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         int error = bio->bi_status;
 162 #else
 163 static void dio_complete_routine(struct bio *bio, int error)
 164 {
 165 #endif
 166         struct osd_iobuf *iobuf = bio->bi_private;
 167         int iter;
 168         struct bio_vec *bvl;
 169
 170         /* CAVEAT EMPTOR: possibly in IRQ context
 171          * DO NOT record procfs stats here!!! */
 172
 173         if (unlikely(iobuf == NULL)) {
 174                 CERROR("***** bio->bi_private is NULL!  This should never "
 175                        "happen.  Normally, I would crash here, but instead I "
 176                        "will dump the bio contents to the console.  Please "
 177                        "report this to <https://jira.whamcloud.com/> , along "
 178                        "with any interesting messages leading up to this point "
 179                        "(like SCSI errors, perhaps).  Because bi_private is "
 180                        "NULL, I can't wake up the thread that initiated this "
 181                        "IO - you will probably have to reboot this node.\n");
 182                 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
 183                        ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
 184                        bio->bi_next, (unsigned long)bio->bi_flags,
 185                        (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
 186                        bio_sectors(bio) << 9, bio->bi_end_io,
 187                        atomic_read(&bio->__bi_cnt),
 188                        bio->bi_private);
 189                 return;
 190         }
 191
 192         /* the check is outside of the cycle for performance reason -bzzz */
 193         if (!bio_data_dir(bio)) {
 194                 bio_for_each_segment_all(bvl, bio, iter) {
 195                         if (likely(error == 0))
 196                                 SetPageUptodate(bvl_to_page(bvl));
 197                         LASSERT(PageLocked(bvl_to_page(bvl)));
 198                 }
 199                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 200         } else {
 201                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 202         }
 203
 204         /* any real error is good enough -bzzz */
 205         if (error != 0 && iobuf->dr_error == 0)
 206                 iobuf->dr_error = error;
 207
 208         /*
 209          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 210          * it's possible that service thread will see dr_numreqs
 211          * is zero, but dr_elapsed is not set yet, leading to lost
 212          * data in this processing and an assertion in a subsequent
 213          * call to OSD.
 214          */
 215         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 216                 ktime_t now = ktime_get();
 217
 218                 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
 219                 iobuf->dr_elapsed_valid = 1;
 220         }
 221         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 222                 wake_up(&iobuf->dr_wait);
 223
 224         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 225          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 226          * mempool when serious on-disk fragmentation was encountered,
 227          * deadlocking the OST.  The bios are now released as soon as complete
 228          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 229         bio_put(bio);
 230 }
 231
 232 static void record_start_io(struct osd_iobuf *iobuf, int size)
 233 {
 234         struct osd_device    *osd = iobuf->dr_dev;
 235         struct obd_histogram *h = osd->od_brw_stats.hist;
 236
 237         iobuf->dr_frags++;
 238         atomic_inc(&iobuf->dr_numreqs);
 239
 240         if (iobuf->dr_rw == 0) {
 241                 atomic_inc(&osd->od_r_in_flight);
 242                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 243                                  atomic_read(&osd->od_r_in_flight));
 244                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 245         } else if (iobuf->dr_rw == 1) {
 246                 atomic_inc(&osd->od_w_in_flight);
 247                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 248                                  atomic_read(&osd->od_w_in_flight));
 249                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 250         } else {
 251                 LBUG();
 252         }
 253 }
 254
 255 static void osd_submit_bio(int rw, struct bio *bio)
 256 {
 257         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 258 #ifdef HAVE_SUBMIT_BIO_2ARGS
 259         submit_bio(rw ? WRITE : READ, bio);
 260 #else
 261         bio->bi_opf |= rw;
 262         submit_bio(bio);
 263 #endif
 264 }
 265
 266 static int can_be_merged(struct bio *bio, sector_t sector)
 267 {
 268         if (bio == NULL)
 269                 return 0;
 270
 271         return bio_end_sector(bio) == sector ? 1 : 0;
 272 }
 273
 274 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
 275 /*
 276  * This function will change the data written, thus it should only be
 277  * used when checking data integrity feature
 278  */
 279 static void bio_integrity_fault_inject(struct bio *bio)
 280 {
 281         struct bio_vec *bvec;
 282         int i;
 283         void *kaddr;
 284         char *addr;
 285
 286         bio_for_each_segment_all(bvec, bio, i) {
 287                 struct page *page = bvec->bv_page;
 288
 289                 kaddr = kmap(page);
 290                 addr = kaddr;
 291                 *addr = ~(*addr);
 292                 kunmap(page);
 293                 break;
 294         }
 295 }
 296
 297 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
 298                            unsigned int sectors, int tuple_size)
 299 {
 300         __u16 *expected_guard;
 301         __u16 *bio_guard;
 302         int i;
 303
 304         expected_guard = expected_guard_buf;
 305         for (i = 0; i < sectors; i++) {
 306                 bio_guard = (__u16 *)bio_prot_buf;
 307                 if (*bio_guard != *expected_guard) {
 308                         CERROR("unexpected guard tags on sector %d "
 309                                "expected guard %u, bio guard "
 310                                "%u, sectors %u, tuple size %d\n",
 311                                i, *expected_guard, *bio_guard, sectors,
 312                                tuple_size);
 313                         return -EIO;
 314                 }
 315                 expected_guard++;
 316                 bio_prot_buf += tuple_size;
 317         }
 318         return 0;
 319 }
 320
 321 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
 322                                      struct osd_iobuf *iobuf, int index)
 323 {
 324         struct blk_integrity *bi = bdev_get_integrity(bdev);
 325         struct bio_integrity_payload *bip = bio->bi_integrity;
 326         struct niobuf_local *lnb;
 327         unsigned short sector_size = blk_integrity_interval(bi);
 328         void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
 329                 bip->bip_vec->bv_offset;
 330         struct bio_vec *bv;
 331         sector_t sector = bio_start_sector(bio);
 332         unsigned int i, sectors, total;
 333         __u16 *expected_guard;
 334         int rc;
 335
 336         total = 0;
 337         bio_for_each_segment_all(bv, bio, i) {
 338                 lnb = iobuf->dr_lnbs[index];
 339                 expected_guard = lnb->lnb_guards;
 340                 sectors = bv->bv_len / sector_size;
 341                 if (lnb->lnb_guard_rpc) {
 342                         rc = bio_dif_compare(expected_guard, bio_prot_buf,
 343                                              sectors, bi->tuple_size);
 344                         if (rc)
 345                                 return rc;
 346                 }
 347
 348                 sector += sectors;
 349                 bio_prot_buf += sectors * bi->tuple_size;
 350                 total += sectors * bi->tuple_size;
 351                 LASSERT(total <= bip_size(bio->bi_integrity));
 352                 index++;
 353         }
 354         return 0;
 355 }
 356
 357 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
 358                                     struct osd_iobuf *iobuf,
 359                                     int start_page_idx, bool fault_inject,
 360                                     bool integrity_enabled)
 361 {
 362         struct super_block *sb = osd_sb(osd);
 363         integrity_gen_fn *generate_fn = NULL;
 364         integrity_vrfy_fn *verify_fn = NULL;
 365         int rc;
 366
 367         ENTRY;
 368
 369         if (!integrity_enabled)
 370                 RETURN(0);
 371
 372         rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
 373         if (rc)
 374                 RETURN(rc);
 375
 376         rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
 377         if (rc)
 378                 RETURN(rc);
 379
 380         /* Verify and inject fault only when writing */
 381         if (iobuf->dr_rw == 1) {
 382                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
 383                         rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
 384                                                        start_page_idx);
 385                         if (rc)
 386                                 RETURN(rc);
 387                 }
 388
 389                 if (unlikely(fault_inject))
 390                         bio_integrity_fault_inject(bio);
 391         }
 392
 393         RETURN(0);
 394 }
 395
 396 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 397 #  ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 398 static void dio_integrity_complete_routine(struct bio *bio)
 399 #  else
 400 static void dio_integrity_complete_routine(struct bio *bio, int error)
 401 #  endif
 402 {
 403         struct osd_bio_private *bio_private = bio->bi_private;
 404
 405         bio->bi_private = bio_private->obp_iobuf;
 406         osd_dio_complete_routine(bio, error);
 407
 408         OBD_FREE_PTR(bio_private);
 409 }
 410 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
 411 #else  /* !CONFIG_BLK_DEV_INTEGRITY */
 412 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
 413                                  fault_inject, integrity_enabled) 0
 414 #endif /* CONFIG_BLK_DEV_INTEGRITY */
 415
 416 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
 417                         bool integrity_enabled, int start_page_idx,
 418                         struct osd_bio_private **pprivate)
 419 {
 420         ENTRY;
 421
 422         *pprivate = NULL;
 423
 424 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 425         if (integrity_enabled) {
 426                 struct osd_bio_private *bio_private = NULL;
 427
 428                 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
 429                 if (bio_private == NULL)
 430                         RETURN(-ENOMEM);
 431                 bio->bi_end_io = dio_integrity_complete_routine;
 432                 bio->bi_private = bio_private;
 433                 bio_private->obp_start_page_idx = start_page_idx;
 434                 bio_private->obp_iobuf = iobuf;
 435                 *pprivate = bio_private;
 436         } else
 437 #endif
 438         {
 439                 bio->bi_end_io = dio_complete_routine;
 440                 bio->bi_private = iobuf;
 441         }
 442
 443         RETURN(0);
 444 }
 445
 446 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 447                       struct osd_iobuf *iobuf)
 448 {
 449         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 450         struct page **pages = iobuf->dr_pages;
 451         int npages = iobuf->dr_npages;
 452         sector_t *blocks = iobuf->dr_blocks;
 453         int total_blocks = npages * blocks_per_page;
 454         struct super_block *sb = inode->i_sb;
 455         int sector_bits = sb->s_blocksize_bits - 9;
 456         unsigned int blocksize = sb->s_blocksize;
 457         struct block_device *bdev = sb->s_bdev;
 458         struct osd_bio_private *bio_private = NULL;
 459         struct bio *bio = NULL;
 460         int bio_start_page_idx;
 461         struct page *page;
 462         unsigned int page_offset;
 463         sector_t sector;
 464         int nblocks;
 465         int block_idx;
 466         int page_idx;
 467         int i;
 468         int rc = 0;
 469         bool fault_inject;
 470         bool integrity_enabled;
 471         struct blk_plug plug;
 472         ENTRY;
 473
 474         fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
 475         LASSERT(iobuf->dr_npages == npages);
 476
 477         integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
 478
 479         osd_brw_stats_update(osd, iobuf);
 480         iobuf->dr_start_time = ktime_get();
 481
 482         blk_start_plug(&plug);
 483         for (page_idx = 0, block_idx = 0;
 484              page_idx < npages;
 485              page_idx++, block_idx += blocks_per_page) {
 486
 487                 page = pages[page_idx];
 488                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 489
 490                 for (i = 0, page_offset = 0;
 491                      i < blocks_per_page;
 492                      i += nblocks, page_offset += blocksize * nblocks) {
 493
 494                         nblocks = 1;
 495
 496                         if (blocks[block_idx + i] == 0) {  /* hole */
 497                                 LASSERTF(iobuf->dr_rw == 0,
 498                                          "page_idx %u, block_idx %u, i %u\n",
 499                                          page_idx, block_idx, i);
 500                                 memset(kmap(page) + page_offset, 0, blocksize);
 501                                 kunmap(page);
 502                                 continue;
 503                         }
 504
 505                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 506
 507                         /* Additional contiguous file blocks? */
 508                         while (i + nblocks < blocks_per_page &&
 509                                (sector + (nblocks << sector_bits)) ==
 510                                ((sector_t)blocks[block_idx + i + nblocks] <<
 511                                 sector_bits))
 512                                 nblocks++;
 513
 514                         if (bio != NULL &&
 515                             can_be_merged(bio, sector) &&
 516                             bio_add_page(bio, page,
 517                                          blocksize * nblocks, page_offset) != 0)
 518                                 continue;       /* added this frag OK */
 519
 520                         if (bio != NULL) {
 521                                 struct request_queue *q = bio_get_queue(bio);
 522                                 unsigned int bi_size = bio_sectors(bio) << 9;
 523
 524                                 /* Dang! I have to fragment this I/O */
 525                                 CDEBUG(D_INODE,
 526                                        "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
 527                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 528                                        bio_sectors(bio),
 529                                        queue_max_sectors(q),
 530                                        bio->bi_phys_segments,
 531                                        queue_max_segments(q));
 532                                 rc = osd_bio_integrity_handle(osd, bio,
 533                                         iobuf, bio_start_page_idx,
 534                                         fault_inject, integrity_enabled);
 535                                 if (rc) {
 536                                         bio_put(bio);
 537                                         goto out;
 538                                 }
 539
 540                                 record_start_io(iobuf, bi_size);
 541                                 osd_submit_bio(iobuf->dr_rw, bio);
 542                         }
 543
 544                         bio_start_page_idx = page_idx;
 545                         /* allocate new bio */
 546                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 547                                                       (npages - page_idx) *
 548                                                       blocks_per_page));
 549                         if (bio == NULL) {
 550                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 551                                        (npages - page_idx), blocks_per_page,
 552                                        (npages - page_idx) * blocks_per_page);
 553                                 rc = -ENOMEM;
 554                                 goto out;
 555                         }
 556
 557                         bio_set_dev(bio, bdev);
 558                         bio_set_sector(bio, sector);
 559                         bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
 560                         rc = osd_bio_init(bio, iobuf, integrity_enabled,
 561                                           bio_start_page_idx, &bio_private);
 562                         if (rc) {
 563                                 bio_put(bio);
 564                                 goto out;
 565                         }
 566
 567                         rc = bio_add_page(bio, page,
 568                                           blocksize * nblocks, page_offset);
 569                         LASSERT(rc != 0);
 570                 }
 571         }
 572
 573         if (bio != NULL) {
 574                 rc = osd_bio_integrity_handle(osd, bio, iobuf,
 575                                               bio_start_page_idx,
 576                                               fault_inject,
 577                                               integrity_enabled);
 578                 if (rc) {
 579                         bio_put(bio);
 580                         goto out;
 581                 }
 582
 583                 record_start_io(iobuf, bio_sectors(bio) << 9);
 584                 osd_submit_bio(iobuf->dr_rw, bio);
 585                 rc = 0;
 586         }
 587
 588 out:
 589         blk_finish_plug(&plug);
 590
 591         /* in order to achieve better IO throughput, we don't wait for writes
 592          * completion here. instead we proceed with transaction commit in
 593          * parallel and wait for IO completion once transaction is stopped
 594          * see osd_trans_stop() for more details -bzzz */
 595         if (iobuf->dr_rw == 0 || fault_inject) {
 596                 wait_event(iobuf->dr_wait,
 597                            atomic_read(&iobuf->dr_numreqs) == 0);
 598                 osd_fini_iobuf(osd, iobuf);
 599         }
 600
 601         if (rc == 0) {
 602                 rc = iobuf->dr_error;
 603         } else {
 604                 if (bio_private)
 605                         OBD_FREE_PTR(bio_private);
 606         }
 607
 608         RETURN(rc);
 609 }
 610
 611 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 612                                    struct niobuf_local *lnb)
 613 {
 614         ENTRY;
 615
 616         *nrpages = 0;
 617
 618         while (len > 0) {
 619                 int poff = offset & (PAGE_SIZE - 1);
 620                 int plen = PAGE_SIZE - poff;
 621
 622                 if (plen > len)
 623                         plen = len;
 624                 lnb->lnb_file_offset = offset;
 625                 lnb->lnb_page_offset = poff;
 626                 lnb->lnb_len = plen;
 627                 /* lnb->lnb_flags = rnb->rnb_flags; */
 628                 lnb->lnb_flags = 0;
 629                 lnb->lnb_page = NULL;
 630                 lnb->lnb_rc = 0;
 631                 lnb->lnb_guard_rpc = 0;
 632                 lnb->lnb_guard_disk = 0;
 633
 634                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 635                          (long long) len);
 636                 offset += plen;
 637                 len -= plen;
 638                 lnb++;
 639                 (*nrpages)++;
 640         }
 641
 642         RETURN(0);
 643 }
 644
 645 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
 646                                  loff_t offset, gfp_t gfp_mask)
 647 {
 648         struct osd_thread_info *oti = osd_oti_get(env);
 649         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 650         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 651         struct page *page;
 652         int cur = oti->oti_dio_pages_used;
 653
 654         LASSERT(inode);
 655
 656         if (osd_use_page_cache(d)) {
 657                 page = find_or_create_page(inode->i_mapping,
 658                                            offset >> PAGE_SHIFT,
 659                                            gfp_mask);
 660
 661                 if (likely(page))
 662                         LASSERT(!test_bit(PG_private_2, &page->flags));
 663                 else
 664                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 665         } else {
 666
 667                 LASSERT(oti->oti_dio_pages);
 668
 669                 if (unlikely(!oti->oti_dio_pages[cur])) {
 670                         LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
 671                         page = alloc_page(gfp_mask);
 672                         if (!page)
 673                                 return NULL;
 674                         oti->oti_dio_pages[cur] = page;
 675                 }
 676
 677                 page = oti->oti_dio_pages[cur];
 678                 LASSERT(!test_bit(PG_private_2, &page->flags));
 679                 set_bit(PG_private_2, &page->flags);
 680                 oti->oti_dio_pages_used++;
 681
 682                 LASSERT(!PageLocked(page));
 683                 lock_page(page);
 684
 685                 LASSERT(!page->mapping);
 686                 LASSERT(!PageWriteback(page));
 687                 ClearPageUptodate(page);
 688
 689                 page->index = offset >> PAGE_SHIFT;
 690         }
 691
 692         return page;
 693 }
 694
 695 /*
 696  * there are following "locks":
 697  * journal_start
 698  * i_mutex
 699  * page lock
 700  *
 701  * osd write path:
 702  *  - lock page(s)
 703  *  - journal_start
 704  *  - truncate_sem
 705  *
 706  * ext4 vmtruncate:
 707  *  - lock pages, unlock
 708  *  - journal_start
 709  *  - lock partial page
 710  *  - i_data_sem
 711  *
 712  */
 713
 714 /**
 715  * Unlock and release pages loaded by osd_bufs_get()
 716  *
 717  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 718  *
 719  * \param env           thread execution environment
 720  * \param dt            dt object undergoing IO (OSD object + methods)
 721  * \param lnb           array of pages undergoing IO
 722  * \param npages        number of pages in \a lnb
 723  *
 724  * \retval 0            always
 725  */
 726 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 727                         struct niobuf_local *lnb, int npages)
 728 {
 729         struct osd_thread_info *oti = osd_oti_get(env);
 730         struct pagevec pvec;
 731         int i;
 732
 733         ll_pagevec_init(&pvec, 0);
 734
 735         for (i = 0; i < npages; i++) {
 736                 struct page *page = lnb[i].lnb_page;
 737
 738                 if (page == NULL)
 739                         continue;
 740                 LASSERT(PageLocked(page));
 741
 742                 /* if the page isn't cached, then reset uptodate
 743                  * to prevent reuse */
 744                 if (test_bit(PG_private_2, &page->flags)) {
 745                         clear_bit(PG_private_2, &page->flags);
 746                         ClearPageUptodate(page);
 747                         unlock_page(page);
 748                         oti->oti_dio_pages_used--;
 749                 } else {
 750                         unlock_page(page);
 751                         if (pagevec_add(&pvec, page) == 0)
 752                                 pagevec_release(&pvec);
 753                 }
 754                 dt_object_put(env, dt);
 755
 756                 lnb[i].lnb_page = NULL;
 757         }
 758
 759         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
 760
 761         /* Release any partial pagevec */
 762         pagevec_release(&pvec);
 763
 764         RETURN(0);
 765 }
 766
 767 /**
 768  * Load and lock pages undergoing IO
 769  *
 770  * Pages as described in the \a lnb array are fetched (from disk or cache)
 771  * and locked for IO by the caller.
 772  *
 773  * DLM locking protects us from write and truncate competing for same region,
 774  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 775  * It's possible the writeout on a such a page is in progress when we access
 776  * it. It's also possible that during this writeout we put new (partial) data
 777  * into the page, but won't be able to proceed in filter_commitrw_write().
 778  * Therefore, just wait for writeout completion as it should be rare enough.
 779  *
 780  * \param env           thread execution environment
 781  * \param dt            dt object undergoing IO (OSD object + methods)
 782  * \param pos           byte offset of IO start
 783  * \param len           number of bytes of IO
 784  * \param lnb           array of extents undergoing IO
 785  * \param rw            read or write operation, and other flags
 786  * \param capa          capabilities
 787  *
 788  * \retval pages        (zero or more) loaded successfully
 789  * \retval -ENOMEM      on memory/page allocation error
 790  */
 791 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 792                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 793                         enum dt_bufs_type rw)
 794 {
 795         struct osd_thread_info *oti = osd_oti_get(env);
 796         struct osd_object *obj = osd_dt_obj(dt);
 797         int npages, i, rc = 0;
 798         gfp_t gfp_mask;
 799
 800         LASSERT(obj->oo_inode);
 801
 802         if (!osd_use_page_cache(osd_obj2dev(obj))) {
 803                 if (unlikely(!oti->oti_dio_pages)) {
 804                         OBD_ALLOC(oti->oti_dio_pages,
 805                                   sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
 806                         if (!oti->oti_dio_pages)
 807                                 return -ENOMEM;
 808                 }
 809         }
 810
 811         osd_map_remote_to_local(pos, len, &npages, lnb);
 812
 813         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 814         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 815                                              GFP_HIGHUSER;
 816         for (i = 0; i < npages; i++, lnb++) {
 817                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
 818                                              gfp_mask);
 819                 if (lnb->lnb_page == NULL)
 820                         GOTO(cleanup, rc = -ENOMEM);
 821
 822                 wait_on_page_writeback(lnb->lnb_page);
 823                 BUG_ON(PageWriteback(lnb->lnb_page));
 824
 825                 lu_object_get(&dt->do_lu);
 826         }
 827
 828         RETURN(i);
 829
 830 cleanup:
 831         if (i > 0)
 832                 osd_bufs_put(env, dt, lnb - i, i);
 833         return rc;
 834 }
 835
 836 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 837                                        int pages, sector_t *blocks,
 838                                        int create)
 839 {
 840         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 841         int rc = 0, i = 0;
 842         struct page *fp = NULL;
 843         int clen = 0;
 844         pgoff_t max_page_index;
 845         handle_t *handle = NULL;
 846
 847         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
 848
 849         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 850                 inode->i_ino, pages, (*page)->index);
 851
 852         if (create) {
 853                 create = LDISKFS_GET_BLOCKS_CREATE;
 854                 handle = ldiskfs_journal_current_handle();
 855                 LASSERT(handle != NULL);
 856                 rc = osd_attach_jinode(inode);
 857                 if (rc)
 858                         return rc;
 859         }
 860         /* pages are sorted already. so, we just have to find
 861          * contig. space and process them properly */
 862         while (i < pages) {
 863                 long blen, total = 0;
 864                 struct ldiskfs_map_blocks map = { 0 };
 865
 866                 if (fp == NULL) { /* start new extent */
 867                         fp = *page++;
 868                         clen = 1;
 869                         if (++i != pages)
 870                                 continue;
 871                 } else if (fp->index + clen == (*page)->index) {
 872                         /* continue the extent */
 873                         page++;
 874                         clen++;
 875                         if (++i != pages)
 876                                 continue;
 877                 }
 878                 if (fp->index + clen >= max_page_index)
 879                         GOTO(cleanup, rc = -EFBIG);
 880                 /* process found extent */
 881                 map.m_lblk = fp->index * blocks_per_page;
 882                 map.m_len = blen = clen * blocks_per_page;
 883 cont_map:
 884                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 885                 if (rc >= 0) {
 886                         int c = 0;
 887                         for (; total < blen && c < map.m_len; c++, total++) {
 888                                 if (rc == 0) {
 889                                         *(blocks + total) = 0;
 890                                         total++;
 891                                         break;
 892                                 } else {
 893                                         *(blocks + total) = map.m_pblk + c;
 894                                         /* unmap any possible underlying
 895                                          * metadata from the block device
 896                                          * mapping.  bug 6998. */
 897                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
 898                                             create)
 899                                                 clean_bdev_aliases(
 900                                                         inode->i_sb->s_bdev,
 901                                                         map.m_pblk + c, 1);
 902                                 }
 903                         }
 904                         rc = 0;
 905                 }
 906                 if (rc == 0 && total < blen) {
 907                         map.m_lblk = fp->index * blocks_per_page + total;
 908                         map.m_len = blen - total;
 909                         goto cont_map;
 910                 }
 911                 if (rc != 0)
 912                         GOTO(cleanup, rc);
 913
 914                 /* look for next extent */
 915                 fp = NULL;
 916                 blocks += blocks_per_page * clen;
 917         }
 918 cleanup:
 919         return rc;
 920 }
 921
 922 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 923                           struct niobuf_local *lnb, int npages)
 924 {
 925         struct osd_thread_info *oti   = osd_oti_get(env);
 926         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 927         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 928         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 929         ktime_t start;
 930         ktime_t end;
 931         s64 timediff;
 932         ssize_t                 isize;
 933         __s64                   maxidx;
 934         int                     rc = 0;
 935         int                     i;
 936         int                     cache = 0;
 937
 938         LASSERT(inode);
 939
 940         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 941         if (unlikely(rc != 0))
 942                 RETURN(rc);
 943
 944         isize = i_size_read(inode);
 945         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
 946
 947         if (osd->od_writethrough_cache)
 948                 cache = 1;
 949         if (isize > osd->od_readcache_max_filesize)
 950                 cache = 0;
 951
 952         start = ktime_get();
 953         for (i = 0; i < npages; i++) {
 954
 955                 if (cache == 0)
 956                         generic_error_remove_page(inode->i_mapping,
 957                                                   lnb[i].lnb_page);
 958
 959                 /*
 960                  * till commit the content of the page is undefined
 961                  * we'll set it uptodate once bulk is done. otherwise
 962                  * subsequent reads can access non-stable data
 963                  */
 964                 ClearPageUptodate(lnb[i].lnb_page);
 965
 966                 if (lnb[i].lnb_len == PAGE_SIZE)
 967                         continue;
 968
 969                 if (maxidx >= lnb[i].lnb_page->index) {
 970                         osd_iobuf_add_page(iobuf, &lnb[i]);
 971                 } else {
 972                         long off;
 973                         char *p = kmap(lnb[i].lnb_page);
 974
 975                         off = lnb[i].lnb_page_offset;
 976                         if (off)
 977                                 memset(p, 0, off);
 978                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
 979                               ~PAGE_MASK;
 980                         if (off)
 981                                 memset(p + off, 0, PAGE_SIZE - off);
 982                         kunmap(lnb[i].lnb_page);
 983                 }
 984         }
 985         end = ktime_get();
 986         timediff = ktime_us_delta(end, start);
 987         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 988
 989         if (iobuf->dr_npages) {
 990                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
 991                                                  iobuf->dr_npages,
 992                                                  iobuf->dr_blocks, 0);
 993                 if (likely(rc == 0)) {
 994                         rc = osd_do_bio(osd, inode, iobuf);
 995                         /* do IO stats for preparation reads */
 996                         osd_fini_iobuf(osd, iobuf);
 997                 }
 998         }
 999         RETURN(rc);
1000 }
1001
1002 struct osd_fextent {
1003         sector_t        start;
1004         sector_t        end;
1005         unsigned int    mapped:1;
1006 };
1007
1008 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1009                          struct osd_fextent *cached_extent)
1010 {
1011         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1012         sector_t block = offset >> inode->i_blkbits;
1013         sector_t start;
1014         struct fiemap_extent_info fei = { 0 };
1015         struct fiemap_extent fe = { 0 };
1016         mm_segment_t saved_fs;
1017         int rc;
1018
1019         if (block >= cached_extent->start && block < cached_extent->end)
1020                 return cached_extent->mapped;
1021
1022         if (i_size_read(inode) == 0)
1023                 return 0;
1024
1025         /* Beyond EOF, must not be mapped */
1026         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1027                 return 0;
1028
1029         fei.fi_extents_max = 1;
1030         fei.fi_extents_start = &fe;
1031
1032         saved_fs = get_fs();
1033         set_fs(KERNEL_DS);
1034         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1035         set_fs(saved_fs);
1036         if (rc != 0)
1037                 return 0;
1038
1039         start = fe.fe_logical >> inode->i_blkbits;
1040
1041         if (start > block) {
1042                 cached_extent->start = block;
1043                 cached_extent->end = start;
1044                 cached_extent->mapped = 0;
1045         } else {
1046                 cached_extent->start = start;
1047                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1048                                       inode->i_blkbits;
1049                 cached_extent->mapped = 1;
1050         }
1051
1052         return cached_extent->mapped;
1053 }
1054
1055 static int osd_declare_write_commit(const struct lu_env *env,
1056                                     struct dt_object *dt,
1057                                     struct niobuf_local *lnb, int npages,
1058                                     struct thandle *handle)
1059 {
1060         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1061         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1062         struct osd_thandle      *oh;
1063         int                     extents = 1;
1064         int                     depth;
1065         int                     i;
1066         int                     newblocks;
1067         int                     rc = 0;
1068         int                     flags = 0;
1069         int                     credits = 0;
1070         long long               quota_space = 0;
1071         struct osd_fextent      extent = { 0 };
1072         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1073         ENTRY;
1074
1075         LASSERT(handle != NULL);
1076         oh = container_of0(handle, struct osd_thandle, ot_super);
1077         LASSERT(oh->ot_handle == NULL);
1078
1079         newblocks = npages;
1080
1081         /* calculate number of extents (probably better to pass nb) */
1082         for (i = 0; i < npages; i++) {
1083                 if (i && lnb[i].lnb_file_offset !=
1084                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1085                         extents++;
1086
1087                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1088                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1089                 else
1090                         quota_space += PAGE_SIZE;
1091
1092                 /* ignore quota for the whole request if any page is from
1093                  * client cache or written by root.
1094                  *
1095                  * XXX once we drop the 1.8 client support, the checking
1096                  * for whether page is from cache can be simplified as:
1097                  * !(lnb[i].flags & OBD_BRW_SYNC)
1098                  *
1099                  * XXX we could handle this on per-lnb basis as done by
1100                  * grant. */
1101                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1102                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1103                     OBD_BRW_FROM_GRANT)
1104                         declare_flags |= OSD_QID_FORCE;
1105         }
1106
1107         /*
1108          * each extent can go into new leaf causing a split
1109          * 5 is max tree depth: inode + 4 index blocks
1110          * with blockmaps, depth is 3 at most
1111          */
1112         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1113                 /*
1114                  * many concurrent threads may grow tree by the time
1115                  * our transaction starts. so, consider 2 is a min depth
1116                  */
1117                 depth = ext_depth(inode);
1118                 depth = max(depth, 1) + 1;
1119                 newblocks += depth;
1120                 credits++; /* inode */
1121                 credits += depth * 2 * extents;
1122         } else {
1123                 depth = 3;
1124                 newblocks += depth;
1125                 credits++; /* inode */
1126                 credits += depth * extents;
1127         }
1128
1129         /* quota space for metadata blocks */
1130         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1131
1132         /* quota space should be reported in 1K blocks */
1133         quota_space = toqb(quota_space);
1134
1135         /* each new block can go in different group (bitmap + gd) */
1136
1137         /* we can't dirty more bitmap blocks than exist */
1138         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1139                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1140         else
1141                 credits += newblocks;
1142
1143         /* we can't dirty more gd blocks than exist */
1144         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1145                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1146         else
1147                 credits += newblocks;
1148
1149         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1150
1151         /* make sure the over quota flags were not set */
1152         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1153
1154         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1155                                    i_projid_read(inode), quota_space, oh,
1156                                    osd_dt_obj(dt), &flags, declare_flags);
1157
1158         /* we need only to store the overquota flags in the first lnb for
1159          * now, once we support multiple objects BRW, this code needs be
1160          * revised. */
1161         if (flags & QUOTA_FL_OVER_USRQUOTA)
1162                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1163         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1164                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1165         if (flags & QUOTA_FL_OVER_PRJQUOTA)
1166                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1167
1168         if (rc == 0)
1169                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1170
1171         RETURN(rc);
1172 }
1173
1174 /* Check if a block is allocated or not */
1175 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1176                             struct niobuf_local *lnb, int npages,
1177                             struct thandle *thandle)
1178 {
1179         struct osd_thread_info *oti = osd_oti_get(env);
1180         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1181         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1182         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1183         loff_t isize;
1184         int rc = 0, i;
1185
1186         LASSERT(inode);
1187
1188         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1189         if (unlikely(rc != 0))
1190                 RETURN(rc);
1191
1192         isize = i_size_read(inode);
1193         dquot_initialize(inode);
1194
1195         for (i = 0; i < npages; i++) {
1196                 if (lnb[i].lnb_rc == -ENOSPC &&
1197                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1198                         /* Allow the write to proceed if overwriting an
1199                          * existing block */
1200                         lnb[i].lnb_rc = 0;
1201                 }
1202
1203                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1204                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1205                                lnb[i].lnb_rc);
1206                         LASSERT(lnb[i].lnb_page);
1207                         generic_error_remove_page(inode->i_mapping,
1208                                                   lnb[i].lnb_page);
1209                         continue;
1210                 }
1211
1212                 LASSERT(PageLocked(lnb[i].lnb_page));
1213                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1214
1215                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1216                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1217
1218                 /*
1219                  * Since write and truncate are serialized by oo_sem, even
1220                  * partial-page truncate should not leave dirty pages in the
1221                  * page cache.
1222                  */
1223                 LASSERT(!PageDirty(lnb[i].lnb_page));
1224
1225                 SetPageUptodate(lnb[i].lnb_page);
1226
1227                 osd_iobuf_add_page(iobuf, &lnb[i]);
1228         }
1229
1230         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1231
1232         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1233                 rc = -ENOSPC;
1234         } else if (iobuf->dr_npages > 0) {
1235                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1236                                                  iobuf->dr_npages,
1237                                                  iobuf->dr_blocks, 1);
1238         } else {
1239                 /* no pages to write, no transno is needed */
1240                 thandle->th_local = 1;
1241         }
1242
1243         if (likely(rc == 0)) {
1244                 spin_lock(&inode->i_lock);
1245                 if (isize > i_size_read(inode)) {
1246                         i_size_write(inode, isize);
1247                         LDISKFS_I(inode)->i_disksize = isize;
1248                         spin_unlock(&inode->i_lock);
1249                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1250                 } else {
1251                         spin_unlock(&inode->i_lock);
1252                 }
1253
1254                 rc = osd_do_bio(osd, inode, iobuf);
1255                 /* we don't do stats here as in read path because
1256                  * write is async: we'll do this in osd_put_bufs() */
1257         } else {
1258                 osd_fini_iobuf(osd, iobuf);
1259         }
1260
1261         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1262
1263         if (unlikely(rc != 0)) {
1264                 /* if write fails, we should drop pages from the cache */
1265                 for (i = 0; i < npages; i++) {
1266                         if (lnb[i].lnb_page == NULL)
1267                                 continue;
1268                         LASSERT(PageLocked(lnb[i].lnb_page));
1269                         generic_error_remove_page(inode->i_mapping,
1270                                                   lnb[i].lnb_page);
1271                 }
1272         }
1273
1274         RETURN(rc);
1275 }
1276
1277 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1278                          struct niobuf_local *lnb, int npages)
1279 {
1280         struct osd_thread_info *oti = osd_oti_get(env);
1281         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1282         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1283         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1284         int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1285         ktime_t start, end;
1286         s64 timediff;
1287         loff_t isize;
1288
1289         LASSERT(inode);
1290
1291         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1292         if (unlikely(rc != 0))
1293                 RETURN(rc);
1294
1295         isize = i_size_read(inode);
1296
1297         if (osd->od_read_cache)
1298                 cache = 1;
1299         if (isize > osd->od_readcache_max_filesize)
1300                 cache = 0;
1301
1302         start = ktime_get();
1303         for (i = 0; i < npages; i++) {
1304
1305                 if (isize <= lnb[i].lnb_file_offset)
1306                         /* If there's no more data, abort early.
1307                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1308                         break;
1309
1310                 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1311                         lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1312                 else
1313                         lnb[i].lnb_rc = lnb[i].lnb_len;
1314
1315                 /* Bypass disk read if fail_loc is set properly */
1316                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1317                         SetPageUptodate(lnb[i].lnb_page);
1318
1319                 if (PageUptodate(lnb[i].lnb_page)) {
1320                         cache_hits++;
1321                 } else {
1322                         cache_misses++;
1323                         osd_iobuf_add_page(iobuf, &lnb[i]);
1324                 }
1325
1326                 if (cache == 0)
1327                         generic_error_remove_page(inode->i_mapping,
1328                                                   lnb[i].lnb_page);
1329         }
1330         end = ktime_get();
1331         timediff = ktime_us_delta(end, start);
1332         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1333
1334         if (cache_hits != 0)
1335                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1336                                     cache_hits);
1337         if (cache_misses != 0)
1338                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1339                                     cache_misses);
1340         if (cache_hits + cache_misses != 0)
1341                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1342                                     cache_hits + cache_misses);
1343
1344         if (iobuf->dr_npages) {
1345                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1346                                                  iobuf->dr_npages,
1347                                                  iobuf->dr_blocks, 0);
1348                 rc = osd_do_bio(osd, inode, iobuf);
1349
1350                 /* IO stats will be done in osd_bufs_put() */
1351         }
1352
1353         RETURN(rc);
1354 }
1355
1356 /*
1357  * XXX: Another layering violation for now.
1358  *
1359  * We don't want to use ->f_op->read methods, because generic file write
1360  *
1361  *         - serializes on ->i_sem, and
1362  *
1363  *         - does a lot of extra work like balance_dirty_pages(),
1364  *
1365  * which doesn't work for globally shared files like /last_rcvd.
1366  */
1367 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1368 {
1369         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1370
1371         memcpy(buffer, (char *)ei->i_data, buflen);
1372
1373         return  buflen;
1374 }
1375
1376 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1377 {
1378         struct buffer_head *bh;
1379         unsigned long block;
1380         int osize;
1381         int blocksize;
1382         int csize;
1383         int boffs;
1384
1385         /* prevent reading after eof */
1386         spin_lock(&inode->i_lock);
1387         if (i_size_read(inode) < *offs + size) {
1388                 loff_t diff = i_size_read(inode) - *offs;
1389                 spin_unlock(&inode->i_lock);
1390                 if (diff < 0) {
1391                         CDEBUG(D_OTHER,
1392                                "size %llu is too short to read @%llu\n",
1393                                i_size_read(inode), *offs);
1394                         return -EBADR;
1395                 } else if (diff == 0) {
1396                         return 0;
1397                 } else {
1398                         size = diff;
1399                 }
1400         } else {
1401                 spin_unlock(&inode->i_lock);
1402         }
1403
1404         blocksize = 1 << inode->i_blkbits;
1405         osize = size;
1406         while (size > 0) {
1407                 block = *offs >> inode->i_blkbits;
1408                 boffs = *offs & (blocksize - 1);
1409                 csize = min(blocksize - boffs, size);
1410                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1411                 if (IS_ERR(bh)) {
1412                         CERROR("%s: can't read %u@%llu on ino %lu: "
1413                                "rc = %ld\n", osd_ino2name(inode),
1414                                csize, *offs, inode->i_ino,
1415                                PTR_ERR(bh));
1416                         return PTR_ERR(bh);
1417                 }
1418
1419                 if (bh != NULL) {
1420                         memcpy(buf, bh->b_data + boffs, csize);
1421                         brelse(bh);
1422                 } else {
1423                         memset(buf, 0, csize);
1424                 }
1425
1426                 *offs += csize;
1427                 buf += csize;
1428                 size -= csize;
1429         }
1430         return osize;
1431 }
1432
1433 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1434                         struct lu_buf *buf, loff_t *pos)
1435 {
1436         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1437         int           rc;
1438
1439         /* Read small symlink from inode body as we need to maintain correct
1440          * on-disk symlinks for ldiskfs.
1441          */
1442         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1443             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1444                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1445         else
1446                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1447
1448         return rc;
1449 }
1450
1451 static inline int osd_extents_enabled(struct super_block *sb,
1452                                       struct inode *inode)
1453 {
1454         if (inode != NULL) {
1455                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1456                         return 1;
1457         } else if (ldiskfs_has_feature_extents(sb)) {
1458                 return 1;
1459         }
1460         return 0;
1461 }
1462
1463 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1464                            const loff_t size, const loff_t pos,
1465                            const int blocks)
1466 {
1467         int credits, bits, bs, i;
1468
1469         bits = sb->s_blocksize_bits;
1470         bs = 1 << bits;
1471
1472         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1473          * we do not expect blockmaps on the large files,
1474          * so let's shrink it to 2 levels (4GB files) */
1475
1476         /* this is default reservation: 2 levels */
1477         credits = (blocks + 2) * 3;
1478
1479         /* actual offset is unknown, hard to optimize */
1480         if (pos == -1)
1481                 return credits;
1482
1483         /* now check for few specific cases to optimize */
1484         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1485                 /* no indirects */
1486                 credits = blocks;
1487                 /* allocate if not allocated */
1488                 if (inode == NULL) {
1489                         credits += blocks * 2;
1490                         return credits;
1491                 }
1492                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1493                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1494                         if (LDISKFS_I(inode)->i_data[i] == 0)
1495                                 credits += 2;
1496                 }
1497         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1498                 /* single indirect */
1499                 credits = blocks * 3;
1500                 if (inode == NULL ||
1501                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1502                         credits += 3;
1503                 else
1504                         /* The indirect block may be modified. */
1505                         credits += 1;
1506         }
1507
1508         return credits;
1509 }
1510
1511 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1512                                  const struct lu_buf *buf, loff_t _pos,
1513                                  struct thandle *handle)
1514 {
1515         struct osd_object  *obj  = osd_dt_obj(dt);
1516         struct inode       *inode = obj->oo_inode;
1517         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1518         struct osd_thandle *oh;
1519         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1520         int                 bits, bs;
1521         int                 depth, size;
1522         loff_t              pos;
1523         ENTRY;
1524
1525         LASSERT(buf != NULL);
1526         LASSERT(handle != NULL);
1527
1528         oh = container_of0(handle, struct osd_thandle, ot_super);
1529         LASSERT(oh->ot_handle == NULL);
1530
1531         size = buf->lb_len;
1532         bits = sb->s_blocksize_bits;
1533         bs = 1 << bits;
1534
1535         if (_pos == -1) {
1536                 /* if this is an append, then we
1537                  * should expect cross-block record */
1538                 pos = 0;
1539         } else {
1540                 pos = _pos;
1541         }
1542
1543         /* blocks to modify */
1544         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1545         LASSERT(blocks > 0);
1546
1547         if (inode != NULL && _pos != -1) {
1548                 /* object size in blocks */
1549                 est = (i_size_read(inode) + bs - 1) >> bits;
1550                 allocated = inode->i_blocks >> (bits - 9);
1551                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1552                         /* looks like an overwrite, no need to modify tree */
1553                         credits = blocks;
1554                         /* no need to modify i_size */
1555                         goto out;
1556                 }
1557         }
1558
1559         if (osd_extents_enabled(sb, inode)) {
1560                 /*
1561                  * many concurrent threads may grow tree by the time
1562                  * our transaction starts. so, consider 2 is a min depth
1563                  * for every level we may need to allocate a new block
1564                  * and take some entries from the old one. so, 3 blocks
1565                  * to allocate (bitmap, gd, itself) + old block - 4 per
1566                  * level.
1567                  */
1568                 depth = inode != NULL ? ext_depth(inode) : 0;
1569                 depth = max(depth, 1) + 1;
1570                 credits = depth;
1571                 /* if not append, then split may need to modify
1572                  * existing blocks moving entries into the new ones */
1573                 if (_pos != -1)
1574                         credits += depth;
1575                 /* blocks to store data: bitmap,gd,itself */
1576                 credits += blocks * 3;
1577         } else {
1578                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1579         }
1580         /* if inode is created as part of the transaction,
1581          * then it's counted already by the creation method */
1582         if (inode != NULL)
1583                 credits++;
1584
1585 out:
1586
1587         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1588
1589         /* dt_declare_write() is usually called for system objects, such
1590          * as llog or last_rcvd files. We needn't enforce quota on those
1591          * objects, so always set the lqi_space as 0. */
1592         if (inode != NULL)
1593                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1594                                            i_gid_read(inode),
1595                                            i_projid_read(inode), 0,
1596                                            oh, obj, NULL, OSD_QID_BLK);
1597
1598         if (rc == 0)
1599                 rc = osd_trunc_lock(obj, oh, true);
1600
1601         RETURN(rc);
1602 }
1603
1604 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1605 {
1606         /* LU-2634: clear the extent format for fast symlink */
1607         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1608
1609         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1610         spin_lock(&inode->i_lock);
1611         LDISKFS_I(inode)->i_disksize = buflen;
1612         i_size_write(inode, buflen);
1613         spin_unlock(&inode->i_lock);
1614         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1615
1616         return 0;
1617 }
1618
1619 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1620                              int write_NUL, loff_t *offs, handle_t *handle)
1621 {
1622         struct buffer_head *bh        = NULL;
1623         loff_t              offset    = *offs;
1624         loff_t              new_size  = i_size_read(inode);
1625         unsigned long       block;
1626         int                 blocksize = 1 << inode->i_blkbits;
1627         int                 err = 0;
1628         int                 size;
1629         int                 boffs;
1630         int                 dirty_inode = 0;
1631
1632         if (write_NUL) {
1633                 /*
1634                  * long symlink write does not count the NUL terminator in
1635                  * bufsize, we write it, and the inode's file size does not
1636                  * count the NUL terminator as well.
1637                  */
1638                 ((char *)buf)[bufsize] = '\0';
1639                 ++bufsize;
1640         }
1641
1642         while (bufsize > 0) {
1643                 int credits = handle->h_buffer_credits;
1644
1645                 if (bh)
1646                         brelse(bh);
1647
1648                 block = offset >> inode->i_blkbits;
1649                 boffs = offset & (blocksize - 1);
1650                 size = min(blocksize - boffs, bufsize);
1651                 bh = __ldiskfs_bread(handle, inode, block, 1);
1652                 if (IS_ERR_OR_NULL(bh)) {
1653                         if (bh == NULL) {
1654                                 err = -EIO;
1655                         } else {
1656                                 err = PTR_ERR(bh);
1657                                 bh = NULL;
1658                         }
1659
1660                         CERROR("%s: error reading offset %llu (block %lu, "
1661                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
1662                                inode->i_sb->s_id, offset, block, bufsize, *offs,
1663                                credits, handle->h_buffer_credits, err);
1664                         break;
1665                 }
1666
1667                 err = ldiskfs_journal_get_write_access(handle, bh);
1668                 if (err) {
1669                         CERROR("journal_get_write_access() returned error %d\n",
1670                                err);
1671                         break;
1672                 }
1673                 LASSERTF(boffs + size <= bh->b_size,
1674                          "boffs %d size %d bh->b_size %lu\n",
1675                          boffs, size, (unsigned long)bh->b_size);
1676                 memcpy(bh->b_data + boffs, buf, size);
1677                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1678                 if (err)
1679                         break;
1680
1681                 if (offset + size > new_size)
1682                         new_size = offset + size;
1683                 offset += size;
1684                 bufsize -= size;
1685                 buf += size;
1686         }
1687         if (bh)
1688                 brelse(bh);
1689
1690         if (write_NUL)
1691                 --new_size;
1692         /* correct in-core and on-disk sizes */
1693         if (new_size > i_size_read(inode)) {
1694                 spin_lock(&inode->i_lock);
1695                 if (new_size > i_size_read(inode))
1696                         i_size_write(inode, new_size);
1697                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1698                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1699                         dirty_inode = 1;
1700                 }
1701                 spin_unlock(&inode->i_lock);
1702                 if (dirty_inode)
1703                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1704         }
1705
1706         if (err == 0)
1707                 *offs = offset;
1708         return err;
1709 }
1710
1711 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1712                          const struct lu_buf *buf, loff_t *pos,
1713                          struct thandle *handle)
1714 {
1715         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1716         struct osd_thandle      *oh;
1717         ssize_t                 result;
1718         int                     is_link;
1719
1720         LASSERT(dt_object_exists(dt));
1721
1722         LASSERT(handle != NULL);
1723         LASSERT(inode != NULL);
1724         dquot_initialize(inode);
1725
1726         /* XXX: don't check: one declared chunk can be used many times */
1727         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1728
1729         oh = container_of(handle, struct osd_thandle, ot_super);
1730         LASSERT(oh->ot_handle->h_transaction != NULL);
1731         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1732
1733         /* Write small symlink to inode body as we need to maintain correct
1734          * on-disk symlinks for ldiskfs.
1735          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1736          * does not count it in.
1737          */
1738         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1739         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1740                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1741         else
1742                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1743                                                   buf->lb_len, is_link, pos,
1744                                                   oh->ot_handle);
1745         if (result == 0)
1746                 result = buf->lb_len;
1747
1748         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1749
1750         return result;
1751 }
1752
1753 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1754                              __u64 start, __u64 end, struct thandle *th)
1755 {
1756         struct osd_thandle *oh;
1757         struct inode       *inode;
1758         int                 rc;
1759         ENTRY;
1760
1761         LASSERT(th);
1762         oh = container_of(th, struct osd_thandle, ot_super);
1763
1764         /*
1765          * we don't need to reserve credits for whole truncate
1766          * it's not possible as truncate may need to free too many
1767          * blocks and that won't fit a single transaction. instead
1768          * we reserve credits to change i_size and put inode onto
1769          * orphan list. if needed truncate will extend or restart
1770          * transaction
1771          */
1772         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1773                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1774
1775         inode = osd_dt_obj(dt)->oo_inode;
1776         LASSERT(inode);
1777
1778         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1779                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1780                                    NULL, OSD_QID_BLK);
1781
1782         if (rc == 0)
1783                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
1784
1785         RETURN(rc);
1786 }
1787
1788 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1789                      __u64 start, __u64 end, struct thandle *th)
1790 {
1791         struct osd_object *obj = osd_dt_obj(dt);
1792         struct osd_device *osd = osd_obj2dev(obj);
1793         struct inode *inode = obj->oo_inode;
1794         struct osd_access_lock *al;
1795         struct osd_thandle *oh;
1796         int rc = 0, found = 0;
1797         bool grow = false;
1798         ENTRY;
1799
1800         LASSERT(end == OBD_OBJECT_EOF);
1801         LASSERT(dt_object_exists(dt));
1802         LASSERT(osd_invariant(obj));
1803         LASSERT(inode != NULL);
1804         dquot_initialize(inode);
1805
1806         LASSERT(th);
1807         oh = container_of(th, struct osd_thandle, ot_super);
1808         LASSERT(oh->ot_handle->h_transaction != NULL);
1809
1810         /* we used to skip truncate to current size to
1811          * optimize truncates on OST. with DoM we can
1812          * get attr_set to set specific size (MDS_REINT)
1813          * and then get truncate RPC which essentially
1814          * would be skipped. this is bad.. so, disable
1815          * this optimization on MDS till the client stop
1816          * to sent MDS_REINT (LU-11033) -bzzz */
1817         if (osd->od_is_ost && i_size_read(inode) == start)
1818                 RETURN(0);
1819
1820         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1821
1822         spin_lock(&inode->i_lock);
1823         if (i_size_read(inode) < start)
1824                 grow = true;
1825         i_size_write(inode, start);
1826         spin_unlock(&inode->i_lock);
1827         ll_truncate_pagecache(inode, start);
1828
1829         /* optimize grow case */
1830         if (grow) {
1831                 osd_execute_truncate(obj);
1832                 GOTO(out, rc);
1833         }
1834
1835         /* add to orphan list to ensure truncate completion
1836          * if this transaction succeed. ldiskfs_truncate()
1837          * will take the inode out of the list */
1838         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
1839         if (rc != 0)
1840                 GOTO(out, rc);
1841
1842         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
1843                 if (obj != al->tl_obj)
1844                         continue;
1845                 LASSERT(al->tl_shared == 0);
1846                 found = 1;
1847                 /* do actual truncate in osd_trans_stop() */
1848                 al->tl_truncate = 1;
1849                 break;
1850         }
1851         LASSERT(found);
1852
1853 out:
1854         RETURN(rc);
1855 }
1856
1857 static int fiemap_check_ranges(struct inode *inode,
1858                                u64 start, u64 len, u64 *new_len)
1859 {
1860         loff_t maxbytes;
1861
1862         *new_len = len;
1863
1864         if (len == 0)
1865                 return -EINVAL;
1866
1867         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1868                 maxbytes = inode->i_sb->s_maxbytes;
1869         else
1870                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1871
1872         if (start > maxbytes)
1873                 return -EFBIG;
1874
1875         /*
1876          * Shrink request scope to what the fs can actually handle.
1877          */
1878         if (len > maxbytes || (maxbytes - len) < start)
1879                 *new_len = maxbytes - start;
1880
1881         return 0;
1882 }
1883
1884 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1885 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
1886
1887 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1888                           struct fiemap *fm)
1889 {
1890         struct fiemap_extent_info fieinfo = {0, };
1891         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1892         u64 len;
1893         int rc;
1894         mm_segment_t cur_fs;
1895
1896         LASSERT(inode);
1897         if (inode->i_op->fiemap == NULL)
1898                 return -EOPNOTSUPP;
1899
1900         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1901                 return -EINVAL;
1902
1903         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1904         if (rc)
1905                 return rc;
1906
1907         fieinfo.fi_flags = fm->fm_flags;
1908         fieinfo.fi_extents_max = fm->fm_extent_count;
1909         fieinfo.fi_extents_start = fm->fm_extents;
1910
1911         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1912                 filemap_write_and_wait(inode->i_mapping);
1913
1914         /* Save previous value address limit */
1915         cur_fs = get_fs();
1916         /* Set the address limit of the kernel */
1917         set_fs(KERNEL_DS);
1918
1919         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
1920         fm->fm_flags = fieinfo.fi_flags;
1921         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
1922
1923         /* Restore the previous address limt */
1924         set_fs(cur_fs);
1925
1926         return rc;
1927 }
1928
1929 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1930                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1931 {
1932         int              rc = 0;
1933         struct inode    *inode = osd_dt_obj(dt)->oo_inode;
1934         ENTRY;
1935
1936         switch (advice) {
1937         case LU_LADVISE_DONTNEED:
1938                 if (end == 0)
1939                         break;
1940                 invalidate_mapping_pages(inode->i_mapping,
1941                                          start >> PAGE_SHIFT,
1942                                          (end - 1) >> PAGE_SHIFT);
1943                 break;
1944         default:
1945                 rc = -ENOTSUPP;
1946                 break;
1947         }
1948
1949         RETURN(rc);
1950 }
1951
1952 /*
1953  * in some cases we may need declare methods for objects being created
1954  * e.g., when we create symlink
1955  */
1956 const struct dt_body_operations osd_body_ops_new = {
1957         .dbo_declare_write = osd_declare_write,
1958 };
1959
1960 const struct dt_body_operations osd_body_ops = {
1961         .dbo_read                       = osd_read,
1962         .dbo_declare_write              = osd_declare_write,
1963         .dbo_write                      = osd_write,
1964         .dbo_bufs_get                   = osd_bufs_get,
1965         .dbo_bufs_put                   = osd_bufs_put,
1966         .dbo_write_prep                 = osd_write_prep,
1967         .dbo_declare_write_commit       = osd_declare_write_commit,
1968         .dbo_write_commit               = osd_write_commit,
1969         .dbo_read_prep                  = osd_read_prep,
1970         .dbo_declare_punch              = osd_declare_punch,
1971         .dbo_punch                      = osd_punch,
1972         .dbo_fiemap_get                 = osd_fiemap_get,
1973         .dbo_ladvise                    = osd_ladvise,
1974 };
1975
1976 /**
1977  * Get a truncate lock
1978  *
1979  * In order to take multi-transaction truncate out of main transaction we let
1980  * the caller grab a lock on the object passed. the lock can be shared (for
1981  * writes) and exclusive (for truncate). It's not allowed to mix truncate
1982  * and write in the same transaction handle (do not confuse with big ldiskfs
1983  * transaction containing lots of handles).
1984  * The lock must be taken at declaration.
1985  *
1986  * \param obj           object to lock
1987  * \oh                  transaction
1988  * \shared              shared or exclusive
1989  *
1990  * \retval 0            lock is granted
1991  * \retval -NOMEM       no memory to allocate lock
1992  */
1993 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
1994 {
1995         struct osd_access_lock *al, *tmp;
1996
1997         LASSERT(obj);
1998         LASSERT(oh);
1999
2000         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2001                 if (tmp->tl_obj != obj)
2002                         continue;
2003                 LASSERT(tmp->tl_shared == shared);
2004                 /* found same lock */
2005                 return 0;
2006         }
2007
2008         OBD_ALLOC_PTR(al);
2009         if (unlikely(al == NULL))
2010                 return -ENOMEM;
2011         al->tl_obj = obj;
2012         al->tl_truncate = false;
2013         if (shared)
2014                 down_read(&obj->oo_ext_idx_sem);
2015         else
2016                 down_write(&obj->oo_ext_idx_sem);
2017         al->tl_shared = shared;
2018
2019         list_add(&al->tl_list, &oh->ot_trunc_locks);
2020
2021         return 0;
2022 }
2023
2024 void osd_trunc_unlock_all(struct list_head *list)
2025 {
2026         struct osd_access_lock *al, *tmp;
2027         list_for_each_entry_safe(al, tmp, list, tl_list) {
2028                 if (al->tl_shared)
2029                         up_read(&al->tl_obj->oo_ext_idx_sem);
2030                 else
2031                         up_write(&al->tl_obj->oo_ext_idx_sem);
2032                 list_del(&al->tl_list);
2033                 OBD_FREE_PTR(al);
2034         }
2035 }
2036
2037 void osd_execute_truncate(struct osd_object *obj)
2038 {
2039         struct osd_device *d = osd_obj2dev(obj);
2040         struct inode *inode = obj->oo_inode;
2041         __u64 size;
2042
2043         /* simulate crash before (in the middle) of delayed truncate */
2044         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2045                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2046                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2047
2048                 mutex_lock(&sbi->s_orphan_lock);
2049                 list_del_init(&ei->i_orphan);
2050                 mutex_unlock(&sbi->s_orphan_lock);
2051                 return;
2052         }
2053
2054         ldiskfs_truncate(inode);
2055
2056         /*
2057          * For a partial-page truncate, flush the page to disk immediately to
2058          * avoid data corruption during direct disk write.  b=17397
2059          */
2060         size = i_size_read(inode);
2061         if ((size & ~PAGE_MASK) == 0)
2062                 return;
2063         if (osd_use_page_cache(d)) {
2064                 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2065         } else {
2066                 /* Notice we use "wait" version to ensure I/O is complete */
2067                 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2068                 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2069                                          size >> PAGE_SHIFT);
2070         }
2071 }
2072
2073 void osd_process_truncates(struct list_head *list)
2074 {
2075         struct osd_access_lock *al;
2076
2077         LASSERT(journal_current_handle() == NULL);
2078
2079         list_for_each_entry(al, list, tl_list) {
2080                 if (al->tl_shared)
2081                         continue;
2082                 if (!al->tl_truncate)
2083                         continue;
2084                 osd_execute_truncate(al->tl_obj);
2085         }
2086 }