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, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Client Lustre Page.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_CLASS
  43
  44 #include <libcfs/libcfs.h>
  45 #include <obd_class.h>
  46 #include <obd_support.h>
  47 #include <libcfs/list.h>
  48
  49 #include <cl_object.h>
  50 #include "cl_internal.h"
  51
  52 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg);
  53
  54 #ifdef LIBCFS_DEBUG
  55 # define PASSERT(env, page, expr)                                       \
  56   do {                                                                    \
  57           if (unlikely(!(expr))) {                                      \
  58                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  59                   LASSERT(0);                                           \
  60           }                                                             \
  61   } while (0)
  62 #else /* !LIBCFS_DEBUG */
  63 # define PASSERT(env, page, exp) \
  64         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  65 #endif /* !LIBCFS_DEBUG */
  66
  67 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
  68 # define PINVRNT(env, page, expr)                                       \
  69   do {                                                                    \
  70           if (unlikely(!(expr))) {                                      \
  71                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  72                   LINVRNT(0);                                           \
  73           }                                                             \
  74   } while (0)
  75 #else /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
  76 # define PINVRNT(env, page, exp) \
  77          ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  78 #endif /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
  79
  80 /* Disable page statistic by default due to huge performance penalty. */
  81 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
  82 #define CS_PAGE_INC(o, item) \
  83         atomic_inc(&cl_object_site(o)->cs_pages.cs_stats[CS_##item])
  84 #define CS_PAGE_DEC(o, item) \
  85         atomic_dec(&cl_object_site(o)->cs_pages.cs_stats[CS_##item])
  86 #define CS_PAGESTATE_INC(o, state) \
  87         atomic_inc(&cl_object_site(o)->cs_pages_state[state])
  88 #define CS_PAGESTATE_DEC(o, state) \
  89         atomic_dec(&cl_object_site(o)->cs_pages_state[state])
  90 #else
  91 #define CS_PAGE_INC(o, item)
  92 #define CS_PAGE_DEC(o, item)
  93 #define CS_PAGESTATE_INC(o, state)
  94 #define CS_PAGESTATE_DEC(o, state)
  95 #endif
  96
  97 /**
  98  * Internal version of cl_page_get().
  99  *
 100  * This function can be used to obtain initial reference to previously
 101  * unreferenced cached object. It can be called only if concurrent page
 102  * reclamation is somehow prevented, e.g., by keeping a lock on a VM page,
 103  * associated with \a page.
 104  *
 105  * Use with care! Not exported.
 106  */
 107 static void cl_page_get_trust(struct cl_page *page)
 108 {
 109         LASSERT(atomic_read(&page->cp_ref) > 0);
 110         atomic_inc(&page->cp_ref);
 111 }
 112
 113 /**
 114  * Returns a slice within a page, corresponding to the given layer in the
 115  * device stack.
 116  *
 117  * \see cl_lock_at()
 118  */
 119 static const struct cl_page_slice *
 120 cl_page_at_trusted(const struct cl_page *page,
 121                    const struct lu_device_type *dtype)
 122 {
 123         const struct cl_page_slice *slice;
 124         ENTRY;
 125
 126         list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 127                 if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
 128                         RETURN(slice);
 129         }
 130         RETURN(NULL);
 131 }
 132
 133 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
 134 {
 135         struct cl_object *obj  = page->cp_obj;
 136         int pagesize = cl_object_header(obj)->coh_page_bufsize;
 137
 138         PASSERT(env, page, list_empty(&page->cp_batch));
 139         PASSERT(env, page, page->cp_owner == NULL);
 140         PASSERT(env, page, page->cp_req == NULL);
 141         PASSERT(env, page, page->cp_state == CPS_FREEING);
 142
 143         ENTRY;
 144         while (!list_empty(&page->cp_layers)) {
 145                 struct cl_page_slice *slice;
 146
 147                 slice = list_entry(page->cp_layers.next,
 148                                    struct cl_page_slice, cpl_linkage);
 149                 list_del_init(page->cp_layers.next);
 150                 if (unlikely(slice->cpl_ops->cpo_fini != NULL))
 151                         slice->cpl_ops->cpo_fini(env, slice);
 152         }
 153         CS_PAGE_DEC(obj, total);
 154         CS_PAGESTATE_DEC(obj, page->cp_state);
 155         lu_object_ref_del_at(&obj->co_lu, &page->cp_obj_ref, "cl_page", page);
 156         cl_object_put(env, obj);
 157         lu_ref_fini(&page->cp_reference);
 158         OBD_FREE(page, pagesize);
 159         EXIT;
 160 }
 161
 162 /**
 163  * Helper function updating page state. This is the only place in the code
 164  * where cl_page::cp_state field is mutated.
 165  */
 166 static inline void cl_page_state_set_trust(struct cl_page *page,
 167                                            enum cl_page_state state)
 168 {
 169         /* bypass const. */
 170         *(enum cl_page_state *)&page->cp_state = state;
 171 }
 172
 173 struct cl_page *cl_page_alloc(const struct lu_env *env,
 174                 struct cl_object *o, pgoff_t ind, struct page *vmpage,
 175                 enum cl_page_type type)
 176 {
 177         struct cl_page          *page;
 178         struct lu_object_header *head;
 179
 180         ENTRY;
 181         OBD_ALLOC_GFP(page, cl_object_header(o)->coh_page_bufsize,
 182                         GFP_NOFS);
 183         if (page != NULL) {
 184                 int result = 0;
 185                 atomic_set(&page->cp_ref, 1);
 186                 page->cp_obj = o;
 187                 cl_object_get(o);
 188                 lu_object_ref_add_at(&o->co_lu, &page->cp_obj_ref, "cl_page",
 189                                      page);
 190                 page->cp_vmpage = vmpage;
 191                 cl_page_state_set_trust(page, CPS_CACHED);
 192                 page->cp_type = type;
 193                 INIT_LIST_HEAD(&page->cp_layers);
 194                 INIT_LIST_HEAD(&page->cp_batch);
 195                 INIT_LIST_HEAD(&page->cp_flight);
 196                 lu_ref_init(&page->cp_reference);
 197                 head = o->co_lu.lo_header;
 198                 list_for_each_entry(o, &head->loh_layers,
 199                                     co_lu.lo_linkage) {
 200                         if (o->co_ops->coo_page_init != NULL) {
 201                                 result = o->co_ops->coo_page_init(env, o, page,
 202                                                                   ind);
 203                                 if (result != 0) {
 204                                         cl_page_delete0(env, page);
 205                                         cl_page_free(env, page);
 206                                         page = ERR_PTR(result);
 207                                         break;
 208                                 }
 209                         }
 210                 }
 211                 if (result == 0) {
 212                         CS_PAGE_INC(o, total);
 213                         CS_PAGE_INC(o, create);
 214                         CS_PAGESTATE_DEC(o, CPS_CACHED);
 215                 }
 216         } else {
 217                 page = ERR_PTR(-ENOMEM);
 218         }
 219         RETURN(page);
 220 }
 221
 222 /**
 223  * Returns a cl_page with index \a idx at the object \a o, and associated with
 224  * the VM page \a vmpage.
 225  *
 226  * This is the main entry point into the cl_page caching interface. First, a
 227  * cache (implemented as a per-object radix tree) is consulted. If page is
 228  * found there, it is returned immediately. Otherwise new page is allocated
 229  * and returned. In any case, additional reference to page is acquired.
 230  *
 231  * \see cl_object_find(), cl_lock_find()
 232  */
 233 struct cl_page *cl_page_find(const struct lu_env *env,
 234                              struct cl_object *o,
 235                              pgoff_t idx, struct page *vmpage,
 236                              enum cl_page_type type)
 237 {
 238         struct cl_page          *page = NULL;
 239         struct cl_object_header *hdr;
 240
 241         LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 242         might_sleep();
 243
 244         ENTRY;
 245
 246         hdr = cl_object_header(o);
 247         CS_PAGE_INC(o, lookup);
 248
 249         CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
 250                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 251         /* fast path. */
 252         if (type == CPT_CACHEABLE) {
 253                 /* vmpage lock is used to protect the child/parent
 254                  * relationship */
 255                 KLASSERT(PageLocked(vmpage));
 256                 /*
 257                  * cl_vmpage_page() can be called here without any locks as
 258                  *
 259                  *     - "vmpage" is locked (which prevents ->private from
 260                  *       concurrent updates), and
 261                  *
 262                  *     - "o" cannot be destroyed while current thread holds a
 263                  *       reference on it.
 264                  */
 265                 page = cl_vmpage_page(vmpage, o);
 266                 if (page != NULL) {
 267                         CS_PAGE_INC(o, hit);
 268                         RETURN(page);
 269                 }
 270         }
 271
 272         /* allocate and initialize cl_page */
 273         page = cl_page_alloc(env, o, idx, vmpage, type);
 274         RETURN(page);
 275 }
 276 EXPORT_SYMBOL(cl_page_find);
 277
 278 static inline int cl_page_invariant(const struct cl_page *pg)
 279 {
 280         return cl_page_in_use_noref(pg);
 281 }
 282
 283 static void cl_page_state_set0(const struct lu_env *env,
 284                                struct cl_page *page, enum cl_page_state state)
 285 {
 286         enum cl_page_state old;
 287
 288         /*
 289          * Matrix of allowed state transitions [old][new], for sanity
 290          * checking.
 291          */
 292         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 293                 [CPS_CACHED] = {
 294                         [CPS_CACHED]  = 0,
 295                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 296                         [CPS_PAGEIN]  = 0,
 297                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 298                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 299                 },
 300                 [CPS_OWNED] = {
 301                         [CPS_CACHED]  = 1, /* release to the cache */
 302                         [CPS_OWNED]   = 0,
 303                         [CPS_PAGEIN]  = 1, /* start read immediately */
 304                         [CPS_PAGEOUT] = 1, /* start write immediately */
 305                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 306                 },
 307                 [CPS_PAGEIN] = {
 308                         [CPS_CACHED]  = 1, /* io completion */
 309                         [CPS_OWNED]   = 0,
 310                         [CPS_PAGEIN]  = 0,
 311                         [CPS_PAGEOUT] = 0,
 312                         [CPS_FREEING] = 0,
 313                 },
 314                 [CPS_PAGEOUT] = {
 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_FREEING] = {
 322                         [CPS_CACHED]  = 0,
 323                         [CPS_OWNED]   = 0,
 324                         [CPS_PAGEIN]  = 0,
 325                         [CPS_PAGEOUT] = 0,
 326                         [CPS_FREEING] = 0,
 327                 }
 328         };
 329
 330         ENTRY;
 331         old = page->cp_state;
 332         PASSERT(env, page, allowed_transitions[old][state]);
 333         CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
 334         PASSERT(env, page, page->cp_state == old);
 335         PASSERT(env, page, equi(state == CPS_OWNED, page->cp_owner != NULL));
 336
 337         CS_PAGESTATE_DEC(page->cp_obj, page->cp_state);
 338         CS_PAGESTATE_INC(page->cp_obj, state);
 339         cl_page_state_set_trust(page, state);
 340         EXIT;
 341 }
 342
 343 static void cl_page_state_set(const struct lu_env *env,
 344                               struct cl_page *page, enum cl_page_state state)
 345 {
 346         cl_page_state_set0(env, page, state);
 347 }
 348
 349 /**
 350  * Acquires an additional reference to a page.
 351  *
 352  * This can be called only by caller already possessing a reference to \a
 353  * page.
 354  *
 355  * \see cl_object_get(), cl_lock_get().
 356  */
 357 void cl_page_get(struct cl_page *page)
 358 {
 359         ENTRY;
 360         cl_page_get_trust(page);
 361         EXIT;
 362 }
 363 EXPORT_SYMBOL(cl_page_get);
 364
 365 /**
 366  * Releases a reference to a page.
 367  *
 368  * When last reference is released, page is returned to the cache, unless it
 369  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 370  * destroyed.
 371  *
 372  * \see cl_object_put(), cl_lock_put().
 373  */
 374 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 375 {
 376         ENTRY;
 377         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 378                        atomic_read(&page->cp_ref));
 379
 380         if (atomic_dec_and_test(&page->cp_ref)) {
 381                 LASSERT(page->cp_state == CPS_FREEING);
 382
 383                 LASSERT(atomic_read(&page->cp_ref) == 0);
 384                 PASSERT(env, page, page->cp_owner == NULL);
 385                 PASSERT(env, page, list_empty(&page->cp_batch));
 386                 /*
 387                  * Page is no longer reachable by other threads. Tear
 388                  * it down.
 389                  */
 390                 cl_page_free(env, page);
 391         }
 392
 393         EXIT;
 394 }
 395 EXPORT_SYMBOL(cl_page_put);
 396
 397 /**
 398  * Returns a cl_page associated with a VM page, and given cl_object.
 399  */
 400 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj)
 401 {
 402         struct cl_page *page;
 403
 404         ENTRY;
 405         KLASSERT(PageLocked(vmpage));
 406
 407         /*
 408          * NOTE: absence of races and liveness of data are guaranteed by page
 409          *       lock on a "vmpage". That works because object destruction has
 410          *       bottom-to-top pass.
 411          */
 412
 413         page = (struct cl_page *)vmpage->private;
 414         if (page != NULL) {
 415                 cl_page_get_trust(page);
 416                 LASSERT(page->cp_type == CPT_CACHEABLE);
 417         }
 418         RETURN(page);
 419 }
 420 EXPORT_SYMBOL(cl_vmpage_page);
 421
 422 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 423                                        const struct lu_device_type *dtype)
 424 {
 425         return cl_page_at_trusted(page, dtype);
 426 }
 427 EXPORT_SYMBOL(cl_page_at);
 428
 429 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 430
 431 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)                   \
 432 ({                                                                      \
 433         const struct lu_env        *__env  = (_env);                    \
 434         struct cl_page             *__page = (_page);                   \
 435         const struct cl_page_slice *__scan;                             \
 436         int                         __result;                           \
 437         ptrdiff_t                   __op   = (_op);                     \
 438         int                        (*__method)_proto;                   \
 439                                                                         \
 440         __result = 0;                                                   \
 441         list_for_each_entry(__scan, &__page->cp_layers, cpl_linkage) {  \
 442                 __method = *(void **)((char *)__scan->cpl_ops +  __op); \
 443                 if (__method != NULL) {                                 \
 444                         __result = (*__method)(__env, __scan, ## __VA_ARGS__); \
 445                         if (__result != 0)                              \
 446                                 break;                                  \
 447                 }                                                       \
 448         }                                                               \
 449         if (__result > 0)                                               \
 450                 __result = 0;                                           \
 451         __result;                                                       \
 452 })
 453
 454 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 455 do {                                                                    \
 456         const struct lu_env        *__env  = (_env);                    \
 457         struct cl_page             *__page = (_page);                   \
 458         const struct cl_page_slice *__scan;                             \
 459         ptrdiff_t                   __op   = (_op);                     \
 460         void                      (*__method)_proto;                    \
 461                                                                         \
 462         list_for_each_entry(__scan, &__page->cp_layers, cpl_linkage) {  \
 463                 __method = *(void **)((char *)__scan->cpl_ops +  __op); \
 464                 if (__method != NULL)                                   \
 465                         (*__method)(__env, __scan, ## __VA_ARGS__);     \
 466         }                                                               \
 467 } while (0)
 468
 469 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)           \
 470 do {                                                                    \
 471         const struct lu_env        *__env  = (_env);                    \
 472         struct cl_page             *__page = (_page);                   \
 473         const struct cl_page_slice *__scan;                             \
 474         ptrdiff_t                   __op   = (_op);                     \
 475         void                      (*__method)_proto;                    \
 476                                                                         \
 477         /* get to the bottom page. */                                   \
 478         list_for_each_entry_reverse(__scan, &__page->cp_layers,         \
 479                                     cpl_linkage) {                      \
 480                 __method = *(void **)((char *)__scan->cpl_ops + __op);  \
 481                 if (__method != NULL)                                   \
 482                         (*__method)(__env, __scan, ## __VA_ARGS__);     \
 483         }                                                               \
 484 } while (0)
 485
 486 static int cl_page_invoke(const struct lu_env *env,
 487                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 488
 489 {
 490         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 491         ENTRY;
 492         RETURN(CL_PAGE_INVOKE(env, page, op,
 493                               (const struct lu_env *,
 494                                const struct cl_page_slice *, struct cl_io *),
 495                               io));
 496 }
 497
 498 static void cl_page_invoid(const struct lu_env *env,
 499                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 500
 501 {
 502         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 503         ENTRY;
 504         CL_PAGE_INVOID(env, page, op,
 505                        (const struct lu_env *,
 506                         const struct cl_page_slice *, struct cl_io *), io);
 507         EXIT;
 508 }
 509
 510 static void cl_page_owner_clear(struct cl_page *page)
 511 {
 512         ENTRY;
 513         if (page->cp_owner != NULL) {
 514                 LASSERT(page->cp_owner->ci_owned_nr > 0);
 515                 page->cp_owner->ci_owned_nr--;
 516                 page->cp_owner = NULL;
 517         }
 518         EXIT;
 519 }
 520
 521 static void cl_page_owner_set(struct cl_page *page)
 522 {
 523         ENTRY;
 524         LASSERT(page->cp_owner != NULL);
 525         page->cp_owner->ci_owned_nr++;
 526         EXIT;
 527 }
 528
 529 void cl_page_disown0(const struct lu_env *env,
 530                      struct cl_io *io, struct cl_page *pg)
 531 {
 532         enum cl_page_state state;
 533
 534         ENTRY;
 535         state = pg->cp_state;
 536         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 537         PINVRNT(env, pg, cl_page_invariant(pg) || state == CPS_FREEING);
 538         cl_page_owner_clear(pg);
 539
 540         if (state == CPS_OWNED)
 541                 cl_page_state_set(env, pg, CPS_CACHED);
 542         /*
 543          * Completion call-backs are executed in the bottom-up order, so that
 544          * uppermost layer (llite), responsible for VFS/VM interaction runs
 545          * last and can release locks safely.
 546          */
 547         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 548                                (const struct lu_env *,
 549                                 const struct cl_page_slice *, struct cl_io *),
 550                                io);
 551         EXIT;
 552 }
 553
 554 /**
 555  * returns true, iff page is owned by the given io.
 556  */
 557 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 558 {
 559         struct cl_io *top = cl_io_top((struct cl_io *)io);
 560         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 561         ENTRY;
 562         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top);
 563 }
 564 EXPORT_SYMBOL(cl_page_is_owned);
 565
 566 /**
 567  * Try to own a page by IO.
 568  *
 569  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 570  * into cl_page_state::CPS_OWNED state.
 571  *
 572  * \pre  !cl_page_is_owned(pg, io)
 573  * \post result == 0 iff cl_page_is_owned(pg, io)
 574  *
 575  * \retval 0   success
 576  *
 577  * \retval -ve failure, e.g., page was destroyed (and landed in
 578  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 579  *             or, page was owned by another thread, or in IO.
 580  *
 581  * \see cl_page_disown()
 582  * \see cl_page_operations::cpo_own()
 583  * \see cl_page_own_try()
 584  * \see cl_page_own
 585  */
 586 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 587                         struct cl_page *pg, int nonblock)
 588 {
 589         int result;
 590
 591         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 592
 593         ENTRY;
 594         io = cl_io_top(io);
 595
 596         if (pg->cp_state == CPS_FREEING) {
 597                 result = -ENOENT;
 598         } else {
 599                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 600                                         (const struct lu_env *,
 601                                          const struct cl_page_slice *,
 602                                          struct cl_io *, int),
 603                                         io, nonblock);
 604                 if (result == 0) {
 605                         PASSERT(env, pg, pg->cp_owner == NULL);
 606                         PASSERT(env, pg, pg->cp_req == NULL);
 607                         pg->cp_owner = cl_io_top(io);;
 608                         cl_page_owner_set(pg);
 609                         if (pg->cp_state != CPS_FREEING) {
 610                                 cl_page_state_set(env, pg, CPS_OWNED);
 611                         } else {
 612                                 cl_page_disown0(env, io, pg);
 613                                 result = -ENOENT;
 614                         }
 615                 }
 616         }
 617         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 618         RETURN(result);
 619 }
 620
 621 /**
 622  * Own a page, might be blocked.
 623  *
 624  * \see cl_page_own0()
 625  */
 626 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 627 {
 628         return cl_page_own0(env, io, pg, 0);
 629 }
 630 EXPORT_SYMBOL(cl_page_own);
 631
 632 /**
 633  * Nonblock version of cl_page_own().
 634  *
 635  * \see cl_page_own0()
 636  */
 637 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
 638                     struct cl_page *pg)
 639 {
 640         return cl_page_own0(env, io, pg, 1);
 641 }
 642 EXPORT_SYMBOL(cl_page_own_try);
 643
 644
 645 /**
 646  * Assume page ownership.
 647  *
 648  * Called when page is already locked by the hosting VM.
 649  *
 650  * \pre !cl_page_is_owned(pg, io)
 651  * \post cl_page_is_owned(pg, io)
 652  *
 653  * \see cl_page_operations::cpo_assume()
 654  */
 655 void cl_page_assume(const struct lu_env *env,
 656                     struct cl_io *io, struct cl_page *pg)
 657 {
 658         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
 659
 660         ENTRY;
 661         io = cl_io_top(io);
 662
 663         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
 664         PASSERT(env, pg, pg->cp_owner == NULL);
 665         pg->cp_owner = cl_io_top(io);
 666         cl_page_owner_set(pg);
 667         cl_page_state_set(env, pg, CPS_OWNED);
 668         EXIT;
 669 }
 670 EXPORT_SYMBOL(cl_page_assume);
 671
 672 /**
 673  * Releases page ownership without unlocking the page.
 674  *
 675  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
 676  * underlying VM page (as VM is supposed to do this itself).
 677  *
 678  * \pre   cl_page_is_owned(pg, io)
 679  * \post !cl_page_is_owned(pg, io)
 680  *
 681  * \see cl_page_assume()
 682  */
 683 void cl_page_unassume(const struct lu_env *env,
 684                       struct cl_io *io, struct cl_page *pg)
 685 {
 686         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 687         PINVRNT(env, pg, cl_page_invariant(pg));
 688
 689         ENTRY;
 690         io = cl_io_top(io);
 691         cl_page_owner_clear(pg);
 692         cl_page_state_set(env, pg, CPS_CACHED);
 693         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
 694                                (const struct lu_env *,
 695                                 const struct cl_page_slice *, struct cl_io *),
 696                                io);
 697         EXIT;
 698 }
 699 EXPORT_SYMBOL(cl_page_unassume);
 700
 701 /**
 702  * Releases page ownership.
 703  *
 704  * Moves page into cl_page_state::CPS_CACHED.
 705  *
 706  * \pre   cl_page_is_owned(pg, io)
 707  * \post !cl_page_is_owned(pg, io)
 708  *
 709  * \see cl_page_own()
 710  * \see cl_page_operations::cpo_disown()
 711  */
 712 void cl_page_disown(const struct lu_env *env,
 713                     struct cl_io *io, struct cl_page *pg)
 714 {
 715         PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
 716                 pg->cp_state == CPS_FREEING);
 717
 718         ENTRY;
 719         io = cl_io_top(io);
 720         cl_page_disown0(env, io, pg);
 721         EXIT;
 722 }
 723 EXPORT_SYMBOL(cl_page_disown);
 724
 725 /**
 726  * Called when page is to be removed from the object, e.g., as a result of
 727  * truncate.
 728  *
 729  * Calls cl_page_operations::cpo_discard() top-to-bottom.
 730  *
 731  * \pre cl_page_is_owned(pg, io)
 732  *
 733  * \see cl_page_operations::cpo_discard()
 734  */
 735 void cl_page_discard(const struct lu_env *env,
 736                      struct cl_io *io, struct cl_page *pg)
 737 {
 738         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 739         PINVRNT(env, pg, cl_page_invariant(pg));
 740
 741         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
 742 }
 743 EXPORT_SYMBOL(cl_page_discard);
 744
 745 /**
 746  * Version of cl_page_delete() that can be called for not fully constructed
 747  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
 748  * path. Doesn't check page invariant.
 749  */
 750 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg)
 751 {
 752         ENTRY;
 753
 754         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
 755
 756         /*
 757          * Severe all ways to obtain new pointers to @pg.
 758          */
 759         cl_page_owner_clear(pg);
 760
 761         cl_page_state_set0(env, pg, CPS_FREEING);
 762
 763         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_delete),
 764                        (const struct lu_env *, const struct cl_page_slice *));
 765
 766         EXIT;
 767 }
 768
 769 /**
 770  * Called when a decision is made to throw page out of memory.
 771  *
 772  * Notifies all layers about page destruction by calling
 773  * cl_page_operations::cpo_delete() method top-to-bottom.
 774  *
 775  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
 776  * where transition to this state happens).
 777  *
 778  * Eliminates all venues through which new references to the page can be
 779  * obtained:
 780  *
 781  *     - removes page from the radix trees,
 782  *
 783  *     - breaks linkage from VM page to cl_page.
 784  *
 785  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
 786  * drain after some time, at which point page will be recycled.
 787  *
 788  * \pre  VM page is locked
 789  * \post pg->cp_state == CPS_FREEING
 790  *
 791  * \see cl_page_operations::cpo_delete()
 792  */
 793 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
 794 {
 795         PINVRNT(env, pg, cl_page_invariant(pg));
 796         ENTRY;
 797         cl_page_delete0(env, pg);
 798         EXIT;
 799 }
 800 EXPORT_SYMBOL(cl_page_delete);
 801
 802 /**
 803  * Marks page up-to-date.
 804  *
 805  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
 806  * layer responsible for VM interaction has to mark/clear page as up-to-date
 807  * by the \a uptodate argument.
 808  *
 809  * \see cl_page_operations::cpo_export()
 810  */
 811 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
 812 {
 813         PINVRNT(env, pg, cl_page_invariant(pg));
 814         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
 815                        (const struct lu_env *,
 816                         const struct cl_page_slice *, int), uptodate);
 817 }
 818 EXPORT_SYMBOL(cl_page_export);
 819
 820 /**
 821  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
 822  * thread.
 823  */
 824 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
 825 {
 826         int result;
 827         const struct cl_page_slice *slice;
 828
 829         ENTRY;
 830         slice = container_of(pg->cp_layers.next,
 831                              const struct cl_page_slice, cpl_linkage);
 832         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
 833         /*
 834          * Call ->cpo_is_vmlocked() directly instead of going through
 835          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
 836          * cl_page_invariant().
 837          */
 838         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
 839         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
 840         RETURN(result == -EBUSY);
 841 }
 842 EXPORT_SYMBOL(cl_page_is_vmlocked);
 843
 844 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
 845 {
 846         ENTRY;
 847         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
 848 }
 849
 850 static void cl_page_io_start(const struct lu_env *env,
 851                              struct cl_page *pg, enum cl_req_type crt)
 852 {
 853         /*
 854          * Page is queued for IO, change its state.
 855          */
 856         ENTRY;
 857         cl_page_owner_clear(pg);
 858         cl_page_state_set(env, pg, cl_req_type_state(crt));
 859         EXIT;
 860 }
 861
 862 /**
 863  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
 864  * called top-to-bottom. Every layer either agrees to submit this page (by
 865  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
 866  * handling interactions with the VM also has to inform VM that page is under
 867  * transfer now.
 868  */
 869 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
 870                  struct cl_page *pg, enum cl_req_type crt)
 871 {
 872         int result;
 873
 874         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 875         PINVRNT(env, pg, cl_page_invariant(pg));
 876         PINVRNT(env, pg, crt < CRT_NR);
 877
 878         /*
 879          * XXX this has to be called bottom-to-top, so that llite can set up
 880          * PG_writeback without risking other layers deciding to skip this
 881          * page.
 882          */
 883         if (crt >= CRT_NR)
 884                 return -EINVAL;
 885         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
 886         if (result == 0)
 887                 cl_page_io_start(env, pg, crt);
 888
 889         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
 890                       equi(result == 0,
 891                            PageWriteback(cl_page_vmpage(pg)))));
 892         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 893         return result;
 894 }
 895 EXPORT_SYMBOL(cl_page_prep);
 896
 897 /**
 898  * Notify layers about transfer completion.
 899  *
 900  * Invoked by transfer sub-system (which is a part of osc) to notify layers
 901  * that a transfer, of which this page is a part of has completed.
 902  *
 903  * Completion call-backs are executed in the bottom-up order, so that
 904  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
 905  * and can release locks safely.
 906  *
 907  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 908  * \post pg->cp_state == CPS_CACHED
 909  *
 910  * \see cl_page_operations::cpo_completion()
 911  */
 912 void cl_page_completion(const struct lu_env *env,
 913                         struct cl_page *pg, enum cl_req_type crt, int ioret)
 914 {
 915         struct cl_sync_io *anchor = pg->cp_sync_io;
 916
 917         PASSERT(env, pg, crt < CRT_NR);
 918         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
 919         PASSERT(env, pg, pg->cp_req == NULL);
 920         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
 921
 922         ENTRY;
 923         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
 924         cl_page_state_set(env, pg, CPS_CACHED);
 925         if (crt >= CRT_NR)
 926                 return;
 927         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
 928                                (const struct lu_env *,
 929                                 const struct cl_page_slice *, int), ioret);
 930         if (anchor) {
 931                 LASSERT(pg->cp_sync_io == anchor);
 932                 pg->cp_sync_io = NULL;
 933         }
 934         /*
 935          * As page->cp_obj is pinned by a reference from page->cp_req, it is
 936          * safe to call cl_page_put() without risking object destruction in a
 937          * non-blocking context.
 938          */
 939         cl_page_put(env, pg);
 940
 941         if (anchor != NULL)
 942                 cl_sync_io_note(env, anchor, ioret);
 943
 944         EXIT;
 945 }
 946 EXPORT_SYMBOL(cl_page_completion);
 947
 948 /**
 949  * Notify layers that transfer formation engine decided to yank this page from
 950  * the cache and to make it a part of a transfer.
 951  *
 952  * \pre  pg->cp_state == CPS_CACHED
 953  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 954  *
 955  * \see cl_page_operations::cpo_make_ready()
 956  */
 957 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
 958                        enum cl_req_type crt)
 959 {
 960         int result;
 961
 962         PINVRNT(env, pg, crt < CRT_NR);
 963
 964         ENTRY;
 965         if (crt >= CRT_NR)
 966                 RETURN(-EINVAL);
 967         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
 968                                 (const struct lu_env *,
 969                                  const struct cl_page_slice *));
 970         if (result == 0) {
 971                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
 972                 cl_page_io_start(env, pg, crt);
 973         }
 974         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 975         RETURN(result);
 976 }
 977 EXPORT_SYMBOL(cl_page_make_ready);
 978
 979 /**
 980  * Called if a pge is being written back by kernel's intention.
 981  *
 982  * \pre  cl_page_is_owned(pg, io)
 983  * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT)
 984  *
 985  * \see cl_page_operations::cpo_flush()
 986  */
 987 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
 988                   struct cl_page *pg)
 989 {
 990         int result;
 991
 992         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 993         PINVRNT(env, pg, cl_page_invariant(pg));
 994
 995         ENTRY;
 996
 997         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_flush));
 998
 999         CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result);
1000         RETURN(result);
1001 }
1002 EXPORT_SYMBOL(cl_page_flush);
1003
1004 /**
1005  * Tells transfer engine that only part of a page is to be transmitted.
1006  *
1007  * \see cl_page_operations::cpo_clip()
1008  */
1009 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1010                   int from, int to)
1011 {
1012         PINVRNT(env, pg, cl_page_invariant(pg));
1013
1014         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1015         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1016                        (const struct lu_env *,
1017                         const struct cl_page_slice *,int, int),
1018                        from, to);
1019 }
1020 EXPORT_SYMBOL(cl_page_clip);
1021
1022 /**
1023  * Prints human readable representation of \a pg to the \a f.
1024  */
1025 void cl_page_header_print(const struct lu_env *env, void *cookie,
1026                           lu_printer_t printer, const struct cl_page *pg)
1027 {
1028         (*printer)(env, cookie,
1029                    "page@%p[%d %p %d %d %d %p %p]\n",
1030                    pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1031                    pg->cp_state, pg->cp_error, pg->cp_type,
1032                    pg->cp_owner, pg->cp_req);
1033 }
1034 EXPORT_SYMBOL(cl_page_header_print);
1035
1036 /**
1037  * Prints human readable representation of \a pg to the \a f.
1038  */
1039 void cl_page_print(const struct lu_env *env, void *cookie,
1040                    lu_printer_t printer, const struct cl_page *pg)
1041 {
1042         cl_page_header_print(env, cookie, printer, pg);
1043         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1044                        (const struct lu_env *env,
1045                         const struct cl_page_slice *slice,
1046                         void *cookie, lu_printer_t p), cookie, printer);
1047         (*printer)(env, cookie, "end page@%p\n", pg);
1048 }
1049 EXPORT_SYMBOL(cl_page_print);
1050
1051 /**
1052  * Cancel a page which is still in a transfer.
1053  */
1054 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1055 {
1056         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1057                               (const struct lu_env *,
1058                                const struct cl_page_slice *));
1059 }
1060
1061 /**
1062  * Converts a byte offset within object \a obj into a page index.
1063  */
1064 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1065 {
1066         return (loff_t)idx << PAGE_CACHE_SHIFT;
1067 }
1068 EXPORT_SYMBOL(cl_offset);
1069
1070 /**
1071  * Converts a page index into a byte offset within object \a obj.
1072  */
1073 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1074 {
1075         return offset >> PAGE_CACHE_SHIFT;
1076 }
1077 EXPORT_SYMBOL(cl_index);
1078
1079 size_t cl_page_size(const struct cl_object *obj)
1080 {
1081         return 1UL << PAGE_CACHE_SHIFT;
1082 }
1083 EXPORT_SYMBOL(cl_page_size);
1084
1085 /**
1086  * Adds page slice to the compound page.
1087  *
1088  * This is called by cl_object_operations::coo_page_init() methods to add a
1089  * per-layer state to the page. New state is added at the end of
1090  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1091  *
1092  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1093  */
1094 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1095                        struct cl_object *obj, pgoff_t index,
1096                        const struct cl_page_operations *ops)
1097 {
1098         ENTRY;
1099         list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1100         slice->cpl_obj  = obj;
1101         slice->cpl_index = index;
1102         slice->cpl_ops  = ops;
1103         slice->cpl_page = page;
1104         EXIT;
1105 }
1106 EXPORT_SYMBOL(cl_page_slice_add);
1107
1108 /**
1109  * Allocate and initialize cl_cache, called by ll_init_sbi().
1110  */
1111 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max)
1112 {
1113         struct cl_client_cache  *cache = NULL;
1114
1115         ENTRY;
1116         OBD_ALLOC(cache, sizeof(*cache));
1117         if (cache == NULL)
1118                 RETURN(NULL);
1119
1120         /* Initialize cache data */
1121         atomic_set(&cache->ccc_users, 1);
1122         cache->ccc_lru_max = lru_page_max;
1123         atomic_long_set(&cache->ccc_lru_left, lru_page_max);
1124         spin_lock_init(&cache->ccc_lru_lock);
1125         INIT_LIST_HEAD(&cache->ccc_lru);
1126
1127         /* turn unstable check off by default as it impacts performance */
1128         cache->ccc_unstable_check = 0;
1129         atomic_long_set(&cache->ccc_unstable_nr, 0);
1130         init_waitqueue_head(&cache->ccc_unstable_waitq);
1131
1132         RETURN(cache);
1133 }
1134 EXPORT_SYMBOL(cl_cache_init);
1135
1136 /**
1137  * Increase cl_cache refcount
1138  */
1139 void cl_cache_incref(struct cl_client_cache *cache)
1140 {
1141         atomic_inc(&cache->ccc_users);
1142 }
1143 EXPORT_SYMBOL(cl_cache_incref);
1144
1145 /**
1146  * Decrease cl_cache refcount and free the cache if refcount=0.
1147  * Since llite, lov and osc all hold cl_cache refcount,
1148  * the free will not cause race. (LU-6173)
1149  */
1150 void cl_cache_decref(struct cl_client_cache *cache)
1151 {
1152         if (atomic_dec_and_test(&cache->ccc_users))
1153                 OBD_FREE(cache, sizeof(*cache));
1154 }
1155 EXPORT_SYMBOL(cl_cache_decref);