lustre/llite/rw.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Lustre Lite I/O page cache routines shared by different kernel revs
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #include <linux/config.h>
  25 #include <linux/kernel.h>
  26 #include <linux/mm.h>
  27 #include <linux/string.h>
  28 #include <linux/stat.h>
  29 #include <linux/errno.h>
  30 #include <linux/smp_lock.h>
  31 #include <linux/unistd.h>
  32 #include <linux/version.h>
  33 #include <asm/system.h>
  34 #include <asm/uaccess.h>
  35
  36 #include <linux/fs.h>
  37 #include <linux/stat.h>
  38 #include <asm/uaccess.h>
  39 #include <asm/segment.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 <linux/lustre_mds.h>
  47 #include <linux/lustre_lite.h>
  48 #include "llite_internal.h"
  49 #include <linux/lustre_compat25.h>
  50
  51 #ifndef list_for_each_prev_safe
  52 #define list_for_each_prev_safe(pos, n, head) \
  53         for (pos = (head)->prev, n = pos->prev; pos != (head); \
  54                 pos = n, n = pos->prev )
  55 #endif
  56
  57 kmem_cache_t *ll_async_page_slab = NULL;
  58 size_t ll_async_page_slab_size = 0;
  59
  60 /* SYNCHRONOUS I/O to object storage for an inode */
  61 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
  62                   struct page *page, int flags)
  63 {
  64         struct ll_inode_info *lli = ll_i2info(inode);
  65         struct lov_stripe_md *lsm = lli->lli_smd;
  66         struct brw_page pg;
  67         int rc;
  68         ENTRY;
  69
  70         pg.pg = page;
  71         pg.off = ((obd_off)page->index) << PAGE_SHIFT;
  72
  73         if (cmd == OBD_BRW_WRITE && (pg.off + PAGE_SIZE > inode->i_size))
  74                 pg.count = inode->i_size % PAGE_SIZE;
  75         else
  76                 pg.count = 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",
  84                        inode->i_ino, inode, inode->i_size, page->mapping->host,
  85                        page->mapping->host->i_size, page->index, pg.off);
  86         }
  87
  88         pg.flag = flags;
  89
  90         if (cmd == OBD_BRW_WRITE)
  91                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
  92                                     LPROC_LL_BRW_WRITE, pg.count);
  93         else
  94                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
  95                                     LPROC_LL_BRW_READ, pg.count);
  96         rc = obd_brw(cmd, ll_i2obdexp(inode), oa, lsm, 1, &pg, NULL);
  97         if (rc == 0)
  98                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
  99         else if (rc != -EIO)
 100                 CERROR("error from obd_brw: rc = %d\n", rc);
 101         RETURN(rc);
 102 }
 103
 104 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
 105
 106 /* this isn't where truncate starts.   roughly:
 107  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate
 108  * we grab the lock back in setattr_raw to avoid races.
 109  *
 110  * must be called with lli_size_sem held */
 111 void ll_truncate(struct inode *inode)
 112 {
 113         struct ll_inode_info *lli = ll_i2info(inode);
 114         struct lov_stripe_md *lsm = lli->lli_smd;
 115         struct obdo oa;
 116         int rc;
 117         ENTRY;
 118         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %llu\n", inode->i_ino,
 119                inode->i_generation, inode, inode->i_size);
 120
 121         if (lli->lli_size_pid != current->pid) {
 122                 EXIT;
 123                 return;
 124         }
 125
 126         if (!lsm) {
 127                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 128                        inode->i_ino);
 129                 GOTO(out_unlock, 0);
 130         }
 131
 132         LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
 133
 134         if (lov_merge_size(lsm, 0) == inode->i_size) {
 135                 CDEBUG(D_VFSTRACE, "skipping punch for "LPX64" (size = %llu)\n",
 136                        lsm->lsm_object_id, inode->i_size);
 137                 GOTO(out_unlock, 0);
 138         }
 139
 140         CDEBUG(D_INFO, "calling punch for "LPX64" (new size %llu)\n",
 141                lsm->lsm_object_id, inode->i_size);
 142
 143         oa.o_id = lsm->lsm_object_id;
 144         oa.o_valid = OBD_MD_FLID;
 145         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
 146                         OBD_MD_FLATIME |OBD_MD_FLMTIME |OBD_MD_FLCTIME);
 147
 148         obd_adjust_kms(ll_i2obdexp(inode), lsm, inode->i_size, 1);
 149
 150         lli->lli_size_pid = 0;
 151         up(&lli->lli_size_sem);
 152
 153         rc = obd_punch(ll_i2obdexp(inode), &oa, lsm, inode->i_size,
 154                        OBD_OBJECT_EOF, NULL);
 155         if (rc)
 156                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 157         else
 158                 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE|OBD_MD_FLBLOCKS|
 159                               OBD_MD_FLATIME | OBD_MD_FLMTIME |
 160                               OBD_MD_FLCTIME);
 161         EXIT;
 162         return;
 163
 164  out_unlock:
 165         lli->lli_size_pid = 0;
 166         up(&lli->lli_size_sem);
 167 } /* ll_truncate */
 168
 169 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
 170 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 171                      unsigned to)
 172 {
 173         struct inode *inode = page->mapping->host;
 174         struct ll_inode_info *lli = ll_i2info(inode);
 175         struct lov_stripe_md *lsm = lli->lli_smd;
 176         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 177         struct brw_page pga;
 178         struct obdo oa;
 179         __u64 kms;
 180         int rc = 0;
 181         ENTRY;
 182
 183         LASSERT(PageLocked(page));
 184         (void)llap_cast_private(page); /* assertion */
 185
 186         /* Check to see if we should return -EIO right away */
 187         pga.pg = page;
 188         pga.off = offset;
 189         pga.count = PAGE_SIZE;
 190         pga.flag = 0;
 191
 192         oa.o_id = lsm->lsm_object_id;
 193         oa.o_mode = inode->i_mode;
 194         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
 195
 196         rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), &oa, lsm, 1, &pga,
 197                      NULL);
 198         if (rc)
 199                 RETURN(rc);
 200
 201         if (PageUptodate(page)) {
 202                 LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n");
 203                 RETURN(0);
 204         }
 205
 206         /* We're completely overwriting an existing page, so _don't_ set it up
 207          * to date until commit_write */
 208         if (from == 0 && to == PAGE_SIZE) {
 209                 LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n");
 210                 POISON_PAGE(page, 0x11);
 211                 RETURN(0);
 212         }
 213
 214         /* If are writing to a new page, no need to read old data.  The extent
 215          * locking will have updated the KMS, and for our purposes here we can
 216          * treat it like i_size. */
 217         down(&lli->lli_size_sem);
 218         kms = lov_merge_size(lsm, 1);
 219         up(&lli->lli_size_sem);
 220         if (kms <= offset) {
 221                 LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n",
 222                                kms, offset);
 223                 memset(kmap(page), 0, PAGE_SIZE);
 224                 kunmap(page);
 225                 GOTO(prepare_done, rc = 0);
 226         }
 227
 228         /* XXX could be an async ocp read.. read-ahead? */
 229         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 230         if (rc == 0) {
 231                 /* bug 1598: don't clobber blksize */
 232                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 233                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 234         }
 235
 236         EXIT;
 237  prepare_done:
 238         if (rc == 0)
 239                 SetPageUptodate(page);
 240
 241         return rc;
 242 }
 243
 244 struct ll_async_page *llap_from_cookie(void *cookie)
 245 {
 246         struct ll_async_page *llap = cookie;
 247         if (llap->llap_magic != LLAP_MAGIC)
 248                 return ERR_PTR(-EINVAL);
 249         return llap;
 250 };
 251
 252 static int ll_ap_make_ready(void *data, int cmd)
 253 {
 254         struct ll_async_page *llap;
 255         struct page *page;
 256         ENTRY;
 257
 258         llap = llap_from_cookie(data);
 259         if (IS_ERR(llap))
 260                 RETURN(-EINVAL);
 261
 262         page = llap->llap_page;
 263
 264         LASSERT(cmd != OBD_BRW_READ);
 265
 266         /* we're trying to write, but the page is locked.. come back later */
 267         if (TryLockPage(page))
 268                 RETURN(-EAGAIN);
 269
 270         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 271         page_cache_get(page);
 272
 273         /* if we left PageDirty we might get another writepage call
 274          * in the future.  list walkers are bright enough
 275          * to check page dirty so we can leave it on whatever list
 276          * its on.  XXX also, we're called with the cli list so if
 277          * we got the page cache list we'd create a lock inversion
 278          * with the removepage path which gets the page lock then the
 279          * cli lock */
 280         clear_page_dirty(page);
 281         RETURN(0);
 282 }
 283
 284 /* We have two reasons for giving llite the opportunity to change the
 285  * write length of a given queued page as it builds the RPC containing
 286  * the page:
 287  *
 288  * 1) Further extending writes may have landed in the page cache
 289  *    since a partial write first queued this page requiring us
 290  *    to write more from the page cache.  (No further races are possible, since
 291  *    by the time this is called, the page is locked.)
 292  * 2) We might have raced with truncate and want to avoid performing
 293  *    write RPCs that are just going to be thrown away by the
 294  *    truncate's punch on the storage targets.
 295  *
 296  * The kms serves these purposes as it is set at both truncate and extending
 297  * writes.
 298  */
 299 static int ll_ap_refresh_count(void *data, int cmd)
 300 {
 301         struct ll_inode_info *lli;
 302         struct ll_async_page *llap;
 303         struct lov_stripe_md *lsm;
 304         struct page *page;
 305         __u64 kms;
 306         ENTRY;
 307
 308         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 309         LASSERT(cmd != OBD_BRW_READ);
 310
 311         llap = llap_from_cookie(data);
 312         if (IS_ERR(llap))
 313                 RETURN(PTR_ERR(llap));
 314
 315         page = llap->llap_page;
 316         lli = ll_i2info(page->mapping->host);
 317         lsm = lli->lli_smd;
 318
 319         //down(&lli->lli_size_sem);
 320         kms = lov_merge_size(lsm, 1);
 321         //up(&lli->lli_size_sem);
 322
 323         /* catch race with truncate */
 324         if (((__u64)page->index << PAGE_SHIFT) >= kms)
 325                 return 0;
 326
 327         /* catch sub-page write at end of file */
 328         if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms)
 329                 return kms % PAGE_SIZE;
 330
 331         return PAGE_SIZE;
 332 }
 333
 334 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 335 {
 336         struct lov_stripe_md *lsm;
 337         obd_flag valid_flags;
 338
 339         lsm = ll_i2info(inode)->lli_smd;
 340
 341         oa->o_id = lsm->lsm_object_id;
 342         oa->o_valid = OBD_MD_FLID;
 343         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 344         if (cmd == OBD_BRW_WRITE) {
 345                 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
 346                 mdc_pack_fid(obdo_fid(oa), inode->i_ino, 0, inode->i_mode);
 347                 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
 348
 349                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 350                                OBD_MD_FLUID | OBD_MD_FLGID;
 351         }
 352
 353         obdo_from_inode(oa, inode, valid_flags);
 354 }
 355
 356 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 357 {
 358         struct ll_async_page *llap;
 359         ENTRY;
 360
 361         llap = llap_from_cookie(data);
 362         if (IS_ERR(llap)) {
 363                 EXIT;
 364                 return;
 365         }
 366
 367         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 368         EXIT;
 369 }
 370
 371 static void ll_ap_get_ucred(void *data, struct obd_ucred *ouc)
 372 {
 373         struct ll_async_page *llap;
 374
 375         llap = llap_from_cookie(data);
 376         if (IS_ERR(llap)) {
 377                 EXIT;
 378                 return;
 379         }
 380
 381         memcpy(ouc, &llap->llap_ouc, sizeof(*ouc));
 382         EXIT;
 383 }
 384
 385 static struct obd_async_page_ops ll_async_page_ops = {
 386         .ap_make_ready =        ll_ap_make_ready,
 387         .ap_refresh_count =     ll_ap_refresh_count,
 388         .ap_fill_obdo =         ll_ap_fill_obdo,
 389         .ap_completion =        ll_ap_completion,
 390         .ap_get_ucred =         ll_ap_get_ucred,
 391 };
 392
 393 struct ll_async_page *llap_cast_private(struct page *page)
 394 {
 395         struct ll_async_page *llap = (struct ll_async_page *)page->private;
 396
 397         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 398                  "page %p private %lu gave magic %d which != %d\n",
 399                  page, page->private, llap->llap_magic, LLAP_MAGIC);
 400
 401         return llap;
 402 }
 403
 404 /* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction.
 405  *
 406  * There is an llap attached onto every page in lustre, linked off @sbi.
 407  * We add an llap to the list so we don't lose our place during list walking.
 408  * If llaps in the list are being moved they will only move to the end
 409  * of the LRU, and we aren't terribly interested in those pages here (we
 410  * start at the beginning of the list where the least-used llaps are.
 411  */
 412 int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction)
 413 {
 414         struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a };
 415         unsigned long total, want, count = 0;
 416
 417         total = sbi->ll_async_page_count;
 418
 419         /* There can be a large number of llaps (600k or more in a large
 420          * memory machine) so the VM 1/6 shrink ratio is likely too much.
 421          * Since we are freeing pages also, we don't necessarily want to
 422          * shrink so much.  Limit to 40MB of pages + llaps per call. */
 423         if (shrink_fraction == 0)
 424                 want = sbi->ll_async_page_count - sbi->ll_async_page_max + 32;
 425         else
 426                 want = (total + shrink_fraction - 1) / shrink_fraction;
 427
 428         if (want > 40 << (20 - PAGE_CACHE_SHIFT))
 429                 want = 40 << (20 - PAGE_CACHE_SHIFT);
 430
 431         CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n",
 432                want, total, shrink_fraction);
 433
 434         spin_lock(&sbi->ll_lock);
 435         list_add(&dummy_llap.llap_pglist_item, &sbi->ll_pglist);
 436
 437         while (--total >= 0 && count < want) {
 438                 struct page *page;
 439
 440                 if (unlikely(need_resched())) {
 441                         spin_unlock(&sbi->ll_lock);
 442                         cond_resched();
 443                         spin_lock(&sbi->ll_lock);
 444                 }
 445
 446                 llap = llite_pglist_next_llap(sbi,&dummy_llap.llap_pglist_item);
 447                 list_del_init(&dummy_llap.llap_pglist_item);
 448                 if (llap == NULL)
 449                         break;
 450
 451                 page = llap->llap_page;
 452                 LASSERT(page != NULL);
 453
 454                 list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item);
 455
 456                 /* Page needs/undergoing IO */
 457                 if (TryLockPage(page)) {
 458                         LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n");
 459                         continue;
 460                 }
 461
 462                 /* If page is dirty or undergoing IO don't discard it */
 463                 if (llap->llap_write_queued || PageDirty(page) ||
 464                     (!PageUptodate(page) &&
 465                      llap->llap_origin != LLAP_ORIGIN_READAHEAD)) {
 466                         unlock_page(page);
 467                         LL_CDEBUG_PAGE(D_PAGE, page, "can't drop from cache: "
 468                                        "%s%s%s%s origin %s\n",
 469                                        llap->llap_write_queued ? "wq " : "",
 470                                        PageDirty(page) ? "pd " : "",
 471                                        PageUptodate(page) ? "" : "!pu ",
 472                                        llap->llap_defer_uptodate ? "" : "!du",
 473                                        llap_origins[llap->llap_origin]);
 474                         continue;
 475                 }
 476
 477                 page_cache_get(page);
 478                 spin_unlock(&sbi->ll_lock);
 479
 480                 ++count;
 481                 LL_CDEBUG_PAGE(D_PAGE, page, "drop from cache %lu/%lu\n",
 482                                count, want);
 483                 if (page->mapping != NULL) {
 484                         ll_ra_accounting(page, page->mapping);
 485                         ll_truncate_complete_page(page);
 486                 }
 487                 unlock_page(page);
 488                 page_cache_release(page);
 489
 490                 spin_lock(&sbi->ll_lock);
 491         }
 492         list_del(&dummy_llap.llap_pglist_item);
 493         spin_unlock(&sbi->ll_lock);
 494
 495         CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n",
 496                count, want, total);
 497
 498         return count;
 499 }
 500
 501 struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
 502 {
 503         struct ll_async_page *llap;
 504         struct obd_export *exp;
 505         struct inode *inode = page->mapping->host;
 506         struct ll_sb_info *sbi = ll_i2sbi(inode);
 507         int rc;
 508         ENTRY;
 509
 510         LASSERT(ll_async_page_slab);
 511         LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
 512
 513         llap = llap_cast_private(page);
 514         if (llap != NULL) {
 515                 /* move to end of LRU list */
 516                 spin_lock(&sbi->ll_lock);
 517                 sbi->ll_pglist_gen++;
 518                 list_del_init(&llap->llap_pglist_item);
 519                 list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 520                 spin_unlock(&sbi->ll_lock);
 521                 GOTO(out, llap);
 522         }
 523
 524         exp = ll_i2obdexp(page->mapping->host);
 525         if (exp == NULL)
 526                 RETURN(ERR_PTR(-EINVAL));
 527
 528         /* limit the number of lustre-cached pages */
 529         if (sbi->ll_async_page_count >= sbi->ll_async_page_max)
 530                 llap_shrink_cache(sbi, 0);
 531
 532         OBD_SLAB_ALLOC(llap, ll_async_page_slab, SLAB_KERNEL,
 533                        ll_async_page_slab_size);
 534         if (llap == NULL)
 535                 RETURN(ERR_PTR(-ENOMEM));
 536         llap->llap_magic = LLAP_MAGIC;
 537         llap->llap_cookie = (void *)llap + size_round(sizeof(*llap));
 538
 539         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 540                                  (obd_off)page->index << PAGE_SHIFT,
 541                                  &ll_async_page_ops, llap, &llap->llap_cookie);
 542         if (rc) {
 543                 OBD_SLAB_FREE(llap, ll_async_page_slab,
 544                               ll_async_page_slab_size);
 545                 RETURN(ERR_PTR(rc));
 546         }
 547
 548         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 549                page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
 550         /* also zeroing the PRIVBITS low order bitflags */
 551         __set_page_ll_data(page, llap);
 552         llap->llap_page = page;
 553
 554         spin_lock(&sbi->ll_lock);
 555         sbi->ll_pglist_gen++;
 556         sbi->ll_async_page_count++;
 557         list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 558         spin_unlock(&sbi->ll_lock);
 559
 560 out:
 561         llap->llap_origin = origin;
 562         RETURN(llap);
 563 }
 564
 565 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
 566                                struct ll_async_page *llap,
 567                                unsigned to, obd_flag async_flags)
 568 {
 569         unsigned long size_index = inode->i_size >> PAGE_SHIFT;
 570         struct obd_io_group *oig;
 571         int rc;
 572         ENTRY;
 573
 574         /* _make_ready only sees llap once we've unlocked the page */
 575         llap->llap_write_queued = 1;
 576         rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
 577                                 llap->llap_cookie, OBD_BRW_WRITE, 0, 0, 0,
 578                                 async_flags);
 579         if (rc == 0) {
 580                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
 581                 //llap_write_pending(inode, llap);
 582                 GOTO(out, 0);
 583         }
 584
 585         llap->llap_write_queued = 0;
 586
 587         rc = oig_init(&oig);
 588         if (rc)
 589                 GOTO(out, rc);
 590
 591         /* make full-page requests if we are not at EOF (bug 4410) */
 592         if (to != PAGE_SIZE && llap->llap_page->index < size_index) {
 593                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 594                                "sync write before EOF: size_index %lu, to %d\n",
 595                                size_index, to);
 596                 to = PAGE_SIZE;
 597         } else if (to != PAGE_SIZE && llap->llap_page->index == size_index) {
 598                 int size_to = inode->i_size & ~PAGE_MASK;
 599                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 600                                "sync write at EOF: size_index %lu, to %d/%d\n",
 601                                size_index, to, size_to);
 602                 if (to < size_to)
 603                         to = size_to;
 604         }
 605
 606         rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
 607                                 llap->llap_cookie, OBD_BRW_WRITE, 0, to, 0,
 608                                 ASYNC_READY | ASYNC_URGENT |
 609                                 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
 610         if (rc)
 611                 GOTO(free_oig, rc);
 612
 613         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
 614         if (rc)
 615                 GOTO(free_oig, rc);
 616
 617         rc = oig_wait(oig);
 618
 619         if (!rc && async_flags & ASYNC_READY)
 620                 unlock_page(llap->llap_page);
 621
 622         LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
 623
 624 free_oig:
 625         oig_release(oig);
 626 out:
 627         RETURN(rc);
 628 }
 629
 630 /* update our write count to account for i_size increases that may have
 631  * happened since we've queued the page for io. */
 632
 633 /* be careful not to return success without setting the page Uptodate or
 634  * the next pass through prepare_write will read in stale data from disk. */
 635 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 636                     unsigned to)
 637 {
 638         struct inode *inode = page->mapping->host;
 639         struct ll_inode_info *lli = ll_i2info(inode);
 640         struct lov_stripe_md *lsm = lli->lli_smd;
 641         struct obd_export *exp;
 642         struct ll_async_page *llap;
 643         struct ll_uctxt ctxt;
 644         loff_t size;
 645         int rc = 0;
 646         ENTRY;
 647
 648         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 649         LASSERT(inode == file->f_dentry->d_inode);
 650         LASSERT(PageLocked(page));
 651
 652         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 653                inode, page, from, to, page->index);
 654
 655         llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
 656         if (IS_ERR(llap))
 657                 RETURN(PTR_ERR(llap));
 658
 659         exp = ll_i2obdexp(inode);
 660         if (exp == NULL)
 661                 RETURN(-EINVAL);
 662
 663         /* set user credit information for this page */
 664         llap->llap_ouc.ouc_fsuid = current->fsuid;
 665         llap->llap_ouc.ouc_fsgid = current->fsgid;
 666         llap->llap_ouc.ouc_cap = current->cap_effective;
 667         ll_i2uctxt(&ctxt, inode, NULL);
 668         llap->llap_ouc.ouc_suppgid1 = ctxt.gid1;
 669
 670         /* queue a write for some time in the future the first time we
 671          * dirty the page */
 672         if (!PageDirty(page)) {
 673                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 674                                      LPROC_LL_DIRTY_MISSES);
 675
 676                 rc = queue_or_sync_write(exp, inode, llap, to, 0);
 677                 if (rc)
 678                         GOTO(out, rc);
 679         } else {
 680                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 681                                      LPROC_LL_DIRTY_HITS);
 682         }
 683
 684         /* put the page in the page cache, from now on ll_removepage is
 685          * responsible for cleaning up the llap.
 686          * only set page dirty when it's queued to be write out */
 687         if (llap->llap_write_queued)
 688                 set_page_dirty(page);
 689
 690 out:
 691         size = (((obd_off)page->index) << PAGE_SHIFT) + to;
 692         down(&lli->lli_size_sem);
 693         if (rc == 0) {
 694                 obd_adjust_kms(exp, lsm, size, 0);
 695                 if (size > inode->i_size)
 696                         inode->i_size = size;
 697                 SetPageUptodate(page);
 698         } else if (size > inode->i_size) {
 699                 /* this page beyond the pales of i_size, so it can't be
 700                  * truncated in ll_p_r_e during lock revoking. we must
 701                  * teardown our book-keeping here. */
 702                 ll_removepage(page);
 703         }
 704         up(&lli->lli_size_sem);
 705         RETURN(rc);
 706 }
 707
 708 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 709 {
 710         struct ll_ra_info *ra = &sbi->ll_ra_info;
 711         unsigned long ret;
 712         ENTRY;
 713
 714         spin_lock(&sbi->ll_lock);
 715         ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
 716         ra->ra_cur_pages += ret;
 717         spin_unlock(&sbi->ll_lock);
 718
 719         RETURN(ret);
 720 }
 721
 722 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 723 {
 724         struct ll_ra_info *ra = &sbi->ll_ra_info;
 725         spin_lock(&sbi->ll_lock);
 726         LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
 727                  ra->ra_cur_pages, len);
 728         ra->ra_cur_pages -= len;
 729         spin_unlock(&sbi->ll_lock);
 730 }
 731
 732 /* called for each page in a completed rpc.*/
 733 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 734 {
 735         struct ll_async_page *llap;
 736         struct page *page;
 737         ENTRY;
 738
 739         llap = llap_from_cookie(data);
 740         if (IS_ERR(llap)) {
 741                 EXIT;
 742                 return;
 743         }
 744
 745         page = llap->llap_page;
 746         LASSERT(PageLocked(page));
 747
 748         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 749
 750         if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
 751                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 752
 753         if (rc == 0)  {
 754                 if (cmd == OBD_BRW_READ) {
 755                         if (!llap->llap_defer_uptodate)
 756                                 SetPageUptodate(page);
 757                 } else {
 758                         llap->llap_write_queued = 0;
 759                 }
 760                 ClearPageError(page);
 761         } else {
 762                 if (cmd == OBD_BRW_READ) {
 763                         llap->llap_defer_uptodate = 0;
 764                 } else {
 765                         ll_redirty_page(page);
 766                 }
 767                 SetPageError(page);
 768         }
 769
 770         unlock_page(page);
 771
 772         if (0 && cmd == OBD_BRW_WRITE) {
 773                 llap_write_complete(page->mapping->host, llap);
 774                 ll_try_done_writing(page->mapping->host);
 775         }
 776
 777         if (PageWriteback(page)) {
 778                 end_page_writeback(page);
 779         }
 780         page_cache_release(page);
 781         EXIT;
 782 }
 783
 784 /* the kernel calls us here when a page is unhashed from the page cache.
 785  * the page will be locked and the kernel is holding a spinlock, so
 786  * we need to be careful.  we're just tearing down our book-keeping
 787  * here. */
 788 void ll_removepage(struct page *page)
 789 {
 790         struct inode *inode = page->mapping->host;
 791         struct obd_export *exp;
 792         struct ll_async_page *llap;
 793         struct ll_sb_info *sbi = ll_i2sbi(inode);
 794         int rc;
 795         ENTRY;
 796
 797         LASSERT(!in_interrupt());
 798
 799         /* sync pages or failed read pages can leave pages in the page
 800          * cache that don't have our data associated with them anymore */
 801         if (page->private == 0) {
 802                 EXIT;
 803                 return;
 804         }
 805
 806         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 807
 808         exp = ll_i2obdexp(inode);
 809         if (exp == NULL) {
 810                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 811                 EXIT;
 812                 return;
 813         }
 814
 815         llap = llap_from_page(page, 0);
 816         if (IS_ERR(llap)) {
 817                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 818                        page->index, PTR_ERR(llap));
 819                 EXIT;
 820                 return;
 821         }
 822
 823         //llap_write_complete(inode, llap);
 824         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 825                                      llap->llap_cookie);
 826         if (rc != 0)
 827                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 828
 829         /* this unconditional free is only safe because the page lock
 830          * is providing exclusivity to memory pressure/truncate/writeback..*/
 831         __clear_page_ll_data(page);
 832
 833         spin_lock(&sbi->ll_lock);
 834         if (!list_empty(&llap->llap_pglist_item))
 835                 list_del_init(&llap->llap_pglist_item);
 836         sbi->ll_pglist_gen++;
 837         sbi->ll_async_page_count--;
 838         spin_unlock(&sbi->ll_lock);
 839         OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
 840         EXIT;
 841 }
 842
 843 static int ll_page_matches(struct page *page, int readahead)
 844 {
 845         struct lustre_handle match_lockh = {0};
 846         struct inode *inode = page->mapping->host;
 847         ldlm_policy_data_t page_extent;
 848         int flags, matches;
 849         ENTRY;
 850
 851         page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
 852         page_extent.l_extent.end =
 853                 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
 854         flags = LDLM_FL_TEST_LOCK;
 855         if (!readahead)
 856                 flags |= LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
 857         matches = obd_match(ll_i2sbi(inode)->ll_osc_exp,
 858                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
 859                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
 860                             &match_lockh);
 861         RETURN(matches);
 862 }
 863
 864 static int ll_issue_page_read(struct obd_export *exp,
 865                               struct ll_async_page *llap,
 866                               struct obd_io_group *oig, int defer)
 867 {
 868         struct page *page = llap->llap_page;
 869         int rc;
 870
 871         page_cache_get(page);
 872         llap->llap_defer_uptodate = defer;
 873         llap->llap_ra_used = 0;
 874         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
 875                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
 876                                 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
 877                                               | ASYNC_URGENT);
 878         if (rc) {
 879                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
 880                 page_cache_release(page);
 881         }
 882         RETURN(rc);
 883 }
 884
 885 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
 886 {
 887         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
 888         ra->ra_stats[which]++;
 889 }
 890
 891 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
 892 {
 893         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
 894         struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
 895
 896         spin_lock(&sbi->ll_lock);
 897         ll_ra_stats_inc_unlocked(ra, which);
 898         spin_unlock(&sbi->ll_lock);
 899 }
 900
 901 void ll_ra_accounting(struct page *page, struct address_space *mapping)
 902 {
 903         struct ll_async_page *llap;
 904
 905         llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
 906         if (IS_ERR(llap))
 907                 return;
 908
 909         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
 910                 return;
 911
 912         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 913 }
 914
 915 #define RAS_CDEBUG(ras) \
 916         CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n",        \
 917                ras->ras_last_readpage, ras->ras_consecutive,            \
 918                ras->ras_window_start, ras->ras_window_len,              \
 919                ras->ras_next_readahead);
 920
 921 static int index_in_window(unsigned long index, unsigned long point,
 922                            unsigned long before, unsigned long after)
 923 {
 924         unsigned long start = point - before, end = point + after;
 925
 926         if (start > point)
 927                start = 0;
 928         if (end < point)
 929                end = ~0;
 930
 931         return start <= index && index <= end;
 932 }
 933
 934 static int ll_readahead(struct ll_readahead_state *ras,
 935                          struct obd_export *exp, struct address_space *mapping,
 936                          struct obd_io_group *oig, int flags)
 937 {
 938         unsigned long i, start = 0, end = 0, reserved;
 939         struct ll_async_page *llap;
 940         struct page *page;
 941         int rc, ret = 0, match_failed = 0;
 942         __u64 kms;
 943         unsigned int gfp_mask;
 944         ENTRY;
 945
 946         kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
 947         if (kms == 0) {
 948                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
 949                 RETURN(0);
 950         }
 951
 952         spin_lock(&ras->ras_lock);
 953         /* reserve a part of the read-ahead window that we'll be issuing */
 954         if (ras->ras_window_len) {
 955                 start = ras->ras_next_readahead;
 956                 end = ras->ras_window_start + ras->ras_window_len - 1;
 957                 end = min(end, (unsigned long)((kms - 1) >> PAGE_CACHE_SHIFT));
 958                 ras->ras_next_readahead = max(end, end + 1);
 959
 960                 RAS_CDEBUG(ras);
 961         }
 962         spin_unlock(&ras->ras_lock);
 963
 964         if (end == 0) {
 965                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
 966                 RETURN(0);
 967         }
 968
 969         reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
 970         if (reserved < end - start + 1)
 971                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
 972
 973         gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
 974 #ifdef __GFP_NOWARN
 975         gfp_mask |= __GFP_NOWARN;
 976 #endif
 977
 978         for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
 979                 /* skip locked pages from previous readpage calls */
 980                 page = grab_cache_page_nowait_gfp(mapping, i, gfp_mask);
 981                 if (page == NULL) {
 982                         CDEBUG(D_READA, "g_c_p_n failed\n");
 983                         continue;
 984                 }
 985
 986                 /* we do this first so that we can see the page in the /proc
 987                  * accounting */
 988                 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
 989                 if (IS_ERR(llap) || llap->llap_defer_uptodate)
 990                         goto next_page;
 991
 992                 /* skip completed pages */
 993                 if (Page_Uptodate(page))
 994                         goto next_page;
 995
 996                 /* bail when we hit the end of the lock. */
 997                 if ((rc = ll_page_matches(page, 1)) <= 0) {
 998                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 999                                        "lock match failed: rc %d\n", rc);
1000                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
1001                         match_failed = 1;
1002                         goto next_page;
1003                 }
1004
1005                 rc = ll_issue_page_read(exp, llap, oig, 1);
1006                 if (rc == 0) {
1007                         reserved--;
1008                         ret++;
1009                         LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
1010                                        "started read-ahead\n");
1011                 }
1012                 if (rc) {
1013         next_page:
1014                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
1015                                        "skipping read-ahead\n");
1016
1017                         unlock_page(page);
1018                 }
1019                 page_cache_release(page);
1020         }
1021
1022         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
1023         if (reserved != 0)
1024                 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
1025         if (i == end + 1 && end == (kms >> PAGE_CACHE_SHIFT))
1026                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
1027
1028         /* if we didn't get to the end of the region we reserved from
1029          * the ras we need to go back and update the ras so that the
1030          * next read-ahead tries from where we left off.  we only do so
1031          * if the region we failed to issue read-ahead on is still ahead
1032          * of the app and behind the next index to start read-ahead from */
1033         if (i != end + 1) {
1034                 spin_lock(&ras->ras_lock);
1035                 if (i < ras->ras_next_readahead &&
1036                     index_in_window(i, ras->ras_window_start, 0,
1037                                     ras->ras_window_len)) {
1038                         ras->ras_next_readahead = i;
1039                         RAS_CDEBUG(ras);
1040                 }
1041                 spin_unlock(&ras->ras_lock);
1042         }
1043
1044         RETURN(ret);
1045 }
1046
1047 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
1048 {
1049         ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
1050 }
1051
1052 /* called with the ras_lock held or from places where it doesn't matter */
1053 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1054 {
1055         ras->ras_last_readpage = index;
1056         ras->ras_consecutive = 1;
1057         ras->ras_window_len = 0;
1058         ras_set_start(ras, index);
1059         ras->ras_next_readahead = ras->ras_window_start;
1060
1061         RAS_CDEBUG(ras);
1062 }
1063
1064 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1065 {
1066         spin_lock_init(&ras->ras_lock);
1067         ras_reset(ras, 0);
1068 }
1069
1070 static void ras_update(struct ll_sb_info *sbi, struct ll_readahead_state *ras,
1071                        unsigned long index, unsigned hit)
1072 {
1073         struct ll_ra_info *ra = &sbi->ll_ra_info;
1074         int zero = 0;
1075         ENTRY;
1076
1077         spin_lock(&sbi->ll_lock);
1078         spin_lock(&ras->ras_lock);
1079
1080         ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
1081
1082         /* reset the read-ahead window in two cases.  First when the app seeks
1083          * or reads to some other part of the file.  Secondly if we get a
1084          * read-ahead miss that we think we've previously issued.  This can
1085          * be a symptom of there being so many read-ahead pages that the VM is
1086          * reclaiming it before we get to it. */
1087         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1088                 zero = 1;
1089                 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
1090         } else if (!hit && ras->ras_window_len &&
1091                    index < ras->ras_next_readahead &&
1092                    index_in_window(index, ras->ras_window_start, 0,
1093                                    ras->ras_window_len)) {
1094                 zero = 1;
1095                 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1096         }
1097
1098         if (zero) {
1099                 ras_reset(ras, index);
1100                 GOTO(out_unlock, 0);
1101         }
1102
1103         ras->ras_last_readpage = index;
1104         ras->ras_consecutive++;
1105         ras_set_start(ras, index);
1106         ras->ras_next_readahead = max(ras->ras_window_start,
1107                                       ras->ras_next_readahead);
1108
1109         /* wait for a few pages to arrive before issuing readahead to avoid
1110          * the worst overutilization */
1111         if (ras->ras_consecutive == 3) {
1112                 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
1113                 GOTO(out_unlock, 0);
1114         }
1115
1116         /* we need to increase the window sometimes.  we'll arbitrarily
1117          * do it half-way through the pages in an rpc */
1118         if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
1119             (PTLRPC_MAX_BRW_PAGES >> 1)) {
1120                 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
1121                 ras->ras_window_len = min(ras->ras_window_len,
1122                                           ra->ra_max_pages);
1123         }
1124
1125         EXIT;
1126 out_unlock:
1127         RAS_CDEBUG(ras);
1128         spin_unlock(&ras->ras_lock);
1129         spin_unlock(&sbi->ll_lock);
1130         return;
1131 }
1132
1133 int ll_writepage(struct page *page)
1134 {
1135         struct inode *inode = page->mapping->host;
1136         struct ll_inode_info *lli = ll_i2info(inode);
1137         struct obd_export *exp;
1138         struct ll_async_page *llap;
1139         int rc = 0;
1140         ENTRY;
1141
1142         LASSERT(!PageDirty(page));
1143         LASSERT(PageLocked(page));
1144
1145         exp = ll_i2obdexp(inode);
1146         if (exp == NULL)
1147                 GOTO(out, rc = -EINVAL);
1148
1149         llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
1150         if (IS_ERR(llap))
1151                 GOTO(out, rc = PTR_ERR(llap));
1152
1153         page_cache_get(page);
1154         if (llap->llap_write_queued) {
1155                 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
1156                 rc = obd_set_async_flags(exp, lli->lli_smd, NULL,
1157                                          llap->llap_cookie,
1158                                          ASYNC_READY | ASYNC_URGENT);
1159         } else {
1160                 rc = queue_or_sync_write(exp, inode, llap,
1161                                          PAGE_SIZE, ASYNC_READY | ASYNC_URGENT);
1162         }
1163         if (rc)
1164                 page_cache_release(page);
1165 out:
1166         if (rc) {
1167                 if (!lli->lli_async_rc)
1168                         lli->lli_async_rc = rc;
1169                 /* re-dirty page on error so it retries write */
1170                 ll_redirty_page(page);
1171                 unlock_page(page);
1172         }
1173         RETURN(rc);
1174 }
1175
1176 /*
1177  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
1178  * read-ahead assumes it is valid to issue readpage all the way up to
1179  * i_size, but our dlm locks make that not the case.  We disable the
1180  * kernel's read-ahead and do our own by walking ahead in the page cache
1181  * checking for dlm lock coverage.  the main difference between 2.4 and
1182  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1183  * so they look the same.
1184  */
1185 int ll_readpage(struct file *filp, struct page *page)
1186 {
1187         struct ll_file_data *fd = filp->private_data;
1188         struct inode *inode = page->mapping->host;
1189         struct obd_export *exp;
1190         struct ll_async_page *llap;
1191         struct obd_io_group *oig = NULL;
1192         int rc;
1193         ENTRY;
1194
1195         LASSERT(PageLocked(page));
1196         LASSERT(!PageUptodate(page));
1197         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
1198                inode->i_ino, inode->i_generation, inode,
1199                (((obd_off)page->index) << PAGE_SHIFT));
1200         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1201
1202         rc = oig_init(&oig);
1203         if (rc < 0)
1204                 GOTO(out, rc);
1205
1206         exp = ll_i2obdexp(inode);
1207         if (exp == NULL)
1208                 GOTO(out, rc = -EINVAL);
1209
1210         llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1211         if (IS_ERR(llap))
1212                 GOTO(out, rc = PTR_ERR(llap));
1213
1214         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1215                 ras_update(ll_i2sbi(inode), &fd->fd_ras, page->index,
1216                            llap->llap_defer_uptodate);
1217
1218         if (llap->llap_defer_uptodate) {
1219                 llap->llap_ra_used = 1;
1220                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1221                                   fd->fd_flags);
1222                 if (rc > 0)
1223                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1224                                              NULL, oig);
1225                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1226                 SetPageUptodate(page);
1227                 unlock_page(page);
1228                 GOTO(out_oig, rc = 0);
1229         }
1230
1231         rc = ll_page_matches(page, 0);
1232         if (rc < 0) {
1233                 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1234                 GOTO(out, rc);
1235         }
1236
1237         if (rc == 0) {
1238                 CWARN("ino %lu page %lu (%llu) not covered by "
1239                       "a lock (mmap?).  check debug logs.\n",
1240                       inode->i_ino, page->index,
1241                       (long long)page->index << PAGE_CACHE_SHIFT);
1242         }
1243
1244         rc = ll_issue_page_read(exp, llap, oig, 0);
1245         if (rc)
1246                 GOTO(out, rc);
1247
1248         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1249         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1250                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1251                              fd->fd_flags);
1252
1253         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1254
1255 out:
1256         if (rc)
1257                 unlock_page(page);
1258 out_oig:
1259         if (oig != NULL)
1260                 oig_release(oig);
1261         RETURN(rc);
1262 }