lustre/llite/rw.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/llite/rw.c
  37  *
  38  * Lustre Lite I/O page cache routines shared by different kernel revs
  39  */
  40
  41 #include <linux/autoconf.h>
  42 #include <linux/kernel.h>
  43 #include <linux/mm.h>
  44 #include <linux/string.h>
  45 #include <linux/stat.h>
  46 #include <linux/errno.h>
  47 #include <linux/smp_lock.h>
  48 #include <linux/unistd.h>
  49 #include <linux/version.h>
  50 #include <asm/system.h>
  51 #include <asm/uaccess.h>
  52
  53 #include <linux/fs.h>
  54 #include <linux/stat.h>
  55 #include <asm/uaccess.h>
  56 #include <linux/mm.h>
  57 #include <linux/pagemap.h>
  58 #include <linux/smp_lock.h>
  59 /* current_is_kswapd() */
  60 #include <linux/swap.h>
  61
  62 #define DEBUG_SUBSYSTEM S_LLITE
  63
  64 //#include <lustre_mdc.h>
  65 #include <lustre_lite.h>
  66 #include <obd_cksum.h>
  67 #include "llite_internal.h"
  68 #include <linux/lustre_compat25.h>
  69
  70 /* this isn't where truncate starts.   roughly:
  71  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
  72  * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
  73  * avoid races.
  74  *
  75  * must be called under ->lli_size_sem */
  76 void ll_truncate(struct inode *inode)
  77 {
  78         struct ll_inode_info *lli = ll_i2info(inode);
  79         loff_t new_size;
  80         ENTRY;
  81         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",inode->i_ino,
  82                inode->i_generation, inode, i_size_read(inode),
  83                i_size_read(inode));
  84
  85         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
  86         if (lli->lli_size_sem_owner != cfs_current()) {
  87                 EXIT;
  88                 return;
  89         }
  90
  91         if (!lli->lli_smd) {
  92                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
  93                        inode->i_ino);
  94                 GOTO(out_unlock, 0);
  95         }
  96         LASSERT_SEM_LOCKED(&lli->lli_size_sem);
  97
  98         if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_CHECKSUM) &&
  99                      (i_size_read(inode) & ~CFS_PAGE_MASK))) {
 100                 /* If the truncate leaves a partial page, update its checksum */
 101                 struct page *page = find_get_page(inode->i_mapping,
 102                                                   i_size_read(inode) >>
 103                                                   CFS_PAGE_SHIFT);
 104                 if (page != NULL) {
 105 #if 0 /* XXX */
 106                         struct ll_async_page *llap = llap_cast_private(page);
 107                         if (llap != NULL) {
 108                                 char *kaddr = kmap_atomic(page, KM_USER0);
 109                                 llap->llap_checksum =
 110                                         init_checksum(OSC_DEFAULT_CKSUM);
 111                                 llap->llap_checksum =
 112                                         compute_checksum(llap->llap_checksum,
 113                                                          kaddr, CFS_PAGE_SIZE,
 114                                                          OSC_DEFAULT_CKSUM);
 115                                 kunmap_atomic(kaddr, KM_USER0);
 116                         }
 117                         page_cache_release(page);
 118 #endif
 119                 }
 120         }
 121
 122         new_size = i_size_read(inode);
 123         ll_inode_size_unlock(inode, 0);
 124
 125         EXIT;
 126         return;
 127
 128  out_unlock:
 129         ll_inode_size_unlock(inode, 0);
 130 } /* ll_truncate */
 131
 132 /**
 133  * Initializes common cl-data at the typical address_space operation entry
 134  * point.
 135  */
 136 static int ll_cl_init(struct file *file, struct page *vmpage,
 137                       struct lu_env **env,
 138                       struct cl_io **io, struct cl_page **page, int *refcheck)
 139 {
 140         struct lu_env    *_env;
 141         struct cl_io     *_io;
 142         struct cl_page   *_page;
 143         struct cl_object *clob;
 144
 145         int result;
 146
 147         *env  = NULL;
 148         *io   = NULL;
 149         *page = NULL;
 150
 151         clob = ll_i2info(vmpage->mapping->host)->lli_clob;
 152         LASSERT(clob != NULL);
 153
 154         _env = cl_env_get(refcheck);
 155         if (!IS_ERR(env)) {
 156                 struct ccc_io *cio = ccc_env_io(_env);
 157
 158                 *env = _env;
 159                 *io  = _io = cio->cui_cl.cis_io;
 160                 if (_io != NULL) {
 161                         LASSERT(_io->ci_state == CIS_IO_GOING);
 162                         LASSERT(cio->cui_fd == LUSTRE_FPRIVATE(file));
 163                         _page = cl_page_find(_env, clob, vmpage->index, vmpage,
 164                                              CPT_CACHEABLE);
 165                         if (!IS_ERR(_page)) {
 166                                 *page = _page;
 167                                 lu_ref_add(&_page->cp_reference, "cl_io", _io);
 168                                 result = 0;
 169                         } else
 170                                 result = PTR_ERR(_page);
 171                 } else
 172                         /*
 173                          * This is for a case where operation can be called
 174                          * either with or without cl_io created by the upper
 175                          * layer (e.g., ->prepare_write() called directly from
 176                          * loop-back driver).
 177                          */
 178                         result = -EALREADY;
 179         } else
 180                 result = PTR_ERR(_env);
 181         CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %i %p %p %p\n",
 182                vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result,
 183                *env, *io, *page);
 184         return result;
 185 }
 186
 187 /**
 188  * Finalizes cl-data before exiting typical address_space operation. Dual to
 189  * ll_cl_init().
 190  */
 191 static void ll_cl_fini(struct lu_env *env,
 192                        struct cl_io *io, struct cl_page *page, int *refcheck)
 193 {
 194         if (page != NULL) {
 195                 lu_ref_del(&page->cp_reference, "cl_io", io);
 196                 cl_page_put(env, page);
 197         }
 198         if (env != NULL) {
 199                 struct vvp_io *vio;
 200
 201                 vio = vvp_env_io(env);
 202                 LASSERT(vio->cui_oneshot >= 0);
 203                 if (vio->cui_oneshot > 0) {
 204                         if (--vio->cui_oneshot == 0) {
 205                                 cl_io_end(env, io);
 206                                 cl_io_unlock(env, io);
 207                                 cl_io_iter_fini(env, io);
 208                                 cl_io_fini(env, io);
 209                                 /* to trigger assertion above, if ll_cl_fini()
 210                                  * is called against freed io. */
 211                                 vio->cui_oneshot = -1;
 212                         }
 213                         /* additional reference on env was acquired by io,
 214                          * disable refcheck */
 215                         refcheck = NULL;
 216                 }
 217                 cl_env_put(env, refcheck);
 218         } else
 219                 LASSERT(io == NULL);
 220 }
 221
 222 /**
 223  * Initializes one-shot cl_io for the case when loop driver calls
 224  * ->{prepare,commit}_write() methods directly.
 225  */
 226 static int ll_prepare_loop(struct lu_env *env, struct cl_io *io,
 227                            struct file *file, struct page *vmpage,
 228                            unsigned from, unsigned to)
 229 {
 230         struct vvp_io *vio;
 231         struct ccc_io *cio;
 232         int result;
 233         loff_t pos;
 234
 235         vio = vvp_env_io(env);
 236         cio = ccc_env_io(env);
 237         ll_io_init(io, file, 1);
 238         pos = (vmpage->index << CFS_PAGE_SHIFT) + from;
 239         /*
 240          * Create IO and quickly drive it through CIS_{INIT,IT_STARTED,LOCKED}
 241          * states. DLM locks are not taken for vio->cui_oneshot IO---we cannot
 242          * take DLM locks here, because page is already locked. With new
 243          * ->write_{being,end}() address_space operations lustre might be
 244          * luckier.
 245          */
 246         result = cl_io_rw_init(env, io, CIT_WRITE, pos, from - to);
 247         if (result == 0) {
 248                 cio->cui_fd = LUSTRE_FPRIVATE(file);
 249                 vio->cui_oneshot = 1;
 250                 result = cl_io_iter_init(env, io);
 251                 if (result == 0) {
 252                         result = cl_io_lock(env, io);
 253                         if (result == 0)
 254                                 result = cl_io_start(env, io);
 255                 }
 256         } else
 257                 result = io->ci_result;
 258         return result;
 259 }
 260
 261 /**
 262  * ->prepare_write() address space operation called by generic_file_write()
 263  * for every page during write.
 264  */
 265 int ll_prepare_write(struct file *file, struct page *vmpage, unsigned from,
 266                      unsigned to)
 267 {
 268         struct lu_env    *env;
 269         struct cl_io     *io;
 270         struct cl_page   *page;
 271         int result;
 272         int refcheck;
 273         ENTRY;
 274
 275         result = ll_cl_init(file, vmpage, &env, &io, &page, &refcheck);
 276         /*
 277          * Loop-back driver calls ->prepare_write() and ->sendfile() methods
 278          * directly, bypassing file system ->write() operation, so cl_io has
 279          * to be created here.
 280          */
 281         if (result == -EALREADY) {
 282                 io = &ccc_env_info(env)->cti_io;
 283                 result = ll_prepare_loop(env, io, file, vmpage, from, to);
 284                 if (result == 0) {
 285                         result = ll_cl_init(file, vmpage,
 286                                             &env, &io, &page, &refcheck);
 287                         cl_env_put(env, NULL);
 288                 }
 289         }
 290         if (result == 0) {
 291                 cl_page_assume(env, io, page);
 292                 if (cl_io_is_append(io)) {
 293                         struct cl_object   *obj   = io->ci_obj;
 294                         struct inode       *inode = ccc_object_inode(obj);
 295                         /**
 296                          * In VFS file->page write loop, for appending, the
 297                          * write offset might be reset according to the new
 298                          * file size before holding i_mutex. So crw_pos should
 299                          * be reset here. BUG:17711.
 300                          */
 301                         io->u.ci_wr.wr.crw_pos = i_size_read(inode);
 302                 }
 303                 result = cl_io_prepare_write(env, io, page, from, to);
 304                 if (result == 0) {
 305                         struct vvp_io *vio;
 306
 307                         /*
 308                          * Add a reference, so that page is not evicted from
 309                          * the cache until ->commit_write() is called.
 310                          */
 311                         cl_page_get(page);
 312                         lu_ref_add(&page->cp_reference, "prepare_write",
 313                                    cfs_current());
 314                         vio = vvp_env_io(env);
 315                         if (vio->cui_oneshot > 0)
 316                                 vio->cui_oneshot++;
 317                 } else
 318                         cl_page_unassume(env, io, page);
 319         }
 320         ll_cl_fini(env, io, page, &refcheck);
 321         RETURN(result);
 322 }
 323
 324 int ll_commit_write(struct file *file, struct page *vmpage, unsigned from,
 325                     unsigned to)
 326 {
 327         struct lu_env    *env;
 328         struct cl_io     *io;
 329         struct cl_page   *page;
 330         int result;
 331         int refcheck;
 332         ENTRY;
 333
 334         result = ll_cl_init(file, vmpage, &env, &io, &page, &refcheck);
 335         LASSERT(result != -EALREADY);
 336         if (result == 0) {
 337                 LASSERT(cl_page_is_owned(page, io));
 338                 result = cl_io_commit_write(env, io, page, from, to);
 339                 if (cl_page_is_owned(page, io))
 340                         cl_page_unassume(env, io, page);
 341                 /*
 342                  * Release reference acquired by cl_io_prepare_write().
 343                  */
 344                 lu_ref_del(&page->cp_reference, "prepare_write", cfs_current());
 345                 cl_page_put(env, page);
 346         }
 347         ll_cl_fini(env, io, page, &refcheck);
 348         RETURN(result);
 349 }
 350
 351 struct obd_capa *cl_capa_lookup(struct inode *inode, enum cl_req_type crt)
 352 {
 353         __u64 opc;
 354
 355         opc = crt == CRT_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
 356         return ll_osscapa_get(inode, opc);
 357 }
 358
 359 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
 360
 361 /* WARNING: This algorithm is used to reduce the contention on
 362  * sbi->ll_lock. It should work well if the ra_max_pages is much
 363  * greater than the single file's read-ahead window.
 364  *
 365  * TODO: There may exist a `global sync problem' in this implementation.
 366  * Considering the global ra window is 100M, and each file's ra window is 10M,
 367  * there are over 10 files trying to get its ra budget and reach
 368  * ll_ra_count_get at the exactly same time. All of them will get a zero ra
 369  * window, although the global window is 100M. -jay
 370  */
 371 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 372 {
 373         struct ll_ra_info *ra = &sbi->ll_ra_info;
 374         unsigned long ret;
 375         ENTRY;
 376
 377         /**
 378          * If read-ahead pages left are less than 1M, do not do read-ahead,
 379          * otherwise it will form small read RPC(< 1M), which hurt server
 380          * performance a lot.
 381          */
 382         ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages), len);
 383         if ((int)ret < 0 || ret < min((unsigned long)PTLRPC_MAX_BRW_PAGES, len))
 384                 GOTO(out, ret = 0);
 385
 386         if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
 387                 atomic_sub(ret, &ra->ra_cur_pages);
 388                 ret = 0;
 389         }
 390 out:
 391         RETURN(ret);
 392 }
 393
 394 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 395 {
 396         struct ll_ra_info *ra = &sbi->ll_ra_info;
 397         atomic_sub(len, &ra->ra_cur_pages);
 398 }
 399
 400 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
 401 {
 402         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
 403         lprocfs_counter_incr(sbi->ll_ra_stats, which);
 404 }
 405
 406 void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
 407 {
 408         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
 409         ll_ra_stats_inc_sbi(sbi, which);
 410 }
 411
 412 #define RAS_CDEBUG(ras) \
 413         CDEBUG(D_READA,                                                      \
 414                "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu"    \
 415                "csr %lu sf %lu sp %lu sl %lu \n",                            \
 416                ras->ras_last_readpage, ras->ras_consecutive_requests,        \
 417                ras->ras_consecutive_pages, ras->ras_window_start,            \
 418                ras->ras_window_len, ras->ras_next_readahead,                 \
 419                ras->ras_requests, ras->ras_request_index,                    \
 420                ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
 421                ras->ras_stride_pages, ras->ras_stride_length)
 422
 423 static int index_in_window(unsigned long index, unsigned long point,
 424                            unsigned long before, unsigned long after)
 425 {
 426         unsigned long start = point - before, end = point + after;
 427
 428         if (start > point)
 429                start = 0;
 430         if (end < point)
 431                end = ~0;
 432
 433         return start <= index && index <= end;
 434 }
 435
 436 static struct ll_readahead_state *ll_ras_get(struct file *f)
 437 {
 438         struct ll_file_data       *fd;
 439
 440         fd = LUSTRE_FPRIVATE(f);
 441         return &fd->fd_ras;
 442 }
 443
 444 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
 445 {
 446         struct ll_readahead_state *ras;
 447
 448         ras = ll_ras_get(f);
 449
 450         spin_lock(&ras->ras_lock);
 451         ras->ras_requests++;
 452         ras->ras_request_index = 0;
 453         ras->ras_consecutive_requests++;
 454         rar->lrr_reader = current;
 455
 456         list_add(&rar->lrr_linkage, &ras->ras_read_beads);
 457         spin_unlock(&ras->ras_lock);
 458 }
 459
 460 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
 461 {
 462         struct ll_readahead_state *ras;
 463
 464         ras = ll_ras_get(f);
 465
 466         spin_lock(&ras->ras_lock);
 467         list_del_init(&rar->lrr_linkage);
 468         spin_unlock(&ras->ras_lock);
 469 }
 470
 471 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
 472 {
 473         struct ll_ra_read *scan;
 474
 475         list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
 476                 if (scan->lrr_reader == current)
 477                         return scan;
 478         }
 479         return NULL;
 480 }
 481
 482 struct ll_ra_read *ll_ra_read_get(struct file *f)
 483 {
 484         struct ll_readahead_state *ras;
 485         struct ll_ra_read         *bead;
 486
 487         ras = ll_ras_get(f);
 488
 489         spin_lock(&ras->ras_lock);
 490         bead = ll_ra_read_get_locked(ras);
 491         spin_unlock(&ras->ras_lock);
 492         return bead;
 493 }
 494
 495 static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io,
 496                               struct cl_page_list *queue, struct cl_page *page,
 497                               struct page *vmpage)
 498 {
 499         struct ccc_page *cp;
 500         int              rc;
 501
 502         ENTRY;
 503
 504         rc = 0;
 505         cl_page_assume(env, io, page);
 506         lu_ref_add(&page->cp_reference, "ra", cfs_current());
 507         cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
 508         if (!cp->cpg_defer_uptodate && !Page_Uptodate(vmpage)) {
 509                 rc = cl_page_is_under_lock(env, io, page);
 510                 if (rc == -EBUSY) {
 511                         cp->cpg_defer_uptodate = 1;
 512                         cp->cpg_ra_used = 0;
 513                         cl_page_list_add(queue, page);
 514                         rc = 1;
 515                 } else {
 516                         cl_page_delete(env, page);
 517                         rc = -ENOLCK;
 518                 }
 519         } else
 520                 /* skip completed pages */
 521                 cl_page_unassume(env, io, page);
 522         lu_ref_del(&page->cp_reference, "ra", cfs_current());
 523         cl_page_put(env, page);
 524         RETURN(rc);
 525 }
 526
 527 /**
 528  * Initiates read-ahead of a page with given index.
 529  *
 530  * \retval     +ve: page was added to \a queue.
 531  *
 532  * \retval -ENOLCK: there is no extent lock for this part of a file, stop
 533  *                  read-ahead.
 534  *
 535  * \retval  -ve, 0: page wasn't added to \a queue for other reason.
 536  */
 537 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
 538                               struct cl_page_list *queue,
 539                               int index, struct address_space *mapping)
 540 {
 541         struct page      *vmpage;
 542         struct cl_object *clob  = ll_i2info(mapping->host)->lli_clob;
 543         struct cl_page   *page;
 544         enum ra_stat      which = _NR_RA_STAT; /* keep gcc happy */
 545         unsigned int      gfp_mask;
 546         int               rc    = 0;
 547         const char       *msg   = NULL;
 548
 549         ENTRY;
 550
 551         gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
 552 #ifdef __GFP_NOWARN
 553         gfp_mask |= __GFP_NOWARN;
 554 #endif
 555         vmpage = grab_cache_page_nowait_gfp(mapping, index, gfp_mask);
 556         if (vmpage != NULL) {
 557                 /* Check if vmpage was truncated or reclaimed */
 558                 if (vmpage->mapping == mapping) {
 559                         page = cl_page_find(env, clob, vmpage->index,
 560                                             vmpage, CPT_CACHEABLE);
 561                         if (!IS_ERR(page)) {
 562                                 rc = cl_read_ahead_page(env, io, queue,
 563                                                         page, vmpage);
 564                                 if (rc == -ENOLCK) {
 565                                         which = RA_STAT_FAILED_MATCH;
 566                                         msg   = "lock match failed";
 567                                 }
 568                         } else {
 569                                 which = RA_STAT_FAILED_GRAB_PAGE;
 570                                 msg   = "cl_page_find failed";
 571                         }
 572                 } else {
 573                         which = RA_STAT_WRONG_GRAB_PAGE;
 574                         msg   = "g_c_p_n returned invalid page";
 575                 }
 576                 if (rc != 1)
 577                         unlock_page(vmpage);
 578                 page_cache_release(vmpage);
 579         } else {
 580                 which = RA_STAT_FAILED_GRAB_PAGE;
 581                 msg   = "g_c_p_n failed";
 582         }
 583         if (msg != NULL) {
 584                 ll_ra_stats_inc(mapping, which);
 585                 CDEBUG(D_READA, "%s\n", msg);
 586         }
 587         RETURN(rc);
 588 }
 589
 590 #define RIA_DEBUG(ria)                                                       \
 591         CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n",       \
 592         ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
 593         ria->ria_pages)
 594
 595 #define RAS_INCREASE_STEP PTLRPC_MAX_BRW_PAGES
 596
 597 static inline int stride_io_mode(struct ll_readahead_state *ras)
 598 {
 599         return ras->ras_consecutive_stride_requests > 1;
 600 }
 601 /* The function calculates how much pages will be read in
 602  * [off, off + length], in such stride IO area,
 603  * stride_offset = st_off, stride_lengh = st_len,
 604  * stride_pages = st_pgs
 605  *
 606  *   |------------------|*****|------------------|*****|------------|*****|....
 607  * st_off
 608  *   |--- st_pgs     ---|
 609  *   |-----     st_len   -----|
 610  *
 611  *              How many pages it should read in such pattern
 612  *              |-------------------------------------------------------------|
 613  *              off
 614  *              |<------                  length                      ------->|
 615  *
 616  *          =   |<----->|  +  |-------------------------------------| +   |---|
 617  *             start_left                 st_pgs * i                    end_left
 618  */
 619 static unsigned long
 620 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
 621                 unsigned long off, unsigned long length)
 622 {
 623         unsigned long start = off > st_off ? off - st_off : 0;
 624         unsigned long end = off + length > st_off ? off + length - st_off : 0;
 625         unsigned long start_left = 0;
 626         unsigned long end_left = 0;
 627         unsigned long pg_count;
 628
 629         if (st_len == 0 || length == 0 || end == 0)
 630                 return length;
 631
 632         start_left = do_div(start, st_len);
 633         if (start_left < st_pgs)
 634                 start_left = st_pgs - start_left;
 635         else
 636                 start_left = 0;
 637
 638         end_left = do_div(end, st_len);
 639         if (end_left > st_pgs)
 640                 end_left = st_pgs;
 641
 642         CDEBUG(D_READA, "start %lu, end %lu start_left %lu end_left %lu \n",
 643                start, end, start_left, end_left);
 644
 645         if (start == end)
 646                 pg_count = end_left - (st_pgs - start_left);
 647         else
 648                 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
 649
 650         CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
 651                "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
 652
 653         return pg_count;
 654 }
 655
 656 static int ria_page_count(struct ra_io_arg *ria)
 657 {
 658         __u64 length = ria->ria_end >= ria->ria_start ?
 659                        ria->ria_end - ria->ria_start + 1 : 0;
 660
 661         return stride_pg_count(ria->ria_stoff, ria->ria_length,
 662                                ria->ria_pages, ria->ria_start,
 663                                length);
 664 }
 665
 666 /*Check whether the index is in the defined ra-window */
 667 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
 668 {
 669         /* If ria_length == ria_pages, it means non-stride I/O mode,
 670          * idx should always inside read-ahead window in this case
 671          * For stride I/O mode, just check whether the idx is inside
 672          * the ria_pages. */
 673         return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
 674                (idx - ria->ria_stoff) % ria->ria_length < ria->ria_pages;
 675 }
 676
 677 static int ll_read_ahead_pages(const struct lu_env *env,
 678                                struct cl_io *io, struct cl_page_list *queue,
 679                                struct ra_io_arg *ria,
 680                                unsigned long *reserved_pages,
 681                                struct address_space *mapping,
 682                                unsigned long *ra_end)
 683 {
 684         int rc, count = 0, stride_ria;
 685         unsigned long page_idx;
 686
 687         LASSERT(ria != NULL);
 688         RIA_DEBUG(ria);
 689
 690         stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
 691         for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
 692                         *reserved_pages > 0; page_idx++) {
 693                 if (ras_inside_ra_window(page_idx, ria)) {
 694                         /* If the page is inside the read-ahead window*/
 695                         rc = ll_read_ahead_page(env, io, queue,
 696                                                 page_idx, mapping);
 697                         if (rc == 1) {
 698                                 (*reserved_pages)--;
 699                                 count ++;
 700                         } else if (rc == -ENOLCK)
 701                                 break;
 702                 } else if (stride_ria) {
 703                         /* If it is not in the read-ahead window, and it is
 704                          * read-ahead mode, then check whether it should skip
 705                          * the stride gap */
 706                         pgoff_t offset;
 707                         /* FIXME: This assertion only is valid when it is for
 708                          * forward read-ahead, it will be fixed when backward
 709                          * read-ahead is implemented */
 710                         LASSERTF(page_idx > ria->ria_stoff, "since %lu in the"
 711                                 " gap of ra window,it should bigger than stride"
 712                                 " offset %lu \n", page_idx, ria->ria_stoff);
 713
 714                         offset = page_idx - ria->ria_stoff;
 715                         offset = offset % (ria->ria_length);
 716                         if (offset > ria->ria_pages) {
 717                                 page_idx += ria->ria_length - offset;
 718                                 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
 719                                        ria->ria_length - offset);
 720                                 continue;
 721                         }
 722                 }
 723         }
 724         *ra_end = page_idx;
 725         return count;
 726 }
 727
 728 int ll_readahead(const struct lu_env *env, struct cl_io *io,
 729                  struct ll_readahead_state *ras, struct address_space *mapping,
 730                  struct cl_page_list *queue, int flags)
 731 {
 732         struct vvp_io *vio = vvp_env_io(env);
 733         struct vvp_thread_info *vti = vvp_env_info(env);
 734         struct cl_attr *attr = ccc_env_thread_attr(env);
 735         unsigned long start = 0, end = 0, reserved;
 736         unsigned long ra_end, len;
 737         struct inode *inode;
 738         struct ll_ra_read *bead;
 739         struct ra_io_arg *ria = &vti->vti_ria;
 740         struct ll_inode_info *lli;
 741         struct cl_object *clob;
 742         int ret = 0;
 743         __u64 kms;
 744         ENTRY;
 745
 746         inode = mapping->host;
 747         lli = ll_i2info(inode);
 748         clob = lli->lli_clob;
 749
 750         memset(ria, 0, sizeof *ria);
 751
 752         cl_object_attr_lock(clob);
 753         ret = cl_object_attr_get(env, clob, attr);
 754         cl_object_attr_unlock(clob);
 755
 756         if (ret != 0)
 757                 RETURN(ret);
 758         kms = attr->cat_kms;
 759         if (kms == 0) {
 760                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
 761                 RETURN(0);
 762         }
 763
 764         spin_lock(&ras->ras_lock);
 765         if (vio->cui_ra_window_set)
 766                 bead = &vio->cui_bead;
 767         else
 768                 bead = NULL;
 769
 770         /* Enlarge the RA window to encompass the full read */
 771         if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
 772             bead->lrr_start + bead->lrr_count) {
 773                 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
 774                                       ras->ras_window_start;
 775         }
 776         /* Reserve a part of the read-ahead window that we'll be issuing */
 777         if (ras->ras_window_len) {
 778                 start = ras->ras_next_readahead;
 779                 end = ras->ras_window_start + ras->ras_window_len - 1;
 780         }
 781         if (end != 0) {
 782                 /* Truncate RA window to end of file */
 783                 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
 784                 ras->ras_next_readahead = max(end, end + 1);
 785                 RAS_CDEBUG(ras);
 786         }
 787         ria->ria_start = start;
 788         ria->ria_end = end;
 789         /* If stride I/O mode is detected, get stride window*/
 790         if (stride_io_mode(ras)) {
 791                 ria->ria_stoff = ras->ras_stride_offset;
 792                 ria->ria_length = ras->ras_stride_length;
 793                 ria->ria_pages = ras->ras_stride_pages;
 794         }
 795         spin_unlock(&ras->ras_lock);
 796
 797         if (end == 0) {
 798                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
 799                 RETURN(0);
 800         }
 801         len = ria_page_count(ria);
 802         if (len == 0)
 803                 RETURN(0);
 804
 805         reserved = ll_ra_count_get(ll_i2sbi(inode), len);
 806
 807         if (reserved < len)
 808                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
 809
 810         CDEBUG(D_READA, "reserved page %lu \n", reserved);
 811
 812         ret = ll_read_ahead_pages(env, io, queue,
 813                                   ria, &reserved, mapping, &ra_end);
 814
 815         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
 816         if (reserved != 0)
 817                 ll_ra_count_put(ll_i2sbi(inode), reserved);
 818
 819         if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
 820                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
 821
 822         /* if we didn't get to the end of the region we reserved from
 823          * the ras we need to go back and update the ras so that the
 824          * next read-ahead tries from where we left off.  we only do so
 825          * if the region we failed to issue read-ahead on is still ahead
 826          * of the app and behind the next index to start read-ahead from */
 827         CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
 828                ra_end, end, ria->ria_end);
 829
 830         if (ra_end != end + 1) {
 831                 spin_lock(&ras->ras_lock);
 832                 if (ra_end < ras->ras_next_readahead &&
 833                     index_in_window(ra_end, ras->ras_window_start, 0,
 834                                     ras->ras_window_len)) {
 835                         ras->ras_next_readahead = ra_end;
 836                                RAS_CDEBUG(ras);
 837                 }
 838                 spin_unlock(&ras->ras_lock);
 839         }
 840
 841         RETURN(ret);
 842 }
 843
 844 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
 845 {
 846         ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1));
 847 }
 848
 849 /* called with the ras_lock held or from places where it doesn't matter */
 850 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
 851 {
 852         ras->ras_last_readpage = index;
 853         ras->ras_consecutive_requests = 0;
 854         ras->ras_consecutive_pages = 0;
 855         ras->ras_window_len = 0;
 856         ras_set_start(ras, index);
 857         ras->ras_next_readahead = max(ras->ras_window_start, index);
 858
 859         RAS_CDEBUG(ras);
 860 }
 861
 862 /* called with the ras_lock held or from places where it doesn't matter */
 863 static void ras_stride_reset(struct ll_readahead_state *ras)
 864 {
 865         ras->ras_consecutive_stride_requests = 0;
 866         ras->ras_stride_length = 0;
 867         ras->ras_stride_pages = 0;
 868         RAS_CDEBUG(ras);
 869 }
 870
 871 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
 872 {
 873         spin_lock_init(&ras->ras_lock);
 874         ras_reset(ras, 0);
 875         ras->ras_requests = 0;
 876         INIT_LIST_HEAD(&ras->ras_read_beads);
 877 }
 878
 879 /*
 880  * Check whether the read request is in the stride window.
 881  * If it is in the stride window, return 1, otherwise return 0.
 882  */
 883 static int index_in_stride_window(unsigned long index,
 884                                   struct ll_readahead_state *ras,
 885                                   struct inode *inode)
 886 {
 887         unsigned long stride_gap = index - ras->ras_last_readpage - 1;
 888
 889         if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
 890             ras->ras_stride_pages == ras->ras_stride_length)
 891                 return 0;
 892
 893         /* If it is contiguous read */
 894         if (stride_gap == 0)
 895                 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
 896
 897         /*Otherwise check the stride by itself */
 898         return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
 899              ras->ras_consecutive_pages == ras->ras_stride_pages;
 900 }
 901
 902 static void ras_update_stride_detector(struct ll_readahead_state *ras,
 903                                        unsigned long index)
 904 {
 905         unsigned long stride_gap = index - ras->ras_last_readpage - 1;
 906
 907         if (!stride_io_mode(ras) && (stride_gap != 0 ||
 908              ras->ras_consecutive_stride_requests == 0)) {
 909                 ras->ras_stride_pages = ras->ras_consecutive_pages;
 910                 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
 911         }
 912         LASSERT(ras->ras_request_index == 0);
 913         LASSERT(ras->ras_consecutive_stride_requests == 0);
 914
 915         if (index <= ras->ras_last_readpage) {
 916                 /*Reset stride window for forward read*/
 917                 ras_stride_reset(ras);
 918                 return;
 919         }
 920
 921         ras->ras_stride_pages = ras->ras_consecutive_pages;
 922         ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
 923
 924         RAS_CDEBUG(ras);
 925         return;
 926 }
 927
 928 static unsigned long
 929 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
 930 {
 931         return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
 932                                ras->ras_stride_pages, ras->ras_stride_offset,
 933                                len);
 934 }
 935
 936 /* Stride Read-ahead window will be increased inc_len according to
 937  * stride I/O pattern */
 938 static void ras_stride_increase_window(struct ll_readahead_state *ras,
 939                                        struct ll_ra_info *ra,
 940                                        unsigned long inc_len)
 941 {
 942         unsigned long left, step, window_len;
 943         unsigned long stride_len;
 944
 945         LASSERT(ras->ras_stride_length > 0);
 946         LASSERTF(ras->ras_window_start + ras->ras_window_len
 947                  >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
 948                  " stride_offset %lu\n", ras->ras_window_start,
 949                  ras->ras_window_len, ras->ras_stride_offset);
 950
 951         stride_len = ras->ras_window_start + ras->ras_window_len -
 952                      ras->ras_stride_offset;
 953
 954         left = stride_len % ras->ras_stride_length;
 955         window_len = ras->ras_window_len - left;
 956
 957         if (left < ras->ras_stride_pages)
 958                 left += inc_len;
 959         else
 960                 left = ras->ras_stride_pages + inc_len;
 961
 962         LASSERT(ras->ras_stride_pages != 0);
 963
 964         step = left / ras->ras_stride_pages;
 965         left %= ras->ras_stride_pages;
 966
 967         window_len += step * ras->ras_stride_length + left;
 968
 969         if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
 970                 ras->ras_window_len = window_len;
 971
 972         RAS_CDEBUG(ras);
 973 }
 974
 975 static void ras_increase_window(struct ll_readahead_state *ras,
 976                                 struct ll_ra_info *ra, struct inode *inode)
 977 {
 978         /* The stretch of ra-window should be aligned with max rpc_size
 979          * but current clio architecture does not support retrieve such
 980          * information from lower layer. FIXME later
 981          */
 982         if (stride_io_mode(ras))
 983                 ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP);
 984         else
 985                 ras->ras_window_len = min(ras->ras_window_len +
 986                                           RAS_INCREASE_STEP,
 987                                           ra->ra_max_pages_per_file);
 988 }
 989
 990 void ras_update(struct ll_sb_info *sbi, struct inode *inode,
 991                 struct ll_readahead_state *ras, unsigned long index,
 992                 unsigned hit)
 993 {
 994         struct ll_ra_info *ra = &sbi->ll_ra_info;
 995         int zero = 0, stride_detect = 0, ra_miss = 0;
 996         ENTRY;
 997
 998         spin_lock(&sbi->ll_lock);
 999         spin_lock(&ras->ras_lock);
1000
1001         ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1002
1003         /* reset the read-ahead window in two cases.  First when the app seeks
1004          * or reads to some other part of the file.  Secondly if we get a
1005          * read-ahead miss that we think we've previously issued.  This can
1006          * be a symptom of there being so many read-ahead pages that the VM is
1007          * reclaiming it before we get to it. */
1008         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1009                 zero = 1;
1010                 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1011         } else if (!hit && ras->ras_window_len &&
1012                    index < ras->ras_next_readahead &&
1013                    index_in_window(index, ras->ras_window_start, 0,
1014                                    ras->ras_window_len)) {
1015                 ra_miss = 1;
1016                 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1017         }
1018
1019         /* On the second access to a file smaller than the tunable
1020          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1021          * file up to ra_max_pages_per_file.  This is simply a best effort
1022          * and only occurs once per open file.  Normal RA behavior is reverted
1023          * to for subsequent IO.  The mmap case does not increment
1024          * ras_requests and thus can never trigger this behavior. */
1025         if (ras->ras_requests == 2 && !ras->ras_request_index) {
1026                 __u64 kms_pages;
1027
1028                 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1029                             CFS_PAGE_SHIFT;
1030
1031                 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1032                        ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file);
1033
1034                 if (kms_pages &&
1035                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1036                         ras->ras_window_start = 0;
1037                         ras->ras_last_readpage = 0;
1038                         ras->ras_next_readahead = 0;
1039                         ras->ras_window_len = min(ra->ra_max_pages_per_file,
1040                                 ra->ra_max_read_ahead_whole_pages);
1041                         GOTO(out_unlock, 0);
1042                 }
1043         }
1044         if (zero) {
1045                 /* check whether it is in stride I/O mode*/
1046                 if (!index_in_stride_window(index, ras, inode)) {
1047                         if (ras->ras_consecutive_stride_requests == 0 &&
1048                             ras->ras_request_index == 0) {
1049                                 ras_update_stride_detector(ras, index);
1050                                 ras->ras_consecutive_stride_requests ++;
1051                         } else {
1052                                 ras_stride_reset(ras);
1053                         }
1054                         ras_reset(ras, index);
1055                         ras->ras_consecutive_pages++;
1056                         GOTO(out_unlock, 0);
1057                 } else {
1058                         ras->ras_consecutive_pages = 0;
1059                         ras->ras_consecutive_requests = 0;
1060                         if (++ras->ras_consecutive_stride_requests > 1)
1061                                 stride_detect = 1;
1062                         RAS_CDEBUG(ras);
1063                 }
1064         } else {
1065                 if (ra_miss) {
1066                         if (index_in_stride_window(index, ras, inode) &&
1067                             stride_io_mode(ras)) {
1068                                 /*If stride-RA hit cache miss, the stride dector
1069                                  *will not be reset to avoid the overhead of
1070                                  *redetecting read-ahead mode */
1071                                 if (index != ras->ras_last_readpage + 1)
1072                                        ras->ras_consecutive_pages = 0;
1073                                 ras_reset(ras, index);
1074                                 RAS_CDEBUG(ras);
1075                         } else {
1076                                 /* Reset both stride window and normal RA
1077                                  * window */
1078                                 ras_reset(ras, index);
1079                                 ras->ras_consecutive_pages++;
1080                                 ras_stride_reset(ras);
1081                                 GOTO(out_unlock, 0);
1082                         }
1083                 } else if (stride_io_mode(ras)) {
1084                         /* If this is contiguous read but in stride I/O mode
1085                          * currently, check whether stride step still is valid,
1086                          * if invalid, it will reset the stride ra window*/
1087                         if (!index_in_stride_window(index, ras, inode)) {
1088                                 /* Shrink stride read-ahead window to be zero */
1089                                 ras_stride_reset(ras);
1090                                 ras->ras_window_len = 0;
1091                                 ras->ras_next_readahead = index;
1092                         }
1093                 }
1094         }
1095         ras->ras_consecutive_pages++;
1096         ras->ras_last_readpage = index;
1097         ras_set_start(ras, index);
1098         ras->ras_next_readahead = max(ras->ras_window_start,
1099                                       ras->ras_next_readahead);
1100         RAS_CDEBUG(ras);
1101
1102         /* Trigger RA in the mmap case where ras_consecutive_requests
1103          * is not incremented and thus can't be used to trigger RA */
1104         if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1105                 ras->ras_window_len = RAS_INCREASE_STEP;
1106                 GOTO(out_unlock, 0);
1107         }
1108
1109         /* Initially reset the stride window offset to next_readahead*/
1110         if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
1111                 /**
1112                  * Once stride IO mode is detected, next_readahead should be
1113                  * reset to make sure next_readahead > stride offset
1114                  */
1115                 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
1116                 ras->ras_stride_offset = index;
1117                 ras->ras_window_len = RAS_INCREASE_STEP;
1118         }
1119
1120         /* The initial ras_window_len is set to the request size.  To avoid
1121          * uselessly reading and discarding pages for random IO the window is
1122          * only increased once per consecutive request received. */
1123         if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
1124             !ras->ras_request_index)
1125                 ras_increase_window(ras, ra, inode);
1126         EXIT;
1127 out_unlock:
1128         RAS_CDEBUG(ras);
1129         ras->ras_request_index++;
1130         spin_unlock(&ras->ras_lock);
1131         spin_unlock(&sbi->ll_lock);
1132         return;
1133 }
1134
1135 int ll_writepage(struct page *vmpage, struct writeback_control *unused)
1136 {
1137         struct inode           *inode = vmpage->mapping->host;
1138         struct lu_env          *env;
1139         struct cl_io           *io;
1140         struct cl_page         *page;
1141         struct cl_object       *clob;
1142         struct cl_2queue       *queue;
1143         struct cl_env_nest      nest;
1144         int result;
1145         ENTRY;
1146
1147         LASSERT(PageLocked(vmpage));
1148         LASSERT(!PageWriteback(vmpage));
1149
1150         if (ll_i2dtexp(inode) == NULL)
1151                 RETURN(-EINVAL);
1152
1153         env = cl_env_nested_get(&nest);
1154         if (IS_ERR(env))
1155                 RETURN(PTR_ERR(env));
1156
1157         io    = &ccc_env_info(env)->cti_io;
1158         queue = &vvp_env_info(env)->vti_queue;
1159         clob  = ll_i2info(inode)->lli_clob;
1160         LASSERT(clob != NULL);
1161
1162         io->ci_obj = clob;
1163         result = cl_io_init(env, io, CIT_MISC, clob);
1164         if (result == 0) {
1165                 page = cl_page_find(env, clob, vmpage->index,
1166                                     vmpage, CPT_CACHEABLE);
1167                 if (!IS_ERR(page)) {
1168                         lu_ref_add(&page->cp_reference, "writepage",
1169                                    cfs_current());
1170                         cl_page_assume(env, io, page);
1171                         /*
1172                          * Mark page dirty, because this is what
1173                          * ->vio_submit()->cpo_prep_write() assumes.
1174                          *
1175                          * XXX better solution is to detect this from within
1176                          * cl_io_submit_rw() somehow.
1177                          */
1178                         set_page_dirty(vmpage);
1179                         cl_2queue_init_page(queue, page);
1180                         result = cl_io_submit_rw(env, io, CRT_WRITE,
1181                                                  queue, CRP_NORMAL);
1182                         cl_page_list_disown(env, io, &queue->c2_qin);
1183                         if (result != 0) {
1184                                 /*
1185                                  * There is no need to clear PG_writeback, as
1186                                  * cl_io_submit_rw() calls completion callback
1187                                  * on failure.
1188                                  */
1189                                 /*
1190                                  * Re-dirty page on error so it retries write,
1191                                  * but not in case when IO has actually
1192                                  * occurred and completed with an error.
1193                                  */
1194                                 if (!PageError(vmpage))
1195                                         set_page_dirty(vmpage);
1196                         }
1197                         LASSERT(!cl_page_is_owned(page, io));
1198                         lu_ref_del(&page->cp_reference,
1199                                    "writepage", cfs_current());
1200                         cl_page_put(env, page);
1201                         cl_2queue_fini(env, queue);
1202                 }
1203         }
1204         cl_io_fini(env, io);
1205         cl_env_nested_put(&nest, env);
1206         RETURN(result);
1207 }
1208
1209 int ll_readpage(struct file *file, struct page *vmpage)
1210 {
1211         struct lu_env    *env;
1212         struct cl_io     *io;
1213         struct cl_page   *page;
1214         int result;
1215         int refcheck;
1216         ENTRY;
1217
1218         result = ll_cl_init(file, vmpage, &env, &io, &page, &refcheck);
1219         if (result == 0) {
1220                 LASSERT(page->cp_type == CPT_CACHEABLE);
1221                 if (likely(!PageUptodate(vmpage))) {
1222                         cl_page_assume(env, io, page);
1223                         result = cl_io_read_page(env, io, page);
1224                 } else {
1225                         /* Page from a non-object file. */
1226                         LASSERT(!ll_i2info(vmpage->mapping->host)->lli_smd);
1227                         unlock_page(vmpage);
1228                         result = 0;
1229                 }
1230         }
1231         LASSERT(!cl_page_is_owned(page, io));
1232         ll_cl_fini(env, io, page, &refcheck);
1233         RETURN(result);
1234 }
1235