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