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.disk_offset = pg.page_offset = ((obd_off)page->index) << PAGE_SHIFT;
  69
  70         if (cmd == OBD_BRW_WRITE &&
  71             (pg.disk_offset + PAGE_SIZE > inode->i_size))
  72                 pg.count = inode->i_size % PAGE_SIZE;
  73         else
  74                 pg.count = PAGE_SIZE;
  75
  76         CDEBUG(D_PAGE, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
  77                cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, inode->i_ino,
  78                pg.disk_offset, pg.disk_offset);
  79         if (pg.count == 0) {
  80                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
  81                        LPU64"\n", inode->i_ino, inode, inode->i_size,
  82                        page->mapping->host, page->mapping->host->i_size,
  83                        page->index, pg.disk_offset);
  84         }
  85
  86         pg.flag = flags;
  87
  88         if (cmd == OBD_BRW_WRITE)
  89                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
  90                                     LPROC_LL_BRW_WRITE, pg.count);
  91         else
  92                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
  93                                     LPROC_LL_BRW_READ, pg.count);
  94         rc = obd_brw(cmd, ll_i2obdexp(inode), oa, lsm, 1, &pg, NULL);
  95         if (rc == 0)
  96                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
  97         else if (rc != -EIO)
  98                 CERROR("error from obd_brw: rc = %d\n", rc);
  99         RETURN(rc);
 100 }
 101
 102 /* this isn't where truncate starts.   roughly:
 103  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate
 104  * we grab the lock back in setattr_raw to avoid races. */
 105 void ll_truncate(struct inode *inode)
 106 {
 107         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
 108         struct obdo oa;
 109         int rc;
 110         ENTRY;
 111         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
 112                inode->i_generation, inode);
 113
 114         if (!lsm) {
 115                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 116                        inode->i_ino);
 117                 EXIT;
 118                 return;
 119         }
 120
 121         oa.o_id = lsm->lsm_object_id;
 122         oa.o_valid = OBD_MD_FLID;
 123         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE|OBD_MD_FLMODE|OBD_MD_FLATIME|
 124                                     OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 125
 126         CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after %Lu)\n",
 127                oa.o_id, inode->i_size);
 128
 129         /* truncate == punch from new size to absolute end of file */
 130         /* NB: obd_punch must be called with i_sem held!  It updates the kms! */
 131         rc = obd_punch(ll_i2obdexp(inode), &oa, lsm, inode->i_size,
 132                        OBD_OBJECT_EOF, NULL);
 133         if (rc)
 134                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 135         else
 136                 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 137                                           OBD_MD_FLATIME | OBD_MD_FLMTIME |
 138                                           OBD_MD_FLCTIME);
 139
 140         EXIT;
 141         return;
 142 } /* ll_truncate */
 143
 144 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
 145 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 146                      unsigned to)
 147 {
 148         struct inode *inode = page->mapping->host;
 149         struct ll_inode_info *lli = ll_i2info(inode);
 150         struct lov_stripe_md *lsm = lli->lli_smd;
 151         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 152         struct brw_page pga;
 153         struct obdo oa;
 154         __u64 kms;
 155         int rc = 0;
 156         ENTRY;
 157
 158         LASSERT(PageLocked(page));
 159         (void)llap_cast_private(page); /* assertion */
 160
 161         /* Check to see if we should return -EIO right away */
 162         pga.pg = page;
 163         pga.disk_offset = pga.page_offset = offset;
 164         pga.count = PAGE_SIZE;
 165         pga.flag = 0;
 166
 167         oa.o_id = lsm->lsm_object_id;
 168         oa.o_mode = inode->i_mode;
 169         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
 170
 171         rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), &oa, lsm, 1, &pga,
 172                      NULL);
 173         if (rc)
 174                 RETURN(rc);
 175
 176         if (PageUptodate(page))
 177                 RETURN(0);
 178
 179         /* We're completely overwriting an existing page, so _don't_ set it up
 180          * to date until commit_write */
 181         if (from == 0 && to == PAGE_SIZE) {
 182                 POISON_PAGE(page, 0x11);
 183                 RETURN(0);
 184         }
 185
 186         /* If are writing to a new page, no need to read old data.  The extent
 187          * locking will have updated the KMS, and for our purposes here we can
 188          * treat it like i_size. */
 189         kms = lov_merge_size(lsm, 1);
 190         if (kms <= offset) {
 191                 memset(kmap(page), 0, PAGE_SIZE);
 192                 kunmap(page);
 193                 GOTO(prepare_done, rc = 0);
 194         }
 195
 196         /* XXX could be an async ocp read.. read-ahead? */
 197         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 198         if (rc == 0) {
 199                 /* bug 1598: don't clobber blksize */
 200                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 201                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 202         }
 203
 204         EXIT;
 205  prepare_done:
 206         if (rc == 0)
 207                 SetPageUptodate(page);
 208
 209         return rc;
 210 }
 211
 212 struct ll_async_page *llap_from_cookie(void *cookie)
 213 {
 214         struct ll_async_page *llap = cookie;
 215         if (llap->llap_magic != LLAP_MAGIC)
 216                 return ERR_PTR(-EINVAL);
 217         return llap;
 218 };
 219
 220 static int ll_ap_make_ready(void *data, int cmd)
 221 {
 222         struct ll_async_page *llap;
 223         struct page *page;
 224         ENTRY;
 225
 226         llap = llap_from_cookie(data);
 227         if (IS_ERR(llap))
 228                 RETURN(-EINVAL);
 229
 230         page = llap->llap_page;
 231
 232         LASSERT(cmd != OBD_BRW_READ);
 233
 234         /* we're trying to write, but the page is locked.. come back later */
 235         if (TryLockPage(page))
 236                 RETURN(-EAGAIN);
 237
 238         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 239         page_cache_get(page);
 240
 241         /* if we left PageDirty we might get another writepage call
 242          * in the future.  list walkers are bright enough
 243          * to check page dirty so we can leave it on whatever list
 244          * its on.  XXX also, we're called with the cli list so if
 245          * we got the page cache list we'd create a lock inversion
 246          * with the removepage path which gets the page lock then the
 247          * cli lock */
 248         clear_page_dirty(page);
 249         RETURN(0);
 250 }
 251
 252 /* We have two reasons for giving llite the opportunity to change the
 253  * write length of a given queued page as it builds the RPC containing
 254  * the page:
 255  *
 256  * 1) Further extending writes may have landed in the page cache
 257  *    since a partial write first queued this page requiring us
 258  *    to write more from the page cache.
 259  * 2) We might have raced with truncate and want to avoid performing
 260  *    write RPCs that are just going to be thrown away by the
 261  *    truncate's punch on the storage targets.
 262  *
 263  * The kms serves these purposes as it is set at both truncate and extending
 264  * writes.
 265  */
 266 static int ll_ap_refresh_count(void *data, int cmd)
 267 {
 268         struct ll_async_page *llap;
 269         struct lov_stripe_md *lsm;
 270         struct page *page;
 271         __u64 kms;
 272         ENTRY;
 273
 274         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 275         LASSERT(cmd != OBD_BRW_READ);
 276
 277         llap = llap_from_cookie(data);
 278         if (IS_ERR(llap))
 279                 RETURN(PTR_ERR(llap));
 280
 281         page = llap->llap_page;
 282         lsm = ll_i2info(page->mapping->host)->lli_smd;
 283         kms = lov_merge_size(lsm, 1);
 284
 285         /* catch race with truncate */
 286         if (((__u64)page->index << PAGE_SHIFT) >= kms)
 287                 return 0;
 288
 289         /* catch sub-page write at end of file */
 290         if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms)
 291                 return kms % PAGE_SIZE;
 292
 293         return PAGE_SIZE;
 294 }
 295
 296 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 297 {
 298         struct lov_stripe_md *lsm;
 299         obd_flag valid_flags;
 300
 301         lsm = ll_i2info(inode)->lli_smd;
 302
 303         oa->o_id = lsm->lsm_object_id;
 304         oa->o_valid = OBD_MD_FLID;
 305         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 306         if (cmd == OBD_BRW_WRITE) {
 307                 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
 308                 mdc_pack_fid(obdo_fid(oa), inode->i_ino, 0, inode->i_mode);
 309                 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
 310
 311                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 312         }
 313
 314         obdo_from_inode(oa, inode, valid_flags);
 315 }
 316
 317 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 318 {
 319         struct ll_async_page *llap;
 320         ENTRY;
 321
 322         llap = llap_from_cookie(data);
 323         if (IS_ERR(llap)) {
 324                 EXIT;
 325                 return;
 326         }
 327
 328         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 329         EXIT;
 330 }
 331
 332 static struct obd_async_page_ops ll_async_page_ops = {
 333         .ap_make_ready =        ll_ap_make_ready,
 334         .ap_refresh_count =     ll_ap_refresh_count,
 335         .ap_fill_obdo =         ll_ap_fill_obdo,
 336         .ap_completion =        ll_ap_completion,
 337 };
 338
 339 struct ll_async_page *llap_cast_private(struct page *page)
 340 {
 341         struct ll_async_page *llap = (struct ll_async_page *)page->private;
 342
 343         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 344                  "page %p private %lu gave magic %d which != %d\n",
 345                  page, page->private, llap->llap_magic, LLAP_MAGIC);
 346
 347         return llap;
 348 }
 349
 350 /* XXX have the exp be an argument? */
 351 struct ll_async_page *llap_from_page(struct page *page)
 352 {
 353         struct ll_async_page *llap;
 354         struct obd_export *exp;
 355         struct inode *inode = page->mapping->host;
 356         struct ll_sb_info *sbi = ll_i2sbi(inode);
 357         int rc;
 358         ENTRY;
 359
 360         llap = llap_cast_private(page);
 361         if (llap != NULL)
 362                 RETURN(llap);
 363
 364         exp = ll_i2obdexp(page->mapping->host);
 365         if (exp == NULL)
 366                 RETURN(ERR_PTR(-EINVAL));
 367
 368         OBD_ALLOC(llap, sizeof(*llap));
 369         if (llap == NULL)
 370                 RETURN(ERR_PTR(-ENOMEM));
 371         llap->llap_magic = LLAP_MAGIC;
 372         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 373                                  (obd_off)page->index << PAGE_SHIFT,
 374                                  &ll_async_page_ops, llap, &llap->llap_cookie);
 375         if (rc) {
 376                 OBD_FREE(llap, sizeof(*llap));
 377                 RETURN(ERR_PTR(rc));
 378         }
 379
 380         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 381                page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
 382         /* also zeroing the PRIVBITS low order bitflags */
 383         page->private = (unsigned long)llap;
 384         llap->llap_page = page;
 385
 386         spin_lock(&sbi->ll_lock);
 387         sbi->ll_pglist_gen++;
 388         list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
 389         spin_unlock(&sbi->ll_lock);
 390
 391         RETURN(llap);
 392 }
 393
 394 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
 395                       obd_off size);
 396 /* update our write count to account for i_size increases that may have
 397  * happened since we've queued the page for io. */
 398
 399 /* be careful not to return success without setting the page Uptodate or
 400  * the next pass through prepare_write will read in stale data from disk. */
 401 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 402                     unsigned to)
 403 {
 404         struct inode *inode = page->mapping->host;
 405         struct ll_inode_info *lli = ll_i2info(inode);
 406         struct lov_stripe_md *lsm = lli->lli_smd;
 407         struct obd_export *exp = NULL;
 408         struct ll_async_page *llap;
 409         loff_t size;
 410         int rc = 0;
 411         ENTRY;
 412
 413         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 414         LASSERT(inode == file->f_dentry->d_inode);
 415         LASSERT(PageLocked(page));
 416
 417         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 418                inode, page, from, to, page->index);
 419
 420         llap = llap_from_page(page);
 421         if (IS_ERR(llap))
 422                 RETURN(PTR_ERR(llap));
 423
 424         /* queue a write for some time in the future the first time we
 425          * dirty the page */
 426         if (!PageDirty(page)) {
 427                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 428                                      LPROC_LL_DIRTY_MISSES);
 429
 430                 exp = ll_i2obdexp(inode);
 431                 if (exp == NULL)
 432                         RETURN(-EINVAL);
 433
 434                 /* _make_ready only sees llap once we've unlocked the page */
 435                 llap->llap_write_queued = 1;
 436                 rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
 437                                         OBD_BRW_WRITE, 0, 0, 0, 0);
 438                 if (rc != 0) { /* async failed, try sync.. */
 439                         struct obd_io_group *oig;
 440                         rc = oig_init(&oig);
 441                         if (rc)
 442                                 GOTO(out, rc);
 443
 444                         llap->llap_write_queued = 0;
 445                         rc = obd_queue_group_io(exp, lsm, NULL, oig,
 446                                                 llap->llap_cookie,
 447                                                 OBD_BRW_WRITE, 0, to, 0,
 448                                                 ASYNC_READY | ASYNC_URGENT |
 449                                                 ASYNC_COUNT_STABLE |
 450                                                 ASYNC_GROUP_SYNC);
 451
 452                         if (rc)
 453                                 GOTO(free_oig, rc);
 454
 455                         rc = obd_trigger_group_io(exp, lsm, NULL, oig);
 456                         if (rc)
 457                                 GOTO(free_oig, rc);
 458
 459                         rc = oig_wait(oig);
 460 free_oig:
 461                         oig_release(oig);
 462                         GOTO(out, rc);
 463                 }
 464                 LL_CDEBUG_PAGE(D_PAGE, page, "write queued\n");
 465                 //llap_write_pending(inode, llap);
 466         } else {
 467                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 468                                      LPROC_LL_DIRTY_HITS);
 469         }
 470
 471         /* put the page in the page cache, from now on ll_removepage is
 472          * responsible for cleaning up the llap */
 473         set_page_dirty(page);
 474
 475 out:
 476         if (rc == 0) {
 477                 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
 478                 lov_increase_kms(exp, lsm, size);
 479                 if (size > inode->i_size)
 480                         inode->i_size = size;
 481                 SetPageUptodate(page);
 482         }
 483         RETURN(rc);
 484 }
 485
 486 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 487 {
 488         unsigned long ret;
 489         ENTRY;
 490
 491         spin_lock(&sbi->ll_lock);
 492         ret = min(sbi->ll_max_read_ahead_pages - sbi->ll_read_ahead_pages,
 493                   len);
 494         sbi->ll_read_ahead_pages += ret;
 495         spin_unlock(&sbi->ll_lock);
 496
 497         RETURN(ret);
 498 }
 499
 500 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 501 {
 502         spin_lock(&sbi->ll_lock);
 503         LASSERTF(sbi->ll_read_ahead_pages >= len, "r_a_p %lu len %lu\n",
 504                  sbi->ll_read_ahead_pages, len);
 505         sbi->ll_read_ahead_pages -= len;
 506         spin_unlock(&sbi->ll_lock);
 507 }
 508
 509 /* called for each page in a completed rpc.*/
 510 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 511 {
 512         struct ll_async_page *llap;
 513         struct page *page;
 514         ENTRY;
 515
 516         llap = llap_from_cookie(data);
 517         if (IS_ERR(llap)) {
 518                 EXIT;
 519                 return;
 520         }
 521
 522         page = llap->llap_page;
 523         LASSERT(PageLocked(page));
 524
 525         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 526
 527         if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
 528                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 529
 530         if (rc == 0)  {
 531                 if (cmd == OBD_BRW_READ) {
 532                         if (!llap->llap_defer_uptodate)
 533                                 SetPageUptodate(page);
 534                 } else {
 535                         llap->llap_write_queued = 0;
 536                 }
 537                 ClearPageError(page);
 538         } else {
 539                 if (cmd == OBD_BRW_READ)
 540                         llap->llap_defer_uptodate = 0;
 541                 SetPageError(page);
 542         }
 543
 544         unlock_page(page);
 545
 546         if (0 && cmd == OBD_BRW_WRITE) {
 547                 llap_write_complete(page->mapping->host, llap);
 548                 ll_try_done_writing(page->mapping->host);
 549         }
 550
 551         page_cache_release(page);
 552         EXIT;
 553 }
 554
 555 /* the kernel calls us here when a page is unhashed from the page cache.
 556  * the page will be locked and the kernel is holding a spinlock, so
 557  * we need to be careful.  we're just tearing down our book-keeping
 558  * here. */
 559 void ll_removepage(struct page *page)
 560 {
 561         struct inode *inode = page->mapping->host;
 562         struct obd_export *exp;
 563         struct ll_async_page *llap;
 564         struct ll_sb_info *sbi = ll_i2sbi(inode);
 565         int rc;
 566         ENTRY;
 567
 568         LASSERT(!in_interrupt());
 569
 570         /* sync pages or failed read pages can leave pages in the page
 571          * cache that don't have our data associated with them anymore */
 572         if (page->private == 0) {
 573                 EXIT;
 574                 return;
 575         }
 576
 577         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 578
 579         exp = ll_i2obdexp(inode);
 580         if (exp == NULL) {
 581                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 582                 EXIT;
 583                 return;
 584         }
 585
 586         llap = llap_from_page(page);
 587         if (IS_ERR(llap)) {
 588                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 589                        page->index, PTR_ERR(llap));
 590                 EXIT;
 591                 return;
 592         }
 593
 594         //llap_write_complete(inode, llap);
 595         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 596                                      llap->llap_cookie);
 597         if (rc != 0)
 598                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 599
 600         /* this unconditional free is only safe because the page lock
 601          * is providing exclusivity to memory pressure/truncate/writeback..*/
 602         page->private = 0;
 603
 604         spin_lock(&sbi->ll_lock);
 605         if (!list_empty(&llap->llap_proc_item))
 606                 list_del_init(&llap->llap_proc_item);
 607         sbi->ll_pglist_gen++;
 608         spin_unlock(&sbi->ll_lock);
 609         OBD_FREE(llap, sizeof(*llap));
 610         EXIT;
 611 }
 612
 613 static int ll_page_matches(struct page *page, int fd_flags)
 614 {
 615         struct lustre_handle match_lockh = {0};
 616         struct inode *inode = page->mapping->host;
 617         ldlm_policy_data_t page_extent;
 618         int flags, matches;
 619         ENTRY;
 620
 621         if (fd_flags & LL_FILE_GROUP_LOCKED)
 622                 RETURN(1);
 623
 624         page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
 625         page_extent.l_extent.end =
 626                 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
 627         flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK;
 628         matches = obd_match(ll_i2sbi(inode)->ll_osc_exp,
 629                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
 630                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
 631                             &match_lockh);
 632         RETURN(matches);
 633 }
 634
 635 static int ll_issue_page_read(struct obd_export *exp,
 636                               struct ll_async_page *llap,
 637                               struct obd_io_group *oig, int defer)
 638 {
 639         struct page *page = llap->llap_page;
 640         int rc;
 641
 642         page_cache_get(page);
 643         llap->llap_defer_uptodate = defer;
 644         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
 645                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
 646                                 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
 647                                               | ASYNC_URGENT);
 648         if (rc) {
 649                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
 650                 page_cache_release(page);
 651         }
 652         RETURN(rc);
 653 }
 654
 655 #define RAS_CDEBUG(ras) \
 656         CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n",        \
 657                ras->ras_last_readpage, ras->ras_consecutive,            \
 658                ras->ras_window_start, ras->ras_window_len,              \
 659                ras->ras_next_readahead);
 660
 661 static int ll_readahead(struct ll_readahead_state *ras,
 662                          struct obd_export *exp, struct address_space *mapping,
 663                          struct obd_io_group *oig, int flags)
 664 {
 665         unsigned long i, start = 0, end = 0, reserved;
 666         struct ll_async_page *llap;
 667         struct page *page;
 668         int rc, ret = 0;
 669         __u64 kms;
 670         ENTRY;
 671
 672         kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
 673         if (kms == 0)
 674                 RETURN(0);
 675
 676         spin_lock(&ras->ras_lock);
 677
 678         if (ras->ras_window_len) {
 679                 start = ras->ras_next_readahead;
 680                 end = ras->ras_window_start + ras->ras_window_len - 1;
 681                 end = min(end, (unsigned long)(kms >> PAGE_CACHE_SHIFT));
 682                 ras->ras_next_readahead = max(end, end + 1);
 683
 684                 RAS_CDEBUG(ras);
 685         }
 686
 687         spin_unlock(&ras->ras_lock);
 688
 689         if (end == 0)
 690                 RETURN(0);
 691
 692         reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
 693
 694         for (i = start; reserved > 0 && i <= end; i++) {
 695                 /* skip locked pages from previous readpage calls */
 696                 page = grab_cache_page_nowait(mapping, i);
 697                 if (page == NULL) {
 698                         CDEBUG(D_READA, "g_c_p_n failed\n");
 699                         continue;
 700                 }
 701
 702                 /* we do this first so that we can see the page in the /proc
 703                  * accounting */
 704                 llap = llap_from_page(page);
 705                 if (IS_ERR(llap) || llap->llap_defer_uptodate)
 706                         goto next_page;
 707
 708                 /* skip completed pages */
 709                 if (Page_Uptodate(page))
 710                         goto next_page;
 711
 712                 /* bail when we hit the end of the lock. */
 713                 if ((rc = ll_page_matches(page, flags)) <= 0) {
 714                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 715                                        "lock match failed: rc %d\n", rc);
 716                         i = end;
 717                         goto next_page;
 718                 }
 719
 720                 rc = ll_issue_page_read(exp, llap, oig, 1);
 721                 if (rc == 0) {
 722                         reserved--;
 723                         ret++;
 724                         LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
 725                                        "started read-ahead\n");
 726                 }
 727                 if (rc) {
 728         next_page:
 729                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 730                                        "skipping read-ahead\n");
 731
 732                         unlock_page(page);
 733                 }
 734                 page_cache_release(page);
 735         }
 736
 737         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
 738         if (reserved != 0)
 739                 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
 740         RETURN(ret);
 741 }
 742
 743 static void ras_set_start(struct ll_readahead_state *ras,
 744                                unsigned long index)
 745 {
 746         ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
 747         ras->ras_next_readahead = max(ras->ras_window_start,
 748                                       ras->ras_next_readahead);
 749 }
 750
 751 /* called with the ras_lock held or from places where it doesn't matter */
 752 static void ras_reset(struct ll_readahead_state *ras,
 753                       unsigned long index)
 754 {
 755         ras->ras_last_readpage = index;
 756         ras->ras_consecutive = 1;
 757         ras->ras_window_len = 0;
 758         ras_set_start(ras, index);
 759         ras->ras_next_readahead = ras->ras_window_start;
 760
 761         RAS_CDEBUG(ras);
 762 }
 763
 764 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
 765 {
 766         spin_lock_init(&ras->ras_lock);
 767         ras_reset(ras, 0);
 768 }
 769
 770 static void ras_update(struct ll_readahead_state *ras,
 771                        unsigned long index, unsigned long max)
 772 {
 773         ENTRY;
 774
 775         spin_lock(&ras->ras_lock);
 776
 777         if (index != ras->ras_last_readpage + 1) {
 778                 ras_reset(ras, index);
 779                 GOTO(out_unlock, 0);
 780         }
 781
 782         ras->ras_last_readpage = index;
 783         ras->ras_consecutive++;
 784         ras_set_start(ras, index);
 785
 786         if (ras->ras_consecutive == 2) {
 787                 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
 788                 GOTO(out_unlock, 0);
 789         }
 790
 791         /* we need to increase the window sometimes.  we'll arbitrarily
 792          * do it half-way through the pages in an rpc */
 793         if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
 794             (PTLRPC_MAX_BRW_PAGES >> 1)) {
 795                 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
 796                 ras->ras_window_len = min(ras->ras_window_len, max);
 797         }
 798
 799         EXIT;
 800 out_unlock:
 801         RAS_CDEBUG(ras);
 802         spin_unlock(&ras->ras_lock);
 803         return;
 804 }
 805
 806 /*
 807  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
 808  * read-ahead assumes it is valid to issue readpage all the way up to
 809  * i_size, but our dlm locks make that not the case.  We disable the
 810  * kernel's read-ahead and do our own by walking ahead in the page cache
 811  * checking for dlm lock coverage.  the main difference between 2.4 and
 812  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
 813  * so they look the same.
 814  */
 815 int ll_readpage(struct file *filp, struct page *page)
 816 {
 817         struct ll_file_data *fd = filp->private_data;
 818         struct inode *inode = page->mapping->host;
 819         struct obd_export *exp;
 820         struct ll_async_page *llap;
 821         struct obd_io_group *oig = NULL;
 822         int rc;
 823         ENTRY;
 824
 825         LASSERT(PageLocked(page));
 826         LASSERT(!PageUptodate(page));
 827         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
 828                inode->i_ino, inode->i_generation, inode,
 829                (((obd_off)page->index) << PAGE_SHIFT));
 830         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
 831
 832         rc = oig_init(&oig);
 833         if (rc < 0)
 834                 GOTO(out, rc);
 835
 836         exp = ll_i2obdexp(inode);
 837         if (exp == NULL)
 838                 GOTO(out, rc = -EINVAL);
 839
 840         llap = llap_from_page(page);
 841         if (IS_ERR(llap))
 842                 GOTO(out, rc = PTR_ERR(llap));
 843
 844         if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
 845                 ras_update(&fd->fd_ras, page->index,
 846                            ll_i2sbi(inode)->ll_max_read_ahead_pages);
 847
 848         if (llap->llap_defer_uptodate) {
 849                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
 850                                   fd->fd_flags);
 851                 if (rc > 0)
 852                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
 853                                              NULL, oig);
 854                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
 855                 SetPageUptodate(page);
 856                 unlock_page(page);
 857                 GOTO(out_oig, rc = 0);
 858         }
 859
 860         rc = ll_page_matches(page, fd->fd_flags);
 861         if (rc < 0) {
 862                 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
 863                 GOTO(out, rc);
 864         }
 865
 866         if (rc == 0) {
 867                 static unsigned long next_print;
 868                 CDEBUG(D_INODE, "ino %lu page %lu (%llu) didn't match a lock\n",
 869                        inode->i_ino, page->index,
 870                        (long long)page->index << PAGE_CACHE_SHIFT);
 871                 if (time_after(jiffies, next_print)) {
 872                         CWARN("ino %lu page %lu (%llu) not covered by "
 873                                "a lock (mmap?).  check debug logs.\n",
 874                                inode->i_ino, page->index,
 875                                (long long)page->index << PAGE_CACHE_SHIFT);
 876                         next_print = jiffies + 30 * HZ;
 877                 }
 878         }
 879
 880         rc = ll_issue_page_read(exp, llap, oig, 0);
 881         if (rc)
 882                 GOTO(out, rc);
 883
 884         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
 885         if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
 886                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
 887                              fd->fd_flags);
 888
 889         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
 890
 891 out:
 892         if (rc)
 893                 unlock_page(page);
 894 out_oig:
 895         if (oig != NULL)
 896                 oig_release(oig);
 897         RETURN(rc);
 898 }