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 brw_page pg;
  64         int rc;
  65         ENTRY;
  66
  67         pg.pg = page;
  68         pg.off = ((obd_off)page->index) << PAGE_SHIFT;
  69
  70         if (cmd == OBD_BRW_WRITE && (pg.off + PAGE_SIZE > inode->i_size))
  71                 pg.count = inode->i_size % PAGE_SIZE;
  72         else
  73                 pg.count = PAGE_SIZE;
  74
  75         CDEBUG(D_PAGE, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
  76                cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, inode->i_ino,
  77                pg.off, pg.off);
  78         if (pg.count == 0) {
  79                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
  80                        LPU64"\n",
  81                        inode->i_ino, inode, inode->i_size, page->mapping->host,
  82                        page->mapping->host->i_size, page->index, pg.off);
  83         }
  84
  85         pg.flag = flags;
  86
  87         if (cmd == OBD_BRW_WRITE)
  88                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
  89                                     LPROC_LL_BRW_WRITE, pg.count);
  90         else
  91                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
  92                                     LPROC_LL_BRW_READ, pg.count);
  93         rc = obd_brw(cmd, ll_i2obdexp(inode), oa, lsm, 1, &pg, NULL);
  94         if (rc == 0)
  95                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
  96         else if (rc != -EIO)
  97                 CERROR("error from obd_brw: rc = %d\n", rc);
  98         RETURN(rc);
  99 }
 100
 101 /* this isn't where truncate starts.   roughly:
 102  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate
 103  * we grab the lock back in setattr_raw to avoid races. */
 104 void ll_truncate(struct inode *inode)
 105 {
 106         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
 107         struct obdo oa;
 108         int rc;
 109         ENTRY;
 110         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
 111                inode->i_generation, inode);
 112
 113         if (!lsm) {
 114                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 115                        inode->i_ino);
 116                 EXIT;
 117                 return;
 118         }
 119
 120         oa.o_id = lsm->lsm_object_id;
 121         oa.o_valid = OBD_MD_FLID;
 122         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE|OBD_MD_FLMODE|OBD_MD_FLATIME|
 123                                     OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 124
 125         CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after %Lu)\n",
 126                oa.o_id, inode->i_size);
 127
 128         /* truncate == punch from new size to absolute end of file */
 129         /* NB: obd_punch must be called with i_sem held!  It updates the kms! */
 130         rc = obd_punch(ll_i2obdexp(inode), &oa, lsm, inode->i_size,
 131                        OBD_OBJECT_EOF, NULL);
 132         if (rc)
 133                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 134         else
 135                 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 136                                           OBD_MD_FLATIME | OBD_MD_FLMTIME |
 137                                           OBD_MD_FLCTIME);
 138
 139         EXIT;
 140         return;
 141 } /* ll_truncate */
 142
 143 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
 144 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 145                      unsigned to)
 146 {
 147         struct inode *inode = page->mapping->host;
 148         struct ll_inode_info *lli = ll_i2info(inode);
 149         struct lov_stripe_md *lsm = lli->lli_smd;
 150         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 151         struct brw_page pga;
 152         struct obdo oa;
 153         __u64 kms;
 154         int rc = 0;
 155         ENTRY;
 156
 157         LASSERT(PageLocked(page));
 158         (void)llap_cast_private(page); /* assertion */
 159
 160         /* Check to see if we should return -EIO right away */
 161         pga.pg = page;
 162         pga.off = offset;
 163         pga.count = PAGE_SIZE;
 164         pga.flag = 0;
 165
 166         oa.o_id = lsm->lsm_object_id;
 167         oa.o_mode = inode->i_mode;
 168         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
 169
 170         rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), &oa, lsm, 1, &pga,
 171                      NULL);
 172         if (rc)
 173                 RETURN(rc);
 174
 175         if (PageUptodate(page))
 176                 RETURN(0);
 177
 178         /* We're completely overwriting an existing page, so _don't_ set it up
 179          * to date until commit_write */
 180         if (from == 0 && to == PAGE_SIZE) {
 181                 POISON_PAGE(page, 0x11);
 182                 RETURN(0);
 183         }
 184
 185         /* If are writing to a new page, no need to read old data.  The extent
 186          * locking will have updated the KMS, and for our purposes here we can
 187          * treat it like i_size. */
 188         kms = lov_merge_size(lsm, 1);
 189         if (kms <= offset) {
 190                 memset(kmap(page), 0, PAGE_SIZE);
 191                 kunmap(page);
 192                 GOTO(prepare_done, rc = 0);
 193         }
 194
 195         /* XXX could be an async ocp read.. read-ahead? */
 196         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 197         if (rc == 0) {
 198                 /* bug 1598: don't clobber blksize */
 199                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 200                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 201         }
 202
 203         EXIT;
 204  prepare_done:
 205         if (rc == 0)
 206                 SetPageUptodate(page);
 207
 208         return rc;
 209 }
 210
 211 struct ll_async_page *llap_from_cookie(void *cookie)
 212 {
 213         struct ll_async_page *llap = cookie;
 214         if (llap->llap_magic != LLAP_MAGIC)
 215                 return ERR_PTR(-EINVAL);
 216         return llap;
 217 };
 218
 219 static int ll_ap_make_ready(void *data, int cmd)
 220 {
 221         struct ll_async_page *llap;
 222         struct page *page;
 223         ENTRY;
 224
 225         llap = llap_from_cookie(data);
 226         if (IS_ERR(llap))
 227                 RETURN(-EINVAL);
 228
 229         page = llap->llap_page;
 230
 231         LASSERT(cmd != OBD_BRW_READ);
 232
 233         /* we're trying to write, but the page is locked.. come back later */
 234         if (TryLockPage(page))
 235                 RETURN(-EAGAIN);
 236
 237         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 238         page_cache_get(page);
 239
 240         /* if we left PageDirty we might get another writepage call
 241          * in the future.  list walkers are bright enough
 242          * to check page dirty so we can leave it on whatever list
 243          * its on.  XXX also, we're called with the cli list so if
 244          * we got the page cache list we'd create a lock inversion
 245          * with the removepage path which gets the page lock then the
 246          * cli lock */
 247         clear_page_dirty(page);
 248         RETURN(0);
 249 }
 250
 251 /* We have two reasons for giving llite the opportunity to change the
 252  * write length of a given queued page as it builds the RPC containing
 253  * the page:
 254  *
 255  * 1) Further extending writes may have landed in the page cache
 256  *    since a partial write first queued this page requiring us
 257  *    to write more from the page cache.
 258  * 2) We might have raced with truncate and want to avoid performing
 259  *    write RPCs that are just going to be thrown away by the
 260  *    truncate's punch on the storage targets.
 261  *
 262  * The kms serves these purposes as it is set at both truncate and extending
 263  * writes.
 264  */
 265 static int ll_ap_refresh_count(void *data, int cmd)
 266 {
 267         struct ll_async_page *llap;
 268         struct lov_stripe_md *lsm;
 269         struct page *page;
 270         __u64 kms;
 271         ENTRY;
 272
 273         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 274         LASSERT(cmd != OBD_BRW_READ);
 275
 276         llap = llap_from_cookie(data);
 277         if (IS_ERR(llap))
 278                 RETURN(PTR_ERR(llap));
 279
 280         page = llap->llap_page;
 281         lsm = ll_i2info(page->mapping->host)->lli_smd;
 282         kms = lov_merge_size(lsm, 1);
 283
 284         /* catch race with truncate */
 285         if (((__u64)page->index << PAGE_SHIFT) >= kms)
 286                 return 0;
 287
 288         /* catch sub-page write at end of file */
 289         if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms)
 290                 return kms % PAGE_SIZE;
 291
 292         return PAGE_SIZE;
 293 }
 294
 295 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 296 {
 297         struct lov_stripe_md *lsm;
 298         obd_flag valid_flags;
 299
 300         lsm = ll_i2info(inode)->lli_smd;
 301
 302         oa->o_id = lsm->lsm_object_id;
 303         oa->o_valid = OBD_MD_FLID;
 304         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 305         if (cmd == OBD_BRW_WRITE) {
 306                 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
 307                 mdc_pack_fid(obdo_fid(oa), inode->i_ino, 0, inode->i_mode);
 308                 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
 309
 310                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 311         }
 312
 313         obdo_from_inode(oa, inode, valid_flags);
 314 }
 315
 316 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 317 {
 318         struct ll_async_page *llap;
 319         ENTRY;
 320
 321         llap = llap_from_cookie(data);
 322         if (IS_ERR(llap)) {
 323                 EXIT;
 324                 return;
 325         }
 326
 327         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 328         EXIT;
 329 }
 330
 331 static struct obd_async_page_ops ll_async_page_ops = {
 332         .ap_make_ready =        ll_ap_make_ready,
 333         .ap_refresh_count =     ll_ap_refresh_count,
 334         .ap_fill_obdo =         ll_ap_fill_obdo,
 335         .ap_completion =        ll_ap_completion,
 336 };
 337
 338 struct ll_async_page *llap_cast_private(struct page *page)
 339 {
 340         struct ll_async_page *llap = (struct ll_async_page *)page->private;
 341
 342         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 343                  "page %p private %lu gave magic %d which != %d\n",
 344                  page, page->private, llap->llap_magic, LLAP_MAGIC);
 345
 346         return llap;
 347 }
 348
 349 /* XXX have the exp be an argument? */
 350 struct ll_async_page *llap_from_page(struct page *page)
 351 {
 352         struct ll_async_page *llap;
 353         struct obd_export *exp;
 354         struct inode *inode = page->mapping->host;
 355         struct ll_sb_info *sbi = ll_i2sbi(inode);
 356         int rc;
 357         ENTRY;
 358
 359         llap = llap_cast_private(page);
 360         if (llap != NULL)
 361                 RETURN(llap);
 362
 363         exp = ll_i2obdexp(page->mapping->host);
 364         if (exp == NULL)
 365                 RETURN(ERR_PTR(-EINVAL));
 366
 367         OBD_ALLOC(llap, sizeof(*llap));
 368         if (llap == NULL)
 369                 RETURN(ERR_PTR(-ENOMEM));
 370         llap->llap_magic = LLAP_MAGIC;
 371         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 372                                  (obd_off)page->index << PAGE_SHIFT,
 373                                  &ll_async_page_ops, llap, &llap->llap_cookie);
 374         if (rc) {
 375                 OBD_FREE(llap, sizeof(*llap));
 376                 RETURN(ERR_PTR(rc));
 377         }
 378
 379         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 380                page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
 381         /* also zeroing the PRIVBITS low order bitflags */
 382         page->private = (unsigned long)llap;
 383         llap->llap_page = page;
 384
 385         spin_lock(&sbi->ll_lock);
 386         sbi->ll_pglist_gen++;
 387         list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
 388         spin_unlock(&sbi->ll_lock);
 389
 390         RETURN(llap);
 391 }
 392
 393 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
 394                       obd_off size);
 395 /* update our write count to account for i_size increases that may have
 396  * happened since we've queued the page for io. */
 397
 398 /* be careful not to return success without setting the page Uptodate or
 399  * the next pass through prepare_write will read in stale data from disk. */
 400 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 401                     unsigned to)
 402 {
 403         struct inode *inode = page->mapping->host;
 404         struct ll_inode_info *lli = ll_i2info(inode);
 405         struct lov_stripe_md *lsm = lli->lli_smd;
 406         struct obd_export *exp = NULL;
 407         struct ll_async_page *llap;
 408         loff_t size;
 409         int rc = 0;
 410         ENTRY;
 411
 412         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 413         LASSERT(inode == file->f_dentry->d_inode);
 414         LASSERT(PageLocked(page));
 415
 416         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 417                inode, page, from, to, page->index);
 418
 419         llap = llap_from_page(page);
 420         if (IS_ERR(llap))
 421                 RETURN(PTR_ERR(llap));
 422
 423         /* queue a write for some time in the future the first time we
 424          * dirty the page */
 425         if (!PageDirty(page)) {
 426                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 427                                      LPROC_LL_DIRTY_MISSES);
 428
 429                 exp = ll_i2obdexp(inode);
 430                 if (exp == NULL)
 431                         RETURN(-EINVAL);
 432
 433                 /* _make_ready only sees llap once we've unlocked the page */
 434                 llap->llap_write_queued = 1;
 435                 rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
 436                                         OBD_BRW_WRITE, 0, 0, 0, 0);
 437                 if (rc != 0) { /* async failed, try sync.. */
 438                         struct obd_io_group *oig;
 439                         rc = oig_init(&oig);
 440                         if (rc)
 441                                 GOTO(out, rc);
 442
 443                         llap->llap_write_queued = 0;
 444                         rc = obd_queue_group_io(exp, lsm, NULL, oig,
 445                                                 llap->llap_cookie,
 446                                                 OBD_BRW_WRITE, 0, to, 0,
 447                                                 ASYNC_READY | ASYNC_URGENT |
 448                                                 ASYNC_COUNT_STABLE |
 449                                                 ASYNC_GROUP_SYNC);
 450
 451                         if (rc)
 452                                 GOTO(free_oig, rc);
 453
 454                         rc = obd_trigger_group_io(exp, lsm, NULL, oig);
 455                         if (rc)
 456                                 GOTO(free_oig, rc);
 457
 458                         rc = oig_wait(oig);
 459 free_oig:
 460                         oig_release(oig);
 461                         GOTO(out, rc);
 462                 }
 463                 LL_CDEBUG_PAGE(D_PAGE, page, "write queued\n");
 464                 //llap_write_pending(inode, llap);
 465         } else {
 466                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 467                                      LPROC_LL_DIRTY_HITS);
 468         }
 469
 470         /* put the page in the page cache, from now on ll_removepage is
 471          * responsible for cleaning up the llap */
 472         set_page_dirty(page);
 473
 474 out:
 475         if (rc == 0) {
 476                 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
 477                 lov_increase_kms(exp, lsm, size);
 478                 if (size > inode->i_size)
 479                         inode->i_size = size;
 480                 SetPageUptodate(page);
 481         }
 482         RETURN(rc);
 483 }
 484
 485 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 486 {
 487         unsigned long ret;
 488         ENTRY;
 489
 490         spin_lock(&sbi->ll_lock);
 491         ret = min(sbi->ll_max_read_ahead_pages - sbi->ll_read_ahead_pages,
 492                   len);
 493         sbi->ll_read_ahead_pages += ret;
 494         spin_unlock(&sbi->ll_lock);
 495
 496         RETURN(ret);
 497 }
 498
 499 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 500 {
 501         spin_lock(&sbi->ll_lock);
 502         LASSERTF(sbi->ll_read_ahead_pages >= len, "r_a_p %lu len %lu\n",
 503                  sbi->ll_read_ahead_pages, len);
 504         sbi->ll_read_ahead_pages -= len;
 505         spin_unlock(&sbi->ll_lock);
 506 }
 507
 508 /* called for each page in a completed rpc.*/
 509 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 510 {
 511         struct ll_async_page *llap;
 512         struct page *page;
 513         ENTRY;
 514
 515         llap = llap_from_cookie(data);
 516         if (IS_ERR(llap)) {
 517                 EXIT;
 518                 return;
 519         }
 520
 521         page = llap->llap_page;
 522         LASSERT(PageLocked(page));
 523
 524         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 525
 526         if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
 527                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 528
 529         if (rc == 0)  {
 530                 if (cmd == OBD_BRW_READ) {
 531                         if (!llap->llap_defer_uptodate)
 532                                 SetPageUptodate(page);
 533                 } else {
 534                         llap->llap_write_queued = 0;
 535                 }
 536                 ClearPageError(page);
 537         } else {
 538                 if (cmd == OBD_BRW_READ)
 539                         llap->llap_defer_uptodate = 0;
 540                 SetPageError(page);
 541         }
 542
 543         unlock_page(page);
 544
 545         if (0 && cmd == OBD_BRW_WRITE) {
 546                 llap_write_complete(page->mapping->host, llap);
 547                 ll_try_done_writing(page->mapping->host);
 548         }
 549
 550         page_cache_release(page);
 551         EXIT;
 552 }
 553
 554 /* the kernel calls us here when a page is unhashed from the page cache.
 555  * the page will be locked and the kernel is holding a spinlock, so
 556  * we need to be careful.  we're just tearing down our book-keeping
 557  * here. */
 558 void ll_removepage(struct page *page)
 559 {
 560         struct inode *inode = page->mapping->host;
 561         struct obd_export *exp;
 562         struct ll_async_page *llap;
 563         struct ll_sb_info *sbi = ll_i2sbi(inode);
 564         int rc;
 565         ENTRY;
 566
 567         LASSERT(!in_interrupt());
 568
 569         /* sync pages or failed read pages can leave pages in the page
 570          * cache that don't have our data associated with them anymore */
 571         if (page->private == 0) {
 572                 EXIT;
 573                 return;
 574         }
 575
 576         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 577
 578         exp = ll_i2obdexp(inode);
 579         if (exp == NULL) {
 580                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 581                 EXIT;
 582                 return;
 583         }
 584
 585         llap = llap_from_page(page);
 586         if (IS_ERR(llap)) {
 587                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 588                        page->index, PTR_ERR(llap));
 589                 EXIT;
 590                 return;
 591         }
 592
 593         //llap_write_complete(inode, llap);
 594         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 595                                      llap->llap_cookie);
 596         if (rc != 0)
 597                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 598
 599         /* this unconditional free is only safe because the page lock
 600          * is providing exclusivity to memory pressure/truncate/writeback..*/
 601         page->private = 0;
 602
 603         spin_lock(&sbi->ll_lock);
 604         if (!list_empty(&llap->llap_proc_item))
 605                 list_del_init(&llap->llap_proc_item);
 606         sbi->ll_pglist_gen++;
 607         spin_unlock(&sbi->ll_lock);
 608         OBD_FREE(llap, sizeof(*llap));
 609         EXIT;
 610 }
 611
 612 static int ll_page_matches(struct page *page, int fd_flags)
 613 {
 614         struct lustre_handle match_lockh = {0};
 615         struct inode *inode = page->mapping->host;
 616         ldlm_policy_data_t page_extent;
 617         int flags, matches;
 618         ENTRY;
 619
 620         if (fd_flags & LL_FILE_GROUP_LOCKED)
 621                 RETURN(1);
 622
 623         page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
 624         page_extent.l_extent.end =
 625                 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
 626         flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK;
 627         matches = obd_match(ll_i2sbi(inode)->ll_osc_exp,
 628                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
 629                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
 630                             &match_lockh);
 631         RETURN(matches);
 632 }
 633
 634 static int ll_issue_page_read(struct obd_export *exp,
 635                               struct ll_async_page *llap,
 636                               struct obd_io_group *oig, int defer)
 637 {
 638         struct page *page = llap->llap_page;
 639         int rc;
 640
 641         page_cache_get(page);
 642         llap->llap_defer_uptodate = defer;
 643         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
 644                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
 645                                 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
 646                                               | ASYNC_URGENT);
 647         if (rc) {
 648                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
 649                 page_cache_release(page);
 650         }
 651         RETURN(rc);
 652 }
 653
 654 #define RAS_CDEBUG(ras) \
 655         CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n",        \
 656                ras->ras_last_readpage, ras->ras_consecutive,            \
 657                ras->ras_window_start, ras->ras_window_len,              \
 658                ras->ras_next_readahead);
 659
 660 static int ll_readahead(struct ll_readahead_state *ras,
 661                          struct obd_export *exp, struct address_space *mapping,
 662                          struct obd_io_group *oig, int flags)
 663 {
 664         unsigned long i, start = 0, end = 0, reserved;
 665         struct ll_async_page *llap;
 666         struct page *page;
 667         int rc, ret = 0;
 668         __u64 kms;
 669         ENTRY;
 670
 671         kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
 672         if (kms == 0)
 673                 RETURN(0);
 674
 675         spin_lock(&ras->ras_lock);
 676
 677         if (ras->ras_window_len) {
 678                 start = ras->ras_next_readahead;
 679                 end = ras->ras_window_start + ras->ras_window_len - 1;
 680                 end = min(end, (unsigned long)(kms >> PAGE_CACHE_SHIFT));
 681                 ras->ras_next_readahead = max(end, end + 1);
 682
 683                 RAS_CDEBUG(ras);
 684         }
 685
 686         spin_unlock(&ras->ras_lock);
 687
 688         if (end == 0)
 689                 RETURN(0);
 690
 691         reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
 692
 693         for (i = start; reserved > 0 && i <= end; i++) {
 694                 /* skip locked pages from previous readpage calls */
 695                 page = grab_cache_page_nowait(mapping, i);
 696                 if (page == NULL) {
 697                         CDEBUG(D_READA, "g_c_p_n failed\n");
 698                         continue;
 699                 }
 700
 701                 /* we do this first so that we can see the page in the /proc
 702                  * accounting */
 703                 llap = llap_from_page(page);
 704                 if (IS_ERR(llap) || llap->llap_defer_uptodate)
 705                         goto next_page;
 706
 707                 /* skip completed pages */
 708                 if (Page_Uptodate(page))
 709                         goto next_page;
 710
 711                 /* bail when we hit the end of the lock. */
 712                 if ((rc = ll_page_matches(page, flags)) <= 0) {
 713                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 714                                        "lock match failed: rc %d\n", rc);
 715                         i = end;
 716                         goto next_page;
 717                 }
 718
 719                 rc = ll_issue_page_read(exp, llap, oig, 1);
 720                 if (rc == 0) {
 721                         reserved--;
 722                         ret++;
 723                         LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
 724                                        "started read-ahead\n");
 725                 }
 726                 if (rc) {
 727         next_page:
 728                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 729                                        "skipping read-ahead\n");
 730
 731                         unlock_page(page);
 732                 }
 733                 page_cache_release(page);
 734         }
 735
 736         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
 737         if (reserved != 0)
 738                 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
 739         RETURN(ret);
 740 }
 741
 742 static void ras_set_start(struct ll_readahead_state *ras,
 743                                unsigned long index)
 744 {
 745         ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
 746         ras->ras_next_readahead = max(ras->ras_window_start,
 747                                       ras->ras_next_readahead);
 748 }
 749
 750 /* called with the ras_lock held or from places where it doesn't matter */
 751 static void ras_reset(struct ll_readahead_state *ras,
 752                       unsigned long index)
 753 {
 754         ras->ras_last_readpage = index;
 755         ras->ras_consecutive = 1;
 756         ras->ras_window_len = 0;
 757         ras_set_start(ras, index);
 758         ras->ras_next_readahead = ras->ras_window_start;
 759
 760         RAS_CDEBUG(ras);
 761 }
 762
 763 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
 764 {
 765         spin_lock_init(&ras->ras_lock);
 766         ras_reset(ras, 0);
 767 }
 768
 769 static void ras_update(struct ll_readahead_state *ras,
 770                        unsigned long index, unsigned long max)
 771 {
 772         ENTRY;
 773
 774         spin_lock(&ras->ras_lock);
 775
 776         if (index != ras->ras_last_readpage + 1) {
 777                 ras_reset(ras, index);
 778                 GOTO(out_unlock, 0);
 779         }
 780
 781         ras->ras_last_readpage = index;
 782         ras->ras_consecutive++;
 783         ras_set_start(ras, index);
 784
 785         if (ras->ras_consecutive == 2) {
 786                 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
 787                 GOTO(out_unlock, 0);
 788         }
 789
 790         /* we need to increase the window sometimes.  we'll arbitrarily
 791          * do it half-way through the pages in an rpc */
 792         if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
 793             (PTLRPC_MAX_BRW_PAGES >> 1)) {
 794                 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
 795                 ras->ras_window_len = min(ras->ras_window_len, max);
 796         }
 797
 798         EXIT;
 799 out_unlock:
 800         RAS_CDEBUG(ras);
 801         spin_unlock(&ras->ras_lock);
 802         return;
 803 }
 804
 805 /*
 806  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
 807  * read-ahead assumes it is valid to issue readpage all the way up to
 808  * i_size, but our dlm locks make that not the case.  We disable the
 809  * kernel's read-ahead and do our own by walking ahead in the page cache
 810  * checking for dlm lock coverage.  the main difference between 2.4 and
 811  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
 812  * so they look the same.
 813  */
 814 int ll_readpage(struct file *filp, struct page *page)
 815 {
 816         struct ll_file_data *fd = filp->private_data;
 817         struct inode *inode = page->mapping->host;
 818         struct obd_export *exp;
 819         struct ll_async_page *llap;
 820         struct obd_io_group *oig = NULL;
 821         int rc;
 822         ENTRY;
 823
 824         LASSERT(PageLocked(page));
 825         LASSERT(!PageUptodate(page));
 826         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
 827                inode->i_ino, inode->i_generation, inode,
 828                (((obd_off)page->index) << PAGE_SHIFT));
 829         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
 830
 831         rc = oig_init(&oig);
 832         if (rc < 0)
 833                 GOTO(out, rc);
 834
 835         exp = ll_i2obdexp(inode);
 836         if (exp == NULL)
 837                 GOTO(out, rc = -EINVAL);
 838
 839         llap = llap_from_page(page);
 840         if (IS_ERR(llap))
 841                 GOTO(out, rc = PTR_ERR(llap));
 842
 843         if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
 844                 ras_update(&fd->fd_ras, page->index,
 845                            ll_i2sbi(inode)->ll_max_read_ahead_pages);
 846
 847         if (llap->llap_defer_uptodate) {
 848                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
 849                                   fd->fd_flags);
 850                 if (rc > 0)
 851                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
 852                                              NULL, oig);
 853                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
 854                 SetPageUptodate(page);
 855                 unlock_page(page);
 856                 GOTO(out_oig, rc = 0);
 857         }
 858
 859         rc = ll_page_matches(page, fd->fd_flags);
 860         if (rc < 0) {
 861                 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
 862                 GOTO(out, rc);
 863         }
 864
 865         if (rc == 0) {
 866                 static unsigned long next_print;
 867                 CDEBUG(D_INODE, "ino %lu page %lu (%llu) didn't match a lock\n",
 868                        inode->i_ino, page->index,
 869                        (long long)page->index << PAGE_CACHE_SHIFT);
 870                 if (time_after(jiffies, next_print)) {
 871                         CWARN("ino %lu page %lu (%llu) not covered by "
 872                                "a lock (mmap?).  check debug logs.\n",
 873                                inode->i_ino, page->index,
 874                                (long long)page->index << PAGE_CACHE_SHIFT);
 875                         next_print = jiffies + 30 * HZ;
 876                 }
 877         }
 878
 879         rc = ll_issue_page_read(exp, llap, oig, 0);
 880         if (rc)
 881                 GOTO(out, rc);
 882
 883         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
 884         if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
 885                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
 886                              fd->fd_flags);
 887
 888         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
 889
 890 out:
 891         if (rc)
 892                 unlock_page(page);
 893 out_oig:
 894         if (oig != NULL)
 895                 oig_release(oig);
 896         RETURN(rc);
 897 }