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 /* SYNCHRONOUS I/O to object storage for an inode */
  58 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
  59                   struct page *page, int flags)
  60 {
  61         struct ll_inode_info *lli = ll_i2info(inode);
  62         struct lov_stripe_md *lsm = lli->lli_smd;
  63         struct timeval start;
  64         struct brw_page pg;
  65         int rc;
  66         ENTRY;
  67
  68         do_gettimeofday(&start);
  69
  70         pg.pg = page;
  71         pg.disk_offset = pg.page_offset = ((obd_off)page->index) << PAGE_SHIFT;
  72
  73         if (cmd == OBD_BRW_WRITE &&
  74             (pg.disk_offset + PAGE_SIZE > inode->i_size))
  75                 pg.count = inode->i_size % PAGE_SIZE;
  76         else
  77                 pg.count = PAGE_SIZE;
  78
  79         CDEBUG(D_PAGE, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
  80                cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, inode->i_ino,
  81                pg.disk_offset, pg.disk_offset);
  82         if (pg.count == 0) {
  83                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
  84                        LPU64"\n", inode->i_ino, inode, inode->i_size,
  85                        page->mapping->host, page->mapping->host->i_size,
  86                        page->index, pg.disk_offset);
  87         }
  88
  89         pg.flag = flags;
  90
  91         if (cmd == OBD_BRW_WRITE)
  92                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
  93                                     LPROC_LL_BRW_WRITE, pg.count);
  94         else
  95                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
  96                                     LPROC_LL_BRW_READ, pg.count);
  97         rc = obd_brw(cmd, ll_i2obdexp(inode), oa, lsm, 1, &pg, NULL);
  98         if (rc == 0)
  99                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
 100         else if (rc != -EIO)
 101                 CERROR("error from obd_brw: rc = %d\n", rc);
 102         ll_stime_record(ll_i2sbi(inode), &start,
 103                         &ll_i2sbi(inode)->ll_brw_stime);
 104         RETURN(rc);
 105 }
 106
 107 /* this isn't where truncate starts.   roughly:
 108  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate
 109  * we grab the lock back in setattr_raw to avoid races. */
 110 void ll_truncate(struct inode *inode)
 111 {
 112         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
 113         struct obdo oa;
 114         int rc;
 115         ENTRY;
 116         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
 117                inode->i_generation, inode);
 118
 119         if (!lsm) {
 120                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 121                        inode->i_ino);
 122                 EXIT;
 123                 return;
 124         }
 125
 126         oa.o_id = lsm->lsm_object_id;
 127         oa.o_gr = lsm->lsm_object_gr;
 128         oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 129         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE|OBD_MD_FLMODE|OBD_MD_FLATIME|
 130                                     OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 131
 132         CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after %Lu)\n",
 133                oa.o_id, inode->i_size);
 134
 135         /* truncate == punch from new size to absolute end of file */
 136         /* NB: obd_punch must be called with i_sem held!  It updates the kms! */
 137         rc = obd_punch(ll_i2obdexp(inode), &oa, lsm, inode->i_size,
 138                        OBD_OBJECT_EOF, NULL);
 139         if (rc)
 140                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 141         else
 142                 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 143                                           OBD_MD_FLATIME | OBD_MD_FLMTIME |
 144                                           OBD_MD_FLCTIME);
 145
 146         EXIT;
 147         return;
 148 } /* ll_truncate */
 149
 150 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
 151 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 152                      unsigned to)
 153 {
 154         struct inode *inode = page->mapping->host;
 155         struct ll_inode_info *lli = ll_i2info(inode);
 156         struct lov_stripe_md *lsm = lli->lli_smd;
 157         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 158         struct brw_page pga;
 159         struct obdo oa;
 160         __u64 kms;
 161         int rc = 0;
 162         ENTRY;
 163
 164         LASSERT(PageLocked(page));
 165         (void)llap_cast_private(page); /* assertion */
 166
 167         /* Check to see if we should return -EIO right away */
 168         pga.pg = page;
 169         pga.disk_offset = pga.page_offset = offset;
 170         pga.count = PAGE_SIZE;
 171         pga.flag = 0;
 172
 173         oa.o_id = lsm->lsm_object_id;
 174         oa.o_gr = lsm->lsm_object_gr;
 175         oa.o_mode = inode->i_mode;
 176         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE
 177                                 | OBD_MD_FLTYPE | OBD_MD_FLGROUP;
 178
 179         rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), &oa, lsm, 1, &pga,
 180                      NULL);
 181         if (rc)
 182                 RETURN(rc);
 183
 184         if (PageUptodate(page))
 185                 RETURN(0);
 186
 187         /* We're completely overwriting an existing page, so _don't_ set it up
 188          * to date until commit_write */
 189         if (from == 0 && to == PAGE_SIZE) {
 190                 POISON_PAGE(page, 0x11);
 191                 RETURN(0);
 192         }
 193
 194         /* If are writing to a new page, no need to read old data.  The extent
 195          * locking will have updated the KMS, and for our purposes here we can
 196          * treat it like i_size. */
 197         kms = lov_merge_size(lsm, 1);
 198         if (kms <= offset) {
 199                 memset(kmap(page), 0, PAGE_SIZE);
 200                 kunmap(page);
 201                 GOTO(prepare_done, rc = 0);
 202         }
 203
 204         /* XXX could be an async ocp read.. read-ahead? */
 205         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 206         if (rc == 0) {
 207                 /* bug 1598: don't clobber blksize */
 208                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 209                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 210         }
 211
 212         EXIT;
 213  prepare_done:
 214         if (rc == 0)
 215                 SetPageUptodate(page);
 216
 217         return rc;
 218 }
 219
 220 struct ll_async_page *llap_from_cookie(void *cookie)
 221 {
 222         struct ll_async_page *llap = cookie;
 223         if (llap->llap_magic != LLAP_MAGIC)
 224                 return ERR_PTR(-EINVAL);
 225         return llap;
 226 };
 227
 228 static int ll_ap_make_ready(void *data, int cmd)
 229 {
 230         struct ll_async_page *llap;
 231         struct page *page;
 232         ENTRY;
 233
 234         llap = llap_from_cookie(data);
 235         if (IS_ERR(llap))
 236                 RETURN(-EINVAL);
 237
 238         page = llap->llap_page;
 239
 240         LASSERT(cmd != OBD_BRW_READ);
 241
 242         /* we're trying to write, but the page is locked.. come back later */
 243         if (TryLockPage(page))
 244                 RETURN(-EAGAIN);
 245
 246         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 247         page_cache_get(page);
 248
 249         /* if we left PageDirty we might get another writepage call
 250          * in the future.  list walkers are bright enough
 251          * to check page dirty so we can leave it on whatever list
 252          * its on.  XXX also, we're called with the cli list so if
 253          * we got the page cache list we'd create a lock inversion
 254          * with the removepage path which gets the page lock then the
 255          * cli lock */
 256         clear_page_dirty(page);
 257         RETURN(0);
 258 }
 259
 260 /* We have two reasons for giving llite the opportunity to change the
 261  * write length of a given queued page as it builds the RPC containing
 262  * the page:
 263  *
 264  * 1) Further extending writes may have landed in the page cache
 265  *    since a partial write first queued this page requiring us
 266  *    to write more from the page cache.
 267  * 2) We might have raced with truncate and want to avoid performing
 268  *    write RPCs that are just going to be thrown away by the
 269  *    truncate's punch on the storage targets.
 270  *
 271  * The kms serves these purposes as it is set at both truncate and extending
 272  * writes.
 273  */
 274 static int ll_ap_refresh_count(void *data, int cmd)
 275 {
 276         struct ll_async_page *llap;
 277         struct lov_stripe_md *lsm;
 278         struct page *page;
 279         __u64 kms;
 280         ENTRY;
 281
 282         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 283         LASSERT(cmd != OBD_BRW_READ);
 284
 285         llap = llap_from_cookie(data);
 286         if (IS_ERR(llap))
 287                 RETURN(PTR_ERR(llap));
 288
 289         page = llap->llap_page;
 290         lsm = ll_i2info(page->mapping->host)->lli_smd;
 291         kms = lov_merge_size(lsm, 1);
 292
 293         /* catch race with truncate */
 294         if (((__u64)page->index << PAGE_SHIFT) >= kms)
 295                 return 0;
 296
 297         /* catch sub-page write at end of file */
 298         if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms)
 299                 return kms % PAGE_SIZE;
 300
 301         return PAGE_SIZE;
 302 }
 303
 304 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 305 {
 306         struct lov_stripe_md *lsm;
 307         obd_flag valid_flags;
 308
 309         lsm = ll_i2info(inode)->lli_smd;
 310
 311         oa->o_id = lsm->lsm_object_id;
 312         oa->o_gr = lsm->lsm_object_gr;
 313         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 314         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 315         if (cmd == OBD_BRW_WRITE) {
 316                 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
 317                 mdc_pack_id(obdo_id(oa), inode->i_ino, 0, inode->i_mode,
 318                             id_group(&ll_i2info(inode)->lli_id),
 319                             id_fid(&ll_i2info(inode)->lli_id));
 320
 321                 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
 322                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 323         }
 324
 325         obdo_from_inode(oa, inode, valid_flags);
 326 }
 327
 328 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 329 {
 330         struct ll_async_page *llap;
 331         ENTRY;
 332
 333         llap = llap_from_cookie(data);
 334         if (IS_ERR(llap)) {
 335                 EXIT;
 336                 return;
 337         }
 338
 339         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 340         EXIT;
 341 }
 342
 343 static struct obd_async_page_ops ll_async_page_ops = {
 344         .ap_make_ready =        ll_ap_make_ready,
 345         .ap_refresh_count =     ll_ap_refresh_count,
 346         .ap_fill_obdo =         ll_ap_fill_obdo,
 347         .ap_completion =        ll_ap_completion,
 348 };
 349
 350 struct ll_async_page *llap_cast_private(struct page *page)
 351 {
 352         struct ll_async_page *llap = (struct ll_async_page *)page->private;
 353
 354         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 355                  "page %p private %lu gave magic %d which != %d\n",
 356                  page, page->private, llap->llap_magic, LLAP_MAGIC);
 357
 358         return llap;
 359 }
 360
 361 /* XXX have the exp be an argument? */
 362 struct ll_async_page *llap_from_page(struct page *page)
 363 {
 364         struct ll_async_page *llap;
 365         struct obd_export *exp;
 366         struct inode *inode = page->mapping->host;
 367         struct ll_sb_info *sbi = ll_i2sbi(inode);
 368         int rc;
 369         ENTRY;
 370
 371         llap = llap_cast_private(page);
 372         if (llap != NULL)
 373                 RETURN(llap);
 374
 375         exp = ll_i2obdexp(page->mapping->host);
 376         if (exp == NULL)
 377                 RETURN(ERR_PTR(-EINVAL));
 378
 379         OBD_ALLOC(llap, sizeof(*llap));
 380         if (llap == NULL)
 381                 RETURN(ERR_PTR(-ENOMEM));
 382         llap->llap_magic = LLAP_MAGIC;
 383         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 384                                  (obd_off)page->index << PAGE_SHIFT,
 385                                  &ll_async_page_ops, llap, &llap->llap_cookie);
 386         if (rc) {
 387                 OBD_FREE(llap, sizeof(*llap));
 388                 RETURN(ERR_PTR(rc));
 389         }
 390
 391         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 392                page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
 393         /* also zeroing the PRIVBITS low order bitflags */
 394         __set_page_ll_data(page, llap);
 395         llap->llap_page = page;
 396
 397         spin_lock(&sbi->ll_lock);
 398         sbi->ll_pglist_gen++;
 399         list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
 400         spin_unlock(&sbi->ll_lock);
 401
 402         RETURN(llap);
 403 }
 404
 405 static int queue_or_sync_write(struct obd_export *exp,
 406                                struct lov_stripe_md *lsm,
 407                                struct ll_async_page *llap,
 408                                unsigned to,
 409                                obd_flag async_flags)
 410 {
 411         struct obd_io_group *oig;
 412         int rc;
 413         ENTRY;
 414
 415         /* _make_ready only sees llap once we've unlocked the page */
 416         llap->llap_write_queued = 1;
 417         rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
 418                                 OBD_BRW_WRITE, 0, 0, 0, async_flags);
 419         if (rc == 0) {
 420                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
 421                 //llap_write_pending(inode, llap);
 422                 GOTO(out, 0);
 423         }
 424
 425         llap->llap_write_queued = 0;
 426
 427         rc = oig_init(&oig);
 428         if (rc)
 429                 GOTO(out, rc);
 430
 431         rc = obd_queue_group_io(exp, lsm, NULL, oig, llap->llap_cookie,
 432                                 OBD_BRW_WRITE, 0, to, 0, ASYNC_READY |
 433                                 ASYNC_URGENT | ASYNC_COUNT_STABLE |
 434                                 ASYNC_GROUP_SYNC);
 435         if (rc)
 436                 GOTO(free_oig, rc);
 437
 438         rc = obd_trigger_group_io(exp, lsm, NULL, oig);
 439         if (rc)
 440                 GOTO(free_oig, rc);
 441
 442         rc = oig_wait(oig);
 443
 444         if (!rc && async_flags & ASYNC_READY)
 445                 unlock_page(llap->llap_page);
 446
 447         LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 448                        "sync write returned %d\n", rc);
 449
 450         EXIT;
 451 free_oig:
 452         oig_release(oig);
 453 out:
 454         return rc;
 455 }
 456
 457 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
 458                       obd_off size);
 459
 460 /* be careful not to return success without setting the page Uptodate or
 461  * the next pass through prepare_write will read in stale data from disk. */
 462 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 463                     unsigned to)
 464 {
 465         struct inode *inode = page->mapping->host;
 466         struct ll_inode_info *lli = ll_i2info(inode);
 467         struct lov_stripe_md *lsm = lli->lli_smd;
 468         struct obd_export *exp = NULL;
 469         struct ll_async_page *llap;
 470         loff_t size;
 471         int rc = 0;
 472         ENTRY;
 473
 474         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 475         LASSERT(inode == file->f_dentry->d_inode);
 476         LASSERT(PageLocked(page));
 477
 478         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 479                inode, page, from, to, page->index);
 480
 481         llap = llap_from_page(page);
 482         if (IS_ERR(llap))
 483                 RETURN(PTR_ERR(llap));
 484
 485         /* queue a write for some time in the future the first time we
 486          * dirty the page */
 487         if (!PageDirty(page)) {
 488                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 489                                      LPROC_LL_DIRTY_MISSES);
 490
 491                 exp = ll_i2obdexp(inode);
 492                 if (exp == NULL)
 493                         RETURN(-EINVAL);
 494
 495                 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd, llap,
 496                                          to, 0);
 497                 if (rc)
 498                         GOTO(out, rc);
 499         } else {
 500                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 501                                      LPROC_LL_DIRTY_HITS);
 502         }
 503
 504         /* put the page in the page cache, from now on ll_removepage is
 505          * responsible for cleaning up the llap.
 506          * don't dirty the page if it has been write out in q_o_s_w */
 507         if (llap->llap_write_queued)
 508                 set_page_dirty(page);
 509         EXIT;
 510 out:
 511         if (rc == 0) {
 512                 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
 513                 lov_increase_kms(exp, lsm, size);
 514                 if (size > inode->i_size)
 515                         inode->i_size = size;
 516                 SetPageUptodate(page);
 517         }
 518         return rc;
 519 }
 520
 521 int ll_writepage(struct page *page)
 522 {
 523         struct inode *inode = page->mapping->host;
 524         struct obd_export *exp;
 525         struct ll_async_page *llap;
 526         int rc = 0;
 527         ENTRY;
 528
 529         LASSERT(!PageDirty(page));
 530         LASSERT(PageLocked(page));
 531
 532         exp = ll_i2obdexp(inode);
 533         if (exp == NULL)
 534                 GOTO(out, rc = -EINVAL);
 535
 536         llap = llap_from_page(page);
 537         if (IS_ERR(llap))
 538                 GOTO(out, rc = PTR_ERR(llap));
 539
 540         page_cache_get(page);
 541         if (llap->llap_write_queued) {
 542                 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
 543                 rc = obd_set_async_flags(exp, ll_i2info(inode)->lli_smd, NULL,
 544                                          llap->llap_cookie,
 545                                          ASYNC_READY | ASYNC_URGENT);
 546         } else {
 547                 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd, llap,
 548                                          PAGE_SIZE, ASYNC_READY |
 549                                          ASYNC_URGENT);
 550         }
 551         if (rc)
 552                 page_cache_release(page);
 553         EXIT;
 554 out:
 555         if (rc)
 556                 unlock_page(page);
 557         return rc;
 558 }
 559
 560 static unsigned long
 561 ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 562 {
 563         struct ll_ra_info *ra = &sbi->ll_ra_info;
 564         unsigned long ret;
 565         ENTRY;
 566
 567         spin_lock(&sbi->ll_lock);
 568         ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
 569         ra->ra_cur_pages += ret;
 570         spin_unlock(&sbi->ll_lock);
 571
 572         RETURN(ret);
 573 }
 574
 575 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 576 {
 577         struct ll_ra_info *ra = &sbi->ll_ra_info;
 578         spin_lock(&sbi->ll_lock);
 579         LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
 580                  ra->ra_cur_pages, len);
 581         ra->ra_cur_pages -= len;
 582         spin_unlock(&sbi->ll_lock);
 583 }
 584
 585 /* called for each page in a completed rpc.*/
 586 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 587 {
 588         struct ll_async_page *llap;
 589         struct page *page;
 590         ENTRY;
 591
 592         llap = llap_from_cookie(data);
 593         if (IS_ERR(llap)) {
 594                 EXIT;
 595                 return;
 596         }
 597
 598         page = llap->llap_page;
 599         LASSERT(PageLocked(page));
 600
 601         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 602
 603         if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
 604                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 605
 606         if (rc == 0)  {
 607                 if (cmd == OBD_BRW_READ) {
 608                         if (!llap->llap_defer_uptodate)
 609                                 SetPageUptodate(page);
 610                 } else {
 611                         llap->llap_write_queued = 0;
 612                 }
 613                 ClearPageError(page);
 614         } else {
 615                 if (cmd == OBD_BRW_READ)
 616                         llap->llap_defer_uptodate = 0;
 617                 SetPageError(page);
 618         }
 619
 620         unlock_page(page);
 621
 622         if (0 && cmd == OBD_BRW_WRITE) {
 623                 llap_write_complete(page->mapping->host, llap);
 624                 ll_try_done_writing(page->mapping->host);
 625         }
 626
 627         if (PageWriteback(page)) {
 628                 end_page_writeback(page);
 629         }
 630         page_cache_release(page);
 631         EXIT;
 632 }
 633
 634 /* the kernel calls us here when a page is unhashed from the page cache.
 635  * the page will be locked and the kernel is holding a spinlock, so
 636  * we need to be careful.  we're just tearing down our book-keeping
 637  * here. */
 638 void ll_removepage(struct page *page)
 639 {
 640         struct inode *inode = page->mapping->host;
 641         struct obd_export *exp;
 642         struct ll_async_page *llap;
 643         struct ll_sb_info *sbi = ll_i2sbi(inode);
 644         int rc;
 645         ENTRY;
 646
 647         LASSERT(!in_interrupt());
 648
 649         /* sync pages or failed read pages can leave pages in the page
 650          * cache that don't have our data associated with them anymore */
 651         if (page->private == 0) {
 652                 EXIT;
 653                 return;
 654         }
 655
 656         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 657
 658         exp = ll_i2obdexp(inode);
 659         if (exp == NULL) {
 660                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 661                 EXIT;
 662                 return;
 663         }
 664
 665         llap = llap_from_page(page);
 666         if (IS_ERR(llap)) {
 667                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 668                        page->index, PTR_ERR(llap));
 669                 EXIT;
 670                 return;
 671         }
 672
 673         //llap_write_complete(inode, llap);
 674         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 675                                      llap->llap_cookie);
 676         if (rc != 0)
 677                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 678
 679         /* this unconditional free is only safe because the page lock
 680          * is providing exclusivity to memory pressure/truncate/writeback..*/
 681         __clear_page_ll_data(page);
 682
 683         spin_lock(&sbi->ll_lock);
 684         if (!list_empty(&llap->llap_proc_item))
 685                 list_del_init(&llap->llap_proc_item);
 686         sbi->ll_pglist_gen++;
 687         spin_unlock(&sbi->ll_lock);
 688         OBD_FREE(llap, sizeof(*llap));
 689         EXIT;
 690 }
 691
 692 static int ll_page_matches(struct page *page, int fd_flags)
 693 {
 694         struct lustre_handle match_lockh = {0};
 695         struct inode *inode = page->mapping->host;
 696         ldlm_policy_data_t page_extent;
 697         int flags, matches;
 698         ENTRY;
 699
 700         if (fd_flags & LL_FILE_GROUP_LOCKED)
 701                 RETURN(1);
 702
 703         page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
 704         page_extent.l_extent.end =
 705                 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
 706         flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK;
 707         matches = obd_match(ll_i2sbi(inode)->ll_lov_exp,
 708                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
 709                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
 710                             &match_lockh);
 711         RETURN(matches);
 712 }
 713
 714 static int ll_issue_page_read(struct obd_export *exp,
 715                               struct ll_async_page *llap,
 716                               struct obd_io_group *oig, int defer)
 717 {
 718         struct page *page = llap->llap_page;
 719         int rc;
 720
 721         page_cache_get(page);
 722         llap->llap_defer_uptodate = defer;
 723         llap->llap_ra_used = 0;
 724         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
 725                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
 726                                 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
 727                                               | ASYNC_URGENT);
 728         if (rc) {
 729                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
 730                 page_cache_release(page);
 731         }
 732         RETURN(rc);
 733 }
 734
 735 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
 736 {
 737         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
 738         ra->ra_stats[which]++;
 739 }
 740
 741 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
 742 {
 743         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
 744         struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
 745
 746         spin_lock(&sbi->ll_lock);
 747         ll_ra_stats_inc_unlocked(ra, which);
 748         spin_unlock(&sbi->ll_lock);
 749 }
 750
 751 void ll_ra_accounting(struct page *page, struct address_space *mapping)
 752 {
 753         struct ll_async_page *llap;
 754
 755         llap = llap_from_page(page);
 756         if (IS_ERR(llap))
 757                 return;
 758
 759         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
 760                 return;
 761
 762         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 763 }
 764
 765 #define RAS_CDEBUG(ras) \
 766         CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n",        \
 767                ras->ras_last_readpage, ras->ras_consecutive,            \
 768                ras->ras_window_start, ras->ras_window_len,              \
 769                ras->ras_next_readahead);
 770
 771 static int index_in_window(unsigned long index, unsigned long point,
 772                            unsigned long before, unsigned long after)
 773 {
 774         unsigned long start = point - before, end = point + after;
 775
 776         if (start > point)
 777                start = 0;
 778         if (end < point)
 779                end = ~0;
 780
 781         return start <= index && index <= end;
 782 }
 783
 784 static int ll_readahead(struct ll_readahead_state *ras,
 785                          struct obd_export *exp, struct address_space *mapping,
 786                          struct obd_io_group *oig, int flags)
 787 {
 788         unsigned long i, start = 0, end = 0, reserved;
 789         struct ll_async_page *llap;
 790         struct page *page;
 791         int rc, ret = 0, match_failed = 0;
 792         __u64 kms;
 793         ENTRY;
 794
 795         kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
 796         if (kms == 0) {
 797                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
 798                 RETURN(0);
 799         }
 800         spin_lock(&ras->ras_lock);
 801
 802         /* reserve a part of the read-ahead window that we'll be issuing */
 803         if (ras->ras_window_len) {
 804                 start = ras->ras_next_readahead;
 805                 end = ras->ras_window_start + ras->ras_window_len - 1;
 806                 end = min(end, (unsigned long)(kms >> PAGE_CACHE_SHIFT));
 807                 ras->ras_next_readahead = max(end, end + 1);
 808
 809                 RAS_CDEBUG(ras);
 810         }
 811
 812         spin_unlock(&ras->ras_lock);
 813
 814         if (end == 0) {
 815                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
 816                 RETURN(0);
 817         }
 818
 819         reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
 820         if (reserved < end - start + 1)
 821                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
 822
 823         for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
 824                 /* skip locked pages from previous readpage calls */
 825                 page = grab_cache_page_nowait(mapping, i);
 826                 if (page == NULL) {
 827                         CDEBUG(D_READA, "g_c_p_n failed\n");
 828                         continue;
 829                 }
 830
 831                 /* we do this first so that we can see the page in the /proc
 832                  * accounting */
 833                 llap = llap_from_page(page);
 834                 if (IS_ERR(llap) || llap->llap_defer_uptodate)
 835                         goto next_page;
 836
 837                 /* skip completed pages */
 838                 if (Page_Uptodate(page))
 839                         goto next_page;
 840
 841                 /* bail when we hit the end of the lock. */
 842                 if ((rc = ll_page_matches(page, flags)) <= 0) {
 843                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 844                                        "lock match failed: rc %d\n", rc);
 845                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
 846                         match_failed = 1;
 847                         goto next_page;
 848                 }
 849
 850                 rc = ll_issue_page_read(exp, llap, oig, 1);
 851                 if (rc == 0) {
 852                         reserved--;
 853                         ret++;
 854                         LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
 855                                        "started read-ahead\n");
 856                 }
 857                 if (rc) {
 858         next_page:
 859                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 860                                        "skipping read-ahead\n");
 861
 862                         unlock_page(page);
 863                 }
 864                 page_cache_release(page);
 865         }
 866
 867         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
 868         if (reserved != 0)
 869                 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
 870
 871         if (i == end + 1 && end == (kms >> PAGE_CACHE_SHIFT))
 872                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
 873
 874         /* if we didn't get to the end of the region we reserved from
 875          * the ras we need to go back and update the ras so that the
 876          * next read-ahead tries from where we left off.  we only do so
 877          * if the region we failed to issue read-ahead on is still ahead
 878          * of the app and behind the next index to start read-ahead from */
 879         if (i != end + 1) {
 880                 spin_lock(&ras->ras_lock);
 881                 if (i < ras->ras_next_readahead &&
 882                     index_in_window(i, ras->ras_window_start, 0,
 883                                     ras->ras_window_len)) {
 884                         ras->ras_next_readahead = i;
 885                         RAS_CDEBUG(ras);
 886                 }
 887                 spin_unlock(&ras->ras_lock);
 888         }
 889
 890         RETURN(ret);
 891 }
 892
 893 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
 894 {
 895         ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
 896 }
 897
 898 /* called with the ras_lock held or from places where it doesn't matter */
 899 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
 900 {
 901         ras->ras_last_readpage = index;
 902         ras->ras_consecutive = 1;
 903         ras->ras_window_len = 0;
 904         ras_set_start(ras, index);
 905         ras->ras_next_readahead = ras->ras_window_start;
 906
 907         RAS_CDEBUG(ras);
 908 }
 909
 910 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
 911 {
 912         spin_lock_init(&ras->ras_lock);
 913         ras_reset(ras, 0);
 914 }
 915
 916 static void ras_update(struct ll_sb_info *sbi, struct ll_readahead_state *ras,
 917                        unsigned long index, unsigned hit)
 918 {
 919         struct ll_ra_info *ra = &sbi->ll_ra_info;
 920         int zero = 0;
 921         ENTRY;
 922
 923         spin_lock(&sbi->ll_lock);
 924         spin_lock(&ras->ras_lock);
 925
 926         ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
 927
 928         /* reset the read-ahead window in two cases.  First when the app seeks
 929          * or reads to some other part of the file.  Secondly if we get a
 930          * read-ahead miss that we think we've previously issued.  This can
 931          * be a symptom of there being so many read-ahead pages that the VM is
 932          * reclaiming it before we get to it. */
 933         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
 934                 zero = 1;
 935                 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
 936         } else if (!hit && ras->ras_window_len &&
 937                    index < ras->ras_next_readahead &&
 938                    index_in_window(index, ras->ras_window_start, 0,
 939                                    ras->ras_window_len)) {
 940                 zero = 1;
 941                 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
 942         }
 943
 944         if (zero) {
 945                 ras_reset(ras, index);
 946                 GOTO(out_unlock, 0);
 947         }
 948
 949         ras->ras_last_readpage = index;
 950         ras->ras_consecutive++;
 951         ras_set_start(ras, index);
 952         ras->ras_next_readahead = max(ras->ras_window_start,
 953                                       ras->ras_next_readahead);
 954
 955         /* wait for a few pages to arrive before issuing readahead to avoid
 956          * the worst overutilization */
 957         if (ras->ras_consecutive == 3) {
 958                 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
 959                 GOTO(out_unlock, 0);
 960         }
 961
 962         /* we need to increase the window sometimes.  we'll arbitrarily
 963          * do it half-way through the pages in an rpc */
 964         if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
 965             (PTLRPC_MAX_BRW_PAGES >> 1)) {
 966                 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
 967                 ras->ras_window_len = min(ras->ras_window_len,
 968                                           ra->ra_max_pages);
 969
 970         }
 971
 972         EXIT;
 973 out_unlock:
 974         RAS_CDEBUG(ras);
 975         spin_unlock(&ras->ras_lock);
 976         spin_unlock(&sbi->ll_lock);
 977 }
 978
 979 /*
 980  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
 981  * read-ahead assumes it is valid to issue readpage all the way up to
 982  * i_size, but our dlm locks make that not the case.  We disable the
 983  * kernel's read-ahead and do our own by walking ahead in the page cache
 984  * checking for dlm lock coverage.  the main difference between 2.4 and
 985  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
 986  * so they look the same.
 987  */
 988 int ll_readpage(struct file *filp, struct page *page)
 989 {
 990         struct ll_file_data *fd = filp->private_data;
 991         struct inode *inode = page->mapping->host;
 992         struct obd_export *exp;
 993         struct ll_async_page *llap;
 994         struct obd_io_group *oig = NULL;
 995         int rc;
 996         ENTRY;
 997
 998         LASSERT(PageLocked(page));
 999         LASSERT(!PageUptodate(page));
1000         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
1001                inode->i_ino, inode->i_generation, inode,
1002                (((obd_off)page->index) << PAGE_SHIFT));
1003         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1004
1005         rc = oig_init(&oig);
1006         if (rc < 0)
1007                 GOTO(out, rc);
1008
1009         exp = ll_i2obdexp(inode);
1010         if (exp == NULL)
1011                 GOTO(out, rc = -EINVAL);
1012
1013         llap = llap_from_page(page);
1014         if (IS_ERR(llap))
1015                 GOTO(out, rc = PTR_ERR(llap));
1016
1017         if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1018                 ras_update(ll_i2sbi(inode), &fd->fd_ras, page->index,
1019                            llap->llap_defer_uptodate);
1020
1021         if (llap->llap_defer_uptodate) {
1022                 llap->llap_ra_used = 1;
1023                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1024                                   fd->fd_flags);
1025                 if (rc > 0)
1026                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1027                                              NULL, oig);
1028                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1029                 SetPageUptodate(page);
1030                 unlock_page(page);
1031                 GOTO(out_oig, rc = 0);
1032         }
1033
1034         rc = ll_page_matches(page, fd->fd_flags);
1035         if (rc < 0) {
1036                 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1037                 GOTO(out, rc);
1038         }
1039
1040         if (rc == 0) {
1041                 CWARN("ino %lu page %lu (%llu) not covered by "
1042                       "a lock (mmap?).  check debug logs.\n",
1043                       inode->i_ino, page->index,
1044                       (long long)page->index << PAGE_CACHE_SHIFT);
1045         }
1046
1047         rc = ll_issue_page_read(exp, llap, oig, 0);
1048         if (rc)
1049                 GOTO(out, rc);
1050
1051         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1052         if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1053                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1054                              fd->fd_flags);
1055
1056         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1057         EXIT;
1058 out:
1059         if (rc)
1060                 unlock_page(page);
1061 out_oig:
1062         if (oig != NULL)
1063                 oig_release(oig);
1064         return rc;
1065 }