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_fid(obdo_fid(oa), inode->i_ino, 0, inode->i_mode);
 318                 obdo_fid(oa)->mds = ll_i2info(inode)->lli_mds;
 319                 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
 320
 321                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 322         }
 323
 324         obdo_from_inode(oa, inode, valid_flags);
 325 }
 326
 327 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 328 {
 329         struct ll_async_page *llap;
 330         ENTRY;
 331
 332         llap = llap_from_cookie(data);
 333         if (IS_ERR(llap)) {
 334                 EXIT;
 335                 return;
 336         }
 337
 338         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 339         EXIT;
 340 }
 341
 342 static struct obd_async_page_ops ll_async_page_ops = {
 343         .ap_make_ready =        ll_ap_make_ready,
 344         .ap_refresh_count =     ll_ap_refresh_count,
 345         .ap_fill_obdo =         ll_ap_fill_obdo,
 346         .ap_completion =        ll_ap_completion,
 347 };
 348
 349 struct ll_async_page *llap_cast_private(struct page *page)
 350 {
 351         struct ll_async_page *llap = (struct ll_async_page *)page->private;
 352
 353         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 354                  "page %p private %lu gave magic %d which != %d\n",
 355                  page, page->private, llap->llap_magic, LLAP_MAGIC);
 356
 357         return llap;
 358 }
 359
 360 /* XXX have the exp be an argument? */
 361 struct ll_async_page *llap_from_page(struct page *page)
 362 {
 363         struct ll_async_page *llap;
 364         struct obd_export *exp;
 365         struct inode *inode = page->mapping->host;
 366         struct ll_sb_info *sbi = ll_i2sbi(inode);
 367         int rc;
 368         ENTRY;
 369
 370         llap = llap_cast_private(page);
 371         if (llap != NULL)
 372                 RETURN(llap);
 373
 374         exp = ll_i2obdexp(page->mapping->host);
 375         if (exp == NULL)
 376                 RETURN(ERR_PTR(-EINVAL));
 377
 378         OBD_ALLOC(llap, sizeof(*llap));
 379         if (llap == NULL)
 380                 RETURN(ERR_PTR(-ENOMEM));
 381         llap->llap_magic = LLAP_MAGIC;
 382         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 383                                  (obd_off)page->index << PAGE_SHIFT,
 384                                  &ll_async_page_ops, llap, &llap->llap_cookie);
 385         if (rc) {
 386                 OBD_FREE(llap, sizeof(*llap));
 387                 RETURN(ERR_PTR(rc));
 388         }
 389
 390         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 391                page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
 392         /* also zeroing the PRIVBITS low order bitflags */
 393         __set_page_ll_data(page, llap);
 394         llap->llap_page = page;
 395
 396         spin_lock(&sbi->ll_lock);
 397         sbi->ll_pglist_gen++;
 398         list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
 399         spin_unlock(&sbi->ll_lock);
 400
 401         RETURN(llap);
 402 }
 403
 404 static int queue_or_sync_write(struct obd_export *exp,
 405                                struct lov_stripe_md *lsm,
 406                                struct ll_async_page *llap,
 407                                unsigned to,
 408                                obd_flag async_flags)
 409 {
 410         struct obd_io_group *oig;
 411         int rc;
 412         ENTRY;
 413
 414         /* _make_ready only sees llap once we've unlocked the page */
 415         llap->llap_write_queued = 1;
 416         rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
 417                                 OBD_BRW_WRITE, 0, 0, 0, async_flags);
 418         if (rc == 0) {
 419                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
 420                 //llap_write_pending(inode, llap);
 421                 GOTO(out, 0);
 422         }
 423
 424         llap->llap_write_queued = 0;
 425
 426         rc = oig_init(&oig);
 427         if (rc)
 428                 GOTO(out, rc);
 429
 430         rc = obd_queue_group_io(exp, lsm, NULL, oig, llap->llap_cookie,
 431                                 OBD_BRW_WRITE, 0, to, 0, ASYNC_READY |
 432                                 ASYNC_URGENT | ASYNC_COUNT_STABLE |
 433                                 ASYNC_GROUP_SYNC);
 434         if (rc)
 435                 GOTO(free_oig, rc);
 436
 437         rc = obd_trigger_group_io(exp, lsm, NULL, oig);
 438         if (rc)
 439                 GOTO(free_oig, rc);
 440
 441         rc = oig_wait(oig);
 442
 443         if (!rc && async_flags & ASYNC_READY)
 444                 unlock_page(llap->llap_page);
 445
 446         LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n",
 447                        rc);
 448
 449 free_oig:
 450         oig_release(oig);
 451 out:
 452         RETURN(rc);
 453 }
 454
 455 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
 456                       obd_off size);
 457
 458 /* be careful not to return success without setting the page Uptodate or
 459  * the next pass through prepare_write will read in stale data from disk. */
 460 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 461                     unsigned to)
 462 {
 463         struct inode *inode = page->mapping->host;
 464         struct ll_inode_info *lli = ll_i2info(inode);
 465         struct lov_stripe_md *lsm = lli->lli_smd;
 466         struct obd_export *exp = NULL;
 467         struct ll_async_page *llap;
 468         loff_t size;
 469         int rc = 0;
 470         ENTRY;
 471
 472         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 473         LASSERT(inode == file->f_dentry->d_inode);
 474         LASSERT(PageLocked(page));
 475
 476         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 477                inode, page, from, to, page->index);
 478
 479         llap = llap_from_page(page);
 480         if (IS_ERR(llap))
 481                 RETURN(PTR_ERR(llap));
 482
 483         /* queue a write for some time in the future the first time we
 484          * dirty the page */
 485         if (!PageDirty(page)) {
 486                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 487                                      LPROC_LL_DIRTY_MISSES);
 488
 489                 exp = ll_i2obdexp(inode);
 490                 if (exp == NULL)
 491                         RETURN(-EINVAL);
 492
 493                 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd, llap,
 494                                          to, 0);
 495                 if (rc)
 496                         GOTO(out, rc);
 497         } else {
 498                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 499                                      LPROC_LL_DIRTY_HITS);
 500         }
 501
 502         /* put the page in the page cache, from now on ll_removepage is
 503          * responsible for cleaning up the llap.
 504          * don't dirty the page if it has been write out in q_o_s_w */
 505         if (llap->llap_write_queued)
 506                 set_page_dirty(page);
 507
 508 out:
 509         if (rc == 0) {
 510                 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
 511                 lov_increase_kms(exp, lsm, size);
 512                 if (size > inode->i_size)
 513                         inode->i_size = size;
 514                 SetPageUptodate(page);
 515         }
 516         RETURN(rc);
 517 }
 518
 519 int ll_writepage(struct page *page)
 520 {
 521         struct inode *inode = page->mapping->host;
 522         struct obd_export *exp;
 523         struct ll_async_page *llap;
 524         int rc = 0;
 525         ENTRY;
 526
 527         LASSERT(!PageDirty(page));
 528         LASSERT(PageLocked(page));
 529
 530         exp = ll_i2obdexp(inode);
 531         if (exp == NULL)
 532                 GOTO(out, rc = -EINVAL);
 533
 534         llap = llap_from_page(page);
 535         if (IS_ERR(llap))
 536                 GOTO(out, rc = PTR_ERR(llap));
 537
 538         page_cache_get(page);
 539         if (llap->llap_write_queued) {
 540                 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
 541                 rc = obd_set_async_flags(exp, ll_i2info(inode)->lli_smd, NULL,
 542                                          llap->llap_cookie,
 543                                          ASYNC_READY | ASYNC_URGENT);
 544         } else {
 545                 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd, llap,
 546                                          PAGE_SIZE, ASYNC_READY |
 547                                          ASYNC_URGENT);
 548         }
 549         if (rc)
 550                 page_cache_release(page);
 551 out:
 552         if (rc)
 553                 unlock_page(page);
 554         RETURN(rc);
 555 }
 556
 557 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 558 {
 559         struct ll_ra_info *ra = &sbi->ll_ra_info;
 560         unsigned long ret;
 561         ENTRY;
 562
 563         spin_lock(&sbi->ll_lock);
 564         ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
 565         ra->ra_cur_pages += ret;
 566         spin_unlock(&sbi->ll_lock);
 567
 568         RETURN(ret);
 569 }
 570
 571 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 572 {
 573         struct ll_ra_info *ra = &sbi->ll_ra_info;
 574         spin_lock(&sbi->ll_lock);
 575         LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
 576                  ra->ra_cur_pages, len);
 577         ra->ra_cur_pages -= len;
 578         spin_unlock(&sbi->ll_lock);
 579 }
 580
 581 /* called for each page in a completed rpc.*/
 582 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 583 {
 584         struct ll_async_page *llap;
 585         struct page *page;
 586         ENTRY;
 587
 588         llap = llap_from_cookie(data);
 589         if (IS_ERR(llap)) {
 590                 EXIT;
 591                 return;
 592         }
 593
 594         page = llap->llap_page;
 595         LASSERT(PageLocked(page));
 596
 597         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 598
 599         if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
 600                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 601
 602         if (rc == 0)  {
 603                 if (cmd == OBD_BRW_READ) {
 604                         if (!llap->llap_defer_uptodate)
 605                                 SetPageUptodate(page);
 606                 } else {
 607                         llap->llap_write_queued = 0;
 608                 }
 609                 ClearPageError(page);
 610         } else {
 611                 if (cmd == OBD_BRW_READ)
 612                         llap->llap_defer_uptodate = 0;
 613                 SetPageError(page);
 614         }
 615
 616         unlock_page(page);
 617
 618         if (0 && cmd == OBD_BRW_WRITE) {
 619                 llap_write_complete(page->mapping->host, llap);
 620                 ll_try_done_writing(page->mapping->host);
 621         }
 622
 623         if (PageWriteback(page)) {
 624                 end_page_writeback(page);
 625         }
 626         page_cache_release(page);
 627         EXIT;
 628 }
 629
 630 /* the kernel calls us here when a page is unhashed from the page cache.
 631  * the page will be locked and the kernel is holding a spinlock, so
 632  * we need to be careful.  we're just tearing down our book-keeping
 633  * here. */
 634 void ll_removepage(struct page *page)
 635 {
 636         struct inode *inode = page->mapping->host;
 637         struct obd_export *exp;
 638         struct ll_async_page *llap;
 639         struct ll_sb_info *sbi = ll_i2sbi(inode);
 640         int rc;
 641         ENTRY;
 642
 643         LASSERT(!in_interrupt());
 644
 645         /* sync pages or failed read pages can leave pages in the page
 646          * cache that don't have our data associated with them anymore */
 647         if (page->private == 0) {
 648                 EXIT;
 649                 return;
 650         }
 651
 652         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 653
 654         exp = ll_i2obdexp(inode);
 655         if (exp == NULL) {
 656                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 657                 EXIT;
 658                 return;
 659         }
 660
 661         llap = llap_from_page(page);
 662         if (IS_ERR(llap)) {
 663                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 664                        page->index, PTR_ERR(llap));
 665                 EXIT;
 666                 return;
 667         }
 668
 669         //llap_write_complete(inode, llap);
 670         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 671                                      llap->llap_cookie);
 672         if (rc != 0)
 673                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 674
 675         /* this unconditional free is only safe because the page lock
 676          * is providing exclusivity to memory pressure/truncate/writeback..*/
 677         __clear_page_ll_data(page);
 678
 679         spin_lock(&sbi->ll_lock);
 680         if (!list_empty(&llap->llap_proc_item))
 681                 list_del_init(&llap->llap_proc_item);
 682         sbi->ll_pglist_gen++;
 683         spin_unlock(&sbi->ll_lock);
 684         OBD_FREE(llap, sizeof(*llap));
 685         EXIT;
 686 }
 687
 688 static int ll_page_matches(struct page *page, int fd_flags)
 689 {
 690         struct lustre_handle match_lockh = {0};
 691         struct inode *inode = page->mapping->host;
 692         ldlm_policy_data_t page_extent;
 693         int flags, matches;
 694         ENTRY;
 695
 696         if (fd_flags & LL_FILE_GROUP_LOCKED)
 697                 RETURN(1);
 698
 699         page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
 700         page_extent.l_extent.end =
 701                 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
 702         flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK;
 703         matches = obd_match(ll_i2sbi(inode)->ll_osc_exp,
 704                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
 705                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
 706                             &match_lockh);
 707         RETURN(matches);
 708 }
 709
 710 static int ll_issue_page_read(struct obd_export *exp,
 711                               struct ll_async_page *llap,
 712                               struct obd_io_group *oig, int defer)
 713 {
 714         struct page *page = llap->llap_page;
 715         int rc;
 716
 717         page_cache_get(page);
 718         llap->llap_defer_uptodate = defer;
 719         llap->llap_ra_used = 0;
 720         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
 721                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
 722                                 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
 723                                               | ASYNC_URGENT);
 724         if (rc) {
 725                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
 726                 page_cache_release(page);
 727         }
 728         RETURN(rc);
 729 }
 730
 731 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
 732 {
 733         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
 734         ra->ra_stats[which]++;
 735 }
 736
 737 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
 738 {
 739         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
 740         struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
 741
 742         spin_lock(&sbi->ll_lock);
 743         ll_ra_stats_inc_unlocked(ra, which);
 744         spin_unlock(&sbi->ll_lock);
 745 }
 746
 747 void ll_ra_accounting(struct page *page, struct address_space *mapping)
 748 {
 749         struct ll_async_page *llap;
 750
 751         llap = llap_from_page(page);
 752         if (IS_ERR(llap))
 753                 return;
 754
 755         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
 756                 return;
 757
 758         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 759 }
 760
 761 #define RAS_CDEBUG(ras) \
 762         CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n",        \
 763                ras->ras_last_readpage, ras->ras_consecutive,            \
 764                ras->ras_window_start, ras->ras_window_len,              \
 765                ras->ras_next_readahead);
 766
 767 static int index_in_window(unsigned long index, unsigned long point,
 768                            unsigned long before, unsigned long after)
 769 {
 770         unsigned long start = point - before, end = point + after;
 771
 772         if (start > point)
 773                start = 0;
 774         if (end < point)
 775                end = ~0;
 776
 777         return start <= index && index <= end;
 778 }
 779
 780 static int ll_readahead(struct ll_readahead_state *ras,
 781                          struct obd_export *exp, struct address_space *mapping,
 782                          struct obd_io_group *oig, int flags)
 783 {
 784         unsigned long i, start = 0, end = 0, reserved;
 785         struct ll_async_page *llap;
 786         struct page *page;
 787         int rc, ret = 0, match_failed = 0;
 788         __u64 kms;
 789         ENTRY;
 790
 791         kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
 792         if (kms == 0) {
 793                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
 794                 RETURN(0);
 795         }
 796         spin_lock(&ras->ras_lock);
 797
 798         /* reserve a part of the read-ahead window that we'll be issuing */
 799         if (ras->ras_window_len) {
 800                 start = ras->ras_next_readahead;
 801                 end = ras->ras_window_start + ras->ras_window_len - 1;
 802                 end = min(end, (unsigned long)(kms >> PAGE_CACHE_SHIFT));
 803                 ras->ras_next_readahead = max(end, end + 1);
 804
 805                 RAS_CDEBUG(ras);
 806         }
 807
 808         spin_unlock(&ras->ras_lock);
 809
 810         if (end == 0) {
 811                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
 812                 RETURN(0);
 813         }
 814
 815         reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
 816         if (reserved < end - start + 1)
 817                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
 818
 819         for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
 820                 /* skip locked pages from previous readpage calls */
 821                 page = grab_cache_page_nowait(mapping, i);
 822                 if (page == NULL) {
 823                         CDEBUG(D_READA, "g_c_p_n failed\n");
 824                         continue;
 825                 }
 826
 827                 /* we do this first so that we can see the page in the /proc
 828                  * accounting */
 829                 llap = llap_from_page(page);
 830                 if (IS_ERR(llap) || llap->llap_defer_uptodate)
 831                         goto next_page;
 832
 833                 /* skip completed pages */
 834                 if (Page_Uptodate(page))
 835                         goto next_page;
 836
 837                 /* bail when we hit the end of the lock. */
 838                 if ((rc = ll_page_matches(page, flags)) <= 0) {
 839                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 840                                        "lock match failed: rc %d\n", rc);
 841                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
 842                         match_failed = 1;
 843                         goto next_page;
 844                 }
 845
 846                 rc = ll_issue_page_read(exp, llap, oig, 1);
 847                 if (rc == 0) {
 848                         reserved--;
 849                         ret++;
 850                         LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
 851                                        "started read-ahead\n");
 852                 }
 853                 if (rc) {
 854         next_page:
 855                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 856                                        "skipping read-ahead\n");
 857
 858                         unlock_page(page);
 859                 }
 860                 page_cache_release(page);
 861         }
 862
 863         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
 864         if (reserved != 0)
 865                 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
 866
 867         if (i == end + 1 && end == (kms >> PAGE_CACHE_SHIFT))
 868                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
 869
 870         /* if we didn't get to the end of the region we reserved from
 871          * the ras we need to go back and update the ras so that the
 872          * next read-ahead tries from where we left off.  we only do so
 873          * if the region we failed to issue read-ahead on is still ahead
 874          * of the app and behind the next index to start read-ahead from */
 875         if (i != end + 1) {
 876                 spin_lock(&ras->ras_lock);
 877                 if (i < ras->ras_next_readahead &&
 878                     index_in_window(i, ras->ras_window_start, 0,
 879                                     ras->ras_window_len)) {
 880                         ras->ras_next_readahead = i;
 881                         RAS_CDEBUG(ras);
 882                 }
 883                 spin_unlock(&ras->ras_lock);
 884         }
 885
 886         RETURN(ret);
 887 }
 888
 889 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
 890 {
 891         ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
 892 }
 893
 894 /* called with the ras_lock held or from places where it doesn't matter */
 895 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
 896 {
 897         ras->ras_last_readpage = index;
 898         ras->ras_consecutive = 1;
 899         ras->ras_window_len = 0;
 900         ras_set_start(ras, index);
 901         ras->ras_next_readahead = ras->ras_window_start;
 902
 903         RAS_CDEBUG(ras);
 904 }
 905
 906 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
 907 {
 908         spin_lock_init(&ras->ras_lock);
 909         ras_reset(ras, 0);
 910 }
 911
 912 static void ras_update(struct ll_sb_info *sbi, struct ll_readahead_state *ras,
 913                        unsigned long index, unsigned hit)
 914 {
 915         struct ll_ra_info *ra = &sbi->ll_ra_info;
 916         int zero = 0;
 917         ENTRY;
 918
 919         spin_lock(&sbi->ll_lock);
 920         spin_lock(&ras->ras_lock);
 921
 922         ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
 923
 924         /* reset the read-ahead window in two cases.  First when the app seeks
 925          * or reads to some other part of the file.  Secondly if we get a
 926          * read-ahead miss that we think we've previously issued.  This can
 927          * be a symptom of there being so many read-ahead pages that the VM is
 928          * reclaiming it before we get to it. */
 929         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
 930                 zero = 1;
 931                 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
 932         } else if (!hit && ras->ras_window_len &&
 933                    index < ras->ras_next_readahead &&
 934                    index_in_window(index, ras->ras_window_start, 0,
 935                                    ras->ras_window_len)) {
 936                 zero = 1;
 937                 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
 938         }
 939
 940         if (zero) {
 941                 ras_reset(ras, index);
 942                 GOTO(out_unlock, 0);
 943         }
 944
 945         ras->ras_last_readpage = index;
 946         ras->ras_consecutive++;
 947         ras_set_start(ras, index);
 948         ras->ras_next_readahead = max(ras->ras_window_start,
 949                                       ras->ras_next_readahead);
 950
 951         /* wait for a few pages to arrive before issuing readahead to avoid
 952          * the worst overutilization */
 953         if (ras->ras_consecutive == 3) {
 954                 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
 955                 GOTO(out_unlock, 0);
 956         }
 957
 958         /* we need to increase the window sometimes.  we'll arbitrarily
 959          * do it half-way through the pages in an rpc */
 960         if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
 961             (PTLRPC_MAX_BRW_PAGES >> 1)) {
 962                 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
 963                 ras->ras_window_len = min(ras->ras_window_len,
 964                                           ra->ra_max_pages);
 965
 966         }
 967
 968         EXIT;
 969 out_unlock:
 970         RAS_CDEBUG(ras);
 971         spin_unlock(&ras->ras_lock);
 972         spin_unlock(&sbi->ll_lock);
 973         return;
 974 }
 975
 976 /*
 977  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
 978  * read-ahead assumes it is valid to issue readpage all the way up to
 979  * i_size, but our dlm locks make that not the case.  We disable the
 980  * kernel's read-ahead and do our own by walking ahead in the page cache
 981  * checking for dlm lock coverage.  the main difference between 2.4 and
 982  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
 983  * so they look the same.
 984  */
 985 int ll_readpage(struct file *filp, struct page *page)
 986 {
 987         struct ll_file_data *fd = filp->private_data;
 988         struct inode *inode = page->mapping->host;
 989         struct obd_export *exp;
 990         struct ll_async_page *llap;
 991         struct obd_io_group *oig = NULL;
 992         int rc;
 993         ENTRY;
 994
 995         LASSERT(PageLocked(page));
 996         LASSERT(!PageUptodate(page));
 997         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
 998                inode->i_ino, inode->i_generation, inode,
 999                (((obd_off)page->index) << PAGE_SHIFT));
1000         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1001
1002         rc = oig_init(&oig);
1003         if (rc < 0)
1004                 GOTO(out, rc);
1005
1006         exp = ll_i2obdexp(inode);
1007         if (exp == NULL)
1008                 GOTO(out, rc = -EINVAL);
1009
1010         llap = llap_from_page(page);
1011         if (IS_ERR(llap))
1012                 GOTO(out, rc = PTR_ERR(llap));
1013
1014         if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1015                 ras_update(ll_i2sbi(inode), &fd->fd_ras, page->index,
1016                            llap->llap_defer_uptodate);
1017
1018         if (llap->llap_defer_uptodate) {
1019                 llap->llap_ra_used = 1;
1020                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1021                                   fd->fd_flags);
1022                 if (rc > 0)
1023                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1024                                              NULL, oig);
1025                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1026                 SetPageUptodate(page);
1027                 unlock_page(page);
1028                 GOTO(out_oig, rc = 0);
1029         }
1030
1031         rc = ll_page_matches(page, fd->fd_flags);
1032         if (rc < 0) {
1033                 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1034                 GOTO(out, rc);
1035         }
1036
1037         if (rc == 0) {
1038                 CWARN("ino %lu page %lu (%llu) not covered by "
1039                       "a lock (mmap?).  check debug logs.\n",
1040                       inode->i_ino, page->index,
1041                       (long long)page->index << PAGE_CACHE_SHIFT);
1042         }
1043
1044         rc = ll_issue_page_read(exp, llap, oig, 0);
1045         if (rc)
1046                 GOTO(out, rc);
1047
1048         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1049         if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1050                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1051                              fd->fd_flags);
1052
1053         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1054
1055 out:
1056         if (rc)
1057                 unlock_page(page);
1058 out_oig:
1059         if (oig != NULL)
1060                 oig_release(oig);
1061         RETURN(rc);
1062 }