kernel/drivers/staging/lustre/lustre/obdclass/cl_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Client Lustre Page.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_CLASS
  42
  43 #include "../../include/linux/libcfs/libcfs.h"
  44 #include "../include/obd_class.h"
  45 #include "../include/obd_support.h"
  46 #include <linux/list.h>
  47
  48 #include "../include/cl_object.h"
  49 #include "cl_internal.h"
  50
  51 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
  52                             int radix);
  53
  54 # define PASSERT(env, page, expr)                                          \
  55         do {                                                               \
  56                 if (unlikely(!(expr))) {                                   \
  57                         CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n"); \
  58                         LASSERT(0);                                        \
  59                 }                                                          \
  60         } while (0)
  61
  62 # define PINVRNT(env, page, exp) \
  63         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  64
  65 /**
  66  * Internal version of cl_page_top, it should be called if the page is
  67  * known to be not freed, says with page referenced, or radix tree lock held,
  68  * or page owned.
  69  */
  70 static struct cl_page *cl_page_top_trusted(struct cl_page *page)
  71 {
  72         while (page->cp_parent != NULL)
  73                 page = page->cp_parent;
  74         return page;
  75 }
  76
  77 /**
  78  * Internal version of cl_page_get().
  79  *
  80  * This function can be used to obtain initial reference to previously
  81  * unreferenced cached object. It can be called only if concurrent page
  82  * reclamation is somehow prevented, e.g., by locking page radix-tree
  83  * (cl_object_header::hdr->coh_page_guard), or by keeping a lock on a VM page,
  84  * associated with \a page.
  85  *
  86  * Use with care! Not exported.
  87  */
  88 static void cl_page_get_trust(struct cl_page *page)
  89 {
  90         LASSERT(atomic_read(&page->cp_ref) > 0);
  91         atomic_inc(&page->cp_ref);
  92 }
  93
  94 /**
  95  * Returns a slice within a page, corresponding to the given layer in the
  96  * device stack.
  97  *
  98  * \see cl_lock_at()
  99  */
 100 static const struct cl_page_slice *
 101 cl_page_at_trusted(const struct cl_page *page,
 102                    const struct lu_device_type *dtype)
 103 {
 104         const struct cl_page_slice *slice;
 105
 106         page = cl_page_top_trusted((struct cl_page *)page);
 107         do {
 108                 list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 109                         if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
 110                                 return slice;
 111                 }
 112                 page = page->cp_child;
 113         } while (page != NULL);
 114         return NULL;
 115 }
 116
 117 /**
 118  * Returns a page with given index in the given object, or NULL if no page is
 119  * found. Acquires a reference on \a page.
 120  *
 121  * Locking: called under cl_object_header::coh_page_guard spin-lock.
 122  */
 123 struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index)
 124 {
 125         struct cl_page *page;
 126
 127         assert_spin_locked(&hdr->coh_page_guard);
 128
 129         page = radix_tree_lookup(&hdr->coh_tree, index);
 130         if (page != NULL)
 131                 cl_page_get_trust(page);
 132         return page;
 133 }
 134 EXPORT_SYMBOL(cl_page_lookup);
 135
 136 /**
 137  * Returns a list of pages by a given [start, end] of \a obj.
 138  *
 139  * \param resched If not NULL, then we give up before hogging CPU for too
 140  * long and set *resched = 1, in that case caller should implement a retry
 141  * logic.
 142  *
 143  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
 144  * crucial in the face of [offset, EOF] locks.
 145  *
 146  * Return at least one page in @queue unless there is no covered page.
 147  */
 148 int cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj,
 149                         struct cl_io *io, pgoff_t start, pgoff_t end,
 150                         cl_page_gang_cb_t cb, void *cbdata)
 151 {
 152         struct cl_object_header *hdr;
 153         struct cl_page    *page;
 154         struct cl_page   **pvec;
 155         const struct cl_page_slice  *slice;
 156         const struct lu_device_type *dtype;
 157         pgoff_t           idx;
 158         unsigned int         nr;
 159         unsigned int         i;
 160         unsigned int         j;
 161         int                   res = CLP_GANG_OKAY;
 162         int                   tree_lock = 1;
 163
 164         idx = start;
 165         hdr = cl_object_header(obj);
 166         pvec = cl_env_info(env)->clt_pvec;
 167         dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type;
 168         spin_lock(&hdr->coh_page_guard);
 169         while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec,
 170                                             idx, CLT_PVEC_SIZE)) > 0) {
 171                 int end_of_region = 0;
 172
 173                 idx = pvec[nr - 1]->cp_index + 1;
 174                 for (i = 0, j = 0; i < nr; ++i) {
 175                         page = pvec[i];
 176                         pvec[i] = NULL;
 177
 178                         LASSERT(page->cp_type == CPT_CACHEABLE);
 179                         if (page->cp_index > end) {
 180                                 end_of_region = 1;
 181                                 break;
 182                         }
 183                         if (page->cp_state == CPS_FREEING)
 184                                 continue;
 185
 186                         slice = cl_page_at_trusted(page, dtype);
 187                         /*
 188                          * Pages for lsm-less file has no underneath sub-page
 189                          * for osc, in case of ...
 190                          */
 191                         PASSERT(env, page, slice != NULL);
 192
 193                         page = slice->cpl_page;
 194                         /*
 195                          * Can safely call cl_page_get_trust() under
 196                          * radix-tree spin-lock.
 197                          *
 198                          * XXX not true, because @page is from object another
 199                          * than @hdr and protected by different tree lock.
 200                          */
 201                         cl_page_get_trust(page);
 202                         lu_ref_add_atomic(&page->cp_reference,
 203                                           "gang_lookup", current);
 204                         pvec[j++] = page;
 205                 }
 206
 207                 /*
 208                  * Here a delicate locking dance is performed. Current thread
 209                  * holds a reference to a page, but has to own it before it
 210                  * can be placed into queue. Owning implies waiting, so
 211                  * radix-tree lock is to be released. After a wait one has to
 212                  * check that pages weren't truncated (cl_page_own() returns
 213                  * error in the latter case).
 214                  */
 215                 spin_unlock(&hdr->coh_page_guard);
 216                 tree_lock = 0;
 217
 218                 for (i = 0; i < j; ++i) {
 219                         page = pvec[i];
 220                         if (res == CLP_GANG_OKAY)
 221                                 res = (*cb)(env, io, page, cbdata);
 222                         lu_ref_del(&page->cp_reference,
 223                                    "gang_lookup", current);
 224                         cl_page_put(env, page);
 225                 }
 226                 if (nr < CLT_PVEC_SIZE || end_of_region)
 227                         break;
 228
 229                 if (res == CLP_GANG_OKAY && need_resched())
 230                         res = CLP_GANG_RESCHED;
 231                 if (res != CLP_GANG_OKAY)
 232                         break;
 233
 234                 spin_lock(&hdr->coh_page_guard);
 235                 tree_lock = 1;
 236         }
 237         if (tree_lock)
 238                 spin_unlock(&hdr->coh_page_guard);
 239         return res;
 240 }
 241 EXPORT_SYMBOL(cl_page_gang_lookup);
 242
 243 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
 244 {
 245         struct cl_object *obj  = page->cp_obj;
 246
 247         PASSERT(env, page, list_empty(&page->cp_batch));
 248         PASSERT(env, page, page->cp_owner == NULL);
 249         PASSERT(env, page, page->cp_req == NULL);
 250         PASSERT(env, page, page->cp_parent == NULL);
 251         PASSERT(env, page, page->cp_state == CPS_FREEING);
 252
 253         might_sleep();
 254         while (!list_empty(&page->cp_layers)) {
 255                 struct cl_page_slice *slice;
 256
 257                 slice = list_entry(page->cp_layers.next,
 258                                        struct cl_page_slice, cpl_linkage);
 259                 list_del_init(page->cp_layers.next);
 260                 slice->cpl_ops->cpo_fini(env, slice);
 261         }
 262         lu_object_ref_del_at(&obj->co_lu, &page->cp_obj_ref, "cl_page", page);
 263         cl_object_put(env, obj);
 264         lu_ref_fini(&page->cp_reference);
 265         kfree(page);
 266 }
 267
 268 /**
 269  * Helper function updating page state. This is the only place in the code
 270  * where cl_page::cp_state field is mutated.
 271  */
 272 static inline void cl_page_state_set_trust(struct cl_page *page,
 273                                            enum cl_page_state state)
 274 {
 275         /* bypass const. */
 276         *(enum cl_page_state *)&page->cp_state = state;
 277 }
 278
 279 static struct cl_page *cl_page_alloc(const struct lu_env *env,
 280                 struct cl_object *o, pgoff_t ind, struct page *vmpage,
 281                 enum cl_page_type type)
 282 {
 283         struct cl_page    *page;
 284         struct lu_object_header *head;
 285
 286         page = kzalloc(cl_object_header(o)->coh_page_bufsize, GFP_NOFS);
 287         if (page != NULL) {
 288                 int result = 0;
 289
 290                 atomic_set(&page->cp_ref, 1);
 291                 if (type == CPT_CACHEABLE) /* for radix tree */
 292                         atomic_inc(&page->cp_ref);
 293                 page->cp_obj = o;
 294                 cl_object_get(o);
 295                 lu_object_ref_add_at(&o->co_lu, &page->cp_obj_ref, "cl_page",
 296                                      page);
 297                 page->cp_index = ind;
 298                 cl_page_state_set_trust(page, CPS_CACHED);
 299                 page->cp_type = type;
 300                 INIT_LIST_HEAD(&page->cp_layers);
 301                 INIT_LIST_HEAD(&page->cp_batch);
 302                 INIT_LIST_HEAD(&page->cp_flight);
 303                 mutex_init(&page->cp_mutex);
 304                 lu_ref_init(&page->cp_reference);
 305                 head = o->co_lu.lo_header;
 306                 list_for_each_entry(o, &head->loh_layers,
 307                                         co_lu.lo_linkage) {
 308                         if (o->co_ops->coo_page_init != NULL) {
 309                                 result = o->co_ops->coo_page_init(env, o,
 310                                                                   page, vmpage);
 311                                 if (result != 0) {
 312                                         cl_page_delete0(env, page, 0);
 313                                         cl_page_free(env, page);
 314                                         page = ERR_PTR(result);
 315                                         break;
 316                                 }
 317                         }
 318                 }
 319         } else {
 320                 page = ERR_PTR(-ENOMEM);
 321         }
 322         return page;
 323 }
 324
 325 /**
 326  * Returns a cl_page with index \a idx at the object \a o, and associated with
 327  * the VM page \a vmpage.
 328  *
 329  * This is the main entry point into the cl_page caching interface. First, a
 330  * cache (implemented as a per-object radix tree) is consulted. If page is
 331  * found there, it is returned immediately. Otherwise new page is allocated
 332  * and returned. In any case, additional reference to page is acquired.
 333  *
 334  * \see cl_object_find(), cl_lock_find()
 335  */
 336 static struct cl_page *cl_page_find0(const struct lu_env *env,
 337                                      struct cl_object *o,
 338                                      pgoff_t idx, struct page *vmpage,
 339                                      enum cl_page_type type,
 340                                      struct cl_page *parent)
 341 {
 342         struct cl_page    *page = NULL;
 343         struct cl_page    *ghost = NULL;
 344         struct cl_object_header *hdr;
 345         int err;
 346
 347         LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 348         might_sleep();
 349
 350         hdr = cl_object_header(o);
 351
 352         CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
 353                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 354         /* fast path. */
 355         if (type == CPT_CACHEABLE) {
 356                 /*
 357                  * vmpage lock is used to protect the child/parent
 358                  * relationship
 359                  */
 360                 KLASSERT(PageLocked(vmpage));
 361                 /*
 362                  * cl_vmpage_page() can be called here without any locks as
 363                  *
 364                  *     - "vmpage" is locked (which prevents ->private from
 365                  *       concurrent updates), and
 366                  *
 367                  *     - "o" cannot be destroyed while current thread holds a
 368                  *       reference on it.
 369                  */
 370                 page = cl_vmpage_page(vmpage, o);
 371                 PINVRNT(env, page,
 372                         ergo(page != NULL,
 373                              cl_page_vmpage(env, page) == vmpage &&
 374                              (void *)radix_tree_lookup(&hdr->coh_tree,
 375                                                        idx) == page));
 376         }
 377
 378         if (page != NULL)
 379                 return page;
 380
 381         /* allocate and initialize cl_page */
 382         page = cl_page_alloc(env, o, idx, vmpage, type);
 383         if (IS_ERR(page))
 384                 return page;
 385
 386         if (type == CPT_TRANSIENT) {
 387                 if (parent) {
 388                         LASSERT(page->cp_parent == NULL);
 389                         page->cp_parent = parent;
 390                         parent->cp_child = page;
 391                 }
 392                 return page;
 393         }
 394
 395         /*
 396          * XXX optimization: use radix_tree_preload() here, and change tree
 397          * gfp mask to GFP_KERNEL in cl_object_header_init().
 398          */
 399         spin_lock(&hdr->coh_page_guard);
 400         err = radix_tree_insert(&hdr->coh_tree, idx, page);
 401         if (err != 0) {
 402                 ghost = page;
 403                 /*
 404                  * Noted by Jay: a lock on \a vmpage protects cl_page_find()
 405                  * from this race, but
 406                  *
 407                  *     0. it's better to have cl_page interface "locally
 408                  *     consistent" so that its correctness can be reasoned
 409                  *     about without appealing to the (obscure world of) VM
 410                  *     locking.
 411                  *
 412                  *     1. handling this race allows ->coh_tree to remain
 413                  *     consistent even when VM locking is somehow busted,
 414                  *     which is very useful during diagnosing and debugging.
 415                  */
 416                 page = ERR_PTR(err);
 417                 CL_PAGE_DEBUG(D_ERROR, env, ghost,
 418                               "fail to insert into radix tree: %d\n", err);
 419         } else {
 420                 if (parent) {
 421                         LASSERT(page->cp_parent == NULL);
 422                         page->cp_parent = parent;
 423                         parent->cp_child = page;
 424                 }
 425                 hdr->coh_pages++;
 426         }
 427         spin_unlock(&hdr->coh_page_guard);
 428
 429         if (unlikely(ghost != NULL)) {
 430                 cl_page_delete0(env, ghost, 0);
 431                 cl_page_free(env, ghost);
 432         }
 433         return page;
 434 }
 435
 436 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o,
 437                              pgoff_t idx, struct page *vmpage,
 438                              enum cl_page_type type)
 439 {
 440         return cl_page_find0(env, o, idx, vmpage, type, NULL);
 441 }
 442 EXPORT_SYMBOL(cl_page_find);
 443
 444 struct cl_page *cl_page_find_sub(const struct lu_env *env, struct cl_object *o,
 445                                  pgoff_t idx, struct page *vmpage,
 446                                  struct cl_page *parent)
 447 {
 448         return cl_page_find0(env, o, idx, vmpage, parent->cp_type, parent);
 449 }
 450 EXPORT_SYMBOL(cl_page_find_sub);
 451
 452 static inline int cl_page_invariant(const struct cl_page *pg)
 453 {
 454         struct cl_object_header *header;
 455         struct cl_page    *parent;
 456         struct cl_page    *child;
 457         struct cl_io        *owner;
 458
 459         /*
 460          * Page invariant is protected by a VM lock.
 461          */
 462         LINVRNT(cl_page_is_vmlocked(NULL, pg));
 463
 464         header = cl_object_header(pg->cp_obj);
 465         parent = pg->cp_parent;
 466         child  = pg->cp_child;
 467         owner  = pg->cp_owner;
 468
 469         return cl_page_in_use(pg) &&
 470                 ergo(parent != NULL, parent->cp_child == pg) &&
 471                 ergo(child != NULL, child->cp_parent == pg) &&
 472                 ergo(child != NULL, pg->cp_obj != child->cp_obj) &&
 473                 ergo(parent != NULL, pg->cp_obj != parent->cp_obj) &&
 474                 ergo(owner != NULL && parent != NULL,
 475                      parent->cp_owner == pg->cp_owner->ci_parent) &&
 476                 ergo(owner != NULL && child != NULL,
 477                      child->cp_owner->ci_parent == owner) &&
 478                 /*
 479                  * Either page is early in initialization (has neither child
 480                  * nor parent yet), or it is in the object radix tree.
 481                  */
 482                 ergo(pg->cp_state < CPS_FREEING && pg->cp_type == CPT_CACHEABLE,
 483                      (void *)radix_tree_lookup(&header->coh_tree,
 484                                                pg->cp_index) == pg ||
 485                      (child == NULL && parent == NULL));
 486 }
 487
 488 static void cl_page_state_set0(const struct lu_env *env,
 489                                struct cl_page *page, enum cl_page_state state)
 490 {
 491         enum cl_page_state old;
 492
 493         /*
 494          * Matrix of allowed state transitions [old][new], for sanity
 495          * checking.
 496          */
 497         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 498                 [CPS_CACHED] = {
 499                         [CPS_CACHED]  = 0,
 500                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 501                         [CPS_PAGEIN]  = 0,
 502                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 503                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 504                 },
 505                 [CPS_OWNED] = {
 506                         [CPS_CACHED]  = 1, /* release to the cache */
 507                         [CPS_OWNED]   = 0,
 508                         [CPS_PAGEIN]  = 1, /* start read immediately */
 509                         [CPS_PAGEOUT] = 1, /* start write immediately */
 510                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 511                 },
 512                 [CPS_PAGEIN] = {
 513                         [CPS_CACHED]  = 1, /* io completion */
 514                         [CPS_OWNED]   = 0,
 515                         [CPS_PAGEIN]  = 0,
 516                         [CPS_PAGEOUT] = 0,
 517                         [CPS_FREEING] = 0,
 518                 },
 519                 [CPS_PAGEOUT] = {
 520                         [CPS_CACHED]  = 1, /* io completion */
 521                         [CPS_OWNED]   = 0,
 522                         [CPS_PAGEIN]  = 0,
 523                         [CPS_PAGEOUT] = 0,
 524                         [CPS_FREEING] = 0,
 525                 },
 526                 [CPS_FREEING] = {
 527                         [CPS_CACHED]  = 0,
 528                         [CPS_OWNED]   = 0,
 529                         [CPS_PAGEIN]  = 0,
 530                         [CPS_PAGEOUT] = 0,
 531                         [CPS_FREEING] = 0,
 532                 }
 533         };
 534
 535         old = page->cp_state;
 536         PASSERT(env, page, allowed_transitions[old][state]);
 537         CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
 538         for (; page != NULL; page = page->cp_child) {
 539                 PASSERT(env, page, page->cp_state == old);
 540                 PASSERT(env, page,
 541                         equi(state == CPS_OWNED, page->cp_owner != NULL));
 542
 543                 cl_page_state_set_trust(page, state);
 544         }
 545 }
 546
 547 static void cl_page_state_set(const struct lu_env *env,
 548                               struct cl_page *page, enum cl_page_state state)
 549 {
 550         cl_page_state_set0(env, page, state);
 551 }
 552
 553 /**
 554  * Acquires an additional reference to a page.
 555  *
 556  * This can be called only by caller already possessing a reference to \a
 557  * page.
 558  *
 559  * \see cl_object_get(), cl_lock_get().
 560  */
 561 void cl_page_get(struct cl_page *page)
 562 {
 563         cl_page_get_trust(page);
 564 }
 565 EXPORT_SYMBOL(cl_page_get);
 566
 567 /**
 568  * Releases a reference to a page.
 569  *
 570  * When last reference is released, page is returned to the cache, unless it
 571  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 572  * destroyed.
 573  *
 574  * \see cl_object_put(), cl_lock_put().
 575  */
 576 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 577 {
 578         PASSERT(env, page, atomic_read(&page->cp_ref) > !!page->cp_parent);
 579
 580         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 581                        atomic_read(&page->cp_ref));
 582
 583         if (atomic_dec_and_test(&page->cp_ref)) {
 584                 LASSERT(page->cp_state == CPS_FREEING);
 585
 586                 LASSERT(atomic_read(&page->cp_ref) == 0);
 587                 PASSERT(env, page, page->cp_owner == NULL);
 588                 PASSERT(env, page, list_empty(&page->cp_batch));
 589                 /*
 590                  * Page is no longer reachable by other threads. Tear
 591                  * it down.
 592                  */
 593                 cl_page_free(env, page);
 594         }
 595 }
 596 EXPORT_SYMBOL(cl_page_put);
 597
 598 /**
 599  * Returns a VM page associated with a given cl_page.
 600  */
 601 struct page *cl_page_vmpage(const struct lu_env *env, struct cl_page *page)
 602 {
 603         const struct cl_page_slice *slice;
 604
 605         /*
 606          * Find uppermost layer with ->cpo_vmpage() method, and return its
 607          * result.
 608          */
 609         page = cl_page_top(page);
 610         do {
 611                 list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 612                         if (slice->cpl_ops->cpo_vmpage != NULL)
 613                                 return slice->cpl_ops->cpo_vmpage(env, slice);
 614                 }
 615                 page = page->cp_child;
 616         } while (page != NULL);
 617         LBUG(); /* ->cpo_vmpage() has to be defined somewhere in the stack */
 618 }
 619 EXPORT_SYMBOL(cl_page_vmpage);
 620
 621 /**
 622  * Returns a cl_page associated with a VM page, and given cl_object.
 623  */
 624 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj)
 625 {
 626         struct cl_page *top;
 627         struct cl_page *page;
 628
 629         KLASSERT(PageLocked(vmpage));
 630
 631         /*
 632          * NOTE: absence of races and liveness of data are guaranteed by page
 633          *       lock on a "vmpage". That works because object destruction has
 634          *       bottom-to-top pass.
 635          */
 636
 637         /*
 638          * This loop assumes that ->private points to the top-most page. This
 639          * can be rectified easily.
 640          */
 641         top = (struct cl_page *)vmpage->private;
 642         if (top == NULL)
 643                 return NULL;
 644
 645         for (page = top; page != NULL; page = page->cp_child) {
 646                 if (cl_object_same(page->cp_obj, obj)) {
 647                         cl_page_get_trust(page);
 648                         break;
 649                 }
 650         }
 651         LASSERT(ergo(page, page->cp_type == CPT_CACHEABLE));
 652         return page;
 653 }
 654 EXPORT_SYMBOL(cl_vmpage_page);
 655
 656 /**
 657  * Returns the top-page for a given page.
 658  *
 659  * \see cl_object_top(), cl_io_top()
 660  */
 661 struct cl_page *cl_page_top(struct cl_page *page)
 662 {
 663         return cl_page_top_trusted(page);
 664 }
 665 EXPORT_SYMBOL(cl_page_top);
 666
 667 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 668                                        const struct lu_device_type *dtype)
 669 {
 670         return cl_page_at_trusted(page, dtype);
 671 }
 672 EXPORT_SYMBOL(cl_page_at);
 673
 674 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 675
 676 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)              \
 677 ({                                                                    \
 678         const struct lu_env     *__env  = (_env);                   \
 679         struct cl_page       *__page = (_page);            \
 680         const struct cl_page_slice *__scan;                          \
 681         int                      __result;                         \
 682         ptrdiff_t                  __op   = (_op);                   \
 683         int                    (*__method)_proto;                   \
 684                                                                         \
 685         __result = 0;                                              \
 686         __page = cl_page_top(__page);                              \
 687         do {                                                        \
 688                 list_for_each_entry(__scan, &__page->cp_layers,     \
 689                                         cpl_linkage) {            \
 690                         __method = *(void **)((char *)__scan->cpl_ops + \
 691                                               __op);                \
 692                         if (__method != NULL) {                  \
 693                                 __result = (*__method)(__env, __scan,   \
 694                                                        ## __VA_ARGS__); \
 695                                 if (__result != 0)                    \
 696                                         break;                    \
 697                         }                                              \
 698                 }                                                      \
 699                 __page = __page->cp_child;                            \
 700         } while (__page != NULL && __result == 0);                    \
 701         if (__result > 0)                                              \
 702                 __result = 0;                                      \
 703         __result;                                                      \
 704 })
 705
 706 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)              \
 707 do {                                                                \
 708         const struct lu_env     *__env  = (_env);                   \
 709         struct cl_page       *__page = (_page);            \
 710         const struct cl_page_slice *__scan;                          \
 711         ptrdiff_t                  __op   = (_op);                   \
 712         void                  (*__method)_proto;                    \
 713                                                                         \
 714         __page = cl_page_top(__page);                              \
 715         do {                                                        \
 716                 list_for_each_entry(__scan, &__page->cp_layers,     \
 717                                         cpl_linkage) {            \
 718                         __method = *(void **)((char *)__scan->cpl_ops + \
 719                                               __op);                \
 720                         if (__method != NULL)                      \
 721                                 (*__method)(__env, __scan,            \
 722                                             ## __VA_ARGS__);        \
 723                 }                                                      \
 724                 __page = __page->cp_child;                            \
 725         } while (__page != NULL);                                      \
 726 } while (0)
 727
 728 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)          \
 729 do {                                                                    \
 730         const struct lu_env     *__env  = (_env);                       \
 731         struct cl_page       *__page = (_page);                \
 732         const struct cl_page_slice *__scan;                              \
 733         ptrdiff_t                  __op   = (_op);                       \
 734         void                  (*__method)_proto;                        \
 735                                                                             \
 736         /* get to the bottom page. */                                  \
 737         while (__page->cp_child != NULL)                                    \
 738                 __page = __page->cp_child;                                \
 739         do {                                                            \
 740                 list_for_each_entry_reverse(__scan, &__page->cp_layers, \
 741                                                 cpl_linkage) {        \
 742                         __method = *(void **)((char *)__scan->cpl_ops +     \
 743                                               __op);                    \
 744                         if (__method != NULL)                          \
 745                                 (*__method)(__env, __scan,                \
 746                                             ## __VA_ARGS__);            \
 747                 }                                                          \
 748                 __page = __page->cp_parent;                              \
 749         } while (__page != NULL);                                          \
 750 } while (0)
 751
 752 static int cl_page_invoke(const struct lu_env *env,
 753                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 754
 755 {
 756         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 757         return CL_PAGE_INVOKE(env, page, op,
 758                               (const struct lu_env *,
 759                                const struct cl_page_slice *, struct cl_io *),
 760                               io);
 761 }
 762
 763 static void cl_page_invoid(const struct lu_env *env,
 764                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 765
 766 {
 767         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 768         CL_PAGE_INVOID(env, page, op,
 769                        (const struct lu_env *,
 770                         const struct cl_page_slice *, struct cl_io *), io);
 771 }
 772
 773 static void cl_page_owner_clear(struct cl_page *page)
 774 {
 775         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 776                 if (page->cp_owner != NULL) {
 777                         LASSERT(page->cp_owner->ci_owned_nr > 0);
 778                         page->cp_owner->ci_owned_nr--;
 779                         page->cp_owner = NULL;
 780                         page->cp_task = NULL;
 781                 }
 782         }
 783 }
 784
 785 static void cl_page_owner_set(struct cl_page *page)
 786 {
 787         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 788                 LASSERT(page->cp_owner != NULL);
 789                 page->cp_owner->ci_owned_nr++;
 790         }
 791 }
 792
 793 void cl_page_disown0(const struct lu_env *env,
 794                      struct cl_io *io, struct cl_page *pg)
 795 {
 796         enum cl_page_state state;
 797
 798         state = pg->cp_state;
 799         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 800         PINVRNT(env, pg, cl_page_invariant(pg));
 801         cl_page_owner_clear(pg);
 802
 803         if (state == CPS_OWNED)
 804                 cl_page_state_set(env, pg, CPS_CACHED);
 805         /*
 806          * Completion call-backs are executed in the bottom-up order, so that
 807          * uppermost layer (llite), responsible for VFS/VM interaction runs
 808          * last and can release locks safely.
 809          */
 810         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 811                                (const struct lu_env *,
 812                                 const struct cl_page_slice *, struct cl_io *),
 813                                io);
 814 }
 815
 816 /**
 817  * returns true, iff page is owned by the given io.
 818  */
 819 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 820 {
 821         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 822         return pg->cp_state == CPS_OWNED && pg->cp_owner == io;
 823 }
 824 EXPORT_SYMBOL(cl_page_is_owned);
 825
 826 /**
 827  * Try to own a page by IO.
 828  *
 829  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 830  * into cl_page_state::CPS_OWNED state.
 831  *
 832  * \pre  !cl_page_is_owned(pg, io)
 833  * \post result == 0 iff cl_page_is_owned(pg, io)
 834  *
 835  * \retval 0   success
 836  *
 837  * \retval -ve failure, e.g., page was destroyed (and landed in
 838  *           cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 839  *           or, page was owned by another thread, or in IO.
 840  *
 841  * \see cl_page_disown()
 842  * \see cl_page_operations::cpo_own()
 843  * \see cl_page_own_try()
 844  * \see cl_page_own
 845  */
 846 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 847                         struct cl_page *pg, int nonblock)
 848 {
 849         int result;
 850
 851         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 852
 853         pg = cl_page_top(pg);
 854         io = cl_io_top(io);
 855
 856         if (pg->cp_state == CPS_FREEING) {
 857                 result = -ENOENT;
 858         } else {
 859                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 860                                         (const struct lu_env *,
 861                                          const struct cl_page_slice *,
 862                                          struct cl_io *, int),
 863                                         io, nonblock);
 864                 if (result == 0) {
 865                         PASSERT(env, pg, pg->cp_owner == NULL);
 866                         PASSERT(env, pg, pg->cp_req == NULL);
 867                         pg->cp_owner = io;
 868                         pg->cp_task  = current;
 869                         cl_page_owner_set(pg);
 870                         if (pg->cp_state != CPS_FREEING) {
 871                                 cl_page_state_set(env, pg, CPS_OWNED);
 872                         } else {
 873                                 cl_page_disown0(env, io, pg);
 874                                 result = -ENOENT;
 875                         }
 876                 }
 877         }
 878         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 879         return result;
 880 }
 881
 882 /**
 883  * Own a page, might be blocked.
 884  *
 885  * \see cl_page_own0()
 886  */
 887 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 888 {
 889         return cl_page_own0(env, io, pg, 0);
 890 }
 891 EXPORT_SYMBOL(cl_page_own);
 892
 893 /**
 894  * Nonblock version of cl_page_own().
 895  *
 896  * \see cl_page_own0()
 897  */
 898 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
 899                     struct cl_page *pg)
 900 {
 901         return cl_page_own0(env, io, pg, 1);
 902 }
 903 EXPORT_SYMBOL(cl_page_own_try);
 904
 905 /**
 906  * Assume page ownership.
 907  *
 908  * Called when page is already locked by the hosting VM.
 909  *
 910  * \pre !cl_page_is_owned(pg, io)
 911  * \post cl_page_is_owned(pg, io)
 912  *
 913  * \see cl_page_operations::cpo_assume()
 914  */
 915 void cl_page_assume(const struct lu_env *env,
 916                     struct cl_io *io, struct cl_page *pg)
 917 {
 918         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
 919
 920         pg = cl_page_top(pg);
 921         io = cl_io_top(io);
 922
 923         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
 924         PASSERT(env, pg, pg->cp_owner == NULL);
 925         pg->cp_owner = io;
 926         pg->cp_task = current;
 927         cl_page_owner_set(pg);
 928         cl_page_state_set(env, pg, CPS_OWNED);
 929 }
 930 EXPORT_SYMBOL(cl_page_assume);
 931
 932 /**
 933  * Releases page ownership without unlocking the page.
 934  *
 935  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
 936  * underlying VM page (as VM is supposed to do this itself).
 937  *
 938  * \pre   cl_page_is_owned(pg, io)
 939  * \post !cl_page_is_owned(pg, io)
 940  *
 941  * \see cl_page_assume()
 942  */
 943 void cl_page_unassume(const struct lu_env *env,
 944                       struct cl_io *io, struct cl_page *pg)
 945 {
 946         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 947         PINVRNT(env, pg, cl_page_invariant(pg));
 948
 949         pg = cl_page_top(pg);
 950         io = cl_io_top(io);
 951         cl_page_owner_clear(pg);
 952         cl_page_state_set(env, pg, CPS_CACHED);
 953         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
 954                                (const struct lu_env *,
 955                                 const struct cl_page_slice *, struct cl_io *),
 956                                io);
 957 }
 958 EXPORT_SYMBOL(cl_page_unassume);
 959
 960 /**
 961  * Releases page ownership.
 962  *
 963  * Moves page into cl_page_state::CPS_CACHED.
 964  *
 965  * \pre   cl_page_is_owned(pg, io)
 966  * \post !cl_page_is_owned(pg, io)
 967  *
 968  * \see cl_page_own()
 969  * \see cl_page_operations::cpo_disown()
 970  */
 971 void cl_page_disown(const struct lu_env *env,
 972                     struct cl_io *io, struct cl_page *pg)
 973 {
 974         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 975
 976         pg = cl_page_top(pg);
 977         io = cl_io_top(io);
 978         cl_page_disown0(env, io, pg);
 979 }
 980 EXPORT_SYMBOL(cl_page_disown);
 981
 982 /**
 983  * Called when page is to be removed from the object, e.g., as a result of
 984  * truncate.
 985  *
 986  * Calls cl_page_operations::cpo_discard() top-to-bottom.
 987  *
 988  * \pre cl_page_is_owned(pg, io)
 989  *
 990  * \see cl_page_operations::cpo_discard()
 991  */
 992 void cl_page_discard(const struct lu_env *env,
 993                      struct cl_io *io, struct cl_page *pg)
 994 {
 995         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 996         PINVRNT(env, pg, cl_page_invariant(pg));
 997
 998         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
 999 }
1000 EXPORT_SYMBOL(cl_page_discard);
1001
1002 /**
1003  * Version of cl_page_delete() that can be called for not fully constructed
1004  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
1005  * path. Doesn't check page invariant.
1006  */
1007 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
1008                             int radix)
1009 {
1010         struct cl_page *tmp = pg;
1011
1012         PASSERT(env, pg, pg == cl_page_top(pg));
1013         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
1014
1015         /*
1016          * Severe all ways to obtain new pointers to @pg.
1017          */
1018         cl_page_owner_clear(pg);
1019
1020         /*
1021          * unexport the page firstly before freeing it so that
1022          * the page content is considered to be invalid.
1023          * We have to do this because a CPS_FREEING cl_page may
1024          * be NOT under the protection of a cl_lock.
1025          * Afterwards, if this page is found by other threads, then this
1026          * page will be forced to reread.
1027          */
1028         cl_page_export(env, pg, 0);
1029         cl_page_state_set0(env, pg, CPS_FREEING);
1030
1031         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete),
1032                        (const struct lu_env *, const struct cl_page_slice *));
1033
1034         if (tmp->cp_type == CPT_CACHEABLE) {
1035                 if (!radix)
1036                         /* !radix means that @pg is not yet in the radix tree,
1037                          * skip removing it.
1038                          */
1039                         tmp = pg->cp_child;
1040                 for (; tmp != NULL; tmp = tmp->cp_child) {
1041                         void                *value;
1042                         struct cl_object_header *hdr;
1043
1044                         hdr = cl_object_header(tmp->cp_obj);
1045                         spin_lock(&hdr->coh_page_guard);
1046                         value = radix_tree_delete(&hdr->coh_tree,
1047                                                   tmp->cp_index);
1048                         PASSERT(env, tmp, value == tmp);
1049                         PASSERT(env, tmp, hdr->coh_pages > 0);
1050                         hdr->coh_pages--;
1051                         spin_unlock(&hdr->coh_page_guard);
1052                         cl_page_put(env, tmp);
1053                 }
1054         }
1055 }
1056
1057 /**
1058  * Called when a decision is made to throw page out of memory.
1059  *
1060  * Notifies all layers about page destruction by calling
1061  * cl_page_operations::cpo_delete() method top-to-bottom.
1062  *
1063  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
1064  * where transition to this state happens).
1065  *
1066  * Eliminates all venues through which new references to the page can be
1067  * obtained:
1068  *
1069  *     - removes page from the radix trees,
1070  *
1071  *     - breaks linkage from VM page to cl_page.
1072  *
1073  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
1074  * drain after some time, at which point page will be recycled.
1075  *
1076  * \pre  pg == cl_page_top(pg)
1077  * \pre  VM page is locked
1078  * \post pg->cp_state == CPS_FREEING
1079  *
1080  * \see cl_page_operations::cpo_delete()
1081  */
1082 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
1083 {
1084         PINVRNT(env, pg, cl_page_invariant(pg));
1085         cl_page_delete0(env, pg, 1);
1086 }
1087 EXPORT_SYMBOL(cl_page_delete);
1088
1089 /**
1090  * Unmaps page from user virtual memory.
1091  *
1092  * Calls cl_page_operations::cpo_unmap() through all layers top-to-bottom. The
1093  * layer responsible for VM interaction has to unmap page from user space
1094  * virtual memory.
1095  *
1096  * \see cl_page_operations::cpo_unmap()
1097  */
1098 int cl_page_unmap(const struct lu_env *env,
1099                   struct cl_io *io, struct cl_page *pg)
1100 {
1101         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1102         PINVRNT(env, pg, cl_page_invariant(pg));
1103
1104         return cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_unmap));
1105 }
1106 EXPORT_SYMBOL(cl_page_unmap);
1107
1108 /**
1109  * Marks page up-to-date.
1110  *
1111  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
1112  * layer responsible for VM interaction has to mark/clear page as up-to-date
1113  * by the \a uptodate argument.
1114  *
1115  * \see cl_page_operations::cpo_export()
1116  */
1117 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
1118 {
1119         PINVRNT(env, pg, cl_page_invariant(pg));
1120         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
1121                        (const struct lu_env *,
1122                         const struct cl_page_slice *, int), uptodate);
1123 }
1124 EXPORT_SYMBOL(cl_page_export);
1125
1126 /**
1127  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
1128  * thread.
1129  */
1130 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
1131 {
1132         int result;
1133         const struct cl_page_slice *slice;
1134
1135         pg = cl_page_top_trusted((struct cl_page *)pg);
1136         slice = container_of(pg->cp_layers.next,
1137                              const struct cl_page_slice, cpl_linkage);
1138         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
1139         /*
1140          * Call ->cpo_is_vmlocked() directly instead of going through
1141          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
1142          * cl_page_invariant().
1143          */
1144         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
1145         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
1146         return result == -EBUSY;
1147 }
1148 EXPORT_SYMBOL(cl_page_is_vmlocked);
1149
1150 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
1151 {
1152         return crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN;
1153 }
1154
1155 static void cl_page_io_start(const struct lu_env *env,
1156                              struct cl_page *pg, enum cl_req_type crt)
1157 {
1158         /*
1159          * Page is queued for IO, change its state.
1160          */
1161         cl_page_owner_clear(pg);
1162         cl_page_state_set(env, pg, cl_req_type_state(crt));
1163 }
1164
1165 /**
1166  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
1167  * called top-to-bottom. Every layer either agrees to submit this page (by
1168  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
1169  * handling interactions with the VM also has to inform VM that page is under
1170  * transfer now.
1171  */
1172 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
1173                  struct cl_page *pg, enum cl_req_type crt)
1174 {
1175         int result;
1176
1177         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1178         PINVRNT(env, pg, cl_page_invariant(pg));
1179         PINVRNT(env, pg, crt < CRT_NR);
1180
1181         /*
1182          * XXX this has to be called bottom-to-top, so that llite can set up
1183          * PG_writeback without risking other layers deciding to skip this
1184          * page.
1185          */
1186         if (crt >= CRT_NR)
1187                 return -EINVAL;
1188         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
1189         if (result == 0)
1190                 cl_page_io_start(env, pg, crt);
1191
1192         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1193         return result;
1194 }
1195 EXPORT_SYMBOL(cl_page_prep);
1196
1197 /**
1198  * Notify layers about transfer completion.
1199  *
1200  * Invoked by transfer sub-system (which is a part of osc) to notify layers
1201  * that a transfer, of which this page is a part of has completed.
1202  *
1203  * Completion call-backs are executed in the bottom-up order, so that
1204  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1205  * and can release locks safely.
1206  *
1207  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1208  * \post pg->cp_state == CPS_CACHED
1209  *
1210  * \see cl_page_operations::cpo_completion()
1211  */
1212 void cl_page_completion(const struct lu_env *env,
1213                         struct cl_page *pg, enum cl_req_type crt, int ioret)
1214 {
1215         struct cl_sync_io *anchor = pg->cp_sync_io;
1216
1217         PASSERT(env, pg, crt < CRT_NR);
1218         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
1219         PASSERT(env, pg, pg->cp_req == NULL);
1220         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
1221
1222         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
1223         if (crt == CRT_READ && ioret == 0) {
1224                 PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED));
1225                 pg->cp_flags |= CPF_READ_COMPLETED;
1226         }
1227
1228         cl_page_state_set(env, pg, CPS_CACHED);
1229         if (crt >= CRT_NR)
1230                 return;
1231         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
1232                                (const struct lu_env *,
1233                                 const struct cl_page_slice *, int), ioret);
1234         if (anchor) {
1235                 LASSERT(cl_page_is_vmlocked(env, pg));
1236                 LASSERT(pg->cp_sync_io == anchor);
1237                 pg->cp_sync_io = NULL;
1238         }
1239         /*
1240          * As page->cp_obj is pinned by a reference from page->cp_req, it is
1241          * safe to call cl_page_put() without risking object destruction in a
1242          * non-blocking context.
1243          */
1244         cl_page_put(env, pg);
1245
1246         if (anchor)
1247                 cl_sync_io_note(anchor, ioret);
1248 }
1249 EXPORT_SYMBOL(cl_page_completion);
1250
1251 /**
1252  * Notify layers that transfer formation engine decided to yank this page from
1253  * the cache and to make it a part of a transfer.
1254  *
1255  * \pre  pg->cp_state == CPS_CACHED
1256  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1257  *
1258  * \see cl_page_operations::cpo_make_ready()
1259  */
1260 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
1261                        enum cl_req_type crt)
1262 {
1263         int result;
1264
1265         PINVRNT(env, pg, crt < CRT_NR);
1266
1267         if (crt >= CRT_NR)
1268                 return -EINVAL;
1269         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
1270                                 (const struct lu_env *,
1271                                  const struct cl_page_slice *));
1272         if (result == 0) {
1273                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
1274                 cl_page_io_start(env, pg, crt);
1275         }
1276         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1277         return result;
1278 }
1279 EXPORT_SYMBOL(cl_page_make_ready);
1280
1281 /**
1282  * Notify layers that high level io decided to place this page into a cache
1283  * for future transfer.
1284  *
1285  * The layer implementing transfer engine (osc) has to register this page in
1286  * its queues.
1287  *
1288  * \pre  cl_page_is_owned(pg, io)
1289  * \post cl_page_is_owned(pg, io)
1290  *
1291  * \see cl_page_operations::cpo_cache_add()
1292  */
1293 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
1294                       struct cl_page *pg, enum cl_req_type crt)
1295 {
1296         const struct cl_page_slice *scan;
1297         int result = 0;
1298
1299         PINVRNT(env, pg, crt < CRT_NR);
1300         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1301         PINVRNT(env, pg, cl_page_invariant(pg));
1302
1303         if (crt >= CRT_NR)
1304                 return -EINVAL;
1305
1306         list_for_each_entry(scan, &pg->cp_layers, cpl_linkage) {
1307                 if (scan->cpl_ops->io[crt].cpo_cache_add == NULL)
1308                         continue;
1309
1310                 result = scan->cpl_ops->io[crt].cpo_cache_add(env, scan, io);
1311                 if (result != 0)
1312                         break;
1313         }
1314         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1315         return result;
1316 }
1317 EXPORT_SYMBOL(cl_page_cache_add);
1318
1319 /**
1320  * Called if a pge is being written back by kernel's intention.
1321  *
1322  * \pre  cl_page_is_owned(pg, io)
1323  * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT)
1324  *
1325  * \see cl_page_operations::cpo_flush()
1326  */
1327 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
1328                   struct cl_page *pg)
1329 {
1330         int result;
1331
1332         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1333         PINVRNT(env, pg, cl_page_invariant(pg));
1334
1335         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_flush));
1336
1337         CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result);
1338         return result;
1339 }
1340 EXPORT_SYMBOL(cl_page_flush);
1341
1342 /**
1343  * Checks whether page is protected by any extent lock is at least required
1344  * mode.
1345  *
1346  * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1347  * \see cl_page_operations::cpo_is_under_lock()
1348  */
1349 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1350                           struct cl_page *page)
1351 {
1352         int rc;
1353
1354         PINVRNT(env, page, cl_page_invariant(page));
1355
1356         rc = CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1357                             (const struct lu_env *,
1358                              const struct cl_page_slice *, struct cl_io *),
1359                             io);
1360         PASSERT(env, page, rc != 0);
1361         return rc;
1362 }
1363 EXPORT_SYMBOL(cl_page_is_under_lock);
1364
1365 static int page_prune_cb(const struct lu_env *env, struct cl_io *io,
1366                          struct cl_page *page, void *cbdata)
1367 {
1368         cl_page_own(env, io, page);
1369         cl_page_unmap(env, io, page);
1370         cl_page_discard(env, io, page);
1371         cl_page_disown(env, io, page);
1372         return CLP_GANG_OKAY;
1373 }
1374
1375 /**
1376  * Purges all cached pages belonging to the object \a obj.
1377  */
1378 int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj)
1379 {
1380         struct cl_thread_info   *info;
1381         struct cl_object        *obj = cl_object_top(clobj);
1382         struct cl_io        *io;
1383         int                   result;
1384
1385         info  = cl_env_info(env);
1386         io    = &info->clt_io;
1387
1388         /*
1389          * initialize the io. This is ugly since we never do IO in this
1390          * function, we just make cl_page_list functions happy. -jay
1391          */
1392         io->ci_obj = obj;
1393         io->ci_ignore_layout = 1;
1394         result = cl_io_init(env, io, CIT_MISC, obj);
1395         if (result != 0) {
1396                 cl_io_fini(env, io);
1397                 return io->ci_result;
1398         }
1399
1400         do {
1401                 result = cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF,
1402                                              page_prune_cb, NULL);
1403                 if (result == CLP_GANG_RESCHED)
1404                         cond_resched();
1405         } while (result != CLP_GANG_OKAY);
1406
1407         cl_io_fini(env, io);
1408         return result;
1409 }
1410 EXPORT_SYMBOL(cl_pages_prune);
1411
1412 /**
1413  * Tells transfer engine that only part of a page is to be transmitted.
1414  *
1415  * \see cl_page_operations::cpo_clip()
1416  */
1417 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1418                   int from, int to)
1419 {
1420         PINVRNT(env, pg, cl_page_invariant(pg));
1421
1422         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1423         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1424                        (const struct lu_env *,
1425                         const struct cl_page_slice *, int, int),
1426                        from, to);
1427 }
1428 EXPORT_SYMBOL(cl_page_clip);
1429
1430 /**
1431  * Prints human readable representation of \a pg to the \a f.
1432  */
1433 void cl_page_header_print(const struct lu_env *env, void *cookie,
1434                           lu_printer_t printer, const struct cl_page *pg)
1435 {
1436         (*printer)(env, cookie,
1437                    "page@%p[%d %p:%lu ^%p_%p %d %d %d %p %p %#x]\n",
1438                    pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1439                    pg->cp_index, pg->cp_parent, pg->cp_child,
1440                    pg->cp_state, pg->cp_error, pg->cp_type,
1441                    pg->cp_owner, pg->cp_req, pg->cp_flags);
1442 }
1443 EXPORT_SYMBOL(cl_page_header_print);
1444
1445 /**
1446  * Prints human readable representation of \a pg to the \a f.
1447  */
1448 void cl_page_print(const struct lu_env *env, void *cookie,
1449                    lu_printer_t printer, const struct cl_page *pg)
1450 {
1451         struct cl_page *scan;
1452
1453         for (scan = cl_page_top((struct cl_page *)pg);
1454              scan != NULL; scan = scan->cp_child)
1455                 cl_page_header_print(env, cookie, printer, scan);
1456         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1457                        (const struct lu_env *env,
1458                         const struct cl_page_slice *slice,
1459                         void *cookie, lu_printer_t p), cookie, printer);
1460         (*printer)(env, cookie, "end page@%p\n", pg);
1461 }
1462 EXPORT_SYMBOL(cl_page_print);
1463
1464 /**
1465  * Cancel a page which is still in a transfer.
1466  */
1467 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1468 {
1469         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1470                               (const struct lu_env *,
1471                                const struct cl_page_slice *));
1472 }
1473 EXPORT_SYMBOL(cl_page_cancel);
1474
1475 /**
1476  * Converts a byte offset within object \a obj into a page index.
1477  */
1478 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1479 {
1480         /*
1481          * XXX for now.
1482          */
1483         return (loff_t)idx << PAGE_CACHE_SHIFT;
1484 }
1485 EXPORT_SYMBOL(cl_offset);
1486
1487 /**
1488  * Converts a page index into a byte offset within object \a obj.
1489  */
1490 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1491 {
1492         /*
1493          * XXX for now.
1494          */
1495         return offset >> PAGE_CACHE_SHIFT;
1496 }
1497 EXPORT_SYMBOL(cl_index);
1498
1499 int cl_page_size(const struct cl_object *obj)
1500 {
1501         return 1 << PAGE_CACHE_SHIFT;
1502 }
1503 EXPORT_SYMBOL(cl_page_size);
1504
1505 /**
1506  * Adds page slice to the compound page.
1507  *
1508  * This is called by cl_object_operations::coo_page_init() methods to add a
1509  * per-layer state to the page. New state is added at the end of
1510  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1511  *
1512  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1513  */
1514 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1515                        struct cl_object *obj,
1516                        const struct cl_page_operations *ops)
1517 {
1518         list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1519         slice->cpl_obj  = obj;
1520         slice->cpl_ops  = ops;
1521         slice->cpl_page = page;
1522 }
1523 EXPORT_SYMBOL(cl_page_slice_add);
1524
1525 int  cl_page_init(void)
1526 {
1527         return 0;
1528 }
1529
1530 void cl_page_fini(void)
1531 {
1532 }