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                                        bio_hw_segments(q, bio),
 339                                        queue_max_hw_segments(q));
 340
 341                                 record_start_io(iobuf, bio->bi_size);
 342                                 osd_submit_bio(iobuf->dr_rw, bio);
 343                         }
 344
 345                         /* allocate new bio */
 346                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 347                                                       (npages - page_idx) *
 348                                                       blocks_per_page));
 349                         if (bio == NULL) {
 350                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 351                                        (npages - page_idx), blocks_per_page,
 352                                        (npages - page_idx) * blocks_per_page);
 353                                 rc = -ENOMEM;
 354                                 goto out;
 355                         }
 356
 357                         bio->bi_bdev = inode->i_sb->s_bdev;
 358                         bio->bi_sector = sector;
 359                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 360                         bio->bi_end_io = dio_complete_routine;
 361                         bio->bi_private = iobuf;
 362
 363                         rc = bio_add_page(bio, page,
 364                                           blocksize * nblocks, page_offset);
 365                         LASSERT(rc != 0);
 366                 }
 367         }
 368
 369         if (bio != NULL) {
 370                 record_start_io(iobuf, bio->bi_size);
 371                 osd_submit_bio(iobuf->dr_rw, bio);
 372                 rc = 0;
 373         }
 374
 375  out:
 376         /* in order to achieve better IO throughput, we don't wait for writes
 377          * completion here. instead we proceed with transaction commit in
 378          * parallel and wait for IO completion once transaction is stopped
 379          * see osd_trans_stop() for more details -bzzz */
 380         if (iobuf->dr_rw == 0) {
 381                 wait_event(iobuf->dr_wait,
 382                                cfs_atomic_read(&iobuf->dr_numreqs) == 0);
 383         }
 384
 385         if (rc == 0)
 386                 rc = iobuf->dr_error;
 387         RETURN(rc);
 388 }
 389
 390 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 391                                    struct niobuf_local *lnb)
 392 {
 393         ENTRY;
 394
 395         *nrpages = 0;
 396
 397         while (len > 0) {
 398                 int poff = offset & (PAGE_CACHE_SIZE - 1);
 399                 int plen = PAGE_CACHE_SIZE - poff;
 400
 401                 if (plen > len)
 402                         plen = len;
 403                 lnb->lnb_file_offset = offset;
 404                 lnb->lnb_page_offset = poff;
 405                 lnb->len = plen;
 406                 /* lb->flags = rnb->flags; */
 407                 lnb->flags = 0;
 408                 lnb->page = NULL;
 409                 lnb->rc = 0;
 410
 411                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 412                          (long long) len);
 413                 offset += plen;
 414                 len -= plen;
 415                 lnb++;
 416                 (*nrpages)++;
 417         }
 418
 419         RETURN(0);
 420 }
 421
 422 struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
 423 {
 424         struct inode      *inode = osd_dt_obj(dt)->oo_inode;
 425         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 426         struct page       *page;
 427
 428         LASSERT(inode);
 429
 430         page = find_or_create_page(inode->i_mapping, offset >> PAGE_CACHE_SHIFT,
 431                                    GFP_NOFS | __GFP_HIGHMEM);
 432         if (unlikely(page == NULL))
 433                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 434
 435         return page;
 436 }
 437
 438 /*
 439  * there are following "locks":
 440  * journal_start
 441  * i_mutex
 442  * page lock
 443
 444  * osd write path
 445     * lock page(s)
 446     * journal_start
 447     * truncate_sem
 448
 449  * ext4 vmtruncate:
 450     * lock pages, unlock
 451     * journal_start
 452     * lock partial page
 453     * i_data_sem
 454
 455 */
 456 int osd_bufs_get(const struct lu_env *env, struct dt_object *d, loff_t pos,
 457                  ssize_t len, struct niobuf_local *lnb, int rw,
 458                  struct lustre_capa *capa)
 459 {
 460         struct osd_object   *obj    = osd_dt_obj(d);
 461         int npages, i, rc = 0;
 462
 463         LASSERT(obj->oo_inode);
 464
 465         osd_map_remote_to_local(pos, len, &npages, lnb);
 466
 467         for (i = 0; i < npages; i++, lnb++) {
 468
 469                 /* We still set up for ungranted pages so that granted pages
 470                  * can be written to disk as they were promised, and portals
 471                  * needs to keep the pages all aligned properly. */
 472                 lnb->dentry = (void *) obj;
 473
 474                 lnb->page = osd_get_page(d, lnb->lnb_file_offset, rw);
 475                 if (lnb->page == NULL)
 476                         GOTO(cleanup, rc = -ENOMEM);
 477
 478                 /* DLM locking protects us from write and truncate competing
 479                  * for same region, but truncate can leave dirty page in the
 480                  * cache. it's possible the writeout on a such a page is in
 481                  * progress when we access it. it's also possible that during
 482                  * this writeout we put new (partial) data, but then won't
 483                  * be able to proceed in filter_commitrw_write(). thus let's
 484                  * just wait for writeout completion, should be rare enough.
 485                  * -bzzz */
 486                 wait_on_page_writeback(lnb->page);
 487                 BUG_ON(PageWriteback(lnb->page));
 488
 489                 lu_object_get(&d->do_lu);
 490         }
 491         rc = i;
 492
 493 cleanup:
 494         RETURN(rc);
 495 }
 496
 497 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 498                         struct niobuf_local *lnb, int npages)
 499 {
 500         struct osd_thread_info *oti = osd_oti_get(env);
 501         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 502         struct osd_device      *d = osd_obj2dev(osd_dt_obj(dt));
 503         int                     i;
 504
 505         /* to do IO stats, notice we do this here because
 506          * osd_do_bio() doesn't wait for write to complete */
 507         osd_fini_iobuf(d, iobuf);
 508
 509         for (i = 0; i < npages; i++) {
 510                 if (lnb[i].page == NULL)
 511                         continue;
 512                 LASSERT(PageLocked(lnb[i].page));
 513                 unlock_page(lnb[i].page);
 514                 page_cache_release(lnb[i].page);
 515                 lu_object_put(env, &dt->do_lu);
 516                 lnb[i].page = NULL;
 517         }
 518         RETURN(0);
 519 }
 520
 521 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 522                           struct niobuf_local *lnb, int npages)
 523 {
 524         struct osd_thread_info *oti   = osd_oti_get(env);
 525         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 526         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 527         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 528         struct timeval          start;
 529         struct timeval          end;
 530         unsigned long           timediff;
 531         ssize_t                 isize;
 532         __s64                   maxidx;
 533         int                     rc = 0;
 534         int                     i;
 535         int                     cache = 0;
 536
 537         LASSERT(inode);
 538
 539         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 540         if (unlikely(rc != 0))
 541                 RETURN(rc);
 542
 543         isize = i_size_read(inode);
 544         maxidx = ((isize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) - 1;
 545
 546         if (osd->od_writethrough_cache)
 547                 cache = 1;
 548         if (isize > osd->od_readcache_max_filesize)
 549                 cache = 0;
 550
 551         do_gettimeofday(&start);
 552         for (i = 0; i < npages; i++) {
 553
 554                 if (cache == 0)
 555                         generic_error_remove_page(inode->i_mapping,
 556                                                   lnb[i].page);
 557
 558                 /*
 559                  * till commit the content of the page is undefined
 560                  * we'll set it uptodate once bulk is done. otherwise
 561                  * subsequent reads can access non-stable data
 562                  */
 563                 ClearPageUptodate(lnb[i].page);
 564
 565                 if (lnb[i].len == PAGE_CACHE_SIZE)
 566                         continue;
 567
 568                 if (maxidx >= lnb[i].page->index) {
 569                         osd_iobuf_add_page(iobuf, lnb[i].page);
 570                 } else {
 571                         long off;
 572                         char *p = kmap(lnb[i].page);
 573
 574                         off = lnb[i].lnb_page_offset;
 575                         if (off)
 576                                 memset(p, 0, off);
 577                         off = (lnb[i].lnb_page_offset + lnb[i].len) &
 578                               ~CFS_PAGE_MASK;
 579                         if (off)
 580                                 memset(p + off, 0, PAGE_CACHE_SIZE - off);
 581                         kunmap(lnb[i].page);
 582                 }
 583         }
 584         do_gettimeofday(&end);
 585         timediff = cfs_timeval_sub(&end, &start, NULL);
 586         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 587
 588         if (iobuf->dr_npages) {
 589                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 590                                                        iobuf->dr_npages,
 591                                                        iobuf->dr_blocks,
 592                                                        0, NULL);
 593                 if (likely(rc == 0)) {
 594                         rc = osd_do_bio(osd, inode, iobuf);
 595                         /* do IO stats for preparation reads */
 596                         osd_fini_iobuf(osd, iobuf);
 597                 }
 598         }
 599         RETURN(rc);
 600 }
 601
 602 /* Check if a block is allocated or not */
 603 static int osd_is_mapped(struct inode *inode, obd_size offset)
 604 {
 605         sector_t (*fs_bmap)(struct address_space *, sector_t);
 606
 607         fs_bmap = inode->i_mapping->a_ops->bmap;
 608
 609         /* We can't know if we are overwriting or not */
 610         if (unlikely(fs_bmap == NULL))
 611                 return 0;
 612
 613         if (i_size_read(inode) == 0)
 614                 return 0;
 615
 616         /* Beyond EOF, must not be mapped */
 617         if (((i_size_read(inode) - 1) >> inode->i_blkbits) <
 618             (offset >> inode->i_blkbits))
 619                 return 0;
 620
 621         if (fs_bmap(inode->i_mapping, offset >> inode->i_blkbits) == 0)
 622                 return 0;
 623
 624         return 1;
 625 }
 626
 627 static int osd_declare_write_commit(const struct lu_env *env,
 628                                     struct dt_object *dt,
 629                                     struct niobuf_local *lnb, int npages,
 630                                     struct thandle *handle)
 631 {
 632         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 633         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
 634         struct osd_thandle      *oh;
 635         int                      extents = 1;
 636         int                      depth;
 637         int                      i;
 638         int                      newblocks;
 639         int                      rc = 0;
 640         int                      flags = 0;
 641         bool                     ignore_quota = false;
 642         long long                quota_space = 0;
 643         ENTRY;
 644
 645         LASSERT(handle != NULL);
 646         oh = container_of0(handle, struct osd_thandle, ot_super);
 647         LASSERT(oh->ot_handle == NULL);
 648
 649         newblocks = npages;
 650
 651         /* calculate number of extents (probably better to pass nb) */
 652         for (i = 0; i < npages; i++) {
 653                 if (i && lnb[i].lnb_file_offset !=
 654                     lnb[i - 1].lnb_file_offset + lnb[i - 1].len)
 655                         extents++;
 656
 657                 if (!osd_is_mapped(inode, lnb[i].lnb_file_offset))
 658                         quota_space += PAGE_CACHE_SIZE;
 659
 660                 /* ignore quota for the whole request if any page is from
 661                  * client cache or written by root.
 662                  *
 663                  * XXX once we drop the 1.8 client support, the checking
 664                  * for whether page is from cache can be simplified as:
 665                  * !(lnb[i].flags & OBD_BRW_SYNC)
 666                  *
 667                  * XXX we could handle this on per-lnb basis as done by
 668                  * grant. */
 669                 if ((lnb[i].flags & OBD_BRW_NOQUOTA) ||
 670                     (lnb[i].flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 671                     OBD_BRW_FROM_GRANT)
 672                         ignore_quota = true;
 673         }
 674
 675         /*
 676          * each extent can go into new leaf causing a split
 677          * 5 is max tree depth: inode + 4 index blocks
 678          * with blockmaps, depth is 3 at most
 679          */
 680         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 681                 /*
 682                  * many concurrent threads may grow tree by the time
 683                  * our transaction starts. so, consider 2 is a min depth
 684                  */
 685                 depth = ext_depth(inode);
 686                 depth = max(depth, 1) + 1;
 687                 newblocks += depth;
 688                 oh->ot_credits++; /* inode */
 689                 oh->ot_credits += depth * 2 * extents;
 690         } else {
 691                 depth = 3;
 692                 newblocks += depth;
 693                 oh->ot_credits++; /* inode */
 694                 oh->ot_credits += depth * extents;
 695         }
 696
 697         /* quota space for metadata blocks */
 698         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
 699
 700         /* quota space should be reported in 1K blocks */
 701         quota_space = toqb(quota_space);
 702
 703         /* each new block can go in different group (bitmap + gd) */
 704
 705         /* we can't dirty more bitmap blocks than exist */
 706         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
 707                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
 708         else
 709                 oh->ot_credits += newblocks;
 710
 711         /* we can't dirty more gd blocks than exist */
 712         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
 713                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
 714         else
 715                 oh->ot_credits += newblocks;
 716
 717         /* make sure the over quota flags were not set */
 718         lnb[0].flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA);
 719
 720         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid,
 721                                    quota_space, oh, true, true, &flags,
 722                                    ignore_quota);
 723
 724         /* we need only to store the overquota flags in the first lnb for
 725          * now, once we support multiple objects BRW, this code needs be
 726          * revised. */
 727         if (flags & QUOTA_FL_OVER_USRQUOTA)
 728                 lnb[0].flags |= OBD_BRW_OVER_USRQUOTA;
 729         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 730                 lnb[0].flags |= OBD_BRW_OVER_GRPQUOTA;
 731
 732         RETURN(rc);
 733 }
 734
 735 /* Check if a block is allocated or not */
 736 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 737                             struct niobuf_local *lnb, int npages,
 738                             struct thandle *thandle)
 739 {
 740         struct osd_thread_info *oti = osd_oti_get(env);
 741         struct osd_iobuf *iobuf = &oti->oti_iobuf;
 742         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 743         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
 744         loff_t isize;
 745         int rc = 0, i;
 746
 747         LASSERT(inode);
 748
 749         rc = osd_init_iobuf(osd, iobuf, 1, npages);
 750         if (unlikely(rc != 0))
 751                 RETURN(rc);
 752
 753         isize = i_size_read(inode);
 754         ll_vfs_dq_init(inode);
 755
 756         for (i = 0; i < npages; i++) {
 757                 if (lnb[i].rc == -ENOSPC &&
 758                     osd_is_mapped(inode, lnb[i].lnb_file_offset)) {
 759                         /* Allow the write to proceed if overwriting an
 760                          * existing block */
 761                         lnb[i].rc = 0;
 762                 }
 763
 764                 if (lnb[i].rc) { /* ENOSPC, network RPC error, etc. */
 765                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
 766                                lnb[i].rc);
 767                         LASSERT(lnb[i].page);
 768                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 769                         continue;
 770                 }
 771
 772                 LASSERT(PageLocked(lnb[i].page));
 773                 LASSERT(!PageWriteback(lnb[i].page));
 774
 775                 if (lnb[i].lnb_file_offset + lnb[i].len > isize)
 776                         isize = lnb[i].lnb_file_offset + lnb[i].len;
 777
 778                 /*
 779                  * Since write and truncate are serialized by oo_sem, even
 780                  * partial-page truncate should not leave dirty pages in the
 781                  * page cache.
 782                  */
 783                 LASSERT(!PageDirty(lnb[i].page));
 784
 785                 SetPageUptodate(lnb[i].page);
 786
 787                 osd_iobuf_add_page(iobuf, lnb[i].page);
 788         }
 789
 790         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
 791                 rc = -ENOSPC;
 792         } else if (iobuf->dr_npages > 0) {
 793                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 794                                                        iobuf->dr_npages,
 795                                                        iobuf->dr_blocks,
 796                                                        1, NULL);
 797         } else {
 798                 /* no pages to write, no transno is needed */
 799                 thandle->th_local = 1;
 800         }
 801
 802         if (likely(rc == 0)) {
 803                 if (isize > i_size_read(inode)) {
 804                         i_size_write(inode, isize);
 805                         LDISKFS_I(inode)->i_disksize = isize;
 806                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
 807                 }
 808
 809                 rc = osd_do_bio(osd, inode, iobuf);
 810                 /* we don't do stats here as in read path because
 811                  * write is async: we'll do this in osd_put_bufs() */
 812         }
 813
 814         if (unlikely(rc != 0)) {
 815                 /* if write fails, we should drop pages from the cache */
 816                 for (i = 0; i < npages; i++) {
 817                         if (lnb[i].page == NULL)
 818                                 continue;
 819                         LASSERT(PageLocked(lnb[i].page));
 820                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 821                 }
 822         }
 823
 824         RETURN(rc);
 825 }
 826
 827 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 828                          struct niobuf_local *lnb, int npages)
 829 {
 830         struct osd_thread_info *oti = osd_oti_get(env);
 831         struct osd_iobuf *iobuf = &oti->oti_iobuf;
 832         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 833         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 834         struct timeval start, end;
 835         unsigned long timediff;
 836         int rc = 0, i, m = 0, cache = 0;
 837
 838         LASSERT(inode);
 839
 840         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 841         if (unlikely(rc != 0))
 842                 RETURN(rc);
 843
 844         if (osd->od_read_cache)
 845                 cache = 1;
 846         if (i_size_read(inode) > osd->od_readcache_max_filesize)
 847                 cache = 0;
 848
 849         do_gettimeofday(&start);
 850         for (i = 0; i < npages; i++) {
 851
 852                 if (i_size_read(inode) <= lnb[i].lnb_file_offset)
 853                         /* If there's no more data, abort early.
 854                          * lnb->rc == 0, so it's easy to detect later. */
 855                         break;
 856
 857                 if (i_size_read(inode) <
 858                     lnb[i].lnb_file_offset + lnb[i].len - 1)
 859                         lnb[i].rc = i_size_read(inode) - lnb[i].lnb_file_offset;
 860                 else
 861                         lnb[i].rc = lnb[i].len;
 862                 m += lnb[i].len;
 863
 864                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS, 1);
 865                 if (PageUptodate(lnb[i].page)) {
 866                         lprocfs_counter_add(osd->od_stats,
 867                                             LPROC_OSD_CACHE_HIT, 1);
 868                 } else {
 869                         lprocfs_counter_add(osd->od_stats,
 870                                             LPROC_OSD_CACHE_MISS, 1);
 871                         osd_iobuf_add_page(iobuf, lnb[i].page);
 872                 }
 873                 if (cache == 0)
 874                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 875         }
 876         do_gettimeofday(&end);
 877         timediff = cfs_timeval_sub(&end, &start, NULL);
 878         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 879
 880         if (iobuf->dr_npages) {
 881                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 882                                                        iobuf->dr_npages,
 883                                                        iobuf->dr_blocks,
 884                                                        0, NULL);
 885                 rc = osd_do_bio(osd, inode, iobuf);
 886
 887                 /* IO stats will be done in osd_bufs_put() */
 888         }
 889
 890         RETURN(rc);
 891 }
 892
 893 /*
 894  * XXX: Another layering violation for now.
 895  *
 896  * We don't want to use ->f_op->read methods, because generic file write
 897  *
 898  *         - serializes on ->i_sem, and
 899  *
 900  *         - does a lot of extra work like balance_dirty_pages(),
 901  *
 902  * which doesn't work for globally shared files like /last_rcvd.
 903  */
 904 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
 905 {
 906         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 907
 908         memcpy(buffer, (char *)ei->i_data, buflen);
 909
 910         return  buflen;
 911 }
 912
 913 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
 914 {
 915         struct buffer_head *bh;
 916         unsigned long block;
 917         int osize;
 918         int blocksize;
 919         int csize;
 920         int boffs;
 921         int err;
 922
 923         /* prevent reading after eof */
 924         spin_lock(&inode->i_lock);
 925         if (i_size_read(inode) < *offs + size) {
 926                 loff_t diff = i_size_read(inode) - *offs;
 927                 spin_unlock(&inode->i_lock);
 928                 if (diff < 0) {
 929                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
 930                                i_size_read(inode), *offs);
 931                         return -EBADR;
 932                 } else if (diff == 0) {
 933                         return 0;
 934                 } else {
 935                         size = diff;
 936                 }
 937         } else {
 938                 spin_unlock(&inode->i_lock);
 939         }
 940
 941         blocksize = 1 << inode->i_blkbits;
 942         osize = size;
 943         while (size > 0) {
 944                 block = *offs >> inode->i_blkbits;
 945                 boffs = *offs & (blocksize - 1);
 946                 csize = min(blocksize - boffs, size);
 947                 bh = ldiskfs_bread(NULL, inode, block, 0, &err);
 948                 if (!bh) {
 949                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %d\n",
 950                                LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
 951                                csize, *offs, inode->i_ino, err);
 952                         return err;
 953                 }
 954
 955                 memcpy(buf, bh->b_data + boffs, csize);
 956                 brelse(bh);
 957
 958                 *offs += csize;
 959                 buf += csize;
 960                 size -= csize;
 961         }
 962         return osize;
 963 }
 964
 965 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 966                         struct lu_buf *buf, loff_t *pos,
 967                         struct lustre_capa *capa)
 968 {
 969         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 970         int           rc;
 971
 972         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_READ))
 973                 RETURN(-EACCES);
 974
 975         /* Read small symlink from inode body as we need to maintain correct
 976          * on-disk symlinks for ldiskfs.
 977          */
 978         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
 979             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
 980                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
 981         else
 982                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
 983
 984         return rc;
 985 }
 986
 987 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 988                                  const loff_t size, loff_t pos,
 989                                  struct thandle *handle)
 990 {
 991         struct osd_thandle *oh;
 992         int                 credits;
 993         struct inode       *inode;
 994         int                 rc;
 995         ENTRY;
 996
 997         LASSERT(handle != NULL);
 998
 999         oh = container_of0(handle, struct osd_thandle, ot_super);
1000         LASSERT(oh->ot_handle == NULL);
1001
1002         credits = osd_dto_credits_noquota[DTO_WRITE_BLOCK];
1003
1004         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1005
1006         inode = osd_dt_obj(dt)->oo_inode;
1007
1008         /* we may declare write to non-exist llog */
1009         if (inode == NULL)
1010                 RETURN(0);
1011
1012         /* dt_declare_write() is usually called for system objects, such
1013          * as llog or last_rcvd files. We needn't enforce quota on those
1014          * objects, so always set the lqi_space as 0. */
1015         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
1016                                    true, true, NULL, false);
1017         RETURN(rc);
1018 }
1019
1020 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1021 {
1022         /* LU-2634: clear the extent format for fast symlink */
1023         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1024
1025         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1026         LDISKFS_I(inode)->i_disksize = buflen;
1027         i_size_write(inode, buflen);
1028         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1029
1030         return 0;
1031 }
1032
1033 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1034                              int write_NUL, loff_t *offs, handle_t *handle)
1035 {
1036         struct buffer_head *bh        = NULL;
1037         loff_t              offset    = *offs;
1038         loff_t              new_size  = i_size_read(inode);
1039         unsigned long       block;
1040         int                 blocksize = 1 << inode->i_blkbits;
1041         int                 err = 0;
1042         int                 size;
1043         int                 boffs;
1044         int                 dirty_inode = 0;
1045
1046         if (write_NUL) {
1047                 /*
1048                  * long symlink write does not count the NUL terminator in
1049                  * bufsize, we write it, and the inode's file size does not
1050                  * count the NUL terminator as well.
1051                  */
1052                 ((char *)buf)[bufsize] = '\0';
1053                 ++bufsize;
1054         }
1055         while (bufsize > 0) {
1056                 if (bh != NULL)
1057                         brelse(bh);
1058
1059                 block = offset >> inode->i_blkbits;
1060                 boffs = offset & (blocksize - 1);
1061                 size = min(blocksize - boffs, bufsize);
1062                 bh = ldiskfs_bread(handle, inode, block, 1, &err);
1063                 if (!bh) {
1064                         CERROR("%s: error reading offset %llu (block %lu): "
1065                                "rc = %d\n",
1066                                inode->i_sb->s_id, offset, block, err);
1067                         break;
1068                 }
1069
1070                 err = ldiskfs_journal_get_write_access(handle, bh);
1071                 if (err) {
1072                         CERROR("journal_get_write_access() returned error %d\n",
1073                                err);
1074                         break;
1075                 }
1076                 LASSERTF(boffs + size <= bh->b_size,
1077                          "boffs %d size %d bh->b_size %lu",
1078                          boffs, size, (unsigned long)bh->b_size);
1079                 memcpy(bh->b_data + boffs, buf, size);
1080                 err = ldiskfs_journal_dirty_metadata(handle, bh);
1081                 if (err)
1082                         break;
1083
1084                 if (offset + size > new_size)
1085                         new_size = offset + size;
1086                 offset += size;
1087                 bufsize -= size;
1088                 buf += size;
1089         }
1090         if (bh)
1091                 brelse(bh);
1092
1093         if (write_NUL)
1094                 --new_size;
1095         /* correct in-core and on-disk sizes */
1096         if (new_size > i_size_read(inode)) {
1097                 spin_lock(&inode->i_lock);
1098                 if (new_size > i_size_read(inode))
1099                         i_size_write(inode, new_size);
1100                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1101                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1102                         dirty_inode = 1;
1103                 }
1104                 spin_unlock(&inode->i_lock);
1105                 if (dirty_inode)
1106                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1107         }
1108
1109         if (err == 0)
1110                 *offs = offset;
1111         return err;
1112 }
1113
1114 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1115                          const struct lu_buf *buf, loff_t *pos,
1116                          struct thandle *handle, struct lustre_capa *capa,
1117                          int ignore_quota)
1118 {
1119         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1120         struct osd_thandle      *oh;
1121         ssize_t                 result;
1122         int                     is_link;
1123
1124         LASSERT(dt_object_exists(dt));
1125
1126         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_WRITE))
1127                 return -EACCES;
1128
1129         LASSERT(handle != NULL);
1130         LASSERT(inode != NULL);
1131         ll_vfs_dq_init(inode);
1132
1133         /* XXX: don't check: one declared chunk can be used many times */
1134         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1135
1136         oh = container_of(handle, struct osd_thandle, ot_super);
1137         LASSERT(oh->ot_handle->h_transaction != NULL);
1138         /* Write small symlink to inode body as we need to maintain correct
1139          * on-disk symlinks for ldiskfs.
1140          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1141          * does not count it in.
1142          */
1143         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1144         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1145                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1146         else
1147                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1148                                                   buf->lb_len, is_link, pos,
1149                                                   oh->ot_handle);
1150         if (result == 0)
1151                 result = buf->lb_len;
1152         return result;
1153 }
1154
1155 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1156                              __u64 start, __u64 end, struct thandle *th)
1157 {
1158         struct osd_thandle *oh;
1159         struct inode       *inode;
1160         int                 rc;
1161         ENTRY;
1162
1163         LASSERT(th);
1164         oh = container_of(th, struct osd_thandle, ot_super);
1165
1166         /*
1167          * we don't need to reserve credits for whole truncate
1168          * it's not possible as truncate may need to free too many
1169          * blocks and that won't fit a single transaction. instead
1170          * we reserve credits to change i_size and put inode onto
1171          * orphan list. if needed truncate will extend or restart
1172          * transaction
1173          */
1174         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1175                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1176
1177         inode = osd_dt_obj(dt)->oo_inode;
1178         LASSERT(inode);
1179
1180         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
1181                                    true, true, NULL, false);
1182         RETURN(rc);
1183 }
1184
1185 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1186                      __u64 start, __u64 end, struct thandle *th,
1187                      struct lustre_capa *capa)
1188 {
1189         struct osd_thandle *oh;
1190         struct osd_object  *obj = osd_dt_obj(dt);
1191         struct inode       *inode = obj->oo_inode;
1192         handle_t           *h;
1193         tid_t               tid;
1194         loff_t             oldsize;
1195         int                rc = 0, rc2 = 0;
1196         ENTRY;
1197
1198         LASSERT(end == OBD_OBJECT_EOF);
1199         LASSERT(dt_object_exists(dt));
1200         LASSERT(osd_invariant(obj));
1201         LASSERT(inode != NULL);
1202         ll_vfs_dq_init(inode);
1203
1204         LASSERT(th);
1205         oh = container_of(th, struct osd_thandle, ot_super);
1206         LASSERT(oh->ot_handle->h_transaction != NULL);
1207
1208         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1209
1210         tid = oh->ot_handle->h_transaction->t_tid;
1211
1212         oldsize=inode->i_size;
1213         i_size_write(inode, start);
1214         truncate_pagecache(inode, oldsize, start);
1215         if (inode->i_op->truncate)
1216                 inode->i_op->truncate(inode);
1217
1218         /*
1219          * For a partial-page truncate, flush the page to disk immediately to
1220          * avoid data corruption during direct disk write.  b=17397
1221          */
1222         if ((start & ~CFS_PAGE_MASK) != 0)
1223                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1224
1225         h = journal_current_handle();
1226         LASSERT(h != NULL);
1227         LASSERT(h == oh->ot_handle);
1228
1229         if (tid != h->h_transaction->t_tid) {
1230                 int credits = oh->ot_credits;
1231                 /*
1232                  * transaction has changed during truncate
1233                  * we need to restart the handle with our credits
1234                  */
1235                 if (h->h_buffer_credits < credits) {
1236                         if (ldiskfs_journal_extend(h, credits))
1237                                 rc2 = ldiskfs_journal_restart(h, credits);
1238                 }
1239         }
1240
1241         RETURN(rc == 0 ? rc2 : rc);
1242 }
1243
1244 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1245                           struct ll_user_fiemap *fm)
1246 {
1247         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1248         struct osd_thread_info *info   = osd_oti_get(env);
1249         struct dentry          *dentry = &info->oti_obj_dentry;
1250         struct file            *file   = &info->oti_file;
1251         mm_segment_t            saved_fs;
1252         int rc;
1253
1254         LASSERT(inode);
1255         dentry->d_inode = inode;
1256         dentry->d_sb = inode->i_sb;
1257         file->f_dentry = dentry;
1258         file->f_mapping = inode->i_mapping;
1259         file->f_op = inode->i_fop;
1260
1261         saved_fs = get_fs();
1262         set_fs(get_ds());
1263         /* ldiskfs_ioctl does not have a inode argument */
1264         if (inode->i_fop->unlocked_ioctl)
1265                 rc = inode->i_fop->unlocked_ioctl(file, FSFILT_IOC_FIEMAP,
1266                                                   (long)fm);
1267         else
1268                 rc = -ENOTTY;
1269         set_fs(saved_fs);
1270         return rc;
1271 }
1272
1273 /*
1274  * in some cases we may need declare methods for objects being created
1275  * e.g., when we create symlink
1276  */
1277 const struct dt_body_operations osd_body_ops_new = {
1278         .dbo_declare_write = osd_declare_write,
1279 };
1280
1281 const struct dt_body_operations osd_body_ops = {
1282         .dbo_read                 = osd_read,
1283         .dbo_declare_write        = osd_declare_write,
1284         .dbo_write                = osd_write,
1285         .dbo_bufs_get             = osd_bufs_get,
1286         .dbo_bufs_put             = osd_bufs_put,
1287         .dbo_write_prep           = osd_write_prep,
1288         .dbo_declare_write_commit = osd_declare_write_commit,
1289         .dbo_write_commit         = osd_write_commit,
1290         .dbo_read_prep            = osd_read_prep,
1291         .dbo_declare_punch         = osd_declare_punch,
1292         .dbo_punch                 = osd_punch,
1293         .dbo_fiemap_get           = osd_fiemap_get,
1294 };
1295