lustre/obdclass/cl_page.c

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