lustre/osd-ldiskfs/osd_io.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/osd/osd_io.c
  37  *
  38  * body operations
  39  *
  40  * Author: Nikita Danilov <nikita@clusterfs.com>
  41  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  42  *
  43  */
  44
  45 /* LUSTRE_VERSION_CODE */
  46 #include <lustre_ver.h>
  47 /* prerequisite for linux/xattr.h */
  48 #include <linux/types.h>
  49 /* prerequisite for linux/xattr.h */
  50 #include <linux/fs.h>
  51
  52 /*
  53  * struct OBD_{ALLOC,FREE}*()
  54  * OBD_FAIL_CHECK
  55  */
  56 #include <obd_support.h>
  57
  58 #include "osd_internal.h"
  59
  60 /* ext_depth() */
  61 #include <ldiskfs/ldiskfs_extents.h>
  62
  63 #ifndef HAVE_PAGE_CONSTANT
  64 #define mapping_cap_page_constant_write(mapping) 0
  65 #define SetPageConstant(page) do {} while (0)
  66 #define ClearPageConstant(page) do {} while (0)
  67 #endif
  68
  69 #ifndef HAS_GENERIC_ERROR_REMOVE_PAGE
  70 int generic_error_remove_page(struct address_space *mapping, struct page *page)
  71 {
  72         if (mapping == NULL)
  73                 return -EINVAL;
  74
  75         if (mapping != page->mapping)
  76                 return -EIO;
  77         /*
  78          * Only punch for normal data pages for now.
  79          * Handling other types like directories would need more auditing.
  80          */
  81         if (!S_ISREG(mapping->host->i_mode))
  82                 return -EIO;
  83
  84         if (page_mapped(page)) {
  85                 unmap_mapping_range(mapping,
  86                                     (loff_t)page->index << PAGE_CACHE_SHIFT,
  87                                     PAGE_CACHE_SIZE, 0);
  88         }
  89         truncate_complete_page(mapping, page);
  90         return 0;
  91 }
  92 #endif
  93
  94 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  95                             int rw, int line, int pages)
  96 {
  97         int blocks, i;
  98
  99         LASSERTF(iobuf->dr_elapsed_valid == 0,
 100                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
 101                  cfs_atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
 102                  iobuf->dr_init_at);
 103         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
 104
 105         cfs_waitq_init(&iobuf->dr_wait);
 106         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 107         iobuf->dr_npages = 0;
 108         iobuf->dr_error = 0;
 109         iobuf->dr_dev = d;
 110         iobuf->dr_frags = 0;
 111         iobuf->dr_elapsed = 0;
 112         /* must be counted before, so assert */
 113         iobuf->dr_rw = rw;
 114         iobuf->dr_init_at = line;
 115
 116         blocks = pages * (PAGE_CACHE_SIZE >> osd_sb(d)->s_blocksize_bits);
 117         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
 118                 LASSERT(iobuf->dr_pg_buf.lb_len >=
 119                         pages * sizeof(iobuf->dr_pages[0]));
 120                 return 0;
 121         }
 122
 123         /* start with 1MB for 4K blocks */
 124         i = 256;
 125         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
 126                 i <<= 1;
 127
 128         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
 129                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
 130         pages = i;
 131         blocks = pages * (PAGE_CACHE_SIZE >> osd_sb(d)->s_blocksize_bits);
 132         iobuf->dr_max_pages = 0;
 133         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
 134                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 135
 136         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 137         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 138         if (unlikely(iobuf->dr_blocks == NULL))
 139                 return -ENOMEM;
 140
 141         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 142         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 143         if (unlikely(iobuf->dr_pages == NULL))
 144                 return -ENOMEM;
 145
 146         iobuf->dr_max_pages = pages;
 147
 148         return 0;
 149 }
 150 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 151         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 152
 153 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
 154 {
 155         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 156         iobuf->dr_pages[iobuf->dr_npages++] = page;
 157 }
 158
 159 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 160 {
 161         int rw = iobuf->dr_rw;
 162
 163         if (iobuf->dr_elapsed_valid) {
 164                 iobuf->dr_elapsed_valid = 0;
 165                 LASSERT(iobuf->dr_dev == d);
 166                 LASSERT(iobuf->dr_frags > 0);
 167                 lprocfs_oh_tally(&d->od_brw_stats.
 168                                  hist[BRW_R_DIO_FRAGS+rw],
 169                                  iobuf->dr_frags);
 170                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 171                                       iobuf->dr_elapsed);
 172         }
 173 }
 174
 175 #ifndef REQ_WRITE /* pre-2.6.35 */
 176 #define __REQ_WRITE BIO_RW
 177 #endif
 178
 179 #ifdef HAVE_BIO_ENDIO_2ARG
 180 #define DIO_RETURN(a)
 181 static void dio_complete_routine(struct bio *bio, int error)
 182 #else
 183 #define DIO_RETURN(a)   return(a)
 184 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 185 #endif
 186 {
 187         struct osd_iobuf *iobuf = bio->bi_private;
 188         struct bio_vec *bvl;
 189         int i;
 190
 191         /* CAVEAT EMPTOR: possibly in IRQ context
 192          * DO NOT record procfs stats here!!! */
 193
 194         if (unlikely(iobuf == NULL)) {
 195                 CERROR("***** bio->bi_private is NULL!  This should never "
 196                        "happen.  Normally, I would crash here, but instead I "
 197                        "will dump the bio contents to the console.  Please "
 198                        "report this to <http://jira.whamcloud.com/> , along "
 199                        "with any interesting messages leading up to this point "
 200                        "(like SCSI errors, perhaps).  Because bi_private is "
 201                        "NULL, I can't wake up the thread that initiated this "
 202                        "IO - you will probably have to reboot this node.\n");
 203                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 204                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 205                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 206                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 207                        bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
 208                        bio->bi_private);
 209                 DIO_RETURN(0);
 210         }
 211
 212         /* the check is outside of the cycle for performance reason -bzzz */
 213         if (!test_bit(__REQ_WRITE, &bio->bi_rw)) {
 214                 bio_for_each_segment(bvl, bio, i) {
 215                         if (likely(error == 0))
 216                                 SetPageUptodate(bvl->bv_page);
 217                         LASSERT(PageLocked(bvl->bv_page));
 218                         ClearPageConstant(bvl->bv_page);
 219                 }
 220                 cfs_atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 221         } else {
 222                 struct page *p = iobuf->dr_pages[0];
 223                 if (p->mapping) {
 224                         if (mapping_cap_page_constant_write(p->mapping)) {
 225                                 bio_for_each_segment(bvl, bio, i) {
 226                                         ClearPageConstant(bvl->bv_page);
 227                                 }
 228                         }
 229                 }
 230                 cfs_atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 231         }
 232
 233         /* any real error is good enough -bzzz */
 234         if (error != 0 && iobuf->dr_error == 0)
 235                 iobuf->dr_error = error;
 236
 237         /*
 238          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 239          * it's possible that service thread will see dr_numreqs
 240          * is zero, but dr_elapsed is not set yet, leading to lost
 241          * data in this processing and an assertion in a subsequent
 242          * call to OSD.
 243          */
 244         if (cfs_atomic_read(&iobuf->dr_numreqs) == 1) {
 245                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
 246                 iobuf->dr_elapsed_valid = 1;
 247         }
 248         if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs))
 249                 cfs_waitq_signal(&iobuf->dr_wait);
 250
 251         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 252          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 253          * mempool when serious on-disk fragmentation was encountered,
 254          * deadlocking the OST.  The bios are now released as soon as complete
 255          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 256         bio_put(bio);
 257         DIO_RETURN(0);
 258 }
 259
 260 static void record_start_io(struct osd_iobuf *iobuf, int size)
 261 {
 262         struct osd_device    *osd = iobuf->dr_dev;
 263         struct obd_histogram *h = osd->od_brw_stats.hist;
 264
 265         iobuf->dr_frags++;
 266         cfs_atomic_inc(&iobuf->dr_numreqs);
 267
 268         if (iobuf->dr_rw == 0) {
 269                 cfs_atomic_inc(&osd->od_r_in_flight);
 270                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 271                                  cfs_atomic_read(&osd->od_r_in_flight));
 272                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 273         } else if (iobuf->dr_rw == 1) {
 274                 cfs_atomic_inc(&osd->od_w_in_flight);
 275                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 276                                  cfs_atomic_read(&osd->od_w_in_flight));
 277                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 278         } else {
 279                 LBUG();
 280         }
 281 }
 282
 283 static void osd_submit_bio(int rw, struct bio *bio)
 284 {
 285         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 286         if (rw == 0)
 287                 submit_bio(READ, bio);
 288         else
 289                 submit_bio(WRITE, bio);
 290 }
 291
 292 static int can_be_merged(struct bio *bio, sector_t sector)
 293 {
 294         unsigned int size;
 295
 296         if (!bio)
 297                 return 0;
 298
 299         size = bio->bi_size >> 9;
 300         return bio->bi_sector + size == sector ? 1 : 0;
 301 }
 302
 303 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 304                       struct osd_iobuf *iobuf)
 305 {
 306         int            blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
 307         struct page  **pages = iobuf->dr_pages;
 308         int            npages = iobuf->dr_npages;
 309         unsigned long *blocks = iobuf->dr_blocks;
 310         int            total_blocks = npages * blocks_per_page;
 311         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 312         unsigned int   blocksize = inode->i_sb->s_blocksize;
 313         struct bio    *bio = NULL;
 314         struct page   *page;
 315         unsigned int   page_offset;
 316         sector_t       sector;
 317         int            nblocks;
 318         int            block_idx;
 319         int            page_idx;
 320         int            i;
 321         int            rc = 0;
 322         ENTRY;
 323
 324         LASSERT(iobuf->dr_npages == npages);
 325
 326         osd_brw_stats_update(osd, iobuf);
 327         iobuf->dr_start_time = cfs_time_current();
 328
 329         for (page_idx = 0, block_idx = 0;
 330              page_idx < npages;
 331              page_idx++, block_idx += blocks_per_page) {
 332
 333                 page = pages[page_idx];
 334                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 335
 336                 for (i = 0, page_offset = 0;
 337                      i < blocks_per_page;
 338                      i += nblocks, page_offset += blocksize * nblocks) {
 339
 340                         nblocks = 1;
 341
 342                         if (blocks[block_idx + i] == 0) {  /* hole */
 343                                 LASSERTF(iobuf->dr_rw == 0,
 344                                          "page_idx %u, block_idx %u, i %u\n",
 345                                          page_idx, block_idx, i);
 346                                 memset(kmap(page) + page_offset, 0, blocksize);
 347                                 kunmap(page);
 348                                 continue;
 349                         }
 350
 351                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 352
 353                         /* Additional contiguous file blocks? */
 354                         while (i + nblocks < blocks_per_page &&
 355                                (sector + (nblocks << sector_bits)) ==
 356                                ((sector_t)blocks[block_idx + i + nblocks] <<
 357                                 sector_bits))
 358                                 nblocks++;
 359
 360                         /* I only set the page to be constant only if it
 361                          * is mapped to a contiguous underlying disk block(s).
 362                          * It will then make sure the corresponding device
 363                          * cache of raid5 will be overwritten by this page.
 364                          * - jay */
 365                         if (iobuf->dr_rw && (nblocks == blocks_per_page) &&
 366                             mapping_cap_page_constant_write(inode->i_mapping))
 367                                 SetPageConstant(page);
 368
 369                         if (bio != NULL &&
 370                             can_be_merged(bio, sector) &&
 371                             bio_add_page(bio, page,
 372                                          blocksize * nblocks, page_offset) != 0)
 373                                 continue;       /* added this frag OK */
 374
 375                         if (bio != NULL) {
 376                                 struct request_queue *q =
 377                                         bdev_get_queue(bio->bi_bdev);
 378
 379                                 /* Dang! I have to fragment this I/O */
 380                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 381                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 382                                        bio->bi_size,
 383                                        bio->bi_vcnt, bio->bi_max_vecs,
 384                                        bio->bi_size >> 9, queue_max_sectors(q),
 385                                        bio_phys_segments(q, bio),
 386                                        queue_max_phys_segments(q),
 387                                        bio_hw_segments(q, bio),
 388                                        queue_max_hw_segments(q));
 389
 390                                 record_start_io(iobuf, bio->bi_size);
 391                                 osd_submit_bio(iobuf->dr_rw, bio);
 392                         }
 393
 394                         /* allocate new bio */
 395                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 396                                                       (npages - page_idx) *
 397                                                       blocks_per_page));
 398                         if (bio == NULL) {
 399                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 400                                        (npages - page_idx), blocks_per_page,
 401                                        (npages - page_idx) * blocks_per_page);
 402                                 rc = -ENOMEM;
 403                                 goto out;
 404                         }
 405
 406                         bio->bi_bdev = inode->i_sb->s_bdev;
 407                         bio->bi_sector = sector;
 408                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 409                         bio->bi_end_io = dio_complete_routine;
 410                         bio->bi_private = iobuf;
 411
 412                         rc = bio_add_page(bio, page,
 413                                           blocksize * nblocks, page_offset);
 414                         LASSERT(rc != 0);
 415                 }
 416         }
 417
 418         if (bio != NULL) {
 419                 record_start_io(iobuf, bio->bi_size);
 420                 osd_submit_bio(iobuf->dr_rw, bio);
 421                 rc = 0;
 422         }
 423
 424  out:
 425         /* in order to achieve better IO throughput, we don't wait for writes
 426          * completion here. instead we proceed with transaction commit in
 427          * parallel and wait for IO completion once transaction is stopped
 428          * see osd_trans_stop() for more details -bzzz */
 429         if (iobuf->dr_rw == 0) {
 430                 cfs_wait_event(iobuf->dr_wait,
 431                                cfs_atomic_read(&iobuf->dr_numreqs) == 0);
 432         }
 433
 434         if (rc == 0)
 435                 rc = iobuf->dr_error;
 436         RETURN(rc);
 437 }
 438
 439 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 440                                    struct niobuf_local *lnb)
 441 {
 442         ENTRY;
 443
 444         *nrpages = 0;
 445
 446         while (len > 0) {
 447                 int poff = offset & (PAGE_CACHE_SIZE - 1);
 448                 int plen = PAGE_CACHE_SIZE - poff;
 449
 450                 if (plen > len)
 451                         plen = len;
 452                 lnb->lnb_file_offset = offset;
 453                 lnb->lnb_page_offset = poff;
 454                 lnb->len = plen;
 455                 /* lb->flags = rnb->flags; */
 456                 lnb->flags = 0;
 457                 lnb->page = NULL;
 458                 lnb->rc = 0;
 459
 460                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 461                          (long long) len);
 462                 offset += plen;
 463                 len -= plen;
 464                 lnb++;
 465                 (*nrpages)++;
 466         }
 467
 468         RETURN(0);
 469 }
 470
 471 struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
 472 {
 473         struct inode      *inode = osd_dt_obj(dt)->oo_inode;
 474         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 475         struct page       *page;
 476
 477         LASSERT(inode);
 478
 479         page = find_or_create_page(inode->i_mapping, offset >> PAGE_CACHE_SHIFT,
 480                                    GFP_NOFS | __GFP_HIGHMEM);
 481         if (unlikely(page == NULL))
 482                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 483
 484         return page;
 485 }
 486
 487 /*
 488  * there are following "locks":
 489  * journal_start
 490  * i_mutex
 491  * page lock
 492
 493  * osd write path
 494     * lock page(s)
 495     * journal_start
 496     * truncate_sem
 497
 498  * ext4 vmtruncate:
 499     * lock pages, unlock
 500     * journal_start
 501     * lock partial page
 502     * i_data_sem
 503
 504 */
 505 int osd_bufs_get(const struct lu_env *env, struct dt_object *d, loff_t pos,
 506                  ssize_t len, struct niobuf_local *lnb, int rw,
 507                  struct lustre_capa *capa)
 508 {
 509         struct osd_object   *obj    = osd_dt_obj(d);
 510         int npages, i, rc = 0;
 511
 512         LASSERT(obj->oo_inode);
 513
 514         osd_map_remote_to_local(pos, len, &npages, lnb);
 515
 516         for (i = 0; i < npages; i++, lnb++) {
 517
 518                 /* We still set up for ungranted pages so that granted pages
 519                  * can be written to disk as they were promised, and portals
 520                  * needs to keep the pages all aligned properly. */
 521                 lnb->dentry = (void *) obj;
 522
 523                 lnb->page = osd_get_page(d, lnb->lnb_file_offset, rw);
 524                 if (lnb->page == NULL)
 525                         GOTO(cleanup, rc = -ENOMEM);
 526
 527                 /* DLM locking protects us from write and truncate competing
 528                  * for same region, but truncate can leave dirty page in the
 529                  * cache. it's possible the writeout on a such a page is in
 530                  * progress when we access it. it's also possible that during
 531                  * this writeout we put new (partial) data, but then won't
 532                  * be able to proceed in filter_commitrw_write(). thus let's
 533                  * just wait for writeout completion, should be rare enough.
 534                  * -bzzz */
 535                 wait_on_page_writeback(lnb->page);
 536                 BUG_ON(PageWriteback(lnb->page));
 537
 538                 lu_object_get(&d->do_lu);
 539         }
 540         rc = i;
 541
 542 cleanup:
 543         RETURN(rc);
 544 }
 545
 546 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 547                         struct niobuf_local *lnb, int npages)
 548 {
 549         struct osd_thread_info *oti = osd_oti_get(env);
 550         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 551         struct osd_device      *d = osd_obj2dev(osd_dt_obj(dt));
 552         int                     i;
 553
 554         /* to do IO stats, notice we do this here because
 555          * osd_do_bio() doesn't wait for write to complete */
 556         osd_fini_iobuf(d, iobuf);
 557
 558         for (i = 0; i < npages; i++) {
 559                 if (lnb[i].page == NULL)
 560                         continue;
 561                 LASSERT(PageLocked(lnb[i].page));
 562                 unlock_page(lnb[i].page);
 563                 page_cache_release(lnb[i].page);
 564                 lu_object_put(env, &dt->do_lu);
 565                 lnb[i].page = NULL;
 566         }
 567         RETURN(0);
 568 }
 569
 570 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 571                           struct niobuf_local *lnb, int npages)
 572 {
 573         struct osd_thread_info *oti   = osd_oti_get(env);
 574         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 575         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 576         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 577         struct timeval          start;
 578         struct timeval          end;
 579         unsigned long           timediff;
 580         ssize_t                 isize;
 581         __s64                   maxidx;
 582         int                     rc = 0;
 583         int                     i;
 584         int                     cache = 0;
 585
 586         LASSERT(inode);
 587
 588         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 589         if (unlikely(rc != 0))
 590                 RETURN(rc);
 591
 592         isize = i_size_read(inode);
 593         maxidx = ((isize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) - 1;
 594
 595         if (osd->od_writethrough_cache)
 596                 cache = 1;
 597         if (isize > osd->od_readcache_max_filesize)
 598                 cache = 0;
 599
 600         cfs_gettimeofday(&start);
 601         for (i = 0; i < npages; i++) {
 602
 603                 if (cache == 0)
 604                         generic_error_remove_page(inode->i_mapping,
 605                                                   lnb[i].page);
 606
 607                 /*
 608                  * till commit the content of the page is undefined
 609                  * we'll set it uptodate once bulk is done. otherwise
 610                  * subsequent reads can access non-stable data
 611                  */
 612                 ClearPageUptodate(lnb[i].page);
 613
 614                 if (lnb[i].len == PAGE_CACHE_SIZE)
 615                         continue;
 616
 617                 if (maxidx >= lnb[i].page->index) {
 618                         osd_iobuf_add_page(iobuf, lnb[i].page);
 619                 } else {
 620                         long off;
 621                         char *p = kmap(lnb[i].page);
 622
 623                         off = lnb[i].lnb_page_offset;
 624                         if (off)
 625                                 memset(p, 0, off);
 626                         off = (lnb[i].lnb_page_offset + lnb[i].len) &
 627                               ~CFS_PAGE_MASK;
 628                         if (off)
 629                                 memset(p + off, 0, PAGE_CACHE_SIZE - off);
 630                         kunmap(lnb[i].page);
 631                 }
 632         }
 633         cfs_gettimeofday(&end);
 634         timediff = cfs_timeval_sub(&end, &start, NULL);
 635         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 636
 637         if (iobuf->dr_npages) {
 638                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 639                                                        iobuf->dr_npages,
 640                                                        iobuf->dr_blocks,
 641                                                        0, NULL);
 642                 if (likely(rc == 0)) {
 643                         rc = osd_do_bio(osd, inode, iobuf);
 644                         /* do IO stats for preparation reads */
 645                         osd_fini_iobuf(osd, iobuf);
 646                 }
 647         }
 648         RETURN(rc);
 649 }
 650
 651 /* Check if a block is allocated or not */
 652 static int osd_is_mapped(struct inode *inode, obd_size offset)
 653 {
 654         sector_t (*fs_bmap)(struct address_space *, sector_t);
 655
 656         fs_bmap = inode->i_mapping->a_ops->bmap;
 657
 658         /* We can't know if we are overwriting or not */
 659         if (unlikely(fs_bmap == NULL))
 660                 return 0;
 661
 662         if (i_size_read(inode) == 0)
 663                 return 0;
 664
 665         /* Beyond EOF, must not be mapped */
 666         if (((i_size_read(inode) - 1) >> inode->i_blkbits) <
 667             (offset >> inode->i_blkbits))
 668                 return 0;
 669
 670         if (fs_bmap(inode->i_mapping, offset >> inode->i_blkbits) == 0)
 671                 return 0;
 672
 673         return 1;
 674 }
 675
 676 static int osd_declare_write_commit(const struct lu_env *env,
 677                                     struct dt_object *dt,
 678                                     struct niobuf_local *lnb, int npages,
 679                                     struct thandle *handle)
 680 {
 681         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 682         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
 683         struct osd_thandle      *oh;
 684         int                      extents = 1;
 685         int                      depth;
 686         int                      i;
 687         int                      newblocks;
 688         int                      rc = 0;
 689         int                      flags = 0;
 690         bool                     ignore_quota = false;
 691         long long                quota_space = 0;
 692         ENTRY;
 693
 694         LASSERT(handle != NULL);
 695         oh = container_of0(handle, struct osd_thandle, ot_super);
 696         LASSERT(oh->ot_handle == NULL);
 697
 698         newblocks = npages;
 699
 700         /* calculate number of extents (probably better to pass nb) */
 701         for (i = 0; i < npages; i++) {
 702                 if (i && lnb[i].lnb_file_offset !=
 703                     lnb[i - 1].lnb_file_offset + lnb[i - 1].len)
 704                         extents++;
 705
 706                 if (!osd_is_mapped(inode, lnb[i].lnb_file_offset))
 707                         quota_space += PAGE_CACHE_SIZE;
 708
 709                 /* ignore quota for the whole request if any page is from
 710                  * client cache or written by root.
 711                  *
 712                  * XXX once we drop the 1.8 client support, the checking
 713                  * for whether page is from cache can be simplified as:
 714                  * !(lnb[i].flags & OBD_BRW_SYNC)
 715                  *
 716                  * XXX we could handle this on per-lnb basis as done by
 717                  * grant. */
 718                 if ((lnb[i].flags & OBD_BRW_NOQUOTA) ||
 719                     (lnb[i].flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 720                     OBD_BRW_FROM_GRANT)
 721                         ignore_quota = true;
 722         }
 723
 724         /*
 725          * each extent can go into new leaf causing a split
 726          * 5 is max tree depth: inode + 4 index blocks
 727          * with blockmaps, depth is 3 at most
 728          */
 729         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 730                 /*
 731                  * many concurrent threads may grow tree by the time
 732                  * our transaction starts. so, consider 2 is a min depth
 733                  */
 734                 depth = ext_depth(inode);
 735                 depth = max(depth, 1) + 1;
 736                 newblocks += depth;
 737                 oh->ot_credits++; /* inode */
 738                 oh->ot_credits += depth * 2 * extents;
 739         } else {
 740                 depth = 3;
 741                 newblocks += depth;
 742                 oh->ot_credits++; /* inode */
 743                 oh->ot_credits += depth * extents;
 744         }
 745
 746         /* quota space for metadata blocks */
 747         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
 748
 749         /* quota space should be reported in 1K blocks */
 750         quota_space = toqb(quota_space);
 751
 752         /* each new block can go in different group (bitmap + gd) */
 753
 754         /* we can't dirty more bitmap blocks than exist */
 755         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
 756                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
 757         else
 758                 oh->ot_credits += newblocks;
 759
 760         /* we can't dirty more gd blocks than exist */
 761         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
 762                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
 763         else
 764                 oh->ot_credits += newblocks;
 765
 766         /* make sure the over quota flags were not set */
 767         lnb[0].flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA);
 768
 769         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid,
 770                                    quota_space, oh, true, true, &flags,
 771                                    ignore_quota);
 772
 773         /* we need only to store the overquota flags in the first lnb for
 774          * now, once we support multiple objects BRW, this code needs be
 775          * revised. */
 776         if (flags & QUOTA_FL_OVER_USRQUOTA)
 777                 lnb[0].flags |= OBD_BRW_OVER_USRQUOTA;
 778         if (flags & QUOTA_FL_OVER_GRPQUOTA)
 779                 lnb[0].flags |= OBD_BRW_OVER_GRPQUOTA;
 780
 781         RETURN(rc);
 782 }
 783
 784 /* Check if a block is allocated or not */
 785 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 786                             struct niobuf_local *lnb, int npages,
 787                             struct thandle *thandle)
 788 {
 789         struct osd_thread_info *oti = osd_oti_get(env);
 790         struct osd_iobuf *iobuf = &oti->oti_iobuf;
 791         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 792         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
 793         loff_t isize;
 794         int rc = 0, i;
 795
 796         LASSERT(inode);
 797
 798         rc = osd_init_iobuf(osd, iobuf, 1, npages);
 799         if (unlikely(rc != 0))
 800                 RETURN(rc);
 801
 802         isize = i_size_read(inode);
 803         ll_vfs_dq_init(inode);
 804
 805         for (i = 0; i < npages; i++) {
 806                 if (lnb[i].rc == -ENOSPC &&
 807                     osd_is_mapped(inode, lnb[i].lnb_file_offset)) {
 808                         /* Allow the write to proceed if overwriting an
 809                          * existing block */
 810                         lnb[i].rc = 0;
 811                 }
 812
 813                 if (lnb[i].rc) { /* ENOSPC, network RPC error, etc. */
 814                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
 815                                lnb[i].rc);
 816                         LASSERT(lnb[i].page);
 817                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 818                         continue;
 819                 }
 820
 821                 LASSERT(PageLocked(lnb[i].page));
 822                 LASSERT(!PageWriteback(lnb[i].page));
 823
 824                 if (lnb[i].lnb_file_offset + lnb[i].len > isize)
 825                         isize = lnb[i].lnb_file_offset + lnb[i].len;
 826
 827                 /*
 828                  * Since write and truncate are serialized by oo_sem, even
 829                  * partial-page truncate should not leave dirty pages in the
 830                  * page cache.
 831                  */
 832                 LASSERT(!PageDirty(lnb[i].page));
 833
 834                 SetPageUptodate(lnb[i].page);
 835
 836                 osd_iobuf_add_page(iobuf, lnb[i].page);
 837         }
 838
 839         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
 840                 rc = -ENOSPC;
 841         } else if (iobuf->dr_npages > 0) {
 842                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 843                                                        iobuf->dr_npages,
 844                                                        iobuf->dr_blocks,
 845                                                        1, NULL);
 846         } else {
 847                 /* no pages to write, no transno is needed */
 848                 thandle->th_local = 1;
 849         }
 850
 851         if (likely(rc == 0)) {
 852                 if (isize > i_size_read(inode)) {
 853                         i_size_write(inode, isize);
 854                         LDISKFS_I(inode)->i_disksize = isize;
 855                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
 856                 }
 857
 858                 rc = osd_do_bio(osd, inode, iobuf);
 859                 /* we don't do stats here as in read path because
 860                  * write is async: we'll do this in osd_put_bufs() */
 861         }
 862
 863         if (unlikely(rc != 0)) {
 864                 /* if write fails, we should drop pages from the cache */
 865                 for (i = 0; i < npages; i++) {
 866                         if (lnb[i].page == NULL)
 867                                 continue;
 868                         LASSERT(PageLocked(lnb[i].page));
 869                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 870                 }
 871         }
 872
 873         RETURN(rc);
 874 }
 875
 876 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 877                          struct niobuf_local *lnb, int npages)
 878 {
 879         struct osd_thread_info *oti = osd_oti_get(env);
 880         struct osd_iobuf *iobuf = &oti->oti_iobuf;
 881         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 882         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 883         struct timeval start, end;
 884         unsigned long timediff;
 885         int rc = 0, i, m = 0, cache = 0;
 886
 887         LASSERT(inode);
 888
 889         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 890         if (unlikely(rc != 0))
 891                 RETURN(rc);
 892
 893         if (osd->od_read_cache)
 894                 cache = 1;
 895         if (i_size_read(inode) > osd->od_readcache_max_filesize)
 896                 cache = 0;
 897
 898         cfs_gettimeofday(&start);
 899         for (i = 0; i < npages; i++) {
 900
 901                 if (i_size_read(inode) <= lnb[i].lnb_file_offset)
 902                         /* If there's no more data, abort early.
 903                          * lnb->rc == 0, so it's easy to detect later. */
 904                         break;
 905
 906                 if (i_size_read(inode) <
 907                     lnb[i].lnb_file_offset + lnb[i].len - 1)
 908                         lnb[i].rc = i_size_read(inode) - lnb[i].lnb_file_offset;
 909                 else
 910                         lnb[i].rc = lnb[i].len;
 911                 m += lnb[i].len;
 912
 913                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS, 1);
 914                 if (PageUptodate(lnb[i].page)) {
 915                         lprocfs_counter_add(osd->od_stats,
 916                                             LPROC_OSD_CACHE_HIT, 1);
 917                 } else {
 918                         lprocfs_counter_add(osd->od_stats,
 919                                             LPROC_OSD_CACHE_MISS, 1);
 920                         osd_iobuf_add_page(iobuf, lnb[i].page);
 921                 }
 922                 if (cache == 0)
 923                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 924         }
 925         cfs_gettimeofday(&end);
 926         timediff = cfs_timeval_sub(&end, &start, NULL);
 927         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 928
 929         if (iobuf->dr_npages) {
 930                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 931                                                        iobuf->dr_npages,
 932                                                        iobuf->dr_blocks,
 933                                                        0, NULL);
 934                 rc = osd_do_bio(osd, inode, iobuf);
 935
 936                 /* IO stats will be done in osd_bufs_put() */
 937         }
 938
 939         RETURN(rc);
 940 }
 941
 942 /*
 943  * XXX: Another layering violation for now.
 944  *
 945  * We don't want to use ->f_op->read methods, because generic file write
 946  *
 947  *         - serializes on ->i_sem, and
 948  *
 949  *         - does a lot of extra work like balance_dirty_pages(),
 950  *
 951  * which doesn't work for globally shared files like /last_rcvd.
 952  */
 953 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
 954 {
 955         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 956
 957         memcpy(buffer, (char *)ei->i_data, buflen);
 958
 959         return  buflen;
 960 }
 961
 962 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
 963 {
 964         struct buffer_head *bh;
 965         unsigned long block;
 966         int osize;
 967         int blocksize;
 968         int csize;
 969         int boffs;
 970         int err;
 971
 972         /* prevent reading after eof */
 973         spin_lock(&inode->i_lock);
 974         if (i_size_read(inode) < *offs + size) {
 975                 loff_t diff = i_size_read(inode) - *offs;
 976                 spin_unlock(&inode->i_lock);
 977                 if (diff < 0) {
 978                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
 979                                i_size_read(inode), *offs);
 980                         return -EBADR;
 981                 } else if (diff == 0) {
 982                         return 0;
 983                 } else {
 984                         size = diff;
 985                 }
 986         } else {
 987                 spin_unlock(&inode->i_lock);
 988         }
 989
 990         blocksize = 1 << inode->i_blkbits;
 991         osize = size;
 992         while (size > 0) {
 993                 block = *offs >> inode->i_blkbits;
 994                 boffs = *offs & (blocksize - 1);
 995                 csize = min(blocksize - boffs, size);
 996                 bh = ldiskfs_bread(NULL, inode, block, 0, &err);
 997                 if (!bh) {
 998                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %d\n",
 999                                LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
1000                                csize, *offs, inode->i_ino, err);
1001                         return err;
1002                 }
1003
1004                 memcpy(buf, bh->b_data + boffs, csize);
1005                 brelse(bh);
1006
1007                 *offs += csize;
1008                 buf += csize;
1009                 size -= csize;
1010         }
1011         return osize;
1012 }
1013
1014 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1015                         struct lu_buf *buf, loff_t *pos,
1016                         struct lustre_capa *capa)
1017 {
1018         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1019         int           rc;
1020
1021         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_READ))
1022                 RETURN(-EACCES);
1023
1024         /* Read small symlink from inode body as we need to maintain correct
1025          * on-disk symlinks for ldiskfs.
1026          */
1027         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1028             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1029                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1030         else
1031                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1032
1033         return rc;
1034 }
1035
1036 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1037                                  const loff_t size, loff_t pos,
1038                                  struct thandle *handle)
1039 {
1040         struct osd_thandle *oh;
1041         int                 credits;
1042         struct inode       *inode;
1043         int                 rc;
1044         ENTRY;
1045
1046         LASSERT(handle != NULL);
1047
1048         oh = container_of0(handle, struct osd_thandle, ot_super);
1049         LASSERT(oh->ot_handle == NULL);
1050
1051         credits = osd_dto_credits_noquota[DTO_WRITE_BLOCK];
1052
1053         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1054
1055         inode = osd_dt_obj(dt)->oo_inode;
1056
1057         /* we may declare write to non-exist llog */
1058         if (inode == NULL)
1059                 RETURN(0);
1060
1061         /* dt_declare_write() is usually called for system objects, such
1062          * as llog or last_rcvd files. We needn't enforce quota on those
1063          * objects, so always set the lqi_space as 0. */
1064         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
1065                                    true, true, NULL, false);
1066         RETURN(rc);
1067 }
1068
1069 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1070 {
1071         /* LU-2634: clear the extent format for fast symlink */
1072         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1073
1074         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1075         LDISKFS_I(inode)->i_disksize = buflen;
1076         i_size_write(inode, buflen);
1077         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1078
1079         return 0;
1080 }
1081
1082 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1083                              int write_NUL, loff_t *offs, handle_t *handle)
1084 {
1085         struct buffer_head *bh        = NULL;
1086         loff_t              offset    = *offs;
1087         loff_t              new_size  = i_size_read(inode);
1088         unsigned long       block;
1089         int                 blocksize = 1 << inode->i_blkbits;
1090         int                 err = 0;
1091         int                 size;
1092         int                 boffs;
1093         int                 dirty_inode = 0;
1094
1095         if (write_NUL) {
1096                 /*
1097                  * long symlink write does not count the NUL terminator in
1098                  * bufsize, we write it, and the inode's file size does not
1099                  * count the NUL terminator as well.
1100                  */
1101                 ((char *)buf)[bufsize] = '\0';
1102                 ++bufsize;
1103         }
1104         while (bufsize > 0) {
1105                 if (bh != NULL)
1106                         brelse(bh);
1107
1108                 block = offset >> inode->i_blkbits;
1109                 boffs = offset & (blocksize - 1);
1110                 size = min(blocksize - boffs, bufsize);
1111                 bh = ldiskfs_bread(handle, inode, block, 1, &err);
1112                 if (!bh) {
1113                         CERROR("%s: error reading offset %llu (block %lu): "
1114                                "rc = %d\n",
1115                                inode->i_sb->s_id, offset, block, err);
1116                         break;
1117                 }
1118
1119                 err = ldiskfs_journal_get_write_access(handle, bh);
1120                 if (err) {
1121                         CERROR("journal_get_write_access() returned error %d\n",
1122                                err);
1123                         break;
1124                 }
1125                 LASSERTF(boffs + size <= bh->b_size,
1126                          "boffs %d size %d bh->b_size %lu",
1127                          boffs, size, (unsigned long)bh->b_size);
1128                 memcpy(bh->b_data + boffs, buf, size);
1129                 err = ldiskfs_journal_dirty_metadata(handle, bh);
1130                 if (err)
1131                         break;
1132
1133                 if (offset + size > new_size)
1134                         new_size = offset + size;
1135                 offset += size;
1136                 bufsize -= size;
1137                 buf += size;
1138         }
1139         if (bh)
1140                 brelse(bh);
1141
1142         if (write_NUL)
1143                 --new_size;
1144         /* correct in-core and on-disk sizes */
1145         if (new_size > i_size_read(inode)) {
1146                 spin_lock(&inode->i_lock);
1147                 if (new_size > i_size_read(inode))
1148                         i_size_write(inode, new_size);
1149                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1150                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1151                         dirty_inode = 1;
1152                 }
1153                 spin_unlock(&inode->i_lock);
1154                 if (dirty_inode)
1155                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1156         }
1157
1158         if (err == 0)
1159                 *offs = offset;
1160         return err;
1161 }
1162
1163 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1164                          const struct lu_buf *buf, loff_t *pos,
1165                          struct thandle *handle, struct lustre_capa *capa,
1166                          int ignore_quota)
1167 {
1168         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1169         struct osd_thandle      *oh;
1170         ssize_t                 result;
1171         int                     is_link;
1172
1173         LASSERT(dt_object_exists(dt));
1174
1175         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_WRITE))
1176                 return -EACCES;
1177
1178         LASSERT(handle != NULL);
1179         LASSERT(inode != NULL);
1180         ll_vfs_dq_init(inode);
1181
1182         /* XXX: don't check: one declared chunk can be used many times */
1183         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1184
1185         oh = container_of(handle, struct osd_thandle, ot_super);
1186         LASSERT(oh->ot_handle->h_transaction != NULL);
1187         /* Write small symlink to inode body as we need to maintain correct
1188          * on-disk symlinks for ldiskfs.
1189          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1190          * does not count it in.
1191          */
1192         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1193         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1194                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1195         else
1196                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1197                                                   buf->lb_len, is_link, pos,
1198                                                   oh->ot_handle);
1199         if (result == 0)
1200                 result = buf->lb_len;
1201         return result;
1202 }
1203
1204 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1205                              __u64 start, __u64 end, struct thandle *th)
1206 {
1207         struct osd_thandle *oh;
1208         struct inode       *inode;
1209         int                 rc;
1210         ENTRY;
1211
1212         LASSERT(th);
1213         oh = container_of(th, struct osd_thandle, ot_super);
1214
1215         /*
1216          * we don't need to reserve credits for whole truncate
1217          * it's not possible as truncate may need to free too many
1218          * blocks and that won't fit a single transaction. instead
1219          * we reserve credits to change i_size and put inode onto
1220          * orphan list. if needed truncate will extend or restart
1221          * transaction
1222          */
1223         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1224                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1225
1226         inode = osd_dt_obj(dt)->oo_inode;
1227         LASSERT(inode);
1228
1229         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
1230                                    true, true, NULL, false);
1231         RETURN(rc);
1232 }
1233
1234 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1235                      __u64 start, __u64 end, struct thandle *th,
1236                      struct lustre_capa *capa)
1237 {
1238         struct osd_thandle *oh;
1239         struct osd_object  *obj = osd_dt_obj(dt);
1240         struct inode       *inode = obj->oo_inode;
1241         handle_t           *h;
1242         tid_t               tid;
1243         loff_t             oldsize;
1244         int                rc = 0, rc2 = 0;
1245         ENTRY;
1246
1247         LASSERT(end == OBD_OBJECT_EOF);
1248         LASSERT(dt_object_exists(dt));
1249         LASSERT(osd_invariant(obj));
1250         LASSERT(inode != NULL);
1251         ll_vfs_dq_init(inode);
1252
1253         LASSERT(th);
1254         oh = container_of(th, struct osd_thandle, ot_super);
1255         LASSERT(oh->ot_handle->h_transaction != NULL);
1256
1257         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1258
1259         tid = oh->ot_handle->h_transaction->t_tid;
1260
1261         oldsize=inode->i_size;
1262         i_size_write(inode, start);
1263         truncate_pagecache(inode, oldsize, start);
1264         if (inode->i_op->truncate)
1265                 inode->i_op->truncate(inode);
1266
1267         /*
1268          * For a partial-page truncate, flush the page to disk immediately to
1269          * avoid data corruption during direct disk write.  b=17397
1270          */
1271         if ((start & ~CFS_PAGE_MASK) != 0)
1272                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1273
1274         h = journal_current_handle();
1275         LASSERT(h != NULL);
1276         LASSERT(h == oh->ot_handle);
1277
1278         if (tid != h->h_transaction->t_tid) {
1279                 int credits = oh->ot_credits;
1280                 /*
1281                  * transaction has changed during truncate
1282                  * we need to restart the handle with our credits
1283                  */
1284                 if (h->h_buffer_credits < credits) {
1285                         if (ldiskfs_journal_extend(h, credits))
1286                                 rc2 = ldiskfs_journal_restart(h, credits);
1287                 }
1288         }
1289
1290         RETURN(rc == 0 ? rc2 : rc);
1291 }
1292
1293 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1294                           struct ll_user_fiemap *fm)
1295 {
1296         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1297         struct osd_thread_info *info   = osd_oti_get(env);
1298         struct dentry          *dentry = &info->oti_obj_dentry;
1299         struct file            *file   = &info->oti_file;
1300         mm_segment_t            saved_fs;
1301         int rc;
1302
1303         LASSERT(inode);
1304         dentry->d_inode = inode;
1305         dentry->d_sb = inode->i_sb;
1306         file->f_dentry = dentry;
1307         file->f_mapping = inode->i_mapping;
1308         file->f_op = inode->i_fop;
1309
1310         saved_fs = get_fs();
1311         set_fs(get_ds());
1312         /* ldiskfs_ioctl does not have a inode argument */
1313         if (inode->i_fop->unlocked_ioctl)
1314                 rc = inode->i_fop->unlocked_ioctl(file, FSFILT_IOC_FIEMAP,
1315                                                   (long)fm);
1316         else
1317                 rc = -ENOTTY;
1318         set_fs(saved_fs);
1319         return rc;
1320 }
1321
1322 /*
1323  * in some cases we may need declare methods for objects being created
1324  * e.g., when we create symlink
1325  */
1326 const struct dt_body_operations osd_body_ops_new = {
1327         .dbo_declare_write = osd_declare_write,
1328 };
1329
1330 const struct dt_body_operations osd_body_ops = {
1331         .dbo_read                 = osd_read,
1332         .dbo_declare_write        = osd_declare_write,
1333         .dbo_write                = osd_write,
1334         .dbo_bufs_get             = osd_bufs_get,
1335         .dbo_bufs_put             = osd_bufs_put,
1336         .dbo_write_prep           = osd_write_prep,
1337         .dbo_declare_write_commit = osd_declare_write_commit,
1338         .dbo_write_commit         = osd_write_commit,
1339         .dbo_read_prep            = osd_read_prep,
1340         .dbo_declare_punch         = osd_declare_punch,
1341         .dbo_punch                 = osd_punch,
1342         .dbo_fiemap_get           = osd_fiemap_get,
1343 };
1344