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 #ifndef AUTOCONF_INCLUDED
  24 #include <linux/config.h>
  25 #endif
  26 #include <linux/kernel.h>
  27 #include <linux/mm.h>
  28 #include <linux/string.h>
  29 #include <linux/stat.h>
  30 #include <linux/errno.h>
  31 #include <linux/smp_lock.h>
  32 #include <linux/unistd.h>
  33 #include <linux/version.h>
  34 #include <asm/system.h>
  35 #include <asm/uaccess.h>
  36
  37 #include <linux/fs.h>
  38 #include <linux/stat.h>
  39 #include <asm/uaccess.h>
  40 #include <linux/mm.h>
  41 #include <linux/pagemap.h>
  42 #include <linux/smp_lock.h>
  43
  44 #define DEBUG_SUBSYSTEM S_LLITE
  45
  46 //#include <lustre_mdc.h>
  47 #include <lustre_lite.h>
  48 #include "llite_internal.h"
  49 #include <linux/lustre_compat25.h>
  50
  51 #ifndef list_for_each_prev_safe
  52 #define list_for_each_prev_safe(pos, n, head) \
  53         for (pos = (head)->prev, n = pos->prev; pos != (head); \
  54                 pos = n, n = pos->prev )
  55 #endif
  56
  57 cfs_mem_cache_t *ll_async_page_slab = NULL;
  58 size_t ll_async_page_slab_size = 0;
  59
  60 /* SYNCHRONOUS I/O to object storage for an inode */
  61 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
  62                   struct page *page, int flags)
  63 {
  64         struct ll_inode_info *lli = ll_i2info(inode);
  65         struct lov_stripe_md *lsm = lli->lli_smd;
  66         struct obd_info oinfo = { { { 0 } } };
  67         struct brw_page pg;
  68         int opc, rc;
  69         ENTRY;
  70
  71         pg.pg = page;
  72         pg.off = ((obd_off)page->index) << CFS_PAGE_SHIFT;
  73
  74         if ((cmd & OBD_BRW_WRITE) && (pg.off+CFS_PAGE_SIZE>i_size_read(inode)))
  75                 pg.count = i_size_read(inode) % CFS_PAGE_SIZE;
  76         else
  77                 pg.count = CFS_PAGE_SIZE;
  78
  79         LL_CDEBUG_PAGE(D_PAGE, page, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
  80                        cmd & OBD_BRW_WRITE ? "write" : "read", pg.count,
  81                        inode->i_ino, pg.off, pg.off);
  82         if (pg.count == 0) {
  83                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
  84                        LPU64"\n", inode->i_ino, inode, i_size_read(inode),
  85                        page->mapping->host, i_size_read(page->mapping->host),
  86                        page->index, pg.off);
  87         }
  88
  89         pg.flag = flags;
  90
  91         if (cmd & OBD_BRW_WRITE)
  92                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_WRITE,
  93                                    pg.count);
  94         else
  95                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_READ,
  96                                    pg.count);
  97         oinfo.oi_oa = oa;
  98         oinfo.oi_md = lsm;
  99         /* NB partial write, so we might not have CAPA_OPC_OSS_READ capa */
 100         opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
 101         oinfo.oi_capa = ll_osscapa_get(inode, opc);
 102         rc = obd_brw(cmd, ll_i2dtexp(inode), &oinfo, 1, &pg, NULL);
 103         capa_put(oinfo.oi_capa);
 104         if (rc == 0)
 105                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
 106         else if (rc != -EIO)
 107                 CERROR("error from obd_brw: rc = %d\n", rc);
 108         RETURN(rc);
 109 }
 110
 111 /* this isn't where truncate starts.   roughly:
 112  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
 113  * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
 114  * avoid races.
 115  *
 116  * must be called under ->lli_size_sem */
 117 void ll_truncate(struct inode *inode)
 118 {
 119         struct ll_inode_info *lli = ll_i2info(inode);
 120         struct obd_info oinfo = { { { 0 } } };
 121         struct ost_lvb lvb;
 122         struct obdo oa;
 123         int rc;
 124         ENTRY;
 125         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",inode->i_ino,
 126                inode->i_generation, inode, i_size_read(inode),
 127                i_size_read(inode));
 128
 129         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
 130         if (lli->lli_size_sem_owner != current) {
 131                 EXIT;
 132                 return;
 133         }
 134
 135         if (!lli->lli_smd) {
 136                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 137                        inode->i_ino);
 138                 GOTO(out_unlock, 0);
 139         }
 140
 141         LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
 142
 143         /* XXX I'm pretty sure this is a hack to paper over a more fundamental
 144          * race condition. */
 145         lov_stripe_lock(lli->lli_smd);
 146         inode_init_lvb(inode, &lvb);
 147         rc = obd_merge_lvb(ll_i2dtexp(inode), lli->lli_smd, &lvb, 0);
 148         if (lvb.lvb_size == i_size_read(inode) && rc == 0) {
 149                 CDEBUG(D_VFSTRACE, "skipping punch for obj "LPX64", %Lu=%#Lx\n",
 150                        lli->lli_smd->lsm_object_id, i_size_read(inode),
 151                        i_size_read(inode));
 152                 lov_stripe_unlock(lli->lli_smd);
 153                 GOTO(out_unlock, 0);
 154         }
 155
 156         obd_adjust_kms(ll_i2dtexp(inode), lli->lli_smd, i_size_read(inode), 1);
 157         lov_stripe_unlock(lli->lli_smd);
 158
 159         if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_CHECKSUM) &&
 160                      (i_size_read(inode) & ~CFS_PAGE_MASK))) {
 161                 /* If the truncate leaves behind a partial page, update its
 162                  * checksum. */
 163                 struct page *page = find_get_page(inode->i_mapping,
 164                                                   i_size_read(inode) >>
 165                                                   CFS_PAGE_SHIFT);
 166                 if (page != NULL) {
 167                         struct ll_async_page *llap = llap_cast_private(page);
 168                         if (llap != NULL) {
 169                                 char *kaddr = kmap_atomic(page, KM_USER0);
 170                                 llap->llap_checksum =
 171                                         init_checksum(OSC_DEFAULT_CKSUM);
 172                                 llap->llap_checksum =
 173                                         compute_checksum(llap->llap_checksum,
 174                                                          kaddr, CFS_PAGE_SIZE,
 175                                                          OSC_DEFAULT_CKSUM);
 176                                 kunmap_atomic(kaddr, KM_USER0);
 177                         }
 178                         page_cache_release(page);
 179                 }
 180         }
 181
 182         CDEBUG(D_INFO, "calling punch for "LPX64" (new size %Lu=%#Lx)\n",
 183                lli->lli_smd->lsm_object_id, i_size_read(inode), i_size_read(inode));
 184
 185         oinfo.oi_md = lli->lli_smd;
 186         oinfo.oi_policy.l_extent.start = i_size_read(inode);
 187         oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
 188         oinfo.oi_oa = &oa;
 189         oa.o_id = lli->lli_smd->lsm_object_id;
 190         oa.o_gr = lli->lli_smd->lsm_object_gr;
 191         oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 192
 193         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
 194                         OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 195                         OBD_MD_FLFID | OBD_MD_FLGENER);
 196
 197         ll_inode_size_unlock(inode, 0);
 198
 199         oinfo.oi_capa = ll_osscapa_get(inode, CAPA_OPC_OSS_TRUNC);
 200         rc = obd_punch_rqset(ll_i2dtexp(inode), &oinfo, NULL);
 201         ll_truncate_free_capa(oinfo.oi_capa);
 202         if (rc)
 203                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 204         else
 205                 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 206                               OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 207         EXIT;
 208         return;
 209
 210  out_unlock:
 211         ll_inode_size_unlock(inode, 0);
 212 } /* ll_truncate */
 213
 214 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 215                      unsigned to)
 216 {
 217         struct inode *inode = page->mapping->host;
 218         struct ll_inode_info *lli = ll_i2info(inode);
 219         struct lov_stripe_md *lsm = lli->lli_smd;
 220         obd_off offset = ((obd_off)page->index) << CFS_PAGE_SHIFT;
 221         struct obd_info oinfo = { { { 0 } } };
 222         struct brw_page pga;
 223         struct obdo oa;
 224         struct ost_lvb lvb;
 225         int rc = 0;
 226         ENTRY;
 227
 228         LASSERT(PageLocked(page));
 229         (void)llap_cast_private(page); /* assertion */
 230
 231         /* Check to see if we should return -EIO right away */
 232         pga.pg = page;
 233         pga.off = offset;
 234         pga.count = CFS_PAGE_SIZE;
 235         pga.flag = 0;
 236
 237         oa.o_mode = inode->i_mode;
 238         oa.o_id = lsm->lsm_object_id;
 239         oa.o_gr = lsm->lsm_object_gr;
 240         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
 241                      OBD_MD_FLTYPE | OBD_MD_FLGROUP;
 242         obdo_from_inode(&oa, inode, OBD_MD_FLFID | OBD_MD_FLGENER);
 243
 244         oinfo.oi_oa = &oa;
 245         oinfo.oi_md = lsm;
 246         rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode), &oinfo, 1, &pga, NULL);
 247         if (rc)
 248                 RETURN(rc);
 249
 250         if (PageUptodate(page)) {
 251                 LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n");
 252                 RETURN(0);
 253         }
 254
 255         /* We're completely overwriting an existing page, so _don't_ set it up
 256          * to date until commit_write */
 257         if (from == 0 && to == CFS_PAGE_SIZE) {
 258                 LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n");
 259                 POISON_PAGE(page, 0x11);
 260                 RETURN(0);
 261         }
 262
 263         /* If are writing to a new page, no need to read old data.  The extent
 264          * locking will have updated the KMS, and for our purposes here we can
 265          * treat it like i_size. */
 266         lov_stripe_lock(lsm);
 267         inode_init_lvb(inode, &lvb);
 268         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
 269         lov_stripe_unlock(lsm);
 270         if (lvb.lvb_size <= offset) {
 271                 char *kaddr = kmap_atomic(page, KM_USER0);
 272                 LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n",
 273                                lvb.lvb_size, offset);
 274                 memset(kaddr, 0, CFS_PAGE_SIZE);
 275                 kunmap_atomic(kaddr, KM_USER0);
 276                 GOTO(prepare_done, rc = 0);
 277         }
 278
 279         /* XXX could be an async ocp read.. read-ahead? */
 280         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 281         if (rc == 0) {
 282                 /* bug 1598: don't clobber blksize */
 283                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 284                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 285         }
 286
 287         EXIT;
 288  prepare_done:
 289         if (rc == 0)
 290                 SetPageUptodate(page);
 291
 292         return rc;
 293 }
 294
 295 static int ll_ap_make_ready(void *data, int cmd)
 296 {
 297         struct ll_async_page *llap;
 298         struct page *page;
 299         ENTRY;
 300
 301         llap = LLAP_FROM_COOKIE(data);
 302         page = llap->llap_page;
 303
 304         LASSERTF(!(cmd & OBD_BRW_READ), "cmd %x page %p ino %lu index %lu\n", cmd, page,
 305                  page->mapping->host->i_ino, page->index);
 306
 307         /* we're trying to write, but the page is locked.. come back later */
 308         if (TryLockPage(page))
 309                 RETURN(-EAGAIN);
 310
 311         LASSERT(!PageWriteback(page));
 312
 313         /* if we left PageDirty we might get another writepage call
 314          * in the future.  list walkers are bright enough
 315          * to check page dirty so we can leave it on whatever list
 316          * its on.  XXX also, we're called with the cli list so if
 317          * we got the page cache list we'd create a lock inversion
 318          * with the removepage path which gets the page lock then the
 319          * cli lock */
 320         LASSERTF(!PageWriteback(page),"cmd %x page %p ino %lu index %lu\n", cmd, page,
 321                  page->mapping->host->i_ino, page->index);
 322         clear_page_dirty_for_io(page);
 323
 324         /* This actually clears the dirty bit in the radix tree.*/
 325         set_page_writeback(page);
 326
 327         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 328         page_cache_get(page);
 329
 330         RETURN(0);
 331 }
 332
 333 /* We have two reasons for giving llite the opportunity to change the
 334  * write length of a given queued page as it builds the RPC containing
 335  * the page:
 336  *
 337  * 1) Further extending writes may have landed in the page cache
 338  *    since a partial write first queued this page requiring us
 339  *    to write more from the page cache.  (No further races are possible, since
 340  *    by the time this is called, the page is locked.)
 341  * 2) We might have raced with truncate and want to avoid performing
 342  *    write RPCs that are just going to be thrown away by the
 343  *    truncate's punch on the storage targets.
 344  *
 345  * The kms serves these purposes as it is set at both truncate and extending
 346  * writes.
 347  */
 348 static int ll_ap_refresh_count(void *data, int cmd)
 349 {
 350         struct ll_inode_info *lli;
 351         struct ll_async_page *llap;
 352         struct lov_stripe_md *lsm;
 353         struct page *page;
 354         struct inode *inode;
 355         struct ost_lvb lvb;
 356         __u64 kms;
 357         ENTRY;
 358
 359         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 360         LASSERT(cmd != OBD_BRW_READ);
 361
 362         llap = LLAP_FROM_COOKIE(data);
 363         page = llap->llap_page;
 364         inode = page->mapping->host;
 365         lli = ll_i2info(inode);
 366         lsm = lli->lli_smd;
 367
 368         lov_stripe_lock(lsm);
 369         inode_init_lvb(inode, &lvb);
 370         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
 371         kms = lvb.lvb_size;
 372         lov_stripe_unlock(lsm);
 373
 374         /* catch race with truncate */
 375         if (((__u64)page->index << CFS_PAGE_SHIFT) >= kms)
 376                 return 0;
 377
 378         /* catch sub-page write at end of file */
 379         if (((__u64)page->index << CFS_PAGE_SHIFT) + CFS_PAGE_SIZE > kms)
 380                 return kms % CFS_PAGE_SIZE;
 381
 382         return CFS_PAGE_SIZE;
 383 }
 384
 385 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 386 {
 387         struct lov_stripe_md *lsm;
 388         obd_flag valid_flags;
 389
 390         lsm = ll_i2info(inode)->lli_smd;
 391
 392         oa->o_id = lsm->lsm_object_id;
 393         oa->o_gr = lsm->lsm_object_gr;
 394         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 395         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 396         if (cmd & OBD_BRW_WRITE) {
 397                 oa->o_valid |= OBD_MD_FLEPOCH;
 398                 oa->o_easize = ll_i2info(inode)->lli_ioepoch;
 399
 400                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 401                         OBD_MD_FLUID | OBD_MD_FLGID |
 402                         OBD_MD_FLFID | OBD_MD_FLGENER;
 403         }
 404
 405         obdo_from_inode(oa, inode, valid_flags);
 406 }
 407
 408 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 409 {
 410         struct ll_async_page *llap;
 411         ENTRY;
 412
 413         llap = LLAP_FROM_COOKIE(data);
 414         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 415
 416         EXIT;
 417 }
 418
 419 static void ll_ap_update_obdo(void *data, int cmd, struct obdo *oa,
 420                               obd_valid valid)
 421 {
 422         struct ll_async_page *llap;
 423         ENTRY;
 424
 425         llap = LLAP_FROM_COOKIE(data);
 426         obdo_from_inode(oa, llap->llap_page->mapping->host, valid);
 427
 428         EXIT;
 429 }
 430
 431 static struct obd_capa *ll_ap_lookup_capa(void *data, int cmd)
 432 {
 433         struct ll_async_page *llap = LLAP_FROM_COOKIE(data);
 434         int opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
 435
 436         return ll_osscapa_get(llap->llap_page->mapping->host, opc);
 437 }
 438
 439 static struct obd_async_page_ops ll_async_page_ops = {
 440         .ap_make_ready =        ll_ap_make_ready,
 441         .ap_refresh_count =     ll_ap_refresh_count,
 442         .ap_fill_obdo =         ll_ap_fill_obdo,
 443         .ap_update_obdo =       ll_ap_update_obdo,
 444         .ap_completion =        ll_ap_completion,
 445         .ap_lookup_capa =       ll_ap_lookup_capa,
 446 };
 447
 448 struct ll_async_page *llap_cast_private(struct page *page)
 449 {
 450         struct ll_async_page *llap = (struct ll_async_page *)page_private(page);
 451
 452         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 453                  "page %p private %lu gave magic %d which != %d\n",
 454                  page, page_private(page), llap->llap_magic, LLAP_MAGIC);
 455
 456         return llap;
 457 }
 458
 459 /* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction.
 460  *
 461  * There is an llap attached onto every page in lustre, linked off @sbi.
 462  * We add an llap to the list so we don't lose our place during list walking.
 463  * If llaps in the list are being moved they will only move to the end
 464  * of the LRU, and we aren't terribly interested in those pages here (we
 465  * start at the beginning of the list where the least-used llaps are.
 466  */
 467 int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction)
 468 {
 469         struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a };
 470         unsigned long total, want, count = 0;
 471
 472         total = sbi->ll_async_page_count;
 473
 474         /* There can be a large number of llaps (600k or more in a large
 475          * memory machine) so the VM 1/6 shrink ratio is likely too much.
 476          * Since we are freeing pages also, we don't necessarily want to
 477          * shrink so much.  Limit to 40MB of pages + llaps per call. */
 478         if (shrink_fraction == 0)
 479                 want = sbi->ll_async_page_count - sbi->ll_async_page_max + 32;
 480         else
 481                 want = (total + shrink_fraction - 1) / shrink_fraction;
 482
 483         if (want > 40 << (20 - CFS_PAGE_SHIFT))
 484                 want = 40 << (20 - CFS_PAGE_SHIFT);
 485
 486         CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n",
 487                want, total, shrink_fraction);
 488
 489         spin_lock(&sbi->ll_lock);
 490         list_add(&dummy_llap.llap_pglist_item, &sbi->ll_pglist);
 491
 492         while (--total >= 0 && count < want) {
 493                 struct page *page;
 494                 int keep;
 495
 496                 if (unlikely(need_resched())) {
 497                         spin_unlock(&sbi->ll_lock);
 498                         cond_resched();
 499                         spin_lock(&sbi->ll_lock);
 500                 }
 501
 502                 llap = llite_pglist_next_llap(sbi,&dummy_llap.llap_pglist_item);
 503                 list_del_init(&dummy_llap.llap_pglist_item);
 504                 if (llap == NULL)
 505                         break;
 506
 507                 page = llap->llap_page;
 508                 LASSERT(page != NULL);
 509
 510                 list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item);
 511
 512                 /* Page needs/undergoing IO */
 513                 if (TryLockPage(page)) {
 514                         LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n");
 515                         continue;
 516                 }
 517
 518                keep = (llap->llap_write_queued || PageDirty(page) ||
 519                       PageWriteback(page) || (!PageUptodate(page) &&
 520                       llap->llap_origin != LLAP_ORIGIN_READAHEAD));
 521
 522                 LL_CDEBUG_PAGE(D_PAGE, page,"%s LRU page: %s%s%s%s%s origin %s\n",
 523                                keep ? "keep" : "drop",
 524                                llap->llap_write_queued ? "wq " : "",
 525                                PageDirty(page) ? "pd " : "",
 526                                PageUptodate(page) ? "" : "!pu ",
 527                                PageWriteback(page) ? "wb" : "",
 528                                llap->llap_defer_uptodate ? "" : "!du",
 529                                llap_origins[llap->llap_origin]);
 530
 531                 /* If page is dirty or undergoing IO don't discard it */
 532                 if (keep) {
 533                         unlock_page(page);
 534                         continue;
 535                 }
 536
 537                 page_cache_get(page);
 538                 spin_unlock(&sbi->ll_lock);
 539
 540                 if (page->mapping != NULL) {
 541                         ll_teardown_mmaps(page->mapping,
 542                                          (__u64)page->index << CFS_PAGE_SHIFT,
 543                                          ((__u64)page->index << CFS_PAGE_SHIFT)|
 544                                           ~CFS_PAGE_MASK);
 545                         if (!PageDirty(page) && !page_mapped(page)) {
 546                                 ll_ra_accounting(llap, page->mapping);
 547                                 ll_truncate_complete_page(page);
 548                                 ++count;
 549                         } else {
 550                                 LL_CDEBUG_PAGE(D_PAGE, page, "Not dropping page"
 551                                                              " because it is "
 552                                                              "%s\n",
 553                                                               PageDirty(page)?
 554                                                               "dirty":"mapped");
 555                         }
 556                 }
 557                 unlock_page(page);
 558                 page_cache_release(page);
 559
 560                 spin_lock(&sbi->ll_lock);
 561         }
 562         list_del(&dummy_llap.llap_pglist_item);
 563         spin_unlock(&sbi->ll_lock);
 564
 565         CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n",
 566                count, want, total);
 567
 568         return count;
 569 }
 570
 571 struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
 572 {
 573         struct ll_async_page *llap;
 574         struct obd_export *exp;
 575         struct inode *inode = page->mapping->host;
 576         struct ll_sb_info *sbi;
 577         int rc;
 578         ENTRY;
 579
 580         if (!inode) {
 581                 static int triggered;
 582
 583                 if (!triggered) {
 584                         LL_CDEBUG_PAGE(D_ERROR, page, "Bug 10047. Wrong anon "
 585                                        "page received\n");
 586                         libcfs_debug_dumpstack(NULL);
 587                         triggered = 1;
 588                 }
 589                 RETURN(ERR_PTR(-EINVAL));
 590         }
 591         sbi = ll_i2sbi(inode);
 592         LASSERT(ll_async_page_slab);
 593         LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
 594
 595         llap = llap_cast_private(page);
 596         if (llap != NULL) {
 597                 /* move to end of LRU list, except when page is just about to
 598                  * die */
 599                 if (origin != LLAP_ORIGIN_REMOVEPAGE) {
 600                         spin_lock(&sbi->ll_lock);
 601                         sbi->ll_pglist_gen++;
 602                         list_del_init(&llap->llap_pglist_item);
 603                         list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 604                         spin_unlock(&sbi->ll_lock);
 605                 }
 606                 GOTO(out, llap);
 607         }
 608
 609         exp = ll_i2dtexp(page->mapping->host);
 610         if (exp == NULL)
 611                 RETURN(ERR_PTR(-EINVAL));
 612
 613         /* limit the number of lustre-cached pages */
 614         if (sbi->ll_async_page_count >= sbi->ll_async_page_max)
 615                 llap_shrink_cache(sbi, 0);
 616
 617         OBD_SLAB_ALLOC(llap, ll_async_page_slab, CFS_ALLOC_STD,
 618                        ll_async_page_slab_size);
 619         if (llap == NULL)
 620                 RETURN(ERR_PTR(-ENOMEM));
 621         llap->llap_magic = LLAP_MAGIC;
 622         llap->llap_cookie = (void *)llap + size_round(sizeof(*llap));
 623
 624         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 625                                  (obd_off)page->index << CFS_PAGE_SHIFT,
 626                                  &ll_async_page_ops, llap, &llap->llap_cookie);
 627         if (rc) {
 628                 OBD_SLAB_FREE(llap, ll_async_page_slab,
 629                               ll_async_page_slab_size);
 630                 RETURN(ERR_PTR(rc));
 631         }
 632
 633         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 634                page, llap->llap_cookie, (obd_off)page->index << CFS_PAGE_SHIFT);
 635         /* also zeroing the PRIVBITS low order bitflags */
 636         __set_page_ll_data(page, llap);
 637         llap->llap_page = page;
 638         spin_lock(&sbi->ll_lock);
 639         sbi->ll_pglist_gen++;
 640         sbi->ll_async_page_count++;
 641         list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 642         INIT_LIST_HEAD(&llap->llap_pending_write);
 643         spin_unlock(&sbi->ll_lock);
 644
 645  out:
 646         if (unlikely(sbi->ll_flags & LL_SBI_CHECKSUM)) {
 647                 __u32 csum;
 648                 char *kaddr = kmap_atomic(page, KM_USER0);
 649                 csum = init_checksum(OSC_DEFAULT_CKSUM);
 650                 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
 651                                         OSC_DEFAULT_CKSUM);
 652                 kunmap_atomic(kaddr, KM_USER0);
 653                 if (origin == LLAP_ORIGIN_READAHEAD ||
 654                     origin == LLAP_ORIGIN_READPAGE) {
 655                         llap->llap_checksum = 0;
 656                 } else if (origin == LLAP_ORIGIN_COMMIT_WRITE ||
 657                            llap->llap_checksum == 0) {
 658                         llap->llap_checksum = csum;
 659                         CDEBUG(D_PAGE, "page %p cksum %x\n", page, csum);
 660                 } else if (llap->llap_checksum == csum) {
 661                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 662                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 663                                page, csum);
 664                 } else {
 665                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 666                         LL_CDEBUG_PAGE(D_ERROR, page, "old cksum %x != new "
 667                                        "%x!\n", llap->llap_checksum, csum);
 668                 }
 669         }
 670
 671         llap->llap_origin = origin;
 672         RETURN(llap);
 673 }
 674
 675 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
 676                                struct ll_async_page *llap,
 677                                unsigned to, obd_flag async_flags)
 678 {
 679         unsigned long size_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
 680         struct obd_io_group *oig;
 681         struct ll_sb_info *sbi = ll_i2sbi(inode);
 682         int rc, noquot = llap->llap_ignore_quota ? OBD_BRW_NOQUOTA : 0;
 683         ENTRY;
 684
 685         /* _make_ready only sees llap once we've unlocked the page */
 686         llap->llap_write_queued = 1;
 687         rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
 688                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 689                                 0, 0, 0, async_flags);
 690         if (rc == 0) {
 691                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
 692                 GOTO(out, 0);
 693         }
 694
 695         llap->llap_write_queued = 0;
 696         /* Do not pass llap here as it is sync write. */
 697         llap_write_pending(inode, NULL);
 698
 699         rc = oig_init(&oig);
 700         if (rc)
 701                 GOTO(out, rc);
 702
 703         /* make full-page requests if we are not at EOF (bug 4410) */
 704         if (to != CFS_PAGE_SIZE && llap->llap_page->index < size_index) {
 705                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 706                                "sync write before EOF: size_index %lu, to %d\n",
 707                                size_index, to);
 708                 to = CFS_PAGE_SIZE;
 709         } else if (to != CFS_PAGE_SIZE && llap->llap_page->index == size_index) {
 710                 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
 711                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 712                                "sync write at EOF: size_index %lu, to %d/%d\n",
 713                                size_index, to, size_to);
 714                 if (to < size_to)
 715                         to = size_to;
 716         }
 717
 718         /* compare the checksum once before the page leaves llite */
 719         if (unlikely((sbi->ll_flags & LL_SBI_CHECKSUM) &&
 720                      llap->llap_checksum != 0)) {
 721                 __u32 csum;
 722                 struct page *page = llap->llap_page;
 723                 char *kaddr = kmap_atomic(page, KM_USER0);
 724                 csum = init_checksum(OSC_DEFAULT_CKSUM);
 725                 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
 726                                         OSC_DEFAULT_CKSUM);
 727                 kunmap_atomic(kaddr, KM_USER0);
 728                 if (llap->llap_checksum == csum) {
 729                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 730                                page, csum);
 731                 } else {
 732                         CERROR("page %p old cksum %x != new cksum %x!\n",
 733                                page, llap->llap_checksum, csum);
 734                 }
 735         }
 736
 737         rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
 738                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 739                                 0, to, 0, ASYNC_READY | ASYNC_URGENT |
 740                                 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
 741         if (rc)
 742                 GOTO(free_oig, rc);
 743
 744         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
 745         if (rc)
 746                 GOTO(free_oig, rc);
 747
 748         rc = oig_wait(oig);
 749
 750         if (!rc && async_flags & ASYNC_READY) {
 751                 unlock_page(llap->llap_page);
 752                 if (PageWriteback(llap->llap_page)) {
 753                         end_page_writeback(llap->llap_page);
 754                 }
 755         }
 756
 757         if (rc == 0 && llap_write_complete(inode, llap))
 758                 ll_queue_done_writing(inode, 0);
 759
 760         LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
 761
 762 free_oig:
 763         oig_release(oig);
 764 out:
 765         RETURN(rc);
 766 }
 767
 768 /* update our write count to account for i_size increases that may have
 769  * happened since we've queued the page for io. */
 770
 771 /* be careful not to return success without setting the page Uptodate or
 772  * the next pass through prepare_write will read in stale data from disk. */
 773 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 774                     unsigned to)
 775 {
 776         struct inode *inode = page->mapping->host;
 777         struct ll_inode_info *lli = ll_i2info(inode);
 778         struct lov_stripe_md *lsm = lli->lli_smd;
 779         struct obd_export *exp;
 780         struct ll_async_page *llap;
 781         loff_t size;
 782         int rc = 0;
 783         ENTRY;
 784
 785         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 786         LASSERT(inode == file->f_dentry->d_inode);
 787         LASSERT(PageLocked(page));
 788
 789         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 790                inode, page, from, to, page->index);
 791
 792         llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
 793         if (IS_ERR(llap))
 794                 RETURN(PTR_ERR(llap));
 795
 796         exp = ll_i2dtexp(inode);
 797         if (exp == NULL)
 798                 RETURN(-EINVAL);
 799
 800         llap->llap_ignore_quota = capable(CAP_SYS_RESOURCE);
 801
 802         /*
 803          * queue a write for some time in the future the first time we
 804          * dirty the page.
 805          *
 806          * This is different from what other file systems do: they usually
 807          * just mark page (and some of its buffers) dirty and rely on
 808          * balance_dirty_pages() to start a write-back. Lustre wants write-back
 809          * to be started earlier for the following reasons:
 810          *
 811          *     (1) with a large number of clients we need to limit the amount
 812          *     of cached data on the clients a lot;
 813          *
 814          *     (2) large compute jobs generally want compute-only then io-only
 815          *     and the IO should complete as quickly as possible;
 816          *
 817          *     (3) IO is batched up to the RPC size and is async until the
 818          *     client max cache is hit
 819          *     (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
 820          *
 821          */
 822         if (!PageDirty(page)) {
 823                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_MISSES, 1);
 824
 825                 rc = queue_or_sync_write(exp, inode, llap, to, 0);
 826                 if (rc)
 827                         GOTO(out, rc);
 828         } else {
 829                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_HITS, 1);
 830         }
 831
 832         /* put the page in the page cache, from now on ll_removepage is
 833          * responsible for cleaning up the llap.
 834          * only set page dirty when it's queued to be write out */
 835         if (llap->llap_write_queued)
 836                 set_page_dirty(page);
 837
 838 out:
 839         size = (((obd_off)page->index) << CFS_PAGE_SHIFT) + to;
 840         ll_inode_size_lock(inode, 0);
 841         if (rc == 0) {
 842                 lov_stripe_lock(lsm);
 843                 obd_adjust_kms(exp, lsm, size, 0);
 844                 lov_stripe_unlock(lsm);
 845                 if (size > i_size_read(inode))
 846                         i_size_write(inode, size);
 847                 SetPageUptodate(page);
 848         } else if (size > i_size_read(inode)) {
 849                 /* this page beyond the pales of i_size, so it can't be
 850                  * truncated in ll_p_r_e during lock revoking. we must
 851                  * teardown our book-keeping here. */
 852                 ll_removepage(page);
 853         }
 854         ll_inode_size_unlock(inode, 0);
 855         RETURN(rc);
 856 }
 857
 858 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 859 {
 860         struct ll_ra_info *ra = &sbi->ll_ra_info;
 861         unsigned long ret;
 862         ENTRY;
 863
 864         spin_lock(&sbi->ll_lock);
 865         ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
 866         ra->ra_cur_pages += ret;
 867         spin_unlock(&sbi->ll_lock);
 868
 869         RETURN(ret);
 870 }
 871
 872 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 873 {
 874         struct ll_ra_info *ra = &sbi->ll_ra_info;
 875         spin_lock(&sbi->ll_lock);
 876         LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
 877                  ra->ra_cur_pages, len);
 878         ra->ra_cur_pages -= len;
 879         spin_unlock(&sbi->ll_lock);
 880 }
 881
 882 /* called for each page in a completed rpc.*/
 883 int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 884 {
 885         struct ll_async_page *llap;
 886         struct page *page;
 887         int ret = 0;
 888         ENTRY;
 889
 890         llap = LLAP_FROM_COOKIE(data);
 891         page = llap->llap_page;
 892         LASSERT(PageLocked(page));
 893         LASSERT(CheckWriteback(page,cmd));
 894
 895         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 896
 897         if (cmd & OBD_BRW_READ && llap->llap_defer_uptodate)
 898                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 899
 900         if (rc == 0)  {
 901                 if (cmd & OBD_BRW_READ) {
 902                         if (!llap->llap_defer_uptodate)
 903                                 SetPageUptodate(page);
 904                 } else {
 905                         llap->llap_write_queued = 0;
 906                 }
 907                 ClearPageError(page);
 908         } else {
 909                 if (cmd & OBD_BRW_READ) {
 910                         llap->llap_defer_uptodate = 0;
 911                 }
 912                 SetPageError(page);
 913                 if (rc == -ENOSPC)
 914                         set_bit(AS_ENOSPC, &page->mapping->flags);
 915                 else
 916                         set_bit(AS_EIO, &page->mapping->flags);
 917         }
 918
 919         unlock_page(page);
 920
 921         if (cmd & OBD_BRW_WRITE) {
 922                 /* Only rc == 0, write succeed, then this page could be deleted
 923                  * from the pending_writing list
 924                  */
 925                 if (rc == 0 && llap_write_complete(page->mapping->host, llap))
 926                         ll_queue_done_writing(page->mapping->host, 0);
 927         }
 928
 929         if (PageWriteback(page)) {
 930                 end_page_writeback(page);
 931         }
 932         page_cache_release(page);
 933
 934         RETURN(ret);
 935 }
 936
 937 /* the kernel calls us here when a page is unhashed from the page cache.
 938  * the page will be locked and the kernel is holding a spinlock, so
 939  * we need to be careful.  we're just tearing down our book-keeping
 940  * here. */
 941 void ll_removepage(struct page *page)
 942 {
 943         struct inode *inode = page->mapping->host;
 944         struct obd_export *exp;
 945         struct ll_async_page *llap;
 946         struct ll_sb_info *sbi = ll_i2sbi(inode);
 947         int rc;
 948         ENTRY;
 949
 950         LASSERT(!in_interrupt());
 951
 952         /* sync pages or failed read pages can leave pages in the page
 953          * cache that don't have our data associated with them anymore */
 954         if (page_private(page) == 0) {
 955                 EXIT;
 956                 return;
 957         }
 958
 959         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 960
 961         exp = ll_i2dtexp(inode);
 962         if (exp == NULL) {
 963                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 964                 EXIT;
 965                 return;
 966         }
 967
 968         llap = llap_from_page(page, LLAP_ORIGIN_REMOVEPAGE);
 969         if (IS_ERR(llap)) {
 970                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 971                        page->index, PTR_ERR(llap));
 972                 EXIT;
 973                 return;
 974         }
 975
 976         if (llap_write_complete(inode, llap))
 977                 ll_queue_done_writing(inode, 0);
 978
 979         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 980                                      llap->llap_cookie);
 981         if (rc != 0)
 982                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 983
 984         /* this unconditional free is only safe because the page lock
 985          * is providing exclusivity to memory pressure/truncate/writeback..*/
 986         __clear_page_ll_data(page);
 987
 988         spin_lock(&sbi->ll_lock);
 989         if (!list_empty(&llap->llap_pglist_item))
 990                 list_del_init(&llap->llap_pglist_item);
 991         sbi->ll_pglist_gen++;
 992         sbi->ll_async_page_count--;
 993         spin_unlock(&sbi->ll_lock);
 994         OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
 995         EXIT;
 996 }
 997
 998 static int ll_page_matches(struct page *page, int fd_flags)
 999 {
1000         struct lustre_handle match_lockh = {0};
1001         struct inode *inode = page->mapping->host;
1002         ldlm_policy_data_t page_extent;
1003         int flags, matches;
1004         ENTRY;
1005
1006         if (unlikely(fd_flags & LL_FILE_GROUP_LOCKED))
1007                 RETURN(1);
1008
1009         page_extent.l_extent.start = (__u64)page->index << CFS_PAGE_SHIFT;
1010         page_extent.l_extent.end =
1011                 page_extent.l_extent.start + CFS_PAGE_SIZE - 1;
1012         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
1013         if (!(fd_flags & LL_FILE_READAHEAD))
1014                 flags |= LDLM_FL_CBPENDING;
1015         matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
1016                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
1017                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
1018                             &match_lockh);
1019         RETURN(matches);
1020 }
1021
1022 static int ll_issue_page_read(struct obd_export *exp,
1023                               struct ll_async_page *llap,
1024                               struct obd_io_group *oig, int defer)
1025 {
1026         struct page *page = llap->llap_page;
1027         int rc;
1028
1029         page_cache_get(page);
1030         llap->llap_defer_uptodate = defer;
1031         llap->llap_ra_used = 0;
1032         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
1033                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
1034                                 CFS_PAGE_SIZE, 0, ASYNC_COUNT_STABLE |
1035                                                   ASYNC_READY | ASYNC_URGENT);
1036         if (rc) {
1037                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
1038                 page_cache_release(page);
1039         }
1040         RETURN(rc);
1041 }
1042
1043 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
1044 {
1045         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
1046         ra->ra_stats[which]++;
1047 }
1048
1049 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
1050 {
1051         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
1052         struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
1053
1054         spin_lock(&sbi->ll_lock);
1055         ll_ra_stats_inc_unlocked(ra, which);
1056         spin_unlock(&sbi->ll_lock);
1057 }
1058
1059 void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping)
1060 {
1061         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
1062                 return;
1063
1064         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
1065 }
1066
1067 #define RAS_CDEBUG(ras) \
1068         CDEBUG(D_READA,                                                      \
1069                "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu"    \
1070                "csr %lu sf %lu sp %lu sl %lu \n",                            \
1071                ras->ras_last_readpage, ras->ras_consecutive_requests,        \
1072                ras->ras_consecutive_pages, ras->ras_window_start,            \
1073                ras->ras_window_len, ras->ras_next_readahead,                 \
1074                ras->ras_requests, ras->ras_request_index,                    \
1075                ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
1076                ras->ras_stride_pages, ras->ras_stride_length)
1077
1078 static int index_in_window(unsigned long index, unsigned long point,
1079                            unsigned long before, unsigned long after)
1080 {
1081         unsigned long start = point - before, end = point + after;
1082
1083         if (start > point)
1084                start = 0;
1085         if (end < point)
1086                end = ~0;
1087
1088         return start <= index && index <= end;
1089 }
1090
1091 static struct ll_readahead_state *ll_ras_get(struct file *f)
1092 {
1093         struct ll_file_data       *fd;
1094
1095         fd = LUSTRE_FPRIVATE(f);
1096         return &fd->fd_ras;
1097 }
1098
1099 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
1100 {
1101         struct ll_readahead_state *ras;
1102
1103         ras = ll_ras_get(f);
1104
1105         spin_lock(&ras->ras_lock);
1106         ras->ras_requests++;
1107         ras->ras_request_index = 0;
1108         ras->ras_consecutive_requests++;
1109         rar->lrr_reader = current;
1110
1111         list_add(&rar->lrr_linkage, &ras->ras_read_beads);
1112         spin_unlock(&ras->ras_lock);
1113 }
1114
1115 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
1116 {
1117         struct ll_readahead_state *ras;
1118
1119         ras = ll_ras_get(f);
1120
1121         spin_lock(&ras->ras_lock);
1122         list_del_init(&rar->lrr_linkage);
1123         spin_unlock(&ras->ras_lock);
1124 }
1125
1126 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
1127 {
1128         struct ll_ra_read *scan;
1129
1130         list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
1131                 if (scan->lrr_reader == current)
1132                         return scan;
1133         }
1134         return NULL;
1135 }
1136
1137 struct ll_ra_read *ll_ra_read_get(struct file *f)
1138 {
1139         struct ll_readahead_state *ras;
1140         struct ll_ra_read         *bead;
1141
1142         ras = ll_ras_get(f);
1143
1144         spin_lock(&ras->ras_lock);
1145         bead = ll_ra_read_get_locked(ras);
1146         spin_unlock(&ras->ras_lock);
1147         return bead;
1148 }
1149
1150 static int ll_read_ahead_page(struct obd_export *exp, struct obd_io_group *oig,
1151                               int index, struct address_space *mapping)
1152 {
1153         struct ll_async_page *llap;
1154         struct page *page;
1155         unsigned int gfp_mask = 0;
1156         int rc = 0;
1157
1158         gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
1159 #ifdef __GFP_NOWARN
1160         gfp_mask |= __GFP_NOWARN;
1161 #endif
1162         page = grab_cache_page_nowait_gfp(mapping, index, gfp_mask);
1163         if (page == NULL) {
1164                 ll_ra_stats_inc(mapping, RA_STAT_FAILED_GRAB_PAGE);
1165                 CDEBUG(D_READA, "g_c_p_n failed\n");
1166                 return 0;
1167         }
1168
1169         /* Check if page was truncated or reclaimed */
1170         if (page->mapping != mapping) {
1171                 ll_ra_stats_inc(mapping, RA_STAT_WRONG_GRAB_PAGE);
1172                 CDEBUG(D_READA, "g_c_p_n returned invalid page\n");
1173                 GOTO(unlock_page, rc = 0);
1174         }
1175
1176         /* we do this first so that we can see the page in the /proc
1177          * accounting */
1178         llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
1179         if (IS_ERR(llap) || llap->llap_defer_uptodate) {
1180                 if (PTR_ERR(llap) == -ENOLCK) {
1181                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
1182                         CDEBUG(D_READA | D_PAGE,
1183                                "Adding page to cache failed index "
1184                                 "%d\n", index);
1185                                 CDEBUG(D_READA, "nolock page\n");
1186                                 GOTO(unlock_page, rc = -ENOLCK);
1187                 }
1188                 CDEBUG(D_READA, "read-ahead page\n");
1189                 GOTO(unlock_page, rc = 0);
1190         }
1191
1192         /* skip completed pages */
1193         if (Page_Uptodate(page))
1194                 GOTO(unlock_page, rc = 0);
1195
1196         /* bail out when we hit the end of the lock. */
1197         rc = ll_issue_page_read(exp, llap, oig, 1);
1198         if (rc == 0) {
1199                 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "started read-ahead\n");
1200                 rc = 1;
1201         } else {
1202 unlock_page:
1203                 unlock_page(page);
1204                 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "skipping read-ahead\n");
1205         }
1206         page_cache_release(page);
1207         return rc;
1208 }
1209
1210 /* ra_io_arg will be filled in the beginning of ll_readahead with
1211  * ras_lock, then the following ll_read_ahead_pages will read RA
1212  * pages according to this arg, all the items in this structure are
1213  * counted by page index.
1214  */
1215 struct ra_io_arg {
1216         unsigned long ria_start;  /* start offset of read-ahead*/
1217         unsigned long ria_end;    /* end offset of read-ahead*/
1218         /* If stride read pattern is detected, ria_stoff means where
1219          * stride read is started. Note: for normal read-ahead, the
1220          * value here is meaningless, and also it will not be accessed*/
1221         pgoff_t ria_stoff;
1222         /* ria_length and ria_pages are the length and pages length in the
1223          * stride I/O mode. And they will also be used to check whether
1224          * it is stride I/O read-ahead in the read-ahead pages*/
1225         unsigned long ria_length;
1226         unsigned long ria_pages;
1227 };
1228
1229 #define RIA_DEBUG(ria)                                                \
1230         CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n",       \
1231         ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
1232         ria->ria_pages)
1233
1234 #define RAS_INCREASE_STEP (1024 * 1024 >> CFS_PAGE_SHIFT)
1235
1236 static inline int stride_io_mode(struct ll_readahead_state *ras)
1237 {
1238         return ras->ras_consecutive_stride_requests > 1;
1239 }
1240
1241 /* The function calculates how much pages will be read in
1242  * [off, off + length], which will be read by stride I/O mode,
1243  * stride_offset = st_off, stride_lengh = st_len,
1244  * stride_pages = st_pgs
1245  */
1246 static unsigned long
1247 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
1248                 unsigned long off, unsigned length)
1249 {
1250         unsigned long cont_len = st_off > off ?  st_off - off : 0;
1251         unsigned long stride_len = length + off > st_off ?
1252                            length + off + 1 - st_off : 0;
1253         unsigned long left, pg_count;
1254
1255         if (st_len == 0 || length == 0)
1256                 return length;
1257
1258         left = do_div(stride_len, st_len);
1259         left = min(left, st_pgs);
1260
1261         pg_count = left + stride_len * st_pgs + cont_len;
1262
1263         LASSERT(pg_count >= left);
1264
1265         CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %u"
1266                "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
1267
1268         return pg_count;
1269 }
1270
1271 static int ria_page_count(struct ra_io_arg *ria)
1272 {
1273         __u64 length = ria->ria_end >= ria->ria_start ?
1274                        ria->ria_end - ria->ria_start + 1 : 0;
1275
1276         return stride_pg_count(ria->ria_stoff, ria->ria_length,
1277                                ria->ria_pages, ria->ria_start,
1278                                length);
1279 }
1280
1281 /*Check whether the index is in the defined ra-window */
1282 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
1283 {
1284         /* If ria_length == ria_pages, it means non-stride I/O mode,
1285          * idx should always inside read-ahead window in this case
1286          * For stride I/O mode, just check whether the idx is inside
1287          * the ria_pages. */
1288         return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
1289                (idx - ria->ria_stoff) % ria->ria_length < ria->ria_pages;
1290 }
1291
1292 static int ll_read_ahead_pages(struct obd_export *exp,
1293                                struct obd_io_group *oig,
1294                                struct ra_io_arg *ria,
1295                                unsigned long *reserved_pages,
1296                                struct address_space *mapping,
1297                                unsigned long *ra_end)
1298 {
1299         int rc, count = 0, stride_ria;
1300         unsigned long page_idx;
1301
1302         LASSERT(ria != NULL);
1303         RIA_DEBUG(ria);
1304
1305         stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
1306         for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
1307                         *reserved_pages > 0; page_idx++) {
1308                 if (ras_inside_ra_window(page_idx, ria)) {
1309                         /* If the page is inside the read-ahead window*/
1310                         rc = ll_read_ahead_page(exp, oig, page_idx, mapping);
1311                         if (rc == 1) {
1312                                 (*reserved_pages)--;
1313                                 count ++;
1314                         } else if (rc == -ENOLCK)
1315                                 break;
1316                 } else if (stride_ria) {
1317                         /* If it is not in the read-ahead window, and it is
1318                          * read-ahead mode, then check whether it should skip
1319                          * the stride gap */
1320                         pgoff_t offset;
1321                         /* FIXME: This assertion only is valid when it is for
1322                          * forward read-ahead, it will be fixed when backward
1323                          * read-ahead is implemented */
1324                         LASSERTF(page_idx > ria->ria_stoff, "since %lu in the"
1325                                 " gap of ra window,it should bigger than stride"
1326                                 " offset %lu \n", page_idx, ria->ria_stoff);
1327
1328                         offset = page_idx - ria->ria_stoff;
1329                         offset = offset % (ria->ria_length);
1330                         if (offset > ria->ria_pages) {
1331                                 page_idx += ria->ria_length - offset;
1332                                 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
1333                                        ria->ria_length - offset);
1334                                 continue;
1335                         }
1336                 }
1337         }
1338         *ra_end = page_idx;
1339         return count;
1340 }
1341
1342 static int ll_readahead(struct ll_readahead_state *ras,
1343                          struct obd_export *exp, struct address_space *mapping,
1344                          struct obd_io_group *oig, int flags)
1345 {
1346         unsigned long start = 0, end = 0, reserved;
1347         unsigned long ra_end, len;
1348         struct inode *inode;
1349         struct lov_stripe_md *lsm;
1350         struct ll_ra_read *bead;
1351         struct ost_lvb lvb;
1352         struct ra_io_arg ria = { 0 };
1353         int ret = 0;
1354         __u64 kms;
1355         ENTRY;
1356
1357         inode = mapping->host;
1358         lsm = ll_i2info(inode)->lli_smd;
1359
1360         lov_stripe_lock(lsm);
1361         inode_init_lvb(inode, &lvb);
1362         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
1363         kms = lvb.lvb_size;
1364         lov_stripe_unlock(lsm);
1365         if (kms == 0) {
1366                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
1367                 RETURN(0);
1368         }
1369
1370         spin_lock(&ras->ras_lock);
1371         bead = ll_ra_read_get_locked(ras);
1372         /* Enlarge the RA window to encompass the full read */
1373         if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
1374             bead->lrr_start + bead->lrr_count) {
1375                 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
1376                                       ras->ras_window_start;
1377         }
1378         /* Reserve a part of the read-ahead window that we'll be issuing */
1379         if (ras->ras_window_len) {
1380                 start = ras->ras_next_readahead;
1381                 end = ras->ras_window_start + ras->ras_window_len - 1;
1382         }
1383         if (end != 0) {
1384                 /* Truncate RA window to end of file */
1385                 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
1386                 ras->ras_next_readahead = max(end, end + 1);
1387                 RAS_CDEBUG(ras);
1388         }
1389         ria.ria_start = start;
1390         ria.ria_end = end;
1391         /* If stride I/O mode is detected, get stride window*/
1392         if (stride_io_mode(ras)) {
1393                 ria.ria_length = ras->ras_stride_length;
1394                 ria.ria_pages = ras->ras_stride_pages;
1395         }
1396         spin_unlock(&ras->ras_lock);
1397
1398         if (end == 0) {
1399                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
1400                 RETURN(0);
1401         }
1402         len = ria_page_count(&ria);
1403         if (len == 0)
1404                 RETURN(0);
1405
1406         reserved = ll_ra_count_get(ll_i2sbi(inode), len);
1407
1408         if (reserved < end - start + 1)
1409                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
1410
1411         CDEBUG(D_READA, "reserved page %lu \n", reserved);
1412
1413         ret = ll_read_ahead_pages(exp, oig, &ria, &reserved, mapping, &ra_end);
1414
1415         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
1416         if (reserved != 0)
1417                 ll_ra_count_put(ll_i2sbi(inode), reserved);
1418
1419         if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
1420                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
1421
1422         /* if we didn't get to the end of the region we reserved from
1423          * the ras we need to go back and update the ras so that the
1424          * next read-ahead tries from where we left off.  we only do so
1425          * if the region we failed to issue read-ahead on is still ahead
1426          * of the app and behind the next index to start read-ahead from */
1427         CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
1428                ra_end, end, ria.ria_end);
1429
1430         if (ra_end != (end + 1)) {
1431                 spin_lock(&ras->ras_lock);
1432                 if (ra_end < ras->ras_next_readahead &&
1433                     index_in_window(ra_end, ras->ras_window_start, 0,
1434                                     ras->ras_window_len)) {
1435                         ras->ras_next_readahead = ra_end;
1436                         RAS_CDEBUG(ras);
1437                 }
1438                 spin_unlock(&ras->ras_lock);
1439         }
1440
1441         RETURN(ret);
1442 }
1443
1444 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
1445 {
1446         ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1));
1447 }
1448
1449 /* called with the ras_lock held or from places where it doesn't matter */
1450 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1451 {
1452         ras->ras_last_readpage = index;
1453         ras->ras_consecutive_requests = 0;
1454         ras->ras_consecutive_pages = 0;
1455         ras->ras_window_len = 0;
1456         ras_set_start(ras, index);
1457         ras->ras_next_readahead = max(ras->ras_window_start, index);
1458
1459         RAS_CDEBUG(ras);
1460 }
1461
1462 /* called with the ras_lock held or from places where it doesn't matter */
1463 static void ras_stride_reset(struct ll_readahead_state *ras)
1464 {
1465         ras->ras_consecutive_stride_requests = 0;
1466         RAS_CDEBUG(ras);
1467 }
1468
1469 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1470 {
1471         spin_lock_init(&ras->ras_lock);
1472         ras_reset(ras, 0);
1473         ras->ras_requests = 0;
1474         INIT_LIST_HEAD(&ras->ras_read_beads);
1475 }
1476
1477 /* Check whether the read request is in the stride window.
1478  * If it is in the stride window, return 1, otherwise return 0.
1479  * and also update stride_gap and stride_pages.
1480  */
1481 static int index_in_stride_window(unsigned long index,
1482                                   struct ll_readahead_state *ras,
1483                                   struct inode *inode)
1484 {
1485         int stride_gap = index - ras->ras_last_readpage - 1;
1486
1487         LASSERT(stride_gap != 0);
1488
1489         if (ras->ras_consecutive_pages == 0)
1490                 return 0;
1491
1492         /*Otherwise check the stride by itself */
1493         if ((ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
1494             ras->ras_consecutive_pages == ras->ras_stride_pages)
1495                 return 1;
1496
1497         if (stride_gap >= 0) {
1498                 /*
1499                  * only set stride_pages, stride_length if
1500                  * it is forward reading ( stride_gap > 0)
1501                  */
1502                 ras->ras_stride_pages = ras->ras_consecutive_pages;
1503                 ras->ras_stride_length = stride_gap + ras->ras_consecutive_pages;
1504         } else {
1505                 /*
1506                  * If stride_gap < 0,(back_forward reading),
1507                  * reset the stride_pages/length.
1508                  * FIXME:back_ward stride I/O read.
1509                  *
1510                  */
1511                 ras->ras_stride_pages = 0;
1512                 ras->ras_stride_length = 0;
1513         }
1514         RAS_CDEBUG(ras);
1515
1516         return 0;
1517 }
1518
1519 static unsigned long
1520 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
1521 {
1522         return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
1523                                ras->ras_stride_pages, ras->ras_stride_offset,
1524                                len);
1525 }
1526
1527 /* Stride Read-ahead window will be increased inc_len according to
1528  * stride I/O pattern */
1529 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1530                                        struct ll_ra_info *ra,
1531                                        unsigned long inc_len)
1532 {
1533         unsigned long left, step, window_len;
1534         unsigned long stride_len;
1535
1536         LASSERT(ras->ras_stride_length > 0);
1537
1538         stride_len = ras->ras_window_start + ras->ras_window_len -
1539                      ras->ras_stride_offset;
1540
1541         LASSERTF(stride_len >= 0, "window_start %lu, window_len %lu"
1542                  " stride_offset %lu\n", ras->ras_window_start,
1543                  ras->ras_window_len, ras->ras_stride_offset);
1544
1545         left = stride_len % ras->ras_stride_length;
1546
1547         window_len = ras->ras_window_len - left;
1548
1549         if (left < ras->ras_stride_pages)
1550                 left += inc_len;
1551         else
1552                 left = ras->ras_stride_pages + inc_len;
1553
1554         LASSERT(ras->ras_stride_pages != 0);
1555
1556         step = left / ras->ras_stride_pages;
1557         left %= ras->ras_stride_pages;
1558
1559         window_len += step * ras->ras_stride_length + left;
1560
1561         if (stride_page_count(ras, window_len) <= ra->ra_max_pages)
1562                 ras->ras_window_len = window_len;
1563
1564         RAS_CDEBUG(ras);
1565 }
1566
1567 /* Set stride I/O read-ahead window start offset */
1568 static void ras_set_stride_offset(struct ll_readahead_state *ras)
1569 {
1570         unsigned long window_len = ras->ras_next_readahead -
1571                                    ras->ras_window_start;
1572         unsigned long left;
1573
1574         LASSERT(ras->ras_stride_length != 0);
1575
1576         left = window_len % ras->ras_stride_length;
1577
1578         ras->ras_stride_offset = ras->ras_next_readahead - left;
1579
1580         RAS_CDEBUG(ras);
1581 }
1582
1583 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1584                        struct ll_readahead_state *ras, unsigned long index,
1585                        unsigned hit)
1586 {
1587         struct ll_ra_info *ra = &sbi->ll_ra_info;
1588         int zero = 0, stride_zero = 0, stride_detect = 0, ra_miss = 0;
1589         ENTRY;
1590
1591         spin_lock(&sbi->ll_lock);
1592         spin_lock(&ras->ras_lock);
1593
1594         ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
1595
1596         /* reset the read-ahead window in two cases.  First when the app seeks
1597          * or reads to some other part of the file.  Secondly if we get a
1598          * read-ahead miss that we think we've previously issued.  This can
1599          * be a symptom of there being so many read-ahead pages that the VM is
1600          * reclaiming it before we get to it. */
1601         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1602                 zero = 1;
1603                 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
1604                 /* check whether it is in stride I/O mode*/
1605                 if (!index_in_stride_window(index, ras, inode))
1606                         stride_zero = 1;
1607         } else if (!hit && ras->ras_window_len &&
1608                    index < ras->ras_next_readahead &&
1609                    index_in_window(index, ras->ras_window_start, 0,
1610                                    ras->ras_window_len)) {
1611                 zero = 1;
1612                 ra_miss = 1;
1613                 /* If it hits read-ahead miss and the stride I/O is still
1614                  * not detected, reset stride stuff to re-detect the whole
1615                  * stride I/O mode to avoid complication */
1616                 if (!stride_io_mode(ras))
1617                         stride_zero = 1;
1618                 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1619         }
1620
1621         /* On the second access to a file smaller than the tunable
1622          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1623          * file up to ra_max_pages.  This is simply a best effort and
1624          * only occurs once per open file.  Normal RA behavior is reverted
1625          * to for subsequent IO.  The mmap case does not increment
1626          * ras_requests and thus can never trigger this behavior. */
1627         if (ras->ras_requests == 2 && !ras->ras_request_index) {
1628                 __u64 kms_pages;
1629
1630                 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1631                             CFS_PAGE_SHIFT;
1632
1633                 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1634                        ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages);
1635
1636                 if (kms_pages &&
1637                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1638                         ras->ras_window_start = 0;
1639                         ras->ras_last_readpage = 0;
1640                         ras->ras_next_readahead = 0;
1641                         ras->ras_window_len = min(ra->ra_max_pages,
1642                                 ra->ra_max_read_ahead_whole_pages);
1643                         GOTO(out_unlock, 0);
1644                 }
1645         }
1646
1647         if (zero) {
1648                 /* If it is discontinuous read, check
1649                  * whether it is stride I/O mode*/
1650                 if (stride_zero) {
1651                         ras_reset(ras, index);
1652                         ras->ras_consecutive_pages++;
1653                         ras_stride_reset(ras);
1654                         RAS_CDEBUG(ras);
1655                         GOTO(out_unlock, 0);
1656                 } else {
1657                         /* The read is still in stride window or
1658                          * it hits read-ahead miss */
1659
1660                         /* If ra-window miss is hitted, which probably means VM
1661                          * pressure, and some read-ahead pages were reclaimed.So
1662                          * the length of ra-window will not increased, but also
1663                          * not reset to avoid redetecting the stride I/O mode.*/
1664                         ras->ras_consecutive_requests = 0;
1665                         if (!ra_miss) {
1666                                 ras->ras_consecutive_pages = 0;
1667                                 if (++ras->ras_consecutive_stride_requests > 1)
1668                                         stride_detect = 1;
1669                         }
1670                         RAS_CDEBUG(ras);
1671                 }
1672         } else if (ras->ras_consecutive_stride_requests > 1) {
1673                 /* If this is contiguous read but in stride I/O mode
1674                  * currently, check whether stride step still is valid,
1675                  * if invalid, it will reset the stride ra window*/
1676                 if (ras->ras_consecutive_pages + 1 > ras->ras_stride_pages)
1677                         ras_stride_reset(ras);
1678         }
1679
1680         ras->ras_last_readpage = index;
1681         ras->ras_consecutive_pages++;
1682         ras_set_start(ras, index);
1683         ras->ras_next_readahead = max(ras->ras_window_start,
1684                                       ras->ras_next_readahead);
1685         RAS_CDEBUG(ras);
1686
1687         /* Trigger RA in the mmap case where ras_consecutive_requests
1688          * is not incremented and thus can't be used to trigger RA */
1689         if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1690                 ras->ras_window_len = RAS_INCREASE_STEP;
1691                 GOTO(out_unlock, 0);
1692         }
1693
1694         /* Initially reset the stride window offset to next_readahead*/
1695         if (ras->ras_consecutive_stride_requests == 2 && stride_detect)
1696                 ras_set_stride_offset(ras);
1697
1698         /* The initial ras_window_len is set to the request size.  To avoid
1699          * uselessly reading and discarding pages for random IO the window is
1700          * only increased once per consecutive request received. */
1701         if ((ras->ras_consecutive_requests > 1 &&
1702             !ras->ras_request_index) || stride_detect) {
1703                 if (stride_io_mode(ras))
1704                         ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP);
1705                 else
1706                         ras->ras_window_len = min(ras->ras_window_len +
1707                                                   RAS_INCREASE_STEP,
1708                                                   ra->ra_max_pages);
1709         }
1710         EXIT;
1711 out_unlock:
1712         RAS_CDEBUG(ras);
1713         ras->ras_request_index++;
1714         spin_unlock(&ras->ras_lock);
1715         spin_unlock(&sbi->ll_lock);
1716         return;
1717 }
1718
1719 int ll_writepage(struct page *page)
1720 {
1721         struct inode *inode = page->mapping->host;
1722         struct ll_inode_info *lli = ll_i2info(inode);
1723         struct obd_export *exp;
1724         struct ll_async_page *llap;
1725         int rc = 0;
1726         ENTRY;
1727
1728         LASSERT(PageLocked(page));
1729
1730         exp = ll_i2dtexp(inode);
1731         if (exp == NULL)
1732                 GOTO(out, rc = -EINVAL);
1733
1734         llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
1735         if (IS_ERR(llap))
1736                 GOTO(out, rc = PTR_ERR(llap));
1737
1738         LASSERT(!PageWriteback(page));
1739         set_page_writeback(page);
1740
1741         page_cache_get(page);
1742         if (llap->llap_write_queued) {
1743                 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
1744                 rc = obd_set_async_flags(exp, lli->lli_smd, NULL,
1745                                          llap->llap_cookie,
1746                                          ASYNC_READY | ASYNC_URGENT);
1747         } else {
1748                 rc = queue_or_sync_write(exp, inode, llap, CFS_PAGE_SIZE,
1749                                          ASYNC_READY | ASYNC_URGENT);
1750         }
1751         if (rc)
1752                 page_cache_release(page);
1753 out:
1754         if (rc) {
1755                 if (!lli->lli_async_rc)
1756                         lli->lli_async_rc = rc;
1757                 /* re-dirty page on error so it retries write */
1758                 if (PageWriteback(page)) {
1759                         end_page_writeback(page);
1760                 }
1761                 /* resend page only for not started IO*/
1762                 if (!PageError(page))
1763                         ll_redirty_page(page);
1764                 unlock_page(page);
1765         }
1766         RETURN(rc);
1767 }
1768
1769 /*
1770  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
1771  * read-ahead assumes it is valid to issue readpage all the way up to
1772  * i_size, but our dlm locks make that not the case.  We disable the
1773  * kernel's read-ahead and do our own by walking ahead in the page cache
1774  * checking for dlm lock coverage.  the main difference between 2.4 and
1775  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1776  * so they look the same.
1777  */
1778 int ll_readpage(struct file *filp, struct page *page)
1779 {
1780         struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
1781         struct inode *inode = page->mapping->host;
1782         struct obd_export *exp;
1783         struct ll_async_page *llap;
1784         struct obd_io_group *oig = NULL;
1785         int rc;
1786         ENTRY;
1787
1788         LASSERT(PageLocked(page));
1789         LASSERT(!PageUptodate(page));
1790         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset=%Lu=%#Lx\n",
1791                inode->i_ino, inode->i_generation, inode,
1792                (((loff_t)page->index) << CFS_PAGE_SHIFT),
1793                (((loff_t)page->index) << CFS_PAGE_SHIFT));
1794         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1795
1796         if (!ll_i2info(inode)->lli_smd) {
1797                 /* File with no objects - one big hole */
1798                 /* We use this just for remove_from_page_cache that is not
1799                  * exported, we'd make page back up to date. */
1800                 ll_truncate_complete_page(page);
1801                 clear_page(kmap(page));
1802                 kunmap(page);
1803                 SetPageUptodate(page);
1804                 unlock_page(page);
1805                 RETURN(0);
1806         }
1807
1808         rc = oig_init(&oig);
1809         if (rc < 0)
1810                 GOTO(out, rc);
1811
1812         exp = ll_i2dtexp(inode);
1813         if (exp == NULL)
1814                 GOTO(out, rc = -EINVAL);
1815
1816         llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1817         if (IS_ERR(llap))
1818                 GOTO(out, rc = PTR_ERR(llap));
1819
1820         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1821                 ras_update(ll_i2sbi(inode), inode, &fd->fd_ras, page->index,
1822                            llap->llap_defer_uptodate);
1823
1824
1825         if (llap->llap_defer_uptodate) {
1826                 /* This is the callpath if we got the page from a readahead */
1827                 llap->llap_ra_used = 1;
1828                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1829                                   fd->fd_flags);
1830                 if (rc > 0)
1831                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1832                                              NULL, oig);
1833                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1834                 SetPageUptodate(page);
1835                 unlock_page(page);
1836                 GOTO(out_oig, rc = 0);
1837         }
1838
1839         if (likely((fd->fd_flags & LL_FILE_IGNORE_LOCK) == 0)) {
1840                 rc = ll_page_matches(page, fd->fd_flags);
1841                 if (rc < 0) {
1842                         LL_CDEBUG_PAGE(D_ERROR, page,
1843                                        "lock match failed: rc %d\n", rc);
1844                         GOTO(out, rc);
1845                 }
1846
1847                 if (rc == 0) {
1848                         CWARN("ino %lu page %lu (%llu) not covered by "
1849                               "a lock (mmap?).  check debug logs.\n",
1850                               inode->i_ino, page->index,
1851                               (long long)page->index << CFS_PAGE_SHIFT);
1852                 }
1853         }
1854
1855         rc = ll_issue_page_read(exp, llap, oig, 0);
1856         if (rc)
1857                 GOTO(out, rc);
1858
1859         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1860         /* We have just requested the actual page we want, see if we can tack
1861          * on some readahead to that page's RPC before it is sent. */
1862         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1863                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1864                              fd->fd_flags);
1865
1866         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1867
1868 out:
1869         if (rc)
1870                 unlock_page(page);
1871 out_oig:
1872         if (oig != NULL)
1873                 oig_release(oig);
1874         RETURN(rc);
1875 }