lustre/obdclass/cl_page.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Client Lustre Page.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_CLASS
  42 #ifndef EXPORT_SYMTAB
  43 # define EXPORT_SYMTAB
  44 #endif
  45
  46 #include <libcfs/libcfs.h>
  47 #include <obd_class.h>
  48 #include <obd_support.h>
  49 #include <libcfs/list.h>
  50
  51 #include <cl_object.h>
  52 #include "cl_internal.h"
  53
  54 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
  55                             int radix);
  56
  57 static cfs_mem_cache_t      *cl_page_kmem = NULL;
  58
  59 static struct lu_kmem_descr cl_page_caches[] = {
  60         {
  61                 .ckd_cache = &cl_page_kmem,
  62                 .ckd_name  = "cl_page_kmem",
  63                 .ckd_size  = sizeof (struct cl_page)
  64         },
  65         {
  66                 .ckd_cache = NULL
  67         }
  68 };
  69
  70 #ifdef LIBCFS_DEBUG
  71 # define PASSERT(env, page, expr)                                       \
  72   do {                                                                    \
  73           if (unlikely(!(expr))) {                                      \
  74                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  75                   LASSERT(0);                                           \
  76           }                                                             \
  77   } while (0)
  78 #else /* !LIBCFS_DEBUG */
  79 # define PASSERT(env, page, exp) \
  80         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  81 #endif /* !LIBCFS_DEBUG */
  82
  83 #ifdef INVARIANT_CHECK
  84 # define PINVRNT(env, page, expr)                                       \
  85   do {                                                                    \
  86           if (unlikely(!(expr))) {                                      \
  87                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  88                   LINVRNT(0);                                           \
  89           }                                                             \
  90   } while (0)
  91 #else /* !INVARIANT_CHECK */
  92 # define PINVRNT(env, page, exp) \
  93         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  94 #endif /* !INVARIANT_CHECK */
  95
  96 /**
  97  * Internal version of cl_page_top, it should be called with page referenced,
  98  * or coh_page_guard held.
  99  */
 100 static struct cl_page *cl_page_top_trusted(struct cl_page *page)
 101 {
 102         LASSERT(cl_is_page(page));
 103         while (page->cp_parent != NULL)
 104                 page = page->cp_parent;
 105         return page;
 106 }
 107
 108 /**
 109  * Internal version of cl_page_get().
 110  *
 111  * This function can be used to obtain initial reference to previously
 112  * unreferenced cached object. It can be called only if concurrent page
 113  * reclamation is somehow prevented, e.g., by locking page radix-tree
 114  * (cl_object_header::hdr->coh_page_guard), or by keeping a lock on a VM page,
 115  * associated with \a page.
 116  *
 117  * Use with care! Not exported.
 118  */
 119 static void cl_page_get_trust(struct cl_page *page)
 120 {
 121         LASSERT(cl_is_page(page));
 122         /*
 123          * Checkless version for trusted users.
 124          */
 125         if (atomic_inc_return(&page->cp_ref) == 1)
 126                 atomic_inc(&cl_object_site(page->cp_obj)->cs_pages.cs_busy);
 127 }
 128
 129 /**
 130  * Returns a slice within a page, corresponding to the given layer in the
 131  * device stack.
 132  *
 133  * \see cl_lock_at()
 134  */
 135 static const struct cl_page_slice *
 136 cl_page_at_trusted(const struct cl_page *page,
 137                    const struct lu_device_type *dtype)
 138 {
 139         const struct cl_page_slice *slice;
 140
 141 #ifdef INVARIANT_CHECK
 142         struct cl_object_header *ch = cl_object_header(page->cp_obj);
 143
 144         if (!atomic_read(&page->cp_ref))
 145                 LASSERT_SPIN_LOCKED(&ch->coh_page_guard);
 146 #endif
 147         ENTRY;
 148
 149         page = cl_page_top_trusted((struct cl_page *)page);
 150         do {
 151                 list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 152                         if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
 153                                 RETURN(slice);
 154                 }
 155                 page = page->cp_child;
 156         } while (page != NULL);
 157         RETURN(NULL);
 158 }
 159
 160 /**
 161  * Returns a page with given index in the given object, or NULL if no page is
 162  * found. Acquires a reference on \a page.
 163  *
 164  * Locking: called under cl_object_header::coh_page_guard spin-lock.
 165  */
 166 struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index)
 167 {
 168         struct cl_page *page;
 169
 170         LASSERT_SPIN_LOCKED(&hdr->coh_page_guard);
 171
 172         page = radix_tree_lookup(&hdr->coh_tree, index);
 173         if (page != NULL) {
 174                 LASSERT(cl_is_page(page));
 175                 cl_page_get_trust(page);
 176         }
 177         return page;
 178 }
 179 EXPORT_SYMBOL(cl_page_lookup);
 180
 181 /**
 182  * Returns a list of pages by a given [start, end] of \a obj.
 183  *
 184  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
 185  * crucial in the face of [offset, EOF] locks.
 186  */
 187 void cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj,
 188                          struct cl_io *io, pgoff_t start, pgoff_t end,
 189                          struct cl_page_list *queue, int nonblock)
 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                    (*page_own)(const struct lu_env *env,
 201                                            struct cl_io *io,
 202                                            struct cl_page *pg);
 203         ENTRY;
 204
 205         page_own = nonblock ? cl_page_own_try : cl_page_own;
 206
 207         idx = start;
 208         hdr = cl_object_header(obj);
 209         pvec = cl_env_info(env)->clt_pvec;
 210         dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type;
 211         spin_lock(&hdr->coh_page_guard);
 212         while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec,
 213                                             idx, CLT_PVEC_SIZE)) > 0) {
 214                 idx = pvec[nr - 1]->cp_index + 1;
 215                 for (i = 0, j = 0; i < nr; ++i) {
 216                         page = pvec[i];
 217                         PASSERT(env, page, cl_is_page(page));
 218                         pvec[i] = NULL;
 219                         if (page->cp_index > end)
 220                                 break;
 221                         if (page->cp_state == CPS_FREEING)
 222                                 continue;
 223                         if (page->cp_type == CPT_TRANSIENT) {
 224                                 /* God, we found a transient page!*/
 225                                 continue;
 226                         }
 227
 228                         slice = cl_page_at_trusted(page, dtype);
 229                         /*
 230                          * Pages for lsm-less file has no underneath sub-page
 231                          * for osc, in case of ...
 232                          */
 233                         PASSERT(env, page, slice != NULL);
 234                         page = slice->cpl_page;
 235                         /*
 236                          * Can safely call cl_page_get_trust() under
 237                          * radix-tree spin-lock.
 238                          *
 239                          * XXX not true, because @page is from object another
 240                          * than @hdr and protected by different tree lock.
 241                          */
 242                         cl_page_get_trust(page);
 243                         lu_ref_add_atomic(&page->cp_reference,
 244                                           "page_list", cfs_current());
 245                         pvec[j++] = page;
 246                 }
 247
 248                 /*
 249                  * Here a delicate locking dance is performed. Current thread
 250                  * holds a reference to a page, but has to own it before it
 251                  * can be placed into queue. Owning implies waiting, so
 252                  * radix-tree lock is to be released. After a wait one has to
 253                  * check that pages weren't truncated (cl_page_own() returns
 254                  * error in the latter case).
 255                  */
 256                 spin_unlock(&hdr->coh_page_guard);
 257                 for (i = 0; i < j; ++i) {
 258                         page = pvec[i];
 259                         if (page_own(env, io, page) == 0)
 260                                 cl_page_list_add(queue, page);
 261                         lu_ref_del(&page->cp_reference,
 262                                    "page_list", cfs_current());
 263                         cl_page_put(env, page);
 264                 }
 265                 spin_lock(&hdr->coh_page_guard);
 266                 if (nr < CLT_PVEC_SIZE)
 267                         break;
 268         }
 269         spin_unlock(&hdr->coh_page_guard);
 270         EXIT;
 271 }
 272 EXPORT_SYMBOL(cl_page_gang_lookup);
 273
 274 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
 275 {
 276         struct cl_object *obj  = page->cp_obj;
 277         struct cl_site   *site = cl_object_site(obj);
 278
 279         PASSERT(env, page, cl_is_page(page));
 280         PASSERT(env, page, list_empty(&page->cp_batch));
 281         PASSERT(env, page, page->cp_owner == NULL);
 282         PASSERT(env, page, page->cp_req == NULL);
 283         PASSERT(env, page, page->cp_parent == NULL);
 284         PASSERT(env, page, page->cp_state == CPS_FREEING);
 285
 286         ENTRY;
 287         might_sleep();
 288         while (!list_empty(&page->cp_layers)) {
 289                 struct cl_page_slice *slice;
 290
 291                 slice = list_entry(page->cp_layers.next, struct cl_page_slice,
 292                                    cpl_linkage);
 293                 list_del_init(page->cp_layers.next);
 294                 slice->cpl_ops->cpo_fini(env, slice);
 295         }
 296         atomic_dec(&site->cs_pages.cs_total);
 297         atomic_dec(&site->cs_pages_state[page->cp_state]);
 298         lu_object_ref_del_at(&obj->co_lu, page->cp_obj_ref, "cl_page", page);
 299         cl_object_put(env, obj);
 300         lu_ref_fini(&page->cp_reference);
 301         OBD_SLAB_FREE_PTR(page, cl_page_kmem);
 302         EXIT;
 303 }
 304
 305 /**
 306  * Helper function updating page state. This is the only place in the code
 307  * where cl_page::cp_state field is mutated.
 308  */
 309 static inline void cl_page_state_set_trust(struct cl_page *page,
 310                                            enum cl_page_state state)
 311 {
 312         /* bypass const. */
 313         *(enum cl_page_state *)&page->cp_state = state;
 314 }
 315
 316 static int cl_page_alloc(const struct lu_env *env, struct cl_object *o,
 317                          pgoff_t ind, struct page *vmpage,
 318                          enum cl_page_type type, struct cl_page **out)
 319 {
 320         struct cl_page          *page;
 321         struct cl_page          *err  = NULL;
 322         struct lu_object_header *head;
 323         struct cl_site          *site = cl_object_site(o);
 324         int                      result;
 325
 326         ENTRY;
 327         result = +1;
 328         OBD_SLAB_ALLOC_PTR_GFP(page, cl_page_kmem, CFS_ALLOC_IO);
 329         if (page != NULL) {
 330                 atomic_set(&page->cp_ref, 1);
 331                 page->cp_obj = o;
 332                 cl_object_get(o);
 333                 page->cp_obj_ref = lu_object_ref_add(&o->co_lu,
 334                                                      "cl_page", page);
 335                 page->cp_index = ind;
 336                 cl_page_state_set_trust(page, CPS_CACHED);
 337                 page->cp_type = type;
 338                 CFS_INIT_LIST_HEAD(&page->cp_layers);
 339                 CFS_INIT_LIST_HEAD(&page->cp_batch);
 340                 CFS_INIT_LIST_HEAD(&page->cp_flight);
 341                 mutex_init(&page->cp_mutex);
 342                 lu_ref_init(&page->cp_reference);
 343                 head = o->co_lu.lo_header;
 344                 list_for_each_entry(o, &head->loh_layers, co_lu.lo_linkage) {
 345                         if (o->co_ops->coo_page_init != NULL) {
 346                                 err = o->co_ops->coo_page_init(env, o,
 347                                                                page, vmpage);
 348                                 if (err != NULL) {
 349                                         cl_page_state_set_trust(page,
 350                                                                 CPS_FREEING);
 351                                         cl_page_free(env, page);
 352                                         page = err;
 353                                         break;
 354                                 }
 355                         }
 356                 }
 357                 if (err == NULL) {
 358                         atomic_inc(&site->cs_pages.cs_busy);
 359                         atomic_inc(&site->cs_pages.cs_total);
 360                         atomic_inc(&site->cs_pages_state[CPS_CACHED]);
 361                         atomic_inc(&site->cs_pages.cs_created);
 362                         result = 0;
 363                 }
 364         } else
 365                 page = ERR_PTR(-ENOMEM);
 366         *out = page;
 367         RETURN(result);
 368 }
 369
 370 /**
 371  * Returns a cl_page with index \a idx at the object \a o, and associated with
 372  * the VM page \a vmpage.
 373  *
 374  * This is the main entry point into the cl_page caching interface. First, a
 375  * cache (implemented as a per-object radix tree) is consulted. If page is
 376  * found there, it is returned immediately. Otherwise new page is allocated
 377  * and returned. In any case, additional reference to page is acquired.
 378  *
 379  * \see cl_object_find(), cl_lock_find()
 380  */
 381 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o,
 382                              pgoff_t idx, struct page *vmpage,
 383                              enum cl_page_type type)
 384 {
 385         struct cl_page          *page;
 386         struct cl_page          *ghost = NULL;
 387         struct cl_object_header *hdr;
 388         struct cl_site          *site = cl_object_site(o);
 389         int err;
 390
 391         LINVRNT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 392         might_sleep();
 393
 394         ENTRY;
 395
 396         hdr = cl_object_header(o);
 397         atomic_inc(&site->cs_pages.cs_lookup);
 398
 399         CDEBUG(D_PAGE, "%lu@"DFID" %p %lu %i\n",
 400                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 401         /* fast path. */
 402         if (type == CPT_CACHEABLE) {
 403                 /*
 404                  * cl_vmpage_page() can be called here without any locks as
 405                  *
 406                  *     - "vmpage" is locked (which prevents ->private from
 407                  *       concurrent updates), and
 408                  *
 409                  *     - "o" cannot be destroyed while current thread holds a
 410                  *       reference on it.
 411                  */
 412                 page = cl_vmpage_page(vmpage, o);
 413                 PINVRNT(env, page,
 414                         ergo(page != NULL,
 415                              cl_page_vmpage(env, page) == vmpage &&
 416                              (void *)radix_tree_lookup(&hdr->coh_tree,
 417                                                        idx) == page));
 418         } else {
 419                 spin_lock(&hdr->coh_page_guard);
 420                 page = cl_page_lookup(hdr, idx);
 421                 spin_unlock(&hdr->coh_page_guard);
 422         }
 423         if (page != NULL) {
 424                 atomic_inc(&site->cs_pages.cs_hit);
 425                 RETURN(page);
 426         }
 427
 428         /* allocate and initialize cl_page */
 429         err = cl_page_alloc(env, o, idx, vmpage, type, &page);
 430         if (err != 0)
 431                 RETURN(page);
 432         /*
 433          * XXX optimization: use radix_tree_preload() here, and change tree
 434          * gfp mask to GFP_KERNEL in cl_object_header_init().
 435          */
 436         spin_lock(&hdr->coh_page_guard);
 437         err = radix_tree_insert(&hdr->coh_tree, idx, page);
 438         if (err != 0) {
 439                 ghost = page;
 440                 /*
 441                  * Noted by Jay: a lock on \a vmpage protects cl_page_find()
 442                  * from this race, but
 443                  *
 444                  *     0. it's better to have cl_page interface "locally
 445                  *     consistent" so that its correctness can be reasoned
 446                  *     about without appealing to the (obscure world of) VM
 447                  *     locking.
 448                  *
 449                  *     1. handling this race allows ->coh_tree to remain
 450                  *     consistent even when VM locking is somehow busted,
 451                  *     which is very useful during diagnosing and debugging.
 452                  */
 453                 if (err == -EEXIST) {
 454                         /*
 455                          * XXX in case of a lookup for CPT_TRANSIENT page,
 456                          * nothing protects a CPT_CACHEABLE page from being
 457                          * concurrently moved into CPS_FREEING state.
 458                          */
 459                         page = cl_page_lookup(hdr, idx);
 460                         PASSERT(env, page, page != NULL);
 461                         if (page->cp_type == CPT_TRANSIENT &&
 462                             type == CPT_CACHEABLE) {
 463                                 /* XXX: We should make sure that inode sem
 464                                  * keeps being held in the lifetime of
 465                                  * transient pages, so it is impossible to
 466                                  * have conflicting transient pages.
 467                                  */
 468                                 spin_unlock(&hdr->coh_page_guard);
 469                                 cl_page_put(env, page);
 470                                 spin_lock(&hdr->coh_page_guard);
 471                                 page = ERR_PTR(-EBUSY);
 472                         }
 473                 } else
 474                         page = ERR_PTR(err);
 475         } else
 476                 hdr->coh_pages++;
 477         spin_unlock(&hdr->coh_page_guard);
 478
 479         if (unlikely(ghost != NULL)) {
 480                 atomic_dec(&site->cs_pages.cs_busy);
 481                 cl_page_delete0(env, ghost, 0);
 482                 cl_page_free(env, ghost);
 483         }
 484         RETURN(page);
 485 }
 486 EXPORT_SYMBOL(cl_page_find);
 487
 488 static inline int cl_page_invariant(const struct cl_page *pg)
 489 {
 490         struct cl_object_header *header;
 491         struct cl_page          *parent;
 492         struct cl_page          *child;
 493         struct cl_io            *owner;
 494
 495         LASSERT(cl_is_page(pg));
 496         /*
 497          * Page invariant is protected by a VM lock.
 498          */
 499         LINVRNT(cl_page_is_vmlocked(NULL, pg));
 500
 501         header = cl_object_header(pg->cp_obj);
 502         parent = pg->cp_parent;
 503         child  = pg->cp_child;
 504         owner  = pg->cp_owner;
 505
 506         return atomic_read(&pg->cp_ref) > 0 &&
 507                 ergo(parent != NULL, parent->cp_child == pg) &&
 508                 ergo(child != NULL, child->cp_parent == pg) &&
 509                 ergo(child != NULL, pg->cp_obj != child->cp_obj) &&
 510                 ergo(parent != NULL, pg->cp_obj != parent->cp_obj) &&
 511                 ergo(owner != NULL && parent != NULL,
 512                      parent->cp_owner == pg->cp_owner->ci_parent) &&
 513                 ergo(owner != NULL && child != NULL,
 514                      child->cp_owner->ci_parent == owner) &&
 515                 /*
 516                  * Either page is early in initialization (has neither child
 517                  * nor parent yet), or it is in the object radix tree.
 518                  */
 519                 ergo(pg->cp_state < CPS_FREEING,
 520                      (void *)radix_tree_lookup(&header->coh_tree,
 521                                                pg->cp_index) == pg ||
 522                      (child == NULL && parent == NULL));
 523 }
 524
 525 static void cl_page_state_set0(const struct lu_env *env,
 526                                struct cl_page *page, enum cl_page_state state)
 527 {
 528         enum cl_page_state old;
 529         struct cl_site *site = cl_object_site(page->cp_obj);
 530
 531         /*
 532          * Matrix of allowed state transitions [old][new], for sanity
 533          * checking.
 534          */
 535         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 536                 [CPS_CACHED] = {
 537                         [CPS_CACHED]  = 0,
 538                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 539                         [CPS_PAGEIN]  = 0,
 540                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 541                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 542                 },
 543                 [CPS_OWNED] = {
 544                         [CPS_CACHED]  = 1, /* release to the cache */
 545                         [CPS_OWNED]   = 0,
 546                         [CPS_PAGEIN]  = 1, /* start read immediately */
 547                         [CPS_PAGEOUT] = 1, /* start write immediately */
 548                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 549                 },
 550                 [CPS_PAGEIN] = {
 551                         [CPS_CACHED]  = 1, /* io completion */
 552                         [CPS_OWNED]   = 0,
 553                         [CPS_PAGEIN]  = 0,
 554                         [CPS_PAGEOUT] = 0,
 555                         [CPS_FREEING] = 0,
 556                 },
 557                 [CPS_PAGEOUT] = {
 558                         [CPS_CACHED]  = 1, /* io completion */
 559                         [CPS_OWNED]   = 0,
 560                         [CPS_PAGEIN]  = 0,
 561                         [CPS_PAGEOUT] = 0,
 562                         [CPS_FREEING] = 0,
 563                 },
 564                 [CPS_FREEING] = {
 565                         [CPS_CACHED]  = 0,
 566                         [CPS_OWNED]   = 0,
 567                         [CPS_PAGEIN]  = 0,
 568                         [CPS_PAGEOUT] = 0,
 569                         [CPS_FREEING] = 0,
 570                 }
 571         };
 572
 573         ENTRY;
 574         old = page->cp_state;
 575         PASSERT(env, page, allowed_transitions[old][state]);
 576         CL_PAGE_HEADER(D_TRACE, env, page, "%i -> %i\n", old, state);
 577         for (; page != NULL; page = page->cp_child) {
 578                 PASSERT(env, page, page->cp_state == old);
 579                 PASSERT(env, page,
 580                         equi(state == CPS_OWNED, page->cp_owner != NULL));
 581
 582                 atomic_dec(&site->cs_pages_state[page->cp_state]);
 583                 atomic_inc(&site->cs_pages_state[state]);
 584                 cl_page_state_set_trust(page, state);
 585         }
 586         EXIT;
 587 }
 588
 589 static void cl_page_state_set(const struct lu_env *env,
 590                               struct cl_page *page, enum cl_page_state state)
 591 {
 592         PINVRNT(env, page, cl_page_invariant(page));
 593         cl_page_state_set0(env, page, state);
 594 }
 595
 596 /**
 597  * Acquires an additional reference to a page.
 598  *
 599  * This can be called only by caller already possessing a reference to \a
 600  * page.
 601  *
 602  * \see cl_object_get(), cl_lock_get().
 603  */
 604 void cl_page_get(struct cl_page *page)
 605 {
 606         ENTRY;
 607         LASSERT(page->cp_state != CPS_FREEING);
 608         cl_page_get_trust(page);
 609         EXIT;
 610 }
 611 EXPORT_SYMBOL(cl_page_get);
 612
 613 /**
 614  * Releases a reference to a page.
 615  *
 616  * When last reference is released, page is returned to the cache, unless it
 617  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 618  * destroyed.
 619  *
 620  * \see cl_object_put(), cl_lock_put().
 621  */
 622 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 623 {
 624         struct cl_object_header *hdr;
 625         struct cl_site *site = cl_object_site(page->cp_obj);
 626
 627         PASSERT(env, page, atomic_read(&page->cp_ref) > !!page->cp_parent);
 628
 629         ENTRY;
 630         CL_PAGE_HEADER(D_TRACE, env, page, "%i\n", atomic_read(&page->cp_ref));
 631
 632         hdr = cl_object_header(cl_object_top(page->cp_obj));
 633         if (atomic_dec_and_lock(&page->cp_ref, &hdr->coh_page_guard)) {
 634                 atomic_dec(&site->cs_pages.cs_busy);
 635                 /* We're going to access the page w/o a reference, but it's
 636                  * ok because we have grabbed the lock coh_page_guard, which
 637                  * means nobody is able to free this page behind us.
 638                  */
 639                 if (page->cp_state == CPS_FREEING) {
 640                         /* We drop the page reference and check the page state
 641                          * inside the coh_page_guard. So that if it gets here,
 642                          * it is the REALLY last reference to this page.
 643                          */
 644                         spin_unlock(&hdr->coh_page_guard);
 645
 646                         LASSERT(atomic_read(&page->cp_ref) == 0);
 647                         PASSERT(env, page, page->cp_owner == NULL);
 648                         PASSERT(env, page, list_empty(&page->cp_batch));
 649                         /*
 650                          * Page is no longer reachable by other threads. Tear
 651                          * it down.
 652                          */
 653                         cl_page_free(env, page);
 654
 655                         EXIT;
 656                         return;
 657                 }
 658                 spin_unlock(&hdr->coh_page_guard);
 659         }
 660
 661         EXIT;
 662 }
 663 EXPORT_SYMBOL(cl_page_put);
 664
 665 /**
 666  * Returns a VM page associated with a given cl_page.
 667  */
 668 cfs_page_t *cl_page_vmpage(const struct lu_env *env, struct cl_page *page)
 669 {
 670         const struct cl_page_slice *slice;
 671
 672         /*
 673          * Find uppermost layer with ->cpo_vmpage() method, and return its
 674          * result.
 675          */
 676         page = cl_page_top(page);
 677         do {
 678                 list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 679                         if (slice->cpl_ops->cpo_vmpage != NULL)
 680                                 RETURN(slice->cpl_ops->cpo_vmpage(env, slice));
 681                 }
 682                 page = page->cp_child;
 683         } while (page != NULL);
 684         LBUG(); /* ->cpo_vmpage() has to be defined somewhere in the stack */
 685 }
 686 EXPORT_SYMBOL(cl_page_vmpage);
 687
 688 /**
 689  * Returns a cl_page associated with a VM page, and given cl_object.
 690  */
 691 struct cl_page *cl_vmpage_page(cfs_page_t *vmpage, struct cl_object *obj)
 692 {
 693         struct cl_page *page;
 694         struct cl_object_header *hdr;
 695
 696         ENTRY;
 697         KLASSERT(PageLocked(vmpage));
 698
 699         /*
 700          * NOTE: absence of races and liveness of data are guaranteed by page
 701          *       lock on a "vmpage". That works because object destruction has
 702          *       bottom-to-top pass.
 703          */
 704
 705         /*
 706          * This loop assumes that ->private points to the top-most page. This
 707          * can be rectified easily.
 708          */
 709         hdr = cl_object_header(cl_object_top(obj));
 710         spin_lock(&hdr->coh_page_guard);
 711         for (page = (void *)vmpage->private;
 712              page != NULL; page = page->cp_child) {
 713                 if (cl_object_same(page->cp_obj, obj)) {
 714                         cl_page_get_trust(page);
 715                         break;
 716                 }
 717         }
 718         spin_unlock(&hdr->coh_page_guard);
 719         LASSERT(ergo(page, cl_is_page(page) && page->cp_type == CPT_CACHEABLE));
 720         RETURN(page);
 721 }
 722 EXPORT_SYMBOL(cl_vmpage_page);
 723
 724 /**
 725  * Returns the top-page for a given page.
 726  *
 727  * \see cl_object_top(), cl_io_top()
 728  */
 729 struct cl_page *cl_page_top(struct cl_page *page)
 730 {
 731         return cl_page_top_trusted(page);
 732 }
 733 EXPORT_SYMBOL(cl_page_top);
 734
 735 /**
 736  * Returns true if \a addr is an address of an allocated cl_page. Used in
 737  * assertions. This check is optimistically imprecise, i.e., it occasionally
 738  * returns true for the incorrect addresses, but if it returns false, then the
 739  * address is guaranteed to be incorrect. (Should be named cl_pagep().)
 740  *
 741  * \see cl_is_lock()
 742  */
 743 int cl_is_page(const void *addr)
 744 {
 745         return cfs_mem_is_in_cache(addr, cl_page_kmem);
 746 }
 747 EXPORT_SYMBOL(cl_is_page);
 748
 749 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 750                                        const struct lu_device_type *dtype)
 751 {
 752         return cl_page_at_trusted(page, dtype);
 753 }
 754 EXPORT_SYMBOL(cl_page_at);
 755
 756 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 757
 758 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)                   \
 759 ({                                                                      \
 760         const struct lu_env        *__env  = (_env);                    \
 761         struct cl_page             *__page = (_page);                   \
 762         const struct cl_page_slice *__scan;                             \
 763         int                         __result;                           \
 764         ptrdiff_t                   __op   = (_op);                     \
 765         int                       (*__method)_proto;                    \
 766                                                                         \
 767         __result = 0;                                                   \
 768         __page = cl_page_top(__page);                                   \
 769         do {                                                            \
 770                 list_for_each_entry(__scan, &__page->cp_layers,         \
 771                                     cpl_linkage) {                      \
 772                         __method = *(void **)((char *)__scan->cpl_ops + \
 773                                               __op);                    \
 774                         if (__method != NULL) {                         \
 775                                 __result = (*__method)(__env, __scan,   \
 776                                                        ## __VA_ARGS__); \
 777                                 if (__result != 0)                      \
 778                                         break;                          \
 779                         }                                               \
 780                 }                                                       \
 781                 __page = __page->cp_child;                              \
 782         } while (__page != NULL && __result == 0);                      \
 783         if (__result > 0)                                               \
 784                 __result = 0;                                           \
 785         __result;                                                       \
 786 })
 787
 788 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 789 do {                                                                    \
 790         const struct lu_env        *__env  = (_env);                    \
 791         struct cl_page             *__page = (_page);                   \
 792         const struct cl_page_slice *__scan;                             \
 793         ptrdiff_t                   __op   = (_op);                     \
 794         void                      (*__method)_proto;                    \
 795                                                                         \
 796         __page = cl_page_top(__page);                                   \
 797         do {                                                            \
 798                 list_for_each_entry(__scan, &__page->cp_layers,         \
 799                                     cpl_linkage) {                      \
 800                         __method = *(void **)((char *)__scan->cpl_ops + \
 801                                               __op);                    \
 802                         if (__method != NULL)                           \
 803                                 (*__method)(__env, __scan,              \
 804                                             ## __VA_ARGS__);            \
 805                 }                                                       \
 806                 __page = __page->cp_child;                              \
 807         } while (__page != NULL);                                       \
 808 } while (0)
 809
 810 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)           \
 811 do {                                                                    \
 812         const struct lu_env        *__env  = (_env);                    \
 813         struct cl_page             *__page = (_page);                   \
 814         const struct cl_page_slice *__scan;                             \
 815         ptrdiff_t                   __op   = (_op);                     \
 816         void                      (*__method)_proto;                    \
 817                                                                         \
 818         /* get to the bottom page. */                                   \
 819         while (__page->cp_child != NULL)                                \
 820                 __page = __page->cp_child;                              \
 821         do {                                                            \
 822                 list_for_each_entry_reverse(__scan, &__page->cp_layers, \
 823                                             cpl_linkage) {              \
 824                         __method = *(void **)((char *)__scan->cpl_ops + \
 825                                               __op);                    \
 826                         if (__method != NULL)                           \
 827                                 (*__method)(__env, __scan,              \
 828                                             ## __VA_ARGS__);            \
 829                 }                                                       \
 830                 __page = __page->cp_parent;                             \
 831         } while (__page != NULL);                                       \
 832 } while (0)
 833
 834 static int cl_page_invoke(const struct lu_env *env,
 835                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 836
 837 {
 838         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 839         ENTRY;
 840         RETURN(CL_PAGE_INVOKE(env, page, op,
 841                               (const struct lu_env *,
 842                                const struct cl_page_slice *, struct cl_io *),
 843                               io));
 844 }
 845
 846 static void cl_page_invoid(const struct lu_env *env,
 847                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 848
 849 {
 850         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 851         ENTRY;
 852         CL_PAGE_INVOID(env, page, op,
 853                        (const struct lu_env *,
 854                         const struct cl_page_slice *, struct cl_io *), io);
 855         EXIT;
 856 }
 857
 858 static void cl_page_owner_clear(struct cl_page *page)
 859 {
 860         ENTRY;
 861         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 862                 if (page->cp_owner != NULL) {
 863                         LASSERT(page->cp_owner->ci_owned_nr > 0);
 864                         page->cp_owner->ci_owned_nr--;
 865                         page->cp_owner = NULL;
 866                         page->cp_task = NULL;
 867                 }
 868         }
 869         EXIT;
 870 }
 871
 872 static void cl_page_owner_set(struct cl_page *page)
 873 {
 874         ENTRY;
 875         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 876                 LASSERT(page->cp_owner != NULL);
 877                 page->cp_owner->ci_owned_nr++;
 878         }
 879         EXIT;
 880 }
 881
 882 void cl_page_disown0(const struct lu_env *env,
 883                      struct cl_io *io, struct cl_page *pg)
 884 {
 885         enum cl_page_state state;
 886
 887         ENTRY;
 888         state = pg->cp_state;
 889         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 890         PINVRNT(env, pg, cl_page_invariant(pg));
 891         cl_page_owner_clear(pg);
 892
 893         if (state == CPS_OWNED)
 894                 cl_page_state_set(env, pg, CPS_CACHED);
 895         /*
 896          * Completion call-backs are executed in the bottom-up order, so that
 897          * uppermost layer (llite), responsible for VFS/VM interaction runs
 898          * last and can release locks safely.
 899          */
 900         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 901                                (const struct lu_env *,
 902                                 const struct cl_page_slice *, struct cl_io *),
 903                                io);
 904         EXIT;
 905 }
 906
 907 /**
 908  * returns true, iff page is owned by the given io.
 909  */
 910 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 911 {
 912         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 913         ENTRY;
 914         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == io);
 915 }
 916 EXPORT_SYMBOL(cl_page_is_owned);
 917
 918 /**
 919  * Try to own a page by IO.
 920  *
 921  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 922  * into cl_page_state::CPS_OWNED state.
 923  *
 924  * \pre  !cl_page_is_owned(pg, io)
 925  * \post result == 0 iff cl_page_is_owned(pg, io)
 926  *
 927  * \retval 0   success
 928  *
 929  * \retval -ve failure, e.g., page was destroyed (and landed in
 930  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 931  *             or, page was owned by another thread, or in IO.
 932  *
 933  * \see cl_page_disown()
 934  * \see cl_page_operations::cpo_own()
 935  * \see cl_page_own_try()
 936  * \see cl_page_own
 937  */
 938 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 939                         struct cl_page *pg, int nonblock)
 940 {
 941         int result;
 942
 943         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 944
 945         ENTRY;
 946         pg = cl_page_top(pg);
 947         io = cl_io_top(io);
 948
 949         if (pg->cp_state == CPS_FREEING) {
 950                 result = -EAGAIN;
 951         } else {
 952                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 953                                         (const struct lu_env *,
 954                                          const struct cl_page_slice *,
 955                                          struct cl_io *, int),
 956                                         io, nonblock);
 957                 if (result == 0) {
 958                         PASSERT(env, pg, pg->cp_owner == NULL);
 959                         PASSERT(env, pg, pg->cp_req == NULL);
 960                         pg->cp_owner = io;
 961                         pg->cp_task  = current;
 962                         cl_page_owner_set(pg);
 963                         if (pg->cp_state != CPS_FREEING) {
 964                                 cl_page_state_set(env, pg, CPS_OWNED);
 965                         } else {
 966                                 cl_page_disown0(env, io, pg);
 967                                 result = -EAGAIN;
 968                         }
 969                 }
 970         }
 971         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 972         RETURN(result);
 973 }
 974
 975 /**
 976  * Own a page, might be blocked.
 977  *
 978  * \see cl_page_own0()
 979  */
 980 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 981 {
 982         return cl_page_own0(env, io, pg, 0);
 983 }
 984 EXPORT_SYMBOL(cl_page_own);
 985
 986 /**
 987  * Nonblock version of cl_page_own().
 988  *
 989  * \see cl_page_own0()
 990  */
 991 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
 992                     struct cl_page *pg)
 993 {
 994         return cl_page_own0(env, io, pg, 1);
 995 }
 996 EXPORT_SYMBOL(cl_page_own_try);
 997
 998
 999 /**
1000  * Assume page ownership.
1001  *
1002  * Called when page is already locked by the hosting VM.
1003  *
1004  * \pre !cl_page_is_owned(pg, io)
1005  * \post cl_page_is_owned(pg, io)
1006  *
1007  * \see cl_page_operations::cpo_assume()
1008  */
1009 void cl_page_assume(const struct lu_env *env,
1010                     struct cl_io *io, struct cl_page *pg)
1011 {
1012         PASSERT(env, pg, pg->cp_state < CPS_OWNED);
1013         PASSERT(env, pg, pg->cp_owner == NULL);
1014         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
1015         PINVRNT(env, pg, cl_page_invariant(pg));
1016
1017         ENTRY;
1018         pg = cl_page_top(pg);
1019         io = cl_io_top(io);
1020
1021         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
1022         pg->cp_owner = io;
1023         pg->cp_task = current;
1024         cl_page_owner_set(pg);
1025         cl_page_state_set(env, pg, CPS_OWNED);
1026         EXIT;
1027 }
1028 EXPORT_SYMBOL(cl_page_assume);
1029
1030 /**
1031  * Releases page ownership without unlocking the page.
1032  *
1033  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
1034  * underlying VM page (as VM is supposed to do this itself).
1035  *
1036  * \pre   cl_page_is_owned(pg, io)
1037  * \post !cl_page_is_owned(pg, io)
1038  *
1039  * \see cl_page_assume()
1040  */
1041 void cl_page_unassume(const struct lu_env *env,
1042                       struct cl_io *io, struct cl_page *pg)
1043 {
1044         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1045         PINVRNT(env, pg, cl_page_invariant(pg));
1046
1047         ENTRY;
1048         pg = cl_page_top(pg);
1049         io = cl_io_top(io);
1050         cl_page_owner_clear(pg);
1051         cl_page_state_set(env, pg, CPS_CACHED);
1052         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
1053                                (const struct lu_env *,
1054                                 const struct cl_page_slice *, struct cl_io *),
1055                                io);
1056         EXIT;
1057 }
1058 EXPORT_SYMBOL(cl_page_unassume);
1059
1060 /**
1061  * Releases page ownership.
1062  *
1063  * Moves page into cl_page_state::CPS_CACHED.
1064  *
1065  * \pre   cl_page_is_owned(pg, io)
1066  * \post !cl_page_is_owned(pg, io)
1067  *
1068  * \see cl_page_own()
1069  * \see cl_page_operations::cpo_disown()
1070  */
1071 void cl_page_disown(const struct lu_env *env,
1072                     struct cl_io *io, struct cl_page *pg)
1073 {
1074         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1075
1076         ENTRY;
1077         pg = cl_page_top(pg);
1078         io = cl_io_top(io);
1079         cl_page_disown0(env, io, pg);
1080         EXIT;
1081 }
1082 EXPORT_SYMBOL(cl_page_disown);
1083
1084 /**
1085  * Called when page is to be removed from the object, e.g., as a result of
1086  * truncate.
1087  *
1088  * Calls cl_page_operations::cpo_discard() top-to-bottom.
1089  *
1090  * \pre cl_page_is_owned(pg, io)
1091  *
1092  * \see cl_page_operations::cpo_discard()
1093  */
1094 void cl_page_discard(const struct lu_env *env,
1095                      struct cl_io *io, struct cl_page *pg)
1096 {
1097         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1098         PINVRNT(env, pg, cl_page_invariant(pg));
1099
1100         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
1101 }
1102 EXPORT_SYMBOL(cl_page_discard);
1103
1104 /**
1105  * Version of cl_page_delete() that can be called for not fully constructed
1106  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
1107  * path. Doesn't check page invariant.
1108  */
1109 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
1110                             int radix)
1111 {
1112         struct cl_page *tmp = pg;
1113         ENTRY;
1114
1115         PASSERT(env, pg, pg == cl_page_top(pg));
1116         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
1117
1118         /*
1119          * Severe all ways to obtain new pointers to @pg.
1120          */
1121         cl_page_owner_clear(pg);
1122
1123         /*
1124          * unexport the page firstly before freeing it so that
1125          * the page content is considered to be invalid.
1126          * We have to do this because a CPS_FREEING cl_page may
1127          * be NOT under the protection of a cl_lock.
1128          * Afterwards, if this page is found by other threads, then this
1129          * page will be forced to reread.
1130          */
1131         cl_page_export(env, pg, 0);
1132         cl_page_state_set0(env, pg, CPS_FREEING);
1133
1134         if (!radix)
1135                 /*
1136                  * !radix means that @pg is not yet in the radix tree, skip
1137                  * removing it.
1138                  */
1139                 tmp = pg->cp_child;
1140         for (; tmp != NULL; tmp = tmp->cp_child) {
1141                 void                    *value;
1142                 struct cl_object_header *hdr;
1143
1144                 hdr = cl_object_header(tmp->cp_obj);
1145                 spin_lock(&hdr->coh_page_guard);
1146                 value = radix_tree_delete(&hdr->coh_tree, tmp->cp_index);
1147                 PASSERT(env, tmp, value == tmp);
1148                 PASSERT(env, tmp, hdr->coh_pages > 0);
1149                 hdr->coh_pages--;
1150                 spin_unlock(&hdr->coh_page_guard);
1151         }
1152
1153         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete),
1154                        (const struct lu_env *, const struct cl_page_slice *));
1155         EXIT;
1156 }
1157
1158 /**
1159  * Called when a decision is made to throw page out of memory.
1160  *
1161  * Notifies all layers about page destruction by calling
1162  * cl_page_operations::cpo_delete() method top-to-bottom.
1163  *
1164  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
1165  * where transition to this state happens).
1166  *
1167  * Eliminates all venues through which new references to the page can be
1168  * obtained:
1169  *
1170  *     - removes page from the radix trees,
1171  *
1172  *     - breaks linkage from VM page to cl_page.
1173  *
1174  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
1175  * drain after some time, at which point page will be recycled.
1176  *
1177  * \pre  pg == cl_page_top(pg)
1178  * \pre  VM page is locked
1179  * \post pg->cp_state == CPS_FREEING
1180  *
1181  * \see cl_page_operations::cpo_delete()
1182  */
1183 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
1184 {
1185         PINVRNT(env, pg, cl_page_invariant(pg));
1186         ENTRY;
1187         cl_page_delete0(env, pg, 1);
1188         EXIT;
1189 }
1190 EXPORT_SYMBOL(cl_page_delete);
1191
1192 /**
1193  * Unmaps page from user virtual memory.
1194  *
1195  * Calls cl_page_operations::cpo_unmap() through all layers top-to-bottom. The
1196  * layer responsible for VM interaction has to unmap page from user space
1197  * virtual memory.
1198  *
1199  * \see cl_page_operations::cpo_unmap()
1200  */
1201 int cl_page_unmap(const struct lu_env *env,
1202                   struct cl_io *io, struct cl_page *pg)
1203 {
1204         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1205         PINVRNT(env, pg, cl_page_invariant(pg));
1206
1207         return cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_unmap));
1208 }
1209 EXPORT_SYMBOL(cl_page_unmap);
1210
1211 /**
1212  * Marks page up-to-date.
1213  *
1214  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
1215  * layer responsible for VM interaction has to mark/clear page as up-to-date
1216  * by the \a uptodate argument.
1217  *
1218  * \see cl_page_operations::cpo_export()
1219  */
1220 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
1221 {
1222         PINVRNT(env, pg, cl_page_invariant(pg));
1223         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
1224                        (const struct lu_env *,
1225                         const struct cl_page_slice *, int), uptodate);
1226 }
1227 EXPORT_SYMBOL(cl_page_export);
1228
1229 /**
1230  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
1231  * thread.
1232  */
1233 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
1234 {
1235         int result;
1236         const struct cl_page_slice *slice;
1237
1238         ENTRY;
1239         pg = cl_page_top_trusted((struct cl_page *)pg);
1240         slice = container_of(pg->cp_layers.next,
1241                              const struct cl_page_slice, cpl_linkage);
1242         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
1243         /*
1244          * Call ->cpo_is_vmlocked() directly instead of going through
1245          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
1246          * cl_page_invariant().
1247          */
1248         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
1249         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
1250         RETURN(result == -EBUSY);
1251 }
1252 EXPORT_SYMBOL(cl_page_is_vmlocked);
1253
1254 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
1255 {
1256         ENTRY;
1257         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
1258 }
1259
1260 static void cl_page_io_start(const struct lu_env *env,
1261                              struct cl_page *pg, enum cl_req_type crt)
1262 {
1263         /*
1264          * Page is queued for IO, change its state.
1265          */
1266         ENTRY;
1267         cl_page_owner_clear(pg);
1268         cl_page_state_set(env, pg, cl_req_type_state(crt));
1269         EXIT;
1270 }
1271
1272 /**
1273  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
1274  * called top-to-bottom. Every layer either agrees to submit this page (by
1275  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
1276  * handling interactions with the VM also has to inform VM that page is under
1277  * transfer now.
1278  */
1279 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
1280                  struct cl_page *pg, enum cl_req_type crt)
1281 {
1282         int result;
1283
1284         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1285         PINVRNT(env, pg, cl_page_invariant(pg));
1286         PINVRNT(env, pg, crt < CRT_NR);
1287
1288         /*
1289          * XXX this has to be called bottom-to-top, so that llite can set up
1290          * PG_writeback without risking other layers deciding to skip this
1291          * page.
1292          */
1293         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
1294         if (result == 0)
1295                 cl_page_io_start(env, pg, crt);
1296
1297         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
1298                       equi(result == 0,
1299                            PageWriteback(cl_page_vmpage(env, pg)))));
1300         CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, result);
1301         return result;
1302 }
1303 EXPORT_SYMBOL(cl_page_prep);
1304
1305 /**
1306  * Notify layers about transfer completion.
1307  *
1308  * Invoked by transfer sub-system (which is a part of osc) to notify layers
1309  * that a transfer, of which this page is a part of has completed.
1310  *
1311  * Completion call-backs are executed in the bottom-up order, so that
1312  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1313  * and can release locks safely.
1314  *
1315  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1316  * \post pg->cp_state == CPS_CACHED
1317  *
1318  * \see cl_page_operations::cpo_completion()
1319  */
1320 void cl_page_completion(const struct lu_env *env,
1321                         struct cl_page *pg, enum cl_req_type crt, int ioret)
1322 {
1323         struct cl_sync_io *anchor = pg->cp_sync_io;
1324
1325         PASSERT(env, pg, crt < CRT_NR);
1326         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
1327         PASSERT(env, pg, pg->cp_req == NULL);
1328         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
1329         PINVRNT(env, pg, cl_page_invariant(pg));
1330
1331         ENTRY;
1332         CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, ioret);
1333         if (crt == CRT_READ && ioret == 0) {
1334                 PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED));
1335                 pg->cp_flags |= CPF_READ_COMPLETED;
1336         }
1337
1338         cl_page_state_set(env, pg, CPS_CACHED);
1339         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
1340                                (const struct lu_env *,
1341                                 const struct cl_page_slice *, int), ioret);
1342         if (anchor) {
1343                 LASSERT(pg->cp_sync_io == anchor);
1344                 pg->cp_sync_io = NULL;
1345                 cl_sync_io_note(anchor, ioret);
1346         }
1347
1348         /* Don't assert the page writeback bit here because the lustre file
1349          * may be as a backend of swap space. in this case, the page writeback
1350          * is set by VM, and obvious we shouldn't clear it at all. Fortunately
1351          * this type of pages are all TRANSIENT pages. */
1352         KLASSERT(ergo(pg->cp_type == CPT_CACHEABLE,
1353                       !PageWriteback(cl_page_vmpage(env, pg))));
1354         EXIT;
1355 }
1356 EXPORT_SYMBOL(cl_page_completion);
1357
1358 /**
1359  * Notify layers that transfer formation engine decided to yank this page from
1360  * the cache and to make it a part of a transfer.
1361  *
1362  * \pre  pg->cp_state == CPS_CACHED
1363  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1364  *
1365  * \see cl_page_operations::cpo_make_ready()
1366  */
1367 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
1368                        enum cl_req_type crt)
1369 {
1370         int result;
1371
1372         PINVRNT(env, pg, crt < CRT_NR);
1373
1374         ENTRY;
1375         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
1376                                 (const struct lu_env *,
1377                                  const struct cl_page_slice *));
1378         if (result == 0) {
1379                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
1380                 cl_page_io_start(env, pg, crt);
1381         }
1382         CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, result);
1383         RETURN(result);
1384 }
1385 EXPORT_SYMBOL(cl_page_make_ready);
1386
1387 /**
1388  * Notify layers that high level io decided to place this page into a cache
1389  * for future transfer.
1390  *
1391  * The layer implementing transfer engine (osc) has to register this page in
1392  * its queues.
1393  *
1394  * \pre  cl_page_is_owned(pg, io)
1395  * \post ergo(result == 0,
1396  *            pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT)
1397  *
1398  * \see cl_page_operations::cpo_cache_add()
1399  */
1400 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
1401                       struct cl_page *pg, enum cl_req_type crt)
1402 {
1403         int result;
1404
1405         PINVRNT(env, pg, crt < CRT_NR);
1406         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1407         PINVRNT(env, pg, cl_page_invariant(pg));
1408
1409         ENTRY;
1410         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_cache_add));
1411         if (result == 0) {
1412                 cl_page_owner_clear(pg);
1413                 cl_page_state_set(env, pg, CPS_CACHED);
1414         }
1415         CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", crt, result);
1416         RETURN(result);
1417 }
1418 EXPORT_SYMBOL(cl_page_cache_add);
1419
1420 /**
1421  * Checks whether page is protected by any extent lock is at least required
1422  * mode.
1423  *
1424  * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1425  * \see cl_page_operations::cpo_is_under_lock()
1426  */
1427 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1428                           struct cl_page *page)
1429 {
1430         int rc;
1431
1432         PINVRNT(env, page, cl_page_invariant(page));
1433
1434         ENTRY;
1435         rc = CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1436                             (const struct lu_env *,
1437                              const struct cl_page_slice *, struct cl_io *),
1438                             io);
1439         PASSERT(env, page, rc != 0);
1440         RETURN(rc);
1441 }
1442 EXPORT_SYMBOL(cl_page_is_under_lock);
1443
1444 /**
1445  * Purges all cached pages belonging to the object \a obj.
1446  */
1447 int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj)
1448 {
1449         struct cl_thread_info   *info;
1450         struct cl_object        *obj = cl_object_top(clobj);
1451         struct cl_io            *io;
1452         struct cl_page_list     *plist;
1453         int                      result;
1454
1455         ENTRY;
1456         info  = cl_env_info(env);
1457         plist = &info->clt_list;
1458         io    = &info->clt_io;
1459
1460         /*
1461          * initialize the io. This is ugly since we never do IO in this
1462          * function, we just make cl_page_list functions happy. -jay
1463          */
1464         io->ci_obj = obj;
1465         result = cl_io_init(env, io, CIT_MISC, obj);
1466         if (result != 0) {
1467                 cl_io_fini(env, io);
1468                 RETURN(io->ci_result);
1469         }
1470
1471         cl_page_list_init(plist);
1472         cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF, plist, 0);
1473         /*
1474          * Since we're purging the pages of an object, we don't care
1475          * the possible outcomes of the following functions.
1476          */
1477         cl_page_list_unmap(env, io, plist);
1478         cl_page_list_discard(env, io, plist);
1479         cl_page_list_disown(env, io, plist);
1480         cl_page_list_fini(env, plist);
1481
1482         cl_io_fini(env, io);
1483         RETURN(result);
1484 }
1485 EXPORT_SYMBOL(cl_pages_prune);
1486
1487 /**
1488  * Tells transfer engine that only part of a page is to be transmitted.
1489  *
1490  * \see cl_page_operations::cpo_clip()
1491  */
1492 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1493                   int from, int to)
1494 {
1495         PINVRNT(env, pg, cl_page_invariant(pg));
1496
1497         CL_PAGE_HEADER(D_TRACE, env, pg, "%i %i\n", from, to);
1498         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1499                        (const struct lu_env *,
1500                         const struct cl_page_slice *,int, int),
1501                        from, to);
1502 }
1503 EXPORT_SYMBOL(cl_page_clip);
1504
1505 /**
1506  * Prints human readable representation of \a pg to the \a f.
1507  */
1508 void cl_page_header_print(const struct lu_env *env, void *cookie,
1509                           lu_printer_t printer, const struct cl_page *pg)
1510 {
1511         (*printer)(env, cookie,
1512                    "page@%p[%d %p:%lu ^%p_%p %d %d %d %p %p %#x]\n",
1513                    pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1514                    pg->cp_index, pg->cp_parent, pg->cp_child,
1515                    pg->cp_state, pg->cp_error, pg->cp_type,
1516                    pg->cp_owner, pg->cp_req, pg->cp_flags);
1517 }
1518 EXPORT_SYMBOL(cl_page_header_print);
1519
1520 /**
1521  * Prints human readable representation of \a pg to the \a f.
1522  */
1523 void cl_page_print(const struct lu_env *env, void *cookie,
1524                    lu_printer_t printer, const struct cl_page *pg)
1525 {
1526         struct cl_page *scan;
1527
1528         for (scan = cl_page_top((struct cl_page *)pg);
1529              scan != NULL; scan = scan->cp_child)
1530                 cl_page_header_print(env, cookie, printer, scan);
1531         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1532                        (const struct lu_env *env,
1533                         const struct cl_page_slice *slice,
1534                         void *cookie, lu_printer_t p), cookie, printer);
1535         (*printer)(env, cookie, "end page@%p\n", pg);
1536 }
1537 EXPORT_SYMBOL(cl_page_print);
1538
1539 /**
1540  * Cancel a page which is still in a transfer.
1541  */
1542 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1543 {
1544         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1545                               (const struct lu_env *,
1546                                const struct cl_page_slice *));
1547 }
1548 EXPORT_SYMBOL(cl_page_cancel);
1549
1550 /**
1551  * Converts a byte offset within object \a obj into a page index.
1552  */
1553 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1554 {
1555         /*
1556          * XXX for now.
1557          */
1558         return (loff_t)idx << CFS_PAGE_SHIFT;
1559 }
1560 EXPORT_SYMBOL(cl_offset);
1561
1562 /**
1563  * Converts a page index into a byte offset within object \a obj.
1564  */
1565 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1566 {
1567         /*
1568          * XXX for now.
1569          */
1570         return offset >> CFS_PAGE_SHIFT;
1571 }
1572 EXPORT_SYMBOL(cl_index);
1573
1574 int cl_page_size(const struct cl_object *obj)
1575 {
1576         return 1 << CFS_PAGE_SHIFT;
1577 }
1578 EXPORT_SYMBOL(cl_page_size);
1579
1580 /**
1581  * Adds page slice to the compound page.
1582  *
1583  * This is called by cl_object_operations::coo_page_init() methods to add a
1584  * per-layer state to the page. New state is added at the end of
1585  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1586  *
1587  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1588  */
1589 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1590                        struct cl_object *obj,
1591                        const struct cl_page_operations *ops)
1592 {
1593         ENTRY;
1594         list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1595         slice->cpl_obj  = obj;
1596         slice->cpl_ops  = ops;
1597         slice->cpl_page = page;
1598         EXIT;
1599 }
1600 EXPORT_SYMBOL(cl_page_slice_add);
1601
1602 int  cl_page_init(void)
1603 {
1604         return lu_kmem_init(cl_page_caches);
1605 }
1606
1607 void cl_page_fini(void)
1608 {
1609         lu_kmem_fini(cl_page_caches);
1610 }