lustre/obdclass/cl_page.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  * Client Lustre Page.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_CLASS
  42 #ifndef EXPORT_SYMTAB
  43 # define EXPORT_SYMTAB
  44 #endif
  45
  46 #include <libcfs/libcfs.h>
  47 #include <obd_class.h>
  48 #include <obd_support.h>
  49 #include <libcfs/list.h>
  50
  51 #include <cl_object.h>
  52 #include "cl_internal.h"
  53
  54 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
  55                             int radix);
  56
  57 static cfs_mem_cache_t      *cl_page_kmem = NULL;
  58
  59 static struct lu_kmem_descr cl_page_caches[] = {
  60         {
  61                 .ckd_cache = &cl_page_kmem,
  62                 .ckd_name  = "cl_page_kmem",
  63                 .ckd_size  = sizeof (struct cl_page)
  64         },
  65         {
  66                 .ckd_cache = NULL
  67         }
  68 };
  69
  70 #ifdef LIBCFS_DEBUG
  71 # define PASSERT(env, page, expr)                                       \
  72   do {                                                                    \
  73           if (unlikely(!(expr))) {                                      \
  74                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  75                   LASSERT(0);                                           \
  76           }                                                             \
  77   } while (0)
  78 #else /* !LIBCFS_DEBUG */
  79 # define PASSERT(env, page, exp) \
  80         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  81 #endif /* !LIBCFS_DEBUG */
  82
  83 #ifdef INVARIANT_CHECK
  84 # define PINVRNT(env, page, expr)                                       \
  85   do {                                                                    \
  86           if (unlikely(!(expr))) {                                      \
  87                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  88                   LINVRNT(0);                                           \
  89           }                                                             \
  90   } while (0)
  91 #else /* !INVARIANT_CHECK */
  92 # define PINVRNT(env, page, exp) \
  93         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  94 #endif /* !INVARIANT_CHECK */
  95
  96 /**
  97  * Internal version of cl_page_top, it should be called with page referenced,
  98  * or coh_page_guard held.
  99  */
 100 static struct cl_page *cl_page_top_trusted(struct cl_page *page)
 101 {
 102         while (page->cp_parent != NULL)
 103                 page = page->cp_parent;
 104         return page;
 105 }
 106
 107 /**
 108  * Internal version of cl_page_get().
 109  *
 110  * This function can be used to obtain initial reference to previously
 111  * unreferenced cached object. It can be called only if concurrent page
 112  * reclamation is somehow prevented, e.g., by locking page radix-tree
 113  * (cl_object_header::hdr->coh_page_guard), or by keeping a lock on a VM page,
 114  * associated with \a page.
 115  *
 116  * Use with care! Not exported.
 117  */
 118 static void cl_page_get_trust(struct cl_page *page)
 119 {
 120         /*
 121          * Checkless version for trusted users.
 122          */
 123         if (cfs_atomic_inc_return(&page->cp_ref) == 1)
 124                 cfs_atomic_inc(&cl_object_site(page->cp_obj)->cs_pages.cs_busy);
 125 }
 126
 127 /**
 128  * Returns a slice within a page, corresponding to the given layer in the
 129  * device stack.
 130  *
 131  * \see cl_lock_at()
 132  */
 133 static const struct cl_page_slice *
 134 cl_page_at_trusted(const struct cl_page *page,
 135                    const struct lu_device_type *dtype)
 136 {
 137         const struct cl_page_slice *slice;
 138
 139 #ifdef INVARIANT_CHECK
 140         struct cl_object_header *ch = cl_object_header(page->cp_obj);
 141
 142         if (!cfs_atomic_read(&page->cp_ref))
 143                 LASSERT_SPIN_LOCKED(&ch->coh_page_guard);
 144 #endif
 145         ENTRY;
 146
 147         page = cl_page_top_trusted((struct cl_page *)page);
 148         do {
 149                 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 150                         if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
 151                                 RETURN(slice);
 152                 }
 153                 page = page->cp_child;
 154         } while (page != NULL);
 155         RETURN(NULL);
 156 }
 157
 158 /**
 159  * Returns a page with given index in the given object, or NULL if no page is
 160  * found. Acquires a reference on \a page.
 161  *
 162  * Locking: called under cl_object_header::coh_page_guard spin-lock.
 163  */
 164 struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index)
 165 {
 166         struct cl_page *page;
 167
 168         LASSERT_SPIN_LOCKED(&hdr->coh_page_guard);
 169
 170         page = radix_tree_lookup(&hdr->coh_tree, index);
 171         if (page != NULL) {
 172                 cl_page_get_trust(page);
 173         }
 174         return page;
 175 }
 176 EXPORT_SYMBOL(cl_page_lookup);
 177
 178 /**
 179  * Returns a list of pages by a given [start, end] of \a obj.
 180  *
 181  * \param resched If not NULL, then we give up before hogging CPU for too
 182  * long and set *resched = 1, in that case caller should implement a retry
 183  * logic.
 184  *
 185  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
 186  * crucial in the face of [offset, EOF] locks.
 187  */
 188 void cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj,
 189                          struct cl_io *io, pgoff_t start, pgoff_t end,
 190                          struct cl_page_list *queue, int nonblock,
 191                          int *resched)
 192 {
 193         struct cl_object_header *hdr;
 194         struct cl_page          *page;
 195         struct cl_page         **pvec;
 196         const struct cl_page_slice  *slice;
 197         const struct lu_device_type *dtype;
 198         pgoff_t                  idx;
 199         unsigned int             nr;
 200         unsigned int             i;
 201         unsigned int             j;
 202         int                    (*page_own)(const struct lu_env *env,
 203                                            struct cl_io *io,
 204                                            struct cl_page *pg);
 205         ENTRY;
 206
 207         if (resched != NULL)
 208                 *resched = 0;
 209         page_own = nonblock ? cl_page_own_try : cl_page_own;
 210
 211         idx = start;
 212         hdr = cl_object_header(obj);
 213         pvec = cl_env_info(env)->clt_pvec;
 214         dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type;
 215         cfs_spin_lock(&hdr->coh_page_guard);
 216         while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec,
 217                                             idx, CLT_PVEC_SIZE)) > 0) {
 218                 idx = pvec[nr - 1]->cp_index + 1;
 219                 for (i = 0, j = 0; i < nr; ++i) {
 220                         page = pvec[i];
 221                         pvec[i] = NULL;
 222
 223                         LASSERT(page->cp_type == CPT_CACHEABLE);
 224                         if (page->cp_index > end)
 225                                 break;
 226                         if (page->cp_state == CPS_FREEING)
 227                                 continue;
 228
 229                         slice = cl_page_at_trusted(page, dtype);
 230                         /*
 231                          * Pages for lsm-less file has no underneath sub-page
 232                          * for osc, in case of ...
 233                          */
 234                         PASSERT(env, page, slice != NULL);
 235
 236                         page = slice->cpl_page;
 237                         /*
 238                          * Can safely call cl_page_get_trust() under
 239                          * radix-tree spin-lock.
 240                          *
 241                          * XXX not true, because @page is from object another
 242                          * than @hdr and protected by different tree lock.
 243                          */
 244                         cl_page_get_trust(page);
 245                         lu_ref_add_atomic(&page->cp_reference,
 246                                           "page_list", cfs_current());
 247                         pvec[j++] = page;
 248                 }
 249
 250                 /*
 251                  * Here a delicate locking dance is performed. Current thread
 252                  * holds a reference to a page, but has to own it before it
 253                  * can be placed into queue. Owning implies waiting, so
 254                  * radix-tree lock is to be released. After a wait one has to
 255                  * check that pages weren't truncated (cl_page_own() returns
 256                  * error in the latter case).
 257                  */
 258                 cfs_spin_unlock(&hdr->coh_page_guard);
 259                 for (i = 0; i < j; ++i) {
 260                         page = pvec[i];
 261                         if (page_own(env, io, page) == 0)
 262                                 cl_page_list_add(queue, page);
 263                         lu_ref_del(&page->cp_reference,
 264                                    "page_list", cfs_current());
 265                         cl_page_put(env, page);
 266                 }
 267                 cfs_spin_lock(&hdr->coh_page_guard);
 268                 if (nr < CLT_PVEC_SIZE)
 269                         break;
 270                 if (resched != NULL && cfs_need_resched()) {
 271                         *resched = 1;
 272                         break;
 273                 }
 274         }
 275         cfs_spin_unlock(&hdr->coh_page_guard);
 276         EXIT;
 277 }
 278 EXPORT_SYMBOL(cl_page_gang_lookup);
 279
 280 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
 281 {
 282         struct cl_object *obj  = page->cp_obj;
 283         struct cl_site   *site = cl_object_site(obj);
 284
 285         PASSERT(env, page, cfs_list_empty(&page->cp_batch));
 286         PASSERT(env, page, page->cp_owner == NULL);
 287         PASSERT(env, page, page->cp_req == NULL);
 288         PASSERT(env, page, page->cp_parent == NULL);
 289         PASSERT(env, page, page->cp_state == CPS_FREEING);
 290
 291         ENTRY;
 292         cfs_might_sleep();
 293         while (!cfs_list_empty(&page->cp_layers)) {
 294                 struct cl_page_slice *slice;
 295
 296                 slice = cfs_list_entry(page->cp_layers.next,
 297                                        struct cl_page_slice, cpl_linkage);
 298                 cfs_list_del_init(page->cp_layers.next);
 299                 slice->cpl_ops->cpo_fini(env, slice);
 300         }
 301         cfs_atomic_dec(&site->cs_pages.cs_total);
 302
 303 #ifdef LUSTRE_PAGESTATE_TRACKING
 304         cfs_atomic_dec(&site->cs_pages_state[page->cp_state]);
 305 #endif
 306         lu_object_ref_del_at(&obj->co_lu, page->cp_obj_ref, "cl_page", page);
 307         cl_object_put(env, obj);
 308         lu_ref_fini(&page->cp_reference);
 309         OBD_SLAB_FREE_PTR(page, cl_page_kmem);
 310         EXIT;
 311 }
 312
 313 /**
 314  * Helper function updating page state. This is the only place in the code
 315  * where cl_page::cp_state field is mutated.
 316  */
 317 static inline void cl_page_state_set_trust(struct cl_page *page,
 318                                            enum cl_page_state state)
 319 {
 320         /* bypass const. */
 321         *(enum cl_page_state *)&page->cp_state = state;
 322 }
 323
 324 static int cl_page_alloc(const struct lu_env *env, struct cl_object *o,
 325                          pgoff_t ind, struct page *vmpage,
 326                          enum cl_page_type type, struct cl_page **out)
 327 {
 328         struct cl_page          *page;
 329         struct cl_page          *err  = NULL;
 330         struct lu_object_header *head;
 331         struct cl_site          *site = cl_object_site(o);
 332         int                      result;
 333
 334         ENTRY;
 335         result = +1;
 336         OBD_SLAB_ALLOC_PTR_GFP(page, cl_page_kmem, CFS_ALLOC_IO);
 337         if (page != NULL) {
 338                 cfs_atomic_set(&page->cp_ref, 1);
 339                 page->cp_obj = o;
 340                 cl_object_get(o);
 341                 page->cp_obj_ref = lu_object_ref_add(&o->co_lu,
 342                                                      "cl_page", page);
 343                 page->cp_index = ind;
 344                 cl_page_state_set_trust(page, CPS_CACHED);
 345                 page->cp_type = type;
 346                 CFS_INIT_LIST_HEAD(&page->cp_layers);
 347                 CFS_INIT_LIST_HEAD(&page->cp_batch);
 348                 CFS_INIT_LIST_HEAD(&page->cp_flight);
 349                 cfs_mutex_init(&page->cp_mutex);
 350                 lu_ref_init(&page->cp_reference);
 351                 head = o->co_lu.lo_header;
 352                 cfs_list_for_each_entry(o, &head->loh_layers,
 353                                         co_lu.lo_linkage) {
 354                         if (o->co_ops->coo_page_init != NULL) {
 355                                 err = o->co_ops->coo_page_init(env, o,
 356                                                                page, vmpage);
 357                                 if (err != NULL) {
 358                                         cl_page_delete0(env, page, 0);
 359                                         cl_page_free(env, page);
 360                                         page = err;
 361                                         break;
 362                                 }
 363                         }
 364                 }
 365                 if (err == NULL) {
 366                         cfs_atomic_inc(&site->cs_pages.cs_busy);
 367                         cfs_atomic_inc(&site->cs_pages.cs_total);
 368
 369 #ifdef LUSTRE_PAGESTATE_TRACKING
 370                         cfs_atomic_inc(&site->cs_pages_state[CPS_CACHED]);
 371 #endif
 372                         cfs_atomic_inc(&site->cs_pages.cs_created);
 373                         result = 0;
 374                 }
 375         } else
 376                 page = ERR_PTR(-ENOMEM);
 377         *out = page;
 378         RETURN(result);
 379 }
 380
 381 /**
 382  * Returns a cl_page with index \a idx at the object \a o, and associated with
 383  * the VM page \a vmpage.
 384  *
 385  * This is the main entry point into the cl_page caching interface. First, a
 386  * cache (implemented as a per-object radix tree) is consulted. If page is
 387  * found there, it is returned immediately. Otherwise new page is allocated
 388  * and returned. In any case, additional reference to page is acquired.
 389  *
 390  * \see cl_object_find(), cl_lock_find()
 391  */
 392 static struct cl_page *cl_page_find0(const struct lu_env *env,
 393                                      struct cl_object *o,
 394                                      pgoff_t idx, struct page *vmpage,
 395                                      enum cl_page_type type,
 396                                      struct cl_page *parent)
 397 {
 398         struct cl_page          *page = NULL;
 399         struct cl_page          *ghost = NULL;
 400         struct cl_object_header *hdr;
 401         struct cl_site          *site = cl_object_site(o);
 402         int err;
 403
 404         LINVRNT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 405         cfs_might_sleep();
 406
 407         ENTRY;
 408
 409         hdr = cl_object_header(o);
 410         cfs_atomic_inc(&site->cs_pages.cs_lookup);
 411
 412         CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
 413                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 414         /* fast path. */
 415         if (type == CPT_CACHEABLE) {
 416                 /*
 417                  * cl_vmpage_page() can be called here without any locks as
 418                  *
 419                  *     - "vmpage" is locked (which prevents ->private from
 420                  *       concurrent updates), and
 421                  *
 422                  *     - "o" cannot be destroyed while current thread holds a
 423                  *       reference on it.
 424                  */
 425                 page = cl_vmpage_page(vmpage, o);
 426                 PINVRNT(env, page,
 427                         ergo(page != NULL,
 428                              cl_page_vmpage(env, page) == vmpage &&
 429                              (void *)radix_tree_lookup(&hdr->coh_tree,
 430                                                        idx) == page));
 431         }
 432
 433         if (page != NULL) {
 434                 cfs_atomic_inc(&site->cs_pages.cs_hit);
 435                 RETURN(page);
 436         }
 437
 438         /* allocate and initialize cl_page */
 439         err = cl_page_alloc(env, o, idx, vmpage, type, &page);
 440         if (err != 0)
 441                 RETURN(page);
 442
 443         if (type == CPT_TRANSIENT) {
 444                 if (parent) {
 445                         LASSERT(page->cp_parent == NULL);
 446                         page->cp_parent = parent;
 447                         parent->cp_child = page;
 448                 }
 449                 RETURN(page);
 450         }
 451
 452         /*
 453          * XXX optimization: use radix_tree_preload() here, and change tree
 454          * gfp mask to GFP_KERNEL in cl_object_header_init().
 455          */
 456         cfs_spin_lock(&hdr->coh_page_guard);
 457         err = radix_tree_insert(&hdr->coh_tree, idx, page);
 458         if (err != 0) {
 459                 ghost = page;
 460                 /*
 461                  * Noted by Jay: a lock on \a vmpage protects cl_page_find()
 462                  * from this race, but
 463                  *
 464                  *     0. it's better to have cl_page interface "locally
 465                  *     consistent" so that its correctness can be reasoned
 466                  *     about without appealing to the (obscure world of) VM
 467                  *     locking.
 468                  *
 469                  *     1. handling this race allows ->coh_tree to remain
 470                  *     consistent even when VM locking is somehow busted,
 471                  *     which is very useful during diagnosing and debugging.
 472                  */
 473                 page = ERR_PTR(err);
 474                 CL_PAGE_DEBUG(D_ERROR, env, ghost,
 475                               "fail to insert into radix tree: %d\n", err);
 476         } else {
 477                 if (parent) {
 478                         LASSERT(page->cp_parent == NULL);
 479                         page->cp_parent = parent;
 480                         parent->cp_child = page;
 481                 }
 482                 hdr->coh_pages++;
 483         }
 484         cfs_spin_unlock(&hdr->coh_page_guard);
 485
 486         if (unlikely(ghost != NULL)) {
 487                 cfs_atomic_dec(&site->cs_pages.cs_busy);
 488                 cl_page_delete0(env, ghost, 0);
 489                 cl_page_free(env, ghost);
 490         }
 491         RETURN(page);
 492 }
 493
 494 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o,
 495                              pgoff_t idx, struct page *vmpage,
 496                              enum cl_page_type type)
 497 {
 498         return cl_page_find0(env, o, idx, vmpage, type, NULL);
 499 }
 500 EXPORT_SYMBOL(cl_page_find);
 501
 502
 503 struct cl_page *cl_page_find_sub(const struct lu_env *env, struct cl_object *o,
 504                                  pgoff_t idx, struct page *vmpage,
 505                                  struct cl_page *parent)
 506 {
 507         return cl_page_find0(env, o, idx, vmpage, parent->cp_type, parent);
 508 }
 509 EXPORT_SYMBOL(cl_page_find_sub);
 510
 511 static inline int cl_page_invariant(const struct cl_page *pg)
 512 {
 513         struct cl_object_header *header;
 514         struct cl_page          *parent;
 515         struct cl_page          *child;
 516         struct cl_io            *owner;
 517
 518         /*
 519          * Page invariant is protected by a VM lock.
 520          */
 521         LINVRNT(cl_page_is_vmlocked(NULL, pg));
 522
 523         header = cl_object_header(pg->cp_obj);
 524         parent = pg->cp_parent;
 525         child  = pg->cp_child;
 526         owner  = pg->cp_owner;
 527
 528         return cfs_atomic_read(&pg->cp_ref) > 0 &&
 529                 ergo(parent != NULL, parent->cp_child == pg) &&
 530                 ergo(child != NULL, child->cp_parent == pg) &&
 531                 ergo(child != NULL, pg->cp_obj != child->cp_obj) &&
 532                 ergo(parent != NULL, pg->cp_obj != parent->cp_obj) &&
 533                 ergo(owner != NULL && parent != NULL,
 534                      parent->cp_owner == pg->cp_owner->ci_parent) &&
 535                 ergo(owner != NULL && child != NULL,
 536                      child->cp_owner->ci_parent == owner) &&
 537                 /*
 538                  * Either page is early in initialization (has neither child
 539                  * nor parent yet), or it is in the object radix tree.
 540                  */
 541                 ergo(pg->cp_state < CPS_FREEING,
 542                      (void *)radix_tree_lookup(&header->coh_tree,
 543                                                pg->cp_index) == pg ||
 544                      (child == NULL && parent == NULL));
 545 }
 546
 547 static void cl_page_state_set0(const struct lu_env *env,
 548                                struct cl_page *page, enum cl_page_state state)
 549 {
 550         enum cl_page_state old;
 551 #ifdef LUSTRE_PAGESTATE_TRACKING
 552         struct cl_site *site = cl_object_site(page->cp_obj);
 553 #endif
 554
 555         /*
 556          * Matrix of allowed state transitions [old][new], for sanity
 557          * checking.
 558          */
 559         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 560                 [CPS_CACHED] = {
 561                         [CPS_CACHED]  = 0,
 562                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 563                         [CPS_PAGEIN]  = 0,
 564                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 565                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 566                 },
 567                 [CPS_OWNED] = {
 568                         [CPS_CACHED]  = 1, /* release to the cache */
 569                         [CPS_OWNED]   = 0,
 570                         [CPS_PAGEIN]  = 1, /* start read immediately */
 571                         [CPS_PAGEOUT] = 1, /* start write immediately */
 572                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 573                 },
 574                 [CPS_PAGEIN] = {
 575                         [CPS_CACHED]  = 1, /* io completion */
 576                         [CPS_OWNED]   = 0,
 577                         [CPS_PAGEIN]  = 0,
 578                         [CPS_PAGEOUT] = 0,
 579                         [CPS_FREEING] = 0,
 580                 },
 581                 [CPS_PAGEOUT] = {
 582                         [CPS_CACHED]  = 1, /* io completion */
 583                         [CPS_OWNED]   = 0,
 584                         [CPS_PAGEIN]  = 0,
 585                         [CPS_PAGEOUT] = 0,
 586                         [CPS_FREEING] = 0,
 587                 },
 588                 [CPS_FREEING] = {
 589                         [CPS_CACHED]  = 0,
 590                         [CPS_OWNED]   = 0,
 591                         [CPS_PAGEIN]  = 0,
 592                         [CPS_PAGEOUT] = 0,
 593                         [CPS_FREEING] = 0,
 594                 }
 595         };
 596
 597         ENTRY;
 598         old = page->cp_state;
 599         PASSERT(env, page, allowed_transitions[old][state]);
 600         CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
 601         for (; page != NULL; page = page->cp_child) {
 602                 PASSERT(env, page, page->cp_state == old);
 603                 PASSERT(env, page,
 604                         equi(state == CPS_OWNED, page->cp_owner != NULL));
 605
 606 #ifdef LUSTRE_PAGESTATE_TRACKING
 607                 cfs_atomic_dec(&site->cs_pages_state[page->cp_state]);
 608                 cfs_atomic_inc(&site->cs_pages_state[state]);
 609 #endif
 610                 cl_page_state_set_trust(page, state);
 611         }
 612         EXIT;
 613 }
 614
 615 static void cl_page_state_set(const struct lu_env *env,
 616                               struct cl_page *page, enum cl_page_state state)
 617 {
 618         PINVRNT(env, page, cl_page_invariant(page));
 619         cl_page_state_set0(env, page, state);
 620 }
 621
 622 /**
 623  * Acquires an additional reference to a page.
 624  *
 625  * This can be called only by caller already possessing a reference to \a
 626  * page.
 627  *
 628  * \see cl_object_get(), cl_lock_get().
 629  */
 630 void cl_page_get(struct cl_page *page)
 631 {
 632         ENTRY;
 633         LASSERT(page->cp_state != CPS_FREEING);
 634         cl_page_get_trust(page);
 635         EXIT;
 636 }
 637 EXPORT_SYMBOL(cl_page_get);
 638
 639 /**
 640  * Releases a reference to a page.
 641  *
 642  * When last reference is released, page is returned to the cache, unless it
 643  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 644  * destroyed.
 645  *
 646  * \see cl_object_put(), cl_lock_put().
 647  */
 648 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 649 {
 650         struct cl_object_header *hdr;
 651         struct cl_site *site = cl_object_site(page->cp_obj);
 652
 653         PASSERT(env, page, cfs_atomic_read(&page->cp_ref) > !!page->cp_parent);
 654
 655         ENTRY;
 656         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 657                        cfs_atomic_read(&page->cp_ref));
 658
 659         hdr = cl_object_header(cl_object_top(page->cp_obj));
 660         if (cfs_atomic_dec_and_lock(&page->cp_ref, &hdr->coh_page_guard)) {
 661                 cfs_atomic_dec(&site->cs_pages.cs_busy);
 662                 /* We're going to access the page w/o a reference, but it's
 663                  * ok because we have grabbed the lock coh_page_guard, which
 664                  * means nobody is able to free this page behind us.
 665                  */
 666                 if (page->cp_state == CPS_FREEING) {
 667                         /* We drop the page reference and check the page state
 668                          * inside the coh_page_guard. So that if it gets here,
 669                          * it is the REALLY last reference to this page.
 670                          */
 671                         cfs_spin_unlock(&hdr->coh_page_guard);
 672
 673                         LASSERT(cfs_atomic_read(&page->cp_ref) == 0);
 674                         PASSERT(env, page, page->cp_owner == NULL);
 675                         PASSERT(env, page, cfs_list_empty(&page->cp_batch));
 676                         /*
 677                          * Page is no longer reachable by other threads. Tear
 678                          * it down.
 679                          */
 680                         cl_page_free(env, page);
 681
 682                         EXIT;
 683                         return;
 684                 }
 685                 cfs_spin_unlock(&hdr->coh_page_guard);
 686         }
 687
 688         EXIT;
 689 }
 690 EXPORT_SYMBOL(cl_page_put);
 691
 692 /**
 693  * Returns a VM page associated with a given cl_page.
 694  */
 695 cfs_page_t *cl_page_vmpage(const struct lu_env *env, struct cl_page *page)
 696 {
 697         const struct cl_page_slice *slice;
 698
 699         /*
 700          * Find uppermost layer with ->cpo_vmpage() method, and return its
 701          * result.
 702          */
 703         page = cl_page_top(page);
 704         do {
 705                 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 706                         if (slice->cpl_ops->cpo_vmpage != NULL)
 707                                 RETURN(slice->cpl_ops->cpo_vmpage(env, slice));
 708                 }
 709                 page = page->cp_child;
 710         } while (page != NULL);
 711         LBUG(); /* ->cpo_vmpage() has to be defined somewhere in the stack */
 712 }
 713 EXPORT_SYMBOL(cl_page_vmpage);
 714
 715 /**
 716  * Returns a cl_page associated with a VM page, and given cl_object.
 717  */
 718 struct cl_page *cl_vmpage_page(cfs_page_t *vmpage, struct cl_object *obj)
 719 {
 720         struct cl_page *page;
 721         struct cl_object_header *hdr;
 722
 723         ENTRY;
 724         KLASSERT(PageLocked(vmpage));
 725
 726         /*
 727          * NOTE: absence of races and liveness of data are guaranteed by page
 728          *       lock on a "vmpage". That works because object destruction has
 729          *       bottom-to-top pass.
 730          */
 731
 732         /*
 733          * This loop assumes that ->private points to the top-most page. This
 734          * can be rectified easily.
 735          */
 736         hdr = cl_object_header(cl_object_top(obj));
 737         cfs_spin_lock(&hdr->coh_page_guard);
 738         for (page = (void *)vmpage->private;
 739              page != NULL; page = page->cp_child) {
 740                 if (cl_object_same(page->cp_obj, obj)) {
 741                         cl_page_get_trust(page);
 742                         break;
 743                 }
 744         }
 745         cfs_spin_unlock(&hdr->coh_page_guard);
 746         LASSERT(ergo(page, page->cp_type == CPT_CACHEABLE));
 747         RETURN(page);
 748 }
 749 EXPORT_SYMBOL(cl_vmpage_page);
 750
 751 /**
 752  * Returns the top-page for a given page.
 753  *
 754  * \see cl_object_top(), cl_io_top()
 755  */
 756 struct cl_page *cl_page_top(struct cl_page *page)
 757 {
 758         return cl_page_top_trusted(page);
 759 }
 760 EXPORT_SYMBOL(cl_page_top);
 761
 762 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 763                                        const struct lu_device_type *dtype)
 764 {
 765         return cl_page_at_trusted(page, dtype);
 766 }
 767 EXPORT_SYMBOL(cl_page_at);
 768
 769 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 770
 771 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)                   \
 772 ({                                                                      \
 773         const struct lu_env        *__env  = (_env);                    \
 774         struct cl_page             *__page = (_page);                   \
 775         const struct cl_page_slice *__scan;                             \
 776         int                         __result;                           \
 777         ptrdiff_t                   __op   = (_op);                     \
 778         int                       (*__method)_proto;                    \
 779                                                                         \
 780         __result = 0;                                                   \
 781         __page = cl_page_top(__page);                                   \
 782         do {                                                            \
 783                 cfs_list_for_each_entry(__scan, &__page->cp_layers,     \
 784                                         cpl_linkage) {                  \
 785                         __method = *(void **)((char *)__scan->cpl_ops + \
 786                                               __op);                    \
 787                         if (__method != NULL) {                         \
 788                                 __result = (*__method)(__env, __scan,   \
 789                                                        ## __VA_ARGS__); \
 790                                 if (__result != 0)                      \
 791                                         break;                          \
 792                         }                                               \
 793                 }                                                       \
 794                 __page = __page->cp_child;                              \
 795         } while (__page != NULL && __result == 0);                      \
 796         if (__result > 0)                                               \
 797                 __result = 0;                                           \
 798         __result;                                                       \
 799 })
 800
 801 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 802 do {                                                                    \
 803         const struct lu_env        *__env  = (_env);                    \
 804         struct cl_page             *__page = (_page);                   \
 805         const struct cl_page_slice *__scan;                             \
 806         ptrdiff_t                   __op   = (_op);                     \
 807         void                      (*__method)_proto;                    \
 808                                                                         \
 809         __page = cl_page_top(__page);                                   \
 810         do {                                                            \
 811                 cfs_list_for_each_entry(__scan, &__page->cp_layers,     \
 812                                         cpl_linkage) {                  \
 813                         __method = *(void **)((char *)__scan->cpl_ops + \
 814                                               __op);                    \
 815                         if (__method != NULL)                           \
 816                                 (*__method)(__env, __scan,              \
 817                                             ## __VA_ARGS__);            \
 818                 }                                                       \
 819                 __page = __page->cp_child;                              \
 820         } while (__page != NULL);                                       \
 821 } while (0)
 822
 823 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)               \
 824 do {                                                                        \
 825         const struct lu_env        *__env  = (_env);                        \
 826         struct cl_page             *__page = (_page);                       \
 827         const struct cl_page_slice *__scan;                                 \
 828         ptrdiff_t                   __op   = (_op);                         \
 829         void                      (*__method)_proto;                        \
 830                                                                             \
 831         /* get to the bottom page. */                                       \
 832         while (__page->cp_child != NULL)                                    \
 833                 __page = __page->cp_child;                                  \
 834         do {                                                                \
 835                 cfs_list_for_each_entry_reverse(__scan, &__page->cp_layers, \
 836                                                 cpl_linkage) {              \
 837                         __method = *(void **)((char *)__scan->cpl_ops +     \
 838                                               __op);                        \
 839                         if (__method != NULL)                               \
 840                                 (*__method)(__env, __scan,                  \
 841                                             ## __VA_ARGS__);                \
 842                 }                                                           \
 843                 __page = __page->cp_parent;                                 \
 844         } while (__page != NULL);                                           \
 845 } while (0)
 846
 847 static int cl_page_invoke(const struct lu_env *env,
 848                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 849
 850 {
 851         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 852         ENTRY;
 853         RETURN(CL_PAGE_INVOKE(env, page, op,
 854                               (const struct lu_env *,
 855                                const struct cl_page_slice *, struct cl_io *),
 856                               io));
 857 }
 858
 859 static void cl_page_invoid(const struct lu_env *env,
 860                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 861
 862 {
 863         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 864         ENTRY;
 865         CL_PAGE_INVOID(env, page, op,
 866                        (const struct lu_env *,
 867                         const struct cl_page_slice *, struct cl_io *), io);
 868         EXIT;
 869 }
 870
 871 static void cl_page_owner_clear(struct cl_page *page)
 872 {
 873         ENTRY;
 874         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 875                 if (page->cp_owner != NULL) {
 876                         LASSERT(page->cp_owner->ci_owned_nr > 0);
 877                         page->cp_owner->ci_owned_nr--;
 878                         page->cp_owner = NULL;
 879                         page->cp_task = NULL;
 880                 }
 881         }
 882         EXIT;
 883 }
 884
 885 static void cl_page_owner_set(struct cl_page *page)
 886 {
 887         ENTRY;
 888         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 889                 LASSERT(page->cp_owner != NULL);
 890                 page->cp_owner->ci_owned_nr++;
 891         }
 892         EXIT;
 893 }
 894
 895 void cl_page_disown0(const struct lu_env *env,
 896                      struct cl_io *io, struct cl_page *pg)
 897 {
 898         enum cl_page_state state;
 899
 900         ENTRY;
 901         state = pg->cp_state;
 902         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 903         PINVRNT(env, pg, cl_page_invariant(pg));
 904         cl_page_owner_clear(pg);
 905
 906         if (state == CPS_OWNED)
 907                 cl_page_state_set(env, pg, CPS_CACHED);
 908         /*
 909          * Completion call-backs are executed in the bottom-up order, so that
 910          * uppermost layer (llite), responsible for VFS/VM interaction runs
 911          * last and can release locks safely.
 912          */
 913         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 914                                (const struct lu_env *,
 915                                 const struct cl_page_slice *, struct cl_io *),
 916                                io);
 917         EXIT;
 918 }
 919
 920 /**
 921  * returns true, iff page is owned by the given io.
 922  */
 923 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 924 {
 925         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 926         ENTRY;
 927         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == io);
 928 }
 929 EXPORT_SYMBOL(cl_page_is_owned);
 930
 931 /**
 932  * Try to own a page by IO.
 933  *
 934  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 935  * into cl_page_state::CPS_OWNED state.
 936  *
 937  * \pre  !cl_page_is_owned(pg, io)
 938  * \post result == 0 iff cl_page_is_owned(pg, io)
 939  *
 940  * \retval 0   success
 941  *
 942  * \retval -ve failure, e.g., page was destroyed (and landed in
 943  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 944  *             or, page was owned by another thread, or in IO.
 945  *
 946  * \see cl_page_disown()
 947  * \see cl_page_operations::cpo_own()
 948  * \see cl_page_own_try()
 949  * \see cl_page_own
 950  */
 951 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 952                         struct cl_page *pg, int nonblock)
 953 {
 954         int result;
 955
 956         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 957
 958         ENTRY;
 959         pg = cl_page_top(pg);
 960         io = cl_io_top(io);
 961
 962         if (pg->cp_state == CPS_FREEING) {
 963                 result = -EAGAIN;
 964         } else {
 965                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 966                                         (const struct lu_env *,
 967                                          const struct cl_page_slice *,
 968                                          struct cl_io *, int),
 969                                         io, nonblock);
 970                 if (result == 0) {
 971                         PASSERT(env, pg, pg->cp_owner == NULL);
 972                         PASSERT(env, pg, pg->cp_req == NULL);
 973                         pg->cp_owner = io;
 974                         pg->cp_task  = current;
 975                         cl_page_owner_set(pg);
 976                         if (pg->cp_state != CPS_FREEING) {
 977                                 cl_page_state_set(env, pg, CPS_OWNED);
 978                         } else {
 979                                 cl_page_disown0(env, io, pg);
 980                                 result = -EAGAIN;
 981                         }
 982                 }
 983         }
 984         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 985         RETURN(result);
 986 }
 987
 988 /**
 989  * Own a page, might be blocked.
 990  *
 991  * \see cl_page_own0()
 992  */
 993 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 994 {
 995         return cl_page_own0(env, io, pg, 0);
 996 }
 997 EXPORT_SYMBOL(cl_page_own);
 998
 999 /**
1000  * Nonblock version of cl_page_own().
1001  *
1002  * \see cl_page_own0()
1003  */
1004 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
1005                     struct cl_page *pg)
1006 {
1007         return cl_page_own0(env, io, pg, 1);
1008 }
1009 EXPORT_SYMBOL(cl_page_own_try);
1010
1011
1012 /**
1013  * Assume page ownership.
1014  *
1015  * Called when page is already locked by the hosting VM.
1016  *
1017  * \pre !cl_page_is_owned(pg, io)
1018  * \post cl_page_is_owned(pg, io)
1019  *
1020  * \see cl_page_operations::cpo_assume()
1021  */
1022 void cl_page_assume(const struct lu_env *env,
1023                     struct cl_io *io, struct cl_page *pg)
1024 {
1025         PASSERT(env, pg, pg->cp_state < CPS_OWNED);
1026         PASSERT(env, pg, pg->cp_owner == NULL);
1027         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
1028         PINVRNT(env, pg, cl_page_invariant(pg));
1029
1030         ENTRY;
1031         pg = cl_page_top(pg);
1032         io = cl_io_top(io);
1033
1034         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
1035         pg->cp_owner = io;
1036         pg->cp_task = current;
1037         cl_page_owner_set(pg);
1038         cl_page_state_set(env, pg, CPS_OWNED);
1039         EXIT;
1040 }
1041 EXPORT_SYMBOL(cl_page_assume);
1042
1043 /**
1044  * Releases page ownership without unlocking the page.
1045  *
1046  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
1047  * underlying VM page (as VM is supposed to do this itself).
1048  *
1049  * \pre   cl_page_is_owned(pg, io)
1050  * \post !cl_page_is_owned(pg, io)
1051  *
1052  * \see cl_page_assume()
1053  */
1054 void cl_page_unassume(const struct lu_env *env,
1055                       struct cl_io *io, struct cl_page *pg)
1056 {
1057         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1058         PINVRNT(env, pg, cl_page_invariant(pg));
1059
1060         ENTRY;
1061         pg = cl_page_top(pg);
1062         io = cl_io_top(io);
1063         cl_page_owner_clear(pg);
1064         cl_page_state_set(env, pg, CPS_CACHED);
1065         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
1066                                (const struct lu_env *,
1067                                 const struct cl_page_slice *, struct cl_io *),
1068                                io);
1069         EXIT;
1070 }
1071 EXPORT_SYMBOL(cl_page_unassume);
1072
1073 /**
1074  * Releases page ownership.
1075  *
1076  * Moves page into cl_page_state::CPS_CACHED.
1077  *
1078  * \pre   cl_page_is_owned(pg, io)
1079  * \post !cl_page_is_owned(pg, io)
1080  *
1081  * \see cl_page_own()
1082  * \see cl_page_operations::cpo_disown()
1083  */
1084 void cl_page_disown(const struct lu_env *env,
1085                     struct cl_io *io, struct cl_page *pg)
1086 {
1087         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1088
1089         ENTRY;
1090         pg = cl_page_top(pg);
1091         io = cl_io_top(io);
1092         cl_page_disown0(env, io, pg);
1093         EXIT;
1094 }
1095 EXPORT_SYMBOL(cl_page_disown);
1096
1097 /**
1098  * Called when page is to be removed from the object, e.g., as a result of
1099  * truncate.
1100  *
1101  * Calls cl_page_operations::cpo_discard() top-to-bottom.
1102  *
1103  * \pre cl_page_is_owned(pg, io)
1104  *
1105  * \see cl_page_operations::cpo_discard()
1106  */
1107 void cl_page_discard(const struct lu_env *env,
1108                      struct cl_io *io, struct cl_page *pg)
1109 {
1110         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1111         PINVRNT(env, pg, cl_page_invariant(pg));
1112
1113         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
1114 }
1115 EXPORT_SYMBOL(cl_page_discard);
1116
1117 /**
1118  * Version of cl_page_delete() that can be called for not fully constructed
1119  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
1120  * path. Doesn't check page invariant.
1121  */
1122 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
1123                             int radix)
1124 {
1125         struct cl_page *tmp = pg;
1126         ENTRY;
1127
1128         PASSERT(env, pg, pg == cl_page_top(pg));
1129         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
1130
1131         /*
1132          * Severe all ways to obtain new pointers to @pg.
1133          */
1134         cl_page_owner_clear(pg);
1135
1136         /*
1137          * unexport the page firstly before freeing it so that
1138          * the page content is considered to be invalid.
1139          * We have to do this because a CPS_FREEING cl_page may
1140          * be NOT under the protection of a cl_lock.
1141          * Afterwards, if this page is found by other threads, then this
1142          * page will be forced to reread.
1143          */
1144         cl_page_export(env, pg, 0);
1145         cl_page_state_set0(env, pg, CPS_FREEING);
1146
1147         if (tmp->cp_type == CPT_CACHEABLE) {
1148                 if (!radix)
1149                         /* !radix means that @pg is not yet in the radix tree,
1150                          * skip removing it.
1151                          */
1152                         tmp = pg->cp_child;
1153                 for (; tmp != NULL; tmp = tmp->cp_child) {
1154                         void                    *value;
1155                         struct cl_object_header *hdr;
1156
1157                         hdr = cl_object_header(tmp->cp_obj);
1158                         cfs_spin_lock(&hdr->coh_page_guard);
1159                         value = radix_tree_delete(&hdr->coh_tree,
1160                                                   tmp->cp_index);
1161                         PASSERT(env, tmp, value == tmp);
1162                         PASSERT(env, tmp, hdr->coh_pages > 0);
1163                         hdr->coh_pages--;
1164                         cfs_spin_unlock(&hdr->coh_page_guard);
1165                 }
1166         }
1167
1168         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete),
1169                        (const struct lu_env *, const struct cl_page_slice *));
1170         EXIT;
1171 }
1172
1173 /**
1174  * Called when a decision is made to throw page out of memory.
1175  *
1176  * Notifies all layers about page destruction by calling
1177  * cl_page_operations::cpo_delete() method top-to-bottom.
1178  *
1179  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
1180  * where transition to this state happens).
1181  *
1182  * Eliminates all venues through which new references to the page can be
1183  * obtained:
1184  *
1185  *     - removes page from the radix trees,
1186  *
1187  *     - breaks linkage from VM page to cl_page.
1188  *
1189  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
1190  * drain after some time, at which point page will be recycled.
1191  *
1192  * \pre  pg == cl_page_top(pg)
1193  * \pre  VM page is locked
1194  * \post pg->cp_state == CPS_FREEING
1195  *
1196  * \see cl_page_operations::cpo_delete()
1197  */
1198 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
1199 {
1200         PINVRNT(env, pg, cl_page_invariant(pg));
1201         ENTRY;
1202         cl_page_delete0(env, pg, 1);
1203         EXIT;
1204 }
1205 EXPORT_SYMBOL(cl_page_delete);
1206
1207 /**
1208  * Unmaps page from user virtual memory.
1209  *
1210  * Calls cl_page_operations::cpo_unmap() through all layers top-to-bottom. The
1211  * layer responsible for VM interaction has to unmap page from user space
1212  * virtual memory.
1213  *
1214  * \see cl_page_operations::cpo_unmap()
1215  */
1216 int cl_page_unmap(const struct lu_env *env,
1217                   struct cl_io *io, struct cl_page *pg)
1218 {
1219         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1220         PINVRNT(env, pg, cl_page_invariant(pg));
1221
1222         return cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_unmap));
1223 }
1224 EXPORT_SYMBOL(cl_page_unmap);
1225
1226 /**
1227  * Marks page up-to-date.
1228  *
1229  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
1230  * layer responsible for VM interaction has to mark/clear page as up-to-date
1231  * by the \a uptodate argument.
1232  *
1233  * \see cl_page_operations::cpo_export()
1234  */
1235 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
1236 {
1237         PINVRNT(env, pg, cl_page_invariant(pg));
1238         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
1239                        (const struct lu_env *,
1240                         const struct cl_page_slice *, int), uptodate);
1241 }
1242 EXPORT_SYMBOL(cl_page_export);
1243
1244 /**
1245  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
1246  * thread.
1247  */
1248 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
1249 {
1250         int result;
1251         const struct cl_page_slice *slice;
1252
1253         ENTRY;
1254         pg = cl_page_top_trusted((struct cl_page *)pg);
1255         slice = container_of(pg->cp_layers.next,
1256                              const struct cl_page_slice, cpl_linkage);
1257         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
1258         /*
1259          * Call ->cpo_is_vmlocked() directly instead of going through
1260          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
1261          * cl_page_invariant().
1262          */
1263         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
1264         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
1265         RETURN(result == -EBUSY);
1266 }
1267 EXPORT_SYMBOL(cl_page_is_vmlocked);
1268
1269 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
1270 {
1271         ENTRY;
1272         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
1273 }
1274
1275 static void cl_page_io_start(const struct lu_env *env,
1276                              struct cl_page *pg, enum cl_req_type crt)
1277 {
1278         /*
1279          * Page is queued for IO, change its state.
1280          */
1281         ENTRY;
1282         cl_page_owner_clear(pg);
1283         cl_page_state_set(env, pg, cl_req_type_state(crt));
1284         EXIT;
1285 }
1286
1287 /**
1288  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
1289  * called top-to-bottom. Every layer either agrees to submit this page (by
1290  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
1291  * handling interactions with the VM also has to inform VM that page is under
1292  * transfer now.
1293  */
1294 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
1295                  struct cl_page *pg, enum cl_req_type crt)
1296 {
1297         int result;
1298
1299         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1300         PINVRNT(env, pg, cl_page_invariant(pg));
1301         PINVRNT(env, pg, crt < CRT_NR);
1302
1303         /*
1304          * XXX this has to be called bottom-to-top, so that llite can set up
1305          * PG_writeback without risking other layers deciding to skip this
1306          * page.
1307          */
1308         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
1309         if (result == 0)
1310                 cl_page_io_start(env, pg, crt);
1311
1312         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
1313                       equi(result == 0,
1314                            PageWriteback(cl_page_vmpage(env, pg)))));
1315         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1316         return result;
1317 }
1318 EXPORT_SYMBOL(cl_page_prep);
1319
1320 /**
1321  * Notify layers about transfer completion.
1322  *
1323  * Invoked by transfer sub-system (which is a part of osc) to notify layers
1324  * that a transfer, of which this page is a part of has completed.
1325  *
1326  * Completion call-backs are executed in the bottom-up order, so that
1327  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1328  * and can release locks safely.
1329  *
1330  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1331  * \post pg->cp_state == CPS_CACHED
1332  *
1333  * \see cl_page_operations::cpo_completion()
1334  */
1335 void cl_page_completion(const struct lu_env *env,
1336                         struct cl_page *pg, enum cl_req_type crt, int ioret)
1337 {
1338         struct cl_sync_io *anchor = pg->cp_sync_io;
1339
1340         PASSERT(env, pg, crt < CRT_NR);
1341         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
1342         PASSERT(env, pg, pg->cp_req == NULL);
1343         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
1344         PINVRNT(env, pg, cl_page_invariant(pg));
1345
1346         ENTRY;
1347         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
1348         if (crt == CRT_READ && ioret == 0) {
1349                 PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED));
1350                 pg->cp_flags |= CPF_READ_COMPLETED;
1351         }
1352
1353         cl_page_state_set(env, pg, CPS_CACHED);
1354         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
1355                                (const struct lu_env *,
1356                                 const struct cl_page_slice *, int), ioret);
1357         if (anchor) {
1358                 LASSERT(pg->cp_sync_io == anchor);
1359                 pg->cp_sync_io = NULL;
1360                 cl_sync_io_note(anchor, ioret);
1361         }
1362
1363         /* Don't assert the page writeback bit here because the lustre file
1364          * may be as a backend of swap space. in this case, the page writeback
1365          * is set by VM, and obvious we shouldn't clear it at all. Fortunately
1366          * this type of pages are all TRANSIENT pages. */
1367         KLASSERT(ergo(pg->cp_type == CPT_CACHEABLE,
1368                       !PageWriteback(cl_page_vmpage(env, pg))));
1369         EXIT;
1370 }
1371 EXPORT_SYMBOL(cl_page_completion);
1372
1373 /**
1374  * Notify layers that transfer formation engine decided to yank this page from
1375  * the cache and to make it a part of a transfer.
1376  *
1377  * \pre  pg->cp_state == CPS_CACHED
1378  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1379  *
1380  * \see cl_page_operations::cpo_make_ready()
1381  */
1382 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
1383                        enum cl_req_type crt)
1384 {
1385         int result;
1386
1387         PINVRNT(env, pg, crt < CRT_NR);
1388
1389         ENTRY;
1390         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
1391                                 (const struct lu_env *,
1392                                  const struct cl_page_slice *));
1393         if (result == 0) {
1394                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
1395                 cl_page_io_start(env, pg, crt);
1396         }
1397         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1398         RETURN(result);
1399 }
1400 EXPORT_SYMBOL(cl_page_make_ready);
1401
1402 /**
1403  * Notify layers that high level io decided to place this page into a cache
1404  * for future transfer.
1405  *
1406  * The layer implementing transfer engine (osc) has to register this page in
1407  * its queues.
1408  *
1409  * \pre  cl_page_is_owned(pg, io)
1410  * \post ergo(result == 0,
1411  *            pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT)
1412  *
1413  * \see cl_page_operations::cpo_cache_add()
1414  */
1415 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
1416                       struct cl_page *pg, enum cl_req_type crt)
1417 {
1418         int result;
1419
1420         PINVRNT(env, pg, crt < CRT_NR);
1421         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1422         PINVRNT(env, pg, cl_page_invariant(pg));
1423
1424         ENTRY;
1425         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_cache_add));
1426         if (result == 0) {
1427                 cl_page_owner_clear(pg);
1428                 cl_page_state_set(env, pg, CPS_CACHED);
1429         }
1430         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1431         RETURN(result);
1432 }
1433 EXPORT_SYMBOL(cl_page_cache_add);
1434
1435 /**
1436  * Checks whether page is protected by any extent lock is at least required
1437  * mode.
1438  *
1439  * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1440  * \see cl_page_operations::cpo_is_under_lock()
1441  */
1442 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1443                           struct cl_page *page)
1444 {
1445         int rc;
1446
1447         PINVRNT(env, page, cl_page_invariant(page));
1448
1449         ENTRY;
1450         rc = CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1451                             (const struct lu_env *,
1452                              const struct cl_page_slice *, struct cl_io *),
1453                             io);
1454         PASSERT(env, page, rc != 0);
1455         RETURN(rc);
1456 }
1457 EXPORT_SYMBOL(cl_page_is_under_lock);
1458
1459 /**
1460  * Purges all cached pages belonging to the object \a obj.
1461  */
1462 int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj)
1463 {
1464         struct cl_thread_info   *info;
1465         struct cl_object        *obj = cl_object_top(clobj);
1466         struct cl_io            *io;
1467         struct cl_page_list     *plist;
1468         int                      resched;
1469         int                      result;
1470
1471         ENTRY;
1472         info  = cl_env_info(env);
1473         plist = &info->clt_list;
1474         io    = &info->clt_io;
1475
1476         /*
1477          * initialize the io. This is ugly since we never do IO in this
1478          * function, we just make cl_page_list functions happy. -jay
1479          */
1480         io->ci_obj = obj;
1481         result = cl_io_init(env, io, CIT_MISC, obj);
1482         if (result != 0) {
1483                 cl_io_fini(env, io);
1484                 RETURN(io->ci_result);
1485         }
1486
1487         do {
1488                 cl_page_list_init(plist);
1489                 cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF, plist, 0,
1490                                     &resched);
1491                 /*
1492                  * Since we're purging the pages of an object, we don't care
1493                  * the possible outcomes of the following functions.
1494                  */
1495                 cl_page_list_unmap(env, io, plist);
1496                 cl_page_list_discard(env, io, plist);
1497                 cl_page_list_disown(env, io, plist);
1498                 cl_page_list_fini(env, plist);
1499
1500                 if (resched)
1501                         cfs_cond_resched();
1502         } while (resched);
1503
1504         cl_io_fini(env, io);
1505         RETURN(result);
1506 }
1507 EXPORT_SYMBOL(cl_pages_prune);
1508
1509 /**
1510  * Tells transfer engine that only part of a page is to be transmitted.
1511  *
1512  * \see cl_page_operations::cpo_clip()
1513  */
1514 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1515                   int from, int to)
1516 {
1517         PINVRNT(env, pg, cl_page_invariant(pg));
1518
1519         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1520         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1521                        (const struct lu_env *,
1522                         const struct cl_page_slice *,int, int),
1523                        from, to);
1524 }
1525 EXPORT_SYMBOL(cl_page_clip);
1526
1527 /**
1528  * Prints human readable representation of \a pg to the \a f.
1529  */
1530 void cl_page_header_print(const struct lu_env *env, void *cookie,
1531                           lu_printer_t printer, const struct cl_page *pg)
1532 {
1533         (*printer)(env, cookie,
1534                    "page@%p[%d %p:%lu ^%p_%p %d %d %d %p %p %#x]\n",
1535                    pg, cfs_atomic_read(&pg->cp_ref), pg->cp_obj,
1536                    pg->cp_index, pg->cp_parent, pg->cp_child,
1537                    pg->cp_state, pg->cp_error, pg->cp_type,
1538                    pg->cp_owner, pg->cp_req, pg->cp_flags);
1539 }
1540 EXPORT_SYMBOL(cl_page_header_print);
1541
1542 /**
1543  * Prints human readable representation of \a pg to the \a f.
1544  */
1545 void cl_page_print(const struct lu_env *env, void *cookie,
1546                    lu_printer_t printer, const struct cl_page *pg)
1547 {
1548         struct cl_page *scan;
1549
1550         for (scan = cl_page_top((struct cl_page *)pg);
1551              scan != NULL; scan = scan->cp_child)
1552                 cl_page_header_print(env, cookie, printer, scan);
1553         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1554                        (const struct lu_env *env,
1555                         const struct cl_page_slice *slice,
1556                         void *cookie, lu_printer_t p), cookie, printer);
1557         (*printer)(env, cookie, "end page@%p\n", pg);
1558 }
1559 EXPORT_SYMBOL(cl_page_print);
1560
1561 /**
1562  * Cancel a page which is still in a transfer.
1563  */
1564 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1565 {
1566         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1567                               (const struct lu_env *,
1568                                const struct cl_page_slice *));
1569 }
1570 EXPORT_SYMBOL(cl_page_cancel);
1571
1572 /**
1573  * Converts a byte offset within object \a obj into a page index.
1574  */
1575 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1576 {
1577         /*
1578          * XXX for now.
1579          */
1580         return (loff_t)idx << CFS_PAGE_SHIFT;
1581 }
1582 EXPORT_SYMBOL(cl_offset);
1583
1584 /**
1585  * Converts a page index into a byte offset within object \a obj.
1586  */
1587 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1588 {
1589         /*
1590          * XXX for now.
1591          */
1592         return offset >> CFS_PAGE_SHIFT;
1593 }
1594 EXPORT_SYMBOL(cl_index);
1595
1596 int cl_page_size(const struct cl_object *obj)
1597 {
1598         return 1 << CFS_PAGE_SHIFT;
1599 }
1600 EXPORT_SYMBOL(cl_page_size);
1601
1602 /**
1603  * Adds page slice to the compound page.
1604  *
1605  * This is called by cl_object_operations::coo_page_init() methods to add a
1606  * per-layer state to the page. New state is added at the end of
1607  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1608  *
1609  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1610  */
1611 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1612                        struct cl_object *obj,
1613                        const struct cl_page_operations *ops)
1614 {
1615         ENTRY;
1616         cfs_list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1617         slice->cpl_obj  = obj;
1618         slice->cpl_ops  = ops;
1619         slice->cpl_page = page;
1620         EXIT;
1621 }
1622 EXPORT_SYMBOL(cl_page_slice_add);
1623
1624 int  cl_page_init(void)
1625 {
1626         return lu_kmem_init(cl_page_caches);
1627 }
1628
1629 void cl_page_fini(void)
1630 {
1631         lu_kmem_fini(cl_page_caches);
1632 }