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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/osd/osd_io.c
  37  *
  38  * body operations
  39  *
  40  * Author: Nikita Danilov <nikita@clusterfs.com>
  41  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  42  *
  43  */
  44
  45 /* LUSTRE_VERSION_CODE */
  46 #include <lustre_ver.h>
  47 /* prerequisite for linux/xattr.h */
  48 #include <linux/types.h>
  49 /* prerequisite for linux/xattr.h */
  50 #include <linux/fs.h>
  51
  52 /*
  53  * struct OBD_{ALLOC,FREE}*()
  54  * OBD_FAIL_CHECK
  55  */
  56 #include <obd_support.h>
  57
  58 #include "osd_internal.h"
  59
  60 /* ext_depth() */
  61 #include <ldiskfs/ldiskfs_extents.h>
  62
  63 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  64                             int rw, int line, int pages)
  65 {
  66         int blocks, i;
  67
  68         LASSERTF(iobuf->dr_elapsed_valid == 0,
  69                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  70                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  71                  iobuf->dr_init_at);
  72         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  73
  74         init_waitqueue_head(&iobuf->dr_wait);
  75         atomic_set(&iobuf->dr_numreqs, 0);
  76         iobuf->dr_npages = 0;
  77         iobuf->dr_error = 0;
  78         iobuf->dr_dev = d;
  79         iobuf->dr_frags = 0;
  80         iobuf->dr_elapsed = 0;
  81         /* must be counted before, so assert */
  82         iobuf->dr_rw = rw;
  83         iobuf->dr_init_at = line;
  84
  85         blocks = pages * (PAGE_CACHE_SIZE >> osd_sb(d)->s_blocksize_bits);
  86         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  87                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  88                         pages * sizeof(iobuf->dr_pages[0]));
  89                 return 0;
  90         }
  91
  92         /* start with 1MB for 4K blocks */
  93         i = 256;
  94         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
  95                 i <<= 1;
  96
  97         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
  98                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
  99         pages = i;
 100         blocks = pages * (PAGE_CACHE_SIZE >> osd_sb(d)->s_blocksize_bits);
 101         iobuf->dr_max_pages = 0;
 102         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
 103                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 104
 105         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 106         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 107         if (unlikely(iobuf->dr_blocks == NULL))
 108                 return -ENOMEM;
 109
 110         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 111         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 112         if (unlikely(iobuf->dr_pages == NULL))
 113                 return -ENOMEM;
 114
 115         iobuf->dr_max_pages = pages;
 116
 117         return 0;
 118 }
 119 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 120         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 121
 122 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
 123 {
 124         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 125         iobuf->dr_pages[iobuf->dr_npages++] = page;
 126 }
 127
 128 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 129 {
 130         int rw = iobuf->dr_rw;
 131
 132         if (iobuf->dr_elapsed_valid) {
 133                 iobuf->dr_elapsed_valid = 0;
 134                 LASSERT(iobuf->dr_dev == d);
 135                 LASSERT(iobuf->dr_frags > 0);
 136                 lprocfs_oh_tally(&d->od_brw_stats.
 137                                  hist[BRW_R_DIO_FRAGS+rw],
 138                                  iobuf->dr_frags);
 139                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 140                                       iobuf->dr_elapsed);
 141         }
 142 }
 143
 144 #ifndef REQ_WRITE /* pre-2.6.35 */
 145 #define __REQ_WRITE BIO_RW
 146 #endif
 147
 148 static void dio_complete_routine(struct bio *bio, int error)
 149 {
 150         struct osd_iobuf *iobuf = bio->bi_private;
 151         struct bio_vec *bvl;
 152         int i;
 153
 154         /* CAVEAT EMPTOR: possibly in IRQ context
 155          * DO NOT record procfs stats here!!! */
 156
 157         if (unlikely(iobuf == NULL)) {
 158                 CERROR("***** bio->bi_private is NULL!  This should never "
 159                        "happen.  Normally, I would crash here, but instead I "
 160                        "will dump the bio contents to the console.  Please "
 161                        "report this to <http://jira.whamcloud.com/> , along "
 162                        "with any interesting messages leading up to this point "
 163                        "(like SCSI errors, perhaps).  Because bi_private is "
 164                        "NULL, I can't wake up the thread that initiated this "
 165                        "IO - you will probably have to reboot this node.\n");
 166                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 167                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 168                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 169                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 170                        bio->bi_end_io, atomic_read(&bio->bi_cnt),
 171                        bio->bi_private);
 172                 return;
 173         }
 174
 175         /* the check is outside of the cycle for performance reason -bzzz */
 176         if (!test_bit(__REQ_WRITE, &bio->bi_rw)) {
 177                 bio_for_each_segment(bvl, bio, i) {
 178                         if (likely(error == 0))
 179                                 SetPageUptodate(bvl->bv_page);
 180                         LASSERT(PageLocked(bvl->bv_page));
 181                 }
 182                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 183         } else {
 184                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 185         }
 186
 187         /* any real error is good enough -bzzz */
 188         if (error != 0 && iobuf->dr_error == 0)
 189                 iobuf->dr_error = error;
 190
 191         /*
 192          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 193          * it's possible that service thread will see dr_numreqs
 194          * is zero, but dr_elapsed is not set yet, leading to lost
 195          * data in this processing and an assertion in a subsequent
 196          * call to OSD.
 197          */
 198         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 199                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
 200                 iobuf->dr_elapsed_valid = 1;
 201         }
 202         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 203                 wake_up(&iobuf->dr_wait);
 204
 205         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 206          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 207          * mempool when serious on-disk fragmentation was encountered,
 208          * deadlocking the OST.  The bios are now released as soon as complete
 209          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 210         bio_put(bio);
 211 }
 212
 213 static void record_start_io(struct osd_iobuf *iobuf, int size)
 214 {
 215         struct osd_device    *osd = iobuf->dr_dev;
 216         struct obd_histogram *h = osd->od_brw_stats.hist;
 217
 218         iobuf->dr_frags++;
 219         atomic_inc(&iobuf->dr_numreqs);
 220
 221         if (iobuf->dr_rw == 0) {
 222                 atomic_inc(&osd->od_r_in_flight);
 223                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 224                                  atomic_read(&osd->od_r_in_flight));
 225                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 226         } else if (iobuf->dr_rw == 1) {
 227                 atomic_inc(&osd->od_w_in_flight);
 228                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 229                                  atomic_read(&osd->od_w_in_flight));
 230                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 231         } else {
 232                 LBUG();
 233         }
 234 }
 235
 236 static void osd_submit_bio(int rw, struct bio *bio)
 237 {
 238         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 239         if (rw == 0)
 240                 submit_bio(READ, bio);
 241         else
 242                 submit_bio(WRITE, bio);
 243 }
 244
 245 static int can_be_merged(struct bio *bio, sector_t sector)
 246 {
 247         unsigned int size;
 248
 249         if (!bio)
 250                 return 0;
 251
 252         size = bio->bi_size >> 9;
 253         return bio->bi_sector + size == sector ? 1 : 0;
 254 }
 255
 256 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 257                       struct osd_iobuf *iobuf)
 258 {
 259         int            blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
 260         struct page  **pages = iobuf->dr_pages;
 261         int            npages = iobuf->dr_npages;
 262         unsigned long *blocks = iobuf->dr_blocks;
 263         int            total_blocks = npages * blocks_per_page;
 264         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 265         unsigned int   blocksize = inode->i_sb->s_blocksize;
 266         struct bio    *bio = NULL;
 267         struct page   *page;
 268         unsigned int   page_offset;
 269         sector_t       sector;
 270         int            nblocks;
 271         int            block_idx;
 272         int            page_idx;
 273         int            i;
 274         int            rc = 0;
 275         ENTRY;
 276
 277         LASSERT(iobuf->dr_npages == npages);
 278
 279         osd_brw_stats_update(osd, iobuf);
 280         iobuf->dr_start_time = cfs_time_current();
 281
 282         for (page_idx = 0, block_idx = 0;
 283              page_idx < npages;
 284              page_idx++, block_idx += blocks_per_page) {
 285
 286                 page = pages[page_idx];
 287                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 288
 289                 for (i = 0, page_offset = 0;
 290                      i < blocks_per_page;
 291                      i += nblocks, page_offset += blocksize * nblocks) {
 292
 293                         nblocks = 1;
 294
 295                         if (blocks[block_idx + i] == 0) {  /* hole */
 296                                 LASSERTF(iobuf->dr_rw == 0,
 297                                          "page_idx %u, block_idx %u, i %u\n",
 298                                          page_idx, block_idx, i);
 299                                 memset(kmap(page) + page_offset, 0, blocksize);
 300                                 kunmap(page);
 301                                 continue;
 302                         }
 303
 304                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 305
 306                         /* Additional contiguous file blocks? */
 307                         while (i + nblocks < blocks_per_page &&
 308                                (sector + (nblocks << sector_bits)) ==
 309                                ((sector_t)blocks[block_idx + i + nblocks] <<
 310                                 sector_bits))
 311                                 nblocks++;
 312
 313                         if (bio != NULL &&
 314                             can_be_merged(bio, sector) &&
 315                             bio_add_page(bio, page,
 316                                          blocksize * nblocks, page_offset) != 0)
 317                                 continue;       /* added this frag OK */
 318
 319                         if (bio != NULL) {
 320                                 struct request_queue *q =
 321                                         bdev_get_queue(bio->bi_bdev);
 322
 323                                 /* Dang! I have to fragment this I/O */
 324                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 325                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 326                                        bio->bi_size,
 327                                        bio->bi_vcnt, bio->bi_max_vecs,
 328                                        bio->bi_size >> 9, queue_max_sectors(q),
 329                                        bio_phys_segments(q, bio),
 330                                        queue_max_phys_segments(q),
 331                                        0, queue_max_hw_segments(q));
 332
 333                                 record_start_io(iobuf, bio->bi_size);
 334                                 osd_submit_bio(iobuf->dr_rw, bio);
 335                         }
 336
 337                         /* allocate new bio */
 338                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 339                                                       (npages - page_idx) *
 340                                                       blocks_per_page));
 341                         if (bio == NULL) {
 342                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 343                                        (npages - page_idx), blocks_per_page,
 344                                        (npages - page_idx) * blocks_per_page);
 345                                 rc = -ENOMEM;
 346                                 goto out;
 347                         }
 348
 349                         bio->bi_bdev = inode->i_sb->s_bdev;
 350                         bio->bi_sector = sector;
 351                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 352                         bio->bi_end_io = dio_complete_routine;
 353                         bio->bi_private = iobuf;
 354
 355                         rc = bio_add_page(bio, page,
 356                                           blocksize * nblocks, page_offset);
 357                         LASSERT(rc != 0);
 358                 }
 359         }
 360
 361         if (bio != NULL) {
 362                 record_start_io(iobuf, bio->bi_size);
 363                 osd_submit_bio(iobuf->dr_rw, bio);
 364                 rc = 0;
 365         }
 366
 367 out:
 368         /* in order to achieve better IO throughput, we don't wait for writes
 369          * completion here. instead we proceed with transaction commit in
 370          * parallel and wait for IO completion once transaction is stopped
 371          * see osd_trans_stop() for more details -bzzz */
 372         if (iobuf->dr_rw == 0) {
 373                 wait_event(iobuf->dr_wait,
 374                            atomic_read(&iobuf->dr_numreqs) == 0);
 375                 osd_fini_iobuf(osd, iobuf);
 376         }
 377
 378         if (rc == 0)
 379                 rc = iobuf->dr_error;
 380         RETURN(rc);
 381 }
 382
 383 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 384                                    struct niobuf_local *lnb)
 385 {
 386         ENTRY;
 387
 388         *nrpages = 0;
 389
 390         while (len > 0) {
 391                 int poff = offset & (PAGE_CACHE_SIZE - 1);
 392                 int plen = PAGE_CACHE_SIZE - poff;
 393
 394                 if (plen > len)
 395                         plen = len;
 396                 lnb->lnb_file_offset = offset;
 397                 lnb->lnb_page_offset = poff;
 398                 lnb->len = plen;
 399                 /* lb->flags = rnb->flags; */
 400                 lnb->flags = 0;
 401                 lnb->page = NULL;
 402                 lnb->rc = 0;
 403
 404                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 405                          (long long) len);
 406                 offset += plen;
 407                 len -= plen;
 408                 lnb++;
 409                 (*nrpages)++;
 410         }
 411
 412         RETURN(0);
 413 }
 414
 415 struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
 416 {
 417         struct inode      *inode = osd_dt_obj(dt)->oo_inode;
 418         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 419         struct page       *page;
 420
 421         LASSERT(inode);
 422
 423         page = find_or_create_page(inode->i_mapping, offset >> PAGE_CACHE_SHIFT,
 424                                    GFP_NOFS | __GFP_HIGHMEM);
 425         if (unlikely(page == NULL))
 426                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 427
 428         return page;
 429 }
 430
 431 /*
 432  * there are following "locks":
 433  * journal_start
 434  * i_mutex
 435  * page lock
 436
 437  * osd write path
 438     * lock page(s)
 439     * journal_start
 440     * truncate_sem
 441
 442  * ext4 vmtruncate:
 443     * lock pages, unlock
 444     * journal_start
 445     * lock partial page
 446     * i_data_sem
 447
 448 */
 449 int osd_bufs_get(const struct lu_env *env, struct dt_object *d, loff_t pos,
 450                  ssize_t len, struct niobuf_local *lnb, int rw,
 451                  struct lustre_capa *capa)
 452 {
 453         struct osd_object   *obj    = osd_dt_obj(d);
 454         int npages, i, rc = 0;
 455
 456         LASSERT(obj->oo_inode);
 457
 458         osd_map_remote_to_local(pos, len, &npages, lnb);
 459
 460         for (i = 0; i < npages; i++, lnb++) {
 461
 462                 /* We still set up for ungranted pages so that granted pages
 463                  * can be written to disk as they were promised, and portals
 464                  * needs to keep the pages all aligned properly. */
 465                 lnb->dentry = (void *) obj;
 466
 467                 lnb->page = osd_get_page(d, lnb->lnb_file_offset, rw);
 468                 if (lnb->page == NULL)
 469                         GOTO(cleanup, rc = -ENOMEM);
 470
 471                 /* DLM locking protects us from write and truncate competing
 472                  * for same region, but truncate can leave dirty page in the
 473                  * cache. it's possible the writeout on a such a page is in
 474                  * progress when we access it. it's also possible that during
 475                  * this writeout we put new (partial) data, but then won't
 476                  * be able to proceed in filter_commitrw_write(). thus let's
 477                  * just wait for writeout completion, should be rare enough.
 478                  * -bzzz */
 479                 wait_on_page_writeback(lnb->page);
 480                 BUG_ON(PageWriteback(lnb->page));
 481
 482                 lu_object_get(&d->do_lu);
 483         }
 484         rc = i;
 485
 486 cleanup:
 487         RETURN(rc);
 488 }
 489
 490 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 491                         struct niobuf_local *lnb, int npages)
 492 {
 493         int                     i;
 494
 495         for (i = 0; i < npages; i++) {
 496                 if (lnb[i].page == NULL)
 497                         continue;
 498                 LASSERT(PageLocked(lnb[i].page));
 499                 unlock_page(lnb[i].page);
 500                 page_cache_release(lnb[i].page);
 501                 lu_object_put(env, &dt->do_lu);
 502                 lnb[i].page = NULL;
 503         }
 504         RETURN(0);
 505 }
 506
 507 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 508 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 509 #endif
 510
 511 struct bpointers {
 512         unsigned long *blocks;
 513         unsigned long start;
 514         int num;
 515         int init_num;
 516         int create;
 517 };
 518
 519 static long ldiskfs_ext_find_goal(struct inode *inode,
 520                                   struct ldiskfs_ext_path *path,
 521                                   unsigned long block, int *aflags)
 522 {
 523         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 524         unsigned long bg_start;
 525         unsigned long colour;
 526         int depth;
 527
 528         if (path) {
 529                 struct ldiskfs_extent *ex;
 530                 depth = path->p_depth;
 531
 532                 /* try to predict block placement */
 533                 if ((ex = path[depth].p_ext))
 534                         return ldiskfs_ext_pblock(ex) +
 535                                 (block - le32_to_cpu(ex->ee_block));
 536
 537                 /* it looks index is empty
 538                  * try to find starting from index itself */
 539                 if (path[depth].p_bh)
 540                         return path[depth].p_bh->b_blocknr;
 541         }
 542
 543         /* OK. use inode's group */
 544         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 545                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 546         colour = (current->pid % 16) *
 547                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 548         return bg_start + colour + block;
 549 }
 550
 551 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 552                                 struct ldiskfs_ext_path *path,
 553                                 unsigned long block, unsigned long *count,
 554                                 int *err)
 555 {
 556         struct ldiskfs_allocation_request ar;
 557         unsigned long pblock;
 558         int aflags;
 559
 560         /* find neighbour allocated blocks */
 561         ar.lleft = block;
 562         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 563         if (*err)
 564                 return 0;
 565         ar.lright = block;
 566         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 567         if (*err)
 568                 return 0;
 569
 570         /* allocate new block */
 571         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 572         ar.inode = inode;
 573         ar.logical = block;
 574         ar.len = *count;
 575         ar.flags = LDISKFS_MB_HINT_DATA;
 576         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 577         *count = ar.len;
 578         return pblock;
 579 }
 580
 581 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 582                                      struct ldiskfs_ext_path *path,
 583                                      struct ldiskfs_ext_cache *cex,
 584 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 585                                      struct ldiskfs_extent *ex,
 586 #endif
 587                                      void *cbdata)
 588 {
 589         struct bpointers *bp = cbdata;
 590         struct ldiskfs_extent nex;
 591         unsigned long pblock;
 592         unsigned long tgen;
 593         int err, i;
 594         unsigned long count;
 595         handle_t *handle;
 596
 597 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 598         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 599 #else
 600         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 601 #endif
 602                 err = EXT_CONTINUE;
 603                 goto map;
 604         }
 605
 606         if (bp->create == 0) {
 607                 i = 0;
 608                 if (cex->ec_block < bp->start)
 609                         i = bp->start - cex->ec_block;
 610                 if (i >= cex->ec_len)
 611                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 612                                         i, cex->ec_len);
 613                 for (; i < cex->ec_len && bp->num; i++) {
 614                         *(bp->blocks) = 0;
 615                         bp->blocks++;
 616                         bp->num--;
 617                         bp->start++;
 618                 }
 619
 620                 return EXT_CONTINUE;
 621         }
 622
 623         tgen = LDISKFS_I(inode)->i_ext_generation;
 624         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
 625
 626         handle = ldiskfs_journal_start(inode, count + LDISKFS_ALLOC_NEEDED + 1);
 627         if (IS_ERR(handle)) {
 628                 return PTR_ERR(handle);
 629         }
 630
 631         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
 632                 /* the tree has changed. so path can be invalid at moment */
 633                 ldiskfs_journal_stop(handle);
 634                 return EXT_REPEAT;
 635         }
 636
 637         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
 638          * protected by i_data_sem as whole. so we patch it to store
 639          * generation to path and now verify the tree hasn't changed */
 640         down_write((&LDISKFS_I(inode)->i_data_sem));
 641
 642         /* validate extent, make sure the extent tree does not changed */
 643         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
 644                 /* cex is invalid, try again */
 645                 up_write(&LDISKFS_I(inode)->i_data_sem);
 646                 ldiskfs_journal_stop(handle);
 647                 return EXT_REPEAT;
 648         }
 649
 650         count = cex->ec_len;
 651         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
 652         if (!pblock)
 653                 goto out;
 654         BUG_ON(count > cex->ec_len);
 655
 656         /* insert new extent */
 657         nex.ee_block = cpu_to_le32(cex->ec_block);
 658         ldiskfs_ext_store_pblock(&nex, pblock);
 659         nex.ee_len = cpu_to_le16(count);
 660         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
 661         if (err) {
 662                 /* free data blocks we just allocated */
 663                 /* not a good idea to call discard here directly,
 664                  * but otherwise we'd need to call it every free() */
 665                 ldiskfs_discard_preallocations(inode);
 666 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
 667                 ldiskfs_free_blocks(handle, inode, NULL,
 668                                     ldiskfs_ext_pblock(&nex),
 669                                     le16_to_cpu(nex.ee_len), 0);
 670 #else
 671                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
 672                                     le16_to_cpu(nex.ee_len), 0);
 673 #endif
 674                 goto out;
 675         }
 676
 677         /*
 678          * Putting len of the actual extent we just inserted,
 679          * we are asking ldiskfs_ext_walk_space() to continue
 680          * scaning after that block
 681          */
 682         cex->ec_len = le16_to_cpu(nex.ee_len);
 683         cex->ec_start = ldiskfs_ext_pblock(&nex);
 684         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
 685         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
 686
 687 out:
 688         up_write((&LDISKFS_I(inode)->i_data_sem));
 689         ldiskfs_journal_stop(handle);
 690 map:
 691         if (err >= 0) {
 692                 /* map blocks */
 693                 if (bp->num == 0) {
 694                         CERROR("hmm. why do we find this extent?\n");
 695                         CERROR("initial space: %lu:%u\n",
 696                                 bp->start, bp->init_num);
 697 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 698                         CERROR("current extent: %u/%u/%llu %d\n",
 699                                 cex->ec_block, cex->ec_len,
 700                                 (unsigned long long)cex->ec_start,
 701                                 cex->ec_type);
 702 #else
 703                         CERROR("current extent: %u/%u/%llu\n",
 704                                 cex->ec_block, cex->ec_len,
 705                                 (unsigned long long)cex->ec_start);
 706 #endif
 707                 }
 708                 i = 0;
 709                 if (cex->ec_block < bp->start)
 710                         i = bp->start - cex->ec_block;
 711                 if (i >= cex->ec_len)
 712                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 713                                         i, cex->ec_len);
 714                 for (; i < cex->ec_len && bp->num; i++) {
 715                         *(bp->blocks) = cex->ec_start + i;
 716 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 717                         if (cex->ec_type != LDISKFS_EXT_CACHE_EXTENT) {
 718 #else
 719                         if ((cex->ec_len == 0) || (cex->ec_start == 0)) {
 720 #endif
 721                                 /* unmap any possible underlying metadata from
 722                                  * the block device mapping.  bug 6998. */
 723                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
 724                                                           *(bp->blocks));
 725                         }
 726                         bp->blocks++;
 727                         bp->num--;
 728                         bp->start++;
 729                 }
 730         }
 731         return err;
 732 }
 733
 734 int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long block,
 735                             unsigned long num, unsigned long *blocks,
 736                             int create)
 737 {
 738         struct bpointers bp;
 739         int err;
 740
 741         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 742                block, block + num - 1, (unsigned) inode->i_ino);
 743
 744         bp.blocks = blocks;
 745         bp.start = block;
 746         bp.init_num = bp.num = num;
 747         bp.create = create;
 748
 749         err = ldiskfs_ext_walk_space(inode, block, num,
 750                                          ldiskfs_ext_new_extent_cb, &bp);
 751         ldiskfs_ext_invalidate_cache(inode);
 752
 753         return err;
 754 }
 755
 756 int osd_ldiskfs_map_ext_inode_pages(struct inode *inode, struct page **page,
 757                                     int pages, unsigned long *blocks,
 758                                     int create)
 759 {
 760         int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
 761         int rc = 0, i = 0;
 762         struct page *fp = NULL;
 763         int clen = 0;
 764
 765         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 766                 inode->i_ino, pages, (*page)->index);
 767
 768         /* pages are sorted already. so, we just have to find
 769          * contig. space and process them properly */
 770         while (i < pages) {
 771                 if (fp == NULL) {
 772                         /* start new extent */
 773                         fp = *page++;
 774                         clen = 1;
 775                         i++;
 776                         continue;
 777                 } else if (fp->index + clen == (*page)->index) {
 778                         /* continue the extent */
 779                         page++;
 780                         clen++;
 781                         i++;
 782                         continue;
 783                 }
 784
 785                 /* process found extent */
 786                 rc = osd_ldiskfs_map_nblocks(inode, fp->index * blocks_per_page,
 787                                              clen * blocks_per_page, blocks,
 788                                              create);
 789                 if (rc)
 790                         GOTO(cleanup, rc);
 791
 792                 /* look for next extent */
 793                 fp = NULL;
 794                 blocks += blocks_per_page * clen;
 795         }
 796
 797         if (fp)
 798                 rc = osd_ldiskfs_map_nblocks(inode, fp->index * blocks_per_page,
 799                                              clen * blocks_per_page, blocks,
 800                                              create);
 801 cleanup:
 802         return rc;
 803 }
 804
 805 int osd_ldiskfs_map_bm_inode_pages(struct inode *inode, struct page **page,
 806                                    int pages, unsigned long *blocks,
 807                                    int create)
 808 {
 809         int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
 810         unsigned long *b;
 811         int rc = 0, i;
 812
 813         for (i = 0, b = blocks; i < pages; i++, page++) {
 814                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
 815                 if (rc) {
 816                         CERROR("ino %lu, blk %lu create %d: rc %d\n",
 817                                inode->i_ino, *b, create, rc);
 818                         break;
 819                 }
 820
 821                 b += blocks_per_page;
 822         }
 823         return rc;
 824 }
 825
 826 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 827                                        int pages, unsigned long *blocks,
 828                                        int create, struct mutex *optional_mutex)
 829 {
 830         int rc;
 831
 832         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 833                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
 834                                                      blocks, create);
 835                 return rc;
 836         }
 837         if (optional_mutex != NULL)
 838                 mutex_lock(optional_mutex);
 839         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
 840         if (optional_mutex != NULL)
 841                 mutex_unlock(optional_mutex);
 842
 843         return rc;
 844 }
 845
 846 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 847                           struct niobuf_local *lnb, int npages)
 848 {
 849         struct osd_thread_info *oti   = osd_oti_get(env);
 850         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 851         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 852         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 853         struct timeval          start;
 854         struct timeval          end;
 855         unsigned long           timediff;
 856         ssize_t                 isize;
 857         __s64                   maxidx;
 858         int                     rc = 0;
 859         int                     i;
 860         int                     cache = 0;
 861
 862         LASSERT(inode);
 863
 864         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 865         if (unlikely(rc != 0))
 866                 RETURN(rc);
 867
 868         isize = i_size_read(inode);
 869         maxidx = ((isize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) - 1;
 870
 871         if (osd->od_writethrough_cache)
 872                 cache = 1;
 873         if (isize > osd->od_readcache_max_filesize)
 874                 cache = 0;
 875
 876         do_gettimeofday(&start);
 877         for (i = 0; i < npages; i++) {
 878
 879                 if (cache == 0)
 880                         generic_error_remove_page(inode->i_mapping,
 881                                                   lnb[i].page);
 882
 883                 /*
 884                  * till commit the content of the page is undefined
 885                  * we'll set it uptodate once bulk is done. otherwise
 886                  * subsequent reads can access non-stable data
 887                  */
 888                 ClearPageUptodate(lnb[i].page);
 889
 890                 if (lnb[i].len == PAGE_CACHE_SIZE)
 891                         continue;
 892
 893                 if (maxidx >= lnb[i].page->index) {
 894                         osd_iobuf_add_page(iobuf, lnb[i].page);
 895                 } else {
 896                         long off;
 897                         char *p = kmap(lnb[i].page);
 898
 899                         off = lnb[i].lnb_page_offset;
 900                         if (off)
 901                                 memset(p, 0, off);
 902                         off = (lnb[i].lnb_page_offset + lnb[i].len) &
 903                               ~CFS_PAGE_MASK;
 904                         if (off)
 905                                 memset(p + off, 0, PAGE_CACHE_SIZE - off);
 906                         kunmap(lnb[i].page);
 907                 }
 908         }
 909         do_gettimeofday(&end);
 910         timediff = cfs_timeval_sub(&end, &start, NULL);
 911         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 912
 913         if (iobuf->dr_npages) {
 914                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
 915                                                  iobuf->dr_npages,
 916                                                  iobuf->dr_blocks,
 917                                                  0, NULL);
 918                 if (likely(rc == 0)) {
 919                         rc = osd_do_bio(osd, inode, iobuf);
 920                         /* do IO stats for preparation reads */
 921                         osd_fini_iobuf(osd, iobuf);
 922                 }
 923         }
 924         RETURN(rc);
 925 }
 926
 927 /* Check if a block is allocated or not */
 928 static int osd_is_mapped(struct inode *inode, obd_size offset)
 929 {
 930         sector_t (*fs_bmap)(struct address_space *, sector_t);
 931
 932         fs_bmap = inode->i_mapping->a_ops->bmap;
 933
 934         /* We can't know if we are overwriting or not */
 935         if (unlikely(fs_bmap == NULL))
 936                 return 0;
 937
 938         if (i_size_read(inode) == 0)
 939                 return 0;
 940
 941         /* Beyond EOF, must not be mapped */
 942         if (((i_size_read(inode) - 1) >> inode->i_blkbits) <
 943             (offset >> inode->i_blkbits))
 944                 return 0;
 945
 946         if (fs_bmap(inode->i_mapping, offset >> inode->i_blkbits) == 0)
 947                 return 0;
 948
 949         return 1;
 950 }
 951
 952 static int osd_declare_write_commit(const struct lu_env *env,
 953                                     struct dt_object *dt,
 954                                     struct niobuf_local *lnb, int npages,
 955                                     struct thandle *handle)
 956 {
 957         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 958         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
 959         struct osd_thandle      *oh;
 960         int                      extents = 1;
 961         int                      depth;
 962         int                      i;
 963         int                      newblocks;
 964         int                      rc = 0;
 965         int                      flags = 0;
 966         bool                     ignore_quota = false;
 967         long long                quota_space = 0;
 968         ENTRY;
 969
 970         LASSERT(handle != NULL);
 971         oh = container_of0(handle, struct osd_thandle, ot_super);
 972         LASSERT(oh->ot_handle == NULL);
 973
 974         newblocks = npages;
 975
 976         /* calculate number of extents (probably better to pass nb) */
 977         for (i = 0; i < npages; i++) {
 978                 if (i && lnb[i].lnb_file_offset !=
 979                     lnb[i - 1].lnb_file_offset + lnb[i - 1].len)
 980                         extents++;
 981
 982                 if (!osd_is_mapped(inode, lnb[i].lnb_file_offset))
 983                         quota_space += PAGE_CACHE_SIZE;
 984
 985                 /* ignore quota for the whole request if any page is from
 986                  * client cache or written by root.
 987                  *
 988                  * XXX once we drop the 1.8 client support, the checking
 989                  * for whether page is from cache can be simplified as:
 990                  * !(lnb[i].flags & OBD_BRW_SYNC)
 991                  *
 992                  * XXX we could handle this on per-lnb basis as done by
 993                  * grant. */
 994                 if ((lnb[i].flags & OBD_BRW_NOQUOTA) ||
 995                     (lnb[i].flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 996                     OBD_BRW_FROM_GRANT)
 997                         ignore_quota = true;
 998         }
 999
1000         /*
1001          * each extent can go into new leaf causing a split
1002          * 5 is max tree depth: inode + 4 index blocks
1003          * with blockmaps, depth is 3 at most
1004          */
1005         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1006                 /*
1007                  * many concurrent threads may grow tree by the time
1008                  * our transaction starts. so, consider 2 is a min depth
1009                  */
1010                 depth = ext_depth(inode);
1011                 depth = max(depth, 1) + 1;
1012                 newblocks += depth;
1013                 oh->ot_credits++; /* inode */
1014                 oh->ot_credits += depth * 2 * extents;
1015         } else {
1016                 depth = 3;
1017                 newblocks += depth;
1018                 oh->ot_credits++; /* inode */
1019                 oh->ot_credits += depth * extents;
1020         }
1021
1022         /* quota space for metadata blocks */
1023         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1024
1025         /* quota space should be reported in 1K blocks */
1026         quota_space = toqb(quota_space);
1027
1028         /* each new block can go in different group (bitmap + gd) */
1029
1030         /* we can't dirty more bitmap blocks than exist */
1031         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1032                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1033         else
1034                 oh->ot_credits += newblocks;
1035
1036         /* we can't dirty more gd blocks than exist */
1037         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1038                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1039         else
1040                 oh->ot_credits += newblocks;
1041
1042         /* make sure the over quota flags were not set */
1043         lnb[0].flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA);
1044
1045         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid,
1046                                    quota_space, oh, true, true, &flags,
1047                                    ignore_quota);
1048
1049         /* we need only to store the overquota flags in the first lnb for
1050          * now, once we support multiple objects BRW, this code needs be
1051          * revised. */
1052         if (flags & QUOTA_FL_OVER_USRQUOTA)
1053                 lnb[0].flags |= OBD_BRW_OVER_USRQUOTA;
1054         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1055                 lnb[0].flags |= OBD_BRW_OVER_GRPQUOTA;
1056
1057         RETURN(rc);
1058 }
1059
1060 /* Check if a block is allocated or not */
1061 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1062                             struct niobuf_local *lnb, int npages,
1063                             struct thandle *thandle)
1064 {
1065         struct osd_thread_info *oti = osd_oti_get(env);
1066         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1067         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1068         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1069         loff_t isize;
1070         int rc = 0, i;
1071
1072         LASSERT(inode);
1073
1074         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1075         if (unlikely(rc != 0))
1076                 RETURN(rc);
1077
1078         isize = i_size_read(inode);
1079         ll_vfs_dq_init(inode);
1080
1081         for (i = 0; i < npages; i++) {
1082                 if (lnb[i].rc == -ENOSPC &&
1083                     osd_is_mapped(inode, lnb[i].lnb_file_offset)) {
1084                         /* Allow the write to proceed if overwriting an
1085                          * existing block */
1086                         lnb[i].rc = 0;
1087                 }
1088
1089                 if (lnb[i].rc) { /* ENOSPC, network RPC error, etc. */
1090                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1091                                lnb[i].rc);
1092                         LASSERT(lnb[i].page);
1093                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
1094                         continue;
1095                 }
1096
1097                 LASSERT(PageLocked(lnb[i].page));
1098                 LASSERT(!PageWriteback(lnb[i].page));
1099
1100                 if (lnb[i].lnb_file_offset + lnb[i].len > isize)
1101                         isize = lnb[i].lnb_file_offset + lnb[i].len;
1102
1103                 /*
1104                  * Since write and truncate are serialized by oo_sem, even
1105                  * partial-page truncate should not leave dirty pages in the
1106                  * page cache.
1107                  */
1108                 LASSERT(!PageDirty(lnb[i].page));
1109
1110                 SetPageUptodate(lnb[i].page);
1111
1112                 osd_iobuf_add_page(iobuf, lnb[i].page);
1113         }
1114
1115         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1116                 rc = -ENOSPC;
1117         } else if (iobuf->dr_npages > 0) {
1118                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1119                                                  iobuf->dr_npages,
1120                                                  iobuf->dr_blocks,
1121                                                  1, NULL);
1122         } else {
1123                 /* no pages to write, no transno is needed */
1124                 thandle->th_local = 1;
1125         }
1126
1127         if (likely(rc == 0)) {
1128                 if (isize > i_size_read(inode)) {
1129                         i_size_write(inode, isize);
1130                         LDISKFS_I(inode)->i_disksize = isize;
1131                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1132                 }
1133
1134                 rc = osd_do_bio(osd, inode, iobuf);
1135                 /* we don't do stats here as in read path because
1136                  * write is async: we'll do this in osd_put_bufs() */
1137         } else {
1138                 osd_fini_iobuf(osd, iobuf);
1139         }
1140
1141         if (unlikely(rc != 0)) {
1142                 /* if write fails, we should drop pages from the cache */
1143                 for (i = 0; i < npages; i++) {
1144                         if (lnb[i].page == NULL)
1145                                 continue;
1146                         LASSERT(PageLocked(lnb[i].page));
1147                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
1148                 }
1149         }
1150
1151         RETURN(rc);
1152 }
1153
1154 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1155                          struct niobuf_local *lnb, int npages)
1156 {
1157         struct osd_thread_info *oti = osd_oti_get(env);
1158         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1159         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1160         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1161         struct timeval start, end;
1162         unsigned long timediff;
1163         int rc = 0, i, m = 0, cache = 0;
1164
1165         LASSERT(inode);
1166
1167         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1168         if (unlikely(rc != 0))
1169                 RETURN(rc);
1170
1171         if (osd->od_read_cache)
1172                 cache = 1;
1173         if (i_size_read(inode) > osd->od_readcache_max_filesize)
1174                 cache = 0;
1175
1176         do_gettimeofday(&start);
1177         for (i = 0; i < npages; i++) {
1178
1179                 if (i_size_read(inode) <= lnb[i].lnb_file_offset)
1180                         /* If there's no more data, abort early.
1181                          * lnb->rc == 0, so it's easy to detect later. */
1182                         break;
1183
1184                 if (i_size_read(inode) <
1185                     lnb[i].lnb_file_offset + lnb[i].len - 1)
1186                         lnb[i].rc = i_size_read(inode) - lnb[i].lnb_file_offset;
1187                 else
1188                         lnb[i].rc = lnb[i].len;
1189                 m += lnb[i].len;
1190
1191                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS, 1);
1192                 if (PageUptodate(lnb[i].page)) {
1193                         lprocfs_counter_add(osd->od_stats,
1194                                             LPROC_OSD_CACHE_HIT, 1);
1195                 } else {
1196                         lprocfs_counter_add(osd->od_stats,
1197                                             LPROC_OSD_CACHE_MISS, 1);
1198                         osd_iobuf_add_page(iobuf, lnb[i].page);
1199                 }
1200                 if (cache == 0)
1201                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
1202         }
1203         do_gettimeofday(&end);
1204         timediff = cfs_timeval_sub(&end, &start, NULL);
1205         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1206
1207         if (iobuf->dr_npages) {
1208                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1209                                                  iobuf->dr_npages,
1210                                                  iobuf->dr_blocks,
1211                                                  0, NULL);
1212                 rc = osd_do_bio(osd, inode, iobuf);
1213
1214                 /* IO stats will be done in osd_bufs_put() */
1215         }
1216
1217         RETURN(rc);
1218 }
1219
1220 /*
1221  * XXX: Another layering violation for now.
1222  *
1223  * We don't want to use ->f_op->read methods, because generic file write
1224  *
1225  *         - serializes on ->i_sem, and
1226  *
1227  *         - does a lot of extra work like balance_dirty_pages(),
1228  *
1229  * which doesn't work for globally shared files like /last_rcvd.
1230  */
1231 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1232 {
1233         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1234
1235         memcpy(buffer, (char *)ei->i_data, buflen);
1236
1237         return  buflen;
1238 }
1239
1240 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1241 {
1242         struct buffer_head *bh;
1243         unsigned long block;
1244         int osize;
1245         int blocksize;
1246         int csize;
1247         int boffs;
1248         int err;
1249
1250         /* prevent reading after eof */
1251         spin_lock(&inode->i_lock);
1252         if (i_size_read(inode) < *offs + size) {
1253                 loff_t diff = i_size_read(inode) - *offs;
1254                 spin_unlock(&inode->i_lock);
1255                 if (diff < 0) {
1256                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1257                                i_size_read(inode), *offs);
1258                         return -EBADR;
1259                 } else if (diff == 0) {
1260                         return 0;
1261                 } else {
1262                         size = diff;
1263                 }
1264         } else {
1265                 spin_unlock(&inode->i_lock);
1266         }
1267
1268         blocksize = 1 << inode->i_blkbits;
1269         osize = size;
1270         while (size > 0) {
1271                 block = *offs >> inode->i_blkbits;
1272                 boffs = *offs & (blocksize - 1);
1273                 csize = min(blocksize - boffs, size);
1274                 bh = ldiskfs_bread(NULL, inode, block, 0, &err);
1275                 if (!bh) {
1276                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %d\n",
1277                                LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
1278                                csize, *offs, inode->i_ino, err);
1279                         return err;
1280                 }
1281
1282                 memcpy(buf, bh->b_data + boffs, csize);
1283                 brelse(bh);
1284
1285                 *offs += csize;
1286                 buf += csize;
1287                 size -= csize;
1288         }
1289         return osize;
1290 }
1291
1292 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1293                         struct lu_buf *buf, loff_t *pos,
1294                         struct lustre_capa *capa)
1295 {
1296         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1297         int           rc;
1298
1299         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_READ))
1300                 RETURN(-EACCES);
1301
1302         /* Read small symlink from inode body as we need to maintain correct
1303          * on-disk symlinks for ldiskfs.
1304          */
1305         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1306             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1307                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1308         else
1309                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1310
1311         return rc;
1312 }
1313
1314 static inline int osd_extents_enabled(struct super_block *sb,
1315                                       struct inode *inode)
1316 {
1317         if (inode != NULL) {
1318                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1319                         return 1;
1320         } else if (test_opt(sb, EXTENTS)) {
1321                 return 1;
1322         }
1323         return 0;
1324 }
1325
1326 static inline int osd_calc_bkmap_credits(struct super_block *sb,
1327                                          struct inode *inode,
1328                                          const loff_t size,
1329                                          const loff_t pos,
1330                                          const int blocks)
1331 {
1332         int credits, bits, bs, i;
1333
1334         bits = sb->s_blocksize_bits;
1335         bs = 1 << bits;
1336
1337         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1338          * we do not expect blockmaps on the large files,
1339          * so let's shrink it to 2 levels (4GB files) */
1340
1341         /* this is default reservation: 2 levels */
1342         credits = (blocks + 2) * 3;
1343
1344         /* actual offset is unknown, hard to optimize */
1345         if (pos == -1)
1346                 return credits;
1347
1348         /* now check for few specific cases to optimize */
1349         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1350                 /* no indirects */
1351                 credits = blocks;
1352                 /* allocate if not allocated */
1353                 if (inode == NULL) {
1354                         credits += blocks * 2;
1355                         return credits;
1356                 }
1357                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1358                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1359                         if (LDISKFS_I(inode)->i_data[i] == 0)
1360                                 credits += 2;
1361                 }
1362         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1363                 /* single indirect */
1364                 credits = blocks * 3;
1365                 /* probably indirect block has been allocated already */
1366                 if (!inode || LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK])
1367                         credits += 3;
1368         }
1369
1370         return credits;
1371 }
1372
1373 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1374                                  const struct lu_buf *buf, loff_t _pos,
1375                                  struct thandle *handle)
1376 {
1377         struct osd_object  *obj  = osd_dt_obj(dt);
1378         struct inode       *inode = obj->oo_inode;
1379         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1380         struct osd_thandle *oh;
1381         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1382         int                 bits, bs;
1383         int                 depth, size;
1384         loff_t              pos;
1385         ENTRY;
1386
1387         LASSERT(buf != NULL);
1388         LASSERT(handle != NULL);
1389
1390         oh = container_of0(handle, struct osd_thandle, ot_super);
1391         LASSERT(oh->ot_handle == NULL);
1392
1393         size = buf->lb_len;
1394         bits = sb->s_blocksize_bits;
1395         bs = 1 << bits;
1396
1397         if (_pos == -1) {
1398                 /* if this is an append, then we
1399                  * should expect cross-block record */
1400                 pos = 0;
1401         } else {
1402                 pos = _pos;
1403         }
1404
1405         /* blocks to modify */
1406         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1407         LASSERT(blocks > 0);
1408
1409         if (inode != NULL && _pos != -1) {
1410                 /* object size in blocks */
1411                 est = (i_size_read(inode) + bs - 1) >> bits;
1412                 allocated = inode->i_blocks >> (bits - 9);
1413                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1414                         /* looks like an overwrite, no need to modify tree */
1415                         credits = blocks;
1416                         /* no need to modify i_size */
1417                         goto out;
1418                 }
1419         }
1420
1421         if (osd_extents_enabled(sb, inode)) {
1422                 /*
1423                  * many concurrent threads may grow tree by the time
1424                  * our transaction starts. so, consider 2 is a min depth
1425                  * for every level we may need to allocate a new block
1426                  * and take some entries from the old one. so, 3 blocks
1427                  * to allocate (bitmap, gd, itself) + old block - 4 per
1428                  * level.
1429                  */
1430                 depth = inode != NULL ? ext_depth(inode) : 0;
1431                 depth = max(depth, 1) + 1;
1432                 credits = depth;
1433                 /* if not append, then split may need to modify
1434                  * existing blocks moving entries into the new ones */
1435                 if (_pos == -1)
1436                         credits += depth;
1437                 /* blocks to store data: bitmap,gd,itself */
1438                 credits += blocks * 3;
1439         } else {
1440                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1441         }
1442         /* if inode is created as part of the transaction,
1443          * then it's counted already by the creation method */
1444         if (inode != NULL)
1445                 credits++;
1446
1447 out:
1448
1449         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1450
1451         /* dt_declare_write() is usually called for system objects, such
1452          * as llog or last_rcvd files. We needn't enforce quota on those
1453          * objects, so always set the lqi_space as 0. */
1454         if (inode != NULL)
1455                 rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid,
1456                                            0, oh, true, true, NULL, false);
1457         RETURN(rc);
1458 }
1459
1460 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1461 {
1462         /* LU-2634: clear the extent format for fast symlink */
1463         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1464
1465         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1466         LDISKFS_I(inode)->i_disksize = buflen;
1467         i_size_write(inode, buflen);
1468         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1469
1470         return 0;
1471 }
1472
1473 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1474                              int write_NUL, loff_t *offs, handle_t *handle)
1475 {
1476         struct buffer_head *bh        = NULL;
1477         loff_t              offset    = *offs;
1478         loff_t              new_size  = i_size_read(inode);
1479         unsigned long       block;
1480         int                 blocksize = 1 << inode->i_blkbits;
1481         int                 err = 0;
1482         int                 size;
1483         int                 boffs;
1484         int                 dirty_inode = 0;
1485
1486         if (write_NUL) {
1487                 /*
1488                  * long symlink write does not count the NUL terminator in
1489                  * bufsize, we write it, and the inode's file size does not
1490                  * count the NUL terminator as well.
1491                  */
1492                 ((char *)buf)[bufsize] = '\0';
1493                 ++bufsize;
1494         }
1495         while (bufsize > 0) {
1496                 if (bh != NULL)
1497                         brelse(bh);
1498
1499                 block = offset >> inode->i_blkbits;
1500                 boffs = offset & (blocksize - 1);
1501                 size = min(blocksize - boffs, bufsize);
1502                 bh = ldiskfs_bread(handle, inode, block, 1, &err);
1503                 if (!bh) {
1504                         CERROR("%s: error reading offset %llu (block %lu): "
1505                                "rc = %d\n",
1506                                inode->i_sb->s_id, offset, block, err);
1507                         break;
1508                 }
1509
1510                 err = ldiskfs_journal_get_write_access(handle, bh);
1511                 if (err) {
1512                         CERROR("journal_get_write_access() returned error %d\n",
1513                                err);
1514                         break;
1515                 }
1516                 LASSERTF(boffs + size <= bh->b_size,
1517                          "boffs %d size %d bh->b_size %lu",
1518                          boffs, size, (unsigned long)bh->b_size);
1519                 memcpy(bh->b_data + boffs, buf, size);
1520                 err = ldiskfs_journal_dirty_metadata(handle, bh);
1521                 if (err)
1522                         break;
1523
1524                 if (offset + size > new_size)
1525                         new_size = offset + size;
1526                 offset += size;
1527                 bufsize -= size;
1528                 buf += size;
1529         }
1530         if (bh)
1531                 brelse(bh);
1532
1533         if (write_NUL)
1534                 --new_size;
1535         /* correct in-core and on-disk sizes */
1536         if (new_size > i_size_read(inode)) {
1537                 spin_lock(&inode->i_lock);
1538                 if (new_size > i_size_read(inode))
1539                         i_size_write(inode, new_size);
1540                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1541                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1542                         dirty_inode = 1;
1543                 }
1544                 spin_unlock(&inode->i_lock);
1545                 if (dirty_inode)
1546                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1547         }
1548
1549         if (err == 0)
1550                 *offs = offset;
1551         return err;
1552 }
1553
1554 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1555                          const struct lu_buf *buf, loff_t *pos,
1556                          struct thandle *handle, struct lustre_capa *capa,
1557                          int ignore_quota)
1558 {
1559         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1560         struct osd_thandle      *oh;
1561         ssize_t                 result;
1562         int                     is_link;
1563
1564         LASSERT(dt_object_exists(dt));
1565
1566         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_WRITE))
1567                 return -EACCES;
1568
1569         LASSERT(handle != NULL);
1570         LASSERT(inode != NULL);
1571         ll_vfs_dq_init(inode);
1572
1573         /* XXX: don't check: one declared chunk can be used many times */
1574         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1575
1576         oh = container_of(handle, struct osd_thandle, ot_super);
1577         LASSERT(oh->ot_handle->h_transaction != NULL);
1578         /* Write small symlink to inode body as we need to maintain correct
1579          * on-disk symlinks for ldiskfs.
1580          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1581          * does not count it in.
1582          */
1583         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1584         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1585                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1586         else
1587                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1588                                                   buf->lb_len, is_link, pos,
1589                                                   oh->ot_handle);
1590         if (result == 0)
1591                 result = buf->lb_len;
1592         return result;
1593 }
1594
1595 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1596                              __u64 start, __u64 end, struct thandle *th)
1597 {
1598         struct osd_thandle *oh;
1599         struct inode       *inode;
1600         int                 rc;
1601         ENTRY;
1602
1603         LASSERT(th);
1604         oh = container_of(th, struct osd_thandle, ot_super);
1605
1606         /*
1607          * we don't need to reserve credits for whole truncate
1608          * it's not possible as truncate may need to free too many
1609          * blocks and that won't fit a single transaction. instead
1610          * we reserve credits to change i_size and put inode onto
1611          * orphan list. if needed truncate will extend or restart
1612          * transaction
1613          */
1614         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1615                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1616
1617         inode = osd_dt_obj(dt)->oo_inode;
1618         LASSERT(inode);
1619
1620         rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
1621                                    true, true, NULL, false);
1622         RETURN(rc);
1623 }
1624
1625 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1626                      __u64 start, __u64 end, struct thandle *th,
1627                      struct lustre_capa *capa)
1628 {
1629         struct osd_thandle *oh;
1630         struct osd_object  *obj = osd_dt_obj(dt);
1631         struct inode       *inode = obj->oo_inode;
1632         handle_t           *h;
1633         tid_t               tid;
1634         int                rc = 0, rc2 = 0;
1635         ENTRY;
1636
1637         LASSERT(end == OBD_OBJECT_EOF);
1638         LASSERT(dt_object_exists(dt));
1639         LASSERT(osd_invariant(obj));
1640         LASSERT(inode != NULL);
1641         ll_vfs_dq_init(inode);
1642
1643         LASSERT(th);
1644         oh = container_of(th, struct osd_thandle, ot_super);
1645         LASSERT(oh->ot_handle->h_transaction != NULL);
1646
1647         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1648
1649         tid = oh->ot_handle->h_transaction->t_tid;
1650
1651         i_size_write(inode, start);
1652         ll_truncate_pagecache(inode, start);
1653 #ifdef HAVE_INODEOPS_TRUNCATE
1654         if (inode->i_op->truncate) {
1655                 inode->i_op->truncate(inode);
1656         } else
1657 #endif
1658                 ldiskfs_truncate(inode);
1659
1660         /*
1661          * For a partial-page truncate, flush the page to disk immediately to
1662          * avoid data corruption during direct disk write.  b=17397
1663          */
1664         if ((start & ~CFS_PAGE_MASK) != 0)
1665                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1666
1667         h = journal_current_handle();
1668         LASSERT(h != NULL);
1669         LASSERT(h == oh->ot_handle);
1670
1671         if (tid != h->h_transaction->t_tid) {
1672                 int credits = oh->ot_credits;
1673                 /*
1674                  * transaction has changed during truncate
1675                  * we need to restart the handle with our credits
1676                  */
1677                 if (h->h_buffer_credits < credits) {
1678                         if (ldiskfs_journal_extend(h, credits))
1679                                 rc2 = ldiskfs_journal_restart(h, credits);
1680                 }
1681         }
1682
1683         RETURN(rc == 0 ? rc2 : rc);
1684 }
1685
1686 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1687                           struct ll_user_fiemap *fm)
1688 {
1689         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1690         struct osd_thread_info *info   = osd_oti_get(env);
1691         struct dentry          *dentry = &info->oti_obj_dentry;
1692         struct file            *file   = &info->oti_file;
1693         mm_segment_t            saved_fs;
1694         int rc;
1695
1696         LASSERT(inode);
1697         dentry->d_inode = inode;
1698         dentry->d_sb = inode->i_sb;
1699         file->f_dentry = dentry;
1700         file->f_mapping = inode->i_mapping;
1701         file->f_op = inode->i_fop;
1702         set_file_inode(file, inode);
1703
1704         saved_fs = get_fs();
1705         set_fs(get_ds());
1706         /* ldiskfs_ioctl does not have a inode argument */
1707         if (inode->i_fop->unlocked_ioctl)
1708                 rc = inode->i_fop->unlocked_ioctl(file, FSFILT_IOC_FIEMAP,
1709                                                   (long)fm);
1710         else
1711                 rc = -ENOTTY;
1712         set_fs(saved_fs);
1713         return rc;
1714 }
1715
1716 /*
1717  * in some cases we may need declare methods for objects being created
1718  * e.g., when we create symlink
1719  */
1720 const struct dt_body_operations osd_body_ops_new = {
1721         .dbo_declare_write = osd_declare_write,
1722 };
1723
1724 const struct dt_body_operations osd_body_ops = {
1725         .dbo_read                 = osd_read,
1726         .dbo_declare_write        = osd_declare_write,
1727         .dbo_write                = osd_write,
1728         .dbo_bufs_get             = osd_bufs_get,
1729         .dbo_bufs_put             = osd_bufs_put,
1730         .dbo_write_prep           = osd_write_prep,
1731         .dbo_declare_write_commit = osd_declare_write_commit,
1732         .dbo_write_commit         = osd_write_commit,
1733         .dbo_read_prep            = osd_read_prep,
1734         .dbo_declare_punch         = osd_declare_punch,
1735         .dbo_punch                 = osd_punch,
1736         .dbo_fiemap_get           = osd_fiemap_get,
1737 };
1738