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(!PagePrivate2(page));
 671                         wait_on_page_writeback(page);
 672                 } else {
 673                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 674                 }
 675
 676                 return page;
 677         }
 678
 679         if (inode->i_mapping->nrpages) {
 680                 /* consult with pagecache, but do not create new pages */
 681                 /* this is normally used once */
 682                 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
 683                 if (page) {
 684                         wait_on_page_writeback(page);
 685                         return page;
 686                 }
 687         }
 688
 689         LASSERT(oti->oti_dio_pages);
 690         cur = oti->oti_dio_pages_used;
 691         page = oti->oti_dio_pages[cur];
 692
 693         if (unlikely(!page)) {
 694                 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
 695                 page = alloc_page(gfp_mask);
 696                 if (!page)
 697                         return NULL;
 698                 oti->oti_dio_pages[cur] = page;
 699                 SetPagePrivate2(page);
 700                 lock_page(page);
 701         }
 702
 703         ClearPageUptodate(page);
 704         page->index = offset >> PAGE_SHIFT;
 705         oti->oti_dio_pages_used++;
 706
 707         return page;
 708 }
 709
 710 /*
 711  * there are following "locks":
 712  * journal_start
 713  * i_mutex
 714  * page lock
 715  *
 716  * osd write path:
 717  *  - lock page(s)
 718  *  - journal_start
 719  *  - truncate_sem
 720  *
 721  * ext4 vmtruncate:
 722  *  - lock pages, unlock
 723  *  - journal_start
 724  *  - lock partial page
 725  *  - i_data_sem
 726  *
 727  */
 728
 729 /**
 730  * Unlock and release pages loaded by osd_bufs_get()
 731  *
 732  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 733  *
 734  * \param env           thread execution environment
 735  * \param dt            dt object undergoing IO (OSD object + methods)
 736  * \param lnb           array of pages undergoing IO
 737  * \param npages        number of pages in \a lnb
 738  *
 739  * \retval 0            always
 740  */
 741 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 742                         struct niobuf_local *lnb, int npages)
 743 {
 744         struct osd_thread_info *oti = osd_oti_get(env);
 745         struct pagevec pvec;
 746         int i;
 747
 748         ll_pagevec_init(&pvec, 0);
 749
 750         for (i = 0; i < npages; i++) {
 751                 struct page *page = lnb[i].lnb_page;
 752
 753                 if (page == NULL)
 754                         continue;
 755
 756                 /* if the page isn't cached, then reset uptodate
 757                  * to prevent reuse */
 758                 if (PagePrivate2(page)) {
 759                         oti->oti_dio_pages_used--;
 760                 } else {
 761                         if (lnb[i].lnb_locked)
 762                                 unlock_page(page);
 763                         if (pagevec_add(&pvec, page) == 0)
 764                                 pagevec_release(&pvec);
 765                 }
 766
 767                 lnb[i].lnb_page = NULL;
 768         }
 769
 770         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
 771
 772         /* Release any partial pagevec */
 773         pagevec_release(&pvec);
 774
 775         RETURN(0);
 776 }
 777
 778 /**
 779  * Load and lock pages undergoing IO
 780  *
 781  * Pages as described in the \a lnb array are fetched (from disk or cache)
 782  * and locked for IO by the caller.
 783  *
 784  * DLM locking protects us from write and truncate competing for same region,
 785  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 786  * It's possible the writeout on a such a page is in progress when we access
 787  * it. It's also possible that during this writeout we put new (partial) data
 788  * into the page, but won't be able to proceed in filter_commitrw_write().
 789  * Therefore, just wait for writeout completion as it should be rare enough.
 790  *
 791  * \param env           thread execution environment
 792  * \param dt            dt object undergoing IO (OSD object + methods)
 793  * \param pos           byte offset of IO start
 794  * \param len           number of bytes of IO
 795  * \param lnb           array of extents undergoing IO
 796  * \param rw            read or write operation, and other flags
 797  * \param capa          capabilities
 798  *
 799  * \retval pages        (zero or more) loaded successfully
 800  * \retval -ENOMEM      on memory/page allocation error
 801  */
 802 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 803                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 804                         int maxlnb, enum dt_bufs_type rw)
 805 {
 806         struct osd_thread_info *oti = osd_oti_get(env);
 807         struct osd_object *obj = osd_dt_obj(dt);
 808         struct osd_device *osd   = osd_obj2dev(obj);
 809         int npages, i, iosize, rc = 0;
 810         bool cache, write;
 811         loff_t fsize;
 812         gfp_t gfp_mask;
 813
 814         LASSERT(obj->oo_inode);
 815
 816         rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
 817         if (rc)
 818                 RETURN(rc);
 819
 820         write = rw & DT_BUFS_TYPE_WRITE;
 821
 822         fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
 823         iosize = fsize - lnb[0].lnb_file_offset;
 824         fsize = max(fsize, i_size_read(obj->oo_inode));
 825
 826         cache = rw & DT_BUFS_TYPE_READAHEAD;
 827         if (cache)
 828                 goto bypass_checks;
 829
 830         cache = osd_use_page_cache(osd);
 831         while (cache) {
 832                 if (write) {
 833                         if (!osd->od_writethrough_cache) {
 834                                 cache = false;
 835                                 break;
 836                         }
 837                         if (iosize > osd->od_writethrough_max_iosize) {
 838                                 cache = false;
 839                                 break;
 840                         }
 841                 } else {
 842                         if (!osd->od_read_cache) {
 843                                 cache = false;
 844                                 break;
 845                         }
 846                         if (iosize > osd->od_readcache_max_iosize) {
 847                                 cache = false;
 848                                 break;
 849                         }
 850                 }
 851                 /* don't use cache on large files */
 852                 if (osd->od_readcache_max_filesize &&
 853                     fsize > osd->od_readcache_max_filesize)
 854                         cache = false;
 855                 break;
 856         }
 857
 858 bypass_checks:
 859         if (!cache && unlikely(!oti->oti_dio_pages)) {
 860                 OBD_ALLOC_PTR_ARRAY(oti->oti_dio_pages, PTLRPC_MAX_BRW_PAGES);
 861                 if (!oti->oti_dio_pages)
 862                         return -ENOMEM;
 863         }
 864
 865         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 866         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 867                                              GFP_HIGHUSER;
 868         for (i = 0; i < npages; i++, lnb++) {
 869                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
 870                                              gfp_mask, cache);
 871                 if (lnb->lnb_page == NULL)
 872                         GOTO(cleanup, rc = -ENOMEM);
 873
 874                 lnb->lnb_locked = 1;
 875         }
 876
 877 #if 0
 878         /* XXX: this version doesn't invalidate cached pages, but use them */
 879         if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
 880                 /* do not allow data aliasing, invalidate pagecache */
 881                 /* XXX: can be quite expensive in mixed case */
 882                 invalidate_mapping_pages(obj->oo_inode->i_mapping,
 883                                 lnb[0].lnb_file_offset >> PAGE_SHIFT,
 884                                 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
 885         }
 886 #endif
 887
 888         RETURN(i);
 889
 890 cleanup:
 891         if (i > 0)
 892                 osd_bufs_put(env, dt, lnb - i, i);
 893         return rc;
 894 }
 895
 896 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 897                                        int pages, sector_t *blocks,
 898                                        int create)
 899 {
 900         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 901         int rc = 0, i = 0;
 902         struct page *fp = NULL;
 903         int clen = 0;
 904         pgoff_t max_page_index;
 905         handle_t *handle = NULL;
 906
 907         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
 908
 909         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 910                 inode->i_ino, pages, (*page)->index);
 911
 912         if (create) {
 913                 create = LDISKFS_GET_BLOCKS_CREATE;
 914                 handle = ldiskfs_journal_current_handle();
 915                 LASSERT(handle != NULL);
 916                 rc = osd_attach_jinode(inode);
 917                 if (rc)
 918                         return rc;
 919         }
 920         /* pages are sorted already. so, we just have to find
 921          * contig. space and process them properly */
 922         while (i < pages) {
 923                 long blen, total = 0;
 924                 struct ldiskfs_map_blocks map = { 0 };
 925
 926                 if (fp == NULL) { /* start new extent */
 927                         fp = *page++;
 928                         clen = 1;
 929                         if (++i != pages)
 930                                 continue;
 931                 } else if (fp->index + clen == (*page)->index) {
 932                         /* continue the extent */
 933                         page++;
 934                         clen++;
 935                         if (++i != pages)
 936                                 continue;
 937                 }
 938                 if (fp->index + clen >= max_page_index)
 939                         GOTO(cleanup, rc = -EFBIG);
 940                 /* process found extent */
 941                 map.m_lblk = fp->index * blocks_per_page;
 942                 map.m_len = blen = clen * blocks_per_page;
 943 cont_map:
 944                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 945                 if (rc >= 0) {
 946                         int c = 0;
 947                         for (; total < blen && c < map.m_len; c++, total++) {
 948                                 if (rc == 0) {
 949                                         *(blocks + total) = 0;
 950                                         total++;
 951                                         break;
 952                                 } else {
 953                                         *(blocks + total) = map.m_pblk + c;
 954                                         /* unmap any possible underlying
 955                                          * metadata from the block device
 956                                          * mapping.  bug 6998. */
 957                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
 958                                             create)
 959                                                 clean_bdev_aliases(
 960                                                         inode->i_sb->s_bdev,
 961                                                         map.m_pblk + c, 1);
 962                                 }
 963                         }
 964                         rc = 0;
 965                 }
 966                 if (rc == 0 && total < blen) {
 967                         map.m_lblk = fp->index * blocks_per_page + total;
 968                         map.m_len = blen - total;
 969                         goto cont_map;
 970                 }
 971                 if (rc != 0)
 972                         GOTO(cleanup, rc);
 973
 974                 /* look for next extent */
 975                 fp = NULL;
 976                 blocks += blocks_per_page * clen;
 977         }
 978 cleanup:
 979         return rc;
 980 }
 981
 982 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 983                           struct niobuf_local *lnb, int npages)
 984 {
 985         struct osd_thread_info *oti   = osd_oti_get(env);
 986         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 987         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 988         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 989         ktime_t start, end;
 990         s64 timediff;
 991         ssize_t isize;
 992         __s64  maxidx;
 993         int i, rc = 0;
 994
 995         LASSERT(inode);
 996
 997         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 998         if (unlikely(rc != 0))
 999                 RETURN(rc);
1000
1001         isize = i_size_read(inode);
1002         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1003
1004         start = ktime_get();
1005         for (i = 0; i < npages; i++) {
1006
1007                 /*
1008                  * till commit the content of the page is undefined
1009                  * we'll set it uptodate once bulk is done. otherwise
1010                  * subsequent reads can access non-stable data
1011                  */
1012                 ClearPageUptodate(lnb[i].lnb_page);
1013
1014                 if (lnb[i].lnb_len == PAGE_SIZE)
1015                         continue;
1016
1017                 if (maxidx >= lnb[i].lnb_page->index) {
1018                         osd_iobuf_add_page(iobuf, &lnb[i]);
1019                 } else {
1020                         long off;
1021                         char *p = kmap(lnb[i].lnb_page);
1022
1023                         off = lnb[i].lnb_page_offset;
1024                         if (off)
1025                                 memset(p, 0, off);
1026                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1027                               ~PAGE_MASK;
1028                         if (off)
1029                                 memset(p + off, 0, PAGE_SIZE - off);
1030                         kunmap(lnb[i].lnb_page);
1031                 }
1032         }
1033         end = ktime_get();
1034         timediff = ktime_us_delta(end, start);
1035         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1036
1037         if (iobuf->dr_npages) {
1038                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1039                                                  iobuf->dr_npages,
1040                                                  iobuf->dr_blocks, 0);
1041                 if (likely(rc == 0)) {
1042                         rc = osd_do_bio(osd, inode, iobuf);
1043                         /* do IO stats for preparation reads */
1044                         osd_fini_iobuf(osd, iobuf);
1045                 }
1046         }
1047         RETURN(rc);
1048 }
1049
1050 struct osd_fextent {
1051         sector_t        start;
1052         sector_t        end;
1053         unsigned int    mapped:1;
1054 };
1055
1056 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1057                          struct osd_fextent *cached_extent)
1058 {
1059         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1060         sector_t block = offset >> inode->i_blkbits;
1061         sector_t start;
1062         struct fiemap_extent_info fei = { 0 };
1063         struct fiemap_extent fe = { 0 };
1064         mm_segment_t saved_fs;
1065         int rc;
1066
1067         if (block >= cached_extent->start && block < cached_extent->end)
1068                 return cached_extent->mapped;
1069
1070         if (i_size_read(inode) == 0)
1071                 return 0;
1072
1073         /* Beyond EOF, must not be mapped */
1074         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1075                 return 0;
1076
1077         fei.fi_extents_max = 1;
1078         fei.fi_extents_start = &fe;
1079
1080         saved_fs = get_fs();
1081         set_fs(KERNEL_DS);
1082         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1083         set_fs(saved_fs);
1084         if (rc != 0)
1085                 return 0;
1086
1087         start = fe.fe_logical >> inode->i_blkbits;
1088
1089         if (start > block) {
1090                 cached_extent->start = block;
1091                 cached_extent->end = start;
1092                 cached_extent->mapped = 0;
1093         } else {
1094                 cached_extent->start = start;
1095                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1096                                       inode->i_blkbits;
1097                 cached_extent->mapped = 1;
1098         }
1099
1100         return cached_extent->mapped;
1101 }
1102
1103 static int osd_declare_write_commit(const struct lu_env *env,
1104                                     struct dt_object *dt,
1105                                     struct niobuf_local *lnb, int npages,
1106                                     struct thandle *handle)
1107 {
1108         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1109         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1110         struct osd_thandle      *oh;
1111         int                     extents = 1;
1112         int                     depth;
1113         int                     i;
1114         int                     newblocks;
1115         int                     rc = 0;
1116         int                     credits = 0;
1117         long long               quota_space = 0;
1118         struct osd_fextent      extent = { 0 };
1119         enum osd_quota_local_flags local_flags = 0;
1120         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1121         ENTRY;
1122
1123         LASSERT(handle != NULL);
1124         oh = container_of0(handle, struct osd_thandle, ot_super);
1125         LASSERT(oh->ot_handle == NULL);
1126
1127         newblocks = npages;
1128
1129         /* calculate number of extents (probably better to pass nb) */
1130         for (i = 0; i < npages; i++) {
1131                 if (i && lnb[i].lnb_file_offset !=
1132                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1133                         extents++;
1134
1135                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1136                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1137                 else
1138                         quota_space += PAGE_SIZE;
1139
1140                 /* ignore quota for the whole request if any page is from
1141                  * client cache or written by root.
1142                  *
1143                  * XXX once we drop the 1.8 client support, the checking
1144                  * for whether page is from cache can be simplified as:
1145                  * !(lnb[i].flags & OBD_BRW_SYNC)
1146                  *
1147                  * XXX we could handle this on per-lnb basis as done by
1148                  * grant. */
1149                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1150                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1151                     OBD_BRW_FROM_GRANT)
1152                         declare_flags |= OSD_QID_FORCE;
1153         }
1154
1155         /*
1156          * each extent can go into new leaf causing a split
1157          * 5 is max tree depth: inode + 4 index blocks
1158          * with blockmaps, depth is 3 at most
1159          */
1160         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1161                 /*
1162                  * many concurrent threads may grow tree by the time
1163                  * our transaction starts. so, consider 2 is a min depth
1164                  */
1165                 depth = ext_depth(inode);
1166                 depth = max(depth, 1) + 1;
1167                 newblocks += depth;
1168                 credits++; /* inode */
1169                 credits += depth * 2 * extents;
1170         } else {
1171                 depth = 3;
1172                 newblocks += depth;
1173                 credits++; /* inode */
1174                 credits += depth * extents;
1175         }
1176
1177         /* quota space for metadata blocks */
1178         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1179
1180         /* quota space should be reported in 1K blocks */
1181         quota_space = toqb(quota_space);
1182
1183         /* each new block can go in different group (bitmap + gd) */
1184
1185         /* we can't dirty more bitmap blocks than exist */
1186         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1187                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1188         else
1189                 credits += newblocks;
1190
1191         /* we can't dirty more gd blocks than exist */
1192         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1193                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1194         else
1195                 credits += newblocks;
1196
1197         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1198
1199         /* make sure the over quota flags were not set */
1200         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1201
1202         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1203                                    i_projid_read(inode), quota_space, oh,
1204                                    osd_dt_obj(dt), &local_flags, declare_flags);
1205
1206         /* we need only to store the overquota flags in the first lnb for
1207          * now, once we support multiple objects BRW, this code needs be
1208          * revised. */
1209         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1210                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1211         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1212                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1213         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1214                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1215
1216         if (rc == 0)
1217                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1218
1219         RETURN(rc);
1220 }
1221
1222 /* Check if a block is allocated or not */
1223 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1224                             struct niobuf_local *lnb, int npages,
1225                             struct thandle *thandle)
1226 {
1227         struct osd_thread_info *oti = osd_oti_get(env);
1228         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1229         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1230         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1231         loff_t isize;
1232         int rc = 0, i;
1233
1234         LASSERT(inode);
1235
1236         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1237         if (unlikely(rc != 0))
1238                 RETURN(rc);
1239
1240         isize = i_size_read(inode);
1241         dquot_initialize(inode);
1242
1243         for (i = 0; i < npages; i++) {
1244                 if (lnb[i].lnb_rc == -ENOSPC &&
1245                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1246                         /* Allow the write to proceed if overwriting an
1247                          * existing block */
1248                         lnb[i].lnb_rc = 0;
1249                 }
1250
1251                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1252                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1253                                lnb[i].lnb_rc);
1254                         LASSERT(lnb[i].lnb_page);
1255                         generic_error_remove_page(inode->i_mapping,
1256                                                   lnb[i].lnb_page);
1257                         continue;
1258                 }
1259
1260                 LASSERT(PageLocked(lnb[i].lnb_page));
1261                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1262
1263                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1264                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1265
1266                 /*
1267                  * Since write and truncate are serialized by oo_sem, even
1268                  * partial-page truncate should not leave dirty pages in the
1269                  * page cache.
1270                  */
1271                 LASSERT(!PageDirty(lnb[i].lnb_page));
1272
1273                 SetPageUptodate(lnb[i].lnb_page);
1274
1275                 osd_iobuf_add_page(iobuf, &lnb[i]);
1276         }
1277
1278         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1279
1280         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1281                 rc = -ENOSPC;
1282         } else if (iobuf->dr_npages > 0) {
1283                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1284                                                  iobuf->dr_npages,
1285                                                  iobuf->dr_blocks, 1);
1286         } else {
1287                 /* no pages to write, no transno is needed */
1288                 thandle->th_local = 1;
1289         }
1290
1291         if (likely(rc == 0)) {
1292                 spin_lock(&inode->i_lock);
1293                 if (isize > i_size_read(inode)) {
1294                         i_size_write(inode, isize);
1295                         LDISKFS_I(inode)->i_disksize = isize;
1296                         spin_unlock(&inode->i_lock);
1297                         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1298                 } else {
1299                         spin_unlock(&inode->i_lock);
1300                 }
1301
1302                 rc = osd_do_bio(osd, inode, iobuf);
1303                 /* we don't do stats here as in read path because
1304                  * write is async: we'll do this in osd_put_bufs() */
1305         } else {
1306                 osd_fini_iobuf(osd, iobuf);
1307         }
1308
1309         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1310
1311         if (unlikely(rc != 0)) {
1312                 /* if write fails, we should drop pages from the cache */
1313                 for (i = 0; i < npages; i++) {
1314                         if (lnb[i].lnb_page == NULL)
1315                                 continue;
1316                         if (!PagePrivate2(lnb[i].lnb_page)) {
1317                                 LASSERT(PageLocked(lnb[i].lnb_page));
1318                                 generic_error_remove_page(inode->i_mapping,
1319                                                           lnb[i].lnb_page);
1320                         }
1321                 }
1322         }
1323
1324         RETURN(rc);
1325 }
1326
1327 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1328                          struct niobuf_local *lnb, int npages)
1329 {
1330         struct osd_thread_info *oti = osd_oti_get(env);
1331         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1332         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1333         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1334         int rc = 0, i, cache_hits = 0, cache_misses = 0;
1335         ktime_t start, end;
1336         s64 timediff;
1337         loff_t isize;
1338
1339         LASSERT(inode);
1340
1341         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1342         if (unlikely(rc != 0))
1343                 RETURN(rc);
1344
1345         isize = i_size_read(inode);
1346
1347         start = ktime_get();
1348         for (i = 0; i < npages; i++) {
1349
1350                 if (isize <= lnb[i].lnb_file_offset)
1351                         /* If there's no more data, abort early.
1352                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1353                         break;
1354
1355                 /* instead of looking if we go beyong isize, send complete
1356                  * pages all the time
1357                  */
1358                 lnb[i].lnb_rc = lnb[i].lnb_len;
1359
1360                 /* Bypass disk read if fail_loc is set properly */
1361                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1362                         SetPageUptodate(lnb[i].lnb_page);
1363
1364                 if (PageUptodate(lnb[i].lnb_page)) {
1365                         cache_hits++;
1366                         unlock_page(lnb[i].lnb_page);
1367                 } else {
1368                         cache_misses++;
1369                         osd_iobuf_add_page(iobuf, &lnb[i]);
1370                 }
1371                 /* no need to unlock in osd_bufs_put(), the sooner page is
1372                  * unlocked, the earlier another client can access it.
1373                  * notice real unlock_page() can be called few lines
1374                  * below after osd_do_bio(). lnb is a per-thread, so it's
1375                  * fine to have PG_locked and lnb_locked inconsistent here */
1376                 lnb[i].lnb_locked = 0;
1377         }
1378         end = ktime_get();
1379         timediff = ktime_us_delta(end, start);
1380         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1381
1382         if (cache_hits != 0)
1383                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1384                                     cache_hits);
1385         if (cache_misses != 0)
1386                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1387                                     cache_misses);
1388         if (cache_hits + cache_misses != 0)
1389                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1390                                     cache_hits + cache_misses);
1391
1392         if (iobuf->dr_npages) {
1393                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1394                                                  iobuf->dr_npages,
1395                                                  iobuf->dr_blocks, 0);
1396                 rc = osd_do_bio(osd, inode, iobuf);
1397
1398                 /* IO stats will be done in osd_bufs_put() */
1399
1400                 /* early release to let others read data during the bulk */
1401                 for (i = 0; i < iobuf->dr_npages; i++) {
1402                         LASSERT(PageLocked(iobuf->dr_pages[i]));
1403                         if (!PagePrivate2(iobuf->dr_pages[i]))
1404                                 unlock_page(iobuf->dr_pages[i]);
1405                 }
1406         }
1407
1408         RETURN(rc);
1409 }
1410
1411 /*
1412  * XXX: Another layering violation for now.
1413  *
1414  * We don't want to use ->f_op->read methods, because generic file write
1415  *
1416  *         - serializes on ->i_sem, and
1417  *
1418  *         - does a lot of extra work like balance_dirty_pages(),
1419  *
1420  * which doesn't work for globally shared files like /last_rcvd.
1421  */
1422 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1423 {
1424         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1425
1426         memcpy(buffer, (char *)ei->i_data, buflen);
1427
1428         return  buflen;
1429 }
1430
1431 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1432 {
1433         struct buffer_head *bh;
1434         unsigned long block;
1435         int osize;
1436         int blocksize;
1437         int csize;
1438         int boffs;
1439
1440         /* prevent reading after eof */
1441         spin_lock(&inode->i_lock);
1442         if (i_size_read(inode) < *offs + size) {
1443                 loff_t diff = i_size_read(inode) - *offs;
1444                 spin_unlock(&inode->i_lock);
1445                 if (diff < 0) {
1446                         CDEBUG(D_OTHER,
1447                                "size %llu is too short to read @%llu\n",
1448                                i_size_read(inode), *offs);
1449                         return -EBADR;
1450                 } else if (diff == 0) {
1451                         return 0;
1452                 } else {
1453                         size = diff;
1454                 }
1455         } else {
1456                 spin_unlock(&inode->i_lock);
1457         }
1458
1459         blocksize = 1 << inode->i_blkbits;
1460         osize = size;
1461         while (size > 0) {
1462                 block = *offs >> inode->i_blkbits;
1463                 boffs = *offs & (blocksize - 1);
1464                 csize = min(blocksize - boffs, size);
1465                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1466                 if (IS_ERR(bh)) {
1467                         CERROR("%s: can't read %u@%llu on ino %lu: "
1468                                "rc = %ld\n", osd_ino2name(inode),
1469                                csize, *offs, inode->i_ino,
1470                                PTR_ERR(bh));
1471                         return PTR_ERR(bh);
1472                 }
1473
1474                 if (bh != NULL) {
1475                         memcpy(buf, bh->b_data + boffs, csize);
1476                         brelse(bh);
1477                 } else {
1478                         memset(buf, 0, csize);
1479                 }
1480
1481                 *offs += csize;
1482                 buf += csize;
1483                 size -= csize;
1484         }
1485         return osize;
1486 }
1487
1488 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1489                         struct lu_buf *buf, loff_t *pos)
1490 {
1491         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1492         int rc;
1493
1494         /* Read small symlink from inode body as we need to maintain correct
1495          * on-disk symlinks for ldiskfs.
1496          */
1497         if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1498                 loff_t size = i_size_read(inode);
1499
1500                 if (buf->lb_len < size)
1501                         return -EOVERFLOW;
1502
1503                 if (size < sizeof(LDISKFS_I(inode)->i_data))
1504                         rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1505                 else
1506                         rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1507         } else {
1508                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1509         }
1510
1511         return rc;
1512 }
1513
1514 static inline int osd_extents_enabled(struct super_block *sb,
1515                                       struct inode *inode)
1516 {
1517         if (inode != NULL) {
1518                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1519                         return 1;
1520         } else if (ldiskfs_has_feature_extents(sb)) {
1521                 return 1;
1522         }
1523         return 0;
1524 }
1525
1526 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1527                            const loff_t size, const loff_t pos,
1528                            const int blocks)
1529 {
1530         int credits, bits, bs, i;
1531
1532         bits = sb->s_blocksize_bits;
1533         bs = 1 << bits;
1534
1535         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1536          * we do not expect blockmaps on the large files,
1537          * so let's shrink it to 2 levels (4GB files) */
1538
1539         /* this is default reservation: 2 levels */
1540         credits = (blocks + 2) * 3;
1541
1542         /* actual offset is unknown, hard to optimize */
1543         if (pos == -1)
1544                 return credits;
1545
1546         /* now check for few specific cases to optimize */
1547         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1548                 /* no indirects */
1549                 credits = blocks;
1550                 /* allocate if not allocated */
1551                 if (inode == NULL) {
1552                         credits += blocks * 2;
1553                         return credits;
1554                 }
1555                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1556                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1557                         if (LDISKFS_I(inode)->i_data[i] == 0)
1558                                 credits += 2;
1559                 }
1560         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1561                 /* single indirect */
1562                 credits = blocks * 3;
1563                 if (inode == NULL ||
1564                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1565                         credits += 3;
1566                 else
1567                         /* The indirect block may be modified. */
1568                         credits += 1;
1569         }
1570
1571         return credits;
1572 }
1573
1574 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1575                                  const struct lu_buf *buf, loff_t _pos,
1576                                  struct thandle *handle)
1577 {
1578         struct osd_object  *obj  = osd_dt_obj(dt);
1579         struct inode       *inode = obj->oo_inode;
1580         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1581         struct osd_thandle *oh;
1582         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1583         int                 bits, bs;
1584         int                 depth, size;
1585         loff_t              pos;
1586         ENTRY;
1587
1588         LASSERT(buf != NULL);
1589         LASSERT(handle != NULL);
1590
1591         oh = container_of0(handle, struct osd_thandle, ot_super);
1592         LASSERT(oh->ot_handle == NULL);
1593
1594         size = buf->lb_len;
1595         bits = sb->s_blocksize_bits;
1596         bs = 1 << bits;
1597
1598         if (_pos == -1) {
1599                 /* if this is an append, then we
1600                  * should expect cross-block record */
1601                 pos = 0;
1602         } else {
1603                 pos = _pos;
1604         }
1605
1606         /* blocks to modify */
1607         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1608         LASSERT(blocks > 0);
1609
1610         if (inode != NULL && _pos != -1) {
1611                 /* object size in blocks */
1612                 est = (i_size_read(inode) + bs - 1) >> bits;
1613                 allocated = inode->i_blocks >> (bits - 9);
1614                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1615                         /* looks like an overwrite, no need to modify tree */
1616                         credits = blocks;
1617                         /* no need to modify i_size */
1618                         goto out;
1619                 }
1620         }
1621
1622         if (osd_extents_enabled(sb, inode)) {
1623                 /*
1624                  * many concurrent threads may grow tree by the time
1625                  * our transaction starts. so, consider 2 is a min depth
1626                  * for every level we may need to allocate a new block
1627                  * and take some entries from the old one. so, 3 blocks
1628                  * to allocate (bitmap, gd, itself) + old block - 4 per
1629                  * level.
1630                  */
1631                 depth = inode != NULL ? ext_depth(inode) : 0;
1632                 depth = max(depth, 1) + 1;
1633                 credits = depth;
1634                 /* if not append, then split may need to modify
1635                  * existing blocks moving entries into the new ones */
1636                 if (_pos != -1)
1637                         credits += depth;
1638                 /* blocks to store data: bitmap,gd,itself */
1639                 credits += blocks * 3;
1640         } else {
1641                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1642         }
1643         /* if inode is created as part of the transaction,
1644          * then it's counted already by the creation method */
1645         if (inode != NULL)
1646                 credits++;
1647
1648 out:
1649
1650         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1651
1652         /* dt_declare_write() is usually called for system objects, such
1653          * as llog or last_rcvd files. We needn't enforce quota on those
1654          * objects, so always set the lqi_space as 0. */
1655         if (inode != NULL)
1656                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1657                                            i_gid_read(inode),
1658                                            i_projid_read(inode), 0,
1659                                            oh, obj, NULL, OSD_QID_BLK);
1660
1661         if (rc == 0)
1662                 rc = osd_trunc_lock(obj, oh, true);
1663
1664         RETURN(rc);
1665 }
1666
1667 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1668 {
1669         /* LU-2634: clear the extent format for fast symlink */
1670         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1671
1672         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1673         spin_lock(&inode->i_lock);
1674         LDISKFS_I(inode)->i_disksize = buflen;
1675         i_size_write(inode, buflen);
1676         spin_unlock(&inode->i_lock);
1677         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1678
1679         return 0;
1680 }
1681
1682 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
1683                                     int bufsize, int write_NUL, loff_t *offs,
1684                                     handle_t *handle)
1685 {
1686         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1687         struct buffer_head *bh        = NULL;
1688         loff_t              offset    = *offs;
1689         loff_t              new_size  = i_size_read(inode);
1690         unsigned long       block;
1691         int                 blocksize = 1 << inode->i_blkbits;
1692         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1693         int                 err = 0;
1694         int                 size;
1695         int                 boffs;
1696         int                 dirty_inode = 0;
1697         bool create, sparse, sync = false;
1698
1699         if (write_NUL) {
1700                 /*
1701                  * long symlink write does not count the NUL terminator in
1702                  * bufsize, we write it, and the inode's file size does not
1703                  * count the NUL terminator as well.
1704                  */
1705                 ((char *)buf)[bufsize] = '\0';
1706                 ++bufsize;
1707         }
1708
1709         dirty_inode = test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA, &ei->i_flags);
1710
1711         /* sparse checking is racy, but sparse is very rare case, leave as is */
1712         sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
1713                   ((new_size - 1) >> inode->i_blkbits) + 1);
1714
1715         while (bufsize > 0) {
1716                 int credits = handle->h_buffer_credits;
1717                 unsigned long last_block = (new_size == 0) ? 0 :
1718                                            (new_size - 1) >> inode->i_blkbits;
1719
1720                 if (bh)
1721                         brelse(bh);
1722
1723                 block = offset >> inode->i_blkbits;
1724                 boffs = offset & (blocksize - 1);
1725                 size = min(blocksize - boffs, bufsize);
1726                 sync = (block > last_block || new_size == 0 || sparse);
1727
1728                 if (sync)
1729                         down(&ei->i_append_sem);
1730
1731                 bh = __ldiskfs_bread(handle, inode, block, 0);
1732
1733                 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
1734                         CWARN("%s: adding bh without locking off %llu (block %lu, "
1735                               "size %d, offs %llu)\n", inode->i_sb->s_id,
1736                               offset, block, bufsize, *offs);
1737
1738                 if (IS_ERR_OR_NULL(bh)) {
1739                         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1740                         int flags = LDISKFS_GET_BLOCKS_CREATE;
1741
1742                         /* while the file system is being mounted, avoid
1743                          * preallocation otherwise mount can take a long
1744                          * time as mballoc cache is cold.
1745                          * XXX: this is a workaround until we have a proper
1746                          *      fix in mballoc
1747                          * XXX: works with extent-based files only */
1748                         if (!osd->od_cl_seq)
1749                                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
1750                         bh = __ldiskfs_bread(handle, inode, block, flags);
1751                         create = true;
1752                 } else {
1753                         if (sync) {
1754                                 up(&ei->i_append_sem);
1755                                 sync = false;
1756                         }
1757                         create = false;
1758                 }
1759                 if (IS_ERR_OR_NULL(bh)) {
1760                         if (bh == NULL) {
1761                                 err = -EIO;
1762                         } else {
1763                                 err = PTR_ERR(bh);
1764                                 bh = NULL;
1765                         }
1766
1767                         CERROR("%s: error reading offset %llu (block %lu, "
1768                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
1769                                inode->i_sb->s_id, offset, block, bufsize, *offs,
1770                                credits, handle->h_buffer_credits, err);
1771                         break;
1772                 }
1773
1774                 err = ldiskfs_journal_get_write_access(handle, bh);
1775                 if (err) {
1776                         CERROR("journal_get_write_access() returned error %d\n",
1777                                err);
1778                         break;
1779                 }
1780                 LASSERTF(boffs + size <= bh->b_size,
1781                          "boffs %d size %d bh->b_size %lu\n",
1782                          boffs, size, (unsigned long)bh->b_size);
1783                 if (create) {
1784                         memset(bh->b_data, 0, bh->b_size);
1785                         if (sync) {
1786                                 up(&ei->i_append_sem);
1787                                 sync = false;
1788                         }
1789                 }
1790                 memcpy(bh->b_data + boffs, buf, size);
1791                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1792                 if (err)
1793                         break;
1794
1795                 if (offset + size > new_size)
1796                         new_size = offset + size;
1797                 offset += size;
1798                 bufsize -= size;
1799                 buf += size;
1800         }
1801         if (sync)
1802                 up(&ei->i_append_sem);
1803
1804         if (bh)
1805                 brelse(bh);
1806
1807         if (write_NUL)
1808                 --new_size;
1809         /* correct in-core and on-disk sizes */
1810         if (new_size > i_size_read(inode)) {
1811                 spin_lock(&inode->i_lock);
1812                 if (new_size > i_size_read(inode))
1813                         i_size_write(inode, new_size);
1814                 if (i_size_read(inode) > ei->i_disksize) {
1815                         ei->i_disksize = i_size_read(inode);
1816                         dirty_inode = 1;
1817                 }
1818                 spin_unlock(&inode->i_lock);
1819         }
1820         if (dirty_inode)
1821                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1822
1823         if (err == 0)
1824                 *offs = offset;
1825         return err;
1826 }
1827
1828 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1829                          const struct lu_buf *buf, loff_t *pos,
1830                          struct thandle *handle)
1831 {
1832         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1833         struct osd_thandle      *oh;
1834         ssize_t                 result;
1835         int                     is_link;
1836
1837         LASSERT(dt_object_exists(dt));
1838
1839         LASSERT(handle != NULL);
1840         LASSERT(inode != NULL);
1841         dquot_initialize(inode);
1842
1843         /* XXX: don't check: one declared chunk can be used many times */
1844         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1845
1846         oh = container_of(handle, struct osd_thandle, ot_super);
1847         LASSERT(oh->ot_handle->h_transaction != NULL);
1848         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1849
1850         /* Write small symlink to inode body as we need to maintain correct
1851          * on-disk symlinks for ldiskfs.
1852          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1853          * does not count it in.
1854          */
1855         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1856         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1857                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1858         else
1859                 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
1860                                                   is_link, pos, oh->ot_handle);
1861         if (result == 0)
1862                 result = buf->lb_len;
1863
1864         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1865
1866         return result;
1867 }
1868
1869 static int osd_declare_fallocate(const struct lu_env *env,
1870                                  struct dt_object *dt, struct thandle *th)
1871 {
1872         struct osd_thandle *oh;
1873         struct inode *inode;
1874         int rc;
1875         ENTRY;
1876
1877         LASSERT(th);
1878         oh = container_of(th, struct osd_thandle, ot_super);
1879
1880         osd_trans_declare_op(env, oh, OSD_OT_PREALLOC,
1881                              osd_dto_credits_noquota[DTO_WRITE_BLOCK]);
1882         inode = osd_dt_obj(dt)->oo_inode;
1883         LASSERT(inode);
1884
1885         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1886                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1887                                    NULL, OSD_QID_BLK);
1888         RETURN(rc);
1889 }
1890
1891 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
1892                          __u64 start, __u64 end, int mode, struct thandle *th)
1893 {
1894         struct osd_object *obj = osd_dt_obj(dt);
1895         struct inode *inode = obj->oo_inode;
1896         int rc = 0;
1897         struct osd_thread_info *info = osd_oti_get(env);
1898         struct dentry *dentry = &info->oti_obj_dentry;
1899         struct file *file = &info->oti_file;
1900
1901         ENTRY;
1902         /*
1903          * Only mode == 0 (which is standard prealloc) is supported now.
1904          * Rest of mode options is not supported yet.
1905          */
1906         if (mode & ~FALLOC_FL_KEEP_SIZE)
1907                 RETURN(-EOPNOTSUPP);
1908
1909         LASSERT(dt_object_exists(dt));
1910         LASSERT(osd_invariant(obj));
1911         LASSERT(inode != NULL);
1912         dquot_initialize(inode);
1913
1914         LASSERT(th);
1915
1916         osd_trans_exec_op(env, th, OSD_OT_PREALLOC);
1917
1918         /*
1919          * Because f_op->fallocate() does not have an inode arg
1920          */
1921         dentry->d_inode = inode;
1922         dentry->d_sb = inode->i_sb;
1923         file->f_path.dentry = dentry;
1924         file->f_mapping = inode->i_mapping;
1925         file->f_op = inode->i_fop;
1926         file->f_inode = inode;
1927         rc = file->f_op->fallocate(file, mode, start, end - start);
1928
1929         RETURN(rc);
1930 }
1931
1932 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1933                              __u64 start, __u64 end, struct thandle *th)
1934 {
1935         struct osd_thandle *oh;
1936         struct inode       *inode;
1937         int                 rc;
1938         ENTRY;
1939
1940         LASSERT(th);
1941         oh = container_of(th, struct osd_thandle, ot_super);
1942
1943         /*
1944          * we don't need to reserve credits for whole truncate
1945          * it's not possible as truncate may need to free too many
1946          * blocks and that won't fit a single transaction. instead
1947          * we reserve credits to change i_size and put inode onto
1948          * orphan list. if needed truncate will extend or restart
1949          * transaction
1950          */
1951         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1952                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1953
1954         inode = osd_dt_obj(dt)->oo_inode;
1955         LASSERT(inode);
1956
1957         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1958                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1959                                    NULL, OSD_QID_BLK);
1960
1961         if (rc == 0)
1962                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
1963
1964         RETURN(rc);
1965 }
1966
1967 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1968                      __u64 start, __u64 end, struct thandle *th)
1969 {
1970         struct osd_object *obj = osd_dt_obj(dt);
1971         struct osd_device *osd = osd_obj2dev(obj);
1972         struct inode *inode = obj->oo_inode;
1973         struct osd_access_lock *al;
1974         struct osd_thandle *oh;
1975         int rc = 0, found = 0;
1976         bool grow = false;
1977         ENTRY;
1978
1979         LASSERT(dt_object_exists(dt));
1980         LASSERT(osd_invariant(obj));
1981         LASSERT(inode != NULL);
1982         dquot_initialize(inode);
1983
1984         LASSERT(th);
1985         oh = container_of(th, struct osd_thandle, ot_super);
1986         LASSERT(oh->ot_handle->h_transaction != NULL);
1987
1988         /* we used to skip truncate to current size to
1989          * optimize truncates on OST. with DoM we can
1990          * get attr_set to set specific size (MDS_REINT)
1991          * and then get truncate RPC which essentially
1992          * would be skipped. this is bad.. so, disable
1993          * this optimization on MDS till the client stop
1994          * to sent MDS_REINT (LU-11033) -bzzz */
1995         if (osd->od_is_ost && i_size_read(inode) == start)
1996                 RETURN(0);
1997
1998         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1999
2000         spin_lock(&inode->i_lock);
2001         if (i_size_read(inode) < start)
2002                 grow = true;
2003         i_size_write(inode, start);
2004         spin_unlock(&inode->i_lock);
2005         ll_truncate_pagecache(inode, start);
2006
2007         /* optimize grow case */
2008         if (grow) {
2009                 osd_execute_truncate(obj);
2010                 GOTO(out, rc);
2011         }
2012
2013         inode_lock(inode);
2014         /* add to orphan list to ensure truncate completion
2015          * if this transaction succeed. ldiskfs_truncate()
2016          * will take the inode out of the list */
2017         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2018         inode_unlock(inode);
2019         if (rc != 0)
2020                 GOTO(out, rc);
2021
2022         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2023                 if (obj != al->tl_obj)
2024                         continue;
2025                 LASSERT(al->tl_shared == 0);
2026                 found = 1;
2027                 /* do actual truncate in osd_trans_stop() */
2028                 al->tl_truncate = 1;
2029                 break;
2030         }
2031         LASSERT(found);
2032
2033 out:
2034         RETURN(rc);
2035 }
2036
2037 static int fiemap_check_ranges(struct inode *inode,
2038                                u64 start, u64 len, u64 *new_len)
2039 {
2040         loff_t maxbytes;
2041
2042         *new_len = len;
2043
2044         if (len == 0)
2045                 return -EINVAL;
2046
2047         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2048                 maxbytes = inode->i_sb->s_maxbytes;
2049         else
2050                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2051
2052         if (start > maxbytes)
2053                 return -EFBIG;
2054
2055         /*
2056          * Shrink request scope to what the fs can actually handle.
2057          */
2058         if (len > maxbytes || (maxbytes - len) < start)
2059                 *new_len = maxbytes - start;
2060
2061         return 0;
2062 }
2063
2064 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2065 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
2066
2067 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2068                           struct fiemap *fm)
2069 {
2070         struct fiemap_extent_info fieinfo = {0, };
2071         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2072         u64 len;
2073         int rc;
2074         mm_segment_t cur_fs;
2075
2076         LASSERT(inode);
2077         if (inode->i_op->fiemap == NULL)
2078                 return -EOPNOTSUPP;
2079
2080         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2081                 return -EINVAL;
2082
2083         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2084         if (rc)
2085                 return rc;
2086
2087         fieinfo.fi_flags = fm->fm_flags;
2088         fieinfo.fi_extents_max = fm->fm_extent_count;
2089         fieinfo.fi_extents_start = fm->fm_extents;
2090
2091         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2092                 filemap_write_and_wait(inode->i_mapping);
2093
2094         /* Save previous value address limit */
2095         cur_fs = get_fs();
2096         /* Set the address limit of the kernel */
2097         set_fs(KERNEL_DS);
2098
2099         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2100         fm->fm_flags = fieinfo.fi_flags;
2101         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2102
2103         /* Restore the previous address limt */
2104         set_fs(cur_fs);
2105
2106         return rc;
2107 }
2108
2109 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2110                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2111 {
2112         struct osd_object *obj = osd_dt_obj(dt);
2113         int rc = 0;
2114         ENTRY;
2115
2116         switch (advice) {
2117         case LU_LADVISE_DONTNEED:
2118                 if (end)
2119                         invalidate_mapping_pages(obj->oo_inode->i_mapping,
2120                                                  start >> PAGE_SHIFT,
2121                                                  (end - 1) >> PAGE_SHIFT);
2122                 break;
2123         default:
2124                 rc = -ENOTSUPP;
2125                 break;
2126         }
2127
2128         RETURN(rc);
2129 }
2130
2131 /*
2132  * in some cases we may need declare methods for objects being created
2133  * e.g., when we create symlink
2134  */
2135 const struct dt_body_operations osd_body_ops_new = {
2136         .dbo_declare_write = osd_declare_write,
2137 };
2138
2139 const struct dt_body_operations osd_body_ops = {
2140         .dbo_read                       = osd_read,
2141         .dbo_declare_write              = osd_declare_write,
2142         .dbo_write                      = osd_write,
2143         .dbo_bufs_get                   = osd_bufs_get,
2144         .dbo_bufs_put                   = osd_bufs_put,
2145         .dbo_write_prep                 = osd_write_prep,
2146         .dbo_declare_write_commit       = osd_declare_write_commit,
2147         .dbo_write_commit               = osd_write_commit,
2148         .dbo_read_prep                  = osd_read_prep,
2149         .dbo_declare_punch              = osd_declare_punch,
2150         .dbo_punch                      = osd_punch,
2151         .dbo_fiemap_get                 = osd_fiemap_get,
2152         .dbo_ladvise                    = osd_ladvise,
2153         .dbo_declare_fallocate          = osd_declare_fallocate,
2154         .dbo_fallocate                  = osd_fallocate,
2155 };
2156
2157 /**
2158  * Get a truncate lock
2159  *
2160  * In order to take multi-transaction truncate out of main transaction we let
2161  * the caller grab a lock on the object passed. the lock can be shared (for
2162  * writes) and exclusive (for truncate). It's not allowed to mix truncate
2163  * and write in the same transaction handle (do not confuse with big ldiskfs
2164  * transaction containing lots of handles).
2165  * The lock must be taken at declaration.
2166  *
2167  * \param obj           object to lock
2168  * \oh                  transaction
2169  * \shared              shared or exclusive
2170  *
2171  * \retval 0            lock is granted
2172  * \retval -NOMEM       no memory to allocate lock
2173  */
2174 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2175 {
2176         struct osd_access_lock *al, *tmp;
2177
2178         LASSERT(obj);
2179         LASSERT(oh);
2180
2181         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2182                 if (tmp->tl_obj != obj)
2183                         continue;
2184                 LASSERT(tmp->tl_shared == shared);
2185                 /* found same lock */
2186                 return 0;
2187         }
2188
2189         OBD_ALLOC_PTR(al);
2190         if (unlikely(al == NULL))
2191                 return -ENOMEM;
2192         al->tl_obj = obj;
2193         al->tl_truncate = false;
2194         if (shared)
2195                 down_read(&obj->oo_ext_idx_sem);
2196         else
2197                 down_write(&obj->oo_ext_idx_sem);
2198         al->tl_shared = shared;
2199         lu_object_get(&obj->oo_dt.do_lu);
2200
2201         list_add(&al->tl_list, &oh->ot_trunc_locks);
2202
2203         return 0;
2204 }
2205
2206 void osd_trunc_unlock_all(const struct lu_env *env, struct list_head *list)
2207 {
2208         struct osd_access_lock *al, *tmp;
2209         list_for_each_entry_safe(al, tmp, list, tl_list) {
2210                 if (al->tl_shared)
2211                         up_read(&al->tl_obj->oo_ext_idx_sem);
2212                 else
2213                         up_write(&al->tl_obj->oo_ext_idx_sem);
2214                 osd_object_put(env, al->tl_obj);
2215                 list_del(&al->tl_list);
2216                 OBD_FREE_PTR(al);
2217         }
2218 }
2219
2220 void osd_execute_truncate(struct osd_object *obj)
2221 {
2222         struct osd_device *d = osd_obj2dev(obj);
2223         struct inode *inode = obj->oo_inode;
2224         __u64 size;
2225
2226         /* simulate crash before (in the middle) of delayed truncate */
2227         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2228                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2229                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2230
2231                 mutex_lock(&sbi->s_orphan_lock);
2232                 list_del_init(&ei->i_orphan);
2233                 mutex_unlock(&sbi->s_orphan_lock);
2234                 return;
2235         }
2236
2237         inode_lock(inode);
2238         ldiskfs_truncate(inode);
2239         inode_unlock(inode);
2240
2241         /*
2242          * For a partial-page truncate, flush the page to disk immediately to
2243          * avoid data corruption during direct disk write.  b=17397
2244          */
2245         size = i_size_read(inode);
2246         if ((size & ~PAGE_MASK) == 0)
2247                 return;
2248         if (osd_use_page_cache(d)) {
2249                 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2250         } else {
2251                 /* Notice we use "wait" version to ensure I/O is complete */
2252                 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2253                 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2254                                          size >> PAGE_SHIFT);
2255         }
2256 }
2257
2258 void osd_process_truncates(struct list_head *list)
2259 {
2260         struct osd_access_lock *al;
2261
2262         LASSERT(journal_current_handle() == NULL);
2263
2264         list_for_each_entry(al, list, tl_list) {
2265                 if (al->tl_shared)
2266                         continue;
2267                 if (!al->tl_truncate)
2268                         continue;
2269                 osd_execute_truncate(al->tl_obj);
2270         }
2271 }