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) 2012, 2013, Intel Corporation.
  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 /*
  53  * struct OBD_{ALLOC,FREE}*()
  54  * OBD_FAIL_CHECK
  55  */
  56 #include <obd_support.h>
  57
  58 #include "osd_internal.h"
  59
  60 /* ext_depth() */
  61 #include <ldiskfs/ldiskfs_extents.h>
  62
  63 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  64                             int rw, int line, int pages)
  65 {
  66         int blocks, i;
  67
  68         LASSERTF(iobuf->dr_elapsed_valid == 0,
  69                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  70                  cfs_atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  71                  iobuf->dr_init_at);
  72         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  73
  74         init_waitqueue_head(&iobuf->dr_wait);
  75         cfs_atomic_set(&iobuf->dr_numreqs, 0);
  76         iobuf->dr_npages = 0;
  77         iobuf->dr_error = 0;
  78         iobuf->dr_dev = d;
  79         iobuf->dr_frags = 0;
  80         iobuf->dr_elapsed = 0;
  81         /* must be counted before, so assert */
  82         iobuf->dr_rw = rw;
  83         iobuf->dr_init_at = line;
  84
  85         blocks = pages * (PAGE_CACHE_SIZE >> osd_sb(d)->s_blocksize_bits);
  86         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  87                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  88                         pages * sizeof(iobuf->dr_pages[0]));
  89                 return 0;
  90         }
  91
  92         /* start with 1MB for 4K blocks */
  93         i = 256;
  94         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
  95                 i <<= 1;
  96
  97         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
  98                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
  99         pages = i;
 100         blocks = pages * (PAGE_CACHE_SIZE >> osd_sb(d)->s_blocksize_bits);
 101         iobuf->dr_max_pages = 0;
 102         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
 103                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 104
 105         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 106         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 107         if (unlikely(iobuf->dr_blocks == NULL))
 108                 return -ENOMEM;
 109
 110         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 111         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 112         if (unlikely(iobuf->dr_pages == NULL))
 113                 return -ENOMEM;
 114
 115         iobuf->dr_max_pages = pages;
 116
 117         return 0;
 118 }
 119 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 120         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 121
 122 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
 123 {
 124         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 125         iobuf->dr_pages[iobuf->dr_npages++] = page;
 126 }
 127
 128 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 129 {
 130         int rw = iobuf->dr_rw;
 131
 132         if (iobuf->dr_elapsed_valid) {
 133                 iobuf->dr_elapsed_valid = 0;
 134                 LASSERT(iobuf->dr_dev == d);
 135                 LASSERT(iobuf->dr_frags > 0);
 136                 lprocfs_oh_tally(&d->od_brw_stats.
 137                                  hist[BRW_R_DIO_FRAGS+rw],
 138                                  iobuf->dr_frags);
 139                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 140                                       iobuf->dr_elapsed);
 141         }
 142 }
 143
 144 #ifndef REQ_WRITE /* pre-2.6.35 */
 145 #define __REQ_WRITE BIO_RW
 146 #endif
 147
 148 #ifdef HAVE_BIO_ENDIO_2ARG
 149 #define DIO_RETURN(a)
 150 static void dio_complete_routine(struct bio *bio, int error)
 151 #else
 152 #define DIO_RETURN(a)   return(a)
 153 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 154 #endif
 155 {
 156         struct osd_iobuf *iobuf = bio->bi_private;
 157         struct bio_vec *bvl;
 158         int i;
 159
 160         /* CAVEAT EMPTOR: possibly in IRQ context
 161          * DO NOT record procfs stats here!!! */
 162
 163         if (unlikely(iobuf == NULL)) {
 164                 CERROR("***** bio->bi_private is NULL!  This should never "
 165                        "happen.  Normally, I would crash here, but instead I "
 166                        "will dump the bio contents to the console.  Please "
 167                        "report this to <http://jira.whamcloud.com/> , along "
 168                        "with any interesting messages leading up to this point "
 169                        "(like SCSI errors, perhaps).  Because bi_private is "
 170                        "NULL, I can't wake up the thread that initiated this "
 171                        "IO - you will probably have to reboot this node.\n");
 172                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 173                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 174                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 175                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 176                        bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
 177                        bio->bi_private);
 178                 DIO_RETURN(0);
 179         }
 180
 181         /* the check is outside of the cycle for performance reason -bzzz */
 182         if (!test_bit(__REQ_WRITE, &bio->bi_rw)) {
 183                 bio_for_each_segment(bvl, bio, i) {
 184                         if (likely(error == 0))
 185                                 SetPageUptodate(bvl->bv_page);
 186                         LASSERT(PageLocked(bvl->bv_page));
 187                 }
 188                 cfs_atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 189         } else {
 190                 cfs_atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 191         }
 192
 193         /* any real error is good enough -bzzz */
 194         if (error != 0 && iobuf->dr_error == 0)
 195                 iobuf->dr_error = error;
 196
 197         /*
 198          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 199          * it's possible that service thread will see dr_numreqs
 200          * is zero, but dr_elapsed is not set yet, leading to lost
 201          * data in this processing and an assertion in a subsequent
 202          * call to OSD.
 203          */
 204         if (cfs_atomic_read(&iobuf->dr_numreqs) == 1) {
 205                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
 206                 iobuf->dr_elapsed_valid = 1;
 207         }
 208         if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs))
 209                 wake_up(&iobuf->dr_wait);
 210
 211         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 212          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 213          * mempool when serious on-disk fragmentation was encountered,
 214          * deadlocking the OST.  The bios are now released as soon as complete
 215          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 216         bio_put(bio);
 217         DIO_RETURN(0);
 218 }
 219
 220 static void record_start_io(struct osd_iobuf *iobuf, int size)
 221 {
 222         struct osd_device    *osd = iobuf->dr_dev;
 223         struct obd_histogram *h = osd->od_brw_stats.hist;
 224
 225         iobuf->dr_frags++;
 226         cfs_atomic_inc(&iobuf->dr_numreqs);
 227
 228         if (iobuf->dr_rw == 0) {
 229                 cfs_atomic_inc(&osd->od_r_in_flight);
 230                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 231                                  cfs_atomic_read(&osd->od_r_in_flight));
 232                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 233         } else if (iobuf->dr_rw == 1) {
 234                 cfs_atomic_inc(&osd->od_w_in_flight);
 235                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 236                                  cfs_atomic_read(&osd->od_w_in_flight));
 237                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 238         } else {
 239                 LBUG();
 240         }
 241 }
 242
 243 static void osd_submit_bio(int rw, struct bio *bio)
 244 {
 245         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 246         if (rw == 0)
 247                 submit_bio(READ, bio);
 248         else
 249                 submit_bio(WRITE, bio);
 250 }
 251
 252 static int can_be_merged(struct bio *bio, sector_t sector)
 253 {
 254         unsigned int size;
 255
 256         if (!bio)
 257                 return 0;
 258
 259         size = bio->bi_size >> 9;
 260         return bio->bi_sector + size == sector ? 1 : 0;
 261 }
 262
 263 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 264                       struct osd_iobuf *iobuf)
 265 {
 266         int            blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
 267         struct page  **pages = iobuf->dr_pages;
 268         int            npages = iobuf->dr_npages;
 269         unsigned long *blocks = iobuf->dr_blocks;
 270         int            total_blocks = npages * blocks_per_page;
 271         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 272         unsigned int   blocksize = inode->i_sb->s_blocksize;
 273         struct bio    *bio = NULL;
 274         struct page   *page;
 275         unsigned int   page_offset;
 276         sector_t       sector;
 277         int            nblocks;
 278         int            block_idx;
 279         int            page_idx;
 280         int            i;
 281         int            rc = 0;
 282         ENTRY;
 283
 284         LASSERT(iobuf->dr_npages == npages);
 285
 286         osd_brw_stats_update(osd, iobuf);
 287         iobuf->dr_start_time = cfs_time_current();
 288
 289         for (page_idx = 0, block_idx = 0;
 290              page_idx < npages;
 291              page_idx++, block_idx += blocks_per_page) {
 292
 293                 page = pages[page_idx];
 294                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 295
 296                 for (i = 0, page_offset = 0;
 297                      i < blocks_per_page;
 298                      i += nblocks, page_offset += blocksize * nblocks) {
 299
 300                         nblocks = 1;
 301
 302                         if (blocks[block_idx + i] == 0) {  /* hole */
 303                                 LASSERTF(iobuf->dr_rw == 0,
 304                                          "page_idx %u, block_idx %u, i %u\n",
 305                                          page_idx, block_idx, i);
 306                                 memset(kmap(page) + page_offset, 0, blocksize);
 307                                 kunmap(page);
 308                                 continue;
 309                         }
 310
 311                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 312
 313                         /* Additional contiguous file blocks? */
 314                         while (i + nblocks < blocks_per_page &&
 315                                (sector + (nblocks << sector_bits)) ==
 316                                ((sector_t)blocks[block_idx + i + nblocks] <<
 317                                 sector_bits))
 318                                 nblocks++;
 319
 320                         if (bio != NULL &&
 321                             can_be_merged(bio, sector) &&
 322                             bio_add_page(bio, page,
 323                                          blocksize * nblocks, page_offset) != 0)
 324                                 continue;       /* added this frag OK */
 325
 326                         if (bio != NULL) {
 327                                 struct request_queue *q =
 328                                         bdev_get_queue(bio->bi_bdev);
 329
 330                                 /* Dang! I have to fragment this I/O */
 331                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 332                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 333                                        bio->bi_size,
 334                                        bio->bi_vcnt, bio->bi_max_vecs,
 335                                        bio->bi_size >> 9, queue_max_sectors(q),
 336                                        bio_phys_segments(q, bio),
 337                                        queue_max_phys_segments(q),
 338                                        0, queue_max_hw_segments(q));
 339
 340                                 record_start_io(iobuf, bio->bi_size);
 341                                 osd_submit_bio(iobuf->dr_rw, bio);
 342                         }
 343
 344                         /* allocate new bio */
 345                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 346                                                       (npages - page_idx) *
 347                                                       blocks_per_page));
 348                         if (bio == NULL) {
 349                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 350                                        (npages - page_idx), blocks_per_page,
 351                                        (npages - page_idx) * blocks_per_page);
 352                                 rc = -ENOMEM;
 353                                 goto out;
 354                         }
 355
 356                         bio->bi_bdev = inode->i_sb->s_bdev;
 357                         bio->bi_sector = sector;
 358                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 359                         bio->bi_end_io = dio_complete_routine;
 360                         bio->bi_private = iobuf;
 361
 362                         rc = bio_add_page(bio, page,
 363                                           blocksize * nblocks, page_offset);
 364                         LASSERT(rc != 0);
 365                 }
 366         }
 367
 368         if (bio != NULL) {
 369                 record_start_io(iobuf, bio->bi_size);
 370                 osd_submit_bio(iobuf->dr_rw, bio);
 371                 rc = 0;
 372         }
 373
 374  out:
 375         /* in order to achieve better IO throughput, we don't wait for writes
 376          * completion here. instead we proceed with transaction commit in
 377          * parallel and wait for IO completion once transaction is stopped
 378          * see osd_trans_stop() for more details -bzzz */
 379         if (iobuf->dr_rw == 0) {
 380                 wait_event(iobuf->dr_wait,
 381                                cfs_atomic_read(&iobuf->dr_numreqs) == 0);
 382         }
 383
 384         if (rc == 0)
 385                 rc = iobuf->dr_error;
 386         RETURN(rc);
 387 }
 388
 389 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 390                                    struct niobuf_local *lnb)
 391 {
 392         ENTRY;
 393
 394         *nrpages = 0;
 395
 396         while (len > 0) {
 397                 int poff = offset & (PAGE_CACHE_SIZE - 1);
 398                 int plen = PAGE_CACHE_SIZE - poff;
 399
 400                 if (plen > len)
 401                         plen = len;
 402                 lnb->lnb_file_offset = offset;
 403                 lnb->lnb_page_offset = poff;
 404                 lnb->len = plen;
 405                 /* lb->flags = rnb->flags; */
 406                 lnb->flags = 0;
 407                 lnb->page = NULL;
 408                 lnb->rc = 0;
 409
 410                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 411                          (long long) len);
 412                 offset += plen;
 413                 len -= plen;
 414                 lnb++;
 415                 (*nrpages)++;
 416         }
 417
 418         RETURN(0);
 419 }
 420
 421 struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
 422 {
 423         struct inode      *inode = osd_dt_obj(dt)->oo_inode;
 424         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 425         struct page       *page;
 426
 427         LASSERT(inode);
 428
 429         page = find_or_create_page(inode->i_mapping, offset >> PAGE_CACHE_SHIFT,
 430                                    GFP_NOFS | __GFP_HIGHMEM);
 431         if (unlikely(page == NULL))
 432                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 433
 434         return page;
 435 }
 436
 437 /*
 438  * there are following "locks":
 439  * journal_start
 440  * i_mutex
 441  * page lock
 442
 443  * osd write path
 444     * lock page(s)
 445     * journal_start
 446     * truncate_sem
 447
 448  * ext4 vmtruncate:
 449     * lock pages, unlock
 450     * journal_start
 451     * lock partial page
 452     * i_data_sem
 453
 454 */
 455 int osd_bufs_get(const struct lu_env *env, struct dt_object *d, loff_t pos,
 456                  ssize_t len, struct niobuf_local *lnb, int rw,
 457                  struct lustre_capa *capa)
 458 {
 459         struct osd_object   *obj    = osd_dt_obj(d);
 460         int npages, i, rc = 0;
 461
 462         LASSERT(obj->oo_inode);
 463
 464         osd_map_remote_to_local(pos, len, &npages, lnb);
 465
 466         for (i = 0; i < npages; i++, lnb++) {
 467
 468                 /* We still set up for ungranted pages so that granted pages
 469                  * can be written to disk as they were promised, and portals
 470                  * needs to keep the pages all aligned properly. */
 471                 lnb->dentry = (void *) obj;
 472
 473                 lnb->page = osd_get_page(d, lnb->lnb_file_offset, rw);
 474                 if (lnb->page == NULL)
 475                         GOTO(cleanup, rc = -ENOMEM);
 476
 477                 /* DLM locking protects us from write and truncate competing
 478                  * for same region, but truncate can leave dirty page in the
 479                  * cache. it's possible the writeout on a such a page is in
 480                  * progress when we access it. it's also possible that during
 481                  * this writeout we put new (partial) data, but then won't
 482                  * be able to proceed in filter_commitrw_write(). thus let's
 483                  * just wait for writeout completion, should be rare enough.
 484                  * -bzzz */
 485                 wait_on_page_writeback(lnb->page);
 486                 BUG_ON(PageWriteback(lnb->page));
 487
 488                 lu_object_get(&d->do_lu);
 489         }
 490         rc = i;
 491
 492 cleanup:
 493         RETURN(rc);
 494 }
 495
 496 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 497                         struct niobuf_local *lnb, int npages)
 498 {
 499         struct osd_thread_info *oti = osd_oti_get(env);
 500         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 501         struct osd_device      *d = osd_obj2dev(osd_dt_obj(dt));
 502         int                     i;
 503
 504         /* to do IO stats, notice we do this here because
 505          * osd_do_bio() doesn't wait for write to complete */
 506         osd_fini_iobuf(d, iobuf);
 507
 508         for (i = 0; i < npages; i++) {
 509                 if (lnb[i].page == NULL)
 510                         continue;
 511                 LASSERT(PageLocked(lnb[i].page));
 512                 unlock_page(lnb[i].page);
 513                 page_cache_release(lnb[i].page);
 514                 lu_object_put(env, &dt->do_lu);
 515                 lnb[i].page = NULL;
 516         }
 517         RETURN(0);
 518 }
 519
 520 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 521                           struct niobuf_local *lnb, int npages)
 522 {
 523         struct osd_thread_info *oti   = osd_oti_get(env);
 524         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 525         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 526         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 527         struct timeval          start;
 528         struct timeval          end;
 529         unsigned long           timediff;
 530         ssize_t                 isize;
 531         __s64                   maxidx;
 532         int                     rc = 0;
 533         int                     i;
 534         int                     cache = 0;
 535
 536         LASSERT(inode);
 537
 538         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 539         if (unlikely(rc != 0))
 540                 RETURN(rc);
 541
 542         isize = i_size_read(inode);
 543         maxidx = ((isize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) - 1;
 544
 545         if (osd->od_writethrough_cache)
 546                 cache = 1;
 547         if (isize > osd->od_readcache_max_filesize)
 548                 cache = 0;
 549
 550         do_gettimeofday(&start);
 551         for (i = 0; i < npages; i++) {
 552
 553                 if (cache == 0)
 554                         generic_error_remove_page(inode->i_mapping,
 555                                                   lnb[i].page);
 556
 557                 /*
 558                  * till commit the content of the page is undefined
 559                  * we'll set it uptodate once bulk is done. otherwise
 560                  * subsequent reads can access non-stable data
 561                  */
 562                 ClearPageUptodate(lnb[i].page);
 563
 564                 if (lnb[i].len == PAGE_CACHE_SIZE)
 565                         continue;
 566
 567                 if (maxidx >= lnb[i].page->index) {
 568                         osd_iobuf_add_page(iobuf, lnb[i].page);
 569                 } else {
 570                         long off;
 571                         char *p = kmap(lnb[i].page);
 572
 573                         off = lnb[i].lnb_page_offset;
 574                         if (off)
 575                                 memset(p, 0, off);
 576                         off = (lnb[i].lnb_page_offset + lnb[i].len) &
 577                               ~CFS_PAGE_MASK;
 578                         if (off)
 579                                 memset(p + off, 0, PAGE_CACHE_SIZE - off);
 580                         kunmap(lnb[i].page);
 581                 }
 582         }
 583         do_gettimeofday(&end);
 584         timediff = cfs_timeval_sub(&end, &start, NULL);
 585         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 586
 587         if (iobuf->dr_npages) {
 588                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 589                                                        iobuf->dr_npages,
 590                                                        iobuf->dr_blocks,
 591                                                        0, NULL);
 592                 if (likely(rc == 0)) {
 593                         rc = osd_do_bio(osd, inode, iobuf);
 594                         /* do IO stats for preparation reads */
 595                         osd_fini_iobuf(osd, iobuf);
 596                 }
 597         }
 598         RETURN(rc);
 599 }
 600
 601 /* Check if a block is allocated or not */
 602 static int osd_is_mapped(struct inode *inode, obd_size offset)
 603 {
 604         sector_t (*fs_bmap)(struct address_space *, sector_t);
 605
 606         fs_bmap = inode->i_mapping->a_ops->bmap;
 607
 608         /* We can't know if we are overwriting or not */
 609         if (unlikely(fs_bmap == NULL))
 610                 return 0;
 611
 612         if (i_size_read(inode) == 0)
 613                 return 0;
 614
 615         /* Beyond EOF, must not be mapped */
 616         if (((i_size_read(inode) - 1) >> inode->i_blkbits) <
 617             (offset >> inode->i_blkbits))
 618                 return 0;
 619
 620         if (fs_bmap(inode->i_mapping, offset >> inode->i_blkbits) == 0)
 621                 return 0;
 622
 623         return 1;
 624 }
 625
 626 static int osd_declare_write_commit(const struct lu_env *env,
 627                                     struct dt_object *dt,
 628                                     struct niobuf_local *lnb, int npages,
 629                                     struct thandle *handle)
 630 {
 631         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 632         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
 633         struct osd_thandle      *oh;
 634         int                      extents = 1;
 635         int                      depth;
 636         int                      i;
 637         int                      newblocks;
 638         int                      rc = 0;
 639         int                      flags = 0;
 640         bool                     ignore_quota = false;
 641         long long                quota_space = 0;
 642         ENTRY;
 643
 644         LASSERT(handle != NULL);
 645         oh = container_of0(handle, struct osd_thandle, ot_super);
 646         LASSERT(oh->ot_handle == NULL);
 647
 648         newblocks = npages;
 649
 650         /* calculate number of extents (probably better to pass nb) */
 651         for (i = 0; i < npages; i++) {
 652                 if (i && lnb[i].lnb_file_offset !=
 653                     lnb[i - 1].lnb_file_offset + lnb[i - 1].len)
 654                         extents++;
 655
 656                 if (!osd_is_mapped(inode, lnb[i].lnb_file_offset))
 657                         quota_space += PAGE_CACHE_SIZE;
 658
 659                 /* ignore quota for the whole request if any page is from
 660                  * client cache or written by root.
 661                  *
 662                  * XXX once we drop the 1.8 client support, the checking
 663                  * for whether page is from cache can be simplified as:
 664                  * !(lnb[i].flags & OBD_BRW_SYNC)
 665                  *
 666                  * XXX we could handle this on per-lnb basis as done by
 667                  * grant. */
 668                 if ((lnb[i].flags & OBD_BRW_NOQUOTA) ||
 669                     (lnb[i].flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 670                     OBD_BRW_FROM_GRANT)
 671                         ignore_quota = true;
 672         }
 673
 674         /*
 675          * each extent can go into new leaf causing a split
 676          * 5 is max tree depth: inode + 4 index blocks
 677          * with blockmaps, depth is 3 at most
 678          */
 679         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 680                 /*
 681                  * many concurrent threads may grow tree by the time
 682                  * our transaction starts. so, consider 2 is a min depth
 683                  */
 684                 depth = ext_depth(inode);
 685                 depth = max(depth, 1) + 1;
 686                 newblocks += depth;
 687                 oh->ot_credits++; /* inode */
 688                 oh->ot_credits += depth * 2 * extents;
 689         } else {
 690                 depth = 3;
 691                 newblocks += depth;
 692                 oh->ot_credits++; /* inode */
 693                 oh->ot_credits += depth * extents;
 694         }
 695
 696         /* quota space for metadata blocks */
 697         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
 698
 699         /* quota space should be reported in 1K blocks */
 700         quota_space = toqb(quota_space);
 701
 702         /* each new block can go in different group (bitmap + gd) */
 703
 704         /* we can't dirty more bitmap blocks than exist */
 705         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
 706                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
 707         else
 708                 oh->ot_credits += newblocks;
 709
 710         /* we can't dirty more gd blocks than exist */
 711         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
 712                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
 713         else
 714                 oh->ot_credits += newblocks;
 715
 716         /* make sure the over quota flags were not set */
 717         lnb[0].flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA);
 718
 719         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid,
 720                                    quota_space, oh, true, true, &flags,
 721                                    ignore_quota);
 722
 723         /* we need only to store the overquota flags in the first lnb for
 724          * now, once we support multiple objects BRW, this code needs be
 725          * revised. */
 726         if (flags & QUOTA_FL_OVER_USRQUOTA)
 727                 lnb[0].flags |= OBD_BRW_OVER_USRQUOTA;
 728         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 729                 lnb[0].flags |= OBD_BRW_OVER_GRPQUOTA;
 730
 731         RETURN(rc);
 732 }
 733
 734 /* Check if a block is allocated or not */
 735 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 736                             struct niobuf_local *lnb, int npages,
 737                             struct thandle *thandle)
 738 {
 739         struct osd_thread_info *oti = osd_oti_get(env);
 740         struct osd_iobuf *iobuf = &oti->oti_iobuf;
 741         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 742         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
 743         loff_t isize;
 744         int rc = 0, i;
 745
 746         LASSERT(inode);
 747
 748         rc = osd_init_iobuf(osd, iobuf, 1, npages);
 749         if (unlikely(rc != 0))
 750                 RETURN(rc);
 751
 752         isize = i_size_read(inode);
 753         ll_vfs_dq_init(inode);
 754
 755         for (i = 0; i < npages; i++) {
 756                 if (lnb[i].rc == -ENOSPC &&
 757                     osd_is_mapped(inode, lnb[i].lnb_file_offset)) {
 758                         /* Allow the write to proceed if overwriting an
 759                          * existing block */
 760                         lnb[i].rc = 0;
 761                 }
 762
 763                 if (lnb[i].rc) { /* ENOSPC, network RPC error, etc. */
 764                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
 765                                lnb[i].rc);
 766                         LASSERT(lnb[i].page);
 767                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 768                         continue;
 769                 }
 770
 771                 LASSERT(PageLocked(lnb[i].page));
 772                 LASSERT(!PageWriteback(lnb[i].page));
 773
 774                 if (lnb[i].lnb_file_offset + lnb[i].len > isize)
 775                         isize = lnb[i].lnb_file_offset + lnb[i].len;
 776
 777                 /*
 778                  * Since write and truncate are serialized by oo_sem, even
 779                  * partial-page truncate should not leave dirty pages in the
 780                  * page cache.
 781                  */
 782                 LASSERT(!PageDirty(lnb[i].page));
 783
 784                 SetPageUptodate(lnb[i].page);
 785
 786                 osd_iobuf_add_page(iobuf, lnb[i].page);
 787         }
 788
 789         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
 790                 rc = -ENOSPC;
 791         } else if (iobuf->dr_npages > 0) {
 792                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 793                                                        iobuf->dr_npages,
 794                                                        iobuf->dr_blocks,
 795                                                        1, NULL);
 796         } else {
 797                 /* no pages to write, no transno is needed */
 798                 thandle->th_local = 1;
 799         }
 800
 801         if (likely(rc == 0)) {
 802                 if (isize > i_size_read(inode)) {
 803                         i_size_write(inode, isize);
 804                         LDISKFS_I(inode)->i_disksize = isize;
 805                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
 806                 }
 807
 808                 rc = osd_do_bio(osd, inode, iobuf);
 809                 /* we don't do stats here as in read path because
 810                  * write is async: we'll do this in osd_put_bufs() */
 811         }
 812
 813         if (unlikely(rc != 0)) {
 814                 /* if write fails, we should drop pages from the cache */
 815                 for (i = 0; i < npages; i++) {
 816                         if (lnb[i].page == NULL)
 817                                 continue;
 818                         LASSERT(PageLocked(lnb[i].page));
 819                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 820                 }
 821         }
 822
 823         RETURN(rc);
 824 }
 825
 826 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 827                          struct niobuf_local *lnb, int npages)
 828 {
 829         struct osd_thread_info *oti = osd_oti_get(env);
 830         struct osd_iobuf *iobuf = &oti->oti_iobuf;
 831         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 832         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 833         struct timeval start, end;
 834         unsigned long timediff;
 835         int rc = 0, i, m = 0, cache = 0;
 836
 837         LASSERT(inode);
 838
 839         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 840         if (unlikely(rc != 0))
 841                 RETURN(rc);
 842
 843         if (osd->od_read_cache)
 844                 cache = 1;
 845         if (i_size_read(inode) > osd->od_readcache_max_filesize)
 846                 cache = 0;
 847
 848         do_gettimeofday(&start);
 849         for (i = 0; i < npages; i++) {
 850
 851                 if (i_size_read(inode) <= lnb[i].lnb_file_offset)
 852                         /* If there's no more data, abort early.
 853                          * lnb->rc == 0, so it's easy to detect later. */
 854                         break;
 855
 856                 if (i_size_read(inode) <
 857                     lnb[i].lnb_file_offset + lnb[i].len - 1)
 858                         lnb[i].rc = i_size_read(inode) - lnb[i].lnb_file_offset;
 859                 else
 860                         lnb[i].rc = lnb[i].len;
 861                 m += lnb[i].len;
 862
 863                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS, 1);
 864                 if (PageUptodate(lnb[i].page)) {
 865                         lprocfs_counter_add(osd->od_stats,
 866                                             LPROC_OSD_CACHE_HIT, 1);
 867                 } else {
 868                         lprocfs_counter_add(osd->od_stats,
 869                                             LPROC_OSD_CACHE_MISS, 1);
 870                         osd_iobuf_add_page(iobuf, lnb[i].page);
 871                 }
 872                 if (cache == 0)
 873                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 874         }
 875         do_gettimeofday(&end);
 876         timediff = cfs_timeval_sub(&end, &start, NULL);
 877         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 878
 879         if (iobuf->dr_npages) {
 880                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 881                                                        iobuf->dr_npages,
 882                                                        iobuf->dr_blocks,
 883                                                        0, NULL);
 884                 rc = osd_do_bio(osd, inode, iobuf);
 885
 886                 /* IO stats will be done in osd_bufs_put() */
 887         }
 888
 889         RETURN(rc);
 890 }
 891
 892 /*
 893  * XXX: Another layering violation for now.
 894  *
 895  * We don't want to use ->f_op->read methods, because generic file write
 896  *
 897  *         - serializes on ->i_sem, and
 898  *
 899  *         - does a lot of extra work like balance_dirty_pages(),
 900  *
 901  * which doesn't work for globally shared files like /last_rcvd.
 902  */
 903 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
 904 {
 905         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 906
 907         memcpy(buffer, (char *)ei->i_data, buflen);
 908
 909         return  buflen;
 910 }
 911
 912 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
 913 {
 914         struct buffer_head *bh;
 915         unsigned long block;
 916         int osize;
 917         int blocksize;
 918         int csize;
 919         int boffs;
 920         int err;
 921
 922         /* prevent reading after eof */
 923         spin_lock(&inode->i_lock);
 924         if (i_size_read(inode) < *offs + size) {
 925                 loff_t diff = i_size_read(inode) - *offs;
 926                 spin_unlock(&inode->i_lock);
 927                 if (diff < 0) {
 928                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
 929                                i_size_read(inode), *offs);
 930                         return -EBADR;
 931                 } else if (diff == 0) {
 932                         return 0;
 933                 } else {
 934                         size = diff;
 935                 }
 936         } else {
 937                 spin_unlock(&inode->i_lock);
 938         }
 939
 940         blocksize = 1 << inode->i_blkbits;
 941         osize = size;
 942         while (size > 0) {
 943                 block = *offs >> inode->i_blkbits;
 944                 boffs = *offs & (blocksize - 1);
 945                 csize = min(blocksize - boffs, size);
 946                 bh = ldiskfs_bread(NULL, inode, block, 0, &err);
 947                 if (!bh) {
 948                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %d\n",
 949                                LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
 950                                csize, *offs, inode->i_ino, err);
 951                         return err;
 952                 }
 953
 954                 memcpy(buf, bh->b_data + boffs, csize);
 955                 brelse(bh);
 956
 957                 *offs += csize;
 958                 buf += csize;
 959                 size -= csize;
 960         }
 961         return osize;
 962 }
 963
 964 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 965                         struct lu_buf *buf, loff_t *pos,
 966                         struct lustre_capa *capa)
 967 {
 968         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 969         int           rc;
 970
 971         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_READ))
 972                 RETURN(-EACCES);
 973
 974         /* Read small symlink from inode body as we need to maintain correct
 975          * on-disk symlinks for ldiskfs.
 976          */
 977         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
 978             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
 979                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
 980         else
 981                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
 982
 983         return rc;
 984 }
 985
 986 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 987                                  const loff_t size, loff_t pos,
 988                                  struct thandle *handle)
 989 {
 990         struct osd_thandle *oh;
 991         int                 credits;
 992         struct inode       *inode;
 993         int                 rc;
 994         ENTRY;
 995
 996         LASSERT(handle != NULL);
 997
 998         oh = container_of0(handle, struct osd_thandle, ot_super);
 999         LASSERT(oh->ot_handle == NULL);
1000
1001         credits = osd_dto_credits_noquota[DTO_WRITE_BLOCK];
1002
1003         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1004
1005         inode = osd_dt_obj(dt)->oo_inode;
1006
1007         /* we may declare write to non-exist llog */
1008         if (inode == NULL)
1009                 RETURN(0);
1010
1011         /* dt_declare_write() is usually called for system objects, such
1012          * as llog or last_rcvd files. We needn't enforce quota on those
1013          * objects, so always set the lqi_space as 0. */
1014         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
1015                                    true, true, NULL, false);
1016         RETURN(rc);
1017 }
1018
1019 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1020 {
1021         /* LU-2634: clear the extent format for fast symlink */
1022         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1023
1024         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1025         LDISKFS_I(inode)->i_disksize = buflen;
1026         i_size_write(inode, buflen);
1027         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1028
1029         return 0;
1030 }
1031
1032 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1033                              int write_NUL, loff_t *offs, handle_t *handle)
1034 {
1035         struct buffer_head *bh        = NULL;
1036         loff_t              offset    = *offs;
1037         loff_t              new_size  = i_size_read(inode);
1038         unsigned long       block;
1039         int                 blocksize = 1 << inode->i_blkbits;
1040         int                 err = 0;
1041         int                 size;
1042         int                 boffs;
1043         int                 dirty_inode = 0;
1044
1045         if (write_NUL) {
1046                 /*
1047                  * long symlink write does not count the NUL terminator in
1048                  * bufsize, we write it, and the inode's file size does not
1049                  * count the NUL terminator as well.
1050                  */
1051                 ((char *)buf)[bufsize] = '\0';
1052                 ++bufsize;
1053         }
1054         while (bufsize > 0) {
1055                 if (bh != NULL)
1056                         brelse(bh);
1057
1058                 block = offset >> inode->i_blkbits;
1059                 boffs = offset & (blocksize - 1);
1060                 size = min(blocksize - boffs, bufsize);
1061                 bh = ldiskfs_bread(handle, inode, block, 1, &err);
1062                 if (!bh) {
1063                         CERROR("%s: error reading offset %llu (block %lu): "
1064                                "rc = %d\n",
1065                                inode->i_sb->s_id, offset, block, err);
1066                         break;
1067                 }
1068
1069                 err = ldiskfs_journal_get_write_access(handle, bh);
1070                 if (err) {
1071                         CERROR("journal_get_write_access() returned error %d\n",
1072                                err);
1073                         break;
1074                 }
1075                 LASSERTF(boffs + size <= bh->b_size,
1076                          "boffs %d size %d bh->b_size %lu",
1077                          boffs, size, (unsigned long)bh->b_size);
1078                 memcpy(bh->b_data + boffs, buf, size);
1079                 err = ldiskfs_journal_dirty_metadata(handle, bh);
1080                 if (err)
1081                         break;
1082
1083                 if (offset + size > new_size)
1084                         new_size = offset + size;
1085                 offset += size;
1086                 bufsize -= size;
1087                 buf += size;
1088         }
1089         if (bh)
1090                 brelse(bh);
1091
1092         if (write_NUL)
1093                 --new_size;
1094         /* correct in-core and on-disk sizes */
1095         if (new_size > i_size_read(inode)) {
1096                 spin_lock(&inode->i_lock);
1097                 if (new_size > i_size_read(inode))
1098                         i_size_write(inode, new_size);
1099                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1100                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1101                         dirty_inode = 1;
1102                 }
1103                 spin_unlock(&inode->i_lock);
1104                 if (dirty_inode)
1105                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1106         }
1107
1108         if (err == 0)
1109                 *offs = offset;
1110         return err;
1111 }
1112
1113 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1114                          const struct lu_buf *buf, loff_t *pos,
1115                          struct thandle *handle, struct lustre_capa *capa,
1116                          int ignore_quota)
1117 {
1118         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1119         struct osd_thandle      *oh;
1120         ssize_t                 result;
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         LASSERT(inode != NULL);
1130         ll_vfs_dq_init(inode);
1131
1132         /* XXX: don't check: one declared chunk can be used many times */
1133         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1134
1135         oh = container_of(handle, struct osd_thandle, ot_super);
1136         LASSERT(oh->ot_handle->h_transaction != NULL);
1137         /* Write small symlink to inode body as we need to maintain correct
1138          * on-disk symlinks for ldiskfs.
1139          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1140          * does not count it in.
1141          */
1142         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1143         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1144                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1145         else
1146                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1147                                                   buf->lb_len, is_link, pos,
1148                                                   oh->ot_handle);
1149         if (result == 0)
1150                 result = buf->lb_len;
1151         return result;
1152 }
1153
1154 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1155                              __u64 start, __u64 end, struct thandle *th)
1156 {
1157         struct osd_thandle *oh;
1158         struct inode       *inode;
1159         int                 rc;
1160         ENTRY;
1161
1162         LASSERT(th);
1163         oh = container_of(th, struct osd_thandle, ot_super);
1164
1165         /*
1166          * we don't need to reserve credits for whole truncate
1167          * it's not possible as truncate may need to free too many
1168          * blocks and that won't fit a single transaction. instead
1169          * we reserve credits to change i_size and put inode onto
1170          * orphan list. if needed truncate will extend or restart
1171          * transaction
1172          */
1173         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1174                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1175
1176         inode = osd_dt_obj(dt)->oo_inode;
1177         LASSERT(inode);
1178
1179         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
1180                                    true, true, NULL, false);
1181         RETURN(rc);
1182 }
1183
1184 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1185                      __u64 start, __u64 end, struct thandle *th,
1186                      struct lustre_capa *capa)
1187 {
1188         struct osd_thandle *oh;
1189         struct osd_object  *obj = osd_dt_obj(dt);
1190         struct inode       *inode = obj->oo_inode;
1191         handle_t           *h;
1192         tid_t               tid;
1193         loff_t             oldsize;
1194         int                rc = 0, rc2 = 0;
1195         ENTRY;
1196
1197         LASSERT(end == OBD_OBJECT_EOF);
1198         LASSERT(dt_object_exists(dt));
1199         LASSERT(osd_invariant(obj));
1200         LASSERT(inode != NULL);
1201         ll_vfs_dq_init(inode);
1202
1203         LASSERT(th);
1204         oh = container_of(th, struct osd_thandle, ot_super);
1205         LASSERT(oh->ot_handle->h_transaction != NULL);
1206
1207         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1208
1209         tid = oh->ot_handle->h_transaction->t_tid;
1210
1211         oldsize=inode->i_size;
1212         i_size_write(inode, start);
1213         truncate_pagecache(inode, oldsize, start);
1214         if (inode->i_op->truncate)
1215                 inode->i_op->truncate(inode);
1216
1217         /*
1218          * For a partial-page truncate, flush the page to disk immediately to
1219          * avoid data corruption during direct disk write.  b=17397
1220          */
1221         if ((start & ~CFS_PAGE_MASK) != 0)
1222                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1223
1224         h = journal_current_handle();
1225         LASSERT(h != NULL);
1226         LASSERT(h == oh->ot_handle);
1227
1228         if (tid != h->h_transaction->t_tid) {
1229                 int credits = oh->ot_credits;
1230                 /*
1231                  * transaction has changed during truncate
1232                  * we need to restart the handle with our credits
1233                  */
1234                 if (h->h_buffer_credits < credits) {
1235                         if (ldiskfs_journal_extend(h, credits))
1236                                 rc2 = ldiskfs_journal_restart(h, credits);
1237                 }
1238         }
1239
1240         RETURN(rc == 0 ? rc2 : rc);
1241 }
1242
1243 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1244                           struct ll_user_fiemap *fm)
1245 {
1246         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1247         struct osd_thread_info *info   = osd_oti_get(env);
1248         struct dentry          *dentry = &info->oti_obj_dentry;
1249         struct file            *file   = &info->oti_file;
1250         mm_segment_t            saved_fs;
1251         int rc;
1252
1253         LASSERT(inode);
1254         dentry->d_inode = inode;
1255         dentry->d_sb = inode->i_sb;
1256         file->f_dentry = dentry;
1257         file->f_mapping = inode->i_mapping;
1258         file->f_op = inode->i_fop;
1259
1260         saved_fs = get_fs();
1261         set_fs(get_ds());
1262         /* ldiskfs_ioctl does not have a inode argument */
1263         if (inode->i_fop->unlocked_ioctl)
1264                 rc = inode->i_fop->unlocked_ioctl(file, FSFILT_IOC_FIEMAP,
1265                                                   (long)fm);
1266         else
1267                 rc = -ENOTTY;
1268         set_fs(saved_fs);
1269         return rc;
1270 }
1271
1272 /*
1273  * in some cases we may need declare methods for objects being created
1274  * e.g., when we create symlink
1275  */
1276 const struct dt_body_operations osd_body_ops_new = {
1277         .dbo_declare_write = osd_declare_write,
1278 };
1279
1280 const struct dt_body_operations osd_body_ops = {
1281         .dbo_read                 = osd_read,
1282         .dbo_declare_write        = osd_declare_write,
1283         .dbo_write                = osd_write,
1284         .dbo_bufs_get             = osd_bufs_get,
1285         .dbo_bufs_put             = osd_bufs_put,
1286         .dbo_write_prep           = osd_write_prep,
1287         .dbo_declare_write_commit = osd_declare_write_commit,
1288         .dbo_write_commit         = osd_write_commit,
1289         .dbo_read_prep            = osd_read_prep,
1290         .dbo_declare_punch         = osd_declare_punch,
1291         .dbo_punch                 = osd_punch,
1292         .dbo_fiemap_get           = osd_fiemap_get,
1293 };
1294