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 /*
  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 (cfs_atomic_inc_return(&page->cp_ref) == 1)
 126                 cfs_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 (!cfs_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                 cfs_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  * \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 void cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj,
 192                          struct cl_io *io, pgoff_t start, pgoff_t end,
 193                          struct cl_page_list *queue, int nonblock,
 194                          int *resched)
 195 {
 196         struct cl_object_header *hdr;
 197         struct cl_page          *page;
 198         struct cl_page         **pvec;
 199         const struct cl_page_slice  *slice;
 200         const struct lu_device_type *dtype;
 201         pgoff_t                  idx;
 202         unsigned int             nr;
 203         unsigned int             i;
 204         unsigned int             j;
 205         int                    (*page_own)(const struct lu_env *env,
 206                                            struct cl_io *io,
 207                                            struct cl_page *pg);
 208         ENTRY;
 209
 210         if (resched != NULL)
 211                 *resched = 0;
 212         page_own = nonblock ? cl_page_own_try : cl_page_own;
 213
 214         idx = start;
 215         hdr = cl_object_header(obj);
 216         pvec = cl_env_info(env)->clt_pvec;
 217         dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type;
 218         cfs_spin_lock(&hdr->coh_page_guard);
 219         while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec,
 220                                             idx, CLT_PVEC_SIZE)) > 0) {
 221                 idx = pvec[nr - 1]->cp_index + 1;
 222                 for (i = 0, j = 0; i < nr; ++i) {
 223                         page = pvec[i];
 224                         PASSERT(env, page, cl_is_page(page));
 225                         pvec[i] = NULL;
 226                         if (page->cp_index > end)
 227                                 break;
 228                         if (page->cp_state == CPS_FREEING)
 229                                 continue;
 230                         if (page->cp_type == CPT_TRANSIENT) {
 231                                 /* God, we found a transient page!*/
 232                                 continue;
 233                         }
 234
 235                         slice = cl_page_at_trusted(page, dtype);
 236                         /*
 237                          * Pages for lsm-less file has no underneath sub-page
 238                          * for osc, in case of ...
 239                          */
 240                         PASSERT(env, page, slice != NULL);
 241
 242                         page = slice->cpl_page;
 243                         /*
 244                          * Can safely call cl_page_get_trust() under
 245                          * radix-tree spin-lock.
 246                          *
 247                          * XXX not true, because @page is from object another
 248                          * than @hdr and protected by different tree lock.
 249                          */
 250                         cl_page_get_trust(page);
 251                         lu_ref_add_atomic(&page->cp_reference,
 252                                           "page_list", cfs_current());
 253                         pvec[j++] = page;
 254                 }
 255
 256                 /*
 257                  * Here a delicate locking dance is performed. Current thread
 258                  * holds a reference to a page, but has to own it before it
 259                  * can be placed into queue. Owning implies waiting, so
 260                  * radix-tree lock is to be released. After a wait one has to
 261                  * check that pages weren't truncated (cl_page_own() returns
 262                  * error in the latter case).
 263                  */
 264                 cfs_spin_unlock(&hdr->coh_page_guard);
 265                 for (i = 0; i < j; ++i) {
 266                         page = pvec[i];
 267                         if (page_own(env, io, page) == 0)
 268                                 cl_page_list_add(queue, page);
 269                         lu_ref_del(&page->cp_reference,
 270                                    "page_list", cfs_current());
 271                         cl_page_put(env, page);
 272                 }
 273                 cfs_spin_lock(&hdr->coh_page_guard);
 274                 if (nr < CLT_PVEC_SIZE)
 275                         break;
 276                 if (resched != NULL && cfs_need_resched()) {
 277                         *resched = 1;
 278                         break;
 279                 }
 280         }
 281         cfs_spin_unlock(&hdr->coh_page_guard);
 282         EXIT;
 283 }
 284 EXPORT_SYMBOL(cl_page_gang_lookup);
 285
 286 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
 287 {
 288         struct cl_object *obj  = page->cp_obj;
 289         struct cl_site   *site = cl_object_site(obj);
 290
 291         PASSERT(env, page, cl_is_page(page));
 292         PASSERT(env, page, cfs_list_empty(&page->cp_batch));
 293         PASSERT(env, page, page->cp_owner == NULL);
 294         PASSERT(env, page, page->cp_req == NULL);
 295         PASSERT(env, page, page->cp_parent == NULL);
 296         PASSERT(env, page, page->cp_state == CPS_FREEING);
 297
 298         ENTRY;
 299         cfs_might_sleep();
 300         while (!cfs_list_empty(&page->cp_layers)) {
 301                 struct cl_page_slice *slice;
 302
 303                 slice = cfs_list_entry(page->cp_layers.next,
 304                                        struct cl_page_slice, cpl_linkage);
 305                 cfs_list_del_init(page->cp_layers.next);
 306                 slice->cpl_ops->cpo_fini(env, slice);
 307         }
 308         cfs_atomic_dec(&site->cs_pages.cs_total);
 309         cfs_atomic_dec(&site->cs_pages_state[page->cp_state]);
 310         lu_object_ref_del_at(&obj->co_lu, page->cp_obj_ref, "cl_page", page);
 311         cl_object_put(env, obj);
 312         lu_ref_fini(&page->cp_reference);
 313         OBD_SLAB_FREE_PTR(page, cl_page_kmem);
 314         EXIT;
 315 }
 316
 317 /**
 318  * Helper function updating page state. This is the only place in the code
 319  * where cl_page::cp_state field is mutated.
 320  */
 321 static inline void cl_page_state_set_trust(struct cl_page *page,
 322                                            enum cl_page_state state)
 323 {
 324         /* bypass const. */
 325         *(enum cl_page_state *)&page->cp_state = state;
 326 }
 327
 328 static int cl_page_alloc(const struct lu_env *env, struct cl_object *o,
 329                          pgoff_t ind, struct page *vmpage,
 330                          enum cl_page_type type, struct cl_page **out)
 331 {
 332         struct cl_page          *page;
 333         struct cl_page          *err  = NULL;
 334         struct lu_object_header *head;
 335         struct cl_site          *site = cl_object_site(o);
 336         int                      result;
 337
 338         ENTRY;
 339         result = +1;
 340         OBD_SLAB_ALLOC_PTR_GFP(page, cl_page_kmem, CFS_ALLOC_IO);
 341         if (page != NULL) {
 342                 cfs_atomic_set(&page->cp_ref, 1);
 343                 page->cp_obj = o;
 344                 cl_object_get(o);
 345                 page->cp_obj_ref = lu_object_ref_add(&o->co_lu,
 346                                                      "cl_page", page);
 347                 page->cp_index = ind;
 348                 cl_page_state_set_trust(page, CPS_CACHED);
 349                 page->cp_type = type;
 350                 CFS_INIT_LIST_HEAD(&page->cp_layers);
 351                 CFS_INIT_LIST_HEAD(&page->cp_batch);
 352                 CFS_INIT_LIST_HEAD(&page->cp_flight);
 353                 cfs_mutex_init(&page->cp_mutex);
 354                 lu_ref_init(&page->cp_reference);
 355                 head = o->co_lu.lo_header;
 356                 cfs_list_for_each_entry(o, &head->loh_layers,
 357                                         co_lu.lo_linkage) {
 358                         if (o->co_ops->coo_page_init != NULL) {
 359                                 err = o->co_ops->coo_page_init(env, o,
 360                                                                page, vmpage);
 361                                 if (err != NULL) {
 362                                         cl_page_state_set_trust(page,
 363                                                                 CPS_FREEING);
 364                                         cl_page_free(env, page);
 365                                         page = err;
 366                                         break;
 367                                 }
 368                         }
 369                 }
 370                 if (err == NULL) {
 371                         cfs_atomic_inc(&site->cs_pages.cs_busy);
 372                         cfs_atomic_inc(&site->cs_pages.cs_total);
 373                         cfs_atomic_inc(&site->cs_pages_state[CPS_CACHED]);
 374                         cfs_atomic_inc(&site->cs_pages.cs_created);
 375                         result = 0;
 376                 }
 377         } else
 378                 page = ERR_PTR(-ENOMEM);
 379         *out = page;
 380         RETURN(result);
 381 }
 382
 383 /**
 384  * Returns a cl_page with index \a idx at the object \a o, and associated with
 385  * the VM page \a vmpage.
 386  *
 387  * This is the main entry point into the cl_page caching interface. First, a
 388  * cache (implemented as a per-object radix tree) is consulted. If page is
 389  * found there, it is returned immediately. Otherwise new page is allocated
 390  * and returned. In any case, additional reference to page is acquired.
 391  *
 392  * \see cl_object_find(), cl_lock_find()
 393  */
 394 static struct cl_page *cl_page_find0(const struct lu_env *env,
 395                                      struct cl_object *o,
 396                                      pgoff_t idx, struct page *vmpage,
 397                                      enum cl_page_type type,
 398                                      struct cl_page *parent)
 399 {
 400         struct cl_page          *page;
 401         struct cl_page          *ghost = NULL;
 402         struct cl_object_header *hdr;
 403         struct cl_site          *site = cl_object_site(o);
 404         int err;
 405
 406         LINVRNT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 407         cfs_might_sleep();
 408
 409         ENTRY;
 410
 411         hdr = cl_object_header(o);
 412         cfs_atomic_inc(&site->cs_pages.cs_lookup);
 413
 414         CDEBUG(D_PAGE, "%lu@"DFID" %p %lu %d\n",
 415                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 416         /* fast path. */
 417         if (type == CPT_CACHEABLE) {
 418                 /*
 419                  * cl_vmpage_page() can be called here without any locks as
 420                  *
 421                  *     - "vmpage" is locked (which prevents ->private from
 422                  *       concurrent updates), and
 423                  *
 424                  *     - "o" cannot be destroyed while current thread holds a
 425                  *       reference on it.
 426                  */
 427                 page = cl_vmpage_page(vmpage, o);
 428                 PINVRNT(env, page,
 429                         ergo(page != NULL,
 430                              cl_page_vmpage(env, page) == vmpage &&
 431                              (void *)radix_tree_lookup(&hdr->coh_tree,
 432                                                        idx) == page));
 433         } else {
 434                 cfs_spin_lock(&hdr->coh_page_guard);
 435                 page = cl_page_lookup(hdr, idx);
 436                 cfs_spin_unlock(&hdr->coh_page_guard);
 437         }
 438         if (page != NULL) {
 439                 cfs_atomic_inc(&site->cs_pages.cs_hit);
 440                 RETURN(page);
 441         }
 442
 443         /* allocate and initialize cl_page */
 444         err = cl_page_alloc(env, o, idx, vmpage, type, &page);
 445         if (err != 0)
 446                 RETURN(page);
 447         /*
 448          * XXX optimization: use radix_tree_preload() here, and change tree
 449          * gfp mask to GFP_KERNEL in cl_object_header_init().
 450          */
 451         cfs_spin_lock(&hdr->coh_page_guard);
 452         err = radix_tree_insert(&hdr->coh_tree, idx, page);
 453         if (err != 0) {
 454                 ghost = page;
 455                 /*
 456                  * Noted by Jay: a lock on \a vmpage protects cl_page_find()
 457                  * from this race, but
 458                  *
 459                  *     0. it's better to have cl_page interface "locally
 460                  *     consistent" so that its correctness can be reasoned
 461                  *     about without appealing to the (obscure world of) VM
 462                  *     locking.
 463                  *
 464                  *     1. handling this race allows ->coh_tree to remain
 465                  *     consistent even when VM locking is somehow busted,
 466                  *     which is very useful during diagnosing and debugging.
 467                  */
 468                 page = ERR_PTR(err);
 469                 if (err == -EEXIST) {
 470                         /*
 471                          * XXX in case of a lookup for CPT_TRANSIENT page,
 472                          * nothing protects a CPT_CACHEABLE page from being
 473                          * concurrently moved into CPS_FREEING state.
 474                          */
 475                         page = cl_page_lookup(hdr, idx);
 476                         PASSERT(env, page, page != NULL);
 477                         if (page->cp_type == CPT_TRANSIENT &&
 478                             type == CPT_CACHEABLE) {
 479                                 /* XXX: We should make sure that inode sem
 480                                  * keeps being held in the lifetime of
 481                                  * transient pages, so it is impossible to
 482                                  * have conflicting transient pages.
 483                                  */
 484                                 cfs_spin_unlock(&hdr->coh_page_guard);
 485                                 cl_page_put(env, page);
 486                                 cfs_spin_lock(&hdr->coh_page_guard);
 487                                 page = ERR_PTR(-EBUSY);
 488                         }
 489                 }
 490         } else {
 491                 if (parent) {
 492                         LASSERT(page->cp_parent == NULL);
 493                         page->cp_parent = parent;
 494                         parent->cp_child = page;
 495                 }
 496                 hdr->coh_pages++;
 497         }
 498         cfs_spin_unlock(&hdr->coh_page_guard);
 499
 500         if (unlikely(ghost != NULL)) {
 501                 cfs_atomic_dec(&site->cs_pages.cs_busy);
 502                 cl_page_delete0(env, ghost, 0);
 503                 cl_page_free(env, ghost);
 504         }
 505         RETURN(page);
 506 }
 507
 508 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o,
 509                              pgoff_t idx, struct page *vmpage,
 510                              enum cl_page_type type)
 511 {
 512         return cl_page_find0(env, o, idx, vmpage, type, NULL);
 513 }
 514 EXPORT_SYMBOL(cl_page_find);
 515
 516
 517 struct cl_page *cl_page_find_sub(const struct lu_env *env, struct cl_object *o,
 518                                  pgoff_t idx, struct page *vmpage,
 519                                  struct cl_page *parent)
 520 {
 521         return cl_page_find0(env, o, idx, vmpage, parent->cp_type, parent);
 522 }
 523 EXPORT_SYMBOL(cl_page_find_sub);
 524
 525 static inline int cl_page_invariant(const struct cl_page *pg)
 526 {
 527         struct cl_object_header *header;
 528         struct cl_page          *parent;
 529         struct cl_page          *child;
 530         struct cl_io            *owner;
 531
 532         LASSERT(cl_is_page(pg));
 533         /*
 534          * Page invariant is protected by a VM lock.
 535          */
 536         LINVRNT(cl_page_is_vmlocked(NULL, pg));
 537
 538         header = cl_object_header(pg->cp_obj);
 539         parent = pg->cp_parent;
 540         child  = pg->cp_child;
 541         owner  = pg->cp_owner;
 542
 543         return cfs_atomic_read(&pg->cp_ref) > 0 &&
 544                 ergo(parent != NULL, parent->cp_child == pg) &&
 545                 ergo(child != NULL, child->cp_parent == pg) &&
 546                 ergo(child != NULL, pg->cp_obj != child->cp_obj) &&
 547                 ergo(parent != NULL, pg->cp_obj != parent->cp_obj) &&
 548                 ergo(owner != NULL && parent != NULL,
 549                      parent->cp_owner == pg->cp_owner->ci_parent) &&
 550                 ergo(owner != NULL && child != NULL,
 551                      child->cp_owner->ci_parent == owner) &&
 552                 /*
 553                  * Either page is early in initialization (has neither child
 554                  * nor parent yet), or it is in the object radix tree.
 555                  */
 556                 ergo(pg->cp_state < CPS_FREEING,
 557                      (void *)radix_tree_lookup(&header->coh_tree,
 558                                                pg->cp_index) == pg ||
 559                      (child == NULL && parent == NULL));
 560 }
 561
 562 static void cl_page_state_set0(const struct lu_env *env,
 563                                struct cl_page *page, enum cl_page_state state)
 564 {
 565         enum cl_page_state old;
 566         struct cl_site *site = cl_object_site(page->cp_obj);
 567
 568         /*
 569          * Matrix of allowed state transitions [old][new], for sanity
 570          * checking.
 571          */
 572         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 573                 [CPS_CACHED] = {
 574                         [CPS_CACHED]  = 0,
 575                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 576                         [CPS_PAGEIN]  = 0,
 577                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 578                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 579                 },
 580                 [CPS_OWNED] = {
 581                         [CPS_CACHED]  = 1, /* release to the cache */
 582                         [CPS_OWNED]   = 0,
 583                         [CPS_PAGEIN]  = 1, /* start read immediately */
 584                         [CPS_PAGEOUT] = 1, /* start write immediately */
 585                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 586                 },
 587                 [CPS_PAGEIN] = {
 588                         [CPS_CACHED]  = 1, /* io completion */
 589                         [CPS_OWNED]   = 0,
 590                         [CPS_PAGEIN]  = 0,
 591                         [CPS_PAGEOUT] = 0,
 592                         [CPS_FREEING] = 0,
 593                 },
 594                 [CPS_PAGEOUT] = {
 595                         [CPS_CACHED]  = 1, /* io completion */
 596                         [CPS_OWNED]   = 0,
 597                         [CPS_PAGEIN]  = 0,
 598                         [CPS_PAGEOUT] = 0,
 599                         [CPS_FREEING] = 0,
 600                 },
 601                 [CPS_FREEING] = {
 602                         [CPS_CACHED]  = 0,
 603                         [CPS_OWNED]   = 0,
 604                         [CPS_PAGEIN]  = 0,
 605                         [CPS_PAGEOUT] = 0,
 606                         [CPS_FREEING] = 0,
 607                 }
 608         };
 609
 610         ENTRY;
 611         old = page->cp_state;
 612         PASSERT(env, page, allowed_transitions[old][state]);
 613         CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
 614         for (; page != NULL; page = page->cp_child) {
 615                 PASSERT(env, page, page->cp_state == old);
 616                 PASSERT(env, page,
 617                         equi(state == CPS_OWNED, page->cp_owner != NULL));
 618
 619                 cfs_atomic_dec(&site->cs_pages_state[page->cp_state]);
 620                 cfs_atomic_inc(&site->cs_pages_state[state]);
 621                 cl_page_state_set_trust(page, state);
 622         }
 623         EXIT;
 624 }
 625
 626 static void cl_page_state_set(const struct lu_env *env,
 627                               struct cl_page *page, enum cl_page_state state)
 628 {
 629         PINVRNT(env, page, cl_page_invariant(page));
 630         cl_page_state_set0(env, page, state);
 631 }
 632
 633 /**
 634  * Acquires an additional reference to a page.
 635  *
 636  * This can be called only by caller already possessing a reference to \a
 637  * page.
 638  *
 639  * \see cl_object_get(), cl_lock_get().
 640  */
 641 void cl_page_get(struct cl_page *page)
 642 {
 643         ENTRY;
 644         LASSERT(page->cp_state != CPS_FREEING);
 645         cl_page_get_trust(page);
 646         EXIT;
 647 }
 648 EXPORT_SYMBOL(cl_page_get);
 649
 650 /**
 651  * Releases a reference to a page.
 652  *
 653  * When last reference is released, page is returned to the cache, unless it
 654  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 655  * destroyed.
 656  *
 657  * \see cl_object_put(), cl_lock_put().
 658  */
 659 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 660 {
 661         struct cl_object_header *hdr;
 662         struct cl_site *site = cl_object_site(page->cp_obj);
 663
 664         PASSERT(env, page, cfs_atomic_read(&page->cp_ref) > !!page->cp_parent);
 665
 666         ENTRY;
 667         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 668                        cfs_atomic_read(&page->cp_ref));
 669
 670         hdr = cl_object_header(cl_object_top(page->cp_obj));
 671         if (cfs_atomic_dec_and_lock(&page->cp_ref, &hdr->coh_page_guard)) {
 672                 cfs_atomic_dec(&site->cs_pages.cs_busy);
 673                 /* We're going to access the page w/o a reference, but it's
 674                  * ok because we have grabbed the lock coh_page_guard, which
 675                  * means nobody is able to free this page behind us.
 676                  */
 677                 if (page->cp_state == CPS_FREEING) {
 678                         /* We drop the page reference and check the page state
 679                          * inside the coh_page_guard. So that if it gets here,
 680                          * it is the REALLY last reference to this page.
 681                          */
 682                         cfs_spin_unlock(&hdr->coh_page_guard);
 683
 684                         LASSERT(cfs_atomic_read(&page->cp_ref) == 0);
 685                         PASSERT(env, page, page->cp_owner == NULL);
 686                         PASSERT(env, page, cfs_list_empty(&page->cp_batch));
 687                         /*
 688                          * Page is no longer reachable by other threads. Tear
 689                          * it down.
 690                          */
 691                         cl_page_free(env, page);
 692
 693                         EXIT;
 694                         return;
 695                 }
 696                 cfs_spin_unlock(&hdr->coh_page_guard);
 697         }
 698
 699         EXIT;
 700 }
 701 EXPORT_SYMBOL(cl_page_put);
 702
 703 /**
 704  * Returns a VM page associated with a given cl_page.
 705  */
 706 cfs_page_t *cl_page_vmpage(const struct lu_env *env, struct cl_page *page)
 707 {
 708         const struct cl_page_slice *slice;
 709
 710         /*
 711          * Find uppermost layer with ->cpo_vmpage() method, and return its
 712          * result.
 713          */
 714         page = cl_page_top(page);
 715         do {
 716                 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 717                         if (slice->cpl_ops->cpo_vmpage != NULL)
 718                                 RETURN(slice->cpl_ops->cpo_vmpage(env, slice));
 719                 }
 720                 page = page->cp_child;
 721         } while (page != NULL);
 722         LBUG(); /* ->cpo_vmpage() has to be defined somewhere in the stack */
 723 }
 724 EXPORT_SYMBOL(cl_page_vmpage);
 725
 726 /**
 727  * Returns a cl_page associated with a VM page, and given cl_object.
 728  */
 729 struct cl_page *cl_vmpage_page(cfs_page_t *vmpage, struct cl_object *obj)
 730 {
 731         struct cl_page *page;
 732         struct cl_object_header *hdr;
 733
 734         ENTRY;
 735         KLASSERT(PageLocked(vmpage));
 736
 737         /*
 738          * NOTE: absence of races and liveness of data are guaranteed by page
 739          *       lock on a "vmpage". That works because object destruction has
 740          *       bottom-to-top pass.
 741          */
 742
 743         /*
 744          * This loop assumes that ->private points to the top-most page. This
 745          * can be rectified easily.
 746          */
 747         hdr = cl_object_header(cl_object_top(obj));
 748         cfs_spin_lock(&hdr->coh_page_guard);
 749         for (page = (void *)vmpage->private;
 750              page != NULL; page = page->cp_child) {
 751                 if (cl_object_same(page->cp_obj, obj)) {
 752                         cl_page_get_trust(page);
 753                         break;
 754                 }
 755         }
 756         cfs_spin_unlock(&hdr->coh_page_guard);
 757         LASSERT(ergo(page, cl_is_page(page) && page->cp_type == CPT_CACHEABLE));
 758         RETURN(page);
 759 }
 760 EXPORT_SYMBOL(cl_vmpage_page);
 761
 762 /**
 763  * Returns the top-page for a given page.
 764  *
 765  * \see cl_object_top(), cl_io_top()
 766  */
 767 struct cl_page *cl_page_top(struct cl_page *page)
 768 {
 769         return cl_page_top_trusted(page);
 770 }
 771 EXPORT_SYMBOL(cl_page_top);
 772
 773 /**
 774  * Returns true if \a addr is an address of an allocated cl_page. Used in
 775  * assertions. This check is optimistically imprecise, i.e., it occasionally
 776  * returns true for the incorrect addresses, but if it returns false, then the
 777  * address is guaranteed to be incorrect. (Should be named cl_pagep().)
 778  *
 779  * \see cl_is_lock()
 780  */
 781 int cl_is_page(const void *addr)
 782 {
 783         return cfs_mem_is_in_cache(addr, cl_page_kmem);
 784 }
 785 EXPORT_SYMBOL(cl_is_page);
 786
 787 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 788                                        const struct lu_device_type *dtype)
 789 {
 790         return cl_page_at_trusted(page, dtype);
 791 }
 792 EXPORT_SYMBOL(cl_page_at);
 793
 794 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 795
 796 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)                   \
 797 ({                                                                      \
 798         const struct lu_env        *__env  = (_env);                    \
 799         struct cl_page             *__page = (_page);                   \
 800         const struct cl_page_slice *__scan;                             \
 801         int                         __result;                           \
 802         ptrdiff_t                   __op   = (_op);                     \
 803         int                       (*__method)_proto;                    \
 804                                                                         \
 805         __result = 0;                                                   \
 806         __page = cl_page_top(__page);                                   \
 807         do {                                                            \
 808                 cfs_list_for_each_entry(__scan, &__page->cp_layers,     \
 809                                         cpl_linkage) {                  \
 810                         __method = *(void **)((char *)__scan->cpl_ops + \
 811                                               __op);                    \
 812                         if (__method != NULL) {                         \
 813                                 __result = (*__method)(__env, __scan,   \
 814                                                        ## __VA_ARGS__); \
 815                                 if (__result != 0)                      \
 816                                         break;                          \
 817                         }                                               \
 818                 }                                                       \
 819                 __page = __page->cp_child;                              \
 820         } while (__page != NULL && __result == 0);                      \
 821         if (__result > 0)                                               \
 822                 __result = 0;                                           \
 823         __result;                                                       \
 824 })
 825
 826 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 827 do {                                                                    \
 828         const struct lu_env        *__env  = (_env);                    \
 829         struct cl_page             *__page = (_page);                   \
 830         const struct cl_page_slice *__scan;                             \
 831         ptrdiff_t                   __op   = (_op);                     \
 832         void                      (*__method)_proto;                    \
 833                                                                         \
 834         __page = cl_page_top(__page);                                   \
 835         do {                                                            \
 836                 cfs_list_for_each_entry(__scan, &__page->cp_layers,     \
 837                                         cpl_linkage) {                  \
 838                         __method = *(void **)((char *)__scan->cpl_ops + \
 839                                               __op);                    \
 840                         if (__method != NULL)                           \
 841                                 (*__method)(__env, __scan,              \
 842                                             ## __VA_ARGS__);            \
 843                 }                                                       \
 844                 __page = __page->cp_child;                              \
 845         } while (__page != NULL);                                       \
 846 } while (0)
 847
 848 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)               \
 849 do {                                                                        \
 850         const struct lu_env        *__env  = (_env);                        \
 851         struct cl_page             *__page = (_page);                       \
 852         const struct cl_page_slice *__scan;                                 \
 853         ptrdiff_t                   __op   = (_op);                         \
 854         void                      (*__method)_proto;                        \
 855                                                                             \
 856         /* get to the bottom page. */                                       \
 857         while (__page->cp_child != NULL)                                    \
 858                 __page = __page->cp_child;                                  \
 859         do {                                                                \
 860                 cfs_list_for_each_entry_reverse(__scan, &__page->cp_layers, \
 861                                                 cpl_linkage) {              \
 862                         __method = *(void **)((char *)__scan->cpl_ops +     \
 863                                               __op);                        \
 864                         if (__method != NULL)                               \
 865                                 (*__method)(__env, __scan,                  \
 866                                             ## __VA_ARGS__);                \
 867                 }                                                           \
 868                 __page = __page->cp_parent;                                 \
 869         } while (__page != NULL);                                           \
 870 } while (0)
 871
 872 static int cl_page_invoke(const struct lu_env *env,
 873                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 874
 875 {
 876         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 877         ENTRY;
 878         RETURN(CL_PAGE_INVOKE(env, page, op,
 879                               (const struct lu_env *,
 880                                const struct cl_page_slice *, struct cl_io *),
 881                               io));
 882 }
 883
 884 static void cl_page_invoid(const struct lu_env *env,
 885                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 886
 887 {
 888         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 889         ENTRY;
 890         CL_PAGE_INVOID(env, page, op,
 891                        (const struct lu_env *,
 892                         const struct cl_page_slice *, struct cl_io *), io);
 893         EXIT;
 894 }
 895
 896 static void cl_page_owner_clear(struct cl_page *page)
 897 {
 898         ENTRY;
 899         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 900                 if (page->cp_owner != NULL) {
 901                         LASSERT(page->cp_owner->ci_owned_nr > 0);
 902                         page->cp_owner->ci_owned_nr--;
 903                         page->cp_owner = NULL;
 904                         page->cp_task = NULL;
 905                 }
 906         }
 907         EXIT;
 908 }
 909
 910 static void cl_page_owner_set(struct cl_page *page)
 911 {
 912         ENTRY;
 913         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 914                 LASSERT(page->cp_owner != NULL);
 915                 page->cp_owner->ci_owned_nr++;
 916         }
 917         EXIT;
 918 }
 919
 920 void cl_page_disown0(const struct lu_env *env,
 921                      struct cl_io *io, struct cl_page *pg)
 922 {
 923         enum cl_page_state state;
 924
 925         ENTRY;
 926         state = pg->cp_state;
 927         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 928         PINVRNT(env, pg, cl_page_invariant(pg));
 929         cl_page_owner_clear(pg);
 930
 931         if (state == CPS_OWNED)
 932                 cl_page_state_set(env, pg, CPS_CACHED);
 933         /*
 934          * Completion call-backs are executed in the bottom-up order, so that
 935          * uppermost layer (llite), responsible for VFS/VM interaction runs
 936          * last and can release locks safely.
 937          */
 938         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 939                                (const struct lu_env *,
 940                                 const struct cl_page_slice *, struct cl_io *),
 941                                io);
 942         EXIT;
 943 }
 944
 945 /**
 946  * returns true, iff page is owned by the given io.
 947  */
 948 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 949 {
 950         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 951         ENTRY;
 952         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == io);
 953 }
 954 EXPORT_SYMBOL(cl_page_is_owned);
 955
 956 /**
 957  * Try to own a page by IO.
 958  *
 959  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 960  * into cl_page_state::CPS_OWNED state.
 961  *
 962  * \pre  !cl_page_is_owned(pg, io)
 963  * \post result == 0 iff cl_page_is_owned(pg, io)
 964  *
 965  * \retval 0   success
 966  *
 967  * \retval -ve failure, e.g., page was destroyed (and landed in
 968  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 969  *             or, page was owned by another thread, or in IO.
 970  *
 971  * \see cl_page_disown()
 972  * \see cl_page_operations::cpo_own()
 973  * \see cl_page_own_try()
 974  * \see cl_page_own
 975  */
 976 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 977                         struct cl_page *pg, int nonblock)
 978 {
 979         int result;
 980
 981         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 982
 983         ENTRY;
 984         pg = cl_page_top(pg);
 985         io = cl_io_top(io);
 986
 987         if (pg->cp_state == CPS_FREEING) {
 988                 result = -EAGAIN;
 989         } else {
 990                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 991                                         (const struct lu_env *,
 992                                          const struct cl_page_slice *,
 993                                          struct cl_io *, int),
 994                                         io, nonblock);
 995                 if (result == 0) {
 996                         PASSERT(env, pg, pg->cp_owner == NULL);
 997                         PASSERT(env, pg, pg->cp_req == NULL);
 998                         pg->cp_owner = io;
 999                         pg->cp_task  = current;
1000                         cl_page_owner_set(pg);
1001                         if (pg->cp_state != CPS_FREEING) {
1002                                 cl_page_state_set(env, pg, CPS_OWNED);
1003                         } else {
1004                                 cl_page_disown0(env, io, pg);
1005                                 result = -EAGAIN;
1006                         }
1007                 }
1008         }
1009         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
1010         RETURN(result);
1011 }
1012
1013 /**
1014  * Own a page, might be blocked.
1015  *
1016  * \see cl_page_own0()
1017  */
1018 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
1019 {
1020         return cl_page_own0(env, io, pg, 0);
1021 }
1022 EXPORT_SYMBOL(cl_page_own);
1023
1024 /**
1025  * Nonblock version of cl_page_own().
1026  *
1027  * \see cl_page_own0()
1028  */
1029 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
1030                     struct cl_page *pg)
1031 {
1032         return cl_page_own0(env, io, pg, 1);
1033 }
1034 EXPORT_SYMBOL(cl_page_own_try);
1035
1036
1037 /**
1038  * Assume page ownership.
1039  *
1040  * Called when page is already locked by the hosting VM.
1041  *
1042  * \pre !cl_page_is_owned(pg, io)
1043  * \post cl_page_is_owned(pg, io)
1044  *
1045  * \see cl_page_operations::cpo_assume()
1046  */
1047 void cl_page_assume(const struct lu_env *env,
1048                     struct cl_io *io, struct cl_page *pg)
1049 {
1050         PASSERT(env, pg, pg->cp_state < CPS_OWNED);
1051         PASSERT(env, pg, pg->cp_owner == NULL);
1052         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
1053         PINVRNT(env, pg, cl_page_invariant(pg));
1054
1055         ENTRY;
1056         pg = cl_page_top(pg);
1057         io = cl_io_top(io);
1058
1059         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
1060         pg->cp_owner = io;
1061         pg->cp_task = current;
1062         cl_page_owner_set(pg);
1063         cl_page_state_set(env, pg, CPS_OWNED);
1064         EXIT;
1065 }
1066 EXPORT_SYMBOL(cl_page_assume);
1067
1068 /**
1069  * Releases page ownership without unlocking the page.
1070  *
1071  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
1072  * underlying VM page (as VM is supposed to do this itself).
1073  *
1074  * \pre   cl_page_is_owned(pg, io)
1075  * \post !cl_page_is_owned(pg, io)
1076  *
1077  * \see cl_page_assume()
1078  */
1079 void cl_page_unassume(const struct lu_env *env,
1080                       struct cl_io *io, struct cl_page *pg)
1081 {
1082         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1083         PINVRNT(env, pg, cl_page_invariant(pg));
1084
1085         ENTRY;
1086         pg = cl_page_top(pg);
1087         io = cl_io_top(io);
1088         cl_page_owner_clear(pg);
1089         cl_page_state_set(env, pg, CPS_CACHED);
1090         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
1091                                (const struct lu_env *,
1092                                 const struct cl_page_slice *, struct cl_io *),
1093                                io);
1094         EXIT;
1095 }
1096 EXPORT_SYMBOL(cl_page_unassume);
1097
1098 /**
1099  * Releases page ownership.
1100  *
1101  * Moves page into cl_page_state::CPS_CACHED.
1102  *
1103  * \pre   cl_page_is_owned(pg, io)
1104  * \post !cl_page_is_owned(pg, io)
1105  *
1106  * \see cl_page_own()
1107  * \see cl_page_operations::cpo_disown()
1108  */
1109 void cl_page_disown(const struct lu_env *env,
1110                     struct cl_io *io, struct cl_page *pg)
1111 {
1112         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1113
1114         ENTRY;
1115         pg = cl_page_top(pg);
1116         io = cl_io_top(io);
1117         cl_page_disown0(env, io, pg);
1118         EXIT;
1119 }
1120 EXPORT_SYMBOL(cl_page_disown);
1121
1122 /**
1123  * Called when page is to be removed from the object, e.g., as a result of
1124  * truncate.
1125  *
1126  * Calls cl_page_operations::cpo_discard() top-to-bottom.
1127  *
1128  * \pre cl_page_is_owned(pg, io)
1129  *
1130  * \see cl_page_operations::cpo_discard()
1131  */
1132 void cl_page_discard(const struct lu_env *env,
1133                      struct cl_io *io, struct cl_page *pg)
1134 {
1135         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1136         PINVRNT(env, pg, cl_page_invariant(pg));
1137
1138         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
1139 }
1140 EXPORT_SYMBOL(cl_page_discard);
1141
1142 /**
1143  * Version of cl_page_delete() that can be called for not fully constructed
1144  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
1145  * path. Doesn't check page invariant.
1146  */
1147 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
1148                             int radix)
1149 {
1150         struct cl_page *tmp = pg;
1151         ENTRY;
1152
1153         PASSERT(env, pg, pg == cl_page_top(pg));
1154         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
1155
1156         /*
1157          * Severe all ways to obtain new pointers to @pg.
1158          */
1159         cl_page_owner_clear(pg);
1160
1161         /*
1162          * unexport the page firstly before freeing it so that
1163          * the page content is considered to be invalid.
1164          * We have to do this because a CPS_FREEING cl_page may
1165          * be NOT under the protection of a cl_lock.
1166          * Afterwards, if this page is found by other threads, then this
1167          * page will be forced to reread.
1168          */
1169         cl_page_export(env, pg, 0);
1170         cl_page_state_set0(env, pg, CPS_FREEING);
1171
1172         if (!radix)
1173                 /*
1174                  * !radix means that @pg is not yet in the radix tree, skip
1175                  * removing it.
1176                  */
1177                 tmp = pg->cp_child;
1178         for (; tmp != NULL; tmp = tmp->cp_child) {
1179                 void                    *value;
1180                 struct cl_object_header *hdr;
1181
1182                 hdr = cl_object_header(tmp->cp_obj);
1183                 cfs_spin_lock(&hdr->coh_page_guard);
1184                 value = radix_tree_delete(&hdr->coh_tree, tmp->cp_index);
1185                 PASSERT(env, tmp, value == tmp);
1186                 PASSERT(env, tmp, hdr->coh_pages > 0);
1187                 hdr->coh_pages--;
1188                 cfs_spin_unlock(&hdr->coh_page_guard);
1189         }
1190
1191         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete),
1192                        (const struct lu_env *, const struct cl_page_slice *));
1193         EXIT;
1194 }
1195
1196 /**
1197  * Called when a decision is made to throw page out of memory.
1198  *
1199  * Notifies all layers about page destruction by calling
1200  * cl_page_operations::cpo_delete() method top-to-bottom.
1201  *
1202  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
1203  * where transition to this state happens).
1204  *
1205  * Eliminates all venues through which new references to the page can be
1206  * obtained:
1207  *
1208  *     - removes page from the radix trees,
1209  *
1210  *     - breaks linkage from VM page to cl_page.
1211  *
1212  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
1213  * drain after some time, at which point page will be recycled.
1214  *
1215  * \pre  pg == cl_page_top(pg)
1216  * \pre  VM page is locked
1217  * \post pg->cp_state == CPS_FREEING
1218  *
1219  * \see cl_page_operations::cpo_delete()
1220  */
1221 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
1222 {
1223         PINVRNT(env, pg, cl_page_invariant(pg));
1224         ENTRY;
1225         cl_page_delete0(env, pg, 1);
1226         EXIT;
1227 }
1228 EXPORT_SYMBOL(cl_page_delete);
1229
1230 /**
1231  * Unmaps page from user virtual memory.
1232  *
1233  * Calls cl_page_operations::cpo_unmap() through all layers top-to-bottom. The
1234  * layer responsible for VM interaction has to unmap page from user space
1235  * virtual memory.
1236  *
1237  * \see cl_page_operations::cpo_unmap()
1238  */
1239 int cl_page_unmap(const struct lu_env *env,
1240                   struct cl_io *io, struct cl_page *pg)
1241 {
1242         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1243         PINVRNT(env, pg, cl_page_invariant(pg));
1244
1245         return cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_unmap));
1246 }
1247 EXPORT_SYMBOL(cl_page_unmap);
1248
1249 /**
1250  * Marks page up-to-date.
1251  *
1252  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
1253  * layer responsible for VM interaction has to mark/clear page as up-to-date
1254  * by the \a uptodate argument.
1255  *
1256  * \see cl_page_operations::cpo_export()
1257  */
1258 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
1259 {
1260         PINVRNT(env, pg, cl_page_invariant(pg));
1261         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
1262                        (const struct lu_env *,
1263                         const struct cl_page_slice *, int), uptodate);
1264 }
1265 EXPORT_SYMBOL(cl_page_export);
1266
1267 /**
1268  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
1269  * thread.
1270  */
1271 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
1272 {
1273         int result;
1274         const struct cl_page_slice *slice;
1275
1276         ENTRY;
1277         pg = cl_page_top_trusted((struct cl_page *)pg);
1278         slice = container_of(pg->cp_layers.next,
1279                              const struct cl_page_slice, cpl_linkage);
1280         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
1281         /*
1282          * Call ->cpo_is_vmlocked() directly instead of going through
1283          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
1284          * cl_page_invariant().
1285          */
1286         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
1287         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
1288         RETURN(result == -EBUSY);
1289 }
1290 EXPORT_SYMBOL(cl_page_is_vmlocked);
1291
1292 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
1293 {
1294         ENTRY;
1295         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
1296 }
1297
1298 static void cl_page_io_start(const struct lu_env *env,
1299                              struct cl_page *pg, enum cl_req_type crt)
1300 {
1301         /*
1302          * Page is queued for IO, change its state.
1303          */
1304         ENTRY;
1305         cl_page_owner_clear(pg);
1306         cl_page_state_set(env, pg, cl_req_type_state(crt));
1307         EXIT;
1308 }
1309
1310 /**
1311  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
1312  * called top-to-bottom. Every layer either agrees to submit this page (by
1313  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
1314  * handling interactions with the VM also has to inform VM that page is under
1315  * transfer now.
1316  */
1317 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
1318                  struct cl_page *pg, enum cl_req_type crt)
1319 {
1320         int result;
1321
1322         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1323         PINVRNT(env, pg, cl_page_invariant(pg));
1324         PINVRNT(env, pg, crt < CRT_NR);
1325
1326         /*
1327          * XXX this has to be called bottom-to-top, so that llite can set up
1328          * PG_writeback without risking other layers deciding to skip this
1329          * page.
1330          */
1331         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
1332         if (result == 0)
1333                 cl_page_io_start(env, pg, crt);
1334
1335         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
1336                       equi(result == 0,
1337                            PageWriteback(cl_page_vmpage(env, pg)))));
1338         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1339         return result;
1340 }
1341 EXPORT_SYMBOL(cl_page_prep);
1342
1343 /**
1344  * Notify layers about transfer completion.
1345  *
1346  * Invoked by transfer sub-system (which is a part of osc) to notify layers
1347  * that a transfer, of which this page is a part of has completed.
1348  *
1349  * Completion call-backs are executed in the bottom-up order, so that
1350  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1351  * and can release locks safely.
1352  *
1353  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1354  * \post pg->cp_state == CPS_CACHED
1355  *
1356  * \see cl_page_operations::cpo_completion()
1357  */
1358 void cl_page_completion(const struct lu_env *env,
1359                         struct cl_page *pg, enum cl_req_type crt, int ioret)
1360 {
1361         struct cl_sync_io *anchor = pg->cp_sync_io;
1362
1363         PASSERT(env, pg, crt < CRT_NR);
1364         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
1365         PASSERT(env, pg, pg->cp_req == NULL);
1366         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
1367         PINVRNT(env, pg, cl_page_invariant(pg));
1368
1369         ENTRY;
1370         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
1371         if (crt == CRT_READ && ioret == 0) {
1372                 PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED));
1373                 pg->cp_flags |= CPF_READ_COMPLETED;
1374         }
1375
1376         cl_page_state_set(env, pg, CPS_CACHED);
1377         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
1378                                (const struct lu_env *,
1379                                 const struct cl_page_slice *, int), ioret);
1380         if (anchor) {
1381                 LASSERT(pg->cp_sync_io == anchor);
1382                 pg->cp_sync_io = NULL;
1383                 cl_sync_io_note(anchor, ioret);
1384         }
1385
1386         /* Don't assert the page writeback bit here because the lustre file
1387          * may be as a backend of swap space. in this case, the page writeback
1388          * is set by VM, and obvious we shouldn't clear it at all. Fortunately
1389          * this type of pages are all TRANSIENT pages. */
1390         KLASSERT(ergo(pg->cp_type == CPT_CACHEABLE,
1391                       !PageWriteback(cl_page_vmpage(env, pg))));
1392         EXIT;
1393 }
1394 EXPORT_SYMBOL(cl_page_completion);
1395
1396 /**
1397  * Notify layers that transfer formation engine decided to yank this page from
1398  * the cache and to make it a part of a transfer.
1399  *
1400  * \pre  pg->cp_state == CPS_CACHED
1401  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1402  *
1403  * \see cl_page_operations::cpo_make_ready()
1404  */
1405 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
1406                        enum cl_req_type crt)
1407 {
1408         int result;
1409
1410         PINVRNT(env, pg, crt < CRT_NR);
1411
1412         ENTRY;
1413         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
1414                                 (const struct lu_env *,
1415                                  const struct cl_page_slice *));
1416         if (result == 0) {
1417                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
1418                 cl_page_io_start(env, pg, crt);
1419         }
1420         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1421         RETURN(result);
1422 }
1423 EXPORT_SYMBOL(cl_page_make_ready);
1424
1425 /**
1426  * Notify layers that high level io decided to place this page into a cache
1427  * for future transfer.
1428  *
1429  * The layer implementing transfer engine (osc) has to register this page in
1430  * its queues.
1431  *
1432  * \pre  cl_page_is_owned(pg, io)
1433  * \post ergo(result == 0,
1434  *            pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT)
1435  *
1436  * \see cl_page_operations::cpo_cache_add()
1437  */
1438 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
1439                       struct cl_page *pg, enum cl_req_type crt)
1440 {
1441         int result;
1442
1443         PINVRNT(env, pg, crt < CRT_NR);
1444         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1445         PINVRNT(env, pg, cl_page_invariant(pg));
1446
1447         ENTRY;
1448         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_cache_add));
1449         if (result == 0) {
1450                 cl_page_owner_clear(pg);
1451                 cl_page_state_set(env, pg, CPS_CACHED);
1452         }
1453         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1454         RETURN(result);
1455 }
1456 EXPORT_SYMBOL(cl_page_cache_add);
1457
1458 /**
1459  * Checks whether page is protected by any extent lock is at least required
1460  * mode.
1461  *
1462  * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1463  * \see cl_page_operations::cpo_is_under_lock()
1464  */
1465 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1466                           struct cl_page *page)
1467 {
1468         int rc;
1469
1470         PINVRNT(env, page, cl_page_invariant(page));
1471
1472         ENTRY;
1473         rc = CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1474                             (const struct lu_env *,
1475                              const struct cl_page_slice *, struct cl_io *),
1476                             io);
1477         PASSERT(env, page, rc != 0);
1478         RETURN(rc);
1479 }
1480 EXPORT_SYMBOL(cl_page_is_under_lock);
1481
1482 /**
1483  * Purges all cached pages belonging to the object \a obj.
1484  */
1485 int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj)
1486 {
1487         struct cl_thread_info   *info;
1488         struct cl_object        *obj = cl_object_top(clobj);
1489         struct cl_io            *io;
1490         struct cl_page_list     *plist;
1491         int                      resched;
1492         int                      result;
1493
1494         ENTRY;
1495         info  = cl_env_info(env);
1496         plist = &info->clt_list;
1497         io    = &info->clt_io;
1498
1499         /*
1500          * initialize the io. This is ugly since we never do IO in this
1501          * function, we just make cl_page_list functions happy. -jay
1502          */
1503         io->ci_obj = obj;
1504         result = cl_io_init(env, io, CIT_MISC, obj);
1505         if (result != 0) {
1506                 cl_io_fini(env, io);
1507                 RETURN(io->ci_result);
1508         }
1509
1510         do {
1511                 cl_page_list_init(plist);
1512                 cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF, plist, 0,
1513                                     &resched);
1514                 /*
1515                  * Since we're purging the pages of an object, we don't care
1516                  * the possible outcomes of the following functions.
1517                  */
1518                 cl_page_list_unmap(env, io, plist);
1519                 cl_page_list_discard(env, io, plist);
1520                 cl_page_list_disown(env, io, plist);
1521                 cl_page_list_fini(env, plist);
1522
1523                 if (resched)
1524                         cfs_cond_resched();
1525         } while (resched);
1526
1527         cl_io_fini(env, io);
1528         RETURN(result);
1529 }
1530 EXPORT_SYMBOL(cl_pages_prune);
1531
1532 /**
1533  * Tells transfer engine that only part of a page is to be transmitted.
1534  *
1535  * \see cl_page_operations::cpo_clip()
1536  */
1537 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1538                   int from, int to)
1539 {
1540         PINVRNT(env, pg, cl_page_invariant(pg));
1541
1542         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1543         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1544                        (const struct lu_env *,
1545                         const struct cl_page_slice *,int, int),
1546                        from, to);
1547 }
1548 EXPORT_SYMBOL(cl_page_clip);
1549
1550 /**
1551  * Prints human readable representation of \a pg to the \a f.
1552  */
1553 void cl_page_header_print(const struct lu_env *env, void *cookie,
1554                           lu_printer_t printer, const struct cl_page *pg)
1555 {
1556         (*printer)(env, cookie,
1557                    "page@%p[%d %p:%lu ^%p_%p %d %d %d %p %p %#x]\n",
1558                    pg, cfs_atomic_read(&pg->cp_ref), pg->cp_obj,
1559                    pg->cp_index, pg->cp_parent, pg->cp_child,
1560                    pg->cp_state, pg->cp_error, pg->cp_type,
1561                    pg->cp_owner, pg->cp_req, pg->cp_flags);
1562 }
1563 EXPORT_SYMBOL(cl_page_header_print);
1564
1565 /**
1566  * Prints human readable representation of \a pg to the \a f.
1567  */
1568 void cl_page_print(const struct lu_env *env, void *cookie,
1569                    lu_printer_t printer, const struct cl_page *pg)
1570 {
1571         struct cl_page *scan;
1572
1573         for (scan = cl_page_top((struct cl_page *)pg);
1574              scan != NULL; scan = scan->cp_child)
1575                 cl_page_header_print(env, cookie, printer, scan);
1576         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1577                        (const struct lu_env *env,
1578                         const struct cl_page_slice *slice,
1579                         void *cookie, lu_printer_t p), cookie, printer);
1580         (*printer)(env, cookie, "end page@%p\n", pg);
1581 }
1582 EXPORT_SYMBOL(cl_page_print);
1583
1584 /**
1585  * Cancel a page which is still in a transfer.
1586  */
1587 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1588 {
1589         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1590                               (const struct lu_env *,
1591                                const struct cl_page_slice *));
1592 }
1593 EXPORT_SYMBOL(cl_page_cancel);
1594
1595 /**
1596  * Converts a byte offset within object \a obj into a page index.
1597  */
1598 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1599 {
1600         /*
1601          * XXX for now.
1602          */
1603         return (loff_t)idx << CFS_PAGE_SHIFT;
1604 }
1605 EXPORT_SYMBOL(cl_offset);
1606
1607 /**
1608  * Converts a page index into a byte offset within object \a obj.
1609  */
1610 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1611 {
1612         /*
1613          * XXX for now.
1614          */
1615         return offset >> CFS_PAGE_SHIFT;
1616 }
1617 EXPORT_SYMBOL(cl_index);
1618
1619 int cl_page_size(const struct cl_object *obj)
1620 {
1621         return 1 << CFS_PAGE_SHIFT;
1622 }
1623 EXPORT_SYMBOL(cl_page_size);
1624
1625 /**
1626  * Adds page slice to the compound page.
1627  *
1628  * This is called by cl_object_operations::coo_page_init() methods to add a
1629  * per-layer state to the page. New state is added at the end of
1630  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1631  *
1632  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1633  */
1634 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1635                        struct cl_object *obj,
1636                        const struct cl_page_operations *ops)
1637 {
1638         ENTRY;
1639         cfs_list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1640         slice->cpl_obj  = obj;
1641         slice->cpl_ops  = ops;
1642         slice->cpl_page = page;
1643         EXIT;
1644 }
1645 EXPORT_SYMBOL(cl_page_slice_add);
1646
1647 int  cl_page_init(void)
1648 {
1649         return lu_kmem_init(cl_page_caches);
1650 }
1651
1652 void cl_page_fini(void)
1653 {
1654         lu_kmem_fini(cl_page_caches);
1655 }