lustre/osd-ldiskfs/osd_io.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/osd/osd_io.c
  37  *
  38  * body operations
  39  *
  40  * Author: Nikita Danilov <nikita@clusterfs.com>
  41  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  42  *
  43  */
  44
  45 /* LUSTRE_VERSION_CODE */
  46 #include <lustre_ver.h>
  47 /* prerequisite for linux/xattr.h */
  48 #include <linux/types.h>
  49 /* prerequisite for linux/xattr.h */
  50 #include <linux/fs.h>
  51
  52 /* ext_depth() */
  53 #include <ldiskfs/ldiskfs.h>
  54 #include <ldiskfs/ldiskfs_jbd2.h>
  55 #include <ldiskfs/ldiskfs_extents.h>
  56
  57 /*
  58  * struct OBD_{ALLOC,FREE}*()
  59  * OBD_FAIL_CHECK
  60  */
  61 #include <obd_support.h>
  62
  63 #include "osd_internal.h"
  64
  65 #ifndef HAVE_PAGE_CONSTANT
  66 #define mapping_cap_page_constant_write(mapping) 0
  67 #define SetPageConstant(page) do {} while (0)
  68 #define ClearPageConstant(page) do {} while (0)
  69 #endif
  70
  71 #ifndef HAS_GENERIC_ERROR_REMOVE_PAGE
  72 int generic_error_remove_page(struct address_space *mapping, struct page *page)
  73 {
  74         if (mapping == NULL)
  75                 return -EINVAL;
  76
  77         if (mapping != page->mapping)
  78                 return -EIO;
  79         /*
  80          * Only punch for normal data pages for now.
  81          * Handling other types like directories would need more auditing.
  82          */
  83         if (!S_ISREG(mapping->host->i_mode))
  84                 return -EIO;
  85
  86         if (page_mapped(page)) {
  87                 unmap_mapping_range(mapping,
  88                                     (loff_t)page->index << PAGE_CACHE_SHIFT,
  89                                     PAGE_CACHE_SIZE, 0);
  90         }
  91         truncate_complete_page(mapping, page);
  92         return 0;
  93 }
  94 #endif
  95
  96 static void osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,int rw)
  97 {
  98         cfs_waitq_init(&iobuf->dr_wait);
  99         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 100         iobuf->dr_max_pages = PTLRPC_MAX_BRW_PAGES;
 101         iobuf->dr_npages = 0;
 102         iobuf->dr_error = 0;
 103         iobuf->dr_dev = d;
 104         iobuf->dr_frags = 0;
 105         iobuf->dr_elapsed = 0;
 106         /* must be counted before, so assert */
 107         LASSERT(iobuf->dr_elapsed_valid == 0);
 108         iobuf->dr_rw = rw;
 109 }
 110
 111 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
 112 {
 113         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 114         iobuf->dr_pages[iobuf->dr_npages++] = page;
 115 }
 116
 117 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 118 {
 119         int rw = iobuf->dr_rw;
 120
 121         if (iobuf->dr_elapsed_valid) {
 122                 iobuf->dr_elapsed_valid = 0;
 123                 LASSERT(iobuf->dr_dev == d);
 124                 LASSERT(iobuf->dr_frags > 0);
 125                 lprocfs_oh_tally(&d->od_brw_stats.
 126                                  hist[BRW_R_DIO_FRAGS+rw],
 127                                  iobuf->dr_frags);
 128                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 129                                       iobuf->dr_elapsed);
 130         }
 131 }
 132
 133 #ifdef HAVE_BIO_ENDIO_2ARG
 134 #define DIO_RETURN(a)
 135 static void dio_complete_routine(struct bio *bio, int error)
 136 #else
 137 #define DIO_RETURN(a)   return(a)
 138 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 139 #endif
 140 {
 141         struct osd_iobuf *iobuf = bio->bi_private;
 142         struct bio_vec *bvl;
 143         int i;
 144
 145         /* CAVEAT EMPTOR: possibly in IRQ context
 146          * DO NOT record procfs stats here!!! */
 147
 148         if (unlikely(iobuf == NULL)) {
 149                 CERROR("***** bio->bi_private is NULL!  This should never "
 150                        "happen.  Normally, I would crash here, but instead I "
 151                        "will dump the bio contents to the console.  Please "
 152                        "report this to <http://jira.whamcloud.com/> , along "
 153                        "with any interesting messages leading up to this point "
 154                        "(like SCSI errors, perhaps).  Because bi_private is "
 155                        "NULL, I can't wake up the thread that initiated this "
 156                        "IO - you will probably have to reboot this node.\n");
 157                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 158                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 159                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 160                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 161                        bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
 162                        bio->bi_private);
 163                 DIO_RETURN(0);
 164         }
 165
 166         /* the check is outside of the cycle for performance reason -bzzz */
 167         if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
 168                 bio_for_each_segment(bvl, bio, i) {
 169                         if (likely(error == 0))
 170                                 SetPageUptodate(bvl->bv_page);
 171                         LASSERT(PageLocked(bvl->bv_page));
 172                         ClearPageConstant(bvl->bv_page);
 173                 }
 174                 cfs_atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 175         } else {
 176                 struct page *p = iobuf->dr_pages[0];
 177                 if (p->mapping) {
 178                         if (mapping_cap_page_constant_write(p->mapping)) {
 179                                 bio_for_each_segment(bvl, bio, i) {
 180                                         ClearPageConstant(bvl->bv_page);
 181                                 }
 182                         }
 183                 }
 184                 cfs_atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 185         }
 186
 187         /* any real error is good enough -bzzz */
 188         if (error != 0 && iobuf->dr_error == 0)
 189                 iobuf->dr_error = error;
 190
 191         if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs)) {
 192                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
 193                 iobuf->dr_elapsed_valid = 1;
 194                 cfs_waitq_signal(&iobuf->dr_wait);
 195         }
 196
 197         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 198          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 199          * mempool when serious on-disk fragmentation was encountered,
 200          * deadlocking the OST.  The bios are now released as soon as complete
 201          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 202         bio_put(bio);
 203         DIO_RETURN(0);
 204 }
 205
 206 static void record_start_io(struct osd_iobuf *iobuf, int size)
 207 {
 208         struct osd_device    *osd = iobuf->dr_dev;
 209         struct obd_histogram *h = osd->od_brw_stats.hist;
 210
 211         iobuf->dr_frags++;
 212         cfs_atomic_inc(&iobuf->dr_numreqs);
 213
 214         if (iobuf->dr_rw == 0) {
 215                 cfs_atomic_inc(&osd->od_r_in_flight);
 216                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 217                                  cfs_atomic_read(&osd->od_r_in_flight));
 218                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 219         } else if (iobuf->dr_rw == 1) {
 220                 cfs_atomic_inc(&osd->od_w_in_flight);
 221                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 222                                  cfs_atomic_read(&osd->od_w_in_flight));
 223                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 224         } else {
 225                 LBUG();
 226         }
 227 }
 228
 229 static void osd_submit_bio(int rw, struct bio *bio)
 230 {
 231         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 232         if (rw == 0)
 233                 submit_bio(READ, bio);
 234         else
 235                 submit_bio(WRITE, bio);
 236 }
 237
 238 static int can_be_merged(struct bio *bio, sector_t sector)
 239 {
 240         unsigned int size;
 241
 242         if (!bio)
 243                 return 0;
 244
 245         size = bio->bi_size >> 9;
 246         return bio->bi_sector + size == sector ? 1 : 0;
 247 }
 248
 249 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 250                       struct osd_iobuf *iobuf)
 251 {
 252         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 253         struct page  **pages = iobuf->dr_pages;
 254         int            npages = iobuf->dr_npages;
 255         unsigned long *blocks = iobuf->dr_blocks;
 256         int            total_blocks = npages * blocks_per_page;
 257         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 258         unsigned int   blocksize = inode->i_sb->s_blocksize;
 259         struct bio    *bio = NULL;
 260         struct page   *page;
 261         unsigned int   page_offset;
 262         sector_t       sector;
 263         int            nblocks;
 264         int            block_idx;
 265         int            page_idx;
 266         int            i;
 267         int            rc = 0;
 268         ENTRY;
 269
 270         LASSERT(iobuf->dr_npages == npages);
 271
 272         osd_brw_stats_update(osd, iobuf);
 273         iobuf->dr_start_time = cfs_time_current();
 274
 275         for (page_idx = 0, block_idx = 0;
 276              page_idx < npages;
 277              page_idx++, block_idx += blocks_per_page) {
 278
 279                 page = pages[page_idx];
 280                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 281
 282                 for (i = 0, page_offset = 0;
 283                      i < blocks_per_page;
 284                      i += nblocks, page_offset += blocksize * nblocks) {
 285
 286                         nblocks = 1;
 287
 288                         if (blocks[block_idx + i] == 0) {  /* hole */
 289                                 LASSERTF(iobuf->dr_rw == 0,
 290                                          "page_idx %u, block_idx %u, i %u\n",
 291                                          page_idx, block_idx, i);
 292                                 memset(kmap(page) + page_offset, 0, blocksize);
 293                                 kunmap(page);
 294                                 continue;
 295                         }
 296
 297                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 298
 299                         /* Additional contiguous file blocks? */
 300                         while (i + nblocks < blocks_per_page &&
 301                                (sector + (nblocks << sector_bits)) ==
 302                                ((sector_t)blocks[block_idx + i + nblocks] <<
 303                                 sector_bits))
 304                                 nblocks++;
 305
 306                         /* I only set the page to be constant only if it
 307                          * is mapped to a contiguous underlying disk block(s).
 308                          * It will then make sure the corresponding device
 309                          * cache of raid5 will be overwritten by this page.
 310                          * - jay */
 311                         if (iobuf->dr_rw && (nblocks == blocks_per_page) &&
 312                             mapping_cap_page_constant_write(inode->i_mapping))
 313                                 SetPageConstant(page);
 314
 315                         if (bio != NULL &&
 316                             can_be_merged(bio, sector) &&
 317                             bio_add_page(bio, page,
 318                                          blocksize * nblocks, page_offset) != 0)
 319                                 continue;       /* added this frag OK */
 320
 321                         if (bio != NULL) {
 322                                 struct request_queue *q =
 323                                         bdev_get_queue(bio->bi_bdev);
 324
 325                                 /* Dang! I have to fragment this I/O */
 326                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 327                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 328                                        bio->bi_size,
 329                                        bio->bi_vcnt, bio->bi_max_vecs,
 330                                        bio->bi_size >> 9, queue_max_sectors(q),
 331                                        bio_phys_segments(q, bio),
 332                                        queue_max_phys_segments(q),
 333                                        bio_hw_segments(q, bio),
 334                                        queue_max_hw_segments(q));
 335
 336                                 record_start_io(iobuf, bio->bi_size);
 337                                 osd_submit_bio(iobuf->dr_rw, bio);
 338                         }
 339
 340                         /* allocate new bio, limited by max BIO size, b=9945 */
 341                         bio = bio_alloc(GFP_NOIO, max(BIO_MAX_PAGES,
 342                                                       (npages - page_idx) *
 343                                                       blocks_per_page));
 344                         if (bio == NULL) {
 345                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 346                                        (npages - page_idx), blocks_per_page,
 347                                        (npages - page_idx) * blocks_per_page);
 348                                 rc = -ENOMEM;
 349                                 goto out;
 350                         }
 351
 352                         bio->bi_bdev = inode->i_sb->s_bdev;
 353                         bio->bi_sector = sector;
 354                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 355                         bio->bi_end_io = dio_complete_routine;
 356                         bio->bi_private = iobuf;
 357
 358                         rc = bio_add_page(bio, page,
 359                                           blocksize * nblocks, page_offset);
 360                         LASSERT(rc != 0);
 361                 }
 362         }
 363
 364         if (bio != NULL) {
 365                 record_start_io(iobuf, bio->bi_size);
 366                 osd_submit_bio(iobuf->dr_rw, bio);
 367                 rc = 0;
 368         }
 369
 370  out:
 371         /* in order to achieve better IO throughput, we don't wait for writes
 372          * completion here. instead we proceed with transaction commit in
 373          * parallel and wait for IO completion once transaction is stopped
 374          * see osd_trans_stop() for more details -bzzz */
 375         if (iobuf->dr_rw == 0) {
 376                 cfs_wait_event(iobuf->dr_wait,
 377                                cfs_atomic_read(&iobuf->dr_numreqs) == 0);
 378         }
 379
 380         if (rc == 0)
 381                 rc = iobuf->dr_error;
 382         RETURN(rc);
 383 }
 384
 385 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 386                                    struct niobuf_local *lnb)
 387 {
 388         ENTRY;
 389
 390         *nrpages = 0;
 391
 392         while (len > 0) {
 393                 int poff = offset & (CFS_PAGE_SIZE - 1);
 394                 int plen = CFS_PAGE_SIZE - poff;
 395
 396                 if (plen > len)
 397                         plen = len;
 398                 lnb->lnb_file_offset = offset;
 399                 lnb->lnb_page_offset = poff;
 400                 lnb->len = plen;
 401                 /* lb->flags = rnb->flags; */
 402                 lnb->flags = 0;
 403                 lnb->page = NULL;
 404                 lnb->rc = 0;
 405
 406                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 407                          (long long) len);
 408                 offset += plen;
 409                 len -= plen;
 410                 lnb++;
 411                 (*nrpages)++;
 412         }
 413
 414         RETURN(0);
 415 }
 416
 417 struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
 418 {
 419         struct inode      *inode = osd_dt_obj(dt)->oo_inode;
 420         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 421         struct page       *page;
 422
 423         LASSERT(inode);
 424
 425         page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
 426                                    GFP_NOFS | __GFP_HIGHMEM);
 427         if (unlikely(page == NULL))
 428                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 429
 430         return page;
 431 }
 432
 433 /*
 434  * there are following "locks":
 435  * journal_start
 436  * i_alloc_sem
 437  * i_mutex
 438  * page lock
 439
 440  * osd write path
 441     * lock page(s)
 442     * journal_start
 443     * truncate_sem
 444
 445  * ext4 vmtruncate:
 446     * lock pages, unlock
 447     * journal_start
 448     * lock partial page
 449     * i_data_sem
 450
 451 */
 452 int osd_bufs_get(const struct lu_env *env, struct dt_object *d, loff_t pos,
 453                  ssize_t len, struct niobuf_local *lnb, int rw,
 454                  struct lustre_capa *capa)
 455 {
 456         struct osd_object   *obj    = osd_dt_obj(d);
 457         int npages, i, rc = 0;
 458
 459         LASSERT(obj->oo_inode);
 460
 461         osd_map_remote_to_local(pos, len, &npages, lnb);
 462
 463         for (i = 0; i < npages; i++, lnb++) {
 464
 465                 /* We still set up for ungranted pages so that granted pages
 466                  * can be written to disk as they were promised, and portals
 467                  * needs to keep the pages all aligned properly. */
 468                 lnb->dentry = (void *) obj;
 469
 470                 lnb->page = osd_get_page(d, lnb->lnb_file_offset, rw);
 471                 if (lnb->page == NULL)
 472                         GOTO(cleanup, rc = -ENOMEM);
 473
 474                 /* DLM locking protects us from write and truncate competing
 475                  * for same region, but truncate can leave dirty page in the
 476                  * cache. it's possible the writeout on a such a page is in
 477                  * progress when we access it. it's also possible that during
 478                  * this writeout we put new (partial) data, but then won't
 479                  * be able to proceed in filter_commitrw_write(). thus let's
 480                  * just wait for writeout completion, should be rare enough.
 481                  * -bzzz */
 482                 wait_on_page_writeback(lnb->page);
 483                 BUG_ON(PageWriteback(lnb->page));
 484
 485                 lu_object_get(&d->do_lu);
 486         }
 487         rc = i;
 488
 489 cleanup:
 490         RETURN(rc);
 491 }
 492
 493 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 494                         struct niobuf_local *lnb, int npages)
 495 {
 496         struct osd_thread_info *oti = osd_oti_get(env);
 497         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 498         struct osd_device      *d = osd_obj2dev(osd_dt_obj(dt));
 499         int                     i;
 500
 501         /* to do IO stats, notice we do this here because
 502          * osd_do_bio() doesn't wait for write to complete */
 503         osd_fini_iobuf(d, iobuf);
 504
 505         for (i = 0; i < npages; i++) {
 506                 if (lnb[i].page == NULL)
 507                         continue;
 508                 LASSERT(PageLocked(lnb[i].page));
 509                 unlock_page(lnb[i].page);
 510                 page_cache_release(lnb[i].page);
 511                 lu_object_put(env, &dt->do_lu);
 512                 lnb[i].page = NULL;
 513         }
 514         RETURN(0);
 515 }
 516
 517 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 518                           struct niobuf_local *lnb, int npages)
 519 {
 520         struct osd_thread_info *oti   = osd_oti_get(env);
 521         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 522         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 523         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 524         struct timeval          start;
 525         struct timeval          end;
 526         unsigned long           timediff;
 527         ssize_t                 isize;
 528         __s64                   maxidx;
 529         int                     rc = 0;
 530         int                     i;
 531         int                     cache = 0;
 532
 533         LASSERT(inode);
 534
 535         osd_init_iobuf(osd, iobuf, 0);
 536
 537         isize = i_size_read(inode);
 538         maxidx = ((isize + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 539
 540         if (osd->od_writethrough_cache)
 541                 cache = 1;
 542         if (isize > osd->od_readcache_max_filesize)
 543                 cache = 0;
 544
 545         cfs_gettimeofday(&start);
 546         for (i = 0; i < npages; i++) {
 547
 548                 if (cache == 0)
 549                         generic_error_remove_page(inode->i_mapping,
 550                                                   lnb[i].page);
 551
 552                 /*
 553                  * till commit the content of the page is undefined
 554                  * we'll set it uptodate once bulk is done. otherwise
 555                  * subsequent reads can access non-stable data
 556                  */
 557                 ClearPageUptodate(lnb[i].page);
 558
 559                 if (lnb[i].len == CFS_PAGE_SIZE)
 560                         continue;
 561
 562                 if (maxidx >= lnb[i].page->index) {
 563                         osd_iobuf_add_page(iobuf, lnb[i].page);
 564                 } else {
 565                         long off;
 566                         char *p = kmap(lnb[i].page);
 567
 568                         off = lnb[i].lnb_page_offset;
 569                         if (off)
 570                                 memset(p, 0, off);
 571                         off = (lnb[i].lnb_page_offset + lnb[i].len) &
 572                               ~CFS_PAGE_MASK;
 573                         if (off)
 574                                 memset(p + off, 0, CFS_PAGE_SIZE - off);
 575                         kunmap(lnb[i].page);
 576                 }
 577         }
 578         cfs_gettimeofday(&end);
 579         timediff = cfs_timeval_sub(&end, &start, NULL);
 580         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 581
 582         if (iobuf->dr_npages) {
 583                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 584                                                        iobuf->dr_npages,
 585                                                        iobuf->dr_blocks,
 586                                                        oti->oti_created,
 587                                                        0, NULL);
 588                 if (likely(rc == 0)) {
 589                         rc = osd_do_bio(osd, inode, iobuf);
 590                         /* do IO stats for preparation reads */
 591                         osd_fini_iobuf(osd, iobuf);
 592                 }
 593         }
 594         RETURN(rc);
 595 }
 596
 597 /* Check if a block is allocated or not */
 598 static int osd_is_mapped(struct inode *inode, obd_size offset)
 599 {
 600         sector_t (*fs_bmap)(struct address_space *, sector_t);
 601
 602         fs_bmap = inode->i_mapping->a_ops->bmap;
 603
 604         /* We can't know if we are overwriting or not */
 605         if (unlikely(fs_bmap == NULL))
 606                 return 0;
 607
 608         if (i_size_read(inode) == 0)
 609                 return 0;
 610
 611         /* Beyond EOF, must not be mapped */
 612         if (((i_size_read(inode) - 1) >> inode->i_blkbits) <
 613             (offset >> inode->i_blkbits))
 614                 return 0;
 615
 616         if (fs_bmap(inode->i_mapping, offset >> inode->i_blkbits) == 0)
 617                 return 0;
 618
 619         return 1;
 620 }
 621
 622 static int osd_declare_write_commit(const struct lu_env *env,
 623                                     struct dt_object *dt,
 624                                     struct niobuf_local *lnb, int npages,
 625                                     struct thandle *handle)
 626 {
 627         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 628         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
 629         struct osd_thandle      *oh;
 630         int                      extents = 1;
 631         int                      depth;
 632         int                      i;
 633         int                      newblocks;
 634         int                      rc = 0;
 635         int                      flags = 0;
 636         bool                     ignore_quota = false;
 637         long long                quota_space = 0;
 638         ENTRY;
 639
 640         LASSERT(handle != NULL);
 641         oh = container_of0(handle, struct osd_thandle, ot_super);
 642         LASSERT(oh->ot_handle == NULL);
 643
 644         newblocks = npages;
 645
 646         /* calculate number of extents (probably better to pass nb) */
 647         for (i = 0; i < npages; i++) {
 648                 if (i && lnb[i].lnb_file_offset !=
 649                     lnb[i - 1].lnb_file_offset + lnb[i - 1].len)
 650                         extents++;
 651
 652                 if (!osd_is_mapped(inode, lnb[i].lnb_file_offset))
 653                         quota_space += CFS_PAGE_SIZE;
 654
 655                 /* ignore quota for the whole request if any page is from
 656                  * client cache or written by root.
 657                  *
 658                  * XXX we could handle this on per-lnb basis as done by
 659                  * grant. */
 660                 if ((lnb[i].flags & OBD_BRW_NOQUOTA) ||
 661                     !(lnb[i].flags & OBD_BRW_SYNC))
 662                         ignore_quota = true;
 663         }
 664
 665         /*
 666          * each extent can go into new leaf causing a split
 667          * 5 is max tree depth: inode + 4 index blocks
 668          * with blockmaps, depth is 3 at most
 669          */
 670         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 671                 /*
 672                  * many concurrent threads may grow tree by the time
 673                  * our transaction starts. so, consider 2 is a min depth
 674                  */
 675                 depth = ext_depth(inode);
 676                 depth = max(depth, 1) + 1;
 677                 newblocks += depth;
 678                 oh->ot_credits++; /* inode */
 679                 oh->ot_credits += depth * 2 * extents;
 680         } else {
 681                 depth = 3;
 682                 newblocks += depth;
 683                 oh->ot_credits++; /* inode */
 684                 oh->ot_credits += depth * extents;
 685         }
 686
 687         /* quota space for metadata blocks */
 688         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
 689
 690         /* quota space should be reported in 1K blocks */
 691         quota_space = toqb(quota_space);
 692
 693         /* each new block can go in different group (bitmap + gd) */
 694
 695         /* we can't dirty more bitmap blocks than exist */
 696         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
 697                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
 698         else
 699                 oh->ot_credits += newblocks;
 700
 701         /* we can't dirty more gd blocks than exist */
 702         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
 703                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
 704         else
 705                 oh->ot_credits += newblocks;
 706
 707         /* make sure the over quota flags were not set */
 708         lnb[0].flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA);
 709
 710         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid,
 711                                    quota_space, oh, true, true, &flags,
 712                                    ignore_quota);
 713
 714         /* we need only to store the overquota flags in the first lnb for
 715          * now, once we support multiple objects BRW, this code needs be
 716          * revised. */
 717         if (flags & QUOTA_FL_OVER_USRQUOTA)
 718                 lnb[0].flags |= OBD_BRW_OVER_USRQUOTA;
 719         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 720                 lnb[0].flags |= OBD_BRW_OVER_GRPQUOTA;
 721
 722         RETURN(rc);
 723 }
 724
 725 /* Check if a block is allocated or not */
 726 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 727                             struct niobuf_local *lnb, int npages,
 728                             struct thandle *thandle)
 729 {
 730         struct osd_thread_info *oti = osd_oti_get(env);
 731         struct osd_iobuf *iobuf = &oti->oti_iobuf;
 732         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 733         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
 734         loff_t isize;
 735         int rc = 0, i;
 736
 737         LASSERT(inode);
 738
 739         osd_init_iobuf(osd, iobuf, 1);
 740         isize = i_size_read(inode);
 741         ll_vfs_dq_init(inode);
 742
 743         for (i = 0; i < npages; i++) {
 744                 if (lnb[i].rc == -ENOSPC &&
 745                     osd_is_mapped(inode, lnb[i].lnb_file_offset)) {
 746                         /* Allow the write to proceed if overwriting an
 747                          * existing block */
 748                         lnb[i].rc = 0;
 749                 }
 750
 751                 if (lnb[i].rc) { /* ENOSPC, network RPC error, etc. */
 752                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
 753                                lnb[i].rc);
 754                         LASSERT(lnb[i].page);
 755                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 756                         continue;
 757                 }
 758
 759                 LASSERT(PageLocked(lnb[i].page));
 760                 LASSERT(!PageWriteback(lnb[i].page));
 761
 762                 if (lnb[i].lnb_file_offset + lnb[i].len > isize)
 763                         isize = lnb[i].lnb_file_offset + lnb[i].len;
 764
 765                 /*
 766                  * Since write and truncate are serialized by oo_sem, even
 767                  * partial-page truncate should not leave dirty pages in the
 768                  * page cache.
 769                  */
 770                 LASSERT(!PageDirty(lnb[i].page));
 771
 772                 SetPageUptodate(lnb[i].page);
 773
 774                 osd_iobuf_add_page(iobuf, lnb[i].page);
 775         }
 776
 777         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
 778                 rc = -ENOSPC;
 779         } else if (iobuf->dr_npages > 0) {
 780                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 781                                                        iobuf->dr_npages,
 782                                                        iobuf->dr_blocks,
 783                                                        oti->oti_created,
 784                                                        1, NULL);
 785         } else {
 786                 /* no pages to write, no transno is needed */
 787                 thandle->th_local = 1;
 788         }
 789
 790         if (likely(rc == 0)) {
 791                 if (isize > i_size_read(inode)) {
 792                         i_size_write(inode, isize);
 793                         LDISKFS_I(inode)->i_disksize = isize;
 794                         inode->i_sb->s_op->dirty_inode(inode);
 795                 }
 796
 797                 rc = osd_do_bio(osd, inode, iobuf);
 798                 /* we don't do stats here as in read path because
 799                  * write is async: we'll do this in osd_put_bufs() */
 800         }
 801
 802         if (unlikely(rc != 0)) {
 803                 /* if write fails, we should drop pages from the cache */
 804                 for (i = 0; i < npages; i++) {
 805                         if (lnb[i].page == NULL)
 806                                 continue;
 807                         LASSERT(PageLocked(lnb[i].page));
 808                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 809                 }
 810         }
 811
 812         RETURN(rc);
 813 }
 814
 815 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 816                          struct niobuf_local *lnb, int npages)
 817 {
 818         struct osd_thread_info *oti = osd_oti_get(env);
 819         struct osd_iobuf *iobuf = &oti->oti_iobuf;
 820         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 821         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 822         struct timeval start, end;
 823         unsigned long timediff;
 824         int rc = 0, i, m = 0, cache = 0;
 825
 826         LASSERT(inode);
 827
 828         osd_init_iobuf(osd, iobuf, 0);
 829
 830         if (osd->od_read_cache)
 831                 cache = 1;
 832         if (i_size_read(inode) > osd->od_readcache_max_filesize)
 833                 cache = 0;
 834
 835         cfs_gettimeofday(&start);
 836         for (i = 0; i < npages; i++) {
 837
 838                 if (i_size_read(inode) <= lnb[i].lnb_file_offset)
 839                         /* If there's no more data, abort early.
 840                          * lnb->rc == 0, so it's easy to detect later. */
 841                         break;
 842
 843                 if (i_size_read(inode) <
 844                     lnb[i].lnb_file_offset + lnb[i].len - 1)
 845                         lnb[i].rc = i_size_read(inode) - lnb[i].lnb_file_offset;
 846                 else
 847                         lnb[i].rc = lnb[i].len;
 848                 m += lnb[i].len;
 849
 850                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS, 1);
 851                 if (PageUptodate(lnb[i].page)) {
 852                         lprocfs_counter_add(osd->od_stats,
 853                                             LPROC_OSD_CACHE_HIT, 1);
 854                 } else {
 855                         lprocfs_counter_add(osd->od_stats,
 856                                             LPROC_OSD_CACHE_MISS, 1);
 857                         osd_iobuf_add_page(iobuf, lnb[i].page);
 858                 }
 859                 if (cache == 0)
 860                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 861         }
 862         cfs_gettimeofday(&end);
 863         timediff = cfs_timeval_sub(&end, &start, NULL);
 864         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 865
 866         if (iobuf->dr_npages) {
 867                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 868                                                        iobuf->dr_npages,
 869                                                        iobuf->dr_blocks,
 870                                                        oti->oti_created,
 871                                                        0, NULL);
 872                 rc = osd_do_bio(osd, inode, iobuf);
 873
 874                 /* IO stats will be done in osd_bufs_put() */
 875         }
 876
 877         RETURN(rc);
 878 }
 879
 880 /*
 881  * XXX: Another layering violation for now.
 882  *
 883  * We don't want to use ->f_op->read methods, because generic file write
 884  *
 885  *         - serializes on ->i_sem, and
 886  *
 887  *         - does a lot of extra work like balance_dirty_pages(),
 888  *
 889  * which doesn't work for globally shared files like /last_rcvd.
 890  */
 891 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
 892 {
 893         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 894
 895         memcpy(buffer, (char *)ei->i_data, buflen);
 896
 897         return  buflen;
 898 }
 899
 900 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
 901 {
 902         struct buffer_head *bh;
 903         unsigned long block;
 904         int osize;
 905         int blocksize;
 906         int csize;
 907         int boffs;
 908         int err;
 909
 910         /* prevent reading after eof */
 911         cfs_spin_lock(&inode->i_lock);
 912         if (i_size_read(inode) < *offs + size) {
 913                 loff_t diff = i_size_read(inode) - *offs;
 914                 cfs_spin_unlock(&inode->i_lock);
 915                 if (diff < 0) {
 916                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
 917                                i_size_read(inode), *offs);
 918                         return -EBADR;
 919                 } else if (diff == 0) {
 920                         return 0;
 921                 } else {
 922                         size = diff;
 923                 }
 924         } else {
 925                 cfs_spin_unlock(&inode->i_lock);
 926         }
 927
 928         blocksize = 1 << inode->i_blkbits;
 929         osize = size;
 930         while (size > 0) {
 931                 block = *offs >> inode->i_blkbits;
 932                 boffs = *offs & (blocksize - 1);
 933                 csize = min(blocksize - boffs, size);
 934                 bh = ldiskfs_bread(NULL, inode, block, 0, &err);
 935                 if (!bh) {
 936                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %d\n",
 937                                LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
 938                                csize, *offs, inode->i_ino, err);
 939                         return err;
 940                 }
 941
 942                 memcpy(buf, bh->b_data + boffs, csize);
 943                 brelse(bh);
 944
 945                 *offs += csize;
 946                 buf += csize;
 947                 size -= csize;
 948         }
 949         return osize;
 950 }
 951
 952 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 953                         struct lu_buf *buf, loff_t *pos,
 954                         struct lustre_capa *capa)
 955 {
 956         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 957         int           rc;
 958
 959         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_READ))
 960                 RETURN(-EACCES);
 961
 962         /* Read small symlink from inode body as we need to maintain correct
 963          * on-disk symlinks for ldiskfs.
 964          */
 965         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
 966             (buf->lb_len <= sizeof(LDISKFS_I(inode)->i_data)))
 967                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
 968         else
 969                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
 970
 971         return rc;
 972 }
 973
 974 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 975                                  const loff_t size, loff_t pos,
 976                                  struct thandle *handle)
 977 {
 978         struct osd_thandle *oh;
 979         int                 credits;
 980         struct inode       *inode;
 981         int                 rc;
 982         ENTRY;
 983
 984         LASSERT(handle != NULL);
 985
 986         oh = container_of0(handle, struct osd_thandle, ot_super);
 987         LASSERT(oh->ot_handle == NULL);
 988
 989         /* XXX: size == 0 or INT_MAX indicating a catalog header update or
 990          *      llog write, see comment in mdd_declare_llog_record().
 991          *
 992          *      This hack will be removed with llog over OSD landing
 993          */
 994         if (size == DECLARE_LLOG_REWRITE)
 995                 credits = 2;
 996         else if (size == DECLARE_LLOG_WRITE)
 997                 credits = 6;
 998         else
 999                 credits = osd_dto_credits_noquota[DTO_WRITE_BLOCK];
1000
1001         OSD_DECLARE_OP(oh, write);
1002         oh->ot_credits += credits;
1003
1004         inode = osd_dt_obj(dt)->oo_inode;
1005
1006         /* we may declare write to non-exist llog */
1007         if (inode == NULL)
1008                 RETURN(0);
1009
1010         /* dt_declare_write() is usually called for system objects, such
1011          * as llog or last_rcvd files. We needn't enforce quota on those
1012          * objects, so always set the lqi_space as 0. */
1013         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
1014                                    true, true, NULL, false);
1015         RETURN(rc);
1016 }
1017
1018 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1019 {
1020
1021         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1022         LDISKFS_I(inode)->i_disksize = buflen;
1023         i_size_write(inode, buflen);
1024         inode->i_sb->s_op->dirty_inode(inode);
1025
1026         return 0;
1027 }
1028
1029 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1030                              int write_NUL, loff_t *offs, handle_t *handle)
1031 {
1032         struct buffer_head *bh        = NULL;
1033         loff_t              offset    = *offs;
1034         loff_t              new_size  = i_size_read(inode);
1035         unsigned long       block;
1036         int                 blocksize = 1 << inode->i_blkbits;
1037         int                 err = 0;
1038         int                 size;
1039         int                 boffs;
1040         int                 dirty_inode = 0;
1041
1042         if (write_NUL) {
1043                 /*
1044                  * long symlink write does not count the NUL terminator in
1045                  * bufsize, we write it, and the inode's file size does not
1046                  * count the NUL terminator as well.
1047                  */
1048                 ((char *)buf)[bufsize] = '\0';
1049                 ++bufsize;
1050         }
1051         while (bufsize > 0) {
1052                 if (bh != NULL)
1053                         brelse(bh);
1054
1055                 block = offset >> inode->i_blkbits;
1056                 boffs = offset & (blocksize - 1);
1057                 size = min(blocksize - boffs, bufsize);
1058                 bh = ldiskfs_bread(handle, inode, block, 1, &err);
1059                 if (!bh) {
1060                         CERROR("%s: error reading offset %llu (block %lu): "
1061                                "rc = %d\n",
1062                                inode->i_sb->s_id, offset, block, err);
1063                         break;
1064                 }
1065
1066                 err = ldiskfs_journal_get_write_access(handle, bh);
1067                 if (err) {
1068                         CERROR("journal_get_write_access() returned error %d\n",
1069                                err);
1070                         break;
1071                 }
1072                 LASSERTF(boffs + size <= bh->b_size,
1073                          "boffs %d size %d bh->b_size %lu",
1074                          boffs, size, (unsigned long)bh->b_size);
1075                 memcpy(bh->b_data + boffs, buf, size);
1076                 err = ldiskfs_journal_dirty_metadata(handle, bh);
1077                 if (err)
1078                         break;
1079
1080                 if (offset + size > new_size)
1081                         new_size = offset + size;
1082                 offset += size;
1083                 bufsize -= size;
1084                 buf += size;
1085         }
1086         if (bh)
1087                 brelse(bh);
1088
1089         if (write_NUL)
1090                 --new_size;
1091         /* correct in-core and on-disk sizes */
1092         if (new_size > i_size_read(inode)) {
1093                 cfs_spin_lock(&inode->i_lock);
1094                 if (new_size > i_size_read(inode))
1095                         i_size_write(inode, new_size);
1096                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1097                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1098                         dirty_inode = 1;
1099                 }
1100                 cfs_spin_unlock(&inode->i_lock);
1101                 if (dirty_inode)
1102                         inode->i_sb->s_op->dirty_inode(inode);
1103         }
1104
1105         if (err == 0)
1106                 *offs = offset;
1107         return err;
1108 }
1109
1110 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1111                          const struct lu_buf *buf, loff_t *pos,
1112                          struct thandle *handle, struct lustre_capa *capa,
1113                          int ignore_quota)
1114 {
1115         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1116         struct osd_thandle      *oh;
1117         ssize_t                 result;
1118 #ifdef HAVE_QUOTA_SUPPORT
1119         cfs_cap_t               save = cfs_curproc_cap_pack();
1120 #endif
1121         int                     is_link;
1122
1123         LASSERT(dt_object_exists(dt));
1124
1125         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_WRITE))
1126                 return -EACCES;
1127
1128         LASSERT(handle != NULL);
1129
1130         /* XXX: don't check: one declared chunk can be used many times */
1131         /* OSD_EXEC_OP(handle, write); */
1132
1133         oh = container_of(handle, struct osd_thandle, ot_super);
1134         LASSERT(oh->ot_handle->h_transaction != NULL);
1135 #ifdef HAVE_QUOTA_SUPPORT
1136         if (ignore_quota)
1137                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1138         else
1139                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1140 #endif
1141         /* Write small symlink to inode body as we need to maintain correct
1142          * on-disk symlinks for ldiskfs.
1143          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1144          * does not count it in.
1145          */
1146         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1147         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1148                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1149         else
1150                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1151                                                   buf->lb_len, is_link, pos,
1152                                                   oh->ot_handle);
1153 #ifdef HAVE_QUOTA_SUPPORT
1154         cfs_curproc_cap_unpack(save);
1155 #endif
1156         if (result == 0)
1157                 result = buf->lb_len;
1158         return result;
1159 }
1160
1161 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1162                              __u64 start, __u64 end, struct thandle *th)
1163 {
1164         struct osd_thandle *oh;
1165         struct inode       *inode;
1166         int                 rc;
1167         ENTRY;
1168
1169         LASSERT(th);
1170         oh = container_of(th, struct osd_thandle, ot_super);
1171
1172         OSD_DECLARE_OP(oh, punch);
1173
1174         /*
1175          * we don't need to reserve credits for whole truncate
1176          * it's not possible as truncate may need to free too many
1177          * blocks and that won't fit a single transaction. instead
1178          * we reserve credits to change i_size and put inode onto
1179          * orphan list. if needed truncate will extend or restart
1180          * transaction
1181          */
1182         oh->ot_credits += osd_dto_credits_noquota[DTO_ATTR_SET_BASE];
1183         oh->ot_credits += 3;
1184
1185         inode = osd_dt_obj(dt)->oo_inode;
1186         LASSERT(inode);
1187
1188         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
1189                                    true, true, NULL, false);
1190         RETURN(rc);
1191 }
1192
1193 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1194                      __u64 start, __u64 end, struct thandle *th,
1195                      struct lustre_capa *capa)
1196 {
1197         struct osd_thandle *oh;
1198         struct osd_object  *obj = osd_dt_obj(dt);
1199         struct inode       *inode = obj->oo_inode;
1200         handle_t           *h;
1201         tid_t               tid;
1202         int                 rc, rc2 = 0;
1203         ENTRY;
1204
1205         LASSERT(end == OBD_OBJECT_EOF);
1206         LASSERT(dt_object_exists(dt));
1207         LASSERT(osd_invariant(obj));
1208
1209         LASSERT(th);
1210         oh = container_of(th, struct osd_thandle, ot_super);
1211         LASSERT(oh->ot_handle->h_transaction != NULL);
1212
1213         OSD_EXEC_OP(th, punch);
1214
1215         tid = oh->ot_handle->h_transaction->t_tid;
1216
1217         rc = vmtruncate(inode, start);
1218
1219         /*
1220          * For a partial-page truncate, flush the page to disk immediately to
1221          * avoid data corruption during direct disk write.  b=17397
1222          */
1223         if (rc == 0 && (start & ~CFS_PAGE_MASK) != 0)
1224                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1225
1226         h = journal_current_handle();
1227         LASSERT(h != NULL);
1228         LASSERT(h == oh->ot_handle);
1229
1230         if (tid != h->h_transaction->t_tid) {
1231                 int credits = oh->ot_credits;
1232                 /*
1233                  * transaction has changed during truncate
1234                  * we need to restart the handle with our credits
1235                  */
1236                 if (h->h_buffer_credits < credits) {
1237                         if (ldiskfs_journal_extend(h, credits))
1238                                 rc2 = ldiskfs_journal_restart(h, credits);
1239                 }
1240         }
1241
1242         RETURN(rc == 0 ? rc2 : rc);
1243 }
1244
1245 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1246                           struct ll_user_fiemap *fm)
1247 {
1248         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1249         struct osd_thread_info *info   = osd_oti_get(env);
1250         struct dentry          *dentry = &info->oti_obj_dentry;
1251         struct file            *file   = &info->oti_file;
1252         mm_segment_t            saved_fs;
1253         int rc;
1254
1255         LASSERT(inode);
1256         dentry->d_inode = inode;
1257         file->f_dentry = dentry;
1258         file->f_mapping = inode->i_mapping;
1259         file->f_op = inode->i_fop;
1260
1261         saved_fs = get_fs();
1262         set_fs(get_ds());
1263         /* ldiskfs_ioctl does not have a inode argument */
1264         if (inode->i_fop->unlocked_ioctl)
1265                 rc = inode->i_fop->unlocked_ioctl(file, FSFILT_IOC_FIEMAP,
1266                                                   (long)fm);
1267         else
1268                 rc = -ENOTTY;
1269         set_fs(saved_fs);
1270         return rc;
1271 }
1272
1273 /*
1274  * in some cases we may need declare methods for objects being created
1275  * e.g., when we create symlink
1276  */
1277 const struct dt_body_operations osd_body_ops_new = {
1278         .dbo_declare_write = osd_declare_write,
1279 };
1280
1281 const struct dt_body_operations osd_body_ops = {
1282         .dbo_read                 = osd_read,
1283         .dbo_declare_write        = osd_declare_write,
1284         .dbo_write                = osd_write,
1285         .dbo_bufs_get             = osd_bufs_get,
1286         .dbo_bufs_put             = osd_bufs_put,
1287         .dbo_write_prep           = osd_write_prep,
1288         .dbo_declare_write_commit = osd_declare_write_commit,
1289         .dbo_write_commit         = osd_write_commit,
1290         .dbo_read_prep            = osd_read_prep,
1291         .do_declare_punch         = osd_declare_punch,
1292         .do_punch                 = osd_punch,
1293         .dbo_fiemap_get           = osd_fiemap_get,
1294 };
1295