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         atomic_inc(&cl_object_site(o)->cs_pages.cs_stats[CS_##item])
  84 #define CS_PAGE_DEC(o, item) \
  85         atomic_dec(&cl_object_site(o)->cs_pages.cs_stats[CS_##item])
  86 #define CS_PAGESTATE_INC(o, state) \
  87         atomic_inc(&cl_object_site(o)->cs_pages_state[state])
  88 #define CS_PAGESTATE_DEC(o, state) \
  89         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(atomic_read(&page->cp_ref) > 0);
 110         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         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, 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 (!list_empty(&page->cp_layers)) {
 145                 struct cl_page_slice *slice;
 146
 147                 slice = list_entry(page->cp_layers.next,
 148                                    struct cl_page_slice, cpl_linkage);
 149                 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_NOFS);
 183         if (page != NULL) {
 184                 int result = 0;
 185                 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                 INIT_LIST_HEAD(&page->cp_layers);
 194                 INIT_LIST_HEAD(&page->cp_batch);
 195                 INIT_LIST_HEAD(&page->cp_flight);
 196                 lu_ref_init(&page->cp_reference);
 197                 head = o->co_lu.lo_header;
 198                 list_for_each_entry(o, &head->loh_layers,
 199                                     co_lu.lo_linkage) {
 200                         if (o->co_ops->coo_page_init != NULL) {
 201                                 result = o->co_ops->coo_page_init(env, o, page,
 202                                                                   ind);
 203                                 if (result != 0) {
 204                                         cl_page_delete0(env, page);
 205                                         cl_page_free(env, page);
 206                                         page = ERR_PTR(result);
 207                                         break;
 208                                 }
 209                         }
 210                 }
 211                 if (result == 0) {
 212                         CS_PAGE_INC(o, total);
 213                         CS_PAGE_INC(o, create);
 214                         CS_PAGESTATE_DEC(o, CPS_CACHED);
 215                 }
 216         } else {
 217                 page = ERR_PTR(-ENOMEM);
 218         }
 219         RETURN(page);
 220 }
 221 EXPORT_SYMBOL(cl_page_alloc);
 222
 223 /**
 224  * Returns a cl_page with index \a idx at the object \a o, and associated with
 225  * the VM page \a vmpage.
 226  *
 227  * This is the main entry point into the cl_page caching interface. First, a
 228  * cache (implemented as a per-object radix tree) is consulted. If page is
 229  * found there, it is returned immediately. Otherwise new page is allocated
 230  * and returned. In any case, additional reference to page is acquired.
 231  *
 232  * \see cl_object_find(), cl_lock_find()
 233  */
 234 struct cl_page *cl_page_find(const struct lu_env *env,
 235                              struct cl_object *o,
 236                              pgoff_t idx, struct page *vmpage,
 237                              enum cl_page_type type)
 238 {
 239         struct cl_page          *page = NULL;
 240         struct cl_object_header *hdr;
 241
 242         LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 243         might_sleep();
 244
 245         ENTRY;
 246
 247         hdr = cl_object_header(o);
 248         CS_PAGE_INC(o, lookup);
 249
 250         CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
 251                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 252         /* fast path. */
 253         if (type == CPT_CACHEABLE) {
 254                 /* vmpage lock is used to protect the child/parent
 255                  * relationship */
 256                 KLASSERT(PageLocked(vmpage));
 257                 /*
 258                  * cl_vmpage_page() can be called here without any locks as
 259                  *
 260                  *     - "vmpage" is locked (which prevents ->private from
 261                  *       concurrent updates), and
 262                  *
 263                  *     - "o" cannot be destroyed while current thread holds a
 264                  *       reference on it.
 265                  */
 266                 page = cl_vmpage_page(vmpage, o);
 267                 if (page != NULL) {
 268                         CS_PAGE_INC(o, hit);
 269                         RETURN(page);
 270                 }
 271         }
 272
 273         /* allocate and initialize cl_page */
 274         page = cl_page_alloc(env, o, idx, vmpage, type);
 275         RETURN(page);
 276 }
 277 EXPORT_SYMBOL(cl_page_find);
 278
 279 static inline int cl_page_invariant(const struct cl_page *pg)
 280 {
 281         /*
 282          * Page invariant is protected by a VM lock.
 283          */
 284         LINVRNT(cl_page_is_vmlocked(NULL, pg));
 285
 286         return cl_page_in_use_noref(pg);
 287 }
 288
 289 static void cl_page_state_set0(const struct lu_env *env,
 290                                struct cl_page *page, enum cl_page_state state)
 291 {
 292         enum cl_page_state old;
 293
 294         /*
 295          * Matrix of allowed state transitions [old][new], for sanity
 296          * checking.
 297          */
 298         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 299                 [CPS_CACHED] = {
 300                         [CPS_CACHED]  = 0,
 301                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 302                         [CPS_PAGEIN]  = 0,
 303                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 304                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 305                 },
 306                 [CPS_OWNED] = {
 307                         [CPS_CACHED]  = 1, /* release to the cache */
 308                         [CPS_OWNED]   = 0,
 309                         [CPS_PAGEIN]  = 1, /* start read immediately */
 310                         [CPS_PAGEOUT] = 1, /* start write immediately */
 311                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 312                 },
 313                 [CPS_PAGEIN] = {
 314                         [CPS_CACHED]  = 1, /* io completion */
 315                         [CPS_OWNED]   = 0,
 316                         [CPS_PAGEIN]  = 0,
 317                         [CPS_PAGEOUT] = 0,
 318                         [CPS_FREEING] = 0,
 319                 },
 320                 [CPS_PAGEOUT] = {
 321                         [CPS_CACHED]  = 1, /* io completion */
 322                         [CPS_OWNED]   = 0,
 323                         [CPS_PAGEIN]  = 0,
 324                         [CPS_PAGEOUT] = 0,
 325                         [CPS_FREEING] = 0,
 326                 },
 327                 [CPS_FREEING] = {
 328                         [CPS_CACHED]  = 0,
 329                         [CPS_OWNED]   = 0,
 330                         [CPS_PAGEIN]  = 0,
 331                         [CPS_PAGEOUT] = 0,
 332                         [CPS_FREEING] = 0,
 333                 }
 334         };
 335
 336         ENTRY;
 337         old = page->cp_state;
 338         PASSERT(env, page, allowed_transitions[old][state]);
 339         CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
 340         PASSERT(env, page, page->cp_state == old);
 341         PASSERT(env, page, equi(state == CPS_OWNED, page->cp_owner != NULL));
 342
 343         CS_PAGESTATE_DEC(page->cp_obj, page->cp_state);
 344         CS_PAGESTATE_INC(page->cp_obj, state);
 345         cl_page_state_set_trust(page, state);
 346         EXIT;
 347 }
 348
 349 static void cl_page_state_set(const struct lu_env *env,
 350                               struct cl_page *page, enum cl_page_state state)
 351 {
 352         cl_page_state_set0(env, page, state);
 353 }
 354
 355 /**
 356  * Acquires an additional reference to a page.
 357  *
 358  * This can be called only by caller already possessing a reference to \a
 359  * page.
 360  *
 361  * \see cl_object_get(), cl_lock_get().
 362  */
 363 void cl_page_get(struct cl_page *page)
 364 {
 365         ENTRY;
 366         cl_page_get_trust(page);
 367         EXIT;
 368 }
 369 EXPORT_SYMBOL(cl_page_get);
 370
 371 /**
 372  * Releases a reference to a page.
 373  *
 374  * When last reference is released, page is returned to the cache, unless it
 375  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 376  * destroyed.
 377  *
 378  * \see cl_object_put(), cl_lock_put().
 379  */
 380 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 381 {
 382         ENTRY;
 383         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 384                        atomic_read(&page->cp_ref));
 385
 386         if (atomic_dec_and_test(&page->cp_ref)) {
 387                 LASSERT(page->cp_state == CPS_FREEING);
 388
 389                 LASSERT(atomic_read(&page->cp_ref) == 0);
 390                 PASSERT(env, page, page->cp_owner == NULL);
 391                 PASSERT(env, page, list_empty(&page->cp_batch));
 392                 /*
 393                  * Page is no longer reachable by other threads. Tear
 394                  * it down.
 395                  */
 396                 cl_page_free(env, page);
 397         }
 398
 399         EXIT;
 400 }
 401 EXPORT_SYMBOL(cl_page_put);
 402
 403 /**
 404  * Returns a cl_page associated with a VM page, and given cl_object.
 405  */
 406 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj)
 407 {
 408         struct cl_page *page;
 409
 410         ENTRY;
 411         KLASSERT(PageLocked(vmpage));
 412
 413         /*
 414          * NOTE: absence of races and liveness of data are guaranteed by page
 415          *       lock on a "vmpage". That works because object destruction has
 416          *       bottom-to-top pass.
 417          */
 418
 419         page = (struct cl_page *)vmpage->private;
 420         if (page != NULL) {
 421                 cl_page_get_trust(page);
 422                 LASSERT(page->cp_type == CPT_CACHEABLE);
 423         }
 424         RETURN(page);
 425 }
 426 EXPORT_SYMBOL(cl_vmpage_page);
 427
 428 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 429                                        const struct lu_device_type *dtype)
 430 {
 431         return cl_page_at_trusted(page, dtype);
 432 }
 433 EXPORT_SYMBOL(cl_page_at);
 434
 435 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 436
 437 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)                   \
 438 ({                                                                      \
 439         const struct lu_env        *__env  = (_env);                    \
 440         struct cl_page             *__page = (_page);                   \
 441         const struct cl_page_slice *__scan;                             \
 442         int                         __result;                           \
 443         ptrdiff_t                   __op   = (_op);                     \
 444         int                        (*__method)_proto;                   \
 445                                                                         \
 446         __result = 0;                                                   \
 447         list_for_each_entry(__scan, &__page->cp_layers, cpl_linkage) {  \
 448                 __method = *(void **)((char *)__scan->cpl_ops +  __op); \
 449                 if (__method != NULL) {                                 \
 450                         __result = (*__method)(__env, __scan, ## __VA_ARGS__); \
 451                         if (__result != 0)                              \
 452                                 break;                                  \
 453                 }                                                       \
 454         }                                                               \
 455         if (__result > 0)                                               \
 456                 __result = 0;                                           \
 457         __result;                                                       \
 458 })
 459
 460 #define CL_PAGE_INVOKE_REVERSE(_env, _page, _op, _proto, ...)           \
 461 ({                                                                      \
 462         const struct lu_env        *__env  = (_env);                    \
 463         struct cl_page             *__page = (_page);                   \
 464         const struct cl_page_slice *__scan;                             \
 465         int                         __result;                           \
 466         ptrdiff_t                   __op   = (_op);                     \
 467         int                       (*__method)_proto;                    \
 468                                                                         \
 469         __result = 0;                                                   \
 470         list_for_each_entry_reverse(__scan, &__page->cp_layers,         \
 471                                     cpl_linkage) {                      \
 472                 __method = *(void **)((char *)__scan->cpl_ops +  __op); \
 473                 if (__method != NULL) {                                 \
 474                         __result = (*__method)(__env, __scan, ## __VA_ARGS__); \
 475                         if (__result != 0)                              \
 476                                 break;                                  \
 477                 }                                                       \
 478         }                                                               \
 479         if (__result > 0)                                               \
 480                 __result = 0;                                           \
 481         __result;                                                       \
 482 })
 483
 484 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 485 do {                                                                    \
 486         const struct lu_env        *__env  = (_env);                    \
 487         struct cl_page             *__page = (_page);                   \
 488         const struct cl_page_slice *__scan;                             \
 489         ptrdiff_t                   __op   = (_op);                     \
 490         void                      (*__method)_proto;                    \
 491                                                                         \
 492         list_for_each_entry(__scan, &__page->cp_layers, cpl_linkage) {  \
 493                 __method = *(void **)((char *)__scan->cpl_ops +  __op); \
 494                 if (__method != NULL)                                   \
 495                         (*__method)(__env, __scan, ## __VA_ARGS__);     \
 496         }                                                               \
 497 } while (0)
 498
 499 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)           \
 500 do {                                                                    \
 501         const struct lu_env        *__env  = (_env);                    \
 502         struct cl_page             *__page = (_page);                   \
 503         const struct cl_page_slice *__scan;                             \
 504         ptrdiff_t                   __op   = (_op);                     \
 505         void                      (*__method)_proto;                    \
 506                                                                         \
 507         /* get to the bottom page. */                                   \
 508         list_for_each_entry_reverse(__scan, &__page->cp_layers,         \
 509                                     cpl_linkage) {                      \
 510                 __method = *(void **)((char *)__scan->cpl_ops + __op);  \
 511                 if (__method != NULL)                                   \
 512                         (*__method)(__env, __scan, ## __VA_ARGS__);     \
 513         }                                                               \
 514 } while (0)
 515
 516 static int cl_page_invoke(const struct lu_env *env,
 517                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 518
 519 {
 520         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 521         ENTRY;
 522         RETURN(CL_PAGE_INVOKE(env, page, op,
 523                               (const struct lu_env *,
 524                                const struct cl_page_slice *, struct cl_io *),
 525                               io));
 526 }
 527
 528 static void cl_page_invoid(const struct lu_env *env,
 529                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 530
 531 {
 532         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 533         ENTRY;
 534         CL_PAGE_INVOID(env, page, op,
 535                        (const struct lu_env *,
 536                         const struct cl_page_slice *, struct cl_io *), io);
 537         EXIT;
 538 }
 539
 540 static void cl_page_owner_clear(struct cl_page *page)
 541 {
 542         ENTRY;
 543         if (page->cp_owner != NULL) {
 544                 LASSERT(page->cp_owner->ci_owned_nr > 0);
 545                 page->cp_owner->ci_owned_nr--;
 546                 page->cp_owner = NULL;
 547         }
 548         EXIT;
 549 }
 550
 551 static void cl_page_owner_set(struct cl_page *page)
 552 {
 553         ENTRY;
 554         LASSERT(page->cp_owner != NULL);
 555         page->cp_owner->ci_owned_nr++;
 556         EXIT;
 557 }
 558
 559 void cl_page_disown0(const struct lu_env *env,
 560                      struct cl_io *io, struct cl_page *pg)
 561 {
 562         enum cl_page_state state;
 563
 564         ENTRY;
 565         state = pg->cp_state;
 566         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 567         PINVRNT(env, pg, cl_page_invariant(pg) || state == CPS_FREEING);
 568         cl_page_owner_clear(pg);
 569
 570         if (state == CPS_OWNED)
 571                 cl_page_state_set(env, pg, CPS_CACHED);
 572         /*
 573          * Completion call-backs are executed in the bottom-up order, so that
 574          * uppermost layer (llite), responsible for VFS/VM interaction runs
 575          * last and can release locks safely.
 576          */
 577         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 578                                (const struct lu_env *,
 579                                 const struct cl_page_slice *, struct cl_io *),
 580                                io);
 581         EXIT;
 582 }
 583
 584 /**
 585  * returns true, iff page is owned by the given io.
 586  */
 587 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 588 {
 589         struct cl_io *top = cl_io_top((struct cl_io *)io);
 590         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 591         ENTRY;
 592         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top);
 593 }
 594 EXPORT_SYMBOL(cl_page_is_owned);
 595
 596 /**
 597  * Try to own a page by IO.
 598  *
 599  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 600  * into cl_page_state::CPS_OWNED state.
 601  *
 602  * \pre  !cl_page_is_owned(pg, io)
 603  * \post result == 0 iff cl_page_is_owned(pg, io)
 604  *
 605  * \retval 0   success
 606  *
 607  * \retval -ve failure, e.g., page was destroyed (and landed in
 608  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 609  *             or, page was owned by another thread, or in IO.
 610  *
 611  * \see cl_page_disown()
 612  * \see cl_page_operations::cpo_own()
 613  * \see cl_page_own_try()
 614  * \see cl_page_own
 615  */
 616 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 617                         struct cl_page *pg, int nonblock)
 618 {
 619         int result;
 620
 621         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 622
 623         ENTRY;
 624         io = cl_io_top(io);
 625
 626         if (pg->cp_state == CPS_FREEING) {
 627                 result = -ENOENT;
 628         } else {
 629                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 630                                         (const struct lu_env *,
 631                                          const struct cl_page_slice *,
 632                                          struct cl_io *, int),
 633                                         io, nonblock);
 634                 if (result == 0) {
 635                         PASSERT(env, pg, pg->cp_owner == NULL);
 636                         PASSERT(env, pg, pg->cp_req == NULL);
 637                         pg->cp_owner = cl_io_top(io);;
 638                         cl_page_owner_set(pg);
 639                         if (pg->cp_state != CPS_FREEING) {
 640                                 cl_page_state_set(env, pg, CPS_OWNED);
 641                         } else {
 642                                 cl_page_disown0(env, io, pg);
 643                                 result = -ENOENT;
 644                         }
 645                 }
 646         }
 647         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 648         RETURN(result);
 649 }
 650
 651 /**
 652  * Own a page, might be blocked.
 653  *
 654  * \see cl_page_own0()
 655  */
 656 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 657 {
 658         return cl_page_own0(env, io, pg, 0);
 659 }
 660 EXPORT_SYMBOL(cl_page_own);
 661
 662 /**
 663  * Nonblock version of cl_page_own().
 664  *
 665  * \see cl_page_own0()
 666  */
 667 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
 668                     struct cl_page *pg)
 669 {
 670         return cl_page_own0(env, io, pg, 1);
 671 }
 672 EXPORT_SYMBOL(cl_page_own_try);
 673
 674
 675 /**
 676  * Assume page ownership.
 677  *
 678  * Called when page is already locked by the hosting VM.
 679  *
 680  * \pre !cl_page_is_owned(pg, io)
 681  * \post cl_page_is_owned(pg, io)
 682  *
 683  * \see cl_page_operations::cpo_assume()
 684  */
 685 void cl_page_assume(const struct lu_env *env,
 686                     struct cl_io *io, struct cl_page *pg)
 687 {
 688         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
 689
 690         ENTRY;
 691         io = cl_io_top(io);
 692
 693         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
 694         PASSERT(env, pg, pg->cp_owner == NULL);
 695         pg->cp_owner = cl_io_top(io);
 696         cl_page_owner_set(pg);
 697         cl_page_state_set(env, pg, CPS_OWNED);
 698         EXIT;
 699 }
 700 EXPORT_SYMBOL(cl_page_assume);
 701
 702 /**
 703  * Releases page ownership without unlocking the page.
 704  *
 705  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
 706  * underlying VM page (as VM is supposed to do this itself).
 707  *
 708  * \pre   cl_page_is_owned(pg, io)
 709  * \post !cl_page_is_owned(pg, io)
 710  *
 711  * \see cl_page_assume()
 712  */
 713 void cl_page_unassume(const struct lu_env *env,
 714                       struct cl_io *io, struct cl_page *pg)
 715 {
 716         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 717         PINVRNT(env, pg, cl_page_invariant(pg));
 718
 719         ENTRY;
 720         io = cl_io_top(io);
 721         cl_page_owner_clear(pg);
 722         cl_page_state_set(env, pg, CPS_CACHED);
 723         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
 724                                (const struct lu_env *,
 725                                 const struct cl_page_slice *, struct cl_io *),
 726                                io);
 727         EXIT;
 728 }
 729 EXPORT_SYMBOL(cl_page_unassume);
 730
 731 /**
 732  * Releases page ownership.
 733  *
 734  * Moves page into cl_page_state::CPS_CACHED.
 735  *
 736  * \pre   cl_page_is_owned(pg, io)
 737  * \post !cl_page_is_owned(pg, io)
 738  *
 739  * \see cl_page_own()
 740  * \see cl_page_operations::cpo_disown()
 741  */
 742 void cl_page_disown(const struct lu_env *env,
 743                     struct cl_io *io, struct cl_page *pg)
 744 {
 745         PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
 746                 pg->cp_state == CPS_FREEING);
 747
 748         ENTRY;
 749         io = cl_io_top(io);
 750         cl_page_disown0(env, io, pg);
 751         EXIT;
 752 }
 753 EXPORT_SYMBOL(cl_page_disown);
 754
 755 /**
 756  * Called when page is to be removed from the object, e.g., as a result of
 757  * truncate.
 758  *
 759  * Calls cl_page_operations::cpo_discard() top-to-bottom.
 760  *
 761  * \pre cl_page_is_owned(pg, io)
 762  *
 763  * \see cl_page_operations::cpo_discard()
 764  */
 765 void cl_page_discard(const struct lu_env *env,
 766                      struct cl_io *io, struct cl_page *pg)
 767 {
 768         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 769         PINVRNT(env, pg, cl_page_invariant(pg));
 770
 771         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
 772 }
 773 EXPORT_SYMBOL(cl_page_discard);
 774
 775 /**
 776  * Version of cl_page_delete() that can be called for not fully constructed
 777  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
 778  * path. Doesn't check page invariant.
 779  */
 780 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg)
 781 {
 782         ENTRY;
 783
 784         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
 785
 786         /*
 787          * Severe all ways to obtain new pointers to @pg.
 788          */
 789         cl_page_owner_clear(pg);
 790
 791         cl_page_state_set0(env, pg, CPS_FREEING);
 792
 793         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_delete),
 794                        (const struct lu_env *, const struct cl_page_slice *));
 795
 796         EXIT;
 797 }
 798
 799 /**
 800  * Called when a decision is made to throw page out of memory.
 801  *
 802  * Notifies all layers about page destruction by calling
 803  * cl_page_operations::cpo_delete() method top-to-bottom.
 804  *
 805  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
 806  * where transition to this state happens).
 807  *
 808  * Eliminates all venues through which new references to the page can be
 809  * obtained:
 810  *
 811  *     - removes page from the radix trees,
 812  *
 813  *     - breaks linkage from VM page to cl_page.
 814  *
 815  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
 816  * drain after some time, at which point page will be recycled.
 817  *
 818  * \pre  VM page is locked
 819  * \post pg->cp_state == CPS_FREEING
 820  *
 821  * \see cl_page_operations::cpo_delete()
 822  */
 823 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
 824 {
 825         PINVRNT(env, pg, cl_page_invariant(pg));
 826         ENTRY;
 827         cl_page_delete0(env, pg);
 828         EXIT;
 829 }
 830 EXPORT_SYMBOL(cl_page_delete);
 831
 832 /**
 833  * Marks page up-to-date.
 834  *
 835  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
 836  * layer responsible for VM interaction has to mark/clear page as up-to-date
 837  * by the \a uptodate argument.
 838  *
 839  * \see cl_page_operations::cpo_export()
 840  */
 841 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
 842 {
 843         PINVRNT(env, pg, cl_page_invariant(pg));
 844         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
 845                        (const struct lu_env *,
 846                         const struct cl_page_slice *, int), uptodate);
 847 }
 848 EXPORT_SYMBOL(cl_page_export);
 849
 850 /**
 851  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
 852  * thread.
 853  */
 854 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
 855 {
 856         int result;
 857         const struct cl_page_slice *slice;
 858
 859         ENTRY;
 860         slice = container_of(pg->cp_layers.next,
 861                              const struct cl_page_slice, cpl_linkage);
 862         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
 863         /*
 864          * Call ->cpo_is_vmlocked() directly instead of going through
 865          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
 866          * cl_page_invariant().
 867          */
 868         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
 869         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
 870         RETURN(result == -EBUSY);
 871 }
 872 EXPORT_SYMBOL(cl_page_is_vmlocked);
 873
 874 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
 875 {
 876         ENTRY;
 877         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
 878 }
 879
 880 static void cl_page_io_start(const struct lu_env *env,
 881                              struct cl_page *pg, enum cl_req_type crt)
 882 {
 883         /*
 884          * Page is queued for IO, change its state.
 885          */
 886         ENTRY;
 887         cl_page_owner_clear(pg);
 888         cl_page_state_set(env, pg, cl_req_type_state(crt));
 889         EXIT;
 890 }
 891
 892 /**
 893  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
 894  * called top-to-bottom. Every layer either agrees to submit this page (by
 895  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
 896  * handling interactions with the VM also has to inform VM that page is under
 897  * transfer now.
 898  */
 899 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
 900                  struct cl_page *pg, enum cl_req_type crt)
 901 {
 902         int result;
 903
 904         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 905         PINVRNT(env, pg, cl_page_invariant(pg));
 906         PINVRNT(env, pg, crt < CRT_NR);
 907
 908         /*
 909          * XXX this has to be called bottom-to-top, so that llite can set up
 910          * PG_writeback without risking other layers deciding to skip this
 911          * page.
 912          */
 913         if (crt >= CRT_NR)
 914                 return -EINVAL;
 915         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
 916         if (result == 0)
 917                 cl_page_io_start(env, pg, crt);
 918
 919         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
 920                       equi(result == 0,
 921                            PageWriteback(cl_page_vmpage(pg)))));
 922         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 923         return result;
 924 }
 925 EXPORT_SYMBOL(cl_page_prep);
 926
 927 /**
 928  * Notify layers about transfer completion.
 929  *
 930  * Invoked by transfer sub-system (which is a part of osc) to notify layers
 931  * that a transfer, of which this page is a part of has completed.
 932  *
 933  * Completion call-backs are executed in the bottom-up order, so that
 934  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
 935  * and can release locks safely.
 936  *
 937  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 938  * \post pg->cp_state == CPS_CACHED
 939  *
 940  * \see cl_page_operations::cpo_completion()
 941  */
 942 void cl_page_completion(const struct lu_env *env,
 943                         struct cl_page *pg, enum cl_req_type crt, int ioret)
 944 {
 945         struct cl_sync_io *anchor = pg->cp_sync_io;
 946
 947         PASSERT(env, pg, crt < CRT_NR);
 948         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
 949         PASSERT(env, pg, pg->cp_req == NULL);
 950         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
 951
 952         ENTRY;
 953         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
 954         cl_page_state_set(env, pg, CPS_CACHED);
 955         if (crt >= CRT_NR)
 956                 return;
 957         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
 958                                (const struct lu_env *,
 959                                 const struct cl_page_slice *, int), ioret);
 960         if (anchor) {
 961                 LASSERT(cl_page_is_vmlocked(env, pg));
 962                 LASSERT(pg->cp_sync_io == anchor);
 963                 pg->cp_sync_io = NULL;
 964         }
 965         /*
 966          * As page->cp_obj is pinned by a reference from page->cp_req, it is
 967          * safe to call cl_page_put() without risking object destruction in a
 968          * non-blocking context.
 969          */
 970         cl_page_put(env, pg);
 971
 972         if (anchor)
 973                 cl_sync_io_note(anchor, ioret);
 974
 975         EXIT;
 976 }
 977 EXPORT_SYMBOL(cl_page_completion);
 978
 979 /**
 980  * Notify layers that transfer formation engine decided to yank this page from
 981  * the cache and to make it a part of a transfer.
 982  *
 983  * \pre  pg->cp_state == CPS_CACHED
 984  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 985  *
 986  * \see cl_page_operations::cpo_make_ready()
 987  */
 988 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
 989                        enum cl_req_type crt)
 990 {
 991         int result;
 992
 993         PINVRNT(env, pg, crt < CRT_NR);
 994
 995         ENTRY;
 996         if (crt >= CRT_NR)
 997                 RETURN(-EINVAL);
 998         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
 999                                 (const struct lu_env *,
1000                                  const struct cl_page_slice *));
1001         if (result == 0) {
1002                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
1003                 cl_page_io_start(env, pg, crt);
1004         }
1005         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1006         RETURN(result);
1007 }
1008 EXPORT_SYMBOL(cl_page_make_ready);
1009
1010 /**
1011  * Called if a pge is being written back by kernel's intention.
1012  *
1013  * \pre  cl_page_is_owned(pg, io)
1014  * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT)
1015  *
1016  * \see cl_page_operations::cpo_flush()
1017  */
1018 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
1019                   struct cl_page *pg)
1020 {
1021         int result;
1022
1023         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1024         PINVRNT(env, pg, cl_page_invariant(pg));
1025
1026         ENTRY;
1027
1028         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_flush));
1029
1030         CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result);
1031         RETURN(result);
1032 }
1033 EXPORT_SYMBOL(cl_page_flush);
1034
1035 /**
1036  * Checks whether page is protected by any extent lock is at least required
1037  * mode.
1038  *
1039  * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1040  * \see cl_page_operations::cpo_is_under_lock()
1041  */
1042 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1043                           struct cl_page *page, pgoff_t *max_index)
1044 {
1045         int rc;
1046
1047         PINVRNT(env, page, cl_page_invariant(page));
1048
1049         ENTRY;
1050         rc = CL_PAGE_INVOKE_REVERSE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1051                                     (const struct lu_env *,
1052                                      const struct cl_page_slice *,
1053                                      struct cl_io *, pgoff_t *),
1054                                     io, max_index);
1055         RETURN(rc);
1056 }
1057 EXPORT_SYMBOL(cl_page_is_under_lock);
1058
1059 /**
1060  * Tells transfer engine that only part of a page is to be transmitted.
1061  *
1062  * \see cl_page_operations::cpo_clip()
1063  */
1064 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1065                   int from, int to)
1066 {
1067         PINVRNT(env, pg, cl_page_invariant(pg));
1068
1069         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1070         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1071                        (const struct lu_env *,
1072                         const struct cl_page_slice *,int, int),
1073                        from, to);
1074 }
1075 EXPORT_SYMBOL(cl_page_clip);
1076
1077 /**
1078  * Prints human readable representation of \a pg to the \a f.
1079  */
1080 void cl_page_header_print(const struct lu_env *env, void *cookie,
1081                           lu_printer_t printer, const struct cl_page *pg)
1082 {
1083         (*printer)(env, cookie,
1084                    "page@%p[%d %p %d %d %d %p %p]\n",
1085                    pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1086                    pg->cp_state, pg->cp_error, pg->cp_type,
1087                    pg->cp_owner, pg->cp_req);
1088 }
1089 EXPORT_SYMBOL(cl_page_header_print);
1090
1091 /**
1092  * Prints human readable representation of \a pg to the \a f.
1093  */
1094 void cl_page_print(const struct lu_env *env, void *cookie,
1095                    lu_printer_t printer, const struct cl_page *pg)
1096 {
1097         cl_page_header_print(env, cookie, printer, pg);
1098         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1099                        (const struct lu_env *env,
1100                         const struct cl_page_slice *slice,
1101                         void *cookie, lu_printer_t p), cookie, printer);
1102         (*printer)(env, cookie, "end page@%p\n", pg);
1103 }
1104 EXPORT_SYMBOL(cl_page_print);
1105
1106 /**
1107  * Cancel a page which is still in a transfer.
1108  */
1109 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1110 {
1111         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1112                               (const struct lu_env *,
1113                                const struct cl_page_slice *));
1114 }
1115 EXPORT_SYMBOL(cl_page_cancel);
1116
1117 /**
1118  * Converts a byte offset within object \a obj into a page index.
1119  */
1120 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1121 {
1122         return (loff_t)idx << PAGE_CACHE_SHIFT;
1123 }
1124 EXPORT_SYMBOL(cl_offset);
1125
1126 /**
1127  * Converts a page index into a byte offset within object \a obj.
1128  */
1129 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1130 {
1131         return offset >> PAGE_CACHE_SHIFT;
1132 }
1133 EXPORT_SYMBOL(cl_index);
1134
1135 int cl_page_size(const struct cl_object *obj)
1136 {
1137         return 1 << PAGE_CACHE_SHIFT;
1138 }
1139 EXPORT_SYMBOL(cl_page_size);
1140
1141 /**
1142  * Adds page slice to the compound page.
1143  *
1144  * This is called by cl_object_operations::coo_page_init() methods to add a
1145  * per-layer state to the page. New state is added at the end of
1146  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1147  *
1148  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1149  */
1150 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1151                        struct cl_object *obj, pgoff_t index,
1152                        const struct cl_page_operations *ops)
1153 {
1154         ENTRY;
1155         list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1156         slice->cpl_obj  = obj;
1157         slice->cpl_index = index;
1158         slice->cpl_ops  = ops;
1159         slice->cpl_page = page;
1160         EXIT;
1161 }
1162 EXPORT_SYMBOL(cl_page_slice_add);
1163
1164 int  cl_page_init(void)
1165 {
1166         return 0;
1167 }
1168
1169 void cl_page_fini(void)
1170 {
1171 }