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 #include <linux/mm.h>
  48 #include <linux/pagevec.h>
  49
  50 /*
  51  * struct OBD_{ALLOC,FREE}*()
  52  * OBD_FAIL_CHECK
  53  */
  54 #include <obd_support.h>
  55
  56 #include "osd_internal.h"
  57
  58 /* ext_depth() */
  59 #include <ldiskfs/ldiskfs_extents.h>
  60
  61 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  62                             int rw, int line, int pages)
  63 {
  64         int blocks, i;
  65
  66         LASSERTF(iobuf->dr_elapsed_valid == 0,
  67                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  68                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  69                  iobuf->dr_init_at);
  70         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  71
  72         init_waitqueue_head(&iobuf->dr_wait);
  73         atomic_set(&iobuf->dr_numreqs, 0);
  74         iobuf->dr_npages = 0;
  75         iobuf->dr_error = 0;
  76         iobuf->dr_dev = d;
  77         iobuf->dr_frags = 0;
  78         iobuf->dr_elapsed = 0;
  79         /* must be counted before, so assert */
  80         iobuf->dr_rw = rw;
  81         iobuf->dr_init_at = line;
  82
  83         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  84         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  85                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  86                         pages * sizeof(iobuf->dr_pages[0]));
  87                 return 0;
  88         }
  89
  90         /* start with 1MB for 4K blocks */
  91         i = 256;
  92         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
  93                 i <<= 1;
  94
  95         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
  96                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
  97         pages = i;
  98         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  99         iobuf->dr_max_pages = 0;
 100         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
 101                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 102
 103         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 104         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 105         if (unlikely(iobuf->dr_blocks == NULL))
 106                 return -ENOMEM;
 107
 108         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 109         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 110         if (unlikely(iobuf->dr_pages == NULL))
 111                 return -ENOMEM;
 112
 113         iobuf->dr_max_pages = pages;
 114
 115         return 0;
 116 }
 117 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 118         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 119
 120 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
 121 {
 122         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 123         iobuf->dr_pages[iobuf->dr_npages++] = page;
 124 }
 125
 126 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 127 {
 128         int rw = iobuf->dr_rw;
 129
 130         if (iobuf->dr_elapsed_valid) {
 131                 iobuf->dr_elapsed_valid = 0;
 132                 LASSERT(iobuf->dr_dev == d);
 133                 LASSERT(iobuf->dr_frags > 0);
 134                 lprocfs_oh_tally(&d->od_brw_stats.
 135                                  hist[BRW_R_DIO_FRAGS+rw],
 136                                  iobuf->dr_frags);
 137                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 138                                       iobuf->dr_elapsed);
 139         }
 140 }
 141
 142 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 143 static void dio_complete_routine(struct bio *bio)
 144 {
 145         int error = bio->bi_error;
 146 #else
 147 static void dio_complete_routine(struct bio *bio, int error)
 148 {
 149 #endif
 150         struct osd_iobuf *iobuf = bio->bi_private;
 151         int iter;
 152         struct bio_vec *bvl;
 153
 154         /* CAVEAT EMPTOR: possibly in IRQ context
 155          * DO NOT record procfs stats here!!! */
 156
 157         if (unlikely(iobuf == NULL)) {
 158                 CERROR("***** bio->bi_private is NULL!  This should never "
 159                        "happen.  Normally, I would crash here, but instead I "
 160                        "will dump the bio contents to the console.  Please "
 161                        "report this to <https://jira.whamcloud.com/> , along "
 162                        "with any interesting messages leading up to this point "
 163                        "(like SCSI errors, perhaps).  Because bi_private is "
 164                        "NULL, I can't wake up the thread that initiated this "
 165                        "IO - you will probably have to reboot this node.\n");
 166                 CERROR("bi_next: %p, bi_flags: %lx, "
 167 #ifdef HAVE_BI_RW
 168                        "bi_rw: %lu,"
 169 #else
 170                        "bi_opf: %u,"
 171 #endif
 172                        "bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p,"
 173                        "bi_cnt: %d, bi_private: %p\n", bio->bi_next,
 174                         (unsigned long)bio->bi_flags,
 175 #ifdef HAVE_BI_RW
 176                         bio->bi_rw,
 177 #else
 178                         bio->bi_opf,
 179 #endif
 180                         bio->bi_vcnt, bio_idx(bio),
 181                         bio_sectors(bio) << 9, bio->bi_end_io,
 182 #ifdef HAVE_BI_CNT
 183                         atomic_read(&bio->bi_cnt),
 184 #else
 185                         atomic_read(&bio->__bi_cnt),
 186 #endif
 187                         bio->bi_private);
 188                 return;
 189         }
 190
 191         /* the check is outside of the cycle for performance reason -bzzz */
 192         if (!bio_data_dir(bio)) {
 193                 bio_for_each_segment_all(bvl, bio, iter) {
 194                         if (likely(error == 0))
 195                                 SetPageUptodate(bvl_to_page(bvl));
 196                         LASSERT(PageLocked(bvl_to_page(bvl)));
 197                 }
 198                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 199         } else {
 200                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 201         }
 202
 203         /* any real error is good enough -bzzz */
 204         if (error != 0 && iobuf->dr_error == 0)
 205                 iobuf->dr_error = error;
 206
 207         /*
 208          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 209          * it's possible that service thread will see dr_numreqs
 210          * is zero, but dr_elapsed is not set yet, leading to lost
 211          * data in this processing and an assertion in a subsequent
 212          * call to OSD.
 213          */
 214         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 215                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
 216                 iobuf->dr_elapsed_valid = 1;
 217         }
 218         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 219                 wake_up(&iobuf->dr_wait);
 220
 221         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 222          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 223          * mempool when serious on-disk fragmentation was encountered,
 224          * deadlocking the OST.  The bios are now released as soon as complete
 225          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 226         bio_put(bio);
 227 }
 228
 229 static void record_start_io(struct osd_iobuf *iobuf, int size)
 230 {
 231         struct osd_device    *osd = iobuf->dr_dev;
 232         struct obd_histogram *h = osd->od_brw_stats.hist;
 233
 234         iobuf->dr_frags++;
 235         atomic_inc(&iobuf->dr_numreqs);
 236
 237         if (iobuf->dr_rw == 0) {
 238                 atomic_inc(&osd->od_r_in_flight);
 239                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 240                                  atomic_read(&osd->od_r_in_flight));
 241                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 242         } else if (iobuf->dr_rw == 1) {
 243                 atomic_inc(&osd->od_w_in_flight);
 244                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 245                                  atomic_read(&osd->od_w_in_flight));
 246                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 247         } else {
 248                 LBUG();
 249         }
 250 }
 251
 252 static void osd_submit_bio(int rw, struct bio *bio)
 253 {
 254         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 255 #ifdef HAVE_SUBMIT_BIO_2ARGS
 256         if (rw == 0)
 257                 submit_bio(READ, bio);
 258         else
 259                 submit_bio(WRITE, bio);
 260 #else
 261         bio->bi_opf |= rw;
 262         submit_bio(bio);
 263 #endif
 264 }
 265
 266 static int can_be_merged(struct bio *bio, sector_t sector)
 267 {
 268         if (bio == NULL)
 269                 return 0;
 270
 271         return bio_end_sector(bio) == sector ? 1 : 0;
 272 }
 273
 274 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 275                       struct osd_iobuf *iobuf)
 276 {
 277         int            blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 278         struct page  **pages = iobuf->dr_pages;
 279         int            npages = iobuf->dr_npages;
 280         sector_t      *blocks = iobuf->dr_blocks;
 281         int            total_blocks = npages * blocks_per_page;
 282         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 283         unsigned int   blocksize = inode->i_sb->s_blocksize;
 284         struct bio    *bio = NULL;
 285         struct page   *page;
 286         unsigned int   page_offset;
 287         sector_t       sector;
 288         int            nblocks;
 289         int            block_idx;
 290         int            page_idx;
 291         int            i;
 292         int            rc = 0;
 293         DECLARE_PLUG(plug);
 294         ENTRY;
 295
 296         LASSERT(iobuf->dr_npages == npages);
 297
 298         osd_brw_stats_update(osd, iobuf);
 299         iobuf->dr_start_time = cfs_time_current();
 300
 301         blk_start_plug(&plug);
 302         for (page_idx = 0, block_idx = 0;
 303              page_idx < npages;
 304              page_idx++, block_idx += blocks_per_page) {
 305
 306                 page = pages[page_idx];
 307                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 308
 309                 for (i = 0, page_offset = 0;
 310                      i < blocks_per_page;
 311                      i += nblocks, page_offset += blocksize * nblocks) {
 312
 313                         nblocks = 1;
 314
 315                         if (blocks[block_idx + i] == 0) {  /* hole */
 316                                 LASSERTF(iobuf->dr_rw == 0,
 317                                          "page_idx %u, block_idx %u, i %u\n",
 318                                          page_idx, block_idx, i);
 319                                 memset(kmap(page) + page_offset, 0, blocksize);
 320                                 kunmap(page);
 321                                 continue;
 322                         }
 323
 324                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 325
 326                         /* Additional contiguous file blocks? */
 327                         while (i + nblocks < blocks_per_page &&
 328                                (sector + (nblocks << sector_bits)) ==
 329                                ((sector_t)blocks[block_idx + i + nblocks] <<
 330                                 sector_bits))
 331                                 nblocks++;
 332
 333                         if (bio != NULL &&
 334                             can_be_merged(bio, sector) &&
 335                             bio_add_page(bio, page,
 336                                          blocksize * nblocks, page_offset) != 0)
 337                                 continue;       /* added this frag OK */
 338
 339                         if (bio != NULL) {
 340                                 struct request_queue *q =
 341                                         bdev_get_queue(bio->bi_bdev);
 342                                 unsigned int bi_size = bio_sectors(bio) << 9;
 343
 344                                 /* Dang! I have to fragment this I/O */
 345                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 346                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 347                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 348                                        bio_sectors(bio),
 349                                        queue_max_sectors(q),
 350                                        bio_phys_segments(q, bio),
 351                                        queue_max_phys_segments(q),
 352                                        0, queue_max_hw_segments(q));
 353                                 record_start_io(iobuf, bi_size);
 354                                 osd_submit_bio(iobuf->dr_rw, bio);
 355                         }
 356
 357                         /* allocate new bio */
 358                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 359                                                       (npages - page_idx) *
 360                                                       blocks_per_page));
 361                         if (bio == NULL) {
 362                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 363                                        (npages - page_idx), blocks_per_page,
 364                                        (npages - page_idx) * blocks_per_page);
 365                                 rc = -ENOMEM;
 366                                 goto out;
 367                         }
 368
 369                         bio->bi_bdev = inode->i_sb->s_bdev;
 370                         bio_set_sector(bio, sector);
 371 #ifdef HAVE_BI_RW
 372                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 373 #else
 374                         bio->bi_opf = (iobuf->dr_rw == 0) ? READ : WRITE;
 375 #endif
 376                         bio->bi_end_io = dio_complete_routine;
 377                         bio->bi_private = iobuf;
 378
 379                         rc = bio_add_page(bio, page,
 380                                           blocksize * nblocks, page_offset);
 381                         LASSERT(rc != 0);
 382                 }
 383         }
 384
 385         if (bio != NULL) {
 386                 record_start_io(iobuf, bio_sectors(bio) << 9);
 387                 osd_submit_bio(iobuf->dr_rw, bio);
 388                 rc = 0;
 389         }
 390
 391 out:
 392         blk_finish_plug(&plug);
 393
 394         /* in order to achieve better IO throughput, we don't wait for writes
 395          * completion here. instead we proceed with transaction commit in
 396          * parallel and wait for IO completion once transaction is stopped
 397          * see osd_trans_stop() for more details -bzzz */
 398         if (iobuf->dr_rw == 0) {
 399                 wait_event(iobuf->dr_wait,
 400                            atomic_read(&iobuf->dr_numreqs) == 0);
 401                 osd_fini_iobuf(osd, iobuf);
 402         }
 403
 404         if (rc == 0)
 405                 rc = iobuf->dr_error;
 406         RETURN(rc);
 407 }
 408
 409 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 410                                    struct niobuf_local *lnb)
 411 {
 412         ENTRY;
 413
 414         *nrpages = 0;
 415
 416         while (len > 0) {
 417                 int poff = offset & (PAGE_SIZE - 1);
 418                 int plen = PAGE_SIZE - poff;
 419
 420                 if (plen > len)
 421                         plen = len;
 422                 lnb->lnb_file_offset = offset;
 423                 lnb->lnb_page_offset = poff;
 424                 lnb->lnb_len = plen;
 425                 /* lnb->lnb_flags = rnb->rnb_flags; */
 426                 lnb->lnb_flags = 0;
 427                 lnb->lnb_page = NULL;
 428                 lnb->lnb_rc = 0;
 429
 430                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 431                          (long long) len);
 432                 offset += plen;
 433                 len -= plen;
 434                 lnb++;
 435                 (*nrpages)++;
 436         }
 437
 438         RETURN(0);
 439 }
 440
 441 static struct page *osd_get_page(struct dt_object *dt, loff_t offset,
 442                                  gfp_t gfp_mask)
 443 {
 444         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 445         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 446         struct page *page;
 447
 448         LASSERT(inode);
 449
 450         page = find_or_create_page(inode->i_mapping, offset >> PAGE_SHIFT,
 451                                    gfp_mask);
 452
 453         if (unlikely(page == NULL))
 454                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 455
 456         return page;
 457 }
 458
 459 /*
 460  * there are following "locks":
 461  * journal_start
 462  * i_mutex
 463  * page lock
 464  *
 465  * osd write path:
 466  *  - lock page(s)
 467  *  - journal_start
 468  *  - truncate_sem
 469  *
 470  * ext4 vmtruncate:
 471  *  - lock pages, unlock
 472  *  - journal_start
 473  *  - lock partial page
 474  *  - i_data_sem
 475  *
 476  */
 477
 478 /**
 479  * Unlock and release pages loaded by osd_bufs_get()
 480  *
 481  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 482  *
 483  * \param env           thread execution environment
 484  * \param dt            dt object undergoing IO (OSD object + methods)
 485  * \param lnb           array of pages undergoing IO
 486  * \param npages        number of pages in \a lnb
 487  *
 488  * \retval 0            always
 489  */
 490 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 491                         struct niobuf_local *lnb, int npages)
 492 {
 493         struct pagevec pvec;
 494         int i;
 495
 496         pagevec_init(&pvec, 0);
 497
 498         for (i = 0; i < npages; i++) {
 499                 if (lnb[i].lnb_page == NULL)
 500                         continue;
 501                 LASSERT(PageLocked(lnb[i].lnb_page));
 502                 unlock_page(lnb[i].lnb_page);
 503                 if (pagevec_add(&pvec, lnb[i].lnb_page) == 0)
 504                         pagevec_release(&pvec);
 505                 dt_object_put(env, dt);
 506                 lnb[i].lnb_page = NULL;
 507         }
 508
 509         /* Release any partial pagevec */
 510         pagevec_release(&pvec);
 511
 512         RETURN(0);
 513 }
 514
 515 /**
 516  * Load and lock pages undergoing IO
 517  *
 518  * Pages as described in the \a lnb array are fetched (from disk or cache)
 519  * and locked for IO by the caller.
 520  *
 521  * DLM locking protects us from write and truncate competing for same region,
 522  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 523  * It's possible the writeout on a such a page is in progress when we access
 524  * it. It's also possible that during this writeout we put new (partial) data
 525  * into the page, but won't be able to proceed in filter_commitrw_write().
 526  * Therefore, just wait for writeout completion as it should be rare enough.
 527  *
 528  * \param env           thread execution environment
 529  * \param dt            dt object undergoing IO (OSD object + methods)
 530  * \param pos           byte offset of IO start
 531  * \param len           number of bytes of IO
 532  * \param lnb           array of extents undergoing IO
 533  * \param rw            read or write operation, and other flags
 534  * \param capa          capabilities
 535  *
 536  * \retval pages        (zero or more) loaded successfully
 537  * \retval -ENOMEM      on memory/page allocation error
 538  */
 539 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 540                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 541                         enum dt_bufs_type rw)
 542 {
 543         struct osd_object *obj = osd_dt_obj(dt);
 544         int npages, i, rc = 0;
 545         gfp_t gfp_mask;
 546
 547         LASSERT(obj->oo_inode);
 548
 549         osd_map_remote_to_local(pos, len, &npages, lnb);
 550
 551         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 552         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 553                                              GFP_HIGHUSER;
 554         for (i = 0; i < npages; i++, lnb++) {
 555                 lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset,
 556                                              gfp_mask);
 557                 if (lnb->lnb_page == NULL)
 558                         GOTO(cleanup, rc = -ENOMEM);
 559
 560                 wait_on_page_writeback(lnb->lnb_page);
 561                 BUG_ON(PageWriteback(lnb->lnb_page));
 562
 563                 lu_object_get(&dt->do_lu);
 564         }
 565
 566         RETURN(i);
 567
 568 cleanup:
 569         if (i > 0)
 570                 osd_bufs_put(env, dt, lnb - i, i);
 571         return rc;
 572 }
 573
 574 #ifndef HAVE_LDISKFS_MAP_BLOCKS
 575
 576 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 577 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 578 #endif
 579
 580 struct bpointers {
 581         sector_t *blocks;
 582         unsigned long start;
 583         int num;
 584         int init_num;
 585         int create;
 586 };
 587
 588 static long ldiskfs_ext_find_goal(struct inode *inode,
 589                                   struct ldiskfs_ext_path *path,
 590                                   unsigned long block, int *aflags)
 591 {
 592         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 593         unsigned long bg_start;
 594         unsigned long colour;
 595         int depth;
 596
 597         if (path) {
 598                 struct ldiskfs_extent *ex;
 599                 depth = path->p_depth;
 600
 601                 /* try to predict block placement */
 602                 if ((ex = path[depth].p_ext))
 603                         return ldiskfs_ext_pblock(ex) +
 604                                 (block - le32_to_cpu(ex->ee_block));
 605
 606                 /* it looks index is empty
 607                  * try to find starting from index itself */
 608                 if (path[depth].p_bh)
 609                         return path[depth].p_bh->b_blocknr;
 610         }
 611
 612         /* OK. use inode's group */
 613         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 614                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 615         colour = (current->pid % 16) *
 616                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 617         return bg_start + colour + block;
 618 }
 619
 620 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 621                                 struct ldiskfs_ext_path *path,
 622                                 unsigned long block, unsigned long *count,
 623                                 int *err)
 624 {
 625         struct ldiskfs_allocation_request ar;
 626         unsigned long pblock;
 627         int aflags;
 628
 629         /* find neighbour allocated blocks */
 630         ar.lleft = block;
 631         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 632         if (*err)
 633                 return 0;
 634         ar.lright = block;
 635         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 636         if (*err)
 637                 return 0;
 638
 639         /* allocate new block */
 640         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 641         ar.inode = inode;
 642         ar.logical = block;
 643         ar.len = *count;
 644         ar.flags = LDISKFS_MB_HINT_DATA;
 645         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 646         *count = ar.len;
 647         return pblock;
 648 }
 649
 650 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 651                                      struct ldiskfs_ext_path *path,
 652                                      struct ldiskfs_ext_cache *cex,
 653 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 654                                      struct ldiskfs_extent *ex,
 655 #endif
 656                                      void *cbdata)
 657 {
 658         struct bpointers *bp = cbdata;
 659         struct ldiskfs_extent nex;
 660         unsigned long pblock = 0;
 661         unsigned long tgen;
 662         int err, i;
 663         unsigned long count;
 664         handle_t *handle;
 665
 666 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 667         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 668 #else
 669         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 670 #endif
 671                 err = EXT_CONTINUE;
 672                 goto map;
 673         }
 674
 675         if (bp->create == 0) {
 676                 i = 0;
 677                 if (cex->ec_block < bp->start)
 678                         i = bp->start - cex->ec_block;
 679                 if (i >= cex->ec_len)
 680                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 681                                         i, cex->ec_len);
 682                 for (; i < cex->ec_len && bp->num; i++) {
 683                         *(bp->blocks) = 0;
 684                         bp->blocks++;
 685                         bp->num--;
 686                         bp->start++;
 687                 }
 688
 689                 return EXT_CONTINUE;
 690         }
 691
 692         tgen = LDISKFS_I(inode)->i_ext_generation;
 693         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
 694
 695         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
 696                                    count + LDISKFS_ALLOC_NEEDED + 1);
 697         if (IS_ERR(handle)) {
 698                 return PTR_ERR(handle);
 699         }
 700
 701         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
 702                 /* the tree has changed. so path can be invalid at moment */
 703                 ldiskfs_journal_stop(handle);
 704                 return EXT_REPEAT;
 705         }
 706
 707         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
 708          * protected by i_data_sem as whole. so we patch it to store
 709          * generation to path and now verify the tree hasn't changed */
 710         down_write((&LDISKFS_I(inode)->i_data_sem));
 711
 712         /* validate extent, make sure the extent tree does not changed */
 713         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
 714                 /* cex is invalid, try again */
 715                 up_write(&LDISKFS_I(inode)->i_data_sem);
 716                 ldiskfs_journal_stop(handle);
 717                 return EXT_REPEAT;
 718         }
 719
 720         count = cex->ec_len;
 721         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
 722         if (!pblock)
 723                 goto out;
 724         BUG_ON(count > cex->ec_len);
 725
 726         /* insert new extent */
 727         nex.ee_block = cpu_to_le32(cex->ec_block);
 728         ldiskfs_ext_store_pblock(&nex, pblock);
 729         nex.ee_len = cpu_to_le16(count);
 730         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
 731         if (err) {
 732                 /* free data blocks we just allocated */
 733                 /* not a good idea to call discard here directly,
 734                  * but otherwise we'd need to call it every free() */
 735                 ldiskfs_discard_preallocations(inode);
 736 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
 737                 ldiskfs_free_blocks(handle, inode, NULL,
 738                                     ldiskfs_ext_pblock(&nex),
 739                                     le16_to_cpu(nex.ee_len), 0);
 740 #else
 741                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
 742                                     le16_to_cpu(nex.ee_len), 0);
 743 #endif
 744                 goto out;
 745         }
 746
 747         /*
 748          * Putting len of the actual extent we just inserted,
 749          * we are asking ldiskfs_ext_walk_space() to continue
 750          * scaning after that block
 751          */
 752         cex->ec_len = le16_to_cpu(nex.ee_len);
 753         cex->ec_start = ldiskfs_ext_pblock(&nex);
 754         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
 755         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
 756
 757 out:
 758         up_write((&LDISKFS_I(inode)->i_data_sem));
 759         ldiskfs_journal_stop(handle);
 760 map:
 761         if (err >= 0) {
 762                 /* map blocks */
 763                 if (bp->num == 0) {
 764                         CERROR("hmm. why do we find this extent?\n");
 765                         CERROR("initial space: %lu:%u\n",
 766                                 bp->start, bp->init_num);
 767 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 768                         CERROR("current extent: %u/%u/%llu %d\n",
 769                                 cex->ec_block, cex->ec_len,
 770                                 (unsigned long long)cex->ec_start,
 771                                 cex->ec_type);
 772 #else
 773                         CERROR("current extent: %u/%u/%llu\n",
 774                                 cex->ec_block, cex->ec_len,
 775                                 (unsigned long long)cex->ec_start);
 776 #endif
 777                 }
 778                 i = 0;
 779                 if (cex->ec_block < bp->start)
 780                         i = bp->start - cex->ec_block;
 781                 if (i >= cex->ec_len)
 782                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 783                                         i, cex->ec_len);
 784                 for (; i < cex->ec_len && bp->num; i++) {
 785                         *(bp->blocks) = cex->ec_start + i;
 786                         if (pblock != 0) {
 787                                 /* unmap any possible underlying metadata from
 788                                  * the block device mapping.  bug 6998. */
 789 #ifndef HAVE_CLEAN_BDEV_ALIASES
 790                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
 791                                                           *(bp->blocks));
 792 #else
 793                                 clean_bdev_aliases(inode->i_sb->s_bdev,
 794                                                    *(bp->blocks), 1);
 795 #endif
 796                         }
 797                         bp->blocks++;
 798                         bp->num--;
 799                         bp->start++;
 800                 }
 801         }
 802         return err;
 803 }
 804
 805 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
 806                                    int clen, sector_t *blocks, int create)
 807 {
 808         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 809         struct bpointers bp;
 810         int err;
 811
 812         if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
 813                 return -EFBIG;
 814
 815         bp.blocks = blocks;
 816         bp.start = index * blocks_per_page;
 817         bp.init_num = bp.num = clen * blocks_per_page;
 818         bp.create = create;
 819
 820         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 821                bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
 822
 823         err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
 824                                      ldiskfs_ext_new_extent_cb, &bp);
 825         ldiskfs_ext_invalidate_cache(inode);
 826
 827         return err;
 828 }
 829
 830 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
 831                                           struct page **page, int pages,
 832                                           sector_t *blocks, int create)
 833 {
 834         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 835         pgoff_t bitmap_max_page_index;
 836         sector_t *b;
 837         int rc = 0, i;
 838
 839         bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
 840                                 PAGE_SHIFT;
 841         for (i = 0, b = blocks; i < pages; i++, page++) {
 842                 if ((*page)->index + 1 >= bitmap_max_page_index) {
 843                         rc = -EFBIG;
 844                         break;
 845                 }
 846                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
 847                 if (rc) {
 848                         CERROR("ino %lu, blk %llu create %d: rc %d\n",
 849                                inode->i_ino,
 850                                (unsigned long long)*b, create, rc);
 851                         break;
 852                 }
 853                 b += blocks_per_page;
 854         }
 855         return rc;
 856 }
 857
 858 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
 859                                            struct page **page,
 860                                            int pages, sector_t *blocks,
 861                                            int create)
 862 {
 863         int rc = 0, i = 0, clen = 0;
 864         struct page *fp = NULL;
 865
 866         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 867                 inode->i_ino, pages, (*page)->index);
 868
 869         /* pages are sorted already. so, we just have to find
 870          * contig. space and process them properly */
 871         while (i < pages) {
 872                 if (fp == NULL) {
 873                         /* start new extent */
 874                         fp = *page++;
 875                         clen = 1;
 876                         i++;
 877                         continue;
 878                 } else if (fp->index + clen == (*page)->index) {
 879                         /* continue the extent */
 880                         page++;
 881                         clen++;
 882                         i++;
 883                         continue;
 884                 }
 885
 886                 /* process found extent */
 887                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 888                                              blocks, create);
 889                 if (rc)
 890                         GOTO(cleanup, rc);
 891
 892                 /* look for next extent */
 893                 fp = NULL;
 894                 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
 895         }
 896
 897         if (fp)
 898                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 899                                              blocks, create);
 900
 901 cleanup:
 902         return rc;
 903 }
 904
 905 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 906                                        int pages, sector_t *blocks,
 907                                        int create)
 908 {
 909         int rc;
 910
 911         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 912                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
 913                                                      blocks, create);
 914                 return rc;
 915         }
 916         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
 917
 918         return rc;
 919 }
 920 #else
 921 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 922                                        int pages, sector_t *blocks,
 923                                        int create)
 924 {
 925         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 926         int rc = 0, i = 0;
 927         struct page *fp = NULL;
 928         int clen = 0;
 929         pgoff_t max_page_index;
 930         handle_t *handle = NULL;
 931
 932         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
 933
 934         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 935                 inode->i_ino, pages, (*page)->index);
 936
 937         if (create) {
 938                 create = LDISKFS_GET_BLOCKS_CREATE;
 939                 handle = ldiskfs_journal_current_handle();
 940                 LASSERT(handle != NULL);
 941                 rc = osd_attach_jinode(inode);
 942                 if (rc)
 943                         return rc;
 944         }
 945         /* pages are sorted already. so, we just have to find
 946          * contig. space and process them properly */
 947         while (i < pages) {
 948                 long blen, total = 0;
 949                 struct ldiskfs_map_blocks map = { 0 };
 950
 951                 if (fp == NULL) { /* start new extent */
 952                         fp = *page++;
 953                         clen = 1;
 954                         if (++i != pages)
 955                                 continue;
 956                 } else if (fp->index + clen == (*page)->index) {
 957                         /* continue the extent */
 958                         page++;
 959                         clen++;
 960                         if (++i != pages)
 961                                 continue;
 962                 }
 963                 if (fp->index + clen >= max_page_index)
 964                         GOTO(cleanup, rc = -EFBIG);
 965                 /* process found extent */
 966                 map.m_lblk = fp->index * blocks_per_page;
 967                 map.m_len = blen = clen * blocks_per_page;
 968 cont_map:
 969                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 970                 if (rc >= 0) {
 971                         int c = 0;
 972                         for (; total < blen && c < map.m_len; c++, total++) {
 973                                 if (rc == 0) {
 974                                         *(blocks + total) = 0;
 975                                         total++;
 976                                         break;
 977                                 } else {
 978                                         *(blocks + total) = map.m_pblk + c;
 979                                         /* unmap any possible underlying
 980                                          * metadata from the block device
 981                                          * mapping.  bug 6998. */
 982                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
 983                                             create)
 984 #ifndef HAVE_CLEAN_BDEV_ALIASES
 985                                                 unmap_underlying_metadata(
 986                                                         inode->i_sb->s_bdev,
 987                                                         map.m_pblk + c);
 988 #else
 989                                                 clean_bdev_aliases(
 990                                                         inode->i_sb->s_bdev,
 991                                                         map.m_pblk + c, 1);
 992 #endif
 993                                 }
 994                         }
 995                         rc = 0;
 996                 }
 997                 if (rc == 0 && total < blen) {
 998                         map.m_lblk = fp->index * blocks_per_page + total;
 999                         map.m_len = blen - total;
1000                         goto cont_map;
1001                 }
1002                 if (rc != 0)
1003                         GOTO(cleanup, rc);
1004
1005                 /* look for next extent */
1006                 fp = NULL;
1007                 blocks += blocks_per_page * clen;
1008         }
1009 cleanup:
1010         return rc;
1011 }
1012 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
1013
1014 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1015                           struct niobuf_local *lnb, int npages)
1016 {
1017         struct osd_thread_info *oti   = osd_oti_get(env);
1018         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1019         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
1020         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
1021         ktime_t start;
1022         ktime_t end;
1023         s64 timediff;
1024         ssize_t                 isize;
1025         __s64                   maxidx;
1026         int                     rc = 0;
1027         int                     i;
1028         int                     cache = 0;
1029
1030         LASSERT(inode);
1031
1032         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1033         if (unlikely(rc != 0))
1034                 RETURN(rc);
1035
1036         isize = i_size_read(inode);
1037         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1038
1039         if (osd->od_writethrough_cache)
1040                 cache = 1;
1041         if (isize > osd->od_readcache_max_filesize)
1042                 cache = 0;
1043
1044         start = ktime_get();
1045         for (i = 0; i < npages; i++) {
1046
1047                 if (cache == 0)
1048                         generic_error_remove_page(inode->i_mapping,
1049                                                   lnb[i].lnb_page);
1050
1051                 /*
1052                  * till commit the content of the page is undefined
1053                  * we'll set it uptodate once bulk is done. otherwise
1054                  * subsequent reads can access non-stable data
1055                  */
1056                 ClearPageUptodate(lnb[i].lnb_page);
1057
1058                 if (lnb[i].lnb_len == PAGE_SIZE)
1059                         continue;
1060
1061                 if (maxidx >= lnb[i].lnb_page->index) {
1062                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1063                 } else {
1064                         long off;
1065                         char *p = kmap(lnb[i].lnb_page);
1066
1067                         off = lnb[i].lnb_page_offset;
1068                         if (off)
1069                                 memset(p, 0, off);
1070                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1071                               ~PAGE_MASK;
1072                         if (off)
1073                                 memset(p + off, 0, PAGE_SIZE - off);
1074                         kunmap(lnb[i].lnb_page);
1075                 }
1076         }
1077         end = ktime_get();
1078         timediff = ktime_us_delta(end, start);
1079         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1080
1081         if (iobuf->dr_npages) {
1082                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1083                                                  iobuf->dr_npages,
1084                                                  iobuf->dr_blocks, 0);
1085                 if (likely(rc == 0)) {
1086                         rc = osd_do_bio(osd, inode, iobuf);
1087                         /* do IO stats for preparation reads */
1088                         osd_fini_iobuf(osd, iobuf);
1089                 }
1090         }
1091         RETURN(rc);
1092 }
1093
1094 struct osd_fextent {
1095         sector_t        start;
1096         sector_t        end;
1097         unsigned int    mapped:1;
1098 };
1099
1100 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1101                          struct osd_fextent *cached_extent)
1102 {
1103         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1104         sector_t block = offset >> inode->i_blkbits;
1105         sector_t start;
1106         struct fiemap_extent_info fei = { 0 };
1107         struct fiemap_extent fe = { 0 };
1108         mm_segment_t saved_fs;
1109         int rc;
1110
1111         if (block >= cached_extent->start && block < cached_extent->end)
1112                 return cached_extent->mapped;
1113
1114         if (i_size_read(inode) == 0)
1115                 return 0;
1116
1117         /* Beyond EOF, must not be mapped */
1118         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1119                 return 0;
1120
1121         fei.fi_extents_max = 1;
1122         fei.fi_extents_start = &fe;
1123
1124         saved_fs = get_fs();
1125         set_fs(get_ds());
1126         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1127         set_fs(saved_fs);
1128         if (rc != 0)
1129                 return 0;
1130
1131         start = fe.fe_logical >> inode->i_blkbits;
1132
1133         if (start > block) {
1134                 cached_extent->start = block;
1135                 cached_extent->end = start;
1136                 cached_extent->mapped = 0;
1137         } else {
1138                 cached_extent->start = start;
1139                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1140                                       inode->i_blkbits;
1141                 cached_extent->mapped = 1;
1142         }
1143
1144         return cached_extent->mapped;
1145 }
1146
1147 static int osd_declare_write_commit(const struct lu_env *env,
1148                                     struct dt_object *dt,
1149                                     struct niobuf_local *lnb, int npages,
1150                                     struct thandle *handle)
1151 {
1152         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1153         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1154         struct osd_thandle      *oh;
1155         int                     extents = 1;
1156         int                     depth;
1157         int                     i;
1158         int                     newblocks;
1159         int                     rc = 0;
1160         int                     flags = 0;
1161         int                     credits = 0;
1162         long long               quota_space = 0;
1163         struct osd_fextent      extent = { 0 };
1164         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1165         ENTRY;
1166
1167         LASSERT(handle != NULL);
1168         oh = container_of0(handle, struct osd_thandle, ot_super);
1169         LASSERT(oh->ot_handle == NULL);
1170
1171         newblocks = npages;
1172
1173         /* calculate number of extents (probably better to pass nb) */
1174         for (i = 0; i < npages; i++) {
1175                 if (i && lnb[i].lnb_file_offset !=
1176                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1177                         extents++;
1178
1179                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1180                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1181                 else
1182                         quota_space += PAGE_SIZE;
1183
1184                 /* ignore quota for the whole request if any page is from
1185                  * client cache or written by root.
1186                  *
1187                  * XXX once we drop the 1.8 client support, the checking
1188                  * for whether page is from cache can be simplified as:
1189                  * !(lnb[i].flags & OBD_BRW_SYNC)
1190                  *
1191                  * XXX we could handle this on per-lnb basis as done by
1192                  * grant. */
1193                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1194                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1195                     OBD_BRW_FROM_GRANT)
1196                         declare_flags |= OSD_QID_FORCE;
1197         }
1198
1199         /*
1200          * each extent can go into new leaf causing a split
1201          * 5 is max tree depth: inode + 4 index blocks
1202          * with blockmaps, depth is 3 at most
1203          */
1204         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1205                 /*
1206                  * many concurrent threads may grow tree by the time
1207                  * our transaction starts. so, consider 2 is a min depth
1208                  */
1209                 depth = ext_depth(inode);
1210                 depth = max(depth, 1) + 1;
1211                 newblocks += depth;
1212                 credits++; /* inode */
1213                 credits += depth * 2 * extents;
1214         } else {
1215                 depth = 3;
1216                 newblocks += depth;
1217                 credits++; /* inode */
1218                 credits += depth * extents;
1219         }
1220
1221         /* quota space for metadata blocks */
1222         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1223
1224         /* quota space should be reported in 1K blocks */
1225         quota_space = toqb(quota_space);
1226
1227         /* each new block can go in different group (bitmap + gd) */
1228
1229         /* we can't dirty more bitmap blocks than exist */
1230         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1231                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1232         else
1233                 credits += newblocks;
1234
1235         /* we can't dirty more gd blocks than exist */
1236         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1237                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1238         else
1239                 credits += newblocks;
1240
1241         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1242
1243         /* make sure the over quota flags were not set */
1244         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1245
1246         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1247                                    i_projid_read(inode), quota_space, oh,
1248                                    osd_dt_obj(dt), &flags, declare_flags);
1249
1250         /* we need only to store the overquota flags in the first lnb for
1251          * now, once we support multiple objects BRW, this code needs be
1252          * revised. */
1253         if (flags & QUOTA_FL_OVER_USRQUOTA)
1254                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1255         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1256                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1257         if (flags & QUOTA_FL_OVER_PRJQUOTA)
1258                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1259
1260         RETURN(rc);
1261 }
1262
1263 /* Check if a block is allocated or not */
1264 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1265                             struct niobuf_local *lnb, int npages,
1266                             struct thandle *thandle)
1267 {
1268         struct osd_thread_info *oti = osd_oti_get(env);
1269         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1270         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1271         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1272         loff_t isize;
1273         int rc = 0, i;
1274
1275         LASSERT(inode);
1276
1277         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1278         if (unlikely(rc != 0))
1279                 RETURN(rc);
1280
1281         isize = i_size_read(inode);
1282         ll_vfs_dq_init(inode);
1283
1284         for (i = 0; i < npages; i++) {
1285                 if (lnb[i].lnb_rc == -ENOSPC &&
1286                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1287                         /* Allow the write to proceed if overwriting an
1288                          * existing block */
1289                         lnb[i].lnb_rc = 0;
1290                 }
1291
1292                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1293                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1294                                lnb[i].lnb_rc);
1295                         LASSERT(lnb[i].lnb_page);
1296                         generic_error_remove_page(inode->i_mapping,
1297                                                   lnb[i].lnb_page);
1298                         continue;
1299                 }
1300
1301                 LASSERT(PageLocked(lnb[i].lnb_page));
1302                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1303
1304                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1305                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1306
1307                 /*
1308                  * Since write and truncate are serialized by oo_sem, even
1309                  * partial-page truncate should not leave dirty pages in the
1310                  * page cache.
1311                  */
1312                 LASSERT(!PageDirty(lnb[i].lnb_page));
1313
1314                 SetPageUptodate(lnb[i].lnb_page);
1315
1316                 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1317         }
1318
1319         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1320
1321         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1322                 rc = -ENOSPC;
1323         } else if (iobuf->dr_npages > 0) {
1324                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1325                                                  iobuf->dr_npages,
1326                                                  iobuf->dr_blocks, 1);
1327         } else {
1328                 /* no pages to write, no transno is needed */
1329                 thandle->th_local = 1;
1330         }
1331
1332         if (likely(rc == 0)) {
1333                 spin_lock(&inode->i_lock);
1334                 if (isize > i_size_read(inode)) {
1335                         i_size_write(inode, isize);
1336                         LDISKFS_I(inode)->i_disksize = isize;
1337                         spin_unlock(&inode->i_lock);
1338                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1339                 } else {
1340                         spin_unlock(&inode->i_lock);
1341                 }
1342
1343                 rc = osd_do_bio(osd, inode, iobuf);
1344                 /* we don't do stats here as in read path because
1345                  * write is async: we'll do this in osd_put_bufs() */
1346         } else {
1347                 osd_fini_iobuf(osd, iobuf);
1348         }
1349
1350         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1351
1352         if (unlikely(rc != 0)) {
1353                 /* if write fails, we should drop pages from the cache */
1354                 for (i = 0; i < npages; i++) {
1355                         if (lnb[i].lnb_page == NULL)
1356                                 continue;
1357                         LASSERT(PageLocked(lnb[i].lnb_page));
1358                         generic_error_remove_page(inode->i_mapping,
1359                                                   lnb[i].lnb_page);
1360                 }
1361         }
1362
1363         RETURN(rc);
1364 }
1365
1366 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1367                          struct niobuf_local *lnb, int npages)
1368 {
1369         struct osd_thread_info *oti = osd_oti_get(env);
1370         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1371         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1372         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1373         int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1374         ktime_t start, end;
1375         s64 timediff;
1376         loff_t isize;
1377
1378         LASSERT(inode);
1379
1380         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1381         if (unlikely(rc != 0))
1382                 RETURN(rc);
1383
1384         isize = i_size_read(inode);
1385
1386         if (osd->od_read_cache)
1387                 cache = 1;
1388         if (isize > osd->od_readcache_max_filesize)
1389                 cache = 0;
1390
1391         start = ktime_get();
1392         for (i = 0; i < npages; i++) {
1393
1394                 if (isize <= lnb[i].lnb_file_offset)
1395                         /* If there's no more data, abort early.
1396                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1397                         break;
1398
1399                 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1400                         lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1401                 else
1402                         lnb[i].lnb_rc = lnb[i].lnb_len;
1403
1404                 /* Bypass disk read if fail_loc is set properly */
1405                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1406                         SetPageUptodate(lnb[i].lnb_page);
1407
1408                 if (PageUptodate(lnb[i].lnb_page)) {
1409                         cache_hits++;
1410                 } else {
1411                         cache_misses++;
1412                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1413                 }
1414
1415                 if (cache == 0)
1416                         generic_error_remove_page(inode->i_mapping,
1417                                                   lnb[i].lnb_page);
1418         }
1419         end = ktime_get();
1420         timediff = ktime_us_delta(end, start);
1421         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1422
1423         if (cache_hits != 0)
1424                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1425                                     cache_hits);
1426         if (cache_misses != 0)
1427                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1428                                     cache_misses);
1429         if (cache_hits + cache_misses != 0)
1430                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1431                                     cache_hits + cache_misses);
1432
1433         if (iobuf->dr_npages) {
1434                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1435                                                  iobuf->dr_npages,
1436                                                  iobuf->dr_blocks, 0);
1437                 rc = osd_do_bio(osd, inode, iobuf);
1438
1439                 /* IO stats will be done in osd_bufs_put() */
1440         }
1441
1442         RETURN(rc);
1443 }
1444
1445 /*
1446  * XXX: Another layering violation for now.
1447  *
1448  * We don't want to use ->f_op->read methods, because generic file write
1449  *
1450  *         - serializes on ->i_sem, and
1451  *
1452  *         - does a lot of extra work like balance_dirty_pages(),
1453  *
1454  * which doesn't work for globally shared files like /last_rcvd.
1455  */
1456 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1457 {
1458         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1459
1460         memcpy(buffer, (char *)ei->i_data, buflen);
1461
1462         return  buflen;
1463 }
1464
1465 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1466 {
1467         struct buffer_head *bh;
1468         unsigned long block;
1469         int osize;
1470         int blocksize;
1471         int csize;
1472         int boffs;
1473
1474         /* prevent reading after eof */
1475         spin_lock(&inode->i_lock);
1476         if (i_size_read(inode) < *offs + size) {
1477                 loff_t diff = i_size_read(inode) - *offs;
1478                 spin_unlock(&inode->i_lock);
1479                 if (diff < 0) {
1480                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1481                                i_size_read(inode), *offs);
1482                         return -EBADR;
1483                 } else if (diff == 0) {
1484                         return 0;
1485                 } else {
1486                         size = diff;
1487                 }
1488         } else {
1489                 spin_unlock(&inode->i_lock);
1490         }
1491
1492         blocksize = 1 << inode->i_blkbits;
1493         osize = size;
1494         while (size > 0) {
1495                 block = *offs >> inode->i_blkbits;
1496                 boffs = *offs & (blocksize - 1);
1497                 csize = min(blocksize - boffs, size);
1498                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1499                 if (IS_ERR(bh)) {
1500                         CERROR("%s: can't read %u@%llu on ino %lu: "
1501                                "rc = %ld\n", osd_ino2name(inode),
1502                                csize, *offs, inode->i_ino,
1503                                PTR_ERR(bh));
1504                         return PTR_ERR(bh);
1505                 }
1506
1507                 if (bh != NULL) {
1508                         memcpy(buf, bh->b_data + boffs, csize);
1509                         brelse(bh);
1510                 } else {
1511                         memset(buf, 0, csize);
1512                 }
1513
1514                 *offs += csize;
1515                 buf += csize;
1516                 size -= csize;
1517         }
1518         return osize;
1519 }
1520
1521 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1522                         struct lu_buf *buf, loff_t *pos)
1523 {
1524         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1525         int           rc;
1526
1527         /* Read small symlink from inode body as we need to maintain correct
1528          * on-disk symlinks for ldiskfs.
1529          */
1530         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1531             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1532                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1533         else
1534                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1535
1536         return rc;
1537 }
1538
1539 static inline int osd_extents_enabled(struct super_block *sb,
1540                                       struct inode *inode)
1541 {
1542         if (inode != NULL) {
1543                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1544                         return 1;
1545         } else if (ldiskfs_has_feature_extents(sb)) {
1546                 return 1;
1547         }
1548         return 0;
1549 }
1550
1551 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1552                            const loff_t size, const loff_t pos,
1553                            const int blocks)
1554 {
1555         int credits, bits, bs, i;
1556
1557         bits = sb->s_blocksize_bits;
1558         bs = 1 << bits;
1559
1560         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1561          * we do not expect blockmaps on the large files,
1562          * so let's shrink it to 2 levels (4GB files) */
1563
1564         /* this is default reservation: 2 levels */
1565         credits = (blocks + 2) * 3;
1566
1567         /* actual offset is unknown, hard to optimize */
1568         if (pos == -1)
1569                 return credits;
1570
1571         /* now check for few specific cases to optimize */
1572         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1573                 /* no indirects */
1574                 credits = blocks;
1575                 /* allocate if not allocated */
1576                 if (inode == NULL) {
1577                         credits += blocks * 2;
1578                         return credits;
1579                 }
1580                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1581                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1582                         if (LDISKFS_I(inode)->i_data[i] == 0)
1583                                 credits += 2;
1584                 }
1585         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1586                 /* single indirect */
1587                 credits = blocks * 3;
1588                 if (inode == NULL ||
1589                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1590                         credits += 3;
1591                 else
1592                         /* The indirect block may be modified. */
1593                         credits += 1;
1594         }
1595
1596         return credits;
1597 }
1598
1599 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1600                                  const struct lu_buf *buf, loff_t _pos,
1601                                  struct thandle *handle)
1602 {
1603         struct osd_object  *obj  = osd_dt_obj(dt);
1604         struct inode       *inode = obj->oo_inode;
1605         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1606         struct osd_thandle *oh;
1607         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1608         int                 bits, bs;
1609         int                 depth, size;
1610         loff_t              pos;
1611         ENTRY;
1612
1613         LASSERT(buf != NULL);
1614         LASSERT(handle != NULL);
1615
1616         oh = container_of0(handle, struct osd_thandle, ot_super);
1617         LASSERT(oh->ot_handle == NULL);
1618
1619         size = buf->lb_len;
1620         bits = sb->s_blocksize_bits;
1621         bs = 1 << bits;
1622
1623         if (_pos == -1) {
1624                 /* if this is an append, then we
1625                  * should expect cross-block record */
1626                 pos = 0;
1627         } else {
1628                 pos = _pos;
1629         }
1630
1631         /* blocks to modify */
1632         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1633         LASSERT(blocks > 0);
1634
1635         if (inode != NULL && _pos != -1) {
1636                 /* object size in blocks */
1637                 est = (i_size_read(inode) + bs - 1) >> bits;
1638                 allocated = inode->i_blocks >> (bits - 9);
1639                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1640                         /* looks like an overwrite, no need to modify tree */
1641                         credits = blocks;
1642                         /* no need to modify i_size */
1643                         goto out;
1644                 }
1645         }
1646
1647         if (osd_extents_enabled(sb, inode)) {
1648                 /*
1649                  * many concurrent threads may grow tree by the time
1650                  * our transaction starts. so, consider 2 is a min depth
1651                  * for every level we may need to allocate a new block
1652                  * and take some entries from the old one. so, 3 blocks
1653                  * to allocate (bitmap, gd, itself) + old block - 4 per
1654                  * level.
1655                  */
1656                 depth = inode != NULL ? ext_depth(inode) : 0;
1657                 depth = max(depth, 1) + 1;
1658                 credits = depth;
1659                 /* if not append, then split may need to modify
1660                  * existing blocks moving entries into the new ones */
1661                 if (_pos != -1)
1662                         credits += depth;
1663                 /* blocks to store data: bitmap,gd,itself */
1664                 credits += blocks * 3;
1665         } else {
1666                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1667         }
1668         /* if inode is created as part of the transaction,
1669          * then it's counted already by the creation method */
1670         if (inode != NULL)
1671                 credits++;
1672
1673 out:
1674
1675         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1676
1677         /* dt_declare_write() is usually called for system objects, such
1678          * as llog or last_rcvd files. We needn't enforce quota on those
1679          * objects, so always set the lqi_space as 0. */
1680         if (inode != NULL)
1681                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1682                                            i_gid_read(inode),
1683                                            i_projid_read(inode), 0,
1684                                            oh, obj, NULL, OSD_QID_BLK);
1685         RETURN(rc);
1686 }
1687
1688 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1689 {
1690         /* LU-2634: clear the extent format for fast symlink */
1691         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1692
1693         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1694         spin_lock(&inode->i_lock);
1695         LDISKFS_I(inode)->i_disksize = buflen;
1696         i_size_write(inode, buflen);
1697         spin_unlock(&inode->i_lock);
1698         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1699
1700         return 0;
1701 }
1702
1703 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1704                              int write_NUL, loff_t *offs, handle_t *handle)
1705 {
1706         struct buffer_head *bh        = NULL;
1707         loff_t              offset    = *offs;
1708         loff_t              new_size  = i_size_read(inode);
1709         unsigned long       block;
1710         int                 blocksize = 1 << inode->i_blkbits;
1711         int                 err = 0;
1712         int                 size;
1713         int                 boffs;
1714         int                 dirty_inode = 0;
1715
1716         if (write_NUL) {
1717                 /*
1718                  * long symlink write does not count the NUL terminator in
1719                  * bufsize, we write it, and the inode's file size does not
1720                  * count the NUL terminator as well.
1721                  */
1722                 ((char *)buf)[bufsize] = '\0';
1723                 ++bufsize;
1724         }
1725
1726         while (bufsize > 0) {
1727                 int credits = handle->h_buffer_credits;
1728
1729                 if (bh)
1730                         brelse(bh);
1731
1732                 block = offset >> inode->i_blkbits;
1733                 boffs = offset & (blocksize - 1);
1734                 size = min(blocksize - boffs, bufsize);
1735                 bh = __ldiskfs_bread(handle, inode, block, 1);
1736                 if (IS_ERR_OR_NULL(bh)) {
1737                         if (bh == NULL) {
1738                                 err = -EIO;
1739                         } else {
1740                                 err = PTR_ERR(bh);
1741                                 bh = NULL;
1742                         }
1743
1744                         CERROR("%s: error reading offset %llu (block %lu, "
1745                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
1746                                inode->i_sb->s_id, offset, block, bufsize, *offs,
1747                                credits, handle->h_buffer_credits, err);
1748                         break;
1749                 }
1750
1751                 err = ldiskfs_journal_get_write_access(handle, bh);
1752                 if (err) {
1753                         CERROR("journal_get_write_access() returned error %d\n",
1754                                err);
1755                         break;
1756                 }
1757                 LASSERTF(boffs + size <= bh->b_size,
1758                          "boffs %d size %d bh->b_size %lu\n",
1759                          boffs, size, (unsigned long)bh->b_size);
1760                 memcpy(bh->b_data + boffs, buf, size);
1761                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1762                 if (err)
1763                         break;
1764
1765                 if (offset + size > new_size)
1766                         new_size = offset + size;
1767                 offset += size;
1768                 bufsize -= size;
1769                 buf += size;
1770         }
1771         if (bh)
1772                 brelse(bh);
1773
1774         if (write_NUL)
1775                 --new_size;
1776         /* correct in-core and on-disk sizes */
1777         if (new_size > i_size_read(inode)) {
1778                 spin_lock(&inode->i_lock);
1779                 if (new_size > i_size_read(inode))
1780                         i_size_write(inode, new_size);
1781                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1782                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1783                         dirty_inode = 1;
1784                 }
1785                 spin_unlock(&inode->i_lock);
1786                 if (dirty_inode)
1787                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1788         }
1789
1790         if (err == 0)
1791                 *offs = offset;
1792         return err;
1793 }
1794
1795 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1796                          const struct lu_buf *buf, loff_t *pos,
1797                          struct thandle *handle, int ignore_quota)
1798 {
1799         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1800         struct osd_thandle      *oh;
1801         ssize_t                 result;
1802         int                     is_link;
1803
1804         LASSERT(dt_object_exists(dt));
1805
1806         LASSERT(handle != NULL);
1807         LASSERT(inode != NULL);
1808         ll_vfs_dq_init(inode);
1809
1810         /* XXX: don't check: one declared chunk can be used many times */
1811         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1812
1813         oh = container_of(handle, struct osd_thandle, ot_super);
1814         LASSERT(oh->ot_handle->h_transaction != NULL);
1815         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1816
1817         /* Write small symlink to inode body as we need to maintain correct
1818          * on-disk symlinks for ldiskfs.
1819          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1820          * does not count it in.
1821          */
1822         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1823         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1824                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1825         else
1826                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1827                                                   buf->lb_len, is_link, pos,
1828                                                   oh->ot_handle);
1829         if (result == 0)
1830                 result = buf->lb_len;
1831
1832         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1833
1834         return result;
1835 }
1836
1837 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1838                              __u64 start, __u64 end, struct thandle *th)
1839 {
1840         struct osd_thandle *oh;
1841         struct inode       *inode;
1842         int                 rc;
1843         ENTRY;
1844
1845         LASSERT(th);
1846         oh = container_of(th, struct osd_thandle, ot_super);
1847
1848         /*
1849          * we don't need to reserve credits for whole truncate
1850          * it's not possible as truncate may need to free too many
1851          * blocks and that won't fit a single transaction. instead
1852          * we reserve credits to change i_size and put inode onto
1853          * orphan list. if needed truncate will extend or restart
1854          * transaction
1855          */
1856         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1857                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1858
1859         inode = osd_dt_obj(dt)->oo_inode;
1860         LASSERT(inode);
1861
1862         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1863                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1864                                    NULL, OSD_QID_BLK);
1865         RETURN(rc);
1866 }
1867
1868 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1869                      __u64 start, __u64 end, struct thandle *th)
1870 {
1871         struct osd_thandle *oh;
1872         struct osd_object  *obj = osd_dt_obj(dt);
1873         struct inode       *inode = obj->oo_inode;
1874         handle_t           *h;
1875         tid_t               tid;
1876         int                rc = 0, rc2 = 0;
1877         ENTRY;
1878
1879         LASSERT(end == OBD_OBJECT_EOF);
1880         LASSERT(dt_object_exists(dt));
1881         LASSERT(osd_invariant(obj));
1882         LASSERT(inode != NULL);
1883         ll_vfs_dq_init(inode);
1884
1885         LASSERT(th);
1886         oh = container_of(th, struct osd_thandle, ot_super);
1887         LASSERT(oh->ot_handle->h_transaction != NULL);
1888
1889         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1890
1891         tid = oh->ot_handle->h_transaction->t_tid;
1892
1893         spin_lock(&inode->i_lock);
1894         i_size_write(inode, start);
1895         spin_unlock(&inode->i_lock);
1896         ll_truncate_pagecache(inode, start);
1897 #ifdef HAVE_INODEOPS_TRUNCATE
1898         if (inode->i_op->truncate) {
1899                 inode->i_op->truncate(inode);
1900         } else
1901 #endif
1902                 ldiskfs_truncate(inode);
1903
1904         /*
1905          * For a partial-page truncate, flush the page to disk immediately to
1906          * avoid data corruption during direct disk write.  b=17397
1907          */
1908         if ((start & ~PAGE_MASK) != 0)
1909                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1910
1911         h = journal_current_handle();
1912         LASSERT(h != NULL);
1913         LASSERT(h == oh->ot_handle);
1914
1915         /* do not check credits with osd_trans_exec_check() as the truncate
1916          * can restart the transaction internally and we restart the
1917          * transaction in this case */
1918
1919         if (tid != h->h_transaction->t_tid) {
1920                 int credits = oh->ot_credits;
1921                 /*
1922                  * transaction has changed during truncate
1923                  * we need to restart the handle with our credits
1924                  */
1925                 if (h->h_buffer_credits < credits) {
1926                         if (ldiskfs_journal_extend(h, credits))
1927                                 rc2 = ldiskfs_journal_restart(h, credits);
1928                 }
1929         }
1930
1931         RETURN(rc == 0 ? rc2 : rc);
1932 }
1933
1934 static int fiemap_check_ranges(struct inode *inode,
1935                                u64 start, u64 len, u64 *new_len)
1936 {
1937         loff_t maxbytes;
1938
1939         *new_len = len;
1940
1941         if (len == 0)
1942                 return -EINVAL;
1943
1944         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1945                 maxbytes = inode->i_sb->s_maxbytes;
1946         else
1947                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1948
1949         if (start > maxbytes)
1950                 return -EFBIG;
1951
1952         /*
1953          * Shrink request scope to what the fs can actually handle.
1954          */
1955         if (len > maxbytes || (maxbytes - len) < start)
1956                 *new_len = maxbytes - start;
1957
1958         return 0;
1959 }
1960
1961 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1962 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
1963
1964 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1965                           struct fiemap *fm)
1966 {
1967         struct fiemap_extent_info fieinfo = {0, };
1968         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1969         u64 len;
1970         int rc;
1971
1972
1973         LASSERT(inode);
1974         if (inode->i_op->fiemap == NULL)
1975                 return -EOPNOTSUPP;
1976
1977         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1978                 return -EINVAL;
1979
1980         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1981         if (rc)
1982                 return rc;
1983
1984         fieinfo.fi_flags = fm->fm_flags;
1985         fieinfo.fi_extents_max = fm->fm_extent_count;
1986         fieinfo.fi_extents_start = fm->fm_extents;
1987
1988         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1989                 filemap_write_and_wait(inode->i_mapping);
1990
1991         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
1992         fm->fm_flags = fieinfo.fi_flags;
1993         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
1994
1995         return rc;
1996 }
1997
1998 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1999                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2000 {
2001         int              rc = 0;
2002         struct inode    *inode = osd_dt_obj(dt)->oo_inode;
2003         ENTRY;
2004
2005         switch (advice) {
2006         case LU_LADVISE_DONTNEED:
2007                 if (end == 0)
2008                         break;
2009                 invalidate_mapping_pages(inode->i_mapping,
2010                                          start >> PAGE_CACHE_SHIFT,
2011                                          (end - 1) >> PAGE_CACHE_SHIFT);
2012                 break;
2013         default:
2014                 rc = -ENOTSUPP;
2015                 break;
2016         }
2017
2018         RETURN(rc);
2019 }
2020
2021 /*
2022  * in some cases we may need declare methods for objects being created
2023  * e.g., when we create symlink
2024  */
2025 const struct dt_body_operations osd_body_ops_new = {
2026         .dbo_declare_write = osd_declare_write,
2027 };
2028
2029 const struct dt_body_operations osd_body_ops = {
2030         .dbo_read                       = osd_read,
2031         .dbo_declare_write              = osd_declare_write,
2032         .dbo_write                      = osd_write,
2033         .dbo_bufs_get                   = osd_bufs_get,
2034         .dbo_bufs_put                   = osd_bufs_put,
2035         .dbo_write_prep                 = osd_write_prep,
2036         .dbo_declare_write_commit       = osd_declare_write_commit,
2037         .dbo_write_commit               = osd_write_commit,
2038         .dbo_read_prep                  = osd_read_prep,
2039         .dbo_declare_punch              = osd_declare_punch,
2040         .dbo_punch                      = osd_punch,
2041         .dbo_fiemap_get                 = osd_fiemap_get,
2042         .dbo_ladvise                    = osd_ladvise,
2043 };