata->sub_descr;
 430                         lov_lock_unlink(env, scan, sub);
 431                         restart = lovsub_lock_delete_one(env, child, lov);
 432                         lovsub_parent_unlock(env, lov);
 433
 434                         if (restart) {
 435                                 cl_lock_mutex_get(env, child);
 436                                 break;
 437                         }
 438                }
 439         } while (restart);
 440         EXIT;
 441 }
 442
 443 static int lovsub_lock_print(const struct lu_env *env, void *cookie,
 444                              lu_printer_t p, const struct cl_lock_slice *slice)
 445 {
 446         struct lovsub_lock   *sub = cl2lovsub_lock(slice);
 447         struct lov_lock      *lov;
 448         struct lov_lock_link *scan;
 449
 450         cfs_list_for_each_entry(scan, &sub->lss_parents, lll_list) {
 451                 lov = scan->lll_super;
 452                 (*p)(env, cookie, "[%d %p ", scan->lll_idx, lov);
 453                 if (lov != NULL)
 454                         cl_lock_descr_print(env, cookie, p,
 455                                             &lov->lls_cl.cls_lock->cll_descr);
 456                 (*p)(env, cookie, "] ");
 457         }
 458         return 0;
 459 }
 460
 461 static const struct cl_lock_operations lovsub_lock_ops = {
 462         .clo_fini    = lovsub_lock_fini,
 463         .clo_state   = lovsub_lock_state,
 464         .clo_delete  = lovsub_lock_delete,
 465         .clo_modify  = lovsub_lock_modify,
 466         .clo_closure = lovsub_lock_closure,
 467         .clo_weigh   = lovsub_lock_weigh,
 468         .clo_print   = lovsub_lock_print
 469 };
 470
 471 int lovsub_lock_init(const struct lu_env *env, struct cl_object *obj,
 472                      struct cl_lock *lock, const struct cl_io *io)
 473 {
 474         struct lovsub_lock *lsk;
 475         int result;
 476
 477         ENTRY;
 478         OBD_SLAB_ALLOC_PTR_GFP(lsk, lovsub_lock_kmem, CFS_ALLOC_IO);
 479         if (lsk != NULL) {
 480                 CFS_INIT_LIST_HEAD(&lsk->lss_parents);
 481                 cl_lock_slice_add(lock, &lsk->lss_cl, obj, &lovsub_lock_ops);
 482                 result = 0;
 483         } else
 484                 result = -ENOMEM;
 485         RETURN(result);
 486 }
 487
 488 /** @} lov */