lustre/obdclass/cl_page.c

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