lustre/obdclass/cl_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Client Lustre Page.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_CLASS
  42
  43 #include <libcfs/libcfs.h>
  44 #include <obd_class.h>
  45 #include <obd_support.h>
  46 #include <libcfs/list.h>
  47
  48 #include <cl_object.h>
  49 #include "cl_internal.h"
  50
  51 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
  52                             int radix);
  53
  54 #ifdef LIBCFS_DEBUG
  55 # define PASSERT(env, page, expr)                                       \
  56   do {                                                                    \
  57           if (unlikely(!(expr))) {                                      \
  58                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  59                   LASSERT(0);                                           \
  60           }                                                             \
  61   } while (0)
  62 #else /* !LIBCFS_DEBUG */
  63 # define PASSERT(env, page, exp) \
  64         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  65 #endif /* !LIBCFS_DEBUG */
  66
  67 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
  68 # define PINVRNT(env, page, expr)                                       \
  69   do {                                                                    \
  70           if (unlikely(!(expr))) {                                      \
  71                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  72                   LINVRNT(0);                                           \
  73           }                                                             \
  74   } while (0)
  75 #else /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
  76 # define PINVRNT(env, page, exp) \
  77          ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  78 #endif /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
  79
  80 /* Disable page statistic by default due to huge performance penalty. */
  81 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
  82 #define CS_PAGE_INC(o, item) \
  83         cfs_atomic_inc(&cl_object_site(o)->cs_pages.cs_stats[CS_##item])
  84 #define CS_PAGE_DEC(o, item) \
  85         cfs_atomic_dec(&cl_object_site(o)->cs_pages.cs_stats[CS_##item])
  86 #define CS_PAGESTATE_INC(o, state) \
  87         cfs_atomic_inc(&cl_object_site(o)->cs_pages_state[state])
  88 #define CS_PAGESTATE_DEC(o, state) \
  89         cfs_atomic_dec(&cl_object_site(o)->cs_pages_state[state])
  90 #else
  91 #define CS_PAGE_INC(o, item)
  92 #define CS_PAGE_DEC(o, item)
  93 #define CS_PAGESTATE_INC(o, state)
  94 #define CS_PAGESTATE_DEC(o, state)
  95 #endif
  96
  97 /**
  98  * Internal version of cl_page_top, it should be called if the page is
  99  * known to be not freed, says with page referenced, or radix tree lock held,
 100  * or page owned.
 101  */
 102 static struct cl_page *cl_page_top_trusted(struct cl_page *page)
 103 {
 104         while (page->cp_parent != NULL)
 105                 page = page->cp_parent;
 106         return page;
 107 }
 108
 109 /**
 110  * Internal version of cl_page_get().
 111  *
 112  * This function can be used to obtain initial reference to previously
 113  * unreferenced cached object. It can be called only if concurrent page
 114  * reclamation is somehow prevented, e.g., by locking page radix-tree
 115  * (cl_object_header::hdr->coh_page_guard), or by keeping a lock on a VM page,
 116  * associated with \a page.
 117  *
 118  * Use with care! Not exported.
 119  */
 120 static void cl_page_get_trust(struct cl_page *page)
 121 {
 122         LASSERT(cfs_atomic_read(&page->cp_ref) > 0);
 123         cfs_atomic_inc(&page->cp_ref);
 124 }
 125
 126 /**
 127  * Returns a slice within a page, corresponding to the given layer in the
 128  * device stack.
 129  *
 130  * \see cl_lock_at()
 131  */
 132 static const struct cl_page_slice *
 133 cl_page_at_trusted(const struct cl_page *page,
 134                    const struct lu_device_type *dtype)
 135 {
 136         const struct cl_page_slice *slice;
 137         ENTRY;
 138
 139         page = cl_page_top_trusted((struct cl_page *)page);
 140         do {
 141                 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 142                         if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
 143                                 RETURN(slice);
 144                 }
 145                 page = page->cp_child;
 146         } while (page != NULL);
 147         RETURN(NULL);
 148 }
 149
 150 /**
 151  * Returns a page with given index in the given object, or NULL if no page is
 152  * found. Acquires a reference on \a page.
 153  *
 154  * Locking: called under cl_object_header::coh_page_guard spin-lock.
 155  */
 156 struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index)
 157 {
 158         struct cl_page *page;
 159
 160         LASSERT_SPIN_LOCKED(&hdr->coh_page_guard);
 161
 162         page = radix_tree_lookup(&hdr->coh_tree, index);
 163         if (page != NULL)
 164                 cl_page_get_trust(page);
 165         return page;
 166 }
 167 EXPORT_SYMBOL(cl_page_lookup);
 168
 169 /**
 170  * Returns a list of pages by a given [start, end] of \a obj.
 171  *
 172  * \param resched If not NULL, then we give up before hogging CPU for too
 173  * long and set *resched = 1, in that case caller should implement a retry
 174  * logic.
 175  *
 176  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
 177  * crucial in the face of [offset, EOF] locks.
 178  *
 179  * Return at least one page in @queue unless there is no covered page.
 180  */
 181 int cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj,
 182                         struct cl_io *io, pgoff_t start, pgoff_t end,
 183                         cl_page_gang_cb_t cb, void *cbdata)
 184 {
 185         struct cl_object_header *hdr;
 186         struct cl_page          *page;
 187         struct cl_page         **pvec;
 188         const struct cl_page_slice  *slice;
 189         const struct lu_device_type *dtype;
 190         pgoff_t                  idx;
 191         unsigned int             nr;
 192         unsigned int             i;
 193         unsigned int             j;
 194         int                      res = CLP_GANG_OKAY;
 195         int                      tree_lock = 1;
 196         ENTRY;
 197
 198         idx = start;
 199         hdr = cl_object_header(obj);
 200         pvec = cl_env_info(env)->clt_pvec;
 201         dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type;
 202         spin_lock(&hdr->coh_page_guard);
 203         while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec,
 204                                             idx, CLT_PVEC_SIZE)) > 0) {
 205                 int end_of_region = 0;
 206                 idx = pvec[nr - 1]->cp_index + 1;
 207                 for (i = 0, j = 0; i < nr; ++i) {
 208                         page = pvec[i];
 209                         pvec[i] = NULL;
 210
 211                         LASSERT(page->cp_type == CPT_CACHEABLE);
 212                         if (page->cp_index > end) {
 213                                 end_of_region = 1;
 214                                 break;
 215                         }
 216                         if (page->cp_state == CPS_FREEING)
 217                                 continue;
 218
 219                         slice = cl_page_at_trusted(page, dtype);
 220                         /*
 221                          * Pages for lsm-less file has no underneath sub-page
 222                          * for osc, in case of ...
 223                          */
 224                         PASSERT(env, page, slice != NULL);
 225
 226                         page = slice->cpl_page;
 227                         /*
 228                          * Can safely call cl_page_get_trust() under
 229                          * radix-tree spin-lock.
 230                          *
 231                          * XXX not true, because @page is from object another
 232                          * than @hdr and protected by different tree lock.
 233                          */
 234                         cl_page_get_trust(page);
 235                         lu_ref_add_atomic(&page->cp_reference,
 236                                           "gang_lookup", cfs_current());
 237                         pvec[j++] = page;
 238                 }
 239
 240                 /*
 241                  * Here a delicate locking dance is performed. Current thread
 242                  * holds a reference to a page, but has to own it before it
 243                  * can be placed into queue. Owning implies waiting, so
 244                  * radix-tree lock is to be released. After a wait one has to
 245                  * check that pages weren't truncated (cl_page_own() returns
 246                  * error in the latter case).
 247                  */
 248                 spin_unlock(&hdr->coh_page_guard);
 249                 tree_lock = 0;
 250
 251                 for (i = 0; i < j; ++i) {
 252                         page = pvec[i];
 253                         if (res == CLP_GANG_OKAY)
 254                                 res = (*cb)(env, io, page, cbdata);
 255                         lu_ref_del(&page->cp_reference,
 256                                    "gang_lookup", cfs_current());
 257                         cl_page_put(env, page);
 258                 }
 259                 if (nr < CLT_PVEC_SIZE || end_of_region)
 260                         break;
 261
 262                 if (res == CLP_GANG_OKAY && cfs_need_resched())
 263                         res = CLP_GANG_RESCHED;
 264                 if (res != CLP_GANG_OKAY)
 265                         break;
 266
 267                 spin_lock(&hdr->coh_page_guard);
 268                 tree_lock = 1;
 269         }
 270         if (tree_lock)
 271                 spin_unlock(&hdr->coh_page_guard);
 272         RETURN(res);
 273 }
 274 EXPORT_SYMBOL(cl_page_gang_lookup);
 275
 276 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
 277 {
 278         struct cl_object *obj  = page->cp_obj;
 279         int pagesize = cl_object_header(obj)->coh_page_bufsize;
 280
 281         PASSERT(env, page, cfs_list_empty(&page->cp_batch));
 282         PASSERT(env, page, page->cp_owner == NULL);
 283         PASSERT(env, page, page->cp_req == NULL);
 284         PASSERT(env, page, page->cp_parent == NULL);
 285         PASSERT(env, page, page->cp_state == CPS_FREEING);
 286
 287         ENTRY;
 288         cfs_might_sleep();
 289         while (!cfs_list_empty(&page->cp_layers)) {
 290                 struct cl_page_slice *slice;
 291
 292                 slice = cfs_list_entry(page->cp_layers.next,
 293                                        struct cl_page_slice, cpl_linkage);
 294                 cfs_list_del_init(page->cp_layers.next);
 295                 slice->cpl_ops->cpo_fini(env, slice);
 296         }
 297         CS_PAGE_DEC(obj, total);
 298         CS_PAGESTATE_DEC(obj, page->cp_state);
 299         lu_object_ref_del_at(&obj->co_lu, &page->cp_obj_ref, "cl_page", page);
 300         cl_object_put(env, obj);
 301         lu_ref_fini(&page->cp_reference);
 302         OBD_FREE(page, pagesize);
 303         EXIT;
 304 }
 305
 306 /**
 307  * Helper function updating page state. This is the only place in the code
 308  * where cl_page::cp_state field is mutated.
 309  */
 310 static inline void cl_page_state_set_trust(struct cl_page *page,
 311                                            enum cl_page_state state)
 312 {
 313         /* bypass const. */
 314         *(enum cl_page_state *)&page->cp_state = state;
 315 }
 316
 317 static struct cl_page *cl_page_alloc(const struct lu_env *env,
 318                 struct cl_object *o, pgoff_t ind, struct page *vmpage,
 319                 enum cl_page_type type)
 320 {
 321         struct cl_page          *page;
 322         struct lu_object_header *head;
 323
 324         ENTRY;
 325         OBD_ALLOC_GFP(page, cl_object_header(o)->coh_page_bufsize,
 326                         CFS_ALLOC_IO);
 327         if (page != NULL) {
 328                 int result = 0;
 329                 cfs_atomic_set(&page->cp_ref, 1);
 330                 if (type == CPT_CACHEABLE) /* for radix tree */
 331                         cfs_atomic_inc(&page->cp_ref);
 332                 page->cp_obj = o;
 333                 cl_object_get(o);
 334                 lu_object_ref_add_at(&o->co_lu, &page->cp_obj_ref, "cl_page",
 335                                      page);
 336                 page->cp_index = ind;
 337                 cl_page_state_set_trust(page, CPS_CACHED);
 338                 page->cp_type = type;
 339                 CFS_INIT_LIST_HEAD(&page->cp_layers);
 340                 CFS_INIT_LIST_HEAD(&page->cp_batch);
 341                 CFS_INIT_LIST_HEAD(&page->cp_flight);
 342                 mutex_init(&page->cp_mutex);
 343                 lu_ref_init(&page->cp_reference);
 344                 head = o->co_lu.lo_header;
 345                 cfs_list_for_each_entry(o, &head->loh_layers,
 346                                         co_lu.lo_linkage) {
 347                         if (o->co_ops->coo_page_init != NULL) {
 348                                 result = o->co_ops->coo_page_init(env, o,
 349                                                                   page, vmpage);
 350                                 if (result != 0) {
 351                                         cl_page_delete0(env, page, 0);
 352                                         cl_page_free(env, page);
 353                                         page = ERR_PTR(result);
 354                                         break;
 355                                 }
 356                         }
 357                 }
 358                 if (result == 0) {
 359                         CS_PAGE_INC(o, total);
 360                         CS_PAGE_INC(o, create);
 361                         CS_PAGESTATE_DEC(o, CPS_CACHED);
 362                 }
 363         } else {
 364                 page = ERR_PTR(-ENOMEM);
 365         }
 366         RETURN(page);
 367 }
 368
 369 /**
 370  * Returns a cl_page with index \a idx at the object \a o, and associated with
 371  * the VM page \a vmpage.
 372  *
 373  * This is the main entry point into the cl_page caching interface. First, a
 374  * cache (implemented as a per-object radix tree) is consulted. If page is
 375  * found there, it is returned immediately. Otherwise new page is allocated
 376  * and returned. In any case, additional reference to page is acquired.
 377  *
 378  * \see cl_object_find(), cl_lock_find()
 379  */
 380 static struct cl_page *cl_page_find0(const struct lu_env *env,
 381                                      struct cl_object *o,
 382                                      pgoff_t idx, struct page *vmpage,
 383                                      enum cl_page_type type,
 384                                      struct cl_page *parent)
 385 {
 386         struct cl_page          *page = NULL;
 387         struct cl_page          *ghost = NULL;
 388         struct cl_object_header *hdr;
 389         int err;
 390
 391         LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 392         cfs_might_sleep();
 393
 394         ENTRY;
 395
 396         hdr = cl_object_header(o);
 397         CS_PAGE_INC(o, lookup);
 398
 399         CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
 400                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 401         /* fast path. */
 402         if (type == CPT_CACHEABLE) {
 403                 /* vmpage lock is used to protect the child/parent
 404                  * relationship */
 405                 KLASSERT(PageLocked(vmpage));
 406                 /*
 407                  * cl_vmpage_page() can be called here without any locks as
 408                  *
 409                  *     - "vmpage" is locked (which prevents ->private from
 410                  *       concurrent updates), and
 411                  *
 412                  *     - "o" cannot be destroyed while current thread holds a
 413                  *       reference on it.
 414                  */
 415                 page = cl_vmpage_page(vmpage, o);
 416                 PINVRNT(env, page,
 417                         ergo(page != NULL,
 418                              cl_page_vmpage(env, page) == vmpage &&
 419                              (void *)radix_tree_lookup(&hdr->coh_tree,
 420                                                        idx) == page));
 421         }
 422
 423         if (page != NULL) {
 424                 CS_PAGE_INC(o, hit);
 425                 RETURN(page);
 426         }
 427
 428         /* allocate and initialize cl_page */
 429         page = cl_page_alloc(env, o, idx, vmpage, type);
 430         if (IS_ERR(page))
 431                 RETURN(page);
 432
 433         if (type == CPT_TRANSIENT) {
 434                 if (parent) {
 435                         LASSERT(page->cp_parent == NULL);
 436                         page->cp_parent = parent;
 437                         parent->cp_child = page;
 438                 }
 439                 RETURN(page);
 440         }
 441
 442         /*
 443          * XXX optimization: use radix_tree_preload() here, and change tree
 444          * gfp mask to GFP_KERNEL in cl_object_header_init().
 445          */
 446         spin_lock(&hdr->coh_page_guard);
 447         err = radix_tree_insert(&hdr->coh_tree, idx, page);
 448         if (err != 0) {
 449                 ghost = page;
 450                 /*
 451                  * Noted by Jay: a lock on \a vmpage protects cl_page_find()
 452                  * from this race, but
 453                  *
 454                  *     0. it's better to have cl_page interface "locally
 455                  *     consistent" so that its correctness can be reasoned
 456                  *     about without appealing to the (obscure world of) VM
 457                  *     locking.
 458                  *
 459                  *     1. handling this race allows ->coh_tree to remain
 460                  *     consistent even when VM locking is somehow busted,
 461                  *     which is very useful during diagnosing and debugging.
 462                  */
 463                 page = ERR_PTR(err);
 464                 CL_PAGE_DEBUG(D_ERROR, env, ghost,
 465                               "fail to insert into radix tree: %d\n", err);
 466         } else {
 467                 if (parent) {
 468                         LASSERT(page->cp_parent == NULL);
 469                         page->cp_parent = parent;
 470                         parent->cp_child = page;
 471                 }
 472                 hdr->coh_pages++;
 473         }
 474         spin_unlock(&hdr->coh_page_guard);
 475
 476         if (unlikely(ghost != NULL)) {
 477                 cl_page_delete0(env, ghost, 0);
 478                 cl_page_free(env, ghost);
 479         }
 480         RETURN(page);
 481 }
 482
 483 struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o,
 484                              pgoff_t idx, struct page *vmpage,
 485                              enum cl_page_type type)
 486 {
 487         return cl_page_find0(env, o, idx, vmpage, type, NULL);
 488 }
 489 EXPORT_SYMBOL(cl_page_find);
 490
 491
 492 struct cl_page *cl_page_find_sub(const struct lu_env *env, struct cl_object *o,
 493                                  pgoff_t idx, struct page *vmpage,
 494                                  struct cl_page *parent)
 495 {
 496         return cl_page_find0(env, o, idx, vmpage, parent->cp_type, parent);
 497 }
 498 EXPORT_SYMBOL(cl_page_find_sub);
 499
 500 static inline int cl_page_invariant(const struct cl_page *pg)
 501 {
 502         struct cl_object_header *header;
 503         struct cl_page          *parent;
 504         struct cl_page          *child;
 505         struct cl_io            *owner;
 506
 507         /*
 508          * Page invariant is protected by a VM lock.
 509          */
 510         LINVRNT(cl_page_is_vmlocked(NULL, pg));
 511
 512         header = cl_object_header(pg->cp_obj);
 513         parent = pg->cp_parent;
 514         child  = pg->cp_child;
 515         owner  = pg->cp_owner;
 516
 517         return cl_page_in_use(pg) &&
 518                 ergo(parent != NULL, parent->cp_child == pg) &&
 519                 ergo(child != NULL, child->cp_parent == pg) &&
 520                 ergo(child != NULL, pg->cp_obj != child->cp_obj) &&
 521                 ergo(parent != NULL, pg->cp_obj != parent->cp_obj) &&
 522                 ergo(owner != NULL && parent != NULL,
 523                      parent->cp_owner == pg->cp_owner->ci_parent) &&
 524                 ergo(owner != NULL && child != NULL,
 525                      child->cp_owner->ci_parent == owner) &&
 526                 /*
 527                  * Either page is early in initialization (has neither child
 528                  * nor parent yet), or it is in the object radix tree.
 529                  */
 530                 ergo(pg->cp_state < CPS_FREEING && pg->cp_type == CPT_CACHEABLE,
 531                      (void *)radix_tree_lookup(&header->coh_tree,
 532                                                pg->cp_index) == pg ||
 533                      (child == NULL && parent == NULL));
 534 }
 535
 536 static void cl_page_state_set0(const struct lu_env *env,
 537                                struct cl_page *page, enum cl_page_state state)
 538 {
 539         enum cl_page_state old;
 540
 541         /*
 542          * Matrix of allowed state transitions [old][new], for sanity
 543          * checking.
 544          */
 545         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 546                 [CPS_CACHED] = {
 547                         [CPS_CACHED]  = 0,
 548                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 549                         [CPS_PAGEIN]  = 0,
 550                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 551                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 552                 },
 553                 [CPS_OWNED] = {
 554                         [CPS_CACHED]  = 1, /* release to the cache */
 555                         [CPS_OWNED]   = 0,
 556                         [CPS_PAGEIN]  = 1, /* start read immediately */
 557                         [CPS_PAGEOUT] = 1, /* start write immediately */
 558                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 559                 },
 560                 [CPS_PAGEIN] = {
 561                         [CPS_CACHED]  = 1, /* io completion */
 562                         [CPS_OWNED]   = 0,
 563                         [CPS_PAGEIN]  = 0,
 564                         [CPS_PAGEOUT] = 0,
 565                         [CPS_FREEING] = 0,
 566                 },
 567                 [CPS_PAGEOUT] = {
 568                         [CPS_CACHED]  = 1, /* io completion */
 569                         [CPS_OWNED]   = 0,
 570                         [CPS_PAGEIN]  = 0,
 571                         [CPS_PAGEOUT] = 0,
 572                         [CPS_FREEING] = 0,
 573                 },
 574                 [CPS_FREEING] = {
 575                         [CPS_CACHED]  = 0,
 576                         [CPS_OWNED]   = 0,
 577                         [CPS_PAGEIN]  = 0,
 578                         [CPS_PAGEOUT] = 0,
 579                         [CPS_FREEING] = 0,
 580                 }
 581         };
 582
 583         ENTRY;
 584         old = page->cp_state;
 585         PASSERT(env, page, allowed_transitions[old][state]);
 586         CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
 587         for (; page != NULL; page = page->cp_child) {
 588                 PASSERT(env, page, page->cp_state == old);
 589                 PASSERT(env, page,
 590                         equi(state == CPS_OWNED, page->cp_owner != NULL));
 591
 592                 CS_PAGESTATE_DEC(page->cp_obj, page->cp_state);
 593                 CS_PAGESTATE_INC(page->cp_obj, state);
 594                 cl_page_state_set_trust(page, state);
 595         }
 596         EXIT;
 597 }
 598
 599 static void cl_page_state_set(const struct lu_env *env,
 600                               struct cl_page *page, enum cl_page_state state)
 601 {
 602         cl_page_state_set0(env, page, state);
 603 }
 604
 605 /**
 606  * Acquires an additional reference to a page.
 607  *
 608  * This can be called only by caller already possessing a reference to \a
 609  * page.
 610  *
 611  * \see cl_object_get(), cl_lock_get().
 612  */
 613 void cl_page_get(struct cl_page *page)
 614 {
 615         ENTRY;
 616         cl_page_get_trust(page);
 617         EXIT;
 618 }
 619 EXPORT_SYMBOL(cl_page_get);
 620
 621 /**
 622  * Releases a reference to a page.
 623  *
 624  * When last reference is released, page is returned to the cache, unless it
 625  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 626  * destroyed.
 627  *
 628  * \see cl_object_put(), cl_lock_put().
 629  */
 630 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 631 {
 632         PASSERT(env, page, cfs_atomic_read(&page->cp_ref) > !!page->cp_parent);
 633
 634         ENTRY;
 635         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 636                        cfs_atomic_read(&page->cp_ref));
 637
 638         if (cfs_atomic_dec_and_test(&page->cp_ref)) {
 639                 LASSERT(page->cp_state == CPS_FREEING);
 640
 641                 LASSERT(cfs_atomic_read(&page->cp_ref) == 0);
 642                 PASSERT(env, page, page->cp_owner == NULL);
 643                 PASSERT(env, page, cfs_list_empty(&page->cp_batch));
 644                 /*
 645                  * Page is no longer reachable by other threads. Tear
 646                  * it down.
 647                  */
 648                 cl_page_free(env, page);
 649         }
 650
 651         EXIT;
 652 }
 653 EXPORT_SYMBOL(cl_page_put);
 654
 655 /**
 656  * Returns a VM page associated with a given cl_page.
 657  */
 658 cfs_page_t *cl_page_vmpage(const struct lu_env *env, struct cl_page *page)
 659 {
 660         const struct cl_page_slice *slice;
 661
 662         /*
 663          * Find uppermost layer with ->cpo_vmpage() method, and return its
 664          * result.
 665          */
 666         page = cl_page_top(page);
 667         do {
 668                 cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 669                         if (slice->cpl_ops->cpo_vmpage != NULL)
 670                                 RETURN(slice->cpl_ops->cpo_vmpage(env, slice));
 671                 }
 672                 page = page->cp_child;
 673         } while (page != NULL);
 674         LBUG(); /* ->cpo_vmpage() has to be defined somewhere in the stack */
 675 }
 676 EXPORT_SYMBOL(cl_page_vmpage);
 677
 678 /**
 679  * Returns a cl_page associated with a VM page, and given cl_object.
 680  */
 681 struct cl_page *cl_vmpage_page(cfs_page_t *vmpage, struct cl_object *obj)
 682 {
 683         struct cl_page *top;
 684         struct cl_page *page;
 685
 686         ENTRY;
 687         KLASSERT(PageLocked(vmpage));
 688
 689         /*
 690          * NOTE: absence of races and liveness of data are guaranteed by page
 691          *       lock on a "vmpage". That works because object destruction has
 692          *       bottom-to-top pass.
 693          */
 694
 695         /*
 696          * This loop assumes that ->private points to the top-most page. This
 697          * can be rectified easily.
 698          */
 699         top = (struct cl_page *)vmpage->private;
 700         if (top == NULL)
 701                 RETURN(NULL);
 702
 703         for (page = top; page != NULL; page = page->cp_child) {
 704                 if (cl_object_same(page->cp_obj, obj)) {
 705                         cl_page_get_trust(page);
 706                         break;
 707                 }
 708         }
 709         LASSERT(ergo(page, page->cp_type == CPT_CACHEABLE));
 710         RETURN(page);
 711 }
 712 EXPORT_SYMBOL(cl_vmpage_page);
 713
 714 /**
 715  * Returns the top-page for a given page.
 716  *
 717  * \see cl_object_top(), cl_io_top()
 718  */
 719 struct cl_page *cl_page_top(struct cl_page *page)
 720 {
 721         return cl_page_top_trusted(page);
 722 }
 723 EXPORT_SYMBOL(cl_page_top);
 724
 725 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 726                                        const struct lu_device_type *dtype)
 727 {
 728         return cl_page_at_trusted(page, dtype);
 729 }
 730 EXPORT_SYMBOL(cl_page_at);
 731
 732 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 733
 734 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)                   \
 735 ({                                                                      \
 736         const struct lu_env        *__env  = (_env);                    \
 737         struct cl_page             *__page = (_page);                   \
 738         const struct cl_page_slice *__scan;                             \
 739         int                         __result;                           \
 740         ptrdiff_t                   __op   = (_op);                     \
 741         int                       (*__method)_proto;                    \
 742                                                                         \
 743         __result = 0;                                                   \
 744         __page = cl_page_top(__page);                                   \
 745         do {                                                            \
 746                 cfs_list_for_each_entry(__scan, &__page->cp_layers,     \
 747                                         cpl_linkage) {                  \
 748                         __method = *(void **)((char *)__scan->cpl_ops + \
 749                                               __op);                    \
 750                         if (__method != NULL) {                         \
 751                                 __result = (*__method)(__env, __scan,   \
 752                                                        ## __VA_ARGS__); \
 753                                 if (__result != 0)                      \
 754                                         break;                          \
 755                         }                                               \
 756                 }                                                       \
 757                 __page = __page->cp_child;                              \
 758         } while (__page != NULL && __result == 0);                      \
 759         if (__result > 0)                                               \
 760                 __result = 0;                                           \
 761         __result;                                                       \
 762 })
 763
 764 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 765 do {                                                                    \
 766         const struct lu_env        *__env  = (_env);                    \
 767         struct cl_page             *__page = (_page);                   \
 768         const struct cl_page_slice *__scan;                             \
 769         ptrdiff_t                   __op   = (_op);                     \
 770         void                      (*__method)_proto;                    \
 771                                                                         \
 772         __page = cl_page_top(__page);                                   \
 773         do {                                                            \
 774                 cfs_list_for_each_entry(__scan, &__page->cp_layers,     \
 775                                         cpl_linkage) {                  \
 776                         __method = *(void **)((char *)__scan->cpl_ops + \
 777                                               __op);                    \
 778                         if (__method != NULL)                           \
 779                                 (*__method)(__env, __scan,              \
 780                                             ## __VA_ARGS__);            \
 781                 }                                                       \
 782                 __page = __page->cp_child;                              \
 783         } while (__page != NULL);                                       \
 784 } while (0)
 785
 786 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)               \
 787 do {                                                                        \
 788         const struct lu_env        *__env  = (_env);                        \
 789         struct cl_page             *__page = (_page);                       \
 790         const struct cl_page_slice *__scan;                                 \
 791         ptrdiff_t                   __op   = (_op);                         \
 792         void                      (*__method)_proto;                        \
 793                                                                             \
 794         /* get to the bottom page. */                                       \
 795         while (__page->cp_child != NULL)                                    \
 796                 __page = __page->cp_child;                                  \
 797         do {                                                                \
 798                 cfs_list_for_each_entry_reverse(__scan, &__page->cp_layers, \
 799                                                 cpl_linkage) {              \
 800                         __method = *(void **)((char *)__scan->cpl_ops +     \
 801                                               __op);                        \
 802                         if (__method != NULL)                               \
 803                                 (*__method)(__env, __scan,                  \
 804                                             ## __VA_ARGS__);                \
 805                 }                                                           \
 806                 __page = __page->cp_parent;                                 \
 807         } while (__page != NULL);                                           \
 808 } while (0)
 809
 810 static int cl_page_invoke(const struct lu_env *env,
 811                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 812
 813 {
 814         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 815         ENTRY;
 816         RETURN(CL_PAGE_INVOKE(env, page, op,
 817                               (const struct lu_env *,
 818                                const struct cl_page_slice *, struct cl_io *),
 819                               io));
 820 }
 821
 822 static void cl_page_invoid(const struct lu_env *env,
 823                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 824
 825 {
 826         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 827         ENTRY;
 828         CL_PAGE_INVOID(env, page, op,
 829                        (const struct lu_env *,
 830                         const struct cl_page_slice *, struct cl_io *), io);
 831         EXIT;
 832 }
 833
 834 static void cl_page_owner_clear(struct cl_page *page)
 835 {
 836         ENTRY;
 837         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 838                 if (page->cp_owner != NULL) {
 839                         LASSERT(page->cp_owner->ci_owned_nr > 0);
 840                         page->cp_owner->ci_owned_nr--;
 841                         page->cp_owner = NULL;
 842                         page->cp_task = NULL;
 843                 }
 844         }
 845         EXIT;
 846 }
 847
 848 static void cl_page_owner_set(struct cl_page *page)
 849 {
 850         ENTRY;
 851         for (page = cl_page_top(page); page != NULL; page = page->cp_child) {
 852                 LASSERT(page->cp_owner != NULL);
 853                 page->cp_owner->ci_owned_nr++;
 854         }
 855         EXIT;
 856 }
 857
 858 void cl_page_disown0(const struct lu_env *env,
 859                      struct cl_io *io, struct cl_page *pg)
 860 {
 861         enum cl_page_state state;
 862
 863         ENTRY;
 864         state = pg->cp_state;
 865         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 866         PINVRNT(env, pg, cl_page_invariant(pg));
 867         cl_page_owner_clear(pg);
 868
 869         if (state == CPS_OWNED)
 870                 cl_page_state_set(env, pg, CPS_CACHED);
 871         /*
 872          * Completion call-backs are executed in the bottom-up order, so that
 873          * uppermost layer (llite), responsible for VFS/VM interaction runs
 874          * last and can release locks safely.
 875          */
 876         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 877                                (const struct lu_env *,
 878                                 const struct cl_page_slice *, struct cl_io *),
 879                                io);
 880         EXIT;
 881 }
 882
 883 /**
 884  * returns true, iff page is owned by the given io.
 885  */
 886 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 887 {
 888         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 889         ENTRY;
 890         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == io);
 891 }
 892 EXPORT_SYMBOL(cl_page_is_owned);
 893
 894 /**
 895  * Try to own a page by IO.
 896  *
 897  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 898  * into cl_page_state::CPS_OWNED state.
 899  *
 900  * \pre  !cl_page_is_owned(pg, io)
 901  * \post result == 0 iff cl_page_is_owned(pg, io)
 902  *
 903  * \retval 0   success
 904  *
 905  * \retval -ve failure, e.g., page was destroyed (and landed in
 906  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 907  *             or, page was owned by another thread, or in IO.
 908  *
 909  * \see cl_page_disown()
 910  * \see cl_page_operations::cpo_own()
 911  * \see cl_page_own_try()
 912  * \see cl_page_own
 913  */
 914 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 915                         struct cl_page *pg, int nonblock)
 916 {
 917         int result;
 918
 919         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 920
 921         ENTRY;
 922         pg = cl_page_top(pg);
 923         io = cl_io_top(io);
 924
 925         if (pg->cp_state == CPS_FREEING) {
 926                 result = -ENOENT;
 927         } else {
 928                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 929                                         (const struct lu_env *,
 930                                          const struct cl_page_slice *,
 931                                          struct cl_io *, int),
 932                                         io, nonblock);
 933                 if (result == 0) {
 934                         PASSERT(env, pg, pg->cp_owner == NULL);
 935                         PASSERT(env, pg, pg->cp_req == NULL);
 936                         pg->cp_owner = io;
 937                         pg->cp_task  = current;
 938                         cl_page_owner_set(pg);
 939                         if (pg->cp_state != CPS_FREEING) {
 940                                 cl_page_state_set(env, pg, CPS_OWNED);
 941                         } else {
 942                                 cl_page_disown0(env, io, pg);
 943                                 result = -ENOENT;
 944                         }
 945                 }
 946         }
 947         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 948         RETURN(result);
 949 }
 950
 951 /**
 952  * Own a page, might be blocked.
 953  *
 954  * \see cl_page_own0()
 955  */
 956 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 957 {
 958         return cl_page_own0(env, io, pg, 0);
 959 }
 960 EXPORT_SYMBOL(cl_page_own);
 961
 962 /**
 963  * Nonblock version of cl_page_own().
 964  *
 965  * \see cl_page_own0()
 966  */
 967 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
 968                     struct cl_page *pg)
 969 {
 970         return cl_page_own0(env, io, pg, 1);
 971 }
 972 EXPORT_SYMBOL(cl_page_own_try);
 973
 974
 975 /**
 976  * Assume page ownership.
 977  *
 978  * Called when page is already locked by the hosting VM.
 979  *
 980  * \pre !cl_page_is_owned(pg, io)
 981  * \post cl_page_is_owned(pg, io)
 982  *
 983  * \see cl_page_operations::cpo_assume()
 984  */
 985 void cl_page_assume(const struct lu_env *env,
 986                     struct cl_io *io, struct cl_page *pg)
 987 {
 988         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
 989
 990         ENTRY;
 991         pg = cl_page_top(pg);
 992         io = cl_io_top(io);
 993
 994         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
 995         PASSERT(env, pg, pg->cp_owner == NULL);
 996         pg->cp_owner = io;
 997         pg->cp_task = current;
 998         cl_page_owner_set(pg);
 999         cl_page_state_set(env, pg, CPS_OWNED);
1000         EXIT;
1001 }
1002 EXPORT_SYMBOL(cl_page_assume);
1003
1004 /**
1005  * Releases page ownership without unlocking the page.
1006  *
1007  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
1008  * underlying VM page (as VM is supposed to do this itself).
1009  *
1010  * \pre   cl_page_is_owned(pg, io)
1011  * \post !cl_page_is_owned(pg, io)
1012  *
1013  * \see cl_page_assume()
1014  */
1015 void cl_page_unassume(const struct lu_env *env,
1016                       struct cl_io *io, struct cl_page *pg)
1017 {
1018         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1019         PINVRNT(env, pg, cl_page_invariant(pg));
1020
1021         ENTRY;
1022         pg = cl_page_top(pg);
1023         io = cl_io_top(io);
1024         cl_page_owner_clear(pg);
1025         cl_page_state_set(env, pg, CPS_CACHED);
1026         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
1027                                (const struct lu_env *,
1028                                 const struct cl_page_slice *, struct cl_io *),
1029                                io);
1030         EXIT;
1031 }
1032 EXPORT_SYMBOL(cl_page_unassume);
1033
1034 /**
1035  * Releases page ownership.
1036  *
1037  * Moves page into cl_page_state::CPS_CACHED.
1038  *
1039  * \pre   cl_page_is_owned(pg, io)
1040  * \post !cl_page_is_owned(pg, io)
1041  *
1042  * \see cl_page_own()
1043  * \see cl_page_operations::cpo_disown()
1044  */
1045 void cl_page_disown(const struct lu_env *env,
1046                     struct cl_io *io, struct cl_page *pg)
1047 {
1048         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1049
1050         ENTRY;
1051         pg = cl_page_top(pg);
1052         io = cl_io_top(io);
1053         cl_page_disown0(env, io, pg);
1054         EXIT;
1055 }
1056 EXPORT_SYMBOL(cl_page_disown);
1057
1058 /**
1059  * Called when page is to be removed from the object, e.g., as a result of
1060  * truncate.
1061  *
1062  * Calls cl_page_operations::cpo_discard() top-to-bottom.
1063  *
1064  * \pre cl_page_is_owned(pg, io)
1065  *
1066  * \see cl_page_operations::cpo_discard()
1067  */
1068 void cl_page_discard(const struct lu_env *env,
1069                      struct cl_io *io, struct cl_page *pg)
1070 {
1071         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1072         PINVRNT(env, pg, cl_page_invariant(pg));
1073
1074         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
1075 }
1076 EXPORT_SYMBOL(cl_page_discard);
1077
1078 /**
1079  * Version of cl_page_delete() that can be called for not fully constructed
1080  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
1081  * path. Doesn't check page invariant.
1082  */
1083 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
1084                             int radix)
1085 {
1086         struct cl_page *tmp = pg;
1087         ENTRY;
1088
1089         PASSERT(env, pg, pg == cl_page_top(pg));
1090         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
1091
1092         /*
1093          * Severe all ways to obtain new pointers to @pg.
1094          */
1095         cl_page_owner_clear(pg);
1096
1097         /*
1098          * unexport the page firstly before freeing it so that
1099          * the page content is considered to be invalid.
1100          * We have to do this because a CPS_FREEING cl_page may
1101          * be NOT under the protection of a cl_lock.
1102          * Afterwards, if this page is found by other threads, then this
1103          * page will be forced to reread.
1104          */
1105         cl_page_export(env, pg, 0);
1106         cl_page_state_set0(env, pg, CPS_FREEING);
1107
1108         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete),
1109                        (const struct lu_env *, const struct cl_page_slice *));
1110
1111         if (tmp->cp_type == CPT_CACHEABLE) {
1112                 if (!radix)
1113                         /* !radix means that @pg is not yet in the radix tree,
1114                          * skip removing it.
1115                          */
1116                         tmp = pg->cp_child;
1117                 for (; tmp != NULL; tmp = tmp->cp_child) {
1118                         void                    *value;
1119                         struct cl_object_header *hdr;
1120
1121                         hdr = cl_object_header(tmp->cp_obj);
1122                         spin_lock(&hdr->coh_page_guard);
1123                         value = radix_tree_delete(&hdr->coh_tree,
1124                                                   tmp->cp_index);
1125                         PASSERT(env, tmp, value == tmp);
1126                         PASSERT(env, tmp, hdr->coh_pages > 0);
1127                         hdr->coh_pages--;
1128                         spin_unlock(&hdr->coh_page_guard);
1129                         cl_page_put(env, tmp);
1130                 }
1131         }
1132
1133         EXIT;
1134 }
1135
1136 /**
1137  * Called when a decision is made to throw page out of memory.
1138  *
1139  * Notifies all layers about page destruction by calling
1140  * cl_page_operations::cpo_delete() method top-to-bottom.
1141  *
1142  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
1143  * where transition to this state happens).
1144  *
1145  * Eliminates all venues through which new references to the page can be
1146  * obtained:
1147  *
1148  *     - removes page from the radix trees,
1149  *
1150  *     - breaks linkage from VM page to cl_page.
1151  *
1152  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
1153  * drain after some time, at which point page will be recycled.
1154  *
1155  * \pre  pg == cl_page_top(pg)
1156  * \pre  VM page is locked
1157  * \post pg->cp_state == CPS_FREEING
1158  *
1159  * \see cl_page_operations::cpo_delete()
1160  */
1161 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
1162 {
1163         PINVRNT(env, pg, cl_page_invariant(pg));
1164         ENTRY;
1165         cl_page_delete0(env, pg, 1);
1166         EXIT;
1167 }
1168 EXPORT_SYMBOL(cl_page_delete);
1169
1170 /**
1171  * Unmaps page from user virtual memory.
1172  *
1173  * Calls cl_page_operations::cpo_unmap() through all layers top-to-bottom. The
1174  * layer responsible for VM interaction has to unmap page from user space
1175  * virtual memory.
1176  *
1177  * \see cl_page_operations::cpo_unmap()
1178  */
1179 int cl_page_unmap(const struct lu_env *env,
1180                   struct cl_io *io, struct cl_page *pg)
1181 {
1182         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1183         PINVRNT(env, pg, cl_page_invariant(pg));
1184
1185         return cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_unmap));
1186 }
1187 EXPORT_SYMBOL(cl_page_unmap);
1188
1189 /**
1190  * Marks page up-to-date.
1191  *
1192  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
1193  * layer responsible for VM interaction has to mark/clear page as up-to-date
1194  * by the \a uptodate argument.
1195  *
1196  * \see cl_page_operations::cpo_export()
1197  */
1198 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
1199 {
1200         PINVRNT(env, pg, cl_page_invariant(pg));
1201         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
1202                        (const struct lu_env *,
1203                         const struct cl_page_slice *, int), uptodate);
1204 }
1205 EXPORT_SYMBOL(cl_page_export);
1206
1207 /**
1208  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
1209  * thread.
1210  */
1211 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
1212 {
1213         int result;
1214         const struct cl_page_slice *slice;
1215
1216         ENTRY;
1217         pg = cl_page_top_trusted((struct cl_page *)pg);
1218         slice = container_of(pg->cp_layers.next,
1219                              const struct cl_page_slice, cpl_linkage);
1220         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
1221         /*
1222          * Call ->cpo_is_vmlocked() directly instead of going through
1223          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
1224          * cl_page_invariant().
1225          */
1226         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
1227         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
1228         RETURN(result == -EBUSY);
1229 }
1230 EXPORT_SYMBOL(cl_page_is_vmlocked);
1231
1232 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
1233 {
1234         ENTRY;
1235         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
1236 }
1237
1238 static void cl_page_io_start(const struct lu_env *env,
1239                              struct cl_page *pg, enum cl_req_type crt)
1240 {
1241         /*
1242          * Page is queued for IO, change its state.
1243          */
1244         ENTRY;
1245         cl_page_owner_clear(pg);
1246         cl_page_state_set(env, pg, cl_req_type_state(crt));
1247         EXIT;
1248 }
1249
1250 /**
1251  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
1252  * called top-to-bottom. Every layer either agrees to submit this page (by
1253  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
1254  * handling interactions with the VM also has to inform VM that page is under
1255  * transfer now.
1256  */
1257 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
1258                  struct cl_page *pg, enum cl_req_type crt)
1259 {
1260         int result;
1261
1262         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1263         PINVRNT(env, pg, cl_page_invariant(pg));
1264         PINVRNT(env, pg, crt < CRT_NR);
1265
1266         /*
1267          * XXX this has to be called bottom-to-top, so that llite can set up
1268          * PG_writeback without risking other layers deciding to skip this
1269          * page.
1270          */
1271         if (crt >= CRT_NR)
1272                 return -EINVAL;
1273         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
1274         if (result == 0)
1275                 cl_page_io_start(env, pg, crt);
1276
1277         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
1278                       equi(result == 0,
1279                            PageWriteback(cl_page_vmpage(env, pg)))));
1280         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1281         return result;
1282 }
1283 EXPORT_SYMBOL(cl_page_prep);
1284
1285 /**
1286  * Notify layers about transfer completion.
1287  *
1288  * Invoked by transfer sub-system (which is a part of osc) to notify layers
1289  * that a transfer, of which this page is a part of has completed.
1290  *
1291  * Completion call-backs are executed in the bottom-up order, so that
1292  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
1293  * and can release locks safely.
1294  *
1295  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1296  * \post pg->cp_state == CPS_CACHED
1297  *
1298  * \see cl_page_operations::cpo_completion()
1299  */
1300 void cl_page_completion(const struct lu_env *env,
1301                         struct cl_page *pg, enum cl_req_type crt, int ioret)
1302 {
1303         struct cl_sync_io *anchor = pg->cp_sync_io;
1304
1305         PASSERT(env, pg, crt < CRT_NR);
1306         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
1307         PASSERT(env, pg, pg->cp_req == NULL);
1308         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
1309
1310         ENTRY;
1311         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
1312         if (crt == CRT_READ && ioret == 0) {
1313                 PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED));
1314                 pg->cp_flags |= CPF_READ_COMPLETED;
1315         }
1316
1317         cl_page_state_set(env, pg, CPS_CACHED);
1318         if (crt >= CRT_NR)
1319                 return;
1320         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
1321                                (const struct lu_env *,
1322                                 const struct cl_page_slice *, int), ioret);
1323         if (anchor) {
1324                 LASSERT(cl_page_is_vmlocked(env, pg));
1325                 LASSERT(pg->cp_sync_io == anchor);
1326                 pg->cp_sync_io = NULL;
1327         }
1328         /*
1329          * As page->cp_obj is pinned by a reference from page->cp_req, it is
1330          * safe to call cl_page_put() without risking object destruction in a
1331          * non-blocking context.
1332          */
1333         cl_page_put(env, pg);
1334
1335         if (anchor)
1336                 cl_sync_io_note(anchor, ioret);
1337
1338         EXIT;
1339 }
1340 EXPORT_SYMBOL(cl_page_completion);
1341
1342 /**
1343  * Notify layers that transfer formation engine decided to yank this page from
1344  * the cache and to make it a part of a transfer.
1345  *
1346  * \pre  pg->cp_state == CPS_CACHED
1347  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
1348  *
1349  * \see cl_page_operations::cpo_make_ready()
1350  */
1351 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
1352                        enum cl_req_type crt)
1353 {
1354         int result;
1355
1356         PINVRNT(env, pg, crt < CRT_NR);
1357
1358         ENTRY;
1359         if (crt >= CRT_NR)
1360                 RETURN(-EINVAL);
1361         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
1362                                 (const struct lu_env *,
1363                                  const struct cl_page_slice *));
1364         if (result == 0) {
1365                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
1366                 cl_page_io_start(env, pg, crt);
1367         }
1368         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1369         RETURN(result);
1370 }
1371 EXPORT_SYMBOL(cl_page_make_ready);
1372
1373 /**
1374  * Notify layers that high level io decided to place this page into a cache
1375  * for future transfer.
1376  *
1377  * The layer implementing transfer engine (osc) has to register this page in
1378  * its queues.
1379  *
1380  * \pre  cl_page_is_owned(pg, io)
1381  * \post cl_page_is_owned(pg, io)
1382  *
1383  * \see cl_page_operations::cpo_cache_add()
1384  */
1385 int cl_page_cache_add(const struct lu_env *env, struct cl_io *io,
1386                       struct cl_page *pg, enum cl_req_type crt)
1387 {
1388         const struct cl_page_slice *scan;
1389         int result = 0;
1390
1391         PINVRNT(env, pg, crt < CRT_NR);
1392         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1393         PINVRNT(env, pg, cl_page_invariant(pg));
1394
1395         ENTRY;
1396
1397         if (crt >= CRT_NR)
1398                 RETURN(-EINVAL);
1399
1400         cfs_list_for_each_entry(scan, &pg->cp_layers, cpl_linkage) {
1401                 if (scan->cpl_ops->io[crt].cpo_cache_add == NULL)
1402                         continue;
1403
1404                 result = scan->cpl_ops->io[crt].cpo_cache_add(env, scan, io);
1405                 if (result != 0)
1406                         break;
1407         }
1408         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1409         RETURN(result);
1410 }
1411 EXPORT_SYMBOL(cl_page_cache_add);
1412
1413 /**
1414  * Called if a pge is being written back by kernel's intention.
1415  *
1416  * \pre  cl_page_is_owned(pg, io)
1417  * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT)
1418  *
1419  * \see cl_page_operations::cpo_flush()
1420  */
1421 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
1422                   struct cl_page *pg)
1423 {
1424         int result;
1425
1426         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1427         PINVRNT(env, pg, cl_page_invariant(pg));
1428
1429         ENTRY;
1430
1431         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_flush));
1432
1433         CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result);
1434         RETURN(result);
1435 }
1436 EXPORT_SYMBOL(cl_page_flush);
1437
1438 /**
1439  * Checks whether page is protected by any extent lock is at least required
1440  * mode.
1441  *
1442  * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1443  * \see cl_page_operations::cpo_is_under_lock()
1444  */
1445 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1446                           struct cl_page *page)
1447 {
1448         int rc;
1449
1450         PINVRNT(env, page, cl_page_invariant(page));
1451
1452         ENTRY;
1453         rc = CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1454                             (const struct lu_env *,
1455                              const struct cl_page_slice *, struct cl_io *),
1456                             io);
1457         PASSERT(env, page, rc != 0);
1458         RETURN(rc);
1459 }
1460 EXPORT_SYMBOL(cl_page_is_under_lock);
1461
1462 static int page_prune_cb(const struct lu_env *env, struct cl_io *io,
1463                          struct cl_page *page, void *cbdata)
1464 {
1465         cl_page_own(env, io, page);
1466         cl_page_unmap(env, io, page);
1467         cl_page_discard(env, io, page);
1468         cl_page_disown(env, io, page);
1469         return CLP_GANG_OKAY;
1470 }
1471
1472 /**
1473  * Purges all cached pages belonging to the object \a obj.
1474  */
1475 int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj)
1476 {
1477         struct cl_thread_info   *info;
1478         struct cl_object        *obj = cl_object_top(clobj);
1479         struct cl_io            *io;
1480         int                      result;
1481
1482         ENTRY;
1483         info  = cl_env_info(env);
1484         io    = &info->clt_io;
1485
1486         /*
1487          * initialize the io. This is ugly since we never do IO in this
1488          * function, we just make cl_page_list functions happy. -jay
1489          */
1490         io->ci_obj = obj;
1491         io->ci_ignore_layout = 1;
1492         result = cl_io_init(env, io, CIT_MISC, obj);
1493         if (result != 0) {
1494                 cl_io_fini(env, io);
1495                 RETURN(io->ci_result);
1496         }
1497
1498         do {
1499                 result = cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF,
1500                                              page_prune_cb, NULL);
1501                 if (result == CLP_GANG_RESCHED)
1502                         cfs_cond_resched();
1503         } while (result != CLP_GANG_OKAY);
1504
1505         cl_io_fini(env, io);
1506         RETURN(result);
1507 }
1508 EXPORT_SYMBOL(cl_pages_prune);
1509
1510 /**
1511  * Tells transfer engine that only part of a page is to be transmitted.
1512  *
1513  * \see cl_page_operations::cpo_clip()
1514  */
1515 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1516                   int from, int to)
1517 {
1518         PINVRNT(env, pg, cl_page_invariant(pg));
1519
1520         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1521         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1522                        (const struct lu_env *,
1523                         const struct cl_page_slice *,int, int),
1524                        from, to);
1525 }
1526 EXPORT_SYMBOL(cl_page_clip);
1527
1528 /**
1529  * Prints human readable representation of \a pg to the \a f.
1530  */
1531 void cl_page_header_print(const struct lu_env *env, void *cookie,
1532                           lu_printer_t printer, const struct cl_page *pg)
1533 {
1534         (*printer)(env, cookie,
1535                    "page@%p[%d %p:%lu ^%p_%p %d %d %d %p %p %#x]\n",
1536                    pg, cfs_atomic_read(&pg->cp_ref), pg->cp_obj,
1537                    pg->cp_index, pg->cp_parent, pg->cp_child,
1538                    pg->cp_state, pg->cp_error, pg->cp_type,
1539                    pg->cp_owner, pg->cp_req, pg->cp_flags);
1540 }
1541 EXPORT_SYMBOL(cl_page_header_print);
1542
1543 /**
1544  * Prints human readable representation of \a pg to the \a f.
1545  */
1546 void cl_page_print(const struct lu_env *env, void *cookie,
1547                    lu_printer_t printer, const struct cl_page *pg)
1548 {
1549         struct cl_page *scan;
1550
1551         for (scan = cl_page_top((struct cl_page *)pg);
1552              scan != NULL; scan = scan->cp_child)
1553                 cl_page_header_print(env, cookie, printer, scan);
1554         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1555                        (const struct lu_env *env,
1556                         const struct cl_page_slice *slice,
1557                         void *cookie, lu_printer_t p), cookie, printer);
1558         (*printer)(env, cookie, "end page@%p\n", pg);
1559 }
1560 EXPORT_SYMBOL(cl_page_print);
1561
1562 /**
1563  * Cancel a page which is still in a transfer.
1564  */
1565 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1566 {
1567         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1568                               (const struct lu_env *,
1569                                const struct cl_page_slice *));
1570 }
1571 EXPORT_SYMBOL(cl_page_cancel);
1572
1573 /**
1574  * Converts a byte offset within object \a obj into a page index.
1575  */
1576 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1577 {
1578         /*
1579          * XXX for now.
1580          */
1581         return (loff_t)idx << CFS_PAGE_SHIFT;
1582 }
1583 EXPORT_SYMBOL(cl_offset);
1584
1585 /**
1586  * Converts a page index into a byte offset within object \a obj.
1587  */
1588 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1589 {
1590         /*
1591          * XXX for now.
1592          */
1593         return offset >> CFS_PAGE_SHIFT;
1594 }
1595 EXPORT_SYMBOL(cl_index);
1596
1597 int cl_page_size(const struct cl_object *obj)
1598 {
1599         return 1 << CFS_PAGE_SHIFT;
1600 }
1601 EXPORT_SYMBOL(cl_page_size);
1602
1603 /**
1604  * Adds page slice to the compound page.
1605  *
1606  * This is called by cl_object_operations::coo_page_init() methods to add a
1607  * per-layer state to the page. New state is added at the end of
1608  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1609  *
1610  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1611  */
1612 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1613                        struct cl_object *obj,
1614                        const struct cl_page_operations *ops)
1615 {
1616         ENTRY;
1617         cfs_list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1618         slice->cpl_obj  = obj;
1619         slice->cpl_ops  = ops;
1620         slice->cpl_page = page;
1621         EXIT;
1622 }
1623 EXPORT_SYMBOL(cl_page_slice_add);
1624
1625 int  cl_page_init(void)
1626 {
1627         return 0;
1628 }
1629
1630 void cl_page_fini(void)
1631 {
1632 }