lustre/llite/rw.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Lustre Lite I/O page cache routines shared by different kernel revs
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23 #ifndef AUTOCONF_INCLUDED
  24 #include <linux/config.h>
  25 #endif
  26 #include <linux/kernel.h>
  27 #include <linux/mm.h>
  28 #include <linux/string.h>
  29 #include <linux/stat.h>
  30 #include <linux/errno.h>
  31 #include <linux/smp_lock.h>
  32 #include <linux/unistd.h>
  33 #include <linux/version.h>
  34 #include <asm/system.h>
  35 #include <asm/uaccess.h>
  36
  37 #include <linux/fs.h>
  38 #include <linux/stat.h>
  39 #include <asm/uaccess.h>
  40 #include <linux/mm.h>
  41 #include <linux/pagemap.h>
  42 #include <linux/smp_lock.h>
  43
  44 #define DEBUG_SUBSYSTEM S_LLITE
  45
  46 #include <lustre_lite.h>
  47 #include "llite_internal.h"
  48 #include <linux/lustre_compat25.h>
  49
  50 #ifndef list_for_each_prev_safe
  51 #define list_for_each_prev_safe(pos, n, head) \
  52         for (pos = (head)->prev, n = pos->prev; pos != (head); \
  53                 pos = n, n = pos->prev )
  54 #endif
  55
  56 cfs_mem_cache_t *ll_async_page_slab = NULL;
  57 size_t ll_async_page_slab_size = 0;
  58
  59 /* SYNCHRONOUS I/O to object storage for an inode */
  60 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
  61                   struct page *page, int flags)
  62 {
  63         struct ll_inode_info *lli = ll_i2info(inode);
  64         struct lov_stripe_md *lsm = lli->lli_smd;
  65         struct obd_info oinfo = { { { 0 } } };
  66         struct brw_page pg;
  67         int rc;
  68         ENTRY;
  69
  70         pg.pg = page;
  71         pg.off = ((obd_off)page->index) << CFS_PAGE_SHIFT;
  72
  73         if ((cmd & OBD_BRW_WRITE) && (pg.off+CFS_PAGE_SIZE>i_size_read(inode)))
  74                 pg.count = i_size_read(inode) % CFS_PAGE_SIZE;
  75         else
  76                 pg.count = CFS_PAGE_SIZE;
  77
  78         LL_CDEBUG_PAGE(D_PAGE, page, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
  79                        cmd & OBD_BRW_WRITE ? "write" : "read", pg.count,
  80                        inode->i_ino, pg.off, pg.off);
  81         if (pg.count == 0) {
  82                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
  83                        LPU64"\n", inode->i_ino, inode, i_size_read(inode),
  84                        page->mapping->host, i_size_read(page->mapping->host),
  85                        page->index, pg.off);
  86         }
  87
  88         pg.flag = flags;
  89
  90         if (cmd & OBD_BRW_WRITE)
  91                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_WRITE,
  92                                    pg.count);
  93         else
  94                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_READ,
  95                            pg.count);
  96         oinfo.oi_oa = oa;
  97         oinfo.oi_md = lsm;
  98         rc = obd_brw(cmd, ll_i2obdexp(inode), &oinfo, 1, &pg, NULL);
  99         if (rc == 0)
 100                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
 101         else if (rc != -EIO)
 102                 CERROR("error from obd_brw: rc = %d\n", rc);
 103         RETURN(rc);
 104 }
 105
 106 int ll_file_punch(struct inode * inode, loff_t new_size, int srvlock)
 107 {
 108         struct ll_inode_info *lli = ll_i2info(inode);
 109         struct obd_info oinfo = { { { 0 } } };
 110         struct obdo oa;
 111         int rc;
 112
 113         ENTRY;
 114         CDEBUG(D_INFO, "calling punch for "LPX64" (new size %Lu=%#Lx)\n",
 115                lli->lli_smd->lsm_object_id, new_size, new_size);
 116
 117         oinfo.oi_md = lli->lli_smd;
 118         oinfo.oi_policy.l_extent.start = new_size;
 119         oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
 120         oinfo.oi_oa = &oa;
 121         oa.o_id = lli->lli_smd->lsm_object_id;
 122         oa.o_valid = OBD_MD_FLID;
 123         oa.o_flags = srvlock ? OBD_FL_TRUNCLOCK : 0;
 124         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |OBD_MD_FLFID|
 125                         OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 126                         OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGENER |
 127                         OBD_MD_FLBLOCKS);
 128         rc = obd_punch_rqset(ll_i2obdexp(inode), &oinfo, NULL);
 129         if (rc) {
 130                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 131                 RETURN(rc);
 132         }
 133         obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 134                       OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 135         RETURN(0);
 136 }
 137 /* this isn't where truncate starts.   roughly:
 138  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
 139  * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
 140  * avoid races.
 141  *
 142  * must be called under ->lli_size_sem */
 143 void ll_truncate(struct inode *inode)
 144 {
 145         struct ll_inode_info *lli = ll_i2info(inode);
 146         int srvlock = test_bit(LLI_F_SRVLOCK, &lli->lli_flags);
 147         loff_t new_size;
 148         ENTRY;
 149         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",inode->i_ino,
 150                inode->i_generation, inode, i_size_read(inode), i_size_read(inode));
 151
 152         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
 153         if (lli->lli_size_sem_owner != current) {
 154                 EXIT;
 155                 return;
 156         }
 157
 158         if (!lli->lli_smd) {
 159                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 160                        inode->i_ino);
 161                 GOTO(out_unlock, 0);
 162         }
 163
 164         LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
 165
 166         if (!srvlock) {
 167                 struct ost_lvb lvb;
 168                 int rc;
 169
 170                 /* XXX I'm pretty sure this is a hack to paper over a more fundamental
 171                  * race condition. */
 172                 lov_stripe_lock(lli->lli_smd);
 173                 inode_init_lvb(inode, &lvb);
 174                 rc = obd_merge_lvb(ll_i2obdexp(inode), lli->lli_smd, &lvb, 0);
 175                 inode->i_blocks = lvb.lvb_blocks;
 176                 if (lvb.lvb_size == i_size_read(inode) && rc == 0) {
 177                         CDEBUG(D_VFSTRACE, "skipping punch for obj "LPX64", %Lu=%#Lx\n",
 178                                lli->lli_smd->lsm_object_id, i_size_read(inode),
 179                                i_size_read(inode));
 180                         lov_stripe_unlock(lli->lli_smd);
 181                         GOTO(out_unlock, 0);
 182                 }
 183
 184                 obd_adjust_kms(ll_i2obdexp(inode), lli->lli_smd,
 185                                i_size_read(inode), 1);
 186                 lov_stripe_unlock(lli->lli_smd);
 187         }
 188
 189         if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_LLITE_CHECKSUM) &&
 190                      (i_size_read(inode) & ~CFS_PAGE_MASK))) {
 191                 /* If the truncate leaves a partial page, update its checksum */
 192                 struct page *page = find_get_page(inode->i_mapping,
 193                                                   i_size_read(inode) >>
 194                                                   CFS_PAGE_SHIFT);
 195                 if (page != NULL) {
 196                         struct ll_async_page *llap = llap_cast_private(page);
 197                         if (llap != NULL) {
 198                                 char *kaddr = kmap_atomic(page, KM_USER0);
 199                                 llap->llap_checksum =
 200                                         init_checksum(OSC_DEFAULT_CKSUM);
 201                                 llap->llap_checksum =
 202                                         compute_checksum(llap->llap_checksum,
 203                                                          kaddr, CFS_PAGE_SIZE,
 204                                                          OSC_DEFAULT_CKSUM);
 205                                 kunmap_atomic(kaddr, KM_USER0);
 206                         }
 207                         page_cache_release(page);
 208                 }
 209         }
 210
 211         new_size = i_size_read(inode);
 212         ll_inode_size_unlock(inode, 0);
 213         if (!srvlock)
 214                 ll_file_punch(inode, new_size, 0);
 215
 216         EXIT;
 217         return;
 218
 219  out_unlock:
 220         ll_inode_size_unlock(inode, 0);
 221 } /* ll_truncate */
 222
 223 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 224                      unsigned to)
 225 {
 226         struct inode *inode = page->mapping->host;
 227         struct ll_inode_info *lli = ll_i2info(inode);
 228         struct lov_stripe_md *lsm = lli->lli_smd;
 229         obd_off offset = ((obd_off)page->index) << CFS_PAGE_SHIFT;
 230         struct obd_info oinfo = { { { 0 } } };
 231         struct brw_page pga;
 232         struct obdo oa;
 233         struct ost_lvb lvb;
 234         int rc = 0;
 235         ENTRY;
 236
 237         LASSERT(PageLocked(page));
 238         (void)llap_cast_private(page); /* assertion */
 239
 240         /* Check to see if we should return -EIO right away */
 241         pga.pg = page;
 242         pga.off = offset;
 243         pga.count = CFS_PAGE_SIZE;
 244         pga.flag = 0;
 245
 246         oa.o_mode = inode->i_mode;
 247         oa.o_id = lsm->lsm_object_id;
 248         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
 249         obdo_from_inode(&oa, inode, OBD_MD_FLFID | OBD_MD_FLGENER);
 250
 251         oinfo.oi_oa = &oa;
 252         oinfo.oi_md = lsm;
 253         rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), &oinfo, 1, &pga, NULL);
 254         if (rc)
 255                 RETURN(rc);
 256
 257         if (PageUptodate(page)) {
 258                 LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n");
 259                 RETURN(0);
 260         }
 261
 262         /* We're completely overwriting an existing page, so _don't_ set it up
 263          * to date until commit_write */
 264         if (from == 0 && to == CFS_PAGE_SIZE) {
 265                 LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n");
 266                 POISON_PAGE(page, 0x11);
 267                 RETURN(0);
 268         }
 269
 270         /* If are writing to a new page, no need to read old data.  The extent
 271          * locking will have updated the KMS, and for our purposes here we can
 272          * treat it like i_size. */
 273         lov_stripe_lock(lsm);
 274         inode_init_lvb(inode, &lvb);
 275         obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
 276         lov_stripe_unlock(lsm);
 277         if (lvb.lvb_size <= offset) {
 278                 char *kaddr = kmap_atomic(page, KM_USER0);
 279                 LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n",
 280                                lvb.lvb_size, offset);
 281                 memset(kaddr, 0, CFS_PAGE_SIZE);
 282                 kunmap_atomic(kaddr, KM_USER0);
 283                 GOTO(prepare_done, rc = 0);
 284         }
 285
 286         /* XXX could be an async ocp read.. read-ahead? */
 287         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 288         if (rc == 0) {
 289                 /* bug 1598: don't clobber blksize */
 290                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 291                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 292         }
 293
 294         EXIT;
 295  prepare_done:
 296         if (rc == 0)
 297                 SetPageUptodate(page);
 298
 299         return rc;
 300 }
 301
 302 static int ll_ap_make_ready(void *data, int cmd)
 303 {
 304         struct ll_async_page *llap;
 305         struct page *page;
 306         ENTRY;
 307
 308         llap = LLAP_FROM_COOKIE(data);
 309         page = llap->llap_page;
 310
 311         LASSERTF(!(cmd & OBD_BRW_READ), "cmd %x page %p ino %lu index %lu\n", cmd, page,
 312                  page->mapping->host->i_ino, page->index);
 313
 314         /* we're trying to write, but the page is locked.. come back later */
 315         if (TryLockPage(page))
 316                 RETURN(-EAGAIN);
 317
 318         LASSERT(!PageWriteback(page));
 319
 320         /* if we left PageDirty we might get another writepage call
 321          * in the future.  list walkers are bright enough
 322          * to check page dirty so we can leave it on whatever list
 323          * its on.  XXX also, we're called with the cli list so if
 324          * we got the page cache list we'd create a lock inversion
 325          * with the removepage path which gets the page lock then the
 326          * cli lock */
 327 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,0)
 328         clear_page_dirty(page);
 329 #else
 330         LASSERTF(!PageWriteback(page),"cmd %x page %p ino %lu index %lu\n", cmd, page,
 331                  page->mapping->host->i_ino, page->index);
 332         clear_page_dirty_for_io(page);
 333
 334         /* This actually clears the dirty bit in the radix tree.*/
 335         set_page_writeback(page);
 336 #endif
 337
 338         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 339         page_cache_get(page);
 340
 341         RETURN(0);
 342 }
 343
 344 /* We have two reasons for giving llite the opportunity to change the
 345  * write length of a given queued page as it builds the RPC containing
 346  * the page:
 347  *
 348  * 1) Further extending writes may have landed in the page cache
 349  *    since a partial write first queued this page requiring us
 350  *    to write more from the page cache.  (No further races are possible, since
 351  *    by the time this is called, the page is locked.)
 352  * 2) We might have raced with truncate and want to avoid performing
 353  *    write RPCs that are just going to be thrown away by the
 354  *    truncate's punch on the storage targets.
 355  *
 356  * The kms serves these purposes as it is set at both truncate and extending
 357  * writes.
 358  */
 359 static int ll_ap_refresh_count(void *data, int cmd)
 360 {
 361         struct ll_inode_info *lli;
 362         struct ll_async_page *llap;
 363         struct lov_stripe_md *lsm;
 364         struct page *page;
 365         struct inode *inode;
 366         struct ost_lvb lvb;
 367         __u64 kms;
 368         ENTRY;
 369
 370         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 371         LASSERT(cmd != OBD_BRW_READ);
 372
 373         llap = LLAP_FROM_COOKIE(data);
 374         page = llap->llap_page;
 375         inode = page->mapping->host;
 376         lli = ll_i2info(inode);
 377         lsm = lli->lli_smd;
 378
 379         lov_stripe_lock(lsm);
 380         inode_init_lvb(inode, &lvb);
 381         obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
 382         kms = lvb.lvb_size;
 383         lov_stripe_unlock(lsm);
 384
 385         /* catch race with truncate */
 386         if (((__u64)page->index << CFS_PAGE_SHIFT) >= kms)
 387                 return 0;
 388
 389         /* catch sub-page write at end of file */
 390         if (((__u64)page->index << CFS_PAGE_SHIFT) + CFS_PAGE_SIZE > kms)
 391                 return kms % CFS_PAGE_SIZE;
 392
 393         return CFS_PAGE_SIZE;
 394 }
 395
 396 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 397 {
 398         struct lov_stripe_md *lsm;
 399         obd_flag valid_flags;
 400
 401         lsm = ll_i2info(inode)->lli_smd;
 402
 403         oa->o_id = lsm->lsm_object_id;
 404         oa->o_valid = OBD_MD_FLID;
 405         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 406         if (cmd & OBD_BRW_WRITE) {
 407                 oa->o_valid |= OBD_MD_FLEPOCH;
 408                 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
 409
 410                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 411                         OBD_MD_FLUID | OBD_MD_FLGID |
 412                         OBD_MD_FLFID | OBD_MD_FLGENER;
 413         }
 414
 415         obdo_from_inode(oa, inode, valid_flags);
 416 }
 417
 418 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 419 {
 420         struct ll_async_page *llap;
 421         ENTRY;
 422
 423         llap = LLAP_FROM_COOKIE(data);
 424         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 425
 426         EXIT;
 427 }
 428
 429 static void ll_ap_update_obdo(void *data, int cmd, struct obdo *oa,
 430                               obd_valid valid)
 431 {
 432         struct ll_async_page *llap;
 433         ENTRY;
 434
 435         llap = LLAP_FROM_COOKIE(data);
 436         obdo_from_inode(oa, llap->llap_page->mapping->host, valid);
 437
 438         EXIT;
 439 }
 440
 441 static struct obd_async_page_ops ll_async_page_ops = {
 442         .ap_make_ready =        ll_ap_make_ready,
 443         .ap_refresh_count =     ll_ap_refresh_count,
 444         .ap_fill_obdo =         ll_ap_fill_obdo,
 445         .ap_update_obdo =       ll_ap_update_obdo,
 446         .ap_completion =        ll_ap_completion,
 447 };
 448
 449 struct ll_async_page *llap_cast_private(struct page *page)
 450 {
 451         struct ll_async_page *llap = (struct ll_async_page *)page_private(page);
 452
 453         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 454                  "page %p private %lu gave magic %d which != %d\n",
 455                  page, page_private(page), llap->llap_magic, LLAP_MAGIC);
 456
 457         return llap;
 458 }
 459
 460 /* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction.
 461  *
 462  * There is an llap attached onto every page in lustre, linked off @sbi.
 463  * We add an llap to the list so we don't lose our place during list walking.
 464  * If llaps in the list are being moved they will only move to the end
 465  * of the LRU, and we aren't terribly interested in those pages here (we
 466  * start at the beginning of the list where the least-used llaps are.
 467  */
 468 int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction)
 469 {
 470         struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a };
 471         unsigned long total, want, count = 0;
 472
 473         total = sbi->ll_async_page_count;
 474
 475         /* There can be a large number of llaps (600k or more in a large
 476          * memory machine) so the VM 1/6 shrink ratio is likely too much.
 477          * Since we are freeing pages also, we don't necessarily want to
 478          * shrink so much.  Limit to 40MB of pages + llaps per call. */
 479         if (shrink_fraction == 0)
 480                 want = sbi->ll_async_page_count - sbi->ll_async_page_max + 32;
 481         else
 482                 want = (total + shrink_fraction - 1) / shrink_fraction;
 483
 484         if (want > 40 << (20 - CFS_PAGE_SHIFT))
 485                 want = 40 << (20 - CFS_PAGE_SHIFT);
 486
 487         CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n",
 488                want, total, shrink_fraction);
 489
 490         spin_lock(&sbi->ll_lock);
 491         list_add(&dummy_llap.llap_pglist_item, &sbi->ll_pglist);
 492
 493         while (--total >= 0 && count < want) {
 494                 struct page *page;
 495                 int keep;
 496
 497                 if (unlikely(need_resched())) {
 498                         spin_unlock(&sbi->ll_lock);
 499                         cond_resched();
 500                         spin_lock(&sbi->ll_lock);
 501                 }
 502
 503                 llap = llite_pglist_next_llap(sbi,&dummy_llap.llap_pglist_item);
 504                 list_del_init(&dummy_llap.llap_pglist_item);
 505                 if (llap == NULL)
 506                         break;
 507
 508                 page = llap->llap_page;
 509                 LASSERT(page != NULL);
 510
 511                 list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item);
 512
 513                 /* Page needs/undergoing IO */
 514                 if (TryLockPage(page)) {
 515                         LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n");
 516                         continue;
 517                 }
 518
 519                keep = (llap->llap_write_queued || PageDirty(page) ||
 520                       PageWriteback(page) || (!PageUptodate(page) &&
 521                       llap->llap_origin != LLAP_ORIGIN_READAHEAD));
 522
 523                 LL_CDEBUG_PAGE(D_PAGE, page,"%s LRU page: %s%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                                PageWriteback(page) ? "wb" : "",
 529                                llap->llap_defer_uptodate ? "" : "!du",
 530                                llap_origins[llap->llap_origin]);
 531
 532                 /* If page is dirty or undergoing IO don't discard it */
 533                 if (keep) {
 534                         unlock_page(page);
 535                         continue;
 536                 }
 537
 538                 page_cache_get(page);
 539                 spin_unlock(&sbi->ll_lock);
 540
 541                 if (page->mapping != NULL) {
 542                         ll_teardown_mmaps(page->mapping,
 543                                          (__u64)page->index << CFS_PAGE_SHIFT,
 544                                          ((__u64)page->index << CFS_PAGE_SHIFT)|
 545                                           ~CFS_PAGE_MASK);
 546                         if (!PageDirty(page) && !page_mapped(page)) {
 547                                 ll_ra_accounting(llap, page->mapping);
 548                                 ll_truncate_complete_page(page);
 549                                 ++count;
 550                         } else {
 551                                 LL_CDEBUG_PAGE(D_PAGE, page, "Not dropping page"
 552                                                              " because it is "
 553                                                              "%s\n",
 554                                                               PageDirty(page)?
 555                                                               "dirty":"mapped");
 556                         }
 557                 }
 558                 unlock_page(page);
 559                 page_cache_release(page);
 560
 561                 spin_lock(&sbi->ll_lock);
 562         }
 563         list_del(&dummy_llap.llap_pglist_item);
 564         spin_unlock(&sbi->ll_lock);
 565
 566         CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n",
 567                count, want, total);
 568
 569         return count;
 570 }
 571
 572 static struct ll_async_page *llap_from_page_with_lockh(struct page *page,
 573                                                        unsigned origin,
 574                                                        struct lustre_handle *lockh)
 575 {
 576         struct ll_async_page *llap;
 577         struct obd_export *exp;
 578         struct inode *inode = page->mapping->host;
 579         struct ll_sb_info *sbi;
 580         int rc;
 581         ENTRY;
 582
 583         if (!inode) {
 584                 static int triggered;
 585
 586                 if (!triggered) {
 587                         LL_CDEBUG_PAGE(D_ERROR, page, "Bug 10047. Wrong anon "
 588                                        "page received\n");
 589                         libcfs_debug_dumpstack(NULL);
 590                         triggered = 1;
 591                 }
 592                 RETURN(ERR_PTR(-EINVAL));
 593         }
 594         sbi = ll_i2sbi(inode);
 595         LASSERT(ll_async_page_slab);
 596         LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
 597
 598         llap = llap_cast_private(page);
 599         if (llap != NULL) {
 600                 /* move to end of LRU list, except when page is just about to
 601                  * die */
 602                 if (origin != LLAP_ORIGIN_REMOVEPAGE) {
 603                         spin_lock(&sbi->ll_lock);
 604                         sbi->ll_pglist_gen++;
 605                         list_del_init(&llap->llap_pglist_item);
 606                         list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 607                         spin_unlock(&sbi->ll_lock);
 608                 }
 609                 GOTO(out, llap);
 610         }
 611
 612         exp = ll_i2obdexp(page->mapping->host);
 613         if (exp == NULL)
 614                 RETURN(ERR_PTR(-EINVAL));
 615
 616         /* limit the number of lustre-cached pages */
 617         if (sbi->ll_async_page_count >= sbi->ll_async_page_max)
 618                 llap_shrink_cache(sbi, 0);
 619
 620         OBD_SLAB_ALLOC(llap, ll_async_page_slab, CFS_ALLOC_STD,
 621                        ll_async_page_slab_size);
 622         if (llap == NULL)
 623                 RETURN(ERR_PTR(-ENOMEM));
 624         llap->llap_magic = LLAP_MAGIC;
 625         llap->llap_cookie = (void *)llap + size_round(sizeof(*llap));
 626
 627         /* XXX: for bug 11270 - check for lockless origin here! */
 628         if (origin == LLAP_ORIGIN_LOCKLESS_IO)
 629                 llap->llap_nocache = 1;
 630
 631         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 632                                  (obd_off)page->index << CFS_PAGE_SHIFT,
 633                                  &ll_async_page_ops, llap, &llap->llap_cookie,
 634                                  llap->llap_nocache, lockh);
 635         if (rc) {
 636                 OBD_SLAB_FREE(llap, ll_async_page_slab,
 637                               ll_async_page_slab_size);
 638                 RETURN(ERR_PTR(rc));
 639         }
 640
 641         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 642                page, llap->llap_cookie, (obd_off)page->index << CFS_PAGE_SHIFT);
 643         /* also zeroing the PRIVBITS low order bitflags */
 644         __set_page_ll_data(page, llap);
 645         llap->llap_page = page;
 646
 647         spin_lock(&sbi->ll_lock);
 648         sbi->ll_pglist_gen++;
 649         sbi->ll_async_page_count++;
 650         list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 651         spin_unlock(&sbi->ll_lock);
 652
 653  out:
 654         if (unlikely(sbi->ll_flags & LL_SBI_LLITE_CHECKSUM)) {
 655                 __u32 csum;
 656                 char *kaddr = kmap_atomic(page, KM_USER0);
 657                 csum = init_checksum(OSC_DEFAULT_CKSUM);
 658                 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
 659                                         OSC_DEFAULT_CKSUM);
 660                 kunmap_atomic(kaddr, KM_USER0);
 661                 if (origin == LLAP_ORIGIN_READAHEAD ||
 662                     origin == LLAP_ORIGIN_READPAGE ||
 663                     origin == LLAP_ORIGIN_LOCKLESS_IO) {
 664                         llap->llap_checksum = 0;
 665                 } else if (origin == LLAP_ORIGIN_COMMIT_WRITE ||
 666                            llap->llap_checksum == 0) {
 667                         llap->llap_checksum = csum;
 668                         CDEBUG(D_PAGE, "page %p cksum %x\n", page, csum);
 669                 } else if (llap->llap_checksum == csum) {
 670                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 671                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 672                                page, csum);
 673                 } else {
 674                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 675                         LL_CDEBUG_PAGE(D_ERROR, page, "old cksum %x != new "
 676                                        "%x!\n", llap->llap_checksum, csum);
 677                 }
 678         }
 679
 680         llap->llap_origin = origin;
 681         RETURN(llap);
 682 }
 683
 684 static inline struct ll_async_page *llap_from_page(struct page *page,
 685                                                    unsigned origin)
 686 {
 687         return llap_from_page_with_lockh(page, origin, NULL);
 688 }
 689
 690 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
 691                                struct ll_async_page *llap,
 692                                unsigned to, obd_flag async_flags)
 693 {
 694         unsigned long size_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
 695         struct obd_io_group *oig;
 696         struct ll_sb_info *sbi = ll_i2sbi(inode);
 697         int rc, noquot = llap->llap_ignore_quota ? OBD_BRW_NOQUOTA : 0;
 698         ENTRY;
 699
 700         /* _make_ready only sees llap once we've unlocked the page */
 701         llap->llap_write_queued = 1;
 702         rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
 703                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 704                                 0, 0, 0, async_flags);
 705         if (rc == 0) {
 706                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
 707                 llap_write_pending(inode, llap);
 708                 GOTO(out, 0);
 709         }
 710
 711         llap->llap_write_queued = 0;
 712
 713         rc = oig_init(&oig);
 714         if (rc)
 715                 GOTO(out, rc);
 716
 717         /* make full-page requests if we are not at EOF (bug 4410) */
 718         if (to != CFS_PAGE_SIZE && llap->llap_page->index < size_index) {
 719                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 720                                "sync write before EOF: size_index %lu, to %d\n",
 721                                size_index, to);
 722                 to = CFS_PAGE_SIZE;
 723         } else if (to != CFS_PAGE_SIZE && llap->llap_page->index == size_index){
 724                 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
 725                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 726                                "sync write at EOF: size_index %lu, to %d/%d\n",
 727                                size_index, to, size_to);
 728                 if (to < size_to)
 729                         to = size_to;
 730         }
 731
 732         /* compare the checksum once before the page leaves llite */
 733         if (unlikely((sbi->ll_flags & LL_SBI_LLITE_CHECKSUM) &&
 734                      llap->llap_checksum != 0)) {
 735                 __u32 csum;
 736                 struct page *page = llap->llap_page;
 737                 char *kaddr = kmap_atomic(page, KM_USER0);
 738                 csum = init_checksum(OSC_DEFAULT_CKSUM);
 739                 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
 740                                         OSC_DEFAULT_CKSUM);
 741                 kunmap_atomic(kaddr, KM_USER0);
 742                 if (llap->llap_checksum == csum) {
 743                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 744                                page, csum);
 745                 } else {
 746                         CERROR("page %p old cksum %x != new cksum %x!\n",
 747                                page, llap->llap_checksum, csum);
 748                 }
 749         }
 750
 751         rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
 752                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 753                                 0, to, 0, ASYNC_READY | ASYNC_URGENT |
 754                                 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
 755         if (rc)
 756                 GOTO(free_oig, rc);
 757
 758         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
 759         if (rc)
 760                 GOTO(free_oig, rc);
 761
 762         rc = oig_wait(oig);
 763
 764         if (!rc && async_flags & ASYNC_READY) {
 765                 unlock_page(llap->llap_page);
 766                 if (PageWriteback(llap->llap_page)) {
 767                         end_page_writeback(llap->llap_page);
 768                 }
 769         }
 770
 771         LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
 772
 773 free_oig:
 774         oig_release(oig);
 775 out:
 776         RETURN(rc);
 777 }
 778
 779 /* update our write count to account for i_size increases that may have
 780  * happened since we've queued the page for io. */
 781
 782 /* be careful not to return success without setting the page Uptodate or
 783  * the next pass through prepare_write will read in stale data from disk. */
 784 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 785                     unsigned to)
 786 {
 787         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
 788         struct inode *inode = page->mapping->host;
 789         struct ll_inode_info *lli = ll_i2info(inode);
 790         struct lov_stripe_md *lsm = lli->lli_smd;
 791         struct obd_export *exp;
 792         struct ll_async_page *llap;
 793         loff_t size;
 794         struct lustre_handle *lockh = NULL;
 795         int rc = 0;
 796         ENTRY;
 797
 798         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 799         LASSERT(inode == file->f_dentry->d_inode);
 800         LASSERT(PageLocked(page));
 801
 802         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 803                inode, page, from, to, page->index);
 804
 805         if (fd->fd_flags & LL_FILE_GROUP_LOCKED)
 806                 lockh = &fd->fd_cwlockh;
 807
 808         llap = llap_from_page_with_lockh(page, LLAP_ORIGIN_COMMIT_WRITE, lockh);
 809         if (IS_ERR(llap))
 810                 RETURN(PTR_ERR(llap));
 811
 812         exp = ll_i2obdexp(inode);
 813         if (exp == NULL)
 814                 RETURN(-EINVAL);
 815
 816         llap->llap_ignore_quota = capable(CAP_SYS_RESOURCE);
 817
 818         /* queue a write for some time in the future the first time we
 819          * dirty the page */
 820         if (!PageDirty(page)) {
 821                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_MISSES, 1);
 822
 823                 rc = queue_or_sync_write(exp, inode, llap, to, 0);
 824                 if (rc)
 825                         GOTO(out, rc);
 826         } else {
 827                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_HITS, 1);
 828         }
 829
 830         /* put the page in the page cache, from now on ll_removepage is
 831          * responsible for cleaning up the llap.
 832          * only set page dirty when it's queued to be write out */
 833         if (llap->llap_write_queued)
 834                 set_page_dirty(page);
 835
 836 out:
 837         size = (((obd_off)page->index) << CFS_PAGE_SHIFT) + to;
 838         ll_inode_size_lock(inode, 0);
 839         if (rc == 0) {
 840                 lov_stripe_lock(lsm);
 841                 obd_adjust_kms(exp, lsm, size, 0);
 842                 lov_stripe_unlock(lsm);
 843                 if (size > i_size_read(inode))
 844                         i_size_write(inode, size);
 845                 SetPageUptodate(page);
 846         } else if (size > i_size_read(inode)) {
 847                 /* this page beyond the pales of i_size, so it can't be
 848                  * truncated in ll_p_r_e during lock revoking. we must
 849                  * teardown our book-keeping here. */
 850                 ll_removepage(page);
 851         }
 852         ll_inode_size_unlock(inode, 0);
 853         RETURN(rc);
 854 }
 855
 856 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 857 {
 858         struct ll_ra_info *ra = &sbi->ll_ra_info;
 859         unsigned long ret;
 860         ENTRY;
 861
 862         spin_lock(&sbi->ll_lock);
 863         ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
 864         ra->ra_cur_pages += ret;
 865         spin_unlock(&sbi->ll_lock);
 866
 867         RETURN(ret);
 868 }
 869
 870 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 871 {
 872         struct ll_ra_info *ra = &sbi->ll_ra_info;
 873         spin_lock(&sbi->ll_lock);
 874         LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
 875                  ra->ra_cur_pages, len);
 876         ra->ra_cur_pages -= len;
 877         spin_unlock(&sbi->ll_lock);
 878 }
 879
 880 /* called for each page in a completed rpc.*/
 881 int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 882 {
 883         struct ll_async_page *llap;
 884         struct page *page;
 885         int ret = 0;
 886         ENTRY;
 887
 888         llap = LLAP_FROM_COOKIE(data);
 889         page = llap->llap_page;
 890         LASSERT(PageLocked(page));
 891         LASSERT(CheckWriteback(page,cmd));
 892
 893         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 894
 895         if (cmd & OBD_BRW_READ && llap->llap_defer_uptodate)
 896                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 897
 898         if (rc == 0)  {
 899                 if (cmd & OBD_BRW_READ) {
 900                         if (!llap->llap_defer_uptodate)
 901                                 SetPageUptodate(page);
 902                 } else {
 903                         llap->llap_write_queued = 0;
 904                 }
 905                 ClearPageError(page);
 906         } else {
 907                 if (cmd & OBD_BRW_READ) {
 908                         llap->llap_defer_uptodate = 0;
 909                 }
 910                 SetPageError(page);
 911 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 912                 if (rc == -ENOSPC)
 913                         set_bit(AS_ENOSPC, &page->mapping->flags);
 914                 else
 915                         set_bit(AS_EIO, &page->mapping->flags);
 916 #else
 917                 page->mapping->gfp_mask |= AS_EIO_MASK;
 918 #endif
 919         }
 920
 921         unlock_page(page);
 922
 923         if (cmd & OBD_BRW_WRITE) {
 924                 llap_write_complete(page->mapping->host, llap);
 925                 ll_try_done_writing(page->mapping->host);
 926         }
 927
 928         if (PageWriteback(page)) {
 929                 end_page_writeback(page);
 930         }
 931         page_cache_release(page);
 932
 933         RETURN(ret);
 934 }
 935
 936 static void __ll_put_llap(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         exp = ll_i2obdexp(inode);
 946         if (exp == NULL) {
 947                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 948                 EXIT;
 949                 return;
 950         }
 951
 952         llap = llap_from_page(page, LLAP_ORIGIN_REMOVEPAGE);
 953         if (IS_ERR(llap)) {
 954                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 955                        page->index, PTR_ERR(llap));
 956                 EXIT;
 957                 return;
 958         }
 959
 960         //llap_write_complete(inode, llap);
 961         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 962                                      llap->llap_cookie);
 963         if (rc != 0)
 964                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 965
 966         /* this unconditional free is only safe because the page lock
 967          * is providing exclusivity to memory pressure/truncate/writeback..*/
 968         __clear_page_ll_data(page);
 969
 970         spin_lock(&sbi->ll_lock);
 971         if (!list_empty(&llap->llap_pglist_item))
 972                 list_del_init(&llap->llap_pglist_item);
 973         sbi->ll_pglist_gen++;
 974         sbi->ll_async_page_count--;
 975         spin_unlock(&sbi->ll_lock);
 976         OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
 977
 978         EXIT;
 979 }
 980
 981 /* the kernel calls us here when a page is unhashed from the page cache.
 982  * the page will be locked and the kernel is holding a spinlock, so
 983  * we need to be careful.  we're just tearing down our book-keeping
 984  * here. */
 985 void ll_removepage(struct page *page)
 986 {
 987         struct ll_async_page *llap = llap_cast_private(page);
 988         ENTRY;
 989
 990         LASSERT(!in_interrupt());
 991
 992         /* sync pages or failed read pages can leave pages in the page
 993          * cache that don't have our data associated with them anymore */
 994         if (page_private(page) == 0) {
 995                 EXIT;
 996                 return;
 997         }
 998
 999         LASSERT(!llap->llap_lockless_io_page);
1000         LASSERT(!llap->llap_nocache);
1001
1002         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
1003         __ll_put_llap(page);
1004
1005         EXIT;
1006 }
1007
1008 static int ll_issue_page_read(struct obd_export *exp,
1009                               struct ll_async_page *llap,
1010                               struct obd_io_group *oig, int defer)
1011 {
1012         struct page *page = llap->llap_page;
1013         int rc;
1014
1015         page_cache_get(page);
1016         llap->llap_defer_uptodate = defer;
1017         llap->llap_ra_used = 0;
1018         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
1019                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
1020                                 CFS_PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY |
1021                                               ASYNC_URGENT);
1022         if (rc) {
1023                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
1024                 page_cache_release(page);
1025         }
1026         RETURN(rc);
1027 }
1028
1029 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
1030 {
1031         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
1032         ra->ra_stats[which]++;
1033 }
1034
1035 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
1036 {
1037         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
1038         struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
1039
1040         spin_lock(&sbi->ll_lock);
1041         ll_ra_stats_inc_unlocked(ra, which);
1042         spin_unlock(&sbi->ll_lock);
1043 }
1044
1045 void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping)
1046 {
1047         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
1048                 return;
1049
1050         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
1051 }
1052
1053 #define RAS_CDEBUG(ras) \
1054         CDEBUG(D_READA,                                                      \
1055                "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu"    \
1056                "csr %lu sf %lu sp %lu sl %lu \n",                            \
1057                ras->ras_last_readpage, ras->ras_consecutive_requests,        \
1058                ras->ras_consecutive_pages, ras->ras_window_start,            \
1059                ras->ras_window_len, ras->ras_next_readahead,                 \
1060                ras->ras_requests, ras->ras_request_index,                    \
1061                ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
1062                ras->ras_stride_pages, ras->ras_stride_length)
1063
1064 static int index_in_window(unsigned long index, unsigned long point,
1065                            unsigned long before, unsigned long after)
1066 {
1067         unsigned long start = point - before, end = point + after;
1068
1069         if (start > point)
1070                start = 0;
1071         if (end < point)
1072                end = ~0;
1073
1074         return start <= index && index <= end;
1075 }
1076
1077 static struct ll_readahead_state *ll_ras_get(struct file *f)
1078 {
1079         struct ll_file_data       *fd;
1080
1081         fd = LUSTRE_FPRIVATE(f);
1082         return &fd->fd_ras;
1083 }
1084
1085 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
1086 {
1087         struct ll_readahead_state *ras;
1088
1089         ras = ll_ras_get(f);
1090
1091         spin_lock(&ras->ras_lock);
1092         ras->ras_requests++;
1093         ras->ras_request_index = 0;
1094         ras->ras_consecutive_requests++;
1095         rar->lrr_reader = current;
1096
1097         list_add(&rar->lrr_linkage, &ras->ras_read_beads);
1098         spin_unlock(&ras->ras_lock);
1099 }
1100
1101 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
1102 {
1103         struct ll_readahead_state *ras;
1104
1105         ras = ll_ras_get(f);
1106
1107         spin_lock(&ras->ras_lock);
1108         list_del_init(&rar->lrr_linkage);
1109         spin_unlock(&ras->ras_lock);
1110 }
1111
1112 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
1113 {
1114         struct ll_ra_read *scan;
1115
1116         list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
1117                 if (scan->lrr_reader == current)
1118                         return scan;
1119         }
1120         return NULL;
1121 }
1122
1123 struct ll_ra_read *ll_ra_read_get(struct file *f)
1124 {
1125         struct ll_readahead_state *ras;
1126         struct ll_ra_read         *bead;
1127
1128         ras = ll_ras_get(f);
1129
1130         spin_lock(&ras->ras_lock);
1131         bead = ll_ra_read_get_locked(ras);
1132         spin_unlock(&ras->ras_lock);
1133         return bead;
1134 }
1135
1136 static int ll_read_ahead_page(struct obd_export *exp, struct obd_io_group *oig,
1137                               int index, struct address_space *mapping)
1138 {
1139         struct ll_async_page *llap;
1140         struct page *page;
1141         unsigned int gfp_mask = 0;
1142         int rc = 0;
1143
1144         gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
1145 #ifdef __GFP_NOWARN
1146         gfp_mask |= __GFP_NOWARN;
1147 #endif
1148         page = grab_cache_page_nowait_gfp(mapping, index, gfp_mask);
1149         if (page == NULL) {
1150                 ll_ra_stats_inc(mapping, RA_STAT_FAILED_GRAB_PAGE);
1151                 CDEBUG(D_READA, "g_c_p_n failed\n");
1152                 return 0;
1153         }
1154
1155         /* Check if page was truncated or reclaimed */
1156         if (page->mapping != mapping) {
1157                 ll_ra_stats_inc(mapping, RA_STAT_WRONG_GRAB_PAGE);
1158                 CDEBUG(D_READA, "g_c_p_n returned invalid page\n");
1159                 GOTO(unlock_page, rc = 0);
1160         }
1161
1162         /* we do this first so that we can see the page in the /proc
1163          * accounting */
1164         llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
1165         if (IS_ERR(llap) || llap->llap_defer_uptodate) {
1166                 if (PTR_ERR(llap) == -ENOLCK) {
1167                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
1168                         CDEBUG(D_READA | D_PAGE,
1169                                "Adding page to cache failed index "
1170                                 "%d\n", index);
1171                                 CDEBUG(D_READA, "nolock page\n");
1172                                 GOTO(unlock_page, rc = -ENOLCK);
1173                 }
1174                 CDEBUG(D_READA, "read-ahead page\n");
1175                 GOTO(unlock_page, rc = 0);
1176         }
1177
1178         /* skip completed pages */
1179         if (Page_Uptodate(page))
1180                 GOTO(unlock_page, rc = 0);
1181
1182         /* bail out when we hit the end of the lock. */
1183         rc = ll_issue_page_read(exp, llap, oig, 1);
1184         if (rc == 0) {
1185                 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "started read-ahead\n");
1186                 rc = 1;
1187         } else {
1188 unlock_page:
1189                 unlock_page(page);
1190                 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "skipping read-ahead\n");
1191         }
1192         page_cache_release(page);
1193         return rc;
1194 }
1195
1196 /* ra_io_arg will be filled in the beginning of ll_readahead with
1197  * ras_lock, then the following ll_read_ahead_pages will read RA
1198  * pages according to this arg, all the items in this structure are
1199  * counted by page index.
1200  */
1201 struct ra_io_arg {
1202         unsigned long ria_start;  /* start offset of read-ahead*/
1203         unsigned long ria_end;    /* end offset of read-ahead*/
1204         /* If stride read pattern is detected, ria_stoff means where
1205          * stride read is started. Note: for normal read-ahead, the
1206          * value here is meaningless, and also it will not be accessed*/
1207         pgoff_t ria_stoff;
1208         /* ria_length and ria_pages are the length and pages length in the
1209          * stride I/O mode. And they will also be used to check whether
1210          * it is stride I/O read-ahead in the read-ahead pages*/
1211         unsigned long ria_length;
1212         unsigned long ria_pages;
1213 };
1214
1215 #define RIA_DEBUG(ria)                                                \
1216         CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n",       \
1217         ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
1218         ria->ria_pages)
1219
1220 #define RAS_INCREASE_STEP (1024 * 1024 >> CFS_PAGE_SHIFT)
1221
1222 static inline int stride_io_mode(struct ll_readahead_state *ras)
1223 {
1224         return ras->ras_consecutive_stride_requests > 1;
1225 }
1226
1227 /* The function calculates how much pages will be read in
1228  * [off, off + length], which will be read by stride I/O mode,
1229  * stride_offset = st_off, stride_lengh = st_len,
1230  * stride_pages = st_pgs
1231  */
1232 static unsigned long
1233 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
1234                 unsigned long off, unsigned length)
1235 {
1236         unsigned long cont_len = st_off > off ?  st_off - off : 0;
1237         unsigned long stride_len = length + off > st_off ?
1238                            length + off + 1 - st_off : 0;
1239         unsigned long left, pg_count;
1240
1241         if (st_len == 0 || length == 0)
1242                 return length;
1243
1244         left = do_div(stride_len, st_len);
1245         left = min(left, st_pgs);
1246
1247         pg_count = left + stride_len * st_pgs + cont_len;
1248
1249         LASSERT(pg_count >= left);
1250
1251         CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %u"
1252                "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
1253
1254         return pg_count;
1255 }
1256
1257 static int ria_page_count(struct ra_io_arg *ria)
1258 {
1259         __u64 length = ria->ria_end >= ria->ria_start ?
1260                        ria->ria_end - ria->ria_start + 1 : 0;
1261
1262         return stride_pg_count(ria->ria_stoff, ria->ria_length,
1263                                ria->ria_pages, ria->ria_start,
1264                                length);
1265 }
1266
1267 /*Check whether the index is in the defined ra-window */
1268 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
1269 {
1270         /* If ria_length == ria_pages, it means non-stride I/O mode,
1271          * idx should always inside read-ahead window in this case
1272          * For stride I/O mode, just check whether the idx is inside
1273          * the ria_pages. */
1274         return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
1275                (idx - ria->ria_stoff) % ria->ria_length < ria->ria_pages;
1276 }
1277
1278 static int ll_read_ahead_pages(struct obd_export *exp,
1279                                struct obd_io_group *oig,
1280                                struct ra_io_arg *ria,
1281                                unsigned long *reserved_pages,
1282                                struct address_space *mapping,
1283                                unsigned long *ra_end)
1284 {
1285         int rc, count = 0, stride_ria;
1286         unsigned long page_idx;
1287
1288         LASSERT(ria != NULL);
1289         RIA_DEBUG(ria);
1290
1291         stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
1292         for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
1293                         *reserved_pages > 0; page_idx++) {
1294                 if (ras_inside_ra_window(page_idx, ria)) {
1295                         /* If the page is inside the read-ahead window*/
1296                         rc = ll_read_ahead_page(exp, oig, page_idx, mapping);
1297                         if (rc == 1) {
1298                                 (*reserved_pages)--;
1299                                 count ++;
1300                         } else if (rc == -ENOLCK)
1301                                 break;
1302                 } else if (stride_ria) {
1303                         /* If it is not in the read-ahead window, and it is
1304                          * read-ahead mode, then check whether it should skip
1305                          * the stride gap */
1306                         pgoff_t offset;
1307                         /* FIXME: This assertion only is valid when it is for
1308                          * forward read-ahead, it will be fixed when backward
1309                          * read-ahead is implemented */
1310                         LASSERTF(page_idx > ria->ria_stoff, "since %lu in the"
1311                                 " gap of ra window,it should bigger than stride"
1312                                 " offset %lu \n", page_idx, ria->ria_stoff);
1313
1314                         offset = page_idx - ria->ria_stoff;
1315                         offset = offset % (ria->ria_length);
1316                         if (offset > ria->ria_pages) {
1317                                 page_idx += ria->ria_length - offset;
1318                                 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
1319                                        ria->ria_length - offset);
1320                                 continue;
1321                         }
1322                 }
1323         }
1324         *ra_end = page_idx;
1325         return count;
1326 }
1327
1328 static int ll_readahead(struct ll_readahead_state *ras,
1329                          struct obd_export *exp, struct address_space *mapping,
1330                          struct obd_io_group *oig, int flags)
1331 {
1332         unsigned long start = 0, end = 0, reserved;
1333         unsigned long ra_end, len;
1334         struct inode *inode;
1335         struct lov_stripe_md *lsm;
1336         struct ll_ra_read *bead;
1337         struct ost_lvb lvb;
1338         struct ra_io_arg ria = { 0 };
1339         int ret = 0;
1340         __u64 kms;
1341         ENTRY;
1342
1343         inode = mapping->host;
1344         lsm = ll_i2info(inode)->lli_smd;
1345
1346         lov_stripe_lock(lsm);
1347         inode_init_lvb(inode, &lvb);
1348         obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
1349         kms = lvb.lvb_size;
1350         lov_stripe_unlock(lsm);
1351         if (kms == 0) {
1352                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
1353                 RETURN(0);
1354         }
1355
1356         spin_lock(&ras->ras_lock);
1357         bead = ll_ra_read_get_locked(ras);
1358         /* Enlarge the RA window to encompass the full read */
1359         if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
1360             bead->lrr_start + bead->lrr_count) {
1361                 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
1362                                       ras->ras_window_start;
1363         }
1364         /* Reserve a part of the read-ahead window that we'll be issuing */
1365         if (ras->ras_window_len) {
1366                 start = ras->ras_next_readahead;
1367                 end = ras->ras_window_start + ras->ras_window_len - 1;
1368         }
1369         if (end != 0) {
1370                 /* Truncate RA window to end of file */
1371                 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
1372                 ras->ras_next_readahead = max(end, end + 1);
1373                 RAS_CDEBUG(ras);
1374         }
1375         ria.ria_start = start;
1376         ria.ria_end = end;
1377         /* If stride I/O mode is detected, get stride window*/
1378         if (stride_io_mode(ras)) {
1379                 ria.ria_length = ras->ras_stride_length;
1380                 ria.ria_pages = ras->ras_stride_pages;
1381         }
1382         spin_unlock(&ras->ras_lock);
1383
1384         if (end == 0) {
1385                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
1386                 RETURN(0);
1387         }
1388
1389         len = ria_page_count(&ria);
1390         if (len == 0)
1391                 RETURN(0);
1392
1393         reserved = ll_ra_count_get(ll_i2sbi(inode), len);
1394         if (reserved < end - start + 1)
1395                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
1396
1397         CDEBUG(D_READA, "reserved page %lu \n", reserved);
1398
1399         ret = ll_read_ahead_pages(exp, oig, &ria, &reserved, mapping, &ra_end);
1400
1401         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
1402         if (reserved != 0)
1403                 ll_ra_count_put(ll_i2sbi(inode), reserved);
1404
1405         if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
1406                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
1407
1408         /* if we didn't get to the end of the region we reserved from
1409          * the ras we need to go back and update the ras so that the
1410          * next read-ahead tries from where we left off.  we only do so
1411          * if the region we failed to issue read-ahead on is still ahead
1412          * of the app and behind the next index to start read-ahead from */
1413         CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
1414                ra_end, end, ria.ria_end);
1415
1416         if (ra_end != (end + 1)) {
1417                 spin_lock(&ras->ras_lock);
1418                 if (ra_end < ras->ras_next_readahead &&
1419                     index_in_window(ra_end, ras->ras_window_start, 0,
1420                                     ras->ras_window_len)) {
1421                         ras->ras_next_readahead = ra_end;
1422                         RAS_CDEBUG(ras);
1423                 }
1424                 spin_unlock(&ras->ras_lock);
1425         }
1426
1427         RETURN(ret);
1428 }
1429
1430 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
1431 {
1432         ras->ras_window_start = index & (~(RAS_INCREASE_STEP - 1));
1433 }
1434
1435 /* called with the ras_lock held or from places where it doesn't matter */
1436 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1437 {
1438         ras->ras_last_readpage = index;
1439         ras->ras_consecutive_requests = 0;
1440         ras->ras_consecutive_pages = 0;
1441         ras->ras_window_len = 0;
1442         ras_set_start(ras, index);
1443         ras->ras_next_readahead = max(ras->ras_window_start, index);
1444
1445         RAS_CDEBUG(ras);
1446 }
1447
1448 /* called with the ras_lock held or from places where it doesn't matter */
1449 static void ras_stride_reset(struct ll_readahead_state *ras)
1450 {
1451         ras->ras_consecutive_stride_requests = 0;
1452         RAS_CDEBUG(ras);
1453 }
1454
1455 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1456 {
1457         spin_lock_init(&ras->ras_lock);
1458         ras_reset(ras, 0);
1459         ras->ras_requests = 0;
1460         INIT_LIST_HEAD(&ras->ras_read_beads);
1461 }
1462
1463 /* Check whether the read request is in the stride window.
1464  * If it is in the stride window, return 1, otherwise return 0.
1465  * and also update stride_gap and stride_pages.
1466  */
1467 static int index_in_stride_window(unsigned long index,
1468                                   struct ll_readahead_state *ras,
1469                                   struct inode *inode)
1470 {
1471         int stride_gap = index - ras->ras_last_readpage - 1;
1472
1473         LASSERT(stride_gap != 0);
1474
1475         if (ras->ras_consecutive_pages == 0)
1476                 return 0;
1477
1478         /*Otherwise check the stride by itself */
1479         if ((ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
1480             ras->ras_consecutive_pages == ras->ras_stride_pages)
1481                 return 1;
1482
1483         if (stride_gap >= 0) {
1484                 /*
1485                  * only set stride_pages, stride_length if
1486                  * it is forward reading ( stride_gap > 0)
1487                  */
1488                 ras->ras_stride_pages = ras->ras_consecutive_pages;
1489                 ras->ras_stride_length = stride_gap + ras->ras_consecutive_pages;
1490         } else {
1491                 /*
1492                  * If stride_gap < 0,(back_forward reading),
1493                  * reset the stride_pages/length.
1494                  * FIXME:back_ward stride I/O read.
1495                  *
1496                  */
1497                 ras->ras_stride_pages = 0;
1498                 ras->ras_stride_length = 0;
1499         }
1500         RAS_CDEBUG(ras);
1501
1502         return 0;
1503 }
1504
1505 static unsigned long
1506 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
1507 {
1508         return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
1509                                ras->ras_stride_pages, ras->ras_stride_offset,
1510                                len);
1511 }
1512
1513 /* Stride Read-ahead window will be increased inc_len according to
1514  * stride I/O pattern */
1515 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1516                                        struct ll_ra_info *ra,
1517                                        unsigned long inc_len)
1518 {
1519         unsigned long left, step, window_len;
1520         unsigned long stride_len;
1521
1522         LASSERT(ras->ras_stride_length > 0);
1523
1524         stride_len = ras->ras_window_start + ras->ras_window_len -
1525                      ras->ras_stride_offset;
1526
1527         LASSERTF(stride_len >= 0, "window_start %lu, window_len %lu"
1528                  " stride_offset %lu\n", ras->ras_window_start,
1529                  ras->ras_window_len, ras->ras_stride_offset);
1530
1531         left = stride_len % ras->ras_stride_length;
1532
1533         window_len = ras->ras_window_len - left;
1534
1535         if (left < ras->ras_stride_pages)
1536                 left += inc_len;
1537         else
1538                 left = ras->ras_stride_pages + inc_len;
1539
1540         LASSERT(ras->ras_stride_pages != 0);
1541
1542         step = left / ras->ras_stride_pages;
1543         left %= ras->ras_stride_pages;
1544
1545         window_len += step * ras->ras_stride_length + left;
1546
1547         if (stride_page_count(ras, window_len) <= ra->ra_max_pages)
1548                 ras->ras_window_len = window_len;
1549
1550         RAS_CDEBUG(ras);
1551 }
1552
1553 /* Set stride I/O read-ahead window start offset */
1554 static void ras_set_stride_offset(struct ll_readahead_state *ras)
1555 {
1556         unsigned long window_len = ras->ras_next_readahead -
1557                                    ras->ras_window_start;
1558         unsigned long left;
1559
1560         LASSERT(ras->ras_stride_length != 0);
1561
1562         left = window_len % ras->ras_stride_length;
1563
1564         ras->ras_stride_offset = ras->ras_next_readahead - left;
1565
1566         RAS_CDEBUG(ras);
1567 }
1568
1569 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1570                        struct ll_readahead_state *ras, unsigned long index,
1571                        unsigned hit)
1572 {
1573         struct ll_ra_info *ra = &sbi->ll_ra_info;
1574         int zero = 0, stride_zero = 0, stride_detect = 0, ra_miss = 0;
1575         ENTRY;
1576
1577         spin_lock(&sbi->ll_lock);
1578         spin_lock(&ras->ras_lock);
1579
1580         ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
1581
1582         /* reset the read-ahead window in two cases.  First when the app seeks
1583          * or reads to some other part of the file.  Secondly if we get a
1584          * read-ahead miss that we think we've previously issued.  This can
1585          * be a symptom of there being so many read-ahead pages that the VM is
1586          * reclaiming it before we get to it. */
1587         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1588                 zero = 1;
1589                 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
1590                 /* check whether it is in stride I/O mode*/
1591                 if (!index_in_stride_window(index, ras, inode))
1592                         stride_zero = 1;
1593         } else if (!hit && ras->ras_window_len &&
1594                    index < ras->ras_next_readahead &&
1595                    index_in_window(index, ras->ras_window_start, 0,
1596                                    ras->ras_window_len)) {
1597                 zero = 1;
1598                 ra_miss = 1;
1599                 /* If it hits read-ahead miss and the stride I/O is still
1600                  * not detected, reset stride stuff to re-detect the whole
1601                  * stride I/O mode to avoid complication */
1602                 if (!stride_io_mode(ras))
1603                         stride_zero = 1;
1604                 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1605         }
1606
1607         /* On the second access to a file smaller than the tunable
1608          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1609          * file up to ra_max_pages.  This is simply a best effort and
1610          * only occurs once per open file.  Normal RA behavior is reverted
1611          * to for subsequent IO.  The mmap case does not increment
1612          * ras_requests and thus can never trigger this behavior. */
1613         if (ras->ras_requests == 2 && !ras->ras_request_index) {
1614                 __u64 kms_pages;
1615
1616                 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1617                             CFS_PAGE_SHIFT;
1618
1619                 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1620                        ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages);
1621
1622                 if (kms_pages &&
1623                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1624                         ras->ras_window_start = 0;
1625                         ras->ras_last_readpage = 0;
1626                         ras->ras_next_readahead = 0;
1627                         ras->ras_window_len = min(ra->ra_max_pages,
1628                                 ra->ra_max_read_ahead_whole_pages);
1629                         GOTO(out_unlock, 0);
1630                 }
1631         }
1632
1633         if (zero) {
1634                 /* If it is discontinuous read, check
1635                  * whether it is stride I/O mode*/
1636                 if (stride_zero) {
1637                         ras_reset(ras, index);
1638                         ras->ras_consecutive_pages++;
1639                         ras_stride_reset(ras);
1640                         RAS_CDEBUG(ras);
1641                         GOTO(out_unlock, 0);
1642                 } else {
1643                         /* The read is still in stride window or
1644                          * it hits read-ahead miss */
1645
1646                         /* If ra-window miss is hitted, which probably means VM
1647                          * pressure, and some read-ahead pages were reclaimed.So
1648                          * the length of ra-window will not increased, but also
1649                          * not reset to avoid redetecting the stride I/O mode.*/
1650                         ras->ras_consecutive_requests = 0;
1651                         if (!ra_miss) {
1652                                 ras->ras_consecutive_pages = 0;
1653                                 if (++ras->ras_consecutive_stride_requests > 1)
1654                                         stride_detect = 1;
1655                         }
1656                         RAS_CDEBUG(ras);
1657                 }
1658         } else if (ras->ras_consecutive_stride_requests > 1) {
1659                 /* If this is contiguous read but in stride I/O mode
1660                  * currently, check whether stride step still is valid,
1661                  * if invalid, it will reset the stride ra window*/
1662                 if (ras->ras_consecutive_pages + 1 > ras->ras_stride_pages)
1663                         ras_stride_reset(ras);
1664         }
1665
1666         ras->ras_last_readpage = index;
1667         ras->ras_consecutive_pages++;
1668         ras_set_start(ras, index);
1669         ras->ras_next_readahead = max(ras->ras_window_start,
1670                                       ras->ras_next_readahead);
1671         RAS_CDEBUG(ras);
1672
1673         /* Trigger RA in the mmap case where ras_consecutive_requests
1674          * is not incremented and thus can't be used to trigger RA */
1675         if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1676                 ras->ras_window_len = RAS_INCREASE_STEP;
1677                 GOTO(out_unlock, 0);
1678         }
1679
1680         /* Initially reset the stride window offset to next_readahead*/
1681         if (ras->ras_consecutive_stride_requests == 2 && stride_detect)
1682                 ras_set_stride_offset(ras);
1683
1684         /* The initial ras_window_len is set to the request size.  To avoid
1685          * uselessly reading and discarding pages for random IO the window is
1686          * only increased once per consecutive request received. */
1687         if ((ras->ras_consecutive_requests > 1 &&
1688             !ras->ras_request_index) || stride_detect) {
1689                 if (stride_io_mode(ras))
1690                         ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP);
1691                 else
1692                         ras->ras_window_len = min(ras->ras_window_len +
1693                                                   RAS_INCREASE_STEP,
1694                                                   ra->ra_max_pages);
1695         }
1696         EXIT;
1697 out_unlock:
1698         RAS_CDEBUG(ras);
1699         ras->ras_request_index++;
1700         spin_unlock(&ras->ras_lock);
1701         spin_unlock(&sbi->ll_lock);
1702         return;
1703 }
1704
1705 int ll_writepage(struct page *page)
1706 {
1707         struct inode *inode = page->mapping->host;
1708         struct ll_inode_info *lli = ll_i2info(inode);
1709         struct obd_export *exp;
1710         struct ll_async_page *llap;
1711         int rc = 0;
1712         ENTRY;
1713
1714         LASSERT(PageLocked(page));
1715
1716         exp = ll_i2obdexp(inode);
1717         if (exp == NULL)
1718                 GOTO(out, rc = -EINVAL);
1719
1720         llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
1721         if (IS_ERR(llap))
1722                 GOTO(out, rc = PTR_ERR(llap));
1723
1724         LASSERT(!llap->llap_nocache);
1725         LASSERT(!PageWriteback(page));
1726         set_page_writeback(page);
1727
1728         page_cache_get(page);
1729         if (llap->llap_write_queued) {
1730                 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
1731                 rc = obd_set_async_flags(exp, lli->lli_smd, NULL,
1732                                          llap->llap_cookie,
1733                                          ASYNC_READY | ASYNC_URGENT);
1734         } else {
1735                 rc = queue_or_sync_write(exp, inode, llap, CFS_PAGE_SIZE,
1736                                          ASYNC_READY | ASYNC_URGENT);
1737         }
1738         if (rc)
1739                 page_cache_release(page);
1740 out:
1741         if (rc) {
1742                 if (!lli->lli_async_rc)
1743                         lli->lli_async_rc = rc;
1744                 /* re-dirty page on error so it retries write */
1745                 if (PageWriteback(page)) {
1746                         end_page_writeback(page);
1747                 }
1748                 /* resend page only for not started IO*/
1749                 if (!PageError(page))
1750                         ll_redirty_page(page);
1751                 unlock_page(page);
1752         }
1753         RETURN(rc);
1754 }
1755
1756 /*
1757  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
1758  * read-ahead assumes it is valid to issue readpage all the way up to
1759  * i_size, but our dlm locks make that not the case.  We disable the
1760  * kernel's read-ahead and do our own by walking ahead in the page cache
1761  * checking for dlm lock coverage.  the main difference between 2.4 and
1762  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1763  * so they look the same.
1764  */
1765 int ll_readpage(struct file *filp, struct page *page)
1766 {
1767         struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
1768         struct inode *inode = page->mapping->host;
1769         struct obd_export *exp;
1770         struct ll_async_page *llap;
1771         struct obd_io_group *oig = NULL;
1772         struct lustre_handle *lockh = NULL;
1773         int rc;
1774         ENTRY;
1775
1776         LASSERT(PageLocked(page));
1777         LASSERT(!PageUptodate(page));
1778         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset=%Lu=%#Lx\n",
1779                inode->i_ino, inode->i_generation, inode,
1780                (((loff_t)page->index) << CFS_PAGE_SHIFT),
1781                (((loff_t)page->index) << CFS_PAGE_SHIFT));
1782         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1783
1784         if (!ll_i2info(inode)->lli_smd) {
1785                 /* File with no objects - one big hole */
1786                 /* We use this just for remove_from_page_cache that is not
1787                  * exported, we'd make page back up to date. */
1788                 ll_truncate_complete_page(page);
1789                 clear_page(kmap(page));
1790                 kunmap(page);
1791                 SetPageUptodate(page);
1792                 unlock_page(page);
1793                 RETURN(0);
1794         }
1795
1796         rc = oig_init(&oig);
1797         if (rc < 0)
1798                 GOTO(out, rc);
1799
1800         exp = ll_i2obdexp(inode);
1801         if (exp == NULL)
1802                 GOTO(out, rc = -EINVAL);
1803
1804         if (fd->fd_flags & LL_FILE_GROUP_LOCKED)
1805                 lockh = &fd->fd_cwlockh;
1806
1807         llap = llap_from_page_with_lockh(page, LLAP_ORIGIN_READPAGE, lockh);
1808         if (IS_ERR(llap)) {
1809                 if (PTR_ERR(llap) == -ENOLCK) {
1810                         CWARN("ino %lu page %lu (%llu) not covered by "
1811                               "a lock (mmap?).  check debug logs.\n",
1812                               inode->i_ino, page->index,
1813                               (long long)page->index << PAGE_CACHE_SHIFT);
1814                 }
1815                 GOTO(out, rc = PTR_ERR(llap));
1816         }
1817
1818         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1819                 ras_update(ll_i2sbi(inode), inode, &fd->fd_ras, page->index,
1820                            llap->llap_defer_uptodate);
1821
1822
1823         if (llap->llap_defer_uptodate) {
1824                 /* This is the callpath if we got the page from a readahead */
1825                 llap->llap_ra_used = 1;
1826                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1827                                   fd->fd_flags);
1828                 if (rc > 0)
1829                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1830                                              NULL, oig);
1831                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1832                 SetPageUptodate(page);
1833                 unlock_page(page);
1834                 GOTO(out_oig, rc = 0);
1835         }
1836
1837         rc = ll_issue_page_read(exp, llap, oig, 0);
1838         if (rc)
1839                 GOTO(out, rc);
1840
1841         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1842         /* We have just requested the actual page we want, see if we can tack
1843          * on some readahead to that page's RPC before it is sent. */
1844         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1845                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1846                              fd->fd_flags);
1847
1848         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1849
1850 out:
1851         if (rc)
1852                 unlock_page(page);
1853 out_oig:
1854         if (oig != NULL)
1855                 oig_release(oig);
1856         RETURN(rc);
1857 }
1858
1859 static void ll_file_put_pages(struct page **pages, int numpages)
1860 {
1861         int i;
1862         struct page **pp;
1863         ENTRY;
1864
1865         for (i = 0, pp = pages; i < numpages; i++, pp++) {
1866                 if (*pp) {
1867                         LL_CDEBUG_PAGE(D_PAGE, (*pp), "free\n");
1868                         __ll_put_llap(*pp);
1869                         if (page_private(*pp))
1870                                 CERROR("the llap wasn't freed\n");
1871                         (*pp)->mapping = NULL;
1872                         if (page_count(*pp) != 1)
1873                                 CERROR("page %p, flags %#lx, count %i, private %p\n",
1874                                 (*pp), (unsigned long)(*pp)->flags, page_count(*pp),
1875                                 (void*)page_private(*pp));
1876                         __free_pages(*pp, 0);
1877                 }
1878         }
1879         OBD_FREE(pages, numpages * sizeof(struct page*));
1880         EXIT;
1881 }
1882
1883 static struct page **ll_file_prepare_pages(int numpages, struct inode *inode,
1884                                            unsigned long first)
1885 {
1886         struct page **pages;
1887         int i;
1888         int rc = 0;
1889         ENTRY;
1890
1891         OBD_ALLOC(pages, sizeof(struct page *) * numpages);
1892         if (pages == NULL)
1893                 RETURN(ERR_PTR(-ENOMEM));
1894         for (i = 0; i < numpages; i++) {
1895                 struct page *page;
1896                 struct ll_async_page *llap;
1897
1898                 page = alloc_pages(GFP_HIGHUSER, 0);
1899                 if (page == NULL)
1900                         GOTO(err, rc = -ENOMEM);
1901                 pages[i] = page;
1902                 /* llap_from_page needs page index and mapping to be set */
1903                 page->index = first++;
1904                 page->mapping = inode->i_mapping;
1905                 llap = llap_from_page(page, LLAP_ORIGIN_LOCKLESS_IO);
1906                 if (IS_ERR(llap))
1907                         GOTO(err, rc = PTR_ERR(llap));
1908                 llap->llap_lockless_io_page = 1;
1909         }
1910         RETURN(pages);
1911 err:
1912         ll_file_put_pages(pages, numpages);
1913         RETURN(ERR_PTR(rc));
1914  }
1915
1916 static ssize_t ll_file_copy_pages(struct page **pages, int numpages,
1917                                   const struct iovec *iov, unsigned long nsegs,
1918                                   ssize_t iov_offset, loff_t pos, size_t count,
1919                                   int rw)
1920 {
1921         ssize_t amount = 0;
1922         int i;
1923         int updatechecksum = ll_i2sbi(pages[0]->mapping->host)->ll_flags &
1924                              LL_SBI_LLITE_CHECKSUM;
1925         ENTRY;
1926
1927         for (i = 0; i < numpages; i++) {
1928                 unsigned offset, bytes, left = 0;
1929                 char *vaddr;
1930
1931                 vaddr = kmap(pages[i]);
1932                 offset = pos & (CFS_PAGE_SIZE - 1);
1933                 bytes = min_t(unsigned, CFS_PAGE_SIZE - offset, count);
1934                 LL_CDEBUG_PAGE(D_PAGE, pages[i], "op = %s, addr = %p, "
1935                                "bytes = %u\n",
1936                                (rw == WRITE) ? "CFU" : "CTU",
1937                                vaddr + offset, bytes);
1938                 while (bytes > 0 && !left && nsegs) {
1939                         unsigned copy = min_t(ssize_t, bytes,
1940                                                iov->iov_len - iov_offset);
1941                         if (rw == WRITE) {
1942                                 left = copy_from_user(vaddr + offset,
1943                                                       iov->iov_base +iov_offset,
1944                                                       copy);
1945                                 if (updatechecksum) {
1946                                         struct ll_async_page *llap;
1947
1948                                         llap = llap_cast_private(pages[i]);
1949                                         llap->llap_checksum =
1950                                                 init_checksum(OSC_DEFAULT_CKSUM);
1951                                         llap->llap_checksum =
1952                                            compute_checksum(llap->llap_checksum,
1953                                                             vaddr,CFS_PAGE_SIZE,
1954                                                             OSC_DEFAULT_CKSUM);
1955                                 }
1956                         } else {
1957                                 left = copy_to_user(iov->iov_base + iov_offset,
1958                                                     vaddr + offset, copy);
1959                         }
1960
1961                         amount += copy;
1962                         count -= copy;
1963                         pos += copy;
1964                         iov_offset += copy;
1965                         bytes -= copy;
1966                         if (iov_offset == iov->iov_len) {
1967                                 iov_offset = 0;
1968                                 iov++;
1969                                 nsegs--;
1970                         }
1971                 }
1972                 kunmap(pages[i]);
1973                 if (left) {
1974                         amount -= left;
1975                         break;
1976                 }
1977         }
1978         if (amount == 0)
1979                 RETURN(-EFAULT);
1980         RETURN(amount);
1981 }
1982
1983 static int ll_file_oig_pages(struct inode * inode, struct page **pages,
1984                              int numpages, loff_t pos, size_t count, int rw)
1985 {
1986         struct obd_io_group *oig;
1987         struct ll_inode_info *lli = ll_i2info(inode);
1988         struct obd_export *exp;
1989         loff_t org_pos = pos;
1990         obd_flag brw_flags;
1991         int rc;
1992         int i;
1993         ENTRY;
1994
1995         exp = ll_i2obdexp(inode);
1996         if (exp == NULL)
1997                 RETURN(-EINVAL);
1998         rc = oig_init(&oig);
1999         if (rc)
2000                 RETURN(rc);
2001         brw_flags = OBD_BRW_SRVLOCK;
2002         if (capable(CAP_SYS_RESOURCE))
2003                 brw_flags |= OBD_BRW_NOQUOTA;
2004
2005         for (i = 0; i < numpages; i++) {
2006                 struct ll_async_page *llap;
2007                 unsigned from, bytes;
2008
2009                 from = pos & (CFS_PAGE_SIZE - 1);
2010                 bytes = min_t(unsigned, CFS_PAGE_SIZE - from,
2011                               count - pos + org_pos);
2012                 llap = llap_cast_private(pages[i]);
2013                 LASSERT(llap);
2014
2015                 lock_page(pages[i]);
2016
2017                 LL_CDEBUG_PAGE(D_PAGE, pages[i], "offset "LPU64","
2018                                " from %u, bytes = %u\n",
2019                                pos, from, bytes);
2020                 LASSERTF(pos >> CFS_PAGE_SHIFT == pages[i]->index,
2021                          "wrong page index %lu (%lu)\n",
2022                          pages[i]->index,
2023                          (unsigned long)(pos >> CFS_PAGE_SHIFT));
2024                 rc = obd_queue_group_io(exp, lli->lli_smd, NULL, oig,
2025                                         llap->llap_cookie,
2026                                         (rw == WRITE) ?
2027                                         OBD_BRW_WRITE:OBD_BRW_READ,
2028                                         from, bytes, brw_flags,
2029                                         ASYNC_READY | ASYNC_URGENT |
2030                                         ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
2031                 if (rc) {
2032                         i++;
2033                         GOTO(out, rc);
2034                 }
2035                 pos += bytes;
2036         }
2037         rc = obd_trigger_group_io(exp, lli->lli_smd, NULL, oig);
2038         if (rc)
2039                 GOTO(out, rc);
2040         rc = oig_wait(oig);
2041 out:
2042         while(--i >= 0)
2043                 unlock_page(pages[i]);
2044         oig_release(oig);
2045         RETURN(rc);
2046 }
2047
2048 /* Advance through passed iov, adjust iov pointer as necessary and return
2049  * starting offset in individual entry we are pointing at. Also reduce
2050  * nr_segs as needed */
2051 static ssize_t ll_iov_advance(const struct iovec **iov, unsigned long *nr_segs,
2052                               ssize_t offset)
2053 {
2054         while (*nr_segs > 0) {
2055                 if ((*iov)->iov_len > offset)
2056                         return ((*iov)->iov_len - offset);
2057                 offset -= (*iov)->iov_len;
2058                 (*iov)++;
2059                 (*nr_segs)--;
2060         }
2061         return 0;
2062 }
2063
2064 ssize_t ll_file_lockless_io(struct file *file, const struct iovec *iov,
2065                             unsigned long nr_segs,
2066                             loff_t *ppos, int rw, ssize_t count)
2067 {
2068         loff_t pos;
2069         struct inode *inode = file->f_dentry->d_inode;
2070         ssize_t rc = 0;
2071         int max_pages;
2072         size_t amount = 0;
2073         unsigned long first, last;
2074         const struct iovec *iv = &iov[0];
2075         unsigned long nsegs = nr_segs;
2076         unsigned long offset = 0;
2077         ENTRY;
2078
2079         if (rw == READ) {
2080                 loff_t isize;
2081
2082                 ll_inode_size_lock(inode, 0);
2083                 isize = i_size_read(inode);
2084                 ll_inode_size_unlock(inode, 0);
2085                 if (*ppos >= isize)
2086                         GOTO(out, rc = 0);
2087                 if (*ppos + count >= isize)
2088                         count -= *ppos + count - isize;
2089                 if (count == 0)
2090                         GOTO(out, rc);
2091         } else {
2092                 rc = generic_write_checks(file, ppos, &count, 0);
2093                 if (rc)
2094                         GOTO(out, rc);
2095                 rc = remove_suid(file->f_dentry);
2096                 if (rc)
2097                         GOTO(out, rc);
2098         }
2099
2100         pos = *ppos;
2101         first = pos >> CFS_PAGE_SHIFT;
2102         last = (pos + count - 1) >> CFS_PAGE_SHIFT;
2103         max_pages = PTLRPC_MAX_BRW_PAGES *
2104                 ll_i2info(inode)->lli_smd->lsm_stripe_count;
2105         CDEBUG(D_INFO, "%u, stripe_count = %u\n",
2106                PTLRPC_MAX_BRW_PAGES /* max_pages_per_rpc */,
2107                ll_i2info(inode)->lli_smd->lsm_stripe_count);
2108
2109         while (first <= last && rc >= 0) {
2110                 int pages_for_io;
2111                 struct page **pages;
2112                 size_t bytes = count - amount;
2113
2114                 pages_for_io = min_t(int, last - first + 1, max_pages);
2115                 pages = ll_file_prepare_pages(pages_for_io, inode, first);
2116                 if (IS_ERR(pages)) {
2117                         rc = PTR_ERR(pages);
2118                         break;
2119                 }
2120                 if (rw == WRITE) {
2121                         rc = ll_file_copy_pages(pages, pages_for_io, iv, nsegs,
2122                                                 offset, pos + amount, bytes,
2123                                                 rw);
2124                         if (rc < 0)
2125                                 GOTO(put_pages, rc);
2126                         offset = ll_iov_advance(&iv, &nsegs, offset + rc);
2127                         bytes = rc;
2128                 }
2129                 rc = ll_file_oig_pages(inode, pages, pages_for_io,
2130                                        pos + amount, bytes, rw);
2131                 if (rc)
2132                         GOTO(put_pages, rc);
2133                 if (rw == READ) {
2134                         rc = ll_file_copy_pages(pages, pages_for_io, iv, nsegs,
2135                                                 offset, pos + amount, bytes, rw);
2136                         if (rc < 0)
2137                                 GOTO(put_pages, rc);
2138                         offset = ll_iov_advance(&iv, &nsegs, offset + rc);
2139                         bytes = rc;
2140                 }
2141                 amount += bytes;
2142 put_pages:
2143                 ll_file_put_pages(pages, pages_for_io);
2144                 first += pages_for_io;
2145                 /* a short read/write check */
2146                 if (pos + amount < ((loff_t)first << CFS_PAGE_SHIFT))
2147                         break;
2148                 /* Check if we are out of userspace buffers. (how that could
2149                    happen?) */
2150                 if (nsegs == 0)
2151                         break;
2152         }
2153         /* NOTE: don't update i_size and KMS in absence of LDLM locks even
2154          * write makes the file large */
2155         file_accessed(file);
2156         if (rw == READ && amount < count && rc == 0) {
2157                 unsigned long not_cleared;
2158
2159                 while (nsegs > 0) {
2160                         ssize_t to_clear = min_t(ssize_t, count - amount,
2161                                                  iv->iov_len - offset);
2162                         not_cleared = clear_user(iv->iov_base + offset,
2163                                                  to_clear);
2164                         amount += to_clear - not_cleared;
2165                         if (not_cleared) {
2166                                 rc = -EFAULT;
2167                                 break;
2168                         }
2169                         offset = 0;
2170                         iv++;
2171                         nsegs--;
2172                 }
2173         }
2174         if (amount > 0) {
2175                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
2176                                     (rw == WRITE) ?
2177                                     LPROC_LL_LOCKLESS_WRITE :
2178                                     LPROC_LL_LOCKLESS_READ,
2179                                     (long)amount);
2180                 *ppos += amount;
2181                 RETURN(amount);
2182         }
2183 out:
2184         RETURN(rc);
2185 }