lustre/obdclass/cl_page.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011 Whamcloud, Inc.
  33  *
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * Client Lustre Page.
  40  *
  41  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_CLASS
  45 #ifndef EXPORT_SYMTAB
  46 # define EXPORT_SYMTAB
  47 #endif
  48
  49 #include <libcfs/libcfs.h>
  50 #include <obd_class.h>
  51 #include <obd_support.h>
  52 #include <libcfs/list.h>
  53
  54 #include <cl_object.h>
  55 #include "cl_internal.h"
  56
  57 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
  58                             int radix);
  59
  60 static cfs_mem_cache_t      *cl_page_kmem = NULL;
  61
  62 static struct lu_kmem_descr cl_page_caches[] = {
  63         {
  64                 .ckd_cache = &cl_page_kmem,
  65                 .ckd_name  = "cl_page_kmem",
  66                 .ckd_size  = sizeof (struct cl_page)
  67         },
  68         {
  69                 .ckd_cache = NULL
  70         }
  71 };
  72
  73 #ifdef LIBCFS_DEBUG
  74 # define PASSERT(env, page, expr)                                       \
  75   do {                                                                    \
  76           if (unlikely(!(expr))) {                                      \
  77                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  78                   LASSERT(0);                                           \
  79           }                                                             \
  80   } while (0)
  81 #else /* !LIBCFS_DEBUG */
  82 # define PASSERT(env, page, exp) \
  83         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  84 #endif /* !LIBCFS_DEBUG */
  85
  86 #ifdef INVARIANT_CHECK
  87 # define PINVRNT(env, page, expr)                                       \
  88   do {                                                                    \
  89           if (unlikely(!(expr))) {                                      \
  90                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  91                   LINVRNT(0);                                           \
  92           }                                                             \
  93   } while (0)
  94 #else /* !INVARIANT_CHECK */
  95 # define PINVRNT(env, page, exp) \
  96         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  97 #endif /* !INVARIANT_CHECK */
  98
  99 /**
 100  * Internal version of cl_page_top, it should be called with page referenced,
 101  * or coh_page_guard held.
 102  */
 103 static struct cl_page *cl_page_top_trusted(struct cl_page *page)
 104 {
 105         while (page->cp_parent != NULL)
 106                 page = page->cp_parent;
 107         return page;
 108 }
 109
 110 /**
 111  * Internal version of cl_page_get().
 112  *
 113  * This function can be used to obtain initial reference to previously
 114  * unreferenced cached object. It can be called only if concurrent page
 115  * reclamation is somehow prevented, e.g., by locking page radix-tree
 116  * (cl_object_header::hdr->coh_page_guard), or by keeping a lock on a VM page,
 117  * associated with \a page.
 118  *
 119  * Use with care! Not exported.
 120  */
 121 static void cl_page_get_trust(struct cl_page *page)
 122 {
 123         /*
 124          * Checkless version for trusted users.
 125          */
 126         if (cfs_atomic_inc_return(&page->cp_ref) == 1)
 127                 cfs_atomic_inc(&cl_object_site(page->cp_obj)->cs_pages.cs_busy);
 128 }
 129
 130 /**
 131  * Returns a slice within a page, corresponding to the given layer in the
 132  * device stack.
 133  *
 134  * \see cl_lock_at()
 135  */
 136 static const struct cl_page_slice *
 137 cl_page_at_trusted(const struct cl_page *page,
 138                    const struct lu_device_type *dtype)
 139 {
 140         const struct cl_page_slice *slice;
 141
 142 #ifdef INVARIANT_CHECK
 143         struct cl_object_header *ch = cl_object_header(page->cp_obj);
 144
 145         if (!cfs_atomic_read(&page->cp_ref))
 146                 LASSERT_SPIN_LOCKED(&ch->coh_page_guard);
 147 #endif
 148         ENTRY;
 149
 150         page = cl_page_top_trusted((struct cl_page *)page);
 151         do {
 152                 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 153                         if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
 154                                 RETURN(slice);
 155                 }
 156                 page = page->cp_child;
 157         } while (page != NULL);
 158         RETURN(NULL);
 159 }
 160
 161 /**
 162  * Returns a page with given index in the given object, or NULL if no page is
 163  * found. Acquires a reference on \a page.
 164  *
 165  * Locking: called under cl_object_header::coh_page_guard spin-lock.
 166  */
 167 struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index)
 168 {
 169         struct cl_page *page;
 170
 171         LASSERT_SPIN_LOCKED(&hdr->coh_page_guard);
 172
 173         page = radix_tree_lookup(&hdr->coh_tree, index);
 174         if (page != NULL) {
 175                 cl_page_get_trust(page);
 176         }
 177         return page;
 178 }
 179 EXPORT_SYMBOL(cl_page_lookup);
 180
 181 /**
 182  * Returns a list of pages by a given [start, end] of \a obj.
 183  *
 184  * \param resched If not NULL, then we give up before hogging CPU for too
 185  * long and set *resched = 1, in that case caller should implement a retry
 186  * logic.
 187  *
 188  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
 189  * crucial in the face of [offset, EOF] locks.
 190  *
 191  * Return at least one page in @queue unless there is no covered page.
 192  */
 193 int cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj,
 194                         struct cl_io *io, pgoff_t start, pgoff_t end,
 195                         struct cl_page_list *queue)
 196 {
 197         struct cl_object_header *hdr;
 198         struct cl_page          *page;
 199         struct cl_page         **pvec;
 200         const struct cl_page_slice  *slice;
 201         const struct lu_device_type *dtype;
 202         pgoff_t                  idx;
 203         unsigned int             nr;
 204         unsigned int             i;
 205         unsigned int             j;
 206         int                      res = CLP_GANG_OKAY;
 207         int                      tree_lock = 1;
 208         ENTRY;
 209
 210         idx = start;
 211         hdr = cl_object_header(obj);
 212         pvec = cl_env_info(env)->clt_pvec;
 213         dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type;
 214         cfs_spin_lock(&hdr->coh_page_guard);
 215         while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec,
 216                                             idx, CLT_PVEC_SIZE)) > 0) {
 217                 int end_of_region = 0;
 218                 idx = pvec[nr - 1]->cp_index + 1;
 219                 for (i = 0, j = 0; i < nr; ++i) {
 220                         page = pvec[i];
 221                         pvec[i] = NULL;
 222
 223                         LASSERT(page->cp_type == CPT_CACHEABLE);
 224                         if (page->cp_index > end) {
 225                                 end_of_region = 1;
 226                                 break;
 227                         }
 228                         if (page->cp_state == CPS_FREEING)
 229                                 continue;
 230
 231                         slice = cl_page_at_trusted(page, dtype);
 232                         /*
 233                          * Pages for lsm-less file has no underneath sub-page
 234                          * for osc, in case of ...
 235                          */
 236                         PASSERT(env, page, slice != NULL);
 237
 238                         page = slice->cpl_page;
 239                         /*
 240                          * Can safely call cl_page_get_trust() under
 241                          * radix-tree spin-lock.
 242                          *
 243                          * XXX not true, because @page is from object another
 244                          * than @hdr and protected by different tree lock.
 245                          */
 246                         cl_page_get_trust(page);
 247                         lu_ref_add_atomic(&page->cp_reference,
 248                                           "page_list", cfs_current());
 249                         pvec[j++] = page;
 250                 }
 251
 252                 /*
 253                  * Here a delicate locking dance is performed. Current thread
 254                  * holds a reference to a page, but has to own it before it
 255                  * can be placed into queue. Owning implies waiting, so
 256                  * radix-tree lock is to be released. After a wait one has to
 257                  * check that pages weren't truncated (cl_page_own() returns
 258                  * error in the latter case).
 259                  */
 260                 cfs_spin_unlock(&hdr->coh_page_guard);
 261                 tree_lock = 0;
 262
 263                 for (i = 0; i < j; ++i) {
 264                         page = pvec[i];
 265                         if (res == CLP_GANG_OKAY) {
 266                                 typeof(cl_page_own) *page_own;
 267
 268                                 page_own = queue->pl_nr ?
 269                                            cl_page_own_try : cl_page_own;
 270                                 if (page_own(env, io, page) == 0) {
 271                                         cl_page_list_add(queue, page);
 272                                 } else if (page->cp_state != CPS_FREEING) {
 273                                         /* cl_page_own() won't fail unless
 274                                          * the page is being freed. */
 275                                         LASSERT(queue->pl_nr != 0);
 276                                         res = CLP_GANG_AGAIN;
 277                                 }
 278                         }
 279                         lu_ref_del(&page->cp_reference,
 280                                    "page_list", cfs_current());
 281                         cl_page_put(env, page);
 282                 }
 283                 if (nr < CLT_PVEC_SIZE || end_of_region)
 284                         break;
 285
 286                 /* if the number of pages is zero, this will mislead the caller
 287                  * that there is no page any more. */
 288                 if (queue->pl_nr && cfs_need_resched())
 289                         res = CLP_GANG_RESCHED;
 290                 if (res != CLP_GANG_OKAY)
 291                         break;
 292
 293                 cfs_spin_lock(&hdr->coh_page_guard);
 294                 tree_lock = 1;
 295         }
 296         if (tree_lock)
 297                 cfs_spin_unlock(&hdr->coh_page_guard);
 298         RETURN(res);
 299 }
 300 EXPORT_SYMBOL(cl_page_gang_lookup);
 301
 302 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
 303 {
 304         struct cl_object *obj  = page->cp_obj;
 305         struct cl_site   *site = cl_object_site(obj);
 306
 307         PASSERT(env, page, cfs_list_empty(&page->cp_batch));
 308         PASSERT(env, page, page->cp_owner == NULL);
 309         PASSERT(env, page, page->cp_req == NULL);
 310         PASSERT(env, page, page->cp_parent == NULL);
 311         PASSERT(env, page, page->cp_state == CPS_FREEING);
 312
 313         ENTRY;
 314         cfs_might_sleep();
 315         while (!cfs_list_empty(&page->cp_layers)) {
 316                 struct cl_page_slice *slice;
 317
 318                 slice = cfs_list_entry(page->cp_layers.next,
 319                                        struct cl_page_slice, cpl_linkage);
 320                 cfs_list_del_init(page->cp_layers.next);
 321                 slice->cpl_ops->cpo_fini(env, slice);
 322         }
 323         cfs_atomic_dec(&site->cs_pages.cs_total);
 324
 325 #ifdef LUSTRE_PAGESTATE_TRACKING
 326         cfs_atomic_dec(&site->cs_pages_state[page->cp_state]);
 327 #endif
 328         lu_object_ref_del_at(&obj->co_lu, page->cp_obj_ref, "cl_page", page);
 329         cl_object_put(env, obj);
 330         lu_ref_fini(&page->cp_reference);
 331         OBD_SLAB_FREE_PTR(page, cl_page_kmem);
 332         EXIT;
 333 }
 334
 335 /**
 336  * Helper function updating page state. This is the only place in the code
 337  * where cl_page::cp_state field is mutated.
 338  */
 339 static inline void cl_page_state_set_trust(struct cl_page *page,
 340                                            enum cl_page_state state)
 341 {
 342         /* bypass const. */
 343         *(enum cl_page_state *)&page->cp_state = state;
 344 }
 345
 346 static int cl_page_alloc(const struct lu_env *env, struct cl_object *o,
 347                          pgoff_t ind, struct page *vmpage,
 348                          enum cl_page_type type, struct cl_page **out)
 349 {
 350         struct cl_page          *page;
 351         struct cl_page          *err  = NULL;
 352         struct lu_object_header *head;
 353         struct cl_site          *site = cl_object_site(o);
 354         int                      result;
 355
 356         ENTRY;
 357         result = +1;
 358         OBD_SLAB_ALLOC_PTR_GFP(page, cl_page_kmem, CFS_ALLOC_IO);
 359         if (page != NULL) {
 360                 cfs_atomic_set(&page->cp_ref, 1);
 361                 page->cp_obj = o;
 362                 cl_object_get(o);
 363                 page->cp_obj_ref = lu_object_ref_add(&o->co_lu,
 364                                                      "cl_page", page);
 365                 page->cp_index = ind;
 366                 cl_page_state_set_trust(page, CPS_CACHED);
 367                 page->cp_type = type;
 368                 CFS_INIT_LIST_HEAD(&page->cp_layers);
 369                 CFS_INIT_LIST_HEAD(&page->cp_batch);
 370                 CFS_INIT_LIST_HEAD(&page->cp_flight);
 371                 cfs_mutex_init(&page->cp_mutex);
 372                 lu_ref_init(&page->cp_reference);
 373                 head = o->co_lu.lo_header;
 374                 cfs_list_for_each_entry(o, &head->loh_layers,
 375                                         co_lu.lo_linkage) {
 376                         if (o->co_ops->coo_page_init != NULL) {
 377                                 err = o->co_ops->coo_page_init(env, o,
 378                                                                page, vmpage);
 379                                 if (err != NULL) {
 380                                         cl_page_delete0(env, page, 0);
 381                                         cl_page_free(env, page);
 382                                         page = err;
 383                                         break;
 384                                 }
 385                         }
 386                 }
 387                 if (err == NULL) {
 388                         cfs_atomic_inc(&site->cs_pages.cs_busy);
 389                         cfs_atomic_inc(&site->cs_pages.cs_total);
 390
 391 #ifdef LUSTRE_PAGESTATE_TRACKING
 392                         cfs_atomic_inc(&site->cs_pages_state[CPS_CACHED]);
 393 #endif
 394                         cfs_atomic_inc(&site->cs_pages.cs_created);
 395                         result = 0;
 396                 }
 397         } else
 398                 page = ERR_PTR(-ENOMEM);
 399         *out = page;
 400         RETURN(result);
 401 }
 402
 403 /**
 404  * Returns a cl_page with index \a idx at the object \a o, and associated with
 405  * the VM page \a vmpage.
 406  *
 407  * This is the main entry point into the cl_page caching interface. First, a
 408  * cache (implemented as a per-object radix tree) is consulted. If page is
 409  * found there, it is returned immediately. Otherwise new page is allocated
 410  * and returned. In any case, additional reference to page is acquired.
 411  *
 412  * \see cl_object_find(), cl_lock_find()
 413  */
 414 static struct cl_page *cl_page_find0(const struct lu_env *env,
 415                                      struct cl_object *o,
 416                                      pgoff_t idx, struct page *vmpage,
 417                                      enum cl_page_type type,
 418                                      struct cl_page *parent)
 419 {
 420         struct cl_page          *page = NULL;
 421         struct cl_page          *ghost = NULL;
 422         struct cl_object_header *hdr;
 423         struct cl_site          *site = cl_object_site(o);
 424         int err;
 425
 426         LINVRNT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 427         cfs_might_sleep();
 428
 429         ENTRY;
 430
 431         hdr = cl_object_header(o);
 432         cfs_atomic_inc(&site->cs_pages.cs_lookup);
 433
 434         CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
 435                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 436         /* fast path. */
 437         if (type == CPT_CACHEABLE) {
 438                 /*
 439                  * cl_vmpage_page() can be called here without any locks as
 440                  *
 441                  *     - "vmpage" is locked (which prevents ->private from
 442                  *       concurrent updates), and
 443                  *
 444                  *     - "o" cannot be destroyed while current thread holds a
 445                  *       reference on it.
 446                  */
 447                 page = cl_vmpage_page(vmpage, o);
 448                 PINVRNT(env, page,
 449                         ergo(page != NULL,
 450                              cl_page_vmpage(env, page) == vmpage &&
 451                              (void *)radix_tree_lookup(&hdr->coh_tree,
 452                                                        idx) == page));
 453         }
 454
 455         if (page != NULL) {
 456                 cfs_atomic_inc(&site->cs_pages.cs_hit);
 457                 RETURN(page);
 458         }
 459
 460         /* allocate and initialize cl_page */
 461         err = cl_page_alloc(env, o, idx, vmpage, type, &page);
 462         if (err != 0)
 463                 RETURN(page);
 464
 465         if (type == CPT_TRANSIENT) {
 466                 if (parent) {
 467                         LASSERT(page->cp_parent == NULL);
 468                         page->cp_parent = parent;
 469                         parent->cp_child = page;
 470                 }
 471                 RETURN(page);
 472         }
 473
 474         /*
 475          * XXX optimization: use radix_tree_preload() here, and change tree
 476          * gfp mask to GFP_KERNEL in cl_object_header_init().
 477          */
 478         cfs_spin_lock(&hdr->coh_page_guard);
 479         err = radix_tree_insert(&hdr->coh_tree, idx, page);
 480         if (err != 0) {
 481                 ghost = page;
 482                 /*
 483                  * Noted by Jay: a lock on \a vmpage protects cl_page_find()
 484                  * from this race, but
 485                  *
 486                  *     0. it's better to have cl_page interface "locally
 487                  *     consistent" so that its correctness can be reasoned
 488                  *     about without appealing to the (obscure world of) VM
 489                  *     locking.
 490                  *
 491                  *     1. handling this race allows ->coh_tree to remain
 492                  *     consistent even when VM locking is somehow busted,
 493                  *     which is very useful during diagnosing and debugging.
 494                  */
 495                 page = ERR_PTR(err);
 496                 CL_PAGE_DEBUG(D_ERROR, env, ghost,
 497                               "fail to insert into radix tree: %d\n", err);
 498         } else {
 499                 if (parent) {
 500                         LASSERT(page->cp_parent == NULL);
 501                         page->cp_parent = parent;
 502                         parent->cp_child = page;
 503                 }
 504                 hdr->coh_pages++;
 505         }
 506         cfs_spin_unlock(&hdr->coh_page_guard);
 507
 508         if (unlikely(ghost != NULL)) {
 509                 cfs_atomic_dec(&site->cs_pages.cs_busy);
 510                 cl_page_delete0(env, ghost, 0);
 511                 cl_page_free(env, ghost);
 512         }
 513         RETURN(page);
 514 }
 515
 516 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o,
 517                              pgoff_t idx, struct page *vmpage,
 518                              enum cl_page_type type)
 519 {
 520         return cl_page_find0(env, o, idx, vmpage, type, NULL);
 521 }
 522 EXPORT_SYMBOL(cl_page_find);
 523
 524
 525 struct cl_page *cl_page_find_sub(const struct lu_env *env, struct cl_object *o,
 526                                  pgoff_t idx, struct page *vmpage,
 527                                  struct cl_page *parent)
 528 {
 529         return cl_page_find0(env, o, idx, vmpage, parent->cp_type, parent);
 530 }
 531 EXPORT_SYMBOL(cl_page_find_sub);
 532
 533 static inline int cl_page_invariant(const struct cl_page *pg)
 534 {
 535         struct cl_object_header *header;
 536         struct cl_page          *parent;
 537         struct cl_page          *child;
 538         struct cl_io            *owner;
 539
 540         /*
 541          * Page invariant is protected by a VM lock.
 542          */
 543         LINVRNT(cl_page_is_vmlocked(NULL, pg));
 544
 545         header = cl_object_header(pg->cp_obj);
 546         parent = pg->cp_parent;
 547         child  = pg->cp_child;
 548         owner  = pg->cp_owner;
 549
 550         return cfs_atomic_read(&pg->cp_ref) > 0 &&
 551                 ergo(parent != NULL, parent->cp_child == pg) &&
 552                 ergo(child != NULL, child->cp_parent == pg) &&
 553                 ergo(child != NULL, pg->cp_obj != child->cp_obj) &&
 554                 ergo(parent != NULL, pg->cp_obj != parent->cp_obj) &&
 555                 ergo(owner != NULL && parent != NULL,
 556                      parent->cp_owner == pg->cp_owner->ci_parent) &&
 557                 ergo(owner != NULL && child != NULL,
 558                      child->cp_owner->ci_parent == owner) &&
 559                 /*
 560                  * Either page is early in initialization (has neither child
 561                  * nor parent yet), or it is in the object radix tree.
 562                  */
 563                 ergo(pg->cp_state < CPS_FREEING,
 564                      (void *)radix_tree_lookup(&header->coh_tree,
 565                                                pg->cp_index) == pg ||
 566                      (child == NULL && parent == NULL));
 567 }
 568
 569 static void cl_page_state_set0(const struct lu_env *env,
 570                                struct cl_page *page, enum cl_page_state state)
 571 {
 572         enum cl_page_state old;
 573 #ifdef LUSTRE_PAGESTATE_TRACKING
 574         struct cl_site *site = cl_object_site(page->cp_obj);
 575 #endif
 576
 577         /*
 578          * Matrix of allowed state transitions [old][new], for sanity
 579          * checking.
 580          */
 581         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 582                 [CPS_CACHED] = {
 583                         [CPS_CACHED]  = 0,
 584                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 585                         [CPS_PAGEIN]  = 0,
 586                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 587                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 588                 },
 589                 [CPS_OWNED] = {
 590                         [CPS_CACHED]  = 1, /* release to the cache */
 591                         [CPS_OWNED]   = 0,
 592                         [CPS_PAGEIN]  = 1, /* start read immediately */
 593                         [CPS_PAGEOUT] = 1, /* start write immediately */
 594                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 595                 },
 596                 [CPS_PAGEIN] = {
 597                         [CPS_CACHED]  = 1, /* io completion */
 598                         [CPS_OWNED]   = 0,
 599                         [CPS_PAGEIN]  = 0,
 600                         [CPS_PAGEOUT] = 0,
 601                         [CPS_FREEING] = 0,
 602                 },
 603                 [CPS_PAGEOUT] = {
 604                         [CPS_CACHED]  = 1, /* io completion */
 605                         [CPS_OWNED]   = 0,
 606                         [CPS_PAGEIN]  = 0,
 607                         [CPS_PAGEOUT] = 0,
 608                         [CPS_FREEING] = 0,
 609                 },
 610                 [CPS_FREEING] = {
 611                         [CPS_CACHED]  = 0,
 612                         [CPS_OWNED]   = 0,
 613                         [CPS_PAGEIN]  = 0,
 614                         [CPS_PAGEOUT] = 0,
 615                         [CPS_FREEING] = 0,
 616                 }
 617         };
 618
 619         ENTRY;
 620         old = page->cp_state;
 621         PASSERT(env, page, allowed_transitions[old][state]);
 622         CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
 623         for (; page != NULL; page = page->cp_child) {
 624                 PASSERT(env, page, page->cp_state == old);
 625                 PASSERT(env, page,
 626                         equi(state == CPS_OWNED, page->cp_owner != NULL));
 627
 628 #ifdef LUSTRE_PAGESTATE_TRACKING
 629                 cfs_atomic_dec(&site->cs_pages_state[page->cp_state]);
 630                 cfs_atomic_inc(&site->cs_pages_state[state]);
 631 #endif
 632                 cl_page_state_set_trust(page, state);
 633         }
 634         EXIT;
 635 }
 636
 637 static void cl_page_state_set(const struct lu_env *env,
 638                               struct cl_page *page, enum cl_page_state state)
 639 {
 640         PINVRNT(env, page, cl_page_invariant(page));
 641         cl_page_state_set0(env, page, state);
 642 }
 643
 644 /**
 645  * Acquires an additional reference to a page.
 646  *
 647  * This can be called only by caller already possessing a reference to \a
 648  * page.
 649  *
 650  * \see cl_object_get(), cl_lock_get().
 651  */
 652 void cl_page_get(struct cl_page *page)
 653 {
 654         ENTRY;
 655         LASSERT(page->cp_state != CPS_FREEING);
 656         cl_page_get_trust(page);
 657         EXIT;
 658 }
 659 EXPORT_SYMBOL(cl_page_get);
 660
 661 /**
 662  * Releases a reference to a page.
 663  *
 664  * When last reference is released, page is returned to the cache, unless it
 665  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 666  * destroyed.
 667  *
 668  * \see cl_object_put(), cl_lock_put().
 669  */
 670 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 671 {
 672         struct cl_object_header *hdr;
 673         struct cl_site *site = cl_object_site(page->cp_obj);
 674
 675         PASSERT(env, page, cfs_atomic_read(&page->cp_ref) > !!page->cp_parent);
 676
 677         ENTRY;
 678         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 679                        cfs_atomic_read(&page->cp_ref));
 680
 681         hdr = cl_object_header(cl_object_top(page->cp_obj));
 682         if (cfs_atomic_dec_and_lock(&page->cp_ref, &hdr->coh_page_guard)) {
 683                 cfs_atomic_dec(&site->cs_pages.cs_busy);
 684                 /* We're going to access the page w/o a reference, but it's
 685                  * ok because we have grabbed the lock coh_page_guard, which
 686                  * means nobody is able to free this page behind us.
 687                  */
 688                 if (page->cp_state == CPS_FREEING) {
 689                         /* We drop the page reference and check the page state
 690                          * inside the coh_page_guard. So that if it gets here,
 691                          * it is the REALLY last reference to this page.
 692                          */
 693                         cfs_spin_unlock(&hdr->coh_page_guard);
 694
 695                         LASSERT(cfs_atomic_read(&page->cp_ref) == 0);
 696                         PASSERT(env, page, page->cp_owner == NULL);
 697                         PASSERT(env, page, cfs_list_empty(&page->cp_batch));
 698                         /*
 699                          * Page is no longer reachable by other threads. Tear
 700                          * it down.
 701                          */
 702                         cl_page_free(env, page);
 703
 704                         EXIT;
 705                         return;
 706                 }
 707                 cfs_spin_unlock(&hdr->coh_page_guard);
 708         }
 709
 710         EXIT;
 711 }
 712 EXPORT_SYMBOL(cl_page_put);
 713
 714 /**
 715  * Returns a VM page associated with a given cl_page.
 716  */
 717 cfs_page_t *cl_page_vmpage(const struct lu_env *env, struct cl_page *page)
 718 {
 719         const struct cl_page_slice *slice;
 720
 721         /*
 722          * Find uppermost layer with ->cpo_vmpage() method, and return its
 723          * result.
 724          */
 725         page = cl_page_top(page);
 726         do {
 727                 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 728                         if (slice->cpl_ops->cpo_vmpage != NULL)
 729                                 RETURN(slice->cpl_ops->cpo_vmpage(env, slice));
 730                 }
 731                 page = page->cp_child;
 732         } while (page != NULL);
 733         LBUG(); /* ->cpo_vmpage() has to be defined somewhere in the stack */
 734 }
 735 EXPORT_SYMBOL(cl_page_vmpage);
 736
 737 /**
 738  * Returns a cl_page associated with a VM page, and given cl_object.
 739  */
 740 struct cl_page *cl_vmpage_page(cfs_page_t *vmpage, struct cl_object *obj)
 741 {
 742         struct cl_page *page;
 743         struct cl_object_header *hdr;
 744
 745         ENTRY;
 746         KLASSERT(PageLocked(vmpage));
 747
 748         /*
 749          * NOTE: absence of races and liveness of data are guaranteed by page
 750          *       lock on a "vmpage". That works because object destruction has
 751          *       bottom-to-top pass.
 752          */
 753
 754         /*
 755          * This loop assumes that ->private points to the top-most page. This
 756          * can be rectified easily.
 757          */
 758         hdr = cl_object_header(cl_object_top(obj));
 759         cfs_spin_lock(&hdr->coh_page_guard);
 760         for (page = (void *)vmpage->private;
 761              page != NULL; page = page->cp_child) {
 762                 if (cl_object_same(page->cp_obj, obj)) {
 763                         cl_page_get_trust(page);
 764                         break;
 765                 }
 766         }
 767         cfs_spin_unlock(&hdr->coh_page_guard);
 768         LASSERT(ergo(page, page->cp_type == CPT_CACHEABLE));
 769         RETURN(page);
 770 }
 771 EXPORT_SYMBOL(cl_vmpage_page);
 772
 773 /**
 774  * Returns the top-page for a given page.
 775  *
 776  * \see cl_object_top(), cl_io_top()
 777  */
 778 struct cl_page *cl_page_top(struct cl_page *page)
 779 {
 780         return cl_page_top_trusted(page);
 781 }
 782 EXPORT_SYMBOL(cl_page_top);
 783
 784 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 785                                        const struct lu_device_type *dtype)
 786 {
 787         return cl_page_at_trusted(page, dtype);
 788 }
 789 EXPORT_SYMBOL(cl_page_at);
 790
 791 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 792
 793 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)                   \
 794 ({                                                                      \
 795         const struct lu_env        *__env  = (_env);                    \
 796         struct cl_page             *__page = (_page);                   \
 797         const struct cl_page_slice *__scan;                             \
 798         int                         __result;                           \
 799         ptrdiff_t                   __op   = (_op);                     \
 800         int                       (*__method)_proto;                    \
 801                                                                         \
 802         __result = 0;                                                   \
 803         __page = cl_page_top(__page);                                   \
 804         do {                                                            \
 805                 cfs_list_for_each_entry(__scan, &__page->cp_layers,     \
 806                                         cpl_linkage) {                  \
 807                         __method = *(void **)((char *)__scan->cpl_ops + \
 808                                               __op);                    \
 809                         if (__method != NULL) {                         \
 810                                 __result = (*__method)(__env, __scan,   \
 811                                                        ## __VA_ARGS__); \
 812                                 if (__result != 0)                      \
 813                                         break;                          \
 814                         }                                               \
 815                 }                                                       \
 816                 __page = __page->cp_child;                              \
 817         } while (__page != NULL && __result == 0);                      \
 818         if (__result > 0)                                               \
 819                 __result = 0;                                           \
 820         __result;                                                       \
 821 })
 822
 823 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 824 do {                                                                    \
 825         const struct lu_env        *__env  = (_env);                    \
 826         struct cl_page             *__page = (_page);                   \
 827         const struct cl_page_slice *__scan;                             \
 828         ptrdiff_t                   __op   = (_op);                     \
 829         void                      (*__method)_proto;                    \
 830                                                                         \
 831         __page = cl_page_top(__page);                                   \
 832         do {                                                            \
 833                 cfs_list_for_each_entry(__scan, &__page->cp_layers,     \
 834                                         cpl_linkage) {                  \
 835                         __method = *(void **)((char *)__scan->cpl_ops + \
 836                                               __op);                    \
 837                         if (__method != NULL)                           \
 838                                 (*__method)(__env, __scan,              \
 839                                             ## __VA_ARGS__);            \
 840                 }                                                       \
 841                 __page = __page->cp_child;                              \
 842         } while (__page != NULL);                                       \
 843 } while (0)
 844
 845 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)               \
 846 do {                                                                        \
 847         const struct lu_env        *__env  = (_env);                        \
 848         struct cl_page             *__page = (_page);                       \
 849         const struct cl_page_slice *__scan;                                 \
 850         ptrdiff_t                   __op   = (_op);                         \
 851         void                      (*__method)_proto;                        \
 852                                                                             \
 853         /* get to the bottom page. */                                       \
 854         while (__page->cp_child != NULL)                                    \
 855                 __page = __page->cp_child;                                  \
 856         do {                                                                \
 857                 cfs_list_for_each_entry_reverse(__scan, &__page->cp_layers, \
 858                                                 cpl_linkage) {              \
 859                         __method = *(void **)((char *)__scan->cpl_ops +     \
 860                                               __op);                        \
 861                         if (__method != NULL)                               \
 862                                 (*__method)(__env, __scan,                  \
 863                                             ## __VA_ARGS__);                \
 864                 }                                                           \
 865                 __page = __page->cp_parent;                                 \
 866         } while (__page != NULL);                                           \
 867 } while (0)
 868
 869 static int cl_page_invoke(const struct lu_env *env,
 870                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 871
 872 {
 873         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 874         ENTRY;
 875         RETURN(CL_PAGE_INVOKE(env, page, op,
 876                               (const struct lu_env *,
 877                                const struct cl_page_slice *, struct cl_io *),
 878                               io));
 879 }
 880
 881 static void cl_page_invoid(const struct lu_env *env,
 882                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 883
 884 {
 885         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 886         ENTRY;
 887         CL_PAGE_INVOID(env, page, op,
 888                        (const struct lu_env *,
 889                         const struct cl_page_slice *, struct cl_io *), io);
 890         EXIT;
 891 }
 892
 893 static void cl_page_owner_clear(struct cl_page *page)
 894 {
 895         ENTRY;
 896         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 897                 if (page->cp_owner != NULL) {
 898                         LASSERT(page->cp_owner->ci_owned_nr > 0);
 899                         page->cp_owner->ci_owned_nr--;
 900                         page->cp_owner = NULL;
 901                         page->cp_task = NULL;
 902                 }
 903         }
 904         EXIT;
 905 }
 906
 907 static void cl_page_owner_set(struct cl_page *page)
 908 {
 909         ENTRY;
 910         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 911                 LASSERT(page->cp_owner != NULL);
 912                 page->cp_owner->ci_owned_nr++;
 913         }
 914         EXIT;
 915 }
 916
 917 void cl_page_disown0(const struct lu_env *env,
 918                      struct cl_io *io, struct cl_page *pg)
 919 {
 920         enum cl_page_state state;
 921
 922         ENTRY;
 923         state = pg->cp_state;
 924         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 925         PINVRNT(env, pg, cl_page_invariant(pg));
 926         cl_page_owner_clear(pg);
 927
 928         if (state == CPS_OWNED)
 929                 cl_page_state_set(env, pg, CPS_CACHED);
 930         /*
 931          * Completion call-backs are executed in the bottom-up order, so that
 932          * uppermost layer (llite), responsible for VFS/VM interaction runs
 933          * last and can release locks safely.
 934          */
 935         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 936                                (const struct lu_env *,
 937                                 const struct cl_page_slice *, struct cl_io *),
 938                                io);
 939         EXIT;
 940 }
 941
 942 /**
 943  * returns true, iff page is owned by the given io.
 944  */
 945 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 946 {
 947         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 948         ENTRY;
 949         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == io);
 950 }
 951 EXPORT_SYMBOL(cl_page_is_owned);
 952
 953 /**
 954  * Try to own a page by IO.
 955  *
 956  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 957  * into cl_page_state::CPS_OWNED state.
 958  *
 959  * \pre  !cl_page_is_owned(pg, io)
 960  * \post result == 0 iff cl_page_is_owned(pg, io)
 961  *
 962  * \retval 0   success
 963  *
 964  * \retval -ve failure, e.g., page was destroyed (and landed in
 965  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 966  *             or, page was owned by another thread, or in IO.
 967  *
 968  * \see cl_page_disown()
 969  * \see cl_page_operations::cpo_own()
 970  * \see cl_page_own_try()
 971  * \see cl_page_own
 972  */
 973 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 974                         struct cl_page *pg, int nonblock)
 975 {
 976         int result;
 977
 978         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 979
 980         ENTRY;
 981         pg = cl_page_top(pg);
 982         io = cl_io_top(io);
 983
 984         if (pg->cp_state == CPS_FREEING) {
 985                 result = -ENOENT;
 986         } else {
 987                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 988                                         (const struct lu_env *,
 989                                          const struct cl_page_slice *,
 990                                          struct cl_io *, int),
 991                                         io, nonblock);
 992                 if (result == 0) {
 993                         PASSERT(env, pg, pg->cp_owner == NULL);
 994                         PASSERT(env, pg, pg->cp_req == NULL);
 995                         pg->cp_owner = io;
 996                         pg->cp_task  = current;
 997                         cl_page_owner_set(pg);
 998                         if (pg->cp_state != CPS_FREEING) {
 999                                 cl_page_state_set(env, pg, CPS_OWNED);
1000                         } else {
1001                                 cl_page_disown0(env, io, pg);
1002                                 result = -ENOENT;
1003                         }
1004                 }
1005         }
1006         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
1007         RETURN(result);
1008 }
1009
1010 /**
1011  * Own a page, might be blocked.
1012  *
1013  * \see cl_page_own0()
1014  */
1015 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
1016 {
1017         return cl_page_own0(env, io, pg, 0);
1018 }
1019 EXPORT_SYMBOL(cl_page_own);
1020
1021 /**
1022  * Nonblock version of cl_page_own().
1023  *
1024  * \see cl_page_own0()
1025  */
1026 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
1027                     struct cl_page *pg)
1028 {
1029         return cl_page_own0(env, io, pg, 1);
1030 }
1031 EXPORT_SYMBOL(cl_page_own_try);
1032
1033
1034 /**
1035  * Assume page ownership.
1036  *
1037  * Called when page is already locked by the hosting VM.
1038  *
1039  * \pre !cl_page_is_owned(pg, io)
1040  * \post cl_page_is_owned(pg, io)
1041  *
1042  * \see cl_page_operations::cpo_assume()
1043  */
1044 void cl_page_assume(const struct lu_env *env,
1045                     struct cl_io *io, struct cl_page *pg)
1046 {
1047         PASSERT(env, pg, pg->cp_state < CPS_OWNED);
1048         PASSERT(env, pg, pg->cp_owner == NULL);
1049         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
1050         PINVRNT(env, pg, cl_page_invariant(pg));
1051
1052         ENTRY;
1053         pg = cl_page_top(pg);
1054         io = cl_io_top(io);
1055
1056         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
1057         pg->cp_owner = io;
1058         pg->cp_task = current;
1059         cl_page_owner_set(pg);
1060         cl_page_state_set(env, pg, CPS_OWNED);
1061         EXIT;
1062 }
1063 EXPORT_SYMBOL(cl_page_assume);
1064
1065 /**
1066  * Releases page ownership without unlocking the page.
1067  *
1068  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
1069  * underlying VM page (as VM is supposed to do this itself).
1070  *
1071  * \pre   cl_page_is_owned(pg, io)
1072  * \post !cl_page_is_owned(pg, io)
1073  *
1074  * \see cl_page_assume()
1075  */
1076 void cl_page_unassume(const struct lu_env *env,
1077                       struct cl_io *io, struct cl_page *pg)
1078 {
1079         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1080         PINVRNT(env, pg, cl_page_invariant(pg));
1081
1082         ENTRY;
1083         pg = cl_page_top(pg);
1084         io = cl_io_top(io);
1085         cl_page_owner_clear(pg);
1086         cl_page_state_set(env, pg, CPS_CACHED);
1087         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
1088                                (const struct lu_env *,
1089                                 const struct cl_page_slice *, struct cl_io *),
1090                                io);
1091         EXIT;
1092 }
1093 EXPORT_SYMBOL(cl_page_unassume);
1094
1095 /**
1096  * Releases page ownership.
1097  *
1098  * Moves page into cl_page_state::CPS_CACHED.
1099  *
1100  * \pre   cl_page_is_owned(pg, io)
1101  * \post !cl_page_is_owned(pg, io)
1102  *
1103  * \see cl_page_own()
1104  * \see cl_page_operations::cpo_disown()
1105  */
1106 void cl_page_disown(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
1111         ENTRY;
1112         pg = cl_page_top(pg);
1113         io = cl_io_top(io);
1114         cl_page_disown0(env, io, pg);
1115         EXIT;
1116 }
1117 EXPORT_SYMBOL(cl_page_disown);
1118
1119 /**
1120  * Called when page is to be removed from the object, e.g., as a result of
1121  * truncate.
1122  *
1123  * Calls cl_page_operations::cpo_discard() top-to-bottom.
1124  *
1125  * \pre cl_page_is_owned(pg, io)
1126  *
1127  * \see cl_page_operations::cpo_discard()
1128  */
1129 void cl_page_discard(const struct lu_env *env,
1130                      struct cl_io *io, struct cl_page *pg)
1131 {
1132         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1133         PINVRNT(env, pg, cl_page_invariant(pg));
1134
1135         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
1136 }
1137 EXPORT_SYMBOL(cl_page_discard);
1138
1139 /**
1140  * Version of cl_page_delete() that can be called for not fully constructed
1141  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
1142  * path. Doesn't check page invariant.
1143  */
1144 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
1145                             int radix)
1146 {
1147         struct cl_page *tmp = pg;
1148         ENTRY;
1149
1150         PASSERT(env, pg, pg == cl_page_top(pg));
1151         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
1152
1153         /*
1154          * Severe all ways to obtain new pointers to @pg.
1155          */
1156         cl_page_owner_clear(pg);
1157
1158         /*
1159          * unexport the page firstly before freeing it so that
1160          * the page content is considered to be invalid.
1161          * We have to do this because a CPS_FREEING cl_page may
1162          * be NOT under the protection of a cl_lock.
1163          * Afterwards, if this page is found by other threads, then this
1164          * page will be forced to reread.
1165          */
1166         cl_page_export(env, pg, 0);
1167         cl_page_state_set0(env, pg, CPS_FREEING);
1168
1169         if (tmp->cp_type == CPT_CACHEABLE) {
1170                 if (!radix)
1171                         /* !radix means that @pg is not yet in the radix tree,
1172                          * skip removing it.
1173                          */
1174                         tmp = pg->cp_child;
1175                 for (; tmp != NULL; tmp = tmp->cp_child) {
1176                         void                    *value;
1177                         struct cl_object_header *hdr;
1178
1179                         hdr = cl_object_header(tmp->cp_obj);
1180                         cfs_spin_lock(&hdr->coh_page_guard);
1181                         value = radix_tree_delete(&hdr->coh_tree,
1182                                                   tmp->cp_index);
1183                         PASSERT(env, tmp, value == tmp);
1184                         PASSERT(env, tmp, hdr->coh_pages > 0);
1185                         hdr->coh_pages--;
1186                         cfs_spin_unlock(&hdr->coh_page_guard);
1187                 }
1188         }
1189
1190         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete),
1191                        (const struct lu_env *, const struct cl_page_slice *));
1192         EXIT;
1193 }
1194
1195 /**
1196  * Called when a decision is made to throw page out of memory.
1197  *
1198  * Notifies all layers about page destruction by calling
1199  * cl_page_operations::cpo_delete() method top-to-bottom.
1200  *
1201  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
1202  * where transition to this state happens).
1203  *
1204  * Eliminates all venues through which new references to the page can be
1205  * obtained:
1206  *
1207  *     - removes page from the radix trees,
1208  *
1209  *     - breaks linkage from VM page to cl_page.
1210  *
1211  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
1212  * drain after some time, at which point page will be recycled.
1213  *
1214  * \pre  pg == cl_page_top(pg)
1215  * \pre  VM page is locked
1216  * \post pg->cp_state == CPS_FREEING
1217  *
1218  * \see cl_page_operations::cpo_delete()
1219  */
1220 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
1221 {
1222         PINVRNT(env, pg, cl_page_invariant(pg));
1223         ENTRY;
1224         cl_page_delete0(env, pg, 1);
1225         EXIT;
1226 }
1227 EXPORT_SYMBOL(cl_page_delete);
1228
1229 /**
1230  * Unmaps page from user virtual memory.
1231  *
1232  * Calls cl_page_operations::cpo_unmap() through all layers top-to-bottom. The
1233  * layer responsible for VM interaction has to unmap page from user space
1234  * virtual memory.
1235  *
1236  * \see cl_page_operations::cpo_unmap()
1237  */
1238 int cl_page_unmap(const struct lu_env *env,
1239                   struct cl_io *io, struct cl_page *pg)
1240 {
1241         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1242         PINVRNT(env, pg, cl_page_invariant(pg));
1243
1244         return cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_unmap));
1245 }
1246 EXPORT_SYMBOL(cl_page_unmap);
1247
1248 /**
1249  * Marks page up-to-date.
1250  *
1251  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
1252  * layer responsible for VM interaction has to mark/clear page as up-to-date
1253  * by the \a uptodate argument.
1254  *
1255  * \see cl_page_operations::cpo_export()
1256  */
1257 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
1258 {
1259         PINVRNT(env, pg, cl_page_invariant(pg));
1260         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
1261                        (const struct lu_env *,
1262                         const struct cl_page_slice *, int), uptodate);
1263 }
1264 EXPORT_SYMBOL(cl_page_export);
1265
1266 /**
1267  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
1268  * thread.
1269  */
1270 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
1271 {
1272         int result;
1273         const struct cl_page_slice *slice;
1274
1275         ENTRY;
1276         pg = cl_page_top_trusted((struct cl_page *)pg);
1277         slice = container_of(pg->cp_layers.next,
1278                              const struct cl_page_slice, cpl_linkage);
1279         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
1280         /*
1281          * Call ->cpo_is_vmlocked() directly instead of going through
1282          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
1283          * cl_page_invariant().
1284          */
1285         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
1286         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
1287         RETURN(result == -EBUSY);
1288 }
1289 EXPORT_SYMBOL(cl_page_is_vmlocked);
1290
1291 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
1292 {
1293         ENTRY;
1294         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
1295 }
1296
1297 static void cl_page_io_start(const struct lu_env *env,
1298                              struct cl_page *pg, enum cl_req_type crt)
1299 {
1300         /*
1301          * Page is queued for IO, change its state.
1302          */
1303         ENTRY;
1304         cl_page_owner_clear(pg);
1305         cl_page_state_set(env, pg, cl_req_type_state(crt));
1306         EXIT;
1307 }
1308
1309 /**
1310  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
1311  * called top-to-bottom. Every layer either agrees to submit this page (by
1312  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
1313  * handling interactions with the VM also has to inform VM that page is under
1314  * transfer now.
1315  */
1316 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
1317                  struct cl_page *pg, enum cl_req_type crt)
1318 {
1319         int result;
1320
1321         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1322         PINVRNT(env, pg, cl_page_invariant(pg));
1323         PINVRNT(env, pg, crt < CRT_NR);
1324
1325         /*
1326          * XXX this has to be called bottom-to-top, so that llite can set up
1327          * PG_writeback without risking other layers deciding to skip this
1328          * page.
1329          */
1330         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
1331         if (result == 0)
1332                 cl_page_io_start(env, pg, crt);
1333
1334         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
1335                       equi(result == 0,
1336                            PageWriteback(cl_page_vmpage(env, pg)))));
1337         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1338         return result;
1339 }
1340 EXPORT_SYMBOL(cl_page_prep);
1341
1342 /**
1343  * Notify layers about transfer completion.
1344  *
1345  * Invoked by transfer sub-system (which is a part of osc) to notify layers
1346  * that a transfer, of which this page is a part of has completed.
1347  *
1348  * Completion call-backs are executed in the bottom-up order, so that
1349  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1350  * and can release locks safely.
1351  *
1352  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1353  * \post pg->cp_state == CPS_CACHED
1354  *
1355  * \see cl_page_operations::cpo_completion()
1356  */
1357 void cl_page_completion(const struct lu_env *env,
1358                         struct cl_page *pg, enum cl_req_type crt, int ioret)
1359 {
1360         struct cl_sync_io *anchor = pg->cp_sync_io;
1361
1362         PASSERT(env, pg, crt < CRT_NR);
1363         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
1364         PASSERT(env, pg, pg->cp_req == NULL);
1365         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
1366         PINVRNT(env, pg, cl_page_invariant(pg));
1367
1368         ENTRY;
1369         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
1370         if (crt == CRT_READ && ioret == 0) {
1371                 PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED));
1372                 pg->cp_flags |= CPF_READ_COMPLETED;
1373         }
1374
1375         cl_page_state_set(env, pg, CPS_CACHED);
1376         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
1377                                (const struct lu_env *,
1378                                 const struct cl_page_slice *, int), ioret);
1379         if (anchor) {
1380                 LASSERT(pg->cp_sync_io == anchor);
1381                 pg->cp_sync_io = NULL;
1382                 cl_sync_io_note(anchor, ioret);
1383         }
1384
1385         /* Don't assert the page writeback bit here because the lustre file
1386          * may be as a backend of swap space. in this case, the page writeback
1387          * is set by VM, and obvious we shouldn't clear it at all. Fortunately
1388          * this type of pages are all TRANSIENT pages. */
1389         KLASSERT(ergo(pg->cp_type == CPT_CACHEABLE,
1390                       !PageWriteback(cl_page_vmpage(env, pg))));
1391         EXIT;
1392 }
1393 EXPORT_SYMBOL(cl_page_completion);
1394
1395 /**
1396  * Notify layers that transfer formation engine decided to yank this page from
1397  * the cache and to make it a part of a transfer.
1398  *
1399  * \pre  pg->cp_state == CPS_CACHED
1400  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1401  *
1402  * \see cl_page_operations::cpo_make_ready()
1403  */
1404 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
1405                        enum cl_req_type crt)
1406 {
1407         int result;
1408
1409         PINVRNT(env, pg, crt < CRT_NR);
1410
1411         ENTRY;
1412         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
1413                                 (const struct lu_env *,
1414                                  const struct cl_page_slice *));
1415         if (result == 0) {
1416                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
1417                 cl_page_io_start(env, pg, crt);
1418         }
1419         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1420         RETURN(result);
1421 }
1422 EXPORT_SYMBOL(cl_page_make_ready);
1423
1424 /**
1425  * Notify layers that high level io decided to place this page into a cache
1426  * for future transfer.
1427  *
1428  * The layer implementing transfer engine (osc) has to register this page in
1429  * its queues.
1430  *
1431  * \pre  cl_page_is_owned(pg, io)
1432  * \post ergo(result == 0,
1433  *            pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT)
1434  *
1435  * \see cl_page_operations::cpo_cache_add()
1436  */
1437 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
1438                       struct cl_page *pg, enum cl_req_type crt)
1439 {
1440         int result;
1441
1442         PINVRNT(env, pg, crt < CRT_NR);
1443         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1444         PINVRNT(env, pg, cl_page_invariant(pg));
1445
1446         ENTRY;
1447         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_cache_add));
1448         if (result == 0) {
1449                 cl_page_owner_clear(pg);
1450                 cl_page_state_set(env, pg, CPS_CACHED);
1451         }
1452         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1453         RETURN(result);
1454 }
1455 EXPORT_SYMBOL(cl_page_cache_add);
1456
1457 /**
1458  * Checks whether page is protected by any extent lock is at least required
1459  * mode.
1460  *
1461  * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1462  * \see cl_page_operations::cpo_is_under_lock()
1463  */
1464 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1465                           struct cl_page *page)
1466 {
1467         int rc;
1468
1469         PINVRNT(env, page, cl_page_invariant(page));
1470
1471         ENTRY;
1472         rc = CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1473                             (const struct lu_env *,
1474                              const struct cl_page_slice *, struct cl_io *),
1475                             io);
1476         PASSERT(env, page, rc != 0);
1477         RETURN(rc);
1478 }
1479 EXPORT_SYMBOL(cl_page_is_under_lock);
1480
1481 /**
1482  * Purges all cached pages belonging to the object \a obj.
1483  */
1484 int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj)
1485 {
1486         struct cl_thread_info   *info;
1487         struct cl_object        *obj = cl_object_top(clobj);
1488         struct cl_io            *io;
1489         struct cl_page_list     *plist;
1490         int                      result;
1491
1492         ENTRY;
1493         info  = cl_env_info(env);
1494         plist = &info->clt_list;
1495         io    = &info->clt_io;
1496
1497         /*
1498          * initialize the io. This is ugly since we never do IO in this
1499          * function, we just make cl_page_list functions happy. -jay
1500          */
1501         io->ci_obj = obj;
1502         result = cl_io_init(env, io, CIT_MISC, obj);
1503         if (result != 0) {
1504                 cl_io_fini(env, io);
1505                 RETURN(io->ci_result);
1506         }
1507
1508         do {
1509                 cl_page_list_init(plist);
1510                 result = cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF,
1511                                              plist);
1512                 /*
1513                  * Since we're purging the pages of an object, we don't care
1514                  * the possible outcomes of the following functions.
1515                  */
1516                 cl_page_list_unmap(env, io, plist);
1517                 cl_page_list_discard(env, io, plist);
1518                 cl_page_list_disown(env, io, plist);
1519                 cl_page_list_fini(env, plist);
1520
1521                 if (result == CLP_GANG_RESCHED)
1522                         cfs_cond_resched();
1523         } while (result != CLP_GANG_OKAY);
1524
1525         cl_io_fini(env, io);
1526         RETURN(result);
1527 }
1528 EXPORT_SYMBOL(cl_pages_prune);
1529
1530 /**
1531  * Tells transfer engine that only part of a page is to be transmitted.
1532  *
1533  * \see cl_page_operations::cpo_clip()
1534  */
1535 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1536                   int from, int to)
1537 {
1538         PINVRNT(env, pg, cl_page_invariant(pg));
1539
1540         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1541         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1542                        (const struct lu_env *,
1543                         const struct cl_page_slice *,int, int),
1544                        from, to);
1545 }
1546 EXPORT_SYMBOL(cl_page_clip);
1547
1548 /**
1549  * Prints human readable representation of \a pg to the \a f.
1550  */
1551 void cl_page_header_print(const struct lu_env *env, void *cookie,
1552                           lu_printer_t printer, const struct cl_page *pg)
1553 {
1554         (*printer)(env, cookie,
1555                    "page@%p[%d %p:%lu ^%p_%p %d %d %d %p %p %#x]\n",
1556                    pg, cfs_atomic_read(&pg->cp_ref), pg->cp_obj,
1557                    pg->cp_index, pg->cp_parent, pg->cp_child,
1558                    pg->cp_state, pg->cp_error, pg->cp_type,
1559                    pg->cp_owner, pg->cp_req, pg->cp_flags);
1560 }
1561 EXPORT_SYMBOL(cl_page_header_print);
1562
1563 /**
1564  * Prints human readable representation of \a pg to the \a f.
1565  */
1566 void cl_page_print(const struct lu_env *env, void *cookie,
1567                    lu_printer_t printer, const struct cl_page *pg)
1568 {
1569         struct cl_page *scan;
1570
1571         for (scan = cl_page_top((struct cl_page *)pg);
1572              scan != NULL; scan = scan->cp_child)
1573                 cl_page_header_print(env, cookie, printer, scan);
1574         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1575                        (const struct lu_env *env,
1576                         const struct cl_page_slice *slice,
1577                         void *cookie, lu_printer_t p), cookie, printer);
1578         (*printer)(env, cookie, "end page@%p\n", pg);
1579 }
1580 EXPORT_SYMBOL(cl_page_print);
1581
1582 /**
1583  * Cancel a page which is still in a transfer.
1584  */
1585 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1586 {
1587         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1588                               (const struct lu_env *,
1589                                const struct cl_page_slice *));
1590 }
1591 EXPORT_SYMBOL(cl_page_cancel);
1592
1593 /**
1594  * Converts a byte offset within object \a obj into a page index.
1595  */
1596 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1597 {
1598         /*
1599          * XXX for now.
1600          */
1601         return (loff_t)idx << CFS_PAGE_SHIFT;
1602 }
1603 EXPORT_SYMBOL(cl_offset);
1604
1605 /**
1606  * Converts a page index into a byte offset within object \a obj.
1607  */
1608 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1609 {
1610         /*
1611          * XXX for now.
1612          */
1613         return offset >> CFS_PAGE_SHIFT;
1614 }
1615 EXPORT_SYMBOL(cl_index);
1616
1617 int cl_page_size(const struct cl_object *obj)
1618 {
1619         return 1 << CFS_PAGE_SHIFT;
1620 }
1621 EXPORT_SYMBOL(cl_page_size);
1622
1623 /**
1624  * Adds page slice to the compound page.
1625  *
1626  * This is called by cl_object_operations::coo_page_init() methods to add a
1627  * per-layer state to the page. New state is added at the end of
1628  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1629  *
1630  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1631  */
1632 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1633                        struct cl_object *obj,
1634                        const struct cl_page_operations *ops)
1635 {
1636         ENTRY;
1637         cfs_list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1638         slice->cpl_obj  = obj;
1639         slice->cpl_ops  = ops;
1640         slice->cpl_page = page;
1641         EXIT;
1642 }
1643 EXPORT_SYMBOL(cl_page_slice_add);
1644
1645 int  cl_page_init(void)
1646 {
1647         return lu_kmem_init(cl_page_caches);
1648 }
1649
1650 void cl_page_fini(void)
1651 {
1652         lu_kmem_fini(cl_page_caches);
1653 }