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