lustre/osd-ldiskfs/osd_io.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/osd/osd_io.c
  37  *
  38  * body operations
  39  *
  40  * Author: Nikita Danilov <nikita@clusterfs.com>
  41  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  42  *
  43  */
  44
  45 /* LUSTRE_VERSION_CODE */
  46 #include <lustre_ver.h>
  47 /* prerequisite for linux/xattr.h */
  48 #include <linux/types.h>
  49 /* prerequisite for linux/xattr.h */
  50 #include <linux/fs.h>
  51
  52 /*
  53  * struct OBD_{ALLOC,FREE}*()
  54  * OBD_FAIL_CHECK
  55  */
  56 #include <obd_support.h>
  57
  58 #include "osd_internal.h"
  59
  60 /* ext_depth() */
  61 #include <ldiskfs/ldiskfs_extents.h>
  62
  63 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  64                             int rw, int line, int pages)
  65 {
  66         int blocks, i;
  67
  68         LASSERTF(iobuf->dr_elapsed_valid == 0,
  69                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  70                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  71                  iobuf->dr_init_at);
  72         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  73
  74         init_waitqueue_head(&iobuf->dr_wait);
  75         atomic_set(&iobuf->dr_numreqs, 0);
  76         iobuf->dr_npages = 0;
  77         iobuf->dr_error = 0;
  78         iobuf->dr_dev = d;
  79         iobuf->dr_frags = 0;
  80         iobuf->dr_elapsed = 0;
  81         /* must be counted before, so assert */
  82         iobuf->dr_rw = rw;
  83         iobuf->dr_init_at = line;
  84
  85         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  86         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  87                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  88                         pages * sizeof(iobuf->dr_pages[0]));
  89                 return 0;
  90         }
  91
  92         /* start with 1MB for 4K blocks */
  93         i = 256;
  94         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
  95                 i <<= 1;
  96
  97         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
  98                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
  99         pages = i;
 100         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
 101         iobuf->dr_max_pages = 0;
 102         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
 103                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 104
 105         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 106         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 107         if (unlikely(iobuf->dr_blocks == NULL))
 108                 return -ENOMEM;
 109
 110         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 111         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 112         if (unlikely(iobuf->dr_pages == NULL))
 113                 return -ENOMEM;
 114
 115         iobuf->dr_max_pages = pages;
 116
 117         return 0;
 118 }
 119 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 120         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 121
 122 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
 123 {
 124         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 125         iobuf->dr_pages[iobuf->dr_npages++] = page;
 126 }
 127
 128 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 129 {
 130         int rw = iobuf->dr_rw;
 131
 132         if (iobuf->dr_elapsed_valid) {
 133                 iobuf->dr_elapsed_valid = 0;
 134                 LASSERT(iobuf->dr_dev == d);
 135                 LASSERT(iobuf->dr_frags > 0);
 136                 lprocfs_oh_tally(&d->od_brw_stats.
 137                                  hist[BRW_R_DIO_FRAGS+rw],
 138                                  iobuf->dr_frags);
 139                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 140                                       iobuf->dr_elapsed);
 141         }
 142 }
 143
 144 #ifndef REQ_WRITE /* pre-2.6.35 */
 145 #define __REQ_WRITE BIO_RW
 146 #endif
 147
 148 static void dio_complete_routine(struct bio *bio, int error)
 149 {
 150         struct osd_iobuf *iobuf = bio->bi_private;
 151 #ifdef HAVE_BVEC_ITER
 152         struct bvec_iter iter;
 153         struct bio_vec bvl;
 154 #else
 155         int iter;
 156         struct bio_vec *bvl;
 157 #endif
 158
 159         /* CAVEAT EMPTOR: possibly in IRQ context
 160          * DO NOT record procfs stats here!!! */
 161
 162         if (unlikely(iobuf == NULL)) {
 163                 CERROR("***** bio->bi_private is NULL!  This should never "
 164                        "happen.  Normally, I would crash here, but instead I "
 165                        "will dump the bio contents to the console.  Please "
 166                        "report this to <https://jira.hpdd.intel.com/> , along "
 167                        "with any interesting messages leading up to this point "
 168                        "(like SCSI errors, perhaps).  Because bi_private is "
 169                        "NULL, I can't wake up the thread that initiated this "
 170                        "IO - you will probably have to reboot this node.\n");
 171                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 172                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 173                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 174                         bio->bi_rw, bio->bi_vcnt, bio_idx(bio),
 175                         bio_sectors(bio) << 9, bio->bi_end_io,
 176                         atomic_read(&bio->bi_cnt), bio->bi_private);
 177                 return;
 178         }
 179
 180         /* the check is outside of the cycle for performance reason -bzzz */
 181         if (!test_bit(__REQ_WRITE, &bio->bi_rw)) {
 182                 bio_for_each_segment_all(bvl, bio, iter) {
 183                         if (likely(error == 0))
 184                                 SetPageUptodate(bvl_to_page(bvl));
 185                         LASSERT(PageLocked(bvl_to_page(bvl)));
 186                 }
 187                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 188         } else {
 189                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 190         }
 191
 192         /* any real error is good enough -bzzz */
 193         if (error != 0 && iobuf->dr_error == 0)
 194                 iobuf->dr_error = error;
 195
 196         /*
 197          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 198          * it's possible that service thread will see dr_numreqs
 199          * is zero, but dr_elapsed is not set yet, leading to lost
 200          * data in this processing and an assertion in a subsequent
 201          * call to OSD.
 202          */
 203         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 204                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
 205                 iobuf->dr_elapsed_valid = 1;
 206         }
 207         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 208                 wake_up(&iobuf->dr_wait);
 209
 210         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 211          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 212          * mempool when serious on-disk fragmentation was encountered,
 213          * deadlocking the OST.  The bios are now released as soon as complete
 214          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 215         bio_put(bio);
 216 }
 217
 218 static void record_start_io(struct osd_iobuf *iobuf, int size)
 219 {
 220         struct osd_device    *osd = iobuf->dr_dev;
 221         struct obd_histogram *h = osd->od_brw_stats.hist;
 222
 223         iobuf->dr_frags++;
 224         atomic_inc(&iobuf->dr_numreqs);
 225
 226         if (iobuf->dr_rw == 0) {
 227                 atomic_inc(&osd->od_r_in_flight);
 228                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 229                                  atomic_read(&osd->od_r_in_flight));
 230                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 231         } else if (iobuf->dr_rw == 1) {
 232                 atomic_inc(&osd->od_w_in_flight);
 233                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 234                                  atomic_read(&osd->od_w_in_flight));
 235                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 236         } else {
 237                 LBUG();
 238         }
 239 }
 240
 241 static void osd_submit_bio(int rw, struct bio *bio)
 242 {
 243         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 244         if (rw == 0)
 245                 submit_bio(READ, bio);
 246         else
 247                 submit_bio(WRITE, bio);
 248 }
 249
 250 static int can_be_merged(struct bio *bio, sector_t sector)
 251 {
 252         if (bio == NULL)
 253                 return 0;
 254
 255         return bio_end_sector(bio) == sector ? 1 : 0;
 256 }
 257
 258 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 259                       struct osd_iobuf *iobuf)
 260 {
 261         int            blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 262         struct page  **pages = iobuf->dr_pages;
 263         int            npages = iobuf->dr_npages;
 264         sector_t      *blocks = iobuf->dr_blocks;
 265         int            total_blocks = npages * blocks_per_page;
 266         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 267         unsigned int   blocksize = inode->i_sb->s_blocksize;
 268         struct bio    *bio = NULL;
 269         struct page   *page;
 270         unsigned int   page_offset;
 271         sector_t       sector;
 272         int            nblocks;
 273         int            block_idx;
 274         int            page_idx;
 275         int            i;
 276         int            rc = 0;
 277         ENTRY;
 278
 279         LASSERT(iobuf->dr_npages == npages);
 280
 281         osd_brw_stats_update(osd, iobuf);
 282         iobuf->dr_start_time = cfs_time_current();
 283
 284         for (page_idx = 0, block_idx = 0;
 285              page_idx < npages;
 286              page_idx++, block_idx += blocks_per_page) {
 287
 288                 page = pages[page_idx];
 289                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 290
 291                 for (i = 0, page_offset = 0;
 292                      i < blocks_per_page;
 293                      i += nblocks, page_offset += blocksize * nblocks) {
 294
 295                         nblocks = 1;
 296
 297                         if (blocks[block_idx + i] == 0) {  /* hole */
 298                                 LASSERTF(iobuf->dr_rw == 0,
 299                                          "page_idx %u, block_idx %u, i %u\n",
 300                                          page_idx, block_idx, i);
 301                                 memset(kmap(page) + page_offset, 0, blocksize);
 302                                 kunmap(page);
 303                                 continue;
 304                         }
 305
 306                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 307
 308                         /* Additional contiguous file blocks? */
 309                         while (i + nblocks < blocks_per_page &&
 310                                (sector + (nblocks << sector_bits)) ==
 311                                ((sector_t)blocks[block_idx + i + nblocks] <<
 312                                 sector_bits))
 313                                 nblocks++;
 314
 315                         if (bio != NULL &&
 316                             can_be_merged(bio, sector) &&
 317                             bio_add_page(bio, page,
 318                                          blocksize * nblocks, page_offset) != 0)
 319                                 continue;       /* added this frag OK */
 320
 321                         if (bio != NULL) {
 322                                 struct request_queue *q =
 323                                         bdev_get_queue(bio->bi_bdev);
 324                                 unsigned int bi_size = bio_sectors(bio) << 9;
 325
 326                                 /* Dang! I have to fragment this I/O */
 327                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 328                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 329                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 330                                        bio_sectors(bio),
 331                                        queue_max_sectors(q),
 332                                        bio_phys_segments(q, bio),
 333                                        queue_max_phys_segments(q),
 334                                        0, queue_max_hw_segments(q));
 335                                 record_start_io(iobuf, bi_size);
 336                                 osd_submit_bio(iobuf->dr_rw, bio);
 337                         }
 338
 339                         /* allocate new bio */
 340                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 341                                                       (npages - page_idx) *
 342                                                       blocks_per_page));
 343                         if (bio == NULL) {
 344                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 345                                        (npages - page_idx), blocks_per_page,
 346                                        (npages - page_idx) * blocks_per_page);
 347                                 rc = -ENOMEM;
 348                                 goto out;
 349                         }
 350
 351                         bio->bi_bdev = inode->i_sb->s_bdev;
 352                         bio_set_sector(bio, sector);
 353                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 354                         bio->bi_end_io = dio_complete_routine;
 355                         bio->bi_private = iobuf;
 356
 357                         rc = bio_add_page(bio, page,
 358                                           blocksize * nblocks, page_offset);
 359                         LASSERT(rc != 0);
 360                 }
 361         }
 362
 363         if (bio != NULL) {
 364                 record_start_io(iobuf, bio_sectors(bio) << 9);
 365                 osd_submit_bio(iobuf->dr_rw, bio);
 366                 rc = 0;
 367         }
 368
 369 out:
 370         /* in order to achieve better IO throughput, we don't wait for writes
 371          * completion here. instead we proceed with transaction commit in
 372          * parallel and wait for IO completion once transaction is stopped
 373          * see osd_trans_stop() for more details -bzzz */
 374         if (iobuf->dr_rw == 0) {
 375                 wait_event(iobuf->dr_wait,
 376                            atomic_read(&iobuf->dr_numreqs) == 0);
 377                 osd_fini_iobuf(osd, iobuf);
 378         }
 379
 380         if (rc == 0)
 381                 rc = iobuf->dr_error;
 382         RETURN(rc);
 383 }
 384
 385 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 386                                    struct niobuf_local *lnb)
 387 {
 388         ENTRY;
 389
 390         *nrpages = 0;
 391
 392         while (len > 0) {
 393                 int poff = offset & (PAGE_SIZE - 1);
 394                 int plen = PAGE_SIZE - poff;
 395
 396                 if (plen > len)
 397                         plen = len;
 398                 lnb->lnb_file_offset = offset;
 399                 lnb->lnb_page_offset = poff;
 400                 lnb->lnb_len = plen;
 401                 /* lnb->lnb_flags = rnb->rnb_flags; */
 402                 lnb->lnb_flags = 0;
 403                 lnb->lnb_page = NULL;
 404                 lnb->lnb_rc = 0;
 405
 406                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 407                          (long long) len);
 408                 offset += plen;
 409                 len -= plen;
 410                 lnb++;
 411                 (*nrpages)++;
 412         }
 413
 414         RETURN(0);
 415 }
 416
 417 static struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
 418 {
 419         struct inode      *inode = osd_dt_obj(dt)->oo_inode;
 420         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 421         struct page       *page;
 422
 423         LASSERT(inode);
 424
 425         page = find_or_create_page(inode->i_mapping, offset >> PAGE_SHIFT,
 426                                    GFP_NOFS | __GFP_HIGHMEM);
 427         if (unlikely(page == NULL))
 428                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 429
 430         return page;
 431 }
 432
 433 /*
 434  * there are following "locks":
 435  * journal_start
 436  * i_mutex
 437  * page lock
 438  *
 439  * osd write path:
 440  *  - lock page(s)
 441  *  - journal_start
 442  *  - truncate_sem
 443  *
 444  * ext4 vmtruncate:
 445  *  - lock pages, unlock
 446  *  - journal_start
 447  *  - lock partial page
 448  *  - i_data_sem
 449  *
 450  */
 451
 452 /**
 453  * Unlock and release pages loaded by osd_bufs_get()
 454  *
 455  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 456  *
 457  * \param env           thread execution environment
 458  * \param dt            dt object undergoing IO (OSD object + methods)
 459  * \param lnb           array of pages undergoing IO
 460  * \param npages        number of pages in \a lnb
 461  *
 462  * \retval 0            always
 463  */
 464 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 465                         struct niobuf_local *lnb, int npages)
 466 {
 467         int i;
 468
 469         for (i = 0; i < npages; i++) {
 470                 if (lnb[i].lnb_page == NULL)
 471                         continue;
 472                 LASSERT(PageLocked(lnb[i].lnb_page));
 473                 unlock_page(lnb[i].lnb_page);
 474                 put_page(lnb[i].lnb_page);
 475                 lu_object_put(env, &dt->do_lu);
 476                 lnb[i].lnb_page = NULL;
 477         }
 478
 479         RETURN(0);
 480 }
 481
 482 /**
 483  * Load and lock pages undergoing IO
 484  *
 485  * Pages as described in the \a lnb array are fetched (from disk or cache)
 486  * and locked for IO by the caller.
 487  *
 488  * DLM locking protects us from write and truncate competing for same region,
 489  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 490  * It's possible the writeout on a such a page is in progress when we access
 491  * it. It's also possible that during this writeout we put new (partial) data
 492  * into the page, but won't be able to proceed in filter_commitrw_write().
 493  * Therefore, just wait for writeout completion as it should be rare enough.
 494  *
 495  * \param env           thread execution environment
 496  * \param dt            dt object undergoing IO (OSD object + methods)
 497  * \param pos           byte offset of IO start
 498  * \param len           number of bytes of IO
 499  * \param lnb           array of extents undergoing IO
 500  * \param rw            read or write operation?
 501  * \param capa          capabilities
 502  *
 503  * \retval pages        (zero or more) loaded successfully
 504  * \retval -ENOMEM      on memory/page allocation error
 505  */
 506 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 507                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 508                         int rw)
 509 {
 510         struct osd_object   *obj    = osd_dt_obj(dt);
 511         int npages, i, rc = 0;
 512
 513         LASSERT(obj->oo_inode);
 514
 515         osd_map_remote_to_local(pos, len, &npages, lnb);
 516
 517         for (i = 0; i < npages; i++, lnb++) {
 518                 lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset, rw);
 519                 if (lnb->lnb_page == NULL)
 520                         GOTO(cleanup, rc = -ENOMEM);
 521
 522                 wait_on_page_writeback(lnb->lnb_page);
 523                 BUG_ON(PageWriteback(lnb->lnb_page));
 524
 525                 lu_object_get(&dt->do_lu);
 526         }
 527
 528         RETURN(i);
 529
 530 cleanup:
 531         if (i > 0)
 532                 osd_bufs_put(env, dt, lnb - i, i);
 533         return rc;
 534 }
 535
 536 #ifndef HAVE_LDISKFS_MAP_BLOCKS
 537
 538 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 539 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 540 #endif
 541
 542 struct bpointers {
 543         sector_t *blocks;
 544         unsigned long start;
 545         int num;
 546         int init_num;
 547         int create;
 548 };
 549
 550 static long ldiskfs_ext_find_goal(struct inode *inode,
 551                                   struct ldiskfs_ext_path *path,
 552                                   unsigned long block, int *aflags)
 553 {
 554         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 555         unsigned long bg_start;
 556         unsigned long colour;
 557         int depth;
 558
 559         if (path) {
 560                 struct ldiskfs_extent *ex;
 561                 depth = path->p_depth;
 562
 563                 /* try to predict block placement */
 564                 if ((ex = path[depth].p_ext))
 565                         return ldiskfs_ext_pblock(ex) +
 566                                 (block - le32_to_cpu(ex->ee_block));
 567
 568                 /* it looks index is empty
 569                  * try to find starting from index itself */
 570                 if (path[depth].p_bh)
 571                         return path[depth].p_bh->b_blocknr;
 572         }
 573
 574         /* OK. use inode's group */
 575         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 576                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 577         colour = (current->pid % 16) *
 578                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 579         return bg_start + colour + block;
 580 }
 581
 582 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 583                                 struct ldiskfs_ext_path *path,
 584                                 unsigned long block, unsigned long *count,
 585                                 int *err)
 586 {
 587         struct ldiskfs_allocation_request ar;
 588         unsigned long pblock;
 589         int aflags;
 590
 591         /* find neighbour allocated blocks */
 592         ar.lleft = block;
 593         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 594         if (*err)
 595                 return 0;
 596         ar.lright = block;
 597         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 598         if (*err)
 599                 return 0;
 600
 601         /* allocate new block */
 602         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 603         ar.inode = inode;
 604         ar.logical = block;
 605         ar.len = *count;
 606         ar.flags = LDISKFS_MB_HINT_DATA;
 607         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 608         *count = ar.len;
 609         return pblock;
 610 }
 611
 612 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 613                                      struct ldiskfs_ext_path *path,
 614                                      struct ldiskfs_ext_cache *cex,
 615 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 616                                      struct ldiskfs_extent *ex,
 617 #endif
 618                                      void *cbdata)
 619 {
 620         struct bpointers *bp = cbdata;
 621         struct ldiskfs_extent nex;
 622         unsigned long pblock = 0;
 623         unsigned long tgen;
 624         int err, i;
 625         unsigned long count;
 626         handle_t *handle;
 627
 628 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 629         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 630 #else
 631         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 632 #endif
 633                 err = EXT_CONTINUE;
 634                 goto map;
 635         }
 636
 637         if (bp->create == 0) {
 638                 i = 0;
 639                 if (cex->ec_block < bp->start)
 640                         i = bp->start - cex->ec_block;
 641                 if (i >= cex->ec_len)
 642                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 643                                         i, cex->ec_len);
 644                 for (; i < cex->ec_len && bp->num; i++) {
 645                         *(bp->blocks) = 0;
 646                         bp->blocks++;
 647                         bp->num--;
 648                         bp->start++;
 649                 }
 650
 651                 return EXT_CONTINUE;
 652         }
 653
 654         tgen = LDISKFS_I(inode)->i_ext_generation;
 655         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
 656
 657         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
 658                                    count + LDISKFS_ALLOC_NEEDED + 1);
 659         if (IS_ERR(handle)) {
 660                 return PTR_ERR(handle);
 661         }
 662
 663         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
 664                 /* the tree has changed. so path can be invalid at moment */
 665                 ldiskfs_journal_stop(handle);
 666                 return EXT_REPEAT;
 667         }
 668
 669         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
 670          * protected by i_data_sem as whole. so we patch it to store
 671          * generation to path and now verify the tree hasn't changed */
 672         down_write((&LDISKFS_I(inode)->i_data_sem));
 673
 674         /* validate extent, make sure the extent tree does not changed */
 675         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
 676                 /* cex is invalid, try again */
 677                 up_write(&LDISKFS_I(inode)->i_data_sem);
 678                 ldiskfs_journal_stop(handle);
 679                 return EXT_REPEAT;
 680         }
 681
 682         count = cex->ec_len;
 683         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
 684         if (!pblock)
 685                 goto out;
 686         BUG_ON(count > cex->ec_len);
 687
 688         /* insert new extent */
 689         nex.ee_block = cpu_to_le32(cex->ec_block);
 690         ldiskfs_ext_store_pblock(&nex, pblock);
 691         nex.ee_len = cpu_to_le16(count);
 692         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
 693         if (err) {
 694                 /* free data blocks we just allocated */
 695                 /* not a good idea to call discard here directly,
 696                  * but otherwise we'd need to call it every free() */
 697                 ldiskfs_discard_preallocations(inode);
 698 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
 699                 ldiskfs_free_blocks(handle, inode, NULL,
 700                                     ldiskfs_ext_pblock(&nex),
 701                                     le16_to_cpu(nex.ee_len), 0);
 702 #else
 703                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
 704                                     le16_to_cpu(nex.ee_len), 0);
 705 #endif
 706                 goto out;
 707         }
 708
 709         /*
 710          * Putting len of the actual extent we just inserted,
 711          * we are asking ldiskfs_ext_walk_space() to continue
 712          * scaning after that block
 713          */
 714         cex->ec_len = le16_to_cpu(nex.ee_len);
 715         cex->ec_start = ldiskfs_ext_pblock(&nex);
 716         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
 717         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
 718
 719 out:
 720         up_write((&LDISKFS_I(inode)->i_data_sem));
 721         ldiskfs_journal_stop(handle);
 722 map:
 723         if (err >= 0) {
 724                 /* map blocks */
 725                 if (bp->num == 0) {
 726                         CERROR("hmm. why do we find this extent?\n");
 727                         CERROR("initial space: %lu:%u\n",
 728                                 bp->start, bp->init_num);
 729 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 730                         CERROR("current extent: %u/%u/%llu %d\n",
 731                                 cex->ec_block, cex->ec_len,
 732                                 (unsigned long long)cex->ec_start,
 733                                 cex->ec_type);
 734 #else
 735                         CERROR("current extent: %u/%u/%llu\n",
 736                                 cex->ec_block, cex->ec_len,
 737                                 (unsigned long long)cex->ec_start);
 738 #endif
 739                 }
 740                 i = 0;
 741                 if (cex->ec_block < bp->start)
 742                         i = bp->start - cex->ec_block;
 743                 if (i >= cex->ec_len)
 744                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 745                                         i, cex->ec_len);
 746                 for (; i < cex->ec_len && bp->num; i++) {
 747                         *(bp->blocks) = cex->ec_start + i;
 748                         if (pblock != 0) {
 749                                 /* unmap any possible underlying metadata from
 750                                  * the block device mapping.  bug 6998. */
 751                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
 752                                                           *(bp->blocks));
 753                         }
 754                         bp->blocks++;
 755                         bp->num--;
 756                         bp->start++;
 757                 }
 758         }
 759         return err;
 760 }
 761
 762 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
 763                                    int clen, sector_t *blocks, int create)
 764 {
 765         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 766         struct bpointers bp;
 767         int err;
 768
 769         if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
 770                 return -EFBIG;
 771
 772         bp.blocks = blocks;
 773         bp.start = index * blocks_per_page;
 774         bp.init_num = bp.num = clen * blocks_per_page;
 775         bp.create = create;
 776
 777         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 778                bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
 779
 780         err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
 781                                      ldiskfs_ext_new_extent_cb, &bp);
 782         ldiskfs_ext_invalidate_cache(inode);
 783
 784         return err;
 785 }
 786
 787 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
 788                                           struct page **page, int pages,
 789                                           sector_t *blocks, int create)
 790 {
 791         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 792         pgoff_t bitmap_max_page_index;
 793         sector_t *b;
 794         int rc = 0, i;
 795
 796         bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
 797                                 PAGE_SHIFT;
 798         for (i = 0, b = blocks; i < pages; i++, page++) {
 799                 if ((*page)->index + 1 >= bitmap_max_page_index) {
 800                         rc = -EFBIG;
 801                         break;
 802                 }
 803                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
 804                 if (rc) {
 805                         CERROR("ino %lu, blk %llu create %d: rc %d\n",
 806                                inode->i_ino,
 807                                (unsigned long long)*b, create, rc);
 808                         break;
 809                 }
 810                 b += blocks_per_page;
 811         }
 812         return rc;
 813 }
 814
 815 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
 816                                            struct page **page,
 817                                            int pages, sector_t *blocks,
 818                                            int create)
 819 {
 820         int rc = 0, i = 0, clen = 0;
 821         struct page *fp = NULL;
 822
 823         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 824                 inode->i_ino, pages, (*page)->index);
 825
 826         /* pages are sorted already. so, we just have to find
 827          * contig. space and process them properly */
 828         while (i < pages) {
 829                 if (fp == NULL) {
 830                         /* start new extent */
 831                         fp = *page++;
 832                         clen = 1;
 833                         i++;
 834                         continue;
 835                 } else if (fp->index + clen == (*page)->index) {
 836                         /* continue the extent */
 837                         page++;
 838                         clen++;
 839                         i++;
 840                         continue;
 841                 }
 842
 843                 /* process found extent */
 844                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 845                                              blocks, create);
 846                 if (rc)
 847                         GOTO(cleanup, rc);
 848
 849                 /* look for next extent */
 850                 fp = NULL;
 851                 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
 852         }
 853
 854         if (fp)
 855                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 856                                              blocks, create);
 857
 858 cleanup:
 859         return rc;
 860 }
 861
 862 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 863                                        int pages, sector_t *blocks,
 864                                        int create)
 865 {
 866         int rc;
 867
 868         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 869                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
 870                                                      blocks, create);
 871                 return rc;
 872         }
 873         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
 874
 875         return rc;
 876 }
 877 #else
 878 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 879                                        int pages, sector_t *blocks,
 880                                        int create)
 881 {
 882         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 883         int rc = 0, i = 0;
 884         struct page *fp = NULL;
 885         int clen = 0;
 886         pgoff_t max_page_index;
 887
 888         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
 889
 890         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 891                 inode->i_ino, pages, (*page)->index);
 892
 893         /* pages are sorted already. so, we just have to find
 894          * contig. space and process them properly */
 895         while (i < pages) {
 896                 long blen, total = 0;
 897                 handle_t *handle = NULL;
 898                 struct ldiskfs_map_blocks map = { 0 };
 899
 900                 if (fp == NULL) { /* start new extent */
 901                         fp = *page++;
 902                         clen = 1;
 903                         if (++i != pages)
 904                                 continue;
 905                 } else if (fp->index + clen == (*page)->index) {
 906                         /* continue the extent */
 907                         page++;
 908                         clen++;
 909                         if (++i != pages)
 910                                 continue;
 911                 }
 912                 if (fp->index + clen >= max_page_index)
 913                         GOTO(cleanup, rc = -EFBIG);
 914                 /* process found extent */
 915                 map.m_lblk = fp->index * blocks_per_page;
 916                 map.m_len = blen = clen * blocks_per_page;
 917                 if (create) {
 918                         create = LDISKFS_GET_BLOCKS_CREATE;
 919                         handle = ldiskfs_journal_current_handle();
 920                         LASSERT(handle != NULL);
 921                 }
 922 cont_map:
 923                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 924                 if (rc >= 0) {
 925                         int c = 0;
 926                         for (; total < blen && c < map.m_len; c++, total++) {
 927                                 if (rc == 0) {
 928                                         *(blocks + total) = 0;
 929                                         total++;
 930                                         break;
 931                                 } else {
 932                                         *(blocks + total) = map.m_pblk + c;
 933                                         /* unmap any possible underlying
 934                                          * metadata from the block device
 935                                          * mapping.  bug 6998. */
 936                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
 937                                             create)
 938                                                 unmap_underlying_metadata(
 939                                                         inode->i_sb->s_bdev,
 940                                                         map.m_pblk + c);
 941                                 }
 942                         }
 943                         rc = 0;
 944                 }
 945                 if (rc == 0 && total < blen) {
 946                         map.m_lblk = fp->index * blocks_per_page + total;
 947                         map.m_len = blen - total;
 948                         goto cont_map;
 949                 }
 950                 if (rc != 0)
 951                         GOTO(cleanup, rc);
 952
 953                 /* look for next extent */
 954                 fp = NULL;
 955                 blocks += blocks_per_page * clen;
 956         }
 957 cleanup:
 958         return rc;
 959 }
 960 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
 961
 962 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 963                           struct niobuf_local *lnb, int npages)
 964 {
 965         struct osd_thread_info *oti   = osd_oti_get(env);
 966         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 967         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 968         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 969         struct timeval          start;
 970         struct timeval          end;
 971         unsigned long           timediff;
 972         ssize_t                 isize;
 973         __s64                   maxidx;
 974         int                     rc = 0;
 975         int                     i;
 976         int                     cache = 0;
 977
 978         LASSERT(inode);
 979
 980         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 981         if (unlikely(rc != 0))
 982                 RETURN(rc);
 983
 984         isize = i_size_read(inode);
 985         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
 986
 987         if (osd->od_writethrough_cache)
 988                 cache = 1;
 989         if (isize > osd->od_readcache_max_filesize)
 990                 cache = 0;
 991
 992         do_gettimeofday(&start);
 993         for (i = 0; i < npages; i++) {
 994
 995                 if (cache == 0)
 996                         generic_error_remove_page(inode->i_mapping,
 997                                                   lnb[i].lnb_page);
 998
 999                 /*
1000                  * till commit the content of the page is undefined
1001                  * we'll set it uptodate once bulk is done. otherwise
1002                  * subsequent reads can access non-stable data
1003                  */
1004                 ClearPageUptodate(lnb[i].lnb_page);
1005
1006                 if (lnb[i].lnb_len == PAGE_SIZE)
1007                         continue;
1008
1009                 if (maxidx >= lnb[i].lnb_page->index) {
1010                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1011                 } else {
1012                         long off;
1013                         char *p = kmap(lnb[i].lnb_page);
1014
1015                         off = lnb[i].lnb_page_offset;
1016                         if (off)
1017                                 memset(p, 0, off);
1018                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1019                               ~PAGE_MASK;
1020                         if (off)
1021                                 memset(p + off, 0, PAGE_SIZE - off);
1022                         kunmap(lnb[i].lnb_page);
1023                 }
1024         }
1025         do_gettimeofday(&end);
1026         timediff = cfs_timeval_sub(&end, &start, NULL);
1027         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1028
1029         if (iobuf->dr_npages) {
1030                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1031                                                  iobuf->dr_npages,
1032                                                  iobuf->dr_blocks, 0);
1033                 if (likely(rc == 0)) {
1034                         rc = osd_do_bio(osd, inode, iobuf);
1035                         /* do IO stats for preparation reads */
1036                         osd_fini_iobuf(osd, iobuf);
1037                 }
1038         }
1039         RETURN(rc);
1040 }
1041
1042 struct osd_fextent {
1043         sector_t        start;
1044         sector_t        end;
1045         unsigned int    mapped:1;
1046 };
1047
1048 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1049                          struct osd_fextent *cached_extent)
1050 {
1051         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1052         sector_t block = offset >> inode->i_blkbits;
1053         sector_t start;
1054         struct fiemap_extent_info fei = { 0 };
1055         struct fiemap_extent fe = { 0 };
1056         mm_segment_t saved_fs;
1057         int rc;
1058
1059         if (block >= cached_extent->start && block < cached_extent->end)
1060                 return cached_extent->mapped;
1061
1062         if (i_size_read(inode) == 0)
1063                 return 0;
1064
1065         /* Beyond EOF, must not be mapped */
1066         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1067                 return 0;
1068
1069         fei.fi_extents_max = 1;
1070         fei.fi_extents_start = &fe;
1071
1072         saved_fs = get_fs();
1073         set_fs(get_ds());
1074         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1075         set_fs(saved_fs);
1076         if (rc != 0)
1077                 return 0;
1078
1079         start = fe.fe_logical >> inode->i_blkbits;
1080
1081         if (start > block) {
1082                 cached_extent->start = block;
1083                 cached_extent->end = start;
1084                 cached_extent->mapped = 0;
1085         } else {
1086                 cached_extent->start = start;
1087                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1088                                       inode->i_blkbits;
1089                 cached_extent->mapped = 1;
1090         }
1091
1092         return cached_extent->mapped;
1093 }
1094
1095 static int osd_declare_write_commit(const struct lu_env *env,
1096                                     struct dt_object *dt,
1097                                     struct niobuf_local *lnb, int npages,
1098                                     struct thandle *handle)
1099 {
1100         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1101         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1102         struct osd_thandle      *oh;
1103         int                      extents = 1;
1104         int                      depth;
1105         int                      i;
1106         int                      newblocks;
1107         int                      rc = 0;
1108         int                      flags = 0;
1109         int                      credits = 0;
1110         bool                     ignore_quota = false;
1111         long long                quota_space = 0;
1112         struct osd_fextent       extent = { 0 };
1113         ENTRY;
1114
1115         LASSERT(handle != NULL);
1116         oh = container_of0(handle, struct osd_thandle, ot_super);
1117         LASSERT(oh->ot_handle == NULL);
1118
1119         newblocks = npages;
1120
1121         /* calculate number of extents (probably better to pass nb) */
1122         for (i = 0; i < npages; i++) {
1123                 if (i && lnb[i].lnb_file_offset !=
1124                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1125                         extents++;
1126
1127                 if (!osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1128                         quota_space += PAGE_SIZE;
1129
1130                 /* ignore quota for the whole request if any page is from
1131                  * client cache or written by root.
1132                  *
1133                  * XXX once we drop the 1.8 client support, the checking
1134                  * for whether page is from cache can be simplified as:
1135                  * !(lnb[i].flags & OBD_BRW_SYNC)
1136                  *
1137                  * XXX we could handle this on per-lnb basis as done by
1138                  * grant. */
1139                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1140                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1141                     OBD_BRW_FROM_GRANT)
1142                         ignore_quota = true;
1143         }
1144
1145         /*
1146          * each extent can go into new leaf causing a split
1147          * 5 is max tree depth: inode + 4 index blocks
1148          * with blockmaps, depth is 3 at most
1149          */
1150         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1151                 /*
1152                  * many concurrent threads may grow tree by the time
1153                  * our transaction starts. so, consider 2 is a min depth
1154                  */
1155                 depth = ext_depth(inode);
1156                 depth = max(depth, 1) + 1;
1157                 newblocks += depth;
1158                 credits++; /* inode */
1159                 credits += depth * 2 * extents;
1160         } else {
1161                 depth = 3;
1162                 newblocks += depth;
1163                 credits++; /* inode */
1164                 credits += depth * extents;
1165         }
1166
1167         /* quota space for metadata blocks */
1168         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1169
1170         /* quota space should be reported in 1K blocks */
1171         quota_space = toqb(quota_space);
1172
1173         /* each new block can go in different group (bitmap + gd) */
1174
1175         /* we can't dirty more bitmap blocks than exist */
1176         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1177                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1178         else
1179                 credits += newblocks;
1180
1181         /* we can't dirty more gd blocks than exist */
1182         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1183                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1184         else
1185                 credits += newblocks;
1186
1187         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1188
1189         /* make sure the over quota flags were not set */
1190         lnb[0].lnb_flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA);
1191
1192         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1193                                    quota_space, oh, osd_dt_obj(dt), true,
1194                                    &flags, ignore_quota);
1195
1196         /* we need only to store the overquota flags in the first lnb for
1197          * now, once we support multiple objects BRW, this code needs be
1198          * revised. */
1199         if (flags & QUOTA_FL_OVER_USRQUOTA)
1200                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1201         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1202                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1203
1204         RETURN(rc);
1205 }
1206
1207 /* Check if a block is allocated or not */
1208 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1209                             struct niobuf_local *lnb, int npages,
1210                             struct thandle *thandle)
1211 {
1212         struct osd_thread_info *oti = osd_oti_get(env);
1213         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1214         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1215         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1216         loff_t isize;
1217         int rc = 0, i;
1218         struct osd_fextent extent = { 0 };
1219
1220         LASSERT(inode);
1221
1222         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1223         if (unlikely(rc != 0))
1224                 RETURN(rc);
1225
1226         isize = i_size_read(inode);
1227         ll_vfs_dq_init(inode);
1228
1229         for (i = 0; i < npages; i++) {
1230                 if (lnb[i].lnb_rc == -ENOSPC &&
1231                     osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent)) {
1232                         /* Allow the write to proceed if overwriting an
1233                          * existing block */
1234                         lnb[i].lnb_rc = 0;
1235                 }
1236
1237                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1238                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1239                                lnb[i].lnb_rc);
1240                         LASSERT(lnb[i].lnb_page);
1241                         generic_error_remove_page(inode->i_mapping,
1242                                                   lnb[i].lnb_page);
1243                         continue;
1244                 }
1245
1246                 LASSERT(PageLocked(lnb[i].lnb_page));
1247                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1248
1249                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1250                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1251
1252                 /*
1253                  * Since write and truncate are serialized by oo_sem, even
1254                  * partial-page truncate should not leave dirty pages in the
1255                  * page cache.
1256                  */
1257                 LASSERT(!PageDirty(lnb[i].lnb_page));
1258
1259                 SetPageUptodate(lnb[i].lnb_page);
1260
1261                 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1262         }
1263
1264         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1265
1266         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1267                 rc = -ENOSPC;
1268         } else if (iobuf->dr_npages > 0) {
1269                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1270                                                  iobuf->dr_npages,
1271                                                  iobuf->dr_blocks, 1);
1272         } else {
1273                 /* no pages to write, no transno is needed */
1274                 thandle->th_local = 1;
1275         }
1276
1277         if (likely(rc == 0)) {
1278                 spin_lock(&inode->i_lock);
1279                 if (isize > i_size_read(inode)) {
1280                         i_size_write(inode, isize);
1281                         LDISKFS_I(inode)->i_disksize = isize;
1282                         spin_unlock(&inode->i_lock);
1283                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1284                 } else {
1285                         spin_unlock(&inode->i_lock);
1286                 }
1287
1288                 rc = osd_do_bio(osd, inode, iobuf);
1289                 /* we don't do stats here as in read path because
1290                  * write is async: we'll do this in osd_put_bufs() */
1291         } else {
1292                 osd_fini_iobuf(osd, iobuf);
1293         }
1294
1295         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1296
1297         if (unlikely(rc != 0)) {
1298                 /* if write fails, we should drop pages from the cache */
1299                 for (i = 0; i < npages; i++) {
1300                         if (lnb[i].lnb_page == NULL)
1301                                 continue;
1302                         LASSERT(PageLocked(lnb[i].lnb_page));
1303                         generic_error_remove_page(inode->i_mapping,
1304                                                   lnb[i].lnb_page);
1305                 }
1306         }
1307
1308         RETURN(rc);
1309 }
1310
1311 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1312                          struct niobuf_local *lnb, int npages)
1313 {
1314         struct osd_thread_info *oti = osd_oti_get(env);
1315         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1316         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1317         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1318         struct timeval start, end;
1319         unsigned long timediff;
1320         int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1321         loff_t isize;
1322
1323         LASSERT(inode);
1324
1325         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1326         if (unlikely(rc != 0))
1327                 RETURN(rc);
1328
1329         isize = i_size_read(inode);
1330
1331         if (osd->od_read_cache)
1332                 cache = 1;
1333         if (isize > osd->od_readcache_max_filesize)
1334                 cache = 0;
1335
1336         do_gettimeofday(&start);
1337         for (i = 0; i < npages; i++) {
1338
1339                 if (isize <= lnb[i].lnb_file_offset)
1340                         /* If there's no more data, abort early.
1341                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1342                         break;
1343
1344                 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1345                         lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1346                 else
1347                         lnb[i].lnb_rc = lnb[i].lnb_len;
1348
1349                 if (PageUptodate(lnb[i].lnb_page)) {
1350                         cache_hits++;
1351                 } else {
1352                         cache_misses++;
1353                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1354                 }
1355
1356                 if (cache == 0)
1357                         generic_error_remove_page(inode->i_mapping,
1358                                                   lnb[i].lnb_page);
1359         }
1360         do_gettimeofday(&end);
1361         timediff = cfs_timeval_sub(&end, &start, NULL);
1362         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1363
1364         if (cache_hits != 0)
1365                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1366                                     cache_hits);
1367         if (cache_misses != 0)
1368                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1369                                     cache_misses);
1370         if (cache_hits + cache_misses != 0)
1371                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1372                                     cache_hits + cache_misses);
1373
1374         if (iobuf->dr_npages) {
1375                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1376                                                  iobuf->dr_npages,
1377                                                  iobuf->dr_blocks, 0);
1378                 rc = osd_do_bio(osd, inode, iobuf);
1379
1380                 /* IO stats will be done in osd_bufs_put() */
1381         }
1382
1383         RETURN(rc);
1384 }
1385
1386 /*
1387  * XXX: Another layering violation for now.
1388  *
1389  * We don't want to use ->f_op->read methods, because generic file write
1390  *
1391  *         - serializes on ->i_sem, and
1392  *
1393  *         - does a lot of extra work like balance_dirty_pages(),
1394  *
1395  * which doesn't work for globally shared files like /last_rcvd.
1396  */
1397 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1398 {
1399         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1400
1401         memcpy(buffer, (char *)ei->i_data, buflen);
1402
1403         return  buflen;
1404 }
1405
1406 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1407 {
1408         struct buffer_head *bh;
1409         unsigned long block;
1410         int osize;
1411         int blocksize;
1412         int csize;
1413         int boffs;
1414
1415         /* prevent reading after eof */
1416         spin_lock(&inode->i_lock);
1417         if (i_size_read(inode) < *offs + size) {
1418                 loff_t diff = i_size_read(inode) - *offs;
1419                 spin_unlock(&inode->i_lock);
1420                 if (diff < 0) {
1421                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1422                                i_size_read(inode), *offs);
1423                         return -EBADR;
1424                 } else if (diff == 0) {
1425                         return 0;
1426                 } else {
1427                         size = diff;
1428                 }
1429         } else {
1430                 spin_unlock(&inode->i_lock);
1431         }
1432
1433         blocksize = 1 << inode->i_blkbits;
1434         osize = size;
1435         while (size > 0) {
1436                 block = *offs >> inode->i_blkbits;
1437                 boffs = *offs & (blocksize - 1);
1438                 csize = min(blocksize - boffs, size);
1439                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1440                 if (IS_ERR(bh)) {
1441                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1442                                LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
1443                                csize, *offs, inode->i_ino,
1444                                PTR_ERR(bh));
1445                         return PTR_ERR(bh);
1446                 }
1447
1448                 if (bh != NULL) {
1449                         memcpy(buf, bh->b_data + boffs, csize);
1450                         brelse(bh);
1451                 } else {
1452                         memset(buf, 0, csize);
1453                 }
1454
1455                 *offs += csize;
1456                 buf += csize;
1457                 size -= csize;
1458         }
1459         return osize;
1460 }
1461
1462 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1463                         struct lu_buf *buf, loff_t *pos)
1464 {
1465         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1466         int           rc;
1467
1468         /* Read small symlink from inode body as we need to maintain correct
1469          * on-disk symlinks for ldiskfs.
1470          */
1471         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1472             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1473                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1474         else
1475                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1476
1477         return rc;
1478 }
1479
1480 static inline int osd_extents_enabled(struct super_block *sb,
1481                                       struct inode *inode)
1482 {
1483         if (inode != NULL) {
1484                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1485                         return 1;
1486         } else if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
1487                                 LDISKFS_FEATURE_INCOMPAT_EXTENTS)) {
1488                 return 1;
1489         }
1490         return 0;
1491 }
1492
1493 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1494                            const loff_t size, const loff_t pos,
1495                            const int blocks)
1496 {
1497         int credits, bits, bs, i;
1498
1499         bits = sb->s_blocksize_bits;
1500         bs = 1 << bits;
1501
1502         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1503          * we do not expect blockmaps on the large files,
1504          * so let's shrink it to 2 levels (4GB files) */
1505
1506         /* this is default reservation: 2 levels */
1507         credits = (blocks + 2) * 3;
1508
1509         /* actual offset is unknown, hard to optimize */
1510         if (pos == -1)
1511                 return credits;
1512
1513         /* now check for few specific cases to optimize */
1514         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1515                 /* no indirects */
1516                 credits = blocks;
1517                 /* allocate if not allocated */
1518                 if (inode == NULL) {
1519                         credits += blocks * 2;
1520                         return credits;
1521                 }
1522                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1523                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1524                         if (LDISKFS_I(inode)->i_data[i] == 0)
1525                                 credits += 2;
1526                 }
1527         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1528                 /* single indirect */
1529                 credits = blocks * 3;
1530                 if (inode == NULL ||
1531                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1532                         credits += 3;
1533                 else
1534                         /* The indirect block may be modified. */
1535                         credits += 1;
1536         }
1537
1538         return credits;
1539 }
1540
1541 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1542                                  const struct lu_buf *buf, loff_t _pos,
1543                                  struct thandle *handle)
1544 {
1545         struct osd_object  *obj  = osd_dt_obj(dt);
1546         struct inode       *inode = obj->oo_inode;
1547         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1548         struct osd_thandle *oh;
1549         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1550         int                 bits, bs;
1551         int                 depth, size;
1552         loff_t              pos;
1553         ENTRY;
1554
1555         LASSERT(buf != NULL);
1556         LASSERT(handle != NULL);
1557
1558         oh = container_of0(handle, struct osd_thandle, ot_super);
1559         LASSERT(oh->ot_handle == NULL);
1560
1561         size = buf->lb_len;
1562         bits = sb->s_blocksize_bits;
1563         bs = 1 << bits;
1564
1565         if (_pos == -1) {
1566                 /* if this is an append, then we
1567                  * should expect cross-block record */
1568                 pos = 0;
1569         } else {
1570                 pos = _pos;
1571         }
1572
1573         /* blocks to modify */
1574         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1575         LASSERT(blocks > 0);
1576
1577         if (inode != NULL && _pos != -1) {
1578                 /* object size in blocks */
1579                 est = (i_size_read(inode) + bs - 1) >> bits;
1580                 allocated = inode->i_blocks >> (bits - 9);
1581                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1582                         /* looks like an overwrite, no need to modify tree */
1583                         credits = blocks;
1584                         /* no need to modify i_size */
1585                         goto out;
1586                 }
1587         }
1588
1589         if (osd_extents_enabled(sb, inode)) {
1590                 /*
1591                  * many concurrent threads may grow tree by the time
1592                  * our transaction starts. so, consider 2 is a min depth
1593                  * for every level we may need to allocate a new block
1594                  * and take some entries from the old one. so, 3 blocks
1595                  * to allocate (bitmap, gd, itself) + old block - 4 per
1596                  * level.
1597                  */
1598                 depth = inode != NULL ? ext_depth(inode) : 0;
1599                 depth = max(depth, 1) + 1;
1600                 credits = depth;
1601                 /* if not append, then split may need to modify
1602                  * existing blocks moving entries into the new ones */
1603                 if (_pos == -1)
1604                         credits += depth;
1605                 /* blocks to store data: bitmap,gd,itself */
1606                 credits += blocks * 3;
1607         } else {
1608                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1609         }
1610         /* if inode is created as part of the transaction,
1611          * then it's counted already by the creation method */
1612         if (inode != NULL)
1613                 credits++;
1614
1615 out:
1616
1617         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1618
1619         /* dt_declare_write() is usually called for system objects, such
1620          * as llog or last_rcvd files. We needn't enforce quota on those
1621          * objects, so always set the lqi_space as 0. */
1622         if (inode != NULL)
1623                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1624                                            i_gid_read(inode), 0, oh, obj, true,
1625                                            NULL, false);
1626         RETURN(rc);
1627 }
1628
1629 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1630 {
1631         /* LU-2634: clear the extent format for fast symlink */
1632         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1633
1634         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1635         spin_lock(&inode->i_lock);
1636         LDISKFS_I(inode)->i_disksize = buflen;
1637         i_size_write(inode, buflen);
1638         spin_unlock(&inode->i_lock);
1639         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1640
1641         return 0;
1642 }
1643
1644 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1645                              int write_NUL, loff_t *offs, handle_t *handle)
1646 {
1647         struct buffer_head *bh        = NULL;
1648         loff_t              offset    = *offs;
1649         loff_t              new_size  = i_size_read(inode);
1650         unsigned long       block;
1651         int                 blocksize = 1 << inode->i_blkbits;
1652         int                 err = 0;
1653         int                 size;
1654         int                 boffs;
1655         int                 dirty_inode = 0;
1656
1657         if (write_NUL) {
1658                 /*
1659                  * long symlink write does not count the NUL terminator in
1660                  * bufsize, we write it, and the inode's file size does not
1661                  * count the NUL terminator as well.
1662                  */
1663                 ((char *)buf)[bufsize] = '\0';
1664                 ++bufsize;
1665         }
1666         while (bufsize > 0) {
1667                 if (bh != NULL)
1668                         brelse(bh);
1669
1670                 block = offset >> inode->i_blkbits;
1671                 boffs = offset & (blocksize - 1);
1672                 size = min(blocksize - boffs, bufsize);
1673                 bh = __ldiskfs_bread(handle, inode, block, 1);
1674                 if (IS_ERR_OR_NULL(bh)) {
1675                         if (bh == NULL) {
1676                                 err = -EIO;
1677                         } else {
1678                                 err = PTR_ERR(bh);
1679                                 bh = NULL;
1680                         }
1681                         CERROR("%s: error reading offset %llu (block %lu): "
1682                                "rc = %d\n",
1683                                inode->i_sb->s_id, offset, block, err);
1684                         break;
1685                 }
1686
1687                 err = ldiskfs_journal_get_write_access(handle, bh);
1688                 if (err) {
1689                         CERROR("journal_get_write_access() returned error %d\n",
1690                                err);
1691                         break;
1692                 }
1693                 LASSERTF(boffs + size <= bh->b_size,
1694                          "boffs %d size %d bh->b_size %lu\n",
1695                          boffs, size, (unsigned long)bh->b_size);
1696                 memcpy(bh->b_data + boffs, buf, size);
1697                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1698                 if (err)
1699                         break;
1700
1701                 if (offset + size > new_size)
1702                         new_size = offset + size;
1703                 offset += size;
1704                 bufsize -= size;
1705                 buf += size;
1706         }
1707         if (bh)
1708                 brelse(bh);
1709
1710         if (write_NUL)
1711                 --new_size;
1712         /* correct in-core and on-disk sizes */
1713         if (new_size > i_size_read(inode)) {
1714                 spin_lock(&inode->i_lock);
1715                 if (new_size > i_size_read(inode))
1716                         i_size_write(inode, new_size);
1717                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1718                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1719                         dirty_inode = 1;
1720                 }
1721                 spin_unlock(&inode->i_lock);
1722                 if (dirty_inode)
1723                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1724         }
1725
1726         if (err == 0)
1727                 *offs = offset;
1728         return err;
1729 }
1730
1731 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1732                          const struct lu_buf *buf, loff_t *pos,
1733                          struct thandle *handle, int ignore_quota)
1734 {
1735         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1736         struct osd_thandle      *oh;
1737         ssize_t                 result;
1738         int                     is_link;
1739
1740         LASSERT(dt_object_exists(dt));
1741
1742         LASSERT(handle != NULL);
1743         LASSERT(inode != NULL);
1744         ll_vfs_dq_init(inode);
1745
1746         /* XXX: don't check: one declared chunk can be used many times */
1747         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1748
1749         oh = container_of(handle, struct osd_thandle, ot_super);
1750         LASSERT(oh->ot_handle->h_transaction != NULL);
1751         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1752
1753         /* Write small symlink to inode body as we need to maintain correct
1754          * on-disk symlinks for ldiskfs.
1755          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1756          * does not count it in.
1757          */
1758         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1759         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1760                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1761         else
1762                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1763                                                   buf->lb_len, is_link, pos,
1764                                                   oh->ot_handle);
1765         if (result == 0)
1766                 result = buf->lb_len;
1767
1768         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1769
1770         return result;
1771 }
1772
1773 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1774                              __u64 start, __u64 end, struct thandle *th)
1775 {
1776         struct osd_thandle *oh;
1777         struct inode       *inode;
1778         int                 rc;
1779         ENTRY;
1780
1781         LASSERT(th);
1782         oh = container_of(th, struct osd_thandle, ot_super);
1783
1784         /*
1785          * we don't need to reserve credits for whole truncate
1786          * it's not possible as truncate may need to free too many
1787          * blocks and that won't fit a single transaction. instead
1788          * we reserve credits to change i_size and put inode onto
1789          * orphan list. if needed truncate will extend or restart
1790          * transaction
1791          */
1792         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1793                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1794
1795         inode = osd_dt_obj(dt)->oo_inode;
1796         LASSERT(inode);
1797
1798         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1799                                    0, oh, osd_dt_obj(dt), true, NULL, false);
1800         RETURN(rc);
1801 }
1802
1803 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1804                      __u64 start, __u64 end, struct thandle *th)
1805 {
1806         struct osd_thandle *oh;
1807         struct osd_object  *obj = osd_dt_obj(dt);
1808         struct inode       *inode = obj->oo_inode;
1809         handle_t           *h;
1810         tid_t               tid;
1811         int                rc = 0, rc2 = 0;
1812         ENTRY;
1813
1814         LASSERT(end == OBD_OBJECT_EOF);
1815         LASSERT(dt_object_exists(dt));
1816         LASSERT(osd_invariant(obj));
1817         LASSERT(inode != NULL);
1818         ll_vfs_dq_init(inode);
1819
1820         LASSERT(th);
1821         oh = container_of(th, struct osd_thandle, ot_super);
1822         LASSERT(oh->ot_handle->h_transaction != NULL);
1823
1824         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1825
1826         tid = oh->ot_handle->h_transaction->t_tid;
1827
1828         spin_lock(&inode->i_lock);
1829         i_size_write(inode, start);
1830         spin_unlock(&inode->i_lock);
1831         ll_truncate_pagecache(inode, start);
1832 #ifdef HAVE_INODEOPS_TRUNCATE
1833         if (inode->i_op->truncate) {
1834                 inode->i_op->truncate(inode);
1835         } else
1836 #endif
1837                 ldiskfs_truncate(inode);
1838
1839         /*
1840          * For a partial-page truncate, flush the page to disk immediately to
1841          * avoid data corruption during direct disk write.  b=17397
1842          */
1843         if ((start & ~PAGE_MASK) != 0)
1844                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1845
1846         h = journal_current_handle();
1847         LASSERT(h != NULL);
1848         LASSERT(h == oh->ot_handle);
1849
1850         /* do not check credits with osd_trans_exec_check() as the truncate
1851          * can restart the transaction internally and we restart the
1852          * transaction in this case */
1853
1854         if (tid != h->h_transaction->t_tid) {
1855                 int credits = oh->ot_credits;
1856                 /*
1857                  * transaction has changed during truncate
1858                  * we need to restart the handle with our credits
1859                  */
1860                 if (h->h_buffer_credits < credits) {
1861                         if (ldiskfs_journal_extend(h, credits))
1862                                 rc2 = ldiskfs_journal_restart(h, credits);
1863                 }
1864         }
1865
1866         RETURN(rc == 0 ? rc2 : rc);
1867 }
1868
1869 static int fiemap_check_ranges(struct inode *inode,
1870                                u64 start, u64 len, u64 *new_len)
1871 {
1872         loff_t maxbytes;
1873
1874         *new_len = len;
1875
1876         if (len == 0)
1877                 return -EINVAL;
1878
1879         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1880                 maxbytes = inode->i_sb->s_maxbytes;
1881         else
1882                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1883
1884         if (start > maxbytes)
1885                 return -EFBIG;
1886
1887         /*
1888          * Shrink request scope to what the fs can actually handle.
1889          */
1890         if (len > maxbytes || (maxbytes - len) < start)
1891                 *new_len = maxbytes - start;
1892
1893         return 0;
1894 }
1895
1896 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1897 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
1898
1899 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1900                           struct fiemap *fm)
1901 {
1902         struct fiemap_extent_info fieinfo = {0, };
1903         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1904         u64 len;
1905         int rc;
1906
1907
1908         LASSERT(inode);
1909         if (inode->i_op->fiemap == NULL)
1910                 return -EOPNOTSUPP;
1911
1912         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1913                 return -EINVAL;
1914
1915         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1916         if (rc)
1917                 return rc;
1918
1919         fieinfo.fi_flags = fm->fm_flags;
1920         fieinfo.fi_extents_max = fm->fm_extent_count;
1921         fieinfo.fi_extents_start = fm->fm_extents;
1922
1923         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1924                 filemap_write_and_wait(inode->i_mapping);
1925
1926         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
1927         fm->fm_flags = fieinfo.fi_flags;
1928         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
1929
1930         return rc;
1931 }
1932
1933 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1934                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1935 {
1936         int              rc = 0;
1937         struct inode    *inode = osd_dt_obj(dt)->oo_inode;
1938         ENTRY;
1939
1940         switch (advice) {
1941         case LU_LADVISE_DONTNEED:
1942                 if (end == 0)
1943                         break;
1944                 invalidate_mapping_pages(inode->i_mapping,
1945                                          start >> PAGE_CACHE_SHIFT,
1946                                          (end - 1) >> PAGE_CACHE_SHIFT);
1947                 break;
1948         default:
1949                 rc = -ENOTSUPP;
1950                 break;
1951         }
1952
1953         RETURN(rc);
1954 }
1955
1956 /*
1957  * in some cases we may need declare methods for objects being created
1958  * e.g., when we create symlink
1959  */
1960 const struct dt_body_operations osd_body_ops_new = {
1961         .dbo_declare_write = osd_declare_write,
1962 };
1963
1964 const struct dt_body_operations osd_body_ops = {
1965         .dbo_read                       = osd_read,
1966         .dbo_declare_write              = osd_declare_write,
1967         .dbo_write                      = osd_write,
1968         .dbo_bufs_get                   = osd_bufs_get,
1969         .dbo_bufs_put                   = osd_bufs_put,
1970         .dbo_write_prep                 = osd_write_prep,
1971         .dbo_declare_write_commit       = osd_declare_write_commit,
1972         .dbo_write_commit               = osd_write_commit,
1973         .dbo_read_prep                  = osd_read_prep,
1974         .dbo_declare_punch              = osd_declare_punch,
1975         .dbo_punch                      = osd_punch,
1976         .dbo_fiemap_get                 = osd_fiemap_get,
1977         .dbo_ladvise                    = osd_ladvise,
1978 };