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 #ifndef AUTOCONF_INCLUDED
  42 #include <linux/config.h>
  43 #endif
  44 #include <linux/kernel.h>
  45 #include <linux/mm.h>
  46 #include <linux/string.h>
  47 #include <linux/stat.h>
  48 #include <linux/errno.h>
  49 #include <linux/smp_lock.h>
  50 #include <linux/unistd.h>
  51 #include <linux/version.h>
  52 #include <asm/system.h>
  53 #include <asm/uaccess.h>
  54
  55 #include <linux/fs.h>
  56 #include <linux/stat.h>
  57 #include <asm/uaccess.h>
  58 #include <linux/mm.h>
  59 #include <linux/pagemap.h>
  60 #include <linux/smp_lock.h>
  61
  62 #define DEBUG_SUBSYSTEM S_LLITE
  63
  64 #include <linux/page-flags.h>
  65
  66 #include <lustre_lite.h>
  67 #include "llite_internal.h"
  68 #include <linux/lustre_compat25.h>
  69
  70 #ifndef list_for_each_prev_safe
  71 #define list_for_each_prev_safe(pos, n, head) \
  72         for (pos = (head)->prev, n = pos->prev; pos != (head); \
  73                 pos = n, n = pos->prev )
  74 #endif
  75
  76 cfs_mem_cache_t *ll_async_page_slab = NULL;
  77 size_t ll_async_page_slab_size = 0;
  78
  79 /* SYNCHRONOUS I/O to object storage for an inode */
  80 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
  81                   struct page *page, int flags)
  82 {
  83         struct ll_inode_info *lli = ll_i2info(inode);
  84         struct lov_stripe_md *lsm = lli->lli_smd;
  85         struct obd_info oinfo = { { { 0 } } };
  86         struct brw_page pg;
  87         int rc;
  88         ENTRY;
  89
  90         pg.pg = page;
  91         pg.off = ((obd_off)page->index) << CFS_PAGE_SHIFT;
  92
  93         if ((cmd & OBD_BRW_WRITE) && (pg.off+CFS_PAGE_SIZE>i_size_read(inode)))
  94                 pg.count = i_size_read(inode) % CFS_PAGE_SIZE;
  95         else
  96                 pg.count = CFS_PAGE_SIZE;
  97
  98         LL_CDEBUG_PAGE(D_PAGE, page, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
  99                        cmd & OBD_BRW_WRITE ? "write" : "read", pg.count,
 100                        inode->i_ino, pg.off, pg.off);
 101         if (pg.count == 0) {
 102                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
 103                        LPU64"\n", inode->i_ino, inode, i_size_read(inode),
 104                        page->mapping->host, i_size_read(page->mapping->host),
 105                        page->index, pg.off);
 106         }
 107
 108         pg.flag = flags;
 109
 110         if (cmd & OBD_BRW_WRITE)
 111                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_WRITE,
 112                                    pg.count);
 113         else
 114                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_READ,
 115                            pg.count);
 116         oinfo.oi_oa = oa;
 117         oinfo.oi_md = lsm;
 118         rc = obd_brw(cmd, ll_i2obdexp(inode), &oinfo, 1, &pg, NULL);
 119         if (rc == 0)
 120                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
 121         else if (rc != -EIO)
 122                 CERROR("error from obd_brw: rc = %d\n", rc);
 123         RETURN(rc);
 124 }
 125
 126 int ll_file_punch(struct inode * inode, loff_t new_size, int srvlock)
 127 {
 128         struct ll_inode_info *lli = ll_i2info(inode);
 129         struct obd_info oinfo = { { { 0 } } };
 130         struct obdo oa = { 0 };
 131         obd_valid valid;
 132         int rc;
 133
 134         ENTRY;
 135         CDEBUG(D_INFO, "calling punch for "LPX64" (new size %Lu=%#Lx)\n",
 136                lli->lli_smd->lsm_object_id, new_size, new_size);
 137
 138         oinfo.oi_md = lli->lli_smd;
 139         oinfo.oi_policy.l_extent.start = new_size;
 140         oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
 141         oinfo.oi_oa = &oa;
 142         oa.o_id = lli->lli_smd->lsm_object_id;
 143         oa.o_gr = lli->lli_smd->lsm_object_gr;
 144         oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 145
 146         valid = OBD_MD_FLTYPE | OBD_MD_FLMODE |OBD_MD_FLFID |
 147                 OBD_MD_FLATIME | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGENER |
 148                 OBD_MD_FLBLOCKS;
 149         if (srvlock) {
 150                 /* set OBD_MD_FLFLAGS in o_valid, only if we
 151                  * set OBD_FL_TRUNCLOCK, otherwise ost_punch
 152                  * and filter_setattr get confused, see the comment
 153                  * in ost_punch */
 154                 oa.o_flags = OBD_FL_TRUNCLOCK;
 155                 oa.o_valid |= OBD_MD_FLFLAGS;
 156
 157                 /* lockless truncate
 158                  *
 159                  * 1. do not use inode's timestamps because concurrent
 160                  * stat might fill the inode with out-of-date times,
 161                  * send current instead
 162                  *
 163                  * 2.do no update lsm, as long as stat (via
 164                  * ll_glimpse_size) will bring attributes from osts
 165                  * anyway */
 166                 oa.o_mtime = oa.o_ctime = LTIME_S(CURRENT_TIME);
 167                 oa.o_valid |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 168         } else {
 169                 struct ost_lvb *xtimes;
 170                 /* truncate under locks
 171                  *
 172                  * 1. update inode's mtime and ctime as long as
 173                  * concurrent stat (via ll_glimpse_size) might bring
 174                  * out-of-date ones
 175                  *
 176                  * 2. update lsm so that next stat (via
 177                  * ll_glimpse_size) could get correct values in lsm */
 178                 OBD_ALLOC_PTR(xtimes);
 179                 if (NULL == xtimes)
 180                         RETURN(-ENOMEM);
 181
 182                 lov_stripe_lock(lli->lli_smd);
 183                 LTIME_S(inode->i_mtime) = LTIME_S(CURRENT_TIME);
 184                 LTIME_S(inode->i_ctime) = LTIME_S(CURRENT_TIME);
 185                 xtimes->lvb_mtime = LTIME_S(inode->i_mtime);
 186                 xtimes->lvb_ctime = LTIME_S(inode->i_ctime);
 187                 obd_update_lvb(ll_i2obdexp(inode), lli->lli_smd, xtimes,
 188                                OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 189                 lov_stripe_unlock(lli->lli_smd);
 190                 OBD_FREE_PTR(xtimes);
 191
 192                 valid |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 193         }
 194         obdo_from_inode(&oa, inode, valid);
 195
 196         rc = obd_punch_rqset(ll_i2obdexp(inode), &oinfo, NULL);
 197         if (rc) {
 198                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 199                 RETURN(rc);
 200         }
 201         obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 202                       OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 203         RETURN(0);
 204 }
 205 /* this isn't where truncate starts.   roughly:
 206  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
 207  * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
 208  * avoid races.
 209  *
 210  * must be called under ->lli_size_sem */
 211 void ll_truncate(struct inode *inode)
 212 {
 213         struct ll_inode_info *lli = ll_i2info(inode);
 214         int srvlock = test_bit(LLI_F_SRVLOCK, &lli->lli_flags);
 215         loff_t new_size;
 216         ENTRY;
 217         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",
 218                inode->i_ino, inode->i_generation, inode, i_size_read(inode),
 219                i_size_read(inode));
 220
 221         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
 222         if (lli->lli_size_sem_owner != current) {
 223                 EXIT;
 224                 return;
 225         }
 226
 227         if (!lli->lli_smd) {
 228                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 229                        inode->i_ino);
 230                 GOTO(out_unlock, 0);
 231         }
 232
 233         LASSERT(SEM_COUNT(&lli->lli_size_sem) <= 0);
 234
 235         if (!srvlock) {
 236                 struct ost_lvb lvb;
 237                 int rc;
 238
 239                 /* XXX I'm pretty sure this is a hack to paper over a more
 240                  * fundamental race condition. */
 241                 lov_stripe_lock(lli->lli_smd);
 242                 inode_init_lvb(inode, &lvb);
 243                 rc = obd_merge_lvb(ll_i2obdexp(inode), lli->lli_smd, &lvb, 0);
 244                 inode->i_blocks = lvb.lvb_blocks;
 245                 if (lvb.lvb_size == i_size_read(inode) && rc == 0) {
 246                         CDEBUG(D_VFSTRACE, "skipping punch for obj "LPX64
 247                                ", %Lu=%#Lx\n",
 248                                lli->lli_smd->lsm_object_id, i_size_read(inode),
 249                                i_size_read(inode));
 250                         lov_stripe_unlock(lli->lli_smd);
 251                         GOTO(out_unlock, 0);
 252                 }
 253
 254                 obd_adjust_kms(ll_i2obdexp(inode), lli->lli_smd,
 255                                i_size_read(inode), 1);
 256                 lov_stripe_unlock(lli->lli_smd);
 257         }
 258
 259         if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_LLITE_CHECKSUM) &&
 260                      (i_size_read(inode) & ~CFS_PAGE_MASK))) {
 261                 /* If the truncate leaves a partial page, update its checksum */
 262                 struct page *page = find_get_page(inode->i_mapping,
 263                                                   i_size_read(inode) >>
 264                                                   CFS_PAGE_SHIFT);
 265                 if (page != NULL) {
 266                         struct ll_async_page *llap = llap_cast_private(page);
 267                         if (llap != NULL) {
 268                                 char *kaddr = kmap_atomic(page, KM_USER0);
 269                                 llap->llap_checksum =
 270                                         init_checksum(OSC_DEFAULT_CKSUM);
 271                                 llap->llap_checksum =
 272                                         compute_checksum(llap->llap_checksum,
 273                                                          kaddr, CFS_PAGE_SIZE,
 274                                                          OSC_DEFAULT_CKSUM);
 275                                 kunmap_atomic(kaddr, KM_USER0);
 276                         }
 277                         page_cache_release(page);
 278                 }
 279         }
 280
 281         new_size = i_size_read(inode);
 282         ll_inode_size_unlock(inode, 0);
 283         if (!srvlock)
 284                 ll_file_punch(inode, new_size, 0);
 285         else
 286                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_LOCKLESS_TRUNC, 1);
 287
 288         EXIT;
 289         return;
 290
 291  out_unlock:
 292         ll_inode_size_unlock(inode, 0);
 293 } /* ll_truncate */
 294
 295 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 296                      unsigned to)
 297 {
 298         struct inode *inode = page->mapping->host;
 299         struct ll_inode_info *lli = ll_i2info(inode);
 300         struct lov_stripe_md *lsm = lli->lli_smd;
 301         obd_off offset = ((obd_off)page->index) << CFS_PAGE_SHIFT;
 302         struct obd_info oinfo = { { { 0 } } };
 303         struct brw_page pga;
 304         struct obdo oa = { 0 };
 305         struct ost_lvb lvb;
 306         int rc = 0;
 307         ENTRY;
 308
 309         LASSERT(PageLocked(page));
 310         (void)llap_cast_private(page); /* assertion */
 311
 312         /* Check to see if we should return -EIO right away */
 313         pga.pg = page;
 314         pga.off = offset;
 315         pga.count = CFS_PAGE_SIZE;
 316         pga.flag = 0;
 317
 318         oa.o_mode = inode->i_mode;
 319         oa.o_id = lsm->lsm_object_id;
 320         oa.o_gr = lsm->lsm_object_gr;
 321         oa.o_valid = OBD_MD_FLID   | OBD_MD_FLMODE |
 322                      OBD_MD_FLTYPE | OBD_MD_FLGROUP;
 323         obdo_from_inode(&oa, inode, OBD_MD_FLFID | OBD_MD_FLGENER);
 324
 325         oinfo.oi_oa = &oa;
 326         oinfo.oi_md = lsm;
 327         rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), &oinfo, 1, &pga, NULL);
 328         if (rc)
 329                 RETURN(rc);
 330
 331         if (PageUptodate(page)) {
 332                 LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n");
 333                 RETURN(0);
 334         }
 335
 336         /* We're completely overwriting an existing page, so _don't_ set it up
 337          * to date until commit_write */
 338         if (from == 0 && to == CFS_PAGE_SIZE) {
 339                 LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n");
 340                 POISON_PAGE(page, 0x11);
 341                 RETURN(0);
 342         }
 343
 344         /* If are writing to a new page, no need to read old data.  The extent
 345          * locking will have updated the KMS, and for our purposes here we can
 346          * treat it like i_size. */
 347         lov_stripe_lock(lsm);
 348         inode_init_lvb(inode, &lvb);
 349         obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
 350         lov_stripe_unlock(lsm);
 351         if (lvb.lvb_size <= offset) {
 352                 char *kaddr = kmap_atomic(page, KM_USER0);
 353                 LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n",
 354                                lvb.lvb_size, offset);
 355                 memset(kaddr, 0, CFS_PAGE_SIZE);
 356                 kunmap_atomic(kaddr, KM_USER0);
 357                 GOTO(prepare_done, rc = 0);
 358         }
 359
 360         /* XXX could be an async ocp read.. read-ahead? */
 361         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 362         if (rc == 0) {
 363                 /* bug 1598: don't clobber blksize */
 364                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 365                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 366         }
 367
 368         EXIT;
 369  prepare_done:
 370         if (rc == 0)
 371                 SetPageUptodate(page);
 372
 373         return rc;
 374 }
 375
 376 /**
 377  * make page ready for ASYNC write
 378  * \param data - pointer to llap cookie
 379  * \param cmd - is OBD_BRW_* macroses
 380  *
 381  * \retval 0 is page successfully prepared to send
 382  * \retval -EAGAIN is page not need to send
 383  */
 384 static int ll_ap_make_ready(void *data, int cmd)
 385 {
 386         struct ll_async_page *llap;
 387         struct page *page;
 388         ENTRY;
 389
 390         llap = LLAP_FROM_COOKIE(data);
 391         page = llap->llap_page;
 392
 393         /* we're trying to write, but the page is locked.. come back later */
 394         if (TryLockPage(page))
 395                 RETURN(-EAGAIN);
 396
 397         LASSERTF(!(cmd & OBD_BRW_READ) || !PageWriteback(page),
 398                 "cmd %x page %p ino %lu index %lu fl %lx\n", cmd, page,
 399                  page->mapping->host->i_ino, page->index, (long)page->flags);
 400
 401         /* if we left PageDirty we might get another writepage call
 402          * in the future.  list walkers are bright enough
 403          * to check page dirty so we can leave it on whatever list
 404          * its on.  XXX also, we're called with the cli list so if
 405          * we got the page cache list we'd create a lock inversion
 406          * with the removepage path which gets the page lock then the
 407          * cli lock */
 408         if(!clear_page_dirty_for_io(page)) {
 409                 unlock_page(page);
 410                 RETURN(-EAGAIN);
 411         }
 412
 413         /* This actually clears the dirty bit in the radix tree.*/
 414         set_page_writeback(page);
 415
 416         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 417         page_cache_get(page);
 418
 419         RETURN(0);
 420 }
 421
 422 /* We have two reasons for giving llite the opportunity to change the
 423  * write length of a given queued page as it builds the RPC containing
 424  * the page:
 425  *
 426  * 1) Further extending writes may have landed in the page cache
 427  *    since a partial write first queued this page requiring us
 428  *    to write more from the page cache.  (No further races are possible,
 429  *    since by the time this is called, the page is locked.)
 430  * 2) We might have raced with truncate and want to avoid performing
 431  *    write RPCs that are just going to be thrown away by the
 432  *    truncate's punch on the storage targets.
 433  *
 434  * The kms serves these purposes as it is set at both truncate and extending
 435  * writes.
 436  */
 437 static int ll_ap_refresh_count(void *data, int cmd)
 438 {
 439         struct ll_inode_info *lli;
 440         struct ll_async_page *llap;
 441         struct lov_stripe_md *lsm;
 442         struct page *page;
 443         struct inode *inode;
 444         struct ost_lvb lvb;
 445         __u64 kms;
 446         ENTRY;
 447
 448         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 449         LASSERT(cmd != OBD_BRW_READ);
 450
 451         llap = LLAP_FROM_COOKIE(data);
 452         page = llap->llap_page;
 453         inode = page->mapping->host;
 454         lli = ll_i2info(inode);
 455         lsm = lli->lli_smd;
 456
 457         lov_stripe_lock(lsm);
 458         inode_init_lvb(inode, &lvb);
 459         obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
 460         kms = lvb.lvb_size;
 461         lov_stripe_unlock(lsm);
 462
 463         /* catch race with truncate */
 464         if (((__u64)page->index << CFS_PAGE_SHIFT) >= kms)
 465                 return 0;
 466
 467         /* catch sub-page write at end of file */
 468         if (((__u64)page->index << CFS_PAGE_SHIFT) + CFS_PAGE_SIZE > kms)
 469                 return kms % CFS_PAGE_SIZE;
 470
 471         return CFS_PAGE_SIZE;
 472 }
 473
 474 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 475 {
 476         struct lov_stripe_md *lsm;
 477         obd_flag valid_flags;
 478
 479         lsm = ll_i2info(inode)->lli_smd;
 480
 481         oa->o_id = lsm->lsm_object_id;
 482         oa->o_gr = lsm->lsm_object_gr;
 483         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 484         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 485         if (cmd & OBD_BRW_WRITE) {
 486                 oa->o_valid |= OBD_MD_FLEPOCH;
 487                 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
 488
 489                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 490                         OBD_MD_FLUID | OBD_MD_FLGID |
 491                         OBD_MD_FLFID | OBD_MD_FLGENER;
 492         }
 493
 494         obdo_from_inode(oa, inode, valid_flags);
 495 }
 496
 497 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 498 {
 499         struct ll_async_page *llap;
 500         ENTRY;
 501
 502         llap = LLAP_FROM_COOKIE(data);
 503         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 504
 505         EXIT;
 506 }
 507
 508 static void ll_ap_update_obdo(void *data, int cmd, struct obdo *oa,
 509                               obd_valid valid)
 510 {
 511         struct ll_async_page *llap;
 512         ENTRY;
 513
 514         llap = LLAP_FROM_COOKIE(data);
 515         obdo_from_inode(oa, llap->llap_page->mapping->host, valid);
 516
 517         EXIT;
 518 }
 519
 520 static struct obd_async_page_ops ll_async_page_ops = {
 521         .ap_make_ready =        ll_ap_make_ready,
 522         .ap_refresh_count =     ll_ap_refresh_count,
 523         .ap_fill_obdo =         ll_ap_fill_obdo,
 524         .ap_update_obdo =       ll_ap_update_obdo,
 525         .ap_completion =        ll_ap_completion,
 526 };
 527
 528 struct ll_async_page *llap_cast_private(struct page *page)
 529 {
 530         struct ll_async_page *llap = (struct ll_async_page *)page_private(page);
 531
 532         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 533                  "page %p private %lu gave magic %d which != %d\n",
 534                  page, page_private(page), llap->llap_magic, LLAP_MAGIC);
 535
 536         return llap;
 537 }
 538
 539 /* Try to reap @target pages in the specific @cpu's async page list.
 540  *
 541  * There is an llap attached onto every page in lustre, linked off @sbi.
 542  * We add an llap to the list so we don't lose our place during list walking.
 543  * If llaps in the list are being moved they will only move to the end
 544  * of the LRU, and we aren't terribly interested in those pages here (we
 545  * start at the beginning of the list where the least-used llaps are. */
 546 static inline int llap_shrink_cache_internal(struct ll_sb_info *sbi,
 547         int cpu, int target)
 548 {
 549         struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a };
 550         struct ll_pglist_data *pd;
 551         struct list_head *head;
 552         int count = 0;
 553
 554         pd = ll_pglist_cpu_lock(sbi, cpu);
 555         head = &pd->llpd_list;
 556         list_add(&dummy_llap.llap_pglist_item, head);
 557         while (count < target) {
 558                 struct page *page;
 559                 int keep;
 560
 561                 if (unlikely(need_resched())) {
 562                         list_del(&dummy_llap.llap_pglist_item);
 563                         ll_pglist_cpu_unlock(sbi, cpu);
 564                         /* vmscan::shrink_slab() have own schedule() */
 565                         return count;
 566                 }
 567
 568                 llap = llite_pglist_next_llap(head,
 569                         &dummy_llap.llap_pglist_item);
 570                 list_del_init(&dummy_llap.llap_pglist_item);
 571                 if (llap == NULL)
 572                         break;
 573
 574                 page = llap->llap_page;
 575                 LASSERT(page != NULL);
 576
 577                 list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item);
 578
 579                 /* Page needs/undergoing IO */
 580                 if (TryLockPage(page)) {
 581                         LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n");
 582                         continue;
 583                 }
 584
 585                keep = (llap->llap_write_queued || PageDirty(page) ||
 586                       PageWriteback(page) || (!PageUptodate(page) &&
 587                       llap->llap_origin != LLAP_ORIGIN_READAHEAD));
 588
 589                 LL_CDEBUG_PAGE(D_PAGE, page,
 590                                "%s LRU page: %s%s%s%s%s origin %s\n",
 591                                keep ? "keep" : "drop",
 592                                llap->llap_write_queued ? "wq " : "",
 593                                PageDirty(page) ? "pd " : "",
 594                                PageUptodate(page) ? "" : "!pu ",
 595                                PageWriteback(page) ? "wb" : "",
 596                                llap->llap_defer_uptodate ? "" : "!du",
 597                                llap_origins[llap->llap_origin]);
 598
 599                 /* If page is dirty or undergoing IO don't discard it */
 600                 if (keep) {
 601                         unlock_page(page);
 602                         continue;
 603                 }
 604
 605                 page_cache_get(page);
 606                 ll_pglist_cpu_unlock(sbi, cpu);
 607
 608                 if (page->mapping != NULL) {
 609                         ll_teardown_mmaps(page->mapping,
 610                                          (__u64)page->index << CFS_PAGE_SHIFT,
 611                                          ((__u64)page->index << CFS_PAGE_SHIFT)|
 612                                           ~CFS_PAGE_MASK);
 613                         if (!PageDirty(page) && !page_mapped(page)) {
 614                                 ll_ra_accounting(llap, page->mapping);
 615                                 ll_truncate_complete_page(page);
 616                                 ++count;
 617                         } else {
 618                                 LL_CDEBUG_PAGE(D_PAGE, page,
 619                                                "Not dropping page because it is"
 620                                                " %s\n", PageDirty(page) ?
 621                                                "dirty" : "mapped");
 622                         }
 623                 }
 624                 unlock_page(page);
 625                 page_cache_release(page);
 626
 627                 ll_pglist_cpu_lock(sbi, cpu);
 628         }
 629         list_del(&dummy_llap.llap_pglist_item);
 630         ll_pglist_cpu_unlock(sbi, cpu);
 631
 632         CDEBUG(D_CACHE, "shrank %d, expected %d however. \n", count, target);
 633         return count;
 634 }
 635
 636
 637 /* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction.
 638  *
 639  * At first, this code calculates total pages wanted by @shrink_fraction, then
 640  * it deduces how many pages should be reaped from each cpu in proportion as
 641  * their own # of page count(llpd_count).
 642  */
 643 int llap_shrink_cache(struct ll_sb_info *sbi, int nr_to_scan)
 644 {
 645         unsigned long total, want, percpu_want, count = 0;
 646         int cpu, nr_cpus;
 647
 648         total = lcounter_read_positive(&sbi->ll_async_page_count);
 649         if (total == 0 || nr_to_scan == 0)
 650                 return total;
 651
 652         want = nr_to_scan;
 653
 654         /* Since we are freeing pages also, we don't necessarily want to
 655          * shrink so much.  Limit to 40MB of pages + llaps per call. */
 656         if (want > 40 << (20 - CFS_PAGE_SHIFT))
 657                 want = 40 << (20 - CFS_PAGE_SHIFT);
 658
 659         CDEBUG(D_CACHE, "shrinking %lu of %lu pages (asked for %u)\n",
 660                want, total, nr_to_scan);
 661
 662         nr_cpus = num_possible_cpus();
 663         cpu = sbi->ll_async_page_clock_hand;
 664         /* we at most do one round */
 665         do {
 666                 int c;
 667
 668                 cpu = (cpu + 1) % nr_cpus;
 669                 c = LL_PGLIST_DATA_CPU(sbi, cpu)->llpd_count;
 670                 if (!cpu_online(cpu))
 671                         percpu_want = c;
 672                 else
 673                         percpu_want = want / ((total / (c + 1)) + 1);
 674                 if (percpu_want == 0)
 675                         continue;
 676
 677                 count += llap_shrink_cache_internal(sbi, cpu, percpu_want);
 678                 if (count >= want)
 679                         sbi->ll_async_page_clock_hand = cpu;
 680         } while (cpu != sbi->ll_async_page_clock_hand);
 681
 682         CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n",
 683                count, want, total);
 684
 685         return lcounter_read_positive(&sbi->ll_async_page_count);
 686 }
 687
 688 /* Rebalance the async page queue len for each cpu. We hope that the cpu
 689  * which do much IO job has a relative longer queue len.
 690  * This function should be called with preempt disabled.
 691  */
 692 static inline int llap_async_cache_rebalance(struct ll_sb_info *sbi)
 693 {
 694         unsigned long sample = 0, *cpu_sample, bias, slice;
 695         struct ll_pglist_data *pd;
 696         cpumask_t mask;
 697         int cpu, surplus;
 698         int w1 = 7, w2 = 3, base = (w1 + w2); /* weight value */
 699         atomic_t *pcnt;
 700
 701         if (!spin_trylock(&sbi->ll_async_page_reblnc_lock)) {
 702                 /* someone else is doing the job */
 703                 return 1;
 704         }
 705
 706         pcnt = &LL_PGLIST_DATA(sbi)->llpd_sample_count;
 707         if (!atomic_read(pcnt)) {
 708                 /* rare case, somebody else has gotten this job done */
 709                 spin_unlock(&sbi->ll_async_page_reblnc_lock);
 710                 return 1;
 711         }
 712
 713         sbi->ll_async_page_reblnc_count++;
 714         cpu_sample = sbi->ll_async_page_sample;
 715         memset(cpu_sample, 0, num_possible_cpus() * sizeof(unsigned long));
 716         for_each_online_cpu(cpu) {
 717                 pcnt = &LL_PGLIST_DATA_CPU(sbi, cpu)->llpd_sample_count;
 718                 cpu_sample[cpu] = atomic_read(pcnt);
 719                 atomic_set(pcnt, 0);
 720                 sample += cpu_sample[cpu];
 721         }
 722
 723         cpus_clear(mask);
 724         surplus = sbi->ll_async_page_max;
 725         slice = surplus / sample + 1;
 726         sample /= num_online_cpus();
 727         bias = sample >> 4;
 728         for_each_online_cpu(cpu) {
 729                 pd = LL_PGLIST_DATA_CPU(sbi, cpu);
 730                 if (labs((long int)sample - cpu_sample[cpu]) > bias) {
 731                         unsigned long budget = pd->llpd_budget;
 732                         /* weighted original queue length and expected queue
 733                          * length to avoid thrashing. */
 734                         pd->llpd_budget = (budget * w1) / base +
 735                                         (slice * cpu_sample[cpu]) * w2 / base;
 736                         cpu_set(cpu, mask);
 737                 }
 738                 surplus -= min_t(int, pd->llpd_budget, surplus);
 739         }
 740         surplus /= cpus_weight(mask) ?: 1;
 741         for_each_cpu_mask(cpu, mask)
 742                 LL_PGLIST_DATA_CPU(sbi, cpu)->llpd_budget += surplus;
 743         spin_unlock(&sbi->ll_async_page_reblnc_lock);
 744
 745         /* We need to call llap_shrink_cache_internal() for every cpu to
 746          * ensure the sbi->ll_async_page_max limit is enforced. */
 747         for_each_cpu_mask(cpu, mask) {
 748                 pd = LL_PGLIST_DATA_CPU(sbi, cpu);
 749                 llap_shrink_cache_internal(sbi, cpu, max_t(int, pd->llpd_count-
 750                                            pd->llpd_budget, 0) + 32);
 751         }
 752
 753         return 0;
 754 }
 755
 756 static struct ll_async_page *llap_from_page_with_lockh(struct page *page,
 757                                                        unsigned origin,
 758                                                        struct lustre_handle *lockh,
 759                                                        int flags)
 760 {
 761         struct ll_async_page *llap;
 762         struct obd_export *exp;
 763         struct inode *inode = page->mapping->host;
 764         struct ll_sb_info *sbi;
 765         struct ll_pglist_data *pd;
 766         int rc, cpu, target;
 767         ENTRY;
 768
 769         if (!inode) {
 770                 static int triggered;
 771
 772                 if (!triggered) {
 773                         LL_CDEBUG_PAGE(D_ERROR, page, "Bug 10047. Wrong anon "
 774                                        "page received\n");
 775                         libcfs_debug_dumpstack(NULL);
 776                         triggered = 1;
 777                 }
 778                 RETURN(ERR_PTR(-EINVAL));
 779         }
 780         sbi = ll_i2sbi(inode);
 781         LASSERT(ll_async_page_slab);
 782         LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
 783
 784         exp = ll_i2obdexp(page->mapping->host);
 785         if (exp == NULL)
 786                 RETURN(ERR_PTR(-EINVAL));
 787
 788         llap = llap_cast_private(page);
 789         if (llap != NULL) {
 790 #if 0 /* disabled since we take lock ref in readahead, see bug 16774/21252 */
 791                 if (origin == LLAP_ORIGIN_READAHEAD && lockh) {
 792                         /* the page could belong to another lock for which
 793                          * we don't hold a reference. We need to check that
 794                          * a reference is taken on a lock covering this page.
 795                          * For readpage origin, this is fine because
 796                          * ll_file_readv() took a reference on lock(s) covering
 797                          * the whole read. However, for readhead, we don't have
 798                          * this guarantee, so we need to check that the lock
 799                          * matched in ll_file_readv() also covers this page */
 800                         __u64 offset = ((loff_t)page->index) << CFS_PAGE_SHIFT;
 801                         if (!obd_get_lock(exp, ll_i2info(inode)->lli_smd,
 802                                           &llap->llap_cookie, OBD_BRW_READ,
 803                                           offset, offset + CFS_PAGE_SIZE - 1,
 804                                           lockh, flags))
 805                                 RETURN(ERR_PTR(-ENOLCK));
 806                 }
 807 #endif
 808                 /* move to end of LRU list, except when page is just about to
 809                  * die */
 810                 if (origin != LLAP_ORIGIN_REMOVEPAGE) {
 811                         int old_cpu = llap->llap_pglist_cpu;
 812                         struct ll_pglist_data *old_pd;
 813
 814                         pd = ll_pglist_double_lock(sbi, old_cpu, &old_pd);
 815                         pd->llpd_hit++;
 816                         while (old_cpu != llap->llap_pglist_cpu) {
 817                                 /* rarely case, someone else is touching this
 818                                  * page too. */
 819                                 ll_pglist_double_unlock(sbi, old_cpu);
 820                                 old_cpu = llap->llap_pglist_cpu;
 821                                 pd=ll_pglist_double_lock(sbi, old_cpu, &old_pd);
 822                         }
 823
 824                         list_move(&llap->llap_pglist_item,
 825                                   &pd->llpd_list);
 826                         old_pd->llpd_gen++;
 827                         if (pd->llpd_cpu != old_cpu) {
 828                                 pd->llpd_count++;
 829                                 old_pd->llpd_count--;
 830                                 old_pd->llpd_gen++;
 831                                 llap->llap_pglist_cpu = pd->llpd_cpu;
 832                                 pd->llpd_cross++;
 833                         }
 834                         ll_pglist_double_unlock(sbi, old_cpu);
 835                 }
 836                 GOTO(out, llap);
 837         }
 838
 839         /* limit the number of lustre-cached pages */
 840         cpu = cfs_get_cpu();
 841         pd = LL_PGLIST_DATA(sbi);
 842         target = pd->llpd_count - pd->llpd_budget;
 843         if (target > 0) {
 844                 rc = 0;
 845                 atomic_inc(&pd->llpd_sample_count);
 846                 if (atomic_read(&pd->llpd_sample_count) >
 847                     sbi->ll_async_page_sample_max) {
 848                         pd->llpd_reblnc_count++;
 849                         rc = llap_async_cache_rebalance(sbi);
 850                         if (rc == 0)
 851                                 target = pd->llpd_count - pd->llpd_budget;
 852                 }
 853                 /* if rc equals 1, it means other cpu is doing the rebalance
 854                  * job, and our budget # would be modified when we read it.
 855                  * Furthermore, it is much likely being increased because
 856                  * we have already reached the rebalance threshold. In this
 857                  * case, we skip to shrink cache here. */
 858                 if ((rc == 0) && target > 0)
 859                         llap_shrink_cache_internal(sbi, cpu, target + 32);
 860         }
 861         cfs_put_cpu();
 862
 863         OBD_SLAB_ALLOC(llap, ll_async_page_slab, CFS_ALLOC_STD,
 864                        ll_async_page_slab_size);
 865         if (llap == NULL)
 866                 RETURN(ERR_PTR(-ENOMEM));
 867         llap->llap_magic = LLAP_MAGIC;
 868         llap->llap_cookie = (void *)llap + size_round(sizeof(*llap));
 869
 870         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 871                                  (obd_off)page->index << CFS_PAGE_SHIFT,
 872                                  &ll_async_page_ops, llap, &llap->llap_cookie,
 873                                  flags, lockh);
 874         if (rc) {
 875                 OBD_SLAB_FREE(llap, ll_async_page_slab,
 876                               ll_async_page_slab_size);
 877                 RETURN(ERR_PTR(rc));
 878         }
 879
 880         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 881                page, llap->llap_cookie, (obd_off)page->index << CFS_PAGE_SHIFT);
 882         /* also zeroing the PRIVBITS low order bitflags */
 883         __set_page_ll_data(page, llap);
 884         llap->llap_page = page;
 885
 886         lcounter_inc(&sbi->ll_async_page_count);
 887         pd = ll_pglist_lock(sbi);
 888         list_add_tail(&llap->llap_pglist_item, &pd->llpd_list);
 889         pd->llpd_count++;
 890         pd->llpd_gen++;
 891         pd->llpd_miss++;
 892         llap->llap_pglist_cpu = pd->llpd_cpu;
 893         ll_pglist_unlock(sbi);
 894
 895  out:
 896         if (unlikely(sbi->ll_flags & LL_SBI_LLITE_CHECKSUM)) {
 897                 __u32 csum;
 898                 char *kaddr = kmap_atomic(page, KM_USER0);
 899                 csum = init_checksum(OSC_DEFAULT_CKSUM);
 900                 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
 901                                         OSC_DEFAULT_CKSUM);
 902                 kunmap_atomic(kaddr, KM_USER0);
 903                 if (origin == LLAP_ORIGIN_READAHEAD ||
 904                     origin == LLAP_ORIGIN_READPAGE) {
 905                         llap->llap_checksum = 0;
 906                 } else if (origin == LLAP_ORIGIN_COMMIT_WRITE ||
 907                            llap->llap_checksum == 0) {
 908                         llap->llap_checksum = csum;
 909                         CDEBUG(D_PAGE, "page %p cksum %x\n", page, csum);
 910                 } else if (llap->llap_checksum == csum) {
 911                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 912                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 913                                page, csum);
 914                 } else {
 915                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 916                         LL_CDEBUG_PAGE(D_ERROR, page, "old cksum %x != new "
 917                                        "%x!\n", llap->llap_checksum, csum);
 918                 }
 919         }
 920
 921         llap->llap_origin = origin;
 922         RETURN(llap);
 923 }
 924
 925 static inline struct ll_async_page *llap_from_page(struct page *page,
 926                                                    unsigned origin)
 927 {
 928         return llap_from_page_with_lockh(page, origin, NULL, 0);
 929 }
 930
 931 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
 932                                struct ll_async_page *llap,
 933                                unsigned to, obd_flag async_flags)
 934 {
 935         unsigned long size_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
 936         struct obd_io_group *oig;
 937         struct ll_sb_info *sbi = ll_i2sbi(inode);
 938         int rc, noquot = llap->llap_ignore_quota ? OBD_BRW_NOQUOTA : 0;
 939         int brwflags = OBD_BRW_ASYNC;
 940         ENTRY;
 941
 942         /* _make_ready only sees llap once we've unlocked the page */
 943         llap->llap_write_queued = 1;
 944         rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
 945                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 946                                 0, 0, brwflags, async_flags);
 947         if (rc == 0) {
 948                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
 949                 llap_write_pending(inode, llap);
 950                 GOTO(out, 0);
 951         }
 952
 953         llap->llap_write_queued = 0;
 954
 955         rc = oig_init(&oig);
 956         if (rc)
 957                 GOTO(out, rc);
 958
 959         /* make full-page requests if we are not at EOF (bug 4410) */
 960         if (to != CFS_PAGE_SIZE && llap->llap_page->index < size_index) {
 961                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 962                                "sync write before EOF: size_index %lu, to %d\n",
 963                                size_index, to);
 964                 to = CFS_PAGE_SIZE;
 965         } else if (to != CFS_PAGE_SIZE && llap->llap_page->index == size_index){
 966                 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
 967                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 968                                "sync write at EOF: size_index %lu, to %d/%d\n",
 969                                size_index, to, size_to);
 970                 if (to < size_to)
 971                         to = size_to;
 972         }
 973
 974         /* compare the checksum once before the page leaves llite */
 975         if (unlikely((sbi->ll_flags & LL_SBI_LLITE_CHECKSUM) &&
 976                      llap->llap_checksum != 0)) {
 977                 __u32 csum;
 978                 struct page *page = llap->llap_page;
 979                 char *kaddr = kmap_atomic(page, KM_USER0);
 980                 csum = init_checksum(OSC_DEFAULT_CKSUM);
 981                 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
 982                                         OSC_DEFAULT_CKSUM);
 983                 kunmap_atomic(kaddr, KM_USER0);
 984                 if (llap->llap_checksum == csum) {
 985                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 986                                page, csum);
 987                 } else {
 988                         CERROR("page %p old cksum %x != new cksum %x!\n",
 989                                page, llap->llap_checksum, csum);
 990                 }
 991         }
 992
 993         rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
 994                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 995                                 0, to, 0, ASYNC_READY | ASYNC_URGENT |
 996                                 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
 997         if (rc)
 998                 GOTO(free_oig, rc);
 999
1000         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1001         if (rc)
1002                 GOTO(free_oig, rc);
1003
1004         rc = oig_wait(oig);
1005
1006         if (!rc && async_flags & ASYNC_READY) {
1007                 unlock_page(llap->llap_page);
1008                 if (PageWriteback(llap->llap_page))
1009                         end_page_writeback(llap->llap_page);
1010         }
1011
1012         LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
1013
1014 free_oig:
1015         oig_release(oig);
1016 out:
1017         RETURN(rc);
1018 }
1019
1020 /* update our write count to account for i_size increases that may have
1021  * happened since we've queued the page for io. */
1022
1023 /* be careful not to return success without setting the page Uptodate or
1024  * the next pass through prepare_write will read in stale data from disk. */
1025 int ll_commit_write(struct file *file, struct page *page, unsigned from,
1026                     unsigned to)
1027 {
1028         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1029         struct inode *inode = page->mapping->host;
1030         struct ll_inode_info *lli = ll_i2info(inode);
1031         struct lov_stripe_md *lsm = lli->lli_smd;
1032         struct obd_export *exp;
1033         struct ll_async_page *llap;
1034         loff_t size;
1035         struct lustre_handle *lockh = NULL;
1036         int rc = 0;
1037         ENTRY;
1038
1039         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1040         LASSERT(inode == file->f_dentry->d_inode);
1041         LASSERT(PageLocked(page));
1042
1043         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
1044                inode, page, from, to, page->index);
1045
1046         if (fd->fd_flags & LL_FILE_GROUP_LOCKED)
1047                 lockh = &fd->fd_cwlockh;
1048
1049         llap = llap_from_page_with_lockh(page, LLAP_ORIGIN_COMMIT_WRITE, lockh,
1050                                          0);
1051         if (IS_ERR(llap))
1052                 RETURN(PTR_ERR(llap));
1053
1054         exp = ll_i2obdexp(inode);
1055         if (exp == NULL)
1056                 RETURN(-EINVAL);
1057
1058         llap->llap_ignore_quota = cfs_capable(CFS_CAP_SYS_RESOURCE);
1059
1060         /* queue a write for some time in the future the first time we
1061          * dirty the page */
1062         if (!PageDirty(page)) {
1063                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_MISSES, 1);
1064
1065                 rc = queue_or_sync_write(exp, inode, llap, to, 0);
1066                 if (rc)
1067                         GOTO(out, rc);
1068         } else {
1069                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_HITS, 1);
1070         }
1071
1072         /* put the page in the page cache, from now on ll_removepage is
1073          * responsible for cleaning up the llap.
1074          * only set page dirty when it's queued to be write out */
1075         if (llap->llap_write_queued)
1076                 set_page_dirty(page);
1077
1078 out:
1079         size = (((obd_off)page->index) << CFS_PAGE_SHIFT) + to;
1080         ll_inode_size_lock(inode, 0);
1081         if (rc == 0) {
1082                 lov_stripe_lock(lsm);
1083                 obd_adjust_kms(exp, lsm, size, 0);
1084                 lov_stripe_unlock(lsm);
1085                 if (size > i_size_read(inode))
1086                         i_size_write(inode, size);
1087                 SetPageUptodate(page);
1088         } else if (size > i_size_read(inode)) {
1089                 /* this page beyond the pales of i_size, so it can't be
1090                  * truncated in ll_p_r_e during lock revoking. we must
1091                  * teardown our book-keeping here. */
1092                 ll_removepage(page);
1093         }
1094         ll_inode_size_unlock(inode, 0);
1095         RETURN(rc);
1096 }
1097
1098 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
1099
1100 /* WARNING: This algorithm is used to reduce the contention on
1101  * sbi->ll_lock. It should work well if the ra_max_pages is much
1102  * greater than the single file's read-ahead window.
1103  *
1104  * TODO: There may exist a `global sync problem' in this implementation.
1105  * Considering the global ra window is 100M, and each file's ra window is 10M,
1106  * there are over 10 files trying to get its ra budget and reach
1107  * ll_ra_count_get at the exactly same time. All of them will get a zero ra
1108  * window, although the global window is 100M. -jay
1109  */
1110 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
1111 {
1112         struct ll_ra_info *ra = &sbi->ll_ra_info;
1113         unsigned long ret = 0;
1114         ENTRY;
1115
1116         /**
1117          * If read-ahead pages left are less than 1M, do not do read-ahead,
1118          * otherwise it will form small read RPC(< 1M), which hurt server
1119          * performance a lot.
1120          */
1121         ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages), len);
1122         if ((int)ret < min((unsigned long)PTLRPC_MAX_BRW_PAGES, len))
1123                 GOTO(out, ret = 0);
1124
1125         if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
1126                 atomic_sub(ret, &ra->ra_cur_pages);
1127                 ret = 0;
1128         }
1129 out:
1130         RETURN(ret);
1131 }
1132
1133 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
1134 {
1135         struct ll_ra_info *ra = &sbi->ll_ra_info;
1136         atomic_sub(len, &ra->ra_cur_pages);
1137 }
1138
1139 /* called for each page in a completed rpc.*/
1140 int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
1141 {
1142         struct ll_async_page *llap;
1143         struct page *page;
1144         struct obd_export *exp;
1145         obd_off end;
1146         int ret = 0;
1147         ENTRY;
1148
1149         llap = LLAP_FROM_COOKIE(data);
1150         page = llap->llap_page;
1151         LASSERT(PageLocked(page));
1152         LASSERT(CheckWriteback(page,cmd));
1153
1154         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
1155
1156         if (cmd & OBD_BRW_READ && llap->llap_defer_uptodate) {
1157                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
1158
1159                 LASSERT(lustre_handle_is_used(&llap->llap_lockh_granted));
1160                 exp = ll_i2obdexp(page->mapping->host);
1161                 end = ((loff_t)page->index) << CFS_PAGE_SHIFT;
1162                 end += CFS_PAGE_SIZE - 1;
1163                 obd_cancel(exp, ll_i2info(page->mapping->host)->lli_smd, LCK_PR,
1164                            &llap->llap_lockh_granted, OBD_FAST_LOCK, end);
1165         }
1166
1167         if (rc == 0)  {
1168                 if (cmd & OBD_BRW_READ) {
1169                         if (!llap->llap_defer_uptodate)
1170                                 SetPageUptodate(page);
1171                 } else {
1172                         llap->llap_write_queued = 0;
1173                 }
1174                 ClearPageError(page);
1175         } else {
1176                 if (cmd & OBD_BRW_READ) {
1177                         llap->llap_defer_uptodate = 0;
1178                 }
1179                 SetPageError(page);
1180                 if (rc == -ENOSPC)
1181                         set_bit(AS_ENOSPC, &page->mapping->flags);
1182                 else
1183                         set_bit(AS_EIO, &page->mapping->flags);
1184         }
1185
1186         /* be carefull about clear WB.
1187          * if WB will cleared after page lock is released - paralel IO can be
1188          * started before ap_make_ready is finished - so we will be have page
1189          * with PG_Writeback set from ->writepage() and completed READ which
1190          * clear this flag */
1191         if ((cmd & OBD_BRW_WRITE) && PageWriteback(page))
1192                 end_page_writeback(page);
1193
1194         unlock_page(page);
1195
1196         if (cmd & OBD_BRW_WRITE) {
1197                 llap_write_complete(page->mapping->host, llap);
1198                 ll_try_done_writing(page->mapping->host);
1199         }
1200
1201         page_cache_release(page);
1202
1203         RETURN(ret);
1204 }
1205
1206 static void __ll_put_llap(struct page *page)
1207 {
1208         struct inode *inode = page->mapping->host;
1209         struct obd_export *exp;
1210         struct ll_async_page *llap;
1211         struct ll_sb_info *sbi = ll_i2sbi(inode);
1212         struct ll_pglist_data *pd;
1213         int rc, cpu;
1214         ENTRY;
1215
1216         exp = ll_i2obdexp(inode);
1217         if (exp == NULL) {
1218                 CERROR("page %p ind %lu gave null export\n", page, page->index);
1219                 EXIT;
1220                 return;
1221         }
1222
1223         llap = llap_from_page(page, LLAP_ORIGIN_REMOVEPAGE);
1224         if (IS_ERR(llap)) {
1225                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
1226                        page->index, PTR_ERR(llap));
1227                 EXIT;
1228                 return;
1229         }
1230
1231         //llap_write_complete(inode, llap);
1232         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
1233                                      llap->llap_cookie);
1234         if (rc != 0)
1235                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
1236
1237         /* this unconditional free is only safe because the page lock
1238          * is providing exclusivity to memory pressure/truncate/writeback..*/
1239         __clear_page_ll_data(page);
1240
1241         lcounter_dec(&sbi->ll_async_page_count);
1242         cpu = llap->llap_pglist_cpu;
1243         pd = ll_pglist_cpu_lock(sbi, cpu);
1244         pd->llpd_gen++;
1245         pd->llpd_count--;
1246         if (!list_empty(&llap->llap_pglist_item))
1247                 list_del_init(&llap->llap_pglist_item);
1248         ll_pglist_cpu_unlock(sbi, cpu);
1249         OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
1250
1251         EXIT;
1252 }
1253
1254 /* the kernel calls us here when a page is unhashed from the page cache.
1255  * the page will be locked and the kernel is holding a spinlock, so
1256  * we need to be careful.  we're just tearing down our book-keeping
1257  * here. */
1258 void ll_removepage(struct page *page)
1259 {
1260         struct ll_async_page *llap = llap_cast_private(page);
1261         ENTRY;
1262
1263         LASSERT(!in_interrupt());
1264
1265         /* sync pages or failed read pages can leave pages in the page
1266          * cache that don't have our data associated with them anymore */
1267         if (page_private(page) == 0) {
1268                 EXIT;
1269                 return;
1270         }
1271
1272         ll_ra_accounting(llap, page->mapping);
1273         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
1274         __ll_put_llap(page);
1275
1276         EXIT;
1277 }
1278
1279 static int ll_issue_page_read(struct obd_export *exp,
1280                               struct ll_async_page *llap,
1281                               struct obd_io_group *oig, int defer)
1282 {
1283         struct page *page = llap->llap_page;
1284         int rc;
1285
1286         page_cache_get(page);
1287         llap->llap_defer_uptodate = defer;
1288         llap->llap_ra_used = 0;
1289         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
1290                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
1291                                 CFS_PAGE_SIZE, 0, ASYNC_COUNT_STABLE |
1292                                 ASYNC_READY | ASYNC_URGENT);
1293         if (rc) {
1294                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
1295                 page_cache_release(page);
1296         }
1297         RETURN(rc);
1298 }
1299
1300 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
1301 {
1302         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
1303         lprocfs_counter_incr(sbi->ll_ra_stats, which);
1304 }
1305
1306 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
1307 {
1308         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
1309         ll_ra_stats_inc_sbi(sbi, which);
1310 }
1311
1312 void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping)
1313 {
1314         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
1315                 return;
1316
1317         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
1318 }
1319
1320 #define RAS_CDEBUG(ras) \
1321         CDEBUG(D_READA,                                                      \
1322                "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu"    \
1323                "csr %lu sf %lu sp %lu sl %lu \n",                            \
1324                ras->ras_last_readpage, ras->ras_consecutive_requests,        \
1325                ras->ras_consecutive_pages, ras->ras_window_start,            \
1326                ras->ras_window_len, ras->ras_next_readahead,                 \
1327                ras->ras_requests, ras->ras_request_index,                    \
1328                ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
1329                ras->ras_stride_pages, ras->ras_stride_length)
1330
1331 static int index_in_window(unsigned long index, unsigned long point,
1332                            unsigned long before, unsigned long after)
1333 {
1334         unsigned long start = point - before, end = point + after;
1335
1336         if (start > point)
1337                start = 0;
1338         if (end < point)
1339                end = ~0;
1340
1341         return start <= index && index <= end;
1342 }
1343
1344 struct ll_thread_data *ll_td_get()
1345 {
1346         struct ll_thread_data *ltd = current->journal_info;
1347
1348         LASSERT(ltd == NULL || ltd->ltd_magic == LTD_MAGIC);
1349         return ltd;
1350 }
1351
1352 void ll_td_set(struct ll_thread_data *ltd)
1353 {
1354         if (ltd == NULL) {
1355                 ltd = current->journal_info;
1356                 LASSERT(ltd == NULL || ltd->ltd_magic == LTD_MAGIC);
1357                 current->journal_info = NULL;
1358                 return;
1359         }
1360
1361         LASSERT(current->journal_info == NULL);
1362         LASSERT(ltd->ltd_magic == LTD_MAGIC);
1363         current->journal_info = ltd;
1364 }
1365
1366 static struct ll_readahead_state *ll_ras_get(struct file *f)
1367 {
1368         struct ll_file_data       *fd;
1369
1370         fd = LUSTRE_FPRIVATE(f);
1371         return &fd->fd_ras;
1372 }
1373
1374 void ll_ra_read_init(struct file *f, struct ll_ra_read *rar,
1375                      loff_t offset, size_t count)
1376 {
1377         struct ll_readahead_state *ras;
1378
1379         ras = ll_ras_get(f);
1380
1381         rar->lrr_start = offset >> CFS_PAGE_SHIFT;
1382         rar->lrr_count = (count + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
1383
1384         spin_lock(&ras->ras_lock);
1385         ras->ras_requests++;
1386         ras->ras_request_index = 0;
1387         ras->ras_consecutive_requests++;
1388         rar->lrr_reader = current;
1389
1390         list_add(&rar->lrr_linkage, &ras->ras_read_beads);
1391         spin_unlock(&ras->ras_lock);
1392 }
1393
1394 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
1395 {
1396         struct ll_readahead_state *ras;
1397
1398         ras = ll_ras_get(f);
1399
1400         spin_lock(&ras->ras_lock);
1401         list_del_init(&rar->lrr_linkage);
1402         spin_unlock(&ras->ras_lock);
1403 }
1404
1405 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
1406 {
1407         struct ll_ra_read *scan;
1408
1409         list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
1410                 if (scan->lrr_reader == current)
1411                         return scan;
1412         }
1413         return NULL;
1414 }
1415
1416 struct ll_ra_read *ll_ra_read_get(struct file *f)
1417 {
1418         struct ll_readahead_state *ras;
1419         struct ll_ra_read         *bead;
1420
1421         ras = ll_ras_get(f);
1422
1423         spin_lock(&ras->ras_lock);
1424         bead = ll_ra_read_get_locked(ras);
1425         spin_unlock(&ras->ras_lock);
1426         return bead;
1427 }
1428
1429 static int ll_read_ahead_page(struct obd_export *exp, struct obd_io_group *oig,
1430                               pgoff_t index, struct address_space *mapping)
1431 {
1432         struct ll_async_page *llap;
1433         struct page *page;
1434         unsigned int gfp_mask = 0;
1435         int rc = 0, flags = 0;
1436         struct lustre_handle lockh = { 0 };
1437         obd_off start, end;
1438
1439         gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
1440 #ifdef __GFP_NOWARN
1441         gfp_mask |= __GFP_NOWARN;
1442 #endif
1443         page = grab_cache_page_nowait_gfp(mapping, index, gfp_mask);
1444         if (page == NULL) {
1445                 ll_ra_stats_inc(mapping, RA_STAT_FAILED_GRAB_PAGE);
1446                 CDEBUG(D_READA, "g_c_p_n failed\n");
1447                 return 0;
1448         }
1449
1450         /* Check if page was truncated or reclaimed */
1451         if (page->mapping != mapping) {
1452                 ll_ra_stats_inc(mapping, RA_STAT_WRONG_GRAB_PAGE);
1453                 CDEBUG(D_READA, "g_c_p_n returned invalid page\n");
1454                 GOTO(unlock_page, rc = 0);
1455         }
1456
1457 #if 0 /* the fast lock stored in ltd can't be guaranteed to be the lock used
1458        * by the llap returned by "llap_from_page_with_lockh" if there is a
1459        * ready llap, for lock check against readahead is disabled.
1460        * see bug 16774/21252 */
1461
1462         ltd = ll_td_get();
1463         if (ltd && ltd->lock_style > 0) {
1464                 __u64 offset = ((loff_t)page->index) << CFS_PAGE_SHIFT;
1465                 lockh = ltd2lockh(ltd, offset,
1466                                   offset + CFS_PAGE_SIZE - 1);
1467                 if (ltd->lock_style == LL_LOCK_STYLE_FASTLOCK)
1468                         flags = OBD_FAST_LOCK;
1469         }
1470 #endif
1471
1472         /* we do this first so that we can see the page in the /proc
1473          * accounting */
1474         llap = llap_from_page_with_lockh(page, LLAP_ORIGIN_READAHEAD, &lockh,
1475                                          flags);
1476         if (IS_ERR(llap) || llap->llap_defer_uptodate) {
1477                 /* bail out when we hit the end of the lock. */
1478                 if (PTR_ERR(llap) == -ENOLCK) {
1479                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
1480                         CDEBUG(D_READA | D_PAGE,
1481                                "Adding page to cache failed index "
1482                                 "%ld\n", index);
1483                                 CDEBUG(D_READA, "nolock page\n");
1484                                 GOTO(unlock_page, rc = -ENOLCK);
1485                 }
1486                 CDEBUG(D_READA, "read-ahead page\n");
1487                 GOTO(unlock_page, rc = 0);
1488         }
1489
1490         /* skip completed pages */
1491         if (Page_Uptodate(page))
1492                 GOTO(unlock_page, rc = 0);
1493
1494         rc = ll_issue_page_read(exp, llap, oig, 1);
1495         if (rc == 0) {
1496                 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "started read-ahead\n");
1497                 rc = 1;
1498
1499                 if (!lustre_handle_is_used(&lockh)) {
1500                         start = ((loff_t)index) << CFS_PAGE_SHIFT;
1501                         end = start + CFS_PAGE_SIZE - 1;
1502                         rc = obd_get_lock(exp,
1503                                           ll_i2info(mapping->host)->lli_smd,
1504                                           &llap->llap_cookie, OBD_BRW_READ,
1505                                           start, end, &lockh, OBD_FAST_LOCK);
1506                         LASSERT(rc);
1507                 }
1508
1509                 llap->llap_lockh_granted = lockh;
1510         } else {
1511 unlock_page:
1512                 if (lustre_handle_is_used(&lockh))
1513                         ldlm_lock_decref(&lockh, LCK_PR);
1514
1515                 unlock_page(page);
1516                 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "skipping read-ahead\n");
1517         }
1518         page_cache_release(page);
1519         return rc;
1520 }
1521
1522 /* ra_io_arg will be filled in the beginning of ll_readahead with
1523  * ras_lock, then the following ll_read_ahead_pages will read RA
1524  * pages according to this arg, all the items in this structure are
1525  * counted by page index.
1526  */
1527 struct ra_io_arg {
1528         unsigned long ria_start;  /* start offset of read-ahead*/
1529         unsigned long ria_end;    /* end offset of read-ahead*/
1530         /* If stride read pattern is detected, ria_stoff means where
1531          * stride read is started. Note: for normal read-ahead, the
1532          * value here is meaningless, and also it will not be accessed*/
1533         pgoff_t ria_stoff;
1534         /* ria_length and ria_pages are the length and pages length in the
1535          * stride I/O mode. And they will also be used to check whether
1536          * it is stride I/O read-ahead in the read-ahead pages*/
1537         unsigned long ria_length;
1538         unsigned long ria_pages;
1539 };
1540
1541 #define RIA_DEBUG(ria)                                                \
1542         CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n",       \
1543         ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
1544         ria->ria_pages)
1545
1546 #define INIT_RAS_WINDOW_PAGES PTLRPC_MAX_BRW_PAGES
1547
1548 static inline int stride_io_mode(struct ll_readahead_state *ras)
1549 {
1550         return ras->ras_consecutive_stride_requests > 1;
1551 }
1552
1553 /* The function calculates how much pages will be read in
1554  * [off, off + length], in such stride IO area,
1555  * stride_offset = st_off, stride_lengh = st_len,
1556  * stride_pages = st_pgs
1557  *
1558  *   |------------------|*****|------------------|*****|------------|*****|....
1559  * st_off
1560  *   |--- st_pgs     ---|
1561  *   |-----     st_len   -----|
1562  *
1563  *              How many pages it should read in such pattern
1564  *              |-------------------------------------------------------------|
1565  *              off
1566  *              |<------                  length                      ------->|
1567  *
1568  *          =   |<----->|  +  |-------------------------------------| +   |---|
1569  *             start_left                 st_pgs * i                    end_left
1570  */
1571 static unsigned long
1572 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
1573                 unsigned long off, unsigned long length)
1574 {
1575         unsigned long start = off > st_off ? off - st_off : 0;
1576         unsigned long end = off + length > st_off ? off + length - st_off : 0;
1577         unsigned long start_left = 0;
1578         unsigned long end_left = 0;
1579         unsigned long pg_count;
1580
1581         if (st_len == 0 || length == 0 || end == 0)
1582                 return length;
1583
1584         start_left = do_div(start, st_len);
1585         if (start_left < st_pgs)
1586                 start_left = st_pgs - start_left;
1587         else
1588                 start_left = 0;
1589
1590         end_left = do_div(end, st_len);
1591         if (end_left > st_pgs)
1592                 end_left = st_pgs;
1593
1594         CDEBUG(D_READA, "start %lu, end %lu start_left %lu end_left %lu \n",
1595                start, end, start_left, end_left);
1596
1597         if (start == end)
1598                 pg_count = end_left - (st_pgs - start_left);
1599         else
1600                 pg_count = start_left + st_pgs * (end - start - 1) + end_left;
1601
1602         CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu"
1603                "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
1604
1605         return pg_count;
1606 }
1607
1608 static int ria_page_count(struct ra_io_arg *ria)
1609 {
1610         __u64 length = ria->ria_end >= ria->ria_start ?
1611                        ria->ria_end - ria->ria_start + 1 : 0;
1612
1613         return stride_pg_count(ria->ria_stoff, ria->ria_length,
1614                                ria->ria_pages, ria->ria_start,
1615                                length);
1616 }
1617
1618 /*Check whether the index is in the defined ra-window */
1619 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
1620 {
1621         /* If ria_length == ria_pages, it means non-stride I/O mode,
1622          * idx should always inside read-ahead window in this case
1623          * For stride I/O mode, just check whether the idx is inside
1624          * the ria_pages. */
1625         return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
1626                (idx - ria->ria_stoff) % ria->ria_length < ria->ria_pages;
1627 }
1628
1629 static int ll_read_ahead_pages(struct obd_export *exp,
1630                                struct obd_io_group *oig,
1631                                struct ra_io_arg *ria,
1632                                unsigned long *reserved_pages,
1633                                struct address_space *mapping,
1634                                unsigned long *ra_end)
1635 {
1636         int rc, count = 0, stride_ria;
1637         unsigned long page_idx;
1638
1639         LASSERT(ria != NULL);
1640         RIA_DEBUG(ria);
1641
1642         stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
1643         for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
1644                         *reserved_pages > 0; page_idx++) {
1645                 if (ras_inside_ra_window(page_idx, ria)) {
1646                         /* If the page is inside the read-ahead window*/
1647                         rc = ll_read_ahead_page(exp, oig, page_idx, mapping);
1648                         if (rc == 1) {
1649                                 (*reserved_pages)--;
1650                                 count ++;
1651                         } else if (rc == -ENOLCK)
1652                                 break;
1653                 } else if (stride_ria) {
1654                         /* If it is not in the read-ahead window, and it is
1655                          * read-ahead mode, then check whether it should skip
1656                          * the stride gap */
1657                         pgoff_t offset;
1658                         /* FIXME: This assertion only is valid when it is for
1659                          * forward read-ahead, it will be fixed when backward
1660                          * read-ahead is implemented */
1661                         LASSERTF(page_idx > ria->ria_stoff, "since %lu in the"
1662                                 " gap of ra window,it should bigger than stride"
1663                                 " offset %lu \n", page_idx, ria->ria_stoff);
1664
1665                         offset = page_idx - ria->ria_stoff;
1666                         offset = offset % (ria->ria_length);
1667                         if (offset > ria->ria_pages) {
1668                                 page_idx += ria->ria_length - offset;
1669                                 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
1670                                        ria->ria_length - offset);
1671                                 continue;
1672                         }
1673                 }
1674         }
1675         *ra_end = page_idx;
1676         return count;
1677 }
1678
1679
1680 /**
1681  * Current readahead process
1682  *          read_syscall
1683  *              |
1684  *       ll_file_readv (init ll_readahead_state for the open file)
1685  *              |
1686  *              |
1687  *   |---> ll_readpage (read page)
1688  *   |          |
1689  *   |          |
1690  *   |     ras_update  (update read-ahead window according to read pattern)
1691  *   |          |
1692  *   |          |
1693  *   |--- ll_readahead  (read_ahead pages)
1694  *
1695  *
1696  *  During this process, ras_update controls how many ahead pages it should
1697  *  read by adjusting read-ahead window(RA window).The window is represented
1698  *  by following three varibles (all these values are counted by pages)
1699  *
1700  *      1. ras_window_start: start offset of the read-ahead window. It is
1701  *                           initialized as the read offset, then as pages
1702  *                           are being read, it will be set as the last
1703  *                           page(Note: it is 1M aligned, so it actually
1704  *                           is last_page_index & ~index & (~(256 - 1));
1705  *
1706  *      2. ras_window_len: length of the read-ahead window. The read-ahead window
1707  *                         length is decided by two factors
1708  *
1709  *                         a. It is at least >= current read syscall length.
1710  *                         b. If continguous read is detected, (Note: it is syscall
1711  *                            continguous, intead of page-read contingous) the
1712  *                            read-ahead window len will be increased by 1M each
1713  *                            time.
1714  *                         c. If stride read pattern is detected, the read-ahead
1715  *                            window will also be increased 1M but by stride pattern.
1716  *                            stride pattern is defined by ras_stride_length,
1717  *                            ras_stride_pages and ras_stride_gap. (see
1718  *                            ll_readahead_states comments)
1719  *
1720  *      3. ras_next_readahead: current offset in the read-ahead window, i.e. where
1721  *                             ll_readahead will start in next next-ahead.
1722  *
1723  *
1724  *  Cache miss: If memory load is very high, it begins to evicted the page from cache,
1725  *  also includes read-ahead pages, once we found read-ahead page is being evicted before
1726  *  it is "really" accessed, it will reset the read-ahead window to the current read extent
1727  *  i.e. from current page to the end of this read.
1728  *
1729  *  In flight read-ahead amount is controlled by 2 varible (read-ahead rate)
1730  *       ra_max_pages: how much max in-flight read-ahead pages on the client.
1731  *       ra_max_pages_per_file: how much max in-flight read-ahead pages per file.
1732  **/
1733
1734
1735
1736 static int ll_readahead(struct ll_readahead_state *ras,
1737                         struct obd_export *exp, struct address_space *mapping,
1738                         struct obd_io_group *oig, int flags)
1739 {
1740         unsigned long start = 0, end = 0, reserved;
1741         unsigned long ra_end, len;
1742         struct inode *inode;
1743         struct lov_stripe_md *lsm;
1744         struct ll_ra_read *bead;
1745         struct ost_lvb lvb;
1746         struct ra_io_arg ria = { 0 };
1747         int ret = 0;
1748         __u64 kms;
1749         ENTRY;
1750
1751         inode = mapping->host;
1752         lsm = ll_i2info(inode)->lli_smd;
1753
1754         lov_stripe_lock(lsm);
1755         inode_init_lvb(inode, &lvb);
1756         obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
1757         kms = lvb.lvb_size;
1758         lov_stripe_unlock(lsm);
1759         if (kms == 0) {
1760                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
1761                 RETURN(0);
1762         }
1763
1764         spin_lock(&ras->ras_lock);
1765         bead = ll_ra_read_get_locked(ras);
1766         /* Enlarge the RA window to encompass the full read */
1767         if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
1768             bead->lrr_start + bead->lrr_count) {
1769                 obd_off read_end = ((obd_off)(bead->lrr_start +
1770                                               bead->lrr_count))<<CFS_PAGE_SHIFT;
1771                 obd_extent_calc(exp, lsm, OBD_CALC_STRIPE_RPC_END_ALIGN,
1772                                 &read_end);
1773                 ras->ras_window_len = ((read_end + 1) >> CFS_PAGE_SHIFT) -
1774                                       ras->ras_window_start;
1775         }
1776         /* Reserve a part of the read-ahead window that we'll be issuing */
1777         if (ras->ras_window_len) {
1778                 start = ras->ras_next_readahead;
1779                 end = ras->ras_window_start + ras->ras_window_len - 1;
1780         }
1781         if (end != 0) {
1782                 /* Truncate RA window to end of file */
1783                 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
1784                 ras->ras_next_readahead = max(end, end + 1);
1785                 RAS_CDEBUG(ras);
1786         }
1787         ria.ria_start = start;
1788         ria.ria_end = end;
1789         /* If stride I/O mode is detected, get stride window*/
1790         if (stride_io_mode(ras)) {
1791                 ria.ria_stoff = ras->ras_stride_offset;
1792                 ria.ria_length = ras->ras_stride_length;
1793                 ria.ria_pages = ras->ras_stride_pages;
1794         }
1795         spin_unlock(&ras->ras_lock);
1796
1797         if (end == 0) {
1798                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
1799                 RETURN(0);
1800         }
1801
1802         len = ria_page_count(&ria);
1803         if (len == 0)
1804                 RETURN(0);
1805
1806         reserved = ll_ra_count_get(ll_i2sbi(inode), len);
1807         if (reserved < len)
1808                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
1809
1810         CDEBUG(D_READA, "reserved page %lu \n", reserved);
1811
1812         ret = ll_read_ahead_pages(exp, oig, &ria, &reserved, mapping, &ra_end);
1813
1814         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
1815         if (reserved != 0)
1816                 ll_ra_count_put(ll_i2sbi(inode), reserved);
1817
1818         if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
1819                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
1820
1821         /* if we didn't get to the end of the region we reserved from
1822          * the ras we need to go back and update the ras so that the
1823          * next read-ahead tries from where we left off.  we only do so
1824          * if the region we failed to issue read-ahead on is still ahead
1825          * of the app and behind the next index to start read-ahead from */
1826         CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
1827                ra_end, end, ria.ria_end);
1828
1829         if (ra_end != (end + 1)) {
1830                 spin_lock(&ras->ras_lock);
1831                 if (ra_end < ras->ras_next_readahead &&
1832                     index_in_window(ra_end, ras->ras_window_start, 0,
1833                                     ras->ras_window_len)) {
1834                         ras->ras_next_readahead = ra_end;
1835                         RAS_CDEBUG(ras);
1836                 }
1837                 spin_unlock(&ras->ras_lock);
1838         }
1839
1840         RETURN(ret);
1841 }
1842
1843 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
1844 {
1845         ras->ras_window_start = index & (~(INIT_RAS_WINDOW_PAGES - 1));
1846 }
1847
1848 /* called with the ras_lock held or from places where it doesn't matter */
1849 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1850 {
1851         ras->ras_last_readpage = index;
1852         ras->ras_consecutive_requests = 0;
1853         ras->ras_consecutive_pages = 0;
1854         ras->ras_window_len = 0;
1855         ras_set_start(ras, index);
1856         ras->ras_next_readahead = max(ras->ras_window_start, index);
1857
1858         RAS_CDEBUG(ras);
1859 }
1860
1861 /* called with the ras_lock held or from places where it doesn't matter */
1862 static void ras_stride_reset(struct ll_readahead_state *ras)
1863 {
1864         ras->ras_consecutive_stride_requests = 0;
1865         ras->ras_stride_length = 0;
1866         ras->ras_stride_pages = 0;
1867         RAS_CDEBUG(ras);
1868 }
1869
1870 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1871 {
1872         spin_lock_init(&ras->ras_lock);
1873         ras_reset(ras, 0);
1874         ras->ras_requests = 0;
1875         INIT_LIST_HEAD(&ras->ras_read_beads);
1876 }
1877
1878 /*
1879  * Check whether the read request is in the stride window.
1880  * If it is in the stride window, return 1, otherwise return 0.
1881  */
1882 static int index_in_stride_window(unsigned long index,
1883                                   struct ll_readahead_state *ras,
1884                                   struct inode *inode)
1885 {
1886         unsigned long stride_gap = index - ras->ras_last_readpage - 1;
1887
1888         if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 ||
1889             ras->ras_stride_pages == ras->ras_stride_length)
1890                 return 0;
1891
1892         /* If it is contiguous read */
1893         if (stride_gap == 0)
1894                 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
1895
1896         /*Otherwise check the stride by itself */
1897         return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
1898              ras->ras_consecutive_pages == ras->ras_stride_pages;
1899 }
1900
1901 static void ras_update_stride_detector(struct ll_readahead_state *ras,
1902                                        unsigned long index)
1903 {
1904         unsigned long stride_gap = index - ras->ras_last_readpage - 1;
1905
1906         if (!stride_io_mode(ras) && (stride_gap != 0 ||
1907              ras->ras_consecutive_stride_requests == 0)) {
1908                 ras->ras_stride_pages = ras->ras_consecutive_pages;
1909                 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
1910         }
1911         LASSERT(ras->ras_request_index == 0);
1912         LASSERT(ras->ras_consecutive_stride_requests == 0);
1913
1914         if (index <= ras->ras_last_readpage) {
1915                 /*Reset stride window for forward read*/
1916                 ras_stride_reset(ras);
1917                 return;
1918         }
1919
1920         ras->ras_stride_pages = ras->ras_consecutive_pages;
1921         ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
1922
1923         RAS_CDEBUG(ras);
1924         return;
1925 }
1926
1927 static unsigned long
1928 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
1929 {
1930         return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
1931                                ras->ras_stride_pages, ras->ras_stride_offset,
1932                                len);
1933 }
1934
1935 /* Stride Read-ahead window will be increased inc_len according to
1936  * stride I/O pattern */
1937 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1938                                        struct ll_ra_info *ra,
1939                                        unsigned long inc_len)
1940 {
1941         unsigned long left, step, window_len;
1942         unsigned long stride_len;
1943
1944         LASSERT(ras->ras_stride_length > 0);
1945         LASSERTF(ras->ras_window_start + ras->ras_window_len
1946                  >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
1947                  " stride_offset %lu\n", ras->ras_window_start,
1948                  ras->ras_window_len, ras->ras_stride_offset);
1949
1950         stride_len = ras->ras_window_start + ras->ras_window_len -
1951                      ras->ras_stride_offset;
1952
1953         left = stride_len % ras->ras_stride_length;
1954         window_len = ras->ras_window_len - left;
1955
1956         if (left < ras->ras_stride_pages)
1957                 left += inc_len;
1958         else
1959                 left = ras->ras_stride_pages + inc_len;
1960
1961         LASSERT(ras->ras_stride_pages != 0);
1962
1963         step = left / ras->ras_stride_pages;
1964         left %= ras->ras_stride_pages;
1965
1966         window_len += step * ras->ras_stride_length + left;
1967
1968         if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
1969                 ras->ras_window_len = window_len;
1970
1971         RAS_CDEBUG(ras);
1972 }
1973
1974 static void ras_increase_window(struct ll_readahead_state *ras,
1975                                 struct ll_ra_info *ra, struct inode *inode)
1976 {
1977         __u64 step;
1978         __u32 size;
1979         int rc;
1980
1981         step = ((loff_t)(ras->ras_window_start +
1982                          ras->ras_window_len)) << CFS_PAGE_SHIFT;
1983         size = sizeof(step);
1984         /*Get rpc_size for this offset (step) */
1985         rc = obd_get_info(ll_i2obdexp(inode), sizeof(KEY_OFF_RPCSIZE),
1986                           KEY_OFF_RPCSIZE, &size, &step,
1987                           ll_i2info(inode)->lli_smd);
1988         if (rc)
1989                 step = INIT_RAS_WINDOW_PAGES;
1990
1991         if (stride_io_mode(ras))
1992                 ras_stride_increase_window(ras, ra, (unsigned long)step);
1993         else
1994                 ras->ras_window_len = min(ras->ras_window_len +
1995                                           (unsigned long)step,
1996                                           ra->ra_max_pages);
1997 }
1998
1999 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
2000                        struct ll_readahead_state *ras, unsigned long index,
2001                        unsigned hit)
2002 {
2003         struct ll_ra_info *ra = &sbi->ll_ra_info;
2004         int zero = 0, stride_detect = 0, ra_miss = 0;
2005         ENTRY;
2006
2007         spin_lock(&ras->ras_lock);
2008
2009         ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
2010
2011         /* reset the read-ahead window in two cases.  First when the app seeks
2012          * or reads to some other part of the file.  Secondly if we get a
2013          * read-ahead miss that we think we've previously issued.  This can
2014          * be a symptom of there being so many read-ahead pages that the VM is
2015          * reclaiming it before we get to it. */
2016         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
2017                 zero = 1;
2018                 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
2019         } else if (!hit && ras->ras_window_len &&
2020                    index < ras->ras_next_readahead &&
2021                    index_in_window(index, ras->ras_window_start, 0,
2022                                    ras->ras_window_len)) {
2023                 ra_miss = 1;
2024                 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
2025         }
2026
2027         /* On the second access to a file smaller than the tunable
2028          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
2029          * file up to ra_max_pages_per_file.  This is simply a best effort
2030          * and only occurs once per open file.  Normal RA behavior is reverted
2031          * to for subsequent IO.  The mmap case does not increment
2032          * ras_requests and thus can never trigger this behavior. */
2033         if (ras->ras_requests == 2 && !ras->ras_request_index) {
2034                 __u64 kms_pages;
2035
2036                 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
2037                             CFS_PAGE_SHIFT;
2038
2039                 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
2040                        ra->ra_max_read_ahead_whole_pages,
2041                        ra->ra_max_pages_per_file);
2042
2043                 if (kms_pages &&
2044                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
2045                         ras->ras_window_start = 0;
2046                         ras->ras_last_readpage = 0;
2047                         ras->ras_next_readahead = 0;
2048                         ras->ras_window_len = min(ra->ra_max_pages_per_file,
2049                                 ra->ra_max_read_ahead_whole_pages);
2050                         GOTO(out_unlock, 0);
2051                 }
2052         }
2053         if (zero) {
2054                 /* check whether it is in stride I/O mode*/
2055                 if (!index_in_stride_window(index, ras, inode)) {
2056                         if (ras->ras_consecutive_stride_requests == 0 &&
2057                             ras->ras_request_index == 0) {
2058                                 ras_update_stride_detector(ras, index);
2059                                 ras->ras_consecutive_stride_requests ++;
2060                         } else {
2061                                 ras_stride_reset(ras);
2062                         }
2063                         ras_reset(ras, index);
2064                         ras->ras_consecutive_pages++;
2065                         GOTO(out_unlock, 0);
2066                 } else {
2067                         ras->ras_consecutive_pages = 0;
2068                         ras->ras_consecutive_requests = 0;
2069                         if (++ras->ras_consecutive_stride_requests > 1)
2070                                 stride_detect = 1;
2071                         RAS_CDEBUG(ras);
2072                 }
2073         } else {
2074                 if (ra_miss) {
2075                         if (index_in_stride_window(index, ras, inode) &&
2076                             stride_io_mode(ras)) {
2077                                 /*If stride-RA hit cache miss, the stride dector
2078                                  *will not be reset to avoid the overhead of
2079                                  *redetecting read-ahead mode */
2080                                 if (index != ras->ras_last_readpage + 1)
2081                                        ras->ras_consecutive_pages = 0;
2082                                 ras_reset(ras, index);
2083                                 RAS_CDEBUG(ras);
2084                         } else {
2085                                 /* Reset both stride window and normal RA
2086                                  * window */
2087                                 ras_reset(ras, index);
2088                                 ras->ras_consecutive_pages++;
2089                                 ras_stride_reset(ras);
2090                                 GOTO(out_unlock, 0);
2091                         }
2092                 } else if (stride_io_mode(ras)) {
2093                         /* If this is contiguous read but in stride I/O mode
2094                          * currently, check whether stride step still is valid,
2095                          * if invalid, it will reset the stride ra window*/
2096                         if (!index_in_stride_window(index, ras, inode)) {
2097                                 /* Shrink stride read-ahead window to be zero */
2098                                 ras_stride_reset(ras);
2099                                 ras->ras_window_len = 0;
2100                                 ras->ras_next_readahead = index;
2101                         }
2102                 }
2103         }
2104         ras->ras_consecutive_pages++;
2105         ras->ras_last_readpage = index;
2106         ras_set_start(ras, index);
2107         ras->ras_next_readahead = max(ras->ras_window_start,
2108                                       ras->ras_next_readahead);
2109         RAS_CDEBUG(ras);
2110
2111         /* Trigger RA in the mmap case where ras_consecutive_requests
2112          * is not incremented and thus can't be used to trigger RA */
2113         if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
2114                 ras->ras_window_len = INIT_RAS_WINDOW_PAGES;
2115                 GOTO(out_unlock, 0);
2116         }
2117
2118         /* Initially reset the stride window offset to next_readahead*/
2119         if (ras->ras_consecutive_stride_requests == 2 && stride_detect) {
2120                 /**
2121                  * Once stride IO mode is detected, next_readahead should be
2122                  * reset to make sure next_readahead > stride offset
2123                  */
2124                 ras->ras_next_readahead = max(index, ras->ras_next_readahead);
2125                 ras->ras_stride_offset = index;
2126                 ras->ras_window_len = INIT_RAS_WINDOW_PAGES;
2127         }
2128
2129         /* The initial ras_window_len is set to the request size.  To avoid
2130          * uselessly reading and discarding pages for random IO the window is
2131          * only increased once per consecutive request received. */
2132         if ((ras->ras_consecutive_requests > 1 || stride_detect) &&
2133             !ras->ras_request_index)
2134                 ras_increase_window(ras, ra, inode);
2135         EXIT;
2136 out_unlock:
2137         RAS_CDEBUG(ras);
2138         ras->ras_request_index++;
2139         spin_unlock(&ras->ras_lock);
2140         return;
2141 }
2142
2143 int ll_writepage(struct page *page)
2144 {
2145         struct inode *inode = page->mapping->host;
2146         struct ll_inode_info *lli = ll_i2info(inode);
2147         struct obd_export *exp;
2148         struct ll_async_page *llap;
2149         struct ll_thread_data *ltd;
2150         struct lustre_handle *lockh = NULL;
2151         int rc = 0;
2152         ENTRY;
2153
2154         LASSERT(PageLocked(page));
2155
2156         exp = ll_i2obdexp(inode);
2157         if (exp == NULL)
2158                 GOTO(out, rc = -EINVAL);
2159
2160         ltd = ll_td_get();
2161         /* currently, no FAST lock in write path */
2162         if (ltd && ltd->lock_style == LL_LOCK_STYLE_TREELOCK) {
2163                 __u64 offset = ((loff_t)page->index) << CFS_PAGE_SHIFT;
2164                 lockh = ltd2lockh(ltd, offset, offset + CFS_PAGE_SIZE - 1);
2165         }
2166
2167         llap = llap_from_page_with_lockh(page, LLAP_ORIGIN_WRITEPAGE, lockh, 0);
2168         if (IS_ERR(llap))
2169                 GOTO(out, rc = PTR_ERR(llap));
2170
2171         LASSERT(!PageWriteback(page));
2172         set_page_writeback(page);
2173
2174         page_cache_get(page);
2175         if (llap->llap_write_queued) {
2176                 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
2177                 rc = obd_set_async_flags(exp, lli->lli_smd, NULL,
2178                                          llap->llap_cookie,
2179                                          ASYNC_READY | ASYNC_URGENT);
2180         } else {
2181                 rc = queue_or_sync_write(exp, inode, llap, CFS_PAGE_SIZE,
2182                                          ASYNC_READY | ASYNC_URGENT);
2183         }
2184         if (rc) {
2185                 /* re-dirty page on error so it retries write */
2186                 if (PageWriteback(page))
2187                         end_page_writeback(page);
2188
2189                 /* resend page only for not started IO*/
2190                 if (!PageError(page))
2191                         ll_redirty_page(page);
2192
2193                 page_cache_release(page);
2194         }
2195 out:
2196         if (rc) {
2197                 if (!lli->lli_async_rc)
2198                         lli->lli_async_rc = rc;
2199                 /* resend page only for not started IO*/
2200                 unlock_page(page);
2201         }
2202         RETURN(rc);
2203 }
2204
2205 /*
2206  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
2207  * read-ahead assumes it is valid to issue readpage all the way up to
2208  * i_size, but our dlm locks make that not the case.  We disable the
2209  * kernel's read-ahead and do our own by walking ahead in the page cache
2210  * checking for dlm lock coverage.  the main difference between 2.4 and
2211  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
2212  * so they look the same.
2213  */
2214 int ll_readpage(struct file *filp, struct page *page)
2215 {
2216         struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
2217         struct inode *inode = page->mapping->host;
2218         struct obd_export *exp;
2219         struct ll_async_page *llap;
2220         struct obd_io_group *oig = NULL;
2221         struct lustre_handle *lockh = NULL;
2222         int rc, flags = 0;
2223         ENTRY;
2224
2225         LASSERT(PageLocked(page));
2226         LASSERT(!PageUptodate(page));
2227         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset=%Lu=%#Lx\n",
2228                inode->i_ino, inode->i_generation, inode,
2229                (((loff_t)page->index) << CFS_PAGE_SHIFT),
2230                (((loff_t)page->index) << CFS_PAGE_SHIFT));
2231         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
2232
2233         if (!ll_i2info(inode)->lli_smd) {
2234                 /* File with no objects - one big hole */
2235                 /* We use this just for remove_from_page_cache that is not
2236                  * exported, we'd make page back up to date. */
2237                 ll_truncate_complete_page(page);
2238                 clear_page(kmap(page));
2239                 kunmap(page);
2240                 SetPageUptodate(page);
2241                 unlock_page(page);
2242                 RETURN(0);
2243         }
2244
2245         rc = oig_init(&oig);
2246         if (rc < 0)
2247                 GOTO(out, rc);
2248
2249         exp = ll_i2obdexp(inode);
2250         if (exp == NULL)
2251                 GOTO(out, rc = -EINVAL);
2252
2253         if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
2254                 lockh = &fd->fd_cwlockh;
2255         } else {
2256                 struct ll_thread_data *ltd;
2257                 ltd = ll_td_get();
2258                 if (ltd && ltd->lock_style > 0) {
2259                         __u64 offset = ((loff_t)page->index) << CFS_PAGE_SHIFT;
2260                         lockh = ltd2lockh(ltd, offset,
2261                                           offset + CFS_PAGE_SIZE - 1);
2262                         if (ltd->lock_style == LL_LOCK_STYLE_FASTLOCK)
2263                                 flags = OBD_FAST_LOCK;
2264                 }
2265         }
2266
2267         llap = llap_from_page_with_lockh(page, LLAP_ORIGIN_READPAGE, lockh,
2268                                          flags);
2269         if (IS_ERR(llap)) {
2270                 if (PTR_ERR(llap) == -ENOLCK) {
2271                         CWARN("ino %lu page %lu (%llu) not covered by "
2272                               "a lock (mmap?).  check debug logs.\n",
2273                               inode->i_ino, page->index,
2274                               (long long)page->index << PAGE_CACHE_SHIFT);
2275                 }
2276                 GOTO(out, rc = PTR_ERR(llap));
2277         }
2278
2279         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages_per_file)
2280                 ras_update(ll_i2sbi(inode), inode, &fd->fd_ras, page->index,
2281                            llap->llap_defer_uptodate);
2282
2283
2284         if (llap->llap_defer_uptodate) {
2285                 /* This is the callpath if we got the page from a readahead */
2286                 llap->llap_ra_used = 1;
2287                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
2288                                   fd->fd_flags);
2289                 if (rc > 0)
2290                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
2291                                              NULL, oig);
2292                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
2293                 SetPageUptodate(page);
2294                 unlock_page(page);
2295                 GOTO(out_oig, rc = 0);
2296         }
2297
2298         rc = ll_issue_page_read(exp, llap, oig, 0);
2299         if (rc)
2300                 GOTO(out, rc);
2301
2302         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
2303         /* We have just requested the actual page we want, see if we can tack
2304          * on some readahead to that page's RPC before it is sent. */
2305         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages_per_file)
2306                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
2307                              fd->fd_flags);
2308
2309         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
2310
2311 out:
2312         if (rc)
2313                 unlock_page(page);
2314 out_oig:
2315         if (oig != NULL)
2316                 oig_release(oig);
2317         RETURN(rc);
2318 }
2319