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, 2017, 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 /* prerequisite for linux/xattr.h */
  42 #include <linux/types.h>
  43 /* prerequisite for linux/xattr.h */
  44 #include <linux/fs.h>
  45 #include <linux/mm.h>
  46 #include <linux/pagevec.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 = ktime_set(0, 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                                       ktime_to_ms(iobuf->dr_elapsed));
 137         }
 138 }
 139
 140 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 141 static void dio_complete_routine(struct bio *bio)
 142 {
 143 # ifdef HAVE_BI_STATUS
 144         int error = bio->bi_status;
 145 # else
 146         int error = bio->bi_error;
 147 # endif
 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, "
 169 #ifdef HAVE_BI_RW
 170                        "bi_rw: %lu,"
 171 #else
 172                        "bi_opf: %u,"
 173 #endif
 174                        "bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p,"
 175                        "bi_cnt: %d, bi_private: %p\n", bio->bi_next,
 176                         (unsigned long)bio->bi_flags,
 177 #ifdef HAVE_BI_RW
 178                         bio->bi_rw,
 179 #else
 180                         bio->bi_opf,
 181 #endif
 182                         bio->bi_vcnt, bio_idx(bio),
 183                         bio_sectors(bio) << 9, bio->bi_end_io,
 184 #ifdef HAVE_BI_CNT
 185                         atomic_read(&bio->bi_cnt),
 186 #else
 187                         atomic_read(&bio->__bi_cnt),
 188 #endif
 189                         bio->bi_private);
 190                 return;
 191         }
 192
 193         /* the check is outside of the cycle for performance reason -bzzz */
 194         if (!bio_data_dir(bio)) {
 195                 bio_for_each_segment_all(bvl, bio, iter) {
 196                         if (likely(error == 0))
 197                                 SetPageUptodate(bvl_to_page(bvl));
 198                         LASSERT(PageLocked(bvl_to_page(bvl)));
 199                 }
 200                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 201         } else {
 202                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 203         }
 204
 205         /* any real error is good enough -bzzz */
 206         if (error != 0 && iobuf->dr_error == 0)
 207                 iobuf->dr_error = error;
 208
 209         /*
 210          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 211          * it's possible that service thread will see dr_numreqs
 212          * is zero, but dr_elapsed is not set yet, leading to lost
 213          * data in this processing and an assertion in a subsequent
 214          * call to OSD.
 215          */
 216         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 217                 ktime_t now = ktime_get();
 218
 219                 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
 220                 iobuf->dr_elapsed_valid = 1;
 221         }
 222         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 223                 wake_up(&iobuf->dr_wait);
 224
 225         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 226          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 227          * mempool when serious on-disk fragmentation was encountered,
 228          * deadlocking the OST.  The bios are now released as soon as complete
 229          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 230         bio_put(bio);
 231 }
 232
 233 static void record_start_io(struct osd_iobuf *iobuf, int size)
 234 {
 235         struct osd_device    *osd = iobuf->dr_dev;
 236         struct obd_histogram *h = osd->od_brw_stats.hist;
 237
 238         iobuf->dr_frags++;
 239         atomic_inc(&iobuf->dr_numreqs);
 240
 241         if (iobuf->dr_rw == 0) {
 242                 atomic_inc(&osd->od_r_in_flight);
 243                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 244                                  atomic_read(&osd->od_r_in_flight));
 245                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 246         } else if (iobuf->dr_rw == 1) {
 247                 atomic_inc(&osd->od_w_in_flight);
 248                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 249                                  atomic_read(&osd->od_w_in_flight));
 250                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 251         } else {
 252                 LBUG();
 253         }
 254 }
 255
 256 static void osd_submit_bio(int rw, struct bio *bio)
 257 {
 258         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 259 #ifdef HAVE_SUBMIT_BIO_2ARGS
 260         if (rw == 0)
 261                 submit_bio(READ, bio);
 262         else
 263                 submit_bio(WRITE, bio);
 264 #else
 265         bio->bi_opf |= rw;
 266         submit_bio(bio);
 267 #endif
 268 }
 269
 270 static int can_be_merged(struct bio *bio, sector_t sector)
 271 {
 272         if (bio == NULL)
 273                 return 0;
 274
 275         return bio_end_sector(bio) == sector ? 1 : 0;
 276 }
 277
 278 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 279                       struct osd_iobuf *iobuf)
 280 {
 281         int            blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 282         struct page  **pages = iobuf->dr_pages;
 283         int            npages = iobuf->dr_npages;
 284         sector_t      *blocks = iobuf->dr_blocks;
 285         int            total_blocks = npages * blocks_per_page;
 286         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 287         unsigned int   blocksize = inode->i_sb->s_blocksize;
 288         struct bio    *bio = NULL;
 289         struct page   *page;
 290         unsigned int   page_offset;
 291         sector_t       sector;
 292         int            nblocks;
 293         int            block_idx;
 294         int            page_idx;
 295         int            i;
 296         int            rc = 0;
 297         DECLARE_PLUG(plug);
 298         ENTRY;
 299
 300         LASSERT(iobuf->dr_npages == npages);
 301
 302         osd_brw_stats_update(osd, iobuf);
 303         iobuf->dr_start_time = ktime_get();
 304
 305         blk_start_plug(&plug);
 306         for (page_idx = 0, block_idx = 0;
 307              page_idx < npages;
 308              page_idx++, block_idx += blocks_per_page) {
 309
 310                 page = pages[page_idx];
 311                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 312
 313                 for (i = 0, page_offset = 0;
 314                      i < blocks_per_page;
 315                      i += nblocks, page_offset += blocksize * nblocks) {
 316
 317                         nblocks = 1;
 318
 319                         if (blocks[block_idx + i] == 0) {  /* hole */
 320                                 LASSERTF(iobuf->dr_rw == 0,
 321                                          "page_idx %u, block_idx %u, i %u\n",
 322                                          page_idx, block_idx, i);
 323                                 memset(kmap(page) + page_offset, 0, blocksize);
 324                                 kunmap(page);
 325                                 continue;
 326                         }
 327
 328                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 329
 330                         /* Additional contiguous file blocks? */
 331                         while (i + nblocks < blocks_per_page &&
 332                                (sector + (nblocks << sector_bits)) ==
 333                                ((sector_t)blocks[block_idx + i + nblocks] <<
 334                                 sector_bits))
 335                                 nblocks++;
 336
 337                         if (bio != NULL &&
 338                             can_be_merged(bio, sector) &&
 339                             bio_add_page(bio, page,
 340                                          blocksize * nblocks, page_offset) != 0)
 341                                 continue;       /* added this frag OK */
 342
 343                         if (bio != NULL) {
 344                                 struct request_queue *q =
 345                                         bdev_get_queue(bio->bi_bdev);
 346                                 unsigned int bi_size = bio_sectors(bio) << 9;
 347
 348                                 /* Dang! I have to fragment this I/O */
 349                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 350                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 351                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 352                                        bio_sectors(bio),
 353                                        queue_max_sectors(q),
 354                                        bio_phys_segments(q, bio),
 355                                        queue_max_phys_segments(q),
 356                                        0, queue_max_hw_segments(q));
 357                                 record_start_io(iobuf, bi_size);
 358                                 osd_submit_bio(iobuf->dr_rw, bio);
 359                         }
 360
 361                         /* allocate new bio */
 362                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 363                                                       (npages - page_idx) *
 364                                                       blocks_per_page));
 365                         if (bio == NULL) {
 366                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 367                                        (npages - page_idx), blocks_per_page,
 368                                        (npages - page_idx) * blocks_per_page);
 369                                 rc = -ENOMEM;
 370                                 goto out;
 371                         }
 372
 373                         bio->bi_bdev = inode->i_sb->s_bdev;
 374                         bio_set_sector(bio, sector);
 375 #ifdef HAVE_BI_RW
 376                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 377 #else
 378                         bio->bi_opf = (iobuf->dr_rw == 0) ? READ : WRITE;
 379 #endif
 380                         bio->bi_end_io = dio_complete_routine;
 381                         bio->bi_private = iobuf;
 382
 383                         rc = bio_add_page(bio, page,
 384                                           blocksize * nblocks, page_offset);
 385                         LASSERT(rc != 0);
 386                 }
 387         }
 388
 389         if (bio != NULL) {
 390                 record_start_io(iobuf, bio_sectors(bio) << 9);
 391                 osd_submit_bio(iobuf->dr_rw, bio);
 392                 rc = 0;
 393         }
 394
 395 out:
 396         blk_finish_plug(&plug);
 397
 398         /* in order to achieve better IO throughput, we don't wait for writes
 399          * completion here. instead we proceed with transaction commit in
 400          * parallel and wait for IO completion once transaction is stopped
 401          * see osd_trans_stop() for more details -bzzz */
 402         if (iobuf->dr_rw == 0) {
 403                 wait_event(iobuf->dr_wait,
 404                            atomic_read(&iobuf->dr_numreqs) == 0);
 405                 osd_fini_iobuf(osd, iobuf);
 406         }
 407
 408         if (rc == 0)
 409                 rc = iobuf->dr_error;
 410         RETURN(rc);
 411 }
 412
 413 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 414                                    struct niobuf_local *lnb)
 415 {
 416         ENTRY;
 417
 418         *nrpages = 0;
 419
 420         while (len > 0) {
 421                 int poff = offset & (PAGE_SIZE - 1);
 422                 int plen = PAGE_SIZE - poff;
 423
 424                 if (plen > len)
 425                         plen = len;
 426                 lnb->lnb_file_offset = offset;
 427                 lnb->lnb_page_offset = poff;
 428                 lnb->lnb_len = plen;
 429                 /* lnb->lnb_flags = rnb->rnb_flags; */
 430                 lnb->lnb_flags = 0;
 431                 lnb->lnb_page = NULL;
 432                 lnb->lnb_rc = 0;
 433
 434                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 435                          (long long) len);
 436                 offset += plen;
 437                 len -= plen;
 438                 lnb++;
 439                 (*nrpages)++;
 440         }
 441
 442         RETURN(0);
 443 }
 444
 445 static struct page *osd_get_page(struct dt_object *dt, loff_t offset,
 446                                  gfp_t gfp_mask)
 447 {
 448         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 449         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 450         struct page *page;
 451
 452         LASSERT(inode);
 453
 454         page = find_or_create_page(inode->i_mapping, offset >> PAGE_SHIFT,
 455                                    gfp_mask);
 456
 457         if (unlikely(page == NULL))
 458                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 459
 460         return page;
 461 }
 462
 463 /*
 464  * there are following "locks":
 465  * journal_start
 466  * i_mutex
 467  * page lock
 468  *
 469  * osd write path:
 470  *  - lock page(s)
 471  *  - journal_start
 472  *  - truncate_sem
 473  *
 474  * ext4 vmtruncate:
 475  *  - lock pages, unlock
 476  *  - journal_start
 477  *  - lock partial page
 478  *  - i_data_sem
 479  *
 480  */
 481
 482 /**
 483  * Unlock and release pages loaded by osd_bufs_get()
 484  *
 485  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 486  *
 487  * \param env           thread execution environment
 488  * \param dt            dt object undergoing IO (OSD object + methods)
 489  * \param lnb           array of pages undergoing IO
 490  * \param npages        number of pages in \a lnb
 491  *
 492  * \retval 0            always
 493  */
 494 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 495                         struct niobuf_local *lnb, int npages)
 496 {
 497         struct pagevec pvec;
 498         int i;
 499
 500 #ifdef HAVE_PAGEVEC_INIT_ONE_PARAM
 501         pagevec_init(&pvec);
 502 #else
 503         pagevec_init(&pvec, 0);
 504 #endif
 505
 506         for (i = 0; i < npages; i++) {
 507                 if (lnb[i].lnb_page == NULL)
 508                         continue;
 509                 LASSERT(PageLocked(lnb[i].lnb_page));
 510                 unlock_page(lnb[i].lnb_page);
 511                 if (pagevec_add(&pvec, lnb[i].lnb_page) == 0)
 512                         pagevec_release(&pvec);
 513                 dt_object_put(env, dt);
 514                 lnb[i].lnb_page = NULL;
 515         }
 516
 517         /* Release any partial pagevec */
 518         pagevec_release(&pvec);
 519
 520         RETURN(0);
 521 }
 522
 523 /**
 524  * Load and lock pages undergoing IO
 525  *
 526  * Pages as described in the \a lnb array are fetched (from disk or cache)
 527  * and locked for IO by the caller.
 528  *
 529  * DLM locking protects us from write and truncate competing for same region,
 530  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 531  * It's possible the writeout on a such a page is in progress when we access
 532  * it. It's also possible that during this writeout we put new (partial) data
 533  * into the page, but won't be able to proceed in filter_commitrw_write().
 534  * Therefore, just wait for writeout completion as it should be rare enough.
 535  *
 536  * \param env           thread execution environment
 537  * \param dt            dt object undergoing IO (OSD object + methods)
 538  * \param pos           byte offset of IO start
 539  * \param len           number of bytes of IO
 540  * \param lnb           array of extents undergoing IO
 541  * \param rw            read or write operation, and other flags
 542  * \param capa          capabilities
 543  *
 544  * \retval pages        (zero or more) loaded successfully
 545  * \retval -ENOMEM      on memory/page allocation error
 546  */
 547 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 548                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 549                         enum dt_bufs_type rw)
 550 {
 551         struct osd_object *obj = osd_dt_obj(dt);
 552         int npages, i, rc = 0;
 553         gfp_t gfp_mask;
 554
 555         LASSERT(obj->oo_inode);
 556
 557         osd_map_remote_to_local(pos, len, &npages, lnb);
 558
 559         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 560         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 561                                              GFP_HIGHUSER;
 562         for (i = 0; i < npages; i++, lnb++) {
 563                 lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset,
 564                                              gfp_mask);
 565                 if (lnb->lnb_page == NULL)
 566                         GOTO(cleanup, rc = -ENOMEM);
 567
 568                 wait_on_page_writeback(lnb->lnb_page);
 569                 BUG_ON(PageWriteback(lnb->lnb_page));
 570
 571                 lu_object_get(&dt->do_lu);
 572         }
 573
 574         RETURN(i);
 575
 576 cleanup:
 577         if (i > 0)
 578                 osd_bufs_put(env, dt, lnb - i, i);
 579         return rc;
 580 }
 581
 582 #ifndef HAVE_LDISKFS_MAP_BLOCKS
 583
 584 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 585 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 586 #endif
 587
 588 struct bpointers {
 589         sector_t *blocks;
 590         unsigned long start;
 591         int num;
 592         int init_num;
 593         int create;
 594 };
 595
 596 static long ldiskfs_ext_find_goal(struct inode *inode,
 597                                   struct ldiskfs_ext_path *path,
 598                                   unsigned long block, int *aflags)
 599 {
 600         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 601         unsigned long bg_start;
 602         unsigned long colour;
 603         int depth;
 604
 605         if (path) {
 606                 struct ldiskfs_extent *ex;
 607                 depth = path->p_depth;
 608
 609                 /* try to predict block placement */
 610                 if ((ex = path[depth].p_ext))
 611                         return ldiskfs_ext_pblock(ex) +
 612                                 (block - le32_to_cpu(ex->ee_block));
 613
 614                 /* it looks index is empty
 615                  * try to find starting from index itself */
 616                 if (path[depth].p_bh)
 617                         return path[depth].p_bh->b_blocknr;
 618         }
 619
 620         /* OK. use inode's group */
 621         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 622                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 623         colour = (current->pid % 16) *
 624                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 625         return bg_start + colour + block;
 626 }
 627
 628 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 629                                 struct ldiskfs_ext_path *path,
 630                                 unsigned long block, unsigned long *count,
 631                                 int *err)
 632 {
 633         struct ldiskfs_allocation_request ar;
 634         unsigned long pblock;
 635         int aflags;
 636
 637         /* find neighbour allocated blocks */
 638         ar.lleft = block;
 639         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 640         if (*err)
 641                 return 0;
 642         ar.lright = block;
 643         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 644         if (*err)
 645                 return 0;
 646
 647         /* allocate new block */
 648         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 649         ar.inode = inode;
 650         ar.logical = block;
 651         ar.len = *count;
 652         ar.flags = LDISKFS_MB_HINT_DATA;
 653         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 654         *count = ar.len;
 655         return pblock;
 656 }
 657
 658 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 659                                      struct ldiskfs_ext_path *path,
 660                                      struct ldiskfs_ext_cache *cex,
 661 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 662                                      struct ldiskfs_extent *ex,
 663 #endif
 664                                      void *cbdata)
 665 {
 666         struct bpointers *bp = cbdata;
 667         struct ldiskfs_extent nex;
 668         unsigned long pblock = 0;
 669         unsigned long tgen;
 670         int err, i;
 671         unsigned long count;
 672         handle_t *handle;
 673
 674 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 675         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 676 #else
 677         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 678 #endif
 679                 err = EXT_CONTINUE;
 680                 goto map;
 681         }
 682
 683         if (bp->create == 0) {
 684                 i = 0;
 685                 if (cex->ec_block < bp->start)
 686                         i = bp->start - cex->ec_block;
 687                 if (i >= cex->ec_len)
 688                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 689                                         i, cex->ec_len);
 690                 for (; i < cex->ec_len && bp->num; i++) {
 691                         *(bp->blocks) = 0;
 692                         bp->blocks++;
 693                         bp->num--;
 694                         bp->start++;
 695                 }
 696
 697                 return EXT_CONTINUE;
 698         }
 699
 700         tgen = LDISKFS_I(inode)->i_ext_generation;
 701         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
 702
 703         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
 704                                    count + LDISKFS_ALLOC_NEEDED + 1);
 705         if (IS_ERR(handle)) {
 706                 return PTR_ERR(handle);
 707         }
 708
 709         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
 710                 /* the tree has changed. so path can be invalid at moment */
 711                 ldiskfs_journal_stop(handle);
 712                 return EXT_REPEAT;
 713         }
 714
 715         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
 716          * protected by i_data_sem as whole. so we patch it to store
 717          * generation to path and now verify the tree hasn't changed */
 718         down_write((&LDISKFS_I(inode)->i_data_sem));
 719
 720         /* validate extent, make sure the extent tree does not changed */
 721         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
 722                 /* cex is invalid, try again */
 723                 up_write(&LDISKFS_I(inode)->i_data_sem);
 724                 ldiskfs_journal_stop(handle);
 725                 return EXT_REPEAT;
 726         }
 727
 728         count = cex->ec_len;
 729         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
 730         if (!pblock)
 731                 goto out;
 732         BUG_ON(count > cex->ec_len);
 733
 734         /* insert new extent */
 735         nex.ee_block = cpu_to_le32(cex->ec_block);
 736         ldiskfs_ext_store_pblock(&nex, pblock);
 737         nex.ee_len = cpu_to_le16(count);
 738         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
 739         if (err) {
 740                 /* free data blocks we just allocated */
 741                 /* not a good idea to call discard here directly,
 742                  * but otherwise we'd need to call it every free() */
 743                 ldiskfs_discard_preallocations(inode);
 744 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
 745                 ldiskfs_free_blocks(handle, inode, NULL,
 746                                     ldiskfs_ext_pblock(&nex),
 747                                     le16_to_cpu(nex.ee_len), 0);
 748 #else
 749                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
 750                                     le16_to_cpu(nex.ee_len), 0);
 751 #endif
 752                 goto out;
 753         }
 754
 755         /*
 756          * Putting len of the actual extent we just inserted,
 757          * we are asking ldiskfs_ext_walk_space() to continue
 758          * scaning after that block
 759          */
 760         cex->ec_len = le16_to_cpu(nex.ee_len);
 761         cex->ec_start = ldiskfs_ext_pblock(&nex);
 762         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
 763         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
 764
 765 out:
 766         up_write((&LDISKFS_I(inode)->i_data_sem));
 767         ldiskfs_journal_stop(handle);
 768 map:
 769         if (err >= 0) {
 770                 /* map blocks */
 771                 if (bp->num == 0) {
 772                         CERROR("hmm. why do we find this extent?\n");
 773                         CERROR("initial space: %lu:%u\n",
 774                                 bp->start, bp->init_num);
 775 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 776                         CERROR("current extent: %u/%u/%llu %d\n",
 777                                 cex->ec_block, cex->ec_len,
 778                                 (unsigned long long)cex->ec_start,
 779                                 cex->ec_type);
 780 #else
 781                         CERROR("current extent: %u/%u/%llu\n",
 782                                 cex->ec_block, cex->ec_len,
 783                                 (unsigned long long)cex->ec_start);
 784 #endif
 785                 }
 786                 i = 0;
 787                 if (cex->ec_block < bp->start)
 788                         i = bp->start - cex->ec_block;
 789                 if (i >= cex->ec_len)
 790                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 791                                         i, cex->ec_len);
 792                 for (; i < cex->ec_len && bp->num; i++) {
 793                         *(bp->blocks) = cex->ec_start + i;
 794                         if (pblock != 0) {
 795                                 /* unmap any possible underlying metadata from
 796                                  * the block device mapping.  bug 6998. */
 797 #ifndef HAVE_CLEAN_BDEV_ALIASES
 798                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
 799                                                           *(bp->blocks));
 800 #else
 801                                 clean_bdev_aliases(inode->i_sb->s_bdev,
 802                                                    *(bp->blocks), 1);
 803 #endif
 804                         }
 805                         bp->blocks++;
 806                         bp->num--;
 807                         bp->start++;
 808                 }
 809         }
 810         return err;
 811 }
 812
 813 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
 814                                    int clen, sector_t *blocks, int create)
 815 {
 816         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 817         struct bpointers bp;
 818         int err;
 819
 820         if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
 821                 return -EFBIG;
 822
 823         bp.blocks = blocks;
 824         bp.start = index * blocks_per_page;
 825         bp.init_num = bp.num = clen * blocks_per_page;
 826         bp.create = create;
 827
 828         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 829                bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
 830
 831         err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
 832                                      ldiskfs_ext_new_extent_cb, &bp);
 833         ldiskfs_ext_invalidate_cache(inode);
 834
 835         return err;
 836 }
 837
 838 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
 839                                           struct page **page, int pages,
 840                                           sector_t *blocks, int create)
 841 {
 842         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 843         pgoff_t bitmap_max_page_index;
 844         sector_t *b;
 845         int rc = 0, i;
 846
 847         bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
 848                                 PAGE_SHIFT;
 849         for (i = 0, b = blocks; i < pages; i++, page++) {
 850                 if ((*page)->index + 1 >= bitmap_max_page_index) {
 851                         rc = -EFBIG;
 852                         break;
 853                 }
 854                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
 855                 if (rc) {
 856                         CERROR("ino %lu, blk %llu create %d: rc %d\n",
 857                                inode->i_ino,
 858                                (unsigned long long)*b, create, rc);
 859                         break;
 860                 }
 861                 b += blocks_per_page;
 862         }
 863         return rc;
 864 }
 865
 866 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
 867                                            struct page **page,
 868                                            int pages, sector_t *blocks,
 869                                            int create)
 870 {
 871         int rc = 0, i = 0, clen = 0;
 872         struct page *fp = NULL;
 873
 874         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 875                 inode->i_ino, pages, (*page)->index);
 876
 877         /* pages are sorted already. so, we just have to find
 878          * contig. space and process them properly */
 879         while (i < pages) {
 880                 if (fp == NULL) {
 881                         /* start new extent */
 882                         fp = *page++;
 883                         clen = 1;
 884                         i++;
 885                         continue;
 886                 } else if (fp->index + clen == (*page)->index) {
 887                         /* continue the extent */
 888                         page++;
 889                         clen++;
 890                         i++;
 891                         continue;
 892                 }
 893
 894                 /* process found extent */
 895                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 896                                              blocks, create);
 897                 if (rc)
 898                         GOTO(cleanup, rc);
 899
 900                 /* look for next extent */
 901                 fp = NULL;
 902                 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
 903         }
 904
 905         if (fp)
 906                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 907                                              blocks, create);
 908
 909 cleanup:
 910         return rc;
 911 }
 912
 913 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 914                                        int pages, sector_t *blocks,
 915                                        int create)
 916 {
 917         int rc;
 918
 919         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 920                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
 921                                                      blocks, create);
 922                 return rc;
 923         }
 924         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
 925
 926         return rc;
 927 }
 928 #else
 929 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 930                                        int pages, sector_t *blocks,
 931                                        int create)
 932 {
 933         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 934         int rc = 0, i = 0;
 935         struct page *fp = NULL;
 936         int clen = 0;
 937         pgoff_t max_page_index;
 938         handle_t *handle = NULL;
 939
 940         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
 941
 942         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 943                 inode->i_ino, pages, (*page)->index);
 944
 945         if (create) {
 946                 create = LDISKFS_GET_BLOCKS_CREATE;
 947                 handle = ldiskfs_journal_current_handle();
 948                 LASSERT(handle != NULL);
 949                 rc = osd_attach_jinode(inode);
 950                 if (rc)
 951                         return rc;
 952         }
 953         /* pages are sorted already. so, we just have to find
 954          * contig. space and process them properly */
 955         while (i < pages) {
 956                 long blen, total = 0;
 957                 struct ldiskfs_map_blocks map = { 0 };
 958
 959                 if (fp == NULL) { /* start new extent */
 960                         fp = *page++;
 961                         clen = 1;
 962                         if (++i != pages)
 963                                 continue;
 964                 } else if (fp->index + clen == (*page)->index) {
 965                         /* continue the extent */
 966                         page++;
 967                         clen++;
 968                         if (++i != pages)
 969                                 continue;
 970                 }
 971                 if (fp->index + clen >= max_page_index)
 972                         GOTO(cleanup, rc = -EFBIG);
 973                 /* process found extent */
 974                 map.m_lblk = fp->index * blocks_per_page;
 975                 map.m_len = blen = clen * blocks_per_page;
 976 cont_map:
 977                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 978                 if (rc >= 0) {
 979                         int c = 0;
 980                         for (; total < blen && c < map.m_len; c++, total++) {
 981                                 if (rc == 0) {
 982                                         *(blocks + total) = 0;
 983                                         total++;
 984                                         break;
 985                                 } else {
 986                                         *(blocks + total) = map.m_pblk + c;
 987                                         /* unmap any possible underlying
 988                                          * metadata from the block device
 989                                          * mapping.  bug 6998. */
 990                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
 991                                             create)
 992 #ifndef HAVE_CLEAN_BDEV_ALIASES
 993                                                 unmap_underlying_metadata(
 994                                                         inode->i_sb->s_bdev,
 995                                                         map.m_pblk + c);
 996 #else
 997                                                 clean_bdev_aliases(
 998                                                         inode->i_sb->s_bdev,
 999                                                         map.m_pblk + c, 1);
1000 #endif
1001                                 }
1002                         }
1003                         rc = 0;
1004                 }
1005                 if (rc == 0 && total < blen) {
1006                         map.m_lblk = fp->index * blocks_per_page + total;
1007                         map.m_len = blen - total;
1008                         goto cont_map;
1009                 }
1010                 if (rc != 0)
1011                         GOTO(cleanup, rc);
1012
1013                 /* look for next extent */
1014                 fp = NULL;
1015                 blocks += blocks_per_page * clen;
1016         }
1017 cleanup:
1018         return rc;
1019 }
1020 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
1021
1022 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1023                           struct niobuf_local *lnb, int npages)
1024 {
1025         struct osd_thread_info *oti   = osd_oti_get(env);
1026         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1027         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
1028         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
1029         ktime_t start;
1030         ktime_t end;
1031         s64 timediff;
1032         ssize_t                 isize;
1033         __s64                   maxidx;
1034         int                     rc = 0;
1035         int                     i;
1036         int                     cache = 0;
1037
1038         LASSERT(inode);
1039
1040         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1041         if (unlikely(rc != 0))
1042                 RETURN(rc);
1043
1044         isize = i_size_read(inode);
1045         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1046
1047         if (osd->od_writethrough_cache)
1048                 cache = 1;
1049         if (isize > osd->od_readcache_max_filesize)
1050                 cache = 0;
1051
1052         start = ktime_get();
1053         for (i = 0; i < npages; i++) {
1054
1055                 if (cache == 0)
1056                         generic_error_remove_page(inode->i_mapping,
1057                                                   lnb[i].lnb_page);
1058
1059                 /*
1060                  * till commit the content of the page is undefined
1061                  * we'll set it uptodate once bulk is done. otherwise
1062                  * subsequent reads can access non-stable data
1063                  */
1064                 ClearPageUptodate(lnb[i].lnb_page);
1065
1066                 if (lnb[i].lnb_len == PAGE_SIZE)
1067                         continue;
1068
1069                 if (maxidx >= lnb[i].lnb_page->index) {
1070                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1071                 } else {
1072                         long off;
1073                         char *p = kmap(lnb[i].lnb_page);
1074
1075                         off = lnb[i].lnb_page_offset;
1076                         if (off)
1077                                 memset(p, 0, off);
1078                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1079                               ~PAGE_MASK;
1080                         if (off)
1081                                 memset(p + off, 0, PAGE_SIZE - off);
1082                         kunmap(lnb[i].lnb_page);
1083                 }
1084         }
1085         end = ktime_get();
1086         timediff = ktime_us_delta(end, start);
1087         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1088
1089         if (iobuf->dr_npages) {
1090                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1091                                                  iobuf->dr_npages,
1092                                                  iobuf->dr_blocks, 0);
1093                 if (likely(rc == 0)) {
1094                         rc = osd_do_bio(osd, inode, iobuf);
1095                         /* do IO stats for preparation reads */
1096                         osd_fini_iobuf(osd, iobuf);
1097                 }
1098         }
1099         RETURN(rc);
1100 }
1101
1102 struct osd_fextent {
1103         sector_t        start;
1104         sector_t        end;
1105         unsigned int    mapped:1;
1106 };
1107
1108 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1109                          struct osd_fextent *cached_extent)
1110 {
1111         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1112         sector_t block = offset >> inode->i_blkbits;
1113         sector_t start;
1114         struct fiemap_extent_info fei = { 0 };
1115         struct fiemap_extent fe = { 0 };
1116         mm_segment_t saved_fs;
1117         int rc;
1118
1119         if (block >= cached_extent->start && block < cached_extent->end)
1120                 return cached_extent->mapped;
1121
1122         if (i_size_read(inode) == 0)
1123                 return 0;
1124
1125         /* Beyond EOF, must not be mapped */
1126         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1127                 return 0;
1128
1129         fei.fi_extents_max = 1;
1130         fei.fi_extents_start = &fe;
1131
1132         saved_fs = get_fs();
1133         set_fs(get_ds());
1134         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1135         set_fs(saved_fs);
1136         if (rc != 0)
1137                 return 0;
1138
1139         start = fe.fe_logical >> inode->i_blkbits;
1140
1141         if (start > block) {
1142                 cached_extent->start = block;
1143                 cached_extent->end = start;
1144                 cached_extent->mapped = 0;
1145         } else {
1146                 cached_extent->start = start;
1147                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1148                                       inode->i_blkbits;
1149                 cached_extent->mapped = 1;
1150         }
1151
1152         return cached_extent->mapped;
1153 }
1154
1155 static int osd_declare_write_commit(const struct lu_env *env,
1156                                     struct dt_object *dt,
1157                                     struct niobuf_local *lnb, int npages,
1158                                     struct thandle *handle)
1159 {
1160         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1161         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1162         struct osd_thandle      *oh;
1163         int                     extents = 1;
1164         int                     depth;
1165         int                     i;
1166         int                     newblocks;
1167         int                     rc = 0;
1168         int                     flags = 0;
1169         int                     credits = 0;
1170         long long               quota_space = 0;
1171         struct osd_fextent      extent = { 0 };
1172         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1173         ENTRY;
1174
1175         LASSERT(handle != NULL);
1176         oh = container_of0(handle, struct osd_thandle, ot_super);
1177         LASSERT(oh->ot_handle == NULL);
1178
1179         newblocks = npages;
1180
1181         /* calculate number of extents (probably better to pass nb) */
1182         for (i = 0; i < npages; i++) {
1183                 if (i && lnb[i].lnb_file_offset !=
1184                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1185                         extents++;
1186
1187                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1188                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1189                 else
1190                         quota_space += PAGE_SIZE;
1191
1192                 /* ignore quota for the whole request if any page is from
1193                  * client cache or written by root.
1194                  *
1195                  * XXX once we drop the 1.8 client support, the checking
1196                  * for whether page is from cache can be simplified as:
1197                  * !(lnb[i].flags & OBD_BRW_SYNC)
1198                  *
1199                  * XXX we could handle this on per-lnb basis as done by
1200                  * grant. */
1201                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1202                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1203                     OBD_BRW_FROM_GRANT)
1204                         declare_flags |= OSD_QID_FORCE;
1205         }
1206
1207         /*
1208          * each extent can go into new leaf causing a split
1209          * 5 is max tree depth: inode + 4 index blocks
1210          * with blockmaps, depth is 3 at most
1211          */
1212         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1213                 /*
1214                  * many concurrent threads may grow tree by the time
1215                  * our transaction starts. so, consider 2 is a min depth
1216                  */
1217                 depth = ext_depth(inode);
1218                 depth = max(depth, 1) + 1;
1219                 newblocks += depth;
1220                 credits++; /* inode */
1221                 credits += depth * 2 * extents;
1222         } else {
1223                 depth = 3;
1224                 newblocks += depth;
1225                 credits++; /* inode */
1226                 credits += depth * extents;
1227         }
1228
1229         /* quota space for metadata blocks */
1230         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1231
1232         /* quota space should be reported in 1K blocks */
1233         quota_space = toqb(quota_space);
1234
1235         /* each new block can go in different group (bitmap + gd) */
1236
1237         /* we can't dirty more bitmap blocks than exist */
1238         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1239                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1240         else
1241                 credits += newblocks;
1242
1243         /* we can't dirty more gd blocks than exist */
1244         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1245                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1246         else
1247                 credits += newblocks;
1248
1249         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1250
1251         /* make sure the over quota flags were not set */
1252         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1253
1254         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1255                                    i_projid_read(inode), quota_space, oh,
1256                                    osd_dt_obj(dt), &flags, declare_flags);
1257
1258         /* we need only to store the overquota flags in the first lnb for
1259          * now, once we support multiple objects BRW, this code needs be
1260          * revised. */
1261         if (flags & QUOTA_FL_OVER_USRQUOTA)
1262                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1263         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1264                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1265         if (flags & QUOTA_FL_OVER_PRJQUOTA)
1266                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1267
1268         RETURN(rc);
1269 }
1270
1271 /* Check if a block is allocated or not */
1272 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1273                             struct niobuf_local *lnb, int npages,
1274                             struct thandle *thandle)
1275 {
1276         struct osd_thread_info *oti = osd_oti_get(env);
1277         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1278         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1279         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1280         loff_t isize;
1281         int rc = 0, i;
1282
1283         LASSERT(inode);
1284
1285         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1286         if (unlikely(rc != 0))
1287                 RETURN(rc);
1288
1289         isize = i_size_read(inode);
1290         ll_vfs_dq_init(inode);
1291
1292         for (i = 0; i < npages; i++) {
1293                 if (lnb[i].lnb_rc == -ENOSPC &&
1294                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1295                         /* Allow the write to proceed if overwriting an
1296                          * existing block */
1297                         lnb[i].lnb_rc = 0;
1298                 }
1299
1300                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1301                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1302                                lnb[i].lnb_rc);
1303                         LASSERT(lnb[i].lnb_page);
1304                         generic_error_remove_page(inode->i_mapping,
1305                                                   lnb[i].lnb_page);
1306                         continue;
1307                 }
1308
1309                 LASSERT(PageLocked(lnb[i].lnb_page));
1310                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1311
1312                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1313                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1314
1315                 /*
1316                  * Since write and truncate are serialized by oo_sem, even
1317                  * partial-page truncate should not leave dirty pages in the
1318                  * page cache.
1319                  */
1320                 LASSERT(!PageDirty(lnb[i].lnb_page));
1321
1322                 SetPageUptodate(lnb[i].lnb_page);
1323
1324                 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1325         }
1326
1327         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1328
1329         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1330                 rc = -ENOSPC;
1331         } else if (iobuf->dr_npages > 0) {
1332                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1333                                                  iobuf->dr_npages,
1334                                                  iobuf->dr_blocks, 1);
1335         } else {
1336                 /* no pages to write, no transno is needed */
1337                 thandle->th_local = 1;
1338         }
1339
1340         if (likely(rc == 0)) {
1341                 spin_lock(&inode->i_lock);
1342                 if (isize > i_size_read(inode)) {
1343                         i_size_write(inode, isize);
1344                         LDISKFS_I(inode)->i_disksize = isize;
1345                         spin_unlock(&inode->i_lock);
1346                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1347                 } else {
1348                         spin_unlock(&inode->i_lock);
1349                 }
1350
1351                 rc = osd_do_bio(osd, inode, iobuf);
1352                 /* we don't do stats here as in read path because
1353                  * write is async: we'll do this in osd_put_bufs() */
1354         } else {
1355                 osd_fini_iobuf(osd, iobuf);
1356         }
1357
1358         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1359
1360         if (unlikely(rc != 0)) {
1361                 /* if write fails, we should drop pages from the cache */
1362                 for (i = 0; i < npages; i++) {
1363                         if (lnb[i].lnb_page == NULL)
1364                                 continue;
1365                         LASSERT(PageLocked(lnb[i].lnb_page));
1366                         generic_error_remove_page(inode->i_mapping,
1367                                                   lnb[i].lnb_page);
1368                 }
1369         }
1370
1371         RETURN(rc);
1372 }
1373
1374 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1375                          struct niobuf_local *lnb, int npages)
1376 {
1377         struct osd_thread_info *oti = osd_oti_get(env);
1378         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1379         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1380         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1381         int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1382         ktime_t start, end;
1383         s64 timediff;
1384         loff_t isize;
1385
1386         LASSERT(inode);
1387
1388         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1389         if (unlikely(rc != 0))
1390                 RETURN(rc);
1391
1392         isize = i_size_read(inode);
1393
1394         if (osd->od_read_cache)
1395                 cache = 1;
1396         if (isize > osd->od_readcache_max_filesize)
1397                 cache = 0;
1398
1399         start = ktime_get();
1400         for (i = 0; i < npages; i++) {
1401
1402                 if (isize <= lnb[i].lnb_file_offset)
1403                         /* If there's no more data, abort early.
1404                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1405                         break;
1406
1407                 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1408                         lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1409                 else
1410                         lnb[i].lnb_rc = lnb[i].lnb_len;
1411
1412                 /* Bypass disk read if fail_loc is set properly */
1413                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1414                         SetPageUptodate(lnb[i].lnb_page);
1415
1416                 if (PageUptodate(lnb[i].lnb_page)) {
1417                         cache_hits++;
1418                 } else {
1419                         cache_misses++;
1420                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1421                 }
1422
1423                 if (cache == 0)
1424                         generic_error_remove_page(inode->i_mapping,
1425                                                   lnb[i].lnb_page);
1426         }
1427         end = ktime_get();
1428         timediff = ktime_us_delta(end, start);
1429         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1430
1431         if (cache_hits != 0)
1432                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1433                                     cache_hits);
1434         if (cache_misses != 0)
1435                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1436                                     cache_misses);
1437         if (cache_hits + cache_misses != 0)
1438                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1439                                     cache_hits + cache_misses);
1440
1441         if (iobuf->dr_npages) {
1442                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1443                                                  iobuf->dr_npages,
1444                                                  iobuf->dr_blocks, 0);
1445                 rc = osd_do_bio(osd, inode, iobuf);
1446
1447                 /* IO stats will be done in osd_bufs_put() */
1448         }
1449
1450         RETURN(rc);
1451 }
1452
1453 /*
1454  * XXX: Another layering violation for now.
1455  *
1456  * We don't want to use ->f_op->read methods, because generic file write
1457  *
1458  *         - serializes on ->i_sem, and
1459  *
1460  *         - does a lot of extra work like balance_dirty_pages(),
1461  *
1462  * which doesn't work for globally shared files like /last_rcvd.
1463  */
1464 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1465 {
1466         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1467
1468         memcpy(buffer, (char *)ei->i_data, buflen);
1469
1470         return  buflen;
1471 }
1472
1473 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1474 {
1475         struct buffer_head *bh;
1476         unsigned long block;
1477         int osize;
1478         int blocksize;
1479         int csize;
1480         int boffs;
1481
1482         /* prevent reading after eof */
1483         spin_lock(&inode->i_lock);
1484         if (i_size_read(inode) < *offs + size) {
1485                 loff_t diff = i_size_read(inode) - *offs;
1486                 spin_unlock(&inode->i_lock);
1487                 if (diff < 0) {
1488                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1489                                i_size_read(inode), *offs);
1490                         return -EBADR;
1491                 } else if (diff == 0) {
1492                         return 0;
1493                 } else {
1494                         size = diff;
1495                 }
1496         } else {
1497                 spin_unlock(&inode->i_lock);
1498         }
1499
1500         blocksize = 1 << inode->i_blkbits;
1501         osize = size;
1502         while (size > 0) {
1503                 block = *offs >> inode->i_blkbits;
1504                 boffs = *offs & (blocksize - 1);
1505                 csize = min(blocksize - boffs, size);
1506                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1507                 if (IS_ERR(bh)) {
1508                         CERROR("%s: can't read %u@%llu on ino %lu: "
1509                                "rc = %ld\n", osd_ino2name(inode),
1510                                csize, *offs, inode->i_ino,
1511                                PTR_ERR(bh));
1512                         return PTR_ERR(bh);
1513                 }
1514
1515                 if (bh != NULL) {
1516                         memcpy(buf, bh->b_data + boffs, csize);
1517                         brelse(bh);
1518                 } else {
1519                         memset(buf, 0, csize);
1520                 }
1521
1522                 *offs += csize;
1523                 buf += csize;
1524                 size -= csize;
1525         }
1526         return osize;
1527 }
1528
1529 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1530                         struct lu_buf *buf, loff_t *pos)
1531 {
1532         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1533         int           rc;
1534
1535         /* Read small symlink from inode body as we need to maintain correct
1536          * on-disk symlinks for ldiskfs.
1537          */
1538         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1539             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1540                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1541         else
1542                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1543
1544         return rc;
1545 }
1546
1547 static inline int osd_extents_enabled(struct super_block *sb,
1548                                       struct inode *inode)
1549 {
1550         if (inode != NULL) {
1551                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1552                         return 1;
1553         } else if (ldiskfs_has_feature_extents(sb)) {
1554                 return 1;
1555         }
1556         return 0;
1557 }
1558
1559 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1560                            const loff_t size, const loff_t pos,
1561                            const int blocks)
1562 {
1563         int credits, bits, bs, i;
1564
1565         bits = sb->s_blocksize_bits;
1566         bs = 1 << bits;
1567
1568         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1569          * we do not expect blockmaps on the large files,
1570          * so let's shrink it to 2 levels (4GB files) */
1571
1572         /* this is default reservation: 2 levels */
1573         credits = (blocks + 2) * 3;
1574
1575         /* actual offset is unknown, hard to optimize */
1576         if (pos == -1)
1577                 return credits;
1578
1579         /* now check for few specific cases to optimize */
1580         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1581                 /* no indirects */
1582                 credits = blocks;
1583                 /* allocate if not allocated */
1584                 if (inode == NULL) {
1585                         credits += blocks * 2;
1586                         return credits;
1587                 }
1588                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1589                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1590                         if (LDISKFS_I(inode)->i_data[i] == 0)
1591                                 credits += 2;
1592                 }
1593         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1594                 /* single indirect */
1595                 credits = blocks * 3;
1596                 if (inode == NULL ||
1597                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1598                         credits += 3;
1599                 else
1600                         /* The indirect block may be modified. */
1601                         credits += 1;
1602         }
1603
1604         return credits;
1605 }
1606
1607 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1608                                  const struct lu_buf *buf, loff_t _pos,
1609                                  struct thandle *handle)
1610 {
1611         struct osd_object  *obj  = osd_dt_obj(dt);
1612         struct inode       *inode = obj->oo_inode;
1613         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1614         struct osd_thandle *oh;
1615         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1616         int                 bits, bs;
1617         int                 depth, size;
1618         loff_t              pos;
1619         ENTRY;
1620
1621         LASSERT(buf != NULL);
1622         LASSERT(handle != NULL);
1623
1624         oh = container_of0(handle, struct osd_thandle, ot_super);
1625         LASSERT(oh->ot_handle == NULL);
1626
1627         size = buf->lb_len;
1628         bits = sb->s_blocksize_bits;
1629         bs = 1 << bits;
1630
1631         if (_pos == -1) {
1632                 /* if this is an append, then we
1633                  * should expect cross-block record */
1634                 pos = 0;
1635         } else {
1636                 pos = _pos;
1637         }
1638
1639         /* blocks to modify */
1640         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1641         LASSERT(blocks > 0);
1642
1643         if (inode != NULL && _pos != -1) {
1644                 /* object size in blocks */
1645                 est = (i_size_read(inode) + bs - 1) >> bits;
1646                 allocated = inode->i_blocks >> (bits - 9);
1647                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1648                         /* looks like an overwrite, no need to modify tree */
1649                         credits = blocks;
1650                         /* no need to modify i_size */
1651                         goto out;
1652                 }
1653         }
1654
1655         if (osd_extents_enabled(sb, inode)) {
1656                 /*
1657                  * many concurrent threads may grow tree by the time
1658                  * our transaction starts. so, consider 2 is a min depth
1659                  * for every level we may need to allocate a new block
1660                  * and take some entries from the old one. so, 3 blocks
1661                  * to allocate (bitmap, gd, itself) + old block - 4 per
1662                  * level.
1663                  */
1664                 depth = inode != NULL ? ext_depth(inode) : 0;
1665                 depth = max(depth, 1) + 1;
1666                 credits = depth;
1667                 /* if not append, then split may need to modify
1668                  * existing blocks moving entries into the new ones */
1669                 if (_pos != -1)
1670                         credits += depth;
1671                 /* blocks to store data: bitmap,gd,itself */
1672                 credits += blocks * 3;
1673         } else {
1674                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1675         }
1676         /* if inode is created as part of the transaction,
1677          * then it's counted already by the creation method */
1678         if (inode != NULL)
1679                 credits++;
1680
1681 out:
1682
1683         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1684
1685         /* dt_declare_write() is usually called for system objects, such
1686          * as llog or last_rcvd files. We needn't enforce quota on those
1687          * objects, so always set the lqi_space as 0. */
1688         if (inode != NULL)
1689                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1690                                            i_gid_read(inode),
1691                                            i_projid_read(inode), 0,
1692                                            oh, obj, NULL, OSD_QID_BLK);
1693         RETURN(rc);
1694 }
1695
1696 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1697 {
1698         /* LU-2634: clear the extent format for fast symlink */
1699         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1700
1701         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1702         spin_lock(&inode->i_lock);
1703         LDISKFS_I(inode)->i_disksize = buflen;
1704         i_size_write(inode, buflen);
1705         spin_unlock(&inode->i_lock);
1706         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1707
1708         return 0;
1709 }
1710
1711 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1712                              int write_NUL, loff_t *offs, handle_t *handle)
1713 {
1714         struct buffer_head *bh        = NULL;
1715         loff_t              offset    = *offs;
1716         loff_t              new_size  = i_size_read(inode);
1717         unsigned long       block;
1718         int                 blocksize = 1 << inode->i_blkbits;
1719         int                 err = 0;
1720         int                 size;
1721         int                 boffs;
1722         int                 dirty_inode = 0;
1723
1724         if (write_NUL) {
1725                 /*
1726                  * long symlink write does not count the NUL terminator in
1727                  * bufsize, we write it, and the inode's file size does not
1728                  * count the NUL terminator as well.
1729                  */
1730                 ((char *)buf)[bufsize] = '\0';
1731                 ++bufsize;
1732         }
1733
1734         while (bufsize > 0) {
1735                 int credits = handle->h_buffer_credits;
1736
1737                 if (bh)
1738                         brelse(bh);
1739
1740                 block = offset >> inode->i_blkbits;
1741                 boffs = offset & (blocksize - 1);
1742                 size = min(blocksize - boffs, bufsize);
1743                 bh = __ldiskfs_bread(handle, inode, block, 1);
1744                 if (IS_ERR_OR_NULL(bh)) {
1745                         if (bh == NULL) {
1746                                 err = -EIO;
1747                         } else {
1748                                 err = PTR_ERR(bh);
1749                                 bh = NULL;
1750                         }
1751
1752                         CERROR("%s: error reading offset %llu (block %lu, "
1753                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
1754                                inode->i_sb->s_id, offset, block, bufsize, *offs,
1755                                credits, handle->h_buffer_credits, err);
1756                         break;
1757                 }
1758
1759                 err = ldiskfs_journal_get_write_access(handle, bh);
1760                 if (err) {
1761                         CERROR("journal_get_write_access() returned error %d\n",
1762                                err);
1763                         break;
1764                 }
1765                 LASSERTF(boffs + size <= bh->b_size,
1766                          "boffs %d size %d bh->b_size %lu\n",
1767                          boffs, size, (unsigned long)bh->b_size);
1768                 memcpy(bh->b_data + boffs, buf, size);
1769                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1770                 if (err)
1771                         break;
1772
1773                 if (offset + size > new_size)
1774                         new_size = offset + size;
1775                 offset += size;
1776                 bufsize -= size;
1777                 buf += size;
1778         }
1779         if (bh)
1780                 brelse(bh);
1781
1782         if (write_NUL)
1783                 --new_size;
1784         /* correct in-core and on-disk sizes */
1785         if (new_size > i_size_read(inode)) {
1786                 spin_lock(&inode->i_lock);
1787                 if (new_size > i_size_read(inode))
1788                         i_size_write(inode, new_size);
1789                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1790                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1791                         dirty_inode = 1;
1792                 }
1793                 spin_unlock(&inode->i_lock);
1794                 if (dirty_inode)
1795                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1796         }
1797
1798         if (err == 0)
1799                 *offs = offset;
1800         return err;
1801 }
1802
1803 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1804                          const struct lu_buf *buf, loff_t *pos,
1805                          struct thandle *handle, int ignore_quota)
1806 {
1807         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1808         struct osd_thandle      *oh;
1809         ssize_t                 result;
1810         int                     is_link;
1811
1812         LASSERT(dt_object_exists(dt));
1813
1814         LASSERT(handle != NULL);
1815         LASSERT(inode != NULL);
1816         ll_vfs_dq_init(inode);
1817
1818         /* XXX: don't check: one declared chunk can be used many times */
1819         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1820
1821         oh = container_of(handle, struct osd_thandle, ot_super);
1822         LASSERT(oh->ot_handle->h_transaction != NULL);
1823         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1824
1825         /* Write small symlink to inode body as we need to maintain correct
1826          * on-disk symlinks for ldiskfs.
1827          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1828          * does not count it in.
1829          */
1830         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1831         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1832                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1833         else
1834                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1835                                                   buf->lb_len, is_link, pos,
1836                                                   oh->ot_handle);
1837         if (result == 0)
1838                 result = buf->lb_len;
1839
1840         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1841
1842         return result;
1843 }
1844
1845 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1846                              __u64 start, __u64 end, struct thandle *th)
1847 {
1848         struct osd_thandle *oh;
1849         struct inode       *inode;
1850         int                 rc;
1851         ENTRY;
1852
1853         LASSERT(th);
1854         oh = container_of(th, struct osd_thandle, ot_super);
1855
1856         /*
1857          * we don't need to reserve credits for whole truncate
1858          * it's not possible as truncate may need to free too many
1859          * blocks and that won't fit a single transaction. instead
1860          * we reserve credits to change i_size and put inode onto
1861          * orphan list. if needed truncate will extend or restart
1862          * transaction
1863          */
1864         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1865                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1866
1867         inode = osd_dt_obj(dt)->oo_inode;
1868         LASSERT(inode);
1869
1870         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1871                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1872                                    NULL, OSD_QID_BLK);
1873         RETURN(rc);
1874 }
1875
1876 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1877                      __u64 start, __u64 end, struct thandle *th)
1878 {
1879         struct osd_thandle *oh;
1880         struct osd_object  *obj = osd_dt_obj(dt);
1881         struct inode       *inode = obj->oo_inode;
1882         handle_t           *h;
1883         tid_t               tid;
1884         int                rc = 0, rc2 = 0;
1885         ENTRY;
1886
1887         LASSERT(end == OBD_OBJECT_EOF);
1888         LASSERT(dt_object_exists(dt));
1889         LASSERT(osd_invariant(obj));
1890         LASSERT(inode != NULL);
1891         ll_vfs_dq_init(inode);
1892
1893         LASSERT(th);
1894         oh = container_of(th, struct osd_thandle, ot_super);
1895         LASSERT(oh->ot_handle->h_transaction != NULL);
1896
1897         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1898
1899         tid = oh->ot_handle->h_transaction->t_tid;
1900
1901         spin_lock(&inode->i_lock);
1902         i_size_write(inode, start);
1903         spin_unlock(&inode->i_lock);
1904         ll_truncate_pagecache(inode, start);
1905 #ifdef HAVE_INODEOPS_TRUNCATE
1906         if (inode->i_op->truncate) {
1907                 inode->i_op->truncate(inode);
1908         } else
1909 #endif
1910                 ldiskfs_truncate(inode);
1911
1912         /*
1913          * For a partial-page truncate, flush the page to disk immediately to
1914          * avoid data corruption during direct disk write.  b=17397
1915          */
1916         if ((start & ~PAGE_MASK) != 0)
1917                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1918
1919         h = journal_current_handle();
1920         LASSERT(h != NULL);
1921         LASSERT(h == oh->ot_handle);
1922
1923         /* do not check credits with osd_trans_exec_check() as the truncate
1924          * can restart the transaction internally and we restart the
1925          * transaction in this case */
1926
1927         if (tid != h->h_transaction->t_tid) {
1928                 int credits = oh->ot_credits;
1929                 /*
1930                  * transaction has changed during truncate
1931                  * we need to restart the handle with our credits
1932                  */
1933                 if (h->h_buffer_credits < credits) {
1934                         if (ldiskfs_journal_extend(h, credits))
1935                                 rc2 = ldiskfs_journal_restart(h, credits);
1936                 }
1937         }
1938
1939         RETURN(rc == 0 ? rc2 : rc);
1940 }
1941
1942 static int fiemap_check_ranges(struct inode *inode,
1943                                u64 start, u64 len, u64 *new_len)
1944 {
1945         loff_t maxbytes;
1946
1947         *new_len = len;
1948
1949         if (len == 0)
1950                 return -EINVAL;
1951
1952         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1953                 maxbytes = inode->i_sb->s_maxbytes;
1954         else
1955                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1956
1957         if (start > maxbytes)
1958                 return -EFBIG;
1959
1960         /*
1961          * Shrink request scope to what the fs can actually handle.
1962          */
1963         if (len > maxbytes || (maxbytes - len) < start)
1964                 *new_len = maxbytes - start;
1965
1966         return 0;
1967 }
1968
1969 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1970 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
1971
1972 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1973                           struct fiemap *fm)
1974 {
1975         struct fiemap_extent_info fieinfo = {0, };
1976         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1977         u64 len;
1978         int rc;
1979
1980
1981         LASSERT(inode);
1982         if (inode->i_op->fiemap == NULL)
1983                 return -EOPNOTSUPP;
1984
1985         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1986                 return -EINVAL;
1987
1988         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1989         if (rc)
1990                 return rc;
1991
1992         fieinfo.fi_flags = fm->fm_flags;
1993         fieinfo.fi_extents_max = fm->fm_extent_count;
1994         fieinfo.fi_extents_start = fm->fm_extents;
1995
1996         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1997                 filemap_write_and_wait(inode->i_mapping);
1998
1999         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2000         fm->fm_flags = fieinfo.fi_flags;
2001         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2002
2003         return rc;
2004 }
2005
2006 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2007                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2008 {
2009         int              rc = 0;
2010         struct inode    *inode = osd_dt_obj(dt)->oo_inode;
2011         ENTRY;
2012
2013         switch (advice) {
2014         case LU_LADVISE_DONTNEED:
2015                 if (end == 0)
2016                         break;
2017                 invalidate_mapping_pages(inode->i_mapping,
2018                                          start >> PAGE_SHIFT,
2019                                          (end - 1) >> PAGE_SHIFT);
2020                 break;
2021         default:
2022                 rc = -ENOTSUPP;
2023                 break;
2024         }
2025
2026         RETURN(rc);
2027 }
2028
2029 /*
2030  * in some cases we may need declare methods for objects being created
2031  * e.g., when we create symlink
2032  */
2033 const struct dt_body_operations osd_body_ops_new = {
2034         .dbo_declare_write = osd_declare_write,
2035 };
2036
2037 const struct dt_body_operations osd_body_ops = {
2038         .dbo_read                       = osd_read,
2039         .dbo_declare_write              = osd_declare_write,
2040         .dbo_write                      = osd_write,
2041         .dbo_bufs_get                   = osd_bufs_get,
2042         .dbo_bufs_put                   = osd_bufs_put,
2043         .dbo_write_prep                 = osd_write_prep,
2044         .dbo_declare_write_commit       = osd_declare_write_commit,
2045         .dbo_write_commit               = osd_write_commit,
2046         .dbo_read_prep                  = osd_read_prep,
2047         .dbo_declare_punch              = osd_declare_punch,
2048         .dbo_punch                      = osd_punch,
2049         .dbo_fiemap_get                 = osd_fiemap_get,
2050         .dbo_ladvise                    = osd_ladvise,
2051 };