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 /* LUSTRE_VERSION_CODE */
  42 #include <lustre_ver.h>
  43 /* prerequisite for linux/xattr.h */
  44 #include <linux/types.h>
  45 /* prerequisite for linux/xattr.h */
  46 #include <linux/fs.h>
  47
  48 /*
  49  * struct OBD_{ALLOC,FREE}*()
  50  * OBD_FAIL_CHECK
  51  */
  52 #include <obd_support.h>
  53
  54 #include "osd_internal.h"
  55
  56 /* ext_depth() */
  57 #include <ldiskfs/ldiskfs_extents.h>
  58
  59 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  60                             int rw, int line, int pages)
  61 {
  62         int blocks, i;
  63
  64         LASSERTF(iobuf->dr_elapsed_valid == 0,
  65                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  66                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  67                  iobuf->dr_init_at);
  68         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  69
  70         init_waitqueue_head(&iobuf->dr_wait);
  71         atomic_set(&iobuf->dr_numreqs, 0);
  72         iobuf->dr_npages = 0;
  73         iobuf->dr_error = 0;
  74         iobuf->dr_dev = d;
  75         iobuf->dr_frags = 0;
  76         iobuf->dr_elapsed = 0;
  77         /* must be counted before, so assert */
  78         iobuf->dr_rw = rw;
  79         iobuf->dr_init_at = line;
  80
  81         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  82         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  83                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  84                         pages * sizeof(iobuf->dr_pages[0]));
  85                 return 0;
  86         }
  87
  88         /* start with 1MB for 4K blocks */
  89         i = 256;
  90         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
  91                 i <<= 1;
  92
  93         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
  94                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
  95         pages = i;
  96         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  97         iobuf->dr_max_pages = 0;
  98         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
  99                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 100
 101         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 102         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 103         if (unlikely(iobuf->dr_blocks == NULL))
 104                 return -ENOMEM;
 105
 106         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 107         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 108         if (unlikely(iobuf->dr_pages == NULL))
 109                 return -ENOMEM;
 110
 111         iobuf->dr_max_pages = pages;
 112
 113         return 0;
 114 }
 115 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 116         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 117
 118 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
 119 {
 120         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 121         iobuf->dr_pages[iobuf->dr_npages++] = page;
 122 }
 123
 124 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 125 {
 126         int rw = iobuf->dr_rw;
 127
 128         if (iobuf->dr_elapsed_valid) {
 129                 iobuf->dr_elapsed_valid = 0;
 130                 LASSERT(iobuf->dr_dev == d);
 131                 LASSERT(iobuf->dr_frags > 0);
 132                 lprocfs_oh_tally(&d->od_brw_stats.
 133                                  hist[BRW_R_DIO_FRAGS+rw],
 134                                  iobuf->dr_frags);
 135                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 136                                       iobuf->dr_elapsed);
 137         }
 138 }
 139
 140 #ifndef REQ_WRITE /* pre-2.6.35 */
 141 #define __REQ_WRITE BIO_RW
 142 #endif
 143
 144 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 145 static void dio_complete_routine(struct bio *bio)
 146 {
 147         int error = bio->bi_error;
 148 #else
 149 static void dio_complete_routine(struct bio *bio, int error)
 150 {
 151 #endif
 152         struct osd_iobuf *iobuf = bio->bi_private;
 153         int iter;
 154         struct bio_vec *bvl;
 155
 156         /* CAVEAT EMPTOR: possibly in IRQ context
 157          * DO NOT record procfs stats here!!! */
 158
 159         if (unlikely(iobuf == NULL)) {
 160                 CERROR("***** bio->bi_private is NULL!  This should never "
 161                        "happen.  Normally, I would crash here, but instead I "
 162                        "will dump the bio contents to the console.  Please "
 163                        "report this to <https://jira.hpdd.intel.com/> , along "
 164                        "with any interesting messages leading up to this point "
 165                        "(like SCSI errors, perhaps).  Because bi_private is "
 166                        "NULL, I can't wake up the thread that initiated this "
 167                        "IO - you will probably have to reboot this node.\n");
 168                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 169                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 170                        "bi_private: %p\n", bio->bi_next,
 171                         (unsigned long)bio->bi_flags,
 172                         bio->bi_rw, bio->bi_vcnt, bio_idx(bio),
 173                         bio_sectors(bio) << 9, bio->bi_end_io,
 174 #ifdef HAVE_BI_CNT
 175                         atomic_read(&bio->bi_cnt),
 176 #else
 177                         atomic_read(&bio->__bi_cnt),
 178 #endif
 179                         bio->bi_private);
 180                 return;
 181         }
 182
 183         /* the check is outside of the cycle for performance reason -bzzz */
 184         if (!test_bit(__REQ_WRITE, &bio->bi_rw)) {
 185                 bio_for_each_segment_all(bvl, bio, iter) {
 186                         if (likely(error == 0))
 187                                 SetPageUptodate(bvl_to_page(bvl));
 188                         LASSERT(PageLocked(bvl_to_page(bvl)));
 189                 }
 190                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 191         } else {
 192                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 193         }
 194
 195         /* any real error is good enough -bzzz */
 196         if (error != 0 && iobuf->dr_error == 0)
 197                 iobuf->dr_error = error;
 198
 199         /*
 200          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 201          * it's possible that service thread will see dr_numreqs
 202          * is zero, but dr_elapsed is not set yet, leading to lost
 203          * data in this processing and an assertion in a subsequent
 204          * call to OSD.
 205          */
 206         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 207                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
 208                 iobuf->dr_elapsed_valid = 1;
 209         }
 210         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 211                 wake_up(&iobuf->dr_wait);
 212
 213         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 214          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 215          * mempool when serious on-disk fragmentation was encountered,
 216          * deadlocking the OST.  The bios are now released as soon as complete
 217          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 218         bio_put(bio);
 219 }
 220
 221 static void record_start_io(struct osd_iobuf *iobuf, int size)
 222 {
 223         struct osd_device    *osd = iobuf->dr_dev;
 224         struct obd_histogram *h = osd->od_brw_stats.hist;
 225
 226         iobuf->dr_frags++;
 227         atomic_inc(&iobuf->dr_numreqs);
 228
 229         if (iobuf->dr_rw == 0) {
 230                 atomic_inc(&osd->od_r_in_flight);
 231                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 232                                  atomic_read(&osd->od_r_in_flight));
 233                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 234         } else if (iobuf->dr_rw == 1) {
 235                 atomic_inc(&osd->od_w_in_flight);
 236                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 237                                  atomic_read(&osd->od_w_in_flight));
 238                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 239         } else {
 240                 LBUG();
 241         }
 242 }
 243
 244 static void osd_submit_bio(int rw, struct bio *bio)
 245 {
 246         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 247         if (rw == 0)
 248                 submit_bio(READ, bio);
 249         else
 250                 submit_bio(WRITE, bio);
 251 }
 252
 253 static int can_be_merged(struct bio *bio, sector_t sector)
 254 {
 255         if (bio == NULL)
 256                 return 0;
 257
 258         return bio_end_sector(bio) == sector ? 1 : 0;
 259 }
 260
 261 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 262                       struct osd_iobuf *iobuf)
 263 {
 264         int            blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 265         struct page  **pages = iobuf->dr_pages;
 266         int            npages = iobuf->dr_npages;
 267         sector_t      *blocks = iobuf->dr_blocks;
 268         int            total_blocks = npages * blocks_per_page;
 269         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 270         unsigned int   blocksize = inode->i_sb->s_blocksize;
 271         struct bio    *bio = NULL;
 272         struct page   *page;
 273         unsigned int   page_offset;
 274         sector_t       sector;
 275         int            nblocks;
 276         int            block_idx;
 277         int            page_idx;
 278         int            i;
 279         int            rc = 0;
 280         ENTRY;
 281
 282         LASSERT(iobuf->dr_npages == npages);
 283
 284         osd_brw_stats_update(osd, iobuf);
 285         iobuf->dr_start_time = cfs_time_current();
 286
 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         /* in order to achieve better IO throughput, we don't wait for writes
 374          * completion here. instead we proceed with transaction commit in
 375          * parallel and wait for IO completion once transaction is stopped
 376          * see osd_trans_stop() for more details -bzzz */
 377         if (iobuf->dr_rw == 0) {
 378                 wait_event(iobuf->dr_wait,
 379                            atomic_read(&iobuf->dr_numreqs) == 0);
 380                 osd_fini_iobuf(osd, iobuf);
 381         }
 382
 383         if (rc == 0)
 384                 rc = iobuf->dr_error;
 385         RETURN(rc);
 386 }
 387
 388 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 389                                    struct niobuf_local *lnb)
 390 {
 391         ENTRY;
 392
 393         *nrpages = 0;
 394
 395         while (len > 0) {
 396                 int poff = offset & (PAGE_SIZE - 1);
 397                 int plen = PAGE_SIZE - poff;
 398
 399                 if (plen > len)
 400                         plen = len;
 401                 lnb->lnb_file_offset = offset;
 402                 lnb->lnb_page_offset = poff;
 403                 lnb->lnb_len = plen;
 404                 /* lnb->lnb_flags = rnb->rnb_flags; */
 405                 lnb->lnb_flags = 0;
 406                 lnb->lnb_page = NULL;
 407                 lnb->lnb_rc = 0;
 408
 409                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 410                          (long long) len);
 411                 offset += plen;
 412                 len -= plen;
 413                 lnb++;
 414                 (*nrpages)++;
 415         }
 416
 417         RETURN(0);
 418 }
 419
 420 static struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
 421 {
 422         struct inode      *inode = osd_dt_obj(dt)->oo_inode;
 423         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 424         struct page       *page;
 425
 426         LASSERT(inode);
 427
 428         page = find_or_create_page(inode->i_mapping, offset >> PAGE_SHIFT,
 429                                    GFP_NOFS | __GFP_HIGHMEM);
 430         if (unlikely(page == NULL))
 431                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 432
 433         return page;
 434 }
 435
 436 /*
 437  * there are following "locks":
 438  * journal_start
 439  * i_mutex
 440  * page lock
 441  *
 442  * osd write path:
 443  *  - lock page(s)
 444  *  - journal_start
 445  *  - truncate_sem
 446  *
 447  * ext4 vmtruncate:
 448  *  - lock pages, unlock
 449  *  - journal_start
 450  *  - lock partial page
 451  *  - i_data_sem
 452  *
 453  */
 454
 455 /**
 456  * Unlock and release pages loaded by osd_bufs_get()
 457  *
 458  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 459  *
 460  * \param env           thread execution environment
 461  * \param dt            dt object undergoing IO (OSD object + methods)
 462  * \param lnb           array of pages undergoing IO
 463  * \param npages        number of pages in \a lnb
 464  *
 465  * \retval 0            always
 466  */
 467 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 468                         struct niobuf_local *lnb, int npages)
 469 {
 470         int i;
 471
 472         for (i = 0; i < npages; i++) {
 473                 if (lnb[i].lnb_page == NULL)
 474                         continue;
 475                 LASSERT(PageLocked(lnb[i].lnb_page));
 476                 unlock_page(lnb[i].lnb_page);
 477                 put_page(lnb[i].lnb_page);
 478                 dt_object_put(env, dt);
 479                 lnb[i].lnb_page = NULL;
 480         }
 481
 482         RETURN(0);
 483 }
 484
 485 /**
 486  * Load and lock pages undergoing IO
 487  *
 488  * Pages as described in the \a lnb array are fetched (from disk or cache)
 489  * and locked for IO by the caller.
 490  *
 491  * DLM locking protects us from write and truncate competing for same region,
 492  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 493  * It's possible the writeout on a such a page is in progress when we access
 494  * it. It's also possible that during this writeout we put new (partial) data
 495  * into the page, but won't be able to proceed in filter_commitrw_write().
 496  * Therefore, just wait for writeout completion as it should be rare enough.
 497  *
 498  * \param env           thread execution environment
 499  * \param dt            dt object undergoing IO (OSD object + methods)
 500  * \param pos           byte offset of IO start
 501  * \param len           number of bytes of IO
 502  * \param lnb           array of extents undergoing IO
 503  * \param rw            read or write operation?
 504  * \param capa          capabilities
 505  *
 506  * \retval pages        (zero or more) loaded successfully
 507  * \retval -ENOMEM      on memory/page allocation error
 508  */
 509 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 510                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 511                         int rw)
 512 {
 513         struct osd_object   *obj    = osd_dt_obj(dt);
 514         int npages, i, rc = 0;
 515
 516         LASSERT(obj->oo_inode);
 517
 518         osd_map_remote_to_local(pos, len, &npages, lnb);
 519
 520         for (i = 0; i < npages; i++, lnb++) {
 521                 lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset, rw);
 522                 if (lnb->lnb_page == NULL)
 523                         GOTO(cleanup, rc = -ENOMEM);
 524
 525                 wait_on_page_writeback(lnb->lnb_page);
 526                 BUG_ON(PageWriteback(lnb->lnb_page));
 527
 528                 lu_object_get(&dt->do_lu);
 529         }
 530
 531         RETURN(i);
 532
 533 cleanup:
 534         if (i > 0)
 535                 osd_bufs_put(env, dt, lnb - i, i);
 536         return rc;
 537 }
 538
 539 #ifndef HAVE_LDISKFS_MAP_BLOCKS
 540
 541 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 542 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 543 #endif
 544
 545 struct bpointers {
 546         sector_t *blocks;
 547         unsigned long start;
 548         int num;
 549         int init_num;
 550         int create;
 551 };
 552
 553 static long ldiskfs_ext_find_goal(struct inode *inode,
 554                                   struct ldiskfs_ext_path *path,
 555                                   unsigned long block, int *aflags)
 556 {
 557         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 558         unsigned long bg_start;
 559         unsigned long colour;
 560         int depth;
 561
 562         if (path) {
 563                 struct ldiskfs_extent *ex;
 564                 depth = path->p_depth;
 565
 566                 /* try to predict block placement */
 567                 if ((ex = path[depth].p_ext))
 568                         return ldiskfs_ext_pblock(ex) +
 569                                 (block - le32_to_cpu(ex->ee_block));
 570
 571                 /* it looks index is empty
 572                  * try to find starting from index itself */
 573                 if (path[depth].p_bh)
 574                         return path[depth].p_bh->b_blocknr;
 575         }
 576
 577         /* OK. use inode's group */
 578         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 579                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 580         colour = (current->pid % 16) *
 581                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 582         return bg_start + colour + block;
 583 }
 584
 585 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 586                                 struct ldiskfs_ext_path *path,
 587                                 unsigned long block, unsigned long *count,
 588                                 int *err)
 589 {
 590         struct ldiskfs_allocation_request ar;
 591         unsigned long pblock;
 592         int aflags;
 593
 594         /* find neighbour allocated blocks */
 595         ar.lleft = block;
 596         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 597         if (*err)
 598                 return 0;
 599         ar.lright = block;
 600         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 601         if (*err)
 602                 return 0;
 603
 604         /* allocate new block */
 605         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 606         ar.inode = inode;
 607         ar.logical = block;
 608         ar.len = *count;
 609         ar.flags = LDISKFS_MB_HINT_DATA;
 610         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 611         *count = ar.len;
 612         return pblock;
 613 }
 614
 615 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 616                                      struct ldiskfs_ext_path *path,
 617                                      struct ldiskfs_ext_cache *cex,
 618 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 619                                      struct ldiskfs_extent *ex,
 620 #endif
 621                                      void *cbdata)
 622 {
 623         struct bpointers *bp = cbdata;
 624         struct ldiskfs_extent nex;
 625         unsigned long pblock = 0;
 626         unsigned long tgen;
 627         int err, i;
 628         unsigned long count;
 629         handle_t *handle;
 630
 631 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 632         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 633 #else
 634         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 635 #endif
 636                 err = EXT_CONTINUE;
 637                 goto map;
 638         }
 639
 640         if (bp->create == 0) {
 641                 i = 0;
 642                 if (cex->ec_block < bp->start)
 643                         i = bp->start - cex->ec_block;
 644                 if (i >= cex->ec_len)
 645                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 646                                         i, cex->ec_len);
 647                 for (; i < cex->ec_len && bp->num; i++) {
 648                         *(bp->blocks) = 0;
 649                         bp->blocks++;
 650                         bp->num--;
 651                         bp->start++;
 652                 }
 653
 654                 return EXT_CONTINUE;
 655         }
 656
 657         tgen = LDISKFS_I(inode)->i_ext_generation;
 658         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
 659
 660         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
 661                                    count + LDISKFS_ALLOC_NEEDED + 1);
 662         if (IS_ERR(handle)) {
 663                 return PTR_ERR(handle);
 664         }
 665
 666         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
 667                 /* the tree has changed. so path can be invalid at moment */
 668                 ldiskfs_journal_stop(handle);
 669                 return EXT_REPEAT;
 670         }
 671
 672         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
 673          * protected by i_data_sem as whole. so we patch it to store
 674          * generation to path and now verify the tree hasn't changed */
 675         down_write((&LDISKFS_I(inode)->i_data_sem));
 676
 677         /* validate extent, make sure the extent tree does not changed */
 678         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
 679                 /* cex is invalid, try again */
 680                 up_write(&LDISKFS_I(inode)->i_data_sem);
 681                 ldiskfs_journal_stop(handle);
 682                 return EXT_REPEAT;
 683         }
 684
 685         count = cex->ec_len;
 686         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
 687         if (!pblock)
 688                 goto out;
 689         BUG_ON(count > cex->ec_len);
 690
 691         /* insert new extent */
 692         nex.ee_block = cpu_to_le32(cex->ec_block);
 693         ldiskfs_ext_store_pblock(&nex, pblock);
 694         nex.ee_len = cpu_to_le16(count);
 695         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
 696         if (err) {
 697                 /* free data blocks we just allocated */
 698                 /* not a good idea to call discard here directly,
 699                  * but otherwise we'd need to call it every free() */
 700                 ldiskfs_discard_preallocations(inode);
 701 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
 702                 ldiskfs_free_blocks(handle, inode, NULL,
 703                                     ldiskfs_ext_pblock(&nex),
 704                                     le16_to_cpu(nex.ee_len), 0);
 705 #else
 706                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
 707                                     le16_to_cpu(nex.ee_len), 0);
 708 #endif
 709                 goto out;
 710         }
 711
 712         /*
 713          * Putting len of the actual extent we just inserted,
 714          * we are asking ldiskfs_ext_walk_space() to continue
 715          * scaning after that block
 716          */
 717         cex->ec_len = le16_to_cpu(nex.ee_len);
 718         cex->ec_start = ldiskfs_ext_pblock(&nex);
 719         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
 720         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
 721
 722 out:
 723         up_write((&LDISKFS_I(inode)->i_data_sem));
 724         ldiskfs_journal_stop(handle);
 725 map:
 726         if (err >= 0) {
 727                 /* map blocks */
 728                 if (bp->num == 0) {
 729                         CERROR("hmm. why do we find this extent?\n");
 730                         CERROR("initial space: %lu:%u\n",
 731                                 bp->start, bp->init_num);
 732 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 733                         CERROR("current extent: %u/%u/%llu %d\n",
 734                                 cex->ec_block, cex->ec_len,
 735                                 (unsigned long long)cex->ec_start,
 736                                 cex->ec_type);
 737 #else
 738                         CERROR("current extent: %u/%u/%llu\n",
 739                                 cex->ec_block, cex->ec_len,
 740                                 (unsigned long long)cex->ec_start);
 741 #endif
 742                 }
 743                 i = 0;
 744                 if (cex->ec_block < bp->start)
 745                         i = bp->start - cex->ec_block;
 746                 if (i >= cex->ec_len)
 747                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 748                                         i, cex->ec_len);
 749                 for (; i < cex->ec_len && bp->num; i++) {
 750                         *(bp->blocks) = cex->ec_start + i;
 751                         if (pblock != 0) {
 752                                 /* unmap any possible underlying metadata from
 753                                  * the block device mapping.  bug 6998. */
 754                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
 755                                                           *(bp->blocks));
 756                         }
 757                         bp->blocks++;
 758                         bp->num--;
 759                         bp->start++;
 760                 }
 761         }
 762         return err;
 763 }
 764
 765 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
 766                                    int clen, sector_t *blocks, int create)
 767 {
 768         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 769         struct bpointers bp;
 770         int err;
 771
 772         if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
 773                 return -EFBIG;
 774
 775         bp.blocks = blocks;
 776         bp.start = index * blocks_per_page;
 777         bp.init_num = bp.num = clen * blocks_per_page;
 778         bp.create = create;
 779
 780         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 781                bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
 782
 783         err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
 784                                      ldiskfs_ext_new_extent_cb, &bp);
 785         ldiskfs_ext_invalidate_cache(inode);
 786
 787         return err;
 788 }
 789
 790 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
 791                                           struct page **page, int pages,
 792                                           sector_t *blocks, int create)
 793 {
 794         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 795         pgoff_t bitmap_max_page_index;
 796         sector_t *b;
 797         int rc = 0, i;
 798
 799         bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
 800                                 PAGE_SHIFT;
 801         for (i = 0, b = blocks; i < pages; i++, page++) {
 802                 if ((*page)->index + 1 >= bitmap_max_page_index) {
 803                         rc = -EFBIG;
 804                         break;
 805                 }
 806                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
 807                 if (rc) {
 808                         CERROR("ino %lu, blk %llu create %d: rc %d\n",
 809                                inode->i_ino,
 810                                (unsigned long long)*b, create, rc);
 811                         break;
 812                 }
 813                 b += blocks_per_page;
 814         }
 815         return rc;
 816 }
 817
 818 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
 819                                            struct page **page,
 820                                            int pages, sector_t *blocks,
 821                                            int create)
 822 {
 823         int rc = 0, i = 0, clen = 0;
 824         struct page *fp = NULL;
 825
 826         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 827                 inode->i_ino, pages, (*page)->index);
 828
 829         /* pages are sorted already. so, we just have to find
 830          * contig. space and process them properly */
 831         while (i < pages) {
 832                 if (fp == NULL) {
 833                         /* start new extent */
 834                         fp = *page++;
 835                         clen = 1;
 836                         i++;
 837                         continue;
 838                 } else if (fp->index + clen == (*page)->index) {
 839                         /* continue the extent */
 840                         page++;
 841                         clen++;
 842                         i++;
 843                         continue;
 844                 }
 845
 846                 /* process found extent */
 847                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 848                                              blocks, create);
 849                 if (rc)
 850                         GOTO(cleanup, rc);
 851
 852                 /* look for next extent */
 853                 fp = NULL;
 854                 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
 855         }
 856
 857         if (fp)
 858                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 859                                              blocks, create);
 860
 861 cleanup:
 862         return rc;
 863 }
 864
 865 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 866                                        int pages, sector_t *blocks,
 867                                        int create)
 868 {
 869         int rc;
 870
 871         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 872                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
 873                                                      blocks, create);
 874                 return rc;
 875         }
 876         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
 877
 878         return rc;
 879 }
 880 #else
 881 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 882                                        int pages, sector_t *blocks,
 883                                        int create)
 884 {
 885         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 886         int rc = 0, i = 0;
 887         struct page *fp = NULL;
 888         int clen = 0;
 889         pgoff_t max_page_index;
 890         handle_t *handle = NULL;
 891
 892         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
 893
 894         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 895                 inode->i_ino, pages, (*page)->index);
 896
 897         if (create) {
 898                 create = LDISKFS_GET_BLOCKS_CREATE;
 899                 handle = ldiskfs_journal_current_handle();
 900                 LASSERT(handle != NULL);
 901                 rc = osd_attach_jinode(inode);
 902                 if (rc)
 903                         return rc;
 904         }
 905         /* pages are sorted already. so, we just have to find
 906          * contig. space and process them properly */
 907         while (i < pages) {
 908                 long blen, total = 0;
 909                 struct ldiskfs_map_blocks map = { 0 };
 910
 911                 if (fp == NULL) { /* start new extent */
 912                         fp = *page++;
 913                         clen = 1;
 914                         if (++i != pages)
 915                                 continue;
 916                 } else if (fp->index + clen == (*page)->index) {
 917                         /* continue the extent */
 918                         page++;
 919                         clen++;
 920                         if (++i != pages)
 921                                 continue;
 922                 }
 923                 if (fp->index + clen >= max_page_index)
 924                         GOTO(cleanup, rc = -EFBIG);
 925                 /* process found extent */
 926                 map.m_lblk = fp->index * blocks_per_page;
 927                 map.m_len = blen = clen * blocks_per_page;
 928 cont_map:
 929                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 930                 if (rc >= 0) {
 931                         int c = 0;
 932                         for (; total < blen && c < map.m_len; c++, total++) {
 933                                 if (rc == 0) {
 934                                         *(blocks + total) = 0;
 935                                         total++;
 936                                         break;
 937                                 } else {
 938                                         *(blocks + total) = map.m_pblk + c;
 939                                         /* unmap any possible underlying
 940                                          * metadata from the block device
 941                                          * mapping.  bug 6998. */
 942                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
 943                                             create)
 944                                                 unmap_underlying_metadata(
 945                                                         inode->i_sb->s_bdev,
 946                                                         map.m_pblk + c);
 947                                 }
 948                         }
 949                         rc = 0;
 950                 }
 951                 if (rc == 0 && total < blen) {
 952                         map.m_lblk = fp->index * blocks_per_page + total;
 953                         map.m_len = blen - total;
 954                         goto cont_map;
 955                 }
 956                 if (rc != 0)
 957                         GOTO(cleanup, rc);
 958
 959                 /* look for next extent */
 960                 fp = NULL;
 961                 blocks += blocks_per_page * clen;
 962         }
 963 cleanup:
 964         return rc;
 965 }
 966 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
 967
 968 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 969                           struct niobuf_local *lnb, int npages)
 970 {
 971         struct osd_thread_info *oti   = osd_oti_get(env);
 972         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 973         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 974         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 975         ktime_t start;
 976         ktime_t end;
 977         s64 timediff;
 978         ssize_t                 isize;
 979         __s64                   maxidx;
 980         int                     rc = 0;
 981         int                     i;
 982         int                     cache = 0;
 983
 984         LASSERT(inode);
 985
 986         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 987         if (unlikely(rc != 0))
 988                 RETURN(rc);
 989
 990         isize = i_size_read(inode);
 991         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
 992
 993         if (osd->od_writethrough_cache)
 994                 cache = 1;
 995         if (isize > osd->od_readcache_max_filesize)
 996                 cache = 0;
 997
 998         start = ktime_get();
 999         for (i = 0; i < npages; i++) {
1000
1001                 if (cache == 0)
1002                         generic_error_remove_page(inode->i_mapping,
1003                                                   lnb[i].lnb_page);
1004
1005                 /*
1006                  * till commit the content of the page is undefined
1007                  * we'll set it uptodate once bulk is done. otherwise
1008                  * subsequent reads can access non-stable data
1009                  */
1010                 ClearPageUptodate(lnb[i].lnb_page);
1011
1012                 if (lnb[i].lnb_len == PAGE_SIZE)
1013                         continue;
1014
1015                 if (maxidx >= lnb[i].lnb_page->index) {
1016                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1017                 } else {
1018                         long off;
1019                         char *p = kmap(lnb[i].lnb_page);
1020
1021                         off = lnb[i].lnb_page_offset;
1022                         if (off)
1023                                 memset(p, 0, off);
1024                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1025                               ~PAGE_MASK;
1026                         if (off)
1027                                 memset(p + off, 0, PAGE_SIZE - off);
1028                         kunmap(lnb[i].lnb_page);
1029                 }
1030         }
1031         end = ktime_get();
1032         timediff = ktime_us_delta(end, start);
1033         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1034
1035         if (iobuf->dr_npages) {
1036                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1037                                                  iobuf->dr_npages,
1038                                                  iobuf->dr_blocks, 0);
1039                 if (likely(rc == 0)) {
1040                         rc = osd_do_bio(osd, inode, iobuf);
1041                         /* do IO stats for preparation reads */
1042                         osd_fini_iobuf(osd, iobuf);
1043                 }
1044         }
1045         RETURN(rc);
1046 }
1047
1048 struct osd_fextent {
1049         sector_t        start;
1050         sector_t        end;
1051         unsigned int    mapped:1;
1052 };
1053
1054 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1055                          struct osd_fextent *cached_extent)
1056 {
1057         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1058         sector_t block = offset >> inode->i_blkbits;
1059         sector_t start;
1060         struct fiemap_extent_info fei = { 0 };
1061         struct fiemap_extent fe = { 0 };
1062         mm_segment_t saved_fs;
1063         int rc;
1064
1065         if (block >= cached_extent->start && block < cached_extent->end)
1066                 return cached_extent->mapped;
1067
1068         if (i_size_read(inode) == 0)
1069                 return 0;
1070
1071         /* Beyond EOF, must not be mapped */
1072         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1073                 return 0;
1074
1075         fei.fi_extents_max = 1;
1076         fei.fi_extents_start = &fe;
1077
1078         saved_fs = get_fs();
1079         set_fs(get_ds());
1080         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1081         set_fs(saved_fs);
1082         if (rc != 0)
1083                 return 0;
1084
1085         start = fe.fe_logical >> inode->i_blkbits;
1086
1087         if (start > block) {
1088                 cached_extent->start = block;
1089                 cached_extent->end = start;
1090                 cached_extent->mapped = 0;
1091         } else {
1092                 cached_extent->start = start;
1093                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1094                                       inode->i_blkbits;
1095                 cached_extent->mapped = 1;
1096         }
1097
1098         return cached_extent->mapped;
1099 }
1100
1101 static int osd_declare_write_commit(const struct lu_env *env,
1102                                     struct dt_object *dt,
1103                                     struct niobuf_local *lnb, int npages,
1104                                     struct thandle *handle)
1105 {
1106         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1107         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1108         struct osd_thandle      *oh;
1109         int                      extents = 1;
1110         int                      depth;
1111         int                      i;
1112         int                      newblocks;
1113         int                      rc = 0;
1114         int                      flags = 0;
1115         int                      credits = 0;
1116         bool                     ignore_quota = false;
1117         long long                quota_space = 0;
1118         struct osd_fextent       extent = { 0 };
1119         ENTRY;
1120
1121         LASSERT(handle != NULL);
1122         oh = container_of0(handle, struct osd_thandle, ot_super);
1123         LASSERT(oh->ot_handle == NULL);
1124
1125         newblocks = npages;
1126
1127         /* calculate number of extents (probably better to pass nb) */
1128         for (i = 0; i < npages; i++) {
1129                 if (i && lnb[i].lnb_file_offset !=
1130                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1131                         extents++;
1132
1133                 if (!osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1134                         quota_space += PAGE_SIZE;
1135
1136                 /* ignore quota for the whole request if any page is from
1137                  * client cache or written by root.
1138                  *
1139                  * XXX once we drop the 1.8 client support, the checking
1140                  * for whether page is from cache can be simplified as:
1141                  * !(lnb[i].flags & OBD_BRW_SYNC)
1142                  *
1143                  * XXX we could handle this on per-lnb basis as done by
1144                  * grant. */
1145                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1146                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1147                     OBD_BRW_FROM_GRANT)
1148                         ignore_quota = true;
1149         }
1150
1151         /*
1152          * each extent can go into new leaf causing a split
1153          * 5 is max tree depth: inode + 4 index blocks
1154          * with blockmaps, depth is 3 at most
1155          */
1156         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1157                 /*
1158                  * many concurrent threads may grow tree by the time
1159                  * our transaction starts. so, consider 2 is a min depth
1160                  */
1161                 depth = ext_depth(inode);
1162                 depth = max(depth, 1) + 1;
1163                 newblocks += depth;
1164                 credits++; /* inode */
1165                 credits += depth * 2 * extents;
1166         } else {
1167                 depth = 3;
1168                 newblocks += depth;
1169                 credits++; /* inode */
1170                 credits += depth * extents;
1171         }
1172
1173         /* quota space for metadata blocks */
1174         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1175
1176         /* quota space should be reported in 1K blocks */
1177         quota_space = toqb(quota_space);
1178
1179         /* each new block can go in different group (bitmap + gd) */
1180
1181         /* we can't dirty more bitmap blocks than exist */
1182         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1183                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1184         else
1185                 credits += newblocks;
1186
1187         /* we can't dirty more gd blocks than exist */
1188         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1189                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1190         else
1191                 credits += newblocks;
1192
1193         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1194
1195         /* make sure the over quota flags were not set */
1196         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1197
1198         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1199                                    quota_space, oh, osd_dt_obj(dt), true,
1200                                    &flags, ignore_quota);
1201
1202         /* we need only to store the overquota flags in the first lnb for
1203          * now, once we support multiple objects BRW, this code needs be
1204          * revised. */
1205         if (flags & QUOTA_FL_OVER_USRQUOTA)
1206                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1207         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1208                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1209
1210         RETURN(rc);
1211 }
1212
1213 /* Check if a block is allocated or not */
1214 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1215                             struct niobuf_local *lnb, int npages,
1216                             struct thandle *thandle)
1217 {
1218         struct osd_thread_info *oti = osd_oti_get(env);
1219         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1220         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1221         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1222         loff_t isize;
1223         int rc = 0, i;
1224         struct osd_fextent extent = { 0 };
1225
1226         LASSERT(inode);
1227
1228         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1229         if (unlikely(rc != 0))
1230                 RETURN(rc);
1231
1232         isize = i_size_read(inode);
1233         ll_vfs_dq_init(inode);
1234
1235         for (i = 0; i < npages; i++) {
1236                 if (lnb[i].lnb_rc == -ENOSPC &&
1237                     osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent)) {
1238                         /* Allow the write to proceed if overwriting an
1239                          * existing block */
1240                         lnb[i].lnb_rc = 0;
1241                 }
1242
1243                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1244                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1245                                lnb[i].lnb_rc);
1246                         LASSERT(lnb[i].lnb_page);
1247                         generic_error_remove_page(inode->i_mapping,
1248                                                   lnb[i].lnb_page);
1249                         continue;
1250                 }
1251
1252                 LASSERT(PageLocked(lnb[i].lnb_page));
1253                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1254
1255                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1256                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1257
1258                 /*
1259                  * Since write and truncate are serialized by oo_sem, even
1260                  * partial-page truncate should not leave dirty pages in the
1261                  * page cache.
1262                  */
1263                 LASSERT(!PageDirty(lnb[i].lnb_page));
1264
1265                 SetPageUptodate(lnb[i].lnb_page);
1266
1267                 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1268         }
1269
1270         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1271
1272         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1273                 rc = -ENOSPC;
1274         } else if (iobuf->dr_npages > 0) {
1275                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1276                                                  iobuf->dr_npages,
1277                                                  iobuf->dr_blocks, 1);
1278         } else {
1279                 /* no pages to write, no transno is needed */
1280                 thandle->th_local = 1;
1281         }
1282
1283         if (likely(rc == 0)) {
1284                 spin_lock(&inode->i_lock);
1285                 if (isize > i_size_read(inode)) {
1286                         i_size_write(inode, isize);
1287                         LDISKFS_I(inode)->i_disksize = isize;
1288                         spin_unlock(&inode->i_lock);
1289                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1290                 } else {
1291                         spin_unlock(&inode->i_lock);
1292                 }
1293
1294                 rc = osd_do_bio(osd, inode, iobuf);
1295                 /* we don't do stats here as in read path because
1296                  * write is async: we'll do this in osd_put_bufs() */
1297         } else {
1298                 osd_fini_iobuf(osd, iobuf);
1299         }
1300
1301         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1302
1303         if (unlikely(rc != 0)) {
1304                 /* if write fails, we should drop pages from the cache */
1305                 for (i = 0; i < npages; i++) {
1306                         if (lnb[i].lnb_page == NULL)
1307                                 continue;
1308                         LASSERT(PageLocked(lnb[i].lnb_page));
1309                         generic_error_remove_page(inode->i_mapping,
1310                                                   lnb[i].lnb_page);
1311                 }
1312         }
1313
1314         RETURN(rc);
1315 }
1316
1317 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1318                          struct niobuf_local *lnb, int npages)
1319 {
1320         struct osd_thread_info *oti = osd_oti_get(env);
1321         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1322         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1323         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1324         int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1325         ktime_t start, end;
1326         s64 timediff;
1327         loff_t isize;
1328
1329         LASSERT(inode);
1330
1331         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1332         if (unlikely(rc != 0))
1333                 RETURN(rc);
1334
1335         isize = i_size_read(inode);
1336
1337         if (osd->od_read_cache)
1338                 cache = 1;
1339         if (isize > osd->od_readcache_max_filesize)
1340                 cache = 0;
1341
1342         start = ktime_get();
1343         for (i = 0; i < npages; i++) {
1344
1345                 if (isize <= lnb[i].lnb_file_offset)
1346                         /* If there's no more data, abort early.
1347                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1348                         break;
1349
1350                 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1351                         lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1352                 else
1353                         lnb[i].lnb_rc = lnb[i].lnb_len;
1354
1355                 /* Bypass disk read if fail_loc is set properly */
1356                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1357                         SetPageUptodate(lnb[i].lnb_page);
1358
1359                 if (PageUptodate(lnb[i].lnb_page)) {
1360                         cache_hits++;
1361                 } else {
1362                         cache_misses++;
1363                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1364                 }
1365
1366                 if (cache == 0)
1367                         generic_error_remove_page(inode->i_mapping,
1368                                                   lnb[i].lnb_page);
1369         }
1370         end = ktime_get();
1371         timediff = ktime_us_delta(end, start);
1372         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1373
1374         if (cache_hits != 0)
1375                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1376                                     cache_hits);
1377         if (cache_misses != 0)
1378                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1379                                     cache_misses);
1380         if (cache_hits + cache_misses != 0)
1381                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1382                                     cache_hits + cache_misses);
1383
1384         if (iobuf->dr_npages) {
1385                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1386                                                  iobuf->dr_npages,
1387                                                  iobuf->dr_blocks, 0);
1388                 rc = osd_do_bio(osd, inode, iobuf);
1389
1390                 /* IO stats will be done in osd_bufs_put() */
1391         }
1392
1393         RETURN(rc);
1394 }
1395
1396 /*
1397  * XXX: Another layering violation for now.
1398  *
1399  * We don't want to use ->f_op->read methods, because generic file write
1400  *
1401  *         - serializes on ->i_sem, and
1402  *
1403  *         - does a lot of extra work like balance_dirty_pages(),
1404  *
1405  * which doesn't work for globally shared files like /last_rcvd.
1406  */
1407 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1408 {
1409         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1410
1411         memcpy(buffer, (char *)ei->i_data, buflen);
1412
1413         return  buflen;
1414 }
1415
1416 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1417 {
1418         struct buffer_head *bh;
1419         unsigned long block;
1420         int osize;
1421         int blocksize;
1422         int csize;
1423         int boffs;
1424
1425         /* prevent reading after eof */
1426         spin_lock(&inode->i_lock);
1427         if (i_size_read(inode) < *offs + size) {
1428                 loff_t diff = i_size_read(inode) - *offs;
1429                 spin_unlock(&inode->i_lock);
1430                 if (diff < 0) {
1431                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1432                                i_size_read(inode), *offs);
1433                         return -EBADR;
1434                 } else if (diff == 0) {
1435                         return 0;
1436                 } else {
1437                         size = diff;
1438                 }
1439         } else {
1440                 spin_unlock(&inode->i_lock);
1441         }
1442
1443         blocksize = 1 << inode->i_blkbits;
1444         osize = size;
1445         while (size > 0) {
1446                 block = *offs >> inode->i_blkbits;
1447                 boffs = *offs & (blocksize - 1);
1448                 csize = min(blocksize - boffs, size);
1449                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1450                 if (IS_ERR(bh)) {
1451                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1452                                LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
1453                                csize, *offs, inode->i_ino,
1454                                PTR_ERR(bh));
1455                         return PTR_ERR(bh);
1456                 }
1457
1458                 if (bh != NULL) {
1459                         memcpy(buf, bh->b_data + boffs, csize);
1460                         brelse(bh);
1461                 } else {
1462                         memset(buf, 0, csize);
1463                 }
1464
1465                 *offs += csize;
1466                 buf += csize;
1467                 size -= csize;
1468         }
1469         return osize;
1470 }
1471
1472 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1473                         struct lu_buf *buf, loff_t *pos)
1474 {
1475         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1476         int           rc;
1477
1478         /* Read small symlink from inode body as we need to maintain correct
1479          * on-disk symlinks for ldiskfs.
1480          */
1481         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1482             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1483                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1484         else
1485                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1486
1487         return rc;
1488 }
1489
1490 static inline int osd_extents_enabled(struct super_block *sb,
1491                                       struct inode *inode)
1492 {
1493         if (inode != NULL) {
1494                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1495                         return 1;
1496         } else if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
1497                                 LDISKFS_FEATURE_INCOMPAT_EXTENTS)) {
1498                 return 1;
1499         }
1500         return 0;
1501 }
1502
1503 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1504                            const loff_t size, const loff_t pos,
1505                            const int blocks)
1506 {
1507         int credits, bits, bs, i;
1508
1509         bits = sb->s_blocksize_bits;
1510         bs = 1 << bits;
1511
1512         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1513          * we do not expect blockmaps on the large files,
1514          * so let's shrink it to 2 levels (4GB files) */
1515
1516         /* this is default reservation: 2 levels */
1517         credits = (blocks + 2) * 3;
1518
1519         /* actual offset is unknown, hard to optimize */
1520         if (pos == -1)
1521                 return credits;
1522
1523         /* now check for few specific cases to optimize */
1524         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1525                 /* no indirects */
1526                 credits = blocks;
1527                 /* allocate if not allocated */
1528                 if (inode == NULL) {
1529                         credits += blocks * 2;
1530                         return credits;
1531                 }
1532                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1533                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1534                         if (LDISKFS_I(inode)->i_data[i] == 0)
1535                                 credits += 2;
1536                 }
1537         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1538                 /* single indirect */
1539                 credits = blocks * 3;
1540                 if (inode == NULL ||
1541                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1542                         credits += 3;
1543                 else
1544                         /* The indirect block may be modified. */
1545                         credits += 1;
1546         }
1547
1548         return credits;
1549 }
1550
1551 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1552                                  const struct lu_buf *buf, loff_t _pos,
1553                                  struct thandle *handle)
1554 {
1555         struct osd_object  *obj  = osd_dt_obj(dt);
1556         struct inode       *inode = obj->oo_inode;
1557         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1558         struct osd_thandle *oh;
1559         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1560         int                 bits, bs;
1561         int                 depth, size;
1562         loff_t              pos;
1563         ENTRY;
1564
1565         LASSERT(buf != NULL);
1566         LASSERT(handle != NULL);
1567
1568         oh = container_of0(handle, struct osd_thandle, ot_super);
1569         LASSERT(oh->ot_handle == NULL);
1570
1571         size = buf->lb_len;
1572         bits = sb->s_blocksize_bits;
1573         bs = 1 << bits;
1574
1575         if (_pos == -1) {
1576                 /* if this is an append, then we
1577                  * should expect cross-block record */
1578                 pos = 0;
1579         } else {
1580                 pos = _pos;
1581         }
1582
1583         /* blocks to modify */
1584         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1585         LASSERT(blocks > 0);
1586
1587         if (inode != NULL && _pos != -1) {
1588                 /* object size in blocks */
1589                 est = (i_size_read(inode) + bs - 1) >> bits;
1590                 allocated = inode->i_blocks >> (bits - 9);
1591                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1592                         /* looks like an overwrite, no need to modify tree */
1593                         credits = blocks;
1594                         /* no need to modify i_size */
1595                         goto out;
1596                 }
1597         }
1598
1599         if (osd_extents_enabled(sb, inode)) {
1600                 /*
1601                  * many concurrent threads may grow tree by the time
1602                  * our transaction starts. so, consider 2 is a min depth
1603                  * for every level we may need to allocate a new block
1604                  * and take some entries from the old one. so, 3 blocks
1605                  * to allocate (bitmap, gd, itself) + old block - 4 per
1606                  * level.
1607                  */
1608                 depth = inode != NULL ? ext_depth(inode) : 0;
1609                 depth = max(depth, 1) + 1;
1610                 credits = depth;
1611                 /* if not append, then split may need to modify
1612                  * existing blocks moving entries into the new ones */
1613                 if (_pos == -1)
1614                         credits += depth;
1615                 /* blocks to store data: bitmap,gd,itself */
1616                 credits += blocks * 3;
1617         } else {
1618                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1619         }
1620         /* if inode is created as part of the transaction,
1621          * then it's counted already by the creation method */
1622         if (inode != NULL)
1623                 credits++;
1624
1625 out:
1626
1627         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1628
1629         /* dt_declare_write() is usually called for system objects, such
1630          * as llog or last_rcvd files. We needn't enforce quota on those
1631          * objects, so always set the lqi_space as 0. */
1632         if (inode != NULL)
1633                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1634                                            i_gid_read(inode), 0, oh, obj, true,
1635                                            NULL, false);
1636         RETURN(rc);
1637 }
1638
1639 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1640 {
1641         /* LU-2634: clear the extent format for fast symlink */
1642         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1643
1644         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1645         spin_lock(&inode->i_lock);
1646         LDISKFS_I(inode)->i_disksize = buflen;
1647         i_size_write(inode, buflen);
1648         spin_unlock(&inode->i_lock);
1649         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1650
1651         return 0;
1652 }
1653
1654 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1655                              int write_NUL, loff_t *offs, handle_t *handle)
1656 {
1657         struct buffer_head *bh        = NULL;
1658         loff_t              offset    = *offs;
1659         loff_t              new_size  = i_size_read(inode);
1660         unsigned long       block;
1661         int                 blocksize = 1 << inode->i_blkbits;
1662         int                 err = 0;
1663         int                 size;
1664         int                 boffs;
1665         int                 dirty_inode = 0;
1666
1667         if (write_NUL) {
1668                 /*
1669                  * long symlink write does not count the NUL terminator in
1670                  * bufsize, we write it, and the inode's file size does not
1671                  * count the NUL terminator as well.
1672                  */
1673                 ((char *)buf)[bufsize] = '\0';
1674                 ++bufsize;
1675         }
1676         while (bufsize > 0) {
1677                 if (bh != NULL)
1678                         brelse(bh);
1679
1680                 block = offset >> inode->i_blkbits;
1681                 boffs = offset & (blocksize - 1);
1682                 size = min(blocksize - boffs, bufsize);
1683                 bh = __ldiskfs_bread(handle, inode, block, 1);
1684                 if (IS_ERR_OR_NULL(bh)) {
1685                         if (bh == NULL) {
1686                                 err = -EIO;
1687                         } else {
1688                                 err = PTR_ERR(bh);
1689                                 bh = NULL;
1690                         }
1691                         CERROR("%s: error reading offset %llu (block %lu): "
1692                                "rc = %d\n",
1693                                inode->i_sb->s_id, offset, block, err);
1694                         break;
1695                 }
1696
1697                 err = ldiskfs_journal_get_write_access(handle, bh);
1698                 if (err) {
1699                         CERROR("journal_get_write_access() returned error %d\n",
1700                                err);
1701                         break;
1702                 }
1703                 LASSERTF(boffs + size <= bh->b_size,
1704                          "boffs %d size %d bh->b_size %lu\n",
1705                          boffs, size, (unsigned long)bh->b_size);
1706                 memcpy(bh->b_data + boffs, buf, size);
1707                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1708                 if (err)
1709                         break;
1710
1711                 if (offset + size > new_size)
1712                         new_size = offset + size;
1713                 offset += size;
1714                 bufsize -= size;
1715                 buf += size;
1716         }
1717         if (bh)
1718                 brelse(bh);
1719
1720         if (write_NUL)
1721                 --new_size;
1722         /* correct in-core and on-disk sizes */
1723         if (new_size > i_size_read(inode)) {
1724                 spin_lock(&inode->i_lock);
1725                 if (new_size > i_size_read(inode))
1726                         i_size_write(inode, new_size);
1727                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1728                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1729                         dirty_inode = 1;
1730                 }
1731                 spin_unlock(&inode->i_lock);
1732                 if (dirty_inode)
1733                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1734         }
1735
1736         if (err == 0)
1737                 *offs = offset;
1738         return err;
1739 }
1740
1741 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1742                          const struct lu_buf *buf, loff_t *pos,
1743                          struct thandle *handle, int ignore_quota)
1744 {
1745         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1746         struct osd_thandle      *oh;
1747         ssize_t                 result;
1748         int                     is_link;
1749
1750         LASSERT(dt_object_exists(dt));
1751
1752         LASSERT(handle != NULL);
1753         LASSERT(inode != NULL);
1754         ll_vfs_dq_init(inode);
1755
1756         /* XXX: don't check: one declared chunk can be used many times */
1757         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1758
1759         oh = container_of(handle, struct osd_thandle, ot_super);
1760         LASSERT(oh->ot_handle->h_transaction != NULL);
1761         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1762
1763         /* Write small symlink to inode body as we need to maintain correct
1764          * on-disk symlinks for ldiskfs.
1765          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1766          * does not count it in.
1767          */
1768         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1769         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1770                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1771         else
1772                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1773                                                   buf->lb_len, is_link, pos,
1774                                                   oh->ot_handle);
1775         if (result == 0)
1776                 result = buf->lb_len;
1777
1778         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1779
1780         return result;
1781 }
1782
1783 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1784                              __u64 start, __u64 end, struct thandle *th)
1785 {
1786         struct osd_thandle *oh;
1787         struct inode       *inode;
1788         int                 rc;
1789         ENTRY;
1790
1791         LASSERT(th);
1792         oh = container_of(th, struct osd_thandle, ot_super);
1793
1794         /*
1795          * we don't need to reserve credits for whole truncate
1796          * it's not possible as truncate may need to free too many
1797          * blocks and that won't fit a single transaction. instead
1798          * we reserve credits to change i_size and put inode onto
1799          * orphan list. if needed truncate will extend or restart
1800          * transaction
1801          */
1802         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1803                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1804
1805         inode = osd_dt_obj(dt)->oo_inode;
1806         LASSERT(inode);
1807
1808         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1809                                    0, oh, osd_dt_obj(dt), true, NULL, false);
1810         RETURN(rc);
1811 }
1812
1813 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1814                      __u64 start, __u64 end, struct thandle *th)
1815 {
1816         struct osd_thandle *oh;
1817         struct osd_object  *obj = osd_dt_obj(dt);
1818         struct inode       *inode = obj->oo_inode;
1819         handle_t           *h;
1820         tid_t               tid;
1821         int                rc = 0, rc2 = 0;
1822         ENTRY;
1823
1824         LASSERT(end == OBD_OBJECT_EOF);
1825         LASSERT(dt_object_exists(dt));
1826         LASSERT(osd_invariant(obj));
1827         LASSERT(inode != NULL);
1828         ll_vfs_dq_init(inode);
1829
1830         LASSERT(th);
1831         oh = container_of(th, struct osd_thandle, ot_super);
1832         LASSERT(oh->ot_handle->h_transaction != NULL);
1833
1834         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1835
1836         tid = oh->ot_handle->h_transaction->t_tid;
1837
1838         spin_lock(&inode->i_lock);
1839         i_size_write(inode, start);
1840         spin_unlock(&inode->i_lock);
1841         ll_truncate_pagecache(inode, start);
1842 #ifdef HAVE_INODEOPS_TRUNCATE
1843         if (inode->i_op->truncate) {
1844                 inode->i_op->truncate(inode);
1845         } else
1846 #endif
1847                 ldiskfs_truncate(inode);
1848
1849         /*
1850          * For a partial-page truncate, flush the page to disk immediately to
1851          * avoid data corruption during direct disk write.  b=17397
1852          */
1853         if ((start & ~PAGE_MASK) != 0)
1854                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1855
1856         h = journal_current_handle();
1857         LASSERT(h != NULL);
1858         LASSERT(h == oh->ot_handle);
1859
1860         /* do not check credits with osd_trans_exec_check() as the truncate
1861          * can restart the transaction internally and we restart the
1862          * transaction in this case */
1863
1864         if (tid != h->h_transaction->t_tid) {
1865                 int credits = oh->ot_credits;
1866                 /*
1867                  * transaction has changed during truncate
1868                  * we need to restart the handle with our credits
1869                  */
1870                 if (h->h_buffer_credits < credits) {
1871                         if (ldiskfs_journal_extend(h, credits))
1872                                 rc2 = ldiskfs_journal_restart(h, credits);
1873                 }
1874         }
1875
1876         RETURN(rc == 0 ? rc2 : rc);
1877 }
1878
1879 static int fiemap_check_ranges(struct inode *inode,
1880                                u64 start, u64 len, u64 *new_len)
1881 {
1882         loff_t maxbytes;
1883
1884         *new_len = len;
1885
1886         if (len == 0)
1887                 return -EINVAL;
1888
1889         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1890                 maxbytes = inode->i_sb->s_maxbytes;
1891         else
1892                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1893
1894         if (start > maxbytes)
1895                 return -EFBIG;
1896
1897         /*
1898          * Shrink request scope to what the fs can actually handle.
1899          */
1900         if (len > maxbytes || (maxbytes - len) < start)
1901                 *new_len = maxbytes - start;
1902
1903         return 0;
1904 }
1905
1906 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1907 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
1908
1909 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1910                           struct fiemap *fm)
1911 {
1912         struct fiemap_extent_info fieinfo = {0, };
1913         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1914         u64 len;
1915         int rc;
1916
1917
1918         LASSERT(inode);
1919         if (inode->i_op->fiemap == NULL)
1920                 return -EOPNOTSUPP;
1921
1922         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1923                 return -EINVAL;
1924
1925         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1926         if (rc)
1927                 return rc;
1928
1929         fieinfo.fi_flags = fm->fm_flags;
1930         fieinfo.fi_extents_max = fm->fm_extent_count;
1931         fieinfo.fi_extents_start = fm->fm_extents;
1932
1933         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1934                 filemap_write_and_wait(inode->i_mapping);
1935
1936         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
1937         fm->fm_flags = fieinfo.fi_flags;
1938         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
1939
1940         return rc;
1941 }
1942
1943 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1944                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1945 {
1946         int              rc = 0;
1947         struct inode    *inode = osd_dt_obj(dt)->oo_inode;
1948         ENTRY;
1949
1950         switch (advice) {
1951         case LU_LADVISE_DONTNEED:
1952                 if (end == 0)
1953                         break;
1954                 invalidate_mapping_pages(inode->i_mapping,
1955                                          start >> PAGE_CACHE_SHIFT,
1956                                          (end - 1) >> PAGE_CACHE_SHIFT);
1957                 break;
1958         default:
1959                 rc = -ENOTSUPP;
1960                 break;
1961         }
1962
1963         RETURN(rc);
1964 }
1965
1966 /*
1967  * in some cases we may need declare methods for objects being created
1968  * e.g., when we create symlink
1969  */
1970 const struct dt_body_operations osd_body_ops_new = {
1971         .dbo_declare_write = osd_declare_write,
1972 };
1973
1974 const struct dt_body_operations osd_body_ops = {
1975         .dbo_read                       = osd_read,
1976         .dbo_declare_write              = osd_declare_write,
1977         .dbo_write                      = osd_write,
1978         .dbo_bufs_get                   = osd_bufs_get,
1979         .dbo_bufs_put                   = osd_bufs_put,
1980         .dbo_write_prep                 = osd_write_prep,
1981         .dbo_declare_write_commit       = osd_declare_write_commit,
1982         .dbo_write_commit               = osd_write_commit,
1983         .dbo_read_prep                  = osd_read_prep,
1984         .dbo_declare_punch              = osd_declare_punch,
1985         .dbo_punch                      = osd_punch,
1986         .dbo_fiemap_get                 = osd_fiemap_get,
1987         .dbo_ladvise                    = osd_ladvise,
1988 };