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>i_size_read(inode)))
  77                 pg.count = i_size_read(inode) % 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", inode->i_ino, inode, i_size_read(inode),
  87                        page->mapping->host, i_size_read(page->mapping->host),
  88                        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, 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, i_size_read(inode),
 129                i_size_read(inode));
 130
 131         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
 132         if (lli->lli_size_sem_owner != current) {
 133                 EXIT;
 134                 return;
 135         }
 136
 137         if (!lli->lli_smd) {
 138                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 139                        inode->i_ino);
 140                 GOTO(out_unlock, 0);
 141         }
 142
 143         LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
 144
 145         /* XXX I'm pretty sure this is a hack to paper over a more fundamental
 146          * race condition. */
 147         lov_stripe_lock(lli->lli_smd);
 148         inode_init_lvb(inode, &lvb);
 149         obd_merge_lvb(ll_i2dtexp(inode), lli->lli_smd, &lvb, 0);
 150         if (lvb.lvb_size == i_size_read(inode)) {
 151                 CDEBUG(D_VFSTRACE, "skipping punch for obj "LPX64", %Lu=%#Lx\n",
 152                        lli->lli_smd->lsm_object_id, i_size_read(inode),
 153                        i_size_read(inode));
 154                 lov_stripe_unlock(lli->lli_smd);
 155                 GOTO(out_unlock, 0);
 156         }
 157
 158         obd_adjust_kms(ll_i2dtexp(inode), lli->lli_smd, i_size_read(inode), 1);
 159         lov_stripe_unlock(lli->lli_smd);
 160
 161         if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_CHECKSUM) &&
 162                      (i_size_read(inode) & ~CFS_PAGE_MASK))) {
 163                 /* If the truncate leaves behind a partial page, update its
 164                  * checksum. */
 165                 struct page *page = find_get_page(inode->i_mapping,
 166                                                   i_size_read(inode) >>
 167                                                   CFS_PAGE_SHIFT);
 168                 if (page != NULL) {
 169                         struct ll_async_page *llap = llap_cast_private(page);
 170                         if (llap != NULL) {
 171                                 llap->llap_checksum =
 172                                         crc32_le(0, kmap(page), CFS_PAGE_SIZE);
 173                                 kunmap(page);
 174                         }
 175                         page_cache_release(page);
 176                 }
 177         }
 178
 179         CDEBUG(D_INFO, "calling punch for "LPX64" (new size %Lu=%#Lx)\n",
 180                lli->lli_smd->lsm_object_id, i_size_read(inode), i_size_read(inode));
 181
 182         oinfo.oi_md = lli->lli_smd;
 183         oinfo.oi_policy.l_extent.start = i_size_read(inode);
 184         oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
 185         oinfo.oi_oa = &oa;
 186         oa.o_id = lli->lli_smd->lsm_object_id;
 187         oa.o_gr = lli->lli_smd->lsm_object_gr;
 188         oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 189
 190         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
 191                         OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 192                         OBD_MD_FLFID | OBD_MD_FLGENER);
 193
 194         ll_inode_size_unlock(inode, 0);
 195
 196         oinfo.oi_capa = ll_osscapa_get(inode, CAPA_OPC_OSS_TRUNC);
 197         rc = obd_punch_rqset(ll_i2dtexp(inode), &oinfo, NULL);
 198         ll_truncate_free_capa(oinfo.oi_capa);
 199         if (rc)
 200                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 201         else
 202                 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 203                               OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 204         EXIT;
 205         return;
 206
 207  out_unlock:
 208         ll_inode_size_unlock(inode, 0);
 209 } /* ll_truncate */
 210
 211 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 212                      unsigned to)
 213 {
 214         struct inode *inode = page->mapping->host;
 215         struct ll_inode_info *lli = ll_i2info(inode);
 216         struct lov_stripe_md *lsm = lli->lli_smd;
 217         obd_off offset = ((obd_off)page->index) << CFS_PAGE_SHIFT;
 218         struct obd_info oinfo = { { { 0 } } };
 219         struct brw_page pga;
 220         struct obdo oa;
 221         struct ost_lvb lvb;
 222         int rc = 0;
 223         ENTRY;
 224
 225         LASSERT(PageLocked(page));
 226         (void)llap_cast_private(page); /* assertion */
 227
 228         /* Check to see if we should return -EIO right away */
 229         pga.pg = page;
 230         pga.off = offset;
 231         pga.count = CFS_PAGE_SIZE;
 232         pga.flag = 0;
 233
 234         oa.o_mode = inode->i_mode;
 235         oa.o_id = lsm->lsm_object_id;
 236         oa.o_gr = lsm->lsm_object_gr;
 237         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
 238                      OBD_MD_FLTYPE | OBD_MD_FLGROUP;
 239         obdo_from_inode(&oa, inode, OBD_MD_FLFID | OBD_MD_FLGENER);
 240
 241         oinfo.oi_oa = &oa;
 242         oinfo.oi_md = lsm;
 243         rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode), &oinfo, 1, &pga, NULL);
 244         if (rc)
 245                 RETURN(rc);
 246
 247         if (PageUptodate(page)) {
 248                 LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n");
 249                 RETURN(0);
 250         }
 251
 252         /* We're completely overwriting an existing page, so _don't_ set it up
 253          * to date until commit_write */
 254         if (from == 0 && to == CFS_PAGE_SIZE) {
 255                 LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n");
 256                 POISON_PAGE(page, 0x11);
 257                 RETURN(0);
 258         }
 259
 260         /* If are writing to a new page, no need to read old data.  The extent
 261          * locking will have updated the KMS, and for our purposes here we can
 262          * treat it like i_size. */
 263         lov_stripe_lock(lsm);
 264         inode_init_lvb(inode, &lvb);
 265         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
 266         lov_stripe_unlock(lsm);
 267         if (lvb.lvb_size <= offset) {
 268                 LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n",
 269                                lvb.lvb_size, offset);
 270                 memset(kmap(page), 0, CFS_PAGE_SIZE);
 271                 kunmap(page);
 272                 GOTO(prepare_done, rc = 0);
 273         }
 274
 275         /* XXX could be an async ocp read.. read-ahead? */
 276         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 277         if (rc == 0) {
 278                 /* bug 1598: don't clobber blksize */
 279                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 280                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 281         }
 282
 283         EXIT;
 284  prepare_done:
 285         if (rc == 0)
 286                 SetPageUptodate(page);
 287
 288         return rc;
 289 }
 290
 291 static int ll_ap_make_ready(void *data, int cmd)
 292 {
 293         struct ll_async_page *llap;
 294         struct page *page;
 295         ENTRY;
 296
 297         llap = LLAP_FROM_COOKIE(data);
 298         page = llap->llap_page;
 299
 300         LASSERTF(!(cmd & OBD_BRW_READ), "cmd %x page %p ino %lu index %lu\n", cmd, page,
 301                  page->mapping->host->i_ino, page->index);
 302
 303         /* we're trying to write, but the page is locked.. come back later */
 304         if (TryLockPage(page))
 305                 RETURN(-EAGAIN);
 306
 307         LASSERT(!PageWriteback(page));
 308
 309         /* if we left PageDirty we might get another writepage call
 310          * in the future.  list walkers are bright enough
 311          * to check page dirty so we can leave it on whatever list
 312          * its on.  XXX also, we're called with the cli list so if
 313          * we got the page cache list we'd create a lock inversion
 314          * with the removepage path which gets the page lock then the
 315          * cli lock */
 316 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,0)
 317         clear_page_dirty(page);
 318 #else
 319         LASSERTF(!PageWriteback(page),"cmd %x page %p ino %lu index %lu\n", cmd, page,
 320                  page->mapping->host->i_ino, page->index);
 321         clear_page_dirty_for_io(page);
 322
 323         /* This actually clears the dirty bit in the radix tree.*/
 324         set_page_writeback(page);
 325 #endif
 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                 if (llap->llap_write_queued || PageDirty(page) ||
 519                     (!PageUptodate(page) &&
 520                      llap->llap_origin != LLAP_ORIGIN_READAHEAD))
 521                         keep = 1;
 522                 else
 523                         keep = 0;
 524
 525                 LL_CDEBUG_PAGE(D_PAGE, page,"%s LRU page: %s%s%s%s origin %s\n",
 526                                keep ? "keep" : "drop",
 527                                llap->llap_write_queued ? "wq " : "",
 528                                PageDirty(page) ? "pd " : "",
 529                                PageUptodate(page) ? "" : "!pu ",
 530                                llap->llap_defer_uptodate ? "" : "!du",
 531                                llap_origins[llap->llap_origin]);
 532
 533                 /* If page is dirty or undergoing IO don't discard it */
 534                 if (keep) {
 535                         unlock_page(page);
 536                         continue;
 537                 }
 538
 539                 page_cache_get(page);
 540                 spin_unlock(&sbi->ll_lock);
 541
 542                 if (page->mapping != NULL) {
 543                         ll_teardown_mmaps(page->mapping,
 544                                          (__u64)page->index << CFS_PAGE_SHIFT,
 545                                          ((__u64)page->index << CFS_PAGE_SHIFT)|
 546                                           ~CFS_PAGE_MASK);
 547                         if (!PageDirty(page) && !page_mapped(page)) {
 548                                 ll_ra_accounting(llap, page->mapping);
 549                                 ll_truncate_complete_page(page);
 550                                 ++count;
 551                         } else {
 552                                 LL_CDEBUG_PAGE(D_PAGE, page, "Not dropping page"
 553                                                              " because it is "
 554                                                              "%s\n",
 555                                                               PageDirty(page)?
 556                                                               "dirty":"mapped");
 557                         }
 558                 }
 559                 unlock_page(page);
 560                 page_cache_release(page);
 561
 562                 spin_lock(&sbi->ll_lock);
 563         }
 564         list_del(&dummy_llap.llap_pglist_item);
 565         spin_unlock(&sbi->ll_lock);
 566
 567         CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n",
 568                count, want, total);
 569
 570         return count;
 571 }
 572
 573 struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
 574 {
 575         struct ll_async_page *llap;
 576         struct obd_export *exp;
 577         struct inode *inode = page->mapping->host;
 578         struct ll_sb_info *sbi;
 579         int rc;
 580         ENTRY;
 581
 582         if (!inode) {
 583                 static int triggered;
 584
 585                 if (!triggered) {
 586                         LL_CDEBUG_PAGE(D_ERROR, page, "Bug 10047. Wrong anon "
 587                                        "page received\n");
 588                         libcfs_debug_dumpstack(NULL);
 589                         triggered = 1;
 590                 }
 591                 RETURN(ERR_PTR(-EINVAL));
 592         }
 593         sbi = ll_i2sbi(inode);
 594         LASSERT(ll_async_page_slab);
 595         LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
 596
 597         llap = llap_cast_private(page);
 598         if (llap != NULL) {
 599                 /* move to end of LRU list, except when page is just about to
 600                  * die */
 601                 if (origin != LLAP_ORIGIN_REMOVEPAGE) {
 602                         spin_lock(&sbi->ll_lock);
 603                         sbi->ll_pglist_gen++;
 604                         list_del_init(&llap->llap_pglist_item);
 605                         list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 606                         spin_unlock(&sbi->ll_lock);
 607                 }
 608                 GOTO(out, llap);
 609         }
 610
 611         exp = ll_i2dtexp(page->mapping->host);
 612         if (exp == NULL)
 613                 RETURN(ERR_PTR(-EINVAL));
 614
 615         /* limit the number of lustre-cached pages */
 616         if (sbi->ll_async_page_count >= sbi->ll_async_page_max)
 617                 llap_shrink_cache(sbi, 0);
 618
 619         OBD_SLAB_ALLOC(llap, ll_async_page_slab, CFS_ALLOC_STD,
 620                        ll_async_page_slab_size);
 621         if (llap == NULL)
 622                 RETURN(ERR_PTR(-ENOMEM));
 623         llap->llap_magic = LLAP_MAGIC;
 624         llap->llap_cookie = (void *)llap + size_round(sizeof(*llap));
 625
 626         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 627                                  (obd_off)page->index << CFS_PAGE_SHIFT,
 628                                  &ll_async_page_ops, llap, &llap->llap_cookie);
 629         if (rc) {
 630                 OBD_SLAB_FREE(llap, ll_async_page_slab,
 631                               ll_async_page_slab_size);
 632                 RETURN(ERR_PTR(rc));
 633         }
 634
 635         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 636                page, llap->llap_cookie, (obd_off)page->index << CFS_PAGE_SHIFT);
 637         /* also zeroing the PRIVBITS low order bitflags */
 638         __set_page_ll_data(page, llap);
 639         llap->llap_page = page;
 640         spin_lock(&sbi->ll_lock);
 641         sbi->ll_pglist_gen++;
 642         sbi->ll_async_page_count++;
 643         list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 644         INIT_LIST_HEAD(&llap->llap_pending_write);
 645         spin_unlock(&sbi->ll_lock);
 646
 647  out:
 648         if (unlikely(sbi->ll_flags & LL_SBI_CHECKSUM)) {
 649                 __u32 csum = 0;
 650                 csum = crc32_le(csum, kmap(page), CFS_PAGE_SIZE);
 651                 kunmap(page);
 652                 if (origin == LLAP_ORIGIN_READAHEAD ||
 653                     origin == LLAP_ORIGIN_READPAGE) {
 654                         llap->llap_checksum = 0;
 655                 } else if (origin == LLAP_ORIGIN_COMMIT_WRITE ||
 656                            llap->llap_checksum == 0) {
 657                         llap->llap_checksum = csum;
 658                         CDEBUG(D_PAGE, "page %p cksum %x\n", page, csum);
 659                 } else if (llap->llap_checksum == csum) {
 660                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 661                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 662                                page, csum);
 663                 } else {
 664                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 665                         LL_CDEBUG_PAGE(D_ERROR, page, "old cksum %x != new "
 666                                        "%x!\n", llap->llap_checksum, csum);
 667                 }
 668         }
 669
 670         llap->llap_origin = origin;
 671         RETURN(llap);
 672 }
 673
 674 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
 675                                struct ll_async_page *llap,
 676                                unsigned to, obd_flag async_flags)
 677 {
 678         unsigned long size_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
 679         struct obd_io_group *oig;
 680         struct ll_sb_info *sbi = ll_i2sbi(inode);
 681         int rc, noquot = llap->llap_ignore_quota ? OBD_BRW_NOQUOTA : 0;
 682         ENTRY;
 683
 684         /* _make_ready only sees llap once we've unlocked the page */
 685         llap->llap_write_queued = 1;
 686         rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
 687                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 688                                 0, 0, 0, async_flags);
 689         if (rc == 0) {
 690                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
 691                 GOTO(out, 0);
 692         }
 693
 694         llap->llap_write_queued = 0;
 695         /* Do not pass llap here as it is sync write. */
 696         llap_write_pending(inode, NULL);
 697
 698         rc = oig_init(&oig);
 699         if (rc)
 700                 GOTO(out, rc);
 701
 702         /* make full-page requests if we are not at EOF (bug 4410) */
 703         if (to != CFS_PAGE_SIZE && llap->llap_page->index < size_index) {
 704                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 705                                "sync write before EOF: size_index %lu, to %d\n",
 706                                size_index, to);
 707                 to = CFS_PAGE_SIZE;
 708         } else if (to != CFS_PAGE_SIZE && llap->llap_page->index == size_index) {
 709                 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
 710                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 711                                "sync write at EOF: size_index %lu, to %d/%d\n",
 712                                size_index, to, size_to);
 713                 if (to < size_to)
 714                         to = size_to;
 715         }
 716
 717         /* compare the checksum once before the page leaves llite */
 718         if (unlikely((sbi->ll_flags & LL_SBI_CHECKSUM) &&
 719                      llap->llap_checksum != 0)) {
 720                 __u32 csum = 0;
 721                 struct page *page = llap->llap_page;
 722                 csum = crc32_le(csum, kmap(page), CFS_PAGE_SIZE);
 723                 kunmap(page);
 724                 if (llap->llap_checksum == csum) {
 725                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 726                                page, csum);
 727                 } else {
 728                         CERROR("page %p old cksum %x != new cksum %x!\n",
 729                                page, llap->llap_checksum, csum);
 730                 }
 731         }
 732
 733         rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
 734                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 735                                 0, to, 0, ASYNC_READY | ASYNC_URGENT |
 736                                 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
 737         if (rc)
 738                 GOTO(free_oig, rc);
 739
 740         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
 741         if (rc)
 742                 GOTO(free_oig, rc);
 743
 744         rc = oig_wait(oig);
 745
 746         if (!rc && async_flags & ASYNC_READY) {
 747                 unlock_page(llap->llap_page);
 748                 if (PageWriteback(llap->llap_page)) {
 749                         end_page_writeback(llap->llap_page);
 750                 }
 751         }
 752
 753         if (rc == 0 && llap_write_complete(inode, llap))
 754                 ll_queue_done_writing(inode, 0);
 755
 756         LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
 757
 758 free_oig:
 759         oig_release(oig);
 760 out:
 761         RETURN(rc);
 762 }
 763
 764 /* update our write count to account for i_size increases that may have
 765  * happened since we've queued the page for io. */
 766
 767 /* be careful not to return success without setting the page Uptodate or
 768  * the next pass through prepare_write will read in stale data from disk. */
 769 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 770                     unsigned to)
 771 {
 772         struct inode *inode = page->mapping->host;
 773         struct ll_inode_info *lli = ll_i2info(inode);
 774         struct lov_stripe_md *lsm = lli->lli_smd;
 775         struct obd_export *exp;
 776         struct ll_async_page *llap;
 777         loff_t size;
 778         int rc = 0;
 779         ENTRY;
 780
 781         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 782         LASSERT(inode == file->f_dentry->d_inode);
 783         LASSERT(PageLocked(page));
 784
 785         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 786                inode, page, from, to, page->index);
 787
 788         llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
 789         if (IS_ERR(llap))
 790                 RETURN(PTR_ERR(llap));
 791
 792         exp = ll_i2dtexp(inode);
 793         if (exp == NULL)
 794                 RETURN(-EINVAL);
 795
 796         llap->llap_ignore_quota = capable(CAP_SYS_RESOURCE);
 797
 798         /*
 799          * queue a write for some time in the future the first time we
 800          * dirty the page.
 801          *
 802          * This is different from what other file systems do: they usually
 803          * just mark page (and some of its buffers) dirty and rely on
 804          * balance_dirty_pages() to start a write-back. Lustre wants write-back
 805          * to be started earlier for the following reasons:
 806          *
 807          *     (1) with a large number of clients we need to limit the amount
 808          *     of cached data on the clients a lot;
 809          *
 810          *     (2) large compute jobs generally want compute-only then io-only
 811          *     and the IO should complete as quickly as possible;
 812          *
 813          *     (3) IO is batched up to the RPC size and is async until the
 814          *     client max cache is hit
 815          *     (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
 816          *
 817          */
 818         if (!PageDirty(page)) {
 819                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_MISSES, 1);
 820
 821                 rc = queue_or_sync_write(exp, inode, llap, to, 0);
 822                 if (rc)
 823                         GOTO(out, rc);
 824         } else {
 825                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_HITS, 1);
 826         }
 827
 828         /* put the page in the page cache, from now on ll_removepage is
 829          * responsible for cleaning up the llap.
 830          * only set page dirty when it's queued to be write out */
 831         if (llap->llap_write_queued)
 832                 set_page_dirty(page);
 833
 834 out:
 835         size = (((obd_off)page->index) << CFS_PAGE_SHIFT) + to;
 836         ll_inode_size_lock(inode, 0);
 837         if (rc == 0) {
 838                 lov_stripe_lock(lsm);
 839                 obd_adjust_kms(exp, lsm, size, 0);
 840                 lov_stripe_unlock(lsm);
 841                 if (size > i_size_read(inode))
 842                         i_size_write(inode, size);
 843                 SetPageUptodate(page);
 844         } else if (size > i_size_read(inode)) {
 845                 /* this page beyond the pales of i_size, so it can't be
 846                  * truncated in ll_p_r_e during lock revoking. we must
 847                  * teardown our book-keeping here. */
 848                 ll_removepage(page);
 849         }
 850         ll_inode_size_unlock(inode, 0);
 851         RETURN(rc);
 852 }
 853
 854 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 855 {
 856         struct ll_ra_info *ra = &sbi->ll_ra_info;
 857         unsigned long ret;
 858         ENTRY;
 859
 860         spin_lock(&sbi->ll_lock);
 861         ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
 862         ra->ra_cur_pages += ret;
 863         spin_unlock(&sbi->ll_lock);
 864
 865         RETURN(ret);
 866 }
 867
 868 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 869 {
 870         struct ll_ra_info *ra = &sbi->ll_ra_info;
 871         spin_lock(&sbi->ll_lock);
 872         LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
 873                  ra->ra_cur_pages, len);
 874         ra->ra_cur_pages -= len;
 875         spin_unlock(&sbi->ll_lock);
 876 }
 877
 878 /* called for each page in a completed rpc.*/
 879 int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 880 {
 881         struct ll_async_page *llap;
 882         struct page *page;
 883         int ret = 0;
 884         ENTRY;
 885
 886         llap = LLAP_FROM_COOKIE(data);
 887         page = llap->llap_page;
 888         LASSERT(PageLocked(page));
 889         LASSERT(CheckWriteback(page,cmd));
 890
 891         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 892
 893         if (cmd & OBD_BRW_READ && llap->llap_defer_uptodate)
 894                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 895
 896         if (rc == 0)  {
 897                 if (cmd & OBD_BRW_READ) {
 898                         if (!llap->llap_defer_uptodate)
 899                                 SetPageUptodate(page);
 900                 } else {
 901                         llap->llap_write_queued = 0;
 902                 }
 903                 ClearPageError(page);
 904         } else {
 905                 if (cmd & OBD_BRW_READ) {
 906                         llap->llap_defer_uptodate = 0;
 907                 } else {
 908                         ll_redirty_page(page);
 909                         ret = 1;
 910                 }
 911                 SetPageError(page);
 912         }
 913
 914         unlock_page(page);
 915
 916         if (cmd & OBD_BRW_WRITE) {
 917                 /* Only rc == 0, write succeed, then this page could be deleted
 918                  * from the pending_writing list
 919                  */
 920                 if (rc == 0 && llap_write_complete(page->mapping->host, llap))
 921                         ll_queue_done_writing(page->mapping->host, 0);
 922         }
 923
 924         if (PageWriteback(page)) {
 925                 end_page_writeback(page);
 926         }
 927         page_cache_release(page);
 928
 929         RETURN(ret);
 930 }
 931
 932 /* the kernel calls us here when a page is unhashed from the page cache.
 933  * the page will be locked and the kernel is holding a spinlock, so
 934  * we need to be careful.  we're just tearing down our book-keeping
 935  * here. */
 936 void ll_removepage(struct page *page)
 937 {
 938         struct inode *inode = page->mapping->host;
 939         struct obd_export *exp;
 940         struct ll_async_page *llap;
 941         struct ll_sb_info *sbi = ll_i2sbi(inode);
 942         int rc;
 943         ENTRY;
 944
 945         LASSERT(!in_interrupt());
 946
 947         /* sync pages or failed read pages can leave pages in the page
 948          * cache that don't have our data associated with them anymore */
 949         if (page_private(page) == 0) {
 950                 EXIT;
 951                 return;
 952         }
 953
 954         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 955
 956         exp = ll_i2dtexp(inode);
 957         if (exp == NULL) {
 958                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 959                 EXIT;
 960                 return;
 961         }
 962
 963         llap = llap_from_page(page, LLAP_ORIGIN_REMOVEPAGE);
 964         if (IS_ERR(llap)) {
 965                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 966                        page->index, PTR_ERR(llap));
 967                 EXIT;
 968                 return;
 969         }
 970
 971         if (llap_write_complete(inode, llap))
 972                 ll_queue_done_writing(inode, 0);
 973
 974         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 975                                      llap->llap_cookie);
 976         if (rc != 0)
 977                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 978
 979         /* this unconditional free is only safe because the page lock
 980          * is providing exclusivity to memory pressure/truncate/writeback..*/
 981         __clear_page_ll_data(page);
 982
 983         spin_lock(&sbi->ll_lock);
 984         if (!list_empty(&llap->llap_pglist_item))
 985                 list_del_init(&llap->llap_pglist_item);
 986         sbi->ll_pglist_gen++;
 987         sbi->ll_async_page_count--;
 988         spin_unlock(&sbi->ll_lock);
 989         OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
 990         EXIT;
 991 }
 992
 993 static int ll_page_matches(struct page *page, int fd_flags)
 994 {
 995         struct lustre_handle match_lockh = {0};
 996         struct inode *inode = page->mapping->host;
 997         ldlm_policy_data_t page_extent;
 998         int flags, matches;
 999         ENTRY;
1000
1001         if (unlikely(fd_flags & LL_FILE_GROUP_LOCKED))
1002                 RETURN(1);
1003
1004         page_extent.l_extent.start = (__u64)page->index << CFS_PAGE_SHIFT;
1005         page_extent.l_extent.end =
1006                 page_extent.l_extent.start + CFS_PAGE_SIZE - 1;
1007         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
1008         if (!(fd_flags & LL_FILE_READAHEAD))
1009                 flags |= LDLM_FL_CBPENDING;
1010         matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
1011                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
1012                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
1013                             &match_lockh);
1014         RETURN(matches);
1015 }
1016
1017 static int ll_issue_page_read(struct obd_export *exp,
1018                               struct ll_async_page *llap,
1019                               struct obd_io_group *oig, int defer)
1020 {
1021         struct page *page = llap->llap_page;
1022         int rc;
1023
1024         page_cache_get(page);
1025         llap->llap_defer_uptodate = defer;
1026         llap->llap_ra_used = 0;
1027         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
1028                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
1029                                 CFS_PAGE_SIZE, 0, ASYNC_COUNT_STABLE |
1030                                                   ASYNC_READY | ASYNC_URGENT);
1031         if (rc) {
1032                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
1033                 page_cache_release(page);
1034         }
1035         RETURN(rc);
1036 }
1037
1038 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
1039 {
1040         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
1041         ra->ra_stats[which]++;
1042 }
1043
1044 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
1045 {
1046         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
1047         struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
1048
1049         spin_lock(&sbi->ll_lock);
1050         ll_ra_stats_inc_unlocked(ra, which);
1051         spin_unlock(&sbi->ll_lock);
1052 }
1053
1054 void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping)
1055 {
1056         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
1057                 return;
1058
1059         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
1060 }
1061
1062 #define RAS_CDEBUG(ras) \
1063         CDEBUG(D_READA,                                                      \
1064                "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu\n", \
1065                ras->ras_last_readpage, ras->ras_consecutive_requests,        \
1066                ras->ras_consecutive_pages, ras->ras_window_start,            \
1067                ras->ras_window_len, ras->ras_next_readahead,                 \
1068                ras->ras_requests, ras->ras_request_index);
1069
1070 static int index_in_window(unsigned long index, unsigned long point,
1071                            unsigned long before, unsigned long after)
1072 {
1073         unsigned long start = point - before, end = point + after;
1074
1075         if (start > point)
1076                start = 0;
1077         if (end < point)
1078                end = ~0;
1079
1080         return start <= index && index <= end;
1081 }
1082
1083 static struct ll_readahead_state *ll_ras_get(struct file *f)
1084 {
1085         struct ll_file_data       *fd;
1086
1087         fd = LUSTRE_FPRIVATE(f);
1088         return &fd->fd_ras;
1089 }
1090
1091 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
1092 {
1093         struct ll_readahead_state *ras;
1094
1095         ras = ll_ras_get(f);
1096
1097         spin_lock(&ras->ras_lock);
1098         ras->ras_requests++;
1099         ras->ras_request_index = 0;
1100         ras->ras_consecutive_requests++;
1101         rar->lrr_reader = current;
1102
1103         list_add(&rar->lrr_linkage, &ras->ras_read_beads);
1104         spin_unlock(&ras->ras_lock);
1105 }
1106
1107 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
1108 {
1109         struct ll_readahead_state *ras;
1110
1111         ras = ll_ras_get(f);
1112
1113         spin_lock(&ras->ras_lock);
1114         list_del_init(&rar->lrr_linkage);
1115         spin_unlock(&ras->ras_lock);
1116 }
1117
1118 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
1119 {
1120         struct ll_ra_read *scan;
1121
1122         list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
1123                 if (scan->lrr_reader == current)
1124                         return scan;
1125         }
1126         return NULL;
1127 }
1128
1129 struct ll_ra_read *ll_ra_read_get(struct file *f)
1130 {
1131         struct ll_readahead_state *ras;
1132         struct ll_ra_read         *bead;
1133
1134         ras = ll_ras_get(f);
1135
1136         spin_lock(&ras->ras_lock);
1137         bead = ll_ra_read_get_locked(ras);
1138         spin_unlock(&ras->ras_lock);
1139         return bead;
1140 }
1141
1142 static int ll_readahead(struct ll_readahead_state *ras,
1143                          struct obd_export *exp, struct address_space *mapping,
1144                          struct obd_io_group *oig, int flags)
1145 {
1146         unsigned long i, start = 0, end = 0, reserved;
1147         struct ll_async_page *llap;
1148         struct page *page;
1149         int rc, ret = 0, match_failed = 0;
1150         __u64 kms;
1151         unsigned int gfp_mask;
1152         struct inode *inode;
1153         struct lov_stripe_md *lsm;
1154         struct ll_ra_read *bead;
1155         struct ost_lvb lvb;
1156         ENTRY;
1157
1158         inode = mapping->host;
1159         lsm = ll_i2info(inode)->lli_smd;
1160
1161         lov_stripe_lock(lsm);
1162         inode_init_lvb(inode, &lvb);
1163         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
1164         kms = lvb.lvb_size;
1165         lov_stripe_unlock(lsm);
1166         if (kms == 0) {
1167                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
1168                 RETURN(0);
1169         }
1170
1171         spin_lock(&ras->ras_lock);
1172         bead = ll_ra_read_get_locked(ras);
1173         /* Enlarge the RA window to encompass the full read */
1174         if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
1175             bead->lrr_start + bead->lrr_count) {
1176                 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
1177                                       ras->ras_window_start;
1178         }
1179         /* Reserve a part of the read-ahead window that we'll be issuing */
1180         if (ras->ras_window_len) {
1181                 start = ras->ras_next_readahead;
1182                 end = ras->ras_window_start + ras->ras_window_len - 1;
1183         }
1184         if (end != 0) {
1185                 /* Truncate RA window to end of file */
1186                 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
1187                 ras->ras_next_readahead = max(end, end + 1);
1188                 RAS_CDEBUG(ras);
1189         }
1190         spin_unlock(&ras->ras_lock);
1191
1192         if (end == 0) {
1193                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
1194                 RETURN(0);
1195         }
1196
1197         reserved = ll_ra_count_get(ll_i2sbi(inode), end - start + 1);
1198         if (reserved < end - start + 1)
1199                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
1200
1201         gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
1202 #ifdef __GFP_NOWARN
1203         gfp_mask |= __GFP_NOWARN;
1204 #endif
1205
1206         for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
1207                 /* skip locked pages from previous readpage calls */
1208                 page = grab_cache_page_nowait_gfp(mapping, i, gfp_mask);
1209                 if (page == NULL) {
1210                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_GRAB_PAGE);
1211                         CDEBUG(D_READA, "g_c_p_n failed\n");
1212                         continue;
1213                 }
1214
1215                 /* Check if page was truncated or reclaimed */
1216                 if (page->mapping != mapping) {
1217                         ll_ra_stats_inc(mapping, RA_STAT_WRONG_GRAB_PAGE);
1218                         CDEBUG(D_READA, "g_c_p_n returned invalid page\n");
1219                         goto next_page;
1220                 }
1221
1222                 /* we do this first so that we can see the page in the /proc
1223                  * accounting */
1224                 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
1225                 if (IS_ERR(llap) || llap->llap_defer_uptodate)
1226                         goto next_page;
1227
1228                 /* skip completed pages */
1229                 if (Page_Uptodate(page))
1230                         goto next_page;
1231
1232                 /* bail when we hit the end of the lock. */
1233                 if ((rc = ll_page_matches(page, flags|LL_FILE_READAHEAD)) <= 0){
1234                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
1235                                        "lock match failed: rc %d\n", rc);
1236                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
1237                         match_failed = 1;
1238                         goto next_page;
1239                 }
1240
1241                 rc = ll_issue_page_read(exp, llap, oig, 1);
1242                 if (rc == 0) {
1243                         reserved--;
1244                         ret++;
1245                         LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
1246                                        "started read-ahead\n");
1247                 } else {
1248         next_page:
1249                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
1250                                        "skipping read-ahead\n");
1251
1252                         unlock_page(page);
1253                 }
1254                 page_cache_release(page);
1255         }
1256
1257         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
1258         if (reserved != 0)
1259                 ll_ra_count_put(ll_i2sbi(inode), reserved);
1260         if (i == end + 1 && end == (kms >> CFS_PAGE_SHIFT))
1261                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
1262
1263         /* if we didn't get to the end of the region we reserved from
1264          * the ras we need to go back and update the ras so that the
1265          * next read-ahead tries from where we left off.  we only do so
1266          * if the region we failed to issue read-ahead on is still ahead
1267          * of the app and behind the next index to start read-ahead from */
1268         if (i != end + 1) {
1269                 spin_lock(&ras->ras_lock);
1270                 if (i < ras->ras_next_readahead &&
1271                     index_in_window(i, ras->ras_window_start, 0,
1272                                     ras->ras_window_len)) {
1273                         ras->ras_next_readahead = i;
1274                         RAS_CDEBUG(ras);
1275                 }
1276                 spin_unlock(&ras->ras_lock);
1277         }
1278
1279         RETURN(ret);
1280 }
1281
1282 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
1283 {
1284         ras->ras_window_start = index & (~((1024 * 1024 >> CFS_PAGE_SHIFT) - 1));
1285 }
1286
1287 /* called with the ras_lock held or from places where it doesn't matter */
1288 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1289 {
1290         ras->ras_last_readpage = index;
1291         ras->ras_consecutive_requests = 0;
1292         ras->ras_consecutive_pages = 0;
1293         ras->ras_window_len = 0;
1294         ras_set_start(ras, index);
1295         ras->ras_next_readahead = max(ras->ras_window_start, index);
1296
1297         RAS_CDEBUG(ras);
1298 }
1299
1300 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1301 {
1302         spin_lock_init(&ras->ras_lock);
1303         ras_reset(ras, 0);
1304         ras->ras_requests = 0;
1305         INIT_LIST_HEAD(&ras->ras_read_beads);
1306 }
1307
1308 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1309                        struct ll_readahead_state *ras, unsigned long index,
1310                        unsigned hit)
1311 {
1312         struct ll_ra_info *ra = &sbi->ll_ra_info;
1313         int zero = 0;
1314         ENTRY;
1315
1316         spin_lock(&sbi->ll_lock);
1317         spin_lock(&ras->ras_lock);
1318
1319         ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
1320
1321         /* reset the read-ahead window in two cases.  First when the app seeks
1322          * or reads to some other part of the file.  Secondly if we get a
1323          * read-ahead miss that we think we've previously issued.  This can
1324          * be a symptom of there being so many read-ahead pages that the VM is
1325          * reclaiming it before we get to it. */
1326         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1327                 zero = 1;
1328                 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
1329         } else if (!hit && ras->ras_window_len &&
1330                    index < ras->ras_next_readahead &&
1331                    index_in_window(index, ras->ras_window_start, 0,
1332                                    ras->ras_window_len)) {
1333                 zero = 1;
1334                 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1335         }
1336
1337         /* On the second access to a file smaller than the tunable
1338          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1339          * file up to ra_max_pages.  This is simply a best effort and
1340          * only occurs once per open file.  Normal RA behavior is reverted
1341          * to for subsequent IO.  The mmap case does not increment
1342          * ras_requests and thus can never trigger this behavior. */
1343         if (ras->ras_requests == 2 && !ras->ras_request_index) {
1344                 __u64 kms_pages;
1345
1346                 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1347                             CFS_PAGE_SHIFT;
1348
1349                 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1350                        ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages);
1351
1352                 if (kms_pages &&
1353                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1354                         ras->ras_window_start = 0;
1355                         ras->ras_last_readpage = 0;
1356                         ras->ras_next_readahead = 0;
1357                         ras->ras_window_len = min(ra->ra_max_pages,
1358                                 ra->ra_max_read_ahead_whole_pages);
1359                         GOTO(out_unlock, 0);
1360                 }
1361         }
1362
1363         if (zero) {
1364                 ras_reset(ras, index);
1365                 GOTO(out_unlock, 0);
1366         }
1367
1368         ras->ras_last_readpage = index;
1369         ras->ras_consecutive_pages++;
1370         ras_set_start(ras, index);
1371         ras->ras_next_readahead = max(ras->ras_window_start,
1372                                       ras->ras_next_readahead);
1373
1374         /* Trigger RA in the mmap case where ras_consecutive_requests
1375          * is not incremented and thus can't be used to trigger RA */
1376         if (!ras->ras_window_len && ras->ras_consecutive_pages == 3) {
1377                 ras->ras_window_len = 1024 * 1024 >> CFS_PAGE_SHIFT;
1378                 GOTO(out_unlock, 0);
1379         }
1380
1381         /* The initial ras_window_len is set to the request size.  To avoid
1382          * uselessly reading and discarding pages for random IO the window is
1383          * only increased once per consecutive request received. */
1384         if (ras->ras_consecutive_requests > 1 && !ras->ras_request_index) {
1385                 ras->ras_window_len = min(ras->ras_window_len +
1386                                           (1024 * 1024 >> CFS_PAGE_SHIFT),
1387                                           ra->ra_max_pages);
1388         }
1389
1390         EXIT;
1391 out_unlock:
1392         RAS_CDEBUG(ras);
1393         ras->ras_request_index++;
1394         spin_unlock(&ras->ras_lock);
1395         spin_unlock(&sbi->ll_lock);
1396         return;
1397 }
1398
1399 int ll_writepage(struct page *page)
1400 {
1401         struct inode *inode = page->mapping->host;
1402         struct ll_inode_info *lli = ll_i2info(inode);
1403         struct obd_export *exp;
1404         struct ll_async_page *llap;
1405         int rc = 0;
1406         ENTRY;
1407
1408         LASSERT(!PageDirty(page));
1409         LASSERT(PageLocked(page));
1410
1411         exp = ll_i2dtexp(inode);
1412         if (exp == NULL)
1413                 GOTO(out, rc = -EINVAL);
1414
1415         llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
1416         if (IS_ERR(llap))
1417                 GOTO(out, rc = PTR_ERR(llap));
1418
1419         LASSERT(!PageWriteback(page));
1420         set_page_writeback(page);
1421
1422         page_cache_get(page);
1423         if (llap->llap_write_queued) {
1424                 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
1425                 rc = obd_set_async_flags(exp, lli->lli_smd, NULL,
1426                                          llap->llap_cookie,
1427                                          ASYNC_READY | ASYNC_URGENT);
1428         } else {
1429                 rc = queue_or_sync_write(exp, inode, llap, CFS_PAGE_SIZE,
1430                                          ASYNC_READY | ASYNC_URGENT);
1431         }
1432         if (rc)
1433                 page_cache_release(page);
1434 out:
1435         if (rc) {
1436                 if (!lli->lli_async_rc)
1437                         lli->lli_async_rc = rc;
1438                 /* re-dirty page on error so it retries write */
1439                 if (PageWriteback(page)) {
1440                         end_page_writeback(page);
1441                 }
1442                 ll_redirty_page(page);
1443                 unlock_page(page);
1444         }
1445         RETURN(rc);
1446 }
1447
1448 /*
1449  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
1450  * read-ahead assumes it is valid to issue readpage all the way up to
1451  * i_size, but our dlm locks make that not the case.  We disable the
1452  * kernel's read-ahead and do our own by walking ahead in the page cache
1453  * checking for dlm lock coverage.  the main difference between 2.4 and
1454  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1455  * so they look the same.
1456  */
1457 int ll_readpage(struct file *filp, struct page *page)
1458 {
1459         struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
1460         struct inode *inode = page->mapping->host;
1461         struct obd_export *exp;
1462         struct ll_async_page *llap;
1463         struct obd_io_group *oig = NULL;
1464         int rc;
1465         ENTRY;
1466
1467         LASSERT(PageLocked(page));
1468         LASSERT(!PageUptodate(page));
1469         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset=%Lu=%#Lx\n",
1470                inode->i_ino, inode->i_generation, inode,
1471                (((loff_t)page->index) << CFS_PAGE_SHIFT),
1472                (((loff_t)page->index) << CFS_PAGE_SHIFT));
1473         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1474
1475         if (!ll_i2info(inode)->lli_smd) {
1476                 /* File with no objects - one big hole */
1477                 /* We use this just for remove_from_page_cache that is not
1478                  * exported, we'd make page back up to date. */
1479                 ll_truncate_complete_page(page);
1480                 clear_page(kmap(page));
1481                 kunmap(page);
1482                 SetPageUptodate(page);
1483                 unlock_page(page);
1484                 RETURN(0);
1485         }
1486
1487         rc = oig_init(&oig);
1488         if (rc < 0)
1489                 GOTO(out, rc);
1490
1491         exp = ll_i2dtexp(inode);
1492         if (exp == NULL)
1493                 GOTO(out, rc = -EINVAL);
1494
1495         llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1496         if (IS_ERR(llap))
1497                 GOTO(out, rc = PTR_ERR(llap));
1498
1499         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1500                 ras_update(ll_i2sbi(inode), inode, &fd->fd_ras, page->index,
1501                            llap->llap_defer_uptodate);
1502
1503         if (llap->llap_defer_uptodate) {
1504                 llap->llap_ra_used = 1;
1505                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1506                                   fd->fd_flags);
1507                 if (rc > 0)
1508                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1509                                              NULL, oig);
1510                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1511                 SetPageUptodate(page);
1512                 unlock_page(page);
1513                 GOTO(out_oig, rc = 0);
1514         }
1515
1516         if (likely((fd->fd_flags & LL_FILE_IGNORE_LOCK) == 0)) {
1517                 rc = ll_page_matches(page, fd->fd_flags);
1518                 if (rc < 0) {
1519                         LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1520                         GOTO(out, rc);
1521                 }
1522
1523                 if (rc == 0) {
1524                         CWARN("ino %lu page %lu (%llu) not covered by "
1525                               "a lock (mmap?).  check debug logs.\n",
1526                               inode->i_ino, page->index,
1527                               (long long)page->index << CFS_PAGE_SHIFT);
1528                 }
1529         }
1530
1531         rc = ll_issue_page_read(exp, llap, oig, 0);
1532         if (rc)
1533                 GOTO(out, rc);
1534
1535         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1536         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1537                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1538                              fd->fd_flags);
1539
1540         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1541
1542 out:
1543         if (rc)
1544                 unlock_page(page);
1545 out_oig:
1546         if (oig != NULL)
1547                 oig_release(oig);
1548         RETURN(rc);
1549 }