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 static void cl_page_owner_clear(struct cl_page *page)
 424 {
 425         ENTRY;
 426         if (page->cp_owner != NULL) {
 427                 LASSERT(page->cp_owner->ci_owned_nr > 0);
 428                 page->cp_owner->ci_owned_nr--;
 429                 page->cp_owner = NULL;
 430         }
 431         EXIT;
 432 }
 433
 434 static void cl_page_owner_set(struct cl_page *page)
 435 {
 436         ENTRY;
 437         LASSERT(page->cp_owner != NULL);
 438         page->cp_owner->ci_owned_nr++;
 439         EXIT;
 440 }
 441
 442 void cl_page_disown0(const struct lu_env *env,
 443                      struct cl_io *io, struct cl_page *pg)
 444 {
 445         const struct cl_page_slice *slice;
 446         enum cl_page_state state;
 447
 448         ENTRY;
 449         state = pg->cp_state;
 450         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 451         PINVRNT(env, pg, cl_page_invariant(pg) || state == CPS_FREEING);
 452         cl_page_owner_clear(pg);
 453
 454         if (state == CPS_OWNED)
 455                 cl_page_state_set(env, pg, CPS_CACHED);
 456         /*
 457          * Completion call-backs are executed in the bottom-up order, so that
 458          * uppermost layer (llite), responsible for VFS/VM interaction runs
 459          * last and can release locks safely.
 460          */
 461         list_for_each_entry_reverse(slice, &pg->cp_layers, cpl_linkage) {
 462                 if (slice->cpl_ops->cpo_disown != NULL)
 463                         (*slice->cpl_ops->cpo_disown)(env, slice, io);
 464         }
 465
 466         EXIT;
 467 }
 468
 469 /**
 470  * returns true, iff page is owned by the given io.
 471  */
 472 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 473 {
 474         struct cl_io *top = cl_io_top((struct cl_io *)io);
 475         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 476         ENTRY;
 477         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top);
 478 }
 479 EXPORT_SYMBOL(cl_page_is_owned);
 480
 481 /**
 482  * Try to own a page by IO.
 483  *
 484  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 485  * into cl_page_state::CPS_OWNED state.
 486  *
 487  * \pre  !cl_page_is_owned(pg, io)
 488  * \post result == 0 iff cl_page_is_owned(pg, io)
 489  *
 490  * \retval 0   success
 491  *
 492  * \retval -ve failure, e.g., page was destroyed (and landed in
 493  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 494  *             or, page was owned by another thread, or in IO.
 495  *
 496  * \see cl_page_disown()
 497  * \see cl_page_operations::cpo_own()
 498  * \see cl_page_own_try()
 499  * \see cl_page_own
 500  */
 501 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 502                         struct cl_page *pg, int nonblock)
 503 {
 504         int result = 0;
 505         const struct cl_page_slice *slice;
 506
 507         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 508
 509         ENTRY;
 510         io = cl_io_top(io);
 511
 512         if (pg->cp_state == CPS_FREEING) {
 513                 result = -ENOENT;
 514                 goto out;
 515         }
 516
 517         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
 518                 if (slice->cpl_ops->cpo_own)
 519                         result = (*slice->cpl_ops->cpo_own)(env, slice,
 520                                                             io, nonblock);
 521
 522                 if (result != 0)
 523                         break;
 524
 525         }
 526         if (result > 0)
 527                 result = 0;
 528
 529         if (result == 0) {
 530                 PASSERT(env, pg, pg->cp_owner == NULL);
 531                 pg->cp_owner = cl_io_top(io);
 532                 cl_page_owner_set(pg);
 533                 if (pg->cp_state != CPS_FREEING) {
 534                         cl_page_state_set(env, pg, CPS_OWNED);
 535                 } else {
 536                         cl_page_disown0(env, io, pg);
 537                         result = -ENOENT;
 538                 }
 539         }
 540
 541 out:
 542         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 543         RETURN(result);
 544 }
 545
 546 /**
 547  * Own a page, might be blocked.
 548  *
 549  * \see cl_page_own0()
 550  */
 551 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 552 {
 553         return cl_page_own0(env, io, pg, 0);
 554 }
 555 EXPORT_SYMBOL(cl_page_own);
 556
 557 /**
 558  * Nonblock version of cl_page_own().
 559  *
 560  * \see cl_page_own0()
 561  */
 562 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
 563                     struct cl_page *pg)
 564 {
 565         return cl_page_own0(env, io, pg, 1);
 566 }
 567 EXPORT_SYMBOL(cl_page_own_try);
 568
 569
 570 /**
 571  * Assume page ownership.
 572  *
 573  * Called when page is already locked by the hosting VM.
 574  *
 575  * \pre !cl_page_is_owned(pg, io)
 576  * \post cl_page_is_owned(pg, io)
 577  *
 578  * \see cl_page_operations::cpo_assume()
 579  */
 580 void cl_page_assume(const struct lu_env *env,
 581                     struct cl_io *io, struct cl_page *pg)
 582 {
 583         const struct cl_page_slice *slice;
 584
 585         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
 586
 587         ENTRY;
 588         io = cl_io_top(io);
 589
 590         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
 591                 if (slice->cpl_ops->cpo_assume != NULL)
 592                         (*slice->cpl_ops->cpo_assume)(env, slice, io);
 593         }
 594
 595         PASSERT(env, pg, pg->cp_owner == NULL);
 596         pg->cp_owner = cl_io_top(io);
 597         cl_page_owner_set(pg);
 598         cl_page_state_set(env, pg, CPS_OWNED);
 599         EXIT;
 600 }
 601 EXPORT_SYMBOL(cl_page_assume);
 602
 603 /**
 604  * Releases page ownership without unlocking the page.
 605  *
 606  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
 607  * underlying VM page (as VM is supposed to do this itself).
 608  *
 609  * \pre   cl_page_is_owned(pg, io)
 610  * \post !cl_page_is_owned(pg, io)
 611  *
 612  * \see cl_page_assume()
 613  */
 614 void cl_page_unassume(const struct lu_env *env,
 615                       struct cl_io *io, struct cl_page *pg)
 616 {
 617         const struct cl_page_slice *slice;
 618
 619         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 620         PINVRNT(env, pg, cl_page_invariant(pg));
 621
 622         ENTRY;
 623         io = cl_io_top(io);
 624         cl_page_owner_clear(pg);
 625         cl_page_state_set(env, pg, CPS_CACHED);
 626
 627         list_for_each_entry_reverse(slice, &pg->cp_layers, cpl_linkage) {
 628                 if (slice->cpl_ops->cpo_unassume != NULL)
 629                         (*slice->cpl_ops->cpo_unassume)(env, slice, io);
 630         }
 631
 632         EXIT;
 633 }
 634 EXPORT_SYMBOL(cl_page_unassume);
 635
 636 /**
 637  * Releases page ownership.
 638  *
 639  * Moves page into cl_page_state::CPS_CACHED.
 640  *
 641  * \pre   cl_page_is_owned(pg, io)
 642  * \post !cl_page_is_owned(pg, io)
 643  *
 644  * \see cl_page_own()
 645  * \see cl_page_operations::cpo_disown()
 646  */
 647 void cl_page_disown(const struct lu_env *env,
 648                     struct cl_io *io, struct cl_page *pg)
 649 {
 650         PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
 651                 pg->cp_state == CPS_FREEING);
 652
 653         ENTRY;
 654         io = cl_io_top(io);
 655         cl_page_disown0(env, io, pg);
 656         EXIT;
 657 }
 658 EXPORT_SYMBOL(cl_page_disown);
 659
 660 /**
 661  * Called when page is to be removed from the object, e.g., as a result of
 662  * truncate.
 663  *
 664  * Calls cl_page_operations::cpo_discard() top-to-bottom.
 665  *
 666  * \pre cl_page_is_owned(pg, io)
 667  *
 668  * \see cl_page_operations::cpo_discard()
 669  */
 670 void cl_page_discard(const struct lu_env *env,
 671                      struct cl_io *io, struct cl_page *pg)
 672 {
 673         const struct cl_page_slice *slice;
 674
 675         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 676         PINVRNT(env, pg, cl_page_invariant(pg));
 677
 678         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
 679                 if (slice->cpl_ops->cpo_discard != NULL)
 680                         (*slice->cpl_ops->cpo_discard)(env, slice, io);
 681         }
 682 }
 683 EXPORT_SYMBOL(cl_page_discard);
 684
 685 /**
 686  * Version of cl_page_delete() that can be called for not fully constructed
 687  * pages, e.g. in an error handling cl_page_find()->cl_page_delete0()
 688  * path. Doesn't check page invariant.
 689  */
 690 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg)
 691 {
 692         const struct cl_page_slice *slice;
 693
 694         ENTRY;
 695
 696         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
 697
 698         /*
 699          * Severe all ways to obtain new pointers to @pg.
 700          */
 701         cl_page_owner_clear(pg);
 702         cl_page_state_set0(env, pg, CPS_FREEING);
 703
 704         list_for_each_entry_reverse(slice, &pg->cp_layers, cpl_linkage) {
 705                 if (slice->cpl_ops->cpo_delete != NULL)
 706                         (*slice->cpl_ops->cpo_delete)(env, slice);
 707         }
 708
 709         EXIT;
 710 }
 711
 712 /**
 713  * Called when a decision is made to throw page out of memory.
 714  *
 715  * Notifies all layers about page destruction by calling
 716  * cl_page_operations::cpo_delete() method top-to-bottom.
 717  *
 718  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
 719  * where transition to this state happens).
 720  *
 721  * Eliminates all venues through which new references to the page can be
 722  * obtained:
 723  *
 724  *     - removes page from the radix trees,
 725  *
 726  *     - breaks linkage from VM page to cl_page.
 727  *
 728  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
 729  * drain after some time, at which point page will be recycled.
 730  *
 731  * \pre  VM page is locked
 732  * \post pg->cp_state == CPS_FREEING
 733  *
 734  * \see cl_page_operations::cpo_delete()
 735  */
 736 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
 737 {
 738         PINVRNT(env, pg, cl_page_invariant(pg));
 739         ENTRY;
 740         cl_page_delete0(env, pg);
 741         EXIT;
 742 }
 743 EXPORT_SYMBOL(cl_page_delete);
 744
 745 /**
 746  * Marks page up-to-date.
 747  *
 748  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
 749  * layer responsible for VM interaction has to mark/clear page as up-to-date
 750  * by the \a uptodate argument.
 751  *
 752  * \see cl_page_operations::cpo_export()
 753  */
 754 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
 755 {
 756         const struct cl_page_slice *slice;
 757
 758         PINVRNT(env, pg, cl_page_invariant(pg));
 759
 760         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
 761                 if (slice->cpl_ops->cpo_export != NULL)
 762                         (*slice->cpl_ops->cpo_export)(env, slice, uptodate);
 763         }
 764 }
 765 EXPORT_SYMBOL(cl_page_export);
 766
 767 /**
 768  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
 769  * thread.
 770  */
 771 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
 772 {
 773         const struct cl_page_slice *slice;
 774         int result;
 775
 776         ENTRY;
 777         slice = container_of(pg->cp_layers.next,
 778                              const struct cl_page_slice, cpl_linkage);
 779         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
 780         /*
 781          * Call ->cpo_is_vmlocked() directly instead of going through
 782          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
 783          * cl_page_invariant().
 784          */
 785         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
 786         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
 787         RETURN(result == -EBUSY);
 788 }
 789 EXPORT_SYMBOL(cl_page_is_vmlocked);
 790
 791 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
 792 {
 793         ENTRY;
 794         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
 795 }
 796
 797 static void cl_page_io_start(const struct lu_env *env,
 798                              struct cl_page *pg, enum cl_req_type crt)
 799 {
 800         /*
 801          * Page is queued for IO, change its state.
 802          */
 803         ENTRY;
 804         cl_page_owner_clear(pg);
 805         cl_page_state_set(env, pg, cl_req_type_state(crt));
 806         EXIT;
 807 }
 808
 809 /**
 810  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
 811  * called top-to-bottom. Every layer either agrees to submit this page (by
 812  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
 813  * handling interactions with the VM also has to inform VM that page is under
 814  * transfer now.
 815  */
 816 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
 817                  struct cl_page *pg, enum cl_req_type crt)
 818 {
 819         const struct cl_page_slice *slice;
 820         int result = 0;
 821
 822         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 823         PINVRNT(env, pg, cl_page_invariant(pg));
 824         PINVRNT(env, pg, crt < CRT_NR);
 825
 826         /*
 827          * XXX this has to be called bottom-to-top, so that llite can set up
 828          * PG_writeback without risking other layers deciding to skip this
 829          * page.
 830          */
 831         if (crt >= CRT_NR)
 832                 return -EINVAL;
 833
 834         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
 835                 if (slice->cpl_ops->cpo_own)
 836                         result = (*slice->cpl_ops->io[crt].cpo_prep)(env,
 837                                                                      slice,
 838                                                                      io);
 839
 840                 if (result != 0)
 841                         break;
 842
 843         }
 844
 845         if (result >= 0) {
 846                 result = 0;
 847                 cl_page_io_start(env, pg, crt);
 848         }
 849
 850         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 851         return result;
 852 }
 853 EXPORT_SYMBOL(cl_page_prep);
 854
 855 /**
 856  * Notify layers about transfer completion.
 857  *
 858  * Invoked by transfer sub-system (which is a part of osc) to notify layers
 859  * that a transfer, of which this page is a part of has completed.
 860  *
 861  * Completion call-backs are executed in the bottom-up order, so that
 862  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
 863  * and can release locks safely.
 864  *
 865  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 866  * \post pg->cp_state == CPS_CACHED
 867  *
 868  * \see cl_page_operations::cpo_completion()
 869  */
 870 void cl_page_completion(const struct lu_env *env,
 871                         struct cl_page *pg, enum cl_req_type crt, int ioret)
 872 {
 873         const struct cl_page_slice *slice;
 874         struct cl_sync_io *anchor = pg->cp_sync_io;
 875
 876         PASSERT(env, pg, crt < CRT_NR);
 877         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
 878
 879         ENTRY;
 880         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
 881         cl_page_state_set(env, pg, CPS_CACHED);
 882         if (crt >= CRT_NR)
 883                 return;
 884
 885         list_for_each_entry_reverse(slice, &pg->cp_layers, cpl_linkage) {
 886                 if (slice->cpl_ops->io[crt].cpo_completion != NULL)
 887                         (*slice->cpl_ops->io[crt].cpo_completion)(env, slice,
 888                                                                   ioret);
 889         }
 890
 891         if (anchor != NULL) {
 892                 LASSERT(pg->cp_sync_io == anchor);
 893                 pg->cp_sync_io = NULL;
 894                 cl_sync_io_note(env, anchor, ioret);
 895         }
 896         EXIT;
 897 }
 898 EXPORT_SYMBOL(cl_page_completion);
 899
 900 /**
 901  * Notify layers that transfer formation engine decided to yank this page from
 902  * the cache and to make it a part of a transfer.
 903  *
 904  * \pre  pg->cp_state == CPS_CACHED
 905  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 906  *
 907  * \see cl_page_operations::cpo_make_ready()
 908  */
 909 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
 910                        enum cl_req_type crt)
 911 {
 912         const struct cl_page_slice *sli;
 913         int result = 0;
 914
 915         PINVRNT(env, pg, crt < CRT_NR);
 916
 917         ENTRY;
 918         if (crt >= CRT_NR)
 919                 RETURN(-EINVAL);
 920
 921         list_for_each_entry(sli, &pg->cp_layers, cpl_linkage) {
 922                 if (sli->cpl_ops->io[crt].cpo_make_ready != NULL)
 923                         result = (*sli->cpl_ops->io[crt].cpo_make_ready)(env,
 924                                                                          sli);
 925                 if (result != 0)
 926                         break;
 927         }
 928
 929         if (result >= 0) {
 930                 result = 0;
 931                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
 932                 cl_page_io_start(env, pg, crt);
 933         }
 934         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 935         RETURN(result);
 936 }
 937 EXPORT_SYMBOL(cl_page_make_ready);
 938
 939 /**
 940  * Called if a pge is being written back by kernel's intention.
 941  *
 942  * \pre  cl_page_is_owned(pg, io)
 943  * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT)
 944  *
 945  * \see cl_page_operations::cpo_flush()
 946  */
 947 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
 948                   struct cl_page *pg)
 949 {
 950         const struct cl_page_slice *slice;
 951         int result = 0;
 952
 953         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 954         PINVRNT(env, pg, cl_page_invariant(pg));
 955
 956         ENTRY;
 957
 958         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
 959                 if (slice->cpl_ops->cpo_flush != NULL)
 960                         result = (*slice->cpl_ops->cpo_flush)(env, slice, io);
 961                 if (result != 0)
 962                         break;
 963         }
 964         if (result > 0)
 965                 result = 0;
 966
 967         CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result);
 968         RETURN(result);
 969 }
 970 EXPORT_SYMBOL(cl_page_flush);
 971
 972 /**
 973  * Tells transfer engine that only part of a page is to be transmitted.
 974  *
 975  * \see cl_page_operations::cpo_clip()
 976  */
 977 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
 978                   int from, int to)
 979 {
 980         const struct cl_page_slice *slice;
 981
 982         PINVRNT(env, pg, cl_page_invariant(pg));
 983
 984         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
 985         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
 986                 if (slice->cpl_ops->cpo_clip != NULL)
 987                         (*slice->cpl_ops->cpo_clip)(env, slice, from, to);
 988         }
 989 }
 990 EXPORT_SYMBOL(cl_page_clip);
 991
 992 /**
 993  * Prints human readable representation of \a pg to the \a f.
 994  */
 995 void cl_page_header_print(const struct lu_env *env, void *cookie,
 996                           lu_printer_t printer, const struct cl_page *pg)
 997 {
 998         (*printer)(env, cookie,
 999                    "page@%p[%d %p %d %d %p]\n",
1000                    pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1001                    pg->cp_state, pg->cp_type,
1002                    pg->cp_owner);
1003 }
1004 EXPORT_SYMBOL(cl_page_header_print);
1005
1006 /**
1007  * Prints human readable representation of \a pg to the \a f.
1008  */
1009 void cl_page_print(const struct lu_env *env, void *cookie,
1010                    lu_printer_t printer, const struct cl_page *pg)
1011 {
1012         const struct cl_page_slice *slice;
1013         int result = 0;
1014
1015         cl_page_header_print(env, cookie, printer, pg);
1016         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
1017                 if (slice->cpl_ops->cpo_print != NULL)
1018                         result = (*slice->cpl_ops->cpo_print)(env, slice,
1019                                                              cookie, printer);
1020                 if (result != 0)
1021                         break;
1022         }
1023         (*printer)(env, cookie, "end page@%p\n", pg);
1024 }
1025 EXPORT_SYMBOL(cl_page_print);
1026
1027 /**
1028  * Cancel a page which is still in a transfer.
1029  */
1030 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1031 {
1032         const struct cl_page_slice *slice;
1033         int                         result = 0;
1034
1035         list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
1036                 if (slice->cpl_ops->cpo_cancel != NULL)
1037                         result = (*slice->cpl_ops->cpo_cancel)(env, slice);
1038                 if (result != 0)
1039                         break;
1040         }
1041         if (result > 0)
1042                 result = 0;
1043
1044         return result;
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);