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, 2015, 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 <linux/list.h>
  45 #include <libcfs/libcfs.h>
  46 #include <obd_class.h>
  47 #include <obd_support.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_state == CPS_FREEING);
 141
 142         ENTRY;
 143         while (!list_empty(&page->cp_layers)) {
 144                 struct cl_page_slice *slice;
 145
 146                 slice = list_entry(page->cp_layers.next,
 147                                    struct cl_page_slice, cpl_linkage);
 148                 list_del_init(page->cp_layers.next);
 149                 if (unlikely(slice->cpl_ops->cpo_fini != NULL))
 150                         slice->cpl_ops->cpo_fini(env, slice);
 151         }
 152         CS_PAGE_DEC(obj, total);
 153         CS_PAGESTATE_DEC(obj, page->cp_state);
 154         lu_object_ref_del_at(&obj->co_lu, &page->cp_obj_ref, "cl_page", page);
 155         cl_object_put(env, obj);
 156         lu_ref_fini(&page->cp_reference);
 157         OBD_FREE(page, pagesize);
 158         EXIT;
 159 }
 160
 161 /**
 162  * Helper function updating page state. This is the only place in the code
 163  * where cl_page::cp_state field is mutated.
 164  */
 165 static inline void cl_page_state_set_trust(struct cl_page *page,
 166                                            enum cl_page_state state)
 167 {
 168         /* bypass const. */
 169         *(enum cl_page_state *)&page->cp_state = state;
 170 }
 171
 172 struct cl_page *cl_page_alloc(const struct lu_env *env,
 173                 struct cl_object *o, pgoff_t ind, struct page *vmpage,
 174                 enum cl_page_type type)
 175 {
 176         struct cl_page          *page;
 177         struct lu_object_header *head;
 178
 179         ENTRY;
 180         OBD_ALLOC_GFP(page, cl_object_header(o)->coh_page_bufsize,
 181                         GFP_NOFS);
 182         if (page != NULL) {
 183                 int result = 0;
 184                 atomic_set(&page->cp_ref, 1);
 185                 page->cp_obj = o;
 186                 cl_object_get(o);
 187                 lu_object_ref_add_at(&o->co_lu, &page->cp_obj_ref, "cl_page",
 188                                      page);
 189                 page->cp_vmpage = vmpage;
 190                 cl_page_state_set_trust(page, CPS_CACHED);
 191                 page->cp_type = type;
 192                 INIT_LIST_HEAD(&page->cp_layers);
 193                 INIT_LIST_HEAD(&page->cp_batch);
 194                 lu_ref_init(&page->cp_reference);
 195                 head = o->co_lu.lo_header;
 196                 list_for_each_entry(o, &head->loh_layers,
 197                                     co_lu.lo_linkage) {
 198                         if (o->co_ops->coo_page_init != NULL) {
 199                                 result = o->co_ops->coo_page_init(env, o, page,
 200                                                                   ind);
 201                                 if (result != 0) {
 202                                         cl_page_delete0(env, page);
 203                                         cl_page_free(env, page);
 204                                         page = ERR_PTR(result);
 205                                         break;
 206                                 }
 207                         }
 208                 }
 209                 if (result == 0) {
 210                         CS_PAGE_INC(o, total);
 211                         CS_PAGE_INC(o, create);
 212                         CS_PAGESTATE_DEC(o, CPS_CACHED);
 213                 }
 214         } else {
 215                 page = ERR_PTR(-ENOMEM);
 216         }
 217         RETURN(page);
 218 }
 219
 220 /**
 221  * Returns a cl_page with index \a idx at the object \a o, and associated with
 222  * the VM page \a vmpage.
 223  *
 224  * This is the main entry point into the cl_page caching interface. First, a
 225  * cache (implemented as a per-object radix tree) is consulted. If page is
 226  * found there, it is returned immediately. Otherwise new page is allocated
 227  * and returned. In any case, additional reference to page is acquired.
 228  *
 229  * \see cl_object_find(), cl_lock_find()
 230  */
 231 struct cl_page *cl_page_find(const struct lu_env *env,
 232                              struct cl_object *o,
 233                              pgoff_t idx, struct page *vmpage,
 234                              enum cl_page_type type)
 235 {
 236         struct cl_page          *page = NULL;
 237         struct cl_object_header *hdr;
 238
 239         LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 240         might_sleep();
 241
 242         ENTRY;
 243
 244         hdr = cl_object_header(o);
 245         CS_PAGE_INC(o, lookup);
 246
 247         CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
 248                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 249         /* fast path. */
 250         if (type == CPT_CACHEABLE) {
 251                 /* vmpage lock is used to protect the child/parent
 252                  * relationship */
 253                 KLASSERT(PageLocked(vmpage));
 254                 /*
 255                  * cl_vmpage_page() can be called here without any locks as
 256                  *
 257                  *     - "vmpage" is locked (which prevents ->private from
 258                  *       concurrent updates), and
 259                  *
 260                  *     - "o" cannot be destroyed while current thread holds a
 261                  *       reference on it.
 262                  */
 263                 page = cl_vmpage_page(vmpage, o);
 264                 if (page != NULL) {
 265                         CS_PAGE_INC(o, hit);
 266                         RETURN(page);
 267                 }
 268         }
 269
 270         /* allocate and initialize cl_page */
 271         page = cl_page_alloc(env, o, idx, vmpage, type);
 272         RETURN(page);
 273 }
 274 EXPORT_SYMBOL(cl_page_find);
 275
 276 static inline int cl_page_invariant(const struct cl_page *pg)
 277 {
 278         return cl_page_in_use_noref(pg);
 279 }
 280
 281 static void cl_page_state_set0(const struct lu_env *env,
 282                                struct cl_page *page, enum cl_page_state state)
 283 {
 284         enum cl_page_state old;
 285
 286         /*
 287          * Matrix of allowed state transitions [old][new], for sanity
 288          * checking.
 289          */
 290         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 291                 [CPS_CACHED] = {
 292                         [CPS_CACHED]  = 0,
 293                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 294                         [CPS_PAGEIN]  = 0,
 295                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 296                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 297                 },
 298                 [CPS_OWNED] = {
 299                         [CPS_CACHED]  = 1, /* release to the cache */
 300                         [CPS_OWNED]   = 0,
 301                         [CPS_PAGEIN]  = 1, /* start read immediately */
 302                         [CPS_PAGEOUT] = 1, /* start write immediately */
 303                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 304                 },
 305                 [CPS_PAGEIN] = {
 306                         [CPS_CACHED]  = 1, /* io completion */
 307                         [CPS_OWNED]   = 0,
 308                         [CPS_PAGEIN]  = 0,
 309                         [CPS_PAGEOUT] = 0,
 310                         [CPS_FREEING] = 0,
 311                 },
 312                 [CPS_PAGEOUT] = {
 313                         [CPS_CACHED]  = 1, /* io completion */
 314                         [CPS_OWNED]   = 0,
 315                         [CPS_PAGEIN]  = 0,
 316                         [CPS_PAGEOUT] = 0,
 317                         [CPS_FREEING] = 0,
 318                 },
 319                 [CPS_FREEING] = {
 320                         [CPS_CACHED]  = 0,
 321                         [CPS_OWNED]   = 0,
 322                         [CPS_PAGEIN]  = 0,
 323                         [CPS_PAGEOUT] = 0,
 324                         [CPS_FREEING] = 0,
 325                 }
 326         };
 327
 328         ENTRY;
 329         old = page->cp_state;
 330         PASSERT(env, page, allowed_transitions[old][state]);
 331         CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
 332         PASSERT(env, page, page->cp_state == old);
 333         PASSERT(env, page, equi(state == CPS_OWNED, page->cp_owner != NULL));
 334
 335         CS_PAGESTATE_DEC(page->cp_obj, page->cp_state);
 336         CS_PAGESTATE_INC(page->cp_obj, state);
 337         cl_page_state_set_trust(page, state);
 338         EXIT;
 339 }
 340
 341 static void cl_page_state_set(const struct lu_env *env,
 342                               struct cl_page *page, enum cl_page_state state)
 343 {
 344         cl_page_state_set0(env, page, state);
 345 }
 346
 347 /**
 348  * Acquires an additional reference to a page.
 349  *
 350  * This can be called only by caller already possessing a reference to \a
 351  * page.
 352  *
 353  * \see cl_object_get(), cl_lock_get().
 354  */
 355 void cl_page_get(struct cl_page *page)
 356 {
 357         ENTRY;
 358         cl_page_get_trust(page);
 359         EXIT;
 360 }
 361 EXPORT_SYMBOL(cl_page_get);
 362
 363 /**
 364  * Releases a reference to a page.
 365  *
 366  * When last reference is released, page is returned to the cache, unless it
 367  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 368  * destroyed.
 369  *
 370  * \see cl_object_put(), cl_lock_put().
 371  */
 372 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 373 {
 374         ENTRY;
 375         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 376                        atomic_read(&page->cp_ref));
 377
 378         if (atomic_dec_and_test(&page->cp_ref)) {
 379                 LASSERT(page->cp_state == CPS_FREEING);
 380
 381                 LASSERT(atomic_read(&page->cp_ref) == 0);
 382                 PASSERT(env, page, page->cp_owner == NULL);
 383                 PASSERT(env, page, list_empty(&page->cp_batch));
 384                 /*
 385                  * Page is no longer reachable by other threads. Tear
 386                  * it down.
 387                  */
 388                 cl_page_free(env, page);
 389         }
 390
 391         EXIT;
 392 }
 393 EXPORT_SYMBOL(cl_page_put);
 394
 395 /**
 396  * Returns a cl_page associated with a VM page, and given cl_object.
 397  */
 398 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj)
 399 {
 400         struct cl_page *page;
 401
 402         ENTRY;
 403         KLASSERT(PageLocked(vmpage));
 404
 405         /*
 406          * NOTE: absence of races and liveness of data are guaranteed by page
 407          *       lock on a "vmpage". That works because object destruction has
 408          *       bottom-to-top pass.
 409          */
 410
 411         page = (struct cl_page *)vmpage->private;
 412         if (page != NULL) {
 413                 cl_page_get_trust(page);
 414                 LASSERT(page->cp_type == CPT_CACHEABLE);
 415         }
 416         RETURN(page);
 417 }
 418 EXPORT_SYMBOL(cl_vmpage_page);
 419
 420 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 421                                        const struct lu_device_type *dtype)
 422 {
 423         return cl_page_at_trusted(page, dtype);
 424 }
 425 EXPORT_SYMBOL(cl_page_at);
 426
 427 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 428
 429 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)                   \
 430 ({                                                                      \
 431         const struct lu_env        *__env  = (_env);                    \
 432         struct cl_page             *__page = (_page);                   \
 433         const struct cl_page_slice *__scan;                             \
 434         int                         __result;                           \
 435         ptrdiff_t                   __op   = (_op);                     \
 436         int                        (*__method)_proto;                   \
 437                                                                         \
 438         __result = 0;                                                   \
 439         list_for_each_entry(__scan, &__page->cp_layers, cpl_linkage) {  \
 440                 __method = *(void **)((char *)__scan->cpl_ops +  __op); \
 441                 if (__method != NULL) {                                 \
 442                         __result = (*__method)(__env, __scan, ## __VA_ARGS__); \
 443                         if (__result != 0)                              \
 444                                 break;                                  \
 445                 }                                                       \
 446         }                                                               \
 447         if (__result > 0)                                               \
 448                 __result = 0;                                           \
 449         __result;                                                       \
 450 })
 451
 452 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 453 do {                                                                    \
 454         const struct lu_env        *__env  = (_env);                    \
 455         struct cl_page             *__page = (_page);                   \
 456         const struct cl_page_slice *__scan;                             \
 457         ptrdiff_t                   __op   = (_op);                     \
 458         void                      (*__method)_proto;                    \
 459                                                                         \
 460         list_for_each_entry(__scan, &__page->cp_layers, cpl_linkage) {  \
 461                 __method = *(void **)((char *)__scan->cpl_ops +  __op); \
 462                 if (__method != NULL)                                   \
 463                         (*__method)(__env, __scan, ## __VA_ARGS__);     \
 464         }                                                               \
 465 } while (0)
 466
 467 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)           \
 468 do {                                                                    \
 469         const struct lu_env        *__env  = (_env);                    \
 470         struct cl_page             *__page = (_page);                   \
 471         const struct cl_page_slice *__scan;                             \
 472         ptrdiff_t                   __op   = (_op);                     \
 473         void                      (*__method)_proto;                    \
 474                                                                         \
 475         /* get to the bottom page. */                                   \
 476         list_for_each_entry_reverse(__scan, &__page->cp_layers,         \
 477                                     cpl_linkage) {                      \
 478                 __method = *(void **)((char *)__scan->cpl_ops + __op);  \
 479                 if (__method != NULL)                                   \
 480                         (*__method)(__env, __scan, ## __VA_ARGS__);     \
 481         }                                                               \
 482 } while (0)
 483
 484 static int cl_page_invoke(const struct lu_env *env,
 485                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 486
 487 {
 488         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 489         ENTRY;
 490         RETURN(CL_PAGE_INVOKE(env, page, op,
 491                               (const struct lu_env *,
 492                                const struct cl_page_slice *, struct cl_io *),
 493                               io));
 494 }
 495
 496 static void cl_page_invoid(const struct lu_env *env,
 497                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 498
 499 {
 500         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 501         ENTRY;
 502         CL_PAGE_INVOID(env, page, op,
 503                        (const struct lu_env *,
 504                         const struct cl_page_slice *, struct cl_io *), io);
 505         EXIT;
 506 }
 507
 508 static void cl_page_owner_clear(struct cl_page *page)
 509 {
 510         ENTRY;
 511         if (page->cp_owner != NULL) {
 512                 LASSERT(page->cp_owner->ci_owned_nr > 0);
 513                 page->cp_owner->ci_owned_nr--;
 514                 page->cp_owner = NULL;
 515         }
 516         EXIT;
 517 }
 518
 519 static void cl_page_owner_set(struct cl_page *page)
 520 {
 521         ENTRY;
 522         LASSERT(page->cp_owner != NULL);
 523         page->cp_owner->ci_owned_nr++;
 524         EXIT;
 525 }
 526
 527 void cl_page_disown0(const struct lu_env *env,
 528                      struct cl_io *io, struct cl_page *pg)
 529 {
 530         enum cl_page_state state;
 531
 532         ENTRY;
 533         state = pg->cp_state;
 534         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 535         PINVRNT(env, pg, cl_page_invariant(pg) || state == CPS_FREEING);
 536         cl_page_owner_clear(pg);
 537
 538         if (state == CPS_OWNED)
 539                 cl_page_state_set(env, pg, CPS_CACHED);
 540         /*
 541          * Completion call-backs are executed in the bottom-up order, so that
 542          * uppermost layer (llite), responsible for VFS/VM interaction runs
 543          * last and can release locks safely.
 544          */
 545         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 546                                (const struct lu_env *,
 547                                 const struct cl_page_slice *, struct cl_io *),
 548                                io);
 549         EXIT;
 550 }
 551
 552 /**
 553  * returns true, iff page is owned by the given io.
 554  */
 555 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 556 {
 557         struct cl_io *top = cl_io_top((struct cl_io *)io);
 558         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 559         ENTRY;
 560         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top);
 561 }
 562 EXPORT_SYMBOL(cl_page_is_owned);
 563
 564 /**
 565  * Try to own a page by IO.
 566  *
 567  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 568  * into cl_page_state::CPS_OWNED state.
 569  *
 570  * \pre  !cl_page_is_owned(pg, io)
 571  * \post result == 0 iff cl_page_is_owned(pg, io)
 572  *
 573  * \retval 0   success
 574  *
 575  * \retval -ve failure, e.g., page was destroyed (and landed in
 576  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 577  *             or, page was owned by another thread, or in IO.
 578  *
 579  * \see cl_page_disown()
 580  * \see cl_page_operations::cpo_own()
 581  * \see cl_page_own_try()
 582  * \see cl_page_own
 583  */
 584 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 585                         struct cl_page *pg, int nonblock)
 586 {
 587         int result;
 588
 589         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 590
 591         ENTRY;
 592         io = cl_io_top(io);
 593
 594         if (pg->cp_state == CPS_FREEING) {
 595                 result = -ENOENT;
 596         } else {
 597                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 598                                         (const struct lu_env *,
 599                                          const struct cl_page_slice *,
 600                                          struct cl_io *, int),
 601                                         io, nonblock);
 602                 if (result == 0) {
 603                         PASSERT(env, pg, pg->cp_owner == NULL);
 604                         pg->cp_owner = cl_io_top(io);;
 605                         cl_page_owner_set(pg);
 606                         if (pg->cp_state != CPS_FREEING) {
 607                                 cl_page_state_set(env, pg, CPS_OWNED);
 608                         } else {
 609                                 cl_page_disown0(env, io, pg);
 610                                 result = -ENOENT;
 611                         }
 612                 }
 613         }
 614         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 615         RETURN(result);
 616 }
 617
 618 /**
 619  * Own a page, might be blocked.
 620  *
 621  * \see cl_page_own0()
 622  */
 623 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 624 {
 625         return cl_page_own0(env, io, pg, 0);
 626 }
 627 EXPORT_SYMBOL(cl_page_own);
 628
 629 /**
 630  * Nonblock version of cl_page_own().
 631  *
 632  * \see cl_page_own0()
 633  */
 634 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
 635                     struct cl_page *pg)
 636 {
 637         return cl_page_own0(env, io, pg, 1);
 638 }
 639 EXPORT_SYMBOL(cl_page_own_try);
 640
 641
 642 /**
 643  * Assume page ownership.
 644  *
 645  * Called when page is already locked by the hosting VM.
 646  *
 647  * \pre !cl_page_is_owned(pg, io)
 648  * \post cl_page_is_owned(pg, io)
 649  *
 650  * \see cl_page_operations::cpo_assume()
 651  */
 652 void cl_page_assume(const struct lu_env *env,
 653                     struct cl_io *io, struct cl_page *pg)
 654 {
 655         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
 656
 657         ENTRY;
 658         io = cl_io_top(io);
 659
 660         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
 661         PASSERT(env, pg, pg->cp_owner == NULL);
 662         pg->cp_owner = cl_io_top(io);
 663         cl_page_owner_set(pg);
 664         cl_page_state_set(env, pg, CPS_OWNED);
 665         EXIT;
 666 }
 667 EXPORT_SYMBOL(cl_page_assume);
 668
 669 /**
 670  * Releases page ownership without unlocking the page.
 671  *
 672  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
 673  * underlying VM page (as VM is supposed to do this itself).
 674  *
 675  * \pre   cl_page_is_owned(pg, io)
 676  * \post !cl_page_is_owned(pg, io)
 677  *
 678  * \see cl_page_assume()
 679  */
 680 void cl_page_unassume(const struct lu_env *env,
 681                       struct cl_io *io, struct cl_page *pg)
 682 {
 683         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 684         PINVRNT(env, pg, cl_page_invariant(pg));
 685
 686         ENTRY;
 687         io = cl_io_top(io);
 688         cl_page_owner_clear(pg);
 689         cl_page_state_set(env, pg, CPS_CACHED);
 690         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
 691                                (const struct lu_env *,
 692                                 const struct cl_page_slice *, struct cl_io *),
 693                                io);
 694         EXIT;
 695 }
 696 EXPORT_SYMBOL(cl_page_unassume);
 697
 698 /**
 699  * Releases page ownership.
 700  *
 701  * Moves page into cl_page_state::CPS_CACHED.
 702  *
 703  * \pre   cl_page_is_owned(pg, io)
 704  * \post !cl_page_is_owned(pg, io)
 705  *
 706  * \see cl_page_own()
 707  * \see cl_page_operations::cpo_disown()
 708  */
 709 void cl_page_disown(const struct lu_env *env,
 710                     struct cl_io *io, struct cl_page *pg)
 711 {
 712         PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
 713                 pg->cp_state == CPS_FREEING);
 714
 715         ENTRY;
 716         io = cl_io_top(io);
 717         cl_page_disown0(env, io, pg);
 718         EXIT;
 719 }
 720 EXPORT_SYMBOL(cl_page_disown);
 721
 722 /**
 723  * Called when page is to be removed from the object, e.g., as a result of
 724  * truncate.
 725  *
 726  * Calls cl_page_operations::cpo_discard() top-to-bottom.
 727  *
 728  * \pre cl_page_is_owned(pg, io)
 729  *
 730  * \see cl_page_operations::cpo_discard()
 731  */
 732 void cl_page_discard(const struct lu_env *env,
 733                      struct cl_io *io, struct cl_page *pg)
 734 {
 735         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 736         PINVRNT(env, pg, cl_page_invariant(pg));
 737
 738         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
 739 }
 740 EXPORT_SYMBOL(cl_page_discard);
 741
 742 /**
 743  * Version of cl_page_delete() that can be called for not fully constructed
 744  * pages, e.g. in an error handling cl_page_find()->cl_page_delete0()
 745  * path. Doesn't check page invariant.
 746  */
 747 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg)
 748 {
 749         ENTRY;
 750
 751         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
 752
 753         /*
 754          * Severe all ways to obtain new pointers to @pg.
 755          */
 756         cl_page_owner_clear(pg);
 757
 758         cl_page_state_set0(env, pg, CPS_FREEING);
 759
 760         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_delete),
 761                        (const struct lu_env *, const struct cl_page_slice *));
 762
 763         EXIT;
 764 }
 765
 766 /**
 767  * Called when a decision is made to throw page out of memory.
 768  *
 769  * Notifies all layers about page destruction by calling
 770  * cl_page_operations::cpo_delete() method top-to-bottom.
 771  *
 772  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
 773  * where transition to this state happens).
 774  *
 775  * Eliminates all venues through which new references to the page can be
 776  * obtained:
 777  *
 778  *     - removes page from the radix trees,
 779  *
 780  *     - breaks linkage from VM page to cl_page.
 781  *
 782  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
 783  * drain after some time, at which point page will be recycled.
 784  *
 785  * \pre  VM page is locked
 786  * \post pg->cp_state == CPS_FREEING
 787  *
 788  * \see cl_page_operations::cpo_delete()
 789  */
 790 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
 791 {
 792         PINVRNT(env, pg, cl_page_invariant(pg));
 793         ENTRY;
 794         cl_page_delete0(env, pg);
 795         EXIT;
 796 }
 797 EXPORT_SYMBOL(cl_page_delete);
 798
 799 /**
 800  * Marks page up-to-date.
 801  *
 802  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
 803  * layer responsible for VM interaction has to mark/clear page as up-to-date
 804  * by the \a uptodate argument.
 805  *
 806  * \see cl_page_operations::cpo_export()
 807  */
 808 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
 809 {
 810         PINVRNT(env, pg, cl_page_invariant(pg));
 811         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
 812                        (const struct lu_env *,
 813                         const struct cl_page_slice *, int), uptodate);
 814 }
 815 EXPORT_SYMBOL(cl_page_export);
 816
 817 /**
 818  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
 819  * thread.
 820  */
 821 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
 822 {
 823         int result;
 824         const struct cl_page_slice *slice;
 825
 826         ENTRY;
 827         slice = container_of(pg->cp_layers.next,
 828                              const struct cl_page_slice, cpl_linkage);
 829         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
 830         /*
 831          * Call ->cpo_is_vmlocked() directly instead of going through
 832          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
 833          * cl_page_invariant().
 834          */
 835         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
 836         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
 837         RETURN(result == -EBUSY);
 838 }
 839 EXPORT_SYMBOL(cl_page_is_vmlocked);
 840
 841 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
 842 {
 843         ENTRY;
 844         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
 845 }
 846
 847 static void cl_page_io_start(const struct lu_env *env,
 848                              struct cl_page *pg, enum cl_req_type crt)
 849 {
 850         /*
 851          * Page is queued for IO, change its state.
 852          */
 853         ENTRY;
 854         cl_page_owner_clear(pg);
 855         cl_page_state_set(env, pg, cl_req_type_state(crt));
 856         EXIT;
 857 }
 858
 859 /**
 860  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
 861  * called top-to-bottom. Every layer either agrees to submit this page (by
 862  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
 863  * handling interactions with the VM also has to inform VM that page is under
 864  * transfer now.
 865  */
 866 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
 867                  struct cl_page *pg, enum cl_req_type crt)
 868 {
 869         int result;
 870
 871         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 872         PINVRNT(env, pg, cl_page_invariant(pg));
 873         PINVRNT(env, pg, crt < CRT_NR);
 874
 875         /*
 876          * XXX this has to be called bottom-to-top, so that llite can set up
 877          * PG_writeback without risking other layers deciding to skip this
 878          * page.
 879          */
 880         if (crt >= CRT_NR)
 881                 return -EINVAL;
 882         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
 883         if (result == 0)
 884                 cl_page_io_start(env, pg, crt);
 885
 886         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
 887                       equi(result == 0,
 888                            PageWriteback(cl_page_vmpage(pg)))));
 889         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 890         return result;
 891 }
 892 EXPORT_SYMBOL(cl_page_prep);
 893
 894 /**
 895  * Notify layers about transfer completion.
 896  *
 897  * Invoked by transfer sub-system (which is a part of osc) to notify layers
 898  * that a transfer, of which this page is a part of has completed.
 899  *
 900  * Completion call-backs are executed in the bottom-up order, so that
 901  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
 902  * and can release locks safely.
 903  *
 904  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 905  * \post pg->cp_state == CPS_CACHED
 906  *
 907  * \see cl_page_operations::cpo_completion()
 908  */
 909 void cl_page_completion(const struct lu_env *env,
 910                         struct cl_page *pg, enum cl_req_type crt, int ioret)
 911 {
 912         struct cl_sync_io *anchor = pg->cp_sync_io;
 913
 914         PASSERT(env, pg, crt < CRT_NR);
 915         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
 916
 917         ENTRY;
 918         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
 919         cl_page_state_set(env, pg, CPS_CACHED);
 920         if (crt >= CRT_NR)
 921                 return;
 922         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
 923                                (const struct lu_env *,
 924                                 const struct cl_page_slice *, int), ioret);
 925         if (anchor != NULL) {
 926                 LASSERT(pg->cp_sync_io == anchor);
 927                 pg->cp_sync_io = NULL;
 928                 cl_sync_io_note(env, anchor, ioret);
 929         }
 930         EXIT;
 931 }
 932 EXPORT_SYMBOL(cl_page_completion);
 933
 934 /**
 935  * Notify layers that transfer formation engine decided to yank this page from
 936  * the cache and to make it a part of a transfer.
 937  *
 938  * \pre  pg->cp_state == CPS_CACHED
 939  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 940  *
 941  * \see cl_page_operations::cpo_make_ready()
 942  */
 943 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
 944                        enum cl_req_type crt)
 945 {
 946         int result;
 947
 948         PINVRNT(env, pg, crt < CRT_NR);
 949
 950         ENTRY;
 951         if (crt >= CRT_NR)
 952                 RETURN(-EINVAL);
 953         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
 954                                 (const struct lu_env *,
 955                                  const struct cl_page_slice *));
 956         if (result == 0) {
 957                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
 958                 cl_page_io_start(env, pg, crt);
 959         }
 960         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 961         RETURN(result);
 962 }
 963 EXPORT_SYMBOL(cl_page_make_ready);
 964
 965 /**
 966  * Called if a pge is being written back by kernel's intention.
 967  *
 968  * \pre  cl_page_is_owned(pg, io)
 969  * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT)
 970  *
 971  * \see cl_page_operations::cpo_flush()
 972  */
 973 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
 974                   struct cl_page *pg)
 975 {
 976         int result;
 977
 978         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 979         PINVRNT(env, pg, cl_page_invariant(pg));
 980
 981         ENTRY;
 982
 983         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_flush));
 984
 985         CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result);
 986         RETURN(result);
 987 }
 988 EXPORT_SYMBOL(cl_page_flush);
 989
 990 /**
 991  * Tells transfer engine that only part of a page is to be transmitted.
 992  *
 993  * \see cl_page_operations::cpo_clip()
 994  */
 995 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
 996                   int from, int to)
 997 {
 998         PINVRNT(env, pg, cl_page_invariant(pg));
 999
1000         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1001         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1002                        (const struct lu_env *,
1003                         const struct cl_page_slice *,int, int),
1004                        from, to);
1005 }
1006 EXPORT_SYMBOL(cl_page_clip);
1007
1008 /**
1009  * Prints human readable representation of \a pg to the \a f.
1010  */
1011 void cl_page_header_print(const struct lu_env *env, void *cookie,
1012                           lu_printer_t printer, const struct cl_page *pg)
1013 {
1014         (*printer)(env, cookie,
1015                    "page@%p[%d %p %d %d %p]\n",
1016                    pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1017                    pg->cp_state, pg->cp_type,
1018                    pg->cp_owner);
1019 }
1020 EXPORT_SYMBOL(cl_page_header_print);
1021
1022 /**
1023  * Prints human readable representation of \a pg to the \a f.
1024  */
1025 void cl_page_print(const struct lu_env *env, void *cookie,
1026                    lu_printer_t printer, const struct cl_page *pg)
1027 {
1028         cl_page_header_print(env, cookie, printer, pg);
1029         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1030                        (const struct lu_env *env,
1031                         const struct cl_page_slice *slice,
1032                         void *cookie, lu_printer_t p), cookie, printer);
1033         (*printer)(env, cookie, "end page@%p\n", pg);
1034 }
1035 EXPORT_SYMBOL(cl_page_print);
1036
1037 /**
1038  * Cancel a page which is still in a transfer.
1039  */
1040 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1041 {
1042         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1043                               (const struct lu_env *,
1044                                const struct cl_page_slice *));
1045 }
1046
1047 /**
1048  * Converts a byte offset within object \a obj into a page index.
1049  */
1050 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1051 {
1052         return (loff_t)idx << PAGE_SHIFT;
1053 }
1054 EXPORT_SYMBOL(cl_offset);
1055
1056 /**
1057  * Converts a page index into a byte offset within object \a obj.
1058  */
1059 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1060 {
1061         return offset >> PAGE_SHIFT;
1062 }
1063 EXPORT_SYMBOL(cl_index);
1064
1065 size_t cl_page_size(const struct cl_object *obj)
1066 {
1067         return 1UL << PAGE_SHIFT;
1068 }
1069 EXPORT_SYMBOL(cl_page_size);
1070
1071 /**
1072  * Adds page slice to the compound page.
1073  *
1074  * This is called by cl_object_operations::coo_page_init() methods to add a
1075  * per-layer state to the page. New state is added at the end of
1076  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1077  *
1078  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1079  */
1080 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1081                        struct cl_object *obj, pgoff_t index,
1082                        const struct cl_page_operations *ops)
1083 {
1084         ENTRY;
1085         list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1086         slice->cpl_obj  = obj;
1087         slice->cpl_index = index;
1088         slice->cpl_ops  = ops;
1089         slice->cpl_page = page;
1090         EXIT;
1091 }
1092 EXPORT_SYMBOL(cl_page_slice_add);
1093
1094 /**
1095  * Allocate and initialize cl_cache, called by ll_init_sbi().
1096  */
1097 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max)
1098 {
1099         struct cl_client_cache  *cache = NULL;
1100
1101         ENTRY;
1102         OBD_ALLOC(cache, sizeof(*cache));
1103         if (cache == NULL)
1104                 RETURN(NULL);
1105
1106         /* Initialize cache data */
1107         atomic_set(&cache->ccc_users, 1);
1108         cache->ccc_lru_max = lru_page_max;
1109         atomic_long_set(&cache->ccc_lru_left, lru_page_max);
1110         spin_lock_init(&cache->ccc_lru_lock);
1111         INIT_LIST_HEAD(&cache->ccc_lru);
1112
1113         /* turn unstable check off by default as it impacts performance */
1114         cache->ccc_unstable_check = 0;
1115         atomic_long_set(&cache->ccc_unstable_nr, 0);
1116         init_waitqueue_head(&cache->ccc_unstable_waitq);
1117
1118         RETURN(cache);
1119 }
1120 EXPORT_SYMBOL(cl_cache_init);
1121
1122 /**
1123  * Increase cl_cache refcount
1124  */
1125 void cl_cache_incref(struct cl_client_cache *cache)
1126 {
1127         atomic_inc(&cache->ccc_users);
1128 }
1129 EXPORT_SYMBOL(cl_cache_incref);
1130
1131 /**
1132  * Decrease cl_cache refcount and free the cache if refcount=0.
1133  * Since llite, lov and osc all hold cl_cache refcount,
1134  * the free will not cause race. (LU-6173)
1135  */
1136 void cl_cache_decref(struct cl_client_cache *cache)
1137 {
1138         if (atomic_dec_and_test(&cache->ccc_users))
1139                 OBD_FREE(cache, sizeof(*cache));
1140 }
1141 EXPORT_SYMBOL(cl_cache_decref);