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, 2016, 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 /* prerequisite for linux/xattr.h */
  42 #include <linux/types.h>
  43 /* prerequisite for linux/xattr.h */
  44 #include <linux/fs.h>
  45
  46 /*
  47  * struct OBD_{ALLOC,FREE}*()
  48  * OBD_FAIL_CHECK
  49  */
  50 #include <obd_support.h>
  51
  52 #include "osd_internal.h"
  53
  54 /* ext_depth() */
  55 #include <ldiskfs/ldiskfs_extents.h>
  56
  57 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  58                             int rw, int line, int pages)
  59 {
  60         int blocks, i;
  61
  62         LASSERTF(iobuf->dr_elapsed_valid == 0,
  63                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  64                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  65                  iobuf->dr_init_at);
  66         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  67
  68         init_waitqueue_head(&iobuf->dr_wait);
  69         atomic_set(&iobuf->dr_numreqs, 0);
  70         iobuf->dr_npages = 0;
  71         iobuf->dr_error = 0;
  72         iobuf->dr_dev = d;
  73         iobuf->dr_frags = 0;
  74         iobuf->dr_elapsed = 0;
  75         /* must be counted before, so assert */
  76         iobuf->dr_rw = rw;
  77         iobuf->dr_init_at = line;
  78
  79         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  80         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  81                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  82                         pages * sizeof(iobuf->dr_pages[0]));
  83                 return 0;
  84         }
  85
  86         /* start with 1MB for 4K blocks */
  87         i = 256;
  88         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
  89                 i <<= 1;
  90
  91         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
  92                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
  93         pages = i;
  94         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  95         iobuf->dr_max_pages = 0;
  96         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
  97                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
  98
  99         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 100         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 101         if (unlikely(iobuf->dr_blocks == NULL))
 102                 return -ENOMEM;
 103
 104         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 105         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 106         if (unlikely(iobuf->dr_pages == NULL))
 107                 return -ENOMEM;
 108
 109         iobuf->dr_max_pages = pages;
 110
 111         return 0;
 112 }
 113 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 114         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 115
 116 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
 117 {
 118         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 119         iobuf->dr_pages[iobuf->dr_npages++] = page;
 120 }
 121
 122 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 123 {
 124         int rw = iobuf->dr_rw;
 125
 126         if (iobuf->dr_elapsed_valid) {
 127                 iobuf->dr_elapsed_valid = 0;
 128                 LASSERT(iobuf->dr_dev == d);
 129                 LASSERT(iobuf->dr_frags > 0);
 130                 lprocfs_oh_tally(&d->od_brw_stats.
 131                                  hist[BRW_R_DIO_FRAGS+rw],
 132                                  iobuf->dr_frags);
 133                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 134                                       iobuf->dr_elapsed);
 135         }
 136 }
 137
 138 #ifndef REQ_WRITE /* pre-2.6.35 */
 139 #define __REQ_WRITE BIO_RW
 140 #endif
 141
 142 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 143 static void dio_complete_routine(struct bio *bio)
 144 {
 145         int error = bio->bi_error;
 146 #else
 147 static void dio_complete_routine(struct bio *bio, int error)
 148 {
 149 #endif
 150         struct osd_iobuf *iobuf = bio->bi_private;
 151         int iter;
 152         struct bio_vec *bvl;
 153
 154         /* CAVEAT EMPTOR: possibly in IRQ context
 155          * DO NOT record procfs stats here!!! */
 156
 157         if (unlikely(iobuf == NULL)) {
 158                 CERROR("***** bio->bi_private is NULL!  This should never "
 159                        "happen.  Normally, I would crash here, but instead I "
 160                        "will dump the bio contents to the console.  Please "
 161                        "report this to <https://jira.hpdd.intel.com/> , along "
 162                        "with any interesting messages leading up to this point "
 163                        "(like SCSI errors, perhaps).  Because bi_private is "
 164                        "NULL, I can't wake up the thread that initiated this "
 165                        "IO - you will probably have to reboot this node.\n");
 166                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 167                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 168                        "bi_private: %p\n", bio->bi_next,
 169                         (unsigned long)bio->bi_flags,
 170                         bio->bi_rw, bio->bi_vcnt, bio_idx(bio),
 171                         bio_sectors(bio) << 9, bio->bi_end_io,
 172 #ifdef HAVE_BI_CNT
 173                         atomic_read(&bio->bi_cnt),
 174 #else
 175                         atomic_read(&bio->__bi_cnt),
 176 #endif
 177                         bio->bi_private);
 178                 return;
 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_all(bvl, bio, iter) {
 184                         if (likely(error == 0))
 185                                 SetPageUptodate(bvl_to_page(bvl));
 186                         LASSERT(PageLocked(bvl_to_page(bvl)));
 187                 }
 188                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 189         } else {
 190                 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 (atomic_read(&iobuf->dr_numreqs) == 1) {
 205                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
 206                 iobuf->dr_elapsed_valid = 1;
 207         }
 208         if (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 }
 218
 219 static void record_start_io(struct osd_iobuf *iobuf, int size)
 220 {
 221         struct osd_device    *osd = iobuf->dr_dev;
 222         struct obd_histogram *h = osd->od_brw_stats.hist;
 223
 224         iobuf->dr_frags++;
 225         atomic_inc(&iobuf->dr_numreqs);
 226
 227         if (iobuf->dr_rw == 0) {
 228                 atomic_inc(&osd->od_r_in_flight);
 229                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 230                                  atomic_read(&osd->od_r_in_flight));
 231                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 232         } else if (iobuf->dr_rw == 1) {
 233                 atomic_inc(&osd->od_w_in_flight);
 234                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 235                                  atomic_read(&osd->od_w_in_flight));
 236                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 237         } else {
 238                 LBUG();
 239         }
 240 }
 241
 242 static void osd_submit_bio(int rw, struct bio *bio)
 243 {
 244         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 245         if (rw == 0)
 246                 submit_bio(READ, bio);
 247         else
 248                 submit_bio(WRITE, bio);
 249 }
 250
 251 static int can_be_merged(struct bio *bio, sector_t sector)
 252 {
 253         if (bio == NULL)
 254                 return 0;
 255
 256         return bio_end_sector(bio) == sector ? 1 : 0;
 257 }
 258
 259 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 260                       struct osd_iobuf *iobuf)
 261 {
 262         int            blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 263         struct page  **pages = iobuf->dr_pages;
 264         int            npages = iobuf->dr_npages;
 265         sector_t      *blocks = iobuf->dr_blocks;
 266         int            total_blocks = npages * blocks_per_page;
 267         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 268         unsigned int   blocksize = inode->i_sb->s_blocksize;
 269         struct bio    *bio = NULL;
 270         struct page   *page;
 271         unsigned int   page_offset;
 272         sector_t       sector;
 273         int            nblocks;
 274         int            block_idx;
 275         int            page_idx;
 276         int            i;
 277         int            rc = 0;
 278         DECLARE_PLUG(plug);
 279         ENTRY;
 280
 281         LASSERT(iobuf->dr_npages == npages);
 282
 283         osd_brw_stats_update(osd, iobuf);
 284         iobuf->dr_start_time = cfs_time_current();
 285
 286         blk_start_plug(&plug);
 287         for (page_idx = 0, block_idx = 0;
 288              page_idx < npages;
 289              page_idx++, block_idx += blocks_per_page) {
 290
 291                 page = pages[page_idx];
 292                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 293
 294                 for (i = 0, page_offset = 0;
 295                      i < blocks_per_page;
 296                      i += nblocks, page_offset += blocksize * nblocks) {
 297
 298                         nblocks = 1;
 299
 300                         if (blocks[block_idx + i] == 0) {  /* hole */
 301                                 LASSERTF(iobuf->dr_rw == 0,
 302                                          "page_idx %u, block_idx %u, i %u\n",
 303                                          page_idx, block_idx, i);
 304                                 memset(kmap(page) + page_offset, 0, blocksize);
 305                                 kunmap(page);
 306                                 continue;
 307                         }
 308
 309                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 310
 311                         /* Additional contiguous file blocks? */
 312                         while (i + nblocks < blocks_per_page &&
 313                                (sector + (nblocks << sector_bits)) ==
 314                                ((sector_t)blocks[block_idx + i + nblocks] <<
 315                                 sector_bits))
 316                                 nblocks++;
 317
 318                         if (bio != NULL &&
 319                             can_be_merged(bio, sector) &&
 320                             bio_add_page(bio, page,
 321                                          blocksize * nblocks, page_offset) != 0)
 322                                 continue;       /* added this frag OK */
 323
 324                         if (bio != NULL) {
 325                                 struct request_queue *q =
 326                                         bdev_get_queue(bio->bi_bdev);
 327                                 unsigned int bi_size = bio_sectors(bio) << 9;
 328
 329                                 /* Dang! I have to fragment this I/O */
 330                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 331                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 332                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 333                                        bio_sectors(bio),
 334                                        queue_max_sectors(q),
 335                                        bio_phys_segments(q, bio),
 336                                        queue_max_phys_segments(q),
 337                                        0, queue_max_hw_segments(q));
 338                                 record_start_io(iobuf, bi_size);
 339                                 osd_submit_bio(iobuf->dr_rw, bio);
 340                         }
 341
 342                         /* allocate new bio */
 343                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 344                                                       (npages - page_idx) *
 345                                                       blocks_per_page));
 346                         if (bio == NULL) {
 347                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 348                                        (npages - page_idx), blocks_per_page,
 349                                        (npages - page_idx) * blocks_per_page);
 350                                 rc = -ENOMEM;
 351                                 goto out;
 352                         }
 353
 354                         bio->bi_bdev = inode->i_sb->s_bdev;
 355                         bio_set_sector(bio, sector);
 356                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 357                         bio->bi_end_io = dio_complete_routine;
 358                         bio->bi_private = iobuf;
 359
 360                         rc = bio_add_page(bio, page,
 361                                           blocksize * nblocks, page_offset);
 362                         LASSERT(rc != 0);
 363                 }
 364         }
 365
 366         if (bio != NULL) {
 367                 record_start_io(iobuf, bio_sectors(bio) << 9);
 368                 osd_submit_bio(iobuf->dr_rw, bio);
 369                 rc = 0;
 370         }
 371
 372 out:
 373         blk_finish_plug(&plug);
 374
 375         /* in order to achieve better IO throughput, we don't wait for writes
 376          * completion here. instead we proceed with transaction commit in
 377          * parallel and wait for IO completion once transaction is stopped
 378          * see osd_trans_stop() for more details -bzzz */
 379         if (iobuf->dr_rw == 0) {
 380                 wait_event(iobuf->dr_wait,
 381                            atomic_read(&iobuf->dr_numreqs) == 0);
 382                 osd_fini_iobuf(osd, iobuf);
 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_SIZE - 1);
 399                 int plen = PAGE_SIZE - poff;
 400
 401                 if (plen > len)
 402                         plen = len;
 403                 lnb->lnb_file_offset = offset;
 404                 lnb->lnb_page_offset = poff;
 405                 lnb->lnb_len = plen;
 406                 /* lnb->lnb_flags = rnb->rnb_flags; */
 407                 lnb->lnb_flags = 0;
 408                 lnb->lnb_page = NULL;
 409                 lnb->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 static struct page *osd_get_page(struct dt_object *dt, loff_t offset,
 423                                  gfp_t gfp_mask)
 424 {
 425         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 426         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 427         struct page *page;
 428
 429         LASSERT(inode);
 430
 431         page = find_or_create_page(inode->i_mapping, offset >> PAGE_SHIFT,
 432                                    gfp_mask);
 433
 434         if (unlikely(page == NULL))
 435                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 436
 437         return page;
 438 }
 439
 440 /*
 441  * there are following "locks":
 442  * journal_start
 443  * i_mutex
 444  * page lock
 445  *
 446  * osd write path:
 447  *  - lock page(s)
 448  *  - journal_start
 449  *  - truncate_sem
 450  *
 451  * ext4 vmtruncate:
 452  *  - lock pages, unlock
 453  *  - journal_start
 454  *  - lock partial page
 455  *  - i_data_sem
 456  *
 457  */
 458
 459 /**
 460  * Unlock and release pages loaded by osd_bufs_get()
 461  *
 462  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 463  *
 464  * \param env           thread execution environment
 465  * \param dt            dt object undergoing IO (OSD object + methods)
 466  * \param lnb           array of pages undergoing IO
 467  * \param npages        number of pages in \a lnb
 468  *
 469  * \retval 0            always
 470  */
 471 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 472                         struct niobuf_local *lnb, int npages)
 473 {
 474         int i;
 475
 476         for (i = 0; i < npages; i++) {
 477                 if (lnb[i].lnb_page == NULL)
 478                         continue;
 479                 LASSERT(PageLocked(lnb[i].lnb_page));
 480                 unlock_page(lnb[i].lnb_page);
 481                 put_page(lnb[i].lnb_page);
 482                 dt_object_put(env, dt);
 483                 lnb[i].lnb_page = NULL;
 484         }
 485
 486         RETURN(0);
 487 }
 488
 489 /**
 490  * Load and lock pages undergoing IO
 491  *
 492  * Pages as described in the \a lnb array are fetched (from disk or cache)
 493  * and locked for IO by the caller.
 494  *
 495  * DLM locking protects us from write and truncate competing for same region,
 496  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 497  * It's possible the writeout on a such a page is in progress when we access
 498  * it. It's also possible that during this writeout we put new (partial) data
 499  * into the page, but won't be able to proceed in filter_commitrw_write().
 500  * Therefore, just wait for writeout completion as it should be rare enough.
 501  *
 502  * \param env           thread execution environment
 503  * \param dt            dt object undergoing IO (OSD object + methods)
 504  * \param pos           byte offset of IO start
 505  * \param len           number of bytes of IO
 506  * \param lnb           array of extents undergoing IO
 507  * \param rw            read or write operation, and other flags
 508  * \param capa          capabilities
 509  *
 510  * \retval pages        (zero or more) loaded successfully
 511  * \retval -ENOMEM      on memory/page allocation error
 512  */
 513 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 514                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 515                         enum dt_bufs_type rw)
 516 {
 517         struct osd_object *obj = osd_dt_obj(dt);
 518         int npages, i, rc = 0;
 519         gfp_t gfp_mask;
 520
 521         LASSERT(obj->oo_inode);
 522
 523         osd_map_remote_to_local(pos, len, &npages, lnb);
 524
 525         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 526         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 527                                              GFP_HIGHUSER;
 528         for (i = 0; i < npages; i++, lnb++) {
 529                 lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset,
 530                                              gfp_mask);
 531                 if (lnb->lnb_page == NULL)
 532                         GOTO(cleanup, rc = -ENOMEM);
 533
 534                 wait_on_page_writeback(lnb->lnb_page);
 535                 BUG_ON(PageWriteback(lnb->lnb_page));
 536
 537                 lu_object_get(&dt->do_lu);
 538         }
 539
 540         RETURN(i);
 541
 542 cleanup:
 543         if (i > 0)
 544                 osd_bufs_put(env, dt, lnb - i, i);
 545         return rc;
 546 }
 547
 548 #ifndef HAVE_LDISKFS_MAP_BLOCKS
 549
 550 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 551 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 552 #endif
 553
 554 struct bpointers {
 555         sector_t *blocks;
 556         unsigned long start;
 557         int num;
 558         int init_num;
 559         int create;
 560 };
 561
 562 static long ldiskfs_ext_find_goal(struct inode *inode,
 563                                   struct ldiskfs_ext_path *path,
 564                                   unsigned long block, int *aflags)
 565 {
 566         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 567         unsigned long bg_start;
 568         unsigned long colour;
 569         int depth;
 570
 571         if (path) {
 572                 struct ldiskfs_extent *ex;
 573                 depth = path->p_depth;
 574
 575                 /* try to predict block placement */
 576                 if ((ex = path[depth].p_ext))
 577                         return ldiskfs_ext_pblock(ex) +
 578                                 (block - le32_to_cpu(ex->ee_block));
 579
 580                 /* it looks index is empty
 581                  * try to find starting from index itself */
 582                 if (path[depth].p_bh)
 583                         return path[depth].p_bh->b_blocknr;
 584         }
 585
 586         /* OK. use inode's group */
 587         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 588                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 589         colour = (current->pid % 16) *
 590                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 591         return bg_start + colour + block;
 592 }
 593
 594 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 595                                 struct ldiskfs_ext_path *path,
 596                                 unsigned long block, unsigned long *count,
 597                                 int *err)
 598 {
 599         struct ldiskfs_allocation_request ar;
 600         unsigned long pblock;
 601         int aflags;
 602
 603         /* find neighbour allocated blocks */
 604         ar.lleft = block;
 605         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 606         if (*err)
 607                 return 0;
 608         ar.lright = block;
 609         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 610         if (*err)
 611                 return 0;
 612
 613         /* allocate new block */
 614         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 615         ar.inode = inode;
 616         ar.logical = block;
 617         ar.len = *count;
 618         ar.flags = LDISKFS_MB_HINT_DATA;
 619         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 620         *count = ar.len;
 621         return pblock;
 622 }
 623
 624 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 625                                      struct ldiskfs_ext_path *path,
 626                                      struct ldiskfs_ext_cache *cex,
 627 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 628                                      struct ldiskfs_extent *ex,
 629 #endif
 630                                      void *cbdata)
 631 {
 632         struct bpointers *bp = cbdata;
 633         struct ldiskfs_extent nex;
 634         unsigned long pblock = 0;
 635         unsigned long tgen;
 636         int err, i;
 637         unsigned long count;
 638         handle_t *handle;
 639
 640 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 641         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 642 #else
 643         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 644 #endif
 645                 err = EXT_CONTINUE;
 646                 goto map;
 647         }
 648
 649         if (bp->create == 0) {
 650                 i = 0;
 651                 if (cex->ec_block < bp->start)
 652                         i = bp->start - cex->ec_block;
 653                 if (i >= cex->ec_len)
 654                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 655                                         i, cex->ec_len);
 656                 for (; i < cex->ec_len && bp->num; i++) {
 657                         *(bp->blocks) = 0;
 658                         bp->blocks++;
 659                         bp->num--;
 660                         bp->start++;
 661                 }
 662
 663                 return EXT_CONTINUE;
 664         }
 665
 666         tgen = LDISKFS_I(inode)->i_ext_generation;
 667         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
 668
 669         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
 670                                    count + LDISKFS_ALLOC_NEEDED + 1);
 671         if (IS_ERR(handle)) {
 672                 return PTR_ERR(handle);
 673         }
 674
 675         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
 676                 /* the tree has changed. so path can be invalid at moment */
 677                 ldiskfs_journal_stop(handle);
 678                 return EXT_REPEAT;
 679         }
 680
 681         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
 682          * protected by i_data_sem as whole. so we patch it to store
 683          * generation to path and now verify the tree hasn't changed */
 684         down_write((&LDISKFS_I(inode)->i_data_sem));
 685
 686         /* validate extent, make sure the extent tree does not changed */
 687         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
 688                 /* cex is invalid, try again */
 689                 up_write(&LDISKFS_I(inode)->i_data_sem);
 690                 ldiskfs_journal_stop(handle);
 691                 return EXT_REPEAT;
 692         }
 693
 694         count = cex->ec_len;
 695         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
 696         if (!pblock)
 697                 goto out;
 698         BUG_ON(count > cex->ec_len);
 699
 700         /* insert new extent */
 701         nex.ee_block = cpu_to_le32(cex->ec_block);
 702         ldiskfs_ext_store_pblock(&nex, pblock);
 703         nex.ee_len = cpu_to_le16(count);
 704         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
 705         if (err) {
 706                 /* free data blocks we just allocated */
 707                 /* not a good idea to call discard here directly,
 708                  * but otherwise we'd need to call it every free() */
 709                 ldiskfs_discard_preallocations(inode);
 710 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
 711                 ldiskfs_free_blocks(handle, inode, NULL,
 712                                     ldiskfs_ext_pblock(&nex),
 713                                     le16_to_cpu(nex.ee_len), 0);
 714 #else
 715                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
 716                                     le16_to_cpu(nex.ee_len), 0);
 717 #endif
 718                 goto out;
 719         }
 720
 721         /*
 722          * Putting len of the actual extent we just inserted,
 723          * we are asking ldiskfs_ext_walk_space() to continue
 724          * scaning after that block
 725          */
 726         cex->ec_len = le16_to_cpu(nex.ee_len);
 727         cex->ec_start = ldiskfs_ext_pblock(&nex);
 728         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
 729         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
 730
 731 out:
 732         up_write((&LDISKFS_I(inode)->i_data_sem));
 733         ldiskfs_journal_stop(handle);
 734 map:
 735         if (err >= 0) {
 736                 /* map blocks */
 737                 if (bp->num == 0) {
 738                         CERROR("hmm. why do we find this extent?\n");
 739                         CERROR("initial space: %lu:%u\n",
 740                                 bp->start, bp->init_num);
 741 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 742                         CERROR("current extent: %u/%u/%llu %d\n",
 743                                 cex->ec_block, cex->ec_len,
 744                                 (unsigned long long)cex->ec_start,
 745                                 cex->ec_type);
 746 #else
 747                         CERROR("current extent: %u/%u/%llu\n",
 748                                 cex->ec_block, cex->ec_len,
 749                                 (unsigned long long)cex->ec_start);
 750 #endif
 751                 }
 752                 i = 0;
 753                 if (cex->ec_block < bp->start)
 754                         i = bp->start - cex->ec_block;
 755                 if (i >= cex->ec_len)
 756                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 757                                         i, cex->ec_len);
 758                 for (; i < cex->ec_len && bp->num; i++) {
 759                         *(bp->blocks) = cex->ec_start + i;
 760                         if (pblock != 0) {
 761                                 /* unmap any possible underlying metadata from
 762                                  * the block device mapping.  bug 6998. */
 763                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
 764                                                           *(bp->blocks));
 765                         }
 766                         bp->blocks++;
 767                         bp->num--;
 768                         bp->start++;
 769                 }
 770         }
 771         return err;
 772 }
 773
 774 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
 775                                    int clen, sector_t *blocks, int create)
 776 {
 777         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 778         struct bpointers bp;
 779         int err;
 780
 781         if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
 782                 return -EFBIG;
 783
 784         bp.blocks = blocks;
 785         bp.start = index * blocks_per_page;
 786         bp.init_num = bp.num = clen * blocks_per_page;
 787         bp.create = create;
 788
 789         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 790                bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
 791
 792         err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
 793                                      ldiskfs_ext_new_extent_cb, &bp);
 794         ldiskfs_ext_invalidate_cache(inode);
 795
 796         return err;
 797 }
 798
 799 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
 800                                           struct page **page, int pages,
 801                                           sector_t *blocks, int create)
 802 {
 803         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 804         pgoff_t bitmap_max_page_index;
 805         sector_t *b;
 806         int rc = 0, i;
 807
 808         bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
 809                                 PAGE_SHIFT;
 810         for (i = 0, b = blocks; i < pages; i++, page++) {
 811                 if ((*page)->index + 1 >= bitmap_max_page_index) {
 812                         rc = -EFBIG;
 813                         break;
 814                 }
 815                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
 816                 if (rc) {
 817                         CERROR("ino %lu, blk %llu create %d: rc %d\n",
 818                                inode->i_ino,
 819                                (unsigned long long)*b, create, rc);
 820                         break;
 821                 }
 822                 b += blocks_per_page;
 823         }
 824         return rc;
 825 }
 826
 827 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
 828                                            struct page **page,
 829                                            int pages, sector_t *blocks,
 830                                            int create)
 831 {
 832         int rc = 0, i = 0, clen = 0;
 833         struct page *fp = NULL;
 834
 835         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 836                 inode->i_ino, pages, (*page)->index);
 837
 838         /* pages are sorted already. so, we just have to find
 839          * contig. space and process them properly */
 840         while (i < pages) {
 841                 if (fp == NULL) {
 842                         /* start new extent */
 843                         fp = *page++;
 844                         clen = 1;
 845                         i++;
 846                         continue;
 847                 } else if (fp->index + clen == (*page)->index) {
 848                         /* continue the extent */
 849                         page++;
 850                         clen++;
 851                         i++;
 852                         continue;
 853                 }
 854
 855                 /* process found extent */
 856                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 857                                              blocks, create);
 858                 if (rc)
 859                         GOTO(cleanup, rc);
 860
 861                 /* look for next extent */
 862                 fp = NULL;
 863                 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
 864         }
 865
 866         if (fp)
 867                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 868                                              blocks, create);
 869
 870 cleanup:
 871         return rc;
 872 }
 873
 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 rc;
 879
 880         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 881                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
 882                                                      blocks, create);
 883                 return rc;
 884         }
 885         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
 886
 887         return rc;
 888 }
 889 #else
 890 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 891                                        int pages, sector_t *blocks,
 892                                        int create)
 893 {
 894         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 895         int rc = 0, i = 0;
 896         struct page *fp = NULL;
 897         int clen = 0;
 898         pgoff_t max_page_index;
 899         handle_t *handle = NULL;
 900
 901         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
 902
 903         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 904                 inode->i_ino, pages, (*page)->index);
 905
 906         if (create) {
 907                 create = LDISKFS_GET_BLOCKS_CREATE;
 908                 handle = ldiskfs_journal_current_handle();
 909                 LASSERT(handle != NULL);
 910                 rc = osd_attach_jinode(inode);
 911                 if (rc)
 912                         return rc;
 913         }
 914         /* pages are sorted already. so, we just have to find
 915          * contig. space and process them properly */
 916         while (i < pages) {
 917                 long blen, total = 0;
 918                 struct ldiskfs_map_blocks map = { 0 };
 919
 920                 if (fp == NULL) { /* start new extent */
 921                         fp = *page++;
 922                         clen = 1;
 923                         if (++i != pages)
 924                                 continue;
 925                 } else if (fp->index + clen == (*page)->index) {
 926                         /* continue the extent */
 927                         page++;
 928                         clen++;
 929                         if (++i != pages)
 930                                 continue;
 931                 }
 932                 if (fp->index + clen >= max_page_index)
 933                         GOTO(cleanup, rc = -EFBIG);
 934                 /* process found extent */
 935                 map.m_lblk = fp->index * blocks_per_page;
 936                 map.m_len = blen = clen * blocks_per_page;
 937 cont_map:
 938                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 939                 if (rc >= 0) {
 940                         int c = 0;
 941                         for (; total < blen && c < map.m_len; c++, total++) {
 942                                 if (rc == 0) {
 943                                         *(blocks + total) = 0;
 944                                         total++;
 945                                         break;
 946                                 } else {
 947                                         *(blocks + total) = map.m_pblk + c;
 948                                         /* unmap any possible underlying
 949                                          * metadata from the block device
 950                                          * mapping.  bug 6998. */
 951                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
 952                                             create)
 953                                                 unmap_underlying_metadata(
 954                                                         inode->i_sb->s_bdev,
 955                                                         map.m_pblk + c);
 956                                 }
 957                         }
 958                         rc = 0;
 959                 }
 960                 if (rc == 0 && total < blen) {
 961                         map.m_lblk = fp->index * blocks_per_page + total;
 962                         map.m_len = blen - total;
 963                         goto cont_map;
 964                 }
 965                 if (rc != 0)
 966                         GOTO(cleanup, rc);
 967
 968                 /* look for next extent */
 969                 fp = NULL;
 970                 blocks += blocks_per_page * clen;
 971         }
 972 cleanup:
 973         return rc;
 974 }
 975 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
 976
 977 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 978                           struct niobuf_local *lnb, int npages)
 979 {
 980         struct osd_thread_info *oti   = osd_oti_get(env);
 981         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 982         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 983         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 984         ktime_t start;
 985         ktime_t end;
 986         s64 timediff;
 987         ssize_t                 isize;
 988         __s64                   maxidx;
 989         int                     rc = 0;
 990         int                     i;
 991         int                     cache = 0;
 992
 993         LASSERT(inode);
 994
 995         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 996         if (unlikely(rc != 0))
 997                 RETURN(rc);
 998
 999         isize = i_size_read(inode);
1000         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1001
1002         if (osd->od_writethrough_cache)
1003                 cache = 1;
1004         if (isize > osd->od_readcache_max_filesize)
1005                 cache = 0;
1006
1007         start = ktime_get();
1008         for (i = 0; i < npages; i++) {
1009
1010                 if (cache == 0)
1011                         generic_error_remove_page(inode->i_mapping,
1012                                                   lnb[i].lnb_page);
1013
1014                 /*
1015                  * till commit the content of the page is undefined
1016                  * we'll set it uptodate once bulk is done. otherwise
1017                  * subsequent reads can access non-stable data
1018                  */
1019                 ClearPageUptodate(lnb[i].lnb_page);
1020
1021                 if (lnb[i].lnb_len == PAGE_SIZE)
1022                         continue;
1023
1024                 if (maxidx >= lnb[i].lnb_page->index) {
1025                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1026                 } else {
1027                         long off;
1028                         char *p = kmap(lnb[i].lnb_page);
1029
1030                         off = lnb[i].lnb_page_offset;
1031                         if (off)
1032                                 memset(p, 0, off);
1033                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1034                               ~PAGE_MASK;
1035                         if (off)
1036                                 memset(p + off, 0, PAGE_SIZE - off);
1037                         kunmap(lnb[i].lnb_page);
1038                 }
1039         }
1040         end = ktime_get();
1041         timediff = ktime_us_delta(end, start);
1042         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1043
1044         if (iobuf->dr_npages) {
1045                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1046                                                  iobuf->dr_npages,
1047                                                  iobuf->dr_blocks, 0);
1048                 if (likely(rc == 0)) {
1049                         rc = osd_do_bio(osd, inode, iobuf);
1050                         /* do IO stats for preparation reads */
1051                         osd_fini_iobuf(osd, iobuf);
1052                 }
1053         }
1054         RETURN(rc);
1055 }
1056
1057 struct osd_fextent {
1058         sector_t        start;
1059         sector_t        end;
1060         unsigned int    mapped:1;
1061 };
1062
1063 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1064                          struct osd_fextent *cached_extent)
1065 {
1066         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1067         sector_t block = offset >> inode->i_blkbits;
1068         sector_t start;
1069         struct fiemap_extent_info fei = { 0 };
1070         struct fiemap_extent fe = { 0 };
1071         mm_segment_t saved_fs;
1072         int rc;
1073
1074         if (block >= cached_extent->start && block < cached_extent->end)
1075                 return cached_extent->mapped;
1076
1077         if (i_size_read(inode) == 0)
1078                 return 0;
1079
1080         /* Beyond EOF, must not be mapped */
1081         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1082                 return 0;
1083
1084         fei.fi_extents_max = 1;
1085         fei.fi_extents_start = &fe;
1086
1087         saved_fs = get_fs();
1088         set_fs(get_ds());
1089         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1090         set_fs(saved_fs);
1091         if (rc != 0)
1092                 return 0;
1093
1094         start = fe.fe_logical >> inode->i_blkbits;
1095
1096         if (start > block) {
1097                 cached_extent->start = block;
1098                 cached_extent->end = start;
1099                 cached_extent->mapped = 0;
1100         } else {
1101                 cached_extent->start = start;
1102                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1103                                       inode->i_blkbits;
1104                 cached_extent->mapped = 1;
1105         }
1106
1107         return cached_extent->mapped;
1108 }
1109
1110 static int osd_declare_write_commit(const struct lu_env *env,
1111                                     struct dt_object *dt,
1112                                     struct niobuf_local *lnb, int npages,
1113                                     struct thandle *handle)
1114 {
1115         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1116         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1117         struct osd_thandle      *oh;
1118         int                     extents = 1;
1119         int                     depth;
1120         int                     i;
1121         int                     newblocks;
1122         int                     rc = 0;
1123         int                     flags = 0;
1124         int                     credits = 0;
1125         long long               quota_space = 0;
1126         struct osd_fextent      extent = { 0 };
1127         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1128         ENTRY;
1129
1130         LASSERT(handle != NULL);
1131         oh = container_of0(handle, struct osd_thandle, ot_super);
1132         LASSERT(oh->ot_handle == NULL);
1133
1134         newblocks = npages;
1135
1136         /* calculate number of extents (probably better to pass nb) */
1137         for (i = 0; i < npages; i++) {
1138                 if (i && lnb[i].lnb_file_offset !=
1139                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1140                         extents++;
1141
1142                 if (!osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1143                         quota_space += PAGE_SIZE;
1144
1145                 /* ignore quota for the whole request if any page is from
1146                  * client cache or written by root.
1147                  *
1148                  * XXX once we drop the 1.8 client support, the checking
1149                  * for whether page is from cache can be simplified as:
1150                  * !(lnb[i].flags & OBD_BRW_SYNC)
1151                  *
1152                  * XXX we could handle this on per-lnb basis as done by
1153                  * grant. */
1154                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1155                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1156                     OBD_BRW_FROM_GRANT)
1157                         declare_flags |= OSD_QID_FORCE;
1158         }
1159
1160         /*
1161          * each extent can go into new leaf causing a split
1162          * 5 is max tree depth: inode + 4 index blocks
1163          * with blockmaps, depth is 3 at most
1164          */
1165         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1166                 /*
1167                  * many concurrent threads may grow tree by the time
1168                  * our transaction starts. so, consider 2 is a min depth
1169                  */
1170                 depth = ext_depth(inode);
1171                 depth = max(depth, 1) + 1;
1172                 newblocks += depth;
1173                 credits++; /* inode */
1174                 credits += depth * 2 * extents;
1175         } else {
1176                 depth = 3;
1177                 newblocks += depth;
1178                 credits++; /* inode */
1179                 credits += depth * extents;
1180         }
1181
1182         /* quota space for metadata blocks */
1183         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1184
1185         /* quota space should be reported in 1K blocks */
1186         quota_space = toqb(quota_space);
1187
1188         /* each new block can go in different group (bitmap + gd) */
1189
1190         /* we can't dirty more bitmap blocks than exist */
1191         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1192                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1193         else
1194                 credits += newblocks;
1195
1196         /* we can't dirty more gd blocks than exist */
1197         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1198                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1199         else
1200                 credits += newblocks;
1201
1202         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1203
1204         /* make sure the over quota flags were not set */
1205         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1206
1207         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1208                                    i_projid_read(inode), quota_space, oh,
1209                                    osd_dt_obj(dt), &flags, declare_flags);
1210
1211         /* we need only to store the overquota flags in the first lnb for
1212          * now, once we support multiple objects BRW, this code needs be
1213          * revised. */
1214         if (flags & QUOTA_FL_OVER_USRQUOTA)
1215                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1216         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1217                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1218         if (flags & QUOTA_FL_OVER_PRJQUOTA)
1219                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1220
1221         RETURN(rc);
1222 }
1223
1224 /* Check if a block is allocated or not */
1225 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1226                             struct niobuf_local *lnb, int npages,
1227                             struct thandle *thandle)
1228 {
1229         struct osd_thread_info *oti = osd_oti_get(env);
1230         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1231         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1232         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1233         loff_t isize;
1234         int rc = 0, i;
1235         struct osd_fextent extent = { 0 };
1236
1237         LASSERT(inode);
1238
1239         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1240         if (unlikely(rc != 0))
1241                 RETURN(rc);
1242
1243         isize = i_size_read(inode);
1244         ll_vfs_dq_init(inode);
1245
1246         for (i = 0; i < npages; i++) {
1247                 if (lnb[i].lnb_rc == -ENOSPC &&
1248                     osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent)) {
1249                         /* Allow the write to proceed if overwriting an
1250                          * existing block */
1251                         lnb[i].lnb_rc = 0;
1252                 }
1253
1254                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1255                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1256                                lnb[i].lnb_rc);
1257                         LASSERT(lnb[i].lnb_page);
1258                         generic_error_remove_page(inode->i_mapping,
1259                                                   lnb[i].lnb_page);
1260                         continue;
1261                 }
1262
1263                 LASSERT(PageLocked(lnb[i].lnb_page));
1264                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1265
1266                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1267                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1268
1269                 /*
1270                  * Since write and truncate are serialized by oo_sem, even
1271                  * partial-page truncate should not leave dirty pages in the
1272                  * page cache.
1273                  */
1274                 LASSERT(!PageDirty(lnb[i].lnb_page));
1275
1276                 SetPageUptodate(lnb[i].lnb_page);
1277
1278                 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1279         }
1280
1281         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1282
1283         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1284                 rc = -ENOSPC;
1285         } else if (iobuf->dr_npages > 0) {
1286                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1287                                                  iobuf->dr_npages,
1288                                                  iobuf->dr_blocks, 1);
1289         } else {
1290                 /* no pages to write, no transno is needed */
1291                 thandle->th_local = 1;
1292         }
1293
1294         if (likely(rc == 0)) {
1295                 spin_lock(&inode->i_lock);
1296                 if (isize > i_size_read(inode)) {
1297                         i_size_write(inode, isize);
1298                         LDISKFS_I(inode)->i_disksize = isize;
1299                         spin_unlock(&inode->i_lock);
1300                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1301                 } else {
1302                         spin_unlock(&inode->i_lock);
1303                 }
1304
1305                 rc = osd_do_bio(osd, inode, iobuf);
1306                 /* we don't do stats here as in read path because
1307                  * write is async: we'll do this in osd_put_bufs() */
1308         } else {
1309                 osd_fini_iobuf(osd, iobuf);
1310         }
1311
1312         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1313
1314         if (unlikely(rc != 0)) {
1315                 /* if write fails, we should drop pages from the cache */
1316                 for (i = 0; i < npages; i++) {
1317                         if (lnb[i].lnb_page == NULL)
1318                                 continue;
1319                         LASSERT(PageLocked(lnb[i].lnb_page));
1320                         generic_error_remove_page(inode->i_mapping,
1321                                                   lnb[i].lnb_page);
1322                 }
1323         }
1324
1325         RETURN(rc);
1326 }
1327
1328 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1329                          struct niobuf_local *lnb, int npages)
1330 {
1331         struct osd_thread_info *oti = osd_oti_get(env);
1332         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1333         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1334         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1335         int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1336         ktime_t start, end;
1337         s64 timediff;
1338         loff_t isize;
1339
1340         LASSERT(inode);
1341
1342         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1343         if (unlikely(rc != 0))
1344                 RETURN(rc);
1345
1346         isize = i_size_read(inode);
1347
1348         if (osd->od_read_cache)
1349                 cache = 1;
1350         if (isize > osd->od_readcache_max_filesize)
1351                 cache = 0;
1352
1353         start = ktime_get();
1354         for (i = 0; i < npages; i++) {
1355
1356                 if (isize <= lnb[i].lnb_file_offset)
1357                         /* If there's no more data, abort early.
1358                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1359                         break;
1360
1361                 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1362                         lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1363                 else
1364                         lnb[i].lnb_rc = lnb[i].lnb_len;
1365
1366                 /* Bypass disk read if fail_loc is set properly */
1367                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1368                         SetPageUptodate(lnb[i].lnb_page);
1369
1370                 if (PageUptodate(lnb[i].lnb_page)) {
1371                         cache_hits++;
1372                 } else {
1373                         cache_misses++;
1374                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1375                 }
1376
1377                 if (cache == 0)
1378                         generic_error_remove_page(inode->i_mapping,
1379                                                   lnb[i].lnb_page);
1380         }
1381         end = ktime_get();
1382         timediff = ktime_us_delta(end, start);
1383         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1384
1385         if (cache_hits != 0)
1386                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1387                                     cache_hits);
1388         if (cache_misses != 0)
1389                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1390                                     cache_misses);
1391         if (cache_hits + cache_misses != 0)
1392                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1393                                     cache_hits + cache_misses);
1394
1395         if (iobuf->dr_npages) {
1396                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1397                                                  iobuf->dr_npages,
1398                                                  iobuf->dr_blocks, 0);
1399                 rc = osd_do_bio(osd, inode, iobuf);
1400
1401                 /* IO stats will be done in osd_bufs_put() */
1402         }
1403
1404         RETURN(rc);
1405 }
1406
1407 /*
1408  * XXX: Another layering violation for now.
1409  *
1410  * We don't want to use ->f_op->read methods, because generic file write
1411  *
1412  *         - serializes on ->i_sem, and
1413  *
1414  *         - does a lot of extra work like balance_dirty_pages(),
1415  *
1416  * which doesn't work for globally shared files like /last_rcvd.
1417  */
1418 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1419 {
1420         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1421
1422         memcpy(buffer, (char *)ei->i_data, buflen);
1423
1424         return  buflen;
1425 }
1426
1427 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1428 {
1429         struct buffer_head *bh;
1430         unsigned long block;
1431         int osize;
1432         int blocksize;
1433         int csize;
1434         int boffs;
1435
1436         /* prevent reading after eof */
1437         spin_lock(&inode->i_lock);
1438         if (i_size_read(inode) < *offs + size) {
1439                 loff_t diff = i_size_read(inode) - *offs;
1440                 spin_unlock(&inode->i_lock);
1441                 if (diff < 0) {
1442                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1443                                i_size_read(inode), *offs);
1444                         return -EBADR;
1445                 } else if (diff == 0) {
1446                         return 0;
1447                 } else {
1448                         size = diff;
1449                 }
1450         } else {
1451                 spin_unlock(&inode->i_lock);
1452         }
1453
1454         blocksize = 1 << inode->i_blkbits;
1455         osize = size;
1456         while (size > 0) {
1457                 block = *offs >> inode->i_blkbits;
1458                 boffs = *offs & (blocksize - 1);
1459                 csize = min(blocksize - boffs, size);
1460                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1461                 if (IS_ERR(bh)) {
1462                         CERROR("%s: can't read %u@%llu on ino %lu: "
1463                                "rc = %ld\n", osd_ino2name(inode),
1464                                csize, *offs, inode->i_ino,
1465                                PTR_ERR(bh));
1466                         return PTR_ERR(bh);
1467                 }
1468
1469                 if (bh != NULL) {
1470                         memcpy(buf, bh->b_data + boffs, csize);
1471                         brelse(bh);
1472                 } else {
1473                         memset(buf, 0, csize);
1474                 }
1475
1476                 *offs += csize;
1477                 buf += csize;
1478                 size -= csize;
1479         }
1480         return osize;
1481 }
1482
1483 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1484                         struct lu_buf *buf, loff_t *pos)
1485 {
1486         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1487         int           rc;
1488
1489         /* Read small symlink from inode body as we need to maintain correct
1490          * on-disk symlinks for ldiskfs.
1491          */
1492         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1493             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1494                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1495         else
1496                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1497
1498         return rc;
1499 }
1500
1501 static inline int osd_extents_enabled(struct super_block *sb,
1502                                       struct inode *inode)
1503 {
1504         if (inode != NULL) {
1505                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1506                         return 1;
1507         } else if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
1508                                 LDISKFS_FEATURE_INCOMPAT_EXTENTS)) {
1509                 return 1;
1510         }
1511         return 0;
1512 }
1513
1514 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1515                            const loff_t size, const loff_t pos,
1516                            const int blocks)
1517 {
1518         int credits, bits, bs, i;
1519
1520         bits = sb->s_blocksize_bits;
1521         bs = 1 << bits;
1522
1523         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1524          * we do not expect blockmaps on the large files,
1525          * so let's shrink it to 2 levels (4GB files) */
1526
1527         /* this is default reservation: 2 levels */
1528         credits = (blocks + 2) * 3;
1529
1530         /* actual offset is unknown, hard to optimize */
1531         if (pos == -1)
1532                 return credits;
1533
1534         /* now check for few specific cases to optimize */
1535         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1536                 /* no indirects */
1537                 credits = blocks;
1538                 /* allocate if not allocated */
1539                 if (inode == NULL) {
1540                         credits += blocks * 2;
1541                         return credits;
1542                 }
1543                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1544                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1545                         if (LDISKFS_I(inode)->i_data[i] == 0)
1546                                 credits += 2;
1547                 }
1548         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1549                 /* single indirect */
1550                 credits = blocks * 3;
1551                 if (inode == NULL ||
1552                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1553                         credits += 3;
1554                 else
1555                         /* The indirect block may be modified. */
1556                         credits += 1;
1557         }
1558
1559         return credits;
1560 }
1561
1562 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1563                                  const struct lu_buf *buf, loff_t _pos,
1564                                  struct thandle *handle)
1565 {
1566         struct osd_object  *obj  = osd_dt_obj(dt);
1567         struct inode       *inode = obj->oo_inode;
1568         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1569         struct osd_thandle *oh;
1570         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1571         int                 bits, bs;
1572         int                 depth, size;
1573         loff_t              pos;
1574         ENTRY;
1575
1576         LASSERT(buf != NULL);
1577         LASSERT(handle != NULL);
1578
1579         oh = container_of0(handle, struct osd_thandle, ot_super);
1580         LASSERT(oh->ot_handle == NULL);
1581
1582         size = buf->lb_len;
1583         bits = sb->s_blocksize_bits;
1584         bs = 1 << bits;
1585
1586         if (_pos == -1) {
1587                 /* if this is an append, then we
1588                  * should expect cross-block record */
1589                 pos = 0;
1590         } else {
1591                 pos = _pos;
1592         }
1593
1594         /* blocks to modify */
1595         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1596         LASSERT(blocks > 0);
1597
1598         if (inode != NULL && _pos != -1) {
1599                 /* object size in blocks */
1600                 est = (i_size_read(inode) + bs - 1) >> bits;
1601                 allocated = inode->i_blocks >> (bits - 9);
1602                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1603                         /* looks like an overwrite, no need to modify tree */
1604                         credits = blocks;
1605                         /* no need to modify i_size */
1606                         goto out;
1607                 }
1608         }
1609
1610         if (osd_extents_enabled(sb, inode)) {
1611                 /*
1612                  * many concurrent threads may grow tree by the time
1613                  * our transaction starts. so, consider 2 is a min depth
1614                  * for every level we may need to allocate a new block
1615                  * and take some entries from the old one. so, 3 blocks
1616                  * to allocate (bitmap, gd, itself) + old block - 4 per
1617                  * level.
1618                  */
1619                 depth = inode != NULL ? ext_depth(inode) : 0;
1620                 depth = max(depth, 1) + 1;
1621                 credits = depth;
1622                 /* if not append, then split may need to modify
1623                  * existing blocks moving entries into the new ones */
1624                 if (_pos != -1)
1625                         credits += depth;
1626                 /* blocks to store data: bitmap,gd,itself */
1627                 credits += blocks * 3;
1628         } else {
1629                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1630         }
1631         /* if inode is created as part of the transaction,
1632          * then it's counted already by the creation method */
1633         if (inode != NULL)
1634                 credits++;
1635
1636 out:
1637
1638         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1639
1640         /* dt_declare_write() is usually called for system objects, such
1641          * as llog or last_rcvd files. We needn't enforce quota on those
1642          * objects, so always set the lqi_space as 0. */
1643         if (inode != NULL)
1644                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1645                                            i_gid_read(inode),
1646                                            i_projid_read(inode), 0,
1647                                            oh, obj, NULL, OSD_QID_BLK);
1648         RETURN(rc);
1649 }
1650
1651 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1652 {
1653         /* LU-2634: clear the extent format for fast symlink */
1654         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1655
1656         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1657         spin_lock(&inode->i_lock);
1658         LDISKFS_I(inode)->i_disksize = buflen;
1659         i_size_write(inode, buflen);
1660         spin_unlock(&inode->i_lock);
1661         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1662
1663         return 0;
1664 }
1665
1666 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1667                              int write_NUL, loff_t *offs, handle_t *handle)
1668 {
1669         struct buffer_head *bh        = NULL;
1670         loff_t              offset    = *offs;
1671         loff_t              new_size  = i_size_read(inode);
1672         unsigned long       block;
1673         int                 blocksize = 1 << inode->i_blkbits;
1674         int                 err = 0;
1675         int                 size;
1676         int                 boffs;
1677         int                 dirty_inode = 0;
1678
1679         if (write_NUL) {
1680                 /*
1681                  * long symlink write does not count the NUL terminator in
1682                  * bufsize, we write it, and the inode's file size does not
1683                  * count the NUL terminator as well.
1684                  */
1685                 ((char *)buf)[bufsize] = '\0';
1686                 ++bufsize;
1687         }
1688
1689         while (bufsize > 0) {
1690                 int credits = handle->h_buffer_credits;
1691
1692                 if (bh)
1693                         brelse(bh);
1694
1695                 block = offset >> inode->i_blkbits;
1696                 boffs = offset & (blocksize - 1);
1697                 size = min(blocksize - boffs, bufsize);
1698                 bh = __ldiskfs_bread(handle, inode, block, 1);
1699                 if (IS_ERR_OR_NULL(bh)) {
1700                         if (bh == NULL) {
1701                                 err = -EIO;
1702                         } else {
1703                                 err = PTR_ERR(bh);
1704                                 bh = NULL;
1705                         }
1706
1707                         CERROR("%s: error reading offset %llu (block %lu, "
1708                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
1709                                inode->i_sb->s_id, offset, block, bufsize, *offs,
1710                                credits, handle->h_buffer_credits, err);
1711                         break;
1712                 }
1713
1714                 err = ldiskfs_journal_get_write_access(handle, bh);
1715                 if (err) {
1716                         CERROR("journal_get_write_access() returned error %d\n",
1717                                err);
1718                         break;
1719                 }
1720                 LASSERTF(boffs + size <= bh->b_size,
1721                          "boffs %d size %d bh->b_size %lu\n",
1722                          boffs, size, (unsigned long)bh->b_size);
1723                 memcpy(bh->b_data + boffs, buf, size);
1724                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1725                 if (err)
1726                         break;
1727
1728                 if (offset + size > new_size)
1729                         new_size = offset + size;
1730                 offset += size;
1731                 bufsize -= size;
1732                 buf += size;
1733         }
1734         if (bh)
1735                 brelse(bh);
1736
1737         if (write_NUL)
1738                 --new_size;
1739         /* correct in-core and on-disk sizes */
1740         if (new_size > i_size_read(inode)) {
1741                 spin_lock(&inode->i_lock);
1742                 if (new_size > i_size_read(inode))
1743                         i_size_write(inode, new_size);
1744                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1745                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1746                         dirty_inode = 1;
1747                 }
1748                 spin_unlock(&inode->i_lock);
1749                 if (dirty_inode)
1750                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1751         }
1752
1753         if (err == 0)
1754                 *offs = offset;
1755         return err;
1756 }
1757
1758 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1759                          const struct lu_buf *buf, loff_t *pos,
1760                          struct thandle *handle, int ignore_quota)
1761 {
1762         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1763         struct osd_thandle      *oh;
1764         ssize_t                 result;
1765         int                     is_link;
1766
1767         LASSERT(dt_object_exists(dt));
1768
1769         LASSERT(handle != NULL);
1770         LASSERT(inode != NULL);
1771         ll_vfs_dq_init(inode);
1772
1773         /* XXX: don't check: one declared chunk can be used many times */
1774         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1775
1776         oh = container_of(handle, struct osd_thandle, ot_super);
1777         LASSERT(oh->ot_handle->h_transaction != NULL);
1778         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1779
1780         /* Write small symlink to inode body as we need to maintain correct
1781          * on-disk symlinks for ldiskfs.
1782          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1783          * does not count it in.
1784          */
1785         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1786         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1787                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1788         else
1789                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1790                                                   buf->lb_len, is_link, pos,
1791                                                   oh->ot_handle);
1792         if (result == 0)
1793                 result = buf->lb_len;
1794
1795         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1796
1797         return result;
1798 }
1799
1800 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1801                              __u64 start, __u64 end, struct thandle *th)
1802 {
1803         struct osd_thandle *oh;
1804         struct inode       *inode;
1805         int                 rc;
1806         ENTRY;
1807
1808         LASSERT(th);
1809         oh = container_of(th, struct osd_thandle, ot_super);
1810
1811         /*
1812          * we don't need to reserve credits for whole truncate
1813          * it's not possible as truncate may need to free too many
1814          * blocks and that won't fit a single transaction. instead
1815          * we reserve credits to change i_size and put inode onto
1816          * orphan list. if needed truncate will extend or restart
1817          * transaction
1818          */
1819         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1820                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1821
1822         inode = osd_dt_obj(dt)->oo_inode;
1823         LASSERT(inode);
1824
1825         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1826                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1827                                    NULL, OSD_QID_BLK);
1828         RETURN(rc);
1829 }
1830
1831 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1832                      __u64 start, __u64 end, struct thandle *th)
1833 {
1834         struct osd_thandle *oh;
1835         struct osd_object  *obj = osd_dt_obj(dt);
1836         struct inode       *inode = obj->oo_inode;
1837         handle_t           *h;
1838         tid_t               tid;
1839         int                rc = 0, rc2 = 0;
1840         ENTRY;
1841
1842         LASSERT(end == OBD_OBJECT_EOF);
1843         LASSERT(dt_object_exists(dt));
1844         LASSERT(osd_invariant(obj));
1845         LASSERT(inode != NULL);
1846         ll_vfs_dq_init(inode);
1847
1848         LASSERT(th);
1849         oh = container_of(th, struct osd_thandle, ot_super);
1850         LASSERT(oh->ot_handle->h_transaction != NULL);
1851
1852         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1853
1854         tid = oh->ot_handle->h_transaction->t_tid;
1855
1856         spin_lock(&inode->i_lock);
1857         i_size_write(inode, start);
1858         spin_unlock(&inode->i_lock);
1859         ll_truncate_pagecache(inode, start);
1860 #ifdef HAVE_INODEOPS_TRUNCATE
1861         if (inode->i_op->truncate) {
1862                 inode->i_op->truncate(inode);
1863         } else
1864 #endif
1865                 ldiskfs_truncate(inode);
1866
1867         /*
1868          * For a partial-page truncate, flush the page to disk immediately to
1869          * avoid data corruption during direct disk write.  b=17397
1870          */
1871         if ((start & ~PAGE_MASK) != 0)
1872                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1873
1874         h = journal_current_handle();
1875         LASSERT(h != NULL);
1876         LASSERT(h == oh->ot_handle);
1877
1878         /* do not check credits with osd_trans_exec_check() as the truncate
1879          * can restart the transaction internally and we restart the
1880          * transaction in this case */
1881
1882         if (tid != h->h_transaction->t_tid) {
1883                 int credits = oh->ot_credits;
1884                 /*
1885                  * transaction has changed during truncate
1886                  * we need to restart the handle with our credits
1887                  */
1888                 if (h->h_buffer_credits < credits) {
1889                         if (ldiskfs_journal_extend(h, credits))
1890                                 rc2 = ldiskfs_journal_restart(h, credits);
1891                 }
1892         }
1893
1894         RETURN(rc == 0 ? rc2 : rc);
1895 }
1896
1897 static int fiemap_check_ranges(struct inode *inode,
1898                                u64 start, u64 len, u64 *new_len)
1899 {
1900         loff_t maxbytes;
1901
1902         *new_len = len;
1903
1904         if (len == 0)
1905                 return -EINVAL;
1906
1907         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1908                 maxbytes = inode->i_sb->s_maxbytes;
1909         else
1910                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1911
1912         if (start > maxbytes)
1913                 return -EFBIG;
1914
1915         /*
1916          * Shrink request scope to what the fs can actually handle.
1917          */
1918         if (len > maxbytes || (maxbytes - len) < start)
1919                 *new_len = maxbytes - start;
1920
1921         return 0;
1922 }
1923
1924 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1925 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
1926
1927 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1928                           struct fiemap *fm)
1929 {
1930         struct fiemap_extent_info fieinfo = {0, };
1931         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1932         u64 len;
1933         int rc;
1934
1935
1936         LASSERT(inode);
1937         if (inode->i_op->fiemap == NULL)
1938                 return -EOPNOTSUPP;
1939
1940         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1941                 return -EINVAL;
1942
1943         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1944         if (rc)
1945                 return rc;
1946
1947         fieinfo.fi_flags = fm->fm_flags;
1948         fieinfo.fi_extents_max = fm->fm_extent_count;
1949         fieinfo.fi_extents_start = fm->fm_extents;
1950
1951         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1952                 filemap_write_and_wait(inode->i_mapping);
1953
1954         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
1955         fm->fm_flags = fieinfo.fi_flags;
1956         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
1957
1958         return rc;
1959 }
1960
1961 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1962                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1963 {
1964         int              rc = 0;
1965         struct inode    *inode = osd_dt_obj(dt)->oo_inode;
1966         ENTRY;
1967
1968         switch (advice) {
1969         case LU_LADVISE_DONTNEED:
1970                 if (end == 0)
1971                         break;
1972                 invalidate_mapping_pages(inode->i_mapping,
1973                                          start >> PAGE_CACHE_SHIFT,
1974                                          (end - 1) >> PAGE_CACHE_SHIFT);
1975                 break;
1976         default:
1977                 rc = -ENOTSUPP;
1978                 break;
1979         }
1980
1981         RETURN(rc);
1982 }
1983
1984 /*
1985  * in some cases we may need declare methods for objects being created
1986  * e.g., when we create symlink
1987  */
1988 const struct dt_body_operations osd_body_ops_new = {
1989         .dbo_declare_write = osd_declare_write,
1990 };
1991
1992 const struct dt_body_operations osd_body_ops = {
1993         .dbo_read                       = osd_read,
1994         .dbo_declare_write              = osd_declare_write,
1995         .dbo_write                      = osd_write,
1996         .dbo_bufs_get                   = osd_bufs_get,
1997         .dbo_bufs_put                   = osd_bufs_put,
1998         .dbo_write_prep                 = osd_write_prep,
1999         .dbo_declare_write_commit       = osd_declare_write_commit,
2000         .dbo_write_commit               = osd_write_commit,
2001         .dbo_read_prep                  = osd_read_prep,
2002         .dbo_declare_punch              = osd_declare_punch,
2003         .dbo_punch                      = osd_punch,
2004         .dbo_fiemap_get                 = osd_fiemap_get,
2005         .dbo_ladvise                    = osd_ladvise,
2006 };