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_INVOKE_REVERSE(_env, _page, _op, _proto, ...)           \
 455 ({                                                                      \
 456         const struct lu_env        *__env  = (_env);                    \
 457         struct cl_page             *__page = (_page);                   \
 458         const struct cl_page_slice *__scan;                             \
 459         int                         __result;                           \
 460         ptrdiff_t                   __op   = (_op);                     \
 461         int                       (*__method)_proto;                    \
 462                                                                         \
 463         __result = 0;                                                   \
 464         list_for_each_entry_reverse(__scan, &__page->cp_layers,         \
 465                                     cpl_linkage) {                      \
 466                 __method = *(void **)((char *)__scan->cpl_ops +  __op); \
 467                 if (__method != NULL) {                                 \
 468                         __result = (*__method)(__env, __scan, ## __VA_ARGS__); \
 469                         if (__result != 0)                              \
 470                                 break;                                  \
 471                 }                                                       \
 472         }                                                               \
 473         if (__result > 0)                                               \
 474                 __result = 0;                                           \
 475         __result;                                                       \
 476 })
 477
 478 #define CL_PAGE_INVOID(_env, _page, _op, _proto, ...)                   \
 479 do {                                                                    \
 480         const struct lu_env        *__env  = (_env);                    \
 481         struct cl_page             *__page = (_page);                   \
 482         const struct cl_page_slice *__scan;                             \
 483         ptrdiff_t                   __op   = (_op);                     \
 484         void                      (*__method)_proto;                    \
 485                                                                         \
 486         list_for_each_entry(__scan, &__page->cp_layers, cpl_linkage) {  \
 487                 __method = *(void **)((char *)__scan->cpl_ops +  __op); \
 488                 if (__method != NULL)                                   \
 489                         (*__method)(__env, __scan, ## __VA_ARGS__);     \
 490         }                                                               \
 491 } while (0)
 492
 493 #define CL_PAGE_INVOID_REVERSE(_env, _page, _op, _proto, ...)           \
 494 do {                                                                    \
 495         const struct lu_env        *__env  = (_env);                    \
 496         struct cl_page             *__page = (_page);                   \
 497         const struct cl_page_slice *__scan;                             \
 498         ptrdiff_t                   __op   = (_op);                     \
 499         void                      (*__method)_proto;                    \
 500                                                                         \
 501         /* get to the bottom page. */                                   \
 502         list_for_each_entry_reverse(__scan, &__page->cp_layers,         \
 503                                     cpl_linkage) {                      \
 504                 __method = *(void **)((char *)__scan->cpl_ops + __op);  \
 505                 if (__method != NULL)                                   \
 506                         (*__method)(__env, __scan, ## __VA_ARGS__);     \
 507         }                                                               \
 508 } while (0)
 509
 510 static int cl_page_invoke(const struct lu_env *env,
 511                           struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 512
 513 {
 514         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 515         ENTRY;
 516         RETURN(CL_PAGE_INVOKE(env, page, op,
 517                               (const struct lu_env *,
 518                                const struct cl_page_slice *, struct cl_io *),
 519                               io));
 520 }
 521
 522 static void cl_page_invoid(const struct lu_env *env,
 523                            struct cl_io *io, struct cl_page *page, ptrdiff_t op)
 524
 525 {
 526         PINVRNT(env, page, cl_object_same(page->cp_obj, io->ci_obj));
 527         ENTRY;
 528         CL_PAGE_INVOID(env, page, op,
 529                        (const struct lu_env *,
 530                         const struct cl_page_slice *, struct cl_io *), io);
 531         EXIT;
 532 }
 533
 534 static void cl_page_owner_clear(struct cl_page *page)
 535 {
 536         ENTRY;
 537         if (page->cp_owner != NULL) {
 538                 LASSERT(page->cp_owner->ci_owned_nr > 0);
 539                 page->cp_owner->ci_owned_nr--;
 540                 page->cp_owner = NULL;
 541         }
 542         EXIT;
 543 }
 544
 545 static void cl_page_owner_set(struct cl_page *page)
 546 {
 547         ENTRY;
 548         LASSERT(page->cp_owner != NULL);
 549         page->cp_owner->ci_owned_nr++;
 550         EXIT;
 551 }
 552
 553 void cl_page_disown0(const struct lu_env *env,
 554                      struct cl_io *io, struct cl_page *pg)
 555 {
 556         enum cl_page_state state;
 557
 558         ENTRY;
 559         state = pg->cp_state;
 560         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 561         PINVRNT(env, pg, cl_page_invariant(pg) || state == CPS_FREEING);
 562         cl_page_owner_clear(pg);
 563
 564         if (state == CPS_OWNED)
 565                 cl_page_state_set(env, pg, CPS_CACHED);
 566         /*
 567          * Completion call-backs are executed in the bottom-up order, so that
 568          * uppermost layer (llite), responsible for VFS/VM interaction runs
 569          * last and can release locks safely.
 570          */
 571         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_disown),
 572                                (const struct lu_env *,
 573                                 const struct cl_page_slice *, struct cl_io *),
 574                                io);
 575         EXIT;
 576 }
 577
 578 /**
 579  * returns true, iff page is owned by the given io.
 580  */
 581 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 582 {
 583         struct cl_io *top = cl_io_top((struct cl_io *)io);
 584         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 585         ENTRY;
 586         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top);
 587 }
 588 EXPORT_SYMBOL(cl_page_is_owned);
 589
 590 /**
 591  * Try to own a page by IO.
 592  *
 593  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 594  * into cl_page_state::CPS_OWNED state.
 595  *
 596  * \pre  !cl_page_is_owned(pg, io)
 597  * \post result == 0 iff cl_page_is_owned(pg, io)
 598  *
 599  * \retval 0   success
 600  *
 601  * \retval -ve failure, e.g., page was destroyed (and landed in
 602  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 603  *             or, page was owned by another thread, or in IO.
 604  *
 605  * \see cl_page_disown()
 606  * \see cl_page_operations::cpo_own()
 607  * \see cl_page_own_try()
 608  * \see cl_page_own
 609  */
 610 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 611                         struct cl_page *pg, int nonblock)
 612 {
 613         int result;
 614
 615         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 616
 617         ENTRY;
 618         io = cl_io_top(io);
 619
 620         if (pg->cp_state == CPS_FREEING) {
 621                 result = -ENOENT;
 622         } else {
 623                 result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(cpo_own),
 624                                         (const struct lu_env *,
 625                                          const struct cl_page_slice *,
 626                                          struct cl_io *, int),
 627                                         io, nonblock);
 628                 if (result == 0) {
 629                         PASSERT(env, pg, pg->cp_owner == NULL);
 630                         PASSERT(env, pg, pg->cp_req == NULL);
 631                         pg->cp_owner = cl_io_top(io);;
 632                         cl_page_owner_set(pg);
 633                         if (pg->cp_state != CPS_FREEING) {
 634                                 cl_page_state_set(env, pg, CPS_OWNED);
 635                         } else {
 636                                 cl_page_disown0(env, io, pg);
 637                                 result = -ENOENT;
 638                         }
 639                 }
 640         }
 641         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 642         RETURN(result);
 643 }
 644
 645 /**
 646  * Own a page, might be blocked.
 647  *
 648  * \see cl_page_own0()
 649  */
 650 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 651 {
 652         return cl_page_own0(env, io, pg, 0);
 653 }
 654 EXPORT_SYMBOL(cl_page_own);
 655
 656 /**
 657  * Nonblock version of cl_page_own().
 658  *
 659  * \see cl_page_own0()
 660  */
 661 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
 662                     struct cl_page *pg)
 663 {
 664         return cl_page_own0(env, io, pg, 1);
 665 }
 666 EXPORT_SYMBOL(cl_page_own_try);
 667
 668
 669 /**
 670  * Assume page ownership.
 671  *
 672  * Called when page is already locked by the hosting VM.
 673  *
 674  * \pre !cl_page_is_owned(pg, io)
 675  * \post cl_page_is_owned(pg, io)
 676  *
 677  * \see cl_page_operations::cpo_assume()
 678  */
 679 void cl_page_assume(const struct lu_env *env,
 680                     struct cl_io *io, struct cl_page *pg)
 681 {
 682         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
 683
 684         ENTRY;
 685         io = cl_io_top(io);
 686
 687         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_assume));
 688         PASSERT(env, pg, pg->cp_owner == NULL);
 689         pg->cp_owner = cl_io_top(io);
 690         cl_page_owner_set(pg);
 691         cl_page_state_set(env, pg, CPS_OWNED);
 692         EXIT;
 693 }
 694 EXPORT_SYMBOL(cl_page_assume);
 695
 696 /**
 697  * Releases page ownership without unlocking the page.
 698  *
 699  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
 700  * underlying VM page (as VM is supposed to do this itself).
 701  *
 702  * \pre   cl_page_is_owned(pg, io)
 703  * \post !cl_page_is_owned(pg, io)
 704  *
 705  * \see cl_page_assume()
 706  */
 707 void cl_page_unassume(const struct lu_env *env,
 708                       struct cl_io *io, struct cl_page *pg)
 709 {
 710         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 711         PINVRNT(env, pg, cl_page_invariant(pg));
 712
 713         ENTRY;
 714         io = cl_io_top(io);
 715         cl_page_owner_clear(pg);
 716         cl_page_state_set(env, pg, CPS_CACHED);
 717         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_unassume),
 718                                (const struct lu_env *,
 719                                 const struct cl_page_slice *, struct cl_io *),
 720                                io);
 721         EXIT;
 722 }
 723 EXPORT_SYMBOL(cl_page_unassume);
 724
 725 /**
 726  * Releases page ownership.
 727  *
 728  * Moves page into cl_page_state::CPS_CACHED.
 729  *
 730  * \pre   cl_page_is_owned(pg, io)
 731  * \post !cl_page_is_owned(pg, io)
 732  *
 733  * \see cl_page_own()
 734  * \see cl_page_operations::cpo_disown()
 735  */
 736 void cl_page_disown(const struct lu_env *env,
 737                     struct cl_io *io, struct cl_page *pg)
 738 {
 739         PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
 740                 pg->cp_state == CPS_FREEING);
 741
 742         ENTRY;
 743         io = cl_io_top(io);
 744         cl_page_disown0(env, io, pg);
 745         EXIT;
 746 }
 747 EXPORT_SYMBOL(cl_page_disown);
 748
 749 /**
 750  * Called when page is to be removed from the object, e.g., as a result of
 751  * truncate.
 752  *
 753  * Calls cl_page_operations::cpo_discard() top-to-bottom.
 754  *
 755  * \pre cl_page_is_owned(pg, io)
 756  *
 757  * \see cl_page_operations::cpo_discard()
 758  */
 759 void cl_page_discard(const struct lu_env *env,
 760                      struct cl_io *io, struct cl_page *pg)
 761 {
 762         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 763         PINVRNT(env, pg, cl_page_invariant(pg));
 764
 765         cl_page_invoid(env, io, pg, CL_PAGE_OP(cpo_discard));
 766 }
 767 EXPORT_SYMBOL(cl_page_discard);
 768
 769 /**
 770  * Version of cl_page_delete() that can be called for not fully constructed
 771  * pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
 772  * path. Doesn't check page invariant.
 773  */
 774 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg)
 775 {
 776         ENTRY;
 777
 778         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
 779
 780         /*
 781          * Severe all ways to obtain new pointers to @pg.
 782          */
 783         cl_page_owner_clear(pg);
 784
 785         cl_page_state_set0(env, pg, CPS_FREEING);
 786
 787         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_delete),
 788                        (const struct lu_env *, const struct cl_page_slice *));
 789
 790         EXIT;
 791 }
 792
 793 /**
 794  * Called when a decision is made to throw page out of memory.
 795  *
 796  * Notifies all layers about page destruction by calling
 797  * cl_page_operations::cpo_delete() method top-to-bottom.
 798  *
 799  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
 800  * where transition to this state happens).
 801  *
 802  * Eliminates all venues through which new references to the page can be
 803  * obtained:
 804  *
 805  *     - removes page from the radix trees,
 806  *
 807  *     - breaks linkage from VM page to cl_page.
 808  *
 809  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
 810  * drain after some time, at which point page will be recycled.
 811  *
 812  * \pre  VM page is locked
 813  * \post pg->cp_state == CPS_FREEING
 814  *
 815  * \see cl_page_operations::cpo_delete()
 816  */
 817 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
 818 {
 819         PINVRNT(env, pg, cl_page_invariant(pg));
 820         ENTRY;
 821         cl_page_delete0(env, pg);
 822         EXIT;
 823 }
 824 EXPORT_SYMBOL(cl_page_delete);
 825
 826 /**
 827  * Marks page up-to-date.
 828  *
 829  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
 830  * layer responsible for VM interaction has to mark/clear page as up-to-date
 831  * by the \a uptodate argument.
 832  *
 833  * \see cl_page_operations::cpo_export()
 834  */
 835 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
 836 {
 837         PINVRNT(env, pg, cl_page_invariant(pg));
 838         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_export),
 839                        (const struct lu_env *,
 840                         const struct cl_page_slice *, int), uptodate);
 841 }
 842 EXPORT_SYMBOL(cl_page_export);
 843
 844 /**
 845  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
 846  * thread.
 847  */
 848 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
 849 {
 850         int result;
 851         const struct cl_page_slice *slice;
 852
 853         ENTRY;
 854         slice = container_of(pg->cp_layers.next,
 855                              const struct cl_page_slice, cpl_linkage);
 856         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
 857         /*
 858          * Call ->cpo_is_vmlocked() directly instead of going through
 859          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
 860          * cl_page_invariant().
 861          */
 862         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
 863         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
 864         RETURN(result == -EBUSY);
 865 }
 866 EXPORT_SYMBOL(cl_page_is_vmlocked);
 867
 868 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
 869 {
 870         ENTRY;
 871         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
 872 }
 873
 874 static void cl_page_io_start(const struct lu_env *env,
 875                              struct cl_page *pg, enum cl_req_type crt)
 876 {
 877         /*
 878          * Page is queued for IO, change its state.
 879          */
 880         ENTRY;
 881         cl_page_owner_clear(pg);
 882         cl_page_state_set(env, pg, cl_req_type_state(crt));
 883         EXIT;
 884 }
 885
 886 /**
 887  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
 888  * called top-to-bottom. Every layer either agrees to submit this page (by
 889  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
 890  * handling interactions with the VM also has to inform VM that page is under
 891  * transfer now.
 892  */
 893 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
 894                  struct cl_page *pg, enum cl_req_type crt)
 895 {
 896         int result;
 897
 898         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 899         PINVRNT(env, pg, cl_page_invariant(pg));
 900         PINVRNT(env, pg, crt < CRT_NR);
 901
 902         /*
 903          * XXX this has to be called bottom-to-top, so that llite can set up
 904          * PG_writeback without risking other layers deciding to skip this
 905          * page.
 906          */
 907         if (crt >= CRT_NR)
 908                 return -EINVAL;
 909         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(io[crt].cpo_prep));
 910         if (result == 0)
 911                 cl_page_io_start(env, pg, crt);
 912
 913         KLASSERT(ergo(crt == CRT_WRITE && pg->cp_type == CPT_CACHEABLE,
 914                       equi(result == 0,
 915                            PageWriteback(cl_page_vmpage(pg)))));
 916         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 917         return result;
 918 }
 919 EXPORT_SYMBOL(cl_page_prep);
 920
 921 /**
 922  * Notify layers about transfer completion.
 923  *
 924  * Invoked by transfer sub-system (which is a part of osc) to notify layers
 925  * that a transfer, of which this page is a part of has completed.
 926  *
 927  * Completion call-backs are executed in the bottom-up order, so that
 928  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
 929  * and can release locks safely.
 930  *
 931  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 932  * \post pg->cp_state == CPS_CACHED
 933  *
 934  * \see cl_page_operations::cpo_completion()
 935  */
 936 void cl_page_completion(const struct lu_env *env,
 937                         struct cl_page *pg, enum cl_req_type crt, int ioret)
 938 {
 939         struct cl_sync_io *anchor = pg->cp_sync_io;
 940
 941         PASSERT(env, pg, crt < CRT_NR);
 942         /* cl_page::cp_req already cleared by the caller (osc_completion()) */
 943         PASSERT(env, pg, pg->cp_req == NULL);
 944         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
 945
 946         ENTRY;
 947         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
 948         cl_page_state_set(env, pg, CPS_CACHED);
 949         if (crt >= CRT_NR)
 950                 return;
 951         CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(io[crt].cpo_completion),
 952                                (const struct lu_env *,
 953                                 const struct cl_page_slice *, int), ioret);
 954         if (anchor) {
 955                 LASSERT(pg->cp_sync_io == anchor);
 956                 pg->cp_sync_io = NULL;
 957         }
 958         /*
 959          * As page->cp_obj is pinned by a reference from page->cp_req, it is
 960          * safe to call cl_page_put() without risking object destruction in a
 961          * non-blocking context.
 962          */
 963         cl_page_put(env, pg);
 964
 965         if (anchor != NULL)
 966                 cl_sync_io_note(env, anchor, ioret);
 967
 968         EXIT;
 969 }
 970 EXPORT_SYMBOL(cl_page_completion);
 971
 972 /**
 973  * Notify layers that transfer formation engine decided to yank this page from
 974  * the cache and to make it a part of a transfer.
 975  *
 976  * \pre  pg->cp_state == CPS_CACHED
 977  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 978  *
 979  * \see cl_page_operations::cpo_make_ready()
 980  */
 981 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
 982                        enum cl_req_type crt)
 983 {
 984         int result;
 985
 986         PINVRNT(env, pg, crt < CRT_NR);
 987
 988         ENTRY;
 989         if (crt >= CRT_NR)
 990                 RETURN(-EINVAL);
 991         result = CL_PAGE_INVOKE(env, pg, CL_PAGE_OP(io[crt].cpo_make_ready),
 992                                 (const struct lu_env *,
 993                                  const struct cl_page_slice *));
 994         if (result == 0) {
 995                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
 996                 cl_page_io_start(env, pg, crt);
 997         }
 998         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 999         RETURN(result);
1000 }
1001 EXPORT_SYMBOL(cl_page_make_ready);
1002
1003 /**
1004  * Called if a pge is being written back by kernel's intention.
1005  *
1006  * \pre  cl_page_is_owned(pg, io)
1007  * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT)
1008  *
1009  * \see cl_page_operations::cpo_flush()
1010  */
1011 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
1012                   struct cl_page *pg)
1013 {
1014         int result;
1015
1016         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1017         PINVRNT(env, pg, cl_page_invariant(pg));
1018
1019         ENTRY;
1020
1021         result = cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_flush));
1022
1023         CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result);
1024         RETURN(result);
1025 }
1026 EXPORT_SYMBOL(cl_page_flush);
1027
1028 /**
1029  * Checks whether page is protected by any extent lock is at least required
1030  * mode.
1031  *
1032  * \return the same as in cl_page_operations::cpo_is_under_lock() method.
1033  * \see cl_page_operations::cpo_is_under_lock()
1034  */
1035 int cl_page_is_under_lock(const struct lu_env *env, struct cl_io *io,
1036                           struct cl_page *page, pgoff_t *max_index)
1037 {
1038         int rc;
1039
1040         PINVRNT(env, page, cl_page_invariant(page));
1041
1042         ENTRY;
1043         rc = CL_PAGE_INVOKE_REVERSE(env, page, CL_PAGE_OP(cpo_is_under_lock),
1044                                     (const struct lu_env *,
1045                                      const struct cl_page_slice *,
1046                                      struct cl_io *, pgoff_t *),
1047                                     io, max_index);
1048         RETURN(rc);
1049 }
1050 EXPORT_SYMBOL(cl_page_is_under_lock);
1051
1052 /**
1053  * Tells transfer engine that only part of a page is to be transmitted.
1054  *
1055  * \see cl_page_operations::cpo_clip()
1056  */
1057 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1058                   int from, int to)
1059 {
1060         PINVRNT(env, pg, cl_page_invariant(pg));
1061
1062         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1063         CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_clip),
1064                        (const struct lu_env *,
1065                         const struct cl_page_slice *,int, int),
1066                        from, to);
1067 }
1068 EXPORT_SYMBOL(cl_page_clip);
1069
1070 /**
1071  * Prints human readable representation of \a pg to the \a f.
1072  */
1073 void cl_page_header_print(const struct lu_env *env, void *cookie,
1074                           lu_printer_t printer, const struct cl_page *pg)
1075 {
1076         (*printer)(env, cookie,
1077                    "page@%p[%d %p %d %d %d %p %p]\n",
1078                    pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1079                    pg->cp_state, pg->cp_error, pg->cp_type,
1080                    pg->cp_owner, pg->cp_req);
1081 }
1082 EXPORT_SYMBOL(cl_page_header_print);
1083
1084 /**
1085  * Prints human readable representation of \a pg to the \a f.
1086  */
1087 void cl_page_print(const struct lu_env *env, void *cookie,
1088                    lu_printer_t printer, const struct cl_page *pg)
1089 {
1090         cl_page_header_print(env, cookie, printer, pg);
1091         CL_PAGE_INVOKE(env, (struct cl_page *)pg, CL_PAGE_OP(cpo_print),
1092                        (const struct lu_env *env,
1093                         const struct cl_page_slice *slice,
1094                         void *cookie, lu_printer_t p), cookie, printer);
1095         (*printer)(env, cookie, "end page@%p\n", pg);
1096 }
1097 EXPORT_SYMBOL(cl_page_print);
1098
1099 /**
1100  * Cancel a page which is still in a transfer.
1101  */
1102 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1103 {
1104         return CL_PAGE_INVOKE(env, page, CL_PAGE_OP(cpo_cancel),
1105                               (const struct lu_env *,
1106                                const struct cl_page_slice *));
1107 }
1108
1109 /**
1110  * Converts a byte offset within object \a obj into a page index.
1111  */
1112 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1113 {
1114         return (loff_t)idx << PAGE_CACHE_SHIFT;
1115 }
1116 EXPORT_SYMBOL(cl_offset);
1117
1118 /**
1119  * Converts a page index into a byte offset within object \a obj.
1120  */
1121 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1122 {
1123         return offset >> PAGE_CACHE_SHIFT;
1124 }
1125 EXPORT_SYMBOL(cl_index);
1126
1127 size_t cl_page_size(const struct cl_object *obj)
1128 {
1129         return 1UL << PAGE_CACHE_SHIFT;
1130 }
1131 EXPORT_SYMBOL(cl_page_size);
1132
1133 /**
1134  * Adds page slice to the compound page.
1135  *
1136  * This is called by cl_object_operations::coo_page_init() methods to add a
1137  * per-layer state to the page. New state is added at the end of
1138  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1139  *
1140  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1141  */
1142 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1143                        struct cl_object *obj, pgoff_t index,
1144                        const struct cl_page_operations *ops)
1145 {
1146         ENTRY;
1147         list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1148         slice->cpl_obj  = obj;
1149         slice->cpl_index = index;
1150         slice->cpl_ops  = ops;
1151         slice->cpl_page = page;
1152         EXIT;
1153 }
1154 EXPORT_SYMBOL(cl_page_slice_add);