lustre/llite/rw.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Lustre Lite I/O page cache routines shared by different kernel revs
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23 #include <linux/autoconf.h>
  24 #include <linux/kernel.h>
  25 #include <linux/mm.h>
  26 #include <linux/string.h>
  27 #include <linux/stat.h>
  28 #include <linux/errno.h>
  29 #include <linux/smp_lock.h>
  30 #include <linux/unistd.h>
  31 #include <linux/version.h>
  32 #include <asm/system.h>
  33 #include <asm/uaccess.h>
  34
  35 #include <linux/fs.h>
  36 #include <linux/stat.h>
  37 #include <asm/uaccess.h>
  38 #include <linux/mm.h>
  39 #include <linux/pagemap.h>
  40 #include <linux/smp_lock.h>
  41
  42 #define DEBUG_SUBSYSTEM S_LLITE
  43
  44 //#include <lustre_mdc.h>
  45 #include <lustre_lite.h>
  46 #include <obd_cksum.h>
  47 #include "llite_internal.h"
  48 #include <linux/lustre_compat25.h>
  49
  50 #ifndef list_for_each_prev_safe
  51 #define list_for_each_prev_safe(pos, n, head) \
  52         for (pos = (head)->prev, n = pos->prev; pos != (head); \
  53                 pos = n, n = pos->prev )
  54 #endif
  55
  56 cfs_mem_cache_t *ll_async_page_slab = NULL;
  57 size_t ll_async_page_slab_size = 0;
  58
  59 /* SYNCHRONOUS I/O to object storage for an inode */
  60 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
  61                   struct page *page, int flags)
  62 {
  63         struct ll_inode_info *lli = ll_i2info(inode);
  64         struct lov_stripe_md *lsm = lli->lli_smd;
  65         struct obd_info oinfo = { { { 0 } } };
  66         struct brw_page pg;
  67         int opc, rc;
  68         ENTRY;
  69
  70         pg.pg = page;
  71         pg.off = ((obd_off)page->index) << CFS_PAGE_SHIFT;
  72
  73         if ((cmd & OBD_BRW_WRITE) && (pg.off+CFS_PAGE_SIZE>i_size_read(inode)))
  74                 pg.count = i_size_read(inode) % CFS_PAGE_SIZE;
  75         else
  76                 pg.count = CFS_PAGE_SIZE;
  77
  78         LL_CDEBUG_PAGE(D_PAGE, page, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
  79                        cmd & OBD_BRW_WRITE ? "write" : "read", pg.count,
  80                        inode->i_ino, pg.off, pg.off);
  81         if (pg.count == 0) {
  82                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
  83                        LPU64"\n", inode->i_ino, inode, i_size_read(inode),
  84                        page->mapping->host, i_size_read(page->mapping->host),
  85                        page->index, pg.off);
  86         }
  87
  88         pg.flag = flags;
  89
  90         if (cmd & OBD_BRW_WRITE)
  91                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_WRITE,
  92                                    pg.count);
  93         else
  94                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_READ,
  95                                    pg.count);
  96         oinfo.oi_oa = oa;
  97         oinfo.oi_md = lsm;
  98         /* NB partial write, so we might not have CAPA_OPC_OSS_READ capa */
  99         opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
 100         oinfo.oi_capa = ll_osscapa_get(inode, opc);
 101         rc = obd_brw(cmd, ll_i2dtexp(inode), &oinfo, 1, &pg, NULL);
 102         capa_put(oinfo.oi_capa);
 103         if (rc == 0)
 104                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
 105         else if (rc != -EIO)
 106                 CERROR("error from obd_brw: rc = %d\n", rc);
 107         RETURN(rc);
 108 }
 109
 110 /* this isn't where truncate starts.   roughly:
 111  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
 112  * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
 113  * avoid races.
 114  *
 115  * must be called under ->lli_size_sem */
 116 void ll_truncate(struct inode *inode)
 117 {
 118         struct ll_inode_info *lli = ll_i2info(inode);
 119         struct obd_info oinfo = { { { 0 } } };
 120         struct ost_lvb lvb;
 121         struct obdo oa;
 122         int rc;
 123         ENTRY;
 124         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",inode->i_ino,
 125                inode->i_generation, inode, i_size_read(inode),
 126                i_size_read(inode));
 127
 128         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
 129         if (lli->lli_size_sem_owner != current) {
 130                 EXIT;
 131                 return;
 132         }
 133
 134         if (!lli->lli_smd) {
 135                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 136                        inode->i_ino);
 137                 GOTO(out_unlock, 0);
 138         }
 139
 140         LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
 141
 142         /* XXX I'm pretty sure this is a hack to paper over a more fundamental
 143          * race condition. */
 144         lov_stripe_lock(lli->lli_smd);
 145         inode_init_lvb(inode, &lvb);
 146         rc = obd_merge_lvb(ll_i2dtexp(inode), lli->lli_smd, &lvb, 0);
 147         if (lvb.lvb_size == i_size_read(inode) && rc == 0) {
 148                 CDEBUG(D_VFSTRACE, "skipping punch for obj "LPX64", %Lu=%#Lx\n",
 149                        lli->lli_smd->lsm_object_id, i_size_read(inode),
 150                        i_size_read(inode));
 151                 lov_stripe_unlock(lli->lli_smd);
 152                 GOTO(out_unlock, 0);
 153         }
 154
 155         obd_adjust_kms(ll_i2dtexp(inode), lli->lli_smd, i_size_read(inode), 1);
 156         lov_stripe_unlock(lli->lli_smd);
 157
 158         if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_CHECKSUM) &&
 159                      (i_size_read(inode) & ~CFS_PAGE_MASK))) {
 160                 /* If the truncate leaves behind a partial page, update its
 161                  * checksum. */
 162                 struct page *page = find_get_page(inode->i_mapping,
 163                                                   i_size_read(inode) >>
 164                                                   CFS_PAGE_SHIFT);
 165                 if (page != NULL) {
 166                         struct ll_async_page *llap = llap_cast_private(page);
 167                         if (llap != NULL) {
 168                                 char *kaddr = kmap_atomic(page, KM_USER0);
 169                                 llap->llap_checksum =
 170                                         init_checksum(OSC_DEFAULT_CKSUM);
 171                                 llap->llap_checksum =
 172                                         compute_checksum(llap->llap_checksum,
 173                                                          kaddr, CFS_PAGE_SIZE,
 174                                                          OSC_DEFAULT_CKSUM);
 175                                 kunmap_atomic(kaddr, KM_USER0);
 176                         }
 177                         page_cache_release(page);
 178                 }
 179         }
 180
 181         CDEBUG(D_INFO, "calling punch for "LPX64" (new size %Lu=%#Lx)\n",
 182                lli->lli_smd->lsm_object_id, i_size_read(inode), i_size_read(inode));
 183
 184         oinfo.oi_md = lli->lli_smd;
 185         oinfo.oi_policy.l_extent.start = i_size_read(inode);
 186         oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
 187         oinfo.oi_oa = &oa;
 188         oa.o_id = lli->lli_smd->lsm_object_id;
 189         oa.o_gr = lli->lli_smd->lsm_object_gr;
 190         oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 191
 192         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
 193                         OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 194                         OBD_MD_FLFID | OBD_MD_FLGENER);
 195
 196         ll_inode_size_unlock(inode, 0);
 197
 198         oinfo.oi_capa = ll_osscapa_get(inode, CAPA_OPC_OSS_TRUNC);
 199         rc = obd_punch_rqset(ll_i2dtexp(inode), &oinfo, NULL);
 200         ll_truncate_free_capa(oinfo.oi_capa);
 201         if (rc)
 202                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 203         else
 204                 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 205                               OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 206         EXIT;
 207         return;
 208
 209  out_unlock:
 210         ll_inode_size_unlock(inode, 0);
 211 } /* ll_truncate */
 212
 213 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 214                      unsigned to)
 215 {
 216         struct inode *inode = page->mapping->host;
 217         struct ll_inode_info *lli = ll_i2info(inode);
 218         struct lov_stripe_md *lsm = lli->lli_smd;
 219         obd_off offset = ((obd_off)page->index) << CFS_PAGE_SHIFT;
 220         struct obd_info oinfo = { { { 0 } } };
 221         struct brw_page pga;
 222         struct obdo oa;
 223         struct ost_lvb lvb;
 224         int rc = 0;
 225         ENTRY;
 226
 227         LASSERT(PageLocked(page));
 228         (void)llap_cast_private(page); /* assertion */
 229
 230         /* Check to see if we should return -EIO right away */
 231         pga.pg = page;
 232         pga.off = offset;
 233         pga.count = CFS_PAGE_SIZE;
 234         pga.flag = 0;
 235
 236         oa.o_mode = inode->i_mode;
 237         oa.o_id = lsm->lsm_object_id;
 238         oa.o_gr = lsm->lsm_object_gr;
 239         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
 240                      OBD_MD_FLTYPE | OBD_MD_FLGROUP;
 241         obdo_from_inode(&oa, inode, OBD_MD_FLFID | OBD_MD_FLGENER);
 242
 243         oinfo.oi_oa = &oa;
 244         oinfo.oi_md = lsm;
 245         rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode), &oinfo, 1, &pga, NULL);
 246         if (rc)
 247                 RETURN(rc);
 248
 249         if (PageUptodate(page)) {
 250                 LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n");
 251                 RETURN(0);
 252         }
 253
 254         /* We're completely overwriting an existing page, so _don't_ set it up
 255          * to date until commit_write */
 256         if (from == 0 && to == CFS_PAGE_SIZE) {
 257                 LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n");
 258                 POISON_PAGE(page, 0x11);
 259                 RETURN(0);
 260         }
 261
 262         /* If are writing to a new page, no need to read old data.  The extent
 263          * locking will have updated the KMS, and for our purposes here we can
 264          * treat it like i_size. */
 265         lov_stripe_lock(lsm);
 266         inode_init_lvb(inode, &lvb);
 267         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
 268         lov_stripe_unlock(lsm);
 269         if (lvb.lvb_size <= offset) {
 270                 char *kaddr = kmap_atomic(page, KM_USER0);
 271                 LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n",
 272                                lvb.lvb_size, offset);
 273                 memset(kaddr, 0, CFS_PAGE_SIZE);
 274                 kunmap_atomic(kaddr, KM_USER0);
 275                 GOTO(prepare_done, rc = 0);
 276         }
 277
 278         /* XXX could be an async ocp read.. read-ahead? */
 279         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 280         if (rc == 0) {
 281                 /* bug 1598: don't clobber blksize */
 282                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 283                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 284         }
 285
 286         EXIT;
 287  prepare_done:
 288         if (rc == 0)
 289                 SetPageUptodate(page);
 290
 291         return rc;
 292 }
 293
 294 static int ll_ap_make_ready(void *data, int cmd)
 295 {
 296         struct ll_async_page *llap;
 297         struct page *page;
 298         ENTRY;
 299
 300         llap = LLAP_FROM_COOKIE(data);
 301         page = llap->llap_page;
 302
 303         LASSERTF(!(cmd & OBD_BRW_READ), "cmd %x page %p ino %lu index %lu\n", cmd, page,
 304                  page->mapping->host->i_ino, page->index);
 305
 306         /* we're trying to write, but the page is locked.. come back later */
 307         if (TryLockPage(page))
 308                 RETURN(-EAGAIN);
 309
 310         LASSERT(!PageWriteback(page));
 311
 312         /* if we left PageDirty we might get another writepage call
 313          * in the future.  list walkers are bright enough
 314          * to check page dirty so we can leave it on whatever list
 315          * its on.  XXX also, we're called with the cli list so if
 316          * we got the page cache list we'd create a lock inversion
 317          * with the removepage path which gets the page lock then the
 318          * cli lock */
 319         LASSERTF(!PageWriteback(page),"cmd %x page %p ino %lu index %lu\n", cmd, page,
 320                  page->mapping->host->i_ino, page->index);
 321         clear_page_dirty_for_io(page);
 322
 323         /* This actually clears the dirty bit in the radix tree.*/
 324         set_page_writeback(page);
 325
 326         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 327         page_cache_get(page);
 328
 329         RETURN(0);
 330 }
 331
 332 /* We have two reasons for giving llite the opportunity to change the
 333  * write length of a given queued page as it builds the RPC containing
 334  * the page:
 335  *
 336  * 1) Further extending writes may have landed in the page cache
 337  *    since a partial write first queued this page requiring us
 338  *    to write more from the page cache.  (No further races are possible, since
 339  *    by the time this is called, the page is locked.)
 340  * 2) We might have raced with truncate and want to avoid performing
 341  *    write RPCs that are just going to be thrown away by the
 342  *    truncate's punch on the storage targets.
 343  *
 344  * The kms serves these purposes as it is set at both truncate and extending
 345  * writes.
 346  */
 347 static int ll_ap_refresh_count(void *data, int cmd)
 348 {
 349         struct ll_inode_info *lli;
 350         struct ll_async_page *llap;
 351         struct lov_stripe_md *lsm;
 352         struct page *page;
 353         struct inode *inode;
 354         struct ost_lvb lvb;
 355         __u64 kms;
 356         ENTRY;
 357
 358         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 359         LASSERT(cmd != OBD_BRW_READ);
 360
 361         llap = LLAP_FROM_COOKIE(data);
 362         page = llap->llap_page;
 363         inode = page->mapping->host;
 364         lli = ll_i2info(inode);
 365         lsm = lli->lli_smd;
 366
 367         lov_stripe_lock(lsm);
 368         inode_init_lvb(inode, &lvb);
 369         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
 370         kms = lvb.lvb_size;
 371         lov_stripe_unlock(lsm);
 372
 373         /* catch race with truncate */
 374         if (((__u64)page->index << CFS_PAGE_SHIFT) >= kms)
 375                 return 0;
 376
 377         /* catch sub-page write at end of file */
 378         if (((__u64)page->index << CFS_PAGE_SHIFT) + CFS_PAGE_SIZE > kms)
 379                 return kms % CFS_PAGE_SIZE;
 380
 381         return CFS_PAGE_SIZE;
 382 }
 383
 384 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 385 {
 386         struct lov_stripe_md *lsm;
 387         obd_flag valid_flags;
 388
 389         lsm = ll_i2info(inode)->lli_smd;
 390
 391         oa->o_id = lsm->lsm_object_id;
 392         oa->o_gr = lsm->lsm_object_gr;
 393         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 394         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 395         if (cmd & OBD_BRW_WRITE) {
 396                 oa->o_valid |= OBD_MD_FLEPOCH;
 397                 oa->o_easize = ll_i2info(inode)->lli_ioepoch;
 398
 399                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 400                         OBD_MD_FLUID | OBD_MD_FLGID |
 401                         OBD_MD_FLFID | OBD_MD_FLGENER;
 402         }
 403
 404         obdo_from_inode(oa, inode, valid_flags);
 405 }
 406
 407 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 408 {
 409         struct ll_async_page *llap;
 410         ENTRY;
 411
 412         llap = LLAP_FROM_COOKIE(data);
 413         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 414
 415         EXIT;
 416 }
 417
 418 static void ll_ap_update_obdo(void *data, int cmd, struct obdo *oa,
 419                               obd_valid valid)
 420 {
 421         struct ll_async_page *llap;
 422         ENTRY;
 423
 424         llap = LLAP_FROM_COOKIE(data);
 425         obdo_from_inode(oa, llap->llap_page->mapping->host, valid);
 426
 427         EXIT;
 428 }
 429
 430 static struct obd_capa *ll_ap_lookup_capa(void *data, int cmd)
 431 {
 432         struct ll_async_page *llap = LLAP_FROM_COOKIE(data);
 433         int opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
 434
 435         return ll_osscapa_get(llap->llap_page->mapping->host, opc);
 436 }
 437
 438 static struct obd_async_page_ops ll_async_page_ops = {
 439         .ap_make_ready =        ll_ap_make_ready,
 440         .ap_refresh_count =     ll_ap_refresh_count,
 441         .ap_fill_obdo =         ll_ap_fill_obdo,
 442         .ap_update_obdo =       ll_ap_update_obdo,
 443         .ap_completion =        ll_ap_completion,
 444         .ap_lookup_capa =       ll_ap_lookup_capa,
 445 };
 446
 447 struct ll_async_page *llap_cast_private(struct page *page)
 448 {
 449         struct ll_async_page *llap = (struct ll_async_page *)page_private(page);
 450
 451         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 452                  "page %p private %lu gave magic %d which != %d\n",
 453                  page, page_private(page), llap->llap_magic, LLAP_MAGIC);
 454
 455         return llap;
 456 }
 457
 458 /* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction.
 459  *
 460  * There is an llap attached onto every page in lustre, linked off @sbi.
 461  * We add an llap to the list so we don't lose our place during list walking.
 462  * If llaps in the list are being moved they will only move to the end
 463  * of the LRU, and we aren't terribly interested in those pages here (we
 464  * start at the beginning of the list where the least-used llaps are.
 465  */
 466 int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction)
 467 {
 468         struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a };
 469         unsigned long total, want, count = 0;
 470
 471         total = sbi->ll_async_page_count;
 472
 473         /* There can be a large number of llaps (600k or more in a large
 474          * memory machine) so the VM 1/6 shrink ratio is likely too much.
 475          * Since we are freeing pages also, we don't necessarily want to
 476          * shrink so much.  Limit to 40MB of pages + llaps per call. */
 477         if (shrink_fraction == 0)
 478                 want = sbi->ll_async_page_count - sbi->ll_async_page_max + 32;
 479         else
 480                 want = (total + shrink_fraction - 1) / shrink_fraction;
 481
 482         if (want > 40 << (20 - CFS_PAGE_SHIFT))
 483                 want = 40 << (20 - CFS_PAGE_SHIFT);
 484
 485         CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n",
 486                want, total, shrink_fraction);
 487
 488         spin_lock(&sbi->ll_lock);
 489         list_add(&dummy_llap.llap_pglist_item, &sbi->ll_pglist);
 490
 491         while (--total >= 0 && count < want) {
 492                 struct page *page;
 493                 int keep;
 494
 495                 if (unlikely(need_resched())) {
 496                         spin_unlock(&sbi->ll_lock);
 497                         cond_resched();
 498                         spin_lock(&sbi->ll_lock);
 499                 }
 500
 501                 llap = llite_pglist_next_llap(sbi,&dummy_llap.llap_pglist_item);
 502                 list_del_init(&dummy_llap.llap_pglist_item);
 503                 if (llap == NULL)
 504                         break;
 505
 506                 page = llap->llap_page;
 507                 LASSERT(page != NULL);
 508
 509                 list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item);
 510
 511                 /* Page needs/undergoing IO */
 512                 if (TryLockPage(page)) {
 513                         LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n");
 514                         continue;
 515                 }
 516
 517                keep = (llap->llap_write_queued || PageDirty(page) ||
 518                       PageWriteback(page) || (!PageUptodate(page) &&
 519                       llap->llap_origin != LLAP_ORIGIN_READAHEAD));
 520
 521                 LL_CDEBUG_PAGE(D_PAGE, page,"%s LRU page: %s%s%s%s%s origin %s\n",
 522                                keep ? "keep" : "drop",
 523                                llap->llap_write_queued ? "wq " : "",
 524                                PageDirty(page) ? "pd " : "",
 525                                PageUptodate(page) ? "" : "!pu ",
 526                                PageWriteback(page) ? "wb" : "",
 527                                llap->llap_defer_uptodate ? "" : "!du",
 528                                llap_origins[llap->llap_origin]);
 529
 530                 /* If page is dirty or undergoing IO don't discard it */
 531                 if (keep) {
 532                         unlock_page(page);
 533                         continue;
 534                 }
 535
 536                 page_cache_get(page);
 537                 spin_unlock(&sbi->ll_lock);
 538
 539                 if (page->mapping != NULL) {
 540                         ll_teardown_mmaps(page->mapping,
 541                                          (__u64)page->index << CFS_PAGE_SHIFT,
 542                                          ((__u64)page->index << CFS_PAGE_SHIFT)|
 543                                           ~CFS_PAGE_MASK);
 544                         if (!PageDirty(page) && !page_mapped(page)) {
 545                                 ll_ra_accounting(llap, page->mapping);
 546                                 ll_truncate_complete_page(page);
 547                                 ++count;
 548                         } else {
 549                                 LL_CDEBUG_PAGE(D_PAGE, page, "Not dropping page"
 550                                                              " because it is "
 551                                                              "%s\n",
 552                                                               PageDirty(page)?
 553                                                               "dirty":"mapped");
 554                         }
 555                 }
 556                 unlock_page(page);
 557                 page_cache_release(page);
 558
 559                 spin_lock(&sbi->ll_lock);
 560         }
 561         list_del(&dummy_llap.llap_pglist_item);
 562         spin_unlock(&sbi->ll_lock);
 563
 564         CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n",
 565                count, want, total);
 566
 567         return count;
 568 }
 569
 570 struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
 571 {
 572         struct ll_async_page *llap;
 573         struct obd_export *exp;
 574         struct inode *inode = page->mapping->host;
 575         struct ll_sb_info *sbi;
 576         int rc;
 577         ENTRY;
 578
 579         if (!inode) {
 580                 static int triggered;
 581
 582                 if (!triggered) {
 583                         LL_CDEBUG_PAGE(D_ERROR, page, "Bug 10047. Wrong anon "
 584                                        "page received\n");
 585                         libcfs_debug_dumpstack(NULL);
 586                         triggered = 1;
 587                 }
 588                 RETURN(ERR_PTR(-EINVAL));
 589         }
 590         sbi = ll_i2sbi(inode);
 591         LASSERT(ll_async_page_slab);
 592         LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
 593
 594         llap = llap_cast_private(page);
 595         if (llap != NULL) {
 596                 /* move to end of LRU list, except when page is just about to
 597                  * die */
 598                 if (origin != LLAP_ORIGIN_REMOVEPAGE) {
 599                         spin_lock(&sbi->ll_lock);
 600                         sbi->ll_pglist_gen++;
 601                         list_del_init(&llap->llap_pglist_item);
 602                         list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 603                         spin_unlock(&sbi->ll_lock);
 604                 }
 605                 GOTO(out, llap);
 606         }
 607
 608         exp = ll_i2dtexp(page->mapping->host);
 609         if (exp == NULL)
 610                 RETURN(ERR_PTR(-EINVAL));
 611
 612         /* limit the number of lustre-cached pages */
 613         if (sbi->ll_async_page_count >= sbi->ll_async_page_max)
 614                 llap_shrink_cache(sbi, 0);
 615
 616         OBD_SLAB_ALLOC(llap, ll_async_page_slab, CFS_ALLOC_STD,
 617                        ll_async_page_slab_size);
 618         if (llap == NULL)
 619                 RETURN(ERR_PTR(-ENOMEM));
 620         llap->llap_magic = LLAP_MAGIC;
 621         llap->llap_cookie = (void *)llap + size_round(sizeof(*llap));
 622
 623         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 624                                  (obd_off)page->index << CFS_PAGE_SHIFT,
 625                                  &ll_async_page_ops, llap, &llap->llap_cookie);
 626         if (rc) {
 627                 OBD_SLAB_FREE(llap, ll_async_page_slab,
 628                               ll_async_page_slab_size);
 629                 RETURN(ERR_PTR(rc));
 630         }
 631
 632         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 633                page, llap->llap_cookie, (obd_off)page->index << CFS_PAGE_SHIFT);
 634         /* also zeroing the PRIVBITS low order bitflags */
 635         __set_page_ll_data(page, llap);
 636         llap->llap_page = page;
 637         spin_lock(&sbi->ll_lock);
 638         sbi->ll_pglist_gen++;
 639         sbi->ll_async_page_count++;
 640         list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 641         INIT_LIST_HEAD(&llap->llap_pending_write);
 642         spin_unlock(&sbi->ll_lock);
 643
 644  out:
 645         if (unlikely(sbi->ll_flags & LL_SBI_CHECKSUM)) {
 646                 __u32 csum;
 647                 char *kaddr = kmap_atomic(page, KM_USER0);
 648                 csum = init_checksum(OSC_DEFAULT_CKSUM);
 649                 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
 650                                         OSC_DEFAULT_CKSUM);
 651                 kunmap_atomic(kaddr, KM_USER0);
 652                 if (origin == LLAP_ORIGIN_READAHEAD ||
 653                     origin == LLAP_ORIGIN_READPAGE) {
 654                         llap->llap_checksum = 0;
 655                 } else if (origin == LLAP_ORIGIN_COMMIT_WRITE ||
 656                            llap->llap_checksum == 0) {
 657                         llap->llap_checksum = csum;
 658                         CDEBUG(D_PAGE, "page %p cksum %x\n", page, csum);
 659                 } else if (llap->llap_checksum == csum) {
 660                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 661                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 662                                page, csum);
 663                 } else {
 664                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 665                         LL_CDEBUG_PAGE(D_ERROR, page, "old cksum %x != new "
 666                                        "%x!\n", llap->llap_checksum, csum);
 667                 }
 668         }
 669
 670         llap->llap_origin = origin;
 671         RETURN(llap);
 672 }
 673
 674 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
 675                                struct ll_async_page *llap,
 676                                unsigned to, obd_flag async_flags)
 677 {
 678         unsigned long size_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
 679         struct obd_io_group *oig;
 680         struct ll_sb_info *sbi = ll_i2sbi(inode);
 681         int rc, noquot = llap->llap_ignore_quota ? OBD_BRW_NOQUOTA : 0;
 682         ENTRY;
 683
 684         /* _make_ready only sees llap once we've unlocked the page */
 685         llap->llap_write_queued = 1;
 686         rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
 687                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 688                                 0, 0, 0, async_flags);
 689         if (rc == 0) {
 690                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
 691                 GOTO(out, 0);
 692         }
 693
 694         llap->llap_write_queued = 0;
 695         /* Do not pass llap here as it is sync write. */
 696         llap_write_pending(inode, NULL);
 697
 698         rc = oig_init(&oig);
 699         if (rc)
 700                 GOTO(out, rc);
 701
 702         /* make full-page requests if we are not at EOF (bug 4410) */
 703         if (to != CFS_PAGE_SIZE && llap->llap_page->index < size_index) {
 704                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 705                                "sync write before EOF: size_index %lu, to %d\n",
 706                                size_index, to);
 707                 to = CFS_PAGE_SIZE;
 708         } else if (to != CFS_PAGE_SIZE && llap->llap_page->index == size_index) {
 709                 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
 710                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 711                                "sync write at EOF: size_index %lu, to %d/%d\n",
 712                                size_index, to, size_to);
 713                 if (to < size_to)
 714                         to = size_to;
 715         }
 716
 717         /* compare the checksum once before the page leaves llite */
 718         if (unlikely((sbi->ll_flags & LL_SBI_CHECKSUM) &&
 719                      llap->llap_checksum != 0)) {
 720                 __u32 csum;
 721                 struct page *page = llap->llap_page;
 722                 char *kaddr = kmap_atomic(page, KM_USER0);
 723                 csum = init_checksum(OSC_DEFAULT_CKSUM);
 724                 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
 725                                         OSC_DEFAULT_CKSUM);
 726                 kunmap_atomic(kaddr, KM_USER0);
 727                 if (llap->llap_checksum == csum) {
 728                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 729                                page, csum);
 730                 } else {
 731                         CERROR("page %p old cksum %x != new cksum %x!\n",
 732                                page, llap->llap_checksum, csum);
 733                 }
 734         }
 735
 736         rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
 737                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 738                                 0, to, 0, ASYNC_READY | ASYNC_URGENT |
 739                                 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
 740         if (rc)
 741                 GOTO(free_oig, rc);
 742
 743         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
 744         if (rc)
 745                 GOTO(free_oig, rc);
 746
 747         rc = oig_wait(oig);
 748
 749         if (!rc && async_flags & ASYNC_READY) {
 750                 unlock_page(llap->llap_page);
 751                 if (PageWriteback(llap->llap_page)) {
 752                         end_page_writeback(llap->llap_page);
 753                 }
 754         }
 755
 756         if (rc == 0 && llap_write_complete(inode, llap))
 757                 ll_queue_done_writing(inode, 0);
 758
 759         LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
 760
 761 free_oig:
 762         oig_release(oig);
 763 out:
 764         RETURN(rc);
 765 }
 766
 767 /* update our write count to account for i_size increases that may have
 768  * happened since we've queued the page for io. */
 769
 770 /* be careful not to return success without setting the page Uptodate or
 771  * the next pass through prepare_write will read in stale data from disk. */
 772 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 773                     unsigned to)
 774 {
 775         struct inode *inode = page->mapping->host;
 776         struct ll_inode_info *lli = ll_i2info(inode);
 777         struct lov_stripe_md *lsm = lli->lli_smd;
 778         struct obd_export *exp;
 779         struct ll_async_page *llap;
 780         loff_t size;
 781         int rc = 0;
 782         ENTRY;
 783
 784         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 785         LASSERT(inode == file->f_dentry->d_inode);
 786         LASSERT(PageLocked(page));
 787
 788         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 789                inode, page, from, to, page->index);
 790
 791         llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
 792         if (IS_ERR(llap))
 793                 RETURN(PTR_ERR(llap));
 794
 795         exp = ll_i2dtexp(inode);
 796         if (exp == NULL)
 797                 RETURN(-EINVAL);
 798
 799         llap->llap_ignore_quota = capable(CAP_SYS_RESOURCE);
 800
 801         /*
 802          * queue a write for some time in the future the first time we
 803          * dirty the page.
 804          *
 805          * This is different from what other file systems do: they usually
 806          * just mark page (and some of its buffers) dirty and rely on
 807          * balance_dirty_pages() to start a write-back. Lustre wants write-back
 808          * to be started earlier for the following reasons:
 809          *
 810          *     (1) with a large number of clients we need to limit the amount
 811          *     of cached data on the clients a lot;
 812          *
 813          *     (2) large compute jobs generally want compute-only then io-only
 814          *     and the IO should complete as quickly as possible;
 815          *
 816          *     (3) IO is batched up to the RPC size and is async until the
 817          *     client max cache is hit
 818          *     (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
 819          *
 820          */
 821         if (!PageDirty(page)) {
 822                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_MISSES, 1);
 823
 824                 rc = queue_or_sync_write(exp, inode, llap, to, 0);
 825                 if (rc)
 826                         GOTO(out, rc);
 827         } else {
 828                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_HITS, 1);
 829         }
 830
 831         /* put the page in the page cache, from now on ll_removepage is
 832          * responsible for cleaning up the llap.
 833          * only set page dirty when it's queued to be write out */
 834         if (llap->llap_write_queued)
 835                 set_page_dirty(page);
 836
 837 out:
 838         size = (((obd_off)page->index) << CFS_PAGE_SHIFT) + to;
 839         ll_inode_size_lock(inode, 0);
 840         if (rc == 0) {
 841                 lov_stripe_lock(lsm);
 842                 obd_adjust_kms(exp, lsm, size, 0);
 843                 lov_stripe_unlock(lsm);
 844                 if (size > i_size_read(inode))
 845                         i_size_write(inode, size);
 846                 SetPageUptodate(page);
 847         } else if (size > i_size_read(inode)) {
 848                 /* this page beyond the pales of i_size, so it can't be
 849                  * truncated in ll_p_r_e during lock revoking. we must
 850                  * teardown our book-keeping here. */
 851                 ll_removepage(page);
 852         }
 853         ll_inode_size_unlock(inode, 0);
 854         RETURN(rc);
 855 }
 856
 857 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 858 {
 859         struct ll_ra_info *ra = &sbi->ll_ra_info;
 860         unsigned long ret;
 861         ENTRY;
 862
 863         spin_lock(&sbi->ll_lock);
 864         ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
 865         ra->ra_cur_pages += ret;
 866         spin_unlock(&sbi->ll_lock);
 867
 868         RETURN(ret);
 869 }
 870
 871 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 872 {
 873         struct ll_ra_info *ra = &sbi->ll_ra_info;
 874         spin_lock(&sbi->ll_lock);
 875         LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
 876                  ra->ra_cur_pages, len);
 877         ra->ra_cur_pages -= len;
 878         spin_unlock(&sbi->ll_lock);
 879 }
 880
 881 /* called for each page in a completed rpc.*/
 882 int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 883 {
 884         struct ll_async_page *llap;
 885         struct page *page;
 886         int ret = 0;
 887         ENTRY;
 888
 889         llap = LLAP_FROM_COOKIE(data);
 890         page = llap->llap_page;
 891         LASSERT(PageLocked(page));
 892         LASSERT(CheckWriteback(page,cmd));
 893
 894         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 895
 896         if (cmd & OBD_BRW_READ && llap->llap_defer_uptodate)
 897                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 898
 899         if (rc == 0)  {
 900                 if (cmd & OBD_BRW_READ) {
 901                         if (!llap->llap_defer_uptodate)
 902                                 SetPageUptodate(page);
 903                 } else {
 904                         llap->llap_write_queued = 0;
 905                 }
 906                 ClearPageError(page);
 907         } else {
 908                 if (cmd & OBD_BRW_READ) {
 909                         llap->llap_defer_uptodate = 0;
 910                 }
 911                 SetPageError(page);
 912                 if (rc == -ENOSPC)
 913                         set_bit(AS_ENOSPC, &page->mapping->flags);
 914                 else
 915                         set_bit(AS_EIO, &page->mapping->flags);
 916         }
 917
 918         unlock_page(page);
 919
 920         if (cmd & OBD_BRW_WRITE) {
 921                 /* Only rc == 0, write succeed, then this page could be deleted
 922                  * from the pending_writing list
 923                  */
 924                 if (rc == 0 && llap_write_complete(page->mapping->host, llap))
 925                         ll_queue_done_writing(page->mapping->host, 0);
 926         }
 927
 928         if (PageWriteback(page)) {
 929                 end_page_writeback(page);
 930         }
 931         page_cache_release(page);
 932
 933         RETURN(ret);
 934 }
 935
 936 /* the kernel calls us here when a page is unhashed from the page cache.
 937  * the page will be locked and the kernel is holding a spinlock, so
 938  * we need to be careful.  we're just tearing down our book-keeping
 939  * here. */
 940 void ll_removepage(struct page *page)
 941 {
 942         struct inode *inode = page->mapping->host;
 943         struct obd_export *exp;
 944         struct ll_async_page *llap;
 945         struct ll_sb_info *sbi = ll_i2sbi(inode);
 946         int rc;
 947         ENTRY;
 948
 949         LASSERT(!in_interrupt());
 950
 951         /* sync pages or failed read pages can leave pages in the page
 952          * cache that don't have our data associated with them anymore */
 953         if (page_private(page) == 0) {
 954                 EXIT;
 955                 return;
 956         }
 957
 958         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 959
 960         exp = ll_i2dtexp(inode);
 961         if (exp == NULL) {
 962                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 963                 EXIT;
 964                 return;
 965         }
 966
 967         llap = llap_from_page(page, LLAP_ORIGIN_REMOVEPAGE);
 968         if (IS_ERR(llap)) {
 969                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 970                        page->index, PTR_ERR(llap));
 971                 EXIT;
 972                 return;
 973         }
 974
 975         if (llap_write_complete(inode, llap))
 976                 ll_queue_done_writing(inode, 0);
 977
 978         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 979                                      llap->llap_cookie);
 980         if (rc != 0)
 981                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 982
 983         /* this unconditional free is only safe because the page lock
 984          * is providing exclusivity to memory pressure/truncate/writeback..*/
 985         __clear_page_ll_data(page);
 986
 987         spin_lock(&sbi->ll_lock);
 988         if (!list_empty(&llap->llap_pglist_item))
 989                 list_del_init(&llap->llap_pglist_item);
 990         sbi->ll_pglist_gen++;
 991         sbi->ll_async_page_count--;
 992         spin_unlock(&sbi->ll_lock);
 993         OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
 994         EXIT;
 995 }
 996
 997 static int ll_page_matches(struct page *page, int fd_flags)
 998 {
 999         struct lustre_handle match_lockh = {0};
1000         struct inode *inode = page->mapping->host;
1001         ldlm_policy_data_t page_extent;
1002         int flags, matches;
1003         ENTRY;
1004
1005         if (unlikely(fd_flags & LL_FILE_GROUP_LOCKED))
1006                 RETURN(1);
1007
1008         page_extent.l_extent.start = (__u64)page->index << CFS_PAGE_SHIFT;
1009         page_extent.l_extent.end =
1010                 page_extent.l_extent.start + CFS_PAGE_SIZE - 1;
1011         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
1012         if (!(fd_flags & LL_FILE_READAHEAD))
1013                 flags |= LDLM_FL_CBPENDING;
1014         matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
1015                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
1016                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
1017                             &match_lockh);
1018         RETURN(matches);
1019 }
1020
1021 static int ll_issue_page_read(struct obd_export *exp,
1022                               struct ll_async_page *llap,
1023                               struct obd_io_group *oig, int defer)
1024 {
1025         struct page *page = llap->llap_page;
1026         int rc;
1027
1028         page_cache_get(page);
1029         llap->llap_defer_uptodate = defer;
1030         llap->llap_ra_used = 0;
1031         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
1032                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
1033                                 CFS_PAGE_SIZE, 0, ASYNC_COUNT_STABLE |
1034                                                   ASYNC_READY | ASYNC_URGENT);
1035         if (rc) {
1036                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
1037                 page_cache_release(page);
1038         }
1039         RETURN(rc);
1040 }
1041
1042 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
1043 {
1044         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
1045         ra->ra_stats[which]++;
1046 }
1047
1048 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
1049 {
1050         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
1051         struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
1052
1053         spin_lock(&sbi->ll_lock);
1054         ll_ra_stats_inc_unlocked(ra, which);
1055         spin_unlock(&sbi->ll_lock);
1056 }
1057
1058 void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping)
1059 {
1060         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
1061                 return;
1062
1063         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
1064 }
1065
1066 #define RAS_CDEBUG(ras) \
1067         CDEBUG(D_READA,                                                      \
1068                "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu"    \
1069                "csr %lu sf %lu sp %lu sl %lu \n",                            \
1070                ras->ras_last_readpage, ras->ras_consecutive_requests,        \
1071                ras->ras_consecutive_pages, ras->ras_window_start,            \
1072                ras->ras_window_len, ras->ras_next_readahead,                 \
1073                ras->ras_requests, ras->ras_request_index,                    \
1074                ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
1075                ras->ras_stride_pages, ras->ras_stride_length)
1076
1077 static int index_in_window(unsigned long index, unsigned long point,
1078                            unsigned long before, unsigned long after)
1079 {
1080         unsigned long start = point - before, end = point + after;
1081
1082         if (start > point)
1083                start = 0;
1084         if (end < point)
1085                end = ~0;
1086
1087         return start <= index && index <= end;
1088 }
1089
1090 static struct ll_readahead_state *ll_ras_get(struct file *f)
1091 {
1092         struct ll_file_data       *fd;
1093
1094         fd = LUSTRE_FPRIVATE(f);
1095         return &fd->fd_ras;
1096 }
1097
1098 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
1099 {
1100         struct ll_readahead_state *ras;
1101
1102         ras = ll_ras_get(f);
1103
1104         spin_lock(&ras->ras_lock);
1105         ras->ras_requests++;
1106         ras->ras_request_index = 0;
1107         ras->ras_consecutive_requests++;
1108         rar->lrr_reader = current;
1109
1110         list_add(&rar->lrr_linkage, &ras->ras_read_beads);
1111         spin_unlock(&ras->ras_lock);
1112 }
1113
1114 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
1115 {
1116         struct ll_readahead_state *ras;
1117
1118         ras = ll_ras_get(f);
1119
1120         spin_lock(&ras->ras_lock);
1121         list_del_init(&rar->lrr_linkage);
1122         spin_unlock(&ras->ras_lock);
1123 }
1124
1125 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
1126 {
1127         struct ll_ra_read *scan;
1128
1129         list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
1130                 if (scan->lrr_reader == current)
1131                         return scan;
1132         }
1133         return NULL;
1134 }
1135
1136 struct ll_ra_read *ll_ra_read_get(struct file *f)
1137 {
1138         struct ll_readahead_state *ras;
1139         struct ll_ra_read         *bead;
1140
1141         ras = ll_ras_get(f);
1142
1143         spin_lock(&ras->ras_lock);
1144         bead = ll_ra_read_get_locked(ras);
1145         spin_unlock(&ras->ras_lock);
1146         return bead;
1147 }
1148
1149 static int ll_read_ahead_page(struct obd_export *exp, struct obd_io_group *oig,
1150                               int index, struct address_space *mapping)
1151 {
1152         struct ll_async_page *llap;
1153         struct page *page;
1154         unsigned int gfp_mask = 0;
1155         int rc = 0;
1156
1157         gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
1158 #ifdef __GFP_NOWARN
1159         gfp_mask |= __GFP_NOWARN;
1160 #endif
1161         page = grab_cache_page_nowait_gfp(mapping, index, gfp_mask);
1162         if (page == NULL) {
1163                 ll_ra_stats_inc(mapping, RA_STAT_FAILED_GRAB_PAGE);
1164                 CDEBUG(D_READA, "g_c_p_n failed\n");
1165                 return 0;
1166         }
1167
1168         /* Check if page was truncated or reclaimed */
1169         if (page->mapping != mapping) {
1170                 ll_ra_stats_inc(mapping, RA_STAT_WRONG_GRAB_PAGE);
1171                 CDEBUG(D_READA, "g_c_p_n returned invalid page\n");
1172                 GOTO(unlock_page, rc = 0);
1173         }
1174
1175         /* we do this first so that we can see the page in the /proc
1176          * accounting */
1177         llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
1178         if (IS_ERR(llap) || llap->llap_defer_uptodate) {
1179                 if (PTR_ERR(llap) == -ENOLCK) {
1180                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
1181                         CDEBUG(D_READA | D_PAGE,
1182                                "Adding page to cache failed index "
1183                                 "%d\n", index);
1184                                 CDEBUG(D_READA, "nolock page\n");
1185                                 GOTO(unlock_page, rc = -ENOLCK);
1186                 }
1187                 CDEBUG(D_READA, "read-ahead page\n");
1188                 GOTO(unlock_page, rc = 0);
1189         }
1190
1191         /* skip completed pages */
1192         if (Page_Uptodate(page))
1193                 GOTO(unlock_page, rc = 0);
1194
1195         /* bail out when we hit the end of the lock. */
1196         rc = ll_issue_page_read(exp, llap, oig, 1);
1197         if (rc == 0) {
1198                 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "started read-ahead\n");
1199                 rc = 1;
1200         } else {
1201 unlock_page:
1202                 unlock_page(page);
1203                 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "skipping read-ahead\n");
1204         }
1205         page_cache_release(page);
1206         return rc;
1207 }
1208
1209 /* ra_io_arg will be filled in the beginning of ll_readahead with
1210  * ras_lock, then the following ll_read_ahead_pages will read RA
1211  * pages according to this arg, all the items in this structure are
1212  * counted by page index.
1213  */
1214 struct ra_io_arg {
1215         unsigned long ria_start;  /* start offset of read-ahead*/
1216         unsigned long ria_end;    /* end offset of read-ahead*/
1217         /* If stride read pattern is detected, ria_stoff means where
1218          * stride read is started. Note: for normal read-ahead, the
1219          * value here is meaningless, and also it will not be accessed*/
1220         pgoff_t ria_stoff;
1221         /* ria_length and ria_pages are the length and pages length in the
1222          * stride I/O mode. And they will also be used to check whether
1223          * it is stride I/O read-ahead in the read-ahead pages*/
1224         unsigned long ria_length;
1225         unsigned long ria_pages;
1226 };
1227
1228 #define RIA_DEBUG(ria)                                                \
1229         CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n",       \
1230         ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
1231         ria->ria_pages)
1232
1233 #define RAS_INCREASE_STEP (1024 * 1024 >> CFS_PAGE_SHIFT)
1234
1235 static inline int stride_io_mode(struct ll_readahead_state *ras)
1236 {
1237         return ras->ras_consecutive_stride_requests > 1;
1238 }
1239
1240 /* The function calculates how much pages will be read in
1241  * [off, off + length], which will be read by stride I/O mode,
1242  * stride_offset = st_off, stride_lengh = st_len,
1243  * stride_pages = st_pgs
1244  */
1245 static unsigned long
1246 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
1247                 unsigned long off, unsigned length)
1248 {
1249         unsigned long cont_len = st_off > off ?  st_off - off : 0;
1250         unsigned long stride_len = length + off > st_off ?
1251                            length + off + 1 - st_off : 0;
1252         unsigned long left, pg_count;
1253
1254         if (st_len == 0 || length == 0)
1255                 return length;
1256
1257         left = do_div(stride_len, st_len);
1258         left = min(left, st_pgs);
1259
1260         pg_count = left + stride_len * st_pgs + cont_len;
1261
1262         LASSERT(pg_count >= left);
1263
1264         CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %u"
1265                "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
1266
1267         return pg_count;
1268 }
1269
1270 static int ria_page_count(struct ra_io_arg *ria)
1271 {
1272         __u64 length = ria->ria_end >= ria->ria_start ?
1273                        ria->ria_end - ria->ria_start + 1 : 0;
1274
1275         return stride_pg_count(ria->ria_stoff, ria->ria_length,
1276                                ria->ria_pages, ria->ria_start,
1277                                length);
1278 }
1279
1280 /*Check whether the index is in the defined ra-window */
1281 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
1282 {
1283         /* If ria_length == ria_pages, it means non-stride I/O mode,
1284          * idx should always inside read-ahead window in this case
1285          * For stride I/O mode, just check whether the idx is inside
1286          * the ria_pages. */
1287         return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
1288                (idx - ria->ria_stoff) % ria->ria_length < ria->ria_pages;
1289 }
1290
1291 static int ll_read_ahead_pages(struct obd_export *exp,
1292                                struct obd_io_group *oig,
1293                                struct ra_io_arg *ria,
1294                                unsigned long *reserved_pages,
1295                                struct address_space *mapping,
1296                                unsigned long *ra_end)
1297 {
1298         int rc, count = 0, stride_ria;
1299         unsigned long page_idx;
1300
1301         LASSERT(ria != NULL);
1302         RIA_DEBUG(ria);
1303
1304         stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
1305         for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
1306                         *reserved_pages > 0; page_idx++) {
1307                 if (ras_inside_ra_window(page_idx, ria)) {
1308                         /* If the page is inside the read-ahead window*/
1309                         rc = ll_read_ahead_page(exp, oig, page_idx, mapping);
1310                         if (rc == 1) {
1311                                 (*reserved_pages)--;
1312                                 count ++;
1313                         } else if (rc == -ENOLCK)
1314                                 break;
1315                 } else if (stride_ria) {
1316                         /* If it is not in the read-ahead window, and it is
1317                          * read-ahead mode, then check whether it should skip
1318                          * the stride gap */
1319                         pgoff_t offset;
1320                         /* FIXME: This assertion only is valid when it is for
1321                          * forward read-ahead, it will be fixed when backward
1322                          * read-ahead is implemented */
1323                         LASSERTF(page_idx > ria->ria_stoff, "since %lu in the"
1324                                 " gap of ra window,it should bigger than stride"
1325                                 " offset %lu \n", page_idx, ria->ria_stoff);
1326
1327                         offset = page_idx - ria->ria_stoff;
1328                         offset = offset % (ria->ria_length);
1329                         if (offset > ria->ria_pages) {
1330                                 page_idx += ria->ria_length - offset;
1331                                 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
1332                                        ria->ria_length - offset);
1333                                 continue;
1334                         }
1335                 }
1336         }
1337         *ra_end = page_idx;
1338         return count;
1339 }
1340
1341 static int ll_readahead(struct ll_readahead_state *ras,
1342                          struct obd_export *exp, struct address_space *mapping,
1343                          struct obd_io_group *oig, int flags)
1344 {
1345         unsigned long start = 0, end = 0, reserved;
1346         unsigned long ra_end, len;
1347         struct inode *inode;
1348         struct lov_stripe_md *lsm;
1349         struct ll_ra_read *bead;
1350         struct ost_lvb lvb;
1351         struct ra_io_arg ria = { 0 };
1352         int ret = 0;
1353         __u64 kms;
1354         ENTRY;
1355
1356         inode = mapping->host;
1357         lsm = ll_i2info(inode)->lli_smd;
1358
1359         lov_stripe_lock(lsm);
1360         inode_init_lvb(inode, &lvb);
1361         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
1362         kms = lvb.lvb_size;
1363         lov_stripe_unlock(lsm);
1364         if (kms == 0) {
1365                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
1366                 RETURN(0);
1367         }
1368
1369         spin_lock(&ras->ras_lock);
1370         bead = ll_ra_read_get_locked(ras);
1371         /* Enlarge the RA window to encompass the full read */
1372         if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
1373             bead->lrr_start + bead->lrr_count) {
1374                 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
1375                                       ras->ras_window_start;
1376         }
1377         /* Reserve a part of the read-ahead window that we'll be issuing */
1378         if (ras->ras_window_len) {
1379                 start = ras->ras_next_readahead;
1380                 end = ras->ras_window_start + ras->ras_window_len - 1;
1381         }
1382         if (end != 0) {
1383                 /* Truncate RA window to end of file */
1384                 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
1385                 ras->ras_next_readahead = max(end, end + 1);
1386                 RAS_CDEBUG(ras);
1387         }
1388         ria.ria_start = start;
1389         ria.ria_end = end;
1390         /* If stride I/O mode is detected, get stride window*/
1391         if (stride_io_mode(ras)) {
1392                 ria.ria_length = ras->ras_stride_length;
1393                 ria.ria_pages = ras->ras_stride_pages;
1394         }
1395         spin_unlock(&ras->ras_lock);
1396
1397         if (end == 0) {
1398                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
1399                 RETURN(0);
1400         }
1401         len = ria_page_count(&ria);
1402         if (len == 0)
1403                 RETURN(0);
1404
1405         reserved = ll_ra_count_get(ll_i2sbi(inode), len);
1406
1407         if (reserved < end - start + 1)
1408                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
1409
1410         CDEBUG(D_READA, "reserved page %lu \n", reserved);
1411
1412         ret = ll_read_ahead_pages(exp, oig, &ria, &reserved, mapping, &ra_end);
1413
1414         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
1415         if (reserved != 0)
1416                 ll_ra_count_put(ll_i2sbi(inode), reserved);
1417
1418         if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
1419                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
1420
1421         /* if we didn't get to the end of the region we reserved from
1422          * the ras we need to go back and update the ras so that the
1423          * next read-ahead tries from where we left off.  we only do so
1424          * if the region we failed to issue read-ahead on is still ahead
1425          * of the app and behind the next index to start read-ahead from */
1426         CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
1427                ra_end, end, ria.ria_end);
1428
1429         if (ra_end != (end + 1)) {
1430                 spin_lock(&ras->ras_lock);
1431                 if (ra_end < ras->ras_next_readahead &&
1432                     index_in_window(ra_end, ras->ras_window_start, 0,
1433                                     ras->ras_window_len)) {
1434                         ras->ras_next_readahead = ra_end;
1435                         RAS_CDEBUG(ras);
1436                 }
1437                 spin_unlock(&ras->ras_lock);
1438         }
1439
1440         RETURN(ret);
1441 }
1442
1443 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
1444 {
1445         ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1));
1446 }
1447
1448 /* called with the ras_lock held or from places where it doesn't matter */
1449 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1450 {
1451         ras->ras_last_readpage = index;
1452         ras->ras_consecutive_requests = 0;
1453         ras->ras_consecutive_pages = 0;
1454         ras->ras_window_len = 0;
1455         ras_set_start(ras, index);
1456         ras->ras_next_readahead = max(ras->ras_window_start, index);
1457
1458         RAS_CDEBUG(ras);
1459 }
1460
1461 /* called with the ras_lock held or from places where it doesn't matter */
1462 static void ras_stride_reset(struct ll_readahead_state *ras)
1463 {
1464         ras->ras_consecutive_stride_requests = 0;
1465         RAS_CDEBUG(ras);
1466 }
1467
1468 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1469 {
1470         spin_lock_init(&ras->ras_lock);
1471         ras_reset(ras, 0);
1472         ras->ras_requests = 0;
1473         INIT_LIST_HEAD(&ras->ras_read_beads);
1474 }
1475
1476 /* Check whether the read request is in the stride window.
1477  * If it is in the stride window, return 1, otherwise return 0.
1478  * and also update stride_gap and stride_pages.
1479  */
1480 static int index_in_stride_window(unsigned long index,
1481                                   struct ll_readahead_state *ras,
1482                                   struct inode *inode)
1483 {
1484         int stride_gap = index - ras->ras_last_readpage - 1;
1485
1486         LASSERT(stride_gap != 0);
1487
1488         if (ras->ras_consecutive_pages == 0)
1489                 return 0;
1490
1491         /*Otherwise check the stride by itself */
1492         if ((ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
1493             ras->ras_consecutive_pages == ras->ras_stride_pages)
1494                 return 1;
1495
1496         if (stride_gap >= 0) {
1497                 /*
1498                  * only set stride_pages, stride_length if
1499                  * it is forward reading ( stride_gap > 0)
1500                  */
1501                 ras->ras_stride_pages = ras->ras_consecutive_pages;
1502                 ras->ras_stride_length = stride_gap + ras->ras_consecutive_pages;
1503         } else {
1504                 /*
1505                  * If stride_gap < 0,(back_forward reading),
1506                  * reset the stride_pages/length.
1507                  * FIXME:back_ward stride I/O read.
1508                  *
1509                  */
1510                 ras->ras_stride_pages = 0;
1511                 ras->ras_stride_length = 0;
1512         }
1513         RAS_CDEBUG(ras);
1514
1515         return 0;
1516 }
1517
1518 static unsigned long
1519 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
1520 {
1521         return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
1522                                ras->ras_stride_pages, ras->ras_stride_offset,
1523                                len);
1524 }
1525
1526 /* Stride Read-ahead window will be increased inc_len according to
1527  * stride I/O pattern */
1528 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1529                                        struct ll_ra_info *ra,
1530                                        unsigned long inc_len)
1531 {
1532         unsigned long left, step, window_len;
1533         unsigned long stride_len;
1534
1535         LASSERT(ras->ras_stride_length > 0);
1536
1537         stride_len = ras->ras_window_start + ras->ras_window_len -
1538                      ras->ras_stride_offset;
1539
1540         LASSERTF(stride_len >= 0, "window_start %lu, window_len %lu"
1541                  " stride_offset %lu\n", ras->ras_window_start,
1542                  ras->ras_window_len, ras->ras_stride_offset);
1543
1544         left = stride_len % ras->ras_stride_length;
1545
1546         window_len = ras->ras_window_len - left;
1547
1548         if (left < ras->ras_stride_pages)
1549                 left += inc_len;
1550         else
1551                 left = ras->ras_stride_pages + inc_len;
1552
1553         LASSERT(ras->ras_stride_pages != 0);
1554
1555         step = left / ras->ras_stride_pages;
1556         left %= ras->ras_stride_pages;
1557
1558         window_len += step * ras->ras_stride_length + left;
1559
1560         if (stride_page_count(ras, window_len) <= ra->ra_max_pages)
1561                 ras->ras_window_len = window_len;
1562
1563         RAS_CDEBUG(ras);
1564 }
1565
1566 /* Set stride I/O read-ahead window start offset */
1567 static void ras_set_stride_offset(struct ll_readahead_state *ras)
1568 {
1569         unsigned long window_len = ras->ras_next_readahead -
1570                                    ras->ras_window_start;
1571         unsigned long left;
1572
1573         LASSERT(ras->ras_stride_length != 0);
1574
1575         left = window_len % ras->ras_stride_length;
1576
1577         ras->ras_stride_offset = ras->ras_next_readahead - left;
1578
1579         RAS_CDEBUG(ras);
1580 }
1581
1582 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1583                        struct ll_readahead_state *ras, unsigned long index,
1584                        unsigned hit)
1585 {
1586         struct ll_ra_info *ra = &sbi->ll_ra_info;
1587         int zero = 0, stride_zero = 0, stride_detect = 0, ra_miss = 0;
1588         ENTRY;
1589
1590         spin_lock(&sbi->ll_lock);
1591         spin_lock(&ras->ras_lock);
1592
1593         ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
1594
1595         /* reset the read-ahead window in two cases.  First when the app seeks
1596          * or reads to some other part of the file.  Secondly if we get a
1597          * read-ahead miss that we think we've previously issued.  This can
1598          * be a symptom of there being so many read-ahead pages that the VM is
1599          * reclaiming it before we get to it. */
1600         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1601                 zero = 1;
1602                 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
1603                 /* check whether it is in stride I/O mode*/
1604                 if (!index_in_stride_window(index, ras, inode))
1605                         stride_zero = 1;
1606         } else if (!hit && ras->ras_window_len &&
1607                    index < ras->ras_next_readahead &&
1608                    index_in_window(index, ras->ras_window_start, 0,
1609                                    ras->ras_window_len)) {
1610                 zero = 1;
1611                 ra_miss = 1;
1612                 /* If it hits read-ahead miss and the stride I/O is still
1613                  * not detected, reset stride stuff to re-detect the whole
1614                  * stride I/O mode to avoid complication */
1615                 if (!stride_io_mode(ras))
1616                         stride_zero = 1;
1617                 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1618         }
1619
1620         /* On the second access to a file smaller than the tunable
1621          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1622          * file up to ra_max_pages.  This is simply a best effort and
1623          * only occurs once per open file.  Normal RA behavior is reverted
1624          * to for subsequent IO.  The mmap case does not increment
1625          * ras_requests and thus can never trigger this behavior. */
1626         if (ras->ras_requests == 2 && !ras->ras_request_index) {
1627                 __u64 kms_pages;
1628
1629                 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1630                             CFS_PAGE_SHIFT;
1631
1632                 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1633                        ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages);
1634
1635                 if (kms_pages &&
1636                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1637                         ras->ras_window_start = 0;
1638                         ras->ras_last_readpage = 0;
1639                         ras->ras_next_readahead = 0;
1640                         ras->ras_window_len = min(ra->ra_max_pages,
1641                                 ra->ra_max_read_ahead_whole_pages);
1642                         GOTO(out_unlock, 0);
1643                 }
1644         }
1645
1646         if (zero) {
1647                 /* If it is discontinuous read, check
1648                  * whether it is stride I/O mode*/
1649                 if (stride_zero) {
1650                         ras_reset(ras, index);
1651                         ras->ras_consecutive_pages++;
1652                         ras_stride_reset(ras);
1653                         RAS_CDEBUG(ras);
1654                         GOTO(out_unlock, 0);
1655                 } else {
1656                         /* The read is still in stride window or
1657                          * it hits read-ahead miss */
1658
1659                         /* If ra-window miss is hitted, which probably means VM
1660                          * pressure, and some read-ahead pages were reclaimed.So
1661                          * the length of ra-window will not increased, but also
1662                          * not reset to avoid redetecting the stride I/O mode.*/
1663                         ras->ras_consecutive_requests = 0;
1664                         if (!ra_miss) {
1665                                 ras->ras_consecutive_pages = 0;
1666                                 if (++ras->ras_consecutive_stride_requests > 1)
1667                                         stride_detect = 1;
1668                         }
1669                         RAS_CDEBUG(ras);
1670                 }
1671         } else if (ras->ras_consecutive_stride_requests > 1) {
1672                 /* If this is contiguous read but in stride I/O mode
1673                  * currently, check whether stride step still is valid,
1674                  * if invalid, it will reset the stride ra window*/
1675                 if (ras->ras_consecutive_pages + 1 > ras->ras_stride_pages)
1676                         ras_stride_reset(ras);
1677         }
1678
1679         ras->ras_last_readpage = index;
1680         ras->ras_consecutive_pages++;
1681         ras_set_start(ras, index);
1682         ras->ras_next_readahead = max(ras->ras_window_start,
1683                                       ras->ras_next_readahead);
1684         RAS_CDEBUG(ras);
1685
1686         /* Trigger RA in the mmap case where ras_consecutive_requests
1687          * is not incremented and thus can't be used to trigger RA */
1688         if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1689                 ras->ras_window_len = RAS_INCREASE_STEP;
1690                 GOTO(out_unlock, 0);
1691         }
1692
1693         /* Initially reset the stride window offset to next_readahead*/
1694         if (ras->ras_consecutive_stride_requests == 2 && stride_detect)
1695                 ras_set_stride_offset(ras);
1696
1697         /* The initial ras_window_len is set to the request size.  To avoid
1698          * uselessly reading and discarding pages for random IO the window is
1699          * only increased once per consecutive request received. */
1700         if ((ras->ras_consecutive_requests > 1 &&
1701             !ras->ras_request_index) || stride_detect) {
1702                 if (stride_io_mode(ras))
1703                         ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP);
1704                 else
1705                         ras->ras_window_len = min(ras->ras_window_len +
1706                                                   RAS_INCREASE_STEP,
1707                                                   ra->ra_max_pages);
1708         }
1709         EXIT;
1710 out_unlock:
1711         RAS_CDEBUG(ras);
1712         ras->ras_request_index++;
1713         spin_unlock(&ras->ras_lock);
1714         spin_unlock(&sbi->ll_lock);
1715         return;
1716 }
1717
1718 int ll_writepage(struct page *page)
1719 {
1720         struct inode *inode = page->mapping->host;
1721         struct ll_inode_info *lli = ll_i2info(inode);
1722         struct obd_export *exp;
1723         struct ll_async_page *llap;
1724         int rc = 0;
1725         ENTRY;
1726
1727         LASSERT(PageLocked(page));
1728
1729         exp = ll_i2dtexp(inode);
1730         if (exp == NULL)
1731                 GOTO(out, rc = -EINVAL);
1732
1733         llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
1734         if (IS_ERR(llap))
1735                 GOTO(out, rc = PTR_ERR(llap));
1736
1737         LASSERT(!PageWriteback(page));
1738         set_page_writeback(page);
1739
1740         page_cache_get(page);
1741         if (llap->llap_write_queued) {
1742                 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
1743                 rc = obd_set_async_flags(exp, lli->lli_smd, NULL,
1744                                          llap->llap_cookie,
1745                                          ASYNC_READY | ASYNC_URGENT);
1746         } else {
1747                 rc = queue_or_sync_write(exp, inode, llap, CFS_PAGE_SIZE,
1748                                          ASYNC_READY | ASYNC_URGENT);
1749         }
1750         if (rc)
1751                 page_cache_release(page);
1752 out:
1753         if (rc) {
1754                 if (!lli->lli_async_rc)
1755                         lli->lli_async_rc = rc;
1756                 /* re-dirty page on error so it retries write */
1757                 if (PageWriteback(page)) {
1758                         end_page_writeback(page);
1759                 }
1760                 /* resend page only for not started IO*/
1761                 if (!PageError(page))
1762                         ll_redirty_page(page);
1763                 unlock_page(page);
1764         }
1765         RETURN(rc);
1766 }
1767
1768 /*
1769  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
1770  * read-ahead assumes it is valid to issue readpage all the way up to
1771  * i_size, but our dlm locks make that not the case.  We disable the
1772  * kernel's read-ahead and do our own by walking ahead in the page cache
1773  * checking for dlm lock coverage.  the main difference between 2.4 and
1774  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1775  * so they look the same.
1776  */
1777 int ll_readpage(struct file *filp, struct page *page)
1778 {
1779         struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
1780         struct inode *inode = page->mapping->host;
1781         struct obd_export *exp;
1782         struct ll_async_page *llap;
1783         struct obd_io_group *oig = NULL;
1784         int rc;
1785         ENTRY;
1786
1787         LASSERT(PageLocked(page));
1788         LASSERT(!PageUptodate(page));
1789         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset=%Lu=%#Lx\n",
1790                inode->i_ino, inode->i_generation, inode,
1791                (((loff_t)page->index) << CFS_PAGE_SHIFT),
1792                (((loff_t)page->index) << CFS_PAGE_SHIFT));
1793         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1794
1795         if (!ll_i2info(inode)->lli_smd) {
1796                 /* File with no objects - one big hole */
1797                 /* We use this just for remove_from_page_cache that is not
1798                  * exported, we'd make page back up to date. */
1799                 ll_truncate_complete_page(page);
1800                 clear_page(kmap(page));
1801                 kunmap(page);
1802                 SetPageUptodate(page);
1803                 unlock_page(page);
1804                 RETURN(0);
1805         }
1806
1807         rc = oig_init(&oig);
1808         if (rc < 0)
1809                 GOTO(out, rc);
1810
1811         exp = ll_i2dtexp(inode);
1812         if (exp == NULL)
1813                 GOTO(out, rc = -EINVAL);
1814
1815         llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1816         if (IS_ERR(llap))
1817                 GOTO(out, rc = PTR_ERR(llap));
1818
1819         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1820                 ras_update(ll_i2sbi(inode), inode, &fd->fd_ras, page->index,
1821                            llap->llap_defer_uptodate);
1822
1823
1824         if (llap->llap_defer_uptodate) {
1825                 /* This is the callpath if we got the page from a readahead */
1826                 llap->llap_ra_used = 1;
1827                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1828                                   fd->fd_flags);
1829                 if (rc > 0)
1830                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1831                                              NULL, oig);
1832                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1833                 SetPageUptodate(page);
1834                 unlock_page(page);
1835                 GOTO(out_oig, rc = 0);
1836         }
1837
1838         if (likely((fd->fd_flags & LL_FILE_IGNORE_LOCK) == 0)) {
1839                 rc = ll_page_matches(page, fd->fd_flags);
1840                 if (rc < 0) {
1841                         LL_CDEBUG_PAGE(D_ERROR, page,
1842                                        "lock match failed: rc %d\n", rc);
1843                         GOTO(out, rc);
1844                 }
1845
1846                 if (rc == 0) {
1847                         CWARN("ino %lu page %lu (%llu) not covered by "
1848                               "a lock (mmap?).  check debug logs.\n",
1849                               inode->i_ino, page->index,
1850                               (long long)page->index << CFS_PAGE_SHIFT);
1851                 }
1852         }
1853
1854         rc = ll_issue_page_read(exp, llap, oig, 0);
1855         if (rc)
1856                 GOTO(out, rc);
1857
1858         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1859         /* We have just requested the actual page we want, see if we can tack
1860          * on some readahead to that page's RPC before it is sent. */
1861         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1862                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1863                              fd->fd_flags);
1864
1865         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1866
1867 out:
1868         if (rc)
1869                 unlock_page(page);
1870 out_oig:
1871         if (oig != NULL)
1872                 oig_release(oig);
1873         RETURN(rc);
1874 }