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

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Client Lustre Page.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_CLASS
  42
  43 #include "../../include/linux/libcfs/libcfs.h"
  44 #include "../include/obd_class.h"
  45 #include "../include/obd_support.h"
  46 #include <linux/list.h>
  47
  48 #include "../include/cl_object.h"
  49 #include "cl_internal.h"
  50
  51 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
  52                             int radix);
  53
  54 # define PASSERT(env, page, expr)                                      \
  55   do {                                                              \
  56           if (unlikely(!(expr))) {                                    \
  57                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  58                   LASSERT(0);                                      \
  59           }                                                          \
  60   } while (0)
  61
  62 # define PINVRNT(env, page, exp) \
  63         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  64
  65 /* Disable page statistic by default due to huge performance penalty. */
  66 #define CS_PAGE_INC(o, item)
  67 #define CS_PAGE_DEC(o, item)
  68 #define CS_PAGESTATE_INC(o, state)
  69 #define CS_PAGESTATE_DEC(o, state)
  70
  71 /**
  72  * Internal version of cl_page_top, it should be called if the page is
  73  * known to be not freed, says with page referenced, or radix tree lock held,
  74  * or page owned.
  75  */
  76 static struct cl_page *cl_page_top_trusted(struct cl_page *page)
  77 {
  78         while (page->cp_parent != NULL)
  79                 page = page->cp_parent;
  80         return page;
  81 }
  82
  83 /**
  84  * Internal version of cl_page_get().
  85  *
  86  * This function can be used to obtain initial reference to previously
  87  * unreferenced cached object. It can be called only if concurrent page
  88  * reclamation is somehow prevented, e.g., by locking page radix-tree
  89  * (cl_object_header::hdr->coh_page_guard), or by keeping a lock on a VM page,
  90  * associated with \a page.
  91  *
  92  * Use with care! Not exported.
  93  */
  94 static void cl_page_get_trust(struct cl_page *page)
  95 {
  96         LASSERT(atomic_read(&page->cp_ref) > 0);
  97         atomic_inc(&page->cp_ref);
  98 }
  99
 100 /**
 101  * Returns a slice within a page, corresponding to the given layer in the
 102  * device stack.
 103  *
 104  * \see cl_lock_at()
 105  */
 106 static const struct cl_page_slice *
 107 cl_page_at_trusted(const struct cl_page *page,
 108                    const struct lu_device_type *dtype)
 109 {
 110         const struct cl_page_slice *slice;
 111
 112         page = cl_page_top_trusted((struct cl_page *)page);
 113         do {
 114                 list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 115                         if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
 116                                 return slice;
 117                 }
 118                 page = page->cp_child;
 119         } while (page != NULL);
 120         return NULL;
 121 }
 122
 123 /**
 124  * Returns a page with given index in the given object, or NULL if no page is
 125  * found. Acquires a reference on \a page.
 126  *
 127  * Locking: called under cl_object_header::coh_page_guard spin-lock.
 128  */
 129 struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index)
 130 {
 131         struct cl_page *page;
 132
 133         assert_spin_locked(&hdr->coh_page_guard);
 134
 135         page = radix_tree_lookup(&hdr->coh_tree, index);
 136         if (page != NULL)
 137                 cl_page_get_trust(page);
 138         return page;
 139 }
 140 EXPORT_SYMBOL(cl_page_lookup);
 141
 142 /**
 143  * Returns a list of pages by a given [start, end] of \a obj.
 144  *
 145  * \param resched If not NULL, then we give up before hogging CPU for too
 146  * long and set *resched = 1, in that case caller should implement a retry
 147  * logic.
 148  *
 149  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
 150  * crucial in the face of [offset, EOF] locks.
 151  *
 152  * Return at least one page in @queue unless there is no covered page.
 153  */
 154 int cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj,
 155                         struct cl_io *io, pgoff_t start, pgoff_t end,
 156                         cl_page_gang_cb_t cb, void *cbdata)
 157 {
 158         struct cl_object_header *hdr;
 159         struct cl_page    *page;
 160         struct cl_page   **pvec;
 161         const struct cl_page_slice  *slice;
 162         const struct lu_device_type *dtype;
 163         pgoff_t           idx;
 164         unsigned int         nr;
 165         unsigned int         i;
 166         unsigned int         j;
 167         int                   res = CLP_GANG_OKAY;
 168         int                   tree_lock = 1;
 169
 170         idx = start;
 171         hdr = cl_object_header(obj);
 172         pvec = cl_env_info(env)->clt_pvec;
 173         dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type;
 174         spin_lock(&hdr->coh_page_guard);
 175         while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec,
 176                                             idx, CLT_PVEC_SIZE)) > 0) {
 177                 int end_of_region = 0;
 178                 idx = pvec[nr - 1]->cp_index + 1;
 179                 for (i = 0, j = 0; i < nr; ++i) {
 180                         page = pvec[i];
 181                         pvec[i] = NULL;
 182
 183                         LASSERT(page->cp_type == CPT_CACHEABLE);
 184                         if (page->cp_index > end) {
 185                                 end_of_region = 1;
 186                                 break;
 187                         }
 188                         if (page->cp_state == CPS_FREEING)
 189                                 continue;
 190
 191                         slice = cl_page_at_trusted(page, dtype);
 192                         /*
 193                          * Pages for lsm-less file has no underneath sub-page
 194                          * for osc, in case of ...
 195                          */
 196                         PASSERT(env, page, slice != NULL);
 197
 198                         page = slice->cpl_page;
 199                         /*
 200                          * Can safely call cl_page_get_trust() under
 201                          * radix-tree spin-lock.
 202                          *
 203                          * XXX not true, because @page is from object another
 204                          * than @hdr and protected by different tree lock.
 205                          */
 206                         cl_page_get_trust(page);
 207                         lu_ref_add_atomic(&page->cp_reference,
 208                                           "gang_lookup", current);
 209                         pvec[j++] = page;
 210                 }
 211
 212                 /*
 213                  * Here a delicate locking dance is performed. Current thread
 214                  * holds a reference to a page, but has to own it before it
 215                  * can be placed into queue. Owning implies waiting, so
 216                  * radix-tree lock is to be released. After a wait one has to
 217                  * check that pages weren't truncated (cl_page_own() returns
 218                  * error in the latter case).
 219                  */
 220                 spin_unlock(&hdr->coh_page_guard);
 221                 tree_lock = 0;
 222
 223                 for (i = 0; i < j; ++i) {
 224                         page = pvec[i];
 225                         if (res == CLP_GANG_OKAY)
 226                                 res = (*cb)(env, io, page, cbdata);
 227                         lu_ref_del(&page->cp_reference,
 228                                    "gang_lookup", current);
 229                         cl_page_put(env, page);
 230                 }
 231                 if (nr < CLT_PVEC_SIZE || end_of_region)
 232                         break;
 233
 234                 if (res == CLP_GANG_OKAY && need_resched())
 235                         res = CLP_GANG_RESCHED;
 236                 if (res != CLP_GANG_OKAY)
 237                         break;
 238
 239                 spin_lock(&hdr->coh_page_guard);
 240                 tree_lock = 1;
 241         }
 242         if (tree_lock)
 243                 spin_unlock(&hdr->coh_page_guard);
 244         return res;
 245 }
 246 EXPORT_SYMBOL(cl_page_gang_lookup);
 247
 248 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
 249 {
 250         struct cl_object *obj  = page->cp_obj;
 251         int pagesize = cl_object_header(obj)->coh_page_bufsize;
 252
 253         PASSERT(env, page, list_empty(&page->cp_batch));
 254         PASSERT(env, page, page->cp_owner == NULL);
 255         PASSERT(env, page, page->cp_req == NULL);
 256         PASSERT(env, page, page->cp_parent == NULL);
 257         PASSERT(env, page, page->cp_state == CPS_FREEING);
 258
 259         might_sleep();
 260         while (!list_empty(&page->cp_layers)) {
 261                 struct cl_page_slice *slice;
 262
 263                 slice = list_entry(page->cp_layers.next,
 264                                        struct cl_page_slice, cpl_linkage);
 265                 list_del_init(page->cp_layers.next);
 266                 slice->cpl_ops->cpo_fini(env, slice);
 267         }
 268         CS_PAGE_DEC(obj, total);
 269         CS_PAGESTATE_DEC(obj, page->cp_state);
 270         lu_object_ref_del_at(&obj->co_lu, &page->cp_obj_ref, "cl_page", page);
 271         cl_object_put(env, obj);
 272         lu_ref_fini(&page->cp_reference);
 273         OBD_FREE(page, pagesize);
 274 }
 275
 276 /**
 277  * Helper function updating page state. This is the only place in the code
 278  * where cl_page::cp_state field is mutated.
 279  */
 280 static inline void cl_page_state_set_trust(struct cl_page *page,
 281                                            enum cl_page_state state)
 282 {
 283         /* bypass const. */
 284         *(enum cl_page_state *)&page->cp_state = state;
 285 }
 286
 287 static struct cl_page *cl_page_alloc(const struct lu_env *env,
 288                 struct cl_object *o, pgoff_t ind, struct page *vmpage,
 289                 enum cl_page_type type)
 290 {
 291         struct cl_page    *page;
 292         struct lu_object_header *head;
 293
 294         OBD_ALLOC_GFP(page, cl_object_header(o)->coh_page_bufsize,
 295                         GFP_NOFS);
 296         if (page != NULL) {
 297                 int result = 0;
 298                 atomic_set(&page->cp_ref, 1);
 299                 if (type == CPT_CACHEABLE) /* for radix tree */
 300                         atomic_inc(&page->cp_ref);
 301                 page->cp_obj = o;
 302                 cl_object_get(o);
 303                 lu_object_ref_add_at(&o->co_lu, &page->cp_obj_ref, "cl_page",
 304                                      page);
 305                 page->cp_index = ind;
 306                 cl_page_state_set_trust(page, CPS_CACHED);
 307                 page->cp_type = type;
 308                 INIT_LIST_HEAD(&page->cp_layers);
 309                 INIT_LIST_HEAD(&page->cp_batch);
 310                 INIT_LIST_HEAD(&page->cp_flight);
 311                 mutex_init(&page->cp_mutex);
 312                 lu_ref_init(&page->cp_reference);
 313                 head = o->co_lu.lo_header;
 314                 list_for_each_entry(o, &head->loh_layers,
 315                                         co_lu.lo_linkage) {
 316                         if (o->co_ops->coo_page_init != NULL) {
 317                                 result = o->co_ops->coo_page_init(env, o,
 318                                                                   page, vmpage);
 319                                 if (result != 0) {
 320                                         cl_page_delete0(env, page, 0);
 321                                         cl_page_free(env, page);
 322                                         page = ERR_PTR(result);
 323                                         break;
 324                                 }
 325                         }
 326                 }
 327                 if (result == 0) {
 328                         CS_PAGE_INC(o, total);
 329                         CS_PAGE_INC(o, create);
 330                         CS_PAGESTATE_DEC(o, CPS_CACHED);
 331                 }
 332         } else {
 333                 page = ERR_PTR(-ENOMEM);
 334         }
 335         return page;
 336 }
 337
 338 /**
 339  * Returns a cl_page with index \a idx at the object \a o, and associated with
 340  * the VM page \a vmpage.
 341  *
 342  * This is the main entry point into the cl_page caching interface. First, a
 343  * cache (implemented as a per-object radix tree) is consulted. If page is
 344  * found there, it is returned immediately. Otherwise new page is allocated
 345  * and returned. In any case, additional reference to page is acquired.
 346  *
 347  * \see cl_object_find(), cl_lock_find()
 348  */
 349 static struct cl_page *cl_page_find0(const struct lu_env *env,
 350                                      struct cl_object *o,
 351                                      pgoff_t idx, struct page *vmpage,
 352                                      enum cl_page_type type,
 353                                      struct cl_page *parent)
 354 {
 355         struct cl_page    *page = NULL;
 356         struct cl_page    *ghost = NULL;
 357         struct cl_object_header *hdr;
 358         int err;
 359
 360         LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 361         might_sleep();
 362
 363         hdr = cl_object_header(o);
 364         CS_PAGE_INC(o, lookup);
 365
 366         CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
 367                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 368         /* fast path. */
 369         if (type == CPT_CACHEABLE) {
 370                 /* vmpage lock is used to protect the child/parent
 371                  * relationship */
 372                 KLASSERT(PageLocked(vmpage));
 373                 /*
 374                  * cl_vmpage_page() can be called here without any locks as
 375                  *
 376                  *     - "vmpage" is locked (which prevents ->private from
 377                  *       concurrent updates), and
 378                  *
 379                  *     - "o" cannot be destroyed while current thread holds a
 380                  *       reference on it.
 381                  */
 382                 page = cl_vmpage_page(vmpage, o);
 383                 PINVRNT(env, page,
 384                         ergo(page != NULL,
 385                              cl_page_vmpage(env, page) == vmpage &&
 386                              (void *)radix_tree_lookup(&hdr->coh_tree,
 387                                                        idx) == page));
 388         }
 389
 390         if (page != NULL) {
 391                 CS_PAGE_INC(o, hit);
 392                 return page;
 393         }
 394
 395         /* allocate and initialize cl_page */
 396         page = cl_page_alloc(env, o, idx, vmpage, type);
 397         if (IS_ERR(page))
 398                 return page;
 399
 400         if (type == CPT_TRANSIENT) {
 401                 if (parent) {
 402                         LASSERT(page->cp_parent == NULL);
 403                         page->cp_parent = parent;
 404                         parent->cp_child = page;
 405                 }
 406                 return page;
 407         }
 408
 409         /*
 410          * XXX optimization: use radix_tree_preload() here, and change tree
 411          * gfp mask to GFP_KERNEL in cl_object_header_init().
 412          */
 413         spin_lock(&hdr->coh_page_guard);
 414         err = radix_tree_insert(&hdr->coh_tree, idx, page);
 415         if (err != 0) {
 416                 ghost = page;
 417                 /*
 418                  * Noted by Jay: a lock on \a vmpage protects cl_page_find()
 419                  * from this race, but
 420                  *
 421                  *     0. it's better to have cl_page interface "locally
 422                  *     consistent" so that its correctness can be reasoned
 423                  *     about without appealing to the (obscure world of) VM
 424                  *     locking.
 425                  *
 426                  *     1. handling this race allows ->coh_tree to remain
 427                  *     consistent even when VM locking is somehow busted,
 428                  *     which is very useful during diagnosing and debugging.
 429                  */
 430                 page = ERR_PTR(err);
 431                 CL_PAGE_DEBUG(D_ERROR, env, ghost,
 432                               "fail to insert into radix tree: %d\n", err);
 433         } else {
 434                 if (parent) {
 435                         LASSERT(page->cp_parent == NULL);
 436                         page->cp_parent = parent;
 437                         parent->cp_child = page;
 438                 }
 439                 hdr->coh_pages++;
 440         }
 441         spin_unlock(&hdr->coh_page_guard);
 442
 443         if (unlikely(ghost != NULL)) {
 444                 cl_page_delete0(env, ghost, 0);
 445                 cl_page_free(env, ghost);
 446         }
 447         return page;
 448 }
 449
 450 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o,
 451                              pgoff_t idx, struct page *vmpage,
 452                              enum cl_page_type type)
 453 {
 454         return cl_page_find0(env, o, idx, vmpage, type, NULL);
 455 }
 456 EXPORT_SYMBOL(cl_page_find);
 457
 458
 459 struct cl_page *cl_page_find_sub(const struct lu_env *env, struct cl_object *o,
 460                                  pgoff_t idx, struct page *vmpage,
 461                                  struct cl_page *parent)
 462 {
 463         return cl_page_find0(env, o, idx, vmpage, parent->cp_type, parent);
 464 }
 465 EXPORT_SYMBOL(cl_page_find_sub);
 466
 467 static inline int cl_page_invariant(const struct cl_page *pg)
 468 {
 469         struct cl_object_header *header;
 470         struct cl_page    *parent;
 471         struct cl_page    *child;
 472         struct cl_io        *owner;
 473
 474         /*
 475          * Page invariant is protected by a VM lock.
 476          */
 477         LINVRNT(cl_page_is_vmlocked(NULL, pg));
 478
 479         header = cl_object_header(pg->cp_obj);
 480         parent = pg->cp_parent;
 481         child  = pg->cp_child;
 482         owner  = pg->cp_owner;
 483
 484         return cl_page_in_use(pg) &&
 485                 ergo(parent != NULL, parent->cp_child == pg) &&
 486                 ergo(child != NULL, child->cp_parent == pg) &&
 487                 ergo(child != NULL, pg->cp_obj != child->cp_obj) &&
 488                 ergo(parent != NULL, pg->cp_obj != parent->cp_obj) &&
 489                 ergo(owner != NULL && parent != NULL,
 490                      parent->cp_owner == pg->cp_owner->ci_parent) &&
 491                 ergo(owner != NULL && child != NULL,
 492                      child->cp_owner->ci_parent == owner) &&
 493                 /*
 494                  * Either page is early in initialization (has neither child
 495                  * nor parent yet), or it is in the object radix tree.
 496                  */
 497                 ergo(pg->cp_state < CPS_FREEING && pg->cp_type == CPT_CACHEABLE,
 498                      (void *)radix_tree_lookup(&header->coh_tree,
 499                                                pg->cp_index) == pg ||
 500                      (child == NULL && parent == NULL));
 501 }
 502
 503 static void cl_page_state_set0(const struct lu_env *env,
 504                                struct cl_page *page, enum cl_page_state state)
 505 {
 506         enum cl_page_state old;
 507
 508         /*
 509          * Matrix of allowed state transitions [old][new], for sanity
 510          * checking.
 511          */
 512         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 513                 [CPS_CACHED] = {
 514                         [CPS_CACHED]  = 0,
 515                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 516                         [CPS_PAGEIN]  = 0,
 517                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 518                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 519                 },
 520                 [CPS_OWNED] = {
 521                         [CPS_CACHED]  = 1, /* release to the cache */
 522                         [CPS_OWNED]   = 0,
 523                         [CPS_PAGEIN]  = 1, /* start read immediately */
 524                         [CPS_PAGEOUT] = 1, /* start write immediately */
 525                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 526                 },
 527                 [CPS_PAGEIN] = {
 528                         [CPS_CACHED]  = 1, /* io completion */
 529                         [CPS_OWNED]   = 0,
 530                         [CPS_PAGEIN]  = 0,
 531                         [CPS_PAGEOUT] = 0,
 532                         [CPS_FREEING] = 0,
 533                 },
 534                 [CPS_PAGEOUT] = {
 535                         [CPS_CACHED]  = 1, /* io completion */
 536                         [CPS_OWNED]   = 0,
 537                         [CPS_PAGEIN]  = 0,
 538                         [CPS_PAGEOUT] = 0,
 539                         [CPS_FREEING] = 0,
 540                 },
 541                 [CPS_FREEING] = {
 542                         [CPS_CACHED]  = 0,
 543                         [CPS_OWNED]   = 0,
 544                         [CPS_PAGEIN]  = 0,
 545                         [CPS_PAGEOUT] = 0,
 546                         [CPS_FREEING] = 0,
 547                 }
 548         };
 549
 550         old = page->cp_state;
 551         PASSERT(env, page, allowed_transitions[old][state]);
 552         CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
 553         for (; page != NULL; page = page->cp_child) {
 554                 PASSERT(env, page, page->cp_state == old);
 555                 PASSERT(env, page,
 556                         equi(state == CPS_OWNED, page->cp_owner != NULL));
 557
 558                 CS_PAGESTATE_DEC(page->cp_obj, page->cp_state);
 559                 CS_PAGESTATE_INC(page->cp_obj, state);
 560                 cl_page_state_set_trust(page, state);
 561         }
 562 }
 563
 564 static void cl_page_state_set(const struct lu_env *env,
 565                               struct cl_page *page, enum cl_page_state state)
 566 {
 567         cl_page_state_set0(env, page, state);
 568 }
 569
 570 /**
 571  * Acquires an additional reference to a page.
 572  *
 573  * This can be called only by caller already possessing a reference to \a
 574  * page.
 575  *
 576  * \see cl_object_get(), cl_lock_get().
 577  */
 578 void cl_page_get(struct cl_page *page)
 579 {
 580         cl_page_get_trust(page);
 581 }
 582 EXPORT_SYMBOL(cl_page_get);
 583
 584 /**
 585  * Releases a reference to a page.
 586  *
 587  * When last reference is released, page is returned to the cache, unless it
 588  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 589  * destroyed.
 590  *
 591  * \see cl_object_put(), cl_lock_put().
 592  */
 593 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 594 {
 595         PASSERT(env, page, atomic_read(&page->cp_ref) > !!page->cp_parent);
 596
 597         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 598                        atomic_read(&page->cp_ref));
 599
 600         if (atomic_dec_and_test(&page->cp_ref)) {
 601                 LASSERT(page->cp_state == CPS_FREEING);
 602
 603                 LASSERT(atomic_read(&page->cp_ref) == 0);
 604                 PASSERT(env, page, page->cp_owner == NULL);
 605                 PASSERT(env, page, list_empty(&page->cp_batch));
 606                 /*
 607                  * Page is no longer reachable by other threads. Tear
 608                  * it down.
 609                  */
 610                 cl_page_free(env, page);
 611         }
 612 }
 613 EXPORT_SYMBOL(cl_page_put);
 614
 615 /**
 616  * Returns a VM page associated with a given cl_page.
 617  */
 618 struct page *cl_page_vmpage(const struct lu_env *env, struct cl_page *page)
 619 {
 620         const struct cl_page_slice *slice;
 621
 622         /*
 623          * Find uppermost layer with ->cpo_vmpage() method, and return its
 624          * result.
 625          */
 626         page = cl_page_top(page);
 627         do {
 628                 list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 629                         if (slice->cpl_ops->cpo_vmpage != NULL)
 630                                 return slice->cpl_ops->cpo_vmpage(env, slice);
 631                 }
 632                 page = page->cp_child;
 633         } while (page != NULL);
 634         LBUG(); /* ->cpo_vmpage() has to be defined somewhere in the stack */
 635 }
 636 EXPORT_SYMBOL(cl_page_vmpage);
 637
 638 /**
 639  * Returns a cl_page associated with a VM page, and given cl_object.
 640  */
 641 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj)
 642 {
 643         struct cl_page *top;
 644         struct cl_page *page;
 645
 646         KLASSERT(PageLocked(vmpage));
 647
 648         /*
 649          * NOTE: absence of races and liveness of data are guaranteed by page
 650          *       lock on a "vmpage". That works because object destruction has
 651          *       bottom-to-top pass.
 652          */
 653
 654         /*
 655          * This loop assumes that ->private points to the top-most page. This
 656          * can be rectified easily.
 657          */
 658         top = (struct cl_page *)vmpage->private;
 659         if (top == NULL)
 660                 return NULL;
 661
 662         for (page = top; page != NULL; page = page->cp_child) {
 663                 if (cl_object_same(page->cp_obj, obj)) {
 664                         cl_page_get_trust(page);
 665                         break;
 666                 }
 667         }
 668         LASSERT(ergo(page, page->cp_type == CPT_CACHEABLE));
 669         return page;
 670 }
 671 EXPORT_SYMBOL(cl_vmpage_page);
 672
 673 /**
 674  * Returns the top-page for a given page.
 675  *
 676  * \see cl_object_top(), cl_io_top()
 677  */
 678 struct cl_page *cl_page_top(struct cl_page *page)
 679 {
 680         return cl_page_top_trusted(page);
 681 }
 682 EXPORT_SYMBOL(cl_page_top);
 683
 684 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 685                                        const struct lu_device_type *dtype)
 686 {
 687         return cl_page_at_trusted(page, dtype);
 688 }
 689 EXPORT_SYMBOL(cl_page_at);
 690
 691 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 692
 693 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)              \
 694 ({                                                                    \
 695         const struct lu_env     *__env  = (_env);                   \
 696         struct cl_page       *__page = (_page);            \
 697         const struct cl_page_slice *__scan;                          \
 698         int                      __result;                         \
 699         ptrdiff_t                  __op   = (_op);                   \
 700         int                    (*__method)_proto;                   \
 701                                                                         \
 702         __result = 0;                                              \
 703         __page = cl_page_top(__page);                              \
 704         do {                                                        \
 705                 list_for_each_entry(__scan, &__page->cp_layers,     \
 706                                         cpl_linkage) {            \
 707                         __method = *(void **)((char *)__scan->cpl_ops + \
 708                                               __op);                \
 709                         if (__method != NULL) {                  \
 710                                 __result = (*__method)(__env, __scan,   \
 711                                                        ## __VA_ARGS__); \
 712                                 if (__result != 0)                    \
 713                                         break;                    \
 714                         }                                              \
 715                 }                                                      \
 716                 __page = __page->cp_child;                            \
 717         } while (__page != NULL && __result == 0);                    \
 718         if (__result > 0)                                              \
 719                 __result = 0;                                      \
 720         __result;                                                      \
 721 })
 722
 723 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)              \
 724 do {                                                                \
 725         const struct lu_env     *__env  = (_env);                   \
 726         struct cl_page       *__page = (_page);            \
 727         const struct cl_page_slice *__scan;                          \
 728         ptrdiff_t                  __op   = (_op);                   \
 729         void                  (*__method)_proto;                    \
 730                                                                         \
 731         __page = cl_page_top(__page);                              \
 732         do {                                                        \
 733                 list_for_each_entry(__scan, &__page->cp_layers,     \
 734                                         cpl_linkage) {            \
 735                         __method = *(void **)((char *)__scan->cpl_ops + \
 736                                               __op);                \
 737                         if (__method != NULL)                      \
 738                                 (*__method)(__env, __scan,            \
 739                                             ## __VA_ARGS__);        \
 740                 }                                                      \
 741                 __page = __page->cp_child;                            \
 742         } while (__page != NULL);                                      \
 743 } while (0)
 744
 745 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)          \
 746 do {                                                                    \
 747         const struct lu_env     *__env  = (_env);                       \
 748         struct cl_page       *__page = (_page);                \
 749         const struct cl_page_slice *__scan;                              \
 750         ptrdiff_t                  __op   = (_op);                       \
 751         void                  (*__method)_proto;                        \
 752                                                                             \
 753         /* get to the bottom page. */                                  \
 754         while (__page->cp_child != NULL)                                    \
 755                 __page = __page->cp_child;                                \
 756         do {                                                            \
 757                 list_for_each_entry_reverse(__scan, &__page->cp_layers, \
 758                                                 cpl_linkage) {        \
 759                         __method = *(void **)((char *)__scan->cpl_ops +     \
 760                                               __op);                    \
 761                         if (__method != NULL)                          \
 762                                 (*__method)(__env, __scan,                \
 763                                             ## __VA_ARGS__);            \
 764                 }                                                          \
 765                 __page = __page->cp_parent;                              \
 766         } while (__page != NULL);                                          \
 767 } while (0)
 768
 769 static int cl_page_invoke(const struct lu_env *env,
 770                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 771
 772 {
 773         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 774         return CL_PAGE_INVOKE(env, page, op,
 775                               (const struct lu_env *,
 776                                const struct cl_page_slice *, struct cl_io *),
 777                               io);
 778 }
 779
 780 static void cl_page_invoid(const struct lu_env *env,
 781                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 782
 783 {
 784         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 785         CL_PAGE_INVOID(env, page, op,
 786                        (const struct lu_env *,
 787                         const struct cl_page_slice *, struct cl_io *), io);
 788 }
 789
 790 static void cl_page_owner_clear(struct cl_page *page)
 791 {
 792         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 793                 if (page->cp_owner != NULL) {
 794                         LASSERT(page->cp_owner->ci_owned_nr > 0);
 795                         page->cp_owner->ci_owned_nr--;
 796                         page->cp_owner = NULL;
 797                         page->cp_task = NULL;
 798                 }
 799         }
 800 }
 801
 802 static void cl_page_owner_set(struct cl_page *page)
 803 {
 804         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 805                 LASSERT(page->cp_owner != NULL);
 806                 page->cp_owner->ci_owned_nr++;
 807         }
 808 }
 809
 810 void cl_page_disown0(const struct lu_env *env,
 811                      struct cl_io *io, struct cl_page *pg)
 812 {
 813         enum cl_page_state state;
 814
 815         state = pg->cp_state;
 816         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 817         PINVRNT(env, pg, cl_page_invariant(pg));
 818         cl_page_owner_clear(pg);
 819
 820         if (state == CPS_OWNED)
 821                 cl_page_state_set(env, pg, CPS_CACHED);
 822         /*
 823          * Completion call-backs are executed in the bottom-up order, so that
 824          * uppermost layer (llite), responsible for VFS/VM interaction runs
 825          * last and can release locks safely.
 826          */
 827         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 828                                (const struct lu_env *,
 829                                 const struct cl_page_slice *, struct cl_io *),
 830                                io);
 831 }
 832
 833 /**
 834  * returns true, iff page is owned by the given io.
 835  */
 836 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 837 {
 838         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 839         return pg->cp_state == CPS_OWNED && pg->cp_owner == io;
 840 }
 841 EXPORT_SYMBOL(cl_page_is_owned);
 842
 843 /**
 844  * Try to own a page by IO.
 845  *
 846  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 847  * into cl_page_state::CPS_OWNED state.
 848  *
 849  * \pre  !cl_page_is_owned(pg, io)
 850  * \post result == 0 iff cl_page_is_owned(pg, io)
 851  *
 852  * \retval 0   success
 853  *
 854  * \retval -ve failure, e.g., page was destroyed (and landed in
 855  *           cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 856  *           or, page was owned by another thread, or in IO.
 857  *
 858  * \see cl_page_disown()
 859  * \see cl_page_operations::cpo_own()
 860  * \see cl_page_own_try()
 861  * \see cl_page_own
 862  */
 863 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 864                         struct cl_page *pg, int nonblock)
 865 {
 866         int result;
 867
 868         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 869
 870         pg = cl_page_top(pg);
 871         io = cl_io_top(io);
 872
 873         if (pg->cp_state == CPS_FREEING) {
 874                 result = -ENOENT;
 875         } else {
 876                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 877                                         (const struct lu_env *,
 878                                          const struct cl_page_slice *,
 879                                          struct cl_io *, int),
 880                                         io, nonblock);
 881                 if (result == 0) {
 882                         PASSERT(env, pg, pg->cp_owner == NULL);
 883                         PASSERT(env, pg, pg->cp_req == NULL);
 884                         pg->cp_owner = io;
 885                         pg->cp_task  = current;
 886                         cl_page_owner_set(pg);
 887                         if (pg->cp_state != CPS_FREEING) {
 888                                 cl_page_state_set(env, pg, CPS_OWNED);
 889                         } else {
 890                                 cl_page_disown0(env, io, pg);
 891                                 result = -ENOENT;
 892                         }
 893                 }
 894         }
 895         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 896         return result;
 897 }
 898
 899 /**
 900  * Own a page, might be blocked.
 901  *
 902  * \see cl_page_own0()
 903  */
 904 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 905 {
 906         return cl_page_own0(env, io, pg, 0);
 907 }
 908 EXPORT_SYMBOL(cl_page_own);
 909
 910 /**
 911  * Nonblock version of cl_page_own().
 912  *
 913  * \see cl_page_own0()
 914  */
 915 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
 916                     struct cl_page *pg)
 917 {
 918         return cl_page_own0(env, io, pg, 1);
 919 }
 920 EXPORT_SYMBOL(cl_page_own_try);
 921
 922
 923 /**
 924  * Assume page ownership.
 925  *
 926  * Called when page is already locked by the hosting VM.
 927  *
 928  * \pre !cl_page_is_owned(pg, io)
 929  * \post cl_page_is_owned(pg, io)
 930  *
 931  * \see cl_page_operations::cpo_assume()
 932  */
 933 void cl_page_assume(const struct lu_env *env,
 934                     struct cl_io *io, struct cl_page *pg)
 935 {
 936         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
 937
 938         pg = cl_page_top(pg);
 939         io = cl_io_top(io);
 940
 941         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
 942         PASSERT(env, pg, pg->cp_owner == NULL);
 943         pg->cp_owner = io;
 944         pg->cp_task = current;
 945         cl_page_owner_set(pg);
 946         cl_page_state_set(env, pg, CPS_OWNED);
 947 }
 948 EXPORT_SYMBOL(cl_page_assume);
 949
 950 /**
 951  * Releases page ownership without unlocking the page.
 952  *
 953  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
 954  * underlying VM page (as VM is supposed to do this itself).
 955  *
 956  * \pre   cl_page_is_owned(pg, io)
 957  * \post !cl_page_is_owned(pg, io)
 958  *
 959  * \see cl_page_assume()
 960  */
 961 void cl_page_unassume(const struct lu_env *env,
 962                       struct cl_io *io, struct cl_page *pg)
 963 {
 964         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 965         PINVRNT(env, pg, cl_page_invariant(pg));
 966
 967         pg = cl_page_top(pg);
 968         io = cl_io_top(io);
 969         cl_page_owner_clear(pg);
 970         cl_page_state_set(env, pg, CPS_CACHED);
 971         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
 972                                (const struct lu_env *,
 973                                 const struct cl_page_slice *, struct cl_io *),
 974                                io);
 975 }
 976 EXPORT_SYMBOL(cl_page_unassume);
 977
 978 /**
 979  * Releases page ownership.
 980  *
 981  * Moves page into cl_page_state::CPS_CACHED.
 982  *
 983  * \pre   cl_page_is_owned(pg, io)
 984  * \post !cl_page_is_owned(pg, io)
 985  *
 986  * \see cl_page_own()
 987  * \see cl_page_operations::cpo_disown()
 988  */
 989 void cl_page_disown(const struct lu_env *env,
 990                     struct cl_io *io, struct cl_page *pg)
 991 {
 992         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 993
 994         pg = cl_page_top(pg);
 995         io = cl_io_top(io);
 996         cl_page_disown0(env, io, pg);
 997 }
 998 EXPORT_SYMBOL(cl_page_disown);
 999
1000 /**
1001  * Called when page is to be removed from the object, e.g., as a result of
1002  * truncate.
1003  *
1004  * Calls cl_page_operations::cpo_discard() top-to-bottom.
1005  *
1006  * \pre cl_page_is_owned(pg, io)
1007  *
1008  * \see cl_page_operations::cpo_discard()
1009  */
1010 void cl_page_discard(const struct lu_env *env,
1011                      struct cl_io *io, struct cl_page *pg)
1012 {
1013         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1014         PINVRNT(env, pg, cl_page_invariant(pg));
1015
1016         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
1017 }
1018 EXPORT_SYMBOL(cl_page_discard);
1019
1020 /**
1021  * Version of cl_page_delete() that can be called for not fully constructed
1022  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
1023  * path. Doesn't check page invariant.
1024  */
1025 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
1026                             int radix)
1027 {
1028         struct cl_page *tmp = pg;
1029
1030         PASSERT(env, pg, pg == cl_page_top(pg));
1031         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
1032
1033         /*
1034          * Severe all ways to obtain new pointers to @pg.
1035          */
1036         cl_page_owner_clear(pg);
1037
1038         /*
1039          * unexport the page firstly before freeing it so that
1040          * the page content is considered to be invalid.
1041          * We have to do this because a CPS_FREEING cl_page may
1042          * be NOT under the protection of a cl_lock.
1043          * Afterwards, if this page is found by other threads, then this
1044          * page will be forced to reread.
1045          */
1046         cl_page_export(env, pg, 0);
1047         cl_page_state_set0(env, pg, CPS_FREEING);
1048
1049         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete),
1050                        (const struct lu_env *, const struct cl_page_slice *));
1051
1052         if (tmp->cp_type == CPT_CACHEABLE) {
1053                 if (!radix)
1054                         /* !radix means that @pg is not yet in the radix tree,
1055                          * skip removing it.
1056                          */
1057                         tmp = pg->cp_child;
1058                 for (; tmp != NULL; tmp = tmp->cp_child) {
1059                         void                *value;
1060                         struct cl_object_header *hdr;
1061
1062                         hdr = cl_object_header(tmp->cp_obj);
1063                         spin_lock(&hdr->coh_page_guard);
1064                         value = radix_tree_delete(&hdr->coh_tree,
1065                                                   tmp->cp_index);
1066                         PASSERT(env, tmp, value == tmp);
1067                         PASSERT(env, tmp, hdr->coh_pages > 0);
1068                         hdr->coh_pages--;
1069                         spin_unlock(&hdr->coh_page_guard);
1070                         cl_page_put(env, tmp);
1071                 }
1072         }
1073 }
1074
1075 /**
1076  * Called when a decision is made to throw page out of memory.
1077  *
1078  * Notifies all layers about page destruction by calling
1079  * cl_page_operations::cpo_delete() method top-to-bottom.
1080  *
1081  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
1082  * where transition to this state happens).
1083  *
1084  * Eliminates all venues through which new references to the page can be
1085  * obtained:
1086  *
1087  *     - removes page from the radix trees,
1088  *
1089  *     - breaks linkage from VM page to cl_page.
1090  *
1091  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
1092  * drain after some time, at which point page will be recycled.
1093  *
1094  * \pre  pg == cl_page_top(pg)
1095  * \pre  VM page is locked
1096  * \post pg->cp_state == CPS_FREEING
1097  *
1098  * \see cl_page_operations::cpo_delete()
1099  */
1100 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
1101 {
1102         PINVRNT(env, pg, cl_page_invariant(pg));
1103         cl_page_delete0(env, pg, 1);
1104 }
1105 EXPORT_SYMBOL(cl_page_delete);
1106
1107 /**
1108  * Unmaps page from user virtual memory.
1109  *
1110  * Calls cl_page_operations::cpo_unmap() through all layers top-to-bottom. The
1111  * layer responsible for VM interaction has to unmap page from user space
1112  * virtual memory.
1113  *
1114  * \see cl_page_operations::cpo_unmap()
1115  */
1116 int cl_page_unmap(const struct lu_env *env,
1117                   struct cl_io *io, struct cl_page *pg)
1118 {
1119         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1120         PINVRNT(env, pg, cl_page_invariant(pg));
1121
1122         return cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_unmap));
1123 }
1124 EXPORT_SYMBOL(cl_page_unmap);
1125
1126 /**
1127  * Marks page up-to-date.
1128  *
1129  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
1130  * layer responsible for VM interaction has to mark/clear page as up-to-date
1131  * by the \a uptodate argument.
1132  *
1133  * \see cl_page_operations::cpo_export()
1134  */
1135 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
1136 {
1137         PINVRNT(env, pg, cl_page_invariant(pg));
1138         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
1139                        (const struct lu_env *,
1140                         const struct cl_page_slice *, int), uptodate);
1141 }
1142 EXPORT_SYMBOL(cl_page_export);
1143
1144 /**
1145  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
1146  * thread.
1147  */
1148 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
1149 {
1150         int result;
1151         const struct cl_page_slice *slice;
1152
1153         pg = cl_page_top_trusted((struct cl_page *)pg);
1154         slice = container_of(pg->cp_layers.next,
1155                              const struct cl_page_slice, cpl_linkage);
1156         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
1157         /*
1158          * Call ->cpo_is_vmlocked() directly instead of going through
1159          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
1160          * cl_page_invariant().
1161          */
1162         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
1163         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
1164         return result == -EBUSY;
1165 }
1166 EXPORT_SYMBOL(cl_page_is_vmlocked);
1167
1168 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
1169 {
1170         return crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN;
1171 }
1172
1173 static void cl_page_io_start(const struct lu_env *env,
1174                              struct cl_page *pg, enum cl_req_type crt)
1175 {
1176         /*
1177          * Page is queued for IO, change its state.
1178          */
1179         cl_page_owner_clear(pg);
1180         cl_page_state_set(env, pg, cl_req_type_state(crt));
1181 }
1182
1183 /**
1184  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
1185  * called top-to-bottom. Every layer either agrees to submit this page (by
1186  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
1187  * handling interactions with the VM also has to inform VM that page is under
1188  * transfer now.
1189  */
1190 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
1191                  struct cl_page *pg, enum cl_req_type crt)
1192 {
1193         int result;
1194
1195         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1196         PINVRNT(env, pg, cl_page_invariant(pg));
1197         PINVRNT(env, pg, crt < CRT_NR);
1198
1199         /*
1200          * XXX this has to be called bottom-to-top, so that llite can set up
1201          * PG_writeback without risking other layers deciding to skip this
1202          * page.
1203          */
1204         if (crt >= CRT_NR)
1205                 return -EINVAL;
1206         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
1207         if (result == 0)
1208                 cl_page_io_start(env, pg, crt);
1209
1210         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
1211                       equi(result == 0,
1212                            PageWriteback(cl_page_vmpage(env, pg)))));
1213         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1214         return result;
1215 }
1216 EXPORT_SYMBOL(cl_page_prep);
1217
1218 /**
1219  * Notify layers about transfer completion.
1220  *
1221  * Invoked by transfer sub-system (which is a part of osc) to notify layers
1222  * that a transfer, of which this page is a part of has completed.
1223  *
1224  * Completion call-backs are executed in the bottom-up order, so that
1225  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1226  * and can release locks safely.
1227  *
1228  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1229  * \post pg->cp_state == CPS_CACHED
1230  *
1231  * \see cl_page_operations::cpo_completion()
1232  */
1233 void cl_page_completion(const struct lu_env *env,
1234                         struct cl_page *pg, enum cl_req_type crt, int ioret)
1235 {
1236         struct cl_sync_io *anchor = pg->cp_sync_io;
1237
1238         PASSERT(env, pg, crt < CRT_NR);
1239         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
1240         PASSERT(env, pg, pg->cp_req == NULL);
1241         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
1242
1243         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
1244         if (crt == CRT_READ && ioret == 0) {
1245                 PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED));
1246                 pg->cp_flags |= CPF_READ_COMPLETED;
1247         }
1248
1249         cl_page_state_set(env, pg, CPS_CACHED);
1250         if (crt >= CRT_NR)
1251                 return;
1252         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
1253                                (const struct lu_env *,
1254                                 const struct cl_page_slice *, int), ioret);
1255         if (anchor) {
1256                 LASSERT(cl_page_is_vmlocked(env, pg));
1257                 LASSERT(pg->cp_sync_io == anchor);
1258                 pg->cp_sync_io = NULL;
1259         }
1260         /*
1261          * As page->cp_obj is pinned by a reference from page->cp_req, it is
1262          * safe to call cl_page_put() without risking object destruction in a
1263          * non-blocking context.
1264          */
1265         cl_page_put(env, pg);
1266
1267         if (anchor)
1268                 cl_sync_io_note(anchor, ioret);
1269 }
1270 EXPORT_SYMBOL(cl_page_completion);
1271
1272 /**
1273  * Notify layers that transfer formation engine decided to yank this page from
1274  * the cache and to make it a part of a transfer.
1275  *
1276  * \pre  pg->cp_state == CPS_CACHED
1277  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1278  *
1279  * \see cl_page_operations::cpo_make_ready()
1280  */
1281 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
1282                        enum cl_req_type crt)
1283 {
1284         int result;
1285
1286         PINVRNT(env, pg, crt < CRT_NR);
1287
1288         if (crt >= CRT_NR)
1289                 return -EINVAL;
1290         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
1291                                 (const struct lu_env *,
1292                                  const struct cl_page_slice *));
1293         if (result == 0) {
1294                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
1295                 cl_page_io_start(env, pg, crt);
1296         }
1297         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1298         return result;
1299 }
1300 EXPORT_SYMBOL(cl_page_make_ready);
1301
1302 /**
1303  * Notify layers that high level io decided to place this page into a cache
1304  * for future transfer.
1305  *
1306  * The layer implementing transfer engine (osc) has to register this page in
1307  * its queues.
1308  *
1309  * \pre  cl_page_is_owned(pg, io)
1310  * \post cl_page_is_owned(pg, io)
1311  *
1312  * \see cl_page_operations::cpo_cache_add()
1313  */
1314 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
1315                       struct cl_page *pg, enum cl_req_type crt)
1316 {
1317         const struct cl_page_slice *scan;
1318         int result = 0;
1319
1320         PINVRNT(env, pg, crt < CRT_NR);
1321         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1322         PINVRNT(env, pg, cl_page_invariant(pg));
1323
1324         if (crt >= CRT_NR)
1325                 return -EINVAL;
1326
1327         list_for_each_entry(scan, &pg->cp_layers, cpl_linkage) {
1328                 if (scan->cpl_ops->io[crt].cpo_cache_add == NULL)
1329                         continue;
1330
1331                 result = scan->cpl_ops->io[crt].cpo_cache_add(env, scan, io);
1332                 if (result != 0)
1333                         break;
1334         }
1335         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1336         return result;
1337 }
1338 EXPORT_SYMBOL(cl_page_cache_add);
1339
1340 /**
1341  * Called if a pge is being written back by kernel's intention.
1342  *
1343  * \pre  cl_page_is_owned(pg, io)
1344  * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT)
1345  *
1346  * \see cl_page_operations::cpo_flush()
1347  */
1348 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
1349                   struct cl_page *pg)
1350 {
1351         int result;
1352
1353         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1354         PINVRNT(env, pg, cl_page_invariant(pg));
1355
1356         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_flush));
1357
1358         CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result);
1359         return result;
1360 }
1361 EXPORT_SYMBOL(cl_page_flush);
1362
1363 /**
1364  * Checks whether page is protected by any extent lock is at least required
1365  * mode.
1366  *
1367  * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1368  * \see cl_page_operations::cpo_is_under_lock()
1369  */
1370 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1371                           struct cl_page *page)
1372 {
1373         int rc;
1374
1375         PINVRNT(env, page, cl_page_invariant(page));
1376
1377         rc = CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1378                             (const struct lu_env *,
1379                              const struct cl_page_slice *, struct cl_io *),
1380                             io);
1381         PASSERT(env, page, rc != 0);
1382         return rc;
1383 }
1384 EXPORT_SYMBOL(cl_page_is_under_lock);
1385
1386 static int page_prune_cb(const struct lu_env *env, struct cl_io *io,
1387                          struct cl_page *page, void *cbdata)
1388 {
1389         cl_page_own(env, io, page);
1390         cl_page_unmap(env, io, page);
1391         cl_page_discard(env, io, page);
1392         cl_page_disown(env, io, page);
1393         return CLP_GANG_OKAY;
1394 }
1395
1396 /**
1397  * Purges all cached pages belonging to the object \a obj.
1398  */
1399 int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj)
1400 {
1401         struct cl_thread_info   *info;
1402         struct cl_object        *obj = cl_object_top(clobj);
1403         struct cl_io        *io;
1404         int                   result;
1405
1406         info  = cl_env_info(env);
1407         io    = &info->clt_io;
1408
1409         /*
1410          * initialize the io. This is ugly since we never do IO in this
1411          * function, we just make cl_page_list functions happy. -jay
1412          */
1413         io->ci_obj = obj;
1414         io->ci_ignore_layout = 1;
1415         result = cl_io_init(env, io, CIT_MISC, obj);
1416         if (result != 0) {
1417                 cl_io_fini(env, io);
1418                 return io->ci_result;
1419         }
1420
1421         do {
1422                 result = cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF,
1423                                              page_prune_cb, NULL);
1424                 if (result == CLP_GANG_RESCHED)
1425                         cond_resched();
1426         } while (result != CLP_GANG_OKAY);
1427
1428         cl_io_fini(env, io);
1429         return result;
1430 }
1431 EXPORT_SYMBOL(cl_pages_prune);
1432
1433 /**
1434  * Tells transfer engine that only part of a page is to be transmitted.
1435  *
1436  * \see cl_page_operations::cpo_clip()
1437  */
1438 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1439                   int from, int to)
1440 {
1441         PINVRNT(env, pg, cl_page_invariant(pg));
1442
1443         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1444         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1445                        (const struct lu_env *,
1446                         const struct cl_page_slice *,int, int),
1447                        from, to);
1448 }
1449 EXPORT_SYMBOL(cl_page_clip);
1450
1451 /**
1452  * Prints human readable representation of \a pg to the \a f.
1453  */
1454 void cl_page_header_print(const struct lu_env *env, void *cookie,
1455                           lu_printer_t printer, const struct cl_page *pg)
1456 {
1457         (*printer)(env, cookie,
1458                    "page@%p[%d %p:%lu ^%p_%p %d %d %d %p %p %#x]\n",
1459                    pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1460                    pg->cp_index, pg->cp_parent, pg->cp_child,
1461                    pg->cp_state, pg->cp_error, pg->cp_type,
1462                    pg->cp_owner, pg->cp_req, pg->cp_flags);
1463 }
1464 EXPORT_SYMBOL(cl_page_header_print);
1465
1466 /**
1467  * Prints human readable representation of \a pg to the \a f.
1468  */
1469 void cl_page_print(const struct lu_env *env, void *cookie,
1470                    lu_printer_t printer, const struct cl_page *pg)
1471 {
1472         struct cl_page *scan;
1473
1474         for (scan = cl_page_top((struct cl_page *)pg);
1475              scan != NULL; scan = scan->cp_child)
1476                 cl_page_header_print(env, cookie, printer, scan);
1477         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1478                        (const struct lu_env *env,
1479                         const struct cl_page_slice *slice,
1480                         void *cookie, lu_printer_t p), cookie, printer);
1481         (*printer)(env, cookie, "end page@%p\n", pg);
1482 }
1483 EXPORT_SYMBOL(cl_page_print);
1484
1485 /**
1486  * Cancel a page which is still in a transfer.
1487  */
1488 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1489 {
1490         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1491                               (const struct lu_env *,
1492                                const struct cl_page_slice *));
1493 }
1494 EXPORT_SYMBOL(cl_page_cancel);
1495
1496 /**
1497  * Converts a byte offset within object \a obj into a page index.
1498  */
1499 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1500 {
1501         /*
1502          * XXX for now.
1503          */
1504         return (loff_t)idx << PAGE_CACHE_SHIFT;
1505 }
1506 EXPORT_SYMBOL(cl_offset);
1507
1508 /**
1509  * Converts a page index into a byte offset within object \a obj.
1510  */
1511 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1512 {
1513         /*
1514          * XXX for now.
1515          */
1516         return offset >> PAGE_CACHE_SHIFT;
1517 }
1518 EXPORT_SYMBOL(cl_index);
1519
1520 int cl_page_size(const struct cl_object *obj)
1521 {
1522         return 1 << PAGE_CACHE_SHIFT;
1523 }
1524 EXPORT_SYMBOL(cl_page_size);
1525
1526 /**
1527  * Adds page slice to the compound page.
1528  *
1529  * This is called by cl_object_operations::coo_page_init() methods to add a
1530  * per-layer state to the page. New state is added at the end of
1531  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1532  *
1533  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1534  */
1535 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1536                        struct cl_object *obj,
1537                        const struct cl_page_operations *ops)
1538 {
1539         list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1540         slice->cpl_obj  = obj;
1541         slice->cpl_ops  = ops;
1542         slice->cpl_page = page;
1543 }
1544 EXPORT_SYMBOL(cl_page_slice_add);
1545
1546 int  cl_page_init(void)
1547 {
1548         return 0;
1549 }
1550
1551 void cl_page_fini(void)
1552 {
1553 }