lustre/osd-ldiskfs/osd_io.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/osd/osd_io.c
  33  *
  34  * body operations
  35  *
  36  * Author: Nikita Danilov <nikita@clusterfs.com>
  37  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  38  *
  39  */
  40
  41 /* LUSTRE_VERSION_CODE */
  42 #include <lustre_ver.h>
  43 /* prerequisite for linux/xattr.h */
  44 #include <linux/types.h>
  45 /* prerequisite for linux/xattr.h */
  46 #include <linux/fs.h>
  47
  48 /*
  49  * struct OBD_{ALLOC,FREE}*()
  50  * OBD_FAIL_CHECK
  51  */
  52 #include <obd_support.h>
  53
  54 #include "osd_internal.h"
  55
  56 /* ext_depth() */
  57 #include <ldiskfs/ldiskfs_extents.h>
  58
  59 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  60                             int rw, int line, int pages)
  61 {
  62         int blocks, i;
  63
  64         LASSERTF(iobuf->dr_elapsed_valid == 0,
  65                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  66                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  67                  iobuf->dr_init_at);
  68         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  69
  70         init_waitqueue_head(&iobuf->dr_wait);
  71         atomic_set(&iobuf->dr_numreqs, 0);
  72         iobuf->dr_npages = 0;
  73         iobuf->dr_error = 0;
  74         iobuf->dr_dev = d;
  75         iobuf->dr_frags = 0;
  76         iobuf->dr_elapsed = 0;
  77         /* must be counted before, so assert */
  78         iobuf->dr_rw = rw;
  79         iobuf->dr_init_at = line;
  80
  81         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  82         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  83                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  84                         pages * sizeof(iobuf->dr_pages[0]));
  85                 return 0;
  86         }
  87
  88         /* start with 1MB for 4K blocks */
  89         i = 256;
  90         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
  91                 i <<= 1;
  92
  93         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
  94                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
  95         pages = i;
  96         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  97         iobuf->dr_max_pages = 0;
  98         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
  99                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 100
 101         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 102         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 103         if (unlikely(iobuf->dr_blocks == NULL))
 104                 return -ENOMEM;
 105
 106         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 107         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 108         if (unlikely(iobuf->dr_pages == NULL))
 109                 return -ENOMEM;
 110
 111         iobuf->dr_max_pages = pages;
 112
 113         return 0;
 114 }
 115 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 116         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 117
 118 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
 119 {
 120         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 121         iobuf->dr_pages[iobuf->dr_npages++] = page;
 122 }
 123
 124 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 125 {
 126         int rw = iobuf->dr_rw;
 127
 128         if (iobuf->dr_elapsed_valid) {
 129                 iobuf->dr_elapsed_valid = 0;
 130                 LASSERT(iobuf->dr_dev == d);
 131                 LASSERT(iobuf->dr_frags > 0);
 132                 lprocfs_oh_tally(&d->od_brw_stats.
 133                                  hist[BRW_R_DIO_FRAGS+rw],
 134                                  iobuf->dr_frags);
 135                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 136                                       iobuf->dr_elapsed);
 137         }
 138 }
 139
 140 #ifndef REQ_WRITE /* pre-2.6.35 */
 141 #define __REQ_WRITE BIO_RW
 142 #endif
 143
 144 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 145 static void dio_complete_routine(struct bio *bio)
 146 {
 147         int error = bio->bi_error;
 148 #else
 149 static void dio_complete_routine(struct bio *bio, int error)
 150 {
 151 #endif
 152         struct osd_iobuf *iobuf = bio->bi_private;
 153         int iter;
 154         struct bio_vec *bvl;
 155
 156         /* CAVEAT EMPTOR: possibly in IRQ context
 157          * DO NOT record procfs stats here!!! */
 158
 159         if (unlikely(iobuf == NULL)) {
 160                 CERROR("***** bio->bi_private is NULL!  This should never "
 161                        "happen.  Normally, I would crash here, but instead I "
 162                        "will dump the bio contents to the console.  Please "
 163                        "report this to <https://jira.hpdd.intel.com/> , along "
 164                        "with any interesting messages leading up to this point "
 165                        "(like SCSI errors, perhaps).  Because bi_private is "
 166                        "NULL, I can't wake up the thread that initiated this "
 167                        "IO - you will probably have to reboot this node.\n");
 168                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 169                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 170                        "bi_private: %p\n", bio->bi_next,
 171                         (unsigned long)bio->bi_flags,
 172                         bio->bi_rw, bio->bi_vcnt, bio_idx(bio),
 173                         bio_sectors(bio) << 9, bio->bi_end_io,
 174 #ifdef HAVE_BI_CNT
 175                         atomic_read(&bio->bi_cnt),
 176 #else
 177                         atomic_read(&bio->__bi_cnt),
 178 #endif
 179                         bio->bi_private);
 180                 return;
 181         }
 182
 183         /* the check is outside of the cycle for performance reason -bzzz */
 184         if (!test_bit(__REQ_WRITE, &bio->bi_rw)) {
 185                 bio_for_each_segment_all(bvl, bio, iter) {
 186                         if (likely(error == 0))
 187                                 SetPageUptodate(bvl_to_page(bvl));
 188                         LASSERT(PageLocked(bvl_to_page(bvl)));
 189                 }
 190                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 191         } else {
 192                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 193         }
 194
 195         /* any real error is good enough -bzzz */
 196         if (error != 0 && iobuf->dr_error == 0)
 197                 iobuf->dr_error = error;
 198
 199         /*
 200          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 201          * it's possible that service thread will see dr_numreqs
 202          * is zero, but dr_elapsed is not set yet, leading to lost
 203          * data in this processing and an assertion in a subsequent
 204          * call to OSD.
 205          */
 206         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 207                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
 208                 iobuf->dr_elapsed_valid = 1;
 209         }
 210         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 211                 wake_up(&iobuf->dr_wait);
 212
 213         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 214          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 215          * mempool when serious on-disk fragmentation was encountered,
 216          * deadlocking the OST.  The bios are now released as soon as complete
 217          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 218         bio_put(bio);
 219 }
 220
 221 static void record_start_io(struct osd_iobuf *iobuf, int size)
 222 {
 223         struct osd_device    *osd = iobuf->dr_dev;
 224         struct obd_histogram *h = osd->od_brw_stats.hist;
 225
 226         iobuf->dr_frags++;
 227         atomic_inc(&iobuf->dr_numreqs);
 228
 229         if (iobuf->dr_rw == 0) {
 230                 atomic_inc(&osd->od_r_in_flight);
 231                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 232                                  atomic_read(&osd->od_r_in_flight));
 233                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 234         } else if (iobuf->dr_rw == 1) {
 235                 atomic_inc(&osd->od_w_in_flight);
 236                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 237                                  atomic_read(&osd->od_w_in_flight));
 238                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 239         } else {
 240                 LBUG();
 241         }
 242 }
 243
 244 static void osd_submit_bio(int rw, struct bio *bio)
 245 {
 246         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 247         if (rw == 0)
 248                 submit_bio(READ, bio);
 249         else
 250                 submit_bio(WRITE, bio);
 251 }
 252
 253 static int can_be_merged(struct bio *bio, sector_t sector)
 254 {
 255         if (bio == NULL)
 256                 return 0;
 257
 258         return bio_end_sector(bio) == sector ? 1 : 0;
 259 }
 260
 261 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 262                       struct osd_iobuf *iobuf)
 263 {
 264         int            blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 265         struct page  **pages = iobuf->dr_pages;
 266         int            npages = iobuf->dr_npages;
 267         sector_t      *blocks = iobuf->dr_blocks;
 268         int            total_blocks = npages * blocks_per_page;
 269         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 270         unsigned int   blocksize = inode->i_sb->s_blocksize;
 271         struct bio    *bio = NULL;
 272         struct page   *page;
 273         unsigned int   page_offset;
 274         sector_t       sector;
 275         int            nblocks;
 276         int            block_idx;
 277         int            page_idx;
 278         int            i;
 279         int            rc = 0;
 280         DECLARE_PLUG(plug);
 281         ENTRY;
 282
 283         LASSERT(iobuf->dr_npages == npages);
 284
 285         osd_brw_stats_update(osd, iobuf);
 286         iobuf->dr_start_time = cfs_time_current();
 287
 288         blk_start_plug(&plug);
 289         for (page_idx = 0, block_idx = 0;
 290              page_idx < npages;
 291              page_idx++, block_idx += blocks_per_page) {
 292
 293                 page = pages[page_idx];
 294                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 295
 296                 for (i = 0, page_offset = 0;
 297                      i < blocks_per_page;
 298                      i += nblocks, page_offset += blocksize * nblocks) {
 299
 300                         nblocks = 1;
 301
 302                         if (blocks[block_idx + i] == 0) {  /* hole */
 303                                 LASSERTF(iobuf->dr_rw == 0,
 304                                          "page_idx %u, block_idx %u, i %u\n",
 305                                          page_idx, block_idx, i);
 306                                 memset(kmap(page) + page_offset, 0, blocksize);
 307                                 kunmap(page);
 308                                 continue;
 309                         }
 310
 311                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 312
 313                         /* Additional contiguous file blocks? */
 314                         while (i + nblocks < blocks_per_page &&
 315                                (sector + (nblocks << sector_bits)) ==
 316                                ((sector_t)blocks[block_idx + i + nblocks] <<
 317                                 sector_bits))
 318                                 nblocks++;
 319
 320                         if (bio != NULL &&
 321                             can_be_merged(bio, sector) &&
 322                             bio_add_page(bio, page,
 323                                          blocksize * nblocks, page_offset) != 0)
 324                                 continue;       /* added this frag OK */
 325
 326                         if (bio != NULL) {
 327                                 struct request_queue *q =
 328                                         bdev_get_queue(bio->bi_bdev);
 329                                 unsigned int bi_size = bio_sectors(bio) << 9;
 330
 331                                 /* Dang! I have to fragment this I/O */
 332                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 333                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 334                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 335                                        bio_sectors(bio),
 336                                        queue_max_sectors(q),
 337                                        bio_phys_segments(q, bio),
 338                                        queue_max_phys_segments(q),
 339                                        0, queue_max_hw_segments(q));
 340                                 record_start_io(iobuf, bi_size);
 341                                 osd_submit_bio(iobuf->dr_rw, bio);
 342                         }
 343
 344                         /* allocate new bio */
 345                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 346                                                       (npages - page_idx) *
 347                                                       blocks_per_page));
 348                         if (bio == NULL) {
 349                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 350                                        (npages - page_idx), blocks_per_page,
 351                                        (npages - page_idx) * blocks_per_page);
 352                                 rc = -ENOMEM;
 353                                 goto out;
 354                         }
 355
 356                         bio->bi_bdev = inode->i_sb->s_bdev;
 357                         bio_set_sector(bio, sector);
 358                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 359                         bio->bi_end_io = dio_complete_routine;
 360                         bio->bi_private = iobuf;
 361
 362                         rc = bio_add_page(bio, page,
 363                                           blocksize * nblocks, page_offset);
 364                         LASSERT(rc != 0);
 365                 }
 366         }
 367
 368         if (bio != NULL) {
 369                 record_start_io(iobuf, bio_sectors(bio) << 9);
 370                 osd_submit_bio(iobuf->dr_rw, bio);
 371                 rc = 0;
 372         }
 373
 374 out:
 375         blk_finish_plug(&plug);
 376
 377         /* in order to achieve better IO throughput, we don't wait for writes
 378          * completion here. instead we proceed with transaction commit in
 379          * parallel and wait for IO completion once transaction is stopped
 380          * see osd_trans_stop() for more details -bzzz */
 381         if (iobuf->dr_rw == 0) {
 382                 wait_event(iobuf->dr_wait,
 383                            atomic_read(&iobuf->dr_numreqs) == 0);
 384                 osd_fini_iobuf(osd, iobuf);
 385         }
 386
 387         if (rc == 0)
 388                 rc = iobuf->dr_error;
 389         RETURN(rc);
 390 }
 391
 392 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 393                                    struct niobuf_local *lnb)
 394 {
 395         ENTRY;
 396
 397         *nrpages = 0;
 398
 399         while (len > 0) {
 400                 int poff = offset & (PAGE_SIZE - 1);
 401                 int plen = PAGE_SIZE - poff;
 402
 403                 if (plen > len)
 404                         plen = len;
 405                 lnb->lnb_file_offset = offset;
 406                 lnb->lnb_page_offset = poff;
 407                 lnb->lnb_len = plen;
 408                 /* lnb->lnb_flags = rnb->rnb_flags; */
 409                 lnb->lnb_flags = 0;
 410                 lnb->lnb_page = NULL;
 411                 lnb->lnb_rc = 0;
 412
 413                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 414                          (long long) len);
 415                 offset += plen;
 416                 len -= plen;
 417                 lnb++;
 418                 (*nrpages)++;
 419         }
 420
 421         RETURN(0);
 422 }
 423
 424 static struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
 425 {
 426         struct inode      *inode = osd_dt_obj(dt)->oo_inode;
 427         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 428         struct page       *page;
 429
 430         LASSERT(inode);
 431
 432         page = find_or_create_page(inode->i_mapping, offset >> PAGE_SHIFT,
 433                                    GFP_NOFS | __GFP_HIGHMEM);
 434         if (unlikely(page == NULL))
 435                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 436
 437         return page;
 438 }
 439
 440 /*
 441  * there are following "locks":
 442  * journal_start
 443  * i_mutex
 444  * page lock
 445  *
 446  * osd write path:
 447  *  - lock page(s)
 448  *  - journal_start
 449  *  - truncate_sem
 450  *
 451  * ext4 vmtruncate:
 452  *  - lock pages, unlock
 453  *  - journal_start
 454  *  - lock partial page
 455  *  - i_data_sem
 456  *
 457  */
 458
 459 /**
 460  * Unlock and release pages loaded by osd_bufs_get()
 461  *
 462  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 463  *
 464  * \param env           thread execution environment
 465  * \param dt            dt object undergoing IO (OSD object + methods)
 466  * \param lnb           array of pages undergoing IO
 467  * \param npages        number of pages in \a lnb
 468  *
 469  * \retval 0            always
 470  */
 471 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 472                         struct niobuf_local *lnb, int npages)
 473 {
 474         int i;
 475
 476         for (i = 0; i < npages; i++) {
 477                 if (lnb[i].lnb_page == NULL)
 478                         continue;
 479                 LASSERT(PageLocked(lnb[i].lnb_page));
 480                 unlock_page(lnb[i].lnb_page);
 481                 put_page(lnb[i].lnb_page);
 482                 dt_object_put(env, dt);
 483                 lnb[i].lnb_page = NULL;
 484         }
 485
 486         RETURN(0);
 487 }
 488
 489 /**
 490  * Load and lock pages undergoing IO
 491  *
 492  * Pages as described in the \a lnb array are fetched (from disk or cache)
 493  * and locked for IO by the caller.
 494  *
 495  * DLM locking protects us from write and truncate competing for same region,
 496  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 497  * It's possible the writeout on a such a page is in progress when we access
 498  * it. It's also possible that during this writeout we put new (partial) data
 499  * into the page, but won't be able to proceed in filter_commitrw_write().
 500  * Therefore, just wait for writeout completion as it should be rare enough.
 501  *
 502  * \param env           thread execution environment
 503  * \param dt            dt object undergoing IO (OSD object + methods)
 504  * \param pos           byte offset of IO start
 505  * \param len           number of bytes of IO
 506  * \param lnb           array of extents undergoing IO
 507  * \param rw            read or write operation?
 508  * \param capa          capabilities
 509  *
 510  * \retval pages        (zero or more) loaded successfully
 511  * \retval -ENOMEM      on memory/page allocation error
 512  */
 513 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 514                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 515                         int rw)
 516 {
 517         struct osd_object   *obj    = osd_dt_obj(dt);
 518         int npages, i, rc = 0;
 519
 520         LASSERT(obj->oo_inode);
 521
 522         osd_map_remote_to_local(pos, len, &npages, lnb);
 523
 524         for (i = 0; i < npages; i++, lnb++) {
 525                 lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset, rw);
 526                 if (lnb->lnb_page == NULL)
 527                         GOTO(cleanup, rc = -ENOMEM);
 528
 529                 wait_on_page_writeback(lnb->lnb_page);
 530                 BUG_ON(PageWriteback(lnb->lnb_page));
 531
 532                 lu_object_get(&dt->do_lu);
 533         }
 534
 535         RETURN(i);
 536
 537 cleanup:
 538         if (i > 0)
 539                 osd_bufs_put(env, dt, lnb - i, i);
 540         return rc;
 541 }
 542
 543 #ifndef HAVE_LDISKFS_MAP_BLOCKS
 544
 545 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 546 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 547 #endif
 548
 549 struct bpointers {
 550         sector_t *blocks;
 551         unsigned long start;
 552         int num;
 553         int init_num;
 554         int create;
 555 };
 556
 557 static long ldiskfs_ext_find_goal(struct inode *inode,
 558                                   struct ldiskfs_ext_path *path,
 559                                   unsigned long block, int *aflags)
 560 {
 561         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 562         unsigned long bg_start;
 563         unsigned long colour;
 564         int depth;
 565
 566         if (path) {
 567                 struct ldiskfs_extent *ex;
 568                 depth = path->p_depth;
 569
 570                 /* try to predict block placement */
 571                 if ((ex = path[depth].p_ext))
 572                         return ldiskfs_ext_pblock(ex) +
 573                                 (block - le32_to_cpu(ex->ee_block));
 574
 575                 /* it looks index is empty
 576                  * try to find starting from index itself */
 577                 if (path[depth].p_bh)
 578                         return path[depth].p_bh->b_blocknr;
 579         }
 580
 581         /* OK. use inode's group */
 582         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 583                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 584         colour = (current->pid % 16) *
 585                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 586         return bg_start + colour + block;
 587 }
 588
 589 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 590                                 struct ldiskfs_ext_path *path,
 591                                 unsigned long block, unsigned long *count,
 592                                 int *err)
 593 {
 594         struct ldiskfs_allocation_request ar;
 595         unsigned long pblock;
 596         int aflags;
 597
 598         /* find neighbour allocated blocks */
 599         ar.lleft = block;
 600         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 601         if (*err)
 602                 return 0;
 603         ar.lright = block;
 604         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 605         if (*err)
 606                 return 0;
 607
 608         /* allocate new block */
 609         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 610         ar.inode = inode;
 611         ar.logical = block;
 612         ar.len = *count;
 613         ar.flags = LDISKFS_MB_HINT_DATA;
 614         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 615         *count = ar.len;
 616         return pblock;
 617 }
 618
 619 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 620                                      struct ldiskfs_ext_path *path,
 621                                      struct ldiskfs_ext_cache *cex,
 622 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 623                                      struct ldiskfs_extent *ex,
 624 #endif
 625                                      void *cbdata)
 626 {
 627         struct bpointers *bp = cbdata;
 628         struct ldiskfs_extent nex;
 629         unsigned long pblock = 0;
 630         unsigned long tgen;
 631         int err, i;
 632         unsigned long count;
 633         handle_t *handle;
 634
 635 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 636         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 637 #else
 638         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 639 #endif
 640                 err = EXT_CONTINUE;
 641                 goto map;
 642         }
 643
 644         if (bp->create == 0) {
 645                 i = 0;
 646                 if (cex->ec_block < bp->start)
 647                         i = bp->start - cex->ec_block;
 648                 if (i >= cex->ec_len)
 649                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 650                                         i, cex->ec_len);
 651                 for (; i < cex->ec_len && bp->num; i++) {
 652                         *(bp->blocks) = 0;
 653                         bp->blocks++;
 654                         bp->num--;
 655                         bp->start++;
 656                 }
 657
 658                 return EXT_CONTINUE;
 659         }
 660
 661         tgen = LDISKFS_I(inode)->i_ext_generation;
 662         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
 663
 664         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
 665                                    count + LDISKFS_ALLOC_NEEDED + 1);
 666         if (IS_ERR(handle)) {
 667                 return PTR_ERR(handle);
 668         }
 669
 670         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
 671                 /* the tree has changed. so path can be invalid at moment */
 672                 ldiskfs_journal_stop(handle);
 673                 return EXT_REPEAT;
 674         }
 675
 676         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
 677          * protected by i_data_sem as whole. so we patch it to store
 678          * generation to path and now verify the tree hasn't changed */
 679         down_write((&LDISKFS_I(inode)->i_data_sem));
 680
 681         /* validate extent, make sure the extent tree does not changed */
 682         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
 683                 /* cex is invalid, try again */
 684                 up_write(&LDISKFS_I(inode)->i_data_sem);
 685                 ldiskfs_journal_stop(handle);
 686                 return EXT_REPEAT;
 687         }
 688
 689         count = cex->ec_len;
 690         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
 691         if (!pblock)
 692                 goto out;
 693         BUG_ON(count > cex->ec_len);
 694
 695         /* insert new extent */
 696         nex.ee_block = cpu_to_le32(cex->ec_block);
 697         ldiskfs_ext_store_pblock(&nex, pblock);
 698         nex.ee_len = cpu_to_le16(count);
 699         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
 700         if (err) {
 701                 /* free data blocks we just allocated */
 702                 /* not a good idea to call discard here directly,
 703                  * but otherwise we'd need to call it every free() */
 704                 ldiskfs_discard_preallocations(inode);
 705 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
 706                 ldiskfs_free_blocks(handle, inode, NULL,
 707                                     ldiskfs_ext_pblock(&nex),
 708                                     le16_to_cpu(nex.ee_len), 0);
 709 #else
 710                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
 711                                     le16_to_cpu(nex.ee_len), 0);
 712 #endif
 713                 goto out;
 714         }
 715
 716         /*
 717          * Putting len of the actual extent we just inserted,
 718          * we are asking ldiskfs_ext_walk_space() to continue
 719          * scaning after that block
 720          */
 721         cex->ec_len = le16_to_cpu(nex.ee_len);
 722         cex->ec_start = ldiskfs_ext_pblock(&nex);
 723         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
 724         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
 725
 726 out:
 727         up_write((&LDISKFS_I(inode)->i_data_sem));
 728         ldiskfs_journal_stop(handle);
 729 map:
 730         if (err >= 0) {
 731                 /* map blocks */
 732                 if (bp->num == 0) {
 733                         CERROR("hmm. why do we find this extent?\n");
 734                         CERROR("initial space: %lu:%u\n",
 735                                 bp->start, bp->init_num);
 736 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 737                         CERROR("current extent: %u/%u/%llu %d\n",
 738                                 cex->ec_block, cex->ec_len,
 739                                 (unsigned long long)cex->ec_start,
 740                                 cex->ec_type);
 741 #else
 742                         CERROR("current extent: %u/%u/%llu\n",
 743                                 cex->ec_block, cex->ec_len,
 744                                 (unsigned long long)cex->ec_start);
 745 #endif
 746                 }
 747                 i = 0;
 748                 if (cex->ec_block < bp->start)
 749                         i = bp->start - cex->ec_block;
 750                 if (i >= cex->ec_len)
 751                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 752                                         i, cex->ec_len);
 753                 for (; i < cex->ec_len && bp->num; i++) {
 754                         *(bp->blocks) = cex->ec_start + i;
 755                         if (pblock != 0) {
 756                                 /* unmap any possible underlying metadata from
 757                                  * the block device mapping.  bug 6998. */
 758                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
 759                                                           *(bp->blocks));
 760                         }
 761                         bp->blocks++;
 762                         bp->num--;
 763                         bp->start++;
 764                 }
 765         }
 766         return err;
 767 }
 768
 769 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
 770                                    int clen, sector_t *blocks, int create)
 771 {
 772         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 773         struct bpointers bp;
 774         int err;
 775
 776         if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
 777                 return -EFBIG;
 778
 779         bp.blocks = blocks;
 780         bp.start = index * blocks_per_page;
 781         bp.init_num = bp.num = clen * blocks_per_page;
 782         bp.create = create;
 783
 784         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 785                bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
 786
 787         err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
 788                                      ldiskfs_ext_new_extent_cb, &bp);
 789         ldiskfs_ext_invalidate_cache(inode);
 790
 791         return err;
 792 }
 793
 794 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
 795                                           struct page **page, int pages,
 796                                           sector_t *blocks, int create)
 797 {
 798         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 799         pgoff_t bitmap_max_page_index;
 800         sector_t *b;
 801         int rc = 0, i;
 802
 803         bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
 804                                 PAGE_SHIFT;
 805         for (i = 0, b = blocks; i < pages; i++, page++) {
 806                 if ((*page)->index + 1 >= bitmap_max_page_index) {
 807                         rc = -EFBIG;
 808                         break;
 809                 }
 810                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
 811                 if (rc) {
 812                         CERROR("ino %lu, blk %llu create %d: rc %d\n",
 813                                inode->i_ino,
 814                                (unsigned long long)*b, create, rc);
 815                         break;
 816                 }
 817                 b += blocks_per_page;
 818         }
 819         return rc;
 820 }
 821
 822 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
 823                                            struct page **page,
 824                                            int pages, sector_t *blocks,
 825                                            int create)
 826 {
 827         int rc = 0, i = 0, clen = 0;
 828         struct page *fp = NULL;
 829
 830         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 831                 inode->i_ino, pages, (*page)->index);
 832
 833         /* pages are sorted already. so, we just have to find
 834          * contig. space and process them properly */
 835         while (i < pages) {
 836                 if (fp == NULL) {
 837                         /* start new extent */
 838                         fp = *page++;
 839                         clen = 1;
 840                         i++;
 841                         continue;
 842                 } else if (fp->index + clen == (*page)->index) {
 843                         /* continue the extent */
 844                         page++;
 845                         clen++;
 846                         i++;
 847                         continue;
 848                 }
 849
 850                 /* process found extent */
 851                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 852                                              blocks, create);
 853                 if (rc)
 854                         GOTO(cleanup, rc);
 855
 856                 /* look for next extent */
 857                 fp = NULL;
 858                 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
 859         }
 860
 861         if (fp)
 862                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 863                                              blocks, create);
 864
 865 cleanup:
 866         return rc;
 867 }
 868
 869 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 870                                        int pages, sector_t *blocks,
 871                                        int create)
 872 {
 873         int rc;
 874
 875         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 876                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
 877                                                      blocks, create);
 878                 return rc;
 879         }
 880         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
 881
 882         return rc;
 883 }
 884 #else
 885 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 886                                        int pages, sector_t *blocks,
 887                                        int create)
 888 {
 889         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 890         int rc = 0, i = 0;
 891         struct page *fp = NULL;
 892         int clen = 0;
 893         pgoff_t max_page_index;
 894         handle_t *handle = NULL;
 895
 896         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
 897
 898         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 899                 inode->i_ino, pages, (*page)->index);
 900
 901         if (create) {
 902                 create = LDISKFS_GET_BLOCKS_CREATE;
 903                 handle = ldiskfs_journal_current_handle();
 904                 LASSERT(handle != NULL);
 905                 rc = osd_attach_jinode(inode);
 906                 if (rc)
 907                         return rc;
 908         }
 909         /* pages are sorted already. so, we just have to find
 910          * contig. space and process them properly */
 911         while (i < pages) {
 912                 long blen, total = 0;
 913                 struct ldiskfs_map_blocks map = { 0 };
 914
 915                 if (fp == NULL) { /* start new extent */
 916                         fp = *page++;
 917                         clen = 1;
 918                         if (++i != pages)
 919                                 continue;
 920                 } else if (fp->index + clen == (*page)->index) {
 921                         /* continue the extent */
 922                         page++;
 923                         clen++;
 924                         if (++i != pages)
 925                                 continue;
 926                 }
 927                 if (fp->index + clen >= max_page_index)
 928                         GOTO(cleanup, rc = -EFBIG);
 929                 /* process found extent */
 930                 map.m_lblk = fp->index * blocks_per_page;
 931                 map.m_len = blen = clen * blocks_per_page;
 932 cont_map:
 933                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 934                 if (rc >= 0) {
 935                         int c = 0;
 936                         for (; total < blen && c < map.m_len; c++, total++) {
 937                                 if (rc == 0) {
 938                                         *(blocks + total) = 0;
 939                                         total++;
 940                                         break;
 941                                 } else {
 942                                         *(blocks + total) = map.m_pblk + c;
 943                                         /* unmap any possible underlying
 944                                          * metadata from the block device
 945                                          * mapping.  bug 6998. */
 946                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
 947                                             create)
 948                                                 unmap_underlying_metadata(
 949                                                         inode->i_sb->s_bdev,
 950                                                         map.m_pblk + c);
 951                                 }
 952                         }
 953                         rc = 0;
 954                 }
 955                 if (rc == 0 && total < blen) {
 956                         map.m_lblk = fp->index * blocks_per_page + total;
 957                         map.m_len = blen - total;
 958                         goto cont_map;
 959                 }
 960                 if (rc != 0)
 961                         GOTO(cleanup, rc);
 962
 963                 /* look for next extent */
 964                 fp = NULL;
 965                 blocks += blocks_per_page * clen;
 966         }
 967 cleanup:
 968         return rc;
 969 }
 970 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
 971
 972 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 973                           struct niobuf_local *lnb, int npages)
 974 {
 975         struct osd_thread_info *oti   = osd_oti_get(env);
 976         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 977         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 978         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 979         ktime_t start;
 980         ktime_t end;
 981         s64 timediff;
 982         ssize_t                 isize;
 983         __s64                   maxidx;
 984         int                     rc = 0;
 985         int                     i;
 986         int                     cache = 0;
 987
 988         LASSERT(inode);
 989
 990         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 991         if (unlikely(rc != 0))
 992                 RETURN(rc);
 993
 994         isize = i_size_read(inode);
 995         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
 996
 997         if (osd->od_writethrough_cache)
 998                 cache = 1;
 999         if (isize > osd->od_readcache_max_filesize)
1000                 cache = 0;
1001
1002         start = ktime_get();
1003         for (i = 0; i < npages; i++) {
1004
1005                 if (cache == 0)
1006                         generic_error_remove_page(inode->i_mapping,
1007                                                   lnb[i].lnb_page);
1008
1009                 /*
1010                  * till commit the content of the page is undefined
1011                  * we'll set it uptodate once bulk is done. otherwise
1012                  * subsequent reads can access non-stable data
1013                  */
1014                 ClearPageUptodate(lnb[i].lnb_page);
1015
1016                 if (lnb[i].lnb_len == PAGE_SIZE)
1017                         continue;
1018
1019                 if (maxidx >= lnb[i].lnb_page->index) {
1020                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1021                 } else {
1022                         long off;
1023                         char *p = kmap(lnb[i].lnb_page);
1024
1025                         off = lnb[i].lnb_page_offset;
1026                         if (off)
1027                                 memset(p, 0, off);
1028                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1029                               ~PAGE_MASK;
1030                         if (off)
1031                                 memset(p + off, 0, PAGE_SIZE - off);
1032                         kunmap(lnb[i].lnb_page);
1033                 }
1034         }
1035         end = ktime_get();
1036         timediff = ktime_us_delta(end, start);
1037         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1038
1039         if (iobuf->dr_npages) {
1040                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1041                                                  iobuf->dr_npages,
1042                                                  iobuf->dr_blocks, 0);
1043                 if (likely(rc == 0)) {
1044                         rc = osd_do_bio(osd, inode, iobuf);
1045                         /* do IO stats for preparation reads */
1046                         osd_fini_iobuf(osd, iobuf);
1047                 }
1048         }
1049         RETURN(rc);
1050 }
1051
1052 struct osd_fextent {
1053         sector_t        start;
1054         sector_t        end;
1055         unsigned int    mapped:1;
1056 };
1057
1058 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1059                          struct osd_fextent *cached_extent)
1060 {
1061         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1062         sector_t block = offset >> inode->i_blkbits;
1063         sector_t start;
1064         struct fiemap_extent_info fei = { 0 };
1065         struct fiemap_extent fe = { 0 };
1066         mm_segment_t saved_fs;
1067         int rc;
1068
1069         if (block >= cached_extent->start && block < cached_extent->end)
1070                 return cached_extent->mapped;
1071
1072         if (i_size_read(inode) == 0)
1073                 return 0;
1074
1075         /* Beyond EOF, must not be mapped */
1076         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1077                 return 0;
1078
1079         fei.fi_extents_max = 1;
1080         fei.fi_extents_start = &fe;
1081
1082         saved_fs = get_fs();
1083         set_fs(get_ds());
1084         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1085         set_fs(saved_fs);
1086         if (rc != 0)
1087                 return 0;
1088
1089         start = fe.fe_logical >> inode->i_blkbits;
1090
1091         if (start > block) {
1092                 cached_extent->start = block;
1093                 cached_extent->end = start;
1094                 cached_extent->mapped = 0;
1095         } else {
1096                 cached_extent->start = start;
1097                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1098                                       inode->i_blkbits;
1099                 cached_extent->mapped = 1;
1100         }
1101
1102         return cached_extent->mapped;
1103 }
1104
1105 static int osd_declare_write_commit(const struct lu_env *env,
1106                                     struct dt_object *dt,
1107                                     struct niobuf_local *lnb, int npages,
1108                                     struct thandle *handle)
1109 {
1110         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1111         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1112         struct osd_thandle      *oh;
1113         int                     extents = 1;
1114         int                     depth;
1115         int                     i;
1116         int                     newblocks;
1117         int                     rc = 0;
1118         int                     flags = 0;
1119         int                     credits = 0;
1120         long long               quota_space = 0;
1121         struct osd_fextent      extent = { 0 };
1122         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1123         ENTRY;
1124
1125         LASSERT(handle != NULL);
1126         oh = container_of0(handle, struct osd_thandle, ot_super);
1127         LASSERT(oh->ot_handle == NULL);
1128
1129         newblocks = npages;
1130
1131         /* calculate number of extents (probably better to pass nb) */
1132         for (i = 0; i < npages; i++) {
1133                 if (i && lnb[i].lnb_file_offset !=
1134                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1135                         extents++;
1136
1137                 if (!osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1138                         quota_space += PAGE_SIZE;
1139
1140                 /* ignore quota for the whole request if any page is from
1141                  * client cache or written by root.
1142                  *
1143                  * XXX once we drop the 1.8 client support, the checking
1144                  * for whether page is from cache can be simplified as:
1145                  * !(lnb[i].flags & OBD_BRW_SYNC)
1146                  *
1147                  * XXX we could handle this on per-lnb basis as done by
1148                  * grant. */
1149                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1150                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1151                     OBD_BRW_FROM_GRANT)
1152                         declare_flags |= OSD_QID_FORCE;
1153         }
1154
1155         /*
1156          * each extent can go into new leaf causing a split
1157          * 5 is max tree depth: inode + 4 index blocks
1158          * with blockmaps, depth is 3 at most
1159          */
1160         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1161                 /*
1162                  * many concurrent threads may grow tree by the time
1163                  * our transaction starts. so, consider 2 is a min depth
1164                  */
1165                 depth = ext_depth(inode);
1166                 depth = max(depth, 1) + 1;
1167                 newblocks += depth;
1168                 credits++; /* inode */
1169                 credits += depth * 2 * extents;
1170         } else {
1171                 depth = 3;
1172                 newblocks += depth;
1173                 credits++; /* inode */
1174                 credits += depth * extents;
1175         }
1176
1177         /* quota space for metadata blocks */
1178         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1179
1180         /* quota space should be reported in 1K blocks */
1181         quota_space = toqb(quota_space);
1182
1183         /* each new block can go in different group (bitmap + gd) */
1184
1185         /* we can't dirty more bitmap blocks than exist */
1186         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1187                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1188         else
1189                 credits += newblocks;
1190
1191         /* we can't dirty more gd blocks than exist */
1192         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1193                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1194         else
1195                 credits += newblocks;
1196
1197         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1198
1199         /* make sure the over quota flags were not set */
1200         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1201
1202         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1203                                    i_projid_read(inode), quota_space, oh,
1204                                    osd_dt_obj(dt), &flags, declare_flags);
1205
1206         /* we need only to store the overquota flags in the first lnb for
1207          * now, once we support multiple objects BRW, this code needs be
1208          * revised. */
1209         if (flags & QUOTA_FL_OVER_USRQUOTA)
1210                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1211         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1212                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1213         if (flags & QUOTA_FL_OVER_PRJQUOTA)
1214                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1215
1216         RETURN(rc);
1217 }
1218
1219 /* Check if a block is allocated or not */
1220 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1221                             struct niobuf_local *lnb, int npages,
1222                             struct thandle *thandle)
1223 {
1224         struct osd_thread_info *oti = osd_oti_get(env);
1225         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1226         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1227         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1228         loff_t isize;
1229         int rc = 0, i;
1230         struct osd_fextent extent = { 0 };
1231
1232         LASSERT(inode);
1233
1234         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1235         if (unlikely(rc != 0))
1236                 RETURN(rc);
1237
1238         isize = i_size_read(inode);
1239         ll_vfs_dq_init(inode);
1240
1241         for (i = 0; i < npages; i++) {
1242                 if (lnb[i].lnb_rc == -ENOSPC &&
1243                     osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent)) {
1244                         /* Allow the write to proceed if overwriting an
1245                          * existing block */
1246                         lnb[i].lnb_rc = 0;
1247                 }
1248
1249                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1250                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1251                                lnb[i].lnb_rc);
1252                         LASSERT(lnb[i].lnb_page);
1253                         generic_error_remove_page(inode->i_mapping,
1254                                                   lnb[i].lnb_page);
1255                         continue;
1256                 }
1257
1258                 LASSERT(PageLocked(lnb[i].lnb_page));
1259                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1260
1261                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1262                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1263
1264                 /*
1265                  * Since write and truncate are serialized by oo_sem, even
1266                  * partial-page truncate should not leave dirty pages in the
1267                  * page cache.
1268                  */
1269                 LASSERT(!PageDirty(lnb[i].lnb_page));
1270
1271                 SetPageUptodate(lnb[i].lnb_page);
1272
1273                 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1274         }
1275
1276         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1277
1278         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1279                 rc = -ENOSPC;
1280         } else if (iobuf->dr_npages > 0) {
1281                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1282                                                  iobuf->dr_npages,
1283                                                  iobuf->dr_blocks, 1);
1284         } else {
1285                 /* no pages to write, no transno is needed */
1286                 thandle->th_local = 1;
1287         }
1288
1289         if (likely(rc == 0)) {
1290                 spin_lock(&inode->i_lock);
1291                 if (isize > i_size_read(inode)) {
1292                         i_size_write(inode, isize);
1293                         LDISKFS_I(inode)->i_disksize = isize;
1294                         spin_unlock(&inode->i_lock);
1295                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1296                 } else {
1297                         spin_unlock(&inode->i_lock);
1298                 }
1299
1300                 rc = osd_do_bio(osd, inode, iobuf);
1301                 /* we don't do stats here as in read path because
1302                  * write is async: we'll do this in osd_put_bufs() */
1303         } else {
1304                 osd_fini_iobuf(osd, iobuf);
1305         }
1306
1307         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1308
1309         if (unlikely(rc != 0)) {
1310                 /* if write fails, we should drop pages from the cache */
1311                 for (i = 0; i < npages; i++) {
1312                         if (lnb[i].lnb_page == NULL)
1313                                 continue;
1314                         LASSERT(PageLocked(lnb[i].lnb_page));
1315                         generic_error_remove_page(inode->i_mapping,
1316                                                   lnb[i].lnb_page);
1317                 }
1318         }
1319
1320         RETURN(rc);
1321 }
1322
1323 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1324                          struct niobuf_local *lnb, int npages)
1325 {
1326         struct osd_thread_info *oti = osd_oti_get(env);
1327         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1328         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1329         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1330         int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1331         ktime_t start, end;
1332         s64 timediff;
1333         loff_t isize;
1334
1335         LASSERT(inode);
1336
1337         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1338         if (unlikely(rc != 0))
1339                 RETURN(rc);
1340
1341         isize = i_size_read(inode);
1342
1343         if (osd->od_read_cache)
1344                 cache = 1;
1345         if (isize > osd->od_readcache_max_filesize)
1346                 cache = 0;
1347
1348         start = ktime_get();
1349         for (i = 0; i < npages; i++) {
1350
1351                 if (isize <= lnb[i].lnb_file_offset)
1352                         /* If there's no more data, abort early.
1353                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1354                         break;
1355
1356                 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1357                         lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1358                 else
1359                         lnb[i].lnb_rc = lnb[i].lnb_len;
1360
1361                 /* Bypass disk read if fail_loc is set properly */
1362                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1363                         SetPageUptodate(lnb[i].lnb_page);
1364
1365                 if (PageUptodate(lnb[i].lnb_page)) {
1366                         cache_hits++;
1367                 } else {
1368                         cache_misses++;
1369                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1370                 }
1371
1372                 if (cache == 0)
1373                         generic_error_remove_page(inode->i_mapping,
1374                                                   lnb[i].lnb_page);
1375         }
1376         end = ktime_get();
1377         timediff = ktime_us_delta(end, start);
1378         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1379
1380         if (cache_hits != 0)
1381                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1382                                     cache_hits);
1383         if (cache_misses != 0)
1384                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1385                                     cache_misses);
1386         if (cache_hits + cache_misses != 0)
1387                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1388                                     cache_hits + cache_misses);
1389
1390         if (iobuf->dr_npages) {
1391                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1392                                                  iobuf->dr_npages,
1393                                                  iobuf->dr_blocks, 0);
1394                 rc = osd_do_bio(osd, inode, iobuf);
1395
1396                 /* IO stats will be done in osd_bufs_put() */
1397         }
1398
1399         RETURN(rc);
1400 }
1401
1402 /*
1403  * XXX: Another layering violation for now.
1404  *
1405  * We don't want to use ->f_op->read methods, because generic file write
1406  *
1407  *         - serializes on ->i_sem, and
1408  *
1409  *         - does a lot of extra work like balance_dirty_pages(),
1410  *
1411  * which doesn't work for globally shared files like /last_rcvd.
1412  */
1413 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1414 {
1415         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1416
1417         memcpy(buffer, (char *)ei->i_data, buflen);
1418
1419         return  buflen;
1420 }
1421
1422 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1423 {
1424         struct buffer_head *bh;
1425         unsigned long block;
1426         int osize;
1427         int blocksize;
1428         int csize;
1429         int boffs;
1430
1431         /* prevent reading after eof */
1432         spin_lock(&inode->i_lock);
1433         if (i_size_read(inode) < *offs + size) {
1434                 loff_t diff = i_size_read(inode) - *offs;
1435                 spin_unlock(&inode->i_lock);
1436                 if (diff < 0) {
1437                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1438                                i_size_read(inode), *offs);
1439                         return -EBADR;
1440                 } else if (diff == 0) {
1441                         return 0;
1442                 } else {
1443                         size = diff;
1444                 }
1445         } else {
1446                 spin_unlock(&inode->i_lock);
1447         }
1448
1449         blocksize = 1 << inode->i_blkbits;
1450         osize = size;
1451         while (size > 0) {
1452                 block = *offs >> inode->i_blkbits;
1453                 boffs = *offs & (blocksize - 1);
1454                 csize = min(blocksize - boffs, size);
1455                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1456                 if (IS_ERR(bh)) {
1457                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1458                                LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
1459                                csize, *offs, inode->i_ino,
1460                                PTR_ERR(bh));
1461                         return PTR_ERR(bh);
1462                 }
1463
1464                 if (bh != NULL) {
1465                         memcpy(buf, bh->b_data + boffs, csize);
1466                         brelse(bh);
1467                 } else {
1468                         memset(buf, 0, csize);
1469                 }
1470
1471                 *offs += csize;
1472                 buf += csize;
1473                 size -= csize;
1474         }
1475         return osize;
1476 }
1477
1478 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1479                         struct lu_buf *buf, loff_t *pos)
1480 {
1481         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1482         int           rc;
1483
1484         /* Read small symlink from inode body as we need to maintain correct
1485          * on-disk symlinks for ldiskfs.
1486          */
1487         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1488             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1489                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1490         else
1491                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1492
1493         return rc;
1494 }
1495
1496 static inline int osd_extents_enabled(struct super_block *sb,
1497                                       struct inode *inode)
1498 {
1499         if (inode != NULL) {
1500                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1501                         return 1;
1502         } else if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
1503                                 LDISKFS_FEATURE_INCOMPAT_EXTENTS)) {
1504                 return 1;
1505         }
1506         return 0;
1507 }
1508
1509 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1510                            const loff_t size, const loff_t pos,
1511                            const int blocks)
1512 {
1513         int credits, bits, bs, i;
1514
1515         bits = sb->s_blocksize_bits;
1516         bs = 1 << bits;
1517
1518         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1519          * we do not expect blockmaps on the large files,
1520          * so let's shrink it to 2 levels (4GB files) */
1521
1522         /* this is default reservation: 2 levels */
1523         credits = (blocks + 2) * 3;
1524
1525         /* actual offset is unknown, hard to optimize */
1526         if (pos == -1)
1527                 return credits;
1528
1529         /* now check for few specific cases to optimize */
1530         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1531                 /* no indirects */
1532                 credits = blocks;
1533                 /* allocate if not allocated */
1534                 if (inode == NULL) {
1535                         credits += blocks * 2;
1536                         return credits;
1537                 }
1538                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1539                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1540                         if (LDISKFS_I(inode)->i_data[i] == 0)
1541                                 credits += 2;
1542                 }
1543         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1544                 /* single indirect */
1545                 credits = blocks * 3;
1546                 if (inode == NULL ||
1547                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1548                         credits += 3;
1549                 else
1550                         /* The indirect block may be modified. */
1551                         credits += 1;
1552         }
1553
1554         return credits;
1555 }
1556
1557 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1558                                  const struct lu_buf *buf, loff_t _pos,
1559                                  struct thandle *handle)
1560 {
1561         struct osd_object  *obj  = osd_dt_obj(dt);
1562         struct inode       *inode = obj->oo_inode;
1563         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1564         struct osd_thandle *oh;
1565         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1566         int                 bits, bs;
1567         int                 depth, size;
1568         loff_t              pos;
1569         ENTRY;
1570
1571         LASSERT(buf != NULL);
1572         LASSERT(handle != NULL);
1573
1574         oh = container_of0(handle, struct osd_thandle, ot_super);
1575         LASSERT(oh->ot_handle == NULL);
1576
1577         size = buf->lb_len;
1578         bits = sb->s_blocksize_bits;
1579         bs = 1 << bits;
1580
1581         if (_pos == -1) {
1582                 /* if this is an append, then we
1583                  * should expect cross-block record */
1584                 pos = 0;
1585         } else {
1586                 pos = _pos;
1587         }
1588
1589         /* blocks to modify */
1590         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1591         LASSERT(blocks > 0);
1592
1593         if (inode != NULL && _pos != -1) {
1594                 /* object size in blocks */
1595                 est = (i_size_read(inode) + bs - 1) >> bits;
1596                 allocated = inode->i_blocks >> (bits - 9);
1597                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1598                         /* looks like an overwrite, no need to modify tree */
1599                         credits = blocks;
1600                         /* no need to modify i_size */
1601                         goto out;
1602                 }
1603         }
1604
1605         if (osd_extents_enabled(sb, inode)) {
1606                 /*
1607                  * many concurrent threads may grow tree by the time
1608                  * our transaction starts. so, consider 2 is a min depth
1609                  * for every level we may need to allocate a new block
1610                  * and take some entries from the old one. so, 3 blocks
1611                  * to allocate (bitmap, gd, itself) + old block - 4 per
1612                  * level.
1613                  */
1614                 depth = inode != NULL ? ext_depth(inode) : 0;
1615                 depth = max(depth, 1) + 1;
1616                 credits = depth;
1617                 /* if not append, then split may need to modify
1618                  * existing blocks moving entries into the new ones */
1619                 if (_pos == -1)
1620                         credits += depth;
1621                 /* blocks to store data: bitmap,gd,itself */
1622                 credits += blocks * 3;
1623         } else {
1624                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1625         }
1626         /* if inode is created as part of the transaction,
1627          * then it's counted already by the creation method */
1628         if (inode != NULL)
1629                 credits++;
1630
1631 out:
1632
1633         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1634
1635         /* dt_declare_write() is usually called for system objects, such
1636          * as llog or last_rcvd files. We needn't enforce quota on those
1637          * objects, so always set the lqi_space as 0. */
1638         if (inode != NULL)
1639                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1640                                            i_gid_read(inode),
1641                                            i_projid_read(inode), 0,
1642                                            oh, obj, NULL, OSD_QID_BLK);
1643         RETURN(rc);
1644 }
1645
1646 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1647 {
1648         /* LU-2634: clear the extent format for fast symlink */
1649         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1650
1651         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1652         spin_lock(&inode->i_lock);
1653         LDISKFS_I(inode)->i_disksize = buflen;
1654         i_size_write(inode, buflen);
1655         spin_unlock(&inode->i_lock);
1656         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1657
1658         return 0;
1659 }
1660
1661 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1662                              int write_NUL, loff_t *offs, handle_t *handle)
1663 {
1664         struct buffer_head *bh        = NULL;
1665         loff_t              offset    = *offs;
1666         loff_t              new_size  = i_size_read(inode);
1667         unsigned long       block;
1668         int                 blocksize = 1 << inode->i_blkbits;
1669         int                 err = 0;
1670         int                 size;
1671         int                 boffs;
1672         int                 dirty_inode = 0;
1673
1674         if (write_NUL) {
1675                 /*
1676                  * long symlink write does not count the NUL terminator in
1677                  * bufsize, we write it, and the inode's file size does not
1678                  * count the NUL terminator as well.
1679                  */
1680                 ((char *)buf)[bufsize] = '\0';
1681                 ++bufsize;
1682         }
1683         while (bufsize > 0) {
1684                 if (bh != NULL)
1685                         brelse(bh);
1686
1687                 block = offset >> inode->i_blkbits;
1688                 boffs = offset & (blocksize - 1);
1689                 size = min(blocksize - boffs, bufsize);
1690                 bh = __ldiskfs_bread(handle, inode, block, 1);
1691                 if (IS_ERR_OR_NULL(bh)) {
1692                         if (bh == NULL) {
1693                                 err = -EIO;
1694                         } else {
1695                                 err = PTR_ERR(bh);
1696                                 bh = NULL;
1697                         }
1698                         CERROR("%s: error reading offset %llu (block %lu): "
1699                                "rc = %d\n",
1700                                inode->i_sb->s_id, offset, block, err);
1701                         break;
1702                 }
1703
1704                 err = ldiskfs_journal_get_write_access(handle, bh);
1705                 if (err) {
1706                         CERROR("journal_get_write_access() returned error %d\n",
1707                                err);
1708                         break;
1709                 }
1710                 LASSERTF(boffs + size <= bh->b_size,
1711                          "boffs %d size %d bh->b_size %lu\n",
1712                          boffs, size, (unsigned long)bh->b_size);
1713                 memcpy(bh->b_data + boffs, buf, size);
1714                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1715                 if (err)
1716                         break;
1717
1718                 if (offset + size > new_size)
1719                         new_size = offset + size;
1720                 offset += size;
1721                 bufsize -= size;
1722                 buf += size;
1723         }
1724         if (bh)
1725                 brelse(bh);
1726
1727         if (write_NUL)
1728                 --new_size;
1729         /* correct in-core and on-disk sizes */
1730         if (new_size > i_size_read(inode)) {
1731                 spin_lock(&inode->i_lock);
1732                 if (new_size > i_size_read(inode))
1733                         i_size_write(inode, new_size);
1734                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1735                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1736                         dirty_inode = 1;
1737                 }
1738                 spin_unlock(&inode->i_lock);
1739                 if (dirty_inode)
1740                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1741         }
1742
1743         if (err == 0)
1744                 *offs = offset;
1745         return err;
1746 }
1747
1748 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1749                          const struct lu_buf *buf, loff_t *pos,
1750                          struct thandle *handle, int ignore_quota)
1751 {
1752         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1753         struct osd_thandle      *oh;
1754         ssize_t                 result;
1755         int                     is_link;
1756
1757         LASSERT(dt_object_exists(dt));
1758
1759         LASSERT(handle != NULL);
1760         LASSERT(inode != NULL);
1761         ll_vfs_dq_init(inode);
1762
1763         /* XXX: don't check: one declared chunk can be used many times */
1764         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1765
1766         oh = container_of(handle, struct osd_thandle, ot_super);
1767         LASSERT(oh->ot_handle->h_transaction != NULL);
1768         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1769
1770         /* Write small symlink to inode body as we need to maintain correct
1771          * on-disk symlinks for ldiskfs.
1772          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1773          * does not count it in.
1774          */
1775         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1776         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1777                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1778         else
1779                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1780                                                   buf->lb_len, is_link, pos,
1781                                                   oh->ot_handle);
1782         if (result == 0)
1783                 result = buf->lb_len;
1784
1785         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1786
1787         return result;
1788 }
1789
1790 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1791                              __u64 start, __u64 end, struct thandle *th)
1792 {
1793         struct osd_thandle *oh;
1794         struct inode       *inode;
1795         int                 rc;
1796         ENTRY;
1797
1798         LASSERT(th);
1799         oh = container_of(th, struct osd_thandle, ot_super);
1800
1801         /*
1802          * we don't need to reserve credits for whole truncate
1803          * it's not possible as truncate may need to free too many
1804          * blocks and that won't fit a single transaction. instead
1805          * we reserve credits to change i_size and put inode onto
1806          * orphan list. if needed truncate will extend or restart
1807          * transaction
1808          */
1809         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1810                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1811
1812         inode = osd_dt_obj(dt)->oo_inode;
1813         LASSERT(inode);
1814
1815         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1816                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1817                                    NULL, OSD_QID_BLK);
1818         RETURN(rc);
1819 }
1820
1821 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1822                      __u64 start, __u64 end, struct thandle *th)
1823 {
1824         struct osd_thandle *oh;
1825         struct osd_object  *obj = osd_dt_obj(dt);
1826         struct inode       *inode = obj->oo_inode;
1827         handle_t           *h;
1828         tid_t               tid;
1829         int                rc = 0, rc2 = 0;
1830         ENTRY;
1831
1832         LASSERT(end == OBD_OBJECT_EOF);
1833         LASSERT(dt_object_exists(dt));
1834         LASSERT(osd_invariant(obj));
1835         LASSERT(inode != NULL);
1836         ll_vfs_dq_init(inode);
1837
1838         LASSERT(th);
1839         oh = container_of(th, struct osd_thandle, ot_super);
1840         LASSERT(oh->ot_handle->h_transaction != NULL);
1841
1842         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1843
1844         tid = oh->ot_handle->h_transaction->t_tid;
1845
1846         spin_lock(&inode->i_lock);
1847         i_size_write(inode, start);
1848         spin_unlock(&inode->i_lock);
1849         ll_truncate_pagecache(inode, start);
1850 #ifdef HAVE_INODEOPS_TRUNCATE
1851         if (inode->i_op->truncate) {
1852                 inode->i_op->truncate(inode);
1853         } else
1854 #endif
1855                 ldiskfs_truncate(inode);
1856
1857         /*
1858          * For a partial-page truncate, flush the page to disk immediately to
1859          * avoid data corruption during direct disk write.  b=17397
1860          */
1861         if ((start & ~PAGE_MASK) != 0)
1862                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1863
1864         h = journal_current_handle();
1865         LASSERT(h != NULL);
1866         LASSERT(h == oh->ot_handle);
1867
1868         /* do not check credits with osd_trans_exec_check() as the truncate
1869          * can restart the transaction internally and we restart the
1870          * transaction in this case */
1871
1872         if (tid != h->h_transaction->t_tid) {
1873                 int credits = oh->ot_credits;
1874                 /*
1875                  * transaction has changed during truncate
1876                  * we need to restart the handle with our credits
1877                  */
1878                 if (h->h_buffer_credits < credits) {
1879                         if (ldiskfs_journal_extend(h, credits))
1880                                 rc2 = ldiskfs_journal_restart(h, credits);
1881                 }
1882         }
1883
1884         RETURN(rc == 0 ? rc2 : rc);
1885 }
1886
1887 static int fiemap_check_ranges(struct inode *inode,
1888                                u64 start, u64 len, u64 *new_len)
1889 {
1890         loff_t maxbytes;
1891
1892         *new_len = len;
1893
1894         if (len == 0)
1895                 return -EINVAL;
1896
1897         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1898                 maxbytes = inode->i_sb->s_maxbytes;
1899         else
1900                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1901
1902         if (start > maxbytes)
1903                 return -EFBIG;
1904
1905         /*
1906          * Shrink request scope to what the fs can actually handle.
1907          */
1908         if (len > maxbytes || (maxbytes - len) < start)
1909                 *new_len = maxbytes - start;
1910
1911         return 0;
1912 }
1913
1914 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1915 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
1916
1917 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1918                           struct fiemap *fm)
1919 {
1920         struct fiemap_extent_info fieinfo = {0, };
1921         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1922         u64 len;
1923         int rc;
1924
1925
1926         LASSERT(inode);
1927         if (inode->i_op->fiemap == NULL)
1928                 return -EOPNOTSUPP;
1929
1930         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1931                 return -EINVAL;
1932
1933         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1934         if (rc)
1935                 return rc;
1936
1937         fieinfo.fi_flags = fm->fm_flags;
1938         fieinfo.fi_extents_max = fm->fm_extent_count;
1939         fieinfo.fi_extents_start = fm->fm_extents;
1940
1941         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1942                 filemap_write_and_wait(inode->i_mapping);
1943
1944         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
1945         fm->fm_flags = fieinfo.fi_flags;
1946         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
1947
1948         return rc;
1949 }
1950
1951 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1952                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1953 {
1954         int              rc = 0;
1955         struct inode    *inode = osd_dt_obj(dt)->oo_inode;
1956         ENTRY;
1957
1958         switch (advice) {
1959         case LU_LADVISE_DONTNEED:
1960                 if (end == 0)
1961                         break;
1962                 invalidate_mapping_pages(inode->i_mapping,
1963                                          start >> PAGE_CACHE_SHIFT,
1964                                          (end - 1) >> PAGE_CACHE_SHIFT);
1965                 break;
1966         default:
1967                 rc = -ENOTSUPP;
1968                 break;
1969         }
1970
1971         RETURN(rc);
1972 }
1973
1974 /*
1975  * in some cases we may need declare methods for objects being created
1976  * e.g., when we create symlink
1977  */
1978 const struct dt_body_operations osd_body_ops_new = {
1979         .dbo_declare_write = osd_declare_write,
1980 };
1981
1982 const struct dt_body_operations osd_body_ops = {
1983         .dbo_read                       = osd_read,
1984         .dbo_declare_write              = osd_declare_write,
1985         .dbo_write                      = osd_write,
1986         .dbo_bufs_get                   = osd_bufs_get,
1987         .dbo_bufs_put                   = osd_bufs_put,
1988         .dbo_write_prep                 = osd_write_prep,
1989         .dbo_declare_write_commit       = osd_declare_write_commit,
1990         .dbo_write_commit               = osd_write_commit,
1991         .dbo_read_prep                  = osd_read_prep,
1992         .dbo_declare_punch              = osd_declare_punch,
1993         .dbo_punch                      = osd_punch,
1994         .dbo_fiemap_get                 = osd_fiemap_get,
1995         .dbo_ladvise                    = osd_ladvise,
1996 };