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  *
  31  * Client Lustre Page.
  32  *
  33  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  34  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_CLASS
  38
  39 #include <linux/list.h>
  40 #include <libcfs/libcfs.h>
  41 #include <obd_class.h>
  42 #include <obd_support.h>
  43
  44 #include <cl_object.h>
  45 #include "cl_internal.h"
  46
  47 static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg);
  48 static DEFINE_MUTEX(cl_page_kmem_mutex);
  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 static struct cl_page_slice *
 126 cl_page_slice_get(const struct cl_page *cl_page, int index)
 127 {
 128         if (index < 0 || index >= cl_page->cp_layer_count)
 129                 return NULL;
 130
 131         /* To get the cp_layer_offset values fit under 256 bytes, we
 132          * use the offset beyond the end of struct cl_page.
 133          */
 134         return (struct cl_page_slice *)((char *)cl_page + sizeof(*cl_page) +
 135                                         cl_page->cp_layer_offset[index]);
 136 }
 137
 138 #define cl_page_slice_for_each(cl_page, slice, i)               \
 139         for (i = 0, slice = cl_page_slice_get(cl_page, 0);      \
 140              i < (cl_page)->cp_layer_count;                     \
 141              slice = cl_page_slice_get(cl_page, ++i))
 142
 143 #define cl_page_slice_for_each_reverse(cl_page, slice, i)       \
 144         for (i = (cl_page)->cp_layer_count - 1,                 \
 145              slice = cl_page_slice_get(cl_page, i); i >= 0;     \
 146              slice = cl_page_slice_get(cl_page, --i))
 147
 148 /**
 149  * Returns a slice within a cl_page, corresponding to the given layer in the
 150  * device stack.
 151  *
 152  * \see cl_lock_at()
 153  */
 154 static const struct cl_page_slice *
 155 cl_page_at_trusted(const struct cl_page *cl_page,
 156                    const struct lu_device_type *dtype)
 157 {
 158         const struct cl_page_slice *slice;
 159         int i;
 160
 161         ENTRY;
 162
 163         cl_page_slice_for_each(cl_page, slice, i) {
 164                 if (slice->cpl_obj->co_lu.lo_dev->ld_type == dtype)
 165                         RETURN(slice);
 166         }
 167
 168         RETURN(NULL);
 169 }
 170
 171 static void __cl_page_free(struct cl_page *cl_page, unsigned short bufsize)
 172 {
 173         int index = cl_page->cp_kmem_index;
 174
 175         if (index >= 0) {
 176                 LASSERT(index < ARRAY_SIZE(cl_page_kmem_array));
 177                 LASSERT(cl_page_kmem_size_array[index] == bufsize);
 178                 OBD_SLAB_FREE(cl_page, cl_page_kmem_array[index], bufsize);
 179         } else {
 180                 OBD_FREE(cl_page, bufsize);
 181         }
 182 }
 183
 184 static void cl_page_free(const struct lu_env *env, struct cl_page *cl_page,
 185                          struct pagevec *pvec)
 186 {
 187         struct cl_object *obj  = cl_page->cp_obj;
 188         unsigned short bufsize = cl_object_header(obj)->coh_page_bufsize;
 189         struct cl_page_slice *slice;
 190         int i;
 191
 192         ENTRY;
 193         PASSERT(env, cl_page, list_empty(&cl_page->cp_batch));
 194         PASSERT(env, cl_page, cl_page->cp_owner == NULL);
 195         PASSERT(env, cl_page, cl_page->cp_state == CPS_FREEING);
 196
 197         cl_page_slice_for_each(cl_page, slice, i) {
 198                 if (unlikely(slice->cpl_ops->cpo_fini != NULL))
 199                         slice->cpl_ops->cpo_fini(env, slice, pvec);
 200         }
 201         cl_page->cp_layer_count = 0;
 202         cs_page_dec(obj, CS_total);
 203         cs_pagestate_dec(obj, cl_page->cp_state);
 204         lu_object_ref_del_at(&obj->co_lu, &cl_page->cp_obj_ref,
 205                              "cl_page", cl_page);
 206         if (cl_page->cp_type != CPT_TRANSIENT)
 207                 cl_object_put(env, obj);
 208         lu_ref_fini(&cl_page->cp_reference);
 209         __cl_page_free(cl_page, bufsize);
 210         EXIT;
 211 }
 212
 213 static struct cl_page *__cl_page_alloc(struct cl_object *o)
 214 {
 215         int i = 0;
 216         struct cl_page *cl_page = NULL;
 217         unsigned short bufsize = cl_object_header(o)->coh_page_bufsize;
 218
 219         if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_PAGE_ALLOC))
 220                 return NULL;
 221
 222 check:
 223         /* the number of entries in cl_page_kmem_array is expected to
 224          * only be 2-3 entries, so the lookup overhead should be low.
 225          */
 226         for ( ; i < ARRAY_SIZE(cl_page_kmem_array); i++) {
 227                 if (smp_load_acquire(&cl_page_kmem_size_array[i])
 228                     == bufsize) {
 229                         OBD_SLAB_ALLOC_GFP(cl_page, cl_page_kmem_array[i],
 230                                            bufsize, GFP_NOFS);
 231                         if (cl_page)
 232                                 cl_page->cp_kmem_index = i;
 233                         return cl_page;
 234                 }
 235                 if (cl_page_kmem_size_array[i] == 0)
 236                         break;
 237         }
 238
 239         if (i < ARRAY_SIZE(cl_page_kmem_array)) {
 240                 char cache_name[32];
 241
 242                 mutex_lock(&cl_page_kmem_mutex);
 243                 if (cl_page_kmem_size_array[i]) {
 244                         mutex_unlock(&cl_page_kmem_mutex);
 245                         goto check;
 246                 }
 247                 snprintf(cache_name, sizeof(cache_name),
 248                          "cl_page_kmem-%u", bufsize);
 249                 cl_page_kmem_array[i] =
 250                         kmem_cache_create(cache_name, bufsize,
 251                                           0, 0, NULL);
 252                 if (cl_page_kmem_array[i] == NULL) {
 253                         mutex_unlock(&cl_page_kmem_mutex);
 254                         return NULL;
 255                 }
 256                 smp_store_release(&cl_page_kmem_size_array[i],
 257                                   bufsize);
 258                 mutex_unlock(&cl_page_kmem_mutex);
 259                 goto check;
 260         } else {
 261                 OBD_ALLOC_GFP(cl_page, bufsize, GFP_NOFS);
 262                 if (cl_page)
 263                         cl_page->cp_kmem_index = -1;
 264         }
 265
 266         return cl_page;
 267 }
 268
 269 struct cl_page *cl_page_alloc(const struct lu_env *env, struct cl_object *o,
 270                               pgoff_t ind, struct page *vmpage,
 271                               enum cl_page_type type)
 272 {
 273         struct cl_page *cl_page;
 274         struct cl_object *head;
 275
 276         ENTRY;
 277
 278         cl_page = __cl_page_alloc(o);
 279         if (cl_page != NULL) {
 280                 int result = 0;
 281
 282                 /*
 283                  * Please fix cl_page:cp_state/type declaration if
 284                  * these assertions fail in the future.
 285                  */
 286                 BUILD_BUG_ON((1 << CP_STATE_BITS) < CPS_NR); /* cp_state */
 287                 BUILD_BUG_ON((1 << CP_TYPE_BITS) < CPT_NR); /* cp_type */
 288                 atomic_set(&cl_page->cp_ref, 1);
 289                 cl_page->cp_obj = o;
 290                 if (type != CPT_TRANSIENT)
 291                         cl_object_get(o);
 292                 lu_object_ref_add_at(&o->co_lu, &cl_page->cp_obj_ref,
 293                                      "cl_page", cl_page);
 294                 cl_page->cp_vmpage = vmpage;
 295                 cl_page->cp_state = CPS_CACHED;
 296                 cl_page->cp_type = type;
 297                 if (type == CPT_TRANSIENT)
 298                         /* ref to correct inode will be added
 299                          * in ll_direct_rw_pages
 300                          */
 301                         cl_page->cp_inode = NULL;
 302                 else
 303                         cl_page->cp_inode = page2inode(vmpage);
 304                 INIT_LIST_HEAD(&cl_page->cp_batch);
 305                 lu_ref_init(&cl_page->cp_reference);
 306                 head = o;
 307                 cl_page->cp_page_index = ind;
 308                 cl_object_for_each(o, head) {
 309                         if (o->co_ops->coo_page_init != NULL) {
 310                                 result = o->co_ops->coo_page_init(env, o,
 311                                                         cl_page, ind);
 312                                 if (result != 0) {
 313                                         cl_page_delete0(env, cl_page);
 314                                         cl_page_free(env, cl_page, NULL);
 315                                         cl_page = ERR_PTR(result);
 316                                         break;
 317                                 }
 318                         }
 319                 }
 320                 if (result == 0) {
 321                         cs_page_inc(o, CS_total);
 322                         cs_page_inc(o, CS_create);
 323                         cs_pagestate_dec(o, CPS_CACHED);
 324                 }
 325         } else {
 326                 cl_page = ERR_PTR(-ENOMEM);
 327         }
 328         RETURN(cl_page);
 329 }
 330
 331 /**
 332  * Returns a cl_page with index \a idx at the object \a o, and associated with
 333  * the VM page \a vmpage.
 334  *
 335  * This is the main entry point into the cl_page caching interface. First, a
 336  * cache (implemented as a per-object radix tree) is consulted. If page is
 337  * found there, it is returned immediately. Otherwise new page is allocated
 338  * and returned. In any case, additional reference to page is acquired.
 339  *
 340  * \see cl_object_find(), cl_lock_find()
 341  */
 342 struct cl_page *cl_page_find(const struct lu_env *env,
 343                              struct cl_object *o,
 344                              pgoff_t idx, struct page *vmpage,
 345                              enum cl_page_type type)
 346 {
 347         struct cl_page          *page = NULL;
 348         struct cl_object_header *hdr;
 349
 350         LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
 351         might_sleep();
 352
 353         ENTRY;
 354
 355         hdr = cl_object_header(o);
 356         cs_page_inc(o, CS_lookup);
 357
 358         CDEBUG(D_PAGE, "%lu@"DFID" %p %lx %d\n",
 359                idx, PFID(&hdr->coh_lu.loh_fid), vmpage, vmpage->private, type);
 360         /* fast path. */
 361         if (type == CPT_CACHEABLE) {
 362                 /* vmpage lock is used to protect the child/parent
 363                  * relationship */
 364                 LASSERT(PageLocked(vmpage));
 365                 /*
 366                  * cl_vmpage_page() can be called here without any locks as
 367                  *
 368                  *     - "vmpage" is locked (which prevents ->private from
 369                  *       concurrent updates), and
 370                  *
 371                  *     - "o" cannot be destroyed while current thread holds a
 372                  *       reference on it.
 373                  */
 374                 page = cl_vmpage_page(vmpage, o);
 375                 if (page != NULL) {
 376                         cs_page_inc(o, CS_hit);
 377                         RETURN(page);
 378                 }
 379         }
 380
 381         /* allocate and initialize cl_page */
 382         page = cl_page_alloc(env, o, idx, vmpage, type);
 383         RETURN(page);
 384 }
 385 EXPORT_SYMBOL(cl_page_find);
 386
 387 static inline int cl_page_invariant(const struct cl_page *pg)
 388 {
 389         return cl_page_in_use_noref(pg);
 390 }
 391
 392 static void cl_page_state_set0(const struct lu_env *env,
 393                                struct cl_page *cl_page,
 394                                enum cl_page_state state)
 395 {
 396         enum cl_page_state old;
 397
 398         /*
 399          * Matrix of allowed state transitions [old][new], for sanity
 400          * checking.
 401          */
 402         static const int allowed_transitions[CPS_NR][CPS_NR] = {
 403                 [CPS_CACHED] = {
 404                         [CPS_CACHED]  = 0,
 405                         [CPS_OWNED]   = 1, /* io finds existing cached page */
 406                         [CPS_PAGEIN]  = 0,
 407                         [CPS_PAGEOUT] = 1, /* write-out from the cache */
 408                         [CPS_FREEING] = 1, /* eviction on the memory pressure */
 409                 },
 410                 [CPS_OWNED] = {
 411                         [CPS_CACHED]  = 1, /* release to the cache */
 412                         [CPS_OWNED]   = 0,
 413                         [CPS_PAGEIN]  = 1, /* start read immediately */
 414                         [CPS_PAGEOUT] = 1, /* start write immediately */
 415                         [CPS_FREEING] = 1, /* lock invalidation or truncate */
 416                 },
 417                 [CPS_PAGEIN] = {
 418                         [CPS_CACHED]  = 1, /* io completion */
 419                         [CPS_OWNED]   = 0,
 420                         [CPS_PAGEIN]  = 0,
 421                         [CPS_PAGEOUT] = 0,
 422                         [CPS_FREEING] = 0,
 423                 },
 424                 [CPS_PAGEOUT] = {
 425                         [CPS_CACHED]  = 1, /* io completion */
 426                         [CPS_OWNED]   = 0,
 427                         [CPS_PAGEIN]  = 0,
 428                         [CPS_PAGEOUT] = 0,
 429                         [CPS_FREEING] = 0,
 430                 },
 431                 [CPS_FREEING] = {
 432                         [CPS_CACHED]  = 0,
 433                         [CPS_OWNED]   = 0,
 434                         [CPS_PAGEIN]  = 0,
 435                         [CPS_PAGEOUT] = 0,
 436                         [CPS_FREEING] = 0,
 437                 }
 438         };
 439
 440         ENTRY;
 441         old = cl_page->cp_state;
 442         PASSERT(env, cl_page, allowed_transitions[old][state]);
 443         CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d -> %d\n", old, state);
 444         PASSERT(env, cl_page, cl_page->cp_state == old);
 445         PASSERT(env, cl_page, equi(state == CPS_OWNED,
 446                                    cl_page->cp_owner != NULL));
 447
 448         cs_pagestate_dec(cl_page->cp_obj, cl_page->cp_state);
 449         cs_pagestate_inc(cl_page->cp_obj, state);
 450         cl_page->cp_state = state;
 451         EXIT;
 452 }
 453
 454 static void cl_page_state_set(const struct lu_env *env,
 455                               struct cl_page *page, enum cl_page_state state)
 456 {
 457         cl_page_state_set0(env, page, state);
 458 }
 459
 460 /**
 461  * Acquires an additional reference to a page.
 462  *
 463  * This can be called only by caller already possessing a reference to \a
 464  * page.
 465  *
 466  * \see cl_object_get(), cl_lock_get().
 467  */
 468 void cl_page_get(struct cl_page *page)
 469 {
 470         ENTRY;
 471         cl_page_get_trust(page);
 472         EXIT;
 473 }
 474 EXPORT_SYMBOL(cl_page_get);
 475
 476 /**
 477  * Releases a reference to a page, use the pagevec to release the pages
 478  * in batch if provided.
 479  *
 480  * Users need to do a final pagevec_release() to release any trailing pages.
 481  */
 482 void cl_pagevec_put(const struct lu_env *env, struct cl_page *page,
 483                   struct pagevec *pvec)
 484 {
 485         ENTRY;
 486         CL_PAGE_HEADER(D_TRACE, env, page, "%d\n",
 487                        atomic_read(&page->cp_ref));
 488
 489         if (atomic_dec_and_test(&page->cp_ref)) {
 490                 LASSERT(page->cp_state == CPS_FREEING);
 491
 492                 LASSERT(atomic_read(&page->cp_ref) == 0);
 493                 PASSERT(env, page, page->cp_owner == NULL);
 494                 PASSERT(env, page, list_empty(&page->cp_batch));
 495                 /*
 496                  * Page is no longer reachable by other threads. Tear
 497                  * it down.
 498                  */
 499                 cl_page_free(env, page, pvec);
 500         }
 501
 502         EXIT;
 503 }
 504 EXPORT_SYMBOL(cl_pagevec_put);
 505
 506 /**
 507  * Releases a reference to a page, wrapper to cl_pagevec_put
 508  *
 509  * When last reference is released, page is returned to the cache, unless it
 510  * is in cl_page_state::CPS_FREEING state, in which case it is immediately
 511  * destroyed.
 512  *
 513  * \see cl_object_put(), cl_lock_put().
 514  */
 515 void cl_page_put(const struct lu_env *env, struct cl_page *page)
 516 {
 517         cl_pagevec_put(env, page, NULL);
 518 }
 519 EXPORT_SYMBOL(cl_page_put);
 520
 521 /**
 522  * Returns a cl_page associated with a VM page, and given cl_object.
 523  */
 524 struct cl_page *cl_vmpage_page(struct page *vmpage, struct cl_object *obj)
 525 {
 526         struct cl_page *page;
 527
 528         ENTRY;
 529         LASSERT(PageLocked(vmpage));
 530
 531         /*
 532          * NOTE: absence of races and liveness of data are guaranteed by page
 533          *       lock on a "vmpage". That works because object destruction has
 534          *       bottom-to-top pass.
 535          */
 536
 537         page = (struct cl_page *)vmpage->private;
 538         if (page != NULL) {
 539                 cl_page_get_trust(page);
 540                 LASSERT(page->cp_type == CPT_CACHEABLE);
 541         }
 542         RETURN(page);
 543 }
 544 EXPORT_SYMBOL(cl_vmpage_page);
 545
 546 const struct cl_page_slice *cl_page_at(const struct cl_page *page,
 547                                        const struct lu_device_type *dtype)
 548 {
 549         return cl_page_at_trusted(page, dtype);
 550 }
 551 EXPORT_SYMBOL(cl_page_at);
 552
 553 static void cl_page_owner_clear(struct cl_page *page)
 554 {
 555         ENTRY;
 556         if (page->cp_owner != NULL) {
 557                 LASSERT(page->cp_owner->ci_owned_nr > 0);
 558                 page->cp_owner->ci_owned_nr--;
 559                 page->cp_owner = NULL;
 560         }
 561         EXIT;
 562 }
 563
 564 static void cl_page_owner_set(struct cl_page *page)
 565 {
 566         ENTRY;
 567         LASSERT(page->cp_owner != NULL);
 568         page->cp_owner->ci_owned_nr++;
 569         EXIT;
 570 }
 571
 572 void cl_page_disown0(const struct lu_env *env, struct cl_page *cp)
 573 {
 574         struct page *vmpage;
 575         enum cl_page_state state;
 576
 577         ENTRY;
 578         state = cp->cp_state;
 579         PINVRNT(env, cp, state == CPS_OWNED || state == CPS_FREEING);
 580         PINVRNT(env, cp, cl_page_invariant(cp) || state == CPS_FREEING);
 581         cl_page_owner_clear(cp);
 582
 583         if (state == CPS_OWNED)
 584                 cl_page_state_set(env, cp, CPS_CACHED);
 585
 586         if (cp->cp_type == CPT_CACHEABLE) {
 587                 vmpage = cp->cp_vmpage;
 588                 LASSERT(vmpage != NULL);
 589                 LASSERT(PageLocked(vmpage));
 590                 unlock_page(vmpage);
 591         }
 592
 593         EXIT;
 594 }
 595
 596 /**
 597  * returns true, iff page is owned by the given io.
 598  */
 599 int cl_page_is_owned(const struct cl_page *pg, const struct cl_io *io)
 600 {
 601         struct cl_io *top = cl_io_top((struct cl_io *)io);
 602         LINVRNT(cl_object_same(pg->cp_obj, io->ci_obj));
 603         ENTRY;
 604         RETURN(pg->cp_state == CPS_OWNED && pg->cp_owner == top);
 605 }
 606 EXPORT_SYMBOL(cl_page_is_owned);
 607
 608 /**
 609  * Try to own a page by IO.
 610  *
 611  * Waits until page is in cl_page_state::CPS_CACHED state, and then switch it
 612  * into cl_page_state::CPS_OWNED state.
 613  *
 614  * \pre  !cl_page_is_owned(cl_page, io)
 615  * \post result == 0 iff cl_page_is_owned(cl_page, io)
 616  *
 617  * \retval 0   success
 618  *
 619  * \retval -ve failure, e.g., cl_page was destroyed (and landed in
 620  *             cl_page_state::CPS_FREEING instead of cl_page_state::CPS_CACHED).
 621  *             or, page was owned by another thread, or in IO.
 622  *
 623  * \see cl_page_disown()
 624  * \see cl_page_own_try()
 625  * \see cl_page_own
 626  */
 627 static int cl_page_own0(const struct lu_env *env, struct cl_io *io,
 628                         struct cl_page *cl_page, int nonblock)
 629 {
 630         struct page *vmpage = cl_page->cp_vmpage;
 631         int result;
 632
 633         ENTRY;
 634         PINVRNT(env, cl_page, !cl_page_is_owned(cl_page, io));
 635
 636         if (cl_page->cp_state == CPS_FREEING) {
 637                 result = -ENOENT;
 638                 goto out;
 639         }
 640
 641         LASSERT(vmpage != NULL);
 642
 643         if (cl_page->cp_type == CPT_TRANSIENT) {
 644                 /* OK */
 645         } else if (nonblock) {
 646                 if (!trylock_page(vmpage)) {
 647                         result = -EAGAIN;
 648                         goto out;
 649                 }
 650
 651                 if (unlikely(PageWriteback(vmpage))) {
 652                         unlock_page(vmpage);
 653                         result = -EAGAIN;
 654                         goto out;
 655                 }
 656         } else {
 657                 lock_page(vmpage);
 658                 wait_on_page_writeback(vmpage);
 659         }
 660
 661         PASSERT(env, cl_page, cl_page->cp_owner == NULL);
 662         cl_page->cp_owner = cl_io_top(io);
 663         cl_page_owner_set(cl_page);
 664
 665         if (cl_page->cp_state == CPS_FREEING) {
 666                 cl_page_disown0(env, cl_page);
 667                 result = -ENOENT;
 668                 goto out;
 669         }
 670
 671         cl_page_state_set(env, cl_page, CPS_OWNED);
 672         result = 0;
 673 out:
 674         PINVRNT(env, cl_page, ergo(result == 0,
 675                 cl_page_invariant(cl_page)));
 676         RETURN(result);
 677 }
 678
 679 /**
 680  * Own a page, might be blocked.
 681  *
 682  * \see cl_page_own0()
 683  */
 684 int cl_page_own(const struct lu_env *env, struct cl_io *io, struct cl_page *pg)
 685 {
 686         return cl_page_own0(env, io, pg, 0);
 687 }
 688 EXPORT_SYMBOL(cl_page_own);
 689
 690 /**
 691  * Nonblock version of cl_page_own().
 692  *
 693  * \see cl_page_own0()
 694  */
 695 int cl_page_own_try(const struct lu_env *env, struct cl_io *io,
 696                     struct cl_page *pg)
 697 {
 698         return cl_page_own0(env, io, pg, 1);
 699 }
 700 EXPORT_SYMBOL(cl_page_own_try);
 701
 702
 703 /**
 704  * Assume page ownership.
 705  *
 706  * Called when page is already locked by the hosting VM.
 707  *
 708  * \pre !cl_page_is_owned(cl_page, io)
 709  * \post cl_page_is_owned(cl_page, io)
 710  *
 711  * \see cl_page_operations::cpo_assume()
 712  */
 713 void cl_page_assume(const struct lu_env *env,
 714                     struct cl_io *io, struct cl_page *cl_page)
 715 {
 716         const struct cl_page_slice *slice;
 717         int i;
 718
 719         ENTRY;
 720
 721         PINVRNT(env, cl_page,
 722                 cl_object_same(cl_page->cp_obj, io->ci_obj));
 723         io = cl_io_top(io);
 724
 725         cl_page_slice_for_each(cl_page, slice, i) {
 726                 if (slice->cpl_ops->cpo_assume != NULL)
 727                         (*slice->cpl_ops->cpo_assume)(env, slice, io);
 728         }
 729
 730         PASSERT(env, cl_page, cl_page->cp_owner == NULL);
 731         cl_page->cp_owner = cl_io_top(io);
 732         cl_page_owner_set(cl_page);
 733         cl_page_state_set(env, cl_page, CPS_OWNED);
 734         EXIT;
 735 }
 736 EXPORT_SYMBOL(cl_page_assume);
 737
 738 /**
 739  * Releases page ownership without unlocking the page.
 740  *
 741  * Moves cl_page into cl_page_state::CPS_CACHED without releasing a lock
 742  * on the underlying VM page (as VM is supposed to do this itself).
 743  *
 744  * \pre   cl_page_is_owned(cl_page, io)
 745  * \post !cl_page_is_owned(cl_page, io)
 746  *
 747  * \see cl_page_assume()
 748  */
 749 void cl_page_unassume(const struct lu_env *env,
 750                       struct cl_io *io, struct cl_page *cl_page)
 751 {
 752         const struct cl_page_slice *slice;
 753         int i;
 754
 755         ENTRY;
 756         PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
 757         PINVRNT(env, cl_page, cl_page_invariant(cl_page));
 758
 759         io = cl_io_top(io);
 760         cl_page_owner_clear(cl_page);
 761         cl_page_state_set(env, cl_page, CPS_CACHED);
 762
 763         cl_page_slice_for_each_reverse(cl_page, slice, i) {
 764                 if (slice->cpl_ops->cpo_unassume != NULL)
 765                         (*slice->cpl_ops->cpo_unassume)(env, slice, io);
 766         }
 767
 768         EXIT;
 769 }
 770 EXPORT_SYMBOL(cl_page_unassume);
 771
 772 /**
 773  * Releases page ownership.
 774  *
 775  * Moves page into cl_page_state::CPS_CACHED.
 776  *
 777  * \pre   cl_page_is_owned(pg, io)
 778  * \post !cl_page_is_owned(pg, io)
 779  *
 780  * \see cl_page_own()
 781  */
 782 void cl_page_disown(const struct lu_env *env,
 783                     struct cl_io *io, struct cl_page *pg)
 784 {
 785         PINVRNT(env, pg, cl_page_is_owned(pg, io) ||
 786                 pg->cp_state == CPS_FREEING);
 787
 788         cl_page_disown0(env, pg);
 789 }
 790 EXPORT_SYMBOL(cl_page_disown);
 791
 792 /**
 793  * Called when cl_page is to be removed from the object, e.g.,
 794  * as a result of truncate.
 795  *
 796  * Calls cl_page_operations::cpo_discard() top-to-bottom.
 797  *
 798  * \pre cl_page_is_owned(cl_page, io)
 799  *
 800  * \see cl_page_operations::cpo_discard()
 801  */
 802 void cl_page_discard(const struct lu_env *env,
 803                      struct cl_io *io, struct cl_page *cp)
 804 {
 805         struct page *vmpage;
 806         const struct cl_page_slice *slice;
 807         int i;
 808
 809         PINVRNT(env, cp, cl_page_is_owned(cp, io));
 810         PINVRNT(env, cp, cl_page_invariant(cp));
 811
 812         cl_page_slice_for_each(cp, slice, i) {
 813                 if (slice->cpl_ops->cpo_discard != NULL)
 814                         (*slice->cpl_ops->cpo_discard)(env, slice, io);
 815         }
 816
 817         if (cp->cp_type == CPT_CACHEABLE) {
 818                 vmpage = cp->cp_vmpage;
 819                 LASSERT(vmpage != NULL);
 820                 LASSERT(PageLocked(vmpage));
 821                 generic_error_remove_page(vmpage->mapping, vmpage);
 822         } else {
 823                 cl_page_delete(env, cp);
 824         }
 825 }
 826 EXPORT_SYMBOL(cl_page_discard);
 827
 828 /**
 829  * Version of cl_page_delete() that can be called for not fully constructed
 830  * cl_pages, e.g. in an error handling cl_page_find()->cl_page_delete0()
 831  * path. Doesn't check cl_page invariant.
 832  */
 833 static void cl_page_delete0(const struct lu_env *env,
 834                             struct cl_page *cl_page)
 835 {
 836         const struct cl_page_slice *slice;
 837         int i;
 838
 839         ENTRY;
 840
 841         PASSERT(env, cl_page, cl_page->cp_state != CPS_FREEING);
 842
 843         /*
 844          * Severe all ways to obtain new pointers to @pg.
 845          */
 846         cl_page_owner_clear(cl_page);
 847         cl_page_state_set0(env, cl_page, CPS_FREEING);
 848
 849         cl_page_slice_for_each_reverse(cl_page, slice, i) {
 850                 if (slice->cpl_ops->cpo_delete != NULL)
 851                         (*slice->cpl_ops->cpo_delete)(env, slice);
 852         }
 853
 854         EXIT;
 855 }
 856
 857 /**
 858  * Called when a decision is made to throw page out of memory.
 859  *
 860  * Notifies all layers about page destruction by calling
 861  * cl_page_operations::cpo_delete() method top-to-bottom.
 862  *
 863  * Moves page into cl_page_state::CPS_FREEING state (this is the only place
 864  * where transition to this state happens).
 865  *
 866  * Eliminates all venues through which new references to the page can be
 867  * obtained:
 868  *
 869  *     - removes page from the radix trees,
 870  *
 871  *     - breaks linkage from VM page to cl_page.
 872  *
 873  * Once page reaches cl_page_state::CPS_FREEING, all remaining references will
 874  * drain after some time, at which point page will be recycled.
 875  *
 876  * \pre  VM page is locked
 877  * \post pg->cp_state == CPS_FREEING
 878  *
 879  * \see cl_page_operations::cpo_delete()
 880  */
 881 void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
 882 {
 883         PINVRNT(env, pg, cl_page_invariant(pg));
 884         ENTRY;
 885         cl_page_delete0(env, pg);
 886         EXIT;
 887 }
 888 EXPORT_SYMBOL(cl_page_delete);
 889
 890 void cl_page_touch(const struct lu_env *env,
 891                    const struct cl_page *cl_page, size_t to)
 892 {
 893         const struct cl_page_slice *slice;
 894         int i;
 895
 896         ENTRY;
 897
 898         cl_page_slice_for_each(cl_page, slice, i) {
 899                 if (slice->cpl_ops->cpo_page_touch != NULL)
 900                         (*slice->cpl_ops->cpo_page_touch)(env, slice, to);
 901         }
 902
 903         EXIT;
 904 }
 905 EXPORT_SYMBOL(cl_page_touch);
 906
 907 static enum cl_page_state cl_req_type_state(enum cl_req_type crt)
 908 {
 909         ENTRY;
 910         RETURN(crt == CRT_WRITE ? CPS_PAGEOUT : CPS_PAGEIN);
 911 }
 912
 913 static void cl_page_io_start(const struct lu_env *env,
 914                              struct cl_page *pg, enum cl_req_type crt)
 915 {
 916         /*
 917          * Page is queued for IO, change its state.
 918          */
 919         ENTRY;
 920         cl_page_owner_clear(pg);
 921         cl_page_state_set(env, pg, cl_req_type_state(crt));
 922         EXIT;
 923 }
 924
 925 /**
 926  * Prepares page for immediate transfer. cl_page_operations::cpo_prep() is
 927  * called top-to-bottom. Every layer either agrees to submit this page (by
 928  * returning 0), or requests to omit this page (by returning -EALREADY). Layer
 929  * handling interactions with the VM also has to inform VM that page is under
 930  * transfer now.
 931  */
 932 int cl_page_prep(const struct lu_env *env, struct cl_io *io,
 933                  struct cl_page *cl_page, enum cl_req_type crt)
 934 {
 935         const struct cl_page_slice *slice;
 936         int result = 0;
 937         int i;
 938
 939         PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
 940         PINVRNT(env, cl_page, cl_page_invariant(cl_page));
 941         PINVRNT(env, cl_page, crt < CRT_NR);
 942
 943         /*
 944          * this has to be called bottom-to-top, so that llite can set up
 945          * PG_writeback without risking other layers deciding to skip this
 946          * page.
 947          */
 948         if (crt >= CRT_NR)
 949                 return -EINVAL;
 950
 951         if (cl_page->cp_type != CPT_TRANSIENT) {
 952                 cl_page_slice_for_each(cl_page, slice, i) {
 953                         if (slice->cpl_ops->io[crt].cpo_prep)
 954                                 result =
 955                                  (*slice->cpl_ops->io[crt].cpo_prep)(env,
 956                                                                      slice,
 957                                                                      io);
 958                         if (result != 0)
 959                                 break;
 960                 }
 961         }
 962
 963         if (result >= 0) {
 964                 result = 0;
 965                 cl_page_io_start(env, cl_page, crt);
 966         }
 967
 968         CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", crt, result);
 969         return result;
 970 }
 971 EXPORT_SYMBOL(cl_page_prep);
 972
 973 /**
 974  * Notify layers about transfer completion.
 975  *
 976  * Invoked by transfer sub-system (which is a part of osc) to notify layers
 977  * that a transfer, of which this page is a part of has completed.
 978  *
 979  * Completion call-backs are executed in the bottom-up order, so that
 980  * uppermost layer (llite), responsible for the VFS/VM interaction runs last
 981  * and can release locks safely.
 982  *
 983  * \pre  cl_page->cp_state == CPS_PAGEIN || cl_page->cp_state == CPS_PAGEOUT
 984  * \post cl_page->cl_page_state == CPS_CACHED
 985  *
 986  * \see cl_page_operations::cpo_completion()
 987  */
 988 void cl_page_completion(const struct lu_env *env,
 989                         struct cl_page *cl_page, enum cl_req_type crt,
 990                         int ioret)
 991 {
 992         const struct cl_page_slice *slice;
 993         struct cl_sync_io *anchor = cl_page->cp_sync_io;
 994         int i;
 995
 996         ENTRY;
 997         PASSERT(env, cl_page, crt < CRT_NR);
 998         PASSERT(env, cl_page, cl_page->cp_state == cl_req_type_state(crt));
 999
1000         CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", crt, ioret);
1001         cl_page_state_set(env, cl_page, CPS_CACHED);
1002         if (crt >= CRT_NR)
1003                 return;
1004
1005         cl_page_slice_for_each_reverse(cl_page, slice, i) {
1006                 if (slice->cpl_ops->io[crt].cpo_completion != NULL)
1007                         (*slice->cpl_ops->io[crt].cpo_completion)(env, slice,
1008                                                                   ioret);
1009         }
1010
1011         if (anchor != NULL) {
1012                 LASSERT(cl_page->cp_sync_io == anchor);
1013                 cl_page->cp_sync_io = NULL;
1014                 cl_sync_io_note(env, anchor, ioret);
1015         }
1016         EXIT;
1017 }
1018 EXPORT_SYMBOL(cl_page_completion);
1019
1020 /**
1021  * Notify layers that transfer formation engine decided to yank this page from
1022  * the cache and to make it a part of a transfer.
1023  *
1024  * \pre  cl_page->cp_state == CPS_CACHED
1025  * \post cl_page->cp_state == CPS_PAGEIN || cl_page->cp_state == CPS_PAGEOUT
1026  *
1027  * \see cl_page_operations::cpo_make_ready()
1028  */
1029 int cl_page_make_ready(const struct lu_env *env, struct cl_page *cl_page,
1030                        enum cl_req_type crt)
1031 {
1032         const struct cl_page_slice *slice;
1033         int result = 0;
1034         int i;
1035
1036         ENTRY;
1037         PINVRNT(env, cl_page, crt < CRT_NR);
1038         if (crt >= CRT_NR)
1039                 RETURN(-EINVAL);
1040
1041         cl_page_slice_for_each(cl_page, slice, i) {
1042                 if (slice->cpl_ops->io[crt].cpo_make_ready != NULL)
1043                         result = (*slice->cpl_ops->io[crt].cpo_make_ready)(env, slice);
1044                 if (result != 0)
1045                         break;
1046         }
1047
1048         if (result >= 0) {
1049                 result = 0;
1050                 PASSERT(env, cl_page, cl_page->cp_state == CPS_CACHED);
1051                 cl_page_io_start(env, cl_page, crt);
1052         }
1053         CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", crt, result);
1054
1055         RETURN(result);
1056 }
1057 EXPORT_SYMBOL(cl_page_make_ready);
1058
1059 /**
1060  * Called if a page is being written back by kernel's intention.
1061  *
1062  * \pre  cl_page_is_owned(cl_page, io)
1063  * \post ergo(result == 0, cl_page->cp_state == CPS_PAGEOUT)
1064  *
1065  * \see cl_page_operations::cpo_flush()
1066  */
1067 int cl_page_flush(const struct lu_env *env, struct cl_io *io,
1068                   struct cl_page *cl_page)
1069 {
1070         const struct cl_page_slice *slice;
1071         int result = 0;
1072         int i;
1073
1074         ENTRY;
1075         PINVRNT(env, cl_page, cl_page_is_owned(cl_page, io));
1076         PINVRNT(env, cl_page, cl_page_invariant(cl_page));
1077
1078         cl_page_slice_for_each(cl_page, slice, i) {
1079                 if (slice->cpl_ops->cpo_flush != NULL)
1080                         result = (*slice->cpl_ops->cpo_flush)(env, slice, io);
1081                 if (result != 0)
1082                         break;
1083         }
1084         if (result > 0)
1085                 result = 0;
1086
1087         CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d\n", result);
1088         RETURN(result);
1089 }
1090 EXPORT_SYMBOL(cl_page_flush);
1091
1092 /**
1093  * Tells transfer engine that only part of a page is to be transmitted.
1094  *
1095  * \see cl_page_operations::cpo_clip()
1096  */
1097 void cl_page_clip(const struct lu_env *env, struct cl_page *cl_page,
1098                   int from, int to)
1099 {
1100         const struct cl_page_slice *slice;
1101         int i;
1102
1103         PINVRNT(env, cl_page, cl_page_invariant(cl_page));
1104
1105         CL_PAGE_HEADER(D_TRACE, env, cl_page, "%d %d\n", from, to);
1106         cl_page_slice_for_each(cl_page, slice, i) {
1107                 if (slice->cpl_ops->cpo_clip != NULL)
1108                         (*slice->cpl_ops->cpo_clip)(env, slice, from, to);
1109         }
1110 }
1111 EXPORT_SYMBOL(cl_page_clip);
1112
1113 /**
1114  * Prints human readable representation of \a pg to the \a f.
1115  */
1116 void cl_page_header_print(const struct lu_env *env, void *cookie,
1117                           lu_printer_t printer, const struct cl_page *pg)
1118 {
1119         (*printer)(env, cookie,
1120                    "page@%p[%d %p %d %d %p]\n",
1121                    pg, atomic_read(&pg->cp_ref), pg->cp_obj,
1122                    pg->cp_state, pg->cp_type,
1123                    pg->cp_owner);
1124 }
1125 EXPORT_SYMBOL(cl_page_header_print);
1126
1127 /**
1128  * Prints human readable representation of \a cl_page to the \a f.
1129  */
1130 void cl_page_print(const struct lu_env *env, void *cookie,
1131                    lu_printer_t printer, const struct cl_page *cl_page)
1132 {
1133         const struct cl_page_slice *slice;
1134         int result = 0;
1135         int i;
1136
1137         cl_page_header_print(env, cookie, printer, cl_page);
1138         cl_page_slice_for_each(cl_page, slice, i) {
1139                 if (slice->cpl_ops->cpo_print != NULL)
1140                         result = (*slice->cpl_ops->cpo_print)(env, slice,
1141                                                              cookie, printer);
1142                 if (result != 0)
1143                         break;
1144         }
1145         (*printer)(env, cookie, "end page@%p\n", cl_page);
1146 }
1147 EXPORT_SYMBOL(cl_page_print);
1148
1149 /**
1150  * Converts a byte offset within object \a obj into a page index.
1151  */
1152 loff_t cl_offset(const struct cl_object *obj, pgoff_t idx)
1153 {
1154         return (loff_t)idx << PAGE_SHIFT;
1155 }
1156 EXPORT_SYMBOL(cl_offset);
1157
1158 /**
1159  * Converts a page index into a byte offset within object \a obj.
1160  */
1161 pgoff_t cl_index(const struct cl_object *obj, loff_t offset)
1162 {
1163         return offset >> PAGE_SHIFT;
1164 }
1165 EXPORT_SYMBOL(cl_index);
1166
1167 size_t cl_page_size(const struct cl_object *obj)
1168 {
1169         return 1UL << PAGE_SHIFT;
1170 }
1171 EXPORT_SYMBOL(cl_page_size);
1172
1173 /**
1174  * Adds page slice to the compound page.
1175  *
1176  * This is called by cl_object_operations::coo_page_init() methods to add a
1177  * per-layer state to the page. New state is added at the end of
1178  * cl_page::cp_layers list, that is, it is at the bottom of the stack.
1179  *
1180  * \see cl_lock_slice_add(), cl_req_slice_add(), cl_io_slice_add()
1181  */
1182 void cl_page_slice_add(struct cl_page *cl_page, struct cl_page_slice *slice,
1183                        struct cl_object *obj,
1184                        const struct cl_page_operations *ops)
1185 {
1186         unsigned int offset = (char *)slice -
1187                         ((char *)cl_page + sizeof(*cl_page));
1188
1189         ENTRY;
1190         LASSERT(cl_page->cp_layer_count < CP_MAX_LAYER);
1191         LASSERT(offset < (1 << sizeof(cl_page->cp_layer_offset[0]) * 8));
1192         cl_page->cp_layer_offset[cl_page->cp_layer_count++] = offset;
1193         slice->cpl_obj  = obj;
1194         slice->cpl_ops  = ops;
1195         slice->cpl_page = cl_page;
1196
1197         EXIT;
1198 }
1199 EXPORT_SYMBOL(cl_page_slice_add);
1200
1201 /**
1202  * Allocate and initialize cl_cache, called by ll_init_sbi().
1203  */
1204 struct cl_client_cache *cl_cache_init(unsigned long lru_page_max)
1205 {
1206         struct cl_client_cache  *cache = NULL;
1207
1208         ENTRY;
1209         OBD_ALLOC(cache, sizeof(*cache));
1210         if (cache == NULL)
1211                 RETURN(NULL);
1212
1213         /* Initialize cache data */
1214         atomic_set(&cache->ccc_users, 1);
1215         cache->ccc_lru_max = lru_page_max;
1216         atomic_long_set(&cache->ccc_lru_left, lru_page_max);
1217         spin_lock_init(&cache->ccc_lru_lock);
1218         INIT_LIST_HEAD(&cache->ccc_lru);
1219
1220         /* turn unstable check off by default as it impacts performance */
1221         cache->ccc_unstable_check = 0;
1222         atomic_long_set(&cache->ccc_unstable_nr, 0);
1223         init_waitqueue_head(&cache->ccc_unstable_waitq);
1224         mutex_init(&cache->ccc_max_cache_mb_lock);
1225
1226         RETURN(cache);
1227 }
1228 EXPORT_SYMBOL(cl_cache_init);
1229
1230 /**
1231  * Increase cl_cache refcount
1232  */
1233 void cl_cache_incref(struct cl_client_cache *cache)
1234 {
1235         atomic_inc(&cache->ccc_users);
1236 }
1237 EXPORT_SYMBOL(cl_cache_incref);
1238
1239 /**
1240  * Decrease cl_cache refcount and free the cache if refcount=0.
1241  * Since llite, lov and osc all hold cl_cache refcount,
1242  * the free will not cause race. (LU-6173)
1243  */
1244 void cl_cache_decref(struct cl_client_cache *cache)
1245 {
1246         if (atomic_dec_and_test(&cache->ccc_users))
1247                 OBD_FREE(cache, sizeof(*cache));
1248 }
1249 EXPORT_SYMBOL(cl_cache_decref);