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 2008 Sun Microsystems, Inc.  All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * 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         LASSERT(cl_is_page(page));
 103         while (page->cp_parent != NULL)
 104                 page = page->cp_parent;
 105         return page;
 106 }
 107
 108 /**
 109  * Internal version of cl_page_get().
 110  *
 111  * This function can be used to obtain initial reference to previously
 112  * unreferenced cached object. It can be called only if concurrent page
 113  * reclamation is somehow prevented, e.g., by locking page radix-tree
 114  * (cl_object_header::hdr->coh_page_guard), or by keeping a lock on a VM page,
 115  * associated with \a page.
 116  *
 117  * Use with care! Not exported.
 118  */
 119 static void cl_page_get_trust(struct cl_page *page)
 120 {
 121         LASSERT(cl_is_page(page));
 122         /*
 123          * Checkless version for trusted users.
 124          */
 125         if (atomic_inc_return(&page->cp_ref) == 1)
 126                 atomic_inc(&cl_object_site(page->cp_obj)->cs_pages.cs_busy);
 127 }
 128
 129 /**
 130  * Returns a slice within a page, corresponding to the given layer in the
 131  * device stack.
 132  *
 133  * \see cl_lock_at()
 134  */
 135 static const struct cl_page_slice *
 136 cl_page_at_trusted(const struct cl_page *page,
 137                    const struct lu_device_type *dtype)
 138 {
 139         const struct cl_page_slice *slice;
 140
 141 #ifdef INVARIANT_CHECK
 142         struct cl_object_header *ch = cl_object_header(page->cp_obj);
 143
 144         if (!atomic_read(&page->cp_ref))
 145                 LASSERT_SPIN_LOCKED(&ch->coh_page_guard);
 146 #endif
 147         ENTRY;
 148
 149         page = cl_page_top_trusted((struct cl_page *)page);
 150         do {
 151                 list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 152                         if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
 153                                 RETURN(slice);
 154                 }
 155                 page = page->cp_child;
 156         } while (page != NULL);
 157         RETURN(NULL);
 158 }
 159
 160 /**
 161  * Returns a page with given index in the given object, or NULL if no page is
 162  * found. Acquires a reference on \a page.
 163  *
 164  * Locking: called under cl_object_header::coh_page_guard spin-lock.
 165  */
 166 struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index)
 167 {
 168         struct cl_page *page;
 169
 170         LASSERT_SPIN_LOCKED(&hdr->coh_page_guard);
 171
 172         page = radix_tree_lookup(&hdr->coh_tree, index);
 173         if (page != NULL) {
 174                 LASSERT(cl_is_page(page));
 175                 cl_page_get_trust(page);
 176         }
 177         return page;
 178 }
 179 EXPORT_SYMBOL(cl_page_lookup);
 180
 181 /**
 182  * Returns a list of pages by a given [start, end] of @obj.
 183  *
 184  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
 185  * crucial in the face of [offset, EOF] locks.
 186  */
 187 void cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj,
 188                          struct cl_io *io, pgoff_t start, pgoff_t end,
 189                          struct cl_page_list *queue)
 190 {
 191         struct cl_object_header *hdr;
 192         struct cl_page          *page;
 193         struct cl_page         **pvec;
 194         const struct cl_page_slice  *slice;
 195         const struct lu_device_type *dtype;
 196         pgoff_t                  idx;
 197         unsigned int             nr;
 198         unsigned int             i;
 199         unsigned int             j;
 200         ENTRY;
 201
 202         idx = start;
 203         hdr = cl_object_header(obj);
 204         pvec = cl_env_info(env)->clt_pvec;
 205         dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type;
 206         spin_lock(&hdr->coh_page_guard);
 207         while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec,
 208                                             idx, CLT_PVEC_SIZE)) > 0) {
 209                 idx = pvec[nr - 1]->cp_index + 1;
 210                 for (i = 0, j = 0; i < nr; ++i) {
 211                         page = pvec[i];
 212                         PASSERT(env, page, cl_is_page(page));
 213                         pvec[i] = NULL;
 214                         if (page->cp_index > end)
 215                                 break;
 216                         if (page->cp_state == CPS_FREEING)
 217                                 continue;
 218                         if (page->cp_type == CPT_TRANSIENT) {
 219                                 /* God, we found a transient page!*/
 220                                 continue;
 221                         }
 222
 223                         slice = cl_page_at_trusted(page, dtype);
 224                         /*
 225                          * Pages for lsm-less file has no underneath sub-page
 226                          * for osc, in case of ...
 227                          */
 228                         PASSERT(env, page, slice != NULL);
 229                         page = slice->cpl_page;
 230                         /*
 231                          * Can safely call cl_page_get_trust() under
 232                          * radix-tree spin-lock.
 233                          *
 234                          * XXX not true, because @page is from object another
 235                          * than @hdr and protected by different tree lock.
 236                          */
 237                         cl_page_get_trust(page);
 238                         lu_ref_add_atomic(&page->cp_reference,
 239                                           "page_list", cfs_current());
 240                         pvec[j++] = page;
 241                 }
 242
 243                 /*
 244                  * Here a delicate locking dance is performed. Current thread
 245                  * holds a reference to a page, but has to own it before it
 246                  * can be placed into queue. Owning implies waiting, so
 247                  * radix-tree lock is to be released. After a wait one has to
 248                  * check that pages weren't truncated (cl_page_own() returns
 249                  * error in the latter case).
 250                  */
 251                 spin_unlock(&hdr->coh_page_guard);
 252                 for (i = 0; i < j; ++i) {
 253                         page = pvec[i];
 254                         if (cl_page_own(env, io, page) == 0)
 255                                 cl_page_list_add(queue, page);
 256                         lu_ref_del(&page->cp_reference,
 257                                    "page_list", cfs_current());
 258                         cl_page_put(env, page);
 259                 }
 260                 spin_lock(&hdr->coh_page_guard);
 261                 if (nr < CLT_PVEC_SIZE)
 262                         break;
 263         }
 264         spin_unlock(&hdr->coh_page_guard);
 265         EXIT;
 266 }
 267 EXPORT_SYMBOL(cl_page_gang_lookup);
 268
 269 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
 270 {
 271         struct cl_object *obj  = page->cp_obj;
 272         struct cl_site   *site = cl_object_site(obj);
 273
 274         PASSERT(env, page, cl_is_page(page));
 275         PASSERT(env, page, list_empty(&page->cp_batch));
 276         PASSERT(env, page, page->cp_owner == NULL);
 277         PASSERT(env, page, page->cp_req == NULL);
 278         PASSERT(env, page, page->cp_parent == NULL);
 279         PASSERT(env, page, page->cp_state == CPS_FREEING);
 280
 281         ENTRY;
 282         might_sleep();
 283         while (!list_empty(&page->cp_layers)) {
 284                 struct cl_page_slice *slice;
 285
 286                 slice = list_entry(page->cp_layers.next, struct cl_page_slice,
 287                                    cpl_linkage);
 288                 list_del_init(page->cp_layers.next);
 289                 slice->cpl_ops->cpo_fini(env, slice);
 290         }
 291         atomic_dec(&site->cs_pages.cs_total);
 292         atomic_dec(&site->cs_pages_state[page->cp_state]);
 293         lu_object_ref_del_at(&obj->co_lu, page->cp_obj_ref, "cl_page", page);
 294         cl_object_put(env, obj);
 295         lu_ref_fini(&page->cp_reference);
 296         OBD_SLAB_FREE_PTR(page, cl_page_kmem);
 297         EXIT;
 298 }
 299
 300 /**
 301  * Helper function updating page state. This is the only place in the code
 302  * where cl_page::cp_state field is mutated.
 303  */
 304 static inline void cl_page_state_set_trust(struct cl_page *page,
 305                                            enum cl_page_state state)
 306 {
 307         /* bypass const. */
 308         *(enum cl_page_state *)&page->cp_state = state;
 309 }
 310
 311 static int cl_page_alloc(const struct lu_env *env, struct cl_object *o,
 312                          pgoff_t ind, struct page *vmpage,
 313                          enum cl_page_type type, struct cl_page **out)
 314 {
 315         struct cl_page          *page;
 316         struct cl_page          *err  = NULL;
 317         struct lu_object_header *head;
 318         struct cl_site          *site = cl_object_site(o);
 319         int                      result;
 320
 321         ENTRY;
 322         result = +1;
 323         OBD_SLAB_ALLOC_PTR_GFP(page, cl_page_kmem, CFS_ALLOC_IO);
 324         if (page != NULL) {
 325                 atomic_set(&page->cp_ref, 1);
 326                 page->cp_obj = o;
 327                 cl_object_get(o);
 328                 page->cp_obj_ref = lu_object_ref_add(&o->co_lu,
 329                                                      "cl_page", page);
 330                 page->cp_index = ind;
 331                 cl_page_state_set_trust(page, CPS_CACHED);
 332                 page->cp_type = type;
 333                 CFS_INIT_LIST_HEAD(&page->cp_layers);
 334                 CFS_INIT_LIST_HEAD(&page->cp_batch);
 335                 CFS_INIT_LIST_HEAD(&page->cp_flight);
 336                 mutex_init(&page->cp_mutex);
 337                 lu_ref_init(&page->cp_reference);
 338                 head = o->co_lu.lo_header;
 339                 list_for_each_entry(o, &head->loh_layers, co_lu.lo_linkage) {
 340                         if (o->co_ops->coo_page_init != NULL) {
 341                                 err = o->co_ops->coo_page_init(env, o,
 342                                                                page, vmpage);
 343                                 if (err != NULL) {
 344                                         cl_page_state_set_trust(page,
 345                                                                 CPS_FREEING);
 346                                         cl_page_free(env, page);
 347                                         page = err;
 348                                         break;
 349                                 }
 350                         }
 351                 }
 352                 if (err == NULL) {
 353                         atomic_inc(&site->cs_pages.cs_busy);
 354                         atomic_inc(&site->cs_pages.cs_total);
 355                         atomic_inc(&site->cs_pages_state[CPS_CACHED]);
 356                         atomic_inc(&site->cs_pages.cs_created);
 357                         result = 0;
 358                 }
 359         } else
 360                 page = ERR_PTR(-ENOMEM);
 361         *out = page;
 362         RETURN(result);
 363 }
 364
 365 /**
 366  * Returns a cl_page with index \a idx at the object \a o, and associated with
 367  * the VM page \a vmpage.
 368  *
 369  * This is the main entry point into the cl_page caching interface. First, a
 370  * cache (implemented as a per-object radix tree) is consulted. If page is
 371  * found there, it is returned immediately. Otherwise new page is allocated
 372  * and returned. In any case, additional reference to page is acquired.
 373  *
 374  * \see cl_object_find(), cl_lock_find()
 375  */
 376 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o,
 377                              pgoff_t idx, struct page *vmpage,
 378                              enum cl_page_type type)
 379 {
 380         struct cl_page          *page;
 381         struct cl_page          *ghost = NULL;
 382         struct cl_object_header *hdr;
 383         struct cl_site          *site = cl_object_site(o);
 384         int err;
 385
 386         LINVRNT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 387         might_sleep();
 388
 389         ENTRY;
 390
 391         hdr = cl_object_header(o);
 392         atomic_inc(&site->cs_pages.cs_lookup);
 393
 394         CDEBUG(D_PAGE, "%lu@"DFID" %p %lu %i\n",
 395                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 396         /* fast path. */
 397         if (type == CPT_CACHEABLE) {
 398                 /*
 399                  * cl_vmpage_page() can be called here without any locks as
 400                  *
 401                  *     - "vmpage" is locked (which prevents ->private from
 402                  *       concurrent updates), and
 403                  *
 404                  *     - "o" cannot be destroyed while current thread holds a
 405                  *       reference on it.
 406                  */
 407                 page = cl_vmpage_page(vmpage, o);
 408                 PINVRNT(env, page,
 409                         ergo(page != NULL,
 410                              cl_page_vmpage(env, page) == vmpage &&
 411                              (void *)radix_tree_lookup(&hdr->coh_tree,
 412                                                        idx) == page));
 413         } else {
 414                 spin_lock(&hdr->coh_page_guard);
 415                 page = cl_page_lookup(hdr, idx);
 416                 spin_unlock(&hdr->coh_page_guard);
 417         }
 418         if (page != NULL) {
 419                 atomic_inc(&site->cs_pages.cs_hit);
 420                 RETURN(page);
 421         }
 422
 423         /* allocate and initialize cl_page */
 424         err = cl_page_alloc(env, o, idx, vmpage, type, &page);
 425         if (err != 0)
 426                 RETURN(page);
 427         /*
 428          * XXX optimization: use radix_tree_preload() here, and change tree
 429          * gfp mask to GFP_KERNEL in cl_object_header_init().
 430          */
 431         spin_lock(&hdr->coh_page_guard);
 432         err = radix_tree_insert(&hdr->coh_tree, idx, page);
 433         if (err != 0) {
 434                 ghost = page;
 435                 /*
 436                  * Noted by Jay: a lock on \a vmpage protects cl_page_find()
 437                  * from this race, but
 438                  *
 439                  *     0. it's better to have cl_page interface "locally
 440                  *     consistent" so that its correctness can be reasoned
 441                  *     about without appealing to the (obscure world of) VM
 442                  *     locking.
 443                  *
 444                  *     1. handling this race allows ->coh_tree to remain
 445                  *     consistent even when VM locking is somehow busted,
 446                  *     which is very useful during diagnosing and debugging.
 447                  */
 448                 if (err == -EEXIST) {
 449                         /*
 450                          * XXX in case of a lookup for CPT_TRANSIENT page,
 451                          * nothing protects a CPT_CACHEABLE page from being
 452                          * concurrently moved into CPS_FREEING state.
 453                          */
 454                         page = cl_page_lookup(hdr, idx);
 455                         PASSERT(env, page, page != NULL);
 456                         if (page->cp_type == CPT_TRANSIENT &&
 457                             type == CPT_CACHEABLE) {
 458                                 /* XXX: We should make sure that inode sem
 459                                  * keeps being held in the lifetime of
 460                                  * transient pages, so it is impossible to
 461                                  * have conflicting transient pages.
 462                                  */
 463                                 spin_unlock(&hdr->coh_page_guard);
 464                                 cl_page_put(env, page);
 465                                 spin_lock(&hdr->coh_page_guard);
 466                                 page = ERR_PTR(-EBUSY);
 467                         }
 468                 } else
 469                         page = ERR_PTR(err);
 470         } else
 471                 hdr->coh_pages++;
 472         spin_unlock(&hdr->coh_page_guard);
 473
 474         if (unlikely(ghost != NULL)) {
 475                 atomic_dec(&site->cs_pages.cs_busy);
 476                 cl_page_delete0(env, ghost, 0);
 477                 cl_page_free(env, ghost);
 478         }
 479         RETURN(page);
 480 }
 481 EXPORT_SYMBOL(cl_page_find);
 482
 483 static inline int cl_page_invariant(const struct cl_page *pg)
 484 {
 485         struct cl_object_header *header;
 486         struct cl_page          *parent;
 487         struct cl_page          *child;
 488         struct cl_io            *owner;
 489
 490         LASSERT(cl_is_page(pg));
 491         /*
 492          * Page invariant is protected by a VM lock.
 493          */
 494         LINVRNT(cl_page_is_vmlocked(NULL, pg));
 495
 496         header = cl_object_header(pg->cp_obj);
 497         parent = pg->cp_parent;
 498         child  = pg->cp_child;
 499         owner  = pg->cp_owner;
 500
 501         return atomic_read(&pg->cp_ref) > 0 &&
 502                 ergo(parent != NULL, parent->cp_child == pg) &&
 503                 ergo(child != NULL, child->cp_parent == pg) &&
 504                 ergo(child != NULL, pg->cp_obj != child->cp_obj) &&
 505                 ergo(parent != NULL, pg->cp_obj != parent->cp_obj) &&
 506                 ergo(owner != NULL && parent != NULL,
 507                      parent->cp_owner == pg->cp_owner->ci_parent) &&
 508                 ergo(owner != NULL && child != NULL,
 509                      child->cp_owner->ci_parent == owner) &&
 510                 /*
 511                  * Either page is early in initialization (has neither child
 512                  * nor parent yet), or it is in the object radix tree.
 513                  */
 514                 ergo(pg->cp_state < CPS_FREEING,
 515                      (void *)radix_tree_lookup(&header->coh_tree,
 516                                                pg->cp_index) == pg ||
 517                      (child == NULL && parent == NULL));
 518 }
 519
 520 static void cl_page_state_set0(const struct lu_env *env,
 521                                struct cl_page *page, enum cl_page_state state)
 522 {
 523         enum cl_page_state old;
 524         struct cl_site *site = cl_object_site(page->cp_obj);
 525
 526         /*
 527          * Matrix of allowed state transitions [old][new], for sanity
 528          * checking.
 529          */
 530         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 531                 [CPS_CACHED] = {
 532                         [CPS_CACHED]  = 0,
 533                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 534                         [CPS_PAGEIN]  = 0,
 535                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 536                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 537                 },
 538                 [CPS_OWNED] = {
 539                         [CPS_CACHED]  = 1, /* release to the cache */
 540                         [CPS_OWNED]   = 0,
 541                         [CPS_PAGEIN]  = 1, /* start read immediately */
 542                         [CPS_PAGEOUT] = 1, /* start write immediately */
 543                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 544                 },
 545                 [CPS_PAGEIN] = {
 546                         [CPS_CACHED]  = 1, /* io completion */
 547                         [CPS_OWNED]   = 0,
 548                         [CPS_PAGEIN]  = 0,
 549                         [CPS_PAGEOUT] = 0,
 550                         [CPS_FREEING] = 0,
 551                 },
 552                 [CPS_PAGEOUT] = {
 553                         [CPS_CACHED]  = 1, /* io completion */
 554                         [CPS_OWNED]   = 0,
 555                         [CPS_PAGEIN]  = 0,
 556                         [CPS_PAGEOUT] = 0,
 557                         [CPS_FREEING] = 0,
 558                 },
 559                 [CPS_FREEING] = {
 560                         [CPS_CACHED]  = 0,
 561                         [CPS_OWNED]   = 0,
 562                         [CPS_PAGEIN]  = 0,
 563                         [CPS_PAGEOUT] = 0,
 564                         [CPS_FREEING] = 0,
 565                 }
 566         };
 567
 568         ENTRY;
 569         old = page->cp_state;
 570         PASSERT(env, page, allowed_transitions[old][state]);
 571         CL_PAGE_HEADER(D_TRACE, env, page, "%i -> %i\n", old, state);
 572         for (; page != NULL; page = page->cp_child) {
 573                 PASSERT(env, page, page->cp_state == old);
 574                 PASSERT(env, page,
 575                         equi(state == CPS_OWNED, page->cp_owner != NULL));
 576
 577                 atomic_dec(&site->cs_pages_state[page->cp_state]);
 578                 atomic_inc(&site->cs_pages_state[state]);
 579                 cl_page_state_set_trust(page, state);
 580         }
 581         EXIT;
 582 }
 583
 584 static void cl_page_state_set(const struct lu_env *env,
 585                               struct cl_page *page, enum cl_page_state state)
 586 {
 587         PINVRNT(env, page, cl_page_invariant(page));
 588         cl_page_state_set0(env, page, state);
 589 }
 590
 591 /**
 592  * Acquires an additional reference to a page.
 593  *
 594  * This can be called only by caller already possessing a reference to \a
 595  * page.
 596  *
 597  * \see cl_object_get(), cl_lock_get().
 598  */
 599 void cl_page_get(struct cl_page *page)
 600 {
 601         ENTRY;
 602         LASSERT(page->cp_state != CPS_FREEING);
 603         cl_page_get_trust(page);
 604         EXIT;
 605 }
 606 EXPORT_SYMBOL(cl_page_get);
 607
 608 /**
 609  * Releases a reference to a page.
 610  *
 611  * When last reference is released, page is returned to the cache, unless it
 612  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 613  * destroyed.
 614  *
 615  * \see cl_object_put(), cl_lock_put().
 616  */
 617 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 618 {
 619         struct cl_object_header *hdr;
 620         struct cl_site *site = cl_object_site(page->cp_obj);
 621
 622         PASSERT(env, page, atomic_read(&page->cp_ref) > !!page->cp_parent);
 623
 624         ENTRY;
 625         CL_PAGE_HEADER(D_TRACE, env, page, "%i\n", atomic_read(&page->cp_ref));
 626         hdr = cl_object_header(page->cp_obj);
 627         if (atomic_dec_and_test(&page->cp_ref)) {
 628                 atomic_dec(&site->cs_pages.cs_busy);
 629                 if (page->cp_state == CPS_FREEING) {
 630                         PASSERT(env, page, page->cp_owner == NULL);
 631                         PASSERT(env, page, list_empty(&page->cp_batch));
 632                         /*
 633                          * Page is no longer reachable by other threads. Tear
 634                          * it down.
 635                          */
 636                         cl_page_free(env, page);
 637                 }
 638         }
 639         EXIT;
 640 }
 641 EXPORT_SYMBOL(cl_page_put);
 642
 643 /**
 644  * Returns a VM page associated with a given cl_page.
 645  */
 646 cfs_page_t *cl_page_vmpage(const struct lu_env *env, struct cl_page *page)
 647 {
 648         const struct cl_page_slice *slice;
 649
 650         /*
 651          * Find uppermost layer with ->cpo_vmpage() method, and return its
 652          * result.
 653          */
 654         page = cl_page_top(page);
 655         do {
 656                 list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 657                         if (slice->cpl_ops->cpo_vmpage != NULL)
 658                                 RETURN(slice->cpl_ops->cpo_vmpage(env, slice));
 659                 }
 660                 page = page->cp_child;
 661         } while (page != NULL);
 662         LBUG(); /* ->cpo_vmpage() has to be defined somewhere in the stack */
 663 }
 664 EXPORT_SYMBOL(cl_page_vmpage);
 665
 666 /**
 667  * Returns a cl_page associated with a VM page, and given cl_object.
 668  */
 669 struct cl_page *cl_vmpage_page(cfs_page_t *vmpage, struct cl_object *obj)
 670 {
 671         struct cl_page *page;
 672
 673         ENTRY;
 674         KLASSERT(PageLocked(vmpage));
 675
 676         /*
 677          * NOTE: absence of races and liveness of data are guaranteed by page
 678          *       lock on a "vmpage". That works because object destruction has
 679          *       bottom-to-top pass.
 680          */
 681
 682         /*
 683          * This loop assumes that ->private points to the top-most page. This
 684          * can be rectified easily.
 685          */
 686         for (page = (void *)vmpage->private;
 687              page != NULL; page = page->cp_child) {
 688                 if (cl_object_same(page->cp_obj, obj)) {
 689                         cl_page_get_trust(page);
 690                         break;
 691                 }
 692         }
 693         LASSERT(ergo(page, cl_is_page(page) && page->cp_type == CPT_CACHEABLE));
 694         RETURN(page);
 695 }
 696 EXPORT_SYMBOL(cl_vmpage_page);
 697
 698 /**
 699  * Returns the top-page for a given page.
 700  *
 701  * \see cl_object_top(), cl_io_top()
 702  */
 703 struct cl_page *cl_page_top(struct cl_page *page)
 704 {
 705         return cl_page_top_trusted(page);
 706 }
 707 EXPORT_SYMBOL(cl_page_top);
 708
 709 /**
 710  * Returns true if \a addr is an address of an allocated cl_page. Used in
 711  * assertions. This check is optimistically imprecise, i.e., it occasionally
 712  * returns true for the incorrect addresses, but if it returns false, then the
 713  * address is guaranteed to be incorrect. (Should be named cl_pagep().)
 714  *
 715  * \see cl_is_lock()
 716  */
 717 int cl_is_page(const void *addr)
 718 {
 719         return cfs_mem_is_in_cache(addr, cl_page_kmem);
 720 }
 721 EXPORT_SYMBOL(cl_is_page);
 722
 723 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 724                                        const struct lu_device_type *dtype)
 725 {
 726         return cl_page_at_trusted(page, dtype);
 727 }
 728 EXPORT_SYMBOL(cl_page_at);
 729
 730 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 731
 732 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)                   \
 733 ({                                                                      \
 734         const struct lu_env        *__env  = (_env);                    \
 735         struct cl_page             *__page = (_page);                   \
 736         const struct cl_page_slice *__scan;                             \
 737         int                         __result;                           \
 738         ptrdiff_t                   __op   = (_op);                     \
 739         int                       (*__method)_proto;                    \
 740                                                                         \
 741         __result = 0;                                                   \
 742         __page = cl_page_top(__page);                                   \
 743         do {                                                            \
 744                 list_for_each_entry(__scan, &__page->cp_layers,         \
 745                                     cpl_linkage) {                      \
 746                         __method = *(void **)((char *)__scan->cpl_ops + \
 747                                               __op);                    \
 748                         if (__method != NULL) {                         \
 749                                 __result = (*__method)(__env, __scan,   \
 750                                                        ## __VA_ARGS__); \
 751                                 if (__result != 0)                      \
 752                                         break;                          \
 753                         }                                               \
 754                 }                                                       \
 755                 __page = __page->cp_child;                              \
 756         } while (__page != NULL && __result == 0);                      \
 757         if (__result > 0)                                               \
 758                 __result = 0;                                           \
 759         __result;                                                       \
 760 })
 761
 762 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 763 do {                                                                    \
 764         const struct lu_env        *__env  = (_env);                    \
 765         struct cl_page             *__page = (_page);                   \
 766         const struct cl_page_slice *__scan;                             \
 767         ptrdiff_t                   __op   = (_op);                     \
 768         void                      (*__method)_proto;                    \
 769                                                                         \
 770         __page = cl_page_top(__page);                                   \
 771         do {                                                            \
 772                 list_for_each_entry(__scan, &__page->cp_layers,         \
 773                                     cpl_linkage) {                      \
 774                         __method = *(void **)((char *)__scan->cpl_ops + \
 775                                               __op);                    \
 776                         if (__method != NULL)                           \
 777                                 (*__method)(__env, __scan,              \
 778                                             ## __VA_ARGS__);            \
 779                 }                                                       \
 780                 __page = __page->cp_child;                              \
 781         } while (__page != NULL);                                       \
 782 } while (0)
 783
 784 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)           \
 785 do {                                                                    \
 786         const struct lu_env        *__env  = (_env);                    \
 787         struct cl_page             *__page = (_page);                   \
 788         const struct cl_page_slice *__scan;                             \
 789         ptrdiff_t                   __op   = (_op);                     \
 790         void                      (*__method)_proto;                    \
 791                                                                         \
 792         /* get to the bottom page. */                                   \
 793         while (__page->cp_child != NULL)                                \
 794                 __page = __page->cp_child;                              \
 795         do {                                                            \
 796                 list_for_each_entry_reverse(__scan, &__page->cp_layers, \
 797                                             cpl_linkage) {              \
 798                         __method = *(void **)((char *)__scan->cpl_ops + \
 799                                               __op);                    \
 800                         if (__method != NULL)                           \
 801                                 (*__method)(__env, __scan,              \
 802                                             ## __VA_ARGS__);            \
 803                 }                                                       \
 804                 __page = __page->cp_parent;                             \
 805         } while (__page != NULL);                                       \
 806 } while (0)
 807
 808 static int cl_page_invoke(const struct lu_env *env,
 809                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 810
 811 {
 812         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 813         ENTRY;
 814         RETURN(CL_PAGE_INVOKE(env, page, op,
 815                               (const struct lu_env *,
 816                                const struct cl_page_slice *, struct cl_io *),
 817                               io));
 818 }
 819
 820 static void cl_page_invoid(const struct lu_env *env,
 821                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 822
 823 {
 824         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 825         ENTRY;
 826         CL_PAGE_INVOID(env, page, op,
 827                        (const struct lu_env *,
 828                         const struct cl_page_slice *, struct cl_io *), io);
 829         EXIT;
 830 }
 831
 832 static void cl_page_owner_clear(struct cl_page *page)
 833 {
 834         ENTRY;
 835         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 836                 if (page->cp_owner != NULL) {
 837                         LASSERT(page->cp_owner->ci_owned_nr > 0);
 838                         page->cp_owner->ci_owned_nr--;
 839                         page->cp_owner = NULL;
 840                         page->cp_task = NULL;
 841                 }
 842         }
 843         EXIT;
 844 }
 845
 846 static void cl_page_owner_set(struct cl_page *page)
 847 {
 848         ENTRY;
 849         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 850                 LASSERT(page->cp_owner != NULL);
 851                 page->cp_owner->ci_owned_nr++;
 852         }
 853         EXIT;
 854 }
 855
 856 void cl_page_disown0(const struct lu_env *env,
 857                      struct cl_io *io, struct cl_page *pg)
 858 {
 859         enum cl_page_state state;
 860
 861         ENTRY;
 862         state = pg->cp_state;
 863         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 864         PINVRNT(env, pg, cl_page_invariant(pg));
 865         cl_page_owner_clear(pg);
 866
 867         if (state == CPS_OWNED)
 868                 cl_page_state_set(env, pg, CPS_CACHED);
 869         /*
 870          * Completion call-backs are executed in the bottom-up order, so that
 871          * uppermost layer (llite), responsible for VFS/VM interaction runs
 872          * last and can release locks safely.
 873          */
 874         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 875                                (const struct lu_env *,
 876                                 const struct cl_page_slice *, struct cl_io *),
 877                                io);
 878         EXIT;
 879 }
 880
 881 /**
 882  * returns true, iff page is owned by the given io.
 883  */
 884 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 885 {
 886         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 887         ENTRY;
 888         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == io);
 889 }
 890 EXPORT_SYMBOL(cl_page_is_owned);
 891
 892 /**
 893  * Owns a page by IO.
 894  *
 895  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 896  * into cl_page_state::CPS_OWNED state.
 897  *
 898  * \pre  !cl_page_is_owned(pg, io)
 899  * \post result == 0 iff cl_page_is_owned(pg, io)
 900  *
 901  * \retval 0   success
 902  *
 903  * \retval -ve failure, e.g., page was destroyed (and landed in
 904  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 905  *
 906  * \see cl_page_disown()
 907  * \see cl_page_operations::cpo_own()
 908  */
 909 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 910 {
 911         int result;
 912
 913         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 914
 915         ENTRY;
 916         pg = cl_page_top(pg);
 917         io = cl_io_top(io);
 918
 919         if (pg->cp_state == CPS_FREEING) {
 920                 result = -EAGAIN;
 921         } else {
 922                 cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_own));
 923                 PASSERT(env, pg, pg->cp_owner == NULL);
 924                 PASSERT(env, pg, pg->cp_req == NULL);
 925                 pg->cp_owner = io;
 926                 pg->cp_task  = current;
 927                 cl_page_owner_set(pg);
 928                 if (pg->cp_state != CPS_FREEING) {
 929                         cl_page_state_set(env, pg, CPS_OWNED);
 930                         result = 0;
 931                 } else {
 932                         cl_page_disown0(env, io, pg);
 933                         result = -EAGAIN;
 934                 }
 935         }
 936         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 937         RETURN(result);
 938 }
 939 EXPORT_SYMBOL(cl_page_own);
 940
 941 /**
 942  * Assume page ownership.
 943  *
 944  * Called when page is already locked by the hosting VM.
 945  *
 946  * \pre !cl_page_is_owned(pg, io)
 947  * \post cl_page_is_owned(pg, io)
 948  *
 949  * \see cl_page_operations::cpo_assume()
 950  */
 951 void cl_page_assume(const struct lu_env *env,
 952                     struct cl_io *io, struct cl_page *pg)
 953 {
 954         PASSERT(env, pg, pg->cp_state < CPS_OWNED);
 955         PASSERT(env, pg, pg->cp_owner == NULL);
 956         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
 957         PINVRNT(env, pg, cl_page_invariant(pg));
 958
 959         ENTRY;
 960         pg = cl_page_top(pg);
 961         io = cl_io_top(io);
 962
 963         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
 964         pg->cp_owner = io;
 965         pg->cp_task = current;
 966         cl_page_owner_set(pg);
 967         cl_page_state_set(env, pg, CPS_OWNED);
 968         EXIT;
 969 }
 970 EXPORT_SYMBOL(cl_page_assume);
 971
 972 /**
 973  * Releases page ownership without unlocking the page.
 974  *
 975  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
 976  * underlying VM page (as VM is supposed to do this itself).
 977  *
 978  * \pre   cl_page_is_owned(pg, io)
 979  * \post !cl_page_is_owned(pg, io)
 980  *
 981  * \see cl_page_assume()
 982  */
 983 void cl_page_unassume(const struct lu_env *env,
 984                       struct cl_io *io, struct cl_page *pg)
 985 {
 986         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 987         PINVRNT(env, pg, cl_page_invariant(pg));
 988
 989         ENTRY;
 990         pg = cl_page_top(pg);
 991         io = cl_io_top(io);
 992         cl_page_owner_clear(pg);
 993         cl_page_state_set(env, pg, CPS_CACHED);
 994         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
 995                                (const struct lu_env *,
 996                                 const struct cl_page_slice *, struct cl_io *),
 997                                io);
 998         EXIT;
 999 }
1000 EXPORT_SYMBOL(cl_page_unassume);
1001
1002 /**
1003  * Releases page ownership.
1004  *
1005  * Moves page into cl_page_state::CPS_CACHED.
1006  *
1007  * \pre   cl_page_is_owned(pg, io)
1008  * \post !cl_page_is_owned(pg, io)
1009  *
1010  * \see cl_page_own()
1011  * \see cl_page_operations::cpo_disown()
1012  */
1013 void cl_page_disown(const struct lu_env *env,
1014                     struct cl_io *io, struct cl_page *pg)
1015 {
1016         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1017
1018         ENTRY;
1019         pg = cl_page_top(pg);
1020         io = cl_io_top(io);
1021         cl_page_disown0(env, io, pg);
1022         EXIT;
1023 }
1024 EXPORT_SYMBOL(cl_page_disown);
1025
1026 /**
1027  * Called when page is to be removed from the object, e.g., as a result of
1028  * truncate.
1029  *
1030  * Calls cl_page_operations::cpo_discard() top-to-bottom.
1031  *
1032  * \pre cl_page_is_owned(pg, io)
1033  *
1034  * \see cl_page_operations::cpo_discard()
1035  */
1036 void cl_page_discard(const struct lu_env *env,
1037                      struct cl_io *io, struct cl_page *pg)
1038 {
1039         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1040         PINVRNT(env, pg, cl_page_invariant(pg));
1041
1042         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
1043 }
1044 EXPORT_SYMBOL(cl_page_discard);
1045
1046 /**
1047  * Version of cl_page_delete() that can be called for not fully constructed
1048  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
1049  * path. Doesn't check page invariant.
1050  */
1051 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
1052                             int radix)
1053 {
1054         struct cl_page *tmp = pg;
1055         ENTRY;
1056
1057         PASSERT(env, pg, pg == cl_page_top(pg));
1058         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
1059
1060         /*
1061          * Severe all ways to obtain new pointers to @pg.
1062          */
1063         cl_page_owner_clear(pg);
1064
1065         /*
1066          * unexport the page firstly before freeing it so that
1067          * the page content is considered to be invalid.
1068          * We have to do this because a CPS_FREEING cl_page may
1069          * be NOT under the protection of a cl_lock.
1070          * Afterwards, if this page is found by other threads, then this
1071          * page will be forced to reread.
1072          */
1073         cl_page_export(env, pg, 0);
1074         cl_page_state_set0(env, pg, CPS_FREEING);
1075
1076         if (!radix)
1077                 /*
1078                  * !radix means that @pg is not yet in the radix tree, skip
1079                  * removing it.
1080                  */
1081                 tmp = pg->cp_child;
1082         for (; tmp != NULL; tmp = tmp->cp_child) {
1083                 void                    *value;
1084                 struct cl_object_header *hdr;
1085
1086                 hdr = cl_object_header(tmp->cp_obj);
1087                 spin_lock(&hdr->coh_page_guard);
1088                 value = radix_tree_delete(&hdr->coh_tree, tmp->cp_index);
1089                 PASSERT(env, tmp, value == tmp);
1090                 PASSERT(env, tmp, hdr->coh_pages > 0);
1091                 hdr->coh_pages--;
1092                 spin_unlock(&hdr->coh_page_guard);
1093         }
1094
1095         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete),
1096                        (const struct lu_env *, const struct cl_page_slice *));
1097         EXIT;
1098 }
1099
1100 /**
1101  * Called when a decision is made to throw page out of memory.
1102  *
1103  * Notifies all layers about page destruction by calling
1104  * cl_page_operations::cpo_delete() method top-to-bottom.
1105  *
1106  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
1107  * where transition to this state happens).
1108  *
1109  * Eliminates all venues through which new references to the page can be
1110  * obtained:
1111  *
1112  *     - removes page from the radix trees,
1113  *
1114  *     - breaks linkage from VM page to cl_page.
1115  *
1116  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
1117  * drain after some time, at which point page will be recycled.
1118  *
1119  * \pre  pg == cl_page_top(pg)
1120  * \pre  VM page is locked
1121  * \post pg->cp_state == CPS_FREEING
1122  *
1123  * \see cl_page_operations::cpo_delete()
1124  */
1125 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
1126 {
1127         PINVRNT(env, pg, cl_page_invariant(pg));
1128         ENTRY;
1129         cl_page_delete0(env, pg, 1);
1130         EXIT;
1131 }
1132 EXPORT_SYMBOL(cl_page_delete);
1133
1134 /**
1135  * Unmaps page from user virtual memory.
1136  *
1137  * Calls cl_page_operations::cpo_unmap() through all layers top-to-bottom. The
1138  * layer responsible for VM interaction has to unmap page from user space
1139  * virtual memory.
1140  *
1141  * \see cl_page_operations::cpo_unmap()
1142  */
1143 int cl_page_unmap(const struct lu_env *env,
1144                   struct cl_io *io, struct cl_page *pg)
1145 {
1146         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1147         PINVRNT(env, pg, cl_page_invariant(pg));
1148
1149         return cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_unmap));
1150 }
1151 EXPORT_SYMBOL(cl_page_unmap);
1152
1153 /**
1154  * Marks page up-to-date.
1155  *
1156  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
1157  * layer responsible for VM interaction has to mark/clear page as up-to-date
1158  * by the @uptodate argument.
1159  *
1160  * \see cl_page_operations::cpo_export()
1161  */
1162 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
1163 {
1164         PINVRNT(env, pg, cl_page_invariant(pg));
1165         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
1166                        (const struct lu_env *,
1167                         const struct cl_page_slice *, int), uptodate);
1168 }
1169 EXPORT_SYMBOL(cl_page_export);
1170
1171 /**
1172  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
1173  * thread.
1174  */
1175 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
1176 {
1177         int result;
1178         const struct cl_page_slice *slice;
1179
1180         ENTRY;
1181         pg = cl_page_top_trusted((struct cl_page *)pg);
1182         slice = container_of(pg->cp_layers.next,
1183                              const struct cl_page_slice, cpl_linkage);
1184         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
1185         /*
1186          * Call ->cpo_is_vmlocked() directly instead of going through
1187          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
1188          * cl_page_invariant().
1189          */
1190         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
1191         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
1192         RETURN(result == -EBUSY);
1193 }
1194 EXPORT_SYMBOL(cl_page_is_vmlocked);
1195
1196 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
1197 {
1198         ENTRY;
1199         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
1200 }
1201
1202 static void cl_page_io_start(const struct lu_env *env,
1203                              struct cl_page *pg, enum cl_req_type crt)
1204 {
1205         /*
1206          * Page is queued for IO, change its state.
1207          */
1208         ENTRY;
1209         cl_page_owner_clear(pg);
1210         cl_page_state_set(env, pg, cl_req_type_state(crt));
1211         EXIT;
1212 }
1213
1214 /**
1215  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
1216  * called top-to-bottom. Every layer either agrees to submit this page (by
1217  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
1218  * handling interactions with the VM also has to inform VM that page is under
1219  * transfer now.
1220  */
1221 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
1222                  struct cl_page *pg, enum cl_req_type crt)
1223 {
1224         int result;
1225
1226         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1227         PINVRNT(env, pg, cl_page_invariant(pg));
1228         PINVRNT(env, pg, crt < CRT_NR);
1229
1230         /*
1231          * XXX this has to be called bottom-to-top, so that llite can set up
1232          * PG_writeback without risking other layers deciding to skip this
1233          * page.
1234          */
1235         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
1236         if (result == 0)
1237                 cl_page_io_start(env, pg, crt);
1238
1239         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
1240                       equi(result == 0,
1241                            PageWriteback(cl_page_vmpage(env, pg)))));
1242         CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, result);
1243         return result;
1244 }
1245 EXPORT_SYMBOL(cl_page_prep);
1246
1247 /**
1248  * Notify layers about transfer completion.
1249  *
1250  * Invoked by transfer sub-system (which is a part of osc) to notify layers
1251  * that a transfer, of which this page is a part of has completed.
1252  *
1253  * Completion call-backs are executed in the bottom-up order, so that
1254  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1255  * and can release locks safely.
1256  *
1257  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1258  * \post pg->cp_state == CPS_CACHED
1259  *
1260  * \see cl_page_operations::cpo_completion()
1261  */
1262 void cl_page_completion(const struct lu_env *env,
1263                         struct cl_page *pg, enum cl_req_type crt, int ioret)
1264 {
1265         PASSERT(env, pg, crt < CRT_NR);
1266         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
1267         PASSERT(env, pg, pg->cp_req == NULL);
1268         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
1269         PINVRNT(env, pg, cl_page_invariant(pg));
1270
1271         ENTRY;
1272         CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, ioret);
1273         if (crt == CRT_READ && ioret == 0) {
1274                 PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED));
1275                 pg->cp_flags |= CPF_READ_COMPLETED;
1276         }
1277
1278         cl_page_state_set(env, pg, CPS_CACHED);
1279         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
1280                                (const struct lu_env *,
1281                                 const struct cl_page_slice *, int), ioret);
1282
1283         /* Don't assert the page writeback bit here because the lustre file
1284          * may be as a backend of swap space. in this case, the page writeback
1285          * is set by VM, and obvious we shouldn't clear it at all. Fortunately
1286          * this type of pages are all TRANSIENT pages. */
1287         KLASSERT(ergo(pg->cp_type == CPT_CACHEABLE,
1288                       !PageWriteback(cl_page_vmpage(env, pg))));
1289         EXIT;
1290 }
1291 EXPORT_SYMBOL(cl_page_completion);
1292
1293 /**
1294  * Notify layers that transfer formation engine decided to yank this page from
1295  * the cache and to make it a part of a transfer.
1296  *
1297  * \pre  pg->cp_state == CPS_CACHED
1298  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1299  *
1300  * \see cl_page_operations::cpo_make_ready()
1301  */
1302 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
1303                        enum cl_req_type crt)
1304 {
1305         int result;
1306
1307         PINVRNT(env, pg, crt < CRT_NR);
1308
1309         ENTRY;
1310         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
1311                                 (const struct lu_env *,
1312                                  const struct cl_page_slice *));
1313         if (result == 0) {
1314                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
1315                 cl_page_io_start(env, pg, crt);
1316         }
1317         CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, result);
1318         RETURN(result);
1319 }
1320 EXPORT_SYMBOL(cl_page_make_ready);
1321
1322 /**
1323  * Notify layers that high level io decided to place this page into a cache
1324  * for future transfer.
1325  *
1326  * The layer implementing transfer engine (osc) has to register this page in
1327  * its queues.
1328  *
1329  * \pre  cl_page_is_owned(pg, io)
1330  * \post ergo(result == 0,
1331  *            pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT)
1332  *
1333  * \see cl_page_operations::cpo_cache_add()
1334  */
1335 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
1336                       struct cl_page *pg, enum cl_req_type crt)
1337 {
1338         int result;
1339
1340         PINVRNT(env, pg, crt < CRT_NR);
1341         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1342         PINVRNT(env, pg, cl_page_invariant(pg));
1343
1344         ENTRY;
1345         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_cache_add));
1346         if (result == 0) {
1347                 cl_page_owner_clear(pg);
1348                 cl_page_state_set(env, pg, CPS_CACHED);
1349         }
1350         CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, result);
1351         RETURN(result);
1352 }
1353 EXPORT_SYMBOL(cl_page_cache_add);
1354
1355 /**
1356  * Checks whether page is protected by any extent lock is at least required
1357  * mode.
1358  *
1359  * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1360  * \see cl_page_operations::cpo_is_under_lock()
1361  */
1362 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1363                           struct cl_page *page)
1364 {
1365         int rc;
1366
1367         PINVRNT(env, page, cl_page_invariant(page));
1368
1369         ENTRY;
1370         rc = CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1371                             (const struct lu_env *,
1372                              const struct cl_page_slice *, struct cl_io *),
1373                             io);
1374         PASSERT(env, page, rc != 0);
1375         RETURN(rc);
1376 }
1377 EXPORT_SYMBOL(cl_page_is_under_lock);
1378
1379 /**
1380  * Purges all cached pages belonging to the object \a obj.
1381  */
1382 int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj)
1383 {
1384         struct cl_thread_info   *info;
1385         struct cl_object        *obj = cl_object_top(clobj);
1386         struct cl_io            *io;
1387         struct cl_page_list     *plist;
1388         int                      result;
1389
1390         ENTRY;
1391         info  = cl_env_info(env);
1392         plist = &info->clt_list;
1393         io    = &info->clt_io;
1394
1395         /*
1396          * initialize the io. This is ugly since we never do IO in this
1397          * function, we just make cl_page_list functions happy. -jay
1398          */
1399         io->ci_obj = obj;
1400         result = cl_io_init(env, io, CIT_MISC, obj);
1401         if (result != 0) {
1402                 cl_io_fini(env, io);
1403                 RETURN(io->ci_result);
1404         }
1405
1406         cl_page_list_init(plist);
1407         cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF, plist);
1408         /*
1409          * Since we're purging the pages of an object, we don't care
1410          * the possible outcomes of the following functions.
1411          */
1412         cl_page_list_unmap(env, io, plist);
1413         cl_page_list_discard(env, io, plist);
1414         cl_page_list_disown(env, io, plist);
1415         cl_page_list_fini(env, plist);
1416
1417         cl_io_fini(env, io);
1418         RETURN(result);
1419 }
1420 EXPORT_SYMBOL(cl_pages_prune);
1421
1422 /**
1423  * Tells transfer engine that only part of a page is to be transmitted.
1424  *
1425  * \see cl_page_operations::cpo_clip()
1426  */
1427 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1428                   int from, int to)
1429 {
1430         PINVRNT(env, pg, cl_page_invariant(pg));
1431
1432         CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", from, to);
1433         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1434                        (const struct lu_env *,
1435                         const struct cl_page_slice *,int, int),
1436                        from, to);
1437 }
1438 EXPORT_SYMBOL(cl_page_clip);
1439
1440 /**
1441  * Prints human readable representation of \a pg to the \a f.
1442  */
1443 void cl_page_header_print(const struct lu_env *env, void *cookie,
1444                           lu_printer_t printer, const struct cl_page *pg)
1445 {
1446         (*printer)(env, cookie,
1447                    "page@%p[%d %p:%lu ^%p_%p %d %d %d %p %p %#x]\n",
1448                    pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1449                    pg->cp_index, pg->cp_parent, pg->cp_child,
1450                    pg->cp_state, pg->cp_error, pg->cp_type,
1451                    pg->cp_owner, pg->cp_req, pg->cp_flags);
1452 }
1453 EXPORT_SYMBOL(cl_page_header_print);
1454
1455 /**
1456  * Prints human readable representation of \a pg to the \a f.
1457  */
1458 void cl_page_print(const struct lu_env *env, void *cookie,
1459                    lu_printer_t printer, const struct cl_page *pg)
1460 {
1461         struct cl_page *scan;
1462
1463         for (scan = cl_page_top((struct cl_page *)pg);
1464              scan != NULL; scan = scan->cp_child)
1465                 cl_page_header_print(env, cookie, printer, scan);
1466         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1467                        (const struct lu_env *env,
1468                         const struct cl_page_slice *slice,
1469                         void *cookie, lu_printer_t p), cookie, printer);
1470         (*printer)(env, cookie, "end page@%p\n", pg);
1471 }
1472 EXPORT_SYMBOL(cl_page_print);
1473
1474 /**
1475  * Cancel a page which is still in a transfer.
1476  */
1477 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1478 {
1479         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1480                               (const struct lu_env *,
1481                                const struct cl_page_slice *));
1482 }
1483 EXPORT_SYMBOL(cl_page_cancel);
1484
1485 /**
1486  * Converts a byte offset within object \a obj into a page index.
1487  */
1488 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1489 {
1490         /*
1491          * XXX for now.
1492          */
1493         return (loff_t)idx << CFS_PAGE_SHIFT;
1494 }
1495 EXPORT_SYMBOL(cl_offset);
1496
1497 /**
1498  * Converts a page index into a byte offset within object \a obj.
1499  */
1500 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1501 {
1502         /*
1503          * XXX for now.
1504          */
1505         return offset >> CFS_PAGE_SHIFT;
1506 }
1507 EXPORT_SYMBOL(cl_index);
1508
1509 int cl_page_size(const struct cl_object *obj)
1510 {
1511         return 1 << CFS_PAGE_SHIFT;
1512 }
1513 EXPORT_SYMBOL(cl_page_size);
1514
1515 /**
1516  * Adds page slice to the compound page.
1517  *
1518  * This is called by cl_object_operations::coo_page_init() methods to add a
1519  * per-layer state to the page. New state is added at the end of
1520  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1521  *
1522  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1523  */
1524 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1525                        struct cl_object *obj,
1526                        const struct cl_page_operations *ops)
1527 {
1528         ENTRY;
1529         list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1530         slice->cpl_obj  = obj;
1531         slice->cpl_ops  = ops;
1532         slice->cpl_page = page;
1533         EXIT;
1534 }
1535 EXPORT_SYMBOL(cl_page_slice_add);
1536
1537 int  cl_page_init(void)
1538 {
1539         return lu_kmem_init(cl_page_caches);
1540 }
1541
1542 void cl_page_fini(void)
1543 {
1544         lu_kmem_fini(cl_page_caches);
1545 }