lustre/obdclass/cl_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Client Lustre Page.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_CLASS
  43
  44 #include <libcfs/libcfs.h>
  45 #include <obd_class.h>
  46 #include <obd_support.h>
  47 #include <libcfs/list.h>
  48
  49 #include <cl_object.h>
  50 #include "cl_internal.h"
  51
  52 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg);
  53
  54 #ifdef LIBCFS_DEBUG
  55 # define PASSERT(env, page, expr)                                       \
  56   do {                                                                    \
  57           if (unlikely(!(expr))) {                                      \
  58                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  59                   LASSERT(0);                                           \
  60           }                                                             \
  61   } while (0)
  62 #else /* !LIBCFS_DEBUG */
  63 # define PASSERT(env, page, exp) \
  64         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  65 #endif /* !LIBCFS_DEBUG */
  66
  67 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
  68 # define PINVRNT(env, page, expr)                                       \
  69   do {                                                                    \
  70           if (unlikely(!(expr))) {                                      \
  71                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  72                   LINVRNT(0);                                           \
  73           }                                                             \
  74   } while (0)
  75 #else /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
  76 # define PINVRNT(env, page, exp) \
  77          ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  78 #endif /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
  79
  80 /* Disable page statistic by default due to huge performance penalty. */
  81 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
  82 #define CS_PAGE_INC(o, item) \
  83         cfs_atomic_inc(&cl_object_site(o)->cs_pages.cs_stats[CS_##item])
  84 #define CS_PAGE_DEC(o, item) \
  85         cfs_atomic_dec(&cl_object_site(o)->cs_pages.cs_stats[CS_##item])
  86 #define CS_PAGESTATE_INC(o, state) \
  87         cfs_atomic_inc(&cl_object_site(o)->cs_pages_state[state])
  88 #define CS_PAGESTATE_DEC(o, state) \
  89         cfs_atomic_dec(&cl_object_site(o)->cs_pages_state[state])
  90 #else
  91 #define CS_PAGE_INC(o, item)
  92 #define CS_PAGE_DEC(o, item)
  93 #define CS_PAGESTATE_INC(o, state)
  94 #define CS_PAGESTATE_DEC(o, state)
  95 #endif
  96
  97 /**
  98  * Internal version of cl_page_get().
  99  *
 100  * This function can be used to obtain initial reference to previously
 101  * unreferenced cached object. It can be called only if concurrent page
 102  * reclamation is somehow prevented, e.g., by keeping a lock on a VM page,
 103  * associated with \a page.
 104  *
 105  * Use with care! Not exported.
 106  */
 107 static void cl_page_get_trust(struct cl_page *page)
 108 {
 109         LASSERT(cfs_atomic_read(&page->cp_ref) > 0);
 110         cfs_atomic_inc(&page->cp_ref);
 111 }
 112
 113 /**
 114  * Returns a slice within a page, corresponding to the given layer in the
 115  * device stack.
 116  *
 117  * \see cl_lock_at()
 118  */
 119 static const struct cl_page_slice *
 120 cl_page_at_trusted(const struct cl_page *page,
 121                    const struct lu_device_type *dtype)
 122 {
 123         const struct cl_page_slice *slice;
 124         ENTRY;
 125
 126         cfs_list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 127                 if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
 128                         RETURN(slice);
 129         }
 130         RETURN(NULL);
 131 }
 132
 133 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
 134 {
 135         struct cl_object *obj  = page->cp_obj;
 136         int pagesize = cl_object_header(obj)->coh_page_bufsize;
 137
 138         PASSERT(env, page, cfs_list_empty(&page->cp_batch));
 139         PASSERT(env, page, page->cp_owner == NULL);
 140         PASSERT(env, page, page->cp_req == NULL);
 141         PASSERT(env, page, page->cp_state == CPS_FREEING);
 142
 143         ENTRY;
 144         while (!cfs_list_empty(&page->cp_layers)) {
 145                 struct cl_page_slice *slice;
 146
 147                 slice = cfs_list_entry(page->cp_layers.next,
 148                                        struct cl_page_slice, cpl_linkage);
 149                 cfs_list_del_init(page->cp_layers.next);
 150                 if (unlikely(slice->cpl_ops->cpo_fini != NULL))
 151                         slice->cpl_ops->cpo_fini(env, slice);
 152         }
 153         CS_PAGE_DEC(obj, total);
 154         CS_PAGESTATE_DEC(obj, page->cp_state);
 155         lu_object_ref_del_at(&obj->co_lu, &page->cp_obj_ref, "cl_page", page);
 156         cl_object_put(env, obj);
 157         lu_ref_fini(&page->cp_reference);
 158         OBD_FREE(page, pagesize);
 159         EXIT;
 160 }
 161
 162 /**
 163  * Helper function updating page state. This is the only place in the code
 164  * where cl_page::cp_state field is mutated.
 165  */
 166 static inline void cl_page_state_set_trust(struct cl_page *page,
 167                                            enum cl_page_state state)
 168 {
 169         /* bypass const. */
 170         *(enum cl_page_state *)&page->cp_state = state;
 171 }
 172
 173 struct cl_page *cl_page_alloc(const struct lu_env *env,
 174                 struct cl_object *o, pgoff_t ind, struct page *vmpage,
 175                 enum cl_page_type type)
 176 {
 177         struct cl_page          *page;
 178         struct lu_object_header *head;
 179
 180         ENTRY;
 181         OBD_ALLOC_GFP(page, cl_object_header(o)->coh_page_bufsize,
 182                         __GFP_IO);
 183         if (page != NULL) {
 184                 int result = 0;
 185                 cfs_atomic_set(&page->cp_ref, 1);
 186                 page->cp_obj = o;
 187                 cl_object_get(o);
 188                 lu_object_ref_add_at(&o->co_lu, &page->cp_obj_ref, "cl_page",
 189                                      page);
 190                 page->cp_vmpage = vmpage;
 191                 cl_page_state_set_trust(page, CPS_CACHED);
 192                 page->cp_type = type;
 193                 CFS_INIT_LIST_HEAD(&page->cp_layers);
 194                 CFS_INIT_LIST_HEAD(&page->cp_batch);
 195                 CFS_INIT_LIST_HEAD(&page->cp_flight);
 196                 mutex_init(&page->cp_mutex);
 197                 lu_ref_init(&page->cp_reference);
 198                 head = o->co_lu.lo_header;
 199                 cfs_list_for_each_entry(o, &head->loh_layers,
 200                                         co_lu.lo_linkage) {
 201                         if (o->co_ops->coo_page_init != NULL) {
 202                                 result = o->co_ops->coo_page_init(env, o, page,
 203                                                                   ind);
 204                                 if (result != 0) {
 205                                         cl_page_delete0(env, page);
 206                                         cl_page_free(env, page);
 207                                         page = ERR_PTR(result);
 208                                         break;
 209                                 }
 210                         }
 211                 }
 212                 if (result == 0) {
 213                         CS_PAGE_INC(o, total);
 214                         CS_PAGE_INC(o, create);
 215                         CS_PAGESTATE_DEC(o, CPS_CACHED);
 216                 }
 217         } else {
 218                 page = ERR_PTR(-ENOMEM);
 219         }
 220         RETURN(page);
 221 }
 222 EXPORT_SYMBOL(cl_page_alloc);
 223
 224 /**
 225  * Returns a cl_page with index \a idx at the object \a o, and associated with
 226  * the VM page \a vmpage.
 227  *
 228  * This is the main entry point into the cl_page caching interface. First, a
 229  * cache (implemented as a per-object radix tree) is consulted. If page is
 230  * found there, it is returned immediately. Otherwise new page is allocated
 231  * and returned. In any case, additional reference to page is acquired.
 232  *
 233  * \see cl_object_find(), cl_lock_find()
 234  */
 235 struct cl_page *cl_page_find(const struct lu_env *env,
 236                              struct cl_object *o,
 237                              pgoff_t idx, struct page *vmpage,
 238                              enum cl_page_type type)
 239 {
 240         struct cl_page          *page = NULL;
 241         struct cl_object_header *hdr;
 242
 243         LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 244         might_sleep();
 245
 246         ENTRY;
 247
 248         hdr = cl_object_header(o);
 249         CS_PAGE_INC(o, lookup);
 250
 251         CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
 252                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 253         /* fast path. */
 254         if (type == CPT_CACHEABLE) {
 255                 /* vmpage lock is used to protect the child/parent
 256                  * relationship */
 257                 KLASSERT(PageLocked(vmpage));
 258                 /*
 259                  * cl_vmpage_page() can be called here without any locks as
 260                  *
 261                  *     - "vmpage" is locked (which prevents ->private from
 262                  *       concurrent updates), and
 263                  *
 264                  *     - "o" cannot be destroyed while current thread holds a
 265                  *       reference on it.
 266                  */
 267                 page = cl_vmpage_page(vmpage, o);
 268                 if (page != NULL) {
 269                         CS_PAGE_INC(o, hit);
 270                         RETURN(page);
 271                 }
 272         }
 273
 274         /* allocate and initialize cl_page */
 275         page = cl_page_alloc(env, o, idx, vmpage, type);
 276         RETURN(page);
 277 }
 278 EXPORT_SYMBOL(cl_page_find);
 279
 280 static inline int cl_page_invariant(const struct cl_page *pg)
 281 {
 282         /*
 283          * Page invariant is protected by a VM lock.
 284          */
 285         LINVRNT(cl_page_is_vmlocked(NULL, pg));
 286
 287         return cl_page_in_use_noref(pg);
 288 }
 289
 290 static void cl_page_state_set0(const struct lu_env *env,
 291                                struct cl_page *page, enum cl_page_state state)
 292 {
 293         enum cl_page_state old;
 294
 295         /*
 296          * Matrix of allowed state transitions [old][new], for sanity
 297          * checking.
 298          */
 299         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 300                 [CPS_CACHED] = {
 301                         [CPS_CACHED]  = 0,
 302                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 303                         [CPS_PAGEIN]  = 0,
 304                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 305                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 306                 },
 307                 [CPS_OWNED] = {
 308                         [CPS_CACHED]  = 1, /* release to the cache */
 309                         [CPS_OWNED]   = 0,
 310                         [CPS_PAGEIN]  = 1, /* start read immediately */
 311                         [CPS_PAGEOUT] = 1, /* start write immediately */
 312                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 313                 },
 314                 [CPS_PAGEIN] = {
 315                         [CPS_CACHED]  = 1, /* io completion */
 316                         [CPS_OWNED]   = 0,
 317                         [CPS_PAGEIN]  = 0,
 318                         [CPS_PAGEOUT] = 0,
 319                         [CPS_FREEING] = 0,
 320                 },
 321                 [CPS_PAGEOUT] = {
 322                         [CPS_CACHED]  = 1, /* io completion */
 323                         [CPS_OWNED]   = 0,
 324                         [CPS_PAGEIN]  = 0,
 325                         [CPS_PAGEOUT] = 0,
 326                         [CPS_FREEING] = 0,
 327                 },
 328                 [CPS_FREEING] = {
 329                         [CPS_CACHED]  = 0,
 330                         [CPS_OWNED]   = 0,
 331                         [CPS_PAGEIN]  = 0,
 332                         [CPS_PAGEOUT] = 0,
 333                         [CPS_FREEING] = 0,
 334                 }
 335         };
 336
 337         ENTRY;
 338         old = page->cp_state;
 339         PASSERT(env, page, allowed_transitions[old][state]);
 340         CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
 341         PASSERT(env, page, page->cp_state == old);
 342         PASSERT(env, page, equi(state == CPS_OWNED, page->cp_owner != NULL));
 343
 344         CS_PAGESTATE_DEC(page->cp_obj, page->cp_state);
 345         CS_PAGESTATE_INC(page->cp_obj, state);
 346         cl_page_state_set_trust(page, state);
 347         EXIT;
 348 }
 349
 350 static void cl_page_state_set(const struct lu_env *env,
 351                               struct cl_page *page, enum cl_page_state state)
 352 {
 353         cl_page_state_set0(env, page, state);
 354 }
 355
 356 /**
 357  * Acquires an additional reference to a page.
 358  *
 359  * This can be called only by caller already possessing a reference to \a
 360  * page.
 361  *
 362  * \see cl_object_get(), cl_lock_get().
 363  */
 364 void cl_page_get(struct cl_page *page)
 365 {
 366         ENTRY;
 367         cl_page_get_trust(page);
 368         EXIT;
 369 }
 370 EXPORT_SYMBOL(cl_page_get);
 371
 372 /**
 373  * Releases a reference to a page.
 374  *
 375  * When last reference is released, page is returned to the cache, unless it
 376  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 377  * destroyed.
 378  *
 379  * \see cl_object_put(), cl_lock_put().
 380  */
 381 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 382 {
 383         ENTRY;
 384         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 385                        cfs_atomic_read(&page->cp_ref));
 386
 387         if (cfs_atomic_dec_and_test(&page->cp_ref)) {
 388                 LASSERT(page->cp_state == CPS_FREEING);
 389
 390                 LASSERT(cfs_atomic_read(&page->cp_ref) == 0);
 391                 PASSERT(env, page, page->cp_owner == NULL);
 392                 PASSERT(env, page, cfs_list_empty(&page->cp_batch));
 393                 /*
 394                  * Page is no longer reachable by other threads. Tear
 395                  * it down.
 396                  */
 397                 cl_page_free(env, page);
 398         }
 399
 400         EXIT;
 401 }
 402 EXPORT_SYMBOL(cl_page_put);
 403
 404 /**
 405  * Returns a cl_page associated with a VM page, and given cl_object.
 406  */
 407 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj)
 408 {
 409         struct cl_page *page;
 410
 411         ENTRY;
 412         KLASSERT(PageLocked(vmpage));
 413
 414         /*
 415          * NOTE: absence of races and liveness of data are guaranteed by page
 416          *       lock on a "vmpage". That works because object destruction has
 417          *       bottom-to-top pass.
 418          */
 419
 420         page = (struct cl_page *)vmpage->private;
 421         if (page != NULL) {
 422                 cl_page_get_trust(page);
 423                 LASSERT(page->cp_type == CPT_CACHEABLE);
 424         }
 425         RETURN(page);
 426 }
 427 EXPORT_SYMBOL(cl_vmpage_page);
 428
 429 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 430                                        const struct lu_device_type *dtype)
 431 {
 432         return cl_page_at_trusted(page, dtype);
 433 }
 434 EXPORT_SYMBOL(cl_page_at);
 435
 436 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 437
 438 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)                   \
 439 ({                                                                      \
 440         const struct lu_env        *__env  = (_env);                    \
 441         struct cl_page             *__page = (_page);                   \
 442         const struct cl_page_slice *__scan;                             \
 443         int                         __result;                           \
 444         ptrdiff_t                   __op   = (_op);                     \
 445         int                       (*__method)_proto;                    \
 446                                                                         \
 447         __result = 0;                                                   \
 448         cfs_list_for_each_entry(__scan, &__page->cp_layers, cpl_linkage) {     \
 449                 __method = *(void **)((char *)__scan->cpl_ops +  __op);        \
 450                 if (__method != NULL) {                                        \
 451                         __result = (*__method)(__env, __scan, ## __VA_ARGS__); \
 452                         if (__result != 0)                              \
 453                                 break;                                  \
 454                 }                                                       \
 455         }                                                               \
 456         if (__result > 0)                                               \
 457                 __result = 0;                                           \
 458         __result;                                                       \
 459 })
 460
 461 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 462 do {                                                                    \
 463         const struct lu_env        *__env  = (_env);                    \
 464         struct cl_page             *__page = (_page);                   \
 465         const struct cl_page_slice *__scan;                             \
 466         ptrdiff_t                   __op   = (_op);                     \
 467         void                      (*__method)_proto;                    \
 468                                                                         \
 469         cfs_list_for_each_entry(__scan, &__page->cp_layers,             \
 470                                 cpl_linkage) {                          \
 471                 __method = *(void **)((char *)__scan->cpl_ops +  __op); \
 472                 if (__method != NULL)                                   \
 473                         (*__method)(__env, __scan, ## __VA_ARGS__);     \
 474         }                                                               \
 475 } while (0)
 476
 477 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)           \
 478 do {                                                                    \
 479         const struct lu_env        *__env  = (_env);                    \
 480         struct cl_page             *__page = (_page);                   \
 481         const struct cl_page_slice *__scan;                             \
 482         ptrdiff_t                   __op   = (_op);                     \
 483         void                      (*__method)_proto;                    \
 484                                                                         \
 485         /* get to the bottom page. */                                   \
 486         cfs_list_for_each_entry_reverse(__scan, &__page->cp_layers,     \
 487                                         cpl_linkage) {                  \
 488                 __method = *(void **)((char *)__scan->cpl_ops + __op);  \
 489                 if (__method != NULL)                                   \
 490                         (*__method)(__env, __scan, ## __VA_ARGS__);     \
 491         }                                                               \
 492 } while (0)
 493
 494 static int cl_page_invoke(const struct lu_env *env,
 495                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 496
 497 {
 498         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 499         ENTRY;
 500         RETURN(CL_PAGE_INVOKE(env, page, op,
 501                               (const struct lu_env *,
 502                                const struct cl_page_slice *, struct cl_io *),
 503                               io));
 504 }
 505
 506 static void cl_page_invoid(const struct lu_env *env,
 507                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 508
 509 {
 510         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 511         ENTRY;
 512         CL_PAGE_INVOID(env, page, op,
 513                        (const struct lu_env *,
 514                         const struct cl_page_slice *, struct cl_io *), io);
 515         EXIT;
 516 }
 517
 518 static void cl_page_owner_clear(struct cl_page *page)
 519 {
 520         ENTRY;
 521         if (page->cp_owner != NULL) {
 522                 LASSERT(page->cp_owner->ci_owned_nr > 0);
 523                 page->cp_owner->ci_owned_nr--;
 524                 page->cp_owner = NULL;
 525                 page->cp_task = NULL;
 526         }
 527         EXIT;
 528 }
 529
 530 static void cl_page_owner_set(struct cl_page *page)
 531 {
 532         ENTRY;
 533         LASSERT(page->cp_owner != NULL);
 534         page->cp_owner->ci_owned_nr++;
 535         EXIT;
 536 }
 537
 538 void cl_page_disown0(const struct lu_env *env,
 539                      struct cl_io *io, struct cl_page *pg)
 540 {
 541         enum cl_page_state state;
 542
 543         ENTRY;
 544         state = pg->cp_state;
 545         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 546         PINVRNT(env, pg, cl_page_invariant(pg) || state == CPS_FREEING);
 547         cl_page_owner_clear(pg);
 548
 549         if (state == CPS_OWNED)
 550                 cl_page_state_set(env, pg, CPS_CACHED);
 551         /*
 552          * Completion call-backs are executed in the bottom-up order, so that
 553          * uppermost layer (llite), responsible for VFS/VM interaction runs
 554          * last and can release locks safely.
 555          */
 556         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 557                                (const struct lu_env *,
 558                                 const struct cl_page_slice *, struct cl_io *),
 559                                io);
 560         EXIT;
 561 }
 562
 563 /**
 564  * returns true, iff page is owned by the given io.
 565  */
 566 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 567 {
 568         struct cl_io *top = cl_io_top((struct cl_io *)io);
 569         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 570         ENTRY;
 571         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top);
 572 }
 573 EXPORT_SYMBOL(cl_page_is_owned);
 574
 575 /**
 576  * Try to own a page by IO.
 577  *
 578  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 579  * into cl_page_state::CPS_OWNED state.
 580  *
 581  * \pre  !cl_page_is_owned(pg, io)
 582  * \post result == 0 iff cl_page_is_owned(pg, io)
 583  *
 584  * \retval 0   success
 585  *
 586  * \retval -ve failure, e.g., page was destroyed (and landed in
 587  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 588  *             or, page was owned by another thread, or in IO.
 589  *
 590  * \see cl_page_disown()
 591  * \see cl_page_operations::cpo_own()
 592  * \see cl_page_own_try()
 593  * \see cl_page_own
 594  */
 595 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 596                         struct cl_page *pg, int nonblock)
 597 {
 598         int result;
 599
 600         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 601
 602         ENTRY;
 603         io = cl_io_top(io);
 604
 605         if (pg->cp_state == CPS_FREEING) {
 606                 result = -ENOENT;
 607         } else {
 608                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 609                                         (const struct lu_env *,
 610                                          const struct cl_page_slice *,
 611                                          struct cl_io *, int),
 612                                         io, nonblock);
 613                 if (result == 0) {
 614                         PASSERT(env, pg, pg->cp_owner == NULL);
 615                         PASSERT(env, pg, pg->cp_req == NULL);
 616                         pg->cp_owner = cl_io_top(io);;
 617                         pg->cp_task  = current;
 618                         cl_page_owner_set(pg);
 619                         if (pg->cp_state != CPS_FREEING) {
 620                                 cl_page_state_set(env, pg, CPS_OWNED);
 621                         } else {
 622                                 cl_page_disown0(env, io, pg);
 623                                 result = -ENOENT;
 624                         }
 625                 }
 626         }
 627         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 628         RETURN(result);
 629 }
 630
 631 /**
 632  * Own a page, might be blocked.
 633  *
 634  * \see cl_page_own0()
 635  */
 636 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 637 {
 638         return cl_page_own0(env, io, pg, 0);
 639 }
 640 EXPORT_SYMBOL(cl_page_own);
 641
 642 /**
 643  * Nonblock version of cl_page_own().
 644  *
 645  * \see cl_page_own0()
 646  */
 647 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
 648                     struct cl_page *pg)
 649 {
 650         return cl_page_own0(env, io, pg, 1);
 651 }
 652 EXPORT_SYMBOL(cl_page_own_try);
 653
 654
 655 /**
 656  * Assume page ownership.
 657  *
 658  * Called when page is already locked by the hosting VM.
 659  *
 660  * \pre !cl_page_is_owned(pg, io)
 661  * \post cl_page_is_owned(pg, io)
 662  *
 663  * \see cl_page_operations::cpo_assume()
 664  */
 665 void cl_page_assume(const struct lu_env *env,
 666                     struct cl_io *io, struct cl_page *pg)
 667 {
 668         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
 669
 670         ENTRY;
 671         io = cl_io_top(io);
 672
 673         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
 674         PASSERT(env, pg, pg->cp_owner == NULL);
 675         pg->cp_owner = cl_io_top(io);
 676         pg->cp_task = current;
 677         cl_page_owner_set(pg);
 678         cl_page_state_set(env, pg, CPS_OWNED);
 679         EXIT;
 680 }
 681 EXPORT_SYMBOL(cl_page_assume);
 682
 683 /**
 684  * Releases page ownership without unlocking the page.
 685  *
 686  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
 687  * underlying VM page (as VM is supposed to do this itself).
 688  *
 689  * \pre   cl_page_is_owned(pg, io)
 690  * \post !cl_page_is_owned(pg, io)
 691  *
 692  * \see cl_page_assume()
 693  */
 694 void cl_page_unassume(const struct lu_env *env,
 695                       struct cl_io *io, struct cl_page *pg)
 696 {
 697         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 698         PINVRNT(env, pg, cl_page_invariant(pg));
 699
 700         ENTRY;
 701         io = cl_io_top(io);
 702         cl_page_owner_clear(pg);
 703         cl_page_state_set(env, pg, CPS_CACHED);
 704         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
 705                                (const struct lu_env *,
 706                                 const struct cl_page_slice *, struct cl_io *),
 707                                io);
 708         EXIT;
 709 }
 710 EXPORT_SYMBOL(cl_page_unassume);
 711
 712 /**
 713  * Releases page ownership.
 714  *
 715  * Moves page into cl_page_state::CPS_CACHED.
 716  *
 717  * \pre   cl_page_is_owned(pg, io)
 718  * \post !cl_page_is_owned(pg, io)
 719  *
 720  * \see cl_page_own()
 721  * \see cl_page_operations::cpo_disown()
 722  */
 723 void cl_page_disown(const struct lu_env *env,
 724                     struct cl_io *io, struct cl_page *pg)
 725 {
 726         PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
 727                 pg->cp_state == CPS_FREEING);
 728
 729         ENTRY;
 730         io = cl_io_top(io);
 731         cl_page_disown0(env, io, pg);
 732         EXIT;
 733 }
 734 EXPORT_SYMBOL(cl_page_disown);
 735
 736 /**
 737  * Called when page is to be removed from the object, e.g., as a result of
 738  * truncate.
 739  *
 740  * Calls cl_page_operations::cpo_discard() top-to-bottom.
 741  *
 742  * \pre cl_page_is_owned(pg, io)
 743  *
 744  * \see cl_page_operations::cpo_discard()
 745  */
 746 void cl_page_discard(const struct lu_env *env,
 747                      struct cl_io *io, struct cl_page *pg)
 748 {
 749         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 750         PINVRNT(env, pg, cl_page_invariant(pg));
 751
 752         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
 753 }
 754 EXPORT_SYMBOL(cl_page_discard);
 755
 756 /**
 757  * Version of cl_page_delete() that can be called for not fully constructed
 758  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
 759  * path. Doesn't check page invariant.
 760  */
 761 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg)
 762 {
 763         ENTRY;
 764
 765         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
 766
 767         /*
 768          * Severe all ways to obtain new pointers to @pg.
 769          */
 770         cl_page_owner_clear(pg);
 771
 772         cl_page_state_set0(env, pg, CPS_FREEING);
 773
 774         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_delete),
 775                        (const struct lu_env *, const struct cl_page_slice *));
 776
 777         EXIT;
 778 }
 779
 780 /**
 781  * Called when a decision is made to throw page out of memory.
 782  *
 783  * Notifies all layers about page destruction by calling
 784  * cl_page_operations::cpo_delete() method top-to-bottom.
 785  *
 786  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
 787  * where transition to this state happens).
 788  *
 789  * Eliminates all venues through which new references to the page can be
 790  * obtained:
 791  *
 792  *     - removes page from the radix trees,
 793  *
 794  *     - breaks linkage from VM page to cl_page.
 795  *
 796  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
 797  * drain after some time, at which point page will be recycled.
 798  *
 799  * \pre  VM page is locked
 800  * \post pg->cp_state == CPS_FREEING
 801  *
 802  * \see cl_page_operations::cpo_delete()
 803  */
 804 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
 805 {
 806         PINVRNT(env, pg, cl_page_invariant(pg));
 807         ENTRY;
 808         cl_page_delete0(env, pg);
 809         EXIT;
 810 }
 811 EXPORT_SYMBOL(cl_page_delete);
 812
 813 /**
 814  * Marks page up-to-date.
 815  *
 816  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
 817  * layer responsible for VM interaction has to mark/clear page as up-to-date
 818  * by the \a uptodate argument.
 819  *
 820  * \see cl_page_operations::cpo_export()
 821  */
 822 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
 823 {
 824         PINVRNT(env, pg, cl_page_invariant(pg));
 825         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
 826                        (const struct lu_env *,
 827                         const struct cl_page_slice *, int), uptodate);
 828 }
 829 EXPORT_SYMBOL(cl_page_export);
 830
 831 /**
 832  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
 833  * thread.
 834  */
 835 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
 836 {
 837         int result;
 838         const struct cl_page_slice *slice;
 839
 840         ENTRY;
 841         slice = container_of(pg->cp_layers.next,
 842                              const struct cl_page_slice, cpl_linkage);
 843         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
 844         /*
 845          * Call ->cpo_is_vmlocked() directly instead of going through
 846          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
 847          * cl_page_invariant().
 848          */
 849         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
 850         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
 851         RETURN(result == -EBUSY);
 852 }
 853 EXPORT_SYMBOL(cl_page_is_vmlocked);
 854
 855 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
 856 {
 857         ENTRY;
 858         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
 859 }
 860
 861 static void cl_page_io_start(const struct lu_env *env,
 862                              struct cl_page *pg, enum cl_req_type crt)
 863 {
 864         /*
 865          * Page is queued for IO, change its state.
 866          */
 867         ENTRY;
 868         cl_page_owner_clear(pg);
 869         cl_page_state_set(env, pg, cl_req_type_state(crt));
 870         EXIT;
 871 }
 872
 873 /**
 874  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
 875  * called top-to-bottom. Every layer either agrees to submit this page (by
 876  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
 877  * handling interactions with the VM also has to inform VM that page is under
 878  * transfer now.
 879  */
 880 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
 881                  struct cl_page *pg, enum cl_req_type crt)
 882 {
 883         int result;
 884
 885         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 886         PINVRNT(env, pg, cl_page_invariant(pg));
 887         PINVRNT(env, pg, crt < CRT_NR);
 888
 889         /*
 890          * XXX this has to be called bottom-to-top, so that llite can set up
 891          * PG_writeback without risking other layers deciding to skip this
 892          * page.
 893          */
 894         if (crt >= CRT_NR)
 895                 return -EINVAL;
 896         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
 897         if (result == 0)
 898                 cl_page_io_start(env, pg, crt);
 899
 900         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
 901                       equi(result == 0,
 902                            PageWriteback(cl_page_vmpage(pg)))));
 903         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 904         return result;
 905 }
 906 EXPORT_SYMBOL(cl_page_prep);
 907
 908 /**
 909  * Notify layers about transfer completion.
 910  *
 911  * Invoked by transfer sub-system (which is a part of osc) to notify layers
 912  * that a transfer, of which this page is a part of has completed.
 913  *
 914  * Completion call-backs are executed in the bottom-up order, so that
 915  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
 916  * and can release locks safely.
 917  *
 918  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 919  * \post pg->cp_state == CPS_CACHED
 920  *
 921  * \see cl_page_operations::cpo_completion()
 922  */
 923 void cl_page_completion(const struct lu_env *env,
 924                         struct cl_page *pg, enum cl_req_type crt, int ioret)
 925 {
 926         struct cl_sync_io *anchor = pg->cp_sync_io;
 927
 928         PASSERT(env, pg, crt < CRT_NR);
 929         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
 930         PASSERT(env, pg, pg->cp_req == NULL);
 931         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
 932
 933         ENTRY;
 934         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
 935         if (crt == CRT_READ && ioret == 0) {
 936                 PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED));
 937                 pg->cp_flags |= CPF_READ_COMPLETED;
 938         }
 939
 940         cl_page_state_set(env, pg, CPS_CACHED);
 941         if (crt >= CRT_NR)
 942                 return;
 943         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
 944                                (const struct lu_env *,
 945                                 const struct cl_page_slice *, int), ioret);
 946         if (anchor) {
 947                 LASSERT(cl_page_is_vmlocked(env, pg));
 948                 LASSERT(pg->cp_sync_io == anchor);
 949                 pg->cp_sync_io = NULL;
 950         }
 951         /*
 952          * As page->cp_obj is pinned by a reference from page->cp_req, it is
 953          * safe to call cl_page_put() without risking object destruction in a
 954          * non-blocking context.
 955          */
 956         cl_page_put(env, pg);
 957
 958         if (anchor)
 959                 cl_sync_io_note(anchor, ioret);
 960
 961         EXIT;
 962 }
 963 EXPORT_SYMBOL(cl_page_completion);
 964
 965 /**
 966  * Notify layers that transfer formation engine decided to yank this page from
 967  * the cache and to make it a part of a transfer.
 968  *
 969  * \pre  pg->cp_state == CPS_CACHED
 970  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 971  *
 972  * \see cl_page_operations::cpo_make_ready()
 973  */
 974 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
 975                        enum cl_req_type crt)
 976 {
 977         int result;
 978
 979         PINVRNT(env, pg, crt < CRT_NR);
 980
 981         ENTRY;
 982         if (crt >= CRT_NR)
 983                 RETURN(-EINVAL);
 984         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
 985                                 (const struct lu_env *,
 986                                  const struct cl_page_slice *));
 987         if (result == 0) {
 988                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
 989                 cl_page_io_start(env, pg, crt);
 990         }
 991         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 992         RETURN(result);
 993 }
 994 EXPORT_SYMBOL(cl_page_make_ready);
 995
 996 /**
 997  * Called if a pge is being written back by kernel's intention.
 998  *
 999  * \pre  cl_page_is_owned(pg, io)
1000  * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT)
1001  *
1002  * \see cl_page_operations::cpo_flush()
1003  */
1004 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
1005                   struct cl_page *pg)
1006 {
1007         int result;
1008
1009         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1010         PINVRNT(env, pg, cl_page_invariant(pg));
1011
1012         ENTRY;
1013
1014         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_flush));
1015
1016         CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result);
1017         RETURN(result);
1018 }
1019 EXPORT_SYMBOL(cl_page_flush);
1020
1021 /**
1022  * Checks whether page is protected by any extent lock is at least required
1023  * mode.
1024  *
1025  * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1026  * \see cl_page_operations::cpo_is_under_lock()
1027  */
1028 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1029                           struct cl_page *page)
1030 {
1031         int rc;
1032
1033         PINVRNT(env, page, cl_page_invariant(page));
1034
1035         ENTRY;
1036         rc = CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1037                             (const struct lu_env *,
1038                              const struct cl_page_slice *, struct cl_io *),
1039                             io);
1040         PASSERT(env, page, rc != 0);
1041         RETURN(rc);
1042 }
1043 EXPORT_SYMBOL(cl_page_is_under_lock);
1044
1045 /**
1046  * Tells transfer engine that only part of a page is to be transmitted.
1047  *
1048  * \see cl_page_operations::cpo_clip()
1049  */
1050 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1051                   int from, int to)
1052 {
1053         PINVRNT(env, pg, cl_page_invariant(pg));
1054
1055         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1056         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1057                        (const struct lu_env *,
1058                         const struct cl_page_slice *,int, int),
1059                        from, to);
1060 }
1061 EXPORT_SYMBOL(cl_page_clip);
1062
1063 /**
1064  * Prints human readable representation of \a pg to the \a f.
1065  */
1066 void cl_page_header_print(const struct lu_env *env, void *cookie,
1067                           lu_printer_t printer, const struct cl_page *pg)
1068 {
1069         (*printer)(env, cookie,
1070                    "page@%p[%d %p %d %d %d %p %p %#x]\n",
1071                    pg, cfs_atomic_read(&pg->cp_ref), pg->cp_obj,
1072                    pg->cp_state, pg->cp_error, pg->cp_type,
1073                    pg->cp_owner, pg->cp_req, pg->cp_flags);
1074 }
1075 EXPORT_SYMBOL(cl_page_header_print);
1076
1077 /**
1078  * Prints human readable representation of \a pg to the \a f.
1079  */
1080 void cl_page_print(const struct lu_env *env, void *cookie,
1081                    lu_printer_t printer, const struct cl_page *pg)
1082 {
1083         cl_page_header_print(env, cookie, printer, pg);
1084         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1085                        (const struct lu_env *env,
1086                         const struct cl_page_slice *slice,
1087                         void *cookie, lu_printer_t p), cookie, printer);
1088         (*printer)(env, cookie, "end page@%p\n", pg);
1089 }
1090 EXPORT_SYMBOL(cl_page_print);
1091
1092 /**
1093  * Cancel a page which is still in a transfer.
1094  */
1095 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1096 {
1097         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1098                               (const struct lu_env *,
1099                                const struct cl_page_slice *));
1100 }
1101 EXPORT_SYMBOL(cl_page_cancel);
1102
1103 /**
1104  * Converts a byte offset within object \a obj into a page index.
1105  */
1106 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1107 {
1108         return (loff_t)idx << PAGE_CACHE_SHIFT;
1109 }
1110 EXPORT_SYMBOL(cl_offset);
1111
1112 /**
1113  * Converts a page index into a byte offset within object \a obj.
1114  */
1115 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1116 {
1117         return offset >> PAGE_CACHE_SHIFT;
1118 }
1119 EXPORT_SYMBOL(cl_index);
1120
1121 int cl_page_size(const struct cl_object *obj)
1122 {
1123         return 1 << PAGE_CACHE_SHIFT;
1124 }
1125 EXPORT_SYMBOL(cl_page_size);
1126
1127 /**
1128  * Adds page slice to the compound page.
1129  *
1130  * This is called by cl_object_operations::coo_page_init() methods to add a
1131  * per-layer state to the page. New state is added at the end of
1132  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1133  *
1134  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1135  */
1136 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1137                        struct cl_object *obj,
1138                        const struct cl_page_operations *ops)
1139 {
1140         ENTRY;
1141         cfs_list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1142         slice->cpl_obj  = obj;
1143         slice->cpl_ops  = ops;
1144         slice->cpl_page = page;
1145         EXIT;
1146 }
1147 EXPORT_SYMBOL(cl_page_slice_add);
1148
1149 int  cl_page_init(void)
1150 {
1151         return 0;
1152 }
1153
1154 void cl_page_fini(void)
1155 {
1156 }