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                                        osd_bio_nr_segs(bio),
 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, int maxlnb)
 615 {
 616         int rc = 0;
 617         ENTRY;
 618
 619         *nrpages = 0;
 620
 621         while (len > 0) {
 622                 int poff = offset & (PAGE_SIZE - 1);
 623                 int plen = PAGE_SIZE - poff;
 624
 625                 if (*nrpages >= maxlnb) {
 626                         rc = -EOVERFLOW;
 627                         break;
 628                 }
 629
 630                 if (plen > len)
 631                         plen = len;
 632                 lnb->lnb_file_offset = offset;
 633                 lnb->lnb_page_offset = poff;
 634                 lnb->lnb_len = plen;
 635                 /* lnb->lnb_flags = rnb->rnb_flags; */
 636                 lnb->lnb_flags = 0;
 637                 lnb->lnb_page = NULL;
 638                 lnb->lnb_rc = 0;
 639                 lnb->lnb_guard_rpc = 0;
 640                 lnb->lnb_guard_disk = 0;
 641                 lnb->lnb_locked = 0;
 642
 643                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 644                          (long long) len);
 645                 offset += plen;
 646                 len -= plen;
 647                 lnb++;
 648                 (*nrpages)++;
 649         }
 650
 651         RETURN(rc);
 652 }
 653
 654 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
 655                                  loff_t offset, gfp_t gfp_mask, bool cache)
 656 {
 657         struct osd_thread_info *oti = osd_oti_get(env);
 658         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 659         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 660         struct page *page;
 661         int cur;
 662
 663         LASSERT(inode);
 664
 665         if (cache) {
 666                 page = find_or_create_page(inode->i_mapping,
 667                                            offset >> PAGE_SHIFT, gfp_mask);
 668
 669                 if (likely(page))
 670                         LASSERT(!test_bit(PG_private_2, &page->flags));
 671                 else
 672                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 673
 674                 return page;
 675         }
 676
 677         if (inode->i_mapping->nrpages) {
 678                 /* consult with pagecache, but do not create new pages */
 679                 /* this is normally used once */
 680                 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
 681                 if (page)
 682                         return page;
 683         }
 684
 685         LASSERT(oti->oti_dio_pages);
 686         cur = oti->oti_dio_pages_used;
 687
 688         if (unlikely(!oti->oti_dio_pages[cur])) {
 689                 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
 690                 page = alloc_page(gfp_mask);
 691                 if (!page)
 692                         return NULL;
 693                 oti->oti_dio_pages[cur] = page;
 694         }
 695
 696         page = oti->oti_dio_pages[cur];
 697         LASSERT(!test_bit(PG_private_2, &page->flags));
 698         set_bit(PG_private_2, &page->flags);
 699         oti->oti_dio_pages_used++;
 700
 701         LASSERT(!PageLocked(page));
 702         lock_page(page);
 703
 704         LASSERT(!page->mapping);
 705         LASSERT(!PageWriteback(page));
 706         ClearPageUptodate(page);
 707
 708         page->index = offset >> PAGE_SHIFT;
 709
 710         return page;
 711 }
 712
 713 /*
 714  * there are following "locks":
 715  * journal_start
 716  * i_mutex
 717  * page lock
 718  *
 719  * osd write path:
 720  *  - lock page(s)
 721  *  - journal_start
 722  *  - truncate_sem
 723  *
 724  * ext4 vmtruncate:
 725  *  - lock pages, unlock
 726  *  - journal_start
 727  *  - lock partial page
 728  *  - i_data_sem
 729  *
 730  */
 731
 732 /**
 733  * Unlock and release pages loaded by osd_bufs_get()
 734  *
 735  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 736  *
 737  * \param env           thread execution environment
 738  * \param dt            dt object undergoing IO (OSD object + methods)
 739  * \param lnb           array of pages undergoing IO
 740  * \param npages        number of pages in \a lnb
 741  *
 742  * \retval 0            always
 743  */
 744 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 745                         struct niobuf_local *lnb, int npages)
 746 {
 747         struct osd_thread_info *oti = osd_oti_get(env);
 748         struct pagevec pvec;
 749         int i;
 750
 751         ll_pagevec_init(&pvec, 0);
 752
 753         for (i = 0; i < npages; i++) {
 754                 struct page *page = lnb[i].lnb_page;
 755
 756                 if (page == NULL)
 757                         continue;
 758
 759                 /* if the page isn't cached, then reset uptodate
 760                  * to prevent reuse */
 761                 if (test_bit(PG_private_2, &page->flags)) {
 762                         clear_bit(PG_private_2, &page->flags);
 763                         ClearPageUptodate(page);
 764                         if (lnb[i].lnb_locked)
 765                                 unlock_page(page);
 766                         oti->oti_dio_pages_used--;
 767                 } else {
 768                         if (lnb[i].lnb_locked)
 769                                 unlock_page(page);
 770                         if (pagevec_add(&pvec, page) == 0)
 771                                 pagevec_release(&pvec);
 772                 }
 773
 774                 lnb[i].lnb_page = NULL;
 775         }
 776
 777         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
 778
 779         /* Release any partial pagevec */
 780         pagevec_release(&pvec);
 781
 782         RETURN(0);
 783 }
 784
 785 /**
 786  * Load and lock pages undergoing IO
 787  *
 788  * Pages as described in the \a lnb array are fetched (from disk or cache)
 789  * and locked for IO by the caller.
 790  *
 791  * DLM locking protects us from write and truncate competing for same region,
 792  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 793  * It's possible the writeout on a such a page is in progress when we access
 794  * it. It's also possible that during this writeout we put new (partial) data
 795  * into the page, but won't be able to proceed in filter_commitrw_write().
 796  * Therefore, just wait for writeout completion as it should be rare enough.
 797  *
 798  * \param env           thread execution environment
 799  * \param dt            dt object undergoing IO (OSD object + methods)
 800  * \param pos           byte offset of IO start
 801  * \param len           number of bytes of IO
 802  * \param lnb           array of extents undergoing IO
 803  * \param rw            read or write operation, and other flags
 804  * \param capa          capabilities
 805  *
 806  * \retval pages        (zero or more) loaded successfully
 807  * \retval -ENOMEM      on memory/page allocation error
 808  */
 809 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 810                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 811                         int maxlnb, enum dt_bufs_type rw)
 812 {
 813         struct osd_thread_info *oti = osd_oti_get(env);
 814         struct osd_object *obj = osd_dt_obj(dt);
 815         struct osd_device *osd   = osd_obj2dev(obj);
 816         int npages, i, iosize, rc = 0;
 817         bool cache, write;
 818         loff_t fsize;
 819         gfp_t gfp_mask;
 820
 821         LASSERT(obj->oo_inode);
 822
 823         rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
 824         if (rc)
 825                 RETURN(rc);
 826
 827         write = rw & DT_BUFS_TYPE_WRITE;
 828
 829         fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
 830         iosize = fsize - lnb[0].lnb_file_offset;
 831         fsize = max(fsize, i_size_read(obj->oo_inode));
 832
 833         cache = rw & DT_BUFS_TYPE_READAHEAD;
 834         if (cache)
 835                 goto bypass_checks;
 836
 837         cache = osd_use_page_cache(osd);
 838         while (cache) {
 839                 if (write) {
 840                         if (!osd->od_writethrough_cache) {
 841                                 cache = false;
 842                                 break;
 843                         }
 844                         if (iosize > osd->od_writethrough_max_iosize) {
 845                                 cache = false;
 846                                 break;
 847                         }
 848                 } else {
 849                         if (!osd->od_read_cache) {
 850                                 cache = false;
 851                                 break;
 852                         }
 853                         if (iosize > osd->od_readcache_max_iosize) {
 854                                 cache = false;
 855                                 break;
 856                         }
 857                 }
 858                 /* don't use cache on large files */
 859                 if (osd->od_readcache_max_filesize &&
 860                     fsize > osd->od_readcache_max_filesize)
 861                         cache = false;
 862                 break;
 863         }
 864
 865 bypass_checks:
 866         if (!cache && unlikely(!oti->oti_dio_pages)) {
 867                 OBD_ALLOC(oti->oti_dio_pages,
 868                           sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
 869                 if (!oti->oti_dio_pages)
 870                         return -ENOMEM;
 871         }
 872
 873         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 874         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 875                                              GFP_HIGHUSER;
 876         for (i = 0; i < npages; i++, lnb++) {
 877                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
 878                                              gfp_mask, cache);
 879                 if (lnb->lnb_page == NULL)
 880                         GOTO(cleanup, rc = -ENOMEM);
 881
 882                 lnb->lnb_locked = 1;
 883                 wait_on_page_writeback(lnb->lnb_page);
 884                 BUG_ON(PageWriteback(lnb->lnb_page));
 885         }
 886
 887 #if 0
 888         /* XXX: this version doesn't invalidate cached pages, but use them */
 889         if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
 890                 /* do not allow data aliasing, invalidate pagecache */
 891                 /* XXX: can be quite expensive in mixed case */
 892                 invalidate_mapping_pages(obj->oo_inode->i_mapping,
 893                                 lnb[0].lnb_file_offset >> PAGE_SHIFT,
 894                                 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
 895         }
 896 #endif
 897
 898         RETURN(i);
 899
 900 cleanup:
 901         if (i > 0)
 902                 osd_bufs_put(env, dt, lnb - i, i);
 903         return rc;
 904 }
 905
 906 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 907                                        int pages, sector_t *blocks,
 908                                        int create)
 909 {
 910         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 911         int rc = 0, i = 0;
 912         struct page *fp = NULL;
 913         int clen = 0;
 914         pgoff_t max_page_index;
 915         handle_t *handle = NULL;
 916
 917         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
 918
 919         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 920                 inode->i_ino, pages, (*page)->index);
 921
 922         if (create) {
 923                 create = LDISKFS_GET_BLOCKS_CREATE;
 924                 handle = ldiskfs_journal_current_handle();
 925                 LASSERT(handle != NULL);
 926                 rc = osd_attach_jinode(inode);
 927                 if (rc)
 928                         return rc;
 929         }
 930         /* pages are sorted already. so, we just have to find
 931          * contig. space and process them properly */
 932         while (i < pages) {
 933                 long blen, total = 0;
 934                 struct ldiskfs_map_blocks map = { 0 };
 935
 936                 if (fp == NULL) { /* start new extent */
 937                         fp = *page++;
 938                         clen = 1;
 939                         if (++i != pages)
 940                                 continue;
 941                 } else if (fp->index + clen == (*page)->index) {
 942                         /* continue the extent */
 943                         page++;
 944                         clen++;
 945                         if (++i != pages)
 946                                 continue;
 947                 }
 948                 if (fp->index + clen >= max_page_index)
 949                         GOTO(cleanup, rc = -EFBIG);
 950                 /* process found extent */
 951                 map.m_lblk = fp->index * blocks_per_page;
 952                 map.m_len = blen = clen * blocks_per_page;
 953 cont_map:
 954                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 955                 if (rc >= 0) {
 956                         int c = 0;
 957                         for (; total < blen && c < map.m_len; c++, total++) {
 958                                 if (rc == 0) {
 959                                         *(blocks + total) = 0;
 960                                         total++;
 961                                         break;
 962                                 } else {
 963                                         *(blocks + total) = map.m_pblk + c;
 964                                         /* unmap any possible underlying
 965                                          * metadata from the block device
 966                                          * mapping.  bug 6998. */
 967                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
 968                                             create)
 969                                                 clean_bdev_aliases(
 970                                                         inode->i_sb->s_bdev,
 971                                                         map.m_pblk + c, 1);
 972                                 }
 973                         }
 974                         rc = 0;
 975                 }
 976                 if (rc == 0 && total < blen) {
 977                         map.m_lblk = fp->index * blocks_per_page + total;
 978                         map.m_len = blen - total;
 979                         goto cont_map;
 980                 }
 981                 if (rc != 0)
 982                         GOTO(cleanup, rc);
 983
 984                 /* look for next extent */
 985                 fp = NULL;
 986                 blocks += blocks_per_page * clen;
 987         }
 988 cleanup:
 989         return rc;
 990 }
 991
 992 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 993                           struct niobuf_local *lnb, int npages)
 994 {
 995         struct osd_thread_info *oti   = osd_oti_get(env);
 996         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 997         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 998         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 999         ktime_t start, end;
1000         s64 timediff;
1001         ssize_t isize;
1002         __s64  maxidx;
1003         int i, rc = 0;
1004
1005         LASSERT(inode);
1006
1007         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1008         if (unlikely(rc != 0))
1009                 RETURN(rc);
1010
1011         isize = i_size_read(inode);
1012         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1013
1014         start = ktime_get();
1015         for (i = 0; i < npages; i++) {
1016
1017                 /*
1018                  * till commit the content of the page is undefined
1019                  * we'll set it uptodate once bulk is done. otherwise
1020                  * subsequent reads can access non-stable data
1021                  */
1022                 ClearPageUptodate(lnb[i].lnb_page);
1023
1024                 if (lnb[i].lnb_len == PAGE_SIZE)
1025                         continue;
1026
1027                 if (maxidx >= lnb[i].lnb_page->index) {
1028                         osd_iobuf_add_page(iobuf, &lnb[i]);
1029                 } else {
1030                         long off;
1031                         char *p = kmap(lnb[i].lnb_page);
1032
1033                         off = lnb[i].lnb_page_offset;
1034                         if (off)
1035                                 memset(p, 0, off);
1036                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1037                               ~PAGE_MASK;
1038                         if (off)
1039                                 memset(p + off, 0, PAGE_SIZE - off);
1040                         kunmap(lnb[i].lnb_page);
1041                 }
1042         }
1043         end = ktime_get();
1044         timediff = ktime_us_delta(end, start);
1045         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1046
1047         if (iobuf->dr_npages) {
1048                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1049                                                  iobuf->dr_npages,
1050                                                  iobuf->dr_blocks, 0);
1051                 if (likely(rc == 0)) {
1052                         rc = osd_do_bio(osd, inode, iobuf);
1053                         /* do IO stats for preparation reads */
1054                         osd_fini_iobuf(osd, iobuf);
1055                 }
1056         }
1057         RETURN(rc);
1058 }
1059
1060 struct osd_fextent {
1061         sector_t        start;
1062         sector_t        end;
1063         unsigned int    mapped:1;
1064 };
1065
1066 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1067                          struct osd_fextent *cached_extent)
1068 {
1069         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1070         sector_t block = offset >> inode->i_blkbits;
1071         sector_t start;
1072         struct fiemap_extent_info fei = { 0 };
1073         struct fiemap_extent fe = { 0 };
1074         mm_segment_t saved_fs;
1075         int rc;
1076
1077         if (block >= cached_extent->start && block < cached_extent->end)
1078                 return cached_extent->mapped;
1079
1080         if (i_size_read(inode) == 0)
1081                 return 0;
1082
1083         /* Beyond EOF, must not be mapped */
1084         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1085                 return 0;
1086
1087         fei.fi_extents_max = 1;
1088         fei.fi_extents_start = &fe;
1089
1090         saved_fs = get_fs();
1091         set_fs(KERNEL_DS);
1092         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1093         set_fs(saved_fs);
1094         if (rc != 0)
1095                 return 0;
1096
1097         start = fe.fe_logical >> inode->i_blkbits;
1098
1099         if (start > block) {
1100                 cached_extent->start = block;
1101                 cached_extent->end = start;
1102                 cached_extent->mapped = 0;
1103         } else {
1104                 cached_extent->start = start;
1105                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1106                                       inode->i_blkbits;
1107                 cached_extent->mapped = 1;
1108         }
1109
1110         return cached_extent->mapped;
1111 }
1112
1113 static int osd_declare_write_commit(const struct lu_env *env,
1114                                     struct dt_object *dt,
1115                                     struct niobuf_local *lnb, int npages,
1116                                     struct thandle *handle)
1117 {
1118         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1119         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1120         struct osd_thandle      *oh;
1121         int                     extents = 1;
1122         int                     depth;
1123         int                     i;
1124         int                     newblocks;
1125         int                     rc = 0;
1126         int                     credits = 0;
1127         long long               quota_space = 0;
1128         struct osd_fextent      extent = { 0 };
1129         enum osd_quota_local_flags local_flags = 0;
1130         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1131         ENTRY;
1132
1133         LASSERT(handle != NULL);
1134         oh = container_of0(handle, struct osd_thandle, ot_super);
1135         LASSERT(oh->ot_handle == NULL);
1136
1137         newblocks = npages;
1138
1139         /* calculate number of extents (probably better to pass nb) */
1140         for (i = 0; i < npages; i++) {
1141                 if (i && lnb[i].lnb_file_offset !=
1142                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1143                         extents++;
1144
1145                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1146                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1147                 else
1148                         quota_space += PAGE_SIZE;
1149
1150                 /* ignore quota for the whole request if any page is from
1151                  * client cache or written by root.
1152                  *
1153                  * XXX once we drop the 1.8 client support, the checking
1154                  * for whether page is from cache can be simplified as:
1155                  * !(lnb[i].flags & OBD_BRW_SYNC)
1156                  *
1157                  * XXX we could handle this on per-lnb basis as done by
1158                  * grant. */
1159                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1160                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1161                     OBD_BRW_FROM_GRANT)
1162                         declare_flags |= OSD_QID_FORCE;
1163         }
1164
1165         /*
1166          * each extent can go into new leaf causing a split
1167          * 5 is max tree depth: inode + 4 index blocks
1168          * with blockmaps, depth is 3 at most
1169          */
1170         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1171                 /*
1172                  * many concurrent threads may grow tree by the time
1173                  * our transaction starts. so, consider 2 is a min depth
1174                  */
1175                 depth = ext_depth(inode);
1176                 depth = max(depth, 1) + 1;
1177                 newblocks += depth;
1178                 credits++; /* inode */
1179                 credits += depth * 2 * extents;
1180         } else {
1181                 depth = 3;
1182                 newblocks += depth;
1183                 credits++; /* inode */
1184                 credits += depth * extents;
1185         }
1186
1187         /* quota space for metadata blocks */
1188         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1189
1190         /* quota space should be reported in 1K blocks */
1191         quota_space = toqb(quota_space);
1192
1193         /* each new block can go in different group (bitmap + gd) */
1194
1195         /* we can't dirty more bitmap blocks than exist */
1196         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1197                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1198         else
1199                 credits += newblocks;
1200
1201         /* we can't dirty more gd blocks than exist */
1202         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1203                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1204         else
1205                 credits += newblocks;
1206
1207         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1208
1209         /* make sure the over quota flags were not set */
1210         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1211
1212         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1213                                    i_projid_read(inode), quota_space, oh,
1214                                    osd_dt_obj(dt), &local_flags, declare_flags);
1215
1216         /* we need only to store the overquota flags in the first lnb for
1217          * now, once we support multiple objects BRW, this code needs be
1218          * revised. */
1219         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1220                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1221         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1222                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1223         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1224                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1225
1226         if (rc == 0)
1227                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1228
1229         RETURN(rc);
1230 }
1231
1232 /* Check if a block is allocated or not */
1233 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1234                             struct niobuf_local *lnb, int npages,
1235                             struct thandle *thandle)
1236 {
1237         struct osd_thread_info *oti = osd_oti_get(env);
1238         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1239         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1240         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1241         loff_t isize;
1242         int rc = 0, i;
1243
1244         LASSERT(inode);
1245
1246         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1247         if (unlikely(rc != 0))
1248                 RETURN(rc);
1249
1250         isize = i_size_read(inode);
1251         dquot_initialize(inode);
1252
1253         for (i = 0; i < npages; i++) {
1254                 if (lnb[i].lnb_rc == -ENOSPC &&
1255                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1256                         /* Allow the write to proceed if overwriting an
1257                          * existing block */
1258                         lnb[i].lnb_rc = 0;
1259                 }
1260
1261                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1262                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1263                                lnb[i].lnb_rc);
1264                         LASSERT(lnb[i].lnb_page);
1265                         generic_error_remove_page(inode->i_mapping,
1266                                                   lnb[i].lnb_page);
1267                         continue;
1268                 }
1269
1270                 LASSERT(PageLocked(lnb[i].lnb_page));
1271                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1272
1273                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1274                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1275
1276                 /*
1277                  * Since write and truncate are serialized by oo_sem, even
1278                  * partial-page truncate should not leave dirty pages in the
1279                  * page cache.
1280                  */
1281                 LASSERT(!PageDirty(lnb[i].lnb_page));
1282
1283                 SetPageUptodate(lnb[i].lnb_page);
1284
1285                 osd_iobuf_add_page(iobuf, &lnb[i]);
1286         }
1287
1288         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1289
1290         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1291                 rc = -ENOSPC;
1292         } else if (iobuf->dr_npages > 0) {
1293                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1294                                                  iobuf->dr_npages,
1295                                                  iobuf->dr_blocks, 1);
1296         } else {
1297                 /* no pages to write, no transno is needed */
1298                 thandle->th_local = 1;
1299         }
1300
1301         if (likely(rc == 0)) {
1302                 spin_lock(&inode->i_lock);
1303                 if (isize > i_size_read(inode)) {
1304                         i_size_write(inode, isize);
1305                         LDISKFS_I(inode)->i_disksize = isize;
1306                         spin_unlock(&inode->i_lock);
1307                         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1308                 } else {
1309                         spin_unlock(&inode->i_lock);
1310                 }
1311
1312                 rc = osd_do_bio(osd, inode, iobuf);
1313                 /* we don't do stats here as in read path because
1314                  * write is async: we'll do this in osd_put_bufs() */
1315         } else {
1316                 osd_fini_iobuf(osd, iobuf);
1317         }
1318
1319         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1320
1321         if (unlikely(rc != 0)) {
1322                 /* if write fails, we should drop pages from the cache */
1323                 for (i = 0; i < npages; i++) {
1324                         if (lnb[i].lnb_page == NULL)
1325                                 continue;
1326                         LASSERT(PageLocked(lnb[i].lnb_page));
1327                         generic_error_remove_page(inode->i_mapping,
1328                                                   lnb[i].lnb_page);
1329                 }
1330         }
1331
1332         RETURN(rc);
1333 }
1334
1335 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1336                          struct niobuf_local *lnb, int npages)
1337 {
1338         struct osd_thread_info *oti = osd_oti_get(env);
1339         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1340         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1341         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1342         int rc = 0, i, cache_hits = 0, cache_misses = 0;
1343         ktime_t start, end;
1344         s64 timediff;
1345         loff_t isize;
1346
1347         LASSERT(inode);
1348
1349         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1350         if (unlikely(rc != 0))
1351                 RETURN(rc);
1352
1353         isize = i_size_read(inode);
1354
1355         start = ktime_get();
1356         for (i = 0; i < npages; i++) {
1357
1358                 if (isize <= lnb[i].lnb_file_offset)
1359                         /* If there's no more data, abort early.
1360                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1361                         break;
1362
1363                 /* instead of looking if we go beyong isize, send complete
1364                  * pages all the time
1365                  */
1366                 lnb[i].lnb_rc = lnb[i].lnb_len;
1367
1368                 /* Bypass disk read if fail_loc is set properly */
1369                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1370                         SetPageUptodate(lnb[i].lnb_page);
1371
1372                 if (PageUptodate(lnb[i].lnb_page)) {
1373                         cache_hits++;
1374                         unlock_page(lnb[i].lnb_page);
1375                 } else {
1376                         cache_misses++;
1377                         osd_iobuf_add_page(iobuf, &lnb[i]);
1378                 }
1379                 /* no need to unlock in osd_bufs_put(), the sooner page is
1380                  * unlocked, the earlier another client can access it.
1381                  * notice real unlock_page() can be called few lines
1382                  * below after osd_do_bio(). lnb is a per-thread, so it's
1383                  * fine to have PG_locked and lnb_locked inconsistent here */
1384                 lnb[i].lnb_locked = 0;
1385         }
1386         end = ktime_get();
1387         timediff = ktime_us_delta(end, start);
1388         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1389
1390         if (cache_hits != 0)
1391                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1392                                     cache_hits);
1393         if (cache_misses != 0)
1394                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1395                                     cache_misses);
1396         if (cache_hits + cache_misses != 0)
1397                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1398                                     cache_hits + cache_misses);
1399
1400         if (iobuf->dr_npages) {
1401                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1402                                                  iobuf->dr_npages,
1403                                                  iobuf->dr_blocks, 0);
1404                 rc = osd_do_bio(osd, inode, iobuf);
1405
1406                 /* IO stats will be done in osd_bufs_put() */
1407
1408                 /* early release to let others read data during the bulk */
1409                 for (i = 0; i < iobuf->dr_npages; i++) {
1410                         LASSERT(PageLocked(iobuf->dr_pages[i]));
1411                         unlock_page(iobuf->dr_pages[i]);
1412                 }
1413         }
1414
1415         RETURN(rc);
1416 }
1417
1418 /*
1419  * XXX: Another layering violation for now.
1420  *
1421  * We don't want to use ->f_op->read methods, because generic file write
1422  *
1423  *         - serializes on ->i_sem, and
1424  *
1425  *         - does a lot of extra work like balance_dirty_pages(),
1426  *
1427  * which doesn't work for globally shared files like /last_rcvd.
1428  */
1429 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1430 {
1431         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1432
1433         memcpy(buffer, (char *)ei->i_data, buflen);
1434
1435         return  buflen;
1436 }
1437
1438 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1439 {
1440         struct buffer_head *bh;
1441         unsigned long block;
1442         int osize;
1443         int blocksize;
1444         int csize;
1445         int boffs;
1446
1447         /* prevent reading after eof */
1448         spin_lock(&inode->i_lock);
1449         if (i_size_read(inode) < *offs + size) {
1450                 loff_t diff = i_size_read(inode) - *offs;
1451                 spin_unlock(&inode->i_lock);
1452                 if (diff < 0) {
1453                         CDEBUG(D_OTHER,
1454                                "size %llu is too short to read @%llu\n",
1455                                i_size_read(inode), *offs);
1456                         return -EBADR;
1457                 } else if (diff == 0) {
1458                         return 0;
1459                 } else {
1460                         size = diff;
1461                 }
1462         } else {
1463                 spin_unlock(&inode->i_lock);
1464         }
1465
1466         blocksize = 1 << inode->i_blkbits;
1467         osize = size;
1468         while (size > 0) {
1469                 block = *offs >> inode->i_blkbits;
1470                 boffs = *offs & (blocksize - 1);
1471                 csize = min(blocksize - boffs, size);
1472                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1473                 if (IS_ERR(bh)) {
1474                         CERROR("%s: can't read %u@%llu on ino %lu: "
1475                                "rc = %ld\n", osd_ino2name(inode),
1476                                csize, *offs, inode->i_ino,
1477                                PTR_ERR(bh));
1478                         return PTR_ERR(bh);
1479                 }
1480
1481                 if (bh != NULL) {
1482                         memcpy(buf, bh->b_data + boffs, csize);
1483                         brelse(bh);
1484                 } else {
1485                         memset(buf, 0, csize);
1486                 }
1487
1488                 *offs += csize;
1489                 buf += csize;
1490                 size -= csize;
1491         }
1492         return osize;
1493 }
1494
1495 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1496                         struct lu_buf *buf, loff_t *pos)
1497 {
1498         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1499         int           rc;
1500
1501         /* Read small symlink from inode body as we need to maintain correct
1502          * on-disk symlinks for ldiskfs.
1503          */
1504         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1505             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1506                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1507         else
1508                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1509
1510         return rc;
1511 }
1512
1513 static inline int osd_extents_enabled(struct super_block *sb,
1514                                       struct inode *inode)
1515 {
1516         if (inode != NULL) {
1517                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1518                         return 1;
1519         } else if (ldiskfs_has_feature_extents(sb)) {
1520                 return 1;
1521         }
1522         return 0;
1523 }
1524
1525 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1526                            const loff_t size, const loff_t pos,
1527                            const int blocks)
1528 {
1529         int credits, bits, bs, i;
1530
1531         bits = sb->s_blocksize_bits;
1532         bs = 1 << bits;
1533
1534         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1535          * we do not expect blockmaps on the large files,
1536          * so let's shrink it to 2 levels (4GB files) */
1537
1538         /* this is default reservation: 2 levels */
1539         credits = (blocks + 2) * 3;
1540
1541         /* actual offset is unknown, hard to optimize */
1542         if (pos == -1)
1543                 return credits;
1544
1545         /* now check for few specific cases to optimize */
1546         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1547                 /* no indirects */
1548                 credits = blocks;
1549                 /* allocate if not allocated */
1550                 if (inode == NULL) {
1551                         credits += blocks * 2;
1552                         return credits;
1553                 }
1554                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1555                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1556                         if (LDISKFS_I(inode)->i_data[i] == 0)
1557                                 credits += 2;
1558                 }
1559         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1560                 /* single indirect */
1561                 credits = blocks * 3;
1562                 if (inode == NULL ||
1563                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1564                         credits += 3;
1565                 else
1566                         /* The indirect block may be modified. */
1567                         credits += 1;
1568         }
1569
1570         return credits;
1571 }
1572
1573 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1574                                  const struct lu_buf *buf, loff_t _pos,
1575                                  struct thandle *handle)
1576 {
1577         struct osd_object  *obj  = osd_dt_obj(dt);
1578         struct inode       *inode = obj->oo_inode;
1579         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1580         struct osd_thandle *oh;
1581         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1582         int                 bits, bs;
1583         int                 depth, size;
1584         loff_t              pos;
1585         ENTRY;
1586
1587         LASSERT(buf != NULL);
1588         LASSERT(handle != NULL);
1589
1590         oh = container_of0(handle, struct osd_thandle, ot_super);
1591         LASSERT(oh->ot_handle == NULL);
1592
1593         size = buf->lb_len;
1594         bits = sb->s_blocksize_bits;
1595         bs = 1 << bits;
1596
1597         if (_pos == -1) {
1598                 /* if this is an append, then we
1599                  * should expect cross-block record */
1600                 pos = 0;
1601         } else {
1602                 pos = _pos;
1603         }
1604
1605         /* blocks to modify */
1606         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1607         LASSERT(blocks > 0);
1608
1609         if (inode != NULL && _pos != -1) {
1610                 /* object size in blocks */
1611                 est = (i_size_read(inode) + bs - 1) >> bits;
1612                 allocated = inode->i_blocks >> (bits - 9);
1613                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1614                         /* looks like an overwrite, no need to modify tree */
1615                         credits = blocks;
1616                         /* no need to modify i_size */
1617                         goto out;
1618                 }
1619         }
1620
1621         if (osd_extents_enabled(sb, inode)) {
1622                 /*
1623                  * many concurrent threads may grow tree by the time
1624                  * our transaction starts. so, consider 2 is a min depth
1625                  * for every level we may need to allocate a new block
1626                  * and take some entries from the old one. so, 3 blocks
1627                  * to allocate (bitmap, gd, itself) + old block - 4 per
1628                  * level.
1629                  */
1630                 depth = inode != NULL ? ext_depth(inode) : 0;
1631                 depth = max(depth, 1) + 1;
1632                 credits = depth;
1633                 /* if not append, then split may need to modify
1634                  * existing blocks moving entries into the new ones */
1635                 if (_pos != -1)
1636                         credits += depth;
1637                 /* blocks to store data: bitmap,gd,itself */
1638                 credits += blocks * 3;
1639         } else {
1640                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1641         }
1642         /* if inode is created as part of the transaction,
1643          * then it's counted already by the creation method */
1644         if (inode != NULL)
1645                 credits++;
1646
1647 out:
1648
1649         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1650
1651         /* dt_declare_write() is usually called for system objects, such
1652          * as llog or last_rcvd files. We needn't enforce quota on those
1653          * objects, so always set the lqi_space as 0. */
1654         if (inode != NULL)
1655                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1656                                            i_gid_read(inode),
1657                                            i_projid_read(inode), 0,
1658                                            oh, obj, NULL, OSD_QID_BLK);
1659
1660         if (rc == 0)
1661                 rc = osd_trunc_lock(obj, oh, true);
1662
1663         RETURN(rc);
1664 }
1665
1666 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1667 {
1668         /* LU-2634: clear the extent format for fast symlink */
1669         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1670
1671         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1672         spin_lock(&inode->i_lock);
1673         LDISKFS_I(inode)->i_disksize = buflen;
1674         i_size_write(inode, buflen);
1675         spin_unlock(&inode->i_lock);
1676         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1677
1678         return 0;
1679 }
1680
1681 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
1682                                     int bufsize, int write_NUL, loff_t *offs,
1683                                     handle_t *handle)
1684 {
1685         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1686         struct buffer_head *bh        = NULL;
1687         loff_t              offset    = *offs;
1688         loff_t              new_size  = i_size_read(inode);
1689         unsigned long       block;
1690         int                 blocksize = 1 << inode->i_blkbits;
1691         int                 err = 0;
1692         int                 size;
1693         int                 boffs;
1694         int                 dirty_inode = 0;
1695         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1696         bool create, sparse, sync = false;
1697
1698         if (write_NUL) {
1699                 /*
1700                  * long symlink write does not count the NUL terminator in
1701                  * bufsize, we write it, and the inode's file size does not
1702                  * count the NUL terminator as well.
1703                  */
1704                 ((char *)buf)[bufsize] = '\0';
1705                 ++bufsize;
1706         }
1707
1708         /* sparse checking is racy, but sparse is very rare case, leave as is */
1709         sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
1710                   ((new_size - 1) >> inode->i_blkbits) + 1);
1711
1712         while (bufsize > 0) {
1713                 int credits = handle->h_buffer_credits;
1714                 unsigned long last_block = (new_size == 0) ? 0 :
1715                                            (new_size - 1) >> inode->i_blkbits;
1716
1717                 if (bh)
1718                         brelse(bh);
1719
1720                 block = offset >> inode->i_blkbits;
1721                 boffs = offset & (blocksize - 1);
1722                 size = min(blocksize - boffs, bufsize);
1723                 sync = (block > last_block || new_size == 0 || sparse);
1724
1725                 if (sync)
1726                         down(&ei->i_append_sem);
1727
1728                 bh = __ldiskfs_bread(handle, inode, block, 0);
1729
1730                 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
1731                         CWARN("%s: adding bh without locking off %llu (block %lu, "
1732                               "size %d, offs %llu)\n", inode->i_sb->s_id,
1733                               offset, block, bufsize, *offs);
1734
1735                 if (IS_ERR_OR_NULL(bh)) {
1736                         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1737                         int flags = LDISKFS_GET_BLOCKS_CREATE;
1738
1739                         /* while the file system is being mounted, avoid
1740                          * preallocation otherwise mount can take a long
1741                          * time as mballoc cache is cold.
1742                          * XXX: this is a workaround until we have a proper
1743                          *      fix in mballoc
1744                          * XXX: works with extent-based files only */
1745                         if (!osd->od_cl_seq)
1746                                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
1747                         bh = __ldiskfs_bread(handle, inode, block, flags);
1748                         create = true;
1749                 } else {
1750                         if (sync) {
1751                                 up(&ei->i_append_sem);
1752                                 sync = false;
1753                         }
1754                         create = false;
1755                 }
1756                 if (IS_ERR_OR_NULL(bh)) {
1757                         if (bh == NULL) {
1758                                 err = -EIO;
1759                         } else {
1760                                 err = PTR_ERR(bh);
1761                                 bh = NULL;
1762                         }
1763
1764                         CERROR("%s: error reading offset %llu (block %lu, "
1765                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
1766                                inode->i_sb->s_id, offset, block, bufsize, *offs,
1767                                credits, handle->h_buffer_credits, err);
1768                         break;
1769                 }
1770
1771                 err = ldiskfs_journal_get_write_access(handle, bh);
1772                 if (err) {
1773                         CERROR("journal_get_write_access() returned error %d\n",
1774                                err);
1775                         break;
1776                 }
1777                 LASSERTF(boffs + size <= bh->b_size,
1778                          "boffs %d size %d bh->b_size %lu\n",
1779                          boffs, size, (unsigned long)bh->b_size);
1780                 if (create) {
1781                         memset(bh->b_data, 0, bh->b_size);
1782                         if (sync) {
1783                                 up(&ei->i_append_sem);
1784                                 sync = false;
1785                         }
1786                 }
1787                 memcpy(bh->b_data + boffs, buf, size);
1788                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1789                 if (err)
1790                         break;
1791
1792                 if (offset + size > new_size)
1793                         new_size = offset + size;
1794                 offset += size;
1795                 bufsize -= size;
1796                 buf += size;
1797         }
1798         if (sync)
1799                 up(&ei->i_append_sem);
1800
1801         if (bh)
1802                 brelse(bh);
1803
1804         if (write_NUL)
1805                 --new_size;
1806         /* correct in-core and on-disk sizes */
1807         if (new_size > i_size_read(inode)) {
1808                 spin_lock(&inode->i_lock);
1809                 if (new_size > i_size_read(inode))
1810                         i_size_write(inode, new_size);
1811                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1812                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1813                         dirty_inode = 1;
1814                 }
1815                 spin_unlock(&inode->i_lock);
1816                 if (dirty_inode)
1817                         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1818         }
1819
1820         if (err == 0)
1821                 *offs = offset;
1822         return err;
1823 }
1824
1825 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1826                          const struct lu_buf *buf, loff_t *pos,
1827                          struct thandle *handle)
1828 {
1829         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1830         struct osd_thandle      *oh;
1831         ssize_t                 result;
1832         int                     is_link;
1833
1834         LASSERT(dt_object_exists(dt));
1835
1836         LASSERT(handle != NULL);
1837         LASSERT(inode != NULL);
1838         dquot_initialize(inode);
1839
1840         /* XXX: don't check: one declared chunk can be used many times */
1841         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1842
1843         oh = container_of(handle, struct osd_thandle, ot_super);
1844         LASSERT(oh->ot_handle->h_transaction != NULL);
1845         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1846
1847         /* Write small symlink to inode body as we need to maintain correct
1848          * on-disk symlinks for ldiskfs.
1849          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1850          * does not count it in.
1851          */
1852         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1853         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1854                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1855         else
1856                 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
1857                                                   is_link, pos, oh->ot_handle);
1858         if (result == 0)
1859                 result = buf->lb_len;
1860
1861         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1862
1863         return result;
1864 }
1865
1866 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1867                              __u64 start, __u64 end, struct thandle *th)
1868 {
1869         struct osd_thandle *oh;
1870         struct inode       *inode;
1871         int                 rc;
1872         ENTRY;
1873
1874         LASSERT(th);
1875         oh = container_of(th, struct osd_thandle, ot_super);
1876
1877         /*
1878          * we don't need to reserve credits for whole truncate
1879          * it's not possible as truncate may need to free too many
1880          * blocks and that won't fit a single transaction. instead
1881          * we reserve credits to change i_size and put inode onto
1882          * orphan list. if needed truncate will extend or restart
1883          * transaction
1884          */
1885         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1886                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1887
1888         inode = osd_dt_obj(dt)->oo_inode;
1889         LASSERT(inode);
1890
1891         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1892                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1893                                    NULL, OSD_QID_BLK);
1894
1895         if (rc == 0)
1896                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
1897
1898         RETURN(rc);
1899 }
1900
1901 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1902                      __u64 start, __u64 end, struct thandle *th)
1903 {
1904         struct osd_object *obj = osd_dt_obj(dt);
1905         struct osd_device *osd = osd_obj2dev(obj);
1906         struct inode *inode = obj->oo_inode;
1907         struct osd_access_lock *al;
1908         struct osd_thandle *oh;
1909         int rc = 0, found = 0;
1910         bool grow = false;
1911         ENTRY;
1912
1913         LASSERT(end == OBD_OBJECT_EOF);
1914         LASSERT(dt_object_exists(dt));
1915         LASSERT(osd_invariant(obj));
1916         LASSERT(inode != NULL);
1917         dquot_initialize(inode);
1918
1919         LASSERT(th);
1920         oh = container_of(th, struct osd_thandle, ot_super);
1921         LASSERT(oh->ot_handle->h_transaction != NULL);
1922
1923         /* we used to skip truncate to current size to
1924          * optimize truncates on OST. with DoM we can
1925          * get attr_set to set specific size (MDS_REINT)
1926          * and then get truncate RPC which essentially
1927          * would be skipped. this is bad.. so, disable
1928          * this optimization on MDS till the client stop
1929          * to sent MDS_REINT (LU-11033) -bzzz */
1930         if (osd->od_is_ost && i_size_read(inode) == start)
1931                 RETURN(0);
1932
1933         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1934
1935         spin_lock(&inode->i_lock);
1936         if (i_size_read(inode) < start)
1937                 grow = true;
1938         i_size_write(inode, start);
1939         spin_unlock(&inode->i_lock);
1940         ll_truncate_pagecache(inode, start);
1941
1942         /* optimize grow case */
1943         if (grow) {
1944                 osd_execute_truncate(obj);
1945                 GOTO(out, rc);
1946         }
1947
1948         inode_lock(inode);
1949         /* add to orphan list to ensure truncate completion
1950          * if this transaction succeed. ldiskfs_truncate()
1951          * will take the inode out of the list */
1952         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
1953         inode_unlock(inode);
1954         if (rc != 0)
1955                 GOTO(out, rc);
1956
1957         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
1958                 if (obj != al->tl_obj)
1959                         continue;
1960                 LASSERT(al->tl_shared == 0);
1961                 found = 1;
1962                 /* do actual truncate in osd_trans_stop() */
1963                 al->tl_truncate = 1;
1964                 break;
1965         }
1966         LASSERT(found);
1967
1968 out:
1969         RETURN(rc);
1970 }
1971
1972 static int fiemap_check_ranges(struct inode *inode,
1973                                u64 start, u64 len, u64 *new_len)
1974 {
1975         loff_t maxbytes;
1976
1977         *new_len = len;
1978
1979         if (len == 0)
1980                 return -EINVAL;
1981
1982         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1983                 maxbytes = inode->i_sb->s_maxbytes;
1984         else
1985                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1986
1987         if (start > maxbytes)
1988                 return -EFBIG;
1989
1990         /*
1991          * Shrink request scope to what the fs can actually handle.
1992          */
1993         if (len > maxbytes || (maxbytes - len) < start)
1994                 *new_len = maxbytes - start;
1995
1996         return 0;
1997 }
1998
1999 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2000 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
2001
2002 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2003                           struct fiemap *fm)
2004 {
2005         struct fiemap_extent_info fieinfo = {0, };
2006         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2007         u64 len;
2008         int rc;
2009         mm_segment_t cur_fs;
2010
2011         LASSERT(inode);
2012         if (inode->i_op->fiemap == NULL)
2013                 return -EOPNOTSUPP;
2014
2015         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2016                 return -EINVAL;
2017
2018         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2019         if (rc)
2020                 return rc;
2021
2022         fieinfo.fi_flags = fm->fm_flags;
2023         fieinfo.fi_extents_max = fm->fm_extent_count;
2024         fieinfo.fi_extents_start = fm->fm_extents;
2025
2026         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2027                 filemap_write_and_wait(inode->i_mapping);
2028
2029         /* Save previous value address limit */
2030         cur_fs = get_fs();
2031         /* Set the address limit of the kernel */
2032         set_fs(KERNEL_DS);
2033
2034         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2035         fm->fm_flags = fieinfo.fi_flags;
2036         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2037
2038         /* Restore the previous address limt */
2039         set_fs(cur_fs);
2040
2041         return rc;
2042 }
2043
2044 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2045                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2046 {
2047         struct osd_object *obj = osd_dt_obj(dt);
2048         int rc = 0;
2049         ENTRY;
2050
2051         switch (advice) {
2052         case LU_LADVISE_DONTNEED:
2053                 if (end)
2054                         invalidate_mapping_pages(obj->oo_inode->i_mapping,
2055                                                  start >> PAGE_SHIFT,
2056                                                  (end - 1) >> PAGE_SHIFT);
2057                 break;
2058         default:
2059                 rc = -ENOTSUPP;
2060                 break;
2061         }
2062
2063         RETURN(rc);
2064 }
2065
2066 /*
2067  * in some cases we may need declare methods for objects being created
2068  * e.g., when we create symlink
2069  */
2070 const struct dt_body_operations osd_body_ops_new = {
2071         .dbo_declare_write = osd_declare_write,
2072 };
2073
2074 const struct dt_body_operations osd_body_ops = {
2075         .dbo_read                       = osd_read,
2076         .dbo_declare_write              = osd_declare_write,
2077         .dbo_write                      = osd_write,
2078         .dbo_bufs_get                   = osd_bufs_get,
2079         .dbo_bufs_put                   = osd_bufs_put,
2080         .dbo_write_prep                 = osd_write_prep,
2081         .dbo_declare_write_commit       = osd_declare_write_commit,
2082         .dbo_write_commit               = osd_write_commit,
2083         .dbo_read_prep                  = osd_read_prep,
2084         .dbo_declare_punch              = osd_declare_punch,
2085         .dbo_punch                      = osd_punch,
2086         .dbo_fiemap_get                 = osd_fiemap_get,
2087         .dbo_ladvise                    = osd_ladvise,
2088 };
2089
2090 /**
2091  * Get a truncate lock
2092  *
2093  * In order to take multi-transaction truncate out of main transaction we let
2094  * the caller grab a lock on the object passed. the lock can be shared (for
2095  * writes) and exclusive (for truncate). It's not allowed to mix truncate
2096  * and write in the same transaction handle (do not confuse with big ldiskfs
2097  * transaction containing lots of handles).
2098  * The lock must be taken at declaration.
2099  *
2100  * \param obj           object to lock
2101  * \oh                  transaction
2102  * \shared              shared or exclusive
2103  *
2104  * \retval 0            lock is granted
2105  * \retval -NOMEM       no memory to allocate lock
2106  */
2107 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2108 {
2109         struct osd_access_lock *al, *tmp;
2110
2111         LASSERT(obj);
2112         LASSERT(oh);
2113
2114         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2115                 if (tmp->tl_obj != obj)
2116                         continue;
2117                 LASSERT(tmp->tl_shared == shared);
2118                 /* found same lock */
2119                 return 0;
2120         }
2121
2122         OBD_ALLOC_PTR(al);
2123         if (unlikely(al == NULL))
2124                 return -ENOMEM;
2125         al->tl_obj = obj;
2126         al->tl_truncate = false;
2127         if (shared)
2128                 down_read(&obj->oo_ext_idx_sem);
2129         else
2130                 down_write(&obj->oo_ext_idx_sem);
2131         al->tl_shared = shared;
2132
2133         list_add(&al->tl_list, &oh->ot_trunc_locks);
2134
2135         return 0;
2136 }
2137
2138 void osd_trunc_unlock_all(struct list_head *list)
2139 {
2140         struct osd_access_lock *al, *tmp;
2141         list_for_each_entry_safe(al, tmp, list, tl_list) {
2142                 if (al->tl_shared)
2143                         up_read(&al->tl_obj->oo_ext_idx_sem);
2144                 else
2145                         up_write(&al->tl_obj->oo_ext_idx_sem);
2146                 list_del(&al->tl_list);
2147                 OBD_FREE_PTR(al);
2148         }
2149 }
2150
2151 void osd_execute_truncate(struct osd_object *obj)
2152 {
2153         struct osd_device *d = osd_obj2dev(obj);
2154         struct inode *inode = obj->oo_inode;
2155         __u64 size;
2156
2157         /* simulate crash before (in the middle) of delayed truncate */
2158         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2159                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2160                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2161
2162                 mutex_lock(&sbi->s_orphan_lock);
2163                 list_del_init(&ei->i_orphan);
2164                 mutex_unlock(&sbi->s_orphan_lock);
2165                 return;
2166         }
2167
2168         inode_lock(inode);
2169         ldiskfs_truncate(inode);
2170         inode_unlock(inode);
2171
2172         /*
2173          * For a partial-page truncate, flush the page to disk immediately to
2174          * avoid data corruption during direct disk write.  b=17397
2175          */
2176         size = i_size_read(inode);
2177         if ((size & ~PAGE_MASK) == 0)
2178                 return;
2179         if (osd_use_page_cache(d)) {
2180                 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2181         } else {
2182                 /* Notice we use "wait" version to ensure I/O is complete */
2183                 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2184                 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2185                                          size >> PAGE_SHIFT);
2186         }
2187 }
2188
2189 void osd_process_truncates(struct list_head *list)
2190 {
2191         struct osd_access_lock *al;
2192
2193         LASSERT(journal_current_handle() == NULL);
2194
2195         list_for_each_entry(al, list, tl_list) {
2196                 if (al->tl_shared)
2197                         continue;
2198                 if (!al->tl_truncate)
2199                         continue;
2200                 osd_execute_truncate(al->tl_obj);
2201         }
2202 }