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         iobuf->dr_max_pages = pages;
 121
 122         return 0;
 123 }
 124 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 125         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 126
 127 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
 128 {
 129         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 130         iobuf->dr_pages[iobuf->dr_npages++] = page;
 131 }
 132
 133 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 134 {
 135         int rw = iobuf->dr_rw;
 136
 137         if (iobuf->dr_elapsed_valid) {
 138                 iobuf->dr_elapsed_valid = 0;
 139                 LASSERT(iobuf->dr_dev == d);
 140                 LASSERT(iobuf->dr_frags > 0);
 141                 lprocfs_oh_tally(&d->od_brw_stats.
 142                                  hist[BRW_R_DIO_FRAGS+rw],
 143                                  iobuf->dr_frags);
 144                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 145                                       ktime_to_ms(iobuf->dr_elapsed));
 146         }
 147 }
 148
 149 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 150 static void dio_complete_routine(struct bio *bio)
 151 {
 152 # ifdef HAVE_BI_STATUS
 153         int error = bio->bi_status;
 154 # else
 155         int error = bio->bi_error;
 156 # endif
 157 #else
 158 static void dio_complete_routine(struct bio *bio, int error)
 159 {
 160 #endif
 161         struct osd_iobuf *iobuf = bio->bi_private;
 162         int iter;
 163         struct bio_vec *bvl;
 164
 165         /* CAVEAT EMPTOR: possibly in IRQ context
 166          * DO NOT record procfs stats here!!! */
 167
 168         if (unlikely(iobuf == NULL)) {
 169                 CERROR("***** bio->bi_private is NULL!  This should never "
 170                        "happen.  Normally, I would crash here, but instead I "
 171                        "will dump the bio contents to the console.  Please "
 172                        "report this to <https://jira.whamcloud.com/> , along "
 173                        "with any interesting messages leading up to this point "
 174                        "(like SCSI errors, perhaps).  Because bi_private is "
 175                        "NULL, I can't wake up the thread that initiated this "
 176                        "IO - you will probably have to reboot this node.\n");
 177                 CERROR("bi_next: %p, bi_flags: %lx, "
 178 #ifdef HAVE_BI_RW
 179                        "bi_rw: %lu,"
 180 #else
 181                        "bi_opf: %u,"
 182 #endif
 183                        "bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p,"
 184                        "bi_cnt: %d, bi_private: %p\n", bio->bi_next,
 185                         (unsigned long)bio->bi_flags,
 186 #ifdef HAVE_BI_RW
 187                         bio->bi_rw,
 188 #else
 189                         bio->bi_opf,
 190 #endif
 191                         bio->bi_vcnt, bio_idx(bio),
 192                         bio_sectors(bio) << 9, bio->bi_end_io,
 193 #ifdef HAVE_BI_CNT
 194                         atomic_read(&bio->bi_cnt),
 195 #else
 196                         atomic_read(&bio->__bi_cnt),
 197 #endif
 198                         bio->bi_private);
 199                 return;
 200         }
 201
 202         /* the check is outside of the cycle for performance reason -bzzz */
 203         if (!bio_data_dir(bio)) {
 204                 bio_for_each_segment_all(bvl, bio, iter) {
 205                         if (likely(error == 0))
 206                                 SetPageUptodate(bvl_to_page(bvl));
 207                         LASSERT(PageLocked(bvl_to_page(bvl)));
 208                 }
 209                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 210         } else {
 211                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 212         }
 213
 214         /* any real error is good enough -bzzz */
 215         if (error != 0 && iobuf->dr_error == 0)
 216                 iobuf->dr_error = error;
 217
 218         /*
 219          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 220          * it's possible that service thread will see dr_numreqs
 221          * is zero, but dr_elapsed is not set yet, leading to lost
 222          * data in this processing and an assertion in a subsequent
 223          * call to OSD.
 224          */
 225         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 226                 ktime_t now = ktime_get();
 227
 228                 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
 229                 iobuf->dr_elapsed_valid = 1;
 230         }
 231         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 232                 wake_up(&iobuf->dr_wait);
 233
 234         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 235          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 236          * mempool when serious on-disk fragmentation was encountered,
 237          * deadlocking the OST.  The bios are now released as soon as complete
 238          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 239         bio_put(bio);
 240 }
 241
 242 static void record_start_io(struct osd_iobuf *iobuf, int size)
 243 {
 244         struct osd_device    *osd = iobuf->dr_dev;
 245         struct obd_histogram *h = osd->od_brw_stats.hist;
 246
 247         iobuf->dr_frags++;
 248         atomic_inc(&iobuf->dr_numreqs);
 249
 250         if (iobuf->dr_rw == 0) {
 251                 atomic_inc(&osd->od_r_in_flight);
 252                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 253                                  atomic_read(&osd->od_r_in_flight));
 254                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 255         } else if (iobuf->dr_rw == 1) {
 256                 atomic_inc(&osd->od_w_in_flight);
 257                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 258                                  atomic_read(&osd->od_w_in_flight));
 259                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 260         } else {
 261                 LBUG();
 262         }
 263 }
 264
 265 static void osd_submit_bio(int rw, struct bio *bio)
 266 {
 267         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 268 #ifdef HAVE_SUBMIT_BIO_2ARGS
 269         if (rw == 0)
 270                 submit_bio(READ, bio);
 271         else
 272                 submit_bio(WRITE, bio);
 273 #else
 274         bio->bi_opf |= rw;
 275         submit_bio(bio);
 276 #endif
 277 }
 278
 279 static int can_be_merged(struct bio *bio, sector_t sector)
 280 {
 281         if (bio == NULL)
 282                 return 0;
 283
 284         return bio_end_sector(bio) == sector ? 1 : 0;
 285 }
 286
 287 /*
 288  * This function will change the data written, thus it should only be
 289  * used when checking data integrity feature
 290  */
 291 static void bio_integrity_fault_inject(struct bio *bio)
 292 {
 293         struct bio_vec *bvec;
 294         int i;
 295         void *kaddr;
 296         char *addr;
 297
 298         bio_for_each_segment_all(bvec, bio, i) {
 299                 struct page *page = bvec->bv_page;
 300
 301                 kaddr = kmap(page);
 302                 addr = kaddr;
 303                 *addr = ~(*addr);
 304                 kunmap(page);
 305                 break;
 306         }
 307 }
 308
 309 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 310                       struct osd_iobuf *iobuf)
 311 {
 312         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 313         struct page **pages = iobuf->dr_pages;
 314         int npages = iobuf->dr_npages;
 315         sector_t *blocks = iobuf->dr_blocks;
 316         int total_blocks = npages * blocks_per_page;
 317         int sector_bits = inode->i_sb->s_blocksize_bits - 9;
 318         unsigned int blocksize = inode->i_sb->s_blocksize;
 319         struct bio *bio = NULL;
 320         struct page *page;
 321         unsigned int page_offset;
 322         sector_t sector;
 323         int nblocks;
 324         int block_idx;
 325         int page_idx;
 326         int i;
 327         int rc = 0;
 328         bool fault_inject;
 329         DECLARE_PLUG(plug);
 330         ENTRY;
 331
 332         fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
 333         LASSERT(iobuf->dr_npages == npages);
 334
 335         osd_brw_stats_update(osd, iobuf);
 336         iobuf->dr_start_time = ktime_get();
 337
 338         blk_start_plug(&plug);
 339         for (page_idx = 0, block_idx = 0;
 340              page_idx < npages;
 341              page_idx++, block_idx += blocks_per_page) {
 342
 343                 page = pages[page_idx];
 344                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 345
 346                 for (i = 0, page_offset = 0;
 347                      i < blocks_per_page;
 348                      i += nblocks, page_offset += blocksize * nblocks) {
 349
 350                         nblocks = 1;
 351
 352                         if (blocks[block_idx + i] == 0) {  /* hole */
 353                                 LASSERTF(iobuf->dr_rw == 0,
 354                                          "page_idx %u, block_idx %u, i %u\n",
 355                                          page_idx, block_idx, i);
 356                                 memset(kmap(page) + page_offset, 0, blocksize);
 357                                 kunmap(page);
 358                                 continue;
 359                         }
 360
 361                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 362
 363                         /* Additional contiguous file blocks? */
 364                         while (i + nblocks < blocks_per_page &&
 365                                (sector + (nblocks << sector_bits)) ==
 366                                ((sector_t)blocks[block_idx + i + nblocks] <<
 367                                 sector_bits))
 368                                 nblocks++;
 369
 370                         if (bio != NULL &&
 371                             can_be_merged(bio, sector) &&
 372                             bio_add_page(bio, page,
 373                                          blocksize * nblocks, page_offset) != 0)
 374                                 continue;       /* added this frag OK */
 375
 376                         if (bio != NULL) {
 377                                 struct request_queue *q = bio_get_queue(bio);
 378                                 unsigned int bi_size = bio_sectors(bio) << 9;
 379
 380                                 /* Dang! I have to fragment this I/O */
 381                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 382                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 383                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 384                                        bio_sectors(bio),
 385                                        queue_max_sectors(q),
 386                                        bio_phys_segments(q, bio),
 387                                        queue_max_phys_segments(q),
 388                                        0, queue_max_hw_segments(q));
 389                                 if (bio_integrity_enabled(bio)) {
 390                                         if (bio_integrity_prep(bio)) {
 391                                                 bio_put(bio);
 392                                                 rc = -EIO;
 393                                                 goto out;
 394                                         }
 395                                         if (unlikely(fault_inject))
 396                                                 bio_integrity_fault_inject(bio);
 397                                 }
 398
 399                                 record_start_io(iobuf, bi_size);
 400                                 osd_submit_bio(iobuf->dr_rw, bio);
 401                         }
 402
 403                         /* allocate new bio */
 404                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 405                                                       (npages - page_idx) *
 406                                                       blocks_per_page));
 407                         if (bio == NULL) {
 408                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 409                                        (npages - page_idx), blocks_per_page,
 410                                        (npages - page_idx) * blocks_per_page);
 411                                 rc = -ENOMEM;
 412                                 goto out;
 413                         }
 414
 415                         bio_set_dev(bio, inode->i_sb->s_bdev);
 416                         bio_set_sector(bio, sector);
 417 #ifdef HAVE_BI_RW
 418                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 419 #else
 420                         bio->bi_opf = (iobuf->dr_rw == 0) ? READ : WRITE;
 421 #endif
 422                         bio->bi_end_io = dio_complete_routine;
 423                         bio->bi_private = iobuf;
 424
 425                         rc = bio_add_page(bio, page,
 426                                           blocksize * nblocks, page_offset);
 427                         LASSERT(rc != 0);
 428                 }
 429         }
 430
 431         if (bio != NULL) {
 432                 if (bio_integrity_enabled(bio)) {
 433                         if (bio_integrity_prep(bio)) {
 434                                 bio_put(bio);
 435                                 rc = -EIO;
 436                                 goto out;
 437                         }
 438                         if (unlikely(fault_inject))
 439                                 bio_integrity_fault_inject(bio);
 440                 }
 441
 442                 record_start_io(iobuf, bio_sectors(bio) << 9);
 443                 osd_submit_bio(iobuf->dr_rw, bio);
 444                 rc = 0;
 445         }
 446
 447 out:
 448         blk_finish_plug(&plug);
 449
 450         /* in order to achieve better IO throughput, we don't wait for writes
 451          * completion here. instead we proceed with transaction commit in
 452          * parallel and wait for IO completion once transaction is stopped
 453          * see osd_trans_stop() for more details -bzzz */
 454         if (iobuf->dr_rw == 0 || fault_inject) {
 455                 wait_event(iobuf->dr_wait,
 456                            atomic_read(&iobuf->dr_numreqs) == 0);
 457                 osd_fini_iobuf(osd, iobuf);
 458         }
 459
 460         if (rc == 0)
 461                 rc = iobuf->dr_error;
 462         RETURN(rc);
 463 }
 464
 465 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 466                                    struct niobuf_local *lnb)
 467 {
 468         ENTRY;
 469
 470         *nrpages = 0;
 471
 472         while (len > 0) {
 473                 int poff = offset & (PAGE_SIZE - 1);
 474                 int plen = PAGE_SIZE - poff;
 475
 476                 if (plen > len)
 477                         plen = len;
 478                 lnb->lnb_file_offset = offset;
 479                 lnb->lnb_page_offset = poff;
 480                 lnb->lnb_len = plen;
 481                 /* lnb->lnb_flags = rnb->rnb_flags; */
 482                 lnb->lnb_flags = 0;
 483                 lnb->lnb_page = NULL;
 484                 lnb->lnb_rc = 0;
 485
 486                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 487                          (long long) len);
 488                 offset += plen;
 489                 len -= plen;
 490                 lnb++;
 491                 (*nrpages)++;
 492         }
 493
 494         RETURN(0);
 495 }
 496
 497 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
 498                                  loff_t offset, gfp_t gfp_mask)
 499 {
 500         struct osd_thread_info *oti = osd_oti_get(env);
 501         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 502         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 503         struct page *page;
 504         int cur = oti->oti_dio_pages_used;
 505
 506         LASSERT(inode);
 507
 508         if (osd_use_page_cache(d)) {
 509                 page = find_or_create_page(inode->i_mapping,
 510                                            offset >> PAGE_SHIFT,
 511                                            gfp_mask);
 512
 513                 if (likely(page))
 514                         LASSERT(!test_bit(PG_private_2, &page->flags));
 515                 else
 516                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 517         } else {
 518
 519                 LASSERT(oti->oti_dio_pages);
 520
 521                 if (unlikely(!oti->oti_dio_pages[cur])) {
 522                         LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
 523                         page = alloc_page(gfp_mask);
 524                         if (!page)
 525                                 return NULL;
 526                         oti->oti_dio_pages[cur] = page;
 527                 }
 528
 529                 page = oti->oti_dio_pages[cur];
 530                 LASSERT(!test_bit(PG_private_2, &page->flags));
 531                 set_bit(PG_private_2, &page->flags);
 532                 oti->oti_dio_pages_used++;
 533
 534                 LASSERT(!PageLocked(page));
 535                 lock_page(page);
 536
 537                 LASSERT(!page->mapping);
 538                 LASSERT(!PageWriteback(page));
 539                 ClearPageUptodate(page);
 540
 541                 page->index = offset >> PAGE_SHIFT;
 542         }
 543
 544         return page;
 545 }
 546
 547 /*
 548  * there are following "locks":
 549  * journal_start
 550  * i_mutex
 551  * page lock
 552  *
 553  * osd write path:
 554  *  - lock page(s)
 555  *  - journal_start
 556  *  - truncate_sem
 557  *
 558  * ext4 vmtruncate:
 559  *  - lock pages, unlock
 560  *  - journal_start
 561  *  - lock partial page
 562  *  - i_data_sem
 563  *
 564  */
 565
 566 /**
 567  * Unlock and release pages loaded by osd_bufs_get()
 568  *
 569  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 570  *
 571  * \param env           thread execution environment
 572  * \param dt            dt object undergoing IO (OSD object + methods)
 573  * \param lnb           array of pages undergoing IO
 574  * \param npages        number of pages in \a lnb
 575  *
 576  * \retval 0            always
 577  */
 578 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 579                         struct niobuf_local *lnb, int npages)
 580 {
 581         struct osd_thread_info *oti = osd_oti_get(env);
 582         struct pagevec pvec;
 583         int i;
 584
 585 #ifdef HAVE_PAGEVEC_INIT_ONE_PARAM
 586         pagevec_init(&pvec);
 587 #else
 588         pagevec_init(&pvec, 0);
 589 #endif
 590
 591         for (i = 0; i < npages; i++) {
 592                 struct page *page = lnb[i].lnb_page;
 593
 594                 if (page == NULL)
 595                         continue;
 596                 LASSERT(PageLocked(page));
 597
 598                 /* if the page isn't cached, then reset uptodate
 599                  * to prevent reuse */
 600                 if (test_bit(PG_private_2, &page->flags)) {
 601                         clear_bit(PG_private_2, &page->flags);
 602                         ClearPageUptodate(page);
 603                         unlock_page(page);
 604                         oti->oti_dio_pages_used--;
 605                 } else {
 606                         unlock_page(page);
 607                         if (pagevec_add(&pvec, page) == 0)
 608                                 pagevec_release(&pvec);
 609                 }
 610                 dt_object_put(env, dt);
 611
 612                 lnb[i].lnb_page = NULL;
 613         }
 614
 615         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
 616
 617         /* Release any partial pagevec */
 618         pagevec_release(&pvec);
 619
 620         RETURN(0);
 621 }
 622
 623 /**
 624  * Load and lock pages undergoing IO
 625  *
 626  * Pages as described in the \a lnb array are fetched (from disk or cache)
 627  * and locked for IO by the caller.
 628  *
 629  * DLM locking protects us from write and truncate competing for same region,
 630  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 631  * It's possible the writeout on a such a page is in progress when we access
 632  * it. It's also possible that during this writeout we put new (partial) data
 633  * into the page, but won't be able to proceed in filter_commitrw_write().
 634  * Therefore, just wait for writeout completion as it should be rare enough.
 635  *
 636  * \param env           thread execution environment
 637  * \param dt            dt object undergoing IO (OSD object + methods)
 638  * \param pos           byte offset of IO start
 639  * \param len           number of bytes of IO
 640  * \param lnb           array of extents undergoing IO
 641  * \param rw            read or write operation, and other flags
 642  * \param capa          capabilities
 643  *
 644  * \retval pages        (zero or more) loaded successfully
 645  * \retval -ENOMEM      on memory/page allocation error
 646  */
 647 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 648                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 649                         enum dt_bufs_type rw)
 650 {
 651         struct osd_thread_info *oti = osd_oti_get(env);
 652         struct osd_object *obj = osd_dt_obj(dt);
 653         int npages, i, rc = 0;
 654         gfp_t gfp_mask;
 655
 656         LASSERT(obj->oo_inode);
 657
 658         if (!osd_use_page_cache(osd_obj2dev(obj))) {
 659                 if (unlikely(!oti->oti_dio_pages)) {
 660                         OBD_ALLOC(oti->oti_dio_pages,
 661                                   sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
 662                         if (!oti->oti_dio_pages)
 663                                 return -ENOMEM;
 664                 }
 665         }
 666
 667         osd_map_remote_to_local(pos, len, &npages, lnb);
 668
 669         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 670         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 671                                              GFP_HIGHUSER;
 672         for (i = 0; i < npages; i++, lnb++) {
 673                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
 674                                              gfp_mask);
 675                 if (lnb->lnb_page == NULL)
 676                         GOTO(cleanup, rc = -ENOMEM);
 677
 678                 wait_on_page_writeback(lnb->lnb_page);
 679                 BUG_ON(PageWriteback(lnb->lnb_page));
 680
 681                 lu_object_get(&dt->do_lu);
 682         }
 683
 684         RETURN(i);
 685
 686 cleanup:
 687         if (i > 0)
 688                 osd_bufs_put(env, dt, lnb - i, i);
 689         return rc;
 690 }
 691
 692 #ifndef HAVE_LDISKFS_MAP_BLOCKS
 693
 694 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 695 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 696 #endif
 697
 698 struct bpointers {
 699         sector_t *blocks;
 700         unsigned long start;
 701         int num;
 702         int init_num;
 703         int create;
 704 };
 705
 706 static long ldiskfs_ext_find_goal(struct inode *inode,
 707                                   struct ldiskfs_ext_path *path,
 708                                   unsigned long block, int *aflags)
 709 {
 710         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 711         unsigned long bg_start;
 712         unsigned long colour;
 713         int depth;
 714
 715         if (path) {
 716                 struct ldiskfs_extent *ex;
 717                 depth = path->p_depth;
 718
 719                 /* try to predict block placement */
 720                 if ((ex = path[depth].p_ext))
 721                         return ldiskfs_ext_pblock(ex) +
 722                                 (block - le32_to_cpu(ex->ee_block));
 723
 724                 /* it looks index is empty
 725                  * try to find starting from index itself */
 726                 if (path[depth].p_bh)
 727                         return path[depth].p_bh->b_blocknr;
 728         }
 729
 730         /* OK. use inode's group */
 731         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 732                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 733         colour = (current->pid % 16) *
 734                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 735         return bg_start + colour + block;
 736 }
 737
 738 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 739                                 struct ldiskfs_ext_path *path,
 740                                 unsigned long block, unsigned long *count,
 741                                 int *err)
 742 {
 743         struct ldiskfs_allocation_request ar;
 744         unsigned long pblock;
 745         int aflags;
 746
 747         /* find neighbour allocated blocks */
 748         ar.lleft = block;
 749         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 750         if (*err)
 751                 return 0;
 752         ar.lright = block;
 753         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 754         if (*err)
 755                 return 0;
 756
 757         /* allocate new block */
 758         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 759         ar.inode = inode;
 760         ar.logical = block;
 761         ar.len = *count;
 762         ar.flags = LDISKFS_MB_HINT_DATA;
 763         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 764         *count = ar.len;
 765         return pblock;
 766 }
 767
 768 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 769                                      struct ldiskfs_ext_path *path,
 770                                      struct ldiskfs_ext_cache *cex,
 771 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 772                                      struct ldiskfs_extent *ex,
 773 #endif
 774                                      void *cbdata)
 775 {
 776         struct bpointers *bp = cbdata;
 777         struct ldiskfs_extent nex;
 778         unsigned long pblock = 0;
 779         unsigned long tgen;
 780         int err, i;
 781         unsigned long count;
 782         handle_t *handle;
 783
 784 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 785         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 786 #else
 787         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 788 #endif
 789                 err = EXT_CONTINUE;
 790                 goto map;
 791         }
 792
 793         if (bp->create == 0) {
 794                 i = 0;
 795                 if (cex->ec_block < bp->start)
 796                         i = bp->start - cex->ec_block;
 797                 if (i >= cex->ec_len)
 798                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 799                                         i, cex->ec_len);
 800                 for (; i < cex->ec_len && bp->num; i++) {
 801                         *(bp->blocks) = 0;
 802                         bp->blocks++;
 803                         bp->num--;
 804                         bp->start++;
 805                 }
 806
 807                 return EXT_CONTINUE;
 808         }
 809
 810         tgen = LDISKFS_I(inode)->i_ext_generation;
 811         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
 812
 813         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
 814                                    count + LDISKFS_ALLOC_NEEDED + 1);
 815         if (IS_ERR(handle)) {
 816                 return PTR_ERR(handle);
 817         }
 818
 819         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
 820                 /* the tree has changed. so path can be invalid at moment */
 821                 ldiskfs_journal_stop(handle);
 822                 return EXT_REPEAT;
 823         }
 824
 825         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
 826          * protected by i_data_sem as whole. so we patch it to store
 827          * generation to path and now verify the tree hasn't changed */
 828         down_write((&LDISKFS_I(inode)->i_data_sem));
 829
 830         /* validate extent, make sure the extent tree does not changed */
 831         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
 832                 /* cex is invalid, try again */
 833                 up_write(&LDISKFS_I(inode)->i_data_sem);
 834                 ldiskfs_journal_stop(handle);
 835                 return EXT_REPEAT;
 836         }
 837
 838         count = cex->ec_len;
 839         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
 840         if (!pblock)
 841                 goto out;
 842         BUG_ON(count > cex->ec_len);
 843
 844         /* insert new extent */
 845         nex.ee_block = cpu_to_le32(cex->ec_block);
 846         ldiskfs_ext_store_pblock(&nex, pblock);
 847         nex.ee_len = cpu_to_le16(count);
 848         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
 849         if (err) {
 850                 /* free data blocks we just allocated */
 851                 /* not a good idea to call discard here directly,
 852                  * but otherwise we'd need to call it every free() */
 853                 ldiskfs_discard_preallocations(inode);
 854 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
 855                 ldiskfs_free_blocks(handle, inode, NULL,
 856                                     ldiskfs_ext_pblock(&nex),
 857                                     le16_to_cpu(nex.ee_len), 0);
 858 #else
 859                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
 860                                     le16_to_cpu(nex.ee_len), 0);
 861 #endif
 862                 goto out;
 863         }
 864
 865         /*
 866          * Putting len of the actual extent we just inserted,
 867          * we are asking ldiskfs_ext_walk_space() to continue
 868          * scaning after that block
 869          */
 870         cex->ec_len = le16_to_cpu(nex.ee_len);
 871         cex->ec_start = ldiskfs_ext_pblock(&nex);
 872         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
 873         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
 874
 875 out:
 876         up_write((&LDISKFS_I(inode)->i_data_sem));
 877         ldiskfs_journal_stop(handle);
 878 map:
 879         if (err >= 0) {
 880                 /* map blocks */
 881                 if (bp->num == 0) {
 882                         CERROR("hmm. why do we find this extent?\n");
 883                         CERROR("initial space: %lu:%u\n",
 884                                 bp->start, bp->init_num);
 885 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 886                         CERROR("current extent: %u/%u/%llu %d\n",
 887                                 cex->ec_block, cex->ec_len,
 888                                 (unsigned long long)cex->ec_start,
 889                                 cex->ec_type);
 890 #else
 891                         CERROR("current extent: %u/%u/%llu\n",
 892                                 cex->ec_block, cex->ec_len,
 893                                 (unsigned long long)cex->ec_start);
 894 #endif
 895                 }
 896                 i = 0;
 897                 if (cex->ec_block < bp->start)
 898                         i = bp->start - cex->ec_block;
 899                 if (i >= cex->ec_len)
 900                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 901                                         i, cex->ec_len);
 902                 for (; i < cex->ec_len && bp->num; i++) {
 903                         *(bp->blocks) = cex->ec_start + i;
 904                         if (pblock != 0) {
 905                                 /* unmap any possible underlying metadata from
 906                                  * the block device mapping.  bug 6998. */
 907 #ifndef HAVE_CLEAN_BDEV_ALIASES
 908                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
 909                                                           *(bp->blocks));
 910 #else
 911                                 clean_bdev_aliases(inode->i_sb->s_bdev,
 912                                                    *(bp->blocks), 1);
 913 #endif
 914                         }
 915                         bp->blocks++;
 916                         bp->num--;
 917                         bp->start++;
 918                 }
 919         }
 920         return err;
 921 }
 922
 923 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
 924                                    int clen, sector_t *blocks, int create)
 925 {
 926         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 927         struct bpointers bp;
 928         int err;
 929
 930         if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
 931                 return -EFBIG;
 932
 933         bp.blocks = blocks;
 934         bp.start = index * blocks_per_page;
 935         bp.init_num = bp.num = clen * blocks_per_page;
 936         bp.create = create;
 937
 938         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 939                bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
 940
 941         err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
 942                                      ldiskfs_ext_new_extent_cb, &bp);
 943         ldiskfs_ext_invalidate_cache(inode);
 944
 945         return err;
 946 }
 947
 948 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
 949                                           struct page **page, int pages,
 950                                           sector_t *blocks, int create)
 951 {
 952         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 953         pgoff_t bitmap_max_page_index;
 954         sector_t *b;
 955         int rc = 0, i;
 956
 957         bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
 958                                 PAGE_SHIFT;
 959         for (i = 0, b = blocks; i < pages; i++, page++) {
 960                 if ((*page)->index + 1 >= bitmap_max_page_index) {
 961                         rc = -EFBIG;
 962                         break;
 963                 }
 964                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
 965                 if (rc) {
 966                         CERROR("ino %lu, blk %llu create %d: rc %d\n",
 967                                inode->i_ino,
 968                                (unsigned long long)*b, create, rc);
 969                         break;
 970                 }
 971                 b += blocks_per_page;
 972         }
 973         return rc;
 974 }
 975
 976 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
 977                                            struct page **page,
 978                                            int pages, sector_t *blocks,
 979                                            int create)
 980 {
 981         int rc = 0, i = 0, clen = 0;
 982         struct page *fp = NULL;
 983
 984         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 985                 inode->i_ino, pages, (*page)->index);
 986
 987         /* pages are sorted already. so, we just have to find
 988          * contig. space and process them properly */
 989         while (i < pages) {
 990                 if (fp == NULL) {
 991                         /* start new extent */
 992                         fp = *page++;
 993                         clen = 1;
 994                         i++;
 995                         continue;
 996                 } else if (fp->index + clen == (*page)->index) {
 997                         /* continue the extent */
 998                         page++;
 999                         clen++;
1000                         i++;
1001                         continue;
1002                 }
1003
1004                 /* process found extent */
1005                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
1006                                              blocks, create);
1007                 if (rc)
1008                         GOTO(cleanup, rc);
1009
1010                 /* look for next extent */
1011                 fp = NULL;
1012                 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
1013         }
1014
1015         if (fp)
1016                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
1017                                              blocks, create);
1018
1019 cleanup:
1020         return rc;
1021 }
1022
1023 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
1024                                        int pages, sector_t *blocks,
1025                                        int create)
1026 {
1027         int rc;
1028
1029         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1030                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
1031                                                      blocks, create);
1032                 return rc;
1033         }
1034         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
1035
1036         return rc;
1037 }
1038 #else
1039 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
1040                                        int pages, sector_t *blocks,
1041                                        int create)
1042 {
1043         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1044         int rc = 0, i = 0;
1045         struct page *fp = NULL;
1046         int clen = 0;
1047         pgoff_t max_page_index;
1048         handle_t *handle = NULL;
1049
1050         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
1051
1052         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1053                 inode->i_ino, pages, (*page)->index);
1054
1055         if (create) {
1056                 create = LDISKFS_GET_BLOCKS_CREATE;
1057                 handle = ldiskfs_journal_current_handle();
1058                 LASSERT(handle != NULL);
1059                 rc = osd_attach_jinode(inode);
1060                 if (rc)
1061                         return rc;
1062         }
1063         /* pages are sorted already. so, we just have to find
1064          * contig. space and process them properly */
1065         while (i < pages) {
1066                 long blen, total = 0;
1067                 struct ldiskfs_map_blocks map = { 0 };
1068
1069                 if (fp == NULL) { /* start new extent */
1070                         fp = *page++;
1071                         clen = 1;
1072                         if (++i != pages)
1073                                 continue;
1074                 } else if (fp->index + clen == (*page)->index) {
1075                         /* continue the extent */
1076                         page++;
1077                         clen++;
1078                         if (++i != pages)
1079                                 continue;
1080                 }
1081                 if (fp->index + clen >= max_page_index)
1082                         GOTO(cleanup, rc = -EFBIG);
1083                 /* process found extent */
1084                 map.m_lblk = fp->index * blocks_per_page;
1085                 map.m_len = blen = clen * blocks_per_page;
1086 cont_map:
1087                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
1088                 if (rc >= 0) {
1089                         int c = 0;
1090                         for (; total < blen && c < map.m_len; c++, total++) {
1091                                 if (rc == 0) {
1092                                         *(blocks + total) = 0;
1093                                         total++;
1094                                         break;
1095                                 } else {
1096                                         *(blocks + total) = map.m_pblk + c;
1097                                         /* unmap any possible underlying
1098                                          * metadata from the block device
1099                                          * mapping.  bug 6998. */
1100                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
1101                                             create)
1102 #ifndef HAVE_CLEAN_BDEV_ALIASES
1103                                                 unmap_underlying_metadata(
1104                                                         inode->i_sb->s_bdev,
1105                                                         map.m_pblk + c);
1106 #else
1107                                                 clean_bdev_aliases(
1108                                                         inode->i_sb->s_bdev,
1109                                                         map.m_pblk + c, 1);
1110 #endif
1111                                 }
1112                         }
1113                         rc = 0;
1114                 }
1115                 if (rc == 0 && total < blen) {
1116                         map.m_lblk = fp->index * blocks_per_page + total;
1117                         map.m_len = blen - total;
1118                         goto cont_map;
1119                 }
1120                 if (rc != 0)
1121                         GOTO(cleanup, rc);
1122
1123                 /* look for next extent */
1124                 fp = NULL;
1125                 blocks += blocks_per_page * clen;
1126         }
1127 cleanup:
1128         return rc;
1129 }
1130 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
1131
1132 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1133                           struct niobuf_local *lnb, int npages)
1134 {
1135         struct osd_thread_info *oti   = osd_oti_get(env);
1136         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1137         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
1138         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
1139         ktime_t start;
1140         ktime_t end;
1141         s64 timediff;
1142         ssize_t                 isize;
1143         __s64                   maxidx;
1144         int                     rc = 0;
1145         int                     i;
1146         int                     cache = 0;
1147
1148         LASSERT(inode);
1149
1150         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1151         if (unlikely(rc != 0))
1152                 RETURN(rc);
1153
1154         isize = i_size_read(inode);
1155         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1156
1157         if (osd->od_writethrough_cache)
1158                 cache = 1;
1159         if (isize > osd->od_readcache_max_filesize)
1160                 cache = 0;
1161
1162         start = ktime_get();
1163         for (i = 0; i < npages; i++) {
1164
1165                 if (cache == 0)
1166                         generic_error_remove_page(inode->i_mapping,
1167                                                   lnb[i].lnb_page);
1168
1169                 /*
1170                  * till commit the content of the page is undefined
1171                  * we'll set it uptodate once bulk is done. otherwise
1172                  * subsequent reads can access non-stable data
1173                  */
1174                 ClearPageUptodate(lnb[i].lnb_page);
1175
1176                 if (lnb[i].lnb_len == PAGE_SIZE)
1177                         continue;
1178
1179                 if (maxidx >= lnb[i].lnb_page->index) {
1180                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1181                 } else {
1182                         long off;
1183                         char *p = kmap(lnb[i].lnb_page);
1184
1185                         off = lnb[i].lnb_page_offset;
1186                         if (off)
1187                                 memset(p, 0, off);
1188                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1189                               ~PAGE_MASK;
1190                         if (off)
1191                                 memset(p + off, 0, PAGE_SIZE - off);
1192                         kunmap(lnb[i].lnb_page);
1193                 }
1194         }
1195         end = ktime_get();
1196         timediff = ktime_us_delta(end, start);
1197         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1198
1199         if (iobuf->dr_npages) {
1200                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1201                                                  iobuf->dr_npages,
1202                                                  iobuf->dr_blocks, 0);
1203                 if (likely(rc == 0)) {
1204                         rc = osd_do_bio(osd, inode, iobuf);
1205                         /* do IO stats for preparation reads */
1206                         osd_fini_iobuf(osd, iobuf);
1207                 }
1208         }
1209         RETURN(rc);
1210 }
1211
1212 struct osd_fextent {
1213         sector_t        start;
1214         sector_t        end;
1215         unsigned int    mapped:1;
1216 };
1217
1218 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1219                          struct osd_fextent *cached_extent)
1220 {
1221         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1222         sector_t block = offset >> inode->i_blkbits;
1223         sector_t start;
1224         struct fiemap_extent_info fei = { 0 };
1225         struct fiemap_extent fe = { 0 };
1226         mm_segment_t saved_fs;
1227         int rc;
1228
1229         if (block >= cached_extent->start && block < cached_extent->end)
1230                 return cached_extent->mapped;
1231
1232         if (i_size_read(inode) == 0)
1233                 return 0;
1234
1235         /* Beyond EOF, must not be mapped */
1236         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1237                 return 0;
1238
1239         fei.fi_extents_max = 1;
1240         fei.fi_extents_start = &fe;
1241
1242         saved_fs = get_fs();
1243         set_fs(get_ds());
1244         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1245         set_fs(saved_fs);
1246         if (rc != 0)
1247                 return 0;
1248
1249         start = fe.fe_logical >> inode->i_blkbits;
1250
1251         if (start > block) {
1252                 cached_extent->start = block;
1253                 cached_extent->end = start;
1254                 cached_extent->mapped = 0;
1255         } else {
1256                 cached_extent->start = start;
1257                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1258                                       inode->i_blkbits;
1259                 cached_extent->mapped = 1;
1260         }
1261
1262         return cached_extent->mapped;
1263 }
1264
1265 static int osd_declare_write_commit(const struct lu_env *env,
1266                                     struct dt_object *dt,
1267                                     struct niobuf_local *lnb, int npages,
1268                                     struct thandle *handle)
1269 {
1270         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1271         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1272         struct osd_thandle      *oh;
1273         int                     extents = 1;
1274         int                     depth;
1275         int                     i;
1276         int                     newblocks;
1277         int                     rc = 0;
1278         int                     flags = 0;
1279         int                     credits = 0;
1280         long long               quota_space = 0;
1281         struct osd_fextent      extent = { 0 };
1282         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1283         ENTRY;
1284
1285         LASSERT(handle != NULL);
1286         oh = container_of0(handle, struct osd_thandle, ot_super);
1287         LASSERT(oh->ot_handle == NULL);
1288
1289         newblocks = npages;
1290
1291         /* calculate number of extents (probably better to pass nb) */
1292         for (i = 0; i < npages; i++) {
1293                 if (i && lnb[i].lnb_file_offset !=
1294                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1295                         extents++;
1296
1297                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1298                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1299                 else
1300                         quota_space += PAGE_SIZE;
1301
1302                 /* ignore quota for the whole request if any page is from
1303                  * client cache or written by root.
1304                  *
1305                  * XXX once we drop the 1.8 client support, the checking
1306                  * for whether page is from cache can be simplified as:
1307                  * !(lnb[i].flags & OBD_BRW_SYNC)
1308                  *
1309                  * XXX we could handle this on per-lnb basis as done by
1310                  * grant. */
1311                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1312                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1313                     OBD_BRW_FROM_GRANT)
1314                         declare_flags |= OSD_QID_FORCE;
1315         }
1316
1317         /*
1318          * each extent can go into new leaf causing a split
1319          * 5 is max tree depth: inode + 4 index blocks
1320          * with blockmaps, depth is 3 at most
1321          */
1322         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1323                 /*
1324                  * many concurrent threads may grow tree by the time
1325                  * our transaction starts. so, consider 2 is a min depth
1326                  */
1327                 depth = ext_depth(inode);
1328                 depth = max(depth, 1) + 1;
1329                 newblocks += depth;
1330                 credits++; /* inode */
1331                 credits += depth * 2 * extents;
1332         } else {
1333                 depth = 3;
1334                 newblocks += depth;
1335                 credits++; /* inode */
1336                 credits += depth * extents;
1337         }
1338
1339         /* quota space for metadata blocks */
1340         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1341
1342         /* quota space should be reported in 1K blocks */
1343         quota_space = toqb(quota_space);
1344
1345         /* each new block can go in different group (bitmap + gd) */
1346
1347         /* we can't dirty more bitmap blocks than exist */
1348         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1349                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1350         else
1351                 credits += newblocks;
1352
1353         /* we can't dirty more gd blocks than exist */
1354         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1355                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1356         else
1357                 credits += newblocks;
1358
1359         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1360
1361         /* make sure the over quota flags were not set */
1362         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1363
1364         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1365                                    i_projid_read(inode), quota_space, oh,
1366                                    osd_dt_obj(dt), &flags, declare_flags);
1367
1368         /* we need only to store the overquota flags in the first lnb for
1369          * now, once we support multiple objects BRW, this code needs be
1370          * revised. */
1371         if (flags & QUOTA_FL_OVER_USRQUOTA)
1372                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1373         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1374                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1375         if (flags & QUOTA_FL_OVER_PRJQUOTA)
1376                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1377
1378         if (rc == 0)
1379                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1380
1381         RETURN(rc);
1382 }
1383
1384 /* Check if a block is allocated or not */
1385 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1386                             struct niobuf_local *lnb, int npages,
1387                             struct thandle *thandle)
1388 {
1389         struct osd_thread_info *oti = osd_oti_get(env);
1390         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1391         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1392         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1393         loff_t isize;
1394         int rc = 0, i;
1395
1396         LASSERT(inode);
1397
1398         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1399         if (unlikely(rc != 0))
1400                 RETURN(rc);
1401
1402         isize = i_size_read(inode);
1403         ll_vfs_dq_init(inode);
1404
1405         for (i = 0; i < npages; i++) {
1406                 if (lnb[i].lnb_rc == -ENOSPC &&
1407                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1408                         /* Allow the write to proceed if overwriting an
1409                          * existing block */
1410                         lnb[i].lnb_rc = 0;
1411                 }
1412
1413                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1414                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1415                                lnb[i].lnb_rc);
1416                         LASSERT(lnb[i].lnb_page);
1417                         generic_error_remove_page(inode->i_mapping,
1418                                                   lnb[i].lnb_page);
1419                         continue;
1420                 }
1421
1422                 LASSERT(PageLocked(lnb[i].lnb_page));
1423                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1424
1425                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1426                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1427
1428                 /*
1429                  * Since write and truncate are serialized by oo_sem, even
1430                  * partial-page truncate should not leave dirty pages in the
1431                  * page cache.
1432                  */
1433                 LASSERT(!PageDirty(lnb[i].lnb_page));
1434
1435                 SetPageUptodate(lnb[i].lnb_page);
1436
1437                 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1438         }
1439
1440         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1441
1442         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1443                 rc = -ENOSPC;
1444         } else if (iobuf->dr_npages > 0) {
1445                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1446                                                  iobuf->dr_npages,
1447                                                  iobuf->dr_blocks, 1);
1448         } else {
1449                 /* no pages to write, no transno is needed */
1450                 thandle->th_local = 1;
1451         }
1452
1453         if (likely(rc == 0)) {
1454                 spin_lock(&inode->i_lock);
1455                 if (isize > i_size_read(inode)) {
1456                         i_size_write(inode, isize);
1457                         LDISKFS_I(inode)->i_disksize = isize;
1458                         spin_unlock(&inode->i_lock);
1459                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1460                 } else {
1461                         spin_unlock(&inode->i_lock);
1462                 }
1463
1464                 rc = osd_do_bio(osd, inode, iobuf);
1465                 /* we don't do stats here as in read path because
1466                  * write is async: we'll do this in osd_put_bufs() */
1467         } else {
1468                 osd_fini_iobuf(osd, iobuf);
1469         }
1470
1471         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1472
1473         if (unlikely(rc != 0)) {
1474                 /* if write fails, we should drop pages from the cache */
1475                 for (i = 0; i < npages; i++) {
1476                         if (lnb[i].lnb_page == NULL)
1477                                 continue;
1478                         LASSERT(PageLocked(lnb[i].lnb_page));
1479                         generic_error_remove_page(inode->i_mapping,
1480                                                   lnb[i].lnb_page);
1481                 }
1482         }
1483
1484         RETURN(rc);
1485 }
1486
1487 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1488                          struct niobuf_local *lnb, int npages)
1489 {
1490         struct osd_thread_info *oti = osd_oti_get(env);
1491         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1492         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1493         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1494         int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1495         ktime_t start, end;
1496         s64 timediff;
1497         loff_t isize;
1498
1499         LASSERT(inode);
1500
1501         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1502         if (unlikely(rc != 0))
1503                 RETURN(rc);
1504
1505         isize = i_size_read(inode);
1506
1507         if (osd->od_read_cache)
1508                 cache = 1;
1509         if (isize > osd->od_readcache_max_filesize)
1510                 cache = 0;
1511
1512         start = ktime_get();
1513         for (i = 0; i < npages; i++) {
1514
1515                 if (isize <= lnb[i].lnb_file_offset)
1516                         /* If there's no more data, abort early.
1517                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1518                         break;
1519
1520                 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1521                         lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1522                 else
1523                         lnb[i].lnb_rc = lnb[i].lnb_len;
1524
1525                 /* Bypass disk read if fail_loc is set properly */
1526                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1527                         SetPageUptodate(lnb[i].lnb_page);
1528
1529                 if (PageUptodate(lnb[i].lnb_page)) {
1530                         cache_hits++;
1531                 } else {
1532                         cache_misses++;
1533                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1534                 }
1535
1536                 if (cache == 0)
1537                         generic_error_remove_page(inode->i_mapping,
1538                                                   lnb[i].lnb_page);
1539         }
1540         end = ktime_get();
1541         timediff = ktime_us_delta(end, start);
1542         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1543
1544         if (cache_hits != 0)
1545                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1546                                     cache_hits);
1547         if (cache_misses != 0)
1548                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1549                                     cache_misses);
1550         if (cache_hits + cache_misses != 0)
1551                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1552                                     cache_hits + cache_misses);
1553
1554         if (iobuf->dr_npages) {
1555                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1556                                                  iobuf->dr_npages,
1557                                                  iobuf->dr_blocks, 0);
1558                 rc = osd_do_bio(osd, inode, iobuf);
1559
1560                 /* IO stats will be done in osd_bufs_put() */
1561         }
1562
1563         RETURN(rc);
1564 }
1565
1566 /*
1567  * XXX: Another layering violation for now.
1568  *
1569  * We don't want to use ->f_op->read methods, because generic file write
1570  *
1571  *         - serializes on ->i_sem, and
1572  *
1573  *         - does a lot of extra work like balance_dirty_pages(),
1574  *
1575  * which doesn't work for globally shared files like /last_rcvd.
1576  */
1577 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1578 {
1579         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1580
1581         memcpy(buffer, (char *)ei->i_data, buflen);
1582
1583         return  buflen;
1584 }
1585
1586 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1587 {
1588         struct buffer_head *bh;
1589         unsigned long block;
1590         int osize;
1591         int blocksize;
1592         int csize;
1593         int boffs;
1594
1595         /* prevent reading after eof */
1596         spin_lock(&inode->i_lock);
1597         if (i_size_read(inode) < *offs + size) {
1598                 loff_t diff = i_size_read(inode) - *offs;
1599                 spin_unlock(&inode->i_lock);
1600                 if (diff < 0) {
1601                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1602                                i_size_read(inode), *offs);
1603                         return -EBADR;
1604                 } else if (diff == 0) {
1605                         return 0;
1606                 } else {
1607                         size = diff;
1608                 }
1609         } else {
1610                 spin_unlock(&inode->i_lock);
1611         }
1612
1613         blocksize = 1 << inode->i_blkbits;
1614         osize = size;
1615         while (size > 0) {
1616                 block = *offs >> inode->i_blkbits;
1617                 boffs = *offs & (blocksize - 1);
1618                 csize = min(blocksize - boffs, size);
1619                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1620                 if (IS_ERR(bh)) {
1621                         CERROR("%s: can't read %u@%llu on ino %lu: "
1622                                "rc = %ld\n", osd_ino2name(inode),
1623                                csize, *offs, inode->i_ino,
1624                                PTR_ERR(bh));
1625                         return PTR_ERR(bh);
1626                 }
1627
1628                 if (bh != NULL) {
1629                         memcpy(buf, bh->b_data + boffs, csize);
1630                         brelse(bh);
1631                 } else {
1632                         memset(buf, 0, csize);
1633                 }
1634
1635                 *offs += csize;
1636                 buf += csize;
1637                 size -= csize;
1638         }
1639         return osize;
1640 }
1641
1642 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1643                         struct lu_buf *buf, loff_t *pos)
1644 {
1645         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1646         int           rc;
1647
1648         /* Read small symlink from inode body as we need to maintain correct
1649          * on-disk symlinks for ldiskfs.
1650          */
1651         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1652             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1653                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1654         else
1655                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1656
1657         return rc;
1658 }
1659
1660 static inline int osd_extents_enabled(struct super_block *sb,
1661                                       struct inode *inode)
1662 {
1663         if (inode != NULL) {
1664                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1665                         return 1;
1666         } else if (ldiskfs_has_feature_extents(sb)) {
1667                 return 1;
1668         }
1669         return 0;
1670 }
1671
1672 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1673                            const loff_t size, const loff_t pos,
1674                            const int blocks)
1675 {
1676         int credits, bits, bs, i;
1677
1678         bits = sb->s_blocksize_bits;
1679         bs = 1 << bits;
1680
1681         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1682          * we do not expect blockmaps on the large files,
1683          * so let's shrink it to 2 levels (4GB files) */
1684
1685         /* this is default reservation: 2 levels */
1686         credits = (blocks + 2) * 3;
1687
1688         /* actual offset is unknown, hard to optimize */
1689         if (pos == -1)
1690                 return credits;
1691
1692         /* now check for few specific cases to optimize */
1693         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1694                 /* no indirects */
1695                 credits = blocks;
1696                 /* allocate if not allocated */
1697                 if (inode == NULL) {
1698                         credits += blocks * 2;
1699                         return credits;
1700                 }
1701                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1702                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1703                         if (LDISKFS_I(inode)->i_data[i] == 0)
1704                                 credits += 2;
1705                 }
1706         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1707                 /* single indirect */
1708                 credits = blocks * 3;
1709                 if (inode == NULL ||
1710                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1711                         credits += 3;
1712                 else
1713                         /* The indirect block may be modified. */
1714                         credits += 1;
1715         }
1716
1717         return credits;
1718 }
1719
1720 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1721                                  const struct lu_buf *buf, loff_t _pos,
1722                                  struct thandle *handle)
1723 {
1724         struct osd_object  *obj  = osd_dt_obj(dt);
1725         struct inode       *inode = obj->oo_inode;
1726         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1727         struct osd_thandle *oh;
1728         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1729         int                 bits, bs;
1730         int                 depth, size;
1731         loff_t              pos;
1732         ENTRY;
1733
1734         LASSERT(buf != NULL);
1735         LASSERT(handle != NULL);
1736
1737         oh = container_of0(handle, struct osd_thandle, ot_super);
1738         LASSERT(oh->ot_handle == NULL);
1739
1740         size = buf->lb_len;
1741         bits = sb->s_blocksize_bits;
1742         bs = 1 << bits;
1743
1744         if (_pos == -1) {
1745                 /* if this is an append, then we
1746                  * should expect cross-block record */
1747                 pos = 0;
1748         } else {
1749                 pos = _pos;
1750         }
1751
1752         /* blocks to modify */
1753         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1754         LASSERT(blocks > 0);
1755
1756         if (inode != NULL && _pos != -1) {
1757                 /* object size in blocks */
1758                 est = (i_size_read(inode) + bs - 1) >> bits;
1759                 allocated = inode->i_blocks >> (bits - 9);
1760                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1761                         /* looks like an overwrite, no need to modify tree */
1762                         credits = blocks;
1763                         /* no need to modify i_size */
1764                         goto out;
1765                 }
1766         }
1767
1768         if (osd_extents_enabled(sb, inode)) {
1769                 /*
1770                  * many concurrent threads may grow tree by the time
1771                  * our transaction starts. so, consider 2 is a min depth
1772                  * for every level we may need to allocate a new block
1773                  * and take some entries from the old one. so, 3 blocks
1774                  * to allocate (bitmap, gd, itself) + old block - 4 per
1775                  * level.
1776                  */
1777                 depth = inode != NULL ? ext_depth(inode) : 0;
1778                 depth = max(depth, 1) + 1;
1779                 credits = depth;
1780                 /* if not append, then split may need to modify
1781                  * existing blocks moving entries into the new ones */
1782                 if (_pos != -1)
1783                         credits += depth;
1784                 /* blocks to store data: bitmap,gd,itself */
1785                 credits += blocks * 3;
1786         } else {
1787                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1788         }
1789         /* if inode is created as part of the transaction,
1790          * then it's counted already by the creation method */
1791         if (inode != NULL)
1792                 credits++;
1793
1794 out:
1795
1796         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1797
1798         /* dt_declare_write() is usually called for system objects, such
1799          * as llog or last_rcvd files. We needn't enforce quota on those
1800          * objects, so always set the lqi_space as 0. */
1801         if (inode != NULL)
1802                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1803                                            i_gid_read(inode),
1804                                            i_projid_read(inode), 0,
1805                                            oh, obj, NULL, OSD_QID_BLK);
1806
1807         if (rc == 0)
1808                 rc = osd_trunc_lock(obj, oh, true);
1809
1810         RETURN(rc);
1811 }
1812
1813 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1814 {
1815         /* LU-2634: clear the extent format for fast symlink */
1816         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1817
1818         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1819         spin_lock(&inode->i_lock);
1820         LDISKFS_I(inode)->i_disksize = buflen;
1821         i_size_write(inode, buflen);
1822         spin_unlock(&inode->i_lock);
1823         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1824
1825         return 0;
1826 }
1827
1828 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1829                              int write_NUL, loff_t *offs, handle_t *handle)
1830 {
1831         struct buffer_head *bh        = NULL;
1832         loff_t              offset    = *offs;
1833         loff_t              new_size  = i_size_read(inode);
1834         unsigned long       block;
1835         int                 blocksize = 1 << inode->i_blkbits;
1836         int                 err = 0;
1837         int                 size;
1838         int                 boffs;
1839         int                 dirty_inode = 0;
1840
1841         if (write_NUL) {
1842                 /*
1843                  * long symlink write does not count the NUL terminator in
1844                  * bufsize, we write it, and the inode's file size does not
1845                  * count the NUL terminator as well.
1846                  */
1847                 ((char *)buf)[bufsize] = '\0';
1848                 ++bufsize;
1849         }
1850
1851         while (bufsize > 0) {
1852                 int credits = handle->h_buffer_credits;
1853
1854                 if (bh)
1855                         brelse(bh);
1856
1857                 block = offset >> inode->i_blkbits;
1858                 boffs = offset & (blocksize - 1);
1859                 size = min(blocksize - boffs, bufsize);
1860                 bh = __ldiskfs_bread(handle, inode, block, 1);
1861                 if (IS_ERR_OR_NULL(bh)) {
1862                         if (bh == NULL) {
1863                                 err = -EIO;
1864                         } else {
1865                                 err = PTR_ERR(bh);
1866                                 bh = NULL;
1867                         }
1868
1869                         CERROR("%s: error reading offset %llu (block %lu, "
1870                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
1871                                inode->i_sb->s_id, offset, block, bufsize, *offs,
1872                                credits, handle->h_buffer_credits, err);
1873                         break;
1874                 }
1875
1876                 err = ldiskfs_journal_get_write_access(handle, bh);
1877                 if (err) {
1878                         CERROR("journal_get_write_access() returned error %d\n",
1879                                err);
1880                         break;
1881                 }
1882                 LASSERTF(boffs + size <= bh->b_size,
1883                          "boffs %d size %d bh->b_size %lu\n",
1884                          boffs, size, (unsigned long)bh->b_size);
1885                 memcpy(bh->b_data + boffs, buf, size);
1886                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1887                 if (err)
1888                         break;
1889
1890                 if (offset + size > new_size)
1891                         new_size = offset + size;
1892                 offset += size;
1893                 bufsize -= size;
1894                 buf += size;
1895         }
1896         if (bh)
1897                 brelse(bh);
1898
1899         if (write_NUL)
1900                 --new_size;
1901         /* correct in-core and on-disk sizes */
1902         if (new_size > i_size_read(inode)) {
1903                 spin_lock(&inode->i_lock);
1904                 if (new_size > i_size_read(inode))
1905                         i_size_write(inode, new_size);
1906                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1907                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1908                         dirty_inode = 1;
1909                 }
1910                 spin_unlock(&inode->i_lock);
1911                 if (dirty_inode)
1912                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1913         }
1914
1915         if (err == 0)
1916                 *offs = offset;
1917         return err;
1918 }
1919
1920 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1921                          const struct lu_buf *buf, loff_t *pos,
1922                          struct thandle *handle)
1923 {
1924         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1925         struct osd_thandle      *oh;
1926         ssize_t                 result;
1927         int                     is_link;
1928
1929         LASSERT(dt_object_exists(dt));
1930
1931         LASSERT(handle != NULL);
1932         LASSERT(inode != NULL);
1933         ll_vfs_dq_init(inode);
1934
1935         /* XXX: don't check: one declared chunk can be used many times */
1936         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1937
1938         oh = container_of(handle, struct osd_thandle, ot_super);
1939         LASSERT(oh->ot_handle->h_transaction != NULL);
1940         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1941
1942         /* Write small symlink to inode body as we need to maintain correct
1943          * on-disk symlinks for ldiskfs.
1944          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1945          * does not count it in.
1946          */
1947         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1948         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1949                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1950         else
1951                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1952                                                   buf->lb_len, is_link, pos,
1953                                                   oh->ot_handle);
1954         if (result == 0)
1955                 result = buf->lb_len;
1956
1957         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1958
1959         return result;
1960 }
1961
1962 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1963                              __u64 start, __u64 end, struct thandle *th)
1964 {
1965         struct osd_thandle *oh;
1966         struct inode       *inode;
1967         int                 rc;
1968         ENTRY;
1969
1970         LASSERT(th);
1971         oh = container_of(th, struct osd_thandle, ot_super);
1972
1973         /*
1974          * we don't need to reserve credits for whole truncate
1975          * it's not possible as truncate may need to free too many
1976          * blocks and that won't fit a single transaction. instead
1977          * we reserve credits to change i_size and put inode onto
1978          * orphan list. if needed truncate will extend or restart
1979          * transaction
1980          */
1981         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1982                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1983
1984         inode = osd_dt_obj(dt)->oo_inode;
1985         LASSERT(inode);
1986
1987         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1988                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1989                                    NULL, OSD_QID_BLK);
1990
1991         if (rc == 0)
1992                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
1993
1994         RETURN(rc);
1995 }
1996
1997 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1998                      __u64 start, __u64 end, struct thandle *th)
1999 {
2000         struct osd_object *obj = osd_dt_obj(dt);
2001         struct osd_device *osd = osd_obj2dev(obj);
2002         struct inode *inode = obj->oo_inode;
2003         struct osd_access_lock *al;
2004         struct osd_thandle *oh;
2005         int rc = 0, found = 0;
2006         bool grow = false;
2007         ENTRY;
2008
2009         LASSERT(end == OBD_OBJECT_EOF);
2010         LASSERT(dt_object_exists(dt));
2011         LASSERT(osd_invariant(obj));
2012         LASSERT(inode != NULL);
2013         ll_vfs_dq_init(inode);
2014
2015         LASSERT(th);
2016         oh = container_of(th, struct osd_thandle, ot_super);
2017         LASSERT(oh->ot_handle->h_transaction != NULL);
2018
2019         /* we used to skip truncate to current size to
2020          * optimize truncates on OST. with DoM we can
2021          * get attr_set to set specific size (MDS_REINT)
2022          * and then get truncate RPC which essentially
2023          * would be skipped. this is bad.. so, disable
2024          * this optimization on MDS till the client stop
2025          * to sent MDS_REINT (LU-11033) -bzzz */
2026         if (osd->od_is_ost && i_size_read(inode) == start)
2027                 RETURN(0);
2028
2029         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2030
2031         spin_lock(&inode->i_lock);
2032         if (i_size_read(inode) < start)
2033                 grow = true;
2034         i_size_write(inode, start);
2035         spin_unlock(&inode->i_lock);
2036         ll_truncate_pagecache(inode, start);
2037
2038         /* optimize grow case */
2039         if (grow) {
2040                 osd_execute_truncate(obj);
2041                 GOTO(out, rc);
2042         }
2043
2044         /* add to orphan list to ensure truncate completion
2045          * if this transaction succeed. ldiskfs_truncate()
2046          * will take the inode out of the list */
2047         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2048         if (rc != 0)
2049                 GOTO(out, rc);
2050
2051         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2052                 if (obj != al->tl_obj)
2053                         continue;
2054                 LASSERT(al->tl_shared == 0);
2055                 found = 1;
2056                 /* do actual truncate in osd_trans_stop() */
2057                 al->tl_truncate = 1;
2058                 break;
2059         }
2060         LASSERT(found);
2061
2062 out:
2063         RETURN(rc);
2064 }
2065
2066 static int fiemap_check_ranges(struct inode *inode,
2067                                u64 start, u64 len, u64 *new_len)
2068 {
2069         loff_t maxbytes;
2070
2071         *new_len = len;
2072
2073         if (len == 0)
2074                 return -EINVAL;
2075
2076         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2077                 maxbytes = inode->i_sb->s_maxbytes;
2078         else
2079                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2080
2081         if (start > maxbytes)
2082                 return -EFBIG;
2083
2084         /*
2085          * Shrink request scope to what the fs can actually handle.
2086          */
2087         if (len > maxbytes || (maxbytes - len) < start)
2088                 *new_len = maxbytes - start;
2089
2090         return 0;
2091 }
2092
2093 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2094 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
2095
2096 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2097                           struct fiemap *fm)
2098 {
2099         struct fiemap_extent_info fieinfo = {0, };
2100         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2101         u64 len;
2102         int rc;
2103
2104
2105         LASSERT(inode);
2106         if (inode->i_op->fiemap == NULL)
2107                 return -EOPNOTSUPP;
2108
2109         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2110                 return -EINVAL;
2111
2112         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2113         if (rc)
2114                 return rc;
2115
2116         fieinfo.fi_flags = fm->fm_flags;
2117         fieinfo.fi_extents_max = fm->fm_extent_count;
2118         fieinfo.fi_extents_start = fm->fm_extents;
2119
2120         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2121                 filemap_write_and_wait(inode->i_mapping);
2122
2123         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2124         fm->fm_flags = fieinfo.fi_flags;
2125         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2126
2127         return rc;
2128 }
2129
2130 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2131                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2132 {
2133         int              rc = 0;
2134         struct inode    *inode = osd_dt_obj(dt)->oo_inode;
2135         ENTRY;
2136
2137         switch (advice) {
2138         case LU_LADVISE_DONTNEED:
2139                 if (end == 0)
2140                         break;
2141                 invalidate_mapping_pages(inode->i_mapping,
2142                                          start >> PAGE_SHIFT,
2143                                          (end - 1) >> PAGE_SHIFT);
2144                 break;
2145         default:
2146                 rc = -ENOTSUPP;
2147                 break;
2148         }
2149
2150         RETURN(rc);
2151 }
2152
2153 /*
2154  * in some cases we may need declare methods for objects being created
2155  * e.g., when we create symlink
2156  */
2157 const struct dt_body_operations osd_body_ops_new = {
2158         .dbo_declare_write = osd_declare_write,
2159 };
2160
2161 const struct dt_body_operations osd_body_ops = {
2162         .dbo_read                       = osd_read,
2163         .dbo_declare_write              = osd_declare_write,
2164         .dbo_write                      = osd_write,
2165         .dbo_bufs_get                   = osd_bufs_get,
2166         .dbo_bufs_put                   = osd_bufs_put,
2167         .dbo_write_prep                 = osd_write_prep,
2168         .dbo_declare_write_commit       = osd_declare_write_commit,
2169         .dbo_write_commit               = osd_write_commit,
2170         .dbo_read_prep                  = osd_read_prep,
2171         .dbo_declare_punch              = osd_declare_punch,
2172         .dbo_punch                      = osd_punch,
2173         .dbo_fiemap_get                 = osd_fiemap_get,
2174         .dbo_ladvise                    = osd_ladvise,
2175 };
2176
2177 /**
2178  * Get a truncate lock
2179  *
2180  * In order to take multi-transaction truncate out of main transaction we let
2181  * the caller grab a lock on the object passed. the lock can be shared (for
2182  * writes) and exclusive (for truncate). It's not allowed to mix truncate
2183  * and write in the same transaction handle (do not confuse with big ldiskfs
2184  * transaction containing lots of handles).
2185  * The lock must be taken at declaration.
2186  *
2187  * \param obj           object to lock
2188  * \oh                  transaction
2189  * \shared              shared or exclusive
2190  *
2191  * \retval 0            lock is granted
2192  * \retval -NOMEM       no memory to allocate lock
2193  */
2194 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2195 {
2196         struct osd_access_lock *al, *tmp;
2197
2198         LASSERT(obj);
2199         LASSERT(oh);
2200
2201         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2202                 if (tmp->tl_obj != obj)
2203                         continue;
2204                 LASSERT(tmp->tl_shared == shared);
2205                 /* found same lock */
2206                 return 0;
2207         }
2208
2209         OBD_ALLOC_PTR(al);
2210         if (unlikely(al == NULL))
2211                 return -ENOMEM;
2212         al->tl_obj = obj;
2213         al->tl_truncate = false;
2214         if (shared)
2215                 down_read(&obj->oo_ext_idx_sem);
2216         else
2217                 down_write(&obj->oo_ext_idx_sem);
2218         al->tl_shared = shared;
2219
2220         list_add(&al->tl_list, &oh->ot_trunc_locks);
2221
2222         return 0;
2223 }
2224
2225 void osd_trunc_unlock_all(struct list_head *list)
2226 {
2227         struct osd_access_lock *al, *tmp;
2228         list_for_each_entry_safe(al, tmp, list, tl_list) {
2229                 if (al->tl_shared)
2230                         up_read(&al->tl_obj->oo_ext_idx_sem);
2231                 else
2232                         up_write(&al->tl_obj->oo_ext_idx_sem);
2233                 list_del(&al->tl_list);
2234                 OBD_FREE_PTR(al);
2235         }
2236 }
2237
2238 void osd_execute_truncate(struct osd_object *obj)
2239 {
2240         struct osd_device *d = osd_obj2dev(obj);
2241         struct inode *inode = obj->oo_inode;
2242         __u64 size;
2243
2244         /* simulate crash before (in the middle) of delayed truncate */
2245         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2246                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2247                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2248
2249                 mutex_lock(&sbi->s_orphan_lock);
2250                 list_del_init(&ei->i_orphan);
2251                 mutex_unlock(&sbi->s_orphan_lock);
2252                 return;
2253         }
2254
2255 #ifdef HAVE_INODEOPS_TRUNCATE
2256         if (inode->i_op->truncate)
2257                 inode->i_op->truncate(inode);
2258         else
2259 #endif
2260                 ldiskfs_truncate(inode);
2261
2262         /*
2263          * For a partial-page truncate, flush the page to disk immediately to
2264          * avoid data corruption during direct disk write.  b=17397
2265          */
2266         size = i_size_read(inode);
2267         if ((size & ~PAGE_MASK) == 0)
2268                 return;
2269         if (osd_use_page_cache(d)) {
2270                 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2271         } else {
2272                 /* Notice we use "wait" version to ensure I/O is complete */
2273                 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2274                 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2275                                          size >> PAGE_SHIFT);
2276         }
2277 }
2278
2279 void osd_process_truncates(struct list_head *list)
2280 {
2281         struct osd_access_lock *al;
2282
2283         LASSERT(journal_current_handle() == NULL);
2284
2285         list_for_each_entry(al, list, tl_list) {
2286                 if (al->tl_shared)
2287                         continue;
2288                 if (!al->tl_truncate)
2289                         continue;
2290                 osd_execute_truncate(al->tl_obj);
2291         }
2292 }