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
   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 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
  38 #include <linux/buffer_head.h>
  39 #include <linux/mpage.h>
  40 #include <linux/writeback.h>
  41 #else
  42 #include <linux/iobuf.h>
  43 #endif
  44 #include <linux/stat.h>
  45 #include <asm/uaccess.h>
  46 #include <asm/segment.h>
  47 #include <linux/mm.h>
  48 #include <linux/pagemap.h>
  49 #include <linux/smp_lock.h>
  50
  51 #define DEBUG_SUBSYSTEM S_LLITE
  52
  53 #include <linux/lustre_mds.h>
  54 #include <linux/lustre_lite.h>
  55 #include "llite_internal.h"
  56 #include <linux/lustre_compat25.h>
  57
  58 /*
  59  * Remove page from dirty list
  60  */
  61 static void __set_page_clean(struct page *page)
  62 {
  63         struct address_space *mapping = page->mapping;
  64         struct inode *inode;
  65
  66         if (!mapping)
  67                 return;
  68
  69         PGCACHE_WRLOCK(mapping);
  70
  71         list_del(&page->list);
  72         list_add(&page->list, &mapping->clean_pages);
  73
  74         /* XXX doesn't inode_lock protect i_state ? */
  75         inode = mapping->host;
  76         if (list_empty(&mapping->dirty_pages)) {
  77                 CDEBUG(D_INODE, "inode clean\n");
  78                 inode->i_state &= ~I_DIRTY_PAGES;
  79         }
  80
  81         PGCACHE_WRUNLOCK(mapping);
  82         EXIT;
  83 }
  84
  85 void set_page_clean(struct page *page)
  86 {
  87         if (PageDirty(page)) {
  88                 ClearPageDirty(page);
  89                 __set_page_clean(page);
  90         }
  91 }
  92
  93 /* SYNCHRONOUS I/O to object storage for an inode */
  94 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
  95                   struct page *page, int flags)
  96 {
  97         struct ll_inode_info *lli = ll_i2info(inode);
  98         struct lov_stripe_md *lsm = lli->lli_smd;
  99         struct brw_page pg;
 100         int rc;
 101         ENTRY;
 102
 103         pg.pg = page;
 104         pg.off = ((obd_off)page->index) << PAGE_SHIFT;
 105
 106         if (cmd == OBD_BRW_WRITE && (pg.off + PAGE_SIZE > inode->i_size))
 107                 pg.count = inode->i_size % PAGE_SIZE;
 108         else
 109                 pg.count = PAGE_SIZE;
 110
 111         CDEBUG(D_PAGE, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
 112                cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, inode->i_ino,
 113                pg.off, pg.off);
 114         if (pg.count == 0) {
 115                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
 116                        LPU64"\n",
 117                        inode->i_ino, inode, inode->i_size, page->mapping->host,
 118                        page->mapping->host->i_size, page->index, pg.off);
 119         }
 120
 121         pg.flag = flags;
 122
 123         if (cmd == OBD_BRW_WRITE)
 124                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
 125                                     LPROC_LL_BRW_WRITE, pg.count);
 126         else
 127                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
 128                                     LPROC_LL_BRW_READ, pg.count);
 129         rc = obd_brw(cmd, ll_i2obdconn(inode), oa, lsm, 1, &pg, NULL);
 130         if (rc != 0 && rc != -EIO)
 131                 CERROR("error from obd_brw: rc = %d\n", rc);
 132
 133         RETURN(rc);
 134 }
 135
 136 /*
 137  * we were asked to read a single page but we're going to try and read a batch
 138  * of pages all at once.  this vaguely simulates 2.5's readpages.
 139  */
 140 static int ll_readpage(struct file *file, struct page *first_page)
 141 {
 142         struct inode *inode = first_page->mapping->host;
 143         struct ll_inode_info *lli = ll_i2info(inode);
 144         struct page *page = first_page;
 145         struct list_head *pos;
 146         struct brw_page *pgs;
 147         struct obdo *oa;
 148         unsigned long end_index, extent_end = 0;
 149         struct ptlrpc_request_set *set;
 150         int npgs = 0, rc = 0, max_pages;
 151         ENTRY;
 152
 153         LASSERT(PageLocked(page));
 154         LASSERT(!PageUptodate(page));
 155         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
 156                inode->i_ino, inode->i_generation, inode,
 157                (((obd_off)page->index) << PAGE_SHIFT));
 158         LASSERT(atomic_read(&file->f_dentry->d_inode->i_count) > 0);
 159
 160         if (inode->i_size <= ((obd_off)page->index) << PAGE_SHIFT) {
 161                 CERROR("reading beyond EOF\n");
 162                 memset(kmap(page), 0, PAGE_SIZE);
 163                 kunmap(page);
 164                 SetPageUptodate(page);
 165                 unlock_page(page);
 166                 RETURN(rc);
 167         }
 168
 169         /* try to read the file's preferred block size in a one-er */
 170         end_index = first_page->index +
 171                 (inode->i_blksize >> PAGE_CACHE_SHIFT);
 172         if (end_index > (inode->i_size >> PAGE_CACHE_SHIFT))
 173                 end_index = inode->i_size >> PAGE_CACHE_SHIFT;
 174
 175         max_pages = ((end_index - first_page->index) << PAGE_CACHE_SHIFT) >>
 176                 PAGE_SHIFT;
 177         OBD_ALLOC_GFP(pgs, max_pages * sizeof(*pgs), GFP_USER);
 178         if (pgs == NULL)
 179                 RETURN(-ENOMEM);
 180
 181         /*
 182          * find how far we're allowed to read under the extent ll_file_read
 183          * is passing us..
 184          */
 185         spin_lock(&lli->lli_read_extent_lock);
 186         list_for_each(pos, &lli->lli_read_extents) {
 187                 struct ll_read_extent *rextent;
 188                 rextent = list_entry(pos, struct ll_read_extent, re_lli_item);
 189                 if (rextent->re_task != current)
 190                         continue;
 191
 192                 if (rextent->re_extent.end + PAGE_SIZE < rextent->re_extent.end)
 193                         /* extent wrapping */
 194                         extent_end = ~0;
 195                 else {
 196                         extent_end = (rextent->re_extent.end + PAGE_SIZE)
 197                                                         << PAGE_CACHE_SHIFT;
 198                         /* 32bit indexes, 64bit extents.. */
 199                         if (((u64)extent_end >> PAGE_CACHE_SHIFT) <
 200                                         rextent->re_extent.end)
 201                                 extent_end = ~0;
 202                 }
 203                 break;
 204         }
 205         spin_unlock(&lli->lli_read_extent_lock);
 206
 207         if (extent_end == 0) {
 208                 static long next_print;
 209                 if (time_after(jiffies, next_print)) {
 210                         next_print = jiffies + 30 * HZ;
 211                         CDEBUG(D_INODE, "mmap readpage - check locks\n");
 212                 }
 213                 end_index = page->index + 1;
 214         } else if (extent_end < end_index)
 215                 end_index = extent_end;
 216
 217         CDEBUG(D_INFO, "max_pages: %d, extent_end: %lu, end_index: %lu, "
 218                "i_size: "LPU64"\n",
 219                max_pages, extent_end, end_index, inode->i_size);
 220
 221         /* to balance the find_get_page ref the other pages get that is
 222          * decrefed on teardown.. */
 223         page_cache_get(page);
 224         do {
 225                 unsigned long index ;
 226
 227                 pgs[npgs].pg = page;
 228                 pgs[npgs].off = ((obd_off)page->index) << PAGE_CACHE_SHIFT;
 229                 pgs[npgs].flag = 0;
 230                 pgs[npgs].count = PAGE_SIZE;
 231                 /* XXX Workaround for BA OSTs returning short reads at EOF.
 232                  * The linux OST will return the full page, zero-filled at the
 233                  * end, which will just overwrite the data we set here.  Bug
 234                  * 593 relates to fixing this properly.
 235                  */
 236                 if (inode->i_size < pgs[npgs].off + PAGE_SIZE) {
 237                         int count = inode->i_size - pgs[npgs].off;
 238                         void *addr = kmap(page);
 239                         pgs[npgs].count = count;
 240                         //POISON(addr, 0x7c, count);
 241                         memset(addr + count, 0, PAGE_SIZE - count);
 242                         kunmap(page);
 243                 }
 244
 245                 npgs++;
 246                 if (npgs == max_pages)
 247                         break;
 248
 249                 /*
 250                  * find pages ahead of us that we can read in.
 251                  * grab_cache_page waits on pages that are locked so
 252                  * we first try find_get_page, which doesn't.  this stops
 253                  * the worst case behaviour of racing threads waiting on
 254                  * each other, but doesn't remove it entirely.
 255                  */
 256                 for (index = page->index + 1, page = NULL;
 257                      page == NULL && index < end_index; index++) {
 258
 259                         /* see if the page already exists and needs updating */
 260                         page = find_get_page(inode->i_mapping, index);
 261                         if (page) {
 262                                 if (Page_Uptodate(page) || TryLockPage(page))
 263                                         goto out_release;
 264                                 if (!page->mapping || Page_Uptodate(page))
 265                                         goto out_unlock;
 266                         } else {
 267                                 /* ok, we have to create it.. */
 268                                 page = grab_cache_page(inode->i_mapping, index);
 269                                 if (page == NULL)
 270                                         continue;
 271                                 if (Page_Uptodate(page))
 272                                         goto out_unlock;
 273                         }
 274
 275                         break;
 276
 277                 out_unlock:
 278                         unlock_page(page);
 279                 out_release:
 280                         page_cache_release(page);
 281                         page = NULL;
 282                 }
 283
 284         } while (page);
 285
 286         if ((oa = obdo_alloc()) == NULL) {
 287                 CERROR("ENOMEM allocing obdo\n");
 288                 rc = -ENOMEM;
 289         } else if ((set = ptlrpc_prep_set()) == NULL) {
 290                 CERROR("ENOMEM allocing request set\n");
 291                 obdo_free(oa);
 292                 rc = -ENOMEM;
 293         } else {
 294                 struct ll_file_data *fd = file->private_data;
 295
 296                 oa->o_id = lli->lli_smd->lsm_object_id;
 297                 memcpy(obdo_handle(oa), &fd->fd_ost_och.och_fh,
 298                        sizeof(fd->fd_ost_och.och_fh));
 299                 oa->o_valid = OBD_MD_FLID | OBD_MD_FLHANDLE;
 300                 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME);
 301
 302                 rc = obd_brw_async(OBD_BRW_READ, ll_i2obdconn(inode), oa,
 303                                    ll_i2info(inode)->lli_smd, npgs, pgs,
 304                                    set, NULL);
 305                 if (rc == 0)
 306                         rc = ptlrpc_set_wait(set);
 307                 ptlrpc_set_destroy(set);
 308                 if (rc == 0)
 309                         obdo_refresh_inode(inode, oa, oa->o_valid);
 310                 if (rc && rc != -EIO)
 311                         CERROR("error from obd_brw_async: rc = %d\n", rc);
 312                 obdo_free(oa);
 313         }
 314
 315         while (npgs-- > 0) {
 316                 page = pgs[npgs].pg;
 317
 318                 if (rc == 0)
 319                         SetPageUptodate(page);
 320                 unlock_page(page);
 321                 page_cache_release(page);
 322         }
 323
 324         OBD_FREE(pgs, max_pages * sizeof(*pgs));
 325         RETURN(rc);
 326 } /* ll_readpage */
 327
 328 /* this isn't where truncate starts.   roughly:
 329  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate
 330  * we grab the lock back in setattr_raw to avoid races. */
 331 void ll_truncate(struct inode *inode)
 332 {
 333         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
 334         struct obdo oa;
 335         int err;
 336         ENTRY;
 337         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
 338                inode->i_generation, inode);
 339
 340         /* object not yet allocated */
 341         if (!lsm) {
 342                 CERROR("truncate on inode %lu with no objects\n", inode->i_ino);
 343                 EXIT;
 344                 return;
 345         }
 346
 347         /* vmtruncate will just throw away our dirty pages, make sure
 348          * we don't think they're still dirty, being careful to round
 349          * i_size to the first whole page that was tossed */
 350         err = ll_clear_dirty_pages(ll_i2obdconn(inode), lsm,
 351                         (inode->i_size + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT,
 352                         ~0);
 353
 354         oa.o_id = lsm->lsm_object_id;
 355         oa.o_valid = OBD_MD_FLID;
 356         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE|OBD_MD_FLMODE|OBD_MD_FLATIME|
 357                                     OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 358
 359         CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after %Lu)\n",
 360                oa.o_id, inode->i_size);
 361
 362         /* truncate == punch from new size to absolute end of file */
 363         err = obd_punch(ll_i2obdconn(inode), &oa, lsm, inode->i_size,
 364                         OBD_OBJECT_EOF, NULL);
 365         if (err)
 366                 CERROR("obd_truncate fails (%d) ino %lu\n", err, inode->i_ino);
 367         else
 368                 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 369                                           OBD_MD_FLATIME | OBD_MD_FLMTIME |
 370                                           OBD_MD_FLCTIME);
 371
 372         EXIT;
 373         return;
 374 } /* ll_truncate */
 375
 376 //#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 377
 378 static int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 379                             unsigned to)
 380 {
 381         struct inode *inode = page->mapping->host;
 382         struct ll_inode_info *lli = ll_i2info(inode);
 383         struct ll_file_data *fd = file->private_data;
 384         struct lov_stripe_md *lsm = lli->lli_smd;
 385         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 386         struct brw_page pg;
 387         struct obdo oa;
 388         int rc = 0;
 389         ENTRY;
 390
 391         if (!PageLocked(page))
 392                 LBUG();
 393
 394         if (PageUptodate(page))
 395                 RETURN(0);
 396
 397         //POISON(addr + from, 0xca, to - from);
 398
 399         /* Check to see if we should return -EIO right away */
 400         pg.pg = page;
 401         pg.off = offset;
 402         pg.count = PAGE_SIZE;
 403         pg.flag = 0;
 404         rc = obd_brw(OBD_BRW_CHECK, ll_i2obdconn(inode), NULL, lsm, 1,&pg,NULL);
 405         if (rc)
 406                 RETURN(rc);
 407
 408         /* We're completely overwriting an existing page, so _don't_ set it up
 409          * to date until commit_write */
 410         if (from == 0 && to == PAGE_SIZE)
 411                 RETURN(0);
 412
 413         /* If are writing to a new page, no need to read old data.
 414          * the extent locking and getattr procedures in ll_file_write have
 415          * guaranteed that i_size is stable enough for our zeroing needs */
 416         if (inode->i_size <= offset) {
 417                 memset(kmap(page), 0, PAGE_SIZE);
 418                 kunmap(page);
 419                 GOTO(prepare_done, rc = 0);
 420         }
 421
 422         oa.o_id = lsm->lsm_object_id;
 423         oa.o_mode = inode->i_mode;
 424         memcpy(obdo_handle(&oa), &fd->fd_ost_och.och_fh,
 425                sizeof(fd->fd_ost_och.och_fh));
 426         oa.o_valid = OBD_MD_FLID |OBD_MD_FLMODE |OBD_MD_FLTYPE |OBD_MD_FLHANDLE;
 427
 428         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 429         if (rc == 0)
 430                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 431
 432         EXIT;
 433  prepare_done:
 434         if (rc == 0)
 435                 SetPageUptodate(page);
 436
 437         return rc;
 438 }
 439
 440 /*
 441  * background file writeback.  This is called regularly from kupdated to write
 442  * dirty data, from kswapd when memory is low, and from filemap_fdatasync when
 443  * super blocks or inodes are synced..
 444  *
 445  * obd_brw errors down in _batch_writepage are ignored, so pages are always
 446  * unlocked.  Also, there is nobody to return an error code to from here - the
 447  * application may not even be running anymore.
 448  *
 449  * this should be async so that things like kswapd can have a chance to
 450  * free some more pages that our allocating writeback may need, but it isn't
 451  * yet.
 452  */
 453 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 454 static unsigned long ll_local_cache_dirty_pages;
 455 static unsigned long ll_max_dirty_pages = 20 * 1024 * 1024 / PAGE_SIZE;
 456
 457 static spinlock_t ll_local_cache_page_count_lock = SPIN_LOCK_UNLOCKED;
 458
 459 int ll_rd_dirty_pages(char *page, char **start, off_t off, int count, int *eof,
 460                       void *data)
 461 {
 462         unsigned long dirty_count;
 463         spin_lock(&ll_local_cache_page_count_lock);
 464         dirty_count = ll_local_cache_dirty_pages;
 465         spin_unlock(&ll_local_cache_page_count_lock);
 466         return snprintf(page, count, "%lu\n", dirty_count);
 467 }
 468
 469 int ll_rd_max_dirty_pages(char *page, char **start, off_t off, int count,
 470                           int *eof, void *data)
 471 {
 472         unsigned long max_dirty;
 473         spin_lock(&ll_local_cache_page_count_lock);
 474         max_dirty = ll_max_dirty_pages;
 475         spin_unlock(&ll_local_cache_page_count_lock);
 476         return snprintf(page, count, "%lu\n", max_dirty);
 477 }
 478
 479 int ll_wr_max_dirty_pages(struct file *file, const char *buffer,
 480                           unsigned long count, void *data)
 481 {
 482         unsigned long max_dirty;
 483         signed long max_dirty_signed;
 484         char kernbuf[20], *end;
 485
 486         if (count > (sizeof(kernbuf) - 1))
 487                 return -EINVAL;
 488
 489         if (copy_from_user(kernbuf, buffer, count))
 490                 return -EFAULT;
 491
 492         kernbuf[count] = '\0';
 493
 494         max_dirty_signed = simple_strtol(kernbuf, &end, 0);
 495         if (kernbuf == end)
 496                 return -EINVAL;
 497         max_dirty = (unsigned long)max_dirty_signed;
 498
 499 #if 0
 500         if (max_dirty < ll_local_cache_dirty_pages)
 501                 flush_to_new_max_dirty();
 502 #endif
 503
 504         spin_lock(&ll_local_cache_page_count_lock);
 505         CDEBUG(D_CACHE, "changing max_dirty from %lu to %lu\n",
 506                ll_max_dirty_pages, max_dirty);
 507         ll_max_dirty_pages = max_dirty;
 508         spin_unlock(&ll_local_cache_page_count_lock);
 509         return count;
 510 }
 511
 512 static int ll_local_cache_full(void)
 513 {
 514         int full = 0;
 515         spin_lock(&ll_local_cache_page_count_lock);
 516         if (ll_max_dirty_pages &&
 517             ll_local_cache_dirty_pages >= ll_max_dirty_pages) {
 518                 full = 1;
 519         }
 520         spin_unlock(&ll_local_cache_page_count_lock);
 521         /* XXX instrument? */
 522         /* XXX trigger async writeback when full, or 75% of full? */
 523         return full;
 524 }
 525
 526 static void ll_local_cache_flushed_pages(unsigned long pgcount)
 527 {
 528         unsigned long dirty_count;
 529         spin_lock(&ll_local_cache_page_count_lock);
 530         dirty_count = ll_local_cache_dirty_pages;
 531         ll_local_cache_dirty_pages -= pgcount;
 532         CDEBUG(D_CACHE, "dirty pages: %lu->%lu)\n",
 533                dirty_count, ll_local_cache_dirty_pages);
 534         spin_unlock(&ll_local_cache_page_count_lock);
 535         LASSERT(dirty_count >= pgcount);
 536 }
 537
 538 static void ll_local_cache_dirtied_pages(unsigned long pgcount)
 539 {
 540         unsigned long dirty_count;
 541         spin_lock(&ll_local_cache_page_count_lock);
 542         dirty_count = ll_local_cache_dirty_pages;
 543         ll_local_cache_dirty_pages += pgcount;
 544         CDEBUG(D_CACHE, "dirty pages: %lu->%lu\n",
 545                dirty_count, ll_local_cache_dirty_pages);
 546         spin_unlock(&ll_local_cache_page_count_lock);
 547         /* XXX track maximum cached, report to lprocfs */
 548 }
 549
 550 int ll_clear_dirty_pages(struct lustre_handle *conn, struct lov_stripe_md *lsm,
 551                          unsigned long start, unsigned long end)
 552 {
 553         unsigned long cleared;
 554         int rc;
 555
 556         ENTRY;
 557         rc = obd_clear_dirty_pages(conn, lsm, start, end, &cleared);
 558         if (!rc)
 559                 ll_local_cache_flushed_pages(cleared);
 560         RETURN(rc);
 561 }
 562
 563 int ll_mark_dirty_page(struct lustre_handle *conn, struct lov_stripe_md *lsm,
 564                        unsigned long index)
 565 {
 566         int rc;
 567
 568         ENTRY;
 569         if (ll_local_cache_full())
 570                 RETURN(-EDQUOT);
 571
 572         rc = obd_mark_page_dirty(conn, lsm, index);
 573         if (!rc)
 574                 ll_local_cache_dirtied_pages(1);
 575         RETURN(rc);
 576 }
 577
 578 static int ll_writepage(struct page *page)
 579 {
 580         struct inode *inode = page->mapping->host;
 581         struct obdo oa;
 582         ENTRY;
 583
 584         CDEBUG(D_CACHE, "page %p [lau %d] inode %p\n", page,
 585                PageLaunder(page), inode);
 586         LASSERT(PageLocked(page));
 587
 588         oa.o_id = ll_i2info(inode)->lli_smd->lsm_object_id;
 589         oa.o_valid = OBD_MD_FLID;
 590         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 591                                     OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 592
 593         RETURN(ll_batch_writepage(inode, &oa, page));
 594 }
 595
 596 /*
 597  * we really don't want to start writeback here, we want to give callers some
 598  * time to further dirty the pages before we write them out.
 599  */
 600 static int ll_commit_write(struct file *file, struct page *page,
 601                            unsigned from, unsigned to)
 602 {
 603         struct inode *inode = page->mapping->host;
 604         loff_t size;
 605         int rc = 0;
 606         ENTRY;
 607
 608         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 609         LASSERT(inode == file->f_dentry->d_inode);
 610         LASSERT(PageLocked(page));
 611
 612         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 613                inode, page, from, to, page->index);
 614         if (!PageDirty(page)) {
 615                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 616                                      LPROC_LL_DIRTY_MISSES);
 617                 rc = ll_mark_dirty_page(ll_i2obdconn(inode),
 618                                         ll_i2info(inode)->lli_smd,
 619                                         page->index);
 620                 if (rc < 0 && rc != -EDQUOT)
 621                         RETURN(rc); /* XXX lproc counter here? */
 622         } else {
 623                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 624                                      LPROC_LL_DIRTY_HITS);
 625         }
 626
 627         size = (((obd_off)page->index) << PAGE_SHIFT) + to;
 628         if (size > inode->i_size)
 629                 inode->i_size = size;
 630
 631         SetPageUptodate(page);
 632         set_page_dirty(page);
 633
 634         /* This means that we've hit either the local cache limit or the limit
 635          * of the OST's grant. */
 636         if (rc == -EDQUOT) {
 637                 struct ll_file_data *fd = file->private_data;
 638                 struct obdo oa;
 639                 int rc;
 640
 641                 oa.o_id = ll_i2info(inode)->lli_smd->lsm_object_id;
 642                 memcpy(obdo_handle(&oa), &fd->fd_ost_och.och_fh,
 643                        sizeof(fd->fd_ost_och.och_fh));
 644                 oa.o_valid = OBD_MD_FLID | OBD_MD_FLHANDLE;
 645                 obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 646                                             OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 647
 648                 rc = ll_batch_writepage(inode, &oa, page);
 649                 lock_page(page); /* caller expects to unlock */
 650                 RETURN(rc);
 651         }
 652
 653         RETURN(0);
 654 } /* ll_commit_write */
 655 #else
 656 static int ll_writepage(struct page *page,
 657                         struct writeback_control *wbc)
 658 {
 659
 660         return 0;
 661 }
 662 static int ll_commit_write(struct file *file, struct page *page,
 663                            unsigned from, unsigned to)
 664 {
 665         return 0;
 666 }
 667 #endif
 668
 669 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 670 static int ll_direct_IO(int rw, struct inode *inode, struct kiobuf *iobuf,
 671                         unsigned long blocknr, int blocksize)
 672 {
 673         struct ll_inode_info *lli = ll_i2info(inode);
 674         struct lov_stripe_md *lsm = lli->lli_smd;
 675         struct brw_page *pga;
 676         struct ptlrpc_request_set *set;
 677         struct obdo oa;
 678         int length, i, flags, rc = 0;
 679         loff_t offset;
 680         ENTRY;
 681
 682         if (!lsm || !lsm->lsm_object_id)
 683                 RETURN(-EBADF);
 684
 685         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 */
 686         if ((iobuf->offset & (PAGE_SIZE - 1)) ||
 687             (iobuf->length & (PAGE_SIZE - 1)))
 688                 RETURN(-EINVAL);
 689
 690         set = ptlrpc_prep_set();
 691         if (set == NULL)
 692                 RETURN(-ENOMEM);
 693
 694         OBD_ALLOC(pga, sizeof(*pga) * iobuf->nr_pages);
 695         if (!pga) {
 696                 ptlrpc_set_destroy(set);
 697                 RETURN(-ENOMEM);
 698         }
 699
 700         flags = (rw == WRITE ? OBD_BRW_CREATE : 0) /* | OBD_BRW_DIRECTIO */;
 701         offset = ((obd_off)blocknr << inode->i_blkbits);
 702         length = iobuf->length;
 703
 704         for (i = 0, length = iobuf->length; length > 0;
 705              length -= pga[i].count, offset += pga[i].count, i++) { /*i last!*/
 706                 pga[i].pg = iobuf->maplist[i];
 707                 pga[i].off = offset;
 708                 /* To the end of the page, or the length, whatever is less */
 709                 pga[i].count = min_t(int, PAGE_SIZE - (offset & ~PAGE_MASK),
 710                                      length);
 711                 pga[i].flag = flags;
 712                 if (rw == READ) {
 713                         //POISON(kmap(iobuf->maplist[i]), 0xc5, PAGE_SIZE);
 714                         //kunmap(iobuf->maplist[i]);
 715                 }
 716         }
 717
 718         oa.o_id = lsm->lsm_object_id;
 719         oa.o_valid = OBD_MD_FLID;
 720         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 721                                     OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 722
 723         if (rw == WRITE)
 724                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
 725                                     LPROC_LL_DIRECT_WRITE, iobuf->length);
 726         else
 727                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
 728                                     LPROC_LL_DIRECT_READ, iobuf->length);
 729         rc = obd_brw_async(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 730                            ll_i2obdconn(inode), &oa, lsm, iobuf->nr_pages, pga,
 731                            set, NULL);
 732         if (rc) {
 733                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 734                        "error from obd_brw_async: rc = %d\n", rc);
 735         } else {
 736                 rc = ptlrpc_set_wait(set);
 737                 if (rc)
 738                         CERROR("error from callback: rc = %d\n", rc);
 739         }
 740         ptlrpc_set_destroy(set);
 741         if (rc == 0)
 742                 rc = iobuf->length;
 743
 744         OBD_FREE(pga, sizeof(*pga) * iobuf->nr_pages);
 745         RETURN(rc);
 746 }
 747 #endif
 748
 749 //#endif
 750
 751 struct address_space_operations ll_aops = {
 752         readpage: ll_readpage,
 753 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
 754         direct_IO: ll_direct_IO,
 755 #endif
 756         writepage: ll_writepage,
 757         sync_page: block_sync_page,
 758         prepare_write: ll_prepare_write,
 759         commit_write: ll_commit_write,
 760         bmap: NULL
 761 //#endif
 762 };