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