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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/osd/osd_io.c
  33  *
  34  * body operations
  35  *
  36  * Author: Nikita Danilov <nikita@clusterfs.com>
  37  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  38  *
  39  */
  40
  41 /* LUSTRE_VERSION_CODE */
  42 #include <lustre_ver.h>
  43 /* prerequisite for linux/xattr.h */
  44 #include <linux/types.h>
  45 /* prerequisite for linux/xattr.h */
  46 #include <linux/fs.h>
  47
  48 /*
  49  * struct OBD_{ALLOC,FREE}*()
  50  * OBD_FAIL_CHECK
  51  */
  52 #include <obd_support.h>
  53
  54 #include "osd_internal.h"
  55
  56 /* ext_depth() */
  57 #include <ldiskfs/ldiskfs_extents.h>
  58
  59 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  60                             int rw, int line, int pages)
  61 {
  62         int blocks, i;
  63
  64         LASSERTF(iobuf->dr_elapsed_valid == 0,
  65                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  66                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  67                  iobuf->dr_init_at);
  68         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  69
  70         init_waitqueue_head(&iobuf->dr_wait);
  71         atomic_set(&iobuf->dr_numreqs, 0);
  72         iobuf->dr_npages = 0;
  73         iobuf->dr_error = 0;
  74         iobuf->dr_dev = d;
  75         iobuf->dr_frags = 0;
  76         iobuf->dr_elapsed = 0;
  77         /* must be counted before, so assert */
  78         iobuf->dr_rw = rw;
  79         iobuf->dr_init_at = line;
  80
  81         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  82         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  83                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  84                         pages * sizeof(iobuf->dr_pages[0]));
  85                 return 0;
  86         }
  87
  88         /* start with 1MB for 4K blocks */
  89         i = 256;
  90         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
  91                 i <<= 1;
  92
  93         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
  94                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
  95         pages = i;
  96         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  97         iobuf->dr_max_pages = 0;
  98         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
  99                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 100
 101         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 102         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 103         if (unlikely(iobuf->dr_blocks == NULL))
 104                 return -ENOMEM;
 105
 106         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 107         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 108         if (unlikely(iobuf->dr_pages == NULL))
 109                 return -ENOMEM;
 110
 111         iobuf->dr_max_pages = pages;
 112
 113         return 0;
 114 }
 115 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 116         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 117
 118 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
 119 {
 120         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 121         iobuf->dr_pages[iobuf->dr_npages++] = page;
 122 }
 123
 124 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 125 {
 126         int rw = iobuf->dr_rw;
 127
 128         if (iobuf->dr_elapsed_valid) {
 129                 iobuf->dr_elapsed_valid = 0;
 130                 LASSERT(iobuf->dr_dev == d);
 131                 LASSERT(iobuf->dr_frags > 0);
 132                 lprocfs_oh_tally(&d->od_brw_stats.
 133                                  hist[BRW_R_DIO_FRAGS+rw],
 134                                  iobuf->dr_frags);
 135                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 136                                       iobuf->dr_elapsed);
 137         }
 138 }
 139
 140 #ifndef REQ_WRITE /* pre-2.6.35 */
 141 #define __REQ_WRITE BIO_RW
 142 #endif
 143
 144 static void dio_complete_routine(struct bio *bio, int error)
 145 {
 146         struct osd_iobuf *iobuf = bio->bi_private;
 147 #ifdef HAVE_BVEC_ITER
 148         struct bvec_iter iter;
 149         struct bio_vec bvl;
 150 #else
 151         int iter;
 152         struct bio_vec *bvl;
 153 #endif
 154
 155         /* CAVEAT EMPTOR: possibly in IRQ context
 156          * DO NOT record procfs stats here!!! */
 157
 158         if (unlikely(iobuf == NULL)) {
 159                 CERROR("***** bio->bi_private is NULL!  This should never "
 160                        "happen.  Normally, I would crash here, but instead I "
 161                        "will dump the bio contents to the console.  Please "
 162                        "report this to <https://jira.hpdd.intel.com/> , along "
 163                        "with any interesting messages leading up to this point "
 164                        "(like SCSI errors, perhaps).  Because bi_private is "
 165                        "NULL, I can't wake up the thread that initiated this "
 166                        "IO - you will probably have to reboot this node.\n");
 167                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 168                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 169                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 170                         bio->bi_rw, bio->bi_vcnt, bio_idx(bio),
 171                         bio_sectors(bio) << 9, bio->bi_end_io,
 172                         atomic_read(&bio->bi_cnt), bio->bi_private);
 173                 return;
 174         }
 175
 176         /* the check is outside of the cycle for performance reason -bzzz */
 177         if (!test_bit(__REQ_WRITE, &bio->bi_rw)) {
 178                 bio_for_each_segment_all(bvl, bio, iter) {
 179                         if (likely(error == 0))
 180                                 SetPageUptodate(bvl_to_page(bvl));
 181                         LASSERT(PageLocked(bvl_to_page(bvl)));
 182                 }
 183                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 184         } else {
 185                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 186         }
 187
 188         /* any real error is good enough -bzzz */
 189         if (error != 0 && iobuf->dr_error == 0)
 190                 iobuf->dr_error = error;
 191
 192         /*
 193          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 194          * it's possible that service thread will see dr_numreqs
 195          * is zero, but dr_elapsed is not set yet, leading to lost
 196          * data in this processing and an assertion in a subsequent
 197          * call to OSD.
 198          */
 199         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 200                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
 201                 iobuf->dr_elapsed_valid = 1;
 202         }
 203         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 204                 wake_up(&iobuf->dr_wait);
 205
 206         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 207          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 208          * mempool when serious on-disk fragmentation was encountered,
 209          * deadlocking the OST.  The bios are now released as soon as complete
 210          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 211         bio_put(bio);
 212 }
 213
 214 static void record_start_io(struct osd_iobuf *iobuf, int size)
 215 {
 216         struct osd_device    *osd = iobuf->dr_dev;
 217         struct obd_histogram *h = osd->od_brw_stats.hist;
 218
 219         iobuf->dr_frags++;
 220         atomic_inc(&iobuf->dr_numreqs);
 221
 222         if (iobuf->dr_rw == 0) {
 223                 atomic_inc(&osd->od_r_in_flight);
 224                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 225                                  atomic_read(&osd->od_r_in_flight));
 226                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 227         } else if (iobuf->dr_rw == 1) {
 228                 atomic_inc(&osd->od_w_in_flight);
 229                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 230                                  atomic_read(&osd->od_w_in_flight));
 231                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 232         } else {
 233                 LBUG();
 234         }
 235 }
 236
 237 static void osd_submit_bio(int rw, struct bio *bio)
 238 {
 239         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 240         if (rw == 0)
 241                 submit_bio(READ, bio);
 242         else
 243                 submit_bio(WRITE, bio);
 244 }
 245
 246 static int can_be_merged(struct bio *bio, sector_t sector)
 247 {
 248         if (bio == NULL)
 249                 return 0;
 250
 251         return bio_end_sector(bio) == sector ? 1 : 0;
 252 }
 253
 254 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 255                       struct osd_iobuf *iobuf)
 256 {
 257         int            blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 258         struct page  **pages = iobuf->dr_pages;
 259         int            npages = iobuf->dr_npages;
 260         sector_t      *blocks = iobuf->dr_blocks;
 261         int            total_blocks = npages * blocks_per_page;
 262         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 263         unsigned int   blocksize = inode->i_sb->s_blocksize;
 264         struct bio    *bio = NULL;
 265         struct page   *page;
 266         unsigned int   page_offset;
 267         sector_t       sector;
 268         int            nblocks;
 269         int            block_idx;
 270         int            page_idx;
 271         int            i;
 272         int            rc = 0;
 273         ENTRY;
 274
 275         LASSERT(iobuf->dr_npages == npages);
 276
 277         osd_brw_stats_update(osd, iobuf);
 278         iobuf->dr_start_time = cfs_time_current();
 279
 280         for (page_idx = 0, block_idx = 0;
 281              page_idx < npages;
 282              page_idx++, block_idx += blocks_per_page) {
 283
 284                 page = pages[page_idx];
 285                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 286
 287                 for (i = 0, page_offset = 0;
 288                      i < blocks_per_page;
 289                      i += nblocks, page_offset += blocksize * nblocks) {
 290
 291                         nblocks = 1;
 292
 293                         if (blocks[block_idx + i] == 0) {  /* hole */
 294                                 LASSERTF(iobuf->dr_rw == 0,
 295                                          "page_idx %u, block_idx %u, i %u\n",
 296                                          page_idx, block_idx, i);
 297                                 memset(kmap(page) + page_offset, 0, blocksize);
 298                                 kunmap(page);
 299                                 continue;
 300                         }
 301
 302                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 303
 304                         /* Additional contiguous file blocks? */
 305                         while (i + nblocks < blocks_per_page &&
 306                                (sector + (nblocks << sector_bits)) ==
 307                                ((sector_t)blocks[block_idx + i + nblocks] <<
 308                                 sector_bits))
 309                                 nblocks++;
 310
 311                         if (bio != NULL &&
 312                             can_be_merged(bio, sector) &&
 313                             bio_add_page(bio, page,
 314                                          blocksize * nblocks, page_offset) != 0)
 315                                 continue;       /* added this frag OK */
 316
 317                         if (bio != NULL) {
 318                                 struct request_queue *q =
 319                                         bdev_get_queue(bio->bi_bdev);
 320                                 unsigned int bi_size = bio_sectors(bio) << 9;
 321
 322                                 /* Dang! I have to fragment this I/O */
 323                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 324                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 325                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 326                                        bio_sectors(bio),
 327                                        queue_max_sectors(q),
 328                                        bio_phys_segments(q, bio),
 329                                        queue_max_phys_segments(q),
 330                                        0, queue_max_hw_segments(q));
 331                                 record_start_io(iobuf, bi_size);
 332                                 osd_submit_bio(iobuf->dr_rw, bio);
 333                         }
 334
 335                         /* allocate new bio */
 336                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 337                                                       (npages - page_idx) *
 338                                                       blocks_per_page));
 339                         if (bio == NULL) {
 340                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 341                                        (npages - page_idx), blocks_per_page,
 342                                        (npages - page_idx) * blocks_per_page);
 343                                 rc = -ENOMEM;
 344                                 goto out;
 345                         }
 346
 347                         bio->bi_bdev = inode->i_sb->s_bdev;
 348                         bio_set_sector(bio, sector);
 349                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 350                         bio->bi_end_io = dio_complete_routine;
 351                         bio->bi_private = iobuf;
 352
 353                         rc = bio_add_page(bio, page,
 354                                           blocksize * nblocks, page_offset);
 355                         LASSERT(rc != 0);
 356                 }
 357         }
 358
 359         if (bio != NULL) {
 360                 record_start_io(iobuf, bio_sectors(bio) << 9);
 361                 osd_submit_bio(iobuf->dr_rw, bio);
 362                 rc = 0;
 363         }
 364
 365 out:
 366         /* in order to achieve better IO throughput, we don't wait for writes
 367          * completion here. instead we proceed with transaction commit in
 368          * parallel and wait for IO completion once transaction is stopped
 369          * see osd_trans_stop() for more details -bzzz */
 370         if (iobuf->dr_rw == 0) {
 371                 wait_event(iobuf->dr_wait,
 372                            atomic_read(&iobuf->dr_numreqs) == 0);
 373                 osd_fini_iobuf(osd, iobuf);
 374         }
 375
 376         if (rc == 0)
 377                 rc = iobuf->dr_error;
 378         RETURN(rc);
 379 }
 380
 381 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 382                                    struct niobuf_local *lnb)
 383 {
 384         ENTRY;
 385
 386         *nrpages = 0;
 387
 388         while (len > 0) {
 389                 int poff = offset & (PAGE_SIZE - 1);
 390                 int plen = PAGE_SIZE - poff;
 391
 392                 if (plen > len)
 393                         plen = len;
 394                 lnb->lnb_file_offset = offset;
 395                 lnb->lnb_page_offset = poff;
 396                 lnb->lnb_len = plen;
 397                 /* lnb->lnb_flags = rnb->rnb_flags; */
 398                 lnb->lnb_flags = 0;
 399                 lnb->lnb_page = NULL;
 400                 lnb->lnb_rc = 0;
 401
 402                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 403                          (long long) len);
 404                 offset += plen;
 405                 len -= plen;
 406                 lnb++;
 407                 (*nrpages)++;
 408         }
 409
 410         RETURN(0);
 411 }
 412
 413 static struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
 414 {
 415         struct inode      *inode = osd_dt_obj(dt)->oo_inode;
 416         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 417         struct page       *page;
 418
 419         LASSERT(inode);
 420
 421         page = find_or_create_page(inode->i_mapping, offset >> PAGE_SHIFT,
 422                                    GFP_NOFS | __GFP_HIGHMEM);
 423         if (unlikely(page == NULL))
 424                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 425
 426         return page;
 427 }
 428
 429 /*
 430  * there are following "locks":
 431  * journal_start
 432  * i_mutex
 433  * page lock
 434  *
 435  * osd write path:
 436  *  - lock page(s)
 437  *  - journal_start
 438  *  - truncate_sem
 439  *
 440  * ext4 vmtruncate:
 441  *  - lock pages, unlock
 442  *  - journal_start
 443  *  - lock partial page
 444  *  - i_data_sem
 445  *
 446  */
 447
 448 /**
 449  * Unlock and release pages loaded by osd_bufs_get()
 450  *
 451  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 452  *
 453  * \param env           thread execution environment
 454  * \param dt            dt object undergoing IO (OSD object + methods)
 455  * \param lnb           array of pages undergoing IO
 456  * \param npages        number of pages in \a lnb
 457  *
 458  * \retval 0            always
 459  */
 460 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 461                         struct niobuf_local *lnb, int npages)
 462 {
 463         int i;
 464
 465         for (i = 0; i < npages; i++) {
 466                 if (lnb[i].lnb_page == NULL)
 467                         continue;
 468                 LASSERT(PageLocked(lnb[i].lnb_page));
 469                 unlock_page(lnb[i].lnb_page);
 470                 put_page(lnb[i].lnb_page);
 471                 lu_object_put(env, &dt->do_lu);
 472                 lnb[i].lnb_page = NULL;
 473         }
 474
 475         RETURN(0);
 476 }
 477
 478 /**
 479  * Load and lock pages undergoing IO
 480  *
 481  * Pages as described in the \a lnb array are fetched (from disk or cache)
 482  * and locked for IO by the caller.
 483  *
 484  * DLM locking protects us from write and truncate competing for same region,
 485  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 486  * It's possible the writeout on a such a page is in progress when we access
 487  * it. It's also possible that during this writeout we put new (partial) data
 488  * into the page, but won't be able to proceed in filter_commitrw_write().
 489  * Therefore, just wait for writeout completion as it should be rare enough.
 490  *
 491  * \param env           thread execution environment
 492  * \param dt            dt object undergoing IO (OSD object + methods)
 493  * \param pos           byte offset of IO start
 494  * \param len           number of bytes of IO
 495  * \param lnb           array of extents undergoing IO
 496  * \param rw            read or write operation?
 497  * \param capa          capabilities
 498  *
 499  * \retval pages        (zero or more) loaded successfully
 500  * \retval -ENOMEM      on memory/page allocation error
 501  */
 502 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 503                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 504                         int rw)
 505 {
 506         struct osd_object   *obj    = osd_dt_obj(dt);
 507         int npages, i, rc = 0;
 508
 509         LASSERT(obj->oo_inode);
 510
 511         osd_map_remote_to_local(pos, len, &npages, lnb);
 512
 513         for (i = 0; i < npages; i++, lnb++) {
 514                 lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset, rw);
 515                 if (lnb->lnb_page == NULL)
 516                         GOTO(cleanup, rc = -ENOMEM);
 517
 518                 wait_on_page_writeback(lnb->lnb_page);
 519                 BUG_ON(PageWriteback(lnb->lnb_page));
 520
 521                 lu_object_get(&dt->do_lu);
 522         }
 523
 524         RETURN(i);
 525
 526 cleanup:
 527         if (i > 0)
 528                 osd_bufs_put(env, dt, lnb - i, i);
 529         return rc;
 530 }
 531
 532 #ifndef HAVE_LDISKFS_MAP_BLOCKS
 533
 534 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 535 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 536 #endif
 537
 538 struct bpointers {
 539         sector_t *blocks;
 540         unsigned long start;
 541         int num;
 542         int init_num;
 543         int create;
 544 };
 545
 546 static long ldiskfs_ext_find_goal(struct inode *inode,
 547                                   struct ldiskfs_ext_path *path,
 548                                   unsigned long block, int *aflags)
 549 {
 550         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 551         unsigned long bg_start;
 552         unsigned long colour;
 553         int depth;
 554
 555         if (path) {
 556                 struct ldiskfs_extent *ex;
 557                 depth = path->p_depth;
 558
 559                 /* try to predict block placement */
 560                 if ((ex = path[depth].p_ext))
 561                         return ldiskfs_ext_pblock(ex) +
 562                                 (block - le32_to_cpu(ex->ee_block));
 563
 564                 /* it looks index is empty
 565                  * try to find starting from index itself */
 566                 if (path[depth].p_bh)
 567                         return path[depth].p_bh->b_blocknr;
 568         }
 569
 570         /* OK. use inode's group */
 571         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 572                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 573         colour = (current->pid % 16) *
 574                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 575         return bg_start + colour + block;
 576 }
 577
 578 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 579                                 struct ldiskfs_ext_path *path,
 580                                 unsigned long block, unsigned long *count,
 581                                 int *err)
 582 {
 583         struct ldiskfs_allocation_request ar;
 584         unsigned long pblock;
 585         int aflags;
 586
 587         /* find neighbour allocated blocks */
 588         ar.lleft = block;
 589         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 590         if (*err)
 591                 return 0;
 592         ar.lright = block;
 593         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 594         if (*err)
 595                 return 0;
 596
 597         /* allocate new block */
 598         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 599         ar.inode = inode;
 600         ar.logical = block;
 601         ar.len = *count;
 602         ar.flags = LDISKFS_MB_HINT_DATA;
 603         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 604         *count = ar.len;
 605         return pblock;
 606 }
 607
 608 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 609                                      struct ldiskfs_ext_path *path,
 610                                      struct ldiskfs_ext_cache *cex,
 611 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 612                                      struct ldiskfs_extent *ex,
 613 #endif
 614                                      void *cbdata)
 615 {
 616         struct bpointers *bp = cbdata;
 617         struct ldiskfs_extent nex;
 618         unsigned long pblock = 0;
 619         unsigned long tgen;
 620         int err, i;
 621         unsigned long count;
 622         handle_t *handle;
 623
 624 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 625         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 626 #else
 627         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 628 #endif
 629                 err = EXT_CONTINUE;
 630                 goto map;
 631         }
 632
 633         if (bp->create == 0) {
 634                 i = 0;
 635                 if (cex->ec_block < bp->start)
 636                         i = bp->start - cex->ec_block;
 637                 if (i >= cex->ec_len)
 638                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 639                                         i, cex->ec_len);
 640                 for (; i < cex->ec_len && bp->num; i++) {
 641                         *(bp->blocks) = 0;
 642                         bp->blocks++;
 643                         bp->num--;
 644                         bp->start++;
 645                 }
 646
 647                 return EXT_CONTINUE;
 648         }
 649
 650         tgen = LDISKFS_I(inode)->i_ext_generation;
 651         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
 652
 653         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
 654                                    count + LDISKFS_ALLOC_NEEDED + 1);
 655         if (IS_ERR(handle)) {
 656                 return PTR_ERR(handle);
 657         }
 658
 659         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
 660                 /* the tree has changed. so path can be invalid at moment */
 661                 ldiskfs_journal_stop(handle);
 662                 return EXT_REPEAT;
 663         }
 664
 665         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
 666          * protected by i_data_sem as whole. so we patch it to store
 667          * generation to path and now verify the tree hasn't changed */
 668         down_write((&LDISKFS_I(inode)->i_data_sem));
 669
 670         /* validate extent, make sure the extent tree does not changed */
 671         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
 672                 /* cex is invalid, try again */
 673                 up_write(&LDISKFS_I(inode)->i_data_sem);
 674                 ldiskfs_journal_stop(handle);
 675                 return EXT_REPEAT;
 676         }
 677
 678         count = cex->ec_len;
 679         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
 680         if (!pblock)
 681                 goto out;
 682         BUG_ON(count > cex->ec_len);
 683
 684         /* insert new extent */
 685         nex.ee_block = cpu_to_le32(cex->ec_block);
 686         ldiskfs_ext_store_pblock(&nex, pblock);
 687         nex.ee_len = cpu_to_le16(count);
 688         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
 689         if (err) {
 690                 /* free data blocks we just allocated */
 691                 /* not a good idea to call discard here directly,
 692                  * but otherwise we'd need to call it every free() */
 693                 ldiskfs_discard_preallocations(inode);
 694 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
 695                 ldiskfs_free_blocks(handle, inode, NULL,
 696                                     ldiskfs_ext_pblock(&nex),
 697                                     le16_to_cpu(nex.ee_len), 0);
 698 #else
 699                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
 700                                     le16_to_cpu(nex.ee_len), 0);
 701 #endif
 702                 goto out;
 703         }
 704
 705         /*
 706          * Putting len of the actual extent we just inserted,
 707          * we are asking ldiskfs_ext_walk_space() to continue
 708          * scaning after that block
 709          */
 710         cex->ec_len = le16_to_cpu(nex.ee_len);
 711         cex->ec_start = ldiskfs_ext_pblock(&nex);
 712         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
 713         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
 714
 715 out:
 716         up_write((&LDISKFS_I(inode)->i_data_sem));
 717         ldiskfs_journal_stop(handle);
 718 map:
 719         if (err >= 0) {
 720                 /* map blocks */
 721                 if (bp->num == 0) {
 722                         CERROR("hmm. why do we find this extent?\n");
 723                         CERROR("initial space: %lu:%u\n",
 724                                 bp->start, bp->init_num);
 725 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 726                         CERROR("current extent: %u/%u/%llu %d\n",
 727                                 cex->ec_block, cex->ec_len,
 728                                 (unsigned long long)cex->ec_start,
 729                                 cex->ec_type);
 730 #else
 731                         CERROR("current extent: %u/%u/%llu\n",
 732                                 cex->ec_block, cex->ec_len,
 733                                 (unsigned long long)cex->ec_start);
 734 #endif
 735                 }
 736                 i = 0;
 737                 if (cex->ec_block < bp->start)
 738                         i = bp->start - cex->ec_block;
 739                 if (i >= cex->ec_len)
 740                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 741                                         i, cex->ec_len);
 742                 for (; i < cex->ec_len && bp->num; i++) {
 743                         *(bp->blocks) = cex->ec_start + i;
 744                         if (pblock != 0) {
 745                                 /* unmap any possible underlying metadata from
 746                                  * the block device mapping.  bug 6998. */
 747                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
 748                                                           *(bp->blocks));
 749                         }
 750                         bp->blocks++;
 751                         bp->num--;
 752                         bp->start++;
 753                 }
 754         }
 755         return err;
 756 }
 757
 758 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
 759                                    int clen, sector_t *blocks, int create)
 760 {
 761         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 762         struct bpointers bp;
 763         int err;
 764
 765         if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
 766                 return -EFBIG;
 767
 768         bp.blocks = blocks;
 769         bp.start = index * blocks_per_page;
 770         bp.init_num = bp.num = clen * blocks_per_page;
 771         bp.create = create;
 772
 773         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 774                bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
 775
 776         err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
 777                                      ldiskfs_ext_new_extent_cb, &bp);
 778         ldiskfs_ext_invalidate_cache(inode);
 779
 780         return err;
 781 }
 782
 783 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
 784                                           struct page **page, int pages,
 785                                           sector_t *blocks, int create)
 786 {
 787         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 788         pgoff_t bitmap_max_page_index;
 789         sector_t *b;
 790         int rc = 0, i;
 791
 792         bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
 793                                 PAGE_SHIFT;
 794         for (i = 0, b = blocks; i < pages; i++, page++) {
 795                 if ((*page)->index + 1 >= bitmap_max_page_index) {
 796                         rc = -EFBIG;
 797                         break;
 798                 }
 799                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
 800                 if (rc) {
 801                         CERROR("ino %lu, blk %llu create %d: rc %d\n",
 802                                inode->i_ino,
 803                                (unsigned long long)*b, create, rc);
 804                         break;
 805                 }
 806                 b += blocks_per_page;
 807         }
 808         return rc;
 809 }
 810
 811 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
 812                                            struct page **page,
 813                                            int pages, sector_t *blocks,
 814                                            int create)
 815 {
 816         int rc = 0, i = 0, clen = 0;
 817         struct page *fp = NULL;
 818
 819         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 820                 inode->i_ino, pages, (*page)->index);
 821
 822         /* pages are sorted already. so, we just have to find
 823          * contig. space and process them properly */
 824         while (i < pages) {
 825                 if (fp == NULL) {
 826                         /* start new extent */
 827                         fp = *page++;
 828                         clen = 1;
 829                         i++;
 830                         continue;
 831                 } else if (fp->index + clen == (*page)->index) {
 832                         /* continue the extent */
 833                         page++;
 834                         clen++;
 835                         i++;
 836                         continue;
 837                 }
 838
 839                 /* process found extent */
 840                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 841                                              blocks, create);
 842                 if (rc)
 843                         GOTO(cleanup, rc);
 844
 845                 /* look for next extent */
 846                 fp = NULL;
 847                 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
 848         }
 849
 850         if (fp)
 851                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 852                                              blocks, create);
 853
 854 cleanup:
 855         return rc;
 856 }
 857
 858 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 859                                        int pages, sector_t *blocks,
 860                                        int create)
 861 {
 862         int rc;
 863
 864         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 865                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
 866                                                      blocks, create);
 867                 return rc;
 868         }
 869         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
 870
 871         return rc;
 872 }
 873 #else
 874 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 875                                        int pages, sector_t *blocks,
 876                                        int create)
 877 {
 878         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 879         int rc = 0, i = 0;
 880         struct page *fp = NULL;
 881         int clen = 0;
 882         pgoff_t max_page_index;
 883
 884         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
 885
 886         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 887                 inode->i_ino, pages, (*page)->index);
 888
 889         /* pages are sorted already. so, we just have to find
 890          * contig. space and process them properly */
 891         while (i < pages) {
 892                 long blen, total = 0;
 893                 handle_t *handle = NULL;
 894                 struct ldiskfs_map_blocks map = { 0 };
 895
 896                 if (fp == NULL) { /* start new extent */
 897                         fp = *page++;
 898                         clen = 1;
 899                         if (++i != pages)
 900                                 continue;
 901                 } else if (fp->index + clen == (*page)->index) {
 902                         /* continue the extent */
 903                         page++;
 904                         clen++;
 905                         if (++i != pages)
 906                                 continue;
 907                 }
 908                 if (fp->index + clen >= max_page_index)
 909                         GOTO(cleanup, rc = -EFBIG);
 910                 /* process found extent */
 911                 map.m_lblk = fp->index * blocks_per_page;
 912                 map.m_len = blen = clen * blocks_per_page;
 913                 if (create) {
 914                         create = LDISKFS_GET_BLOCKS_CREATE;
 915                         handle = ldiskfs_journal_current_handle();
 916                         LASSERT(handle != NULL);
 917                 }
 918 cont_map:
 919                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 920                 if (rc >= 0) {
 921                         int c = 0;
 922                         for (; total < blen && c < map.m_len; c++, total++) {
 923                                 if (rc == 0) {
 924                                         *(blocks + total) = 0;
 925                                         total++;
 926                                         break;
 927                                 } else {
 928                                         *(blocks + total) = map.m_pblk + c;
 929                                         /* unmap any possible underlying
 930                                          * metadata from the block device
 931                                          * mapping.  bug 6998. */
 932                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
 933                                             create)
 934                                                 unmap_underlying_metadata(
 935                                                         inode->i_sb->s_bdev,
 936                                                         map.m_pblk + c);
 937                                 }
 938                         }
 939                         rc = 0;
 940                 }
 941                 if (rc == 0 && total < blen) {
 942                         map.m_lblk = fp->index * blocks_per_page + total;
 943                         map.m_len = blen - total;
 944                         goto cont_map;
 945                 }
 946                 if (rc != 0)
 947                         GOTO(cleanup, rc);
 948
 949                 /* look for next extent */
 950                 fp = NULL;
 951                 blocks += blocks_per_page * clen;
 952         }
 953 cleanup:
 954         return rc;
 955 }
 956 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
 957
 958 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 959                           struct niobuf_local *lnb, int npages)
 960 {
 961         struct osd_thread_info *oti   = osd_oti_get(env);
 962         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 963         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 964         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 965         struct timeval          start;
 966         struct timeval          end;
 967         unsigned long           timediff;
 968         ssize_t                 isize;
 969         __s64                   maxidx;
 970         int                     rc = 0;
 971         int                     i;
 972         int                     cache = 0;
 973
 974         LASSERT(inode);
 975
 976         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 977         if (unlikely(rc != 0))
 978                 RETURN(rc);
 979
 980         isize = i_size_read(inode);
 981         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
 982
 983         if (osd->od_writethrough_cache)
 984                 cache = 1;
 985         if (isize > osd->od_readcache_max_filesize)
 986                 cache = 0;
 987
 988         do_gettimeofday(&start);
 989         for (i = 0; i < npages; i++) {
 990
 991                 if (cache == 0)
 992                         generic_error_remove_page(inode->i_mapping,
 993                                                   lnb[i].lnb_page);
 994
 995                 /*
 996                  * till commit the content of the page is undefined
 997                  * we'll set it uptodate once bulk is done. otherwise
 998                  * subsequent reads can access non-stable data
 999                  */
1000                 ClearPageUptodate(lnb[i].lnb_page);
1001
1002                 if (lnb[i].lnb_len == PAGE_SIZE)
1003                         continue;
1004
1005                 if (maxidx >= lnb[i].lnb_page->index) {
1006                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1007                 } else {
1008                         long off;
1009                         char *p = kmap(lnb[i].lnb_page);
1010
1011                         off = lnb[i].lnb_page_offset;
1012                         if (off)
1013                                 memset(p, 0, off);
1014                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1015                               ~PAGE_MASK;
1016                         if (off)
1017                                 memset(p + off, 0, PAGE_SIZE - off);
1018                         kunmap(lnb[i].lnb_page);
1019                 }
1020         }
1021         do_gettimeofday(&end);
1022         timediff = cfs_timeval_sub(&end, &start, NULL);
1023         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1024
1025         if (iobuf->dr_npages) {
1026                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1027                                                  iobuf->dr_npages,
1028                                                  iobuf->dr_blocks, 0);
1029                 if (likely(rc == 0)) {
1030                         rc = osd_do_bio(osd, inode, iobuf);
1031                         /* do IO stats for preparation reads */
1032                         osd_fini_iobuf(osd, iobuf);
1033                 }
1034         }
1035         RETURN(rc);
1036 }
1037
1038 struct osd_fextent {
1039         sector_t        start;
1040         sector_t        end;
1041         unsigned int    mapped:1;
1042 };
1043
1044 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1045                          struct osd_fextent *cached_extent)
1046 {
1047         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1048         sector_t block = offset >> inode->i_blkbits;
1049         sector_t start;
1050         struct fiemap_extent_info fei = { 0 };
1051         struct fiemap_extent fe = { 0 };
1052         mm_segment_t saved_fs;
1053         int rc;
1054
1055         if (block >= cached_extent->start && block < cached_extent->end)
1056                 return cached_extent->mapped;
1057
1058         if (i_size_read(inode) == 0)
1059                 return 0;
1060
1061         /* Beyond EOF, must not be mapped */
1062         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1063                 return 0;
1064
1065         fei.fi_extents_max = 1;
1066         fei.fi_extents_start = &fe;
1067
1068         saved_fs = get_fs();
1069         set_fs(get_ds());
1070         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1071         set_fs(saved_fs);
1072         if (rc != 0)
1073                 return 0;
1074
1075         start = fe.fe_logical >> inode->i_blkbits;
1076
1077         if (start > block) {
1078                 cached_extent->start = block;
1079                 cached_extent->end = start;
1080                 cached_extent->mapped = 0;
1081         } else {
1082                 cached_extent->start = start;
1083                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1084                                       inode->i_blkbits;
1085                 cached_extent->mapped = 1;
1086         }
1087
1088         return cached_extent->mapped;
1089 }
1090
1091 static int osd_declare_write_commit(const struct lu_env *env,
1092                                     struct dt_object *dt,
1093                                     struct niobuf_local *lnb, int npages,
1094                                     struct thandle *handle)
1095 {
1096         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1097         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1098         struct osd_thandle      *oh;
1099         int                      extents = 1;
1100         int                      depth;
1101         int                      i;
1102         int                      newblocks;
1103         int                      rc = 0;
1104         int                      flags = 0;
1105         int                      credits = 0;
1106         bool                     ignore_quota = false;
1107         long long                quota_space = 0;
1108         struct osd_fextent       extent = { 0 };
1109         ENTRY;
1110
1111         LASSERT(handle != NULL);
1112         oh = container_of0(handle, struct osd_thandle, ot_super);
1113         LASSERT(oh->ot_handle == NULL);
1114
1115         newblocks = npages;
1116
1117         /* calculate number of extents (probably better to pass nb) */
1118         for (i = 0; i < npages; i++) {
1119                 if (i && lnb[i].lnb_file_offset !=
1120                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1121                         extents++;
1122
1123                 if (!osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1124                         quota_space += PAGE_SIZE;
1125
1126                 /* ignore quota for the whole request if any page is from
1127                  * client cache or written by root.
1128                  *
1129                  * XXX once we drop the 1.8 client support, the checking
1130                  * for whether page is from cache can be simplified as:
1131                  * !(lnb[i].flags & OBD_BRW_SYNC)
1132                  *
1133                  * XXX we could handle this on per-lnb basis as done by
1134                  * grant. */
1135                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1136                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1137                     OBD_BRW_FROM_GRANT)
1138                         ignore_quota = true;
1139         }
1140
1141         /*
1142          * each extent can go into new leaf causing a split
1143          * 5 is max tree depth: inode + 4 index blocks
1144          * with blockmaps, depth is 3 at most
1145          */
1146         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1147                 /*
1148                  * many concurrent threads may grow tree by the time
1149                  * our transaction starts. so, consider 2 is a min depth
1150                  */
1151                 depth = ext_depth(inode);
1152                 depth = max(depth, 1) + 1;
1153                 newblocks += depth;
1154                 credits++; /* inode */
1155                 credits += depth * 2 * extents;
1156         } else {
1157                 depth = 3;
1158                 newblocks += depth;
1159                 credits++; /* inode */
1160                 credits += depth * extents;
1161         }
1162
1163         /* quota space for metadata blocks */
1164         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1165
1166         /* quota space should be reported in 1K blocks */
1167         quota_space = toqb(quota_space);
1168
1169         /* each new block can go in different group (bitmap + gd) */
1170
1171         /* we can't dirty more bitmap blocks than exist */
1172         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1173                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1174         else
1175                 credits += newblocks;
1176
1177         /* we can't dirty more gd blocks than exist */
1178         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1179                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1180         else
1181                 credits += newblocks;
1182
1183         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1184
1185         /* make sure the over quota flags were not set */
1186         lnb[0].lnb_flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA);
1187
1188         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1189                                    quota_space, oh, osd_dt_obj(dt), true,
1190                                    &flags, ignore_quota);
1191
1192         /* we need only to store the overquota flags in the first lnb for
1193          * now, once we support multiple objects BRW, this code needs be
1194          * revised. */
1195         if (flags & QUOTA_FL_OVER_USRQUOTA)
1196                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1197         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1198                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1199
1200         RETURN(rc);
1201 }
1202
1203 /* Check if a block is allocated or not */
1204 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1205                             struct niobuf_local *lnb, int npages,
1206                             struct thandle *thandle)
1207 {
1208         struct osd_thread_info *oti = osd_oti_get(env);
1209         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1210         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1211         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1212         loff_t isize;
1213         int rc = 0, i;
1214         struct osd_fextent extent = { 0 };
1215
1216         LASSERT(inode);
1217
1218         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1219         if (unlikely(rc != 0))
1220                 RETURN(rc);
1221
1222         isize = i_size_read(inode);
1223         ll_vfs_dq_init(inode);
1224
1225         for (i = 0; i < npages; i++) {
1226                 if (lnb[i].lnb_rc == -ENOSPC &&
1227                     osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent)) {
1228                         /* Allow the write to proceed if overwriting an
1229                          * existing block */
1230                         lnb[i].lnb_rc = 0;
1231                 }
1232
1233                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1234                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1235                                lnb[i].lnb_rc);
1236                         LASSERT(lnb[i].lnb_page);
1237                         generic_error_remove_page(inode->i_mapping,
1238                                                   lnb[i].lnb_page);
1239                         continue;
1240                 }
1241
1242                 LASSERT(PageLocked(lnb[i].lnb_page));
1243                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1244
1245                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1246                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1247
1248                 /*
1249                  * Since write and truncate are serialized by oo_sem, even
1250                  * partial-page truncate should not leave dirty pages in the
1251                  * page cache.
1252                  */
1253                 LASSERT(!PageDirty(lnb[i].lnb_page));
1254
1255                 SetPageUptodate(lnb[i].lnb_page);
1256
1257                 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1258         }
1259
1260         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1261
1262         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1263                 rc = -ENOSPC;
1264         } else if (iobuf->dr_npages > 0) {
1265                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1266                                                  iobuf->dr_npages,
1267                                                  iobuf->dr_blocks, 1);
1268         } else {
1269                 /* no pages to write, no transno is needed */
1270                 thandle->th_local = 1;
1271         }
1272
1273         if (likely(rc == 0)) {
1274                 spin_lock(&inode->i_lock);
1275                 if (isize > i_size_read(inode)) {
1276                         i_size_write(inode, isize);
1277                         LDISKFS_I(inode)->i_disksize = isize;
1278                         spin_unlock(&inode->i_lock);
1279                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1280                 } else {
1281                         spin_unlock(&inode->i_lock);
1282                 }
1283
1284                 rc = osd_do_bio(osd, inode, iobuf);
1285                 /* we don't do stats here as in read path because
1286                  * write is async: we'll do this in osd_put_bufs() */
1287         } else {
1288                 osd_fini_iobuf(osd, iobuf);
1289         }
1290
1291         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1292
1293         if (unlikely(rc != 0)) {
1294                 /* if write fails, we should drop pages from the cache */
1295                 for (i = 0; i < npages; i++) {
1296                         if (lnb[i].lnb_page == NULL)
1297                                 continue;
1298                         LASSERT(PageLocked(lnb[i].lnb_page));
1299                         generic_error_remove_page(inode->i_mapping,
1300                                                   lnb[i].lnb_page);
1301                 }
1302         }
1303
1304         RETURN(rc);
1305 }
1306
1307 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1308                          struct niobuf_local *lnb, int npages)
1309 {
1310         struct osd_thread_info *oti = osd_oti_get(env);
1311         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1312         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1313         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1314         struct timeval start, end;
1315         unsigned long timediff;
1316         int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1317         loff_t isize;
1318
1319         LASSERT(inode);
1320
1321         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1322         if (unlikely(rc != 0))
1323                 RETURN(rc);
1324
1325         isize = i_size_read(inode);
1326
1327         if (osd->od_read_cache)
1328                 cache = 1;
1329         if (isize > osd->od_readcache_max_filesize)
1330                 cache = 0;
1331
1332         do_gettimeofday(&start);
1333         for (i = 0; i < npages; i++) {
1334
1335                 if (isize <= lnb[i].lnb_file_offset)
1336                         /* If there's no more data, abort early.
1337                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1338                         break;
1339
1340                 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1341                         lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1342                 else
1343                         lnb[i].lnb_rc = lnb[i].lnb_len;
1344
1345                 if (PageUptodate(lnb[i].lnb_page)) {
1346                         cache_hits++;
1347                 } else {
1348                         cache_misses++;
1349                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1350                 }
1351
1352                 if (cache == 0)
1353                         generic_error_remove_page(inode->i_mapping,
1354                                                   lnb[i].lnb_page);
1355         }
1356         do_gettimeofday(&end);
1357         timediff = cfs_timeval_sub(&end, &start, NULL);
1358         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1359
1360         if (cache_hits != 0)
1361                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1362                                     cache_hits);
1363         if (cache_misses != 0)
1364                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1365                                     cache_misses);
1366         if (cache_hits + cache_misses != 0)
1367                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1368                                     cache_hits + cache_misses);
1369
1370         if (iobuf->dr_npages) {
1371                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1372                                                  iobuf->dr_npages,
1373                                                  iobuf->dr_blocks, 0);
1374                 rc = osd_do_bio(osd, inode, iobuf);
1375
1376                 /* IO stats will be done in osd_bufs_put() */
1377         }
1378
1379         RETURN(rc);
1380 }
1381
1382 /*
1383  * XXX: Another layering violation for now.
1384  *
1385  * We don't want to use ->f_op->read methods, because generic file write
1386  *
1387  *         - serializes on ->i_sem, and
1388  *
1389  *         - does a lot of extra work like balance_dirty_pages(),
1390  *
1391  * which doesn't work for globally shared files like /last_rcvd.
1392  */
1393 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1394 {
1395         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1396
1397         memcpy(buffer, (char *)ei->i_data, buflen);
1398
1399         return  buflen;
1400 }
1401
1402 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1403 {
1404         struct buffer_head *bh;
1405         unsigned long block;
1406         int osize;
1407         int blocksize;
1408         int csize;
1409         int boffs;
1410
1411         /* prevent reading after eof */
1412         spin_lock(&inode->i_lock);
1413         if (i_size_read(inode) < *offs + size) {
1414                 loff_t diff = i_size_read(inode) - *offs;
1415                 spin_unlock(&inode->i_lock);
1416                 if (diff < 0) {
1417                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1418                                i_size_read(inode), *offs);
1419                         return -EBADR;
1420                 } else if (diff == 0) {
1421                         return 0;
1422                 } else {
1423                         size = diff;
1424                 }
1425         } else {
1426                 spin_unlock(&inode->i_lock);
1427         }
1428
1429         blocksize = 1 << inode->i_blkbits;
1430         osize = size;
1431         while (size > 0) {
1432                 block = *offs >> inode->i_blkbits;
1433                 boffs = *offs & (blocksize - 1);
1434                 csize = min(blocksize - boffs, size);
1435                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1436                 if (IS_ERR(bh)) {
1437                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1438                                LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
1439                                csize, *offs, inode->i_ino,
1440                                PTR_ERR(bh));
1441                         return PTR_ERR(bh);
1442                 }
1443
1444                 if (bh != NULL) {
1445                         memcpy(buf, bh->b_data + boffs, csize);
1446                         brelse(bh);
1447                 } else {
1448                         memset(buf, 0, csize);
1449                 }
1450
1451                 *offs += csize;
1452                 buf += csize;
1453                 size -= csize;
1454         }
1455         return osize;
1456 }
1457
1458 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1459                         struct lu_buf *buf, loff_t *pos)
1460 {
1461         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1462         int           rc;
1463
1464         /* Read small symlink from inode body as we need to maintain correct
1465          * on-disk symlinks for ldiskfs.
1466          */
1467         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1468             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1469                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1470         else
1471                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1472
1473         return rc;
1474 }
1475
1476 static inline int osd_extents_enabled(struct super_block *sb,
1477                                       struct inode *inode)
1478 {
1479         if (inode != NULL) {
1480                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1481                         return 1;
1482         } else if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
1483                                 LDISKFS_FEATURE_INCOMPAT_EXTENTS)) {
1484                 return 1;
1485         }
1486         return 0;
1487 }
1488
1489 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1490                            const loff_t size, const loff_t pos,
1491                            const int blocks)
1492 {
1493         int credits, bits, bs, i;
1494
1495         bits = sb->s_blocksize_bits;
1496         bs = 1 << bits;
1497
1498         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1499          * we do not expect blockmaps on the large files,
1500          * so let's shrink it to 2 levels (4GB files) */
1501
1502         /* this is default reservation: 2 levels */
1503         credits = (blocks + 2) * 3;
1504
1505         /* actual offset is unknown, hard to optimize */
1506         if (pos == -1)
1507                 return credits;
1508
1509         /* now check for few specific cases to optimize */
1510         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1511                 /* no indirects */
1512                 credits = blocks;
1513                 /* allocate if not allocated */
1514                 if (inode == NULL) {
1515                         credits += blocks * 2;
1516                         return credits;
1517                 }
1518                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1519                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1520                         if (LDISKFS_I(inode)->i_data[i] == 0)
1521                                 credits += 2;
1522                 }
1523         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1524                 /* single indirect */
1525                 credits = blocks * 3;
1526                 if (inode == NULL ||
1527                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1528                         credits += 3;
1529                 else
1530                         /* The indirect block may be modified. */
1531                         credits += 1;
1532         }
1533
1534         return credits;
1535 }
1536
1537 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1538                                  const struct lu_buf *buf, loff_t _pos,
1539                                  struct thandle *handle)
1540 {
1541         struct osd_object  *obj  = osd_dt_obj(dt);
1542         struct inode       *inode = obj->oo_inode;
1543         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1544         struct osd_thandle *oh;
1545         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1546         int                 bits, bs;
1547         int                 depth, size;
1548         loff_t              pos;
1549         ENTRY;
1550
1551         LASSERT(buf != NULL);
1552         LASSERT(handle != NULL);
1553
1554         oh = container_of0(handle, struct osd_thandle, ot_super);
1555         LASSERT(oh->ot_handle == NULL);
1556
1557         size = buf->lb_len;
1558         bits = sb->s_blocksize_bits;
1559         bs = 1 << bits;
1560
1561         if (_pos == -1) {
1562                 /* if this is an append, then we
1563                  * should expect cross-block record */
1564                 pos = 0;
1565         } else {
1566                 pos = _pos;
1567         }
1568
1569         /* blocks to modify */
1570         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1571         LASSERT(blocks > 0);
1572
1573         if (inode != NULL && _pos != -1) {
1574                 /* object size in blocks */
1575                 est = (i_size_read(inode) + bs - 1) >> bits;
1576                 allocated = inode->i_blocks >> (bits - 9);
1577                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1578                         /* looks like an overwrite, no need to modify tree */
1579                         credits = blocks;
1580                         /* no need to modify i_size */
1581                         goto out;
1582                 }
1583         }
1584
1585         if (osd_extents_enabled(sb, inode)) {
1586                 /*
1587                  * many concurrent threads may grow tree by the time
1588                  * our transaction starts. so, consider 2 is a min depth
1589                  * for every level we may need to allocate a new block
1590                  * and take some entries from the old one. so, 3 blocks
1591                  * to allocate (bitmap, gd, itself) + old block - 4 per
1592                  * level.
1593                  */
1594                 depth = inode != NULL ? ext_depth(inode) : 0;
1595                 depth = max(depth, 1) + 1;
1596                 credits = depth;
1597                 /* if not append, then split may need to modify
1598                  * existing blocks moving entries into the new ones */
1599                 if (_pos == -1)
1600                         credits += depth;
1601                 /* blocks to store data: bitmap,gd,itself */
1602                 credits += blocks * 3;
1603         } else {
1604                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1605         }
1606         /* if inode is created as part of the transaction,
1607          * then it's counted already by the creation method */
1608         if (inode != NULL)
1609                 credits++;
1610
1611 out:
1612
1613         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1614
1615         /* dt_declare_write() is usually called for system objects, such
1616          * as llog or last_rcvd files. We needn't enforce quota on those
1617          * objects, so always set the lqi_space as 0. */
1618         if (inode != NULL)
1619                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1620                                            i_gid_read(inode), 0, oh, obj, true,
1621                                            NULL, false);
1622         RETURN(rc);
1623 }
1624
1625 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1626 {
1627         /* LU-2634: clear the extent format for fast symlink */
1628         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1629
1630         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1631         spin_lock(&inode->i_lock);
1632         LDISKFS_I(inode)->i_disksize = buflen;
1633         i_size_write(inode, buflen);
1634         spin_unlock(&inode->i_lock);
1635         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1636
1637         return 0;
1638 }
1639
1640 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1641                              int write_NUL, loff_t *offs, handle_t *handle)
1642 {
1643         struct buffer_head *bh        = NULL;
1644         loff_t              offset    = *offs;
1645         loff_t              new_size  = i_size_read(inode);
1646         unsigned long       block;
1647         int                 blocksize = 1 << inode->i_blkbits;
1648         int                 err = 0;
1649         int                 size;
1650         int                 boffs;
1651         int                 dirty_inode = 0;
1652
1653         if (write_NUL) {
1654                 /*
1655                  * long symlink write does not count the NUL terminator in
1656                  * bufsize, we write it, and the inode's file size does not
1657                  * count the NUL terminator as well.
1658                  */
1659                 ((char *)buf)[bufsize] = '\0';
1660                 ++bufsize;
1661         }
1662         while (bufsize > 0) {
1663                 if (bh != NULL)
1664                         brelse(bh);
1665
1666                 block = offset >> inode->i_blkbits;
1667                 boffs = offset & (blocksize - 1);
1668                 size = min(blocksize - boffs, bufsize);
1669                 bh = __ldiskfs_bread(handle, inode, block, 1);
1670                 if (IS_ERR_OR_NULL(bh)) {
1671                         if (bh == NULL) {
1672                                 err = -EIO;
1673                         } else {
1674                                 err = PTR_ERR(bh);
1675                                 bh = NULL;
1676                         }
1677                         CERROR("%s: error reading offset %llu (block %lu): "
1678                                "rc = %d\n",
1679                                inode->i_sb->s_id, offset, block, err);
1680                         break;
1681                 }
1682
1683                 err = ldiskfs_journal_get_write_access(handle, bh);
1684                 if (err) {
1685                         CERROR("journal_get_write_access() returned error %d\n",
1686                                err);
1687                         break;
1688                 }
1689                 LASSERTF(boffs + size <= bh->b_size,
1690                          "boffs %d size %d bh->b_size %lu\n",
1691                          boffs, size, (unsigned long)bh->b_size);
1692                 memcpy(bh->b_data + boffs, buf, size);
1693                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1694                 if (err)
1695                         break;
1696
1697                 if (offset + size > new_size)
1698                         new_size = offset + size;
1699                 offset += size;
1700                 bufsize -= size;
1701                 buf += size;
1702         }
1703         if (bh)
1704                 brelse(bh);
1705
1706         if (write_NUL)
1707                 --new_size;
1708         /* correct in-core and on-disk sizes */
1709         if (new_size > i_size_read(inode)) {
1710                 spin_lock(&inode->i_lock);
1711                 if (new_size > i_size_read(inode))
1712                         i_size_write(inode, new_size);
1713                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1714                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1715                         dirty_inode = 1;
1716                 }
1717                 spin_unlock(&inode->i_lock);
1718                 if (dirty_inode)
1719                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1720         }
1721
1722         if (err == 0)
1723                 *offs = offset;
1724         return err;
1725 }
1726
1727 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1728                          const struct lu_buf *buf, loff_t *pos,
1729                          struct thandle *handle, int ignore_quota)
1730 {
1731         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1732         struct osd_thandle      *oh;
1733         ssize_t                 result;
1734         int                     is_link;
1735
1736         LASSERT(dt_object_exists(dt));
1737
1738         LASSERT(handle != NULL);
1739         LASSERT(inode != NULL);
1740         ll_vfs_dq_init(inode);
1741
1742         /* XXX: don't check: one declared chunk can be used many times */
1743         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1744
1745         oh = container_of(handle, struct osd_thandle, ot_super);
1746         LASSERT(oh->ot_handle->h_transaction != NULL);
1747         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1748
1749         /* Write small symlink to inode body as we need to maintain correct
1750          * on-disk symlinks for ldiskfs.
1751          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1752          * does not count it in.
1753          */
1754         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1755         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1756                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1757         else
1758                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1759                                                   buf->lb_len, is_link, pos,
1760                                                   oh->ot_handle);
1761         if (result == 0)
1762                 result = buf->lb_len;
1763
1764         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1765
1766         return result;
1767 }
1768
1769 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1770                              __u64 start, __u64 end, struct thandle *th)
1771 {
1772         struct osd_thandle *oh;
1773         struct inode       *inode;
1774         int                 rc;
1775         ENTRY;
1776
1777         LASSERT(th);
1778         oh = container_of(th, struct osd_thandle, ot_super);
1779
1780         /*
1781          * we don't need to reserve credits for whole truncate
1782          * it's not possible as truncate may need to free too many
1783          * blocks and that won't fit a single transaction. instead
1784          * we reserve credits to change i_size and put inode onto
1785          * orphan list. if needed truncate will extend or restart
1786          * transaction
1787          */
1788         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1789                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1790
1791         inode = osd_dt_obj(dt)->oo_inode;
1792         LASSERT(inode);
1793
1794         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1795                                    0, oh, osd_dt_obj(dt), true, NULL, false);
1796         RETURN(rc);
1797 }
1798
1799 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1800                      __u64 start, __u64 end, struct thandle *th)
1801 {
1802         struct osd_thandle *oh;
1803         struct osd_object  *obj = osd_dt_obj(dt);
1804         struct inode       *inode = obj->oo_inode;
1805         handle_t           *h;
1806         tid_t               tid;
1807         int                rc = 0, rc2 = 0;
1808         ENTRY;
1809
1810         LASSERT(end == OBD_OBJECT_EOF);
1811         LASSERT(dt_object_exists(dt));
1812         LASSERT(osd_invariant(obj));
1813         LASSERT(inode != NULL);
1814         ll_vfs_dq_init(inode);
1815
1816         LASSERT(th);
1817         oh = container_of(th, struct osd_thandle, ot_super);
1818         LASSERT(oh->ot_handle->h_transaction != NULL);
1819
1820         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1821
1822         tid = oh->ot_handle->h_transaction->t_tid;
1823
1824         spin_lock(&inode->i_lock);
1825         i_size_write(inode, start);
1826         spin_unlock(&inode->i_lock);
1827         ll_truncate_pagecache(inode, start);
1828 #ifdef HAVE_INODEOPS_TRUNCATE
1829         if (inode->i_op->truncate) {
1830                 inode->i_op->truncate(inode);
1831         } else
1832 #endif
1833                 ldiskfs_truncate(inode);
1834
1835         /*
1836          * For a partial-page truncate, flush the page to disk immediately to
1837          * avoid data corruption during direct disk write.  b=17397
1838          */
1839         if ((start & ~PAGE_MASK) != 0)
1840                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1841
1842         h = journal_current_handle();
1843         LASSERT(h != NULL);
1844         LASSERT(h == oh->ot_handle);
1845
1846         /* do not check credits with osd_trans_exec_check() as the truncate
1847          * can restart the transaction internally and we restart the
1848          * transaction in this case */
1849
1850         if (tid != h->h_transaction->t_tid) {
1851                 int credits = oh->ot_credits;
1852                 /*
1853                  * transaction has changed during truncate
1854                  * we need to restart the handle with our credits
1855                  */
1856                 if (h->h_buffer_credits < credits) {
1857                         if (ldiskfs_journal_extend(h, credits))
1858                                 rc2 = ldiskfs_journal_restart(h, credits);
1859                 }
1860         }
1861
1862         RETURN(rc == 0 ? rc2 : rc);
1863 }
1864
1865 static int fiemap_check_ranges(struct inode *inode,
1866                                u64 start, u64 len, u64 *new_len)
1867 {
1868         loff_t maxbytes;
1869
1870         *new_len = len;
1871
1872         if (len == 0)
1873                 return -EINVAL;
1874
1875         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1876                 maxbytes = inode->i_sb->s_maxbytes;
1877         else
1878                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1879
1880         if (start > maxbytes)
1881                 return -EFBIG;
1882
1883         /*
1884          * Shrink request scope to what the fs can actually handle.
1885          */
1886         if (len > maxbytes || (maxbytes - len) < start)
1887                 *new_len = maxbytes - start;
1888
1889         return 0;
1890 }
1891
1892 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1893 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
1894
1895 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1896                           struct fiemap *fm)
1897 {
1898         struct fiemap_extent_info fieinfo = {0, };
1899         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1900         u64 len;
1901         int rc;
1902
1903
1904         LASSERT(inode);
1905         if (inode->i_op->fiemap == NULL)
1906                 return -EOPNOTSUPP;
1907
1908         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1909                 return -EINVAL;
1910
1911         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1912         if (rc)
1913                 return rc;
1914
1915         fieinfo.fi_flags = fm->fm_flags;
1916         fieinfo.fi_extents_max = fm->fm_extent_count;
1917         fieinfo.fi_extents_start = fm->fm_extents;
1918
1919         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1920                 filemap_write_and_wait(inode->i_mapping);
1921
1922         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
1923         fm->fm_flags = fieinfo.fi_flags;
1924         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
1925
1926         return rc;
1927 }
1928
1929 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1930                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1931 {
1932         int              rc = 0;
1933         struct inode    *inode = osd_dt_obj(dt)->oo_inode;
1934         ENTRY;
1935
1936         switch (advice) {
1937         case LU_LADVISE_DONTNEED:
1938                 if (end == 0)
1939                         break;
1940                 invalidate_mapping_pages(inode->i_mapping,
1941                                          start >> PAGE_CACHE_SHIFT,
1942                                          (end - 1) >> PAGE_CACHE_SHIFT);
1943                 break;
1944         default:
1945                 rc = -ENOTSUPP;
1946                 break;
1947         }
1948
1949         RETURN(rc);
1950 }
1951
1952 /*
1953  * in some cases we may need declare methods for objects being created
1954  * e.g., when we create symlink
1955  */
1956 const struct dt_body_operations osd_body_ops_new = {
1957         .dbo_declare_write = osd_declare_write,
1958 };
1959
1960 const struct dt_body_operations osd_body_ops = {
1961         .dbo_read                       = osd_read,
1962         .dbo_declare_write              = osd_declare_write,
1963         .dbo_write                      = osd_write,
1964         .dbo_bufs_get                   = osd_bufs_get,
1965         .dbo_bufs_put                   = osd_bufs_put,
1966         .dbo_write_prep                 = osd_write_prep,
1967         .dbo_declare_write_commit       = osd_declare_write_commit,
1968         .dbo_write_commit               = osd_write_commit,
1969         .dbo_read_prep                  = osd_read_prep,
1970         .dbo_declare_punch              = osd_declare_punch,
1971         .dbo_punch                      = osd_punch,
1972         .dbo_fiemap_get                 = osd_fiemap_get,
1973         .dbo_ladvise                    = osd_ladvise,
1974 };