lustre/obdclass/cl_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Client Lustre Page.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_CLASS
  42
  43 #include <libcfs/libcfs.h>
  44 #include <obd_class.h>
  45 #include <obd_support.h>
  46 #include <libcfs/list.h>
  47
  48 #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 static cfs_mem_cache_t      *cl_page_kmem = NULL;
  55
  56 static struct lu_kmem_descr cl_page_caches[] = {
  57         {
  58                 .ckd_cache = &cl_page_kmem,
  59                 .ckd_name  = "cl_page_kmem",
  60                 .ckd_size  = sizeof (struct cl_page)
  61         },
  62         {
  63                 .ckd_cache = NULL
  64         }
  65 };
  66
  67 #ifdef LIBCFS_DEBUG
  68 # define PASSERT(env, page, expr)                                       \
  69   do {                                                                    \
  70           if (unlikely(!(expr))) {                                      \
  71                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  72                   LASSERT(0);                                           \
  73           }                                                             \
  74   } while (0)
  75 #else /* !LIBCFS_DEBUG */
  76 # define PASSERT(env, page, exp) \
  77         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  78 #endif /* !LIBCFS_DEBUG */
  79
  80 #ifdef INVARIANT_CHECK
  81 # define PINVRNT(env, page, expr)                                       \
  82   do {                                                                    \
  83           if (unlikely(!(expr))) {                                      \
  84                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  85                   LINVRNT(0);                                           \
  86           }                                                             \
  87   } while (0)
  88 #else /* !INVARIANT_CHECK */
  89 # define PINVRNT(env, page, exp) \
  90         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  91 #endif /* !INVARIANT_CHECK */
  92
  93 /**
  94  * Internal version of cl_page_top, it should be called with page referenced,
  95  * or coh_page_guard held.
  96  */
  97 static struct cl_page *cl_page_top_trusted(struct cl_page *page)
  98 {
  99         while (page->cp_parent != NULL)
 100                 page = page->cp_parent;
 101         return page;
 102 }
 103
 104 /**
 105  * Internal version of cl_page_get().
 106  *
 107  * This function can be used to obtain initial reference to previously
 108  * unreferenced cached object. It can be called only if concurrent page
 109  * reclamation is somehow prevented, e.g., by locking page radix-tree
 110  * (cl_object_header::hdr->coh_page_guard), or by keeping a lock on a VM page,
 111  * associated with \a page.
 112  *
 113  * Use with care! Not exported.
 114  */
 115 static void cl_page_get_trust(struct cl_page *page)
 116 {
 117         /*
 118          * Checkless version for trusted users.
 119          */
 120         if (cfs_atomic_inc_return(&page->cp_ref) == 1)
 121                 cfs_atomic_inc(&cl_object_site(page->cp_obj)->cs_pages.cs_busy);
 122 }
 123
 124 /**
 125  * Returns a slice within a page, corresponding to the given layer in the
 126  * device stack.
 127  *
 128  * \see cl_lock_at()
 129  */
 130 static const struct cl_page_slice *
 131 cl_page_at_trusted(const struct cl_page *page,
 132                    const struct lu_device_type *dtype)
 133 {
 134         const struct cl_page_slice *slice;
 135
 136 #ifdef INVARIANT_CHECK
 137         struct cl_object_header *ch = cl_object_header(page->cp_obj);
 138
 139         if (!cfs_atomic_read(&page->cp_ref))
 140                 LASSERT_SPIN_LOCKED(&ch->coh_page_guard);
 141 #endif
 142         ENTRY;
 143
 144         page = cl_page_top_trusted((struct cl_page *)page);
 145         do {
 146                 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 147                         if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
 148                                 RETURN(slice);
 149                 }
 150                 page = page->cp_child;
 151         } while (page != NULL);
 152         RETURN(NULL);
 153 }
 154
 155 /**
 156  * Returns a page with given index in the given object, or NULL if no page is
 157  * found. Acquires a reference on \a page.
 158  *
 159  * Locking: called under cl_object_header::coh_page_guard spin-lock.
 160  */
 161 struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index)
 162 {
 163         struct cl_page *page;
 164
 165         LASSERT_SPIN_LOCKED(&hdr->coh_page_guard);
 166
 167         page = radix_tree_lookup(&hdr->coh_tree, index);
 168         if (page != NULL) {
 169                 cl_page_get_trust(page);
 170         }
 171         return page;
 172 }
 173 EXPORT_SYMBOL(cl_page_lookup);
 174
 175 /**
 176  * Returns a list of pages by a given [start, end] of \a obj.
 177  *
 178  * \param resched If not NULL, then we give up before hogging CPU for too
 179  * long and set *resched = 1, in that case caller should implement a retry
 180  * logic.
 181  *
 182  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
 183  * crucial in the face of [offset, EOF] locks.
 184  *
 185  * Return at least one page in @queue unless there is no covered page.
 186  */
 187 int 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                         cl_page_gang_cb_t cb, void *cbdata)
 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         int                      res = CLP_GANG_OKAY;
 201         int                      tree_lock = 1;
 202         ENTRY;
 203
 204         idx = start;
 205         hdr = cl_object_header(obj);
 206         pvec = cl_env_info(env)->clt_pvec;
 207         dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type;
 208         cfs_spin_lock(&hdr->coh_page_guard);
 209         while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec,
 210                                             idx, CLT_PVEC_SIZE)) > 0) {
 211                 int end_of_region = 0;
 212                 idx = pvec[nr - 1]->cp_index + 1;
 213                 for (i = 0, j = 0; i < nr; ++i) {
 214                         page = pvec[i];
 215                         pvec[i] = NULL;
 216
 217                         LASSERT(page->cp_type == CPT_CACHEABLE);
 218                         if (page->cp_index > end) {
 219                                 end_of_region = 1;
 220                                 break;
 221                         }
 222                         if (page->cp_state == CPS_FREEING)
 223                                 continue;
 224
 225                         slice = cl_page_at_trusted(page, dtype);
 226                         /*
 227                          * Pages for lsm-less file has no underneath sub-page
 228                          * for osc, in case of ...
 229                          */
 230                         PASSERT(env, page, slice != NULL);
 231
 232                         page = slice->cpl_page;
 233                         /*
 234                          * Can safely call cl_page_get_trust() under
 235                          * radix-tree spin-lock.
 236                          *
 237                          * XXX not true, because @page is from object another
 238                          * than @hdr and protected by different tree lock.
 239                          */
 240                         cl_page_get_trust(page);
 241                         lu_ref_add_atomic(&page->cp_reference,
 242                                           "gang_lookup", cfs_current());
 243                         pvec[j++] = page;
 244                 }
 245
 246                 /*
 247                  * Here a delicate locking dance is performed. Current thread
 248                  * holds a reference to a page, but has to own it before it
 249                  * can be placed into queue. Owning implies waiting, so
 250                  * radix-tree lock is to be released. After a wait one has to
 251                  * check that pages weren't truncated (cl_page_own() returns
 252                  * error in the latter case).
 253                  */
 254                 cfs_spin_unlock(&hdr->coh_page_guard);
 255                 tree_lock = 0;
 256
 257                 for (i = 0; i < j; ++i) {
 258                         page = pvec[i];
 259                         if (res == CLP_GANG_OKAY)
 260                                 res = (*cb)(env, io, page, cbdata);
 261                         lu_ref_del(&page->cp_reference,
 262                                    "gang_lookup", cfs_current());
 263                         cl_page_put(env, page);
 264                 }
 265                 if (nr < CLT_PVEC_SIZE || end_of_region)
 266                         break;
 267
 268                 if (res == CLP_GANG_OKAY && cfs_need_resched())
 269                         res = CLP_GANG_RESCHED;
 270                 if (res != CLP_GANG_OKAY)
 271                         break;
 272
 273                 cfs_spin_lock(&hdr->coh_page_guard);
 274                 tree_lock = 1;
 275         }
 276         if (tree_lock)
 277                 cfs_spin_unlock(&hdr->coh_page_guard);
 278         RETURN(res);
 279 }
 280 EXPORT_SYMBOL(cl_page_gang_lookup);
 281
 282 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
 283 {
 284         struct cl_object *obj  = page->cp_obj;
 285         struct cl_site   *site = cl_object_site(obj);
 286
 287         PASSERT(env, page, cfs_list_empty(&page->cp_batch));
 288         PASSERT(env, page, page->cp_owner == NULL);
 289         PASSERT(env, page, page->cp_req == NULL);
 290         PASSERT(env, page, page->cp_parent == NULL);
 291         PASSERT(env, page, page->cp_state == CPS_FREEING);
 292
 293         ENTRY;
 294         cfs_might_sleep();
 295         while (!cfs_list_empty(&page->cp_layers)) {
 296                 struct cl_page_slice *slice;
 297
 298                 slice = cfs_list_entry(page->cp_layers.next,
 299                                        struct cl_page_slice, cpl_linkage);
 300                 cfs_list_del_init(page->cp_layers.next);
 301                 slice->cpl_ops->cpo_fini(env, slice);
 302         }
 303         cfs_atomic_dec(&site->cs_pages.cs_total);
 304
 305 #ifdef LUSTRE_PAGESTATE_TRACKING
 306         cfs_atomic_dec(&site->cs_pages_state[page->cp_state]);
 307 #endif
 308         lu_object_ref_del_at(&obj->co_lu, page->cp_obj_ref, "cl_page", page);
 309         cl_object_put(env, obj);
 310         lu_ref_fini(&page->cp_reference);
 311         OBD_SLAB_FREE_PTR(page, cl_page_kmem);
 312         EXIT;
 313 }
 314
 315 /**
 316  * Helper function updating page state. This is the only place in the code
 317  * where cl_page::cp_state field is mutated.
 318  */
 319 static inline void cl_page_state_set_trust(struct cl_page *page,
 320                                            enum cl_page_state state)
 321 {
 322         /* bypass const. */
 323         *(enum cl_page_state *)&page->cp_state = state;
 324 }
 325
 326 static int cl_page_alloc(const struct lu_env *env, struct cl_object *o,
 327                          pgoff_t ind, struct page *vmpage,
 328                          enum cl_page_type type, struct cl_page **out)
 329 {
 330         struct cl_page          *page;
 331         struct cl_page          *err  = NULL;
 332         struct lu_object_header *head;
 333         struct cl_site          *site = cl_object_site(o);
 334         int                      result;
 335
 336         ENTRY;
 337         result = +1;
 338         OBD_SLAB_ALLOC_PTR_GFP(page, cl_page_kmem, CFS_ALLOC_IO);
 339         if (page != NULL) {
 340                 cfs_atomic_set(&page->cp_ref, 1);
 341                 page->cp_obj = o;
 342                 cl_object_get(o);
 343                 page->cp_obj_ref = lu_object_ref_add(&o->co_lu,
 344                                                      "cl_page", page);
 345                 page->cp_index = ind;
 346                 cl_page_state_set_trust(page, CPS_CACHED);
 347                 page->cp_type = type;
 348                 CFS_INIT_LIST_HEAD(&page->cp_layers);
 349                 CFS_INIT_LIST_HEAD(&page->cp_batch);
 350                 CFS_INIT_LIST_HEAD(&page->cp_flight);
 351                 cfs_mutex_init(&page->cp_mutex);
 352                 lu_ref_init(&page->cp_reference);
 353                 head = o->co_lu.lo_header;
 354                 cfs_list_for_each_entry(o, &head->loh_layers,
 355                                         co_lu.lo_linkage) {
 356                         if (o->co_ops->coo_page_init != NULL) {
 357                                 err = o->co_ops->coo_page_init(env, o,
 358                                                                page, vmpage);
 359                                 if (err != NULL) {
 360                                         cl_page_delete0(env, page, 0);
 361                                         cl_page_free(env, page);
 362                                         page = err;
 363                                         break;
 364                                 }
 365                         }
 366                 }
 367                 if (err == NULL) {
 368                         cfs_atomic_inc(&site->cs_pages.cs_busy);
 369                         cfs_atomic_inc(&site->cs_pages.cs_total);
 370
 371 #ifdef LUSTRE_PAGESTATE_TRACKING
 372                         cfs_atomic_inc(&site->cs_pages_state[CPS_CACHED]);
 373 #endif
 374                         cfs_atomic_inc(&site->cs_pages.cs_created);
 375                         result = 0;
 376                 }
 377         } else
 378                 page = ERR_PTR(-ENOMEM);
 379         *out = page;
 380         RETURN(result);
 381 }
 382
 383 /**
 384  * Returns a cl_page with index \a idx at the object \a o, and associated with
 385  * the VM page \a vmpage.
 386  *
 387  * This is the main entry point into the cl_page caching interface. First, a
 388  * cache (implemented as a per-object radix tree) is consulted. If page is
 389  * found there, it is returned immediately. Otherwise new page is allocated
 390  * and returned. In any case, additional reference to page is acquired.
 391  *
 392  * \see cl_object_find(), cl_lock_find()
 393  */
 394 static struct cl_page *cl_page_find0(const struct lu_env *env,
 395                                      struct cl_object *o,
 396                                      pgoff_t idx, struct page *vmpage,
 397                                      enum cl_page_type type,
 398                                      struct cl_page *parent)
 399 {
 400         struct cl_page          *page = NULL;
 401         struct cl_page          *ghost = NULL;
 402         struct cl_object_header *hdr;
 403         struct cl_site          *site = cl_object_site(o);
 404         int err;
 405
 406         LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 407         cfs_might_sleep();
 408
 409         ENTRY;
 410
 411         hdr = cl_object_header(o);
 412         cfs_atomic_inc(&site->cs_pages.cs_lookup);
 413
 414         CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
 415                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 416         /* fast path. */
 417         if (type == CPT_CACHEABLE) {
 418                 /*
 419                  * cl_vmpage_page() can be called here without any locks as
 420                  *
 421                  *     - "vmpage" is locked (which prevents ->private from
 422                  *       concurrent updates), and
 423                  *
 424                  *     - "o" cannot be destroyed while current thread holds a
 425                  *       reference on it.
 426                  */
 427                 page = cl_vmpage_page(vmpage, o);
 428                 PINVRNT(env, page,
 429                         ergo(page != NULL,
 430                              cl_page_vmpage(env, page) == vmpage &&
 431                              (void *)radix_tree_lookup(&hdr->coh_tree,
 432                                                        idx) == page));
 433         }
 434
 435         if (page != NULL) {
 436                 cfs_atomic_inc(&site->cs_pages.cs_hit);
 437                 RETURN(page);
 438         }
 439
 440         /* allocate and initialize cl_page */
 441         err = cl_page_alloc(env, o, idx, vmpage, type, &page);
 442         if (err != 0)
 443                 RETURN(page);
 444
 445         if (type == CPT_TRANSIENT) {
 446                 if (parent) {
 447                         LASSERT(page->cp_parent == NULL);
 448                         page->cp_parent = parent;
 449                         parent->cp_child = page;
 450                 }
 451                 RETURN(page);
 452         }
 453
 454         /*
 455          * XXX optimization: use radix_tree_preload() here, and change tree
 456          * gfp mask to GFP_KERNEL in cl_object_header_init().
 457          */
 458         cfs_spin_lock(&hdr->coh_page_guard);
 459         err = radix_tree_insert(&hdr->coh_tree, idx, page);
 460         if (err != 0) {
 461                 ghost = page;
 462                 /*
 463                  * Noted by Jay: a lock on \a vmpage protects cl_page_find()
 464                  * from this race, but
 465                  *
 466                  *     0. it's better to have cl_page interface "locally
 467                  *     consistent" so that its correctness can be reasoned
 468                  *     about without appealing to the (obscure world of) VM
 469                  *     locking.
 470                  *
 471                  *     1. handling this race allows ->coh_tree to remain
 472                  *     consistent even when VM locking is somehow busted,
 473                  *     which is very useful during diagnosing and debugging.
 474                  */
 475                 page = ERR_PTR(err);
 476                 CL_PAGE_DEBUG(D_ERROR, env, ghost,
 477                               "fail to insert into radix tree: %d\n", err);
 478         } else {
 479                 if (parent) {
 480                         LASSERT(page->cp_parent == NULL);
 481                         page->cp_parent = parent;
 482                         parent->cp_child = page;
 483                 }
 484                 hdr->coh_pages++;
 485         }
 486         cfs_spin_unlock(&hdr->coh_page_guard);
 487
 488         if (unlikely(ghost != NULL)) {
 489                 cfs_atomic_dec(&site->cs_pages.cs_busy);
 490                 cl_page_delete0(env, ghost, 0);
 491                 cl_page_free(env, ghost);
 492         }
 493         RETURN(page);
 494 }
 495
 496 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o,
 497                              pgoff_t idx, struct page *vmpage,
 498                              enum cl_page_type type)
 499 {
 500         return cl_page_find0(env, o, idx, vmpage, type, NULL);
 501 }
 502 EXPORT_SYMBOL(cl_page_find);
 503
 504
 505 struct cl_page *cl_page_find_sub(const struct lu_env *env, struct cl_object *o,
 506                                  pgoff_t idx, struct page *vmpage,
 507                                  struct cl_page *parent)
 508 {
 509         return cl_page_find0(env, o, idx, vmpage, parent->cp_type, parent);
 510 }
 511 EXPORT_SYMBOL(cl_page_find_sub);
 512
 513 static inline int cl_page_invariant(const struct cl_page *pg)
 514 {
 515         struct cl_object_header *header;
 516         struct cl_page          *parent;
 517         struct cl_page          *child;
 518         struct cl_io            *owner;
 519
 520         /*
 521          * Page invariant is protected by a VM lock.
 522          */
 523         LINVRNT(cl_page_is_vmlocked(NULL, pg));
 524
 525         header = cl_object_header(pg->cp_obj);
 526         parent = pg->cp_parent;
 527         child  = pg->cp_child;
 528         owner  = pg->cp_owner;
 529
 530         return cfs_atomic_read(&pg->cp_ref) > 0 &&
 531                 ergo(parent != NULL, parent->cp_child == pg) &&
 532                 ergo(child != NULL, child->cp_parent == pg) &&
 533                 ergo(child != NULL, pg->cp_obj != child->cp_obj) &&
 534                 ergo(parent != NULL, pg->cp_obj != parent->cp_obj) &&
 535                 ergo(owner != NULL && parent != NULL,
 536                      parent->cp_owner == pg->cp_owner->ci_parent) &&
 537                 ergo(owner != NULL && child != NULL,
 538                      child->cp_owner->ci_parent == owner) &&
 539                 /*
 540                  * Either page is early in initialization (has neither child
 541                  * nor parent yet), or it is in the object radix tree.
 542                  */
 543                 ergo(pg->cp_state < CPS_FREEING && pg->cp_type == CPT_CACHEABLE,
 544                      (void *)radix_tree_lookup(&header->coh_tree,
 545                                                pg->cp_index) == pg ||
 546                      (child == NULL && parent == NULL));
 547 }
 548
 549 static void cl_page_state_set0(const struct lu_env *env,
 550                                struct cl_page *page, enum cl_page_state state)
 551 {
 552         enum cl_page_state old;
 553 #ifdef LUSTRE_PAGESTATE_TRACKING
 554         struct cl_site *site = cl_object_site(page->cp_obj);
 555 #endif
 556
 557         /*
 558          * Matrix of allowed state transitions [old][new], for sanity
 559          * checking.
 560          */
 561         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 562                 [CPS_CACHED] = {
 563                         [CPS_CACHED]  = 0,
 564                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 565                         [CPS_PAGEIN]  = 0,
 566                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 567                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 568                 },
 569                 [CPS_OWNED] = {
 570                         [CPS_CACHED]  = 1, /* release to the cache */
 571                         [CPS_OWNED]   = 0,
 572                         [CPS_PAGEIN]  = 1, /* start read immediately */
 573                         [CPS_PAGEOUT] = 1, /* start write immediately */
 574                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 575                 },
 576                 [CPS_PAGEIN] = {
 577                         [CPS_CACHED]  = 1, /* io completion */
 578                         [CPS_OWNED]   = 0,
 579                         [CPS_PAGEIN]  = 0,
 580                         [CPS_PAGEOUT] = 0,
 581                         [CPS_FREEING] = 0,
 582                 },
 583                 [CPS_PAGEOUT] = {
 584                         [CPS_CACHED]  = 1, /* io completion */
 585                         [CPS_OWNED]   = 0,
 586                         [CPS_PAGEIN]  = 0,
 587                         [CPS_PAGEOUT] = 0,
 588                         [CPS_FREEING] = 0,
 589                 },
 590                 [CPS_FREEING] = {
 591                         [CPS_CACHED]  = 0,
 592                         [CPS_OWNED]   = 0,
 593                         [CPS_PAGEIN]  = 0,
 594                         [CPS_PAGEOUT] = 0,
 595                         [CPS_FREEING] = 0,
 596                 }
 597         };
 598
 599         ENTRY;
 600         old = page->cp_state;
 601         PASSERT(env, page, allowed_transitions[old][state]);
 602         CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
 603         for (; page != NULL; page = page->cp_child) {
 604                 PASSERT(env, page, page->cp_state == old);
 605                 PASSERT(env, page,
 606                         equi(state == CPS_OWNED, page->cp_owner != NULL));
 607
 608 #ifdef LUSTRE_PAGESTATE_TRACKING
 609                 cfs_atomic_dec(&site->cs_pages_state[page->cp_state]);
 610                 cfs_atomic_inc(&site->cs_pages_state[state]);
 611 #endif
 612                 cl_page_state_set_trust(page, state);
 613         }
 614         EXIT;
 615 }
 616
 617 static void cl_page_state_set(const struct lu_env *env,
 618                               struct cl_page *page, enum cl_page_state state)
 619 {
 620         cl_page_state_set0(env, page, state);
 621 }
 622
 623 /**
 624  * Acquires an additional reference to a page.
 625  *
 626  * This can be called only by caller already possessing a reference to \a
 627  * page.
 628  *
 629  * \see cl_object_get(), cl_lock_get().
 630  */
 631 void cl_page_get(struct cl_page *page)
 632 {
 633         ENTRY;
 634         LASSERT(page->cp_state != CPS_FREEING);
 635         cl_page_get_trust(page);
 636         EXIT;
 637 }
 638 EXPORT_SYMBOL(cl_page_get);
 639
 640 /**
 641  * Releases a reference to a page.
 642  *
 643  * When last reference is released, page is returned to the cache, unless it
 644  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 645  * destroyed.
 646  *
 647  * \see cl_object_put(), cl_lock_put().
 648  */
 649 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 650 {
 651         struct cl_object_header *hdr;
 652         struct cl_site *site = cl_object_site(page->cp_obj);
 653
 654         PASSERT(env, page, cfs_atomic_read(&page->cp_ref) > !!page->cp_parent);
 655
 656         ENTRY;
 657         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 658                        cfs_atomic_read(&page->cp_ref));
 659
 660         hdr = cl_object_header(cl_object_top(page->cp_obj));
 661         if (cfs_atomic_dec_and_lock(&page->cp_ref, &hdr->coh_page_guard)) {
 662                 cfs_atomic_dec(&site->cs_pages.cs_busy);
 663                 /* We're going to access the page w/o a reference, but it's
 664                  * ok because we have grabbed the lock coh_page_guard, which
 665                  * means nobody is able to free this page behind us.
 666                  */
 667                 if (page->cp_state == CPS_FREEING) {
 668                         /* We drop the page reference and check the page state
 669                          * inside the coh_page_guard. So that if it gets here,
 670                          * it is the REALLY last reference to this page.
 671                          */
 672                         cfs_spin_unlock(&hdr->coh_page_guard);
 673
 674                         LASSERT(cfs_atomic_read(&page->cp_ref) == 0);
 675                         PASSERT(env, page, page->cp_owner == NULL);
 676                         PASSERT(env, page, cfs_list_empty(&page->cp_batch));
 677                         /*
 678                          * Page is no longer reachable by other threads. Tear
 679                          * it down.
 680                          */
 681                         cl_page_free(env, page);
 682
 683                         EXIT;
 684                         return;
 685                 }
 686                 cfs_spin_unlock(&hdr->coh_page_guard);
 687         }
 688
 689         EXIT;
 690 }
 691 EXPORT_SYMBOL(cl_page_put);
 692
 693 /**
 694  * Returns a VM page associated with a given cl_page.
 695  */
 696 cfs_page_t *cl_page_vmpage(const struct lu_env *env, struct cl_page *page)
 697 {
 698         const struct cl_page_slice *slice;
 699
 700         /*
 701          * Find uppermost layer with ->cpo_vmpage() method, and return its
 702          * result.
 703          */
 704         page = cl_page_top(page);
 705         do {
 706                 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 707                         if (slice->cpl_ops->cpo_vmpage != NULL)
 708                                 RETURN(slice->cpl_ops->cpo_vmpage(env, slice));
 709                 }
 710                 page = page->cp_child;
 711         } while (page != NULL);
 712         LBUG(); /* ->cpo_vmpage() has to be defined somewhere in the stack */
 713 }
 714 EXPORT_SYMBOL(cl_page_vmpage);
 715
 716 /**
 717  * Returns a cl_page associated with a VM page, and given cl_object.
 718  */
 719 struct cl_page *cl_vmpage_page(cfs_page_t *vmpage, struct cl_object *obj)
 720 {
 721         struct cl_page *page;
 722         struct cl_object_header *hdr;
 723
 724         ENTRY;
 725         KLASSERT(PageLocked(vmpage));
 726
 727         /*
 728          * NOTE: absence of races and liveness of data are guaranteed by page
 729          *       lock on a "vmpage". That works because object destruction has
 730          *       bottom-to-top pass.
 731          */
 732
 733         /*
 734          * This loop assumes that ->private points to the top-most page. This
 735          * can be rectified easily.
 736          */
 737         hdr = cl_object_header(cl_object_top(obj));
 738         cfs_spin_lock(&hdr->coh_page_guard);
 739         for (page = (void *)vmpage->private;
 740              page != NULL; page = page->cp_child) {
 741                 if (cl_object_same(page->cp_obj, obj)) {
 742                         cl_page_get_trust(page);
 743                         break;
 744                 }
 745         }
 746         cfs_spin_unlock(&hdr->coh_page_guard);
 747         LASSERT(ergo(page, page->cp_type == CPT_CACHEABLE));
 748         RETURN(page);
 749 }
 750 EXPORT_SYMBOL(cl_vmpage_page);
 751
 752 /**
 753  * Returns the top-page for a given page.
 754  *
 755  * \see cl_object_top(), cl_io_top()
 756  */
 757 struct cl_page *cl_page_top(struct cl_page *page)
 758 {
 759         return cl_page_top_trusted(page);
 760 }
 761 EXPORT_SYMBOL(cl_page_top);
 762
 763 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 764                                        const struct lu_device_type *dtype)
 765 {
 766         return cl_page_at_trusted(page, dtype);
 767 }
 768 EXPORT_SYMBOL(cl_page_at);
 769
 770 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 771
 772 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)                   \
 773 ({                                                                      \
 774         const struct lu_env        *__env  = (_env);                    \
 775         struct cl_page             *__page = (_page);                   \
 776         const struct cl_page_slice *__scan;                             \
 777         int                         __result;                           \
 778         ptrdiff_t                   __op   = (_op);                     \
 779         int                       (*__method)_proto;                    \
 780                                                                         \
 781         __result = 0;                                                   \
 782         __page = cl_page_top(__page);                                   \
 783         do {                                                            \
 784                 cfs_list_for_each_entry(__scan, &__page->cp_layers,     \
 785                                         cpl_linkage) {                  \
 786                         __method = *(void **)((char *)__scan->cpl_ops + \
 787                                               __op);                    \
 788                         if (__method != NULL) {                         \
 789                                 __result = (*__method)(__env, __scan,   \
 790                                                        ## __VA_ARGS__); \
 791                                 if (__result != 0)                      \
 792                                         break;                          \
 793                         }                                               \
 794                 }                                                       \
 795                 __page = __page->cp_child;                              \
 796         } while (__page != NULL && __result == 0);                      \
 797         if (__result > 0)                                               \
 798                 __result = 0;                                           \
 799         __result;                                                       \
 800 })
 801
 802 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 803 do {                                                                    \
 804         const struct lu_env        *__env  = (_env);                    \
 805         struct cl_page             *__page = (_page);                   \
 806         const struct cl_page_slice *__scan;                             \
 807         ptrdiff_t                   __op   = (_op);                     \
 808         void                      (*__method)_proto;                    \
 809                                                                         \
 810         __page = cl_page_top(__page);                                   \
 811         do {                                                            \
 812                 cfs_list_for_each_entry(__scan, &__page->cp_layers,     \
 813                                         cpl_linkage) {                  \
 814                         __method = *(void **)((char *)__scan->cpl_ops + \
 815                                               __op);                    \
 816                         if (__method != NULL)                           \
 817                                 (*__method)(__env, __scan,              \
 818                                             ## __VA_ARGS__);            \
 819                 }                                                       \
 820                 __page = __page->cp_child;                              \
 821         } while (__page != NULL);                                       \
 822 } while (0)
 823
 824 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)               \
 825 do {                                                                        \
 826         const struct lu_env        *__env  = (_env);                        \
 827         struct cl_page             *__page = (_page);                       \
 828         const struct cl_page_slice *__scan;                                 \
 829         ptrdiff_t                   __op   = (_op);                         \
 830         void                      (*__method)_proto;                        \
 831                                                                             \
 832         /* get to the bottom page. */                                       \
 833         while (__page->cp_child != NULL)                                    \
 834                 __page = __page->cp_child;                                  \
 835         do {                                                                \
 836                 cfs_list_for_each_entry_reverse(__scan, &__page->cp_layers, \
 837                                                 cpl_linkage) {              \
 838                         __method = *(void **)((char *)__scan->cpl_ops +     \
 839                                               __op);                        \
 840                         if (__method != NULL)                               \
 841                                 (*__method)(__env, __scan,                  \
 842                                             ## __VA_ARGS__);                \
 843                 }                                                           \
 844                 __page = __page->cp_parent;                                 \
 845         } while (__page != NULL);                                           \
 846 } while (0)
 847
 848 static int cl_page_invoke(const struct lu_env *env,
 849                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 850
 851 {
 852         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 853         ENTRY;
 854         RETURN(CL_PAGE_INVOKE(env, page, op,
 855                               (const struct lu_env *,
 856                                const struct cl_page_slice *, struct cl_io *),
 857                               io));
 858 }
 859
 860 static void cl_page_invoid(const struct lu_env *env,
 861                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 862
 863 {
 864         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 865         ENTRY;
 866         CL_PAGE_INVOID(env, page, op,
 867                        (const struct lu_env *,
 868                         const struct cl_page_slice *, struct cl_io *), io);
 869         EXIT;
 870 }
 871
 872 static void cl_page_owner_clear(struct cl_page *page)
 873 {
 874         ENTRY;
 875         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 876                 if (page->cp_owner != NULL) {
 877                         LASSERT(page->cp_owner->ci_owned_nr > 0);
 878                         page->cp_owner->ci_owned_nr--;
 879                         page->cp_owner = NULL;
 880                         page->cp_task = NULL;
 881                 }
 882         }
 883         EXIT;
 884 }
 885
 886 static void cl_page_owner_set(struct cl_page *page)
 887 {
 888         ENTRY;
 889         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 890                 LASSERT(page->cp_owner != NULL);
 891                 page->cp_owner->ci_owned_nr++;
 892         }
 893         EXIT;
 894 }
 895
 896 void cl_page_disown0(const struct lu_env *env,
 897                      struct cl_io *io, struct cl_page *pg)
 898 {
 899         enum cl_page_state state;
 900
 901         ENTRY;
 902         state = pg->cp_state;
 903         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 904         PINVRNT(env, pg, cl_page_invariant(pg));
 905         cl_page_owner_clear(pg);
 906
 907         if (state == CPS_OWNED)
 908                 cl_page_state_set(env, pg, CPS_CACHED);
 909         /*
 910          * Completion call-backs are executed in the bottom-up order, so that
 911          * uppermost layer (llite), responsible for VFS/VM interaction runs
 912          * last and can release locks safely.
 913          */
 914         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 915                                (const struct lu_env *,
 916                                 const struct cl_page_slice *, struct cl_io *),
 917                                io);
 918         EXIT;
 919 }
 920
 921 /**
 922  * returns true, iff page is owned by the given io.
 923  */
 924 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 925 {
 926         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 927         ENTRY;
 928         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == io);
 929 }
 930 EXPORT_SYMBOL(cl_page_is_owned);
 931
 932 /**
 933  * Try to own a page by IO.
 934  *
 935  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 936  * into cl_page_state::CPS_OWNED state.
 937  *
 938  * \pre  !cl_page_is_owned(pg, io)
 939  * \post result == 0 iff cl_page_is_owned(pg, io)
 940  *
 941  * \retval 0   success
 942  *
 943  * \retval -ve failure, e.g., page was destroyed (and landed in
 944  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 945  *             or, page was owned by another thread, or in IO.
 946  *
 947  * \see cl_page_disown()
 948  * \see cl_page_operations::cpo_own()
 949  * \see cl_page_own_try()
 950  * \see cl_page_own
 951  */
 952 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 953                         struct cl_page *pg, int nonblock)
 954 {
 955         int result;
 956
 957         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 958
 959         ENTRY;
 960         pg = cl_page_top(pg);
 961         io = cl_io_top(io);
 962
 963         if (pg->cp_state == CPS_FREEING) {
 964                 result = -ENOENT;
 965         } else {
 966                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 967                                         (const struct lu_env *,
 968                                          const struct cl_page_slice *,
 969                                          struct cl_io *, int),
 970                                         io, nonblock);
 971                 if (result == 0) {
 972                         PASSERT(env, pg, pg->cp_owner == NULL);
 973                         PASSERT(env, pg, pg->cp_req == NULL);
 974                         pg->cp_owner = io;
 975                         pg->cp_task  = current;
 976                         cl_page_owner_set(pg);
 977                         if (pg->cp_state != CPS_FREEING) {
 978                                 cl_page_state_set(env, pg, CPS_OWNED);
 979                         } else {
 980                                 cl_page_disown0(env, io, pg);
 981                                 result = -ENOENT;
 982                         }
 983                 }
 984         }
 985         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 986         RETURN(result);
 987 }
 988
 989 /**
 990  * Own a page, might be blocked.
 991  *
 992  * \see cl_page_own0()
 993  */
 994 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 995 {
 996         return cl_page_own0(env, io, pg, 0);
 997 }
 998 EXPORT_SYMBOL(cl_page_own);
 999
1000 /**
1001  * Nonblock version of cl_page_own().
1002  *
1003  * \see cl_page_own0()
1004  */
1005 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
1006                     struct cl_page *pg)
1007 {
1008         return cl_page_own0(env, io, pg, 1);
1009 }
1010 EXPORT_SYMBOL(cl_page_own_try);
1011
1012
1013 /**
1014  * Assume page ownership.
1015  *
1016  * Called when page is already locked by the hosting VM.
1017  *
1018  * \pre !cl_page_is_owned(pg, io)
1019  * \post cl_page_is_owned(pg, io)
1020  *
1021  * \see cl_page_operations::cpo_assume()
1022  */
1023 void cl_page_assume(const struct lu_env *env,
1024                     struct cl_io *io, struct cl_page *pg)
1025 {
1026         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
1027
1028         ENTRY;
1029         pg = cl_page_top(pg);
1030         io = cl_io_top(io);
1031
1032         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
1033         PASSERT(env, pg, pg->cp_owner == NULL);
1034         pg->cp_owner = io;
1035         pg->cp_task = current;
1036         cl_page_owner_set(pg);
1037         cl_page_state_set(env, pg, CPS_OWNED);
1038         EXIT;
1039 }
1040 EXPORT_SYMBOL(cl_page_assume);
1041
1042 /**
1043  * Releases page ownership without unlocking the page.
1044  *
1045  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
1046  * underlying VM page (as VM is supposed to do this itself).
1047  *
1048  * \pre   cl_page_is_owned(pg, io)
1049  * \post !cl_page_is_owned(pg, io)
1050  *
1051  * \see cl_page_assume()
1052  */
1053 void cl_page_unassume(const struct lu_env *env,
1054                       struct cl_io *io, struct cl_page *pg)
1055 {
1056         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1057         PINVRNT(env, pg, cl_page_invariant(pg));
1058
1059         ENTRY;
1060         pg = cl_page_top(pg);
1061         io = cl_io_top(io);
1062         cl_page_owner_clear(pg);
1063         cl_page_state_set(env, pg, CPS_CACHED);
1064         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
1065                                (const struct lu_env *,
1066                                 const struct cl_page_slice *, struct cl_io *),
1067                                io);
1068         EXIT;
1069 }
1070 EXPORT_SYMBOL(cl_page_unassume);
1071
1072 /**
1073  * Releases page ownership.
1074  *
1075  * Moves page into cl_page_state::CPS_CACHED.
1076  *
1077  * \pre   cl_page_is_owned(pg, io)
1078  * \post !cl_page_is_owned(pg, io)
1079  *
1080  * \see cl_page_own()
1081  * \see cl_page_operations::cpo_disown()
1082  */
1083 void cl_page_disown(const struct lu_env *env,
1084                     struct cl_io *io, struct cl_page *pg)
1085 {
1086         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1087
1088         ENTRY;
1089         pg = cl_page_top(pg);
1090         io = cl_io_top(io);
1091         cl_page_disown0(env, io, pg);
1092         EXIT;
1093 }
1094 EXPORT_SYMBOL(cl_page_disown);
1095
1096 /**
1097  * Called when page is to be removed from the object, e.g., as a result of
1098  * truncate.
1099  *
1100  * Calls cl_page_operations::cpo_discard() top-to-bottom.
1101  *
1102  * \pre cl_page_is_owned(pg, io)
1103  *
1104  * \see cl_page_operations::cpo_discard()
1105  */
1106 void cl_page_discard(const struct lu_env *env,
1107                      struct cl_io *io, struct cl_page *pg)
1108 {
1109         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1110         PINVRNT(env, pg, cl_page_invariant(pg));
1111
1112         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
1113 }
1114 EXPORT_SYMBOL(cl_page_discard);
1115
1116 /**
1117  * Version of cl_page_delete() that can be called for not fully constructed
1118  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
1119  * path. Doesn't check page invariant.
1120  */
1121 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
1122                             int radix)
1123 {
1124         struct cl_page *tmp = pg;
1125         ENTRY;
1126
1127         PASSERT(env, pg, pg == cl_page_top(pg));
1128         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
1129
1130         /*
1131          * Severe all ways to obtain new pointers to @pg.
1132          */
1133         cl_page_owner_clear(pg);
1134
1135         /*
1136          * unexport the page firstly before freeing it so that
1137          * the page content is considered to be invalid.
1138          * We have to do this because a CPS_FREEING cl_page may
1139          * be NOT under the protection of a cl_lock.
1140          * Afterwards, if this page is found by other threads, then this
1141          * page will be forced to reread.
1142          */
1143         cl_page_export(env, pg, 0);
1144         cl_page_state_set0(env, pg, CPS_FREEING);
1145
1146         if (tmp->cp_type == CPT_CACHEABLE) {
1147                 if (!radix)
1148                         /* !radix means that @pg is not yet in the radix tree,
1149                          * skip removing it.
1150                          */
1151                         tmp = pg->cp_child;
1152                 for (; tmp != NULL; tmp = tmp->cp_child) {
1153                         void                    *value;
1154                         struct cl_object_header *hdr;
1155
1156                         hdr = cl_object_header(tmp->cp_obj);
1157                         cfs_spin_lock(&hdr->coh_page_guard);
1158                         value = radix_tree_delete(&hdr->coh_tree,
1159                                                   tmp->cp_index);
1160                         PASSERT(env, tmp, value == tmp);
1161                         PASSERT(env, tmp, hdr->coh_pages > 0);
1162                         hdr->coh_pages--;
1163                         cfs_spin_unlock(&hdr->coh_page_guard);
1164                 }
1165         }
1166
1167         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete),
1168                        (const struct lu_env *, const struct cl_page_slice *));
1169         EXIT;
1170 }
1171
1172 /**
1173  * Called when a decision is made to throw page out of memory.
1174  *
1175  * Notifies all layers about page destruction by calling
1176  * cl_page_operations::cpo_delete() method top-to-bottom.
1177  *
1178  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
1179  * where transition to this state happens).
1180  *
1181  * Eliminates all venues through which new references to the page can be
1182  * obtained:
1183  *
1184  *     - removes page from the radix trees,
1185  *
1186  *     - breaks linkage from VM page to cl_page.
1187  *
1188  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
1189  * drain after some time, at which point page will be recycled.
1190  *
1191  * \pre  pg == cl_page_top(pg)
1192  * \pre  VM page is locked
1193  * \post pg->cp_state == CPS_FREEING
1194  *
1195  * \see cl_page_operations::cpo_delete()
1196  */
1197 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
1198 {
1199         PINVRNT(env, pg, cl_page_invariant(pg));
1200         ENTRY;
1201         cl_page_delete0(env, pg, 1);
1202         EXIT;
1203 }
1204 EXPORT_SYMBOL(cl_page_delete);
1205
1206 /**
1207  * Unmaps page from user virtual memory.
1208  *
1209  * Calls cl_page_operations::cpo_unmap() through all layers top-to-bottom. The
1210  * layer responsible for VM interaction has to unmap page from user space
1211  * virtual memory.
1212  *
1213  * \see cl_page_operations::cpo_unmap()
1214  */
1215 int cl_page_unmap(const struct lu_env *env,
1216                   struct cl_io *io, struct cl_page *pg)
1217 {
1218         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1219         PINVRNT(env, pg, cl_page_invariant(pg));
1220
1221         return cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_unmap));
1222 }
1223 EXPORT_SYMBOL(cl_page_unmap);
1224
1225 /**
1226  * Marks page up-to-date.
1227  *
1228  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
1229  * layer responsible for VM interaction has to mark/clear page as up-to-date
1230  * by the \a uptodate argument.
1231  *
1232  * \see cl_page_operations::cpo_export()
1233  */
1234 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
1235 {
1236         PINVRNT(env, pg, cl_page_invariant(pg));
1237         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
1238                        (const struct lu_env *,
1239                         const struct cl_page_slice *, int), uptodate);
1240 }
1241 EXPORT_SYMBOL(cl_page_export);
1242
1243 /**
1244  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
1245  * thread.
1246  */
1247 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
1248 {
1249         int result;
1250         const struct cl_page_slice *slice;
1251
1252         ENTRY;
1253         pg = cl_page_top_trusted((struct cl_page *)pg);
1254         slice = container_of(pg->cp_layers.next,
1255                              const struct cl_page_slice, cpl_linkage);
1256         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
1257         /*
1258          * Call ->cpo_is_vmlocked() directly instead of going through
1259          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
1260          * cl_page_invariant().
1261          */
1262         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
1263         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
1264         RETURN(result == -EBUSY);
1265 }
1266 EXPORT_SYMBOL(cl_page_is_vmlocked);
1267
1268 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
1269 {
1270         ENTRY;
1271         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
1272 }
1273
1274 static void cl_page_io_start(const struct lu_env *env,
1275                              struct cl_page *pg, enum cl_req_type crt)
1276 {
1277         /*
1278          * Page is queued for IO, change its state.
1279          */
1280         ENTRY;
1281         cl_page_owner_clear(pg);
1282         cl_page_state_set(env, pg, cl_req_type_state(crt));
1283         EXIT;
1284 }
1285
1286 /**
1287  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
1288  * called top-to-bottom. Every layer either agrees to submit this page (by
1289  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
1290  * handling interactions with the VM also has to inform VM that page is under
1291  * transfer now.
1292  */
1293 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
1294                  struct cl_page *pg, enum cl_req_type crt)
1295 {
1296         int result;
1297
1298         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1299         PINVRNT(env, pg, cl_page_invariant(pg));
1300         PINVRNT(env, pg, crt < CRT_NR);
1301
1302         /*
1303          * XXX this has to be called bottom-to-top, so that llite can set up
1304          * PG_writeback without risking other layers deciding to skip this
1305          * page.
1306          */
1307         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
1308         if (result == 0)
1309                 cl_page_io_start(env, pg, crt);
1310
1311         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
1312                       equi(result == 0,
1313                            PageWriteback(cl_page_vmpage(env, pg)))));
1314         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1315         return result;
1316 }
1317 EXPORT_SYMBOL(cl_page_prep);
1318
1319 /**
1320  * Notify layers about transfer completion.
1321  *
1322  * Invoked by transfer sub-system (which is a part of osc) to notify layers
1323  * that a transfer, of which this page is a part of has completed.
1324  *
1325  * Completion call-backs are executed in the bottom-up order, so that
1326  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1327  * and can release locks safely.
1328  *
1329  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1330  * \post pg->cp_state == CPS_CACHED
1331  *
1332  * \see cl_page_operations::cpo_completion()
1333  */
1334 void cl_page_completion(const struct lu_env *env,
1335                         struct cl_page *pg, enum cl_req_type crt, int ioret)
1336 {
1337         struct cl_sync_io *anchor = pg->cp_sync_io;
1338
1339         PASSERT(env, pg, crt < CRT_NR);
1340         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
1341         PASSERT(env, pg, pg->cp_req == NULL);
1342         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
1343
1344         ENTRY;
1345         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
1346         if (crt == CRT_READ && ioret == 0) {
1347                 PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED));
1348                 pg->cp_flags |= CPF_READ_COMPLETED;
1349         }
1350
1351         cl_page_state_set(env, pg, CPS_CACHED);
1352         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
1353                                (const struct lu_env *,
1354                                 const struct cl_page_slice *, int), ioret);
1355         if (anchor) {
1356                 LASSERT(cl_page_is_vmlocked(env, pg));
1357                 LASSERT(pg->cp_sync_io == anchor);
1358                 pg->cp_sync_io = NULL;
1359                 cl_sync_io_note(anchor, ioret);
1360         }
1361
1362         /* Don't assert the page writeback bit here because the lustre file
1363          * may be as a backend of swap space. in this case, the page writeback
1364          * is set by VM, and obvious we shouldn't clear it at all. Fortunately
1365          * this type of pages are all TRANSIENT pages. */
1366         KLASSERT(ergo(pg->cp_type == CPT_CACHEABLE,
1367                       !PageWriteback(cl_page_vmpage(env, pg))));
1368         EXIT;
1369 }
1370 EXPORT_SYMBOL(cl_page_completion);
1371
1372 /**
1373  * Notify layers that transfer formation engine decided to yank this page from
1374  * the cache and to make it a part of a transfer.
1375  *
1376  * \pre  pg->cp_state == CPS_CACHED
1377  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1378  *
1379  * \see cl_page_operations::cpo_make_ready()
1380  */
1381 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
1382                        enum cl_req_type crt)
1383 {
1384         int result;
1385
1386         PINVRNT(env, pg, crt < CRT_NR);
1387
1388         ENTRY;
1389         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
1390                                 (const struct lu_env *,
1391                                  const struct cl_page_slice *));
1392         if (result == 0) {
1393                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
1394                 cl_page_io_start(env, pg, crt);
1395         }
1396         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1397         RETURN(result);
1398 }
1399 EXPORT_SYMBOL(cl_page_make_ready);
1400
1401 /**
1402  * Notify layers that high level io decided to place this page into a cache
1403  * for future transfer.
1404  *
1405  * The layer implementing transfer engine (osc) has to register this page in
1406  * its queues.
1407  *
1408  * \pre  cl_page_is_owned(pg, io)
1409  * \post cl_page_is_owned(pg, io)
1410  *
1411  * \see cl_page_operations::cpo_cache_add()
1412  */
1413 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
1414                       struct cl_page *pg, enum cl_req_type crt)
1415 {
1416         const struct cl_page_slice *scan;
1417         int result = 0;
1418
1419         PINVRNT(env, pg, crt < CRT_NR);
1420         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1421         PINVRNT(env, pg, cl_page_invariant(pg));
1422
1423         ENTRY;
1424
1425         cfs_list_for_each_entry(scan, &pg->cp_layers, cpl_linkage) {
1426                 if (scan->cpl_ops->io[crt].cpo_cache_add == NULL)
1427                         continue;
1428
1429                 result = scan->cpl_ops->io[crt].cpo_cache_add(env, scan, io);
1430                 if (result != 0)
1431                         break;
1432         }
1433         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1434         RETURN(result);
1435 }
1436 EXPORT_SYMBOL(cl_page_cache_add);
1437
1438 /**
1439  * Called if a pge is being written back by kernel's intention.
1440  *
1441  * \pre  cl_page_is_owned(pg, io)
1442  * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT)
1443  *
1444  * \see cl_page_operations::cpo_flush()
1445  */
1446 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
1447                   struct cl_page *pg)
1448 {
1449         int result;
1450
1451         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1452         PINVRNT(env, pg, cl_page_invariant(pg));
1453
1454         ENTRY;
1455
1456         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_flush));
1457
1458         CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result);
1459         RETURN(result);
1460 }
1461 EXPORT_SYMBOL(cl_page_flush);
1462
1463 /**
1464  * Checks whether page is protected by any extent lock is at least required
1465  * mode.
1466  *
1467  * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1468  * \see cl_page_operations::cpo_is_under_lock()
1469  */
1470 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1471                           struct cl_page *page)
1472 {
1473         int rc;
1474
1475         PINVRNT(env, page, cl_page_invariant(page));
1476
1477         ENTRY;
1478         rc = CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1479                             (const struct lu_env *,
1480                              const struct cl_page_slice *, struct cl_io *),
1481                             io);
1482         PASSERT(env, page, rc != 0);
1483         RETURN(rc);
1484 }
1485 EXPORT_SYMBOL(cl_page_is_under_lock);
1486
1487 static int page_prune_cb(const struct lu_env *env, struct cl_io *io,
1488                          struct cl_page *page, void *cbdata)
1489 {
1490         cl_page_own(env, io, page);
1491         cl_page_unmap(env, io, page);
1492         cl_page_discard(env, io, page);
1493         cl_page_disown(env, io, page);
1494         return CLP_GANG_OKAY;
1495 }
1496
1497 /**
1498  * Purges all cached pages belonging to the object \a obj.
1499  */
1500 int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj)
1501 {
1502         struct cl_thread_info   *info;
1503         struct cl_object        *obj = cl_object_top(clobj);
1504         struct cl_io            *io;
1505         int                      result;
1506
1507         ENTRY;
1508         info  = cl_env_info(env);
1509         io    = &info->clt_io;
1510
1511         /*
1512          * initialize the io. This is ugly since we never do IO in this
1513          * function, we just make cl_page_list functions happy. -jay
1514          */
1515         io->ci_obj = obj;
1516         io->ci_ignore_layout = 1;
1517         result = cl_io_init(env, io, CIT_MISC, obj);
1518         if (result != 0) {
1519                 cl_io_fini(env, io);
1520                 RETURN(io->ci_result);
1521         }
1522
1523         do {
1524                 result = cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF,
1525                                              page_prune_cb, NULL);
1526                 if (result == CLP_GANG_RESCHED)
1527                         cfs_cond_resched();
1528         } while (result != CLP_GANG_OKAY);
1529
1530         cl_io_fini(env, io);
1531         RETURN(result);
1532 }
1533 EXPORT_SYMBOL(cl_pages_prune);
1534
1535 /**
1536  * Tells transfer engine that only part of a page is to be transmitted.
1537  *
1538  * \see cl_page_operations::cpo_clip()
1539  */
1540 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1541                   int from, int to)
1542 {
1543         PINVRNT(env, pg, cl_page_invariant(pg));
1544
1545         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1546         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1547                        (const struct lu_env *,
1548                         const struct cl_page_slice *,int, int),
1549                        from, to);
1550 }
1551 EXPORT_SYMBOL(cl_page_clip);
1552
1553 /**
1554  * Prints human readable representation of \a pg to the \a f.
1555  */
1556 void cl_page_header_print(const struct lu_env *env, void *cookie,
1557                           lu_printer_t printer, const struct cl_page *pg)
1558 {
1559         (*printer)(env, cookie,
1560                    "page@%p[%d %p:%lu ^%p_%p %d %d %d %p %p %#x]\n",
1561                    pg, cfs_atomic_read(&pg->cp_ref), pg->cp_obj,
1562                    pg->cp_index, pg->cp_parent, pg->cp_child,
1563                    pg->cp_state, pg->cp_error, pg->cp_type,
1564                    pg->cp_owner, pg->cp_req, pg->cp_flags);
1565 }
1566 EXPORT_SYMBOL(cl_page_header_print);
1567
1568 /**
1569  * Prints human readable representation of \a pg to the \a f.
1570  */
1571 void cl_page_print(const struct lu_env *env, void *cookie,
1572                    lu_printer_t printer, const struct cl_page *pg)
1573 {
1574         struct cl_page *scan;
1575
1576         for (scan = cl_page_top((struct cl_page *)pg);
1577              scan != NULL; scan = scan->cp_child)
1578                 cl_page_header_print(env, cookie, printer, scan);
1579         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1580                        (const struct lu_env *env,
1581                         const struct cl_page_slice *slice,
1582                         void *cookie, lu_printer_t p), cookie, printer);
1583         (*printer)(env, cookie, "end page@%p\n", pg);
1584 }
1585 EXPORT_SYMBOL(cl_page_print);
1586
1587 /**
1588  * Cancel a page which is still in a transfer.
1589  */
1590 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1591 {
1592         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1593                               (const struct lu_env *,
1594                                const struct cl_page_slice *));
1595 }
1596 EXPORT_SYMBOL(cl_page_cancel);
1597
1598 /**
1599  * Converts a byte offset within object \a obj into a page index.
1600  */
1601 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1602 {
1603         /*
1604          * XXX for now.
1605          */
1606         return (loff_t)idx << CFS_PAGE_SHIFT;
1607 }
1608 EXPORT_SYMBOL(cl_offset);
1609
1610 /**
1611  * Converts a page index into a byte offset within object \a obj.
1612  */
1613 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1614 {
1615         /*
1616          * XXX for now.
1617          */
1618         return offset >> CFS_PAGE_SHIFT;
1619 }
1620 EXPORT_SYMBOL(cl_index);
1621
1622 int cl_page_size(const struct cl_object *obj)
1623 {
1624         return 1 << CFS_PAGE_SHIFT;
1625 }
1626 EXPORT_SYMBOL(cl_page_size);
1627
1628 /**
1629  * Adds page slice to the compound page.
1630  *
1631  * This is called by cl_object_operations::coo_page_init() methods to add a
1632  * per-layer state to the page. New state is added at the end of
1633  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1634  *
1635  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1636  */
1637 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1638                        struct cl_object *obj,
1639                        const struct cl_page_operations *ops)
1640 {
1641         ENTRY;
1642         cfs_list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1643         slice->cpl_obj  = obj;
1644         slice->cpl_ops  = ops;
1645         slice->cpl_page = page;
1646         EXIT;
1647 }
1648 EXPORT_SYMBOL(cl_page_slice_add);
1649
1650 int  cl_page_init(void)
1651 {
1652         return lu_kmem_init(cl_page_caches);
1653 }
1654
1655 void cl_page_fini(void)
1656 {
1657         lu_kmem_fini(cl_page_caches);
1658 }