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, 2017, 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 static void cs_page_inc(const struct cl_object *obj,
  78                         enum cache_stats_item item)
  79 {
  80 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
  81         atomic_inc(&cl_object_site(obj)->cs_pages.cs_stats[item]);
  82 #endif
  83 }
  84
  85 static void cs_page_dec(const struct cl_object *obj,
  86                         enum cache_stats_item item)
  87 {
  88 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
  89         atomic_dec(&cl_object_site(obj)->cs_pages.cs_stats[item]);
  90 #endif
  91 }
  92
  93 static void cs_pagestate_inc(const struct cl_object *obj,
  94                              enum cl_page_state state)
  95 {
  96 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
  97         atomic_inc(&cl_object_site(obj)->cs_pages_state[state]);
  98 #endif
  99 }
 100
 101 static void cs_pagestate_dec(const struct cl_object *obj,
 102                               enum cl_page_state state)
 103 {
 104 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
 105         atomic_dec(&cl_object_site(obj)->cs_pages_state[state]);
 106 #endif
 107 }
 108
 109 /**
 110  * Internal version of cl_page_get().
 111  *
 112  * This function can be used to obtain initial reference to previously
 113  * unreferenced cached object. It can be called only if concurrent page
 114  * reclamation is somehow prevented, e.g., by keeping a lock on a VM page,
 115  * associated with \a page.
 116  *
 117  * Use with care! Not exported.
 118  */
 119 static void cl_page_get_trust(struct cl_page *page)
 120 {
 121         LASSERT(atomic_read(&page->cp_ref) > 0);
 122         atomic_inc(&page->cp_ref);
 123 }
 124
 125 /**
 126  * Returns a slice within a page, corresponding to the given layer in the
 127  * device stack.
 128  *
 129  * \see cl_lock_at()
 130  */
 131 static const struct cl_page_slice *
 132 cl_page_at_trusted(const struct cl_page *page,
 133                    const struct lu_device_type *dtype)
 134 {
 135         const struct cl_page_slice *slice;
 136         ENTRY;
 137
 138         list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
 139                 if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
 140                         RETURN(slice);
 141         }
 142         RETURN(NULL);
 143 }
 144
 145 static void cl_page_free(const struct lu_env *env, struct cl_page *page,
 146                          struct pagevec *pvec)
 147 {
 148         struct cl_object *obj  = page->cp_obj;
 149         int pagesize = cl_object_header(obj)->coh_page_bufsize;
 150
 151         PASSERT(env, page, list_empty(&page->cp_batch));
 152         PASSERT(env, page, page->cp_owner == NULL);
 153         PASSERT(env, page, page->cp_state == CPS_FREEING);
 154
 155         ENTRY;
 156         while (!list_empty(&page->cp_layers)) {
 157                 struct cl_page_slice *slice;
 158
 159                 slice = list_entry(page->cp_layers.next,
 160                                    struct cl_page_slice, cpl_linkage);
 161                 list_del_init(page->cp_layers.next);
 162                 if (unlikely(slice->cpl_ops->cpo_fini != NULL))
 163                         slice->cpl_ops->cpo_fini(env, slice, pvec);
 164         }
 165         cs_page_dec(obj, CS_total);
 166         cs_pagestate_dec(obj, page->cp_state);
 167         lu_object_ref_del_at(&obj->co_lu, &page->cp_obj_ref, "cl_page", page);
 168         cl_object_put(env, obj);
 169         lu_ref_fini(&page->cp_reference);
 170         OBD_FREE(page, pagesize);
 171         EXIT;
 172 }
 173
 174 /**
 175  * Helper function updating page state. This is the only place in the code
 176  * where cl_page::cp_state field is mutated.
 177  */
 178 static inline void cl_page_state_set_trust(struct cl_page *page,
 179                                            enum cl_page_state state)
 180 {
 181         /* bypass const. */
 182         *(enum cl_page_state *)&page->cp_state = state;
 183 }
 184
 185 struct cl_page *cl_page_alloc(const struct lu_env *env,
 186                 struct cl_object *o, pgoff_t ind, struct page *vmpage,
 187                 enum cl_page_type type)
 188 {
 189         struct cl_page          *page;
 190         struct lu_object_header *head;
 191
 192         ENTRY;
 193         OBD_ALLOC_GFP(page, cl_object_header(o)->coh_page_bufsize,
 194                         GFP_NOFS);
 195         if (page != NULL) {
 196                 int result = 0;
 197                 atomic_set(&page->cp_ref, 1);
 198                 page->cp_obj = o;
 199                 cl_object_get(o);
 200                 lu_object_ref_add_at(&o->co_lu, &page->cp_obj_ref, "cl_page",
 201                                      page);
 202                 page->cp_vmpage = vmpage;
 203                 cl_page_state_set_trust(page, CPS_CACHED);
 204                 page->cp_type = type;
 205                 INIT_LIST_HEAD(&page->cp_layers);
 206                 INIT_LIST_HEAD(&page->cp_batch);
 207                 lu_ref_init(&page->cp_reference);
 208                 head = o->co_lu.lo_header;
 209                 list_for_each_entry(o, &head->loh_layers,
 210                                     co_lu.lo_linkage) {
 211                         if (o->co_ops->coo_page_init != NULL) {
 212                                 result = o->co_ops->coo_page_init(env, o, page,
 213                                                                   ind);
 214                                 if (result != 0) {
 215                                         cl_page_delete0(env, page);
 216                                         cl_page_free(env, page, NULL);
 217                                         page = ERR_PTR(result);
 218                                         break;
 219                                 }
 220                         }
 221                 }
 222                 if (result == 0) {
 223                         cs_page_inc(o, CS_total);
 224                         cs_page_inc(o, CS_create);
 225                         cs_pagestate_dec(o, CPS_CACHED);
 226                 }
 227         } else {
 228                 page = ERR_PTR(-ENOMEM);
 229         }
 230         RETURN(page);
 231 }
 232
 233 /**
 234  * Returns a cl_page with index \a idx at the object \a o, and associated with
 235  * the VM page \a vmpage.
 236  *
 237  * This is the main entry point into the cl_page caching interface. First, a
 238  * cache (implemented as a per-object radix tree) is consulted. If page is
 239  * found there, it is returned immediately. Otherwise new page is allocated
 240  * and returned. In any case, additional reference to page is acquired.
 241  *
 242  * \see cl_object_find(), cl_lock_find()
 243  */
 244 struct cl_page *cl_page_find(const struct lu_env *env,
 245                              struct cl_object *o,
 246                              pgoff_t idx, struct page *vmpage,
 247                              enum cl_page_type type)
 248 {
 249         struct cl_page          *page = NULL;
 250         struct cl_object_header *hdr;
 251
 252         LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 253         might_sleep();
 254
 255         ENTRY;
 256
 257         hdr = cl_object_header(o);
 258         cs_page_inc(o, CS_lookup);
 259
 260         CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
 261                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 262         /* fast path. */
 263         if (type == CPT_CACHEABLE) {
 264                 /* vmpage lock is used to protect the child/parent
 265                  * relationship */
 266                 KLASSERT(PageLocked(vmpage));
 267                 /*
 268                  * cl_vmpage_page() can be called here without any locks as
 269                  *
 270                  *     - "vmpage" is locked (which prevents ->private from
 271                  *       concurrent updates), and
 272                  *
 273                  *     - "o" cannot be destroyed while current thread holds a
 274                  *       reference on it.
 275                  */
 276                 page = cl_vmpage_page(vmpage, o);
 277                 if (page != NULL) {
 278                         cs_page_inc(o, CS_hit);
 279                         RETURN(page);
 280                 }
 281         }
 282
 283         /* allocate and initialize cl_page */
 284         page = cl_page_alloc(env, o, idx, vmpage, type);
 285         RETURN(page);
 286 }
 287 EXPORT_SYMBOL(cl_page_find);
 288
 289 static inline int cl_page_invariant(const struct cl_page *pg)
 290 {
 291         return cl_page_in_use_noref(pg);
 292 }
 293
 294 static void cl_page_state_set0(const struct lu_env *env,
 295                                struct cl_page *page, enum cl_page_state state)
 296 {
 297         enum cl_page_state old;
 298
 299         /*
 300          * Matrix of allowed state transitions [old][new], for sanity
 301          * checking.
 302          */
 303         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 304                 [CPS_CACHED] = {
 305                         [CPS_CACHED]  = 0,
 306                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 307                         [CPS_PAGEIN]  = 0,
 308                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 309                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 310                 },
 311                 [CPS_OWNED] = {
 312                         [CPS_CACHED]  = 1, /* release to the cache */
 313                         [CPS_OWNED]   = 0,
 314                         [CPS_PAGEIN]  = 1, /* start read immediately */
 315                         [CPS_PAGEOUT] = 1, /* start write immediately */
 316                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 317                 },
 318                 [CPS_PAGEIN] = {
 319                         [CPS_CACHED]  = 1, /* io completion */
 320                         [CPS_OWNED]   = 0,
 321                         [CPS_PAGEIN]  = 0,
 322                         [CPS_PAGEOUT] = 0,
 323                         [CPS_FREEING] = 0,
 324                 },
 325                 [CPS_PAGEOUT] = {
 326                         [CPS_CACHED]  = 1, /* io completion */
 327                         [CPS_OWNED]   = 0,
 328                         [CPS_PAGEIN]  = 0,
 329                         [CPS_PAGEOUT] = 0,
 330                         [CPS_FREEING] = 0,
 331                 },
 332                 [CPS_FREEING] = {
 333                         [CPS_CACHED]  = 0,
 334                         [CPS_OWNED]   = 0,
 335                         [CPS_PAGEIN]  = 0,
 336                         [CPS_PAGEOUT] = 0,
 337                         [CPS_FREEING] = 0,
 338                 }
 339         };
 340
 341         ENTRY;
 342         old = page->cp_state;
 343         PASSERT(env, page, allowed_transitions[old][state]);
 344         CL_PAGE_HEADER(D_TRACE, env, page, "%d -> %d\n", old, state);
 345         PASSERT(env, page, page->cp_state == old);
 346         PASSERT(env, page, equi(state == CPS_OWNED, page->cp_owner != NULL));
 347
 348         cs_pagestate_dec(page->cp_obj, page->cp_state);
 349         cs_pagestate_inc(page->cp_obj, state);
 350         cl_page_state_set_trust(page, state);
 351         EXIT;
 352 }
 353
 354 static void cl_page_state_set(const struct lu_env *env,
 355                               struct cl_page *page, enum cl_page_state state)
 356 {
 357         cl_page_state_set0(env, page, state);
 358 }
 359
 360 /**
 361  * Acquires an additional reference to a page.
 362  *
 363  * This can be called only by caller already possessing a reference to \a
 364  * page.
 365  *
 366  * \see cl_object_get(), cl_lock_get().
 367  */
 368 void cl_page_get(struct cl_page *page)
 369 {
 370         ENTRY;
 371         cl_page_get_trust(page);
 372         EXIT;
 373 }
 374 EXPORT_SYMBOL(cl_page_get);
 375
 376 /**
 377  * Releases a reference to a page, use the pagevec to release the pages
 378  * in batch if provided.
 379  *
 380  * Users need to do a final pagevec_release() to release any trailing pages.
 381  */
 382 void cl_pagevec_put(const struct lu_env *env, struct cl_page *page,
 383                   struct pagevec *pvec)
 384 {
 385         ENTRY;
 386         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 387                        atomic_read(&page->cp_ref));
 388
 389         if (atomic_dec_and_test(&page->cp_ref)) {
 390                 LASSERT(page->cp_state == CPS_FREEING);
 391
 392                 LASSERT(atomic_read(&page->cp_ref) == 0);
 393                 PASSERT(env, page, page->cp_owner == NULL);
 394                 PASSERT(env, page, list_empty(&page->cp_batch));
 395                 /*
 396                  * Page is no longer reachable by other threads. Tear
 397                  * it down.
 398                  */
 399                 cl_page_free(env, page, pvec);
 400         }
 401
 402         EXIT;
 403 }
 404 EXPORT_SYMBOL(cl_pagevec_put);
 405
 406 /**
 407  * Releases a reference to a page, wrapper to cl_pagevec_put
 408  *
 409  * When last reference is released, page is returned to the cache, unless it
 410  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 411  * destroyed.
 412  *
 413  * \see cl_object_put(), cl_lock_put().
 414  */
 415 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 416 {
 417         cl_pagevec_put(env, page, NULL);
 418 }
 419 EXPORT_SYMBOL(cl_page_put);
 420
 421 /**
 422  * Returns a cl_page associated with a VM page, and given cl_object.
 423  */
 424 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj)
 425 {
 426         struct cl_page *page;
 427
 428         ENTRY;
 429         KLASSERT(PageLocked(vmpage));
 430
 431         /*
 432          * NOTE: absence of races and liveness of data are guaranteed by page
 433          *       lock on a "vmpage". That works because object destruction has
 434          *       bottom-to-top pass.
 435          */
 436
 437         page = (struct cl_page *)vmpage->private;
 438         if (page != NULL) {
 439                 cl_page_get_trust(page);
 440                 LASSERT(page->cp_type == CPT_CACHEABLE);
 441         }
 442         RETURN(page);
 443 }
 444 EXPORT_SYMBOL(cl_vmpage_page);
 445
 446 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 447                                        const struct lu_device_type *dtype)
 448 {
 449         return cl_page_at_trusted(page, dtype);
 450 }
 451 EXPORT_SYMBOL(cl_page_at);
 452
 453 static void cl_page_owner_clear(struct cl_page *page)
 454 {
 455         ENTRY;
 456         if (page->cp_owner != NULL) {
 457                 LASSERT(page->cp_owner->ci_owned_nr > 0);
 458                 page->cp_owner->ci_owned_nr--;
 459                 page->cp_owner = NULL;
 460         }
 461         EXIT;
 462 }
 463
 464 static void cl_page_owner_set(struct cl_page *page)
 465 {
 466         ENTRY;
 467         LASSERT(page->cp_owner != NULL);
 468         page->cp_owner->ci_owned_nr++;
 469         EXIT;
 470 }
 471
 472 void cl_page_disown0(const struct lu_env *env,
 473                      struct cl_io *io, struct cl_page *pg)
 474 {
 475         const struct cl_page_slice *slice;
 476         enum cl_page_state state;
 477
 478         ENTRY;
 479         state = pg->cp_state;
 480         PINVRNT(env, pg, state == CPS_OWNED || state == CPS_FREEING);
 481         PINVRNT(env, pg, cl_page_invariant(pg) || state == CPS_FREEING);
 482         cl_page_owner_clear(pg);
 483
 484         if (state == CPS_OWNED)
 485                 cl_page_state_set(env, pg, CPS_CACHED);
 486         /*
 487          * Completion call-backs are executed in the bottom-up order, so that
 488          * uppermost layer (llite), responsible for VFS/VM interaction runs
 489          * last and can release locks safely.
 490          */
 491         list_for_each_entry_reverse(slice, &pg->cp_layers, cpl_linkage) {
 492                 if (slice->cpl_ops->cpo_disown != NULL)
 493                         (*slice->cpl_ops->cpo_disown)(env, slice, io);
 494         }
 495
 496         EXIT;
 497 }
 498
 499 /**
 500  * returns true, iff page is owned by the given io.
 501  */
 502 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 503 {
 504         struct cl_io *top = cl_io_top((struct cl_io *)io);
 505         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 506         ENTRY;
 507         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top);
 508 }
 509 EXPORT_SYMBOL(cl_page_is_owned);
 510
 511 /**
 512  * Try to own a page by IO.
 513  *
 514  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 515  * into cl_page_state::CPS_OWNED state.
 516  *
 517  * \pre  !cl_page_is_owned(pg, io)
 518  * \post result == 0 iff cl_page_is_owned(pg, io)
 519  *
 520  * \retval 0   success
 521  *
 522  * \retval -ve failure, e.g., page was destroyed (and landed in
 523  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 524  *             or, page was owned by another thread, or in IO.
 525  *
 526  * \see cl_page_disown()
 527  * \see cl_page_operations::cpo_own()
 528  * \see cl_page_own_try()
 529  * \see cl_page_own
 530  */
 531 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 532                         struct cl_page *pg, int nonblock)
 533 {
 534         int result = 0;
 535         const struct cl_page_slice *slice;
 536
 537         PINVRNT(env, pg, !cl_page_is_owned(pg, io));
 538
 539         ENTRY;
 540         io = cl_io_top(io);
 541
 542         if (pg->cp_state == CPS_FREEING) {
 543                 result = -ENOENT;
 544                 goto out;
 545         }
 546
 547         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
 548                 if (slice->cpl_ops->cpo_own)
 549                         result = (*slice->cpl_ops->cpo_own)(env, slice,
 550                                                             io, nonblock);
 551
 552                 if (result != 0)
 553                         break;
 554
 555         }
 556         if (result > 0)
 557                 result = 0;
 558
 559         if (result == 0) {
 560                 PASSERT(env, pg, pg->cp_owner == NULL);
 561                 pg->cp_owner = cl_io_top(io);
 562                 cl_page_owner_set(pg);
 563                 if (pg->cp_state != CPS_FREEING) {
 564                         cl_page_state_set(env, pg, CPS_OWNED);
 565                 } else {
 566                         cl_page_disown0(env, io, pg);
 567                         result = -ENOENT;
 568                 }
 569         }
 570
 571 out:
 572         PINVRNT(env, pg, ergo(result == 0, cl_page_invariant(pg)));
 573         RETURN(result);
 574 }
 575
 576 /**
 577  * Own a page, might be blocked.
 578  *
 579  * \see cl_page_own0()
 580  */
 581 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 582 {
 583         return cl_page_own0(env, io, pg, 0);
 584 }
 585 EXPORT_SYMBOL(cl_page_own);
 586
 587 /**
 588  * Nonblock version of cl_page_own().
 589  *
 590  * \see cl_page_own0()
 591  */
 592 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
 593                     struct cl_page *pg)
 594 {
 595         return cl_page_own0(env, io, pg, 1);
 596 }
 597 EXPORT_SYMBOL(cl_page_own_try);
 598
 599
 600 /**
 601  * Assume page ownership.
 602  *
 603  * Called when page is already locked by the hosting VM.
 604  *
 605  * \pre !cl_page_is_owned(pg, io)
 606  * \post cl_page_is_owned(pg, io)
 607  *
 608  * \see cl_page_operations::cpo_assume()
 609  */
 610 void cl_page_assume(const struct lu_env *env,
 611                     struct cl_io *io, struct cl_page *pg)
 612 {
 613         const struct cl_page_slice *slice;
 614
 615         PINVRNT(env, pg, cl_object_same(pg->cp_obj, io->ci_obj));
 616
 617         ENTRY;
 618         io = cl_io_top(io);
 619
 620         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
 621                 if (slice->cpl_ops->cpo_assume != NULL)
 622                         (*slice->cpl_ops->cpo_assume)(env, slice, io);
 623         }
 624
 625         PASSERT(env, pg, pg->cp_owner == NULL);
 626         pg->cp_owner = cl_io_top(io);
 627         cl_page_owner_set(pg);
 628         cl_page_state_set(env, pg, CPS_OWNED);
 629         EXIT;
 630 }
 631 EXPORT_SYMBOL(cl_page_assume);
 632
 633 /**
 634  * Releases page ownership without unlocking the page.
 635  *
 636  * Moves page into cl_page_state::CPS_CACHED without releasing a lock on the
 637  * underlying VM page (as VM is supposed to do this itself).
 638  *
 639  * \pre   cl_page_is_owned(pg, io)
 640  * \post !cl_page_is_owned(pg, io)
 641  *
 642  * \see cl_page_assume()
 643  */
 644 void cl_page_unassume(const struct lu_env *env,
 645                       struct cl_io *io, struct cl_page *pg)
 646 {
 647         const struct cl_page_slice *slice;
 648
 649         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 650         PINVRNT(env, pg, cl_page_invariant(pg));
 651
 652         ENTRY;
 653         io = cl_io_top(io);
 654         cl_page_owner_clear(pg);
 655         cl_page_state_set(env, pg, CPS_CACHED);
 656
 657         list_for_each_entry_reverse(slice, &pg->cp_layers, cpl_linkage) {
 658                 if (slice->cpl_ops->cpo_unassume != NULL)
 659                         (*slice->cpl_ops->cpo_unassume)(env, slice, io);
 660         }
 661
 662         EXIT;
 663 }
 664 EXPORT_SYMBOL(cl_page_unassume);
 665
 666 /**
 667  * Releases page ownership.
 668  *
 669  * Moves page into cl_page_state::CPS_CACHED.
 670  *
 671  * \pre   cl_page_is_owned(pg, io)
 672  * \post !cl_page_is_owned(pg, io)
 673  *
 674  * \see cl_page_own()
 675  * \see cl_page_operations::cpo_disown()
 676  */
 677 void cl_page_disown(const struct lu_env *env,
 678                     struct cl_io *io, struct cl_page *pg)
 679 {
 680         PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
 681                 pg->cp_state == CPS_FREEING);
 682
 683         ENTRY;
 684         io = cl_io_top(io);
 685         cl_page_disown0(env, io, pg);
 686         EXIT;
 687 }
 688 EXPORT_SYMBOL(cl_page_disown);
 689
 690 /**
 691  * Called when page is to be removed from the object, e.g., as a result of
 692  * truncate.
 693  *
 694  * Calls cl_page_operations::cpo_discard() top-to-bottom.
 695  *
 696  * \pre cl_page_is_owned(pg, io)
 697  *
 698  * \see cl_page_operations::cpo_discard()
 699  */
 700 void cl_page_discard(const struct lu_env *env,
 701                      struct cl_io *io, struct cl_page *pg)
 702 {
 703         const struct cl_page_slice *slice;
 704
 705         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 706         PINVRNT(env, pg, cl_page_invariant(pg));
 707
 708         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
 709                 if (slice->cpl_ops->cpo_discard != NULL)
 710                         (*slice->cpl_ops->cpo_discard)(env, slice, io);
 711         }
 712 }
 713 EXPORT_SYMBOL(cl_page_discard);
 714
 715 /**
 716  * Version of cl_page_delete() that can be called for not fully constructed
 717  * pages, e.g. in an error handling cl_page_find()->cl_page_delete0()
 718  * path. Doesn't check page invariant.
 719  */
 720 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg)
 721 {
 722         const struct cl_page_slice *slice;
 723
 724         ENTRY;
 725
 726         PASSERT(env, pg, pg->cp_state != CPS_FREEING);
 727
 728         /*
 729          * Severe all ways to obtain new pointers to @pg.
 730          */
 731         cl_page_owner_clear(pg);
 732         cl_page_state_set0(env, pg, CPS_FREEING);
 733
 734         list_for_each_entry_reverse(slice, &pg->cp_layers, cpl_linkage) {
 735                 if (slice->cpl_ops->cpo_delete != NULL)
 736                         (*slice->cpl_ops->cpo_delete)(env, slice);
 737         }
 738
 739         EXIT;
 740 }
 741
 742 /**
 743  * Called when a decision is made to throw page out of memory.
 744  *
 745  * Notifies all layers about page destruction by calling
 746  * cl_page_operations::cpo_delete() method top-to-bottom.
 747  *
 748  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
 749  * where transition to this state happens).
 750  *
 751  * Eliminates all venues through which new references to the page can be
 752  * obtained:
 753  *
 754  *     - removes page from the radix trees,
 755  *
 756  *     - breaks linkage from VM page to cl_page.
 757  *
 758  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
 759  * drain after some time, at which point page will be recycled.
 760  *
 761  * \pre  VM page is locked
 762  * \post pg->cp_state == CPS_FREEING
 763  *
 764  * \see cl_page_operations::cpo_delete()
 765  */
 766 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
 767 {
 768         PINVRNT(env, pg, cl_page_invariant(pg));
 769         ENTRY;
 770         cl_page_delete0(env, pg);
 771         EXIT;
 772 }
 773 EXPORT_SYMBOL(cl_page_delete);
 774
 775 /**
 776  * Marks page up-to-date.
 777  *
 778  * Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
 779  * layer responsible for VM interaction has to mark/clear page as up-to-date
 780  * by the \a uptodate argument.
 781  *
 782  * \see cl_page_operations::cpo_export()
 783  */
 784 void cl_page_export(const struct lu_env *env, struct cl_page *pg, int uptodate)
 785 {
 786         const struct cl_page_slice *slice;
 787
 788         PINVRNT(env, pg, cl_page_invariant(pg));
 789
 790         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
 791                 if (slice->cpl_ops->cpo_export != NULL)
 792                         (*slice->cpl_ops->cpo_export)(env, slice, uptodate);
 793         }
 794 }
 795 EXPORT_SYMBOL(cl_page_export);
 796
 797 /**
 798  * Returns true, iff \a pg is VM locked in a suitable sense by the calling
 799  * thread.
 800  */
 801 int cl_page_is_vmlocked(const struct lu_env *env, const struct cl_page *pg)
 802 {
 803         const struct cl_page_slice *slice;
 804         int result;
 805
 806         ENTRY;
 807         slice = container_of(pg->cp_layers.next,
 808                              const struct cl_page_slice, cpl_linkage);
 809         PASSERT(env, pg, slice->cpl_ops->cpo_is_vmlocked != NULL);
 810         /*
 811          * Call ->cpo_is_vmlocked() directly instead of going through
 812          * CL_PAGE_INVOKE(), because cl_page_is_vmlocked() is used by
 813          * cl_page_invariant().
 814          */
 815         result = slice->cpl_ops->cpo_is_vmlocked(env, slice);
 816         PASSERT(env, pg, result == -EBUSY || result == -ENODATA);
 817         RETURN(result == -EBUSY);
 818 }
 819 EXPORT_SYMBOL(cl_page_is_vmlocked);
 820
 821 void cl_page_touch(const struct lu_env *env, const struct cl_page *pg,
 822                   size_t to)
 823 {
 824         const struct cl_page_slice *slice;
 825
 826         ENTRY;
 827
 828         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
 829                 if (slice->cpl_ops->cpo_page_touch != NULL)
 830                         (*slice->cpl_ops->cpo_page_touch)(env, slice, to);
 831         }
 832
 833         EXIT;
 834 }
 835 EXPORT_SYMBOL(cl_page_touch);
 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         const struct cl_page_slice *slice;
 866         int result = 0;
 867
 868         PINVRNT(env, pg, cl_page_is_owned(pg, io));
 869         PINVRNT(env, pg, cl_page_invariant(pg));
 870         PINVRNT(env, pg, crt < CRT_NR);
 871
 872         /*
 873          * XXX this has to be called bottom-to-top, so that llite can set up
 874          * PG_writeback without risking other layers deciding to skip this
 875          * page.
 876          */
 877         if (crt >= CRT_NR)
 878                 return -EINVAL;
 879
 880         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
 881                 if (slice->cpl_ops->cpo_own)
 882                         result = (*slice->cpl_ops->io[crt].cpo_prep)(env,
 883                                                                      slice,
 884                                                                      io);
 885
 886                 if (result != 0)
 887                         break;
 888
 889         }
 890
 891         if (result >= 0) {
 892                 result = 0;
 893                 cl_page_io_start(env, pg, crt);
 894         }
 895
 896         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 897         return result;
 898 }
 899 EXPORT_SYMBOL(cl_page_prep);
 900
 901 /**
 902  * Notify layers about transfer completion.
 903  *
 904  * Invoked by transfer sub-system (which is a part of osc) to notify layers
 905  * that a transfer, of which this page is a part of has completed.
 906  *
 907  * Completion call-backs are executed in the bottom-up order, so that
 908  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
 909  * and can release locks safely.
 910  *
 911  * \pre  pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 912  * \post pg->cp_state == CPS_CACHED
 913  *
 914  * \see cl_page_operations::cpo_completion()
 915  */
 916 void cl_page_completion(const struct lu_env *env,
 917                         struct cl_page *pg, enum cl_req_type crt, int ioret)
 918 {
 919         const struct cl_page_slice *slice;
 920         struct cl_sync_io *anchor = pg->cp_sync_io;
 921
 922         PASSERT(env, pg, crt < CRT_NR);
 923         PASSERT(env, pg, pg->cp_state == cl_req_type_state(crt));
 924
 925         ENTRY;
 926         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, ioret);
 927         cl_page_state_set(env, pg, CPS_CACHED);
 928         if (crt >= CRT_NR)
 929                 return;
 930
 931         list_for_each_entry_reverse(slice, &pg->cp_layers, cpl_linkage) {
 932                 if (slice->cpl_ops->io[crt].cpo_completion != NULL)
 933                         (*slice->cpl_ops->io[crt].cpo_completion)(env, slice,
 934                                                                   ioret);
 935         }
 936
 937         if (anchor != NULL) {
 938                 LASSERT(pg->cp_sync_io == anchor);
 939                 pg->cp_sync_io = NULL;
 940                 cl_sync_io_note(env, anchor, ioret);
 941         }
 942         EXIT;
 943 }
 944 EXPORT_SYMBOL(cl_page_completion);
 945
 946 /**
 947  * Notify layers that transfer formation engine decided to yank this page from
 948  * the cache and to make it a part of a transfer.
 949  *
 950  * \pre  pg->cp_state == CPS_CACHED
 951  * \post pg->cp_state == CPS_PAGEIN || pg->cp_state == CPS_PAGEOUT
 952  *
 953  * \see cl_page_operations::cpo_make_ready()
 954  */
 955 int cl_page_make_ready(const struct lu_env *env, struct cl_page *pg,
 956                        enum cl_req_type crt)
 957 {
 958         const struct cl_page_slice *sli;
 959         int result = 0;
 960
 961         PINVRNT(env, pg, crt < CRT_NR);
 962
 963         ENTRY;
 964         if (crt >= CRT_NR)
 965                 RETURN(-EINVAL);
 966
 967         list_for_each_entry(sli, &pg->cp_layers, cpl_linkage) {
 968                 if (sli->cpl_ops->io[crt].cpo_make_ready != NULL)
 969                         result = (*sli->cpl_ops->io[crt].cpo_make_ready)(env,
 970                                                                          sli);
 971                 if (result != 0)
 972                         break;
 973         }
 974
 975         if (result >= 0) {
 976                 result = 0;
 977                 PASSERT(env, pg, pg->cp_state == CPS_CACHED);
 978                 cl_page_io_start(env, pg, crt);
 979         }
 980         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", crt, result);
 981         RETURN(result);
 982 }
 983 EXPORT_SYMBOL(cl_page_make_ready);
 984
 985 /**
 986  * Called if a pge is being written back by kernel's intention.
 987  *
 988  * \pre  cl_page_is_owned(pg, io)
 989  * \post ergo(result == 0, pg->cp_state == CPS_PAGEOUT)
 990  *
 991  * \see cl_page_operations::cpo_flush()
 992  */
 993 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
 994                   struct cl_page *pg)
 995 {
 996         const struct cl_page_slice *slice;
 997         int result = 0;
 998
 999         PINVRNT(env, pg, cl_page_is_owned(pg, io));
1000         PINVRNT(env, pg, cl_page_invariant(pg));
1001
1002         ENTRY;
1003
1004         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
1005                 if (slice->cpl_ops->cpo_flush != NULL)
1006                         result = (*slice->cpl_ops->cpo_flush)(env, slice, io);
1007                 if (result != 0)
1008                         break;
1009         }
1010         if (result > 0)
1011                 result = 0;
1012
1013         CL_PAGE_HEADER(D_TRACE, env, pg, "%d\n", result);
1014         RETURN(result);
1015 }
1016 EXPORT_SYMBOL(cl_page_flush);
1017
1018 /**
1019  * Tells transfer engine that only part of a page is to be transmitted.
1020  *
1021  * \see cl_page_operations::cpo_clip()
1022  */
1023 void cl_page_clip(const struct lu_env *env, struct cl_page *pg,
1024                   int from, int to)
1025 {
1026         const struct cl_page_slice *slice;
1027
1028         PINVRNT(env, pg, cl_page_invariant(pg));
1029
1030         CL_PAGE_HEADER(D_TRACE, env, pg, "%d %d\n", from, to);
1031         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
1032                 if (slice->cpl_ops->cpo_clip != NULL)
1033                         (*slice->cpl_ops->cpo_clip)(env, slice, from, to);
1034         }
1035 }
1036 EXPORT_SYMBOL(cl_page_clip);
1037
1038 /**
1039  * Prints human readable representation of \a pg to the \a f.
1040  */
1041 void cl_page_header_print(const struct lu_env *env, void *cookie,
1042                           lu_printer_t printer, const struct cl_page *pg)
1043 {
1044         (*printer)(env, cookie,
1045                    "page@%p[%d %p %d %d %p]\n",
1046                    pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1047                    pg->cp_state, pg->cp_type,
1048                    pg->cp_owner);
1049 }
1050 EXPORT_SYMBOL(cl_page_header_print);
1051
1052 /**
1053  * Prints human readable representation of \a pg to the \a f.
1054  */
1055 void cl_page_print(const struct lu_env *env, void *cookie,
1056                    lu_printer_t printer, const struct cl_page *pg)
1057 {
1058         const struct cl_page_slice *slice;
1059         int result = 0;
1060
1061         cl_page_header_print(env, cookie, printer, pg);
1062         list_for_each_entry(slice, &pg->cp_layers, cpl_linkage) {
1063                 if (slice->cpl_ops->cpo_print != NULL)
1064                         result = (*slice->cpl_ops->cpo_print)(env, slice,
1065                                                              cookie, printer);
1066                 if (result != 0)
1067                         break;
1068         }
1069         (*printer)(env, cookie, "end page@%p\n", pg);
1070 }
1071 EXPORT_SYMBOL(cl_page_print);
1072
1073 /**
1074  * Cancel a page which is still in a transfer.
1075  */
1076 int cl_page_cancel(const struct lu_env *env, struct cl_page *page)
1077 {
1078         const struct cl_page_slice *slice;
1079         int                         result = 0;
1080
1081         list_for_each_entry(slice, &page->cp_layers, cpl_linkage) {
1082                 if (slice->cpl_ops->cpo_cancel != NULL)
1083                         result = (*slice->cpl_ops->cpo_cancel)(env, slice);
1084                 if (result != 0)
1085                         break;
1086         }
1087         if (result > 0)
1088                 result = 0;
1089
1090         return result;
1091 }
1092
1093 /**
1094  * Converts a byte offset within object \a obj into a page index.
1095  */
1096 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1097 {
1098         return (loff_t)idx << PAGE_SHIFT;
1099 }
1100 EXPORT_SYMBOL(cl_offset);
1101
1102 /**
1103  * Converts a page index into a byte offset within object \a obj.
1104  */
1105 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1106 {
1107         return offset >> PAGE_SHIFT;
1108 }
1109 EXPORT_SYMBOL(cl_index);
1110
1111 size_t cl_page_size(const struct cl_object *obj)
1112 {
1113         return 1UL << PAGE_SHIFT;
1114 }
1115 EXPORT_SYMBOL(cl_page_size);
1116
1117 /**
1118  * Adds page slice to the compound page.
1119  *
1120  * This is called by cl_object_operations::coo_page_init() methods to add a
1121  * per-layer state to the page. New state is added at the end of
1122  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1123  *
1124  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1125  */
1126 void cl_page_slice_add(struct cl_page *page, struct cl_page_slice *slice,
1127                        struct cl_object *obj, pgoff_t index,
1128                        const struct cl_page_operations *ops)
1129 {
1130         ENTRY;
1131         list_add_tail(&slice->cpl_linkage, &page->cp_layers);
1132         slice->cpl_obj  = obj;
1133         slice->cpl_index = index;
1134         slice->cpl_ops  = ops;
1135         slice->cpl_page = page;
1136         EXIT;
1137 }
1138 EXPORT_SYMBOL(cl_page_slice_add);
1139
1140 /**
1141  * Allocate and initialize cl_cache, called by ll_init_sbi().
1142  */
1143 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max)
1144 {
1145         struct cl_client_cache  *cache = NULL;
1146
1147         ENTRY;
1148         OBD_ALLOC(cache, sizeof(*cache));
1149         if (cache == NULL)
1150                 RETURN(NULL);
1151
1152         /* Initialize cache data */
1153         atomic_set(&cache->ccc_users, 1);
1154         cache->ccc_lru_max = lru_page_max;
1155         atomic_long_set(&cache->ccc_lru_left, lru_page_max);
1156         spin_lock_init(&cache->ccc_lru_lock);
1157         INIT_LIST_HEAD(&cache->ccc_lru);
1158
1159         /* turn unstable check off by default as it impacts performance */
1160         cache->ccc_unstable_check = 0;
1161         atomic_long_set(&cache->ccc_unstable_nr, 0);
1162         init_waitqueue_head(&cache->ccc_unstable_waitq);
1163
1164         RETURN(cache);
1165 }
1166 EXPORT_SYMBOL(cl_cache_init);
1167
1168 /**
1169  * Increase cl_cache refcount
1170  */
1171 void cl_cache_incref(struct cl_client_cache *cache)
1172 {
1173         atomic_inc(&cache->ccc_users);
1174 }
1175 EXPORT_SYMBOL(cl_cache_incref);
1176
1177 /**
1178  * Decrease cl_cache refcount and free the cache if refcount=0.
1179  * Since llite, lov and osc all hold cl_cache refcount,
1180  * the free will not cause race. (LU-6173)
1181  */
1182 void cl_cache_decref(struct cl_client_cache *cache)
1183 {
1184         if (atomic_dec_and_test(&cache->ccc_users))
1185                 OBD_FREE(cache, sizeof(*cache));
1186 }
1187 EXPORT_SYMBOL(cl_cache_decref);