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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Client Lustre Page.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_CLASS
  39
  40 #include <linux/list.h>
  41 #include <libcfs/libcfs.h>
  42 #include <obd_class.h>
  43 #include <obd_support.h>
  44
  45 #include <cl_object.h>
  46 #include "cl_internal.h"
  47
  48 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg);
  49
  50 #ifdef LIBCFS_DEBUG
  51 # define PASSERT(env, page, expr)                                       \
  52   do {                                                                    \
  53           if (unlikely(!(expr))) {                                      \
  54                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  55                   LASSERT(0);                                           \
  56           }                                                             \
  57   } while (0)
  58 #else /* !LIBCFS_DEBUG */
  59 # define PASSERT(env, page, exp) \
  60         ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  61 #endif /* !LIBCFS_DEBUG */
  62
  63 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
  64 # define PINVRNT(env, page, expr)                                       \
  65   do {                                                                    \
  66           if (unlikely(!(expr))) {                                      \
  67                   CL_PAGE_DEBUG(D_ERROR, (env), (page), #expr "\n");    \
  68                   LINVRNT(0);                                           \
  69           }                                                             \
  70   } while (0)
  71 #else /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
  72 # define PINVRNT(env, page, exp) \
  73          ((void)sizeof(env), (void)sizeof(page), (void)sizeof !!(exp))
  74 #endif /* !CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
  75
  76 /* Disable page statistic by default due to huge performance penalty. */
  77 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
  78 #define CS_PAGE_INC(o, item) \
  79         atomic_inc(&cl_object_site(o)->cs_pages.cs_stats[CS_##item])
  80 #define CS_PAGE_DEC(o, item) \
  81         atomic_dec(&cl_object_site(o)->cs_pages.cs_stats[CS_##item])
  82 #define CS_PAGESTATE_INC(o, state) \
  83         atomic_inc(&cl_object_site(o)->cs_pages_state[state])
  84 #define CS_PAGESTATE_DEC(o, state) \
  85         atomic_dec(&cl_object_site(o)->cs_pages_state[state])
  86 #else
  87 #define CS_PAGE_INC(o, item)
  88 #define CS_PAGE_DEC(o, item)
  89 #define CS_PAGESTATE_INC(o, state)
  90 #define CS_PAGESTATE_DEC(o, state)
  91 #endif
  92
  93 /**
  94  * Internal version of cl_page_get().
  95  *
  96  * This function can be used to obtain initial reference to previously
  97  * unreferenced cached object. It can be called only if concurrent page
  98  * reclamation is somehow prevented, e.g., by keeping a lock on a VM page,
  99  * associated with \a page.
 100  *
 101  * Use with care! Not exported.
 102  */
 103 static void cl_page_get_trust(struct cl_page *page)
 104 {
 105         LASSERT(atomic_read(&page->cp_ref) > 0);
 106         atomic_inc(&page->cp_ref);
 107 }
 108
 109 /**
 110  * Returns a slice within a page, corresponding to the given layer in the
 111  * device stack.
 112  *
 113  * \see cl_lock_at()
 114  */
 115 static const struct cl_page_slice *
 116 cl_page_at_trusted(const struct cl_page *page,
 117                    const struct lu_device_type *dtype)
 118 {
 119         const struct cl_page_slice *slice;
 120         ENTRY;
 121
 122         list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 123                 if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
 124                         RETURN(slice);
 125         }
 126         RETURN(NULL);
 127 }
 128
 129 static void cl_page_free(const struct lu_env *env, struct cl_page *page)
 130 {
 131         struct cl_object *obj  = page->cp_obj;
 132         int pagesize = cl_object_header(obj)->coh_page_bufsize;
 133
 134         PASSERT(env, page, list_empty(&page->cp_batch));
 135         PASSERT(env, page, page->cp_owner == NULL);
 136         PASSERT(env, page, page->cp_state == CPS_FREEING);
 137
 138         ENTRY;
 139         while (!list_empty(&page->cp_layers)) {
 140                 struct cl_page_slice *slice;
 141
 142                 slice = list_entry(page->cp_layers.next,
 143                                    struct cl_page_slice, cpl_linkage);
 144                 list_del_init(page->cp_layers.next);
 145                 if (unlikely(slice->cpl_ops->cpo_fini != NULL))
 146                         slice->cpl_ops->cpo_fini(env, slice);
 147         }
 148         CS_PAGE_DEC(obj, total);
 149         CS_PAGESTATE_DEC(obj, page->cp_state);
 150         lu_object_ref_del_at(&obj->co_lu, &page->cp_obj_ref, "cl_page", page);
 151         cl_object_put(env, obj);
 152         lu_ref_fini(&page->cp_reference);
 153         OBD_FREE(page, pagesize);
 154         EXIT;
 155 }
 156
 157 /**
 158  * Helper function updating page state. This is the only place in the code
 159  * where cl_page::cp_state field is mutated.
 160  */
 161 static inline void cl_page_state_set_trust(struct cl_page *page,
 162                                            enum cl_page_state state)
 163 {
 164         /* bypass const. */
 165         *(enum cl_page_state *)&page->cp_state = state;
 166 }
 167
 168 struct cl_page *cl_page_alloc(const struct lu_env *env,
 169                 struct cl_object *o, pgoff_t ind, struct page *vmpage,
 170                 enum cl_page_type type)
 171 {
 172         struct cl_page          *page;
 173         struct lu_object_header *head;
 174
 175         ENTRY;
 176         OBD_ALLOC_GFP(page, cl_object_header(o)->coh_page_bufsize,
 177                         GFP_NOFS);
 178         if (page != NULL) {
 179                 int result = 0;
 180                 atomic_set(&page->cp_ref, 1);
 181                 page->cp_obj = o;
 182                 cl_object_get(o);
 183                 lu_object_ref_add_at(&o->co_lu, &page->cp_obj_ref, "cl_page",
 184                                      page);
 185                 page->cp_vmpage = vmpage;
 186                 cl_page_state_set_trust(page, CPS_CACHED);
 187                 page->cp_type = type;
 188                 INIT_LIST_HEAD(&page->cp_layers);
 189                 INIT_LIST_HEAD(&page->cp_batch);
 190                 lu_ref_init(&page->cp_reference);
 191                 head = o->co_lu.lo_header;
 192                 list_for_each_entry(o, &head->loh_layers,
 193                                     co_lu.lo_linkage) {
 194                         if (o->co_ops->coo_page_init != NULL) {
 195                                 result = o->co_ops->coo_page_init(env, o, page,
 196                                                                   ind);
 197                                 if (result != 0) {
 198                                         cl_page_delete0(env, page);
 199                                         cl_page_free(env, page);
 200                                         page = ERR_PTR(result);
 201                                         break;
 202                                 }
 203                         }
 204                 }
 205                 if (result == 0) {
 206                         CS_PAGE_INC(o, total);
 207                         CS_PAGE_INC(o, create);
 208                         CS_PAGESTATE_DEC(o, CPS_CACHED);
 209                 }
 210         } else {
 211                 page = ERR_PTR(-ENOMEM);
 212         }
 213         RETURN(page);
 214 }
 215
 216 /**
 217  * Returns a cl_page with index \a idx at the object \a o, and associated with
 218  * the VM page \a vmpage.
 219  *
 220  * This is the main entry point into the cl_page caching interface. First, a
 221  * cache (implemented as a per-object radix tree) is consulted. If page is
 222  * found there, it is returned immediately. Otherwise new page is allocated
 223  * and returned. In any case, additional reference to page is acquired.
 224  *
 225  * \see cl_object_find(), cl_lock_find()
 226  */
 227 struct cl_page *cl_page_find(const struct lu_env *env,
 228                              struct cl_object *o,
 229                              pgoff_t idx, struct page *vmpage,
 230                              enum cl_page_type type)
 231 {
 232         struct cl_page          *page = NULL;
 233         struct cl_object_header *hdr;
 234
 235         LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 236         might_sleep();
 237
 238         ENTRY;
 239
 240         hdr = cl_object_header(o);
 241         CS_PAGE_INC(o, lookup);
 242
 243         CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
 244                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 245         /* fast path. */
 246         if (type == CPT_CACHEABLE) {
 247                 /* vmpage lock is used to protect the child/parent
 248                  * relationship */
 249                 KLASSERT(PageLocked(vmpage));
 250                 /*
 251                  * cl_vmpage_page() can be called here without any locks as
 252                  *
 253                  *     - "vmpage" is locked (which prevents ->private from
 254                  *       concurrent updates), and
 255                  *
 256                  *     - "o" cannot be destroyed while current thread holds a
 257                  *       reference on it.
 258                  */
 259                 page = cl_vmpage_page(vmpage, o);
 260                 if (page != NULL) {
 261                         CS_PAGE_INC(o, hit);
 262                         RETURN(page);
 263                 }
 264         }
 265
 266         /* allocate and initialize cl_page */
 267         page = cl_page_alloc(env, o, idx, vmpage, type);
 268         RETURN(page);
 269 }
 270 EXPORT_SYMBOL(cl_page_find);
 271
 272 static inline int cl_page_invariant(const struct cl_page *pg)
 273 {
 274         return cl_page_in_use_noref(pg);
 275 }
 276
 277 static void cl_page_state_set0(const struct lu_env *env,
 278                                struct cl_page *page, enum cl_page_state state)
 279 {
 280         enum cl_page_state old;
 281
 282         /*
 283          * Matrix of allowed state transitions [old][new], for sanity
 284          * checking.
 285          */
 286         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 287                 [CPS_CACHED] = {
 288                         [CPS_CACHED]  = 0,
 289                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 290                         [CPS_PAGEIN]  = 0,
 291                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 292                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 293                 },
 294                 [CPS_OWNED] = {
 295                         [CPS_CACHED]  = 1, /* release to the cache */
 296                         [CPS_OWNED]   = 0,
 297                         [CPS_PAGEIN]  = 1, /* start read immediately */
 298                         [CPS_PAGEOUT] = 1, /* start write immediately */
 299                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 300                 },
 301                 [CPS_PAGEIN] = {
 302                         [CPS_CACHED]  = 1, /* io completion */
 303                         [CPS_OWNED]   = 0,
 304                         [CPS_PAGEIN]  = 0,
 305                         [CPS_PAGEOUT] = 0,
 306                         [CPS_FREEING] = 0,
 307                 },
 308                 [CPS_PAGEOUT] = {
 309                         [CPS_CACHED]  = 1, /* io completion */
 310                         [CPS_OWNED]   = 0,
 311                         [CPS_PAGEIN]  = 0,
 312                         [CPS_PAGEOUT] = 0,
 313                         [CPS_FREEING] = 0,
 314                 },
 315                 [CPS_FREEING] = {
 316                         [CPS_CACHED]  = 0,
 317                         [CPS_OWNED]   = 0,
 318                         [CPS_PAGEIN]  = 0,
 319                         [CPS_PAGEOUT] = 0,
 320                         [CPS_FREEING] = 0,
 321                 }
 322         };
 323
 324         ENTRY;
 325         old = page->cp_state;
 326         PASSERT(env, page, allowed_transitions[old][state]);
 327         CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
 328         PASSERT(env, page, page->cp_state == old);
 329         PASSERT(env, page, equi(state == CPS_OWNED, page->cp_owner != NULL));
 330
 331         CS_PAGESTATE_DEC(page->cp_obj, page->cp_state);
 332         CS_PAGESTATE_INC(page->cp_obj, state);
 333         cl_page_state_set_trust(page, state);
 334         EXIT;
 335 }
 336
 337 static void cl_page_state_set(const struct lu_env *env,
 338                               struct cl_page *page, enum cl_page_state state)
 339 {
 340         cl_page_state_set0(env, page, state);
 341 }
 342
 343 /**
 344  * Acquires an additional reference to a page.
 345  *
 346  * This can be called only by caller already possessing a reference to \a
 347  * page.
 348  *
 349  * \see cl_object_get(), cl_lock_get().
 350  */
 351 void cl_page_get(struct cl_page *page)
 352 {
 353         ENTRY;
 354         cl_page_get_trust(page);
 355         EXIT;
 356 }
 357 EXPORT_SYMBOL(cl_page_get);
 358
 359 /**
 360  * Releases a reference to a page.
 361  *
 362  * When last reference is released, page is returned to the cache, unless it
 363  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 364  * destroyed.
 365  *
 366  * \see cl_object_put(), cl_lock_put().
 367  */
 368 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 369 {
 370         ENTRY;
 371         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 372                        atomic_read(&page->cp_ref));
 373
 374         if (atomic_dec_and_test(&page->cp_ref)) {
 375                 LASSERT(page->cp_state == CPS_FREEING);
 376
 377                 LASSERT(atomic_read(&page->cp_ref) == 0);
 378                 PASSERT(env, page, page->cp_owner == NULL);
 379                 PASSERT(env, page, list_empty(&page->cp_batch));
 380                 /*
 381                  * Page is no longer reachable by other threads. Tear
 382                  * it down.
 383                  */
 384                 cl_page_free(env, page);
 385         }
 386
 387         EXIT;
 388 }
 389 EXPORT_SYMBOL(cl_page_put);
 390
 391 /**
 392  * Returns a cl_page associated with a VM page, and given cl_object.
 393  */
 394 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj)
 395 {
 396         struct cl_page *page;
 397
 398         ENTRY;
 399         KLASSERT(PageLocked(vmpage));
 400
 401         /*
 402          * NOTE: absence of races and liveness of data are guaranteed by page
 403          *       lock on a "vmpage". That works because object destruction has
 404          *       bottom-to-top pass.
 405          */
 406
 407         page = (struct cl_page *)vmpage->private;
 408         if (page != NULL) {
 409                 cl_page_get_trust(page);
 410                 LASSERT(page->cp_type == CPT_CACHEABLE);
 411         }
 412         RETURN(page);
 413 }
 414 EXPORT_SYMBOL(cl_vmpage_page);
 415
 416 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 417                                        const struct lu_device_type *dtype)
 418 {
 419         return cl_page_at_trusted(page, dtype);
 420 }
 421 EXPORT_SYMBOL(cl_page_at);
 422
 423 #define CL_PAGE_OP(opname) offsetof(struct cl_page_operations, opname)
 424
 425 #define CL_PAGE_INVOKE(_env, _page, _op, _proto, ...)                   \
 426 ({                                                                      \
 427         const struct lu_env        *__env  = (_env);                    \
 428         struct cl_page             *__page = (_page);                   \
 429         const struct cl_page_slice *__scan;                             \
 430         int                         __result;                           \
 431         ptrdiff_t                   __op   = (_op);                     \
 432         int                        (*__method)_proto;                   \
 433                                                                         \
 434         __result = 0;                                                   \
 435         list_for_each_entry(__scan, &__page->cp_layers, cpl_linkage) {  \
 436                 __method = *(void **)((char *)__scan->cpl_ops +  __op); \
 437                 if (__method != NULL) {                                 \
 438                         __result = (*__method)(__env, __scan, ## __VA_ARGS__); \
 439                         if (__result != 0)                              \
 440                                 break;                                  \
 441                 }                                                       \
 442         }                                                               \
 443         if (__result > 0)                                               \
 444                 __result = 0;                                           \
 445         __result;                                                       \
 446 })
 447
 448 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 449 do {                                                                    \
 450         const struct lu_env        *__env  = (_env);                    \
 451         struct cl_page             *__page = (_page);                   \
 452         const struct cl_page_slice *__scan;                             \
 453         ptrdiff_t                   __op   = (_op);                     \
 454         void                      (*__method)_proto;                    \
 455                                                                         \
 456         list_for_each_entry(__scan, &__page->cp_layers, cpl_linkage) {  \
 457                 __method = *(void **)((char *)__scan->cpl_ops +  __op); \
 458                 if (__method != NULL)                                   \
 459                         (*__method)(__env, __scan, ## __VA_ARGS__);     \
 460         }                                                               \
 461 } while (0)
 462
 463 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)           \
 464 do {                                                                    \
 465         const struct lu_env        *__env  = (_env);                    \
 466         struct cl_page             *__page = (_page);                   \
 467         const struct cl_page_slice *__scan;                             \
 468         ptrdiff_t                   __op   = (_op);                     \
 469         void                      (*__method)_proto;                    \
 470                                                                         \
 471         /* get to the bottom page. */                                   \
 472         list_for_each_entry_reverse(__scan, &__page->cp_layers,         \
 473                                     cpl_linkage) {                      \
 474                 __method = *(void **)((char *)__scan->cpl_ops + __op);  \
 475                 if (__method != NULL)                                   \
 476                         (*__method)(__env, __scan, ## __VA_ARGS__);     \
 477         }                                                               \
 478 } while (0)
 479
 480 static int cl_page_invoke(const struct lu_env *env,
 481                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 482
 483 {
 484         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 485         ENTRY;
 486         RETURN(CL_PAGE_INVOKE(env, page, op,
 487                               (const struct lu_env *,
 488                                const struct cl_page_slice *, struct cl_io *),
 489                               io));
 490 }
 491
 492 static void cl_page_invoid(const struct lu_env *env,
 493                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 494
 495 {
 496         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 497         ENTRY;
 498         CL_PAGE_INVOID(env, page, op,
 499                        (const struct lu_env *,
 500                         const struct cl_page_slice *, struct cl_io *), io);
 501         EXIT;
 502 }
 503
 504 static void cl_page_owner_clear(struct cl_page *page)
 505 {
 506         ENTRY;
 507         if (page->cp_owner != NULL) {
 508                 LASSERT(page->cp_owner->ci_owned_nr > 0);
 509                 page->cp_owner->ci_owned_nr--;
 510                 page->cp_owner = NULL;
 511         }
 512         EXIT;
 513 }
 514
 515 static void cl_page_owner_set(struct cl_page *page)
 516 {
 517         ENTRY;
 518         LASSERT(page->cp_owner != NULL);
 519         page->cp_owner->ci_owned_nr++;
 520         EXIT;
 521 }
 522
 523 void cl_page_disown0(const struct lu_env *env,
 524                      struct cl_io *io, struct cl_page *pg)
 525 {
 526         enum cl_page_state state;
 527
 528         ENTRY;
 529         state = pg->cp_state;
 530         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 531         PINVRNT(env, pg, cl_page_invariant(pg) || state == CPS_FREEING);
 532         cl_page_owner_clear(pg);
 533
 534         if (state == CPS_OWNED)
 535                 cl_page_state_set(env, pg, CPS_CACHED);
 536         /*
 537          * Completion call-backs are executed in the bottom-up order, so that
 538          * uppermost layer (llite), responsible for VFS/VM interaction runs
 539          * last and can release locks safely.
 540          */
 541         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 542                                (const struct lu_env *,
 543                                 const struct cl_page_slice *, struct cl_io *),
 544                                io);
 545         EXIT;
 546 }
 547
 548 /**
 549  * returns true, iff page is owned by the given io.
 550  */
 551 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 552 {
 553         struct cl_io *top = cl_io_top((struct cl_io *)io);
 554         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 555         ENTRY;
 556         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top);
 557 }
 558 EXPORT_SYMBOL(cl_page_is_owned);
 559
 560 /**
 561  * Try to own a page by IO.
 562  *
 563  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 564  * into cl_page_state::CPS_OWNED state.
 565  *
 566  * \pre  !cl_page_is_owned(pg, io)
 567  * \post result == 0 iff cl_page_is_owned(pg, io)
 568  *
 569  * \retval 0   success
 570  *
 571  * \retval -ve failure, e.g., page was destroyed (and landed in
 572  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 573  *             or, page was owned by another thread, or in IO.
 574  *
 575  * \see cl_page_disown()
 576  * \see cl_page_operations::cpo_own()
 577  * \see cl_page_own_try()
 578  * \see cl_page_own
 579  */
 580 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 581                         struct cl_page *pg, int nonblock)
 582 {
 583         int result;
 584
 585         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 586
 587         ENTRY;
 588         io = cl_io_top(io);
 589
 590         if (pg->cp_state == CPS_FREEING) {
 591                 result = -ENOENT;
 592         } else {
 593                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 594                                         (const struct lu_env *,
 595                                          const struct cl_page_slice *,
 596                                          struct cl_io *, int),
 597                                         io, nonblock);
 598                 if (result == 0) {
 599                         PASSERT(env, pg, pg->cp_owner == NULL);
 600                         pg->cp_owner = cl_io_top(io);;
 601                         cl_page_owner_set(pg);
 602                         if (pg->cp_state != CPS_FREEING) {
 603                                 cl_page_state_set(env, pg, CPS_OWNED);
 604                         } else {
 605                                 cl_page_disown0(env, io, pg);
 606                                 result = -ENOENT;
 607                         }
 608                 }
 609         }
 610         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 611         RETURN(result);
 612 }
 613
 614 /**
 615  * Own a page, might be blocked.
 616  *
 617  * \see cl_page_own0()
 618  */
 619 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 620 {
 621         return cl_page_own0(env, io, pg, 0);
 622 }
 623 EXPORT_SYMBOL(cl_page_own);
 624
 625 /**
 626  * Nonblock version of cl_page_own().
 627  *
 628  * \see cl_page_own0()
 629  */
 630 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
 631                     struct cl_page *pg)
 632 {
 633         return cl_page_own0(env, io, pg, 1);
 634 }
 635 EXPORT_SYMBOL(cl_page_own_try);
 636
 637
 638 /**
 639  * Assume page ownership.
 640  *
 641  * Called when page is already locked by the hosting VM.
 642  *
 643  * \pre !cl_page_is_owned(pg, io)
 644  * \post cl_page_is_owned(pg, io)
 645  *
 646  * \see cl_page_operations::cpo_assume()
 647  */
 648 void cl_page_assume(const struct lu_env *env,
 649                     struct cl_io *io, struct cl_page *pg)
 650 {
 651         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
 652
 653         ENTRY;
 654         io = cl_io_top(io);
 655
 656         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
 657         PASSERT(env, pg, pg->cp_owner == NULL);
 658         pg->cp_owner = cl_io_top(io);
 659         cl_page_owner_set(pg);
 660         cl_page_state_set(env, pg, CPS_OWNED);
 661         EXIT;
 662 }
 663 EXPORT_SYMBOL(cl_page_assume);
 664
 665 /**
 666  * Releases page ownership without unlocking the page.
 667  *
 668  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
 669  * underlying VM page (as VM is supposed to do this itself).
 670  *
 671  * \pre   cl_page_is_owned(pg, io)
 672  * \post !cl_page_is_owned(pg, io)
 673  *
 674  * \see cl_page_assume()
 675  */
 676 void cl_page_unassume(const struct lu_env *env,
 677                       struct cl_io *io, struct cl_page *pg)
 678 {
 679         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 680         PINVRNT(env, pg, cl_page_invariant(pg));
 681
 682         ENTRY;
 683         io = cl_io_top(io);
 684         cl_page_owner_clear(pg);
 685         cl_page_state_set(env, pg, CPS_CACHED);
 686         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
 687                                (const struct lu_env *,
 688                                 const struct cl_page_slice *, struct cl_io *),
 689                                io);
 690         EXIT;
 691 }
 692 EXPORT_SYMBOL(cl_page_unassume);
 693
 694 /**
 695  * Releases page ownership.
 696  *
 697  * Moves page into cl_page_state::CPS_CACHED.
 698  *
 699  * \pre   cl_page_is_owned(pg, io)
 700  * \post !cl_page_is_owned(pg, io)
 701  *
 702  * \see cl_page_own()
 703  * \see cl_page_operations::cpo_disown()
 704  */
 705 void cl_page_disown(const struct lu_env *env,
 706                     struct cl_io *io, struct cl_page *pg)
 707 {
 708         PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
 709                 pg->cp_state == CPS_FREEING);
 710
 711         ENTRY;
 712         io = cl_io_top(io);
 713         cl_page_disown0(env, io, pg);
 714         EXIT;
 715 }
 716 EXPORT_SYMBOL(cl_page_disown);
 717
 718 /**
 719  * Called when page is to be removed from the object, e.g., as a result of
 720  * truncate.
 721  *
 722  * Calls cl_page_operations::cpo_discard() top-to-bottom.
 723  *
 724  * \pre cl_page_is_owned(pg, io)
 725  *
 726  * \see cl_page_operations::cpo_discard()
 727  */
 728 void cl_page_discard(const struct lu_env *env,
 729                      struct cl_io *io, struct cl_page *pg)
 730 {
 731         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 732         PINVRNT(env, pg, cl_page_invariant(pg));
 733
 734         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
 735 }
 736 EXPORT_SYMBOL(cl_page_discard);
 737
 738 /**
 739  * Version of cl_page_delete() that can be called for not fully constructed
 740  * pages, e.g. in an error handling cl_page_find()->cl_page_delete0()
 741  * path. Doesn't check page invariant.
 742  */
 743 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg)
 744 {
 745         ENTRY;
 746
 747         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
 748
 749         /*
 750          * Severe all ways to obtain new pointers to @pg.
 751          */
 752         cl_page_owner_clear(pg);
 753
 754         cl_page_state_set0(env, pg, CPS_FREEING);
 755
 756         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_delete),
 757                        (const struct lu_env *, const struct cl_page_slice *));
 758
 759         EXIT;
 760 }
 761
 762 /**
 763  * Called when a decision is made to throw page out of memory.
 764  *
 765  * Notifies all layers about page destruction by calling
 766  * cl_page_operations::cpo_delete() method top-to-bottom.
 767  *
 768  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
 769  * where transition to this state happens).
 770  *
 771  * Eliminates all venues through which new references to the page can be
 772  * obtained:
 773  *
 774  *     - removes page from the radix trees,
 775  *
 776  *     - breaks linkage from VM page to cl_page.
 777  *
 778  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
 779  * drain after some time, at which point page will be recycled.
 780  *
 781  * \pre  VM page is locked
 782  * \post pg->cp_state == CPS_FREEING
 783  *
 784  * \see cl_page_operations::cpo_delete()
 785  */
 786 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
 787 {
 788         PINVRNT(env, pg, cl_page_invariant(pg));
 789         ENTRY;
 790         cl_page_delete0(env, pg);
 791         EXIT;
 792 }
 793 EXPORT_SYMBOL(cl_page_delete);
 794
 795 /**
 796  * Marks page up-to-date.
 797  *
 798  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
 799  * layer responsible for VM interaction has to mark/clear page as up-to-date
 800  * by the \a uptodate argument.
 801  *
 802  * \see cl_page_operations::cpo_export()
 803  */
 804 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
 805 {
 806         PINVRNT(env, pg, cl_page_invariant(pg));
 807         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
 808                        (const struct lu_env *,
 809                         const struct cl_page_slice *, int), uptodate);
 810 }
 811 EXPORT_SYMBOL(cl_page_export);
 812
 813 /**
 814  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
 815  * thread.
 816  */
 817 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
 818 {
 819         int result;
 820         const struct cl_page_slice *slice;
 821
 822         ENTRY;
 823         slice = container_of(pg->cp_layers.next,
 824                              const struct cl_page_slice, cpl_linkage);
 825         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
 826         /*
 827          * Call ->cpo_is_vmlocked() directly instead of going through
 828          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
 829          * cl_page_invariant().
 830          */
 831         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
 832         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
 833         RETURN(result == -EBUSY);
 834 }
 835 EXPORT_SYMBOL(cl_page_is_vmlocked);
 836
 837 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
 838 {
 839         ENTRY;
 840         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
 841 }
 842
 843 static void cl_page_io_start(const struct lu_env *env,
 844                              struct cl_page *pg, enum cl_req_type crt)
 845 {
 846         /*
 847          * Page is queued for IO, change its state.
 848          */
 849         ENTRY;
 850         cl_page_owner_clear(pg);
 851         cl_page_state_set(env, pg, cl_req_type_state(crt));
 852         EXIT;
 853 }
 854
 855 /**
 856  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
 857  * called top-to-bottom. Every layer either agrees to submit this page (by
 858  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
 859  * handling interactions with the VM also has to inform VM that page is under
 860  * transfer now.
 861  */
 862 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
 863                  struct cl_page *pg, enum cl_req_type crt)
 864 {
 865         int result;
 866
 867         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 868         PINVRNT(env, pg, cl_page_invariant(pg));
 869         PINVRNT(env, pg, crt < CRT_NR);
 870
 871         /*
 872          * XXX this has to be called bottom-to-top, so that llite can set up
 873          * PG_writeback without risking other layers deciding to skip this
 874          * page.
 875          */
 876         if (crt >= CRT_NR)
 877                 return -EINVAL;
 878         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
 879         if (result == 0)
 880                 cl_page_io_start(env, pg, crt);
 881
 882         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 883         return result;
 884 }
 885 EXPORT_SYMBOL(cl_page_prep);
 886
 887 /**
 888  * Notify layers about transfer completion.
 889  *
 890  * Invoked by transfer sub-system (which is a part of osc) to notify layers
 891  * that a transfer, of which this page is a part of has completed.
 892  *
 893  * Completion call-backs are executed in the bottom-up order, so that
 894  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
 895  * and can release locks safely.
 896  *
 897  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 898  * \post pg->cp_state == CPS_CACHED
 899  *
 900  * \see cl_page_operations::cpo_completion()
 901  */
 902 void cl_page_completion(const struct lu_env *env,
 903                         struct cl_page *pg, enum cl_req_type crt, int ioret)
 904 {
 905         struct cl_sync_io *anchor = pg->cp_sync_io;
 906
 907         PASSERT(env, pg, crt < CRT_NR);
 908         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
 909
 910         ENTRY;
 911         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
 912         cl_page_state_set(env, pg, CPS_CACHED);
 913         if (crt >= CRT_NR)
 914                 return;
 915         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
 916                                (const struct lu_env *,
 917                                 const struct cl_page_slice *, int), ioret);
 918         if (anchor != NULL) {
 919                 LASSERT(pg->cp_sync_io == anchor);
 920                 pg->cp_sync_io = NULL;
 921                 cl_sync_io_note(env, anchor, ioret);
 922         }
 923         EXIT;
 924 }
 925 EXPORT_SYMBOL(cl_page_completion);
 926
 927 /**
 928  * Notify layers that transfer formation engine decided to yank this page from
 929  * the cache and to make it a part of a transfer.
 930  *
 931  * \pre  pg->cp_state == CPS_CACHED
 932  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 933  *
 934  * \see cl_page_operations::cpo_make_ready()
 935  */
 936 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
 937                        enum cl_req_type crt)
 938 {
 939         int result;
 940
 941         PINVRNT(env, pg, crt < CRT_NR);
 942
 943         ENTRY;
 944         if (crt >= CRT_NR)
 945                 RETURN(-EINVAL);
 946         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
 947                                 (const struct lu_env *,
 948                                  const struct cl_page_slice *));
 949         if (result == 0) {
 950                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
 951                 cl_page_io_start(env, pg, crt);
 952         }
 953         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 954         RETURN(result);
 955 }
 956 EXPORT_SYMBOL(cl_page_make_ready);
 957
 958 /**
 959  * Called if a pge is being written back by kernel's intention.
 960  *
 961  * \pre  cl_page_is_owned(pg, io)
 962  * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT)
 963  *
 964  * \see cl_page_operations::cpo_flush()
 965  */
 966 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
 967                   struct cl_page *pg)
 968 {
 969         int result;
 970
 971         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 972         PINVRNT(env, pg, cl_page_invariant(pg));
 973
 974         ENTRY;
 975
 976         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_flush));
 977
 978         CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result);
 979         RETURN(result);
 980 }
 981 EXPORT_SYMBOL(cl_page_flush);
 982
 983 /**
 984  * Tells transfer engine that only part of a page is to be transmitted.
 985  *
 986  * \see cl_page_operations::cpo_clip()
 987  */
 988 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
 989                   int from, int to)
 990 {
 991         PINVRNT(env, pg, cl_page_invariant(pg));
 992
 993         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
 994         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
 995                        (const struct lu_env *,
 996                         const struct cl_page_slice *,int, int),
 997                        from, to);
 998 }
 999 EXPORT_SYMBOL(cl_page_clip);
1000
1001 /**
1002  * Prints human readable representation of \a pg to the \a f.
1003  */
1004 void cl_page_header_print(const struct lu_env *env, void *cookie,
1005                           lu_printer_t printer, const struct cl_page *pg)
1006 {
1007         (*printer)(env, cookie,
1008                    "page@%p[%d %p %d %d %p]\n",
1009                    pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1010                    pg->cp_state, pg->cp_type,
1011                    pg->cp_owner);
1012 }
1013 EXPORT_SYMBOL(cl_page_header_print);
1014
1015 /**
1016  * Prints human readable representation of \a pg to the \a f.
1017  */
1018 void cl_page_print(const struct lu_env *env, void *cookie,
1019                    lu_printer_t printer, const struct cl_page *pg)
1020 {
1021         cl_page_header_print(env, cookie, printer, pg);
1022         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1023                        (const struct lu_env *env,
1024                         const struct cl_page_slice *slice,
1025                         void *cookie, lu_printer_t p), cookie, printer);
1026         (*printer)(env, cookie, "end page@%p\n", pg);
1027 }
1028 EXPORT_SYMBOL(cl_page_print);
1029
1030 /**
1031  * Cancel a page which is still in a transfer.
1032  */
1033 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1034 {
1035         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1036                               (const struct lu_env *,
1037                                const struct cl_page_slice *));
1038 }
1039
1040 /**
1041  * Converts a byte offset within object \a obj into a page index.
1042  */
1043 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1044 {
1045         return (loff_t)idx << PAGE_SHIFT;
1046 }
1047 EXPORT_SYMBOL(cl_offset);
1048
1049 /**
1050  * Converts a page index into a byte offset within object \a obj.
1051  */
1052 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1053 {
1054         return offset >> PAGE_SHIFT;
1055 }
1056 EXPORT_SYMBOL(cl_index);
1057
1058 size_t cl_page_size(const struct cl_object *obj)
1059 {
1060         return 1UL << PAGE_SHIFT;
1061 }
1062 EXPORT_SYMBOL(cl_page_size);
1063
1064 /**
1065  * Adds page slice to the compound page.
1066  *
1067  * This is called by cl_object_operations::coo_page_init() methods to add a
1068  * per-layer state to the page. New state is added at the end of
1069  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1070  *
1071  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1072  */
1073 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1074                        struct cl_object *obj, pgoff_t index,
1075                        const struct cl_page_operations *ops)
1076 {
1077         ENTRY;
1078         list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1079         slice->cpl_obj  = obj;
1080         slice->cpl_index = index;
1081         slice->cpl_ops  = ops;
1082         slice->cpl_page = page;
1083         EXIT;
1084 }
1085 EXPORT_SYMBOL(cl_page_slice_add);
1086
1087 /**
1088  * Allocate and initialize cl_cache, called by ll_init_sbi().
1089  */
1090 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max)
1091 {
1092         struct cl_client_cache  *cache = NULL;
1093
1094         ENTRY;
1095         OBD_ALLOC(cache, sizeof(*cache));
1096         if (cache == NULL)
1097                 RETURN(NULL);
1098
1099         /* Initialize cache data */
1100         atomic_set(&cache->ccc_users, 1);
1101         cache->ccc_lru_max = lru_page_max;
1102         atomic_long_set(&cache->ccc_lru_left, lru_page_max);
1103         spin_lock_init(&cache->ccc_lru_lock);
1104         INIT_LIST_HEAD(&cache->ccc_lru);
1105
1106         /* turn unstable check off by default as it impacts performance */
1107         cache->ccc_unstable_check = 0;
1108         atomic_long_set(&cache->ccc_unstable_nr, 0);
1109         init_waitqueue_head(&cache->ccc_unstable_waitq);
1110
1111         RETURN(cache);
1112 }
1113 EXPORT_SYMBOL(cl_cache_init);
1114
1115 /**
1116  * Increase cl_cache refcount
1117  */
1118 void cl_cache_incref(struct cl_client_cache *cache)
1119 {
1120         atomic_inc(&cache->ccc_users);
1121 }
1122 EXPORT_SYMBOL(cl_cache_incref);
1123
1124 /**
1125  * Decrease cl_cache refcount and free the cache if refcount=0.
1126  * Since llite, lov and osc all hold cl_cache refcount,
1127  * the free will not cause race. (LU-6173)
1128  */
1129 void cl_cache_decref(struct cl_client_cache *cache)
1130 {
1131         if (atomic_dec_and_test(&cache->ccc_users))
1132                 OBD_FREE(cache, sizeof(*cache));
1133 }
1134 EXPORT_SYMBOL(cl_cache_decref);