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,
  97                              int rw, int line)
  98 {
  99         LASSERTF(iobuf->dr_elapsed_valid == 0,
 100                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
 101                  cfs_atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
 102                  iobuf->dr_init_at);
 103
 104         cfs_waitq_init(&iobuf->dr_wait);
 105         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 106         iobuf->dr_max_pages = PTLRPC_MAX_BRW_PAGES;
 107         iobuf->dr_npages = 0;
 108         iobuf->dr_error = 0;
 109         iobuf->dr_dev = d;
 110         iobuf->dr_frags = 0;
 111         iobuf->dr_elapsed = 0;
 112         /* must be counted before, so assert */
 113         iobuf->dr_rw = rw;
 114         iobuf->dr_init_at = line;
 115 }
 116 #define osd_init_iobuf(dev,iobuf,rw) __osd_init_iobuf(dev, iobuf, rw, __LINE__)
 117
 118 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
 119 {
 120         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 121         iobuf->dr_pages[iobuf->dr_npages++] = page;
 122 }
 123
 124 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 125 {
 126         int rw = iobuf->dr_rw;
 127
 128         if (iobuf->dr_elapsed_valid) {
 129                 iobuf->dr_elapsed_valid = 0;
 130                 LASSERT(iobuf->dr_dev == d);
 131                 LASSERT(iobuf->dr_frags > 0);
 132                 lprocfs_oh_tally(&d->od_brw_stats.
 133                                  hist[BRW_R_DIO_FRAGS+rw],
 134                                  iobuf->dr_frags);
 135                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 136                                       iobuf->dr_elapsed);
 137         }
 138 }
 139
 140 #ifdef HAVE_BIO_ENDIO_2ARG
 141 #define DIO_RETURN(a)
 142 static void dio_complete_routine(struct bio *bio, int error)
 143 #else
 144 #define DIO_RETURN(a)   return(a)
 145 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 146 #endif
 147 {
 148         struct osd_iobuf *iobuf = bio->bi_private;
 149         struct bio_vec *bvl;
 150         int i;
 151
 152         /* CAVEAT EMPTOR: possibly in IRQ context
 153          * DO NOT record procfs stats here!!! */
 154
 155         if (unlikely(iobuf == NULL)) {
 156                 CERROR("***** bio->bi_private is NULL!  This should never "
 157                        "happen.  Normally, I would crash here, but instead I "
 158                        "will dump the bio contents to the console.  Please "
 159                        "report this to <http://jira.whamcloud.com/> , along "
 160                        "with any interesting messages leading up to this point "
 161                        "(like SCSI errors, perhaps).  Because bi_private is "
 162                        "NULL, I can't wake up the thread that initiated this "
 163                        "IO - you will probably have to reboot this node.\n");
 164                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 165                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 166                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 167                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 168                        bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
 169                        bio->bi_private);
 170                 DIO_RETURN(0);
 171         }
 172
 173         /* the check is outside of the cycle for performance reason -bzzz */
 174         if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
 175                 bio_for_each_segment(bvl, bio, i) {
 176                         if (likely(error == 0))
 177                                 SetPageUptodate(bvl->bv_page);
 178                         LASSERT(PageLocked(bvl->bv_page));
 179                         ClearPageConstant(bvl->bv_page);
 180                 }
 181                 cfs_atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 182         } else {
 183                 struct page *p = iobuf->dr_pages[0];
 184                 if (p->mapping) {
 185                         if (mapping_cap_page_constant_write(p->mapping)) {
 186                                 bio_for_each_segment(bvl, bio, i) {
 187                                         ClearPageConstant(bvl->bv_page);
 188                                 }
 189                         }
 190                 }
 191                 cfs_atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 192         }
 193
 194         /* any real error is good enough -bzzz */
 195         if (error != 0 && iobuf->dr_error == 0)
 196                 iobuf->dr_error = error;
 197
 198         /*
 199          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 200          * it's possible that service thread will see dr_numreqs
 201          * is zero, but dr_elapsed is not set yet, leading to lost
 202          * data in this processing and an assertion in a subsequent
 203          * call to OSD.
 204          */
 205         if (cfs_atomic_read(&iobuf->dr_numreqs) == 1) {
 206                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
 207                 iobuf->dr_elapsed_valid = 1;
 208         }
 209         if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs))
 210                 cfs_waitq_signal(&iobuf->dr_wait);
 211
 212         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 213          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 214          * mempool when serious on-disk fragmentation was encountered,
 215          * deadlocking the OST.  The bios are now released as soon as complete
 216          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 217         bio_put(bio);
 218         DIO_RETURN(0);
 219 }
 220
 221 static void record_start_io(struct osd_iobuf *iobuf, int size)
 222 {
 223         struct osd_device    *osd = iobuf->dr_dev;
 224         struct obd_histogram *h = osd->od_brw_stats.hist;
 225
 226         iobuf->dr_frags++;
 227         cfs_atomic_inc(&iobuf->dr_numreqs);
 228
 229         if (iobuf->dr_rw == 0) {
 230                 cfs_atomic_inc(&osd->od_r_in_flight);
 231                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 232                                  cfs_atomic_read(&osd->od_r_in_flight));
 233                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 234         } else if (iobuf->dr_rw == 1) {
 235                 cfs_atomic_inc(&osd->od_w_in_flight);
 236                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 237                                  cfs_atomic_read(&osd->od_w_in_flight));
 238                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 239         } else {
 240                 LBUG();
 241         }
 242 }
 243
 244 static void osd_submit_bio(int rw, struct bio *bio)
 245 {
 246         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 247         if (rw == 0)
 248                 submit_bio(READ, bio);
 249         else
 250                 submit_bio(WRITE, bio);
 251 }
 252
 253 static int can_be_merged(struct bio *bio, sector_t sector)
 254 {
 255         unsigned int size;
 256
 257         if (!bio)
 258                 return 0;
 259
 260         size = bio->bi_size >> 9;
 261         return bio->bi_sector + size == sector ? 1 : 0;
 262 }
 263
 264 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 265                       struct osd_iobuf *iobuf)
 266 {
 267         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 268         struct page  **pages = iobuf->dr_pages;
 269         int            npages = iobuf->dr_npages;
 270         unsigned long *blocks = iobuf->dr_blocks;
 271         int            total_blocks = npages * blocks_per_page;
 272         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 273         unsigned int   blocksize = inode->i_sb->s_blocksize;
 274         struct bio    *bio = NULL;
 275         struct page   *page;
 276         unsigned int   page_offset;
 277         sector_t       sector;
 278         int            nblocks;
 279         int            block_idx;
 280         int            page_idx;
 281         int            i;
 282         int            rc = 0;
 283         ENTRY;
 284
 285         LASSERT(iobuf->dr_npages == npages);
 286
 287         osd_brw_stats_update(osd, iobuf);
 288         iobuf->dr_start_time = cfs_time_current();
 289
 290         for (page_idx = 0, block_idx = 0;
 291              page_idx < npages;
 292              page_idx++, block_idx += blocks_per_page) {
 293
 294                 page = pages[page_idx];
 295                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 296
 297                 for (i = 0, page_offset = 0;
 298                      i < blocks_per_page;
 299                      i += nblocks, page_offset += blocksize * nblocks) {
 300
 301                         nblocks = 1;
 302
 303                         if (blocks[block_idx + i] == 0) {  /* hole */
 304                                 LASSERTF(iobuf->dr_rw == 0,
 305                                          "page_idx %u, block_idx %u, i %u\n",
 306                                          page_idx, block_idx, i);
 307                                 memset(kmap(page) + page_offset, 0, blocksize);
 308                                 kunmap(page);
 309                                 continue;
 310                         }
 311
 312                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 313
 314                         /* Additional contiguous file blocks? */
 315                         while (i + nblocks < blocks_per_page &&
 316                                (sector + (nblocks << sector_bits)) ==
 317                                ((sector_t)blocks[block_idx + i + nblocks] <<
 318                                 sector_bits))
 319                                 nblocks++;
 320
 321                         /* I only set the page to be constant only if it
 322                          * is mapped to a contiguous underlying disk block(s).
 323                          * It will then make sure the corresponding device
 324                          * cache of raid5 will be overwritten by this page.
 325                          * - jay */
 326                         if (iobuf->dr_rw && (nblocks == blocks_per_page) &&
 327                             mapping_cap_page_constant_write(inode->i_mapping))
 328                                 SetPageConstant(page);
 329
 330                         if (bio != NULL &&
 331                             can_be_merged(bio, sector) &&
 332                             bio_add_page(bio, page,
 333                                          blocksize * nblocks, page_offset) != 0)
 334                                 continue;       /* added this frag OK */
 335
 336                         if (bio != NULL) {
 337                                 struct request_queue *q =
 338                                         bdev_get_queue(bio->bi_bdev);
 339
 340                                 /* Dang! I have to fragment this I/O */
 341                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 342                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 343                                        bio->bi_size,
 344                                        bio->bi_vcnt, bio->bi_max_vecs,
 345                                        bio->bi_size >> 9, queue_max_sectors(q),
 346                                        bio_phys_segments(q, bio),
 347                                        queue_max_phys_segments(q),
 348                                        bio_hw_segments(q, bio),
 349                                        queue_max_hw_segments(q));
 350
 351                                 record_start_io(iobuf, bio->bi_size);
 352                                 osd_submit_bio(iobuf->dr_rw, bio);
 353                         }
 354
 355                         /* allocate new bio, limited by max BIO size, b=9945 */
 356                         bio = bio_alloc(GFP_NOIO, max(BIO_MAX_PAGES,
 357                                                       (npages - page_idx) *
 358                                                       blocks_per_page));
 359                         if (bio == NULL) {
 360                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 361                                        (npages - page_idx), blocks_per_page,
 362                                        (npages - page_idx) * blocks_per_page);
 363                                 rc = -ENOMEM;
 364                                 goto out;
 365                         }
 366
 367                         bio->bi_bdev = inode->i_sb->s_bdev;
 368                         bio->bi_sector = sector;
 369                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 370                         bio->bi_end_io = dio_complete_routine;
 371                         bio->bi_private = iobuf;
 372
 373                         rc = bio_add_page(bio, page,
 374                                           blocksize * nblocks, page_offset);
 375                         LASSERT(rc != 0);
 376                 }
 377         }
 378
 379         if (bio != NULL) {
 380                 record_start_io(iobuf, bio->bi_size);
 381                 osd_submit_bio(iobuf->dr_rw, bio);
 382                 rc = 0;
 383         }
 384
 385  out:
 386         /* in order to achieve better IO throughput, we don't wait for writes
 387          * completion here. instead we proceed with transaction commit in
 388          * parallel and wait for IO completion once transaction is stopped
 389          * see osd_trans_stop() for more details -bzzz */
 390         if (iobuf->dr_rw == 0) {
 391                 cfs_wait_event(iobuf->dr_wait,
 392                                cfs_atomic_read(&iobuf->dr_numreqs) == 0);
 393         }
 394
 395         if (rc == 0)
 396                 rc = iobuf->dr_error;
 397         RETURN(rc);
 398 }
 399
 400 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 401                                    struct niobuf_local *lnb)
 402 {
 403         ENTRY;
 404
 405         *nrpages = 0;
 406
 407         while (len > 0) {
 408                 int poff = offset & (CFS_PAGE_SIZE - 1);
 409                 int plen = CFS_PAGE_SIZE - poff;
 410
 411                 if (plen > len)
 412                         plen = len;
 413                 lnb->lnb_file_offset = offset;
 414                 lnb->lnb_page_offset = poff;
 415                 lnb->len = plen;
 416                 /* lb->flags = rnb->flags; */
 417                 lnb->flags = 0;
 418                 lnb->page = NULL;
 419                 lnb->rc = 0;
 420
 421                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 422                          (long long) len);
 423                 offset += plen;
 424                 len -= plen;
 425                 lnb++;
 426                 (*nrpages)++;
 427         }
 428
 429         RETURN(0);
 430 }
 431
 432 struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
 433 {
 434         struct inode      *inode = osd_dt_obj(dt)->oo_inode;
 435         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 436         struct page       *page;
 437
 438         LASSERT(inode);
 439
 440         page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
 441                                    GFP_NOFS | __GFP_HIGHMEM);
 442         if (unlikely(page == NULL))
 443                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 444
 445         return page;
 446 }
 447
 448 /*
 449  * there are following "locks":
 450  * journal_start
 451  * i_alloc_sem
 452  * i_mutex
 453  * page lock
 454
 455  * osd write path
 456     * lock page(s)
 457     * journal_start
 458     * truncate_sem
 459
 460  * ext4 vmtruncate:
 461     * lock pages, unlock
 462     * journal_start
 463     * lock partial page
 464     * i_data_sem
 465
 466 */
 467 int osd_bufs_get(const struct lu_env *env, struct dt_object *d, loff_t pos,
 468                  ssize_t len, struct niobuf_local *lnb, int rw,
 469                  struct lustre_capa *capa)
 470 {
 471         struct osd_object   *obj    = osd_dt_obj(d);
 472         int npages, i, rc = 0;
 473
 474         LASSERT(obj->oo_inode);
 475
 476         osd_map_remote_to_local(pos, len, &npages, lnb);
 477
 478         for (i = 0; i < npages; i++, lnb++) {
 479
 480                 /* We still set up for ungranted pages so that granted pages
 481                  * can be written to disk as they were promised, and portals
 482                  * needs to keep the pages all aligned properly. */
 483                 lnb->dentry = (void *) obj;
 484
 485                 lnb->page = osd_get_page(d, lnb->lnb_file_offset, rw);
 486                 if (lnb->page == NULL)
 487                         GOTO(cleanup, rc = -ENOMEM);
 488
 489                 /* DLM locking protects us from write and truncate competing
 490                  * for same region, but truncate can leave dirty page in the
 491                  * cache. it's possible the writeout on a such a page is in
 492                  * progress when we access it. it's also possible that during
 493                  * this writeout we put new (partial) data, but then won't
 494                  * be able to proceed in filter_commitrw_write(). thus let's
 495                  * just wait for writeout completion, should be rare enough.
 496                  * -bzzz */
 497                 wait_on_page_writeback(lnb->page);
 498                 BUG_ON(PageWriteback(lnb->page));
 499
 500                 lu_object_get(&d->do_lu);
 501         }
 502         rc = i;
 503
 504 cleanup:
 505         RETURN(rc);
 506 }
 507
 508 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 509                         struct niobuf_local *lnb, int npages)
 510 {
 511         struct osd_thread_info *oti = osd_oti_get(env);
 512         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 513         struct osd_device      *d = osd_obj2dev(osd_dt_obj(dt));
 514         int                     i;
 515
 516         /* to do IO stats, notice we do this here because
 517          * osd_do_bio() doesn't wait for write to complete */
 518         osd_fini_iobuf(d, iobuf);
 519
 520         for (i = 0; i < npages; i++) {
 521                 if (lnb[i].page == NULL)
 522                         continue;
 523                 LASSERT(PageLocked(lnb[i].page));
 524                 unlock_page(lnb[i].page);
 525                 page_cache_release(lnb[i].page);
 526                 lu_object_put(env, &dt->do_lu);
 527                 lnb[i].page = NULL;
 528         }
 529         RETURN(0);
 530 }
 531
 532 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 533                           struct niobuf_local *lnb, int npages)
 534 {
 535         struct osd_thread_info *oti   = osd_oti_get(env);
 536         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 537         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 538         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 539         struct timeval          start;
 540         struct timeval          end;
 541         unsigned long           timediff;
 542         ssize_t                 isize;
 543         __s64                   maxidx;
 544         int                     rc = 0;
 545         int                     i;
 546         int                     cache = 0;
 547
 548         LASSERT(inode);
 549
 550         osd_init_iobuf(osd, iobuf, 0);
 551
 552         isize = i_size_read(inode);
 553         maxidx = ((isize + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 554
 555         if (osd->od_writethrough_cache)
 556                 cache = 1;
 557         if (isize > osd->od_readcache_max_filesize)
 558                 cache = 0;
 559
 560         cfs_gettimeofday(&start);
 561         for (i = 0; i < npages; i++) {
 562
 563                 if (cache == 0)
 564                         generic_error_remove_page(inode->i_mapping,
 565                                                   lnb[i].page);
 566
 567                 /*
 568                  * till commit the content of the page is undefined
 569                  * we'll set it uptodate once bulk is done. otherwise
 570                  * subsequent reads can access non-stable data
 571                  */
 572                 ClearPageUptodate(lnb[i].page);
 573
 574                 if (lnb[i].len == CFS_PAGE_SIZE)
 575                         continue;
 576
 577                 if (maxidx >= lnb[i].page->index) {
 578                         osd_iobuf_add_page(iobuf, lnb[i].page);
 579                 } else {
 580                         long off;
 581                         char *p = kmap(lnb[i].page);
 582
 583                         off = lnb[i].lnb_page_offset;
 584                         if (off)
 585                                 memset(p, 0, off);
 586                         off = (lnb[i].lnb_page_offset + lnb[i].len) &
 587                               ~CFS_PAGE_MASK;
 588                         if (off)
 589                                 memset(p + off, 0, CFS_PAGE_SIZE - off);
 590                         kunmap(lnb[i].page);
 591                 }
 592         }
 593         cfs_gettimeofday(&end);
 594         timediff = cfs_timeval_sub(&end, &start, NULL);
 595         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 596
 597         if (iobuf->dr_npages) {
 598                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 599                                                        iobuf->dr_npages,
 600                                                        iobuf->dr_blocks,
 601                                                        oti->oti_created,
 602                                                        0, NULL);
 603                 if (likely(rc == 0)) {
 604                         rc = osd_do_bio(osd, inode, iobuf);
 605                         /* do IO stats for preparation reads */
 606                         osd_fini_iobuf(osd, iobuf);
 607                 }
 608         }
 609         RETURN(rc);
 610 }
 611
 612 /* Check if a block is allocated or not */
 613 static int osd_is_mapped(struct inode *inode, obd_size offset)
 614 {
 615         sector_t (*fs_bmap)(struct address_space *, sector_t);
 616
 617         fs_bmap = inode->i_mapping->a_ops->bmap;
 618
 619         /* We can't know if we are overwriting or not */
 620         if (unlikely(fs_bmap == NULL))
 621                 return 0;
 622
 623         if (i_size_read(inode) == 0)
 624                 return 0;
 625
 626         /* Beyond EOF, must not be mapped */
 627         if (((i_size_read(inode) - 1) >> inode->i_blkbits) <
 628             (offset >> inode->i_blkbits))
 629                 return 0;
 630
 631         if (fs_bmap(inode->i_mapping, offset >> inode->i_blkbits) == 0)
 632                 return 0;
 633
 634         return 1;
 635 }
 636
 637 static int osd_declare_write_commit(const struct lu_env *env,
 638                                     struct dt_object *dt,
 639                                     struct niobuf_local *lnb, int npages,
 640                                     struct thandle *handle)
 641 {
 642         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 643         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
 644         struct osd_thandle      *oh;
 645         int                      extents = 1;
 646         int                      depth;
 647         int                      i;
 648         int                      newblocks;
 649         int                      rc = 0;
 650         int                      flags = 0;
 651         bool                     ignore_quota = false;
 652         long long                quota_space = 0;
 653         ENTRY;
 654
 655         LASSERT(handle != NULL);
 656         oh = container_of0(handle, struct osd_thandle, ot_super);
 657         LASSERT(oh->ot_handle == NULL);
 658
 659         newblocks = npages;
 660
 661         /* calculate number of extents (probably better to pass nb) */
 662         for (i = 0; i < npages; i++) {
 663                 if (i && lnb[i].lnb_file_offset !=
 664                     lnb[i - 1].lnb_file_offset + lnb[i - 1].len)
 665                         extents++;
 666
 667                 if (!osd_is_mapped(inode, lnb[i].lnb_file_offset))
 668                         quota_space += CFS_PAGE_SIZE;
 669
 670                 /* ignore quota for the whole request if any page is from
 671                  * client cache or written by root.
 672                  *
 673                  * XXX we could handle this on per-lnb basis as done by
 674                  * grant. */
 675                 if ((lnb[i].flags & OBD_BRW_NOQUOTA) ||
 676                     !(lnb[i].flags & OBD_BRW_SYNC))
 677                         ignore_quota = true;
 678         }
 679
 680         /*
 681          * each extent can go into new leaf causing a split
 682          * 5 is max tree depth: inode + 4 index blocks
 683          * with blockmaps, depth is 3 at most
 684          */
 685         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 686                 /*
 687                  * many concurrent threads may grow tree by the time
 688                  * our transaction starts. so, consider 2 is a min depth
 689                  */
 690                 depth = ext_depth(inode);
 691                 depth = max(depth, 1) + 1;
 692                 newblocks += depth;
 693                 oh->ot_credits++; /* inode */
 694                 oh->ot_credits += depth * 2 * extents;
 695         } else {
 696                 depth = 3;
 697                 newblocks += depth;
 698                 oh->ot_credits++; /* inode */
 699                 oh->ot_credits += depth * extents;
 700         }
 701
 702         /* quota space for metadata blocks */
 703         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
 704
 705         /* quota space should be reported in 1K blocks */
 706         quota_space = toqb(quota_space);
 707
 708         /* each new block can go in different group (bitmap + gd) */
 709
 710         /* we can't dirty more bitmap blocks than exist */
 711         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
 712                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
 713         else
 714                 oh->ot_credits += newblocks;
 715
 716         /* we can't dirty more gd blocks than exist */
 717         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
 718                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
 719         else
 720                 oh->ot_credits += newblocks;
 721
 722         /* make sure the over quota flags were not set */
 723         lnb[0].flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA);
 724
 725         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid,
 726                                    quota_space, oh, true, true, &flags,
 727                                    ignore_quota);
 728
 729         /* we need only to store the overquota flags in the first lnb for
 730          * now, once we support multiple objects BRW, this code needs be
 731          * revised. */
 732         if (flags & QUOTA_FL_OVER_USRQUOTA)
 733                 lnb[0].flags |= OBD_BRW_OVER_USRQUOTA;
 734         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 735                 lnb[0].flags |= OBD_BRW_OVER_GRPQUOTA;
 736
 737         RETURN(rc);
 738 }
 739
 740 /* Check if a block is allocated or not */
 741 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 742                             struct niobuf_local *lnb, int npages,
 743                             struct thandle *thandle)
 744 {
 745         struct osd_thread_info *oti = osd_oti_get(env);
 746         struct osd_iobuf *iobuf = &oti->oti_iobuf;
 747         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 748         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
 749         loff_t isize;
 750         int rc = 0, i;
 751
 752         LASSERT(inode);
 753
 754         osd_init_iobuf(osd, iobuf, 1);
 755         isize = i_size_read(inode);
 756         ll_vfs_dq_init(inode);
 757
 758         for (i = 0; i < npages; i++) {
 759                 if (lnb[i].rc == -ENOSPC &&
 760                     osd_is_mapped(inode, lnb[i].lnb_file_offset)) {
 761                         /* Allow the write to proceed if overwriting an
 762                          * existing block */
 763                         lnb[i].rc = 0;
 764                 }
 765
 766                 if (lnb[i].rc) { /* ENOSPC, network RPC error, etc. */
 767                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
 768                                lnb[i].rc);
 769                         LASSERT(lnb[i].page);
 770                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 771                         continue;
 772                 }
 773
 774                 LASSERT(PageLocked(lnb[i].page));
 775                 LASSERT(!PageWriteback(lnb[i].page));
 776
 777                 if (lnb[i].lnb_file_offset + lnb[i].len > isize)
 778                         isize = lnb[i].lnb_file_offset + lnb[i].len;
 779
 780                 /*
 781                  * Since write and truncate are serialized by oo_sem, even
 782                  * partial-page truncate should not leave dirty pages in the
 783                  * page cache.
 784                  */
 785                 LASSERT(!PageDirty(lnb[i].page));
 786
 787                 SetPageUptodate(lnb[i].page);
 788
 789                 osd_iobuf_add_page(iobuf, lnb[i].page);
 790         }
 791
 792         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
 793                 rc = -ENOSPC;
 794         } else if (iobuf->dr_npages > 0) {
 795                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 796                                                        iobuf->dr_npages,
 797                                                        iobuf->dr_blocks,
 798                                                        oti->oti_created,
 799                                                        1, NULL);
 800         } else {
 801                 /* no pages to write, no transno is needed */
 802                 thandle->th_local = 1;
 803         }
 804
 805         if (likely(rc == 0)) {
 806                 if (isize > i_size_read(inode)) {
 807                         i_size_write(inode, isize);
 808                         LDISKFS_I(inode)->i_disksize = isize;
 809                         inode->i_sb->s_op->dirty_inode(inode);
 810                 }
 811
 812                 rc = osd_do_bio(osd, inode, iobuf);
 813                 /* we don't do stats here as in read path because
 814                  * write is async: we'll do this in osd_put_bufs() */
 815         }
 816
 817         if (unlikely(rc != 0)) {
 818                 /* if write fails, we should drop pages from the cache */
 819                 for (i = 0; i < npages; i++) {
 820                         if (lnb[i].page == NULL)
 821                                 continue;
 822                         LASSERT(PageLocked(lnb[i].page));
 823                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 824                 }
 825         }
 826
 827         RETURN(rc);
 828 }
 829
 830 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 831                          struct niobuf_local *lnb, int npages)
 832 {
 833         struct osd_thread_info *oti = osd_oti_get(env);
 834         struct osd_iobuf *iobuf = &oti->oti_iobuf;
 835         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 836         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 837         struct timeval start, end;
 838         unsigned long timediff;
 839         int rc = 0, i, m = 0, cache = 0;
 840
 841         LASSERT(inode);
 842
 843         osd_init_iobuf(osd, iobuf, 0);
 844
 845         if (osd->od_read_cache)
 846                 cache = 1;
 847         if (i_size_read(inode) > osd->od_readcache_max_filesize)
 848                 cache = 0;
 849
 850         cfs_gettimeofday(&start);
 851         for (i = 0; i < npages; i++) {
 852
 853                 if (i_size_read(inode) <= lnb[i].lnb_file_offset)
 854                         /* If there's no more data, abort early.
 855                          * lnb->rc == 0, so it's easy to detect later. */
 856                         break;
 857
 858                 if (i_size_read(inode) <
 859                     lnb[i].lnb_file_offset + lnb[i].len - 1)
 860                         lnb[i].rc = i_size_read(inode) - lnb[i].lnb_file_offset;
 861                 else
 862                         lnb[i].rc = lnb[i].len;
 863                 m += lnb[i].len;
 864
 865                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS, 1);
 866                 if (PageUptodate(lnb[i].page)) {
 867                         lprocfs_counter_add(osd->od_stats,
 868                                             LPROC_OSD_CACHE_HIT, 1);
 869                 } else {
 870                         lprocfs_counter_add(osd->od_stats,
 871                                             LPROC_OSD_CACHE_MISS, 1);
 872                         osd_iobuf_add_page(iobuf, lnb[i].page);
 873                 }
 874                 if (cache == 0)
 875                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 876         }
 877         cfs_gettimeofday(&end);
 878         timediff = cfs_timeval_sub(&end, &start, NULL);
 879         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 880
 881         if (iobuf->dr_npages) {
 882                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 883                                                        iobuf->dr_npages,
 884                                                        iobuf->dr_blocks,
 885                                                        oti->oti_created,
 886                                                        0, NULL);
 887                 rc = osd_do_bio(osd, inode, iobuf);
 888
 889                 /* IO stats will be done in osd_bufs_put() */
 890         }
 891
 892         RETURN(rc);
 893 }
 894
 895 /*
 896  * XXX: Another layering violation for now.
 897  *
 898  * We don't want to use ->f_op->read methods, because generic file write
 899  *
 900  *         - serializes on ->i_sem, and
 901  *
 902  *         - does a lot of extra work like balance_dirty_pages(),
 903  *
 904  * which doesn't work for globally shared files like /last_rcvd.
 905  */
 906 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
 907 {
 908         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 909
 910         memcpy(buffer, (char *)ei->i_data, buflen);
 911
 912         return  buflen;
 913 }
 914
 915 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
 916 {
 917         struct buffer_head *bh;
 918         unsigned long block;
 919         int osize;
 920         int blocksize;
 921         int csize;
 922         int boffs;
 923         int err;
 924
 925         /* prevent reading after eof */
 926         cfs_spin_lock(&inode->i_lock);
 927         if (i_size_read(inode) < *offs + size) {
 928                 loff_t diff = i_size_read(inode) - *offs;
 929                 cfs_spin_unlock(&inode->i_lock);
 930                 if (diff < 0) {
 931                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
 932                                i_size_read(inode), *offs);
 933                         return -EBADR;
 934                 } else if (diff == 0) {
 935                         return 0;
 936                 } else {
 937                         size = diff;
 938                 }
 939         } else {
 940                 cfs_spin_unlock(&inode->i_lock);
 941         }
 942
 943         blocksize = 1 << inode->i_blkbits;
 944         osize = size;
 945         while (size > 0) {
 946                 block = *offs >> inode->i_blkbits;
 947                 boffs = *offs & (blocksize - 1);
 948                 csize = min(blocksize - boffs, size);
 949                 bh = ldiskfs_bread(NULL, inode, block, 0, &err);
 950                 if (!bh) {
 951                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %d\n",
 952                                LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
 953                                csize, *offs, inode->i_ino, err);
 954                         return err;
 955                 }
 956
 957                 memcpy(buf, bh->b_data + boffs, csize);
 958                 brelse(bh);
 959
 960                 *offs += csize;
 961                 buf += csize;
 962                 size -= csize;
 963         }
 964         return osize;
 965 }
 966
 967 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 968                         struct lu_buf *buf, loff_t *pos,
 969                         struct lustre_capa *capa)
 970 {
 971         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 972         int           rc;
 973
 974         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_READ))
 975                 RETURN(-EACCES);
 976
 977         /* Read small symlink from inode body as we need to maintain correct
 978          * on-disk symlinks for ldiskfs.
 979          */
 980         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
 981             (buf->lb_len <= sizeof(LDISKFS_I(inode)->i_data)))
 982                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
 983         else
 984                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
 985
 986         return rc;
 987 }
 988
 989 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 990                                  const loff_t size, loff_t pos,
 991                                  struct thandle *handle)
 992 {
 993         struct osd_thandle *oh;
 994         int                 credits;
 995         struct inode       *inode;
 996         int                 rc;
 997         ENTRY;
 998
 999         LASSERT(handle != NULL);
1000
1001         oh = container_of0(handle, struct osd_thandle, ot_super);
1002         LASSERT(oh->ot_handle == NULL);
1003
1004         /* XXX: size == 0 or INT_MAX indicating a catalog header update or
1005          *      llog write, see comment in mdd_declare_llog_record().
1006          *
1007          *      This hack will be removed with llog over OSD landing
1008          */
1009         if (size == DECLARE_LLOG_REWRITE)
1010                 credits = 2;
1011         else if (size == DECLARE_LLOG_WRITE)
1012                 credits = 6;
1013         else
1014                 credits = osd_dto_credits_noquota[DTO_WRITE_BLOCK];
1015
1016         OSD_DECLARE_OP(oh, write, credits);
1017
1018         inode = osd_dt_obj(dt)->oo_inode;
1019
1020         /* we may declare write to non-exist llog */
1021         if (inode == NULL)
1022                 RETURN(0);
1023
1024         /* dt_declare_write() is usually called for system objects, such
1025          * as llog or last_rcvd files. We needn't enforce quota on those
1026          * objects, so always set the lqi_space as 0. */
1027         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
1028                                    true, true, NULL, false);
1029         RETURN(rc);
1030 }
1031
1032 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1033 {
1034
1035         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1036         LDISKFS_I(inode)->i_disksize = buflen;
1037         i_size_write(inode, buflen);
1038         inode->i_sb->s_op->dirty_inode(inode);
1039
1040         return 0;
1041 }
1042
1043 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1044                              int write_NUL, loff_t *offs, handle_t *handle)
1045 {
1046         struct buffer_head *bh        = NULL;
1047         loff_t              offset    = *offs;
1048         loff_t              new_size  = i_size_read(inode);
1049         unsigned long       block;
1050         int                 blocksize = 1 << inode->i_blkbits;
1051         int                 err = 0;
1052         int                 size;
1053         int                 boffs;
1054         int                 dirty_inode = 0;
1055
1056         if (write_NUL) {
1057                 /*
1058                  * long symlink write does not count the NUL terminator in
1059                  * bufsize, we write it, and the inode's file size does not
1060                  * count the NUL terminator as well.
1061                  */
1062                 ((char *)buf)[bufsize] = '\0';
1063                 ++bufsize;
1064         }
1065         while (bufsize > 0) {
1066                 if (bh != NULL)
1067                         brelse(bh);
1068
1069                 block = offset >> inode->i_blkbits;
1070                 boffs = offset & (blocksize - 1);
1071                 size = min(blocksize - boffs, bufsize);
1072                 bh = ldiskfs_bread(handle, inode, block, 1, &err);
1073                 if (!bh) {
1074                         CERROR("%s: error reading offset %llu (block %lu): "
1075                                "rc = %d\n",
1076                                inode->i_sb->s_id, offset, block, err);
1077                         break;
1078                 }
1079
1080                 err = ldiskfs_journal_get_write_access(handle, bh);
1081                 if (err) {
1082                         CERROR("journal_get_write_access() returned error %d\n",
1083                                err);
1084                         break;
1085                 }
1086                 LASSERTF(boffs + size <= bh->b_size,
1087                          "boffs %d size %d bh->b_size %lu",
1088                          boffs, size, (unsigned long)bh->b_size);
1089                 memcpy(bh->b_data + boffs, buf, size);
1090                 err = ldiskfs_journal_dirty_metadata(handle, bh);
1091                 if (err)
1092                         break;
1093
1094                 if (offset + size > new_size)
1095                         new_size = offset + size;
1096                 offset += size;
1097                 bufsize -= size;
1098                 buf += size;
1099         }
1100         if (bh)
1101                 brelse(bh);
1102
1103         if (write_NUL)
1104                 --new_size;
1105         /* correct in-core and on-disk sizes */
1106         if (new_size > i_size_read(inode)) {
1107                 cfs_spin_lock(&inode->i_lock);
1108                 if (new_size > i_size_read(inode))
1109                         i_size_write(inode, new_size);
1110                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1111                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1112                         dirty_inode = 1;
1113                 }
1114                 cfs_spin_unlock(&inode->i_lock);
1115                 if (dirty_inode)
1116                         inode->i_sb->s_op->dirty_inode(inode);
1117         }
1118
1119         if (err == 0)
1120                 *offs = offset;
1121         return err;
1122 }
1123
1124 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1125                          const struct lu_buf *buf, loff_t *pos,
1126                          struct thandle *handle, struct lustre_capa *capa,
1127                          int ignore_quota)
1128 {
1129         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1130         struct osd_thandle      *oh;
1131         ssize_t                 result;
1132         int                     is_link;
1133
1134         LASSERT(dt_object_exists(dt));
1135
1136         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_WRITE))
1137                 return -EACCES;
1138
1139         LASSERT(handle != NULL);
1140         LASSERT(inode != NULL);
1141         ll_vfs_dq_init(inode);
1142
1143         /* XXX: don't check: one declared chunk can be used many times */
1144         /* OSD_EXEC_OP(handle, write); */
1145
1146         oh = container_of(handle, struct osd_thandle, ot_super);
1147         LASSERT(oh->ot_handle->h_transaction != NULL);
1148         /* Write small symlink to inode body as we need to maintain correct
1149          * on-disk symlinks for ldiskfs.
1150          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1151          * does not count it in.
1152          */
1153         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1154         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1155                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1156         else
1157                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1158                                                   buf->lb_len, is_link, pos,
1159                                                   oh->ot_handle);
1160         if (result == 0)
1161                 result = buf->lb_len;
1162         return result;
1163 }
1164
1165 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1166                              __u64 start, __u64 end, struct thandle *th)
1167 {
1168         struct osd_thandle *oh;
1169         struct inode       *inode;
1170         int                 rc;
1171         ENTRY;
1172
1173         LASSERT(th);
1174         oh = container_of(th, struct osd_thandle, ot_super);
1175
1176         /*
1177          * we don't need to reserve credits for whole truncate
1178          * it's not possible as truncate may need to free too many
1179          * blocks and that won't fit a single transaction. instead
1180          * we reserve credits to change i_size and put inode onto
1181          * orphan list. if needed truncate will extend or restart
1182          * transaction
1183          */
1184         OSD_DECLARE_OP(oh, punch,
1185                        osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1186
1187         inode = osd_dt_obj(dt)->oo_inode;
1188         LASSERT(inode);
1189
1190         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
1191                                    true, true, NULL, false);
1192         RETURN(rc);
1193 }
1194
1195 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1196                      __u64 start, __u64 end, struct thandle *th,
1197                      struct lustre_capa *capa)
1198 {
1199         struct osd_thandle *oh;
1200         struct osd_object  *obj = osd_dt_obj(dt);
1201         struct inode       *inode = obj->oo_inode;
1202         handle_t           *h;
1203         tid_t               tid;
1204         int                 rc, rc2 = 0;
1205         ENTRY;
1206
1207         LASSERT(end == OBD_OBJECT_EOF);
1208         LASSERT(dt_object_exists(dt));
1209         LASSERT(osd_invariant(obj));
1210         LASSERT(inode != NULL);
1211         ll_vfs_dq_init(inode);
1212
1213         LASSERT(th);
1214         oh = container_of(th, struct osd_thandle, ot_super);
1215         LASSERT(oh->ot_handle->h_transaction != NULL);
1216
1217         OSD_EXEC_OP(th, punch);
1218
1219         tid = oh->ot_handle->h_transaction->t_tid;
1220
1221         rc = vmtruncate(inode, start);
1222
1223         /*
1224          * For a partial-page truncate, flush the page to disk immediately to
1225          * avoid data corruption during direct disk write.  b=17397
1226          */
1227         if (rc == 0 && (start & ~CFS_PAGE_MASK) != 0)
1228                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1229
1230         h = journal_current_handle();
1231         LASSERT(h != NULL);
1232         LASSERT(h == oh->ot_handle);
1233
1234         if (tid != h->h_transaction->t_tid) {
1235                 int credits = oh->ot_credits;
1236                 /*
1237                  * transaction has changed during truncate
1238                  * we need to restart the handle with our credits
1239                  */
1240                 if (h->h_buffer_credits < credits) {
1241                         if (ldiskfs_journal_extend(h, credits))
1242                                 rc2 = ldiskfs_journal_restart(h, credits);
1243                 }
1244         }
1245
1246         RETURN(rc == 0 ? rc2 : rc);
1247 }
1248
1249 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1250                           struct ll_user_fiemap *fm)
1251 {
1252         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1253         struct osd_thread_info *info   = osd_oti_get(env);
1254         struct dentry          *dentry = &info->oti_obj_dentry;
1255         struct file            *file   = &info->oti_file;
1256         mm_segment_t            saved_fs;
1257         int rc;
1258
1259         LASSERT(inode);
1260         dentry->d_inode = inode;
1261         file->f_dentry = dentry;
1262         file->f_mapping = inode->i_mapping;
1263         file->f_op = inode->i_fop;
1264
1265         saved_fs = get_fs();
1266         set_fs(get_ds());
1267         /* ldiskfs_ioctl does not have a inode argument */
1268         if (inode->i_fop->unlocked_ioctl)
1269                 rc = inode->i_fop->unlocked_ioctl(file, FSFILT_IOC_FIEMAP,
1270                                                   (long)fm);
1271         else
1272                 rc = -ENOTTY;
1273         set_fs(saved_fs);
1274         return rc;
1275 }
1276
1277 /*
1278  * in some cases we may need declare methods for objects being created
1279  * e.g., when we create symlink
1280  */
1281 const struct dt_body_operations osd_body_ops_new = {
1282         .dbo_declare_write = osd_declare_write,
1283 };
1284
1285 const struct dt_body_operations osd_body_ops = {
1286         .dbo_read                 = osd_read,
1287         .dbo_declare_write        = osd_declare_write,
1288         .dbo_write                = osd_write,
1289         .dbo_bufs_get             = osd_bufs_get,
1290         .dbo_bufs_put             = osd_bufs_put,
1291         .dbo_write_prep           = osd_write_prep,
1292         .dbo_declare_write_commit = osd_declare_write_commit,
1293         .dbo_write_commit         = osd_write_commit,
1294         .dbo_read_prep            = osd_read_prep,
1295         .do_declare_punch         = osd_declare_punch,
1296         .do_punch                 = osd_punch,
1297         .dbo_fiemap_get           = osd_fiemap_get,
1298 };
1299