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         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_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                 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                        atomic_read(&page->cp_ref));
 386
 387         if (atomic_dec_and_test(&page->cp_ref)) {
 388                 LASSERT(page->cp_state == CPS_FREEING);
 389
 390                 LASSERT(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_INVOKE_REVERSE(_env, _page, _op, _proto, ...)           \
 462 ({                                                                      \
 463         const struct lu_env        *__env  = (_env);                    \
 464         struct cl_page             *__page = (_page);                   \
 465         const struct cl_page_slice *__scan;                             \
 466         int                         __result;                           \
 467         ptrdiff_t                   __op   = (_op);                     \
 468         int                       (*__method)_proto;                    \
 469                                                                         \
 470         __result = 0;                                                   \
 471         list_for_each_entry_reverse(__scan, &__page->cp_layers,         \
 472                                         cpl_linkage) {                  \
 473                 __method = *(void **)((char *)__scan->cpl_ops +  __op); \
 474                 if (__method != NULL) {                                 \
 475                         __result = (*__method)(__env, __scan, ## __VA_ARGS__); \
 476                         if (__result != 0)                              \
 477                                 break;                                  \
 478                 }                                                       \
 479         }                                                               \
 480         if (__result > 0)                                               \
 481                 __result = 0;                                           \
 482         __result;                                                       \
 483 })
 484
 485 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 486 do {                                                                    \
 487         const struct lu_env        *__env  = (_env);                    \
 488         struct cl_page             *__page = (_page);                   \
 489         const struct cl_page_slice *__scan;                             \
 490         ptrdiff_t                   __op   = (_op);                     \
 491         void                      (*__method)_proto;                    \
 492                                                                         \
 493         cfs_list_for_each_entry(__scan, &__page->cp_layers,             \
 494                                 cpl_linkage) {                          \
 495                 __method = *(void **)((char *)__scan->cpl_ops +  __op); \
 496                 if (__method != NULL)                                   \
 497                         (*__method)(__env, __scan, ## __VA_ARGS__);     \
 498         }                                                               \
 499 } while (0)
 500
 501 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)           \
 502 do {                                                                    \
 503         const struct lu_env        *__env  = (_env);                    \
 504         struct cl_page             *__page = (_page);                   \
 505         const struct cl_page_slice *__scan;                             \
 506         ptrdiff_t                   __op   = (_op);                     \
 507         void                      (*__method)_proto;                    \
 508                                                                         \
 509         /* get to the bottom page. */                                   \
 510         cfs_list_for_each_entry_reverse(__scan, &__page->cp_layers,     \
 511                                         cpl_linkage) {                  \
 512                 __method = *(void **)((char *)__scan->cpl_ops + __op);  \
 513                 if (__method != NULL)                                   \
 514                         (*__method)(__env, __scan, ## __VA_ARGS__);     \
 515         }                                                               \
 516 } while (0)
 517
 518 static int cl_page_invoke(const struct lu_env *env,
 519                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 520
 521 {
 522         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 523         ENTRY;
 524         RETURN(CL_PAGE_INVOKE(env, page, op,
 525                               (const struct lu_env *,
 526                                const struct cl_page_slice *, struct cl_io *),
 527                               io));
 528 }
 529
 530 static void cl_page_invoid(const struct lu_env *env,
 531                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 532
 533 {
 534         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 535         ENTRY;
 536         CL_PAGE_INVOID(env, page, op,
 537                        (const struct lu_env *,
 538                         const struct cl_page_slice *, struct cl_io *), io);
 539         EXIT;
 540 }
 541
 542 static void cl_page_owner_clear(struct cl_page *page)
 543 {
 544         ENTRY;
 545         if (page->cp_owner != NULL) {
 546                 LASSERT(page->cp_owner->ci_owned_nr > 0);
 547                 page->cp_owner->ci_owned_nr--;
 548                 page->cp_owner = NULL;
 549                 page->cp_task = NULL;
 550         }
 551         EXIT;
 552 }
 553
 554 static void cl_page_owner_set(struct cl_page *page)
 555 {
 556         ENTRY;
 557         LASSERT(page->cp_owner != NULL);
 558         page->cp_owner->ci_owned_nr++;
 559         EXIT;
 560 }
 561
 562 void cl_page_disown0(const struct lu_env *env,
 563                      struct cl_io *io, struct cl_page *pg)
 564 {
 565         enum cl_page_state state;
 566
 567         ENTRY;
 568         state = pg->cp_state;
 569         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 570         PINVRNT(env, pg, cl_page_invariant(pg) || state == CPS_FREEING);
 571         cl_page_owner_clear(pg);
 572
 573         if (state == CPS_OWNED)
 574                 cl_page_state_set(env, pg, CPS_CACHED);
 575         /*
 576          * Completion call-backs are executed in the bottom-up order, so that
 577          * uppermost layer (llite), responsible for VFS/VM interaction runs
 578          * last and can release locks safely.
 579          */
 580         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 581                                (const struct lu_env *,
 582                                 const struct cl_page_slice *, struct cl_io *),
 583                                io);
 584         EXIT;
 585 }
 586
 587 /**
 588  * returns true, iff page is owned by the given io.
 589  */
 590 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 591 {
 592         struct cl_io *top = cl_io_top((struct cl_io *)io);
 593         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 594         ENTRY;
 595         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top);
 596 }
 597 EXPORT_SYMBOL(cl_page_is_owned);
 598
 599 /**
 600  * Try to own a page by IO.
 601  *
 602  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 603  * into cl_page_state::CPS_OWNED state.
 604  *
 605  * \pre  !cl_page_is_owned(pg, io)
 606  * \post result == 0 iff cl_page_is_owned(pg, io)
 607  *
 608  * \retval 0   success
 609  *
 610  * \retval -ve failure, e.g., page was destroyed (and landed in
 611  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 612  *             or, page was owned by another thread, or in IO.
 613  *
 614  * \see cl_page_disown()
 615  * \see cl_page_operations::cpo_own()
 616  * \see cl_page_own_try()
 617  * \see cl_page_own
 618  */
 619 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 620                         struct cl_page *pg, int nonblock)
 621 {
 622         int result;
 623
 624         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 625
 626         ENTRY;
 627         io = cl_io_top(io);
 628
 629         if (pg->cp_state == CPS_FREEING) {
 630                 result = -ENOENT;
 631         } else {
 632                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 633                                         (const struct lu_env *,
 634                                          const struct cl_page_slice *,
 635                                          struct cl_io *, int),
 636                                         io, nonblock);
 637                 if (result == 0) {
 638                         PASSERT(env, pg, pg->cp_owner == NULL);
 639                         PASSERT(env, pg, pg->cp_req == NULL);
 640                         pg->cp_owner = cl_io_top(io);;
 641                         pg->cp_task  = current;
 642                         cl_page_owner_set(pg);
 643                         if (pg->cp_state != CPS_FREEING) {
 644                                 cl_page_state_set(env, pg, CPS_OWNED);
 645                         } else {
 646                                 cl_page_disown0(env, io, pg);
 647                                 result = -ENOENT;
 648                         }
 649                 }
 650         }
 651         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 652         RETURN(result);
 653 }
 654
 655 /**
 656  * Own a page, might be blocked.
 657  *
 658  * \see cl_page_own0()
 659  */
 660 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 661 {
 662         return cl_page_own0(env, io, pg, 0);
 663 }
 664 EXPORT_SYMBOL(cl_page_own);
 665
 666 /**
 667  * Nonblock version of cl_page_own().
 668  *
 669  * \see cl_page_own0()
 670  */
 671 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
 672                     struct cl_page *pg)
 673 {
 674         return cl_page_own0(env, io, pg, 1);
 675 }
 676 EXPORT_SYMBOL(cl_page_own_try);
 677
 678
 679 /**
 680  * Assume page ownership.
 681  *
 682  * Called when page is already locked by the hosting VM.
 683  *
 684  * \pre !cl_page_is_owned(pg, io)
 685  * \post cl_page_is_owned(pg, io)
 686  *
 687  * \see cl_page_operations::cpo_assume()
 688  */
 689 void cl_page_assume(const struct lu_env *env,
 690                     struct cl_io *io, struct cl_page *pg)
 691 {
 692         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
 693
 694         ENTRY;
 695         io = cl_io_top(io);
 696
 697         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
 698         PASSERT(env, pg, pg->cp_owner == NULL);
 699         pg->cp_owner = cl_io_top(io);
 700         pg->cp_task = current;
 701         cl_page_owner_set(pg);
 702         cl_page_state_set(env, pg, CPS_OWNED);
 703         EXIT;
 704 }
 705 EXPORT_SYMBOL(cl_page_assume);
 706
 707 /**
 708  * Releases page ownership without unlocking the page.
 709  *
 710  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
 711  * underlying VM page (as VM is supposed to do this itself).
 712  *
 713  * \pre   cl_page_is_owned(pg, io)
 714  * \post !cl_page_is_owned(pg, io)
 715  *
 716  * \see cl_page_assume()
 717  */
 718 void cl_page_unassume(const struct lu_env *env,
 719                       struct cl_io *io, struct cl_page *pg)
 720 {
 721         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 722         PINVRNT(env, pg, cl_page_invariant(pg));
 723
 724         ENTRY;
 725         io = cl_io_top(io);
 726         cl_page_owner_clear(pg);
 727         cl_page_state_set(env, pg, CPS_CACHED);
 728         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
 729                                (const struct lu_env *,
 730                                 const struct cl_page_slice *, struct cl_io *),
 731                                io);
 732         EXIT;
 733 }
 734 EXPORT_SYMBOL(cl_page_unassume);
 735
 736 /**
 737  * Releases page ownership.
 738  *
 739  * Moves page into cl_page_state::CPS_CACHED.
 740  *
 741  * \pre   cl_page_is_owned(pg, io)
 742  * \post !cl_page_is_owned(pg, io)
 743  *
 744  * \see cl_page_own()
 745  * \see cl_page_operations::cpo_disown()
 746  */
 747 void cl_page_disown(const struct lu_env *env,
 748                     struct cl_io *io, struct cl_page *pg)
 749 {
 750         PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
 751                 pg->cp_state == CPS_FREEING);
 752
 753         ENTRY;
 754         io = cl_io_top(io);
 755         cl_page_disown0(env, io, pg);
 756         EXIT;
 757 }
 758 EXPORT_SYMBOL(cl_page_disown);
 759
 760 /**
 761  * Called when page is to be removed from the object, e.g., as a result of
 762  * truncate.
 763  *
 764  * Calls cl_page_operations::cpo_discard() top-to-bottom.
 765  *
 766  * \pre cl_page_is_owned(pg, io)
 767  *
 768  * \see cl_page_operations::cpo_discard()
 769  */
 770 void cl_page_discard(const struct lu_env *env,
 771                      struct cl_io *io, struct cl_page *pg)
 772 {
 773         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 774         PINVRNT(env, pg, cl_page_invariant(pg));
 775
 776         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
 777 }
 778 EXPORT_SYMBOL(cl_page_discard);
 779
 780 /**
 781  * Version of cl_page_delete() that can be called for not fully constructed
 782  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
 783  * path. Doesn't check page invariant.
 784  */
 785 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg)
 786 {
 787         ENTRY;
 788
 789         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
 790
 791         /*
 792          * Severe all ways to obtain new pointers to @pg.
 793          */
 794         cl_page_owner_clear(pg);
 795
 796         cl_page_state_set0(env, pg, CPS_FREEING);
 797
 798         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_delete),
 799                        (const struct lu_env *, const struct cl_page_slice *));
 800
 801         EXIT;
 802 }
 803
 804 /**
 805  * Called when a decision is made to throw page out of memory.
 806  *
 807  * Notifies all layers about page destruction by calling
 808  * cl_page_operations::cpo_delete() method top-to-bottom.
 809  *
 810  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
 811  * where transition to this state happens).
 812  *
 813  * Eliminates all venues through which new references to the page can be
 814  * obtained:
 815  *
 816  *     - removes page from the radix trees,
 817  *
 818  *     - breaks linkage from VM page to cl_page.
 819  *
 820  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
 821  * drain after some time, at which point page will be recycled.
 822  *
 823  * \pre  VM page is locked
 824  * \post pg->cp_state == CPS_FREEING
 825  *
 826  * \see cl_page_operations::cpo_delete()
 827  */
 828 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
 829 {
 830         PINVRNT(env, pg, cl_page_invariant(pg));
 831         ENTRY;
 832         cl_page_delete0(env, pg);
 833         EXIT;
 834 }
 835 EXPORT_SYMBOL(cl_page_delete);
 836
 837 /**
 838  * Marks page up-to-date.
 839  *
 840  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
 841  * layer responsible for VM interaction has to mark/clear page as up-to-date
 842  * by the \a uptodate argument.
 843  *
 844  * \see cl_page_operations::cpo_export()
 845  */
 846 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
 847 {
 848         PINVRNT(env, pg, cl_page_invariant(pg));
 849         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
 850                        (const struct lu_env *,
 851                         const struct cl_page_slice *, int), uptodate);
 852 }
 853 EXPORT_SYMBOL(cl_page_export);
 854
 855 /**
 856  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
 857  * thread.
 858  */
 859 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
 860 {
 861         int result;
 862         const struct cl_page_slice *slice;
 863
 864         ENTRY;
 865         slice = container_of(pg->cp_layers.next,
 866                              const struct cl_page_slice, cpl_linkage);
 867         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
 868         /*
 869          * Call ->cpo_is_vmlocked() directly instead of going through
 870          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
 871          * cl_page_invariant().
 872          */
 873         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
 874         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
 875         RETURN(result == -EBUSY);
 876 }
 877 EXPORT_SYMBOL(cl_page_is_vmlocked);
 878
 879 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
 880 {
 881         ENTRY;
 882         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
 883 }
 884
 885 static void cl_page_io_start(const struct lu_env *env,
 886                              struct cl_page *pg, enum cl_req_type crt)
 887 {
 888         /*
 889          * Page is queued for IO, change its state.
 890          */
 891         ENTRY;
 892         cl_page_owner_clear(pg);
 893         cl_page_state_set(env, pg, cl_req_type_state(crt));
 894         EXIT;
 895 }
 896
 897 /**
 898  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
 899  * called top-to-bottom. Every layer either agrees to submit this page (by
 900  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
 901  * handling interactions with the VM also has to inform VM that page is under
 902  * transfer now.
 903  */
 904 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
 905                  struct cl_page *pg, enum cl_req_type crt)
 906 {
 907         int result;
 908
 909         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 910         PINVRNT(env, pg, cl_page_invariant(pg));
 911         PINVRNT(env, pg, crt < CRT_NR);
 912
 913         /*
 914          * XXX this has to be called bottom-to-top, so that llite can set up
 915          * PG_writeback without risking other layers deciding to skip this
 916          * page.
 917          */
 918         if (crt >= CRT_NR)
 919                 return -EINVAL;
 920         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
 921         if (result == 0)
 922                 cl_page_io_start(env, pg, crt);
 923
 924         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
 925                       equi(result == 0,
 926                            PageWriteback(cl_page_vmpage(pg)))));
 927         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 928         return result;
 929 }
 930 EXPORT_SYMBOL(cl_page_prep);
 931
 932 /**
 933  * Notify layers about transfer completion.
 934  *
 935  * Invoked by transfer sub-system (which is a part of osc) to notify layers
 936  * that a transfer, of which this page is a part of has completed.
 937  *
 938  * Completion call-backs are executed in the bottom-up order, so that
 939  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
 940  * and can release locks safely.
 941  *
 942  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 943  * \post pg->cp_state == CPS_CACHED
 944  *
 945  * \see cl_page_operations::cpo_completion()
 946  */
 947 void cl_page_completion(const struct lu_env *env,
 948                         struct cl_page *pg, enum cl_req_type crt, int ioret)
 949 {
 950         struct cl_sync_io *anchor = pg->cp_sync_io;
 951
 952         PASSERT(env, pg, crt < CRT_NR);
 953         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
 954         PASSERT(env, pg, pg->cp_req == NULL);
 955         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
 956
 957         ENTRY;
 958         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
 959         if (crt == CRT_READ && ioret == 0) {
 960                 PASSERT(env, pg, !(pg->cp_flags & CPF_READ_COMPLETED));
 961                 pg->cp_flags |= CPF_READ_COMPLETED;
 962         }
 963
 964         cl_page_state_set(env, pg, CPS_CACHED);
 965         if (crt >= CRT_NR)
 966                 return;
 967         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
 968                                (const struct lu_env *,
 969                                 const struct cl_page_slice *, int), ioret);
 970         if (anchor) {
 971                 LASSERT(cl_page_is_vmlocked(env, pg));
 972                 LASSERT(pg->cp_sync_io == anchor);
 973                 pg->cp_sync_io = NULL;
 974         }
 975         /*
 976          * As page->cp_obj is pinned by a reference from page->cp_req, it is
 977          * safe to call cl_page_put() without risking object destruction in a
 978          * non-blocking context.
 979          */
 980         cl_page_put(env, pg);
 981
 982         if (anchor)
 983                 cl_sync_io_note(anchor, ioret);
 984
 985         EXIT;
 986 }
 987 EXPORT_SYMBOL(cl_page_completion);
 988
 989 /**
 990  * Notify layers that transfer formation engine decided to yank this page from
 991  * the cache and to make it a part of a transfer.
 992  *
 993  * \pre  pg->cp_state == CPS_CACHED
 994  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 995  *
 996  * \see cl_page_operations::cpo_make_ready()
 997  */
 998 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
 999                        enum cl_req_type crt)
1000 {
1001         int result;
1002
1003         PINVRNT(env, pg, crt < CRT_NR);
1004
1005         ENTRY;
1006         if (crt >= CRT_NR)
1007                 RETURN(-EINVAL);
1008         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
1009                                 (const struct lu_env *,
1010                                  const struct cl_page_slice *));
1011         if (result == 0) {
1012                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
1013                 cl_page_io_start(env, pg, crt);
1014         }
1015         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
1016         RETURN(result);
1017 }
1018 EXPORT_SYMBOL(cl_page_make_ready);
1019
1020 /**
1021  * Called if a pge is being written back by kernel's intention.
1022  *
1023  * \pre  cl_page_is_owned(pg, io)
1024  * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT)
1025  *
1026  * \see cl_page_operations::cpo_flush()
1027  */
1028 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
1029                   struct cl_page *pg)
1030 {
1031         int result;
1032
1033         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1034         PINVRNT(env, pg, cl_page_invariant(pg));
1035
1036         ENTRY;
1037
1038         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_flush));
1039
1040         CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result);
1041         RETURN(result);
1042 }
1043 EXPORT_SYMBOL(cl_page_flush);
1044
1045 /**
1046  * Checks whether page is protected by any extent lock is at least required
1047  * mode.
1048  *
1049  * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1050  * \see cl_page_operations::cpo_is_under_lock()
1051  */
1052 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1053                           struct cl_page *page, pgoff_t *max_index)
1054 {
1055         int rc;
1056
1057         PINVRNT(env, page, cl_page_invariant(page));
1058
1059         ENTRY;
1060         rc = CL_PAGE_INVOKE_REVERSE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1061                                     (const struct lu_env *,
1062                                      const struct cl_page_slice *,
1063                                      struct cl_io *, pgoff_t *),
1064                                     io, max_index);
1065         RETURN(rc);
1066 }
1067 EXPORT_SYMBOL(cl_page_is_under_lock);
1068
1069 /**
1070  * Tells transfer engine that only part of a page is to be transmitted.
1071  *
1072  * \see cl_page_operations::cpo_clip()
1073  */
1074 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1075                   int from, int to)
1076 {
1077         PINVRNT(env, pg, cl_page_invariant(pg));
1078
1079         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1080         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1081                        (const struct lu_env *,
1082                         const struct cl_page_slice *,int, int),
1083                        from, to);
1084 }
1085 EXPORT_SYMBOL(cl_page_clip);
1086
1087 /**
1088  * Prints human readable representation of \a pg to the \a f.
1089  */
1090 void cl_page_header_print(const struct lu_env *env, void *cookie,
1091                           lu_printer_t printer, const struct cl_page *pg)
1092 {
1093         (*printer)(env, cookie,
1094                    "page@%p[%d %p %d %d %d %p %p %#x]\n",
1095                    pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1096                    pg->cp_state, pg->cp_error, pg->cp_type,
1097                    pg->cp_owner, pg->cp_req, pg->cp_flags);
1098 }
1099 EXPORT_SYMBOL(cl_page_header_print);
1100
1101 /**
1102  * Prints human readable representation of \a pg to the \a f.
1103  */
1104 void cl_page_print(const struct lu_env *env, void *cookie,
1105                    lu_printer_t printer, const struct cl_page *pg)
1106 {
1107         cl_page_header_print(env, cookie, printer, pg);
1108         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1109                        (const struct lu_env *env,
1110                         const struct cl_page_slice *slice,
1111                         void *cookie, lu_printer_t p), cookie, printer);
1112         (*printer)(env, cookie, "end page@%p\n", pg);
1113 }
1114 EXPORT_SYMBOL(cl_page_print);
1115
1116 /**
1117  * Cancel a page which is still in a transfer.
1118  */
1119 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1120 {
1121         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1122                               (const struct lu_env *,
1123                                const struct cl_page_slice *));
1124 }
1125 EXPORT_SYMBOL(cl_page_cancel);
1126
1127 /**
1128  * Converts a byte offset within object \a obj into a page index.
1129  */
1130 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1131 {
1132         return (loff_t)idx << PAGE_CACHE_SHIFT;
1133 }
1134 EXPORT_SYMBOL(cl_offset);
1135
1136 /**
1137  * Converts a page index into a byte offset within object \a obj.
1138  */
1139 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1140 {
1141         return offset >> PAGE_CACHE_SHIFT;
1142 }
1143 EXPORT_SYMBOL(cl_index);
1144
1145 int cl_page_size(const struct cl_object *obj)
1146 {
1147         return 1 << PAGE_CACHE_SHIFT;
1148 }
1149 EXPORT_SYMBOL(cl_page_size);
1150
1151 /**
1152  * Adds page slice to the compound page.
1153  *
1154  * This is called by cl_object_operations::coo_page_init() methods to add a
1155  * per-layer state to the page. New state is added at the end of
1156  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1157  *
1158  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1159  */
1160 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1161                        struct cl_object *obj, pgoff_t index,
1162                        const struct cl_page_operations *ops)
1163 {
1164         ENTRY;
1165         list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1166         slice->cpl_obj  = obj;
1167         slice->cpl_index = index;
1168         slice->cpl_ops  = ops;
1169         slice->cpl_page = page;
1170         EXIT;
1171 }
1172 EXPORT_SYMBOL(cl_page_slice_add);
1173
1174 int  cl_page_init(void)
1175 {
1176         return 0;
1177 }
1178
1179 void cl_page_fini(void)
1180 {
1181 }