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 inline bool osd_use_page_cache(struct osd_device *d)
  60 {
  61         /* do not use pagecache if write and read caching are disabled */
  62         if (d->od_writethrough_cache + d->od_read_cache == 0)
  63                 return false;
  64         /* use pagecache by default */
  65         return true;
  66 }
  67
  68 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  69                             int rw, int line, int pages)
  70 {
  71         int blocks, i;
  72
  73         LASSERTF(iobuf->dr_elapsed_valid == 0,
  74                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  75                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  76                  iobuf->dr_init_at);
  77         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  78
  79         init_waitqueue_head(&iobuf->dr_wait);
  80         atomic_set(&iobuf->dr_numreqs, 0);
  81         iobuf->dr_npages = 0;
  82         iobuf->dr_error = 0;
  83         iobuf->dr_dev = d;
  84         iobuf->dr_frags = 0;
  85         iobuf->dr_elapsed = ktime_set(0, 0);
  86         /* must be counted before, so assert */
  87         iobuf->dr_rw = rw;
  88         iobuf->dr_init_at = line;
  89
  90         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  91         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  92                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  93                         pages * sizeof(iobuf->dr_pages[0]));
  94                 return 0;
  95         }
  96
  97         /* start with 1MB for 4K blocks */
  98         i = 256;
  99         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
 100                 i <<= 1;
 101
 102         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
 103                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
 104         pages = i;
 105         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
 106         iobuf->dr_max_pages = 0;
 107         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
 108                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 109
 110         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 111         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 112         if (unlikely(iobuf->dr_blocks == NULL))
 113                 return -ENOMEM;
 114
 115         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 116         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 117         if (unlikely(iobuf->dr_pages == NULL))
 118                 return -ENOMEM;
 119
 120         lu_buf_realloc(&iobuf->dr_lnb_buf,
 121                        pages * sizeof(iobuf->dr_lnbs[0]));
 122         iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf;
 123         if (unlikely(iobuf->dr_lnbs == NULL))
 124                 return -ENOMEM;
 125
 126         iobuf->dr_max_pages = pages;
 127
 128         return 0;
 129 }
 130 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 131         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 132
 133 static void osd_iobuf_add_page(struct osd_iobuf *iobuf,
 134                                struct niobuf_local *lnb)
 135 {
 136         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 137         iobuf->dr_pages[iobuf->dr_npages] = lnb->lnb_page;
 138         iobuf->dr_lnbs[iobuf->dr_npages] = lnb;
 139         iobuf->dr_npages++;
 140 }
 141
 142 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 143 {
 144         int rw = iobuf->dr_rw;
 145
 146         if (iobuf->dr_elapsed_valid) {
 147                 iobuf->dr_elapsed_valid = 0;
 148                 LASSERT(iobuf->dr_dev == d);
 149                 LASSERT(iobuf->dr_frags > 0);
 150                 lprocfs_oh_tally(&d->od_brw_stats.
 151                                  hist[BRW_R_DIO_FRAGS+rw],
 152                                  iobuf->dr_frags);
 153                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 154                                       ktime_to_ms(iobuf->dr_elapsed));
 155         }
 156 }
 157
 158 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 159 static void dio_complete_routine(struct bio *bio)
 160 {
 161 # ifdef HAVE_BI_STATUS
 162         int error = bio->bi_status;
 163 # else
 164         int error = bio->bi_error;
 165 # endif
 166 #else
 167 static void dio_complete_routine(struct bio *bio, int error)
 168 {
 169 #endif
 170         struct osd_iobuf *iobuf = bio->bi_private;
 171         int iter;
 172         struct bio_vec *bvl;
 173
 174         /* CAVEAT EMPTOR: possibly in IRQ context
 175          * DO NOT record procfs stats here!!! */
 176
 177         if (unlikely(iobuf == NULL)) {
 178                 CERROR("***** bio->bi_private is NULL!  This should never "
 179                        "happen.  Normally, I would crash here, but instead I "
 180                        "will dump the bio contents to the console.  Please "
 181                        "report this to <https://jira.whamcloud.com/> , along "
 182                        "with any interesting messages leading up to this point "
 183                        "(like SCSI errors, perhaps).  Because bi_private is "
 184                        "NULL, I can't wake up the thread that initiated this "
 185                        "IO - you will probably have to reboot this node.\n");
 186                 CERROR("bi_next: %p, bi_flags: %lx, "
 187 #ifdef HAVE_BI_RW
 188                        "bi_rw: %lu,"
 189 #else
 190                        "bi_opf: %u,"
 191 #endif
 192                        "bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p,"
 193                        "bi_cnt: %d, bi_private: %p\n", bio->bi_next,
 194                         (unsigned long)bio->bi_flags,
 195 #ifdef HAVE_BI_RW
 196                         bio->bi_rw,
 197 #else
 198                         bio->bi_opf,
 199 #endif
 200                         bio->bi_vcnt, bio_idx(bio),
 201                         bio_sectors(bio) << 9, bio->bi_end_io,
 202 #ifdef HAVE_BI_CNT
 203                         atomic_read(&bio->bi_cnt),
 204 #else
 205                         atomic_read(&bio->__bi_cnt),
 206 #endif
 207                         bio->bi_private);
 208                 return;
 209         }
 210
 211         /* the check is outside of the cycle for performance reason -bzzz */
 212         if (!bio_data_dir(bio)) {
 213                 bio_for_each_segment_all(bvl, bio, iter) {
 214                         if (likely(error == 0))
 215                                 SetPageUptodate(bvl_to_page(bvl));
 216                         LASSERT(PageLocked(bvl_to_page(bvl)));
 217                 }
 218                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 219         } else {
 220                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 221         }
 222
 223         /* any real error is good enough -bzzz */
 224         if (error != 0 && iobuf->dr_error == 0)
 225                 iobuf->dr_error = error;
 226
 227         /*
 228          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 229          * it's possible that service thread will see dr_numreqs
 230          * is zero, but dr_elapsed is not set yet, leading to lost
 231          * data in this processing and an assertion in a subsequent
 232          * call to OSD.
 233          */
 234         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 235                 ktime_t now = ktime_get();
 236
 237                 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
 238                 iobuf->dr_elapsed_valid = 1;
 239         }
 240         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 241                 wake_up(&iobuf->dr_wait);
 242
 243         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 244          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 245          * mempool when serious on-disk fragmentation was encountered,
 246          * deadlocking the OST.  The bios are now released as soon as complete
 247          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 248         bio_put(bio);
 249 }
 250
 251 static void record_start_io(struct osd_iobuf *iobuf, int size)
 252 {
 253         struct osd_device    *osd = iobuf->dr_dev;
 254         struct obd_histogram *h = osd->od_brw_stats.hist;
 255
 256         iobuf->dr_frags++;
 257         atomic_inc(&iobuf->dr_numreqs);
 258
 259         if (iobuf->dr_rw == 0) {
 260                 atomic_inc(&osd->od_r_in_flight);
 261                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 262                                  atomic_read(&osd->od_r_in_flight));
 263                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 264         } else if (iobuf->dr_rw == 1) {
 265                 atomic_inc(&osd->od_w_in_flight);
 266                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 267                                  atomic_read(&osd->od_w_in_flight));
 268                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 269         } else {
 270                 LBUG();
 271         }
 272 }
 273
 274 static void osd_submit_bio(int rw, struct bio *bio)
 275 {
 276         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 277 #ifdef HAVE_SUBMIT_BIO_2ARGS
 278         if (rw == 0)
 279                 submit_bio(READ, bio);
 280         else
 281                 submit_bio(WRITE, bio);
 282 #else
 283         bio->bi_opf |= rw;
 284         submit_bio(bio);
 285 #endif
 286 }
 287
 288 static int can_be_merged(struct bio *bio, sector_t sector)
 289 {
 290         if (bio == NULL)
 291                 return 0;
 292
 293         return bio_end_sector(bio) == sector ? 1 : 0;
 294 }
 295
 296 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
 297 /*
 298  * This function will change the data written, thus it should only be
 299  * used when checking data integrity feature
 300  */
 301 static void bio_integrity_fault_inject(struct bio *bio)
 302 {
 303         struct bio_vec *bvec;
 304         int i;
 305         void *kaddr;
 306         char *addr;
 307
 308         bio_for_each_segment_all(bvec, bio, i) {
 309                 struct page *page = bvec->bv_page;
 310
 311                 kaddr = kmap(page);
 312                 addr = kaddr;
 313                 *addr = ~(*addr);
 314                 kunmap(page);
 315                 break;
 316         }
 317 }
 318
 319 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
 320                            unsigned int sectors, int tuple_size)
 321 {
 322         __u16 *expected_guard;
 323         __u16 *bio_guard;
 324         int i;
 325
 326         expected_guard = expected_guard_buf;
 327         for (i = 0; i < sectors; i++) {
 328                 bio_guard = (__u16 *)bio_prot_buf;
 329                 if (*bio_guard != *expected_guard) {
 330                         CERROR("unexpected guard tags on sector %d "
 331                                "expected guard %u, bio guard "
 332                                "%u, sectors %u, tuple size %d\n",
 333                                i, *expected_guard, *bio_guard, sectors,
 334                                tuple_size);
 335                         return -EIO;
 336                 }
 337                 expected_guard++;
 338                 bio_prot_buf += tuple_size;
 339         }
 340         return 0;
 341 }
 342
 343 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
 344                                      struct osd_iobuf *iobuf, int index)
 345 {
 346         struct blk_integrity *bi = bdev_get_integrity(bdev);
 347         struct bio_integrity_payload *bip = bio->bi_integrity;
 348         struct niobuf_local *lnb;
 349         unsigned short sector_size = blk_integrity_interval(bi);
 350         void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
 351                 bip->bip_vec->bv_offset;
 352         struct bio_vec *bv;
 353         sector_t sector = bio_start_sector(bio);
 354         unsigned int i, sectors, total;
 355         __u16 *expected_guard;
 356         int rc;
 357
 358         total = 0;
 359         bio_for_each_segment_all(bv, bio, i) {
 360                 lnb = iobuf->dr_lnbs[index];
 361                 expected_guard = lnb->lnb_guards;
 362                 sectors = bv->bv_len / sector_size;
 363                 if (lnb->lnb_guard_rpc) {
 364                         rc = bio_dif_compare(expected_guard, bio_prot_buf,
 365                                              sectors, bi->tuple_size);
 366                         if (rc)
 367                                 return rc;
 368                 }
 369
 370                 sector += sectors;
 371                 bio_prot_buf += sectors * bi->tuple_size;
 372                 total += sectors * bi->tuple_size;
 373                 LASSERT(total <= bip_size(bio->bi_integrity));
 374                 index++;
 375         }
 376         return 0;
 377 }
 378
 379 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
 380                                     struct osd_iobuf *iobuf,
 381                                     int start_page_idx, bool fault_inject,
 382                                     bool integrity_enabled)
 383 {
 384         struct super_block *sb = osd_sb(osd);
 385         int rc;
 386 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 387         integrity_gen_fn *generate_fn = NULL;
 388         integrity_vrfy_fn *verify_fn = NULL;
 389 #endif
 390
 391         ENTRY;
 392
 393         if (!integrity_enabled)
 394                 RETURN(0);
 395
 396 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 397         rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
 398         if (rc)
 399                 RETURN(rc);
 400
 401         rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
 402 #else
 403         rc = bio_integrity_prep(bio);
 404 #endif
 405         if (rc)
 406                 RETURN(rc);
 407
 408         /* Verify and inject fault only when writing */
 409         if (iobuf->dr_rw == 1) {
 410                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
 411                         rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
 412                                                        start_page_idx);
 413                         if (rc)
 414                                 RETURN(rc);
 415                 }
 416
 417                 if (unlikely(fault_inject))
 418                         bio_integrity_fault_inject(bio);
 419         }
 420
 421         RETURN(0);
 422 }
 423
 424 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 425 #  ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 426 static void dio_integrity_complete_routine(struct bio *bio)
 427 {
 428 #  else
 429 static void dio_integrity_complete_routine(struct bio *bio, int error)
 430 {
 431 #  endif
 432         struct osd_bio_private *bio_private = bio->bi_private;
 433
 434         bio->bi_private = bio_private->obp_iobuf;
 435 #  ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 436         dio_complete_routine(bio);
 437 #  else
 438         dio_complete_routine(bio, error);
 439 #  endif
 440
 441         OBD_FREE_PTR(bio_private);
 442 }
 443 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
 444 #else  /* !CONFIG_BLK_DEV_INTEGRITY */
 445 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
 446                                  fault_inject, integrity_enabled) 0
 447 #endif /* CONFIG_BLK_DEV_INTEGRITY */
 448
 449 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
 450                         bool integrity_enabled, int start_page_idx,
 451                         struct osd_bio_private **pprivate)
 452 {
 453 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 454         struct osd_bio_private *bio_private;
 455
 456         ENTRY;
 457
 458         *pprivate = NULL;
 459         if (integrity_enabled) {
 460                 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
 461                 if (bio_private == NULL)
 462                         RETURN(-ENOMEM);
 463                 bio->bi_end_io = dio_integrity_complete_routine;
 464                 bio->bi_private = bio_private;
 465                 bio_private->obp_start_page_idx = start_page_idx;
 466                 bio_private->obp_iobuf = iobuf;
 467                 *pprivate = bio_private;
 468         } else {
 469                 bio->bi_end_io = dio_complete_routine;
 470                 bio->bi_private = iobuf;
 471         }
 472         RETURN(0);
 473 #else
 474         ENTRY;
 475
 476         bio->bi_end_io = dio_complete_routine;
 477         bio->bi_private = iobuf;
 478         RETURN(0);
 479 #endif
 480 }
 481
 482 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 483                       struct osd_iobuf *iobuf)
 484 {
 485         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 486         struct page **pages = iobuf->dr_pages;
 487         int npages = iobuf->dr_npages;
 488         sector_t *blocks = iobuf->dr_blocks;
 489         int total_blocks = npages * blocks_per_page;
 490         struct super_block *sb = inode->i_sb;
 491         int sector_bits = sb->s_blocksize_bits - 9;
 492         unsigned int blocksize = sb->s_blocksize;
 493         struct block_device *bdev = sb->s_bdev;
 494         struct osd_bio_private *bio_private = NULL;
 495         struct bio *bio = NULL;
 496         int bio_start_page_idx;
 497         struct page *page;
 498         unsigned int page_offset;
 499         sector_t sector;
 500         int nblocks;
 501         int block_idx;
 502         int page_idx;
 503         int i;
 504         int rc = 0;
 505         bool fault_inject;
 506         bool integrity_enabled;
 507         DECLARE_PLUG(plug);
 508         ENTRY;
 509
 510         fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
 511         LASSERT(iobuf->dr_npages == npages);
 512
 513         integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
 514
 515         osd_brw_stats_update(osd, iobuf);
 516         iobuf->dr_start_time = ktime_get();
 517
 518         blk_start_plug(&plug);
 519         for (page_idx = 0, block_idx = 0;
 520              page_idx < npages;
 521              page_idx++, block_idx += blocks_per_page) {
 522
 523                 page = pages[page_idx];
 524                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 525
 526                 for (i = 0, page_offset = 0;
 527                      i < blocks_per_page;
 528                      i += nblocks, page_offset += blocksize * nblocks) {
 529
 530                         nblocks = 1;
 531
 532                         if (blocks[block_idx + i] == 0) {  /* hole */
 533                                 LASSERTF(iobuf->dr_rw == 0,
 534                                          "page_idx %u, block_idx %u, i %u\n",
 535                                          page_idx, block_idx, i);
 536                                 memset(kmap(page) + page_offset, 0, blocksize);
 537                                 kunmap(page);
 538                                 continue;
 539                         }
 540
 541                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 542
 543                         /* Additional contiguous file blocks? */
 544                         while (i + nblocks < blocks_per_page &&
 545                                (sector + (nblocks << sector_bits)) ==
 546                                ((sector_t)blocks[block_idx + i + nblocks] <<
 547                                 sector_bits))
 548                                 nblocks++;
 549
 550                         if (bio != NULL &&
 551                             can_be_merged(bio, sector) &&
 552                             bio_add_page(bio, page,
 553                                          blocksize * nblocks, page_offset) != 0)
 554                                 continue;       /* added this frag OK */
 555
 556                         if (bio != NULL) {
 557                                 struct request_queue *q = bio_get_queue(bio);
 558                                 unsigned int bi_size = bio_sectors(bio) << 9;
 559
 560                                 /* Dang! I have to fragment this I/O */
 561                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 562                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 563                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 564                                        bio_sectors(bio),
 565                                        queue_max_sectors(q),
 566                                        bio_phys_segments(q, bio),
 567                                        queue_max_phys_segments(q),
 568                                        0, queue_max_hw_segments(q));
 569                                 rc = osd_bio_integrity_handle(osd, bio,
 570                                         iobuf, bio_start_page_idx,
 571                                         fault_inject, integrity_enabled);
 572                                 if (rc) {
 573                                         bio_put(bio);
 574                                         goto out;
 575                                 }
 576
 577                                 record_start_io(iobuf, bi_size);
 578                                 osd_submit_bio(iobuf->dr_rw, bio);
 579                         }
 580
 581                         bio_start_page_idx = page_idx;
 582                         /* allocate new bio */
 583                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 584                                                       (npages - page_idx) *
 585                                                       blocks_per_page));
 586                         if (bio == NULL) {
 587                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 588                                        (npages - page_idx), blocks_per_page,
 589                                        (npages - page_idx) * blocks_per_page);
 590                                 rc = -ENOMEM;
 591                                 goto out;
 592                         }
 593
 594                         bio_set_dev(bio, bdev);
 595                         bio_set_sector(bio, sector);
 596 #ifdef HAVE_BI_RW
 597                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 598 #else
 599                         bio->bi_opf = (iobuf->dr_rw == 0) ? READ : WRITE;
 600 #endif
 601                         rc = osd_bio_init(bio, iobuf, integrity_enabled,
 602                                           bio_start_page_idx, &bio_private);
 603                         if (rc) {
 604                                 bio_put(bio);
 605                                 goto out;
 606                         }
 607
 608                         rc = bio_add_page(bio, page,
 609                                           blocksize * nblocks, page_offset);
 610                         LASSERT(rc != 0);
 611                 }
 612         }
 613
 614         if (bio != NULL) {
 615                 rc = osd_bio_integrity_handle(osd, bio, iobuf,
 616                                               bio_start_page_idx,
 617                                               fault_inject,
 618                                               integrity_enabled);
 619                 if (rc) {
 620                         bio_put(bio);
 621                         goto out;
 622                 }
 623
 624                 record_start_io(iobuf, bio_sectors(bio) << 9);
 625                 osd_submit_bio(iobuf->dr_rw, bio);
 626                 rc = 0;
 627         }
 628
 629 out:
 630         blk_finish_plug(&plug);
 631
 632         /* in order to achieve better IO throughput, we don't wait for writes
 633          * completion here. instead we proceed with transaction commit in
 634          * parallel and wait for IO completion once transaction is stopped
 635          * see osd_trans_stop() for more details -bzzz */
 636         if (iobuf->dr_rw == 0 || fault_inject) {
 637                 wait_event(iobuf->dr_wait,
 638                            atomic_read(&iobuf->dr_numreqs) == 0);
 639                 osd_fini_iobuf(osd, iobuf);
 640         }
 641
 642         if (rc == 0) {
 643                 rc = iobuf->dr_error;
 644         } else {
 645                 if (bio_private)
 646                         OBD_FREE_PTR(bio_private);
 647         }
 648
 649         RETURN(rc);
 650 }
 651
 652 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 653                                    struct niobuf_local *lnb, int maxlnb)
 654 {
 655         int rc = 0;
 656         ENTRY;
 657
 658         *nrpages = 0;
 659
 660         while (len > 0) {
 661                 int poff = offset & (PAGE_SIZE - 1);
 662                 int plen = PAGE_SIZE - poff;
 663
 664                 if (*nrpages >= maxlnb) {
 665                         rc = -EOVERFLOW;
 666                         break;
 667                 }
 668
 669                 if (plen > len)
 670                         plen = len;
 671                 lnb->lnb_file_offset = offset;
 672                 lnb->lnb_page_offset = poff;
 673                 lnb->lnb_len = plen;
 674                 /* lnb->lnb_flags = rnb->rnb_flags; */
 675                 lnb->lnb_flags = 0;
 676                 lnb->lnb_page = NULL;
 677                 lnb->lnb_rc = 0;
 678                 lnb->lnb_guard_rpc = 0;
 679                 lnb->lnb_guard_disk = 0;
 680                 lnb->lnb_locked = 0;
 681
 682                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 683                          (long long) len);
 684                 offset += plen;
 685                 len -= plen;
 686                 lnb++;
 687                 (*nrpages)++;
 688         }
 689
 690         RETURN(rc);
 691 }
 692
 693 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
 694                                  loff_t offset, gfp_t gfp_mask)
 695 {
 696         struct osd_thread_info *oti = osd_oti_get(env);
 697         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 698         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 699         struct page *page;
 700         int cur = oti->oti_dio_pages_used;
 701
 702         LASSERT(inode);
 703
 704         if (osd_use_page_cache(d)) {
 705                 page = find_or_create_page(inode->i_mapping,
 706                                            offset >> PAGE_SHIFT,
 707                                            gfp_mask);
 708
 709                 if (likely(page))
 710                         LASSERT(!test_bit(PG_private_2, &page->flags));
 711                 else
 712                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 713         } else {
 714
 715                 LASSERT(oti->oti_dio_pages);
 716
 717                 if (unlikely(!oti->oti_dio_pages[cur])) {
 718                         LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
 719                         page = alloc_page(gfp_mask);
 720                         if (!page)
 721                                 return NULL;
 722                         oti->oti_dio_pages[cur] = page;
 723                 }
 724
 725                 page = oti->oti_dio_pages[cur];
 726                 LASSERT(!test_bit(PG_private_2, &page->flags));
 727                 set_bit(PG_private_2, &page->flags);
 728                 oti->oti_dio_pages_used++;
 729
 730                 LASSERT(!PageLocked(page));
 731                 lock_page(page);
 732
 733                 LASSERT(!page->mapping);
 734                 LASSERT(!PageWriteback(page));
 735                 ClearPageUptodate(page);
 736
 737                 page->index = offset >> PAGE_SHIFT;
 738         }
 739
 740         return page;
 741 }
 742
 743 /*
 744  * there are following "locks":
 745  * journal_start
 746  * i_mutex
 747  * page lock
 748  *
 749  * osd write path:
 750  *  - lock page(s)
 751  *  - journal_start
 752  *  - truncate_sem
 753  *
 754  * ext4 vmtruncate:
 755  *  - lock pages, unlock
 756  *  - journal_start
 757  *  - lock partial page
 758  *  - i_data_sem
 759  *
 760  */
 761
 762 /**
 763  * Unlock and release pages loaded by osd_bufs_get()
 764  *
 765  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 766  *
 767  * \param env           thread execution environment
 768  * \param dt            dt object undergoing IO (OSD object + methods)
 769  * \param lnb           array of pages undergoing IO
 770  * \param npages        number of pages in \a lnb
 771  *
 772  * \retval 0            always
 773  */
 774 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 775                         struct niobuf_local *lnb, int npages)
 776 {
 777         struct osd_thread_info *oti = osd_oti_get(env);
 778         struct pagevec pvec;
 779         int i;
 780
 781         ll_pagevec_init(&pvec, 0);
 782
 783         for (i = 0; i < npages; i++) {
 784                 struct page *page = lnb[i].lnb_page;
 785
 786                 if (page == NULL)
 787                         continue;
 788
 789                 /* if the page isn't cached, then reset uptodate
 790                  * to prevent reuse */
 791                 if (test_bit(PG_private_2, &page->flags)) {
 792                         clear_bit(PG_private_2, &page->flags);
 793                         ClearPageUptodate(page);
 794                         if (lnb[i].lnb_locked)
 795                                 unlock_page(page);
 796                         oti->oti_dio_pages_used--;
 797                 } else {
 798                         if (lnb[i].lnb_locked)
 799                                 unlock_page(page);
 800                         if (pagevec_add(&pvec, page) == 0)
 801                                 pagevec_release(&pvec);
 802                 }
 803                 dt_object_put(env, dt);
 804
 805                 lnb[i].lnb_page = NULL;
 806         }
 807
 808         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
 809
 810         /* Release any partial pagevec */
 811         pagevec_release(&pvec);
 812
 813         RETURN(0);
 814 }
 815
 816 /**
 817  * Load and lock pages undergoing IO
 818  *
 819  * Pages as described in the \a lnb array are fetched (from disk or cache)
 820  * and locked for IO by the caller.
 821  *
 822  * DLM locking protects us from write and truncate competing for same region,
 823  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 824  * It's possible the writeout on a such a page is in progress when we access
 825  * it. It's also possible that during this writeout we put new (partial) data
 826  * into the page, but won't be able to proceed in filter_commitrw_write().
 827  * Therefore, just wait for writeout completion as it should be rare enough.
 828  *
 829  * \param env           thread execution environment
 830  * \param dt            dt object undergoing IO (OSD object + methods)
 831  * \param pos           byte offset of IO start
 832  * \param len           number of bytes of IO
 833  * \param lnb           array of extents undergoing IO
 834  * \param rw            read or write operation, and other flags
 835  * \param capa          capabilities
 836  *
 837  * \retval pages        (zero or more) loaded successfully
 838  * \retval -ENOMEM      on memory/page allocation error
 839  */
 840 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 841                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 842                         int maxlnb, enum dt_bufs_type rw)
 843 {
 844         struct osd_thread_info *oti = osd_oti_get(env);
 845         struct osd_object *obj = osd_dt_obj(dt);
 846         int npages, i, rc = 0;
 847         gfp_t gfp_mask;
 848
 849         LASSERT(obj->oo_inode);
 850
 851         if (!osd_use_page_cache(osd_obj2dev(obj))) {
 852                 if (unlikely(!oti->oti_dio_pages)) {
 853                         OBD_ALLOC(oti->oti_dio_pages,
 854                                   sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
 855                         if (!oti->oti_dio_pages)
 856                                 return -ENOMEM;
 857                 }
 858         }
 859
 860         rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
 861         if (rc)
 862                 RETURN(rc);
 863
 864         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 865         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 866                                              GFP_HIGHUSER;
 867         for (i = 0; i < npages; i++, lnb++) {
 868                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
 869                                              gfp_mask);
 870                 if (lnb->lnb_page == NULL)
 871                         GOTO(cleanup, rc = -ENOMEM);
 872
 873                 lnb->lnb_locked = 1;
 874                 wait_on_page_writeback(lnb->lnb_page);
 875                 BUG_ON(PageWriteback(lnb->lnb_page));
 876
 877                 lu_object_get(&dt->do_lu);
 878         }
 879
 880         RETURN(i);
 881
 882 cleanup:
 883         if (i > 0)
 884                 osd_bufs_put(env, dt, lnb - i, i);
 885         return rc;
 886 }
 887
 888 #ifndef HAVE_LDISKFS_MAP_BLOCKS
 889
 890 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 891 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 892 #endif
 893
 894 struct bpointers {
 895         sector_t *blocks;
 896         unsigned long start;
 897         int num;
 898         int init_num;
 899         int create;
 900 };
 901
 902 static long ldiskfs_ext_find_goal(struct inode *inode,
 903                                   struct ldiskfs_ext_path *path,
 904                                   unsigned long block, int *aflags)
 905 {
 906         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 907         unsigned long bg_start;
 908         unsigned long colour;
 909         int depth;
 910
 911         if (path) {
 912                 struct ldiskfs_extent *ex;
 913                 depth = path->p_depth;
 914
 915                 /* try to predict block placement */
 916                 if ((ex = path[depth].p_ext))
 917                         return ldiskfs_ext_pblock(ex) +
 918                                 (block - le32_to_cpu(ex->ee_block));
 919
 920                 /* it looks index is empty
 921                  * try to find starting from index itself */
 922                 if (path[depth].p_bh)
 923                         return path[depth].p_bh->b_blocknr;
 924         }
 925
 926         /* OK. use inode's group */
 927         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 928                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 929         colour = (current->pid % 16) *
 930                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 931         return bg_start + colour + block;
 932 }
 933
 934 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 935                                 struct ldiskfs_ext_path *path,
 936                                 unsigned long block, unsigned long *count,
 937                                 int *err)
 938 {
 939         struct ldiskfs_allocation_request ar;
 940         unsigned long pblock;
 941         int aflags;
 942
 943         /* find neighbour allocated blocks */
 944         ar.lleft = block;
 945         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 946         if (*err)
 947                 return 0;
 948         ar.lright = block;
 949         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 950         if (*err)
 951                 return 0;
 952
 953         /* allocate new block */
 954         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 955         ar.inode = inode;
 956         ar.logical = block;
 957         ar.len = *count;
 958         ar.flags = LDISKFS_MB_HINT_DATA;
 959         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 960         *count = ar.len;
 961         return pblock;
 962 }
 963
 964 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 965                                      struct ldiskfs_ext_path *path,
 966                                      struct ldiskfs_ext_cache *cex,
 967 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 968                                      struct ldiskfs_extent *ex,
 969 #endif
 970                                      void *cbdata)
 971 {
 972         struct bpointers *bp = cbdata;
 973         struct ldiskfs_extent nex;
 974         unsigned long pblock = 0;
 975         unsigned long tgen;
 976         int err, i;
 977         unsigned long count;
 978         handle_t *handle;
 979
 980 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 981         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 982 #else
 983         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 984 #endif
 985                 err = EXT_CONTINUE;
 986                 goto map;
 987         }
 988
 989         if (bp->create == 0) {
 990                 i = 0;
 991                 if (cex->ec_block < bp->start)
 992                         i = bp->start - cex->ec_block;
 993                 if (i >= cex->ec_len)
 994                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 995                                         i, cex->ec_len);
 996                 for (; i < cex->ec_len && bp->num; i++) {
 997                         *(bp->blocks) = 0;
 998                         bp->blocks++;
 999                         bp->num--;
1000                         bp->start++;
1001                 }
1002
1003                 return EXT_CONTINUE;
1004         }
1005
1006         tgen = LDISKFS_I(inode)->i_ext_generation;
1007         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
1008
1009         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
1010                                    count + LDISKFS_ALLOC_NEEDED + 1);
1011         if (IS_ERR(handle)) {
1012                 return PTR_ERR(handle);
1013         }
1014
1015         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
1016                 /* the tree has changed. so path can be invalid at moment */
1017                 ldiskfs_journal_stop(handle);
1018                 return EXT_REPEAT;
1019         }
1020
1021         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
1022          * protected by i_data_sem as whole. so we patch it to store
1023          * generation to path and now verify the tree hasn't changed */
1024         down_write((&LDISKFS_I(inode)->i_data_sem));
1025
1026         /* validate extent, make sure the extent tree does not changed */
1027         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
1028                 /* cex is invalid, try again */
1029                 up_write(&LDISKFS_I(inode)->i_data_sem);
1030                 ldiskfs_journal_stop(handle);
1031                 return EXT_REPEAT;
1032         }
1033
1034         count = cex->ec_len;
1035         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
1036         if (!pblock)
1037                 goto out;
1038         BUG_ON(count > cex->ec_len);
1039
1040         /* insert new extent */
1041         nex.ee_block = cpu_to_le32(cex->ec_block);
1042         ldiskfs_ext_store_pblock(&nex, pblock);
1043         nex.ee_len = cpu_to_le16(count);
1044         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
1045         if (err) {
1046                 /* free data blocks we just allocated */
1047                 /* not a good idea to call discard here directly,
1048                  * but otherwise we'd need to call it every free() */
1049                 ldiskfs_discard_preallocations(inode);
1050 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
1051                 ldiskfs_free_blocks(handle, inode, NULL,
1052                                     ldiskfs_ext_pblock(&nex),
1053                                     le16_to_cpu(nex.ee_len), 0);
1054 #else
1055                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
1056                                     le16_to_cpu(nex.ee_len), 0);
1057 #endif
1058                 goto out;
1059         }
1060
1061         /*
1062          * Putting len of the actual extent we just inserted,
1063          * we are asking ldiskfs_ext_walk_space() to continue
1064          * scaning after that block
1065          */
1066         cex->ec_len = le16_to_cpu(nex.ee_len);
1067         cex->ec_start = ldiskfs_ext_pblock(&nex);
1068         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
1069         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
1070
1071 out:
1072         up_write((&LDISKFS_I(inode)->i_data_sem));
1073         ldiskfs_journal_stop(handle);
1074 map:
1075         if (err >= 0) {
1076                 /* map blocks */
1077                 if (bp->num == 0) {
1078                         CERROR("hmm. why do we find this extent?\n");
1079                         CERROR("initial space: %lu:%u\n",
1080                                 bp->start, bp->init_num);
1081 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
1082                         CERROR("current extent: %u/%u/%llu %d\n",
1083                                 cex->ec_block, cex->ec_len,
1084                                 (unsigned long long)cex->ec_start,
1085                                 cex->ec_type);
1086 #else
1087                         CERROR("current extent: %u/%u/%llu\n",
1088                                 cex->ec_block, cex->ec_len,
1089                                 (unsigned long long)cex->ec_start);
1090 #endif
1091                 }
1092                 i = 0;
1093                 if (cex->ec_block < bp->start)
1094                         i = bp->start - cex->ec_block;
1095                 if (i >= cex->ec_len)
1096                         CERROR("nothing to do?! i = %d, e_num = %u\n",
1097                                         i, cex->ec_len);
1098                 for (; i < cex->ec_len && bp->num; i++) {
1099                         *(bp->blocks) = cex->ec_start + i;
1100                         if (pblock != 0) {
1101                                 /* unmap any possible underlying metadata from
1102                                  * the block device mapping.  bug 6998. */
1103 #ifndef HAVE_CLEAN_BDEV_ALIASES
1104                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
1105                                                           *(bp->blocks));
1106 #else
1107                                 clean_bdev_aliases(inode->i_sb->s_bdev,
1108                                                    *(bp->blocks), 1);
1109 #endif
1110                         }
1111                         bp->blocks++;
1112                         bp->num--;
1113                         bp->start++;
1114                 }
1115         }
1116         return err;
1117 }
1118
1119 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
1120                                    int clen, sector_t *blocks, int create)
1121 {
1122         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1123         struct bpointers bp;
1124         int err;
1125
1126         if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
1127                 return -EFBIG;
1128
1129         bp.blocks = blocks;
1130         bp.start = index * blocks_per_page;
1131         bp.init_num = bp.num = clen * blocks_per_page;
1132         bp.create = create;
1133
1134         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
1135                bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
1136
1137         err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
1138                                      ldiskfs_ext_new_extent_cb, &bp);
1139         ldiskfs_ext_invalidate_cache(inode);
1140
1141         return err;
1142 }
1143
1144 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
1145                                           struct page **page, int pages,
1146                                           sector_t *blocks, int create)
1147 {
1148         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1149         pgoff_t bitmap_max_page_index;
1150         sector_t *b;
1151         int rc = 0, i;
1152
1153         bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
1154                                 PAGE_SHIFT;
1155         for (i = 0, b = blocks; i < pages; i++, page++) {
1156                 if ((*page)->index + 1 >= bitmap_max_page_index) {
1157                         rc = -EFBIG;
1158                         break;
1159                 }
1160                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
1161                 if (rc) {
1162                         CERROR("ino %lu, blk %llu create %d: rc %d\n",
1163                                inode->i_ino,
1164                                (unsigned long long)*b, create, rc);
1165                         break;
1166                 }
1167                 b += blocks_per_page;
1168         }
1169         return rc;
1170 }
1171
1172 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
1173                                            struct page **page,
1174                                            int pages, sector_t *blocks,
1175                                            int create)
1176 {
1177         int rc = 0, i = 0, clen = 0;
1178         struct page *fp = NULL;
1179
1180         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1181                 inode->i_ino, pages, (*page)->index);
1182
1183         /* pages are sorted already. so, we just have to find
1184          * contig. space and process them properly */
1185         while (i < pages) {
1186                 if (fp == NULL) {
1187                         /* start new extent */
1188                         fp = *page++;
1189                         clen = 1;
1190                         i++;
1191                         continue;
1192                 } else if (fp->index + clen == (*page)->index) {
1193                         /* continue the extent */
1194                         page++;
1195                         clen++;
1196                         i++;
1197                         continue;
1198                 }
1199
1200                 /* process found extent */
1201                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
1202                                              blocks, create);
1203                 if (rc)
1204                         GOTO(cleanup, rc);
1205
1206                 /* look for next extent */
1207                 fp = NULL;
1208                 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
1209         }
1210
1211         if (fp)
1212                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
1213                                              blocks, create);
1214
1215 cleanup:
1216         return rc;
1217 }
1218
1219 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
1220                                        int pages, sector_t *blocks,
1221                                        int create)
1222 {
1223         int rc;
1224
1225         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1226                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
1227                                                      blocks, create);
1228                 return rc;
1229         }
1230         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
1231
1232         return rc;
1233 }
1234 #else
1235 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
1236                                        int pages, sector_t *blocks,
1237                                        int create)
1238 {
1239         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1240         int rc = 0, i = 0;
1241         struct page *fp = NULL;
1242         int clen = 0;
1243         pgoff_t max_page_index;
1244         handle_t *handle = NULL;
1245
1246         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
1247
1248         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1249                 inode->i_ino, pages, (*page)->index);
1250
1251         if (create) {
1252                 create = LDISKFS_GET_BLOCKS_CREATE;
1253                 handle = ldiskfs_journal_current_handle();
1254                 LASSERT(handle != NULL);
1255                 rc = osd_attach_jinode(inode);
1256                 if (rc)
1257                         return rc;
1258         }
1259         /* pages are sorted already. so, we just have to find
1260          * contig. space and process them properly */
1261         while (i < pages) {
1262                 long blen, total = 0;
1263                 struct ldiskfs_map_blocks map = { 0 };
1264
1265                 if (fp == NULL) { /* start new extent */
1266                         fp = *page++;
1267                         clen = 1;
1268                         if (++i != pages)
1269                                 continue;
1270                 } else if (fp->index + clen == (*page)->index) {
1271                         /* continue the extent */
1272                         page++;
1273                         clen++;
1274                         if (++i != pages)
1275                                 continue;
1276                 }
1277                 if (fp->index + clen >= max_page_index)
1278                         GOTO(cleanup, rc = -EFBIG);
1279                 /* process found extent */
1280                 map.m_lblk = fp->index * blocks_per_page;
1281                 map.m_len = blen = clen * blocks_per_page;
1282 cont_map:
1283                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
1284                 if (rc >= 0) {
1285                         int c = 0;
1286                         for (; total < blen && c < map.m_len; c++, total++) {
1287                                 if (rc == 0) {
1288                                         *(blocks + total) = 0;
1289                                         total++;
1290                                         break;
1291                                 } else {
1292                                         *(blocks + total) = map.m_pblk + c;
1293                                         /* unmap any possible underlying
1294                                          * metadata from the block device
1295                                          * mapping.  bug 6998. */
1296                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
1297                                             create)
1298 #ifndef HAVE_CLEAN_BDEV_ALIASES
1299                                                 unmap_underlying_metadata(
1300                                                         inode->i_sb->s_bdev,
1301                                                         map.m_pblk + c);
1302 #else
1303                                                 clean_bdev_aliases(
1304                                                         inode->i_sb->s_bdev,
1305                                                         map.m_pblk + c, 1);
1306 #endif
1307                                 }
1308                         }
1309                         rc = 0;
1310                 }
1311                 if (rc == 0 && total < blen) {
1312                         map.m_lblk = fp->index * blocks_per_page + total;
1313                         map.m_len = blen - total;
1314                         goto cont_map;
1315                 }
1316                 if (rc != 0)
1317                         GOTO(cleanup, rc);
1318
1319                 /* look for next extent */
1320                 fp = NULL;
1321                 blocks += blocks_per_page * clen;
1322         }
1323 cleanup:
1324         return rc;
1325 }
1326 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
1327
1328 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1329                           struct niobuf_local *lnb, int npages)
1330 {
1331         struct osd_thread_info *oti   = osd_oti_get(env);
1332         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1333         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
1334         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
1335         ktime_t start;
1336         ktime_t end;
1337         s64 timediff;
1338         ssize_t                 isize;
1339         __s64                   maxidx;
1340         int                     rc = 0;
1341         int                     i;
1342         int                     cache = 0;
1343
1344         LASSERT(inode);
1345
1346         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1347         if (unlikely(rc != 0))
1348                 RETURN(rc);
1349
1350         isize = i_size_read(inode);
1351         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1352
1353         if (osd->od_writethrough_cache)
1354                 cache = 1;
1355         if (isize > osd->od_readcache_max_filesize)
1356                 cache = 0;
1357
1358         start = ktime_get();
1359         for (i = 0; i < npages; i++) {
1360
1361                 if (cache == 0)
1362                         generic_error_remove_page(inode->i_mapping,
1363                                                   lnb[i].lnb_page);
1364
1365                 /*
1366                  * till commit the content of the page is undefined
1367                  * we'll set it uptodate once bulk is done. otherwise
1368                  * subsequent reads can access non-stable data
1369                  */
1370                 ClearPageUptodate(lnb[i].lnb_page);
1371
1372                 if (lnb[i].lnb_len == PAGE_SIZE)
1373                         continue;
1374
1375                 if (maxidx >= lnb[i].lnb_page->index) {
1376                         osd_iobuf_add_page(iobuf, &lnb[i]);
1377                 } else {
1378                         long off;
1379                         char *p = kmap(lnb[i].lnb_page);
1380
1381                         off = lnb[i].lnb_page_offset;
1382                         if (off)
1383                                 memset(p, 0, off);
1384                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1385                               ~PAGE_MASK;
1386                         if (off)
1387                                 memset(p + off, 0, PAGE_SIZE - off);
1388                         kunmap(lnb[i].lnb_page);
1389                 }
1390         }
1391         end = ktime_get();
1392         timediff = ktime_us_delta(end, start);
1393         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1394
1395         if (iobuf->dr_npages) {
1396                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1397                                                  iobuf->dr_npages,
1398                                                  iobuf->dr_blocks, 0);
1399                 if (likely(rc == 0)) {
1400                         rc = osd_do_bio(osd, inode, iobuf);
1401                         /* do IO stats for preparation reads */
1402                         osd_fini_iobuf(osd, iobuf);
1403                 }
1404         }
1405         RETURN(rc);
1406 }
1407
1408 struct osd_fextent {
1409         sector_t        start;
1410         sector_t        end;
1411         unsigned int    mapped:1;
1412 };
1413
1414 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1415                          struct osd_fextent *cached_extent)
1416 {
1417         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1418         sector_t block = offset >> inode->i_blkbits;
1419         sector_t start;
1420         struct fiemap_extent_info fei = { 0 };
1421         struct fiemap_extent fe = { 0 };
1422         mm_segment_t saved_fs;
1423         int rc;
1424
1425         if (block >= cached_extent->start && block < cached_extent->end)
1426                 return cached_extent->mapped;
1427
1428         if (i_size_read(inode) == 0)
1429                 return 0;
1430
1431         /* Beyond EOF, must not be mapped */
1432         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1433                 return 0;
1434
1435         fei.fi_extents_max = 1;
1436         fei.fi_extents_start = &fe;
1437
1438         saved_fs = get_fs();
1439         set_fs(get_ds());
1440         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1441         set_fs(saved_fs);
1442         if (rc != 0)
1443                 return 0;
1444
1445         start = fe.fe_logical >> inode->i_blkbits;
1446
1447         if (start > block) {
1448                 cached_extent->start = block;
1449                 cached_extent->end = start;
1450                 cached_extent->mapped = 0;
1451         } else {
1452                 cached_extent->start = start;
1453                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1454                                       inode->i_blkbits;
1455                 cached_extent->mapped = 1;
1456         }
1457
1458         return cached_extent->mapped;
1459 }
1460
1461 static int osd_declare_write_commit(const struct lu_env *env,
1462                                     struct dt_object *dt,
1463                                     struct niobuf_local *lnb, int npages,
1464                                     struct thandle *handle)
1465 {
1466         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1467         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1468         struct osd_thandle      *oh;
1469         int                     extents = 1;
1470         int                     depth;
1471         int                     i;
1472         int                     newblocks;
1473         int                     rc = 0;
1474         int                     credits = 0;
1475         long long               quota_space = 0;
1476         struct osd_fextent      extent = { 0 };
1477         enum osd_quota_local_flags local_flags = 0;
1478         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1479         ENTRY;
1480
1481         LASSERT(handle != NULL);
1482         oh = container_of0(handle, struct osd_thandle, ot_super);
1483         LASSERT(oh->ot_handle == NULL);
1484
1485         newblocks = npages;
1486
1487         /* calculate number of extents (probably better to pass nb) */
1488         for (i = 0; i < npages; i++) {
1489                 if (i && lnb[i].lnb_file_offset !=
1490                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1491                         extents++;
1492
1493                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1494                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1495                 else
1496                         quota_space += PAGE_SIZE;
1497
1498                 /* ignore quota for the whole request if any page is from
1499                  * client cache or written by root.
1500                  *
1501                  * XXX once we drop the 1.8 client support, the checking
1502                  * for whether page is from cache can be simplified as:
1503                  * !(lnb[i].flags & OBD_BRW_SYNC)
1504                  *
1505                  * XXX we could handle this on per-lnb basis as done by
1506                  * grant. */
1507                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1508                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1509                     OBD_BRW_FROM_GRANT)
1510                         declare_flags |= OSD_QID_FORCE;
1511         }
1512
1513         /*
1514          * each extent can go into new leaf causing a split
1515          * 5 is max tree depth: inode + 4 index blocks
1516          * with blockmaps, depth is 3 at most
1517          */
1518         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1519                 /*
1520                  * many concurrent threads may grow tree by the time
1521                  * our transaction starts. so, consider 2 is a min depth
1522                  */
1523                 depth = ext_depth(inode);
1524                 depth = max(depth, 1) + 1;
1525                 newblocks += depth;
1526                 credits++; /* inode */
1527                 credits += depth * 2 * extents;
1528         } else {
1529                 depth = 3;
1530                 newblocks += depth;
1531                 credits++; /* inode */
1532                 credits += depth * extents;
1533         }
1534
1535         /* quota space for metadata blocks */
1536         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1537
1538         /* quota space should be reported in 1K blocks */
1539         quota_space = toqb(quota_space);
1540
1541         /* each new block can go in different group (bitmap + gd) */
1542
1543         /* we can't dirty more bitmap blocks than exist */
1544         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1545                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1546         else
1547                 credits += newblocks;
1548
1549         /* we can't dirty more gd blocks than exist */
1550         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1551                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1552         else
1553                 credits += newblocks;
1554
1555         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1556
1557         /* make sure the over quota flags were not set */
1558         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1559
1560         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1561                                    i_projid_read(inode), quota_space, oh,
1562                                    osd_dt_obj(dt), &local_flags, declare_flags);
1563
1564         /* we need only to store the overquota flags in the first lnb for
1565          * now, once we support multiple objects BRW, this code needs be
1566          * revised. */
1567         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1568                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1569         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1570                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1571         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1572                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1573
1574         if (rc == 0)
1575                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1576
1577         RETURN(rc);
1578 }
1579
1580 /* Check if a block is allocated or not */
1581 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1582                             struct niobuf_local *lnb, int npages,
1583                             struct thandle *thandle)
1584 {
1585         struct osd_thread_info *oti = osd_oti_get(env);
1586         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1587         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1588         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1589         loff_t isize;
1590         int rc = 0, i;
1591
1592         LASSERT(inode);
1593
1594         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1595         if (unlikely(rc != 0))
1596                 RETURN(rc);
1597
1598         isize = i_size_read(inode);
1599         ll_vfs_dq_init(inode);
1600
1601         for (i = 0; i < npages; i++) {
1602                 if (lnb[i].lnb_rc == -ENOSPC &&
1603                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1604                         /* Allow the write to proceed if overwriting an
1605                          * existing block */
1606                         lnb[i].lnb_rc = 0;
1607                 }
1608
1609                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1610                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1611                                lnb[i].lnb_rc);
1612                         LASSERT(lnb[i].lnb_page);
1613                         generic_error_remove_page(inode->i_mapping,
1614                                                   lnb[i].lnb_page);
1615                         continue;
1616                 }
1617
1618                 LASSERT(PageLocked(lnb[i].lnb_page));
1619                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1620
1621                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1622                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1623
1624                 /*
1625                  * Since write and truncate are serialized by oo_sem, even
1626                  * partial-page truncate should not leave dirty pages in the
1627                  * page cache.
1628                  */
1629                 LASSERT(!PageDirty(lnb[i].lnb_page));
1630
1631                 SetPageUptodate(lnb[i].lnb_page);
1632
1633                 osd_iobuf_add_page(iobuf, &lnb[i]);
1634         }
1635
1636         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1637
1638         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1639                 rc = -ENOSPC;
1640         } else if (iobuf->dr_npages > 0) {
1641                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1642                                                  iobuf->dr_npages,
1643                                                  iobuf->dr_blocks, 1);
1644         } else {
1645                 /* no pages to write, no transno is needed */
1646                 thandle->th_local = 1;
1647         }
1648
1649         if (likely(rc == 0)) {
1650                 spin_lock(&inode->i_lock);
1651                 if (isize > i_size_read(inode)) {
1652                         i_size_write(inode, isize);
1653                         LDISKFS_I(inode)->i_disksize = isize;
1654                         spin_unlock(&inode->i_lock);
1655                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1656                 } else {
1657                         spin_unlock(&inode->i_lock);
1658                 }
1659
1660                 rc = osd_do_bio(osd, inode, iobuf);
1661                 /* we don't do stats here as in read path because
1662                  * write is async: we'll do this in osd_put_bufs() */
1663         } else {
1664                 osd_fini_iobuf(osd, iobuf);
1665         }
1666
1667         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1668
1669         if (unlikely(rc != 0)) {
1670                 /* if write fails, we should drop pages from the cache */
1671                 for (i = 0; i < npages; i++) {
1672                         if (lnb[i].lnb_page == NULL)
1673                                 continue;
1674                         LASSERT(PageLocked(lnb[i].lnb_page));
1675                         generic_error_remove_page(inode->i_mapping,
1676                                                   lnb[i].lnb_page);
1677                 }
1678         }
1679
1680         RETURN(rc);
1681 }
1682
1683 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1684                          struct niobuf_local *lnb, int npages)
1685 {
1686         struct osd_thread_info *oti = osd_oti_get(env);
1687         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1688         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1689         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1690         int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1691         ktime_t start, end;
1692         s64 timediff;
1693         loff_t isize;
1694
1695         LASSERT(inode);
1696
1697         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1698         if (unlikely(rc != 0))
1699                 RETURN(rc);
1700
1701         isize = i_size_read(inode);
1702
1703         if (osd->od_read_cache)
1704                 cache = 1;
1705         if (isize > osd->od_readcache_max_filesize)
1706                 cache = 0;
1707
1708         start = ktime_get();
1709         for (i = 0; i < npages; i++) {
1710
1711                 if (isize <= lnb[i].lnb_file_offset)
1712                         /* If there's no more data, abort early.
1713                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1714                         break;
1715
1716                 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1717                         lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1718                 else
1719                         lnb[i].lnb_rc = lnb[i].lnb_len;
1720
1721                 /* Bypass disk read if fail_loc is set properly */
1722                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1723                         SetPageUptodate(lnb[i].lnb_page);
1724
1725                 if (cache == 0)
1726                         generic_error_remove_page(inode->i_mapping,
1727                                                   lnb[i].lnb_page);
1728
1729                 if (PageUptodate(lnb[i].lnb_page)) {
1730                         cache_hits++;
1731                         unlock_page(lnb[i].lnb_page);
1732                 } else {
1733                         cache_misses++;
1734                         osd_iobuf_add_page(iobuf, &lnb[i]);
1735                 }
1736                 /* no need to unlock in osd_bufs_put(), the sooner page is
1737                  * unlocked, the earlier another client can access it.
1738                  * notice real unlock_page() can be called few lines
1739                  * below after osd_do_bio(). lnb is a per-thread, so it's
1740                  * fine to have PG_locked and lnb_locked inconsistent here */
1741                 lnb[i].lnb_locked = 0;
1742         }
1743         end = ktime_get();
1744         timediff = ktime_us_delta(end, start);
1745         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1746
1747         if (cache_hits != 0)
1748                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1749                                     cache_hits);
1750         if (cache_misses != 0)
1751                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1752                                     cache_misses);
1753         if (cache_hits + cache_misses != 0)
1754                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1755                                     cache_hits + cache_misses);
1756
1757         if (iobuf->dr_npages) {
1758                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1759                                                  iobuf->dr_npages,
1760                                                  iobuf->dr_blocks, 0);
1761                 rc = osd_do_bio(osd, inode, iobuf);
1762
1763                 /* IO stats will be done in osd_bufs_put() */
1764
1765                 /* early release to let others read data during the bulk */
1766                 for (i = 0; i < iobuf->dr_npages; i++) {
1767                         LASSERT(PageLocked(iobuf->dr_pages[i]));
1768                         unlock_page(iobuf->dr_pages[i]);
1769                 }
1770         }
1771
1772         RETURN(rc);
1773 }
1774
1775 /*
1776  * XXX: Another layering violation for now.
1777  *
1778  * We don't want to use ->f_op->read methods, because generic file write
1779  *
1780  *         - serializes on ->i_sem, and
1781  *
1782  *         - does a lot of extra work like balance_dirty_pages(),
1783  *
1784  * which doesn't work for globally shared files like /last_rcvd.
1785  */
1786 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1787 {
1788         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1789
1790         memcpy(buffer, (char *)ei->i_data, buflen);
1791
1792         return  buflen;
1793 }
1794
1795 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1796 {
1797         struct buffer_head *bh;
1798         unsigned long block;
1799         int osize;
1800         int blocksize;
1801         int csize;
1802         int boffs;
1803
1804         /* prevent reading after eof */
1805         spin_lock(&inode->i_lock);
1806         if (i_size_read(inode) < *offs + size) {
1807                 loff_t diff = i_size_read(inode) - *offs;
1808                 spin_unlock(&inode->i_lock);
1809                 if (diff < 0) {
1810                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1811                                i_size_read(inode), *offs);
1812                         return -EBADR;
1813                 } else if (diff == 0) {
1814                         return 0;
1815                 } else {
1816                         size = diff;
1817                 }
1818         } else {
1819                 spin_unlock(&inode->i_lock);
1820         }
1821
1822         blocksize = 1 << inode->i_blkbits;
1823         osize = size;
1824         while (size > 0) {
1825                 block = *offs >> inode->i_blkbits;
1826                 boffs = *offs & (blocksize - 1);
1827                 csize = min(blocksize - boffs, size);
1828                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1829                 if (IS_ERR(bh)) {
1830                         CERROR("%s: can't read %u@%llu on ino %lu: "
1831                                "rc = %ld\n", osd_ino2name(inode),
1832                                csize, *offs, inode->i_ino,
1833                                PTR_ERR(bh));
1834                         return PTR_ERR(bh);
1835                 }
1836
1837                 if (bh != NULL) {
1838                         memcpy(buf, bh->b_data + boffs, csize);
1839                         brelse(bh);
1840                 } else {
1841                         memset(buf, 0, csize);
1842                 }
1843
1844                 *offs += csize;
1845                 buf += csize;
1846                 size -= csize;
1847         }
1848         return osize;
1849 }
1850
1851 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1852                         struct lu_buf *buf, loff_t *pos)
1853 {
1854         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1855         int rc;
1856
1857         /* Read small symlink from inode body as we need to maintain correct
1858          * on-disk symlinks for ldiskfs.
1859          */
1860         if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1861                 loff_t size = i_size_read(inode);
1862
1863                 if (buf->lb_len < size)
1864                         return -EOVERFLOW;
1865
1866                 if (size < sizeof(LDISKFS_I(inode)->i_data))
1867                         rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1868                 else
1869                         rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1870         } else {
1871                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1872         }
1873
1874         return rc;
1875 }
1876
1877 static inline int osd_extents_enabled(struct super_block *sb,
1878                                       struct inode *inode)
1879 {
1880         if (inode != NULL) {
1881                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1882                         return 1;
1883         } else if (ldiskfs_has_feature_extents(sb)) {
1884                 return 1;
1885         }
1886         return 0;
1887 }
1888
1889 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1890                            const loff_t size, const loff_t pos,
1891                            const int blocks)
1892 {
1893         int credits, bits, bs, i;
1894
1895         bits = sb->s_blocksize_bits;
1896         bs = 1 << bits;
1897
1898         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1899          * we do not expect blockmaps on the large files,
1900          * so let's shrink it to 2 levels (4GB files) */
1901
1902         /* this is default reservation: 2 levels */
1903         credits = (blocks + 2) * 3;
1904
1905         /* actual offset is unknown, hard to optimize */
1906         if (pos == -1)
1907                 return credits;
1908
1909         /* now check for few specific cases to optimize */
1910         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1911                 /* no indirects */
1912                 credits = blocks;
1913                 /* allocate if not allocated */
1914                 if (inode == NULL) {
1915                         credits += blocks * 2;
1916                         return credits;
1917                 }
1918                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1919                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1920                         if (LDISKFS_I(inode)->i_data[i] == 0)
1921                                 credits += 2;
1922                 }
1923         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1924                 /* single indirect */
1925                 credits = blocks * 3;
1926                 if (inode == NULL ||
1927                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1928                         credits += 3;
1929                 else
1930                         /* The indirect block may be modified. */
1931                         credits += 1;
1932         }
1933
1934         return credits;
1935 }
1936
1937 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1938                                  const struct lu_buf *buf, loff_t _pos,
1939                                  struct thandle *handle)
1940 {
1941         struct osd_object  *obj  = osd_dt_obj(dt);
1942         struct inode       *inode = obj->oo_inode;
1943         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1944         struct osd_thandle *oh;
1945         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1946         int                 bits, bs;
1947         int                 depth, size;
1948         loff_t              pos;
1949         ENTRY;
1950
1951         LASSERT(buf != NULL);
1952         LASSERT(handle != NULL);
1953
1954         oh = container_of0(handle, struct osd_thandle, ot_super);
1955         LASSERT(oh->ot_handle == NULL);
1956
1957         size = buf->lb_len;
1958         bits = sb->s_blocksize_bits;
1959         bs = 1 << bits;
1960
1961         if (_pos == -1) {
1962                 /* if this is an append, then we
1963                  * should expect cross-block record */
1964                 pos = 0;
1965         } else {
1966                 pos = _pos;
1967         }
1968
1969         /* blocks to modify */
1970         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1971         LASSERT(blocks > 0);
1972
1973         if (inode != NULL && _pos != -1) {
1974                 /* object size in blocks */
1975                 est = (i_size_read(inode) + bs - 1) >> bits;
1976                 allocated = inode->i_blocks >> (bits - 9);
1977                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1978                         /* looks like an overwrite, no need to modify tree */
1979                         credits = blocks;
1980                         /* no need to modify i_size */
1981                         goto out;
1982                 }
1983         }
1984
1985         if (osd_extents_enabled(sb, inode)) {
1986                 /*
1987                  * many concurrent threads may grow tree by the time
1988                  * our transaction starts. so, consider 2 is a min depth
1989                  * for every level we may need to allocate a new block
1990                  * and take some entries from the old one. so, 3 blocks
1991                  * to allocate (bitmap, gd, itself) + old block - 4 per
1992                  * level.
1993                  */
1994                 depth = inode != NULL ? ext_depth(inode) : 0;
1995                 depth = max(depth, 1) + 1;
1996                 credits = depth;
1997                 /* if not append, then split may need to modify
1998                  * existing blocks moving entries into the new ones */
1999                 if (_pos != -1)
2000                         credits += depth;
2001                 /* blocks to store data: bitmap,gd,itself */
2002                 credits += blocks * 3;
2003         } else {
2004                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
2005         }
2006         /* if inode is created as part of the transaction,
2007          * then it's counted already by the creation method */
2008         if (inode != NULL)
2009                 credits++;
2010
2011 out:
2012
2013         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
2014
2015         /* dt_declare_write() is usually called for system objects, such
2016          * as llog or last_rcvd files. We needn't enforce quota on those
2017          * objects, so always set the lqi_space as 0. */
2018         if (inode != NULL)
2019                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
2020                                            i_gid_read(inode),
2021                                            i_projid_read(inode), 0,
2022                                            oh, obj, NULL, OSD_QID_BLK);
2023
2024         if (rc == 0)
2025                 rc = osd_trunc_lock(obj, oh, true);
2026
2027         RETURN(rc);
2028 }
2029
2030 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
2031 {
2032         /* LU-2634: clear the extent format for fast symlink */
2033         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
2034
2035         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
2036         spin_lock(&inode->i_lock);
2037         LDISKFS_I(inode)->i_disksize = buflen;
2038         i_size_write(inode, buflen);
2039         spin_unlock(&inode->i_lock);
2040         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2041
2042         return 0;
2043 }
2044
2045 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
2046                                     int bufsize, int write_NUL, loff_t *offs,
2047                                     handle_t *handle)
2048 {
2049         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2050         struct buffer_head *bh        = NULL;
2051         loff_t              offset    = *offs;
2052         loff_t              new_size  = i_size_read(inode);
2053         unsigned long       block;
2054         int                 blocksize = 1 << inode->i_blkbits;
2055         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2056         int                 err = 0;
2057         int                 size;
2058         int                 boffs;
2059         int                 dirty_inode = 0;
2060         bool create, sparse, sync = false;
2061
2062         if (write_NUL) {
2063                 /*
2064                  * long symlink write does not count the NUL terminator in
2065                  * bufsize, we write it, and the inode's file size does not
2066                  * count the NUL terminator as well.
2067                  */
2068                 ((char *)buf)[bufsize] = '\0';
2069                 ++bufsize;
2070         }
2071
2072         dirty_inode = test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA, &ei->i_flags);
2073
2074         /* sparse checking is racy, but sparse is very rare case, leave as is */
2075         sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
2076                   ((new_size - 1) >> inode->i_blkbits) + 1);
2077
2078         while (bufsize > 0) {
2079                 int credits = handle->h_buffer_credits;
2080                 unsigned long last_block = (new_size == 0) ? 0 :
2081                                            (new_size - 1) >> inode->i_blkbits;
2082
2083                 if (bh)
2084                         brelse(bh);
2085
2086                 block = offset >> inode->i_blkbits;
2087                 boffs = offset & (blocksize - 1);
2088                 size = min(blocksize - boffs, bufsize);
2089                 sync = (block > last_block || new_size == 0 || sparse);
2090
2091                 if (sync)
2092                         down(&ei->i_append_sem);
2093
2094                 bh = __ldiskfs_bread(handle, inode, block, 0);
2095
2096                 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
2097                         CWARN("%s: adding bh without locking off %llu (block %lu, "
2098                               "size %d, offs %llu)\n", inode->i_sb->s_id,
2099                               offset, block, bufsize, *offs);
2100
2101                 if (IS_ERR_OR_NULL(bh)) {
2102                         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
2103                         int flags = LDISKFS_GET_BLOCKS_CREATE;
2104
2105                         /* while the file system is being mounted, avoid
2106                          * preallocation otherwise mount can take a long
2107                          * time as mballoc cache is cold.
2108                          * XXX: this is a workaround until we have a proper
2109                          *      fix in mballoc
2110                          * XXX: works with extent-based files only */
2111                         if (!osd->od_cl_seq)
2112                                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
2113                         bh = __ldiskfs_bread(handle, inode, block, flags);
2114                         create = true;
2115                 } else {
2116                         if (sync) {
2117                                 up(&ei->i_append_sem);
2118                                 sync = false;
2119                         }
2120                         create = false;
2121                 }
2122                 if (IS_ERR_OR_NULL(bh)) {
2123                         if (bh == NULL) {
2124                                 err = -EIO;
2125                         } else {
2126                                 err = PTR_ERR(bh);
2127                                 bh = NULL;
2128                         }
2129
2130                         CERROR("%s: error reading offset %llu (block %lu, "
2131                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
2132                                inode->i_sb->s_id, offset, block, bufsize, *offs,
2133                                credits, handle->h_buffer_credits, err);
2134                         break;
2135                 }
2136
2137                 err = ldiskfs_journal_get_write_access(handle, bh);
2138                 if (err) {
2139                         CERROR("journal_get_write_access() returned error %d\n",
2140                                err);
2141                         break;
2142                 }
2143                 LASSERTF(boffs + size <= bh->b_size,
2144                          "boffs %d size %d bh->b_size %lu\n",
2145                          boffs, size, (unsigned long)bh->b_size);
2146                 if (create) {
2147                         memset(bh->b_data, 0, bh->b_size);
2148                         if (sync) {
2149                                 up(&ei->i_append_sem);
2150                                 sync = false;
2151                         }
2152                 }
2153                 memcpy(bh->b_data + boffs, buf, size);
2154                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2155                 if (err)
2156                         break;
2157
2158                 if (offset + size > new_size)
2159                         new_size = offset + size;
2160                 offset += size;
2161                 bufsize -= size;
2162                 buf += size;
2163         }
2164         if (sync)
2165                 up(&ei->i_append_sem);
2166
2167         if (bh)
2168                 brelse(bh);
2169
2170         if (write_NUL)
2171                 --new_size;
2172         /* correct in-core and on-disk sizes */
2173         if (new_size > i_size_read(inode)) {
2174                 spin_lock(&inode->i_lock);
2175                 if (new_size > i_size_read(inode))
2176                         i_size_write(inode, new_size);
2177                 if (i_size_read(inode) > ei->i_disksize) {
2178                         ei->i_disksize = i_size_read(inode);
2179                         dirty_inode = 1;
2180                 }
2181                 spin_unlock(&inode->i_lock);
2182         }
2183         if (dirty_inode)
2184                 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2185
2186         if (err == 0)
2187                 *offs = offset;
2188         return err;
2189 }
2190
2191 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
2192                          const struct lu_buf *buf, loff_t *pos,
2193                          struct thandle *handle)
2194 {
2195         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
2196         struct osd_thandle      *oh;
2197         ssize_t                 result;
2198         int                     is_link;
2199
2200         LASSERT(dt_object_exists(dt));
2201
2202         LASSERT(handle != NULL);
2203         LASSERT(inode != NULL);
2204         ll_vfs_dq_init(inode);
2205
2206         /* XXX: don't check: one declared chunk can be used many times */
2207         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
2208
2209         oh = container_of(handle, struct osd_thandle, ot_super);
2210         LASSERT(oh->ot_handle->h_transaction != NULL);
2211         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
2212
2213         /* Write small symlink to inode body as we need to maintain correct
2214          * on-disk symlinks for ldiskfs.
2215          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
2216          * does not count it in.
2217          */
2218         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
2219         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
2220                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
2221         else
2222                 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
2223                                                   is_link, pos, oh->ot_handle);
2224         if (result == 0)
2225                 result = buf->lb_len;
2226
2227         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
2228
2229         return result;
2230 }
2231
2232 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
2233                              __u64 start, __u64 end, struct thandle *th)
2234 {
2235         struct osd_thandle *oh;
2236         struct inode       *inode;
2237         int                 rc;
2238         ENTRY;
2239
2240         LASSERT(th);
2241         oh = container_of(th, struct osd_thandle, ot_super);
2242
2243         /*
2244          * we don't need to reserve credits for whole truncate
2245          * it's not possible as truncate may need to free too many
2246          * blocks and that won't fit a single transaction. instead
2247          * we reserve credits to change i_size and put inode onto
2248          * orphan list. if needed truncate will extend or restart
2249          * transaction
2250          */
2251         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
2252                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
2253
2254         inode = osd_dt_obj(dt)->oo_inode;
2255         LASSERT(inode);
2256
2257         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2258                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
2259                                    NULL, OSD_QID_BLK);
2260
2261         if (rc == 0)
2262                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2263
2264         RETURN(rc);
2265 }
2266
2267 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
2268                      __u64 start, __u64 end, struct thandle *th)
2269 {
2270         struct osd_object *obj = osd_dt_obj(dt);
2271         struct osd_device *osd = osd_obj2dev(obj);
2272         struct inode *inode = obj->oo_inode;
2273         struct osd_access_lock *al;
2274         struct osd_thandle *oh;
2275         int rc = 0, found = 0;
2276         bool grow = false;
2277         ENTRY;
2278
2279         LASSERT(end == OBD_OBJECT_EOF);
2280         LASSERT(dt_object_exists(dt));
2281         LASSERT(osd_invariant(obj));
2282         LASSERT(inode != NULL);
2283         ll_vfs_dq_init(inode);
2284
2285         LASSERT(th);
2286         oh = container_of(th, struct osd_thandle, ot_super);
2287         LASSERT(oh->ot_handle->h_transaction != NULL);
2288
2289         /* we used to skip truncate to current size to
2290          * optimize truncates on OST. with DoM we can
2291          * get attr_set to set specific size (MDS_REINT)
2292          * and then get truncate RPC which essentially
2293          * would be skipped. this is bad.. so, disable
2294          * this optimization on MDS till the client stop
2295          * to sent MDS_REINT (LU-11033) -bzzz */
2296         if (osd->od_is_ost && i_size_read(inode) == start)
2297                 RETURN(0);
2298
2299         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2300
2301         spin_lock(&inode->i_lock);
2302         if (i_size_read(inode) < start)
2303                 grow = true;
2304         i_size_write(inode, start);
2305         spin_unlock(&inode->i_lock);
2306         ll_truncate_pagecache(inode, start);
2307
2308         /* optimize grow case */
2309         if (grow) {
2310                 osd_execute_truncate(obj);
2311                 GOTO(out, rc);
2312         }
2313
2314         /* add to orphan list to ensure truncate completion
2315          * if this transaction succeed. ldiskfs_truncate()
2316          * will take the inode out of the list */
2317         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2318         if (rc != 0)
2319                 GOTO(out, rc);
2320
2321         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2322                 if (obj != al->tl_obj)
2323                         continue;
2324                 LASSERT(al->tl_shared == 0);
2325                 found = 1;
2326                 /* do actual truncate in osd_trans_stop() */
2327                 al->tl_truncate = 1;
2328                 break;
2329         }
2330         LASSERT(found);
2331
2332 out:
2333         RETURN(rc);
2334 }
2335
2336 static int fiemap_check_ranges(struct inode *inode,
2337                                u64 start, u64 len, u64 *new_len)
2338 {
2339         loff_t maxbytes;
2340
2341         *new_len = len;
2342
2343         if (len == 0)
2344                 return -EINVAL;
2345
2346         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2347                 maxbytes = inode->i_sb->s_maxbytes;
2348         else
2349                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2350
2351         if (start > maxbytes)
2352                 return -EFBIG;
2353
2354         /*
2355          * Shrink request scope to what the fs can actually handle.
2356          */
2357         if (len > maxbytes || (maxbytes - len) < start)
2358                 *new_len = maxbytes - start;
2359
2360         return 0;
2361 }
2362
2363 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2364 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
2365
2366 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2367                           struct fiemap *fm)
2368 {
2369         struct fiemap_extent_info fieinfo = {0, };
2370         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2371         u64 len;
2372         int rc;
2373
2374
2375         LASSERT(inode);
2376         if (inode->i_op->fiemap == NULL)
2377                 return -EOPNOTSUPP;
2378
2379         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2380                 return -EINVAL;
2381
2382         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2383         if (rc)
2384                 return rc;
2385
2386         fieinfo.fi_flags = fm->fm_flags;
2387         fieinfo.fi_extents_max = fm->fm_extent_count;
2388         fieinfo.fi_extents_start = fm->fm_extents;
2389
2390         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2391                 filemap_write_and_wait(inode->i_mapping);
2392
2393         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2394         fm->fm_flags = fieinfo.fi_flags;
2395         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2396
2397         return rc;
2398 }
2399
2400 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2401                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2402 {
2403         int              rc = 0;
2404         struct inode    *inode = osd_dt_obj(dt)->oo_inode;
2405         ENTRY;
2406
2407         switch (advice) {
2408         case LU_LADVISE_DONTNEED:
2409                 if (end == 0)
2410                         break;
2411                 invalidate_mapping_pages(inode->i_mapping,
2412                                          start >> PAGE_SHIFT,
2413                                          (end - 1) >> PAGE_SHIFT);
2414                 break;
2415         default:
2416                 rc = -ENOTSUPP;
2417                 break;
2418         }
2419
2420         RETURN(rc);
2421 }
2422
2423 /*
2424  * in some cases we may need declare methods for objects being created
2425  * e.g., when we create symlink
2426  */
2427 const struct dt_body_operations osd_body_ops_new = {
2428         .dbo_declare_write = osd_declare_write,
2429 };
2430
2431 const struct dt_body_operations osd_body_ops = {
2432         .dbo_read                       = osd_read,
2433         .dbo_declare_write              = osd_declare_write,
2434         .dbo_write                      = osd_write,
2435         .dbo_bufs_get                   = osd_bufs_get,
2436         .dbo_bufs_put                   = osd_bufs_put,
2437         .dbo_write_prep                 = osd_write_prep,
2438         .dbo_declare_write_commit       = osd_declare_write_commit,
2439         .dbo_write_commit               = osd_write_commit,
2440         .dbo_read_prep                  = osd_read_prep,
2441         .dbo_declare_punch              = osd_declare_punch,
2442         .dbo_punch                      = osd_punch,
2443         .dbo_fiemap_get                 = osd_fiemap_get,
2444         .dbo_ladvise                    = osd_ladvise,
2445 };
2446
2447 /**
2448  * Get a truncate lock
2449  *
2450  * In order to take multi-transaction truncate out of main transaction we let
2451  * the caller grab a lock on the object passed. the lock can be shared (for
2452  * writes) and exclusive (for truncate). It's not allowed to mix truncate
2453  * and write in the same transaction handle (do not confuse with big ldiskfs
2454  * transaction containing lots of handles).
2455  * The lock must be taken at declaration.
2456  *
2457  * \param obj           object to lock
2458  * \oh                  transaction
2459  * \shared              shared or exclusive
2460  *
2461  * \retval 0            lock is granted
2462  * \retval -NOMEM       no memory to allocate lock
2463  */
2464 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2465 {
2466         struct osd_access_lock *al, *tmp;
2467
2468         LASSERT(obj);
2469         LASSERT(oh);
2470
2471         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2472                 if (tmp->tl_obj != obj)
2473                         continue;
2474                 LASSERT(tmp->tl_shared == shared);
2475                 /* found same lock */
2476                 return 0;
2477         }
2478
2479         OBD_ALLOC_PTR(al);
2480         if (unlikely(al == NULL))
2481                 return -ENOMEM;
2482         al->tl_obj = obj;
2483         al->tl_truncate = false;
2484         if (shared)
2485                 down_read(&obj->oo_ext_idx_sem);
2486         else
2487                 down_write(&obj->oo_ext_idx_sem);
2488         al->tl_shared = shared;
2489
2490         list_add(&al->tl_list, &oh->ot_trunc_locks);
2491
2492         return 0;
2493 }
2494
2495 void osd_trunc_unlock_all(struct list_head *list)
2496 {
2497         struct osd_access_lock *al, *tmp;
2498         list_for_each_entry_safe(al, tmp, list, tl_list) {
2499                 if (al->tl_shared)
2500                         up_read(&al->tl_obj->oo_ext_idx_sem);
2501                 else
2502                         up_write(&al->tl_obj->oo_ext_idx_sem);
2503                 list_del(&al->tl_list);
2504                 OBD_FREE_PTR(al);
2505         }
2506 }
2507
2508 void osd_execute_truncate(struct osd_object *obj)
2509 {
2510         struct osd_device *d = osd_obj2dev(obj);
2511         struct inode *inode = obj->oo_inode;
2512         __u64 size;
2513
2514         /* simulate crash before (in the middle) of delayed truncate */
2515         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2516                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2517                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2518
2519                 mutex_lock(&sbi->s_orphan_lock);
2520                 list_del_init(&ei->i_orphan);
2521                 mutex_unlock(&sbi->s_orphan_lock);
2522                 return;
2523         }
2524
2525 #ifdef HAVE_INODEOPS_TRUNCATE
2526         if (inode->i_op->truncate)
2527                 inode->i_op->truncate(inode);
2528         else
2529 #endif
2530                 ldiskfs_truncate(inode);
2531
2532         /*
2533          * For a partial-page truncate, flush the page to disk immediately to
2534          * avoid data corruption during direct disk write.  b=17397
2535          */
2536         size = i_size_read(inode);
2537         if ((size & ~PAGE_MASK) == 0)
2538                 return;
2539         if (osd_use_page_cache(d)) {
2540                 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2541         } else {
2542                 /* Notice we use "wait" version to ensure I/O is complete */
2543                 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2544                 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2545                                          size >> PAGE_SHIFT);
2546         }
2547 }
2548
2549 void osd_process_truncates(struct list_head *list)
2550 {
2551         struct osd_access_lock *al;
2552
2553         LASSERT(journal_current_handle() == NULL);
2554
2555         list_for_each_entry(al, list, tl_list) {
2556                 if (al->tl_shared)
2557                         continue;
2558                 if (!al->tl_truncate)
2559                         continue;
2560                 osd_execute_truncate(al->tl_obj);
2561         }
2562 }