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