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
  24 #ifndef AUTOCONF_INCLUDED
  25 #include <linux/config.h>
  26 #endif
  27 #include <linux/kernel.h>
  28 #include <linux/mm.h>
  29 #include <linux/string.h>
  30 #include <linux/stat.h>
  31 #include <linux/errno.h>
  32 #include <linux/smp_lock.h>
  33 #include <linux/unistd.h>
  34 #include <linux/version.h>
  35 #include <asm/system.h>
  36 #include <asm/uaccess.h>
  37
  38 #include <linux/fs.h>
  39 #include <linux/stat.h>
  40 #include <asm/uaccess.h>
  41 #include <asm/segment.h>
  42 #include <linux/mm.h>
  43 #include <linux/pagemap.h>
  44 #include <linux/smp_lock.h>
  45
  46 #define DEBUG_SUBSYSTEM S_LLITE
  47
  48 //#include <lustre_mdc.h>
  49 #include <lustre_lite.h>
  50 #include "llite_internal.h"
  51 #include <linux/lustre_compat25.h>
  52
  53 #ifndef list_for_each_prev_safe
  54 #define list_for_each_prev_safe(pos, n, head) \
  55         for (pos = (head)->prev, n = pos->prev; pos != (head); \
  56                 pos = n, n = pos->prev )
  57 #endif
  58
  59 cfs_mem_cache_t *ll_async_page_slab = NULL;
  60 size_t ll_async_page_slab_size = 0;
  61
  62 /* SYNCHRONOUS I/O to object storage for an inode */
  63 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
  64                   struct page *page, int flags)
  65 {
  66         struct ll_inode_info *lli = ll_i2info(inode);
  67         struct lov_stripe_md *lsm = lli->lli_smd;
  68         struct obd_info oinfo = { { { 0 } } };
  69         struct brw_page pg;
  70         int opc, rc;
  71         ENTRY;
  72
  73         pg.pg = page;
  74         pg.off = ((obd_off)page->index) << CFS_PAGE_SHIFT;
  75
  76         if ((cmd & OBD_BRW_WRITE) && (pg.off + CFS_PAGE_SIZE > inode->i_size))
  77                 pg.count = inode->i_size % CFS_PAGE_SIZE;
  78         else
  79                 pg.count = CFS_PAGE_SIZE;
  80
  81         LL_CDEBUG_PAGE(D_PAGE, page, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
  82                        cmd & OBD_BRW_WRITE ? "write" : "read", pg.count,
  83                        inode->i_ino, pg.off, pg.off);
  84         if (pg.count == 0) {
  85                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
  86                        LPU64"\n",
  87                        inode->i_ino, inode, inode->i_size, page->mapping->host,
  88                        page->mapping->host->i_size, page->index, pg.off);
  89         }
  90
  91         pg.flag = flags;
  92
  93         if (cmd & OBD_BRW_WRITE)
  94                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_WRITE,
  95                                    pg.count);
  96         else
  97                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_READ,
  98                                    pg.count);
  99         oinfo.oi_oa = oa;
 100         oinfo.oi_md = lsm;
 101         /* NB partial write, so we might not have CAPA_OPC_OSS_READ capa */
 102         opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
 103         oinfo.oi_capa = ll_osscapa_get(inode, current->fsuid, opc);
 104         rc = obd_brw(cmd, ll_i2dtexp(inode), &oinfo, 1, &pg, NULL);
 105         capa_put(oinfo.oi_capa);
 106         if (rc == 0)
 107                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
 108         else if (rc != -EIO)
 109                 CERROR("error from obd_brw: rc = %d\n", rc);
 110         RETURN(rc);
 111 }
 112
 113 /* this isn't where truncate starts.   roughly:
 114  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
 115  * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
 116  * avoid races.
 117  *
 118  * must be called under ->lli_size_sem */
 119 void ll_truncate(struct inode *inode)
 120 {
 121         struct ll_inode_info *lli = ll_i2info(inode);
 122         struct obd_info oinfo = { { { 0 } } };
 123         struct ost_lvb lvb;
 124         struct obdo oa;
 125         int rc;
 126         ENTRY;
 127         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",inode->i_ino,
 128                inode->i_generation, inode, inode->i_size, inode->i_size);
 129
 130         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
 131         if (lli->lli_size_sem_owner != current) {
 132                 EXIT;
 133                 return;
 134         }
 135
 136         if (!lli->lli_smd) {
 137                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 138                        inode->i_ino);
 139                 GOTO(out_unlock, 0);
 140         }
 141
 142         LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
 143
 144         /* XXX I'm pretty sure this is a hack to paper over a more fundamental
 145          * race condition. */
 146         lov_stripe_lock(lli->lli_smd);
 147         inode_init_lvb(inode, &lvb);
 148         obd_merge_lvb(ll_i2dtexp(inode), lli->lli_smd, &lvb, 0);
 149         if (lvb.lvb_size == inode->i_size) {
 150                 CDEBUG(D_VFSTRACE, "skipping punch for obj "LPX64", %Lu=%#Lx\n",
 151                        lli->lli_smd->lsm_object_id, inode->i_size, inode->i_size);
 152                 lov_stripe_unlock(lli->lli_smd);
 153                 GOTO(out_unlock, 0);
 154         }
 155
 156         obd_adjust_kms(ll_i2dtexp(inode), lli->lli_smd, inode->i_size, 1);
 157         lov_stripe_unlock(lli->lli_smd);
 158
 159         if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_CHECKSUM) &&
 160                      (inode->i_size & ~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                                                   inode->i_size >> CFS_PAGE_SHIFT);
 165                 if (page != NULL) {
 166                         struct ll_async_page *llap = llap_cast_private(page);
 167                         if (llap != NULL) {
 168                                 llap->llap_checksum =
 169                                         crc32_le(0, kmap(page), CFS_PAGE_SIZE);
 170                                 kunmap(page);
 171                         }
 172                         page_cache_release(page);
 173                 }
 174         }
 175
 176         CDEBUG(D_INFO, "calling punch for "LPX64" (new size %Lu=%#Lx)\n",
 177                lli->lli_smd->lsm_object_id, inode->i_size, inode->i_size);
 178
 179         oinfo.oi_md = lli->lli_smd;
 180         oinfo.oi_policy.l_extent.start = inode->i_size;
 181         oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
 182         oinfo.oi_oa = &oa;
 183         oa.o_id = lli->lli_smd->lsm_object_id;
 184         oa.o_gr = lli->lli_smd->lsm_object_gr;
 185         oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 186
 187         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
 188                         OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 189                         OBD_MD_FLFID | OBD_MD_FLGENER);
 190
 191         ll_inode_size_unlock(inode, 0);
 192
 193         oinfo.oi_capa = ll_osscapa_get(inode, 0, CAPA_OPC_OSS_TRUNC);
 194         rc = obd_punch_rqset(ll_i2dtexp(inode), &oinfo, NULL);
 195         ll_truncate_free_capa(oinfo.oi_capa);
 196         if (rc)
 197                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 198         else
 199                 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 200                               OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 201         EXIT;
 202         return;
 203
 204  out_unlock:
 205         ll_inode_size_unlock(inode, 0);
 206 } /* ll_truncate */
 207
 208 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 209                      unsigned to)
 210 {
 211         struct inode *inode = page->mapping->host;
 212         struct ll_inode_info *lli = ll_i2info(inode);
 213         struct lov_stripe_md *lsm = lli->lli_smd;
 214         obd_off offset = ((obd_off)page->index) << CFS_PAGE_SHIFT;
 215         struct obd_info oinfo = { { { 0 } } };
 216         struct brw_page pga;
 217         struct obdo oa;
 218         struct ost_lvb lvb;
 219         int rc = 0;
 220         ENTRY;
 221
 222         LASSERT(PageLocked(page));
 223         (void)llap_cast_private(page); /* assertion */
 224
 225         /* Check to see if we should return -EIO right away */
 226         pga.pg = page;
 227         pga.off = offset;
 228         pga.count = CFS_PAGE_SIZE;
 229         pga.flag = 0;
 230
 231         oa.o_mode = inode->i_mode;
 232         oa.o_id = lsm->lsm_object_id;
 233         oa.o_gr = lsm->lsm_object_gr;
 234         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
 235                      OBD_MD_FLTYPE | OBD_MD_FLGROUP;
 236         obdo_from_inode(&oa, inode, OBD_MD_FLFID | OBD_MD_FLGENER);
 237
 238         oinfo.oi_oa = &oa;
 239         oinfo.oi_md = lsm;
 240         rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode), &oinfo, 1, &pga, NULL);
 241         if (rc)
 242                 RETURN(rc);
 243
 244         if (PageUptodate(page)) {
 245                 LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n");
 246                 RETURN(0);
 247         }
 248
 249         /* We're completely overwriting an existing page, so _don't_ set it up
 250          * to date until commit_write */
 251         if (from == 0 && to == CFS_PAGE_SIZE) {
 252                 LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n");
 253                 POISON_PAGE(page, 0x11);
 254                 RETURN(0);
 255         }
 256
 257         /* If are writing to a new page, no need to read old data.  The extent
 258          * locking will have updated the KMS, and for our purposes here we can
 259          * treat it like i_size. */
 260         lov_stripe_lock(lsm);
 261         inode_init_lvb(inode, &lvb);
 262         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
 263         lov_stripe_unlock(lsm);
 264         if (lvb.lvb_size <= offset) {
 265                 LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n",
 266                                lvb.lvb_size, offset);
 267                 memset(kmap(page), 0, CFS_PAGE_SIZE);
 268                 kunmap(page);
 269                 GOTO(prepare_done, rc = 0);
 270         }
 271
 272         /* XXX could be an async ocp read.. read-ahead? */
 273         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 274         if (rc == 0) {
 275                 /* bug 1598: don't clobber blksize */
 276                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 277                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 278         }
 279
 280         EXIT;
 281  prepare_done:
 282         if (rc == 0)
 283                 SetPageUptodate(page);
 284
 285         return rc;
 286 }
 287
 288 static int ll_ap_make_ready(void *data, int cmd)
 289 {
 290         struct ll_async_page *llap;
 291         struct page *page;
 292         ENTRY;
 293
 294         llap = LLAP_FROM_COOKIE(data);
 295         page = llap->llap_page;
 296
 297         LASSERTF(!(cmd & OBD_BRW_READ), "cmd %x page %p ino %lu index %lu\n", cmd, page,
 298                  page->mapping->host->i_ino, page->index);
 299
 300         /* we're trying to write, but the page is locked.. come back later */
 301         if (TryLockPage(page))
 302                 RETURN(-EAGAIN);
 303
 304         LASSERT(!PageWriteback(page));
 305
 306         /* if we left PageDirty we might get another writepage call
 307          * in the future.  list walkers are bright enough
 308          * to check page dirty so we can leave it on whatever list
 309          * its on.  XXX also, we're called with the cli list so if
 310          * we got the page cache list we'd create a lock inversion
 311          * with the removepage path which gets the page lock then the
 312          * cli lock */
 313 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,0)
 314         clear_page_dirty(page);
 315 #else
 316         LASSERTF(!PageWriteback(page),"cmd %x page %p ino %lu index %lu\n", cmd, page,
 317                  page->mapping->host->i_ino, page->index);
 318         clear_page_dirty_for_io(page);
 319
 320         /* This actually clears the dirty bit in the radix tree.*/
 321         set_page_writeback(page);
 322 #endif
 323
 324         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 325         page_cache_get(page);
 326
 327         RETURN(0);
 328 }
 329
 330 /* We have two reasons for giving llite the opportunity to change the
 331  * write length of a given queued page as it builds the RPC containing
 332  * the page:
 333  *
 334  * 1) Further extending writes may have landed in the page cache
 335  *    since a partial write first queued this page requiring us
 336  *    to write more from the page cache.  (No further races are possible, since
 337  *    by the time this is called, the page is locked.)
 338  * 2) We might have raced with truncate and want to avoid performing
 339  *    write RPCs that are just going to be thrown away by the
 340  *    truncate's punch on the storage targets.
 341  *
 342  * The kms serves these purposes as it is set at both truncate and extending
 343  * writes.
 344  */
 345 static int ll_ap_refresh_count(void *data, int cmd)
 346 {
 347         struct ll_inode_info *lli;
 348         struct ll_async_page *llap;
 349         struct lov_stripe_md *lsm;
 350         struct page *page;
 351         struct inode *inode;
 352         struct ost_lvb lvb;
 353         __u64 kms;
 354         ENTRY;
 355
 356         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 357         LASSERT(cmd != OBD_BRW_READ);
 358
 359         llap = LLAP_FROM_COOKIE(data);
 360         page = llap->llap_page;
 361         inode = page->mapping->host;
 362         lli = ll_i2info(inode);
 363         lsm = lli->lli_smd;
 364
 365         lov_stripe_lock(lsm);
 366         inode_init_lvb(inode, &lvb);
 367         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
 368         kms = lvb.lvb_size;
 369         lov_stripe_unlock(lsm);
 370
 371         /* catch race with truncate */
 372         if (((__u64)page->index << CFS_PAGE_SHIFT) >= kms)
 373                 return 0;
 374
 375         /* catch sub-page write at end of file */
 376         if (((__u64)page->index << CFS_PAGE_SHIFT) + CFS_PAGE_SIZE > kms)
 377                 return kms % CFS_PAGE_SIZE;
 378
 379         return CFS_PAGE_SIZE;
 380 }
 381
 382 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 383 {
 384         struct lov_stripe_md *lsm;
 385         obd_flag valid_flags;
 386
 387         lsm = ll_i2info(inode)->lli_smd;
 388
 389         oa->o_id = lsm->lsm_object_id;
 390         oa->o_gr = lsm->lsm_object_gr;
 391         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 392         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 393         if (cmd & OBD_BRW_WRITE) {
 394                 oa->o_valid |= OBD_MD_FLEPOCH;
 395                 oa->o_easize = ll_i2info(inode)->lli_ioepoch;
 396
 397                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 398                         OBD_MD_FLUID | OBD_MD_FLGID |
 399                         OBD_MD_FLFID | OBD_MD_FLGENER;
 400         }
 401
 402         obdo_from_inode(oa, inode, valid_flags);
 403 }
 404
 405 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 406 {
 407         struct ll_async_page *llap;
 408         ENTRY;
 409
 410         llap = LLAP_FROM_COOKIE(data);
 411         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 412
 413         EXIT;
 414 }
 415
 416 static void ll_ap_update_obdo(void *data, int cmd, struct obdo *oa,
 417                               obd_valid valid)
 418 {
 419         struct ll_async_page *llap;
 420         ENTRY;
 421
 422         llap = LLAP_FROM_COOKIE(data);
 423         obdo_from_inode(oa, llap->llap_page->mapping->host, valid);
 424
 425         EXIT;
 426 }
 427
 428 static struct obd_capa *ll_ap_lookup_capa(void *data, int cmd)
 429 {
 430         struct ll_async_page *llap = LLAP_FROM_COOKIE(data);
 431         int opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
 432
 433         return ll_osscapa_get(llap->llap_page->mapping->host, llap->llap_fsuid,
 434                               opc);
 435 }
 436
 437 static struct obd_async_page_ops ll_async_page_ops = {
 438         .ap_make_ready =        ll_ap_make_ready,
 439         .ap_refresh_count =     ll_ap_refresh_count,
 440         .ap_fill_obdo =         ll_ap_fill_obdo,
 441         .ap_update_obdo =       ll_ap_update_obdo,
 442         .ap_completion =        ll_ap_completion,
 443         .ap_lookup_capa =       ll_ap_lookup_capa,
 444 };
 445
 446 struct ll_async_page *llap_cast_private(struct page *page)
 447 {
 448         struct ll_async_page *llap = (struct ll_async_page *)page_private(page);
 449
 450         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 451                  "page %p private %lu gave magic %d which != %d\n",
 452                  page, page_private(page), llap->llap_magic, LLAP_MAGIC);
 453
 454         return llap;
 455 }
 456
 457 /* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction.
 458  *
 459  * There is an llap attached onto every page in lustre, linked off @sbi.
 460  * We add an llap to the list so we don't lose our place during list walking.
 461  * If llaps in the list are being moved they will only move to the end
 462  * of the LRU, and we aren't terribly interested in those pages here (we
 463  * start at the beginning of the list where the least-used llaps are.
 464  */
 465 int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction)
 466 {
 467         struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a };
 468         unsigned long total, want, count = 0;
 469
 470         total = sbi->ll_async_page_count;
 471
 472         /* There can be a large number of llaps (600k or more in a large
 473          * memory machine) so the VM 1/6 shrink ratio is likely too much.
 474          * Since we are freeing pages also, we don't necessarily want to
 475          * shrink so much.  Limit to 40MB of pages + llaps per call. */
 476         if (shrink_fraction == 0)
 477                 want = sbi->ll_async_page_count - sbi->ll_async_page_max + 32;
 478         else
 479                 want = (total + shrink_fraction - 1) / shrink_fraction;
 480
 481         if (want > 40 << (20 - CFS_PAGE_SHIFT))
 482                 want = 40 << (20 - CFS_PAGE_SHIFT);
 483
 484         CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n",
 485                want, total, shrink_fraction);
 486
 487         spin_lock(&sbi->ll_lock);
 488         list_add(&dummy_llap.llap_pglist_item, &sbi->ll_pglist);
 489
 490         while (--total >= 0 && count < want) {
 491                 struct page *page;
 492                 int keep;
 493
 494                 if (unlikely(need_resched())) {
 495                         spin_unlock(&sbi->ll_lock);
 496                         cond_resched();
 497                         spin_lock(&sbi->ll_lock);
 498                 }
 499
 500                 llap = llite_pglist_next_llap(sbi,&dummy_llap.llap_pglist_item);
 501                 list_del_init(&dummy_llap.llap_pglist_item);
 502                 if (llap == NULL)
 503                         break;
 504
 505                 page = llap->llap_page;
 506                 LASSERT(page != NULL);
 507
 508                 list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item);
 509
 510                 /* Page needs/undergoing IO */
 511                 if (TryLockPage(page)) {
 512                         LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n");
 513                         continue;
 514                 }
 515
 516                 if (llap->llap_write_queued || PageDirty(page) ||
 517                     (!PageUptodate(page) &&
 518                      llap->llap_origin != LLAP_ORIGIN_READAHEAD))
 519                         keep = 1;
 520                 else
 521                         keep = 0;
 522
 523                 LL_CDEBUG_PAGE(D_PAGE, page,"%s LRU page: %s%s%s%s origin %s\n",
 524                                keep ? "keep" : "drop",
 525                                llap->llap_write_queued ? "wq " : "",
 526                                PageDirty(page) ? "pd " : "",
 527                                PageUptodate(page) ? "" : "!pu ",
 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 = 0;
 648                 csum = crc32_le(csum, kmap(page), CFS_PAGE_SIZE);
 649                 kunmap(page);
 650                 if (origin == LLAP_ORIGIN_READAHEAD ||
 651                     origin == LLAP_ORIGIN_READPAGE) {
 652                         llap->llap_checksum = 0;
 653                 } else if (origin == LLAP_ORIGIN_COMMIT_WRITE ||
 654                            llap->llap_checksum == 0) {
 655                         llap->llap_checksum = csum;
 656                         CDEBUG(D_PAGE, "page %p cksum %x\n", page, csum);
 657                 } else if (llap->llap_checksum == csum) {
 658                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 659                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 660                                page, csum);
 661                 } else {
 662                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 663                         LL_CDEBUG_PAGE(D_ERROR, page, "old cksum %x != new "
 664                                        "%x!\n", llap->llap_checksum, csum);
 665                 }
 666         }
 667
 668         llap->llap_origin = origin;
 669         RETURN(llap);
 670 }
 671
 672 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
 673                                struct ll_async_page *llap,
 674                                unsigned to, obd_flag async_flags)
 675 {
 676         unsigned long size_index = inode->i_size >> CFS_PAGE_SHIFT;
 677         struct obd_io_group *oig;
 678         struct ll_sb_info *sbi = ll_i2sbi(inode);
 679         int rc, noquot = llap->llap_ignore_quota ? OBD_BRW_NOQUOTA : 0;
 680         ENTRY;
 681
 682         /* _make_ready only sees llap once we've unlocked the page */
 683         llap->llap_write_queued = 1;
 684         rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
 685                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 686                                 0, 0, 0, async_flags);
 687         if (rc == 0) {
 688                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
 689                 GOTO(out, 0);
 690         }
 691
 692         llap->llap_write_queued = 0;
 693         /* Do not pass llap here as it is sync write. */
 694         llap_write_pending(inode, NULL);
 695
 696         rc = oig_init(&oig);
 697         if (rc)
 698                 GOTO(out, rc);
 699
 700         /* make full-page requests if we are not at EOF (bug 4410) */
 701         if (to != CFS_PAGE_SIZE && llap->llap_page->index < size_index) {
 702                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 703                                "sync write before EOF: size_index %lu, to %d\n",
 704                                size_index, to);
 705                 to = CFS_PAGE_SIZE;
 706         } else if (to != CFS_PAGE_SIZE && llap->llap_page->index == size_index) {
 707                 int size_to = inode->i_size & ~CFS_PAGE_MASK;
 708                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 709                                "sync write at EOF: size_index %lu, to %d/%d\n",
 710                                size_index, to, size_to);
 711                 if (to < size_to)
 712                         to = size_to;
 713         }
 714
 715         /* compare the checksum once before the page leaves llite */
 716         if (unlikely((sbi->ll_flags & LL_SBI_CHECKSUM) &&
 717                      llap->llap_checksum != 0)) {
 718                 __u32 csum = 0;
 719                 struct page *page = llap->llap_page;
 720                 csum = crc32_le(csum, kmap(page), CFS_PAGE_SIZE);
 721                 kunmap(page);
 722                 if (llap->llap_checksum == csum) {
 723                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 724                                page, csum);
 725                 } else {
 726                         CERROR("page %p old cksum %x != new cksum %x!\n",
 727                                page, llap->llap_checksum, csum);
 728                 }
 729         }
 730
 731         rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
 732                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 733                                 0, to, 0, ASYNC_READY | ASYNC_URGENT |
 734                                 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
 735         if (rc)
 736                 GOTO(free_oig, rc);
 737
 738         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
 739         if (rc)
 740                 GOTO(free_oig, rc);
 741
 742         rc = oig_wait(oig);
 743
 744         if (!rc && async_flags & ASYNC_READY) {
 745                 unlock_page(llap->llap_page);
 746                 if (PageWriteback(llap->llap_page)) {
 747                         end_page_writeback(llap->llap_page);
 748                 }
 749         }
 750
 751         if (rc == 0 && llap_write_complete(inode, llap))
 752                 ll_queue_done_writing(inode, 0);
 753
 754         LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
 755
 756 free_oig:
 757         oig_release(oig);
 758 out:
 759         RETURN(rc);
 760 }
 761
 762 /* update our write count to account for i_size increases that may have
 763  * happened since we've queued the page for io. */
 764
 765 /* be careful not to return success without setting the page Uptodate or
 766  * the next pass through prepare_write will read in stale data from disk. */
 767 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 768                     unsigned to)
 769 {
 770         struct inode *inode = page->mapping->host;
 771         struct ll_inode_info *lli = ll_i2info(inode);
 772         struct lov_stripe_md *lsm = lli->lli_smd;
 773         struct obd_export *exp;
 774         struct ll_async_page *llap;
 775         loff_t size;
 776         int rc = 0;
 777         ENTRY;
 778
 779         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 780         LASSERT(inode == file->f_dentry->d_inode);
 781         LASSERT(PageLocked(page));
 782
 783         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 784                inode, page, from, to, page->index);
 785
 786         llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
 787         if (IS_ERR(llap))
 788                 RETURN(PTR_ERR(llap));
 789
 790         exp = ll_i2dtexp(inode);
 791         if (exp == NULL)
 792                 RETURN(-EINVAL);
 793
 794         llap->llap_ignore_quota = capable(CAP_SYS_RESOURCE);
 795
 796         /*
 797          * queue a write for some time in the future the first time we
 798          * dirty the page.
 799          *
 800          * This is different from what other file systems do: they usually
 801          * just mark page (and some of its buffers) dirty and rely on
 802          * balance_dirty_pages() to start a write-back. Lustre wants write-back
 803          * to be started earlier for the following reasons:
 804          *
 805          *     (1) with a large number of clients we need to limit the amount
 806          *     of cached data on the clients a lot;
 807          *
 808          *     (2) large compute jobs generally want compute-only then io-only
 809          *     and the IO should complete as quickly as possible;
 810          *
 811          *     (3) IO is batched up to the RPC size and is async until the
 812          *     client max cache is hit
 813          *     (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
 814          *
 815          */
 816         if (!PageDirty(page)) {
 817                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_MISSES, 1);
 818
 819                 rc = queue_or_sync_write(exp, inode, llap, to, 0);
 820                 if (rc)
 821                         GOTO(out, rc);
 822         } else {
 823                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_HITS, 1);
 824         }
 825
 826         /* put the page in the page cache, from now on ll_removepage is
 827          * responsible for cleaning up the llap.
 828          * only set page dirty when it's queued to be write out */
 829         if (llap->llap_write_queued)
 830                 set_page_dirty(page);
 831
 832 out:
 833         size = (((obd_off)page->index) << CFS_PAGE_SHIFT) + to;
 834         ll_inode_size_lock(inode, 0);
 835         if (rc == 0) {
 836                 lov_stripe_lock(lsm);
 837                 obd_adjust_kms(exp, lsm, size, 0);
 838                 lov_stripe_unlock(lsm);
 839                 if (size > inode->i_size)
 840                         inode->i_size = size;
 841                 SetPageUptodate(page);
 842         } else if (size > inode->i_size) {
 843                 /* this page beyond the pales of i_size, so it can't be
 844                  * truncated in ll_p_r_e during lock revoking. we must
 845                  * teardown our book-keeping here. */
 846                 ll_removepage(page);
 847         }
 848         ll_inode_size_unlock(inode, 0);
 849         RETURN(rc);
 850 }
 851
 852 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 853 {
 854         struct ll_ra_info *ra = &sbi->ll_ra_info;
 855         unsigned long ret;
 856         ENTRY;
 857
 858         spin_lock(&sbi->ll_lock);
 859         ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
 860         ra->ra_cur_pages += ret;
 861         spin_unlock(&sbi->ll_lock);
 862
 863         RETURN(ret);
 864 }
 865
 866 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 867 {
 868         struct ll_ra_info *ra = &sbi->ll_ra_info;
 869         spin_lock(&sbi->ll_lock);
 870         LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
 871                  ra->ra_cur_pages, len);
 872         ra->ra_cur_pages -= len;
 873         spin_unlock(&sbi->ll_lock);
 874 }
 875
 876 /* called for each page in a completed rpc.*/
 877 int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 878 {
 879         struct ll_async_page *llap;
 880         struct page *page;
 881         int ret = 0;
 882         ENTRY;
 883
 884         llap = LLAP_FROM_COOKIE(data);
 885         page = llap->llap_page;
 886         LASSERT(PageLocked(page));
 887         LASSERT(CheckWriteback(page,cmd));
 888
 889         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 890
 891         if (cmd & OBD_BRW_READ && llap->llap_defer_uptodate)
 892                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 893
 894         if (rc == 0)  {
 895                 if (cmd & OBD_BRW_READ) {
 896                         if (!llap->llap_defer_uptodate)
 897                                 SetPageUptodate(page);
 898                 } else {
 899                         llap->llap_write_queued = 0;
 900                 }
 901                 ClearPageError(page);
 902         } else {
 903                 if (cmd & OBD_BRW_READ) {
 904                         llap->llap_defer_uptodate = 0;
 905                 } else {
 906                         ll_redirty_page(page);
 907                         ret = 1;
 908                 }
 909                 SetPageError(page);
 910         }
 911
 912         unlock_page(page);
 913
 914         if (cmd & OBD_BRW_WRITE) {
 915                 /* Only rc == 0, write succeed, then this page could be deleted
 916                  * from the pending_writing list
 917                  */
 918                 if (rc == 0 && llap_write_complete(page->mapping->host, llap))
 919                         ll_queue_done_writing(page->mapping->host, 0);
 920         }
 921
 922         if (PageWriteback(page)) {
 923                 end_page_writeback(page);
 924         }
 925         page_cache_release(page);
 926
 927         RETURN(ret);
 928 }
 929
 930 /* the kernel calls us here when a page is unhashed from the page cache.
 931  * the page will be locked and the kernel is holding a spinlock, so
 932  * we need to be careful.  we're just tearing down our book-keeping
 933  * here. */
 934 void ll_removepage(struct page *page)
 935 {
 936         struct inode *inode = page->mapping->host;
 937         struct obd_export *exp;
 938         struct ll_async_page *llap;
 939         struct ll_sb_info *sbi = ll_i2sbi(inode);
 940         int rc;
 941         ENTRY;
 942
 943         LASSERT(!in_interrupt());
 944
 945         /* sync pages or failed read pages can leave pages in the page
 946          * cache that don't have our data associated with them anymore */
 947         if (page_private(page) == 0) {
 948                 EXIT;
 949                 return;
 950         }
 951
 952         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 953
 954         exp = ll_i2dtexp(inode);
 955         if (exp == NULL) {
 956                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 957                 EXIT;
 958                 return;
 959         }
 960
 961         llap = llap_from_page(page, LLAP_ORIGIN_REMOVEPAGE);
 962         if (IS_ERR(llap)) {
 963                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 964                        page->index, PTR_ERR(llap));
 965                 EXIT;
 966                 return;
 967         }
 968
 969         if (llap_write_complete(inode, llap))
 970                 ll_queue_done_writing(inode, 0);
 971
 972         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 973                                      llap->llap_cookie);
 974         if (rc != 0)
 975                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 976
 977         /* this unconditional free is only safe because the page lock
 978          * is providing exclusivity to memory pressure/truncate/writeback..*/
 979         __clear_page_ll_data(page);
 980
 981         spin_lock(&sbi->ll_lock);
 982         if (!list_empty(&llap->llap_pglist_item))
 983                 list_del_init(&llap->llap_pglist_item);
 984         sbi->ll_pglist_gen++;
 985         sbi->ll_async_page_count--;
 986         spin_unlock(&sbi->ll_lock);
 987         OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
 988         EXIT;
 989 }
 990
 991 static int ll_page_matches(struct page *page, int fd_flags)
 992 {
 993         struct lustre_handle match_lockh = {0};
 994         struct inode *inode = page->mapping->host;
 995         ldlm_policy_data_t page_extent;
 996         int flags, matches;
 997         ENTRY;
 998
 999         if (unlikely(fd_flags & LL_FILE_GROUP_LOCKED))
1000                 RETURN(1);
1001
1002         page_extent.l_extent.start = (__u64)page->index << CFS_PAGE_SHIFT;
1003         page_extent.l_extent.end =
1004                 page_extent.l_extent.start + CFS_PAGE_SIZE - 1;
1005         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
1006         if (!(fd_flags & LL_FILE_READAHEAD))
1007                 flags |= LDLM_FL_CBPENDING;
1008         matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
1009                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
1010                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
1011                             &match_lockh);
1012         RETURN(matches);
1013 }
1014
1015 static int ll_issue_page_read(struct obd_export *exp,
1016                               struct ll_async_page *llap,
1017                               struct obd_io_group *oig, int defer)
1018 {
1019         struct page *page = llap->llap_page;
1020         int rc;
1021
1022         page_cache_get(page);
1023         llap->llap_defer_uptodate = defer;
1024         llap->llap_ra_used = 0;
1025         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
1026                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
1027                                 CFS_PAGE_SIZE, 0, ASYNC_COUNT_STABLE |
1028                                                   ASYNC_READY | ASYNC_URGENT);
1029         if (rc) {
1030                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
1031                 page_cache_release(page);
1032         }
1033         RETURN(rc);
1034 }
1035
1036 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
1037 {
1038         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
1039         ra->ra_stats[which]++;
1040 }
1041
1042 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
1043 {
1044         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
1045         struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
1046
1047         spin_lock(&sbi->ll_lock);
1048         ll_ra_stats_inc_unlocked(ra, which);
1049         spin_unlock(&sbi->ll_lock);
1050 }
1051
1052 void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping)
1053 {
1054         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
1055                 return;
1056
1057         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
1058 }
1059
1060 #define RAS_CDEBUG(ras) \
1061         CDEBUG(D_READA,                                                      \
1062                "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu\n", \
1063                ras->ras_last_readpage, ras->ras_consecutive_requests,        \
1064                ras->ras_consecutive_pages, ras->ras_window_start,            \
1065                ras->ras_window_len, ras->ras_next_readahead,                 \
1066                ras->ras_requests, ras->ras_request_index);
1067
1068 static int index_in_window(unsigned long index, unsigned long point,
1069                            unsigned long before, unsigned long after)
1070 {
1071         unsigned long start = point - before, end = point + after;
1072
1073         if (start > point)
1074                start = 0;
1075         if (end < point)
1076                end = ~0;
1077
1078         return start <= index && index <= end;
1079 }
1080
1081 static struct ll_readahead_state *ll_ras_get(struct file *f)
1082 {
1083         struct ll_file_data       *fd;
1084
1085         fd = LUSTRE_FPRIVATE(f);
1086         return &fd->fd_ras;
1087 }
1088
1089 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
1090 {
1091         struct ll_readahead_state *ras;
1092
1093         ras = ll_ras_get(f);
1094
1095         spin_lock(&ras->ras_lock);
1096         ras->ras_requests++;
1097         ras->ras_request_index = 0;
1098         ras->ras_consecutive_requests++;
1099         rar->lrr_reader = current;
1100
1101         list_add(&rar->lrr_linkage, &ras->ras_read_beads);
1102         spin_unlock(&ras->ras_lock);
1103 }
1104
1105 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
1106 {
1107         struct ll_readahead_state *ras;
1108
1109         ras = ll_ras_get(f);
1110
1111         spin_lock(&ras->ras_lock);
1112         list_del_init(&rar->lrr_linkage);
1113         spin_unlock(&ras->ras_lock);
1114 }
1115
1116 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
1117 {
1118         struct ll_ra_read *scan;
1119
1120         list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
1121                 if (scan->lrr_reader == current)
1122                         return scan;
1123         }
1124         return NULL;
1125 }
1126
1127 struct ll_ra_read *ll_ra_read_get(struct file *f)
1128 {
1129         struct ll_readahead_state *ras;
1130         struct ll_ra_read         *bead;
1131
1132         ras = ll_ras_get(f);
1133
1134         spin_lock(&ras->ras_lock);
1135         bead = ll_ra_read_get_locked(ras);
1136         spin_unlock(&ras->ras_lock);
1137         return bead;
1138 }
1139
1140 static int ll_readahead(struct ll_readahead_state *ras,
1141                          struct obd_export *exp, struct address_space *mapping,
1142                          struct obd_io_group *oig, int flags)
1143 {
1144         unsigned long i, start = 0, end = 0, reserved;
1145         struct ll_async_page *llap;
1146         struct page *page;
1147         int rc, ret = 0, match_failed = 0;
1148         __u64 kms;
1149         unsigned int gfp_mask;
1150         struct inode *inode;
1151         struct lov_stripe_md *lsm;
1152         struct ll_ra_read *bead;
1153         struct ost_lvb lvb;
1154         ENTRY;
1155
1156         inode = mapping->host;
1157         lsm = ll_i2info(inode)->lli_smd;
1158
1159         lov_stripe_lock(lsm);
1160         inode_init_lvb(inode, &lvb);
1161         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
1162         kms = lvb.lvb_size;
1163         lov_stripe_unlock(lsm);
1164         if (kms == 0) {
1165                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
1166                 RETURN(0);
1167         }
1168
1169         spin_lock(&ras->ras_lock);
1170         bead = ll_ra_read_get_locked(ras);
1171         /* Enlarge the RA window to encompass the full read */
1172         if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
1173             bead->lrr_start + bead->lrr_count) {
1174                 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
1175                                       ras->ras_window_start;
1176         }
1177         /* Reserve a part of the read-ahead window that we'll be issuing */
1178         if (ras->ras_window_len) {
1179                 start = ras->ras_next_readahead;
1180                 end = ras->ras_window_start + ras->ras_window_len - 1;
1181         }
1182         if (end != 0) {
1183                 /* Truncate RA window to end of file */
1184                 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
1185                 ras->ras_next_readahead = max(end, end + 1);
1186                 RAS_CDEBUG(ras);
1187         }
1188         spin_unlock(&ras->ras_lock);
1189
1190         if (end == 0) {
1191                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
1192                 RETURN(0);
1193         }
1194
1195         reserved = ll_ra_count_get(ll_i2sbi(inode), end - start + 1);
1196         if (reserved < end - start + 1)
1197                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
1198
1199         gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
1200 #ifdef __GFP_NOWARN
1201         gfp_mask |= __GFP_NOWARN;
1202 #endif
1203
1204         for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
1205                 /* skip locked pages from previous readpage calls */
1206                 page = grab_cache_page_nowait_gfp(mapping, i, gfp_mask);
1207                 if (page == NULL) {
1208                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_GRAB_PAGE);
1209                         CDEBUG(D_READA, "g_c_p_n failed\n");
1210                         continue;
1211                 }
1212
1213                 /* Check if page was truncated or reclaimed */
1214                 if (page->mapping != mapping) {
1215                         ll_ra_stats_inc(mapping, RA_STAT_WRONG_GRAB_PAGE);
1216                         CDEBUG(D_READA, "g_c_p_n returned invalid page\n");
1217                         goto next_page;
1218                 }
1219
1220                 /* we do this first so that we can see the page in the /proc
1221                  * accounting */
1222                 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
1223                 if (IS_ERR(llap) || llap->llap_defer_uptodate)
1224                         goto next_page;
1225
1226                 /* skip completed pages */
1227                 if (Page_Uptodate(page))
1228                         goto next_page;
1229
1230                 /* bail when we hit the end of the lock. */
1231                 if ((rc = ll_page_matches(page, flags|LL_FILE_READAHEAD)) <= 0){
1232                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
1233                                        "lock match failed: rc %d\n", rc);
1234                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
1235                         match_failed = 1;
1236                         goto next_page;
1237                 }
1238
1239                 rc = ll_issue_page_read(exp, llap, oig, 1);
1240                 if (rc == 0) {
1241                         reserved--;
1242                         ret++;
1243                         LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
1244                                        "started read-ahead\n");
1245                 } else {
1246         next_page:
1247                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
1248                                        "skipping read-ahead\n");
1249
1250                         unlock_page(page);
1251                 }
1252                 page_cache_release(page);
1253         }
1254
1255         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
1256         if (reserved != 0)
1257                 ll_ra_count_put(ll_i2sbi(inode), reserved);
1258         if (i == end + 1 && end == (kms >> CFS_PAGE_SHIFT))
1259                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
1260
1261         /* if we didn't get to the end of the region we reserved from
1262          * the ras we need to go back and update the ras so that the
1263          * next read-ahead tries from where we left off.  we only do so
1264          * if the region we failed to issue read-ahead on is still ahead
1265          * of the app and behind the next index to start read-ahead from */
1266         if (i != end + 1) {
1267                 spin_lock(&ras->ras_lock);
1268                 if (i < ras->ras_next_readahead &&
1269                     index_in_window(i, ras->ras_window_start, 0,
1270                                     ras->ras_window_len)) {
1271                         ras->ras_next_readahead = i;
1272                         RAS_CDEBUG(ras);
1273                 }
1274                 spin_unlock(&ras->ras_lock);
1275         }
1276
1277         RETURN(ret);
1278 }
1279
1280 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
1281 {
1282         ras->ras_window_start = index & (~((1024 * 1024 >> CFS_PAGE_SHIFT) - 1));
1283 }
1284
1285 /* called with the ras_lock held or from places where it doesn't matter */
1286 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1287 {
1288         ras->ras_last_readpage = index;
1289         ras->ras_consecutive_requests = 0;
1290         ras->ras_consecutive_pages = 0;
1291         ras->ras_window_len = 0;
1292         ras_set_start(ras, index);
1293         ras->ras_next_readahead = max(ras->ras_window_start, index);
1294
1295         RAS_CDEBUG(ras);
1296 }
1297
1298 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1299 {
1300         spin_lock_init(&ras->ras_lock);
1301         ras_reset(ras, 0);
1302         ras->ras_requests = 0;
1303         INIT_LIST_HEAD(&ras->ras_read_beads);
1304 }
1305
1306 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1307                        struct ll_readahead_state *ras, unsigned long index,
1308                        unsigned hit)
1309 {
1310         struct ll_ra_info *ra = &sbi->ll_ra_info;
1311         int zero = 0;
1312         ENTRY;
1313
1314         spin_lock(&sbi->ll_lock);
1315         spin_lock(&ras->ras_lock);
1316
1317         ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
1318
1319         /* reset the read-ahead window in two cases.  First when the app seeks
1320          * or reads to some other part of the file.  Secondly if we get a
1321          * read-ahead miss that we think we've previously issued.  This can
1322          * be a symptom of there being so many read-ahead pages that the VM is
1323          * reclaiming it before we get to it. */
1324         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1325                 zero = 1;
1326                 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
1327         } else if (!hit && ras->ras_window_len &&
1328                    index < ras->ras_next_readahead &&
1329                    index_in_window(index, ras->ras_window_start, 0,
1330                                    ras->ras_window_len)) {
1331                 zero = 1;
1332                 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1333         }
1334
1335         /* On the second access to a file smaller than the tunable
1336          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1337          * file up to ra_max_pages.  This is simply a best effort and
1338          * only occurs once per open file.  Normal RA behavior is reverted
1339          * to for subsequent IO.  The mmap case does not increment
1340          * ras_requests and thus can never trigger this behavior. */
1341         if (ras->ras_requests == 2 && !ras->ras_request_index) {
1342                 __u64 kms_pages;
1343
1344                 kms_pages = (inode->i_size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
1345
1346                 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1347                        ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages);
1348
1349                 if (kms_pages &&
1350                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1351                         ras->ras_window_start = 0;
1352                         ras->ras_last_readpage = 0;
1353                         ras->ras_next_readahead = 0;
1354                         ras->ras_window_len = min(ra->ra_max_pages,
1355                                 ra->ra_max_read_ahead_whole_pages);
1356                         GOTO(out_unlock, 0);
1357                 }
1358         }
1359
1360         if (zero) {
1361                 ras_reset(ras, index);
1362                 GOTO(out_unlock, 0);
1363         }
1364
1365         ras->ras_last_readpage = index;
1366         ras->ras_consecutive_pages++;
1367         ras_set_start(ras, index);
1368         ras->ras_next_readahead = max(ras->ras_window_start,
1369                                       ras->ras_next_readahead);
1370
1371         /* Trigger RA in the mmap case where ras_consecutive_requests
1372          * is not incremented and thus can't be used to trigger RA */
1373         if (!ras->ras_window_len && ras->ras_consecutive_pages == 3) {
1374                 ras->ras_window_len = 1024 * 1024 >> CFS_PAGE_SHIFT;
1375                 GOTO(out_unlock, 0);
1376         }
1377
1378         /* The initial ras_window_len is set to the request size.  To avoid
1379          * uselessly reading and discarding pages for random IO the window is
1380          * only increased once per consecutive request received. */
1381         if (ras->ras_consecutive_requests > 1 && !ras->ras_request_index) {
1382                 ras->ras_window_len = min(ras->ras_window_len +
1383                                           (1024 * 1024 >> CFS_PAGE_SHIFT),
1384                                           ra->ra_max_pages);
1385         }
1386
1387         EXIT;
1388 out_unlock:
1389         RAS_CDEBUG(ras);
1390         ras->ras_request_index++;
1391         spin_unlock(&ras->ras_lock);
1392         spin_unlock(&sbi->ll_lock);
1393         return;
1394 }
1395
1396 int ll_writepage(struct page *page)
1397 {
1398         struct inode *inode = page->mapping->host;
1399         struct ll_inode_info *lli = ll_i2info(inode);
1400         struct obd_export *exp;
1401         struct ll_async_page *llap;
1402         int rc = 0;
1403         ENTRY;
1404
1405         LASSERT(!PageDirty(page));
1406         LASSERT(PageLocked(page));
1407
1408         exp = ll_i2dtexp(inode);
1409         if (exp == NULL)
1410                 GOTO(out, rc = -EINVAL);
1411
1412         llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
1413         if (IS_ERR(llap))
1414                 GOTO(out, rc = PTR_ERR(llap));
1415
1416         LASSERT(!PageWriteback(page));
1417         set_page_writeback(page);
1418
1419         page_cache_get(page);
1420         if (llap->llap_write_queued) {
1421                 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
1422                 rc = obd_set_async_flags(exp, lli->lli_smd, NULL,
1423                                          llap->llap_cookie,
1424                                          ASYNC_READY | ASYNC_URGENT);
1425         } else {
1426                 rc = queue_or_sync_write(exp, inode, llap, CFS_PAGE_SIZE,
1427                                          ASYNC_READY | ASYNC_URGENT);
1428         }
1429         if (rc)
1430                 page_cache_release(page);
1431 out:
1432         if (rc) {
1433                 if (!lli->lli_async_rc)
1434                         lli->lli_async_rc = rc;
1435                 /* re-dirty page on error so it retries write */
1436                 if (PageWriteback(page)) {
1437                         end_page_writeback(page);
1438                 }
1439                 ll_redirty_page(page);
1440                 unlock_page(page);
1441         }
1442         RETURN(rc);
1443 }
1444
1445 /*
1446  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
1447  * read-ahead assumes it is valid to issue readpage all the way up to
1448  * i_size, but our dlm locks make that not the case.  We disable the
1449  * kernel's read-ahead and do our own by walking ahead in the page cache
1450  * checking for dlm lock coverage.  the main difference between 2.4 and
1451  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1452  * so they look the same.
1453  */
1454 int ll_readpage(struct file *filp, struct page *page)
1455 {
1456         struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
1457         struct inode *inode = page->mapping->host;
1458         struct obd_export *exp;
1459         struct ll_async_page *llap;
1460         struct obd_io_group *oig = NULL;
1461         int rc;
1462         ENTRY;
1463
1464         LASSERT(PageLocked(page));
1465         LASSERT(!PageUptodate(page));
1466         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset=%Lu=%#Lx\n",
1467                inode->i_ino, inode->i_generation, inode,
1468                (((loff_t)page->index) << CFS_PAGE_SHIFT),
1469                (((loff_t)page->index) << CFS_PAGE_SHIFT));
1470         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1471
1472         if (!ll_i2info(inode)->lli_smd) {
1473                 /* File with no objects - one big hole */
1474                 /* We use this just for remove_from_page_cache that is not
1475                  * exported, we'd make page back up to date. */
1476                 ll_truncate_complete_page(page);
1477                 clear_page(kmap(page));
1478                 kunmap(page);
1479                 SetPageUptodate(page);
1480                 unlock_page(page);
1481                 RETURN(0);
1482         }
1483
1484         rc = oig_init(&oig);
1485         if (rc < 0)
1486                 GOTO(out, rc);
1487
1488         exp = ll_i2dtexp(inode);
1489         if (exp == NULL)
1490                 GOTO(out, rc = -EINVAL);
1491
1492         llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1493         if (IS_ERR(llap))
1494                 GOTO(out, rc = PTR_ERR(llap));
1495
1496         llap->llap_fsuid = current->fsuid;
1497
1498         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1499                 ras_update(ll_i2sbi(inode), inode, &fd->fd_ras, page->index,
1500                            llap->llap_defer_uptodate);
1501
1502         if (llap->llap_defer_uptodate) {
1503                 llap->llap_ra_used = 1;
1504                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1505                                   fd->fd_flags);
1506                 if (rc > 0)
1507                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1508                                              NULL, oig);
1509                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1510                 SetPageUptodate(page);
1511                 unlock_page(page);
1512                 GOTO(out_oig, rc = 0);
1513         }
1514
1515         if (likely((fd->fd_flags & LL_FILE_IGNORE_LOCK) == 0)) {
1516                 rc = ll_page_matches(page, fd->fd_flags);
1517                 if (rc < 0) {
1518                         LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1519                         GOTO(out, rc);
1520                 }
1521
1522                 if (rc == 0) {
1523                         CWARN("ino %lu page %lu (%llu) not covered by "
1524                               "a lock (mmap?).  check debug logs.\n",
1525                               inode->i_ino, page->index,
1526                               (long long)page->index << CFS_PAGE_SHIFT);
1527                 }
1528         }
1529
1530         rc = ll_issue_page_read(exp, llap, oig, 0);
1531         if (rc)
1532                 GOTO(out, rc);
1533
1534         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1535         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1536                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1537                              fd->fd_flags);
1538
1539         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1540
1541 out:
1542         if (rc)
1543                 unlock_page(page);
1544 out_oig:
1545         if (oig != NULL)
1546                 oig_release(oig);
1547         RETURN(rc);
1548 }