lustre/osd-ldiskfs/osd_io.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/osd/osd_io.c
  33  *
  34  * body operations
  35  *
  36  * Author: Nikita Danilov <nikita@clusterfs.com>
  37  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  38  *
  39  */
  40
  41 /* prerequisite for linux/xattr.h */
  42 #include <linux/types.h>
  43 /* prerequisite for linux/xattr.h */
  44 #include <linux/fs.h>
  45 #include <linux/mm.h>
  46 #include <linux/pagevec.h>
  47
  48 /*
  49  * struct OBD_{ALLOC,FREE}*()
  50  * OBD_FAIL_CHECK
  51  */
  52 #include <obd_support.h>
  53
  54 #include "osd_internal.h"
  55
  56 /* ext_depth() */
  57 #include <ldiskfs/ldiskfs_extents.h>
  58
  59 static inline bool osd_use_page_cache(struct osd_device *d)
  60 {
  61         /* do not use pagecache if write and read caching are disabled */
  62         if (d->od_writethrough_cache + d->od_read_cache == 0)
  63                 return false;
  64         /* use pagecache by default */
  65         return true;
  66 }
  67
  68 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  69                             int rw, int line, int pages)
  70 {
  71         int blocks, i;
  72
  73         LASSERTF(iobuf->dr_elapsed_valid == 0,
  74                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  75                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  76                  iobuf->dr_init_at);
  77         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  78
  79         init_waitqueue_head(&iobuf->dr_wait);
  80         atomic_set(&iobuf->dr_numreqs, 0);
  81         iobuf->dr_npages = 0;
  82         iobuf->dr_error = 0;
  83         iobuf->dr_dev = d;
  84         iobuf->dr_frags = 0;
  85         iobuf->dr_elapsed = ktime_set(0, 0);
  86         /* must be counted before, so assert */
  87         iobuf->dr_rw = rw;
  88         iobuf->dr_init_at = line;
  89
  90         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  91         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  92                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  93                         pages * sizeof(iobuf->dr_pages[0]));
  94                 return 0;
  95         }
  96
  97         /* start with 1MB for 4K blocks */
  98         i = 256;
  99         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
 100                 i <<= 1;
 101
 102         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
 103                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
 104         pages = i;
 105         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
 106         iobuf->dr_max_pages = 0;
 107         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
 108                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 109
 110         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 111         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 112         if (unlikely(iobuf->dr_blocks == NULL))
 113                 return -ENOMEM;
 114
 115         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 116         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 117         if (unlikely(iobuf->dr_pages == NULL))
 118                 return -ENOMEM;
 119
 120         lu_buf_realloc(&iobuf->dr_lnb_buf,
 121                        pages * sizeof(iobuf->dr_lnbs[0]));
 122         iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf;
 123         if (unlikely(iobuf->dr_lnbs == NULL))
 124                 return -ENOMEM;
 125
 126         iobuf->dr_max_pages = pages;
 127
 128         return 0;
 129 }
 130 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 131         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 132
 133 static void osd_iobuf_add_page(struct osd_iobuf *iobuf,
 134                                struct niobuf_local *lnb)
 135 {
 136         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 137         iobuf->dr_pages[iobuf->dr_npages] = lnb->lnb_page;
 138         iobuf->dr_lnbs[iobuf->dr_npages] = lnb;
 139         iobuf->dr_npages++;
 140 }
 141
 142 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 143 {
 144         int rw = iobuf->dr_rw;
 145
 146         if (iobuf->dr_elapsed_valid) {
 147                 iobuf->dr_elapsed_valid = 0;
 148                 LASSERT(iobuf->dr_dev == d);
 149                 LASSERT(iobuf->dr_frags > 0);
 150                 lprocfs_oh_tally(&d->od_brw_stats.
 151                                  hist[BRW_R_DIO_FRAGS+rw],
 152                                  iobuf->dr_frags);
 153                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 154                                       ktime_to_ms(iobuf->dr_elapsed));
 155         }
 156 }
 157
 158 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 159 static void dio_complete_routine(struct bio *bio)
 160 {
 161 # ifdef HAVE_BI_STATUS
 162         int error = bio->bi_status;
 163 # else
 164         int error = bio->bi_error;
 165 # endif
 166 #else
 167 static void dio_complete_routine(struct bio *bio, int error)
 168 {
 169 #endif
 170         struct osd_iobuf *iobuf = bio->bi_private;
 171         int iter;
 172         struct bio_vec *bvl;
 173
 174         /* CAVEAT EMPTOR: possibly in IRQ context
 175          * DO NOT record procfs stats here!!! */
 176
 177         if (unlikely(iobuf == NULL)) {
 178                 CERROR("***** bio->bi_private is NULL!  This should never "
 179                        "happen.  Normally, I would crash here, but instead I "
 180                        "will dump the bio contents to the console.  Please "
 181                        "report this to <https://jira.whamcloud.com/> , along "
 182                        "with any interesting messages leading up to this point "
 183                        "(like SCSI errors, perhaps).  Because bi_private is "
 184                        "NULL, I can't wake up the thread that initiated this "
 185                        "IO - you will probably have to reboot this node.\n");
 186                 CERROR("bi_next: %p, bi_flags: %lx, "
 187 #ifdef HAVE_BI_RW
 188                        "bi_rw: %lu,"
 189 #else
 190                        "bi_opf: %u,"
 191 #endif
 192                        "bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p,"
 193                        "bi_cnt: %d, bi_private: %p\n", bio->bi_next,
 194                         (unsigned long)bio->bi_flags,
 195 #ifdef HAVE_BI_RW
 196                         bio->bi_rw,
 197 #else
 198                         bio->bi_opf,
 199 #endif
 200                         bio->bi_vcnt, bio_idx(bio),
 201                         bio_sectors(bio) << 9, bio->bi_end_io,
 202 #ifdef HAVE_BI_CNT
 203                         atomic_read(&bio->bi_cnt),
 204 #else
 205                         atomic_read(&bio->__bi_cnt),
 206 #endif
 207                         bio->bi_private);
 208                 return;
 209         }
 210
 211         /* the check is outside of the cycle for performance reason -bzzz */
 212         if (!bio_data_dir(bio)) {
 213                 bio_for_each_segment_all(bvl, bio, iter) {
 214                         if (likely(error == 0))
 215                                 SetPageUptodate(bvl_to_page(bvl));
 216                         LASSERT(PageLocked(bvl_to_page(bvl)));
 217                 }
 218                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 219         } else {
 220                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 221         }
 222
 223         /* any real error is good enough -bzzz */
 224         if (error != 0 && iobuf->dr_error == 0)
 225                 iobuf->dr_error = error;
 226
 227         /*
 228          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 229          * it's possible that service thread will see dr_numreqs
 230          * is zero, but dr_elapsed is not set yet, leading to lost
 231          * data in this processing and an assertion in a subsequent
 232          * call to OSD.
 233          */
 234         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 235                 ktime_t now = ktime_get();
 236
 237                 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
 238                 iobuf->dr_elapsed_valid = 1;
 239         }
 240         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 241                 wake_up(&iobuf->dr_wait);
 242
 243         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 244          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 245          * mempool when serious on-disk fragmentation was encountered,
 246          * deadlocking the OST.  The bios are now released as soon as complete
 247          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 248         bio_put(bio);
 249 }
 250
 251 static void record_start_io(struct osd_iobuf *iobuf, int size)
 252 {
 253         struct osd_device    *osd = iobuf->dr_dev;
 254         struct obd_histogram *h = osd->od_brw_stats.hist;
 255
 256         iobuf->dr_frags++;
 257         atomic_inc(&iobuf->dr_numreqs);
 258
 259         if (iobuf->dr_rw == 0) {
 260                 atomic_inc(&osd->od_r_in_flight);
 261                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 262                                  atomic_read(&osd->od_r_in_flight));
 263                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 264         } else if (iobuf->dr_rw == 1) {
 265                 atomic_inc(&osd->od_w_in_flight);
 266                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 267                                  atomic_read(&osd->od_w_in_flight));
 268                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 269         } else {
 270                 LBUG();
 271         }
 272 }
 273
 274 static void osd_submit_bio(int rw, struct bio *bio)
 275 {
 276         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 277 #ifdef HAVE_SUBMIT_BIO_2ARGS
 278         if (rw == 0)
 279                 submit_bio(READ, bio);
 280         else
 281                 submit_bio(WRITE, bio);
 282 #else
 283         bio->bi_opf |= rw;
 284         submit_bio(bio);
 285 #endif
 286 }
 287
 288 static int can_be_merged(struct bio *bio, sector_t sector)
 289 {
 290         if (bio == NULL)
 291                 return 0;
 292
 293         return bio_end_sector(bio) == sector ? 1 : 0;
 294 }
 295
 296 /*
 297  * This function will change the data written, thus it should only be
 298  * used when checking data integrity feature
 299  */
 300 static void bio_integrity_fault_inject(struct bio *bio)
 301 {
 302         struct bio_vec *bvec;
 303         int i;
 304         void *kaddr;
 305         char *addr;
 306
 307         bio_for_each_segment_all(bvec, bio, i) {
 308                 struct page *page = bvec->bv_page;
 309
 310                 kaddr = kmap(page);
 311                 addr = kaddr;
 312                 *addr = ~(*addr);
 313                 kunmap(page);
 314                 break;
 315         }
 316 }
 317
 318 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
 319                            unsigned int sectors, int tuple_size)
 320 {
 321         __u16 *expected_guard;
 322         __u16 *bio_guard;
 323         int i;
 324
 325         expected_guard = expected_guard_buf;
 326         for (i = 0; i < sectors; i++) {
 327                 bio_guard = (__u16 *)bio_prot_buf;
 328                 if (*bio_guard != *expected_guard) {
 329                         CERROR("unexpected guard tags on sector %d "
 330                                "expected guard %u, bio guard "
 331                                "%u, sectors %u, tuple size %d\n",
 332                                i, *expected_guard, *bio_guard, sectors,
 333                                tuple_size);
 334                         return -EIO;
 335                 }
 336                 expected_guard++;
 337                 bio_prot_buf += tuple_size;
 338         }
 339         return 0;
 340 }
 341
 342 static int osd_bio_integrity_compare(struct bio *bio, struct osd_iobuf *iobuf,
 343                                      int index)
 344 {
 345         struct blk_integrity *bi = bdev_get_integrity(bio->bi_bdev);
 346         struct bio_integrity_payload *bip = bio->bi_integrity;
 347         struct niobuf_local *lnb;
 348         unsigned short sector_size = blk_integrity_interval(bi);
 349         void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
 350                 bip->bip_vec->bv_offset;
 351         struct bio_vec *bv;
 352         sector_t sector = bio_start_sector(bio);
 353         unsigned int i, sectors, total;
 354         __u16 *expected_guard;
 355         int rc;
 356
 357         total = 0;
 358         bio_for_each_segment_all(bv, bio, i) {
 359                 lnb = iobuf->dr_lnbs[index];
 360                 expected_guard = lnb->lnb_guards;
 361                 sectors = bv->bv_len / sector_size;
 362                 if (lnb->lnb_guard_rpc) {
 363                         rc = bio_dif_compare(expected_guard, bio_prot_buf,
 364                                              sectors, bi->tuple_size);
 365                         if (rc)
 366                                 return rc;
 367                 }
 368
 369                 sector += sectors;
 370                 bio_prot_buf += sectors * bi->tuple_size;
 371                 total += sectors * bi->tuple_size;
 372                 LASSERT(total <= bip_size(bio->bi_integrity));
 373                 index++;
 374         }
 375         return 0;
 376 }
 377
 378 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
 379                                     struct osd_iobuf *iobuf,
 380                                     int start_page_idx, bool fault_inject,
 381                                     bool integrity_enabled)
 382 {
 383         int rc;
 384 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 385         integrity_gen_fn *generate_fn = NULL;
 386         integrity_vrfy_fn *verify_fn = NULL;
 387 #endif
 388
 389         ENTRY;
 390
 391         if (!integrity_enabled)
 392                 RETURN(0);
 393
 394 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 395         rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
 396         if (rc)
 397                 RETURN(rc);
 398
 399         rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
 400 #else
 401         rc = bio_integrity_prep(bio);
 402 #endif
 403         if (rc)
 404                 RETURN(rc);
 405
 406         /* Verify and inject fault only when writing */
 407         if (iobuf->dr_rw == 1) {
 408                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
 409                         rc = osd_bio_integrity_compare(bio, iobuf,
 410                                                        start_page_idx);
 411                         if (rc)
 412                                 RETURN(rc);
 413                 }
 414
 415                 if (unlikely(fault_inject))
 416                         bio_integrity_fault_inject(bio);
 417         }
 418
 419         RETURN(0);
 420 }
 421
 422 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 423 #  ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 424 static void dio_integrity_complete_routine(struct bio *bio)
 425 {
 426 #  else
 427 static void dio_integrity_complete_routine(struct bio *bio, int error)
 428 {
 429 #  endif
 430         struct osd_bio_private *bio_private = bio->bi_private;
 431
 432         bio->bi_private = bio_private->obp_iobuf;
 433 #  ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 434         dio_complete_routine(bio);
 435 #  else
 436         dio_complete_routine(bio, error);
 437 #  endif
 438
 439         OBD_FREE_PTR(bio_private);
 440 }
 441 #endif
 442
 443 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
 444                         bool integrity_enabled, int start_page_idx,
 445                         struct osd_bio_private **pprivate)
 446 {
 447 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 448         struct osd_bio_private *bio_private;
 449
 450         ENTRY;
 451
 452         *pprivate = NULL;
 453         if (integrity_enabled) {
 454                 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
 455                 if (bio_private == NULL)
 456                         RETURN(-ENOMEM);
 457                 bio->bi_end_io = dio_integrity_complete_routine;
 458                 bio->bi_private = bio_private;
 459                 bio_private->obp_start_page_idx = start_page_idx;
 460                 bio_private->obp_iobuf = iobuf;
 461                 *pprivate = bio_private;
 462         } else {
 463                 bio->bi_end_io = dio_complete_routine;
 464                 bio->bi_private = iobuf;
 465         }
 466         RETURN(0);
 467 #else
 468         ENTRY;
 469
 470         bio->bi_end_io = dio_complete_routine;
 471         bio->bi_private = iobuf;
 472         RETURN(0);
 473 #endif
 474 }
 475
 476 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 477                       struct osd_iobuf *iobuf)
 478 {
 479         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 480         struct page **pages = iobuf->dr_pages;
 481         int npages = iobuf->dr_npages;
 482         sector_t *blocks = iobuf->dr_blocks;
 483         int total_blocks = npages * blocks_per_page;
 484         struct super_block *sb = inode->i_sb;
 485         int sector_bits = sb->s_blocksize_bits - 9;
 486         unsigned int blocksize = sb->s_blocksize;
 487         struct block_device *bdev = sb->s_bdev;
 488         struct osd_bio_private *bio_private = NULL;
 489         struct bio *bio = NULL;
 490         int bio_start_page_idx;
 491         struct page *page;
 492         unsigned int page_offset;
 493         sector_t sector;
 494         int nblocks;
 495         int block_idx;
 496         int page_idx;
 497         int i;
 498         int rc = 0;
 499         bool fault_inject;
 500         bool integrity_enabled;
 501         DECLARE_PLUG(plug);
 502         ENTRY;
 503
 504         fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
 505         LASSERT(iobuf->dr_npages == npages);
 506
 507         integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
 508
 509         osd_brw_stats_update(osd, iobuf);
 510         iobuf->dr_start_time = ktime_get();
 511
 512         blk_start_plug(&plug);
 513         for (page_idx = 0, block_idx = 0;
 514              page_idx < npages;
 515              page_idx++, block_idx += blocks_per_page) {
 516
 517                 page = pages[page_idx];
 518                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 519
 520                 for (i = 0, page_offset = 0;
 521                      i < blocks_per_page;
 522                      i += nblocks, page_offset += blocksize * nblocks) {
 523
 524                         nblocks = 1;
 525
 526                         if (blocks[block_idx + i] == 0) {  /* hole */
 527                                 LASSERTF(iobuf->dr_rw == 0,
 528                                          "page_idx %u, block_idx %u, i %u\n",
 529                                          page_idx, block_idx, i);
 530                                 memset(kmap(page) + page_offset, 0, blocksize);
 531                                 kunmap(page);
 532                                 continue;
 533                         }
 534
 535                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 536
 537                         /* Additional contiguous file blocks? */
 538                         while (i + nblocks < blocks_per_page &&
 539                                (sector + (nblocks << sector_bits)) ==
 540                                ((sector_t)blocks[block_idx + i + nblocks] <<
 541                                 sector_bits))
 542                                 nblocks++;
 543
 544                         if (bio != NULL &&
 545                             can_be_merged(bio, sector) &&
 546                             bio_add_page(bio, page,
 547                                          blocksize * nblocks, page_offset) != 0)
 548                                 continue;       /* added this frag OK */
 549
 550                         if (bio != NULL) {
 551                                 struct request_queue *q = bio_get_queue(bio);
 552                                 unsigned int bi_size = bio_sectors(bio) << 9;
 553
 554                                 /* Dang! I have to fragment this I/O */
 555                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 556                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 557                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 558                                        bio_sectors(bio),
 559                                        queue_max_sectors(q),
 560                                        bio_phys_segments(q, bio),
 561                                        queue_max_phys_segments(q),
 562                                        0, queue_max_hw_segments(q));
 563                                 rc = osd_bio_integrity_handle(osd, bio,
 564                                         iobuf, bio_start_page_idx,
 565                                         fault_inject, integrity_enabled);
 566                                 if (rc) {
 567                                         bio_put(bio);
 568                                         goto out;
 569                                 }
 570
 571                                 record_start_io(iobuf, bi_size);
 572                                 osd_submit_bio(iobuf->dr_rw, bio);
 573                         }
 574
 575                         bio_start_page_idx = page_idx;
 576                         /* allocate new bio */
 577                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 578                                                       (npages - page_idx) *
 579                                                       blocks_per_page));
 580                         if (bio == NULL) {
 581                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 582                                        (npages - page_idx), blocks_per_page,
 583                                        (npages - page_idx) * blocks_per_page);
 584                                 rc = -ENOMEM;
 585                                 goto out;
 586                         }
 587
 588                         bio_set_dev(bio, bdev);
 589                         bio_set_sector(bio, sector);
 590 #ifdef HAVE_BI_RW
 591                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 592 #else
 593                         bio->bi_opf = (iobuf->dr_rw == 0) ? READ : WRITE;
 594 #endif
 595                         rc = osd_bio_init(bio, iobuf, integrity_enabled,
 596                                           bio_start_page_idx, &bio_private);
 597                         if (rc) {
 598                                 bio_put(bio);
 599                                 goto out;
 600                         }
 601
 602                         rc = bio_add_page(bio, page,
 603                                           blocksize * nblocks, page_offset);
 604                         LASSERT(rc != 0);
 605                 }
 606         }
 607
 608         if (bio != NULL) {
 609                 rc = osd_bio_integrity_handle(osd, bio, iobuf,
 610                                               bio_start_page_idx,
 611                                               fault_inject,
 612                                               integrity_enabled);
 613                 if (rc) {
 614                         bio_put(bio);
 615                         goto out;
 616                 }
 617
 618                 record_start_io(iobuf, bio_sectors(bio) << 9);
 619                 osd_submit_bio(iobuf->dr_rw, bio);
 620                 rc = 0;
 621         }
 622
 623 out:
 624         blk_finish_plug(&plug);
 625
 626         /* in order to achieve better IO throughput, we don't wait for writes
 627          * completion here. instead we proceed with transaction commit in
 628          * parallel and wait for IO completion once transaction is stopped
 629          * see osd_trans_stop() for more details -bzzz */
 630         if (iobuf->dr_rw == 0 || fault_inject) {
 631                 wait_event(iobuf->dr_wait,
 632                            atomic_read(&iobuf->dr_numreqs) == 0);
 633                 osd_fini_iobuf(osd, iobuf);
 634         }
 635
 636         if (rc == 0) {
 637                 rc = iobuf->dr_error;
 638         } else {
 639                 if (bio_private)
 640                         OBD_FREE_PTR(bio_private);
 641         }
 642
 643         RETURN(rc);
 644 }
 645
 646 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 647                                    struct niobuf_local *lnb)
 648 {
 649         ENTRY;
 650
 651         *nrpages = 0;
 652
 653         while (len > 0) {
 654                 int poff = offset & (PAGE_SIZE - 1);
 655                 int plen = PAGE_SIZE - poff;
 656
 657                 if (plen > len)
 658                         plen = len;
 659                 lnb->lnb_file_offset = offset;
 660                 lnb->lnb_page_offset = poff;
 661                 lnb->lnb_len = plen;
 662                 /* lnb->lnb_flags = rnb->rnb_flags; */
 663                 lnb->lnb_flags = 0;
 664                 lnb->lnb_page = NULL;
 665                 lnb->lnb_rc = 0;
 666                 lnb->lnb_guard_rpc = 0;
 667                 lnb->lnb_guard_disk = 0;
 668
 669                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 670                          (long long) len);
 671                 offset += plen;
 672                 len -= plen;
 673                 lnb++;
 674                 (*nrpages)++;
 675         }
 676
 677         RETURN(0);
 678 }
 679
 680 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
 681                                  loff_t offset, gfp_t gfp_mask)
 682 {
 683         struct osd_thread_info *oti = osd_oti_get(env);
 684         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 685         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 686         struct page *page;
 687         int cur = oti->oti_dio_pages_used;
 688
 689         LASSERT(inode);
 690
 691         if (osd_use_page_cache(d)) {
 692                 page = find_or_create_page(inode->i_mapping,
 693                                            offset >> PAGE_SHIFT,
 694                                            gfp_mask);
 695
 696                 if (likely(page))
 697                         LASSERT(!test_bit(PG_private_2, &page->flags));
 698                 else
 699                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 700         } else {
 701
 702                 LASSERT(oti->oti_dio_pages);
 703
 704                 if (unlikely(!oti->oti_dio_pages[cur])) {
 705                         LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
 706                         page = alloc_page(gfp_mask);
 707                         if (!page)
 708                                 return NULL;
 709                         oti->oti_dio_pages[cur] = page;
 710                 }
 711
 712                 page = oti->oti_dio_pages[cur];
 713                 LASSERT(!test_bit(PG_private_2, &page->flags));
 714                 set_bit(PG_private_2, &page->flags);
 715                 oti->oti_dio_pages_used++;
 716
 717                 LASSERT(!PageLocked(page));
 718                 lock_page(page);
 719
 720                 LASSERT(!page->mapping);
 721                 LASSERT(!PageWriteback(page));
 722                 ClearPageUptodate(page);
 723
 724                 page->index = offset >> PAGE_SHIFT;
 725         }
 726
 727         return page;
 728 }
 729
 730 /*
 731  * there are following "locks":
 732  * journal_start
 733  * i_mutex
 734  * page lock
 735  *
 736  * osd write path:
 737  *  - lock page(s)
 738  *  - journal_start
 739  *  - truncate_sem
 740  *
 741  * ext4 vmtruncate:
 742  *  - lock pages, unlock
 743  *  - journal_start
 744  *  - lock partial page
 745  *  - i_data_sem
 746  *
 747  */
 748
 749 /**
 750  * Unlock and release pages loaded by osd_bufs_get()
 751  *
 752  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 753  *
 754  * \param env           thread execution environment
 755  * \param dt            dt object undergoing IO (OSD object + methods)
 756  * \param lnb           array of pages undergoing IO
 757  * \param npages        number of pages in \a lnb
 758  *
 759  * \retval 0            always
 760  */
 761 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 762                         struct niobuf_local *lnb, int npages)
 763 {
 764         struct osd_thread_info *oti = osd_oti_get(env);
 765         struct pagevec pvec;
 766         int i;
 767
 768 #ifdef HAVE_PAGEVEC_INIT_ONE_PARAM
 769         pagevec_init(&pvec);
 770 #else
 771         pagevec_init(&pvec, 0);
 772 #endif
 773
 774         for (i = 0; i < npages; i++) {
 775                 struct page *page = lnb[i].lnb_page;
 776
 777                 if (page == NULL)
 778                         continue;
 779                 LASSERT(PageLocked(page));
 780
 781                 /* if the page isn't cached, then reset uptodate
 782                  * to prevent reuse */
 783                 if (test_bit(PG_private_2, &page->flags)) {
 784                         clear_bit(PG_private_2, &page->flags);
 785                         ClearPageUptodate(page);
 786                         unlock_page(page);
 787                         oti->oti_dio_pages_used--;
 788                 } else {
 789                         unlock_page(page);
 790                         if (pagevec_add(&pvec, page) == 0)
 791                                 pagevec_release(&pvec);
 792                 }
 793                 dt_object_put(env, dt);
 794
 795                 lnb[i].lnb_page = NULL;
 796         }
 797
 798         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
 799
 800         /* Release any partial pagevec */
 801         pagevec_release(&pvec);
 802
 803         RETURN(0);
 804 }
 805
 806 /**
 807  * Load and lock pages undergoing IO
 808  *
 809  * Pages as described in the \a lnb array are fetched (from disk or cache)
 810  * and locked for IO by the caller.
 811  *
 812  * DLM locking protects us from write and truncate competing for same region,
 813  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 814  * It's possible the writeout on a such a page is in progress when we access
 815  * it. It's also possible that during this writeout we put new (partial) data
 816  * into the page, but won't be able to proceed in filter_commitrw_write().
 817  * Therefore, just wait for writeout completion as it should be rare enough.
 818  *
 819  * \param env           thread execution environment
 820  * \param dt            dt object undergoing IO (OSD object + methods)
 821  * \param pos           byte offset of IO start
 822  * \param len           number of bytes of IO
 823  * \param lnb           array of extents undergoing IO
 824  * \param rw            read or write operation, and other flags
 825  * \param capa          capabilities
 826  *
 827  * \retval pages        (zero or more) loaded successfully
 828  * \retval -ENOMEM      on memory/page allocation error
 829  */
 830 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 831                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 832                         enum dt_bufs_type rw)
 833 {
 834         struct osd_thread_info *oti = osd_oti_get(env);
 835         struct osd_object *obj = osd_dt_obj(dt);
 836         int npages, i, rc = 0;
 837         gfp_t gfp_mask;
 838
 839         LASSERT(obj->oo_inode);
 840
 841         if (!osd_use_page_cache(osd_obj2dev(obj))) {
 842                 if (unlikely(!oti->oti_dio_pages)) {
 843                         OBD_ALLOC(oti->oti_dio_pages,
 844                                   sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
 845                         if (!oti->oti_dio_pages)
 846                                 return -ENOMEM;
 847                 }
 848         }
 849
 850         osd_map_remote_to_local(pos, len, &npages, lnb);
 851
 852         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 853         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 854                                              GFP_HIGHUSER;
 855         for (i = 0; i < npages; i++, lnb++) {
 856                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
 857                                              gfp_mask);
 858                 if (lnb->lnb_page == NULL)
 859                         GOTO(cleanup, rc = -ENOMEM);
 860
 861                 wait_on_page_writeback(lnb->lnb_page);
 862                 BUG_ON(PageWriteback(lnb->lnb_page));
 863
 864                 lu_object_get(&dt->do_lu);
 865         }
 866
 867         RETURN(i);
 868
 869 cleanup:
 870         if (i > 0)
 871                 osd_bufs_put(env, dt, lnb - i, i);
 872         return rc;
 873 }
 874
 875 #ifndef HAVE_LDISKFS_MAP_BLOCKS
 876
 877 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 878 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 879 #endif
 880
 881 struct bpointers {
 882         sector_t *blocks;
 883         unsigned long start;
 884         int num;
 885         int init_num;
 886         int create;
 887 };
 888
 889 static long ldiskfs_ext_find_goal(struct inode *inode,
 890                                   struct ldiskfs_ext_path *path,
 891                                   unsigned long block, int *aflags)
 892 {
 893         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 894         unsigned long bg_start;
 895         unsigned long colour;
 896         int depth;
 897
 898         if (path) {
 899                 struct ldiskfs_extent *ex;
 900                 depth = path->p_depth;
 901
 902                 /* try to predict block placement */
 903                 if ((ex = path[depth].p_ext))
 904                         return ldiskfs_ext_pblock(ex) +
 905                                 (block - le32_to_cpu(ex->ee_block));
 906
 907                 /* it looks index is empty
 908                  * try to find starting from index itself */
 909                 if (path[depth].p_bh)
 910                         return path[depth].p_bh->b_blocknr;
 911         }
 912
 913         /* OK. use inode's group */
 914         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 915                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 916         colour = (current->pid % 16) *
 917                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 918         return bg_start + colour + block;
 919 }
 920
 921 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 922                                 struct ldiskfs_ext_path *path,
 923                                 unsigned long block, unsigned long *count,
 924                                 int *err)
 925 {
 926         struct ldiskfs_allocation_request ar;
 927         unsigned long pblock;
 928         int aflags;
 929
 930         /* find neighbour allocated blocks */
 931         ar.lleft = block;
 932         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 933         if (*err)
 934                 return 0;
 935         ar.lright = block;
 936         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 937         if (*err)
 938                 return 0;
 939
 940         /* allocate new block */
 941         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 942         ar.inode = inode;
 943         ar.logical = block;
 944         ar.len = *count;
 945         ar.flags = LDISKFS_MB_HINT_DATA;
 946         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 947         *count = ar.len;
 948         return pblock;
 949 }
 950
 951 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 952                                      struct ldiskfs_ext_path *path,
 953                                      struct ldiskfs_ext_cache *cex,
 954 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 955                                      struct ldiskfs_extent *ex,
 956 #endif
 957                                      void *cbdata)
 958 {
 959         struct bpointers *bp = cbdata;
 960         struct ldiskfs_extent nex;
 961         unsigned long pblock = 0;
 962         unsigned long tgen;
 963         int err, i;
 964         unsigned long count;
 965         handle_t *handle;
 966
 967 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 968         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 969 #else
 970         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 971 #endif
 972                 err = EXT_CONTINUE;
 973                 goto map;
 974         }
 975
 976         if (bp->create == 0) {
 977                 i = 0;
 978                 if (cex->ec_block < bp->start)
 979                         i = bp->start - cex->ec_block;
 980                 if (i >= cex->ec_len)
 981                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 982                                         i, cex->ec_len);
 983                 for (; i < cex->ec_len && bp->num; i++) {
 984                         *(bp->blocks) = 0;
 985                         bp->blocks++;
 986                         bp->num--;
 987                         bp->start++;
 988                 }
 989
 990                 return EXT_CONTINUE;
 991         }
 992
 993         tgen = LDISKFS_I(inode)->i_ext_generation;
 994         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
 995
 996         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
 997                                    count + LDISKFS_ALLOC_NEEDED + 1);
 998         if (IS_ERR(handle)) {
 999                 return PTR_ERR(handle);
1000         }
1001
1002         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
1003                 /* the tree has changed. so path can be invalid at moment */
1004                 ldiskfs_journal_stop(handle);
1005                 return EXT_REPEAT;
1006         }
1007
1008         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
1009          * protected by i_data_sem as whole. so we patch it to store
1010          * generation to path and now verify the tree hasn't changed */
1011         down_write((&LDISKFS_I(inode)->i_data_sem));
1012
1013         /* validate extent, make sure the extent tree does not changed */
1014         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
1015                 /* cex is invalid, try again */
1016                 up_write(&LDISKFS_I(inode)->i_data_sem);
1017                 ldiskfs_journal_stop(handle);
1018                 return EXT_REPEAT;
1019         }
1020
1021         count = cex->ec_len;
1022         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
1023         if (!pblock)
1024                 goto out;
1025         BUG_ON(count > cex->ec_len);
1026
1027         /* insert new extent */
1028         nex.ee_block = cpu_to_le32(cex->ec_block);
1029         ldiskfs_ext_store_pblock(&nex, pblock);
1030         nex.ee_len = cpu_to_le16(count);
1031         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
1032         if (err) {
1033                 /* free data blocks we just allocated */
1034                 /* not a good idea to call discard here directly,
1035                  * but otherwise we'd need to call it every free() */
1036                 ldiskfs_discard_preallocations(inode);
1037 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
1038                 ldiskfs_free_blocks(handle, inode, NULL,
1039                                     ldiskfs_ext_pblock(&nex),
1040                                     le16_to_cpu(nex.ee_len), 0);
1041 #else
1042                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
1043                                     le16_to_cpu(nex.ee_len), 0);
1044 #endif
1045                 goto out;
1046         }
1047
1048         /*
1049          * Putting len of the actual extent we just inserted,
1050          * we are asking ldiskfs_ext_walk_space() to continue
1051          * scaning after that block
1052          */
1053         cex->ec_len = le16_to_cpu(nex.ee_len);
1054         cex->ec_start = ldiskfs_ext_pblock(&nex);
1055         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
1056         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
1057
1058 out:
1059         up_write((&LDISKFS_I(inode)->i_data_sem));
1060         ldiskfs_journal_stop(handle);
1061 map:
1062         if (err >= 0) {
1063                 /* map blocks */
1064                 if (bp->num == 0) {
1065                         CERROR("hmm. why do we find this extent?\n");
1066                         CERROR("initial space: %lu:%u\n",
1067                                 bp->start, bp->init_num);
1068 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
1069                         CERROR("current extent: %u/%u/%llu %d\n",
1070                                 cex->ec_block, cex->ec_len,
1071                                 (unsigned long long)cex->ec_start,
1072                                 cex->ec_type);
1073 #else
1074                         CERROR("current extent: %u/%u/%llu\n",
1075                                 cex->ec_block, cex->ec_len,
1076                                 (unsigned long long)cex->ec_start);
1077 #endif
1078                 }
1079                 i = 0;
1080                 if (cex->ec_block < bp->start)
1081                         i = bp->start - cex->ec_block;
1082                 if (i >= cex->ec_len)
1083                         CERROR("nothing to do?! i = %d, e_num = %u\n",
1084                                         i, cex->ec_len);
1085                 for (; i < cex->ec_len && bp->num; i++) {
1086                         *(bp->blocks) = cex->ec_start + i;
1087                         if (pblock != 0) {
1088                                 /* unmap any possible underlying metadata from
1089                                  * the block device mapping.  bug 6998. */
1090 #ifndef HAVE_CLEAN_BDEV_ALIASES
1091                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
1092                                                           *(bp->blocks));
1093 #else
1094                                 clean_bdev_aliases(inode->i_sb->s_bdev,
1095                                                    *(bp->blocks), 1);
1096 #endif
1097                         }
1098                         bp->blocks++;
1099                         bp->num--;
1100                         bp->start++;
1101                 }
1102         }
1103         return err;
1104 }
1105
1106 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
1107                                    int clen, sector_t *blocks, int create)
1108 {
1109         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1110         struct bpointers bp;
1111         int err;
1112
1113         if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
1114                 return -EFBIG;
1115
1116         bp.blocks = blocks;
1117         bp.start = index * blocks_per_page;
1118         bp.init_num = bp.num = clen * blocks_per_page;
1119         bp.create = create;
1120
1121         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
1122                bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
1123
1124         err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
1125                                      ldiskfs_ext_new_extent_cb, &bp);
1126         ldiskfs_ext_invalidate_cache(inode);
1127
1128         return err;
1129 }
1130
1131 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
1132                                           struct page **page, int pages,
1133                                           sector_t *blocks, int create)
1134 {
1135         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1136         pgoff_t bitmap_max_page_index;
1137         sector_t *b;
1138         int rc = 0, i;
1139
1140         bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
1141                                 PAGE_SHIFT;
1142         for (i = 0, b = blocks; i < pages; i++, page++) {
1143                 if ((*page)->index + 1 >= bitmap_max_page_index) {
1144                         rc = -EFBIG;
1145                         break;
1146                 }
1147                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
1148                 if (rc) {
1149                         CERROR("ino %lu, blk %llu create %d: rc %d\n",
1150                                inode->i_ino,
1151                                (unsigned long long)*b, create, rc);
1152                         break;
1153                 }
1154                 b += blocks_per_page;
1155         }
1156         return rc;
1157 }
1158
1159 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
1160                                            struct page **page,
1161                                            int pages, sector_t *blocks,
1162                                            int create)
1163 {
1164         int rc = 0, i = 0, clen = 0;
1165         struct page *fp = NULL;
1166
1167         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1168                 inode->i_ino, pages, (*page)->index);
1169
1170         /* pages are sorted already. so, we just have to find
1171          * contig. space and process them properly */
1172         while (i < pages) {
1173                 if (fp == NULL) {
1174                         /* start new extent */
1175                         fp = *page++;
1176                         clen = 1;
1177                         i++;
1178                         continue;
1179                 } else if (fp->index + clen == (*page)->index) {
1180                         /* continue the extent */
1181                         page++;
1182                         clen++;
1183                         i++;
1184                         continue;
1185                 }
1186
1187                 /* process found extent */
1188                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
1189                                              blocks, create);
1190                 if (rc)
1191                         GOTO(cleanup, rc);
1192
1193                 /* look for next extent */
1194                 fp = NULL;
1195                 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
1196         }
1197
1198         if (fp)
1199                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
1200                                              blocks, create);
1201
1202 cleanup:
1203         return rc;
1204 }
1205
1206 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
1207                                        int pages, sector_t *blocks,
1208                                        int create)
1209 {
1210         int rc;
1211
1212         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1213                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
1214                                                      blocks, create);
1215                 return rc;
1216         }
1217         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
1218
1219         return rc;
1220 }
1221 #else
1222 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
1223                                        int pages, sector_t *blocks,
1224                                        int create)
1225 {
1226         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1227         int rc = 0, i = 0;
1228         struct page *fp = NULL;
1229         int clen = 0;
1230         pgoff_t max_page_index;
1231         handle_t *handle = NULL;
1232
1233         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
1234
1235         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1236                 inode->i_ino, pages, (*page)->index);
1237
1238         if (create) {
1239                 create = LDISKFS_GET_BLOCKS_CREATE;
1240                 handle = ldiskfs_journal_current_handle();
1241                 LASSERT(handle != NULL);
1242                 rc = osd_attach_jinode(inode);
1243                 if (rc)
1244                         return rc;
1245         }
1246         /* pages are sorted already. so, we just have to find
1247          * contig. space and process them properly */
1248         while (i < pages) {
1249                 long blen, total = 0;
1250                 struct ldiskfs_map_blocks map = { 0 };
1251
1252                 if (fp == NULL) { /* start new extent */
1253                         fp = *page++;
1254                         clen = 1;
1255                         if (++i != pages)
1256                                 continue;
1257                 } else if (fp->index + clen == (*page)->index) {
1258                         /* continue the extent */
1259                         page++;
1260                         clen++;
1261                         if (++i != pages)
1262                                 continue;
1263                 }
1264                 if (fp->index + clen >= max_page_index)
1265                         GOTO(cleanup, rc = -EFBIG);
1266                 /* process found extent */
1267                 map.m_lblk = fp->index * blocks_per_page;
1268                 map.m_len = blen = clen * blocks_per_page;
1269 cont_map:
1270                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
1271                 if (rc >= 0) {
1272                         int c = 0;
1273                         for (; total < blen && c < map.m_len; c++, total++) {
1274                                 if (rc == 0) {
1275                                         *(blocks + total) = 0;
1276                                         total++;
1277                                         break;
1278                                 } else {
1279                                         *(blocks + total) = map.m_pblk + c;
1280                                         /* unmap any possible underlying
1281                                          * metadata from the block device
1282                                          * mapping.  bug 6998. */
1283                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
1284                                             create)
1285 #ifndef HAVE_CLEAN_BDEV_ALIASES
1286                                                 unmap_underlying_metadata(
1287                                                         inode->i_sb->s_bdev,
1288                                                         map.m_pblk + c);
1289 #else
1290                                                 clean_bdev_aliases(
1291                                                         inode->i_sb->s_bdev,
1292                                                         map.m_pblk + c, 1);
1293 #endif
1294                                 }
1295                         }
1296                         rc = 0;
1297                 }
1298                 if (rc == 0 && total < blen) {
1299                         map.m_lblk = fp->index * blocks_per_page + total;
1300                         map.m_len = blen - total;
1301                         goto cont_map;
1302                 }
1303                 if (rc != 0)
1304                         GOTO(cleanup, rc);
1305
1306                 /* look for next extent */
1307                 fp = NULL;
1308                 blocks += blocks_per_page * clen;
1309         }
1310 cleanup:
1311         return rc;
1312 }
1313 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
1314
1315 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1316                           struct niobuf_local *lnb, int npages)
1317 {
1318         struct osd_thread_info *oti   = osd_oti_get(env);
1319         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1320         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
1321         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
1322         ktime_t start;
1323         ktime_t end;
1324         s64 timediff;
1325         ssize_t                 isize;
1326         __s64                   maxidx;
1327         int                     rc = 0;
1328         int                     i;
1329         int                     cache = 0;
1330
1331         LASSERT(inode);
1332
1333         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1334         if (unlikely(rc != 0))
1335                 RETURN(rc);
1336
1337         isize = i_size_read(inode);
1338         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1339
1340         if (osd->od_writethrough_cache)
1341                 cache = 1;
1342         if (isize > osd->od_readcache_max_filesize)
1343                 cache = 0;
1344
1345         start = ktime_get();
1346         for (i = 0; i < npages; i++) {
1347
1348                 if (cache == 0)
1349                         generic_error_remove_page(inode->i_mapping,
1350                                                   lnb[i].lnb_page);
1351
1352                 /*
1353                  * till commit the content of the page is undefined
1354                  * we'll set it uptodate once bulk is done. otherwise
1355                  * subsequent reads can access non-stable data
1356                  */
1357                 ClearPageUptodate(lnb[i].lnb_page);
1358
1359                 if (lnb[i].lnb_len == PAGE_SIZE)
1360                         continue;
1361
1362                 if (maxidx >= lnb[i].lnb_page->index) {
1363                         osd_iobuf_add_page(iobuf, &lnb[i]);
1364                 } else {
1365                         long off;
1366                         char *p = kmap(lnb[i].lnb_page);
1367
1368                         off = lnb[i].lnb_page_offset;
1369                         if (off)
1370                                 memset(p, 0, off);
1371                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1372                               ~PAGE_MASK;
1373                         if (off)
1374                                 memset(p + off, 0, PAGE_SIZE - off);
1375                         kunmap(lnb[i].lnb_page);
1376                 }
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 (iobuf->dr_npages) {
1383                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1384                                                  iobuf->dr_npages,
1385                                                  iobuf->dr_blocks, 0);
1386                 if (likely(rc == 0)) {
1387                         rc = osd_do_bio(osd, inode, iobuf);
1388                         /* do IO stats for preparation reads */
1389                         osd_fini_iobuf(osd, iobuf);
1390                 }
1391         }
1392         RETURN(rc);
1393 }
1394
1395 struct osd_fextent {
1396         sector_t        start;
1397         sector_t        end;
1398         unsigned int    mapped:1;
1399 };
1400
1401 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1402                          struct osd_fextent *cached_extent)
1403 {
1404         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1405         sector_t block = offset >> inode->i_blkbits;
1406         sector_t start;
1407         struct fiemap_extent_info fei = { 0 };
1408         struct fiemap_extent fe = { 0 };
1409         mm_segment_t saved_fs;
1410         int rc;
1411
1412         if (block >= cached_extent->start && block < cached_extent->end)
1413                 return cached_extent->mapped;
1414
1415         if (i_size_read(inode) == 0)
1416                 return 0;
1417
1418         /* Beyond EOF, must not be mapped */
1419         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1420                 return 0;
1421
1422         fei.fi_extents_max = 1;
1423         fei.fi_extents_start = &fe;
1424
1425         saved_fs = get_fs();
1426         set_fs(get_ds());
1427         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1428         set_fs(saved_fs);
1429         if (rc != 0)
1430                 return 0;
1431
1432         start = fe.fe_logical >> inode->i_blkbits;
1433
1434         if (start > block) {
1435                 cached_extent->start = block;
1436                 cached_extent->end = start;
1437                 cached_extent->mapped = 0;
1438         } else {
1439                 cached_extent->start = start;
1440                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1441                                       inode->i_blkbits;
1442                 cached_extent->mapped = 1;
1443         }
1444
1445         return cached_extent->mapped;
1446 }
1447
1448 static int osd_declare_write_commit(const struct lu_env *env,
1449                                     struct dt_object *dt,
1450                                     struct niobuf_local *lnb, int npages,
1451                                     struct thandle *handle)
1452 {
1453         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1454         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1455         struct osd_thandle      *oh;
1456         int                     extents = 1;
1457         int                     depth;
1458         int                     i;
1459         int                     newblocks;
1460         int                     rc = 0;
1461         int                     flags = 0;
1462         int                     credits = 0;
1463         long long               quota_space = 0;
1464         struct osd_fextent      extent = { 0 };
1465         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1466         ENTRY;
1467
1468         LASSERT(handle != NULL);
1469         oh = container_of0(handle, struct osd_thandle, ot_super);
1470         LASSERT(oh->ot_handle == NULL);
1471
1472         newblocks = npages;
1473
1474         /* calculate number of extents (probably better to pass nb) */
1475         for (i = 0; i < npages; i++) {
1476                 if (i && lnb[i].lnb_file_offset !=
1477                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1478                         extents++;
1479
1480                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1481                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1482                 else
1483                         quota_space += PAGE_SIZE;
1484
1485                 /* ignore quota for the whole request if any page is from
1486                  * client cache or written by root.
1487                  *
1488                  * XXX once we drop the 1.8 client support, the checking
1489                  * for whether page is from cache can be simplified as:
1490                  * !(lnb[i].flags & OBD_BRW_SYNC)
1491                  *
1492                  * XXX we could handle this on per-lnb basis as done by
1493                  * grant. */
1494                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1495                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1496                     OBD_BRW_FROM_GRANT)
1497                         declare_flags |= OSD_QID_FORCE;
1498         }
1499
1500         /*
1501          * each extent can go into new leaf causing a split
1502          * 5 is max tree depth: inode + 4 index blocks
1503          * with blockmaps, depth is 3 at most
1504          */
1505         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1506                 /*
1507                  * many concurrent threads may grow tree by the time
1508                  * our transaction starts. so, consider 2 is a min depth
1509                  */
1510                 depth = ext_depth(inode);
1511                 depth = max(depth, 1) + 1;
1512                 newblocks += depth;
1513                 credits++; /* inode */
1514                 credits += depth * 2 * extents;
1515         } else {
1516                 depth = 3;
1517                 newblocks += depth;
1518                 credits++; /* inode */
1519                 credits += depth * extents;
1520         }
1521
1522         /* quota space for metadata blocks */
1523         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1524
1525         /* quota space should be reported in 1K blocks */
1526         quota_space = toqb(quota_space);
1527
1528         /* each new block can go in different group (bitmap + gd) */
1529
1530         /* we can't dirty more bitmap blocks than exist */
1531         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1532                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1533         else
1534                 credits += newblocks;
1535
1536         /* we can't dirty more gd blocks than exist */
1537         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1538                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1539         else
1540                 credits += newblocks;
1541
1542         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1543
1544         /* make sure the over quota flags were not set */
1545         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1546
1547         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1548                                    i_projid_read(inode), quota_space, oh,
1549                                    osd_dt_obj(dt), &flags, declare_flags);
1550
1551         /* we need only to store the overquota flags in the first lnb for
1552          * now, once we support multiple objects BRW, this code needs be
1553          * revised. */
1554         if (flags & QUOTA_FL_OVER_USRQUOTA)
1555                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1556         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1557                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1558         if (flags & QUOTA_FL_OVER_PRJQUOTA)
1559                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1560
1561         if (rc == 0)
1562                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1563
1564         RETURN(rc);
1565 }
1566
1567 /* Check if a block is allocated or not */
1568 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1569                             struct niobuf_local *lnb, int npages,
1570                             struct thandle *thandle)
1571 {
1572         struct osd_thread_info *oti = osd_oti_get(env);
1573         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1574         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1575         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1576         loff_t isize;
1577         int rc = 0, i;
1578
1579         LASSERT(inode);
1580
1581         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1582         if (unlikely(rc != 0))
1583                 RETURN(rc);
1584
1585         isize = i_size_read(inode);
1586         ll_vfs_dq_init(inode);
1587
1588         for (i = 0; i < npages; i++) {
1589                 if (lnb[i].lnb_rc == -ENOSPC &&
1590                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1591                         /* Allow the write to proceed if overwriting an
1592                          * existing block */
1593                         lnb[i].lnb_rc = 0;
1594                 }
1595
1596                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1597                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1598                                lnb[i].lnb_rc);
1599                         LASSERT(lnb[i].lnb_page);
1600                         generic_error_remove_page(inode->i_mapping,
1601                                                   lnb[i].lnb_page);
1602                         continue;
1603                 }
1604
1605                 LASSERT(PageLocked(lnb[i].lnb_page));
1606                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1607
1608                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1609                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1610
1611                 /*
1612                  * Since write and truncate are serialized by oo_sem, even
1613                  * partial-page truncate should not leave dirty pages in the
1614                  * page cache.
1615                  */
1616                 LASSERT(!PageDirty(lnb[i].lnb_page));
1617
1618                 SetPageUptodate(lnb[i].lnb_page);
1619
1620                 osd_iobuf_add_page(iobuf, &lnb[i]);
1621         }
1622
1623         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1624
1625         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1626                 rc = -ENOSPC;
1627         } else if (iobuf->dr_npages > 0) {
1628                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1629                                                  iobuf->dr_npages,
1630                                                  iobuf->dr_blocks, 1);
1631         } else {
1632                 /* no pages to write, no transno is needed */
1633                 thandle->th_local = 1;
1634         }
1635
1636         if (likely(rc == 0)) {
1637                 spin_lock(&inode->i_lock);
1638                 if (isize > i_size_read(inode)) {
1639                         i_size_write(inode, isize);
1640                         LDISKFS_I(inode)->i_disksize = isize;
1641                         spin_unlock(&inode->i_lock);
1642                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1643                 } else {
1644                         spin_unlock(&inode->i_lock);
1645                 }
1646
1647                 rc = osd_do_bio(osd, inode, iobuf);
1648                 /* we don't do stats here as in read path because
1649                  * write is async: we'll do this in osd_put_bufs() */
1650         } else {
1651                 osd_fini_iobuf(osd, iobuf);
1652         }
1653
1654         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1655
1656         if (unlikely(rc != 0)) {
1657                 /* if write fails, we should drop pages from the cache */
1658                 for (i = 0; i < npages; i++) {
1659                         if (lnb[i].lnb_page == NULL)
1660                                 continue;
1661                         LASSERT(PageLocked(lnb[i].lnb_page));
1662                         generic_error_remove_page(inode->i_mapping,
1663                                                   lnb[i].lnb_page);
1664                 }
1665         }
1666
1667         RETURN(rc);
1668 }
1669
1670 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1671                          struct niobuf_local *lnb, int npages)
1672 {
1673         struct osd_thread_info *oti = osd_oti_get(env);
1674         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1675         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1676         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1677         int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1678         ktime_t start, end;
1679         s64 timediff;
1680         loff_t isize;
1681
1682         LASSERT(inode);
1683
1684         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1685         if (unlikely(rc != 0))
1686                 RETURN(rc);
1687
1688         isize = i_size_read(inode);
1689
1690         if (osd->od_read_cache)
1691                 cache = 1;
1692         if (isize > osd->od_readcache_max_filesize)
1693                 cache = 0;
1694
1695         start = ktime_get();
1696         for (i = 0; i < npages; i++) {
1697
1698                 if (isize <= lnb[i].lnb_file_offset)
1699                         /* If there's no more data, abort early.
1700                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1701                         break;
1702
1703                 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1704                         lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1705                 else
1706                         lnb[i].lnb_rc = lnb[i].lnb_len;
1707
1708                 /* Bypass disk read if fail_loc is set properly */
1709                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1710                         SetPageUptodate(lnb[i].lnb_page);
1711
1712                 if (PageUptodate(lnb[i].lnb_page)) {
1713                         cache_hits++;
1714                 } else {
1715                         cache_misses++;
1716                         osd_iobuf_add_page(iobuf, &lnb[i]);
1717                 }
1718
1719                 if (cache == 0)
1720                         generic_error_remove_page(inode->i_mapping,
1721                                                   lnb[i].lnb_page);
1722         }
1723         end = ktime_get();
1724         timediff = ktime_us_delta(end, start);
1725         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1726
1727         if (cache_hits != 0)
1728                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1729                                     cache_hits);
1730         if (cache_misses != 0)
1731                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1732                                     cache_misses);
1733         if (cache_hits + cache_misses != 0)
1734                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1735                                     cache_hits + cache_misses);
1736
1737         if (iobuf->dr_npages) {
1738                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1739                                                  iobuf->dr_npages,
1740                                                  iobuf->dr_blocks, 0);
1741                 rc = osd_do_bio(osd, inode, iobuf);
1742
1743                 /* IO stats will be done in osd_bufs_put() */
1744         }
1745
1746         RETURN(rc);
1747 }
1748
1749 /*
1750  * XXX: Another layering violation for now.
1751  *
1752  * We don't want to use ->f_op->read methods, because generic file write
1753  *
1754  *         - serializes on ->i_sem, and
1755  *
1756  *         - does a lot of extra work like balance_dirty_pages(),
1757  *
1758  * which doesn't work for globally shared files like /last_rcvd.
1759  */
1760 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1761 {
1762         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1763
1764         memcpy(buffer, (char *)ei->i_data, buflen);
1765
1766         return  buflen;
1767 }
1768
1769 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1770 {
1771         struct buffer_head *bh;
1772         unsigned long block;
1773         int osize;
1774         int blocksize;
1775         int csize;
1776         int boffs;
1777
1778         /* prevent reading after eof */
1779         spin_lock(&inode->i_lock);
1780         if (i_size_read(inode) < *offs + size) {
1781                 loff_t diff = i_size_read(inode) - *offs;
1782                 spin_unlock(&inode->i_lock);
1783                 if (diff < 0) {
1784                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1785                                i_size_read(inode), *offs);
1786                         return -EBADR;
1787                 } else if (diff == 0) {
1788                         return 0;
1789                 } else {
1790                         size = diff;
1791                 }
1792         } else {
1793                 spin_unlock(&inode->i_lock);
1794         }
1795
1796         blocksize = 1 << inode->i_blkbits;
1797         osize = size;
1798         while (size > 0) {
1799                 block = *offs >> inode->i_blkbits;
1800                 boffs = *offs & (blocksize - 1);
1801                 csize = min(blocksize - boffs, size);
1802                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1803                 if (IS_ERR(bh)) {
1804                         CERROR("%s: can't read %u@%llu on ino %lu: "
1805                                "rc = %ld\n", osd_ino2name(inode),
1806                                csize, *offs, inode->i_ino,
1807                                PTR_ERR(bh));
1808                         return PTR_ERR(bh);
1809                 }
1810
1811                 if (bh != NULL) {
1812                         memcpy(buf, bh->b_data + boffs, csize);
1813                         brelse(bh);
1814                 } else {
1815                         memset(buf, 0, csize);
1816                 }
1817
1818                 *offs += csize;
1819                 buf += csize;
1820                 size -= csize;
1821         }
1822         return osize;
1823 }
1824
1825 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1826                         struct lu_buf *buf, loff_t *pos)
1827 {
1828         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1829         int           rc;
1830
1831         /* Read small symlink from inode body as we need to maintain correct
1832          * on-disk symlinks for ldiskfs.
1833          */
1834         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1835             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1836                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1837         else
1838                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1839
1840         return rc;
1841 }
1842
1843 static inline int osd_extents_enabled(struct super_block *sb,
1844                                       struct inode *inode)
1845 {
1846         if (inode != NULL) {
1847                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1848                         return 1;
1849         } else if (ldiskfs_has_feature_extents(sb)) {
1850                 return 1;
1851         }
1852         return 0;
1853 }
1854
1855 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1856                            const loff_t size, const loff_t pos,
1857                            const int blocks)
1858 {
1859         int credits, bits, bs, i;
1860
1861         bits = sb->s_blocksize_bits;
1862         bs = 1 << bits;
1863
1864         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1865          * we do not expect blockmaps on the large files,
1866          * so let's shrink it to 2 levels (4GB files) */
1867
1868         /* this is default reservation: 2 levels */
1869         credits = (blocks + 2) * 3;
1870
1871         /* actual offset is unknown, hard to optimize */
1872         if (pos == -1)
1873                 return credits;
1874
1875         /* now check for few specific cases to optimize */
1876         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1877                 /* no indirects */
1878                 credits = blocks;
1879                 /* allocate if not allocated */
1880                 if (inode == NULL) {
1881                         credits += blocks * 2;
1882                         return credits;
1883                 }
1884                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1885                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1886                         if (LDISKFS_I(inode)->i_data[i] == 0)
1887                                 credits += 2;
1888                 }
1889         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1890                 /* single indirect */
1891                 credits = blocks * 3;
1892                 if (inode == NULL ||
1893                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1894                         credits += 3;
1895                 else
1896                         /* The indirect block may be modified. */
1897                         credits += 1;
1898         }
1899
1900         return credits;
1901 }
1902
1903 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1904                                  const struct lu_buf *buf, loff_t _pos,
1905                                  struct thandle *handle)
1906 {
1907         struct osd_object  *obj  = osd_dt_obj(dt);
1908         struct inode       *inode = obj->oo_inode;
1909         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1910         struct osd_thandle *oh;
1911         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1912         int                 bits, bs;
1913         int                 depth, size;
1914         loff_t              pos;
1915         ENTRY;
1916
1917         LASSERT(buf != NULL);
1918         LASSERT(handle != NULL);
1919
1920         oh = container_of0(handle, struct osd_thandle, ot_super);
1921         LASSERT(oh->ot_handle == NULL);
1922
1923         size = buf->lb_len;
1924         bits = sb->s_blocksize_bits;
1925         bs = 1 << bits;
1926
1927         if (_pos == -1) {
1928                 /* if this is an append, then we
1929                  * should expect cross-block record */
1930                 pos = 0;
1931         } else {
1932                 pos = _pos;
1933         }
1934
1935         /* blocks to modify */
1936         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1937         LASSERT(blocks > 0);
1938
1939         if (inode != NULL && _pos != -1) {
1940                 /* object size in blocks */
1941                 est = (i_size_read(inode) + bs - 1) >> bits;
1942                 allocated = inode->i_blocks >> (bits - 9);
1943                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1944                         /* looks like an overwrite, no need to modify tree */
1945                         credits = blocks;
1946                         /* no need to modify i_size */
1947                         goto out;
1948                 }
1949         }
1950
1951         if (osd_extents_enabled(sb, inode)) {
1952                 /*
1953                  * many concurrent threads may grow tree by the time
1954                  * our transaction starts. so, consider 2 is a min depth
1955                  * for every level we may need to allocate a new block
1956                  * and take some entries from the old one. so, 3 blocks
1957                  * to allocate (bitmap, gd, itself) + old block - 4 per
1958                  * level.
1959                  */
1960                 depth = inode != NULL ? ext_depth(inode) : 0;
1961                 depth = max(depth, 1) + 1;
1962                 credits = depth;
1963                 /* if not append, then split may need to modify
1964                  * existing blocks moving entries into the new ones */
1965                 if (_pos != -1)
1966                         credits += depth;
1967                 /* blocks to store data: bitmap,gd,itself */
1968                 credits += blocks * 3;
1969         } else {
1970                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1971         }
1972         /* if inode is created as part of the transaction,
1973          * then it's counted already by the creation method */
1974         if (inode != NULL)
1975                 credits++;
1976
1977 out:
1978
1979         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1980
1981         /* dt_declare_write() is usually called for system objects, such
1982          * as llog or last_rcvd files. We needn't enforce quota on those
1983          * objects, so always set the lqi_space as 0. */
1984         if (inode != NULL)
1985                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1986                                            i_gid_read(inode),
1987                                            i_projid_read(inode), 0,
1988                                            oh, obj, NULL, OSD_QID_BLK);
1989
1990         if (rc == 0)
1991                 rc = osd_trunc_lock(obj, oh, true);
1992
1993         RETURN(rc);
1994 }
1995
1996 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1997 {
1998         /* LU-2634: clear the extent format for fast symlink */
1999         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
2000
2001         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
2002         spin_lock(&inode->i_lock);
2003         LDISKFS_I(inode)->i_disksize = buflen;
2004         i_size_write(inode, buflen);
2005         spin_unlock(&inode->i_lock);
2006         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2007
2008         return 0;
2009 }
2010
2011 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
2012                              int write_NUL, loff_t *offs, handle_t *handle)
2013 {
2014         struct buffer_head *bh        = NULL;
2015         loff_t              offset    = *offs;
2016         loff_t              new_size  = i_size_read(inode);
2017         unsigned long       block;
2018         int                 blocksize = 1 << inode->i_blkbits;
2019         int                 err = 0;
2020         int                 size;
2021         int                 boffs;
2022         int                 dirty_inode = 0;
2023
2024         if (write_NUL) {
2025                 /*
2026                  * long symlink write does not count the NUL terminator in
2027                  * bufsize, we write it, and the inode's file size does not
2028                  * count the NUL terminator as well.
2029                  */
2030                 ((char *)buf)[bufsize] = '\0';
2031                 ++bufsize;
2032         }
2033
2034         while (bufsize > 0) {
2035                 int credits = handle->h_buffer_credits;
2036
2037                 if (bh)
2038                         brelse(bh);
2039
2040                 block = offset >> inode->i_blkbits;
2041                 boffs = offset & (blocksize - 1);
2042                 size = min(blocksize - boffs, bufsize);
2043                 bh = __ldiskfs_bread(handle, inode, block, 1);
2044                 if (IS_ERR_OR_NULL(bh)) {
2045                         if (bh == NULL) {
2046                                 err = -EIO;
2047                         } else {
2048                                 err = PTR_ERR(bh);
2049                                 bh = NULL;
2050                         }
2051
2052                         CERROR("%s: error reading offset %llu (block %lu, "
2053                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
2054                                inode->i_sb->s_id, offset, block, bufsize, *offs,
2055                                credits, handle->h_buffer_credits, err);
2056                         break;
2057                 }
2058
2059                 err = ldiskfs_journal_get_write_access(handle, bh);
2060                 if (err) {
2061                         CERROR("journal_get_write_access() returned error %d\n",
2062                                err);
2063                         break;
2064                 }
2065                 LASSERTF(boffs + size <= bh->b_size,
2066                          "boffs %d size %d bh->b_size %lu\n",
2067                          boffs, size, (unsigned long)bh->b_size);
2068                 memcpy(bh->b_data + boffs, buf, size);
2069                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2070                 if (err)
2071                         break;
2072
2073                 if (offset + size > new_size)
2074                         new_size = offset + size;
2075                 offset += size;
2076                 bufsize -= size;
2077                 buf += size;
2078         }
2079         if (bh)
2080                 brelse(bh);
2081
2082         if (write_NUL)
2083                 --new_size;
2084         /* correct in-core and on-disk sizes */
2085         if (new_size > i_size_read(inode)) {
2086                 spin_lock(&inode->i_lock);
2087                 if (new_size > i_size_read(inode))
2088                         i_size_write(inode, new_size);
2089                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
2090                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
2091                         dirty_inode = 1;
2092                 }
2093                 spin_unlock(&inode->i_lock);
2094                 if (dirty_inode)
2095                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2096         }
2097
2098         if (err == 0)
2099                 *offs = offset;
2100         return err;
2101 }
2102
2103 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
2104                          const struct lu_buf *buf, loff_t *pos,
2105                          struct thandle *handle)
2106 {
2107         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
2108         struct osd_thandle      *oh;
2109         ssize_t                 result;
2110         int                     is_link;
2111
2112         LASSERT(dt_object_exists(dt));
2113
2114         LASSERT(handle != NULL);
2115         LASSERT(inode != NULL);
2116         ll_vfs_dq_init(inode);
2117
2118         /* XXX: don't check: one declared chunk can be used many times */
2119         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
2120
2121         oh = container_of(handle, struct osd_thandle, ot_super);
2122         LASSERT(oh->ot_handle->h_transaction != NULL);
2123         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
2124
2125         /* Write small symlink to inode body as we need to maintain correct
2126          * on-disk symlinks for ldiskfs.
2127          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
2128          * does not count it in.
2129          */
2130         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
2131         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
2132                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
2133         else
2134                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
2135                                                   buf->lb_len, is_link, pos,
2136                                                   oh->ot_handle);
2137         if (result == 0)
2138                 result = buf->lb_len;
2139
2140         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
2141
2142         return result;
2143 }
2144
2145 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
2146                              __u64 start, __u64 end, struct thandle *th)
2147 {
2148         struct osd_thandle *oh;
2149         struct inode       *inode;
2150         int                 rc;
2151         ENTRY;
2152
2153         LASSERT(th);
2154         oh = container_of(th, struct osd_thandle, ot_super);
2155
2156         /*
2157          * we don't need to reserve credits for whole truncate
2158          * it's not possible as truncate may need to free too many
2159          * blocks and that won't fit a single transaction. instead
2160          * we reserve credits to change i_size and put inode onto
2161          * orphan list. if needed truncate will extend or restart
2162          * transaction
2163          */
2164         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
2165                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
2166
2167         inode = osd_dt_obj(dt)->oo_inode;
2168         LASSERT(inode);
2169
2170         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2171                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
2172                                    NULL, OSD_QID_BLK);
2173
2174         if (rc == 0)
2175                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2176
2177         RETURN(rc);
2178 }
2179
2180 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
2181                      __u64 start, __u64 end, struct thandle *th)
2182 {
2183         struct osd_object *obj = osd_dt_obj(dt);
2184         struct osd_device *osd = osd_obj2dev(obj);
2185         struct inode *inode = obj->oo_inode;
2186         struct osd_access_lock *al;
2187         struct osd_thandle *oh;
2188         int rc = 0, found = 0;
2189         bool grow = false;
2190         ENTRY;
2191
2192         LASSERT(end == OBD_OBJECT_EOF);
2193         LASSERT(dt_object_exists(dt));
2194         LASSERT(osd_invariant(obj));
2195         LASSERT(inode != NULL);
2196         ll_vfs_dq_init(inode);
2197
2198         LASSERT(th);
2199         oh = container_of(th, struct osd_thandle, ot_super);
2200         LASSERT(oh->ot_handle->h_transaction != NULL);
2201
2202         /* we used to skip truncate to current size to
2203          * optimize truncates on OST. with DoM we can
2204          * get attr_set to set specific size (MDS_REINT)
2205          * and then get truncate RPC which essentially
2206          * would be skipped. this is bad.. so, disable
2207          * this optimization on MDS till the client stop
2208          * to sent MDS_REINT (LU-11033) -bzzz */
2209         if (osd->od_is_ost && i_size_read(inode) == start)
2210                 RETURN(0);
2211
2212         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2213
2214         spin_lock(&inode->i_lock);
2215         if (i_size_read(inode) < start)
2216                 grow = true;
2217         i_size_write(inode, start);
2218         spin_unlock(&inode->i_lock);
2219         ll_truncate_pagecache(inode, start);
2220
2221         /* optimize grow case */
2222         if (grow) {
2223                 osd_execute_truncate(obj);
2224                 GOTO(out, rc);
2225         }
2226
2227         /* add to orphan list to ensure truncate completion
2228          * if this transaction succeed. ldiskfs_truncate()
2229          * will take the inode out of the list */
2230         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2231         if (rc != 0)
2232                 GOTO(out, rc);
2233
2234         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2235                 if (obj != al->tl_obj)
2236                         continue;
2237                 LASSERT(al->tl_shared == 0);
2238                 found = 1;
2239                 /* do actual truncate in osd_trans_stop() */
2240                 al->tl_truncate = 1;
2241                 break;
2242         }
2243         LASSERT(found);
2244
2245 out:
2246         RETURN(rc);
2247 }
2248
2249 static int fiemap_check_ranges(struct inode *inode,
2250                                u64 start, u64 len, u64 *new_len)
2251 {
2252         loff_t maxbytes;
2253
2254         *new_len = len;
2255
2256         if (len == 0)
2257                 return -EINVAL;
2258
2259         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2260                 maxbytes = inode->i_sb->s_maxbytes;
2261         else
2262                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2263
2264         if (start > maxbytes)
2265                 return -EFBIG;
2266
2267         /*
2268          * Shrink request scope to what the fs can actually handle.
2269          */
2270         if (len > maxbytes || (maxbytes - len) < start)
2271                 *new_len = maxbytes - start;
2272
2273         return 0;
2274 }
2275
2276 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2277 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
2278
2279 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2280                           struct fiemap *fm)
2281 {
2282         struct fiemap_extent_info fieinfo = {0, };
2283         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2284         u64 len;
2285         int rc;
2286
2287
2288         LASSERT(inode);
2289         if (inode->i_op->fiemap == NULL)
2290                 return -EOPNOTSUPP;
2291
2292         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2293                 return -EINVAL;
2294
2295         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2296         if (rc)
2297                 return rc;
2298
2299         fieinfo.fi_flags = fm->fm_flags;
2300         fieinfo.fi_extents_max = fm->fm_extent_count;
2301         fieinfo.fi_extents_start = fm->fm_extents;
2302
2303         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2304                 filemap_write_and_wait(inode->i_mapping);
2305
2306         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2307         fm->fm_flags = fieinfo.fi_flags;
2308         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2309
2310         return rc;
2311 }
2312
2313 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2314                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2315 {
2316         int              rc = 0;
2317         struct inode    *inode = osd_dt_obj(dt)->oo_inode;
2318         ENTRY;
2319
2320         switch (advice) {
2321         case LU_LADVISE_DONTNEED:
2322                 if (end == 0)
2323                         break;
2324                 invalidate_mapping_pages(inode->i_mapping,
2325                                          start >> PAGE_SHIFT,
2326                                          (end - 1) >> PAGE_SHIFT);
2327                 break;
2328         default:
2329                 rc = -ENOTSUPP;
2330                 break;
2331         }
2332
2333         RETURN(rc);
2334 }
2335
2336 /*
2337  * in some cases we may need declare methods for objects being created
2338  * e.g., when we create symlink
2339  */
2340 const struct dt_body_operations osd_body_ops_new = {
2341         .dbo_declare_write = osd_declare_write,
2342 };
2343
2344 const struct dt_body_operations osd_body_ops = {
2345         .dbo_read                       = osd_read,
2346         .dbo_declare_write              = osd_declare_write,
2347         .dbo_write                      = osd_write,
2348         .dbo_bufs_get                   = osd_bufs_get,
2349         .dbo_bufs_put                   = osd_bufs_put,
2350         .dbo_write_prep                 = osd_write_prep,
2351         .dbo_declare_write_commit       = osd_declare_write_commit,
2352         .dbo_write_commit               = osd_write_commit,
2353         .dbo_read_prep                  = osd_read_prep,
2354         .dbo_declare_punch              = osd_declare_punch,
2355         .dbo_punch                      = osd_punch,
2356         .dbo_fiemap_get                 = osd_fiemap_get,
2357         .dbo_ladvise                    = osd_ladvise,
2358 };
2359
2360 /**
2361  * Get a truncate lock
2362  *
2363  * In order to take multi-transaction truncate out of main transaction we let
2364  * the caller grab a lock on the object passed. the lock can be shared (for
2365  * writes) and exclusive (for truncate). It's not allowed to mix truncate
2366  * and write in the same transaction handle (do not confuse with big ldiskfs
2367  * transaction containing lots of handles).
2368  * The lock must be taken at declaration.
2369  *
2370  * \param obj           object to lock
2371  * \oh                  transaction
2372  * \shared              shared or exclusive
2373  *
2374  * \retval 0            lock is granted
2375  * \retval -NOMEM       no memory to allocate lock
2376  */
2377 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2378 {
2379         struct osd_access_lock *al, *tmp;
2380
2381         LASSERT(obj);
2382         LASSERT(oh);
2383
2384         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2385                 if (tmp->tl_obj != obj)
2386                         continue;
2387                 LASSERT(tmp->tl_shared == shared);
2388                 /* found same lock */
2389                 return 0;
2390         }
2391
2392         OBD_ALLOC_PTR(al);
2393         if (unlikely(al == NULL))
2394                 return -ENOMEM;
2395         al->tl_obj = obj;
2396         al->tl_truncate = false;
2397         if (shared)
2398                 down_read(&obj->oo_ext_idx_sem);
2399         else
2400                 down_write(&obj->oo_ext_idx_sem);
2401         al->tl_shared = shared;
2402
2403         list_add(&al->tl_list, &oh->ot_trunc_locks);
2404
2405         return 0;
2406 }
2407
2408 void osd_trunc_unlock_all(struct list_head *list)
2409 {
2410         struct osd_access_lock *al, *tmp;
2411         list_for_each_entry_safe(al, tmp, list, tl_list) {
2412                 if (al->tl_shared)
2413                         up_read(&al->tl_obj->oo_ext_idx_sem);
2414                 else
2415                         up_write(&al->tl_obj->oo_ext_idx_sem);
2416                 list_del(&al->tl_list);
2417                 OBD_FREE_PTR(al);
2418         }
2419 }
2420
2421 void osd_execute_truncate(struct osd_object *obj)
2422 {
2423         struct osd_device *d = osd_obj2dev(obj);
2424         struct inode *inode = obj->oo_inode;
2425         __u64 size;
2426
2427         /* simulate crash before (in the middle) of delayed truncate */
2428         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2429                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2430                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2431
2432                 mutex_lock(&sbi->s_orphan_lock);
2433                 list_del_init(&ei->i_orphan);
2434                 mutex_unlock(&sbi->s_orphan_lock);
2435                 return;
2436         }
2437
2438 #ifdef HAVE_INODEOPS_TRUNCATE
2439         if (inode->i_op->truncate)
2440                 inode->i_op->truncate(inode);
2441         else
2442 #endif
2443                 ldiskfs_truncate(inode);
2444
2445         /*
2446          * For a partial-page truncate, flush the page to disk immediately to
2447          * avoid data corruption during direct disk write.  b=17397
2448          */
2449         size = i_size_read(inode);
2450         if ((size & ~PAGE_MASK) == 0)
2451                 return;
2452         if (osd_use_page_cache(d)) {
2453                 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2454         } else {
2455                 /* Notice we use "wait" version to ensure I/O is complete */
2456                 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2457                 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2458                                          size >> PAGE_SHIFT);
2459         }
2460 }
2461
2462 void osd_process_truncates(struct list_head *list)
2463 {
2464         struct osd_access_lock *al;
2465
2466         LASSERT(journal_current_handle() == NULL);
2467
2468         list_for_each_entry(al, list, tl_list) {
2469                 if (al->tl_shared)
2470                         continue;
2471                 if (!al->tl_truncate)
2472                         continue;
2473                 osd_execute_truncate(al->tl_obj);
2474         }
2475 }