lustre/obdclass/cl_page.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011 Whamcloud, Inc.
  33  *
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * Client Lustre Page.
  40  *
  41  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_CLASS
  45 #ifndef EXPORT_SYMTAB
  46 # define EXPORT_SYMTAB
  47 #endif
  48
  49 #include <libcfs/libcfs.h>
  50 #include <obd_class.h>
  51 #include <obd_support.h>
  52 #include <libcfs/list.h>
  53
  54 #include <cl_object.h>
  55 #include "cl_internal.h"
  56
  57 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
  58                             int radix);
  59
  60 static cfs_mem_cache_t      *cl_page_kmem = NULL;
  61
  62 static struct lu_kmem_descr cl_page_caches[] = {
  63         {
  64                 .ckd_cache = &cl_page_kmem,
  65                 .ckd_name  = "cl_page_kmem",
  66                 .ckd_size  = sizeof (struct cl_page)
  67         },
  68         {
  69                 .ckd_cache = NULL
  70         }
  71 };
  72
  73 #ifdef LIBCFS_DEBUG
  74 # define PASSERT(env, page, expr)                                       \
  75   do {                                                                    \
  76           if (unlikely(!(expr))) {                                      \
  77                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  78                   LASSERT(0);                                           \
  79           }                                                             \
  80   } while (0)
  81 #else /* !LIBCFS_DEBUG */
  82 # define PASSERT(env, page, exp) \
  83         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  84 #endif /* !LIBCFS_DEBUG */
  85
  86 #ifdef INVARIANT_CHECK
  87 # define PINVRNT(env, page, expr)                                       \
  88   do {                                                                    \
  89           if (unlikely(!(expr))) {                                      \
  90                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  91                   LINVRNT(0);                                           \
  92           }                                                             \
  93   } while (0)
  94 #else /* !INVARIANT_CHECK */
  95 # define PINVRNT(env, page, exp) \
  96         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  97 #endif /* !INVARIANT_CHECK */
  98
  99 /**
 100  * Internal version of cl_page_top, it should be called with page referenced,
 101  * or coh_page_guard held.
 102  */
 103 static struct cl_page *cl_page_top_trusted(struct cl_page *page)
 104 {
 105         while (page->cp_parent != NULL)
 106                 page = page->cp_parent;
 107         return page;
 108 }
 109
 110 /**
 111  * Internal version of cl_page_get().
 112  *
 113  * This function can be used to obtain initial reference to previously
 114  * unreferenced cached object. It can be called only if concurrent page
 115  * reclamation is somehow prevented, e.g., by locking page radix-tree
 116  * (cl_object_header::hdr->coh_page_guard), or by keeping a lock on a VM page,
 117  * associated with \a page.
 118  *
 119  * Use with care! Not exported.
 120  */
 121 static void cl_page_get_trust(struct cl_page *page)
 122 {
 123         /*
 124          * Checkless version for trusted users.
 125          */
 126         if (cfs_atomic_inc_return(&page->cp_ref) == 1)
 127                 cfs_atomic_inc(&cl_object_site(page->cp_obj)->cs_pages.cs_busy);
 128 }
 129
 130 /**
 131  * Returns a slice within a page, corresponding to the given layer in the
 132  * device stack.
 133  *
 134  * \see cl_lock_at()
 135  */
 136 static const struct cl_page_slice *
 137 cl_page_at_trusted(const struct cl_page *page,
 138                    const struct lu_device_type *dtype)
 139 {
 140         const struct cl_page_slice *slice;
 141
 142 #ifdef INVARIANT_CHECK
 143         struct cl_object_header *ch = cl_object_header(page->cp_obj);
 144
 145         if (!cfs_atomic_read(&page->cp_ref))
 146                 LASSERT_SPIN_LOCKED(&ch->coh_page_guard);
 147 #endif
 148         ENTRY;
 149
 150         page = cl_page_top_trusted((struct cl_page *)page);
 151         do {
 152                 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 153                         if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
 154                                 RETURN(slice);
 155                 }
 156                 page = page->cp_child;
 157         } while (page != NULL);
 158         RETURN(NULL);
 159 }
 160
 161 /**
 162  * Returns a page with given index in the given object, or NULL if no page is
 163  * found. Acquires a reference on \a page.
 164  *
 165  * Locking: called under cl_object_header::coh_page_guard spin-lock.
 166  */
 167 struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index)
 168 {
 169         struct cl_page *page;
 170
 171         LASSERT_SPIN_LOCKED(&hdr->coh_page_guard);
 172
 173         page = radix_tree_lookup(&hdr->coh_tree, index);
 174         if (page != NULL) {
 175                 cl_page_get_trust(page);
 176         }
 177         return page;
 178 }
 179 EXPORT_SYMBOL(cl_page_lookup);
 180
 181 /**
 182  * Returns a list of pages by a given [start, end] of \a obj.
 183  *
 184  * \param resched If not NULL, then we give up before hogging CPU for too
 185  * long and set *resched = 1, in that case caller should implement a retry
 186  * logic.
 187  *
 188  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
 189  * crucial in the face of [offset, EOF] locks.
 190  *
 191  * Return at least one page in @queue unless there is no covered page.
 192  */
 193 int cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj,
 194                         struct cl_io *io, pgoff_t start, pgoff_t end,
 195                         cl_page_gang_cb_t cb, void *cbdata)
 196 {
 197         struct cl_object_header *hdr;
 198         struct cl_page          *page;
 199         struct cl_page         **pvec;
 200         const struct cl_page_slice  *slice;
 201         const struct lu_device_type *dtype;
 202         pgoff_t                  idx;
 203         unsigned int             nr;
 204         unsigned int             i;
 205         unsigned int             j;
 206         int                      res = CLP_GANG_OKAY;
 207         int                      tree_lock = 1;
 208         ENTRY;
 209
 210         idx = start;
 211         hdr = cl_object_header(obj);
 212         pvec = cl_env_info(env)->clt_pvec;
 213         dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type;
 214         cfs_spin_lock(&hdr->coh_page_guard);
 215         while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec,
 216                                             idx, CLT_PVEC_SIZE)) > 0) {
 217                 int end_of_region = 0;
 218                 idx = pvec[nr - 1]->cp_index + 1;
 219                 for (i = 0, j = 0; i < nr; ++i) {
 220                         page = pvec[i];
 221                         pvec[i] = NULL;
 222
 223                         LASSERT(page->cp_type == CPT_CACHEABLE);
 224                         if (page->cp_index > end) {
 225                                 end_of_region = 1;
 226                                 break;
 227                         }
 228                         if (page->cp_state == CPS_FREEING)
 229                                 continue;
 230
 231                         slice = cl_page_at_trusted(page, dtype);
 232                         /*
 233                          * Pages for lsm-less file has no underneath sub-page
 234                          * for osc, in case of ...
 235                          */
 236                         PASSERT(env, page, slice != NULL);
 237
 238                         page = slice->cpl_page;
 239                         /*
 240                          * Can safely call cl_page_get_trust() under
 241                          * radix-tree spin-lock.
 242                          *
 243                          * XXX not true, because @page is from object another
 244                          * than @hdr and protected by different tree lock.
 245                          */
 246                         cl_page_get_trust(page);
 247                         lu_ref_add_atomic(&page->cp_reference,
 248                                           "gang_lookup", cfs_current());
 249                         pvec[j++] = page;
 250                 }
 251
 252                 /*
 253                  * Here a delicate locking dance is performed. Current thread
 254                  * holds a reference to a page, but has to own it before it
 255                  * can be placed into queue. Owning implies waiting, so
 256                  * radix-tree lock is to be released. After a wait one has to
 257                  * check that pages weren't truncated (cl_page_own() returns
 258                  * error in the latter case).
 259                  */
 260                 cfs_spin_unlock(&hdr->coh_page_guard);
 261                 tree_lock = 0;
 262
 263                 for (i = 0; i < j; ++i) {
 264                         page = pvec[i];
 265                         if (res == CLP_GANG_OKAY)
 266                                 res = (*cb)(env, io, page, cbdata);
 267                         lu_ref_del(&page->cp_reference,
 268                                    "gang_lookup", cfs_current());
 269                         cl_page_put(env, page);
 270                 }
 271                 if (nr < CLT_PVEC_SIZE || end_of_region)
 272                         break;
 273
 274                 if (res == CLP_GANG_OKAY && cfs_need_resched())
 275                         res = CLP_GANG_RESCHED;
 276                 if (res != CLP_GANG_OKAY)
 277                         break;
 278
 279                 cfs_spin_lock(&hdr->coh_page_guard);
 280                 tree_lock = 1;
 281         }
 282         if (tree_lock)
 283                 cfs_spin_unlock(&hdr->coh_page_guard);
 284         RETURN(res);
 285 }
 286 EXPORT_SYMBOL(cl_page_gang_lookup);
 287
 288 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
 289 {
 290         struct cl_object *obj  = page->cp_obj;
 291         struct cl_site   *site = cl_object_site(obj);
 292
 293         PASSERT(env, page, cfs_list_empty(&page->cp_batch));
 294         PASSERT(env, page, page->cp_owner == NULL);
 295         PASSERT(env, page, page->cp_req == NULL);
 296         PASSERT(env, page, page->cp_parent == NULL);
 297         PASSERT(env, page, page->cp_state == CPS_FREEING);
 298
 299         ENTRY;
 300         cfs_might_sleep();
 301         while (!cfs_list_empty(&page->cp_layers)) {
 302                 struct cl_page_slice *slice;
 303
 304                 slice = cfs_list_entry(page->cp_layers.next,
 305                                        struct cl_page_slice, cpl_linkage);
 306                 cfs_list_del_init(page->cp_layers.next);
 307                 slice->cpl_ops->cpo_fini(env, slice);
 308         }
 309         cfs_atomic_dec(&site->cs_pages.cs_total);
 310
 311 #ifdef LUSTRE_PAGESTATE_TRACKING
 312         cfs_atomic_dec(&site->cs_pages_state[page->cp_state]);
 313 #endif
 314         lu_object_ref_del_at(&obj->co_lu, page->cp_obj_ref, "cl_page", page);
 315         cl_object_put(env, obj);
 316         lu_ref_fini(&page->cp_reference);
 317         OBD_SLAB_FREE_PTR(page, cl_page_kmem);
 318         EXIT;
 319 }
 320
 321 /**
 322  * Helper function updating page state. This is the only place in the code
 323  * where cl_page::cp_state field is mutated.
 324  */
 325 static inline void cl_page_state_set_trust(struct cl_page *page,
 326                                            enum cl_page_state state)
 327 {
 328         /* bypass const. */
 329         *(enum cl_page_state *)&page->cp_state = state;
 330 }
 331
 332 static int cl_page_alloc(const struct lu_env *env, struct cl_object *o,
 333                          pgoff_t ind, struct page *vmpage,
 334                          enum cl_page_type type, struct cl_page **out)
 335 {
 336         struct cl_page          *page;
 337         struct cl_page          *err  = NULL;
 338         struct lu_object_header *head;
 339         struct cl_site          *site = cl_object_site(o);
 340         int                      result;
 341
 342         ENTRY;
 343         result = +1;
 344         OBD_SLAB_ALLOC_PTR_GFP(page, cl_page_kmem, CFS_ALLOC_IO);
 345         if (page != NULL) {
 346                 cfs_atomic_set(&page->cp_ref, 1);
 347                 page->cp_obj = o;
 348                 cl_object_get(o);
 349                 page->cp_obj_ref = lu_object_ref_add(&o->co_lu,
 350                                                      "cl_page", page);
 351                 page->cp_index = ind;
 352                 cl_page_state_set_trust(page, CPS_CACHED);
 353                 page->cp_type = type;
 354                 CFS_INIT_LIST_HEAD(&page->cp_layers);
 355                 CFS_INIT_LIST_HEAD(&page->cp_batch);
 356                 CFS_INIT_LIST_HEAD(&page->cp_flight);
 357                 cfs_mutex_init(&page->cp_mutex);
 358                 lu_ref_init(&page->cp_reference);
 359                 head = o->co_lu.lo_header;
 360                 cfs_list_for_each_entry(o, &head->loh_layers,
 361                                         co_lu.lo_linkage) {
 362                         if (o->co_ops->coo_page_init != NULL) {
 363                                 err = o->co_ops->coo_page_init(env, o,
 364                                                                page, vmpage);
 365                                 if (err != NULL) {
 366                                         cl_page_delete0(env, page, 0);
 367                                         cl_page_free(env, page);
 368                                         page = err;
 369                                         break;
 370                                 }
 371                         }
 372                 }
 373                 if (err == NULL) {
 374                         cfs_atomic_inc(&site->cs_pages.cs_busy);
 375                         cfs_atomic_inc(&site->cs_pages.cs_total);
 376
 377 #ifdef LUSTRE_PAGESTATE_TRACKING
 378                         cfs_atomic_inc(&site->cs_pages_state[CPS_CACHED]);
 379 #endif
 380                         cfs_atomic_inc(&site->cs_pages.cs_created);
 381                         result = 0;
 382                 }
 383         } else
 384                 page = ERR_PTR(-ENOMEM);
 385         *out = page;
 386         RETURN(result);
 387 }
 388
 389 /**
 390  * Returns a cl_page with index \a idx at the object \a o, and associated with
 391  * the VM page \a vmpage.
 392  *
 393  * This is the main entry point into the cl_page caching interface. First, a
 394  * cache (implemented as a per-object radix tree) is consulted. If page is
 395  * found there, it is returned immediately. Otherwise new page is allocated
 396  * and returned. In any case, additional reference to page is acquired.
 397  *
 398  * \see cl_object_find(), cl_lock_find()
 399  */
 400 static struct cl_page *cl_page_find0(const struct lu_env *env,
 401                                      struct cl_object *o,
 402                                      pgoff_t idx, struct page *vmpage,
 403                                      enum cl_page_type type,
 404                                      struct cl_page *parent)
 405 {
 406         struct cl_page          *page = NULL;
 407         struct cl_page          *ghost = NULL;
 408         struct cl_object_header *hdr;
 409         struct cl_site          *site = cl_object_site(o);
 410         int err;
 411
 412         LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 413         cfs_might_sleep();
 414
 415         ENTRY;
 416
 417         hdr = cl_object_header(o);
 418         cfs_atomic_inc(&site->cs_pages.cs_lookup);
 419
 420         CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
 421                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 422         /* fast path. */
 423         if (type == CPT_CACHEABLE) {
 424                 /*
 425                  * cl_vmpage_page() can be called here without any locks as
 426                  *
 427                  *     - "vmpage" is locked (which prevents ->private from
 428                  *       concurrent updates), and
 429                  *
 430                  *     - "o" cannot be destroyed while current thread holds a
 431                  *       reference on it.
 432                  */
 433                 page = cl_vmpage_page(vmpage, o);
 434                 PINVRNT(env, page,
 435                         ergo(page != NULL,
 436                              cl_page_vmpage(env, page) == vmpage &&
 437                              (void *)radix_tree_lookup(&hdr->coh_tree,
 438                                                        idx) == page));
 439         }
 440
 441         if (page != NULL) {
 442                 cfs_atomic_inc(&site->cs_pages.cs_hit);
 443                 RETURN(page);
 444         }
 445
 446         /* allocate and initialize cl_page */
 447         err = cl_page_alloc(env, o, idx, vmpage, type, &page);
 448         if (err != 0)
 449                 RETURN(page);
 450
 451         if (type == CPT_TRANSIENT) {
 452                 if (parent) {
 453                         LASSERT(page->cp_parent == NULL);
 454                         page->cp_parent = parent;
 455                         parent->cp_child = page;
 456                 }
 457                 RETURN(page);
 458         }
 459
 460         /*
 461          * XXX optimization: use radix_tree_preload() here, and change tree
 462          * gfp mask to GFP_KERNEL in cl_object_header_init().
 463          */
 464         cfs_spin_lock(&hdr->coh_page_guard);
 465         err = radix_tree_insert(&hdr->coh_tree, idx, page);
 466         if (err != 0) {
 467                 ghost = page;
 468                 /*
 469                  * Noted by Jay: a lock on \a vmpage protects cl_page_find()
 470                  * from this race, but
 471                  *
 472                  *     0. it's better to have cl_page interface "locally
 473                  *     consistent" so that its correctness can be reasoned
 474                  *     about without appealing to the (obscure world of) VM
 475                  *     locking.
 476                  *
 477                  *     1. handling this race allows ->coh_tree to remain
 478                  *     consistent even when VM locking is somehow busted,
 479                  *     which is very useful during diagnosing and debugging.
 480                  */
 481                 page = ERR_PTR(err);
 482                 CL_PAGE_DEBUG(D_ERROR, env, ghost,
 483                               "fail to insert into radix tree: %d\n", err);
 484         } else {
 485                 if (parent) {
 486                         LASSERT(page->cp_parent == NULL);
 487                         page->cp_parent = parent;
 488                         parent->cp_child = page;
 489                 }
 490                 hdr->coh_pages++;
 491         }
 492         cfs_spin_unlock(&hdr->coh_page_guard);
 493
 494         if (unlikely(ghost != NULL)) {
 495                 cfs_atomic_dec(&site->cs_pages.cs_busy);
 496                 cl_page_delete0(env, ghost, 0);
 497                 cl_page_free(env, ghost);
 498         }
 499         RETURN(page);
 500 }
 501
 502 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o,
 503                              pgoff_t idx, struct page *vmpage,
 504                              enum cl_page_type type)
 505 {
 506         return cl_page_find0(env, o, idx, vmpage, type, NULL);
 507 }
 508 EXPORT_SYMBOL(cl_page_find);
 509
 510
 511 struct cl_page *cl_page_find_sub(const struct lu_env *env, struct cl_object *o,
 512                                  pgoff_t idx, struct page *vmpage,
 513                                  struct cl_page *parent)
 514 {
 515         return cl_page_find0(env, o, idx, vmpage, parent->cp_type, parent);
 516 }
 517 EXPORT_SYMBOL(cl_page_find_sub);
 518
 519 static inline int cl_page_invariant(const struct cl_page *pg)
 520 {
 521         struct cl_object_header *header;
 522         struct cl_page          *parent;
 523         struct cl_page          *child;
 524         struct cl_io            *owner;
 525
 526         /*
 527          * Page invariant is protected by a VM lock.
 528          */
 529         LINVRNT(cl_page_is_vmlocked(NULL, pg));
 530
 531         header = cl_object_header(pg->cp_obj);
 532         parent = pg->cp_parent;
 533         child  = pg->cp_child;
 534         owner  = pg->cp_owner;
 535
 536         return cfs_atomic_read(&pg->cp_ref) > 0 &&
 537                 ergo(parent != NULL, parent->cp_child == pg) &&
 538                 ergo(child != NULL, child->cp_parent == pg) &&
 539                 ergo(child != NULL, pg->cp_obj != child->cp_obj) &&
 540                 ergo(parent != NULL, pg->cp_obj != parent->cp_obj) &&
 541                 ergo(owner != NULL && parent != NULL,
 542                      parent->cp_owner == pg->cp_owner->ci_parent) &&
 543                 ergo(owner != NULL && child != NULL,
 544                      child->cp_owner->ci_parent == owner) &&
 545                 /*
 546                  * Either page is early in initialization (has neither child
 547                  * nor parent yet), or it is in the object radix tree.
 548                  */
 549                 ergo(pg->cp_state < CPS_FREEING && pg->cp_type == CPT_CACHEABLE,
 550                      (void *)radix_tree_lookup(&header->coh_tree,
 551                                                pg->cp_index) == pg ||
 552                      (child == NULL && parent == NULL));
 553 }
 554
 555 static void cl_page_state_set0(const struct lu_env *env,
 556                                struct cl_page *page, enum cl_page_state state)
 557 {
 558         enum cl_page_state old;
 559 #ifdef LUSTRE_PAGESTATE_TRACKING
 560         struct cl_site *site = cl_object_site(page->cp_obj);
 561 #endif
 562
 563         /*
 564          * Matrix of allowed state transitions [old][new], for sanity
 565          * checking.
 566          */
 567         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 568                 [CPS_CACHED] = {
 569                         [CPS_CACHED]  = 0,
 570                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 571                         [CPS_PAGEIN]  = 0,
 572                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 573                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 574                 },
 575                 [CPS_OWNED] = {
 576                         [CPS_CACHED]  = 1, /* release to the cache */
 577                         [CPS_OWNED]   = 0,
 578                         [CPS_PAGEIN]  = 1, /* start read immediately */
 579                         [CPS_PAGEOUT] = 1, /* start write immediately */
 580                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 581                 },
 582                 [CPS_PAGEIN] = {
 583                         [CPS_CACHED]  = 1, /* io completion */
 584                         [CPS_OWNED]   = 0,
 585                         [CPS_PAGEIN]  = 0,
 586                         [CPS_PAGEOUT] = 0,
 587                         [CPS_FREEING] = 0,
 588                 },
 589                 [CPS_PAGEOUT] = {
 590                         [CPS_CACHED]  = 1, /* io completion */
 591                         [CPS_OWNED]   = 0,
 592                         [CPS_PAGEIN]  = 0,
 593                         [CPS_PAGEOUT] = 0,
 594                         [CPS_FREEING] = 0,
 595                 },
 596                 [CPS_FREEING] = {
 597                         [CPS_CACHED]  = 0,
 598                         [CPS_OWNED]   = 0,
 599                         [CPS_PAGEIN]  = 0,
 600                         [CPS_PAGEOUT] = 0,
 601                         [CPS_FREEING] = 0,
 602                 }
 603         };
 604
 605         ENTRY;
 606         old = page->cp_state;
 607         PASSERT(env, page, allowed_transitions[old][state]);
 608         CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
 609         for (; page != NULL; page = page->cp_child) {
 610                 PASSERT(env, page, page->cp_state == old);
 611                 PASSERT(env, page,
 612                         equi(state == CPS_OWNED, page->cp_owner != NULL));
 613
 614 #ifdef LUSTRE_PAGESTATE_TRACKING
 615                 cfs_atomic_dec(&site->cs_pages_state[page->cp_state]);
 616                 cfs_atomic_inc(&site->cs_pages_state[state]);
 617 #endif
 618                 cl_page_state_set_trust(page, state);
 619         }
 620         EXIT;
 621 }
 622
 623 static void cl_page_state_set(const struct lu_env *env,
 624                               struct cl_page *page, enum cl_page_state state)
 625 {
 626         cl_page_state_set0(env, page, state);
 627 }
 628
 629 /**
 630  * Acquires an additional reference to a page.
 631  *
 632  * This can be called only by caller already possessing a reference to \a
 633  * page.
 634  *
 635  * \see cl_object_get(), cl_lock_get().
 636  */
 637 void cl_page_get(struct cl_page *page)
 638 {
 639         ENTRY;
 640         LASSERT(page->cp_state != CPS_FREEING);
 641         cl_page_get_trust(page);
 642         EXIT;
 643 }
 644 EXPORT_SYMBOL(cl_page_get);
 645
 646 /**
 647  * Releases a reference to a page.
 648  *
 649  * When last reference is released, page is returned to the cache, unless it
 650  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 651  * destroyed.
 652  *
 653  * \see cl_object_put(), cl_lock_put().
 654  */
 655 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 656 {
 657         struct cl_object_header *hdr;
 658         struct cl_site *site = cl_object_site(page->cp_obj);
 659
 660         PASSERT(env, page, cfs_atomic_read(&page->cp_ref) > !!page->cp_parent);
 661
 662         ENTRY;
 663         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 664                        cfs_atomic_read(&page->cp_ref));
 665
 666         hdr = cl_object_header(cl_object_top(page->cp_obj));
 667         if (cfs_atomic_dec_and_lock(&page->cp_ref, &hdr->coh_page_guard)) {
 668                 cfs_atomic_dec(&site->cs_pages.cs_busy);
 669                 /* We're going to access the page w/o a reference, but it's
 670                  * ok because we have grabbed the lock coh_page_guard, which
 671                  * means nobody is able to free this page behind us.
 672                  */
 673                 if (page->cp_state == CPS_FREEING) {
 674                         /* We drop the page reference and check the page state
 675                          * inside the coh_page_guard. So that if it gets here,
 676                          * it is the REALLY last reference to this page.
 677                          */
 678                         cfs_spin_unlock(&hdr->coh_page_guard);
 679
 680                         LASSERT(cfs_atomic_read(&page->cp_ref) == 0);
 681                         PASSERT(env, page, page->cp_owner == NULL);
 682                         PASSERT(env, page, cfs_list_empty(&page->cp_batch));
 683                         /*
 684                          * Page is no longer reachable by other threads. Tear
 685                          * it down.
 686                          */
 687                         cl_page_free(env, page);
 688
 689                         EXIT;
 690                         return;
 691                 }
 692                 cfs_spin_unlock(&hdr->coh_page_guard);
 693         }
 694
 695         EXIT;
 696 }
 697 EXPORT_SYMBOL(cl_page_put);
 698
 699 /**
 700  * Returns a VM page associated with a given cl_page.
 701  */
 702 cfs_page_t *cl_page_vmpage(const struct lu_env *env, struct cl_page *page)
 703 {
 704         const struct cl_page_slice *slice;
 705
 706         /*
 707          * Find uppermost layer with ->cpo_vmpage() method, and return its
 708          * result.
 709          */
 710         page = cl_page_top(page);
 711         do {
 712                 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 713                         if (slice->cpl_ops->cpo_vmpage != NULL)
 714                                 RETURN(slice->cpl_ops->cpo_vmpage(env, slice));
 715                 }
 716                 page = page->cp_child;
 717         } while (page != NULL);
 718         LBUG(); /* ->cpo_vmpage() has to be defined somewhere in the stack */
 719 }
 720 EXPORT_SYMBOL(cl_page_vmpage);
 721
 722 /**
 723  * Returns a cl_page associated with a VM page, and given cl_object.
 724  */
 725 struct cl_page *cl_vmpage_page(cfs_page_t *vmpage, struct cl_object *obj)
 726 {
 727         struct cl_page *page;
 728         struct cl_object_header *hdr;
 729
 730         ENTRY;
 731         KLASSERT(PageLocked(vmpage));
 732
 733         /*
 734          * NOTE: absence of races and liveness of data are guaranteed by page
 735          *       lock on a "vmpage". That works because object destruction has
 736          *       bottom-to-top pass.
 737          */
 738
 739         /*
 740          * This loop assumes that ->private points to the top-most page. This
 741          * can be rectified easily.
 742          */
 743         hdr = cl_object_header(cl_object_top(obj));
 744         cfs_spin_lock(&hdr->coh_page_guard);
 745         for (page = (void *)vmpage->private;
 746              page != NULL; page = page->cp_child) {
 747                 if (cl_object_same(page->cp_obj, obj)) {
 748                         cl_page_get_trust(page);
 749                         break;
 750                 }
 751         }
 752         cfs_spin_unlock(&hdr->coh_page_guard);
 753         LASSERT(ergo(page, page->cp_type == CPT_CACHEABLE));
 754         RETURN(page);
 755 }
 756 EXPORT_SYMBOL(cl_vmpage_page);
 757
 758 /**
 759  * Returns the top-page for a given page.
 760  *
 761  * \see cl_object_top(), cl_io_top()
 762  */
 763 struct cl_page *cl_page_top(struct cl_page *page)
 764 {
 765         return cl_page_top_trusted(page);
 766 }
 767 EXPORT_SYMBOL(cl_page_top);
 768
 769 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 770                                        const struct lu_device_type *dtype)
 771 {
 772         return cl_page_at_trusted(page, dtype);
 773 }
 774 EXPORT_SYMBOL(cl_page_at);
 775
 776 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 777
 778 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)                   \
 779 ({                                                                      \
 780         const struct lu_env        *__env  = (_env);                    \
 781         struct cl_page             *__page = (_page);                   \
 782         const struct cl_page_slice *__scan;                             \
 783         int                         __result;                           \
 784         ptrdiff_t                   __op   = (_op);                     \
 785         int                       (*__method)_proto;                    \
 786                                                                         \
 787         __result = 0;                                                   \
 788         __page = cl_page_top(__page);                                   \
 789         do {                                                            \
 790                 cfs_list_for_each_entry(__scan, &__page->cp_layers,     \
 791                                         cpl_linkage) {                  \
 792                         __method = *(void **)((char *)__scan->cpl_ops + \
 793                                               __op);                    \
 794                         if (__method != NULL) {                         \
 795                                 __result = (*__method)(__env, __scan,   \
 796                                                        ## __VA_ARGS__); \
 797                                 if (__result != 0)                      \
 798                                         break;                          \
 799                         }                                               \
 800                 }                                                       \
 801                 __page = __page->cp_child;                              \
 802         } while (__page != NULL && __result == 0);                      \
 803         if (__result > 0)                                               \
 804                 __result = 0;                                           \
 805         __result;                                                       \
 806 })
 807
 808 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 809 do {                                                                    \
 810         const struct lu_env        *__env  = (_env);                    \
 811         struct cl_page             *__page = (_page);                   \
 812         const struct cl_page_slice *__scan;                             \
 813         ptrdiff_t                   __op   = (_op);                     \
 814         void                      (*__method)_proto;                    \
 815                                                                         \
 816         __page = cl_page_top(__page);                                   \
 817         do {                                                            \
 818                 cfs_list_for_each_entry(__scan, &__page->cp_layers,     \
 819                                         cpl_linkage) {                  \
 820                         __method = *(void **)((char *)__scan->cpl_ops + \
 821                                               __op);                    \
 822                         if (__method != NULL)                           \
 823                                 (*__method)(__env, __scan,              \
 824                                             ## __VA_ARGS__);            \
 825                 }                                                       \
 826                 __page = __page->cp_child;                              \
 827         } while (__page != NULL);                                       \
 828 } while (0)
 829
 830 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)               \
 831 do {                                                                        \
 832         const struct lu_env        *__env  = (_env);                        \
 833         struct cl_page             *__page = (_page);                       \
 834         const struct cl_page_slice *__scan;                                 \
 835         ptrdiff_t                   __op   = (_op);                         \
 836         void                      (*__method)_proto;                        \
 837                                                                             \
 838         /* get to the bottom page. */                                       \
 839         while (__page->cp_child != NULL)                                    \
 840                 __page = __page->cp_child;                                  \
 841         do {                                                                \
 842                 cfs_list_for_each_entry_reverse(__scan, &__page->cp_layers, \
 843                                                 cpl_linkage) {              \
 844                         __method = *(void **)((char *)__scan->cpl_ops +     \
 845                                               __op);                        \
 846                         if (__method != NULL)                               \
 847                                 (*__method)(__env, __scan,                  \
 848                                             ## __VA_ARGS__);                \
 849                 }                                                           \
 850                 __page = __page->cp_parent;                                 \
 851         } while (__page != NULL);                                           \
 852 } while (0)
 853
 854 static int cl_page_invoke(const struct lu_env *env,
 855                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 856
 857 {
 858         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 859         ENTRY;
 860         RETURN(CL_PAGE_INVOKE(env, page, op,
 861                               (const struct lu_env *,
 862                                const struct cl_page_slice *, struct cl_io *),
 863                               io));
 864 }
 865
 866 static void cl_page_invoid(const struct lu_env *env,
 867                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 868
 869 {
 870         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 871         ENTRY;
 872         CL_PAGE_INVOID(env, page, op,
 873                        (const struct lu_env *,
 874                         const struct cl_page_slice *, struct cl_io *), io);
 875         EXIT;
 876 }
 877
 878 static void cl_page_owner_clear(struct cl_page *page)
 879 {
 880         ENTRY;
 881         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 882                 if (page->cp_owner != NULL) {
 883                         LASSERT(page->cp_owner->ci_owned_nr > 0);
 884                         page->cp_owner->ci_owned_nr--;
 885                         page->cp_owner = NULL;
 886                         page->cp_task = NULL;
 887                 }
 888         }
 889         EXIT;
 890 }
 891
 892 static void cl_page_owner_set(struct cl_page *page)
 893 {
 894         ENTRY;
 895         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 896                 LASSERT(page->cp_owner != NULL);
 897                 page->cp_owner->ci_owned_nr++;
 898         }
 899         EXIT;
 900 }
 901
 902 void cl_page_disown0(const struct lu_env *env,
 903                      struct cl_io *io, struct cl_page *pg)
 904 {
 905         enum cl_page_state state;
 906
 907         ENTRY;
 908         state = pg->cp_state;
 909         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 910         PINVRNT(env, pg, cl_page_invariant(pg));
 911         cl_page_owner_clear(pg);
 912
 913         if (state == CPS_OWNED)
 914                 cl_page_state_set(env, pg, CPS_CACHED);
 915         /*
 916          * Completion call-backs are executed in the bottom-up order, so that
 917          * uppermost layer (llite), responsible for VFS/VM interaction runs
 918          * last and can release locks safely.
 919          */
 920         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 921                                (const struct lu_env *,
 922                                 const struct cl_page_slice *, struct cl_io *),
 923                                io);
 924         EXIT;
 925 }
 926
 927 /**
 928  * returns true, iff page is owned by the given io.
 929  */
 930 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 931 {
 932         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 933         ENTRY;
 934         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == io);
 935 }
 936 EXPORT_SYMBOL(cl_page_is_owned);
 937
 938 /**
 939  * Try to own a page by IO.
 940  *
 941  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 942  * into cl_page_state::CPS_OWNED state.
 943  *
 944  * \pre  !cl_page_is_owned(pg, io)
 945  * \post result == 0 iff cl_page_is_owned(pg, io)
 946  *
 947  * \retval 0   success
 948  *
 949  * \retval -ve failure, e.g., page was destroyed (and landed in
 950  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 951  *             or, page was owned by another thread, or in IO.
 952  *
 953  * \see cl_page_disown()
 954  * \see cl_page_operations::cpo_own()
 955  * \see cl_page_own_try()
 956  * \see cl_page_own
 957  */
 958 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 959                         struct cl_page *pg, int nonblock)
 960 {
 961         int result;
 962
 963         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 964
 965         ENTRY;
 966         pg = cl_page_top(pg);
 967         io = cl_io_top(io);
 968
 969         if (pg->cp_state == CPS_FREEING) {
 970                 result = -ENOENT;
 971         } else {
 972                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 973                                         (const struct lu_env *,
 974                                          const struct cl_page_slice *,
 975                                          struct cl_io *, int),
 976                                         io, nonblock);
 977                 if (result == 0) {
 978                         PASSERT(env, pg, pg->cp_owner == NULL);
 979                         PASSERT(env, pg, pg->cp_req == NULL);
 980                         pg->cp_owner = io;
 981                         pg->cp_task  = current;
 982                         cl_page_owner_set(pg);
 983                         if (pg->cp_state != CPS_FREEING) {
 984                                 cl_page_state_set(env, pg, CPS_OWNED);
 985                         } else {
 986                                 cl_page_disown0(env, io, pg);
 987                                 result = -ENOENT;
 988                         }
 989                 }
 990         }
 991         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 992         RETURN(result);
 993 }
 994
 995 /**
 996  * Own a page, might be blocked.
 997  *
 998  * \see cl_page_own0()
 999  */
1000 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
1001 {
1002         return cl_page_own0(env, io, pg, 0);
1003 }
1004 EXPORT_SYMBOL(cl_page_own);
1005
1006 /**
1007  * Nonblock version of cl_page_own().
1008  *
1009  * \see cl_page_own0()
1010  */
1011 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
1012                     struct cl_page *pg)
1013 {
1014         return cl_page_own0(env, io, pg, 1);
1015 }
1016 EXPORT_SYMBOL(cl_page_own_try);
1017
1018
1019 /**
1020  * Assume page ownership.
1021  *
1022  * Called when page is already locked by the hosting VM.
1023  *
1024  * \pre !cl_page_is_owned(pg, io)
1025  * \post cl_page_is_owned(pg, io)
1026  *
1027  * \see cl_page_operations::cpo_assume()
1028  */
1029 void cl_page_assume(const struct lu_env *env,
1030                     struct cl_io *io, struct cl_page *pg)
1031 {
1032         PASSERT(env, pg, pg->cp_owner == NULL);
1033         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
1034         PINVRNT(env, pg, cl_page_invariant(pg));
1035
1036         ENTRY;
1037         pg = cl_page_top(pg);
1038         io = cl_io_top(io);
1039
1040         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
1041         pg->cp_owner = io;
1042         pg->cp_task = current;
1043         cl_page_owner_set(pg);
1044         cl_page_state_set(env, pg, CPS_OWNED);
1045         EXIT;
1046 }
1047 EXPORT_SYMBOL(cl_page_assume);
1048
1049 /**
1050  * Releases page ownership without unlocking the page.
1051  *
1052  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
1053  * underlying VM page (as VM is supposed to do this itself).
1054  *
1055  * \pre   cl_page_is_owned(pg, io)
1056  * \post !cl_page_is_owned(pg, io)
1057  *
1058  * \see cl_page_assume()
1059  */
1060 void cl_page_unassume(const struct lu_env *env,
1061                       struct cl_io *io, struct cl_page *pg)
1062 {
1063         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1064         PINVRNT(env, pg, cl_page_invariant(pg));
1065
1066         ENTRY;
1067         pg = cl_page_top(pg);
1068         io = cl_io_top(io);
1069         cl_page_owner_clear(pg);
1070         cl_page_state_set(env, pg, CPS_CACHED);
1071         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
1072                                (const struct lu_env *,
1073                                 const struct cl_page_slice *, struct cl_io *),
1074                                io);
1075         EXIT;
1076 }
1077 EXPORT_SYMBOL(cl_page_unassume);
1078
1079 /**
1080  * Releases page ownership.
1081  *
1082  * Moves page into cl_page_state::CPS_CACHED.
1083  *
1084  * \pre   cl_page_is_owned(pg, io)
1085  * \post !cl_page_is_owned(pg, io)
1086  *
1087  * \see cl_page_own()
1088  * \see cl_page_operations::cpo_disown()
1089  */
1090 void cl_page_disown(const struct lu_env *env,
1091                     struct cl_io *io, struct cl_page *pg)
1092 {
1093         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1094
1095         ENTRY;
1096         pg = cl_page_top(pg);
1097         io = cl_io_top(io);
1098         cl_page_disown0(env, io, pg);
1099         EXIT;
1100 }
1101 EXPORT_SYMBOL(cl_page_disown);
1102
1103 /**
1104  * Called when page is to be removed from the object, e.g., as a result of
1105  * truncate.
1106  *
1107  * Calls cl_page_operations::cpo_discard() top-to-bottom.
1108  *
1109  * \pre cl_page_is_owned(pg, io)
1110  *
1111  * \see cl_page_operations::cpo_discard()
1112  */
1113 void cl_page_discard(const struct lu_env *env,
1114                      struct cl_io *io, struct cl_page *pg)
1115 {
1116         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1117         PINVRNT(env, pg, cl_page_invariant(pg));
1118
1119         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
1120 }
1121 EXPORT_SYMBOL(cl_page_discard);
1122
1123 /**
1124  * Version of cl_page_delete() that can be called for not fully constructed
1125  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
1126  * path. Doesn't check page invariant.
1127  */
1128 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
1129                             int radix)
1130 {
1131         struct cl_page *tmp = pg;
1132         ENTRY;
1133
1134         PASSERT(env, pg, pg == cl_page_top(pg));
1135         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
1136
1137         /*
1138          * Severe all ways to obtain new pointers to @pg.
1139          */
1140         cl_page_owner_clear(pg);
1141
1142         /*
1143          * unexport the page firstly before freeing it so that
1144          * the page content is considered to be invalid.
1145          * We have to do this because a CPS_FREEING cl_page may
1146          * be NOT under the protection of a cl_lock.
1147          * Afterwards, if this page is found by other threads, then this
1148          * page will be forced to reread.
1149          */
1150         cl_page_export(env, pg, 0);
1151         cl_page_state_set0(env, pg, CPS_FREEING);
1152
1153         if (tmp->cp_type == CPT_CACHEABLE) {
1154                 if (!radix)
1155                         /* !radix means that @pg is not yet in the radix tree,
1156                          * skip removing it.
1157                          */
1158                         tmp = pg->cp_child;
1159                 for (; tmp != NULL; tmp = tmp->cp_child) {
1160                         void                    *value;
1161                         struct cl_object_header *hdr;
1162
1163                         hdr = cl_object_header(tmp->cp_obj);
1164                         cfs_spin_lock(&hdr->coh_page_guard);
1165                         value = radix_tree_delete(&hdr->coh_tree,
1166                                                   tmp->cp_index);
1167                         PASSERT(env, tmp, value == tmp);
1168                         PASSERT(env, tmp, hdr->coh_pages > 0);
1169                         hdr->coh_pages--;
1170                         cfs_spin_unlock(&hdr->coh_page_guard);
1171                 }
1172         }
1173
1174         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete),
1175                        (const struct lu_env *, const struct cl_page_slice *));
1176         EXIT;
1177 }
1178
1179 /**
1180  * Called when a decision is made to throw page out of memory.
1181  *
1182  * Notifies all layers about page destruction by calling
1183  * cl_page_operations::cpo_delete() method top-to-bottom.
1184  *
1185  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
1186  * where transition to this state happens).
1187  *
1188  * Eliminates all venues through which new references to the page can be
1189  * obtained:
1190  *
1191  *     - removes page from the radix trees,
1192  *
1193  *     - breaks linkage from VM page to cl_page.
1194  *
1195  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
1196  * drain after some time, at which point page will be recycled.
1197  *
1198  * \pre  pg == cl_page_top(pg)
1199  * \pre  VM page is locked
1200  * \post pg->cp_state == CPS_FREEING
1201  *
1202  * \see cl_page_operations::cpo_delete()
1203  */
1204 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
1205 {
1206         PINVRNT(env, pg, cl_page_invariant(pg));
1207         ENTRY;
1208         cl_page_delete0(env, pg, 1);
1209         EXIT;
1210 }
1211 EXPORT_SYMBOL(cl_page_delete);
1212
1213 /**
1214  * Unmaps page from user virtual memory.
1215  *
1216  * Calls cl_page_operations::cpo_unmap() through all layers top-to-bottom. The
1217  * layer responsible for VM interaction has to unmap page from user space
1218  * virtual memory.
1219  *
1220  * \see cl_page_operations::cpo_unmap()
1221  */
1222 int cl_page_unmap(const struct lu_env *env,
1223                   struct cl_io *io, struct cl_page *pg)
1224 {
1225         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1226         PINVRNT(env, pg, cl_page_invariant(pg));
1227
1228         return cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_unmap));
1229 }
1230 EXPORT_SYMBOL(cl_page_unmap);
1231
1232 /**
1233  * Marks page up-to-date.
1234  *
1235  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
1236  * layer responsible for VM interaction has to mark/clear page as up-to-date
1237  * by the \a uptodate argument.
1238  *
1239  * \see cl_page_operations::cpo_export()
1240  */
1241 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
1242 {
1243         PINVRNT(env, pg, cl_page_invariant(pg));
1244         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
1245                        (const struct lu_env *,
1246                         const struct cl_page_slice *, int), uptodate);
1247 }
1248 EXPORT_SYMBOL(cl_page_export);
1249
1250 /**
1251  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
1252  * thread.
1253  */
1254 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
1255 {
1256         int result;
1257         const struct cl_page_slice *slice;
1258
1259         ENTRY;
1260         pg = cl_page_top_trusted((struct cl_page *)pg);
1261         slice = container_of(pg->cp_layers.next,
1262                              const struct cl_page_slice, cpl_linkage);
1263         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
1264         /*
1265          * Call ->cpo_is_vmlocked() directly instead of going through
1266          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
1267          * cl_page_invariant().
1268          */
1269         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
1270         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
1271         RETURN(result == -EBUSY);
1272 }
1273 EXPORT_SYMBOL(cl_page_is_vmlocked);
1274
1275 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
1276 {
1277         ENTRY;
1278         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
1279 }
1280
1281 static void cl_page_io_start(const struct lu_env *env,
1282                              struct cl_page *pg, enum cl_req_type crt)
1283 {
1284         /*
1285          * Page is queued for IO, change its state.
1286          */
1287         ENTRY;
1288         cl_page_owner_clear(pg);
1289         cl_page_state_set(env, pg, cl_req_type_state(crt));
1290         EXIT;
1291 }
1292
1293 /**
1294  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
1295  * called top-to-bottom. Every layer either agrees to submit this page (by
1296  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
1297  * handling interactions with the VM also has to inform VM that page is under
1298  * transfer now.
1299  */
1300 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
1301                  struct cl_page *pg, enum cl_req_type crt)
1302 {
1303         int result;
1304
1305         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1306         PINVRNT(env, pg, cl_page_invariant(pg));
1307         PINVRNT(env, pg, crt < CRT_NR);
1308
1309         /*
1310          * XXX this has to be called bottom-to-top, so that llite can set up
1311          * PG_writeback without risking other layers deciding to skip this
1312          * page.
1313          */
1314         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
1315         if (result == 0)
1316                 cl_page_io_start(env, pg, crt);
1317
1318         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
1319                       equi(result == 0,
1320                            PageWriteback(cl_page_vmpage(env, pg)))));
1321         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1322         return result;
1323 }
1324 EXPORT_SYMBOL(cl_page_prep);
1325
1326 /**
1327  * Notify layers about transfer completion.
1328  *
1329  * Invoked by transfer sub-system (which is a part of osc) to notify layers
1330  * that a transfer, of which this page is a part of has completed.
1331  *
1332  * Completion call-backs are executed in the bottom-up order, so that
1333  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1334  * and can release locks safely.
1335  *
1336  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1337  * \post pg->cp_state == CPS_CACHED
1338  *
1339  * \see cl_page_operations::cpo_completion()
1340  */
1341 void cl_page_completion(const struct lu_env *env,
1342                         struct cl_page *pg, enum cl_req_type crt, int ioret)
1343 {
1344         struct cl_sync_io *anchor = pg->cp_sync_io;
1345
1346         PASSERT(env, pg, crt < CRT_NR);
1347         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
1348         PASSERT(env, pg, pg->cp_req == NULL);
1349         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
1350
1351         ENTRY;
1352         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
1353         if (crt == CRT_READ && ioret == 0) {
1354                 PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED));
1355                 pg->cp_flags |= CPF_READ_COMPLETED;
1356         }
1357
1358         cl_page_state_set(env, pg, CPS_CACHED);
1359         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
1360                                (const struct lu_env *,
1361                                 const struct cl_page_slice *, int), ioret);
1362         if (anchor) {
1363                 LASSERT(cl_page_is_vmlocked(env, pg));
1364                 LASSERT(pg->cp_sync_io == anchor);
1365                 pg->cp_sync_io = NULL;
1366                 cl_sync_io_note(anchor, ioret);
1367         }
1368
1369         /* Don't assert the page writeback bit here because the lustre file
1370          * may be as a backend of swap space. in this case, the page writeback
1371          * is set by VM, and obvious we shouldn't clear it at all. Fortunately
1372          * this type of pages are all TRANSIENT pages. */
1373         KLASSERT(ergo(pg->cp_type == CPT_CACHEABLE,
1374                       !PageWriteback(cl_page_vmpage(env, pg))));
1375         EXIT;
1376 }
1377 EXPORT_SYMBOL(cl_page_completion);
1378
1379 /**
1380  * Notify layers that transfer formation engine decided to yank this page from
1381  * the cache and to make it a part of a transfer.
1382  *
1383  * \pre  pg->cp_state == CPS_CACHED
1384  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1385  *
1386  * \see cl_page_operations::cpo_make_ready()
1387  */
1388 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
1389                        enum cl_req_type crt)
1390 {
1391         int result;
1392
1393         PINVRNT(env, pg, crt < CRT_NR);
1394
1395         ENTRY;
1396         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
1397                                 (const struct lu_env *,
1398                                  const struct cl_page_slice *));
1399         if (result == 0) {
1400                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
1401                 cl_page_io_start(env, pg, crt);
1402         }
1403         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1404         RETURN(result);
1405 }
1406 EXPORT_SYMBOL(cl_page_make_ready);
1407
1408 /**
1409  * Notify layers that high level io decided to place this page into a cache
1410  * for future transfer.
1411  *
1412  * The layer implementing transfer engine (osc) has to register this page in
1413  * its queues.
1414  *
1415  * \pre  cl_page_is_owned(pg, io)
1416  * \post ergo(result == 0,
1417  *            pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT)
1418  *
1419  * \see cl_page_operations::cpo_cache_add()
1420  */
1421 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
1422                       struct cl_page *pg, enum cl_req_type crt)
1423 {
1424         int result;
1425
1426         PINVRNT(env, pg, crt < CRT_NR);
1427         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1428         PINVRNT(env, pg, cl_page_invariant(pg));
1429
1430         ENTRY;
1431         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_cache_add));
1432         if (result == 0) {
1433                 cl_page_owner_clear(pg);
1434                 cl_page_state_set(env, pg, CPS_CACHED);
1435         }
1436         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1437         RETURN(result);
1438 }
1439 EXPORT_SYMBOL(cl_page_cache_add);
1440
1441 /**
1442  * Checks whether page is protected by any extent lock is at least required
1443  * mode.
1444  *
1445  * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1446  * \see cl_page_operations::cpo_is_under_lock()
1447  */
1448 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1449                           struct cl_page *page)
1450 {
1451         int rc;
1452
1453         PINVRNT(env, page, cl_page_invariant(page));
1454
1455         ENTRY;
1456         rc = CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1457                             (const struct lu_env *,
1458                              const struct cl_page_slice *, struct cl_io *),
1459                             io);
1460         PASSERT(env, page, rc != 0);
1461         RETURN(rc);
1462 }
1463 EXPORT_SYMBOL(cl_page_is_under_lock);
1464
1465 static int page_prune_cb(const struct lu_env *env, struct cl_io *io,
1466                          struct cl_page *page, void *cbdata)
1467 {
1468         cl_page_own(env, io, page);
1469         cl_page_unmap(env, io, page);
1470         cl_page_discard(env, io, page);
1471         cl_page_disown(env, io, page);
1472         return CLP_GANG_OKAY;
1473 }
1474
1475 /**
1476  * Purges all cached pages belonging to the object \a obj.
1477  */
1478 int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj)
1479 {
1480         struct cl_thread_info   *info;
1481         struct cl_object        *obj = cl_object_top(clobj);
1482         struct cl_io            *io;
1483         int                      result;
1484
1485         ENTRY;
1486         info  = cl_env_info(env);
1487         io    = &info->clt_io;
1488
1489         /*
1490          * initialize the io. This is ugly since we never do IO in this
1491          * function, we just make cl_page_list functions happy. -jay
1492          */
1493         io->ci_obj = obj;
1494         result = cl_io_init(env, io, CIT_MISC, obj);
1495         if (result != 0) {
1496                 cl_io_fini(env, io);
1497                 RETURN(io->ci_result);
1498         }
1499
1500         do {
1501                 result = cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF,
1502                                              page_prune_cb, NULL);
1503                 if (result == CLP_GANG_RESCHED)
1504                         cfs_cond_resched();
1505         } while (result != CLP_GANG_OKAY);
1506
1507         cl_io_fini(env, io);
1508         RETURN(result);
1509 }
1510 EXPORT_SYMBOL(cl_pages_prune);
1511
1512 /**
1513  * Tells transfer engine that only part of a page is to be transmitted.
1514  *
1515  * \see cl_page_operations::cpo_clip()
1516  */
1517 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1518                   int from, int to)
1519 {
1520         PINVRNT(env, pg, cl_page_invariant(pg));
1521
1522         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1523         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1524                        (const struct lu_env *,
1525                         const struct cl_page_slice *,int, int),
1526                        from, to);
1527 }
1528 EXPORT_SYMBOL(cl_page_clip);
1529
1530 /**
1531  * Prints human readable representation of \a pg to the \a f.
1532  */
1533 void cl_page_header_print(const struct lu_env *env, void *cookie,
1534                           lu_printer_t printer, const struct cl_page *pg)
1535 {
1536         (*printer)(env, cookie,
1537                    "page@%p[%d %p:%lu ^%p_%p %d %d %d %p %p %#x]\n",
1538                    pg, cfs_atomic_read(&pg->cp_ref), pg->cp_obj,
1539                    pg->cp_index, pg->cp_parent, pg->cp_child,
1540                    pg->cp_state, pg->cp_error, pg->cp_type,
1541                    pg->cp_owner, pg->cp_req, pg->cp_flags);
1542 }
1543 EXPORT_SYMBOL(cl_page_header_print);
1544
1545 /**
1546  * Prints human readable representation of \a pg to the \a f.
1547  */
1548 void cl_page_print(const struct lu_env *env, void *cookie,
1549                    lu_printer_t printer, const struct cl_page *pg)
1550 {
1551         struct cl_page *scan;
1552
1553         for (scan = cl_page_top((struct cl_page *)pg);
1554              scan != NULL; scan = scan->cp_child)
1555                 cl_page_header_print(env, cookie, printer, scan);
1556         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1557                        (const struct lu_env *env,
1558                         const struct cl_page_slice *slice,
1559                         void *cookie, lu_printer_t p), cookie, printer);
1560         (*printer)(env, cookie, "end page@%p\n", pg);
1561 }
1562 EXPORT_SYMBOL(cl_page_print);
1563
1564 /**
1565  * Cancel a page which is still in a transfer.
1566  */
1567 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1568 {
1569         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1570                               (const struct lu_env *,
1571                                const struct cl_page_slice *));
1572 }
1573 EXPORT_SYMBOL(cl_page_cancel);
1574
1575 /**
1576  * Converts a byte offset within object \a obj into a page index.
1577  */
1578 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1579 {
1580         /*
1581          * XXX for now.
1582          */
1583         return (loff_t)idx << CFS_PAGE_SHIFT;
1584 }
1585 EXPORT_SYMBOL(cl_offset);
1586
1587 /**
1588  * Converts a page index into a byte offset within object \a obj.
1589  */
1590 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1591 {
1592         /*
1593          * XXX for now.
1594          */
1595         return offset >> CFS_PAGE_SHIFT;
1596 }
1597 EXPORT_SYMBOL(cl_index);
1598
1599 int cl_page_size(const struct cl_object *obj)
1600 {
1601         return 1 << CFS_PAGE_SHIFT;
1602 }
1603 EXPORT_SYMBOL(cl_page_size);
1604
1605 /**
1606  * Adds page slice to the compound page.
1607  *
1608  * This is called by cl_object_operations::coo_page_init() methods to add a
1609  * per-layer state to the page. New state is added at the end of
1610  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1611  *
1612  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1613  */
1614 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1615                        struct cl_object *obj,
1616                        const struct cl_page_operations *ops)
1617 {
1618         ENTRY;
1619         cfs_list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1620         slice->cpl_obj  = obj;
1621         slice->cpl_ops  = ops;
1622         slice->cpl_page = page;
1623         EXIT;
1624 }
1625 EXPORT_SYMBOL(cl_page_slice_add);
1626
1627 int  cl_page_init(void)
1628 {
1629         return lu_kmem_init(cl_page_caches);
1630 }
1631
1632 void cl_page_fini(void)
1633 {
1634         lu_kmem_fini(cl_page_caches);
1635 }