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 <https://jira.hpdd.intel.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->lnb_len = plen;
 399                 /* lnb->lnb_flags = rnb->rnb_flags; */
 400                 lnb->lnb_flags = 0;
 401                 lnb->lnb_page = NULL;
 402                 lnb->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                 lnb->lnb_page = osd_get_page(d, lnb->lnb_file_offset, rw);
 462                 if (lnb->lnb_page == NULL)
 463                         GOTO(cleanup, rc = -ENOMEM);
 464
 465                 /* DLM locking protects us from write and truncate competing
 466                  * for same region, but truncate can leave dirty page in the
 467                  * cache. it's possible the writeout on a such a page is in
 468                  * progress when we access it. it's also possible that during
 469                  * this writeout we put new (partial) data, but then won't
 470                  * be able to proceed in filter_commitrw_write(). thus let's
 471                  * just wait for writeout completion, should be rare enough.
 472                  * -bzzz */
 473                 wait_on_page_writeback(lnb->lnb_page);
 474                 BUG_ON(PageWriteback(lnb->lnb_page));
 475
 476                 lu_object_get(&d->do_lu);
 477         }
 478         rc = i;
 479
 480 cleanup:
 481         RETURN(rc);
 482 }
 483
 484 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 485                         struct niobuf_local *lnb, int npages)
 486 {
 487         int i;
 488
 489         for (i = 0; i < npages; i++) {
 490                 if (lnb[i].lnb_page == NULL)
 491                         continue;
 492                 LASSERT(PageLocked(lnb[i].lnb_page));
 493                 unlock_page(lnb[i].lnb_page);
 494                 page_cache_release(lnb[i].lnb_page);
 495                 lu_object_put(env, &dt->do_lu);
 496                 lnb[i].lnb_page = NULL;
 497         }
 498
 499         RETURN(0);
 500 }
 501
 502 #ifndef HAVE_LDISKFS_MAP_BLOCKS
 503
 504 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 505 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 506 #endif
 507
 508 struct bpointers {
 509         unsigned long *blocks;
 510         unsigned long start;
 511         int num;
 512         int init_num;
 513         int create;
 514 };
 515
 516 static long ldiskfs_ext_find_goal(struct inode *inode,
 517                                   struct ldiskfs_ext_path *path,
 518                                   unsigned long block, int *aflags)
 519 {
 520         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 521         unsigned long bg_start;
 522         unsigned long colour;
 523         int depth;
 524
 525         if (path) {
 526                 struct ldiskfs_extent *ex;
 527                 depth = path->p_depth;
 528
 529                 /* try to predict block placement */
 530                 if ((ex = path[depth].p_ext))
 531                         return ldiskfs_ext_pblock(ex) +
 532                                 (block - le32_to_cpu(ex->ee_block));
 533
 534                 /* it looks index is empty
 535                  * try to find starting from index itself */
 536                 if (path[depth].p_bh)
 537                         return path[depth].p_bh->b_blocknr;
 538         }
 539
 540         /* OK. use inode's group */
 541         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 542                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 543         colour = (current->pid % 16) *
 544                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 545         return bg_start + colour + block;
 546 }
 547
 548 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 549                                 struct ldiskfs_ext_path *path,
 550                                 unsigned long block, unsigned long *count,
 551                                 int *err)
 552 {
 553         struct ldiskfs_allocation_request ar;
 554         unsigned long pblock;
 555         int aflags;
 556
 557         /* find neighbour allocated blocks */
 558         ar.lleft = block;
 559         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 560         if (*err)
 561                 return 0;
 562         ar.lright = block;
 563         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 564         if (*err)
 565                 return 0;
 566
 567         /* allocate new block */
 568         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 569         ar.inode = inode;
 570         ar.logical = block;
 571         ar.len = *count;
 572         ar.flags = LDISKFS_MB_HINT_DATA;
 573         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 574         *count = ar.len;
 575         return pblock;
 576 }
 577
 578 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 579                                      struct ldiskfs_ext_path *path,
 580                                      struct ldiskfs_ext_cache *cex,
 581 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 582                                      struct ldiskfs_extent *ex,
 583 #endif
 584                                      void *cbdata)
 585 {
 586         struct bpointers *bp = cbdata;
 587         struct ldiskfs_extent nex;
 588         unsigned long pblock;
 589         unsigned long tgen;
 590         int err, i;
 591         unsigned long count;
 592         handle_t *handle;
 593
 594 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 595         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 596 #else
 597         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 598 #endif
 599                 err = EXT_CONTINUE;
 600                 goto map;
 601         }
 602
 603         if (bp->create == 0) {
 604                 i = 0;
 605                 if (cex->ec_block < bp->start)
 606                         i = bp->start - cex->ec_block;
 607                 if (i >= cex->ec_len)
 608                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 609                                         i, cex->ec_len);
 610                 for (; i < cex->ec_len && bp->num; i++) {
 611                         *(bp->blocks) = 0;
 612                         bp->blocks++;
 613                         bp->num--;
 614                         bp->start++;
 615                 }
 616
 617                 return EXT_CONTINUE;
 618         }
 619
 620         tgen = LDISKFS_I(inode)->i_ext_generation;
 621         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
 622
 623         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
 624                                    count + LDISKFS_ALLOC_NEEDED + 1);
 625         if (IS_ERR(handle)) {
 626                 return PTR_ERR(handle);
 627         }
 628
 629         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
 630                 /* the tree has changed. so path can be invalid at moment */
 631                 ldiskfs_journal_stop(handle);
 632                 return EXT_REPEAT;
 633         }
 634
 635         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
 636          * protected by i_data_sem as whole. so we patch it to store
 637          * generation to path and now verify the tree hasn't changed */
 638         down_write((&LDISKFS_I(inode)->i_data_sem));
 639
 640         /* validate extent, make sure the extent tree does not changed */
 641         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
 642                 /* cex is invalid, try again */
 643                 up_write(&LDISKFS_I(inode)->i_data_sem);
 644                 ldiskfs_journal_stop(handle);
 645                 return EXT_REPEAT;
 646         }
 647
 648         count = cex->ec_len;
 649         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
 650         if (!pblock)
 651                 goto out;
 652         BUG_ON(count > cex->ec_len);
 653
 654         /* insert new extent */
 655         nex.ee_block = cpu_to_le32(cex->ec_block);
 656         ldiskfs_ext_store_pblock(&nex, pblock);
 657         nex.ee_len = cpu_to_le16(count);
 658         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
 659         if (err) {
 660                 /* free data blocks we just allocated */
 661                 /* not a good idea to call discard here directly,
 662                  * but otherwise we'd need to call it every free() */
 663                 ldiskfs_discard_preallocations(inode);
 664 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
 665                 ldiskfs_free_blocks(handle, inode, NULL,
 666                                     ldiskfs_ext_pblock(&nex),
 667                                     le16_to_cpu(nex.ee_len), 0);
 668 #else
 669                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
 670                                     le16_to_cpu(nex.ee_len), 0);
 671 #endif
 672                 goto out;
 673         }
 674
 675         /*
 676          * Putting len of the actual extent we just inserted,
 677          * we are asking ldiskfs_ext_walk_space() to continue
 678          * scaning after that block
 679          */
 680         cex->ec_len = le16_to_cpu(nex.ee_len);
 681         cex->ec_start = ldiskfs_ext_pblock(&nex);
 682         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
 683         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
 684
 685 out:
 686         up_write((&LDISKFS_I(inode)->i_data_sem));
 687         ldiskfs_journal_stop(handle);
 688 map:
 689         if (err >= 0) {
 690                 /* map blocks */
 691                 if (bp->num == 0) {
 692                         CERROR("hmm. why do we find this extent?\n");
 693                         CERROR("initial space: %lu:%u\n",
 694                                 bp->start, bp->init_num);
 695 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 696                         CERROR("current extent: %u/%u/%llu %d\n",
 697                                 cex->ec_block, cex->ec_len,
 698                                 (unsigned long long)cex->ec_start,
 699                                 cex->ec_type);
 700 #else
 701                         CERROR("current extent: %u/%u/%llu\n",
 702                                 cex->ec_block, cex->ec_len,
 703                                 (unsigned long long)cex->ec_start);
 704 #endif
 705                 }
 706                 i = 0;
 707                 if (cex->ec_block < bp->start)
 708                         i = bp->start - cex->ec_block;
 709                 if (i >= cex->ec_len)
 710                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 711                                         i, cex->ec_len);
 712                 for (; i < cex->ec_len && bp->num; i++) {
 713                         *(bp->blocks) = cex->ec_start + i;
 714 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 715                         if (cex->ec_type != LDISKFS_EXT_CACHE_EXTENT) {
 716 #else
 717                         if ((cex->ec_len == 0) || (cex->ec_start == 0)) {
 718 #endif
 719                                 /* unmap any possible underlying metadata from
 720                                  * the block device mapping.  bug 6998. */
 721                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
 722                                                           *(bp->blocks));
 723                         }
 724                         bp->blocks++;
 725                         bp->num--;
 726                         bp->start++;
 727                 }
 728         }
 729         return err;
 730 }
 731
 732 int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long block,
 733                             unsigned long num, unsigned long *blocks,
 734                             int create)
 735 {
 736         struct bpointers bp;
 737         int err;
 738
 739         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 740                block, block + num - 1, (unsigned) inode->i_ino);
 741
 742         bp.blocks = blocks;
 743         bp.start = block;
 744         bp.init_num = bp.num = num;
 745         bp.create = create;
 746
 747         err = ldiskfs_ext_walk_space(inode, block, num,
 748                                          ldiskfs_ext_new_extent_cb, &bp);
 749         ldiskfs_ext_invalidate_cache(inode);
 750
 751         return err;
 752 }
 753
 754 int osd_ldiskfs_map_bm_inode_pages(struct inode *inode, struct page **page,
 755                                    int pages, unsigned long *blocks,
 756                                    int create)
 757 {
 758         int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
 759         unsigned long *b;
 760         int rc = 0, i;
 761
 762         for (i = 0, b = blocks; i < pages; i++, page++) {
 763                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
 764                 if (rc) {
 765                         CERROR("ino %lu, blk %lu create %d: rc %d\n",
 766                                inode->i_ino, *b, create, rc);
 767                         break;
 768                 }
 769                 b += blocks_per_page;
 770         }
 771         return rc;
 772 }
 773
 774 int osd_ldiskfs_map_ext_inode_pages(struct inode *inode, struct page **page,
 775                                     int pages, unsigned long *blocks,
 776                                     int create)
 777 {
 778         int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
 779         int rc = 0, i = 0;
 780         struct page *fp = NULL;
 781         int clen = 0;
 782
 783         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 784                 inode->i_ino, pages, (*page)->index);
 785
 786         /* pages are sorted already. so, we just have to find
 787          * contig. space and process them properly */
 788         while (i < pages) {
 789                 if (fp == NULL) {
 790                         /* start new extent */
 791                         fp = *page++;
 792                         clen = 1;
 793                         i++;
 794                         continue;
 795                 } else if (fp->index + clen == (*page)->index) {
 796                         /* continue the extent */
 797                         page++;
 798                         clen++;
 799                         i++;
 800                         continue;
 801                 }
 802
 803                 /* process found extent */
 804                 rc = osd_ldiskfs_map_nblocks(inode, fp->index * blocks_per_page,
 805                                              clen * blocks_per_page, blocks,
 806                                              create);
 807                 if (rc)
 808                         GOTO(cleanup, rc);
 809
 810                 /* look for next extent */
 811                 fp = NULL;
 812                 blocks += blocks_per_page * clen;
 813         }
 814
 815         if (fp)
 816                 rc = osd_ldiskfs_map_nblocks(inode, fp->index * blocks_per_page,
 817                                              clen * blocks_per_page, blocks,
 818                                              create);
 819 cleanup:
 820         return rc;
 821 }
 822
 823 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 824                                        int pages, unsigned long *blocks,
 825                                        int create)
 826 {
 827         int rc;
 828
 829         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 830                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
 831                                                      blocks, create);
 832                 return rc;
 833         }
 834         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
 835
 836         return rc;
 837 }
 838 #else
 839 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 840                                        int pages, unsigned long *blocks,
 841                                        int create)
 842 {
 843         int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
 844         int rc = 0, i = 0;
 845         struct page *fp = NULL;
 846         int clen = 0;
 847
 848         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 849                 inode->i_ino, pages, (*page)->index);
 850
 851         /* pages are sorted already. so, we just have to find
 852          * contig. space and process them properly */
 853         while (i < pages) {
 854                 long blen, total = 0;
 855                 handle_t *handle = NULL;
 856                 struct ldiskfs_map_blocks map = { 0 };
 857
 858                 if (fp == NULL) { /* start new extent */
 859                         fp = *page++;
 860                         clen = 1;
 861                         if (++i != pages)
 862                                 continue;
 863                 } else if (fp->index + clen == (*page)->index) {
 864                         /* continue the extent */
 865                         page++;
 866                         clen++;
 867                         if (++i != pages)
 868                                 continue;
 869                 }
 870                 /* process found extent */
 871                 map.m_lblk = fp->index * blocks_per_page;
 872                 map.m_len = blen = clen * blocks_per_page;
 873                 if (create) {
 874                         create = LDISKFS_GET_BLOCKS_CREATE;
 875                         handle = ldiskfs_journal_current_handle();
 876                         LASSERT(handle != NULL);
 877                 }
 878 cont_map:
 879                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 880                 if (rc >= 0) {
 881                         int c = 0;
 882                         for (; total < blen && c < map.m_len; c++, total++) {
 883                                 if (rc == 0) {
 884                                         *(blocks + total) = 0;
 885                                         total++;
 886                                         break;
 887                                 } else {
 888                                         *(blocks + total) = map.m_pblk + c;
 889                                         /* unmap any possible underlying
 890                                          * metadata from the block device
 891                                          * mapping.  bug 6998. */
 892                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
 893                                             create)
 894                                                 unmap_underlying_metadata(
 895                                                         inode->i_sb->s_bdev,
 896                                                         map.m_pblk + c);
 897                                 }
 898                         }
 899                         rc = 0;
 900                 }
 901                 if (rc == 0 && total < blen) {
 902                         map.m_lblk = fp->index * blocks_per_page + total;
 903                         map.m_len = blen - total;
 904                         goto cont_map;
 905                 }
 906                 if (rc != 0)
 907                         GOTO(cleanup, rc);
 908
 909                 /* look for next extent */
 910                 fp = NULL;
 911                 blocks += blocks_per_page * clen;
 912         }
 913 cleanup:
 914         return rc;
 915 }
 916 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
 917
 918 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 919                           struct niobuf_local *lnb, int npages)
 920 {
 921         struct osd_thread_info *oti   = osd_oti_get(env);
 922         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 923         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 924         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 925         struct timeval          start;
 926         struct timeval          end;
 927         unsigned long           timediff;
 928         ssize_t                 isize;
 929         __s64                   maxidx;
 930         int                     rc = 0;
 931         int                     i;
 932         int                     cache = 0;
 933
 934         LASSERT(inode);
 935
 936         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 937         if (unlikely(rc != 0))
 938                 RETURN(rc);
 939
 940         isize = i_size_read(inode);
 941         maxidx = ((isize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) - 1;
 942
 943         if (osd->od_writethrough_cache)
 944                 cache = 1;
 945         if (isize > osd->od_readcache_max_filesize)
 946                 cache = 0;
 947
 948         do_gettimeofday(&start);
 949         for (i = 0; i < npages; i++) {
 950
 951                 if (cache == 0)
 952                         generic_error_remove_page(inode->i_mapping,
 953                                                   lnb[i].lnb_page);
 954
 955                 /*
 956                  * till commit the content of the page is undefined
 957                  * we'll set it uptodate once bulk is done. otherwise
 958                  * subsequent reads can access non-stable data
 959                  */
 960                 ClearPageUptodate(lnb[i].lnb_page);
 961
 962                 if (lnb[i].lnb_len == PAGE_CACHE_SIZE)
 963                         continue;
 964
 965                 if (maxidx >= lnb[i].lnb_page->index) {
 966                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
 967                 } else {
 968                         long off;
 969                         char *p = kmap(lnb[i].lnb_page);
 970
 971                         off = lnb[i].lnb_page_offset;
 972                         if (off)
 973                                 memset(p, 0, off);
 974                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
 975                               ~CFS_PAGE_MASK;
 976                         if (off)
 977                                 memset(p + off, 0, PAGE_CACHE_SIZE - off);
 978                         kunmap(lnb[i].lnb_page);
 979                 }
 980         }
 981         do_gettimeofday(&end);
 982         timediff = cfs_timeval_sub(&end, &start, NULL);
 983         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 984
 985         if (iobuf->dr_npages) {
 986                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
 987                                                  iobuf->dr_npages,
 988                                                  iobuf->dr_blocks, 0);
 989                 if (likely(rc == 0)) {
 990                         rc = osd_do_bio(osd, inode, iobuf);
 991                         /* do IO stats for preparation reads */
 992                         osd_fini_iobuf(osd, iobuf);
 993                 }
 994         }
 995         RETURN(rc);
 996 }
 997
 998 /* Check if a block is allocated or not */
 999 static int osd_is_mapped(struct inode *inode, obd_size offset)
1000 {
1001         sector_t (*fs_bmap)(struct address_space *, sector_t);
1002
1003         fs_bmap = inode->i_mapping->a_ops->bmap;
1004
1005         /* We can't know if we are overwriting or not */
1006         if (unlikely(fs_bmap == NULL))
1007                 return 0;
1008
1009         if (i_size_read(inode) == 0)
1010                 return 0;
1011
1012         /* Beyond EOF, must not be mapped */
1013         if (((i_size_read(inode) - 1) >> inode->i_blkbits) <
1014             (offset >> inode->i_blkbits))
1015                 return 0;
1016
1017         if (fs_bmap(inode->i_mapping, offset >> inode->i_blkbits) == 0)
1018                 return 0;
1019
1020         return 1;
1021 }
1022
1023 static int osd_declare_write_commit(const struct lu_env *env,
1024                                     struct dt_object *dt,
1025                                     struct niobuf_local *lnb, int npages,
1026                                     struct thandle *handle)
1027 {
1028         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1029         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1030         struct osd_thandle      *oh;
1031         int                      extents = 1;
1032         int                      depth;
1033         int                      i;
1034         int                      newblocks;
1035         int                      rc = 0;
1036         int                      flags = 0;
1037         bool                     ignore_quota = false;
1038         long long                quota_space = 0;
1039         ENTRY;
1040
1041         LASSERT(handle != NULL);
1042         oh = container_of0(handle, struct osd_thandle, ot_super);
1043         LASSERT(oh->ot_handle == NULL);
1044
1045         newblocks = npages;
1046
1047         /* calculate number of extents (probably better to pass nb) */
1048         for (i = 0; i < npages; i++) {
1049                 if (i && lnb[i].lnb_file_offset !=
1050                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1051                         extents++;
1052
1053                 if (!osd_is_mapped(inode, lnb[i].lnb_file_offset))
1054                         quota_space += PAGE_CACHE_SIZE;
1055
1056                 /* ignore quota for the whole request if any page is from
1057                  * client cache or written by root.
1058                  *
1059                  * XXX once we drop the 1.8 client support, the checking
1060                  * for whether page is from cache can be simplified as:
1061                  * !(lnb[i].flags & OBD_BRW_SYNC)
1062                  *
1063                  * XXX we could handle this on per-lnb basis as done by
1064                  * grant. */
1065                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1066                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1067                     OBD_BRW_FROM_GRANT)
1068                         ignore_quota = true;
1069         }
1070
1071         /*
1072          * each extent can go into new leaf causing a split
1073          * 5 is max tree depth: inode + 4 index blocks
1074          * with blockmaps, depth is 3 at most
1075          */
1076         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1077                 /*
1078                  * many concurrent threads may grow tree by the time
1079                  * our transaction starts. so, consider 2 is a min depth
1080                  */
1081                 depth = ext_depth(inode);
1082                 depth = max(depth, 1) + 1;
1083                 newblocks += depth;
1084                 oh->ot_credits++; /* inode */
1085                 oh->ot_credits += depth * 2 * extents;
1086         } else {
1087                 depth = 3;
1088                 newblocks += depth;
1089                 oh->ot_credits++; /* inode */
1090                 oh->ot_credits += depth * extents;
1091         }
1092
1093         /* quota space for metadata blocks */
1094         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1095
1096         /* quota space should be reported in 1K blocks */
1097         quota_space = toqb(quota_space);
1098
1099         /* each new block can go in different group (bitmap + gd) */
1100
1101         /* we can't dirty more bitmap blocks than exist */
1102         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1103                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1104         else
1105                 oh->ot_credits += newblocks;
1106
1107         /* we can't dirty more gd blocks than exist */
1108         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1109                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1110         else
1111                 oh->ot_credits += newblocks;
1112
1113         /* make sure the over quota flags were not set */
1114         lnb[0].lnb_flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA);
1115
1116         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1117                                    quota_space, oh, true, true, &flags,
1118                                    ignore_quota);
1119
1120         /* we need only to store the overquota flags in the first lnb for
1121          * now, once we support multiple objects BRW, this code needs be
1122          * revised. */
1123         if (flags & QUOTA_FL_OVER_USRQUOTA)
1124                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1125         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1126                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1127
1128         RETURN(rc);
1129 }
1130
1131 /* Check if a block is allocated or not */
1132 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1133                             struct niobuf_local *lnb, int npages,
1134                             struct thandle *thandle)
1135 {
1136         struct osd_thread_info *oti = osd_oti_get(env);
1137         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1138         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1139         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1140         loff_t isize;
1141         int rc = 0, i;
1142
1143         LASSERT(inode);
1144
1145         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1146         if (unlikely(rc != 0))
1147                 RETURN(rc);
1148
1149         isize = i_size_read(inode);
1150         ll_vfs_dq_init(inode);
1151
1152         for (i = 0; i < npages; i++) {
1153                 if (lnb[i].lnb_rc == -ENOSPC &&
1154                     osd_is_mapped(inode, lnb[i].lnb_file_offset)) {
1155                         /* Allow the write to proceed if overwriting an
1156                          * existing block */
1157                         lnb[i].lnb_rc = 0;
1158                 }
1159
1160                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1161                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1162                                lnb[i].lnb_rc);
1163                         LASSERT(lnb[i].lnb_page);
1164                         generic_error_remove_page(inode->i_mapping,
1165                                                   lnb[i].lnb_page);
1166                         continue;
1167                 }
1168
1169                 LASSERT(PageLocked(lnb[i].lnb_page));
1170                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1171
1172                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1173                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1174
1175                 /*
1176                  * Since write and truncate are serialized by oo_sem, even
1177                  * partial-page truncate should not leave dirty pages in the
1178                  * page cache.
1179                  */
1180                 LASSERT(!PageDirty(lnb[i].lnb_page));
1181
1182                 SetPageUptodate(lnb[i].lnb_page);
1183
1184                 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1185         }
1186
1187         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1188                 rc = -ENOSPC;
1189         } else if (iobuf->dr_npages > 0) {
1190                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1191                                                  iobuf->dr_npages,
1192                                                  iobuf->dr_blocks, 1);
1193         } else {
1194                 /* no pages to write, no transno is needed */
1195                 thandle->th_local = 1;
1196         }
1197
1198         if (likely(rc == 0)) {
1199                 if (isize > i_size_read(inode)) {
1200                         i_size_write(inode, isize);
1201                         LDISKFS_I(inode)->i_disksize = isize;
1202                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1203                 }
1204
1205                 rc = osd_do_bio(osd, inode, iobuf);
1206                 /* we don't do stats here as in read path because
1207                  * write is async: we'll do this in osd_put_bufs() */
1208         } else {
1209                 osd_fini_iobuf(osd, iobuf);
1210         }
1211
1212         if (unlikely(rc != 0)) {
1213                 /* if write fails, we should drop pages from the cache */
1214                 for (i = 0; i < npages; i++) {
1215                         if (lnb[i].lnb_page == NULL)
1216                                 continue;
1217                         LASSERT(PageLocked(lnb[i].lnb_page));
1218                         generic_error_remove_page(inode->i_mapping,
1219                                                   lnb[i].lnb_page);
1220                 }
1221         }
1222
1223         RETURN(rc);
1224 }
1225
1226 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1227                          struct niobuf_local *lnb, int npages)
1228 {
1229         struct osd_thread_info *oti = osd_oti_get(env);
1230         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1231         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1232         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1233         struct timeval start, end;
1234         unsigned long timediff;
1235         int rc = 0, i, m = 0, cache = 0, cache_hits = 0, cache_misses = 0;
1236
1237         LASSERT(inode);
1238
1239         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1240         if (unlikely(rc != 0))
1241                 RETURN(rc);
1242
1243         if (osd->od_read_cache)
1244                 cache = 1;
1245         if (i_size_read(inode) > osd->od_readcache_max_filesize)
1246                 cache = 0;
1247
1248         do_gettimeofday(&start);
1249         for (i = 0; i < npages; i++) {
1250
1251                 if (i_size_read(inode) <= lnb[i].lnb_file_offset)
1252                         /* If there's no more data, abort early.
1253                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1254                         break;
1255
1256                 if (i_size_read(inode) <
1257                     lnb[i].lnb_file_offset + lnb[i].lnb_len - 1)
1258                         lnb[i].lnb_rc = i_size_read(inode) -
1259                                 lnb[i].lnb_file_offset;
1260                 else
1261                         lnb[i].lnb_rc = lnb[i].lnb_len;
1262                 m += lnb[i].lnb_len;
1263
1264                 if (PageUptodate(lnb[i].lnb_page)) {
1265                         cache_hits++;
1266                 } else {
1267                         cache_misses++;
1268                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1269                 }
1270
1271                 if (cache == 0)
1272                         generic_error_remove_page(inode->i_mapping,
1273                                                   lnb[i].lnb_page);
1274         }
1275         do_gettimeofday(&end);
1276         timediff = cfs_timeval_sub(&end, &start, NULL);
1277         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1278
1279         if (cache_hits != 0)
1280                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1281                                     cache_hits);
1282         if (cache_misses != 0)
1283                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1284                                     cache_misses);
1285         if (cache_hits + cache_misses != 0)
1286                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1287                                     cache_hits + cache_misses);
1288
1289         if (iobuf->dr_npages) {
1290                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1291                                                  iobuf->dr_npages,
1292                                                  iobuf->dr_blocks, 0);
1293                 rc = osd_do_bio(osd, inode, iobuf);
1294
1295                 /* IO stats will be done in osd_bufs_put() */
1296         }
1297
1298         RETURN(rc);
1299 }
1300
1301 /*
1302  * XXX: Another layering violation for now.
1303  *
1304  * We don't want to use ->f_op->read methods, because generic file write
1305  *
1306  *         - serializes on ->i_sem, and
1307  *
1308  *         - does a lot of extra work like balance_dirty_pages(),
1309  *
1310  * which doesn't work for globally shared files like /last_rcvd.
1311  */
1312 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1313 {
1314         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1315
1316         memcpy(buffer, (char *)ei->i_data, buflen);
1317
1318         return  buflen;
1319 }
1320
1321 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1322 {
1323         struct buffer_head *bh;
1324         unsigned long block;
1325         int osize;
1326         int blocksize;
1327         int csize;
1328         int boffs;
1329         int err;
1330
1331         /* prevent reading after eof */
1332         spin_lock(&inode->i_lock);
1333         if (i_size_read(inode) < *offs + size) {
1334                 loff_t diff = i_size_read(inode) - *offs;
1335                 spin_unlock(&inode->i_lock);
1336                 if (diff < 0) {
1337                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1338                                i_size_read(inode), *offs);
1339                         return -EBADR;
1340                 } else if (diff == 0) {
1341                         return 0;
1342                 } else {
1343                         size = diff;
1344                 }
1345         } else {
1346                 spin_unlock(&inode->i_lock);
1347         }
1348
1349         blocksize = 1 << inode->i_blkbits;
1350         osize = size;
1351         while (size > 0) {
1352                 block = *offs >> inode->i_blkbits;
1353                 boffs = *offs & (blocksize - 1);
1354                 csize = min(blocksize - boffs, size);
1355                 bh = ldiskfs_bread(NULL, inode, block, 0, &err);
1356                 if (!bh) {
1357                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %d\n",
1358                                LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
1359                                csize, *offs, inode->i_ino, err);
1360                         return err;
1361                 }
1362
1363                 memcpy(buf, bh->b_data + boffs, csize);
1364                 brelse(bh);
1365
1366                 *offs += csize;
1367                 buf += csize;
1368                 size -= csize;
1369         }
1370         return osize;
1371 }
1372
1373 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1374                         struct lu_buf *buf, loff_t *pos,
1375                         struct lustre_capa *capa)
1376 {
1377         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1378         int           rc;
1379
1380         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_READ))
1381                 RETURN(-EACCES);
1382
1383         /* Read small symlink from inode body as we need to maintain correct
1384          * on-disk symlinks for ldiskfs.
1385          */
1386         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1387             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1388                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1389         else
1390                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1391
1392         return rc;
1393 }
1394
1395 static inline int osd_extents_enabled(struct super_block *sb,
1396                                       struct inode *inode)
1397 {
1398         if (inode != NULL) {
1399                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1400                         return 1;
1401         } else if (test_opt(sb, EXTENTS)) {
1402                 return 1;
1403         }
1404         return 0;
1405 }
1406
1407 static inline int osd_calc_bkmap_credits(struct super_block *sb,
1408                                          struct inode *inode,
1409                                          const loff_t size,
1410                                          const loff_t pos,
1411                                          const int blocks)
1412 {
1413         int credits, bits, bs, i;
1414
1415         bits = sb->s_blocksize_bits;
1416         bs = 1 << bits;
1417
1418         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1419          * we do not expect blockmaps on the large files,
1420          * so let's shrink it to 2 levels (4GB files) */
1421
1422         /* this is default reservation: 2 levels */
1423         credits = (blocks + 2) * 3;
1424
1425         /* actual offset is unknown, hard to optimize */
1426         if (pos == -1)
1427                 return credits;
1428
1429         /* now check for few specific cases to optimize */
1430         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1431                 /* no indirects */
1432                 credits = blocks;
1433                 /* allocate if not allocated */
1434                 if (inode == NULL) {
1435                         credits += blocks * 2;
1436                         return credits;
1437                 }
1438                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1439                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1440                         if (LDISKFS_I(inode)->i_data[i] == 0)
1441                                 credits += 2;
1442                 }
1443         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1444                 /* single indirect */
1445                 credits = blocks * 3;
1446                 /* probably indirect block has been allocated already */
1447                 if (!inode || LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK])
1448                         credits += 3;
1449         }
1450
1451         return credits;
1452 }
1453
1454 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1455                                  const struct lu_buf *buf, loff_t _pos,
1456                                  struct thandle *handle)
1457 {
1458         struct osd_object  *obj  = osd_dt_obj(dt);
1459         struct inode       *inode = obj->oo_inode;
1460         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1461         struct osd_thandle *oh;
1462         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1463         int                 bits, bs;
1464         int                 depth, size;
1465         loff_t              pos;
1466         ENTRY;
1467
1468         LASSERT(buf != NULL);
1469         LASSERT(handle != NULL);
1470
1471         oh = container_of0(handle, struct osd_thandle, ot_super);
1472         LASSERT(oh->ot_handle == NULL);
1473
1474         size = buf->lb_len;
1475         bits = sb->s_blocksize_bits;
1476         bs = 1 << bits;
1477
1478         if (_pos == -1) {
1479                 /* if this is an append, then we
1480                  * should expect cross-block record */
1481                 pos = 0;
1482         } else {
1483                 pos = _pos;
1484         }
1485
1486         /* blocks to modify */
1487         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1488         LASSERT(blocks > 0);
1489
1490         if (inode != NULL && _pos != -1) {
1491                 /* object size in blocks */
1492                 est = (i_size_read(inode) + bs - 1) >> bits;
1493                 allocated = inode->i_blocks >> (bits - 9);
1494                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1495                         /* looks like an overwrite, no need to modify tree */
1496                         credits = blocks;
1497                         /* no need to modify i_size */
1498                         goto out;
1499                 }
1500         }
1501
1502         if (osd_extents_enabled(sb, inode)) {
1503                 /*
1504                  * many concurrent threads may grow tree by the time
1505                  * our transaction starts. so, consider 2 is a min depth
1506                  * for every level we may need to allocate a new block
1507                  * and take some entries from the old one. so, 3 blocks
1508                  * to allocate (bitmap, gd, itself) + old block - 4 per
1509                  * level.
1510                  */
1511                 depth = inode != NULL ? ext_depth(inode) : 0;
1512                 depth = max(depth, 1) + 1;
1513                 credits = depth;
1514                 /* if not append, then split may need to modify
1515                  * existing blocks moving entries into the new ones */
1516                 if (_pos == -1)
1517                         credits += depth;
1518                 /* blocks to store data: bitmap,gd,itself */
1519                 credits += blocks * 3;
1520         } else {
1521                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1522         }
1523         /* if inode is created as part of the transaction,
1524          * then it's counted already by the creation method */
1525         if (inode != NULL)
1526                 credits++;
1527
1528 out:
1529
1530         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1531
1532         /* dt_declare_write() is usually called for system objects, such
1533          * as llog or last_rcvd files. We needn't enforce quota on those
1534          * objects, so always set the lqi_space as 0. */
1535         if (inode != NULL)
1536                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1537                                            i_gid_read(inode), 0, oh, true,
1538                                            true, NULL, false);
1539         RETURN(rc);
1540 }
1541
1542 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1543 {
1544         /* LU-2634: clear the extent format for fast symlink */
1545         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1546
1547         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1548         LDISKFS_I(inode)->i_disksize = buflen;
1549         i_size_write(inode, buflen);
1550         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1551
1552         return 0;
1553 }
1554
1555 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1556                              int write_NUL, loff_t *offs, handle_t *handle)
1557 {
1558         struct buffer_head *bh        = NULL;
1559         loff_t              offset    = *offs;
1560         loff_t              new_size  = i_size_read(inode);
1561         unsigned long       block;
1562         int                 blocksize = 1 << inode->i_blkbits;
1563         int                 err = 0;
1564         int                 size;
1565         int                 boffs;
1566         int                 dirty_inode = 0;
1567
1568         if (write_NUL) {
1569                 /*
1570                  * long symlink write does not count the NUL terminator in
1571                  * bufsize, we write it, and the inode's file size does not
1572                  * count the NUL terminator as well.
1573                  */
1574                 ((char *)buf)[bufsize] = '\0';
1575                 ++bufsize;
1576         }
1577         while (bufsize > 0) {
1578                 if (bh != NULL)
1579                         brelse(bh);
1580
1581                 block = offset >> inode->i_blkbits;
1582                 boffs = offset & (blocksize - 1);
1583                 size = min(blocksize - boffs, bufsize);
1584                 bh = ldiskfs_bread(handle, inode, block, 1, &err);
1585                 if (!bh) {
1586                         CERROR("%s: error reading offset %llu (block %lu): "
1587                                "rc = %d\n",
1588                                inode->i_sb->s_id, offset, block, err);
1589                         break;
1590                 }
1591
1592                 err = ldiskfs_journal_get_write_access(handle, bh);
1593                 if (err) {
1594                         CERROR("journal_get_write_access() returned error %d\n",
1595                                err);
1596                         break;
1597                 }
1598                 LASSERTF(boffs + size <= bh->b_size,
1599                          "boffs %d size %d bh->b_size %lu\n",
1600                          boffs, size, (unsigned long)bh->b_size);
1601                 memcpy(bh->b_data + boffs, buf, size);
1602                 err = ldiskfs_journal_dirty_metadata(handle, bh);
1603                 if (err)
1604                         break;
1605
1606                 if (offset + size > new_size)
1607                         new_size = offset + size;
1608                 offset += size;
1609                 bufsize -= size;
1610                 buf += size;
1611         }
1612         if (bh)
1613                 brelse(bh);
1614
1615         if (write_NUL)
1616                 --new_size;
1617         /* correct in-core and on-disk sizes */
1618         if (new_size > i_size_read(inode)) {
1619                 spin_lock(&inode->i_lock);
1620                 if (new_size > i_size_read(inode))
1621                         i_size_write(inode, new_size);
1622                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1623                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1624                         dirty_inode = 1;
1625                 }
1626                 spin_unlock(&inode->i_lock);
1627                 if (dirty_inode)
1628                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1629         }
1630
1631         if (err == 0)
1632                 *offs = offset;
1633         return err;
1634 }
1635
1636 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1637                          const struct lu_buf *buf, loff_t *pos,
1638                          struct thandle *handle, struct lustre_capa *capa,
1639                          int ignore_quota)
1640 {
1641         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1642         struct osd_thandle      *oh;
1643         ssize_t                 result;
1644         int                     is_link;
1645
1646         LASSERT(dt_object_exists(dt));
1647
1648         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_WRITE))
1649                 return -EACCES;
1650
1651         LASSERT(handle != NULL);
1652         LASSERT(inode != NULL);
1653         ll_vfs_dq_init(inode);
1654
1655         /* XXX: don't check: one declared chunk can be used many times */
1656         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1657
1658         oh = container_of(handle, struct osd_thandle, ot_super);
1659         LASSERT(oh->ot_handle->h_transaction != NULL);
1660         /* Write small symlink to inode body as we need to maintain correct
1661          * on-disk symlinks for ldiskfs.
1662          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1663          * does not count it in.
1664          */
1665         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1666         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1667                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1668         else
1669                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1670                                                   buf->lb_len, is_link, pos,
1671                                                   oh->ot_handle);
1672         if (result == 0)
1673                 result = buf->lb_len;
1674         return result;
1675 }
1676
1677 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1678                              __u64 start, __u64 end, struct thandle *th)
1679 {
1680         struct osd_thandle *oh;
1681         struct inode       *inode;
1682         int                 rc;
1683         ENTRY;
1684
1685         LASSERT(th);
1686         oh = container_of(th, struct osd_thandle, ot_super);
1687
1688         /*
1689          * we don't need to reserve credits for whole truncate
1690          * it's not possible as truncate may need to free too many
1691          * blocks and that won't fit a single transaction. instead
1692          * we reserve credits to change i_size and put inode onto
1693          * orphan list. if needed truncate will extend or restart
1694          * transaction
1695          */
1696         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1697                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1698
1699         inode = osd_dt_obj(dt)->oo_inode;
1700         LASSERT(inode);
1701
1702         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1703                                    0, oh, true, true, NULL, false);
1704         RETURN(rc);
1705 }
1706
1707 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1708                      __u64 start, __u64 end, struct thandle *th,
1709                      struct lustre_capa *capa)
1710 {
1711         struct osd_thandle *oh;
1712         struct osd_object  *obj = osd_dt_obj(dt);
1713         struct inode       *inode = obj->oo_inode;
1714         handle_t           *h;
1715         tid_t               tid;
1716         int                rc = 0, rc2 = 0;
1717         ENTRY;
1718
1719         LASSERT(end == OBD_OBJECT_EOF);
1720         LASSERT(dt_object_exists(dt));
1721         LASSERT(osd_invariant(obj));
1722         LASSERT(inode != NULL);
1723         ll_vfs_dq_init(inode);
1724
1725         LASSERT(th);
1726         oh = container_of(th, struct osd_thandle, ot_super);
1727         LASSERT(oh->ot_handle->h_transaction != NULL);
1728
1729         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1730
1731         tid = oh->ot_handle->h_transaction->t_tid;
1732
1733         i_size_write(inode, start);
1734         ll_truncate_pagecache(inode, start);
1735 #ifdef HAVE_INODEOPS_TRUNCATE
1736         if (inode->i_op->truncate) {
1737                 inode->i_op->truncate(inode);
1738         } else
1739 #endif
1740                 ldiskfs_truncate(inode);
1741
1742         /*
1743          * For a partial-page truncate, flush the page to disk immediately to
1744          * avoid data corruption during direct disk write.  b=17397
1745          */
1746         if ((start & ~CFS_PAGE_MASK) != 0)
1747                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1748
1749         h = journal_current_handle();
1750         LASSERT(h != NULL);
1751         LASSERT(h == oh->ot_handle);
1752
1753         if (tid != h->h_transaction->t_tid) {
1754                 int credits = oh->ot_credits;
1755                 /*
1756                  * transaction has changed during truncate
1757                  * we need to restart the handle with our credits
1758                  */
1759                 if (h->h_buffer_credits < credits) {
1760                         if (ldiskfs_journal_extend(h, credits))
1761                                 rc2 = ldiskfs_journal_restart(h, credits);
1762                 }
1763         }
1764
1765         RETURN(rc == 0 ? rc2 : rc);
1766 }
1767
1768 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1769                           struct ll_user_fiemap *fm)
1770 {
1771         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1772         struct osd_thread_info *info   = osd_oti_get(env);
1773         struct dentry          *dentry = &info->oti_obj_dentry;
1774         struct file            *file   = &info->oti_file;
1775         mm_segment_t            saved_fs;
1776         int rc;
1777
1778         LASSERT(inode);
1779         dentry->d_inode = inode;
1780         dentry->d_sb = inode->i_sb;
1781         file->f_dentry = dentry;
1782         file->f_mapping = inode->i_mapping;
1783         file->f_op = inode->i_fop;
1784         set_file_inode(file, inode);
1785
1786         saved_fs = get_fs();
1787         set_fs(get_ds());
1788         /* ldiskfs_ioctl does not have a inode argument */
1789         if (inode->i_fop->unlocked_ioctl)
1790                 rc = inode->i_fop->unlocked_ioctl(file, FSFILT_IOC_FIEMAP,
1791                                                   (long)fm);
1792         else
1793                 rc = -ENOTTY;
1794         set_fs(saved_fs);
1795         return rc;
1796 }
1797
1798 /*
1799  * in some cases we may need declare methods for objects being created
1800  * e.g., when we create symlink
1801  */
1802 const struct dt_body_operations osd_body_ops_new = {
1803         .dbo_declare_write = osd_declare_write,
1804 };
1805
1806 const struct dt_body_operations osd_body_ops = {
1807         .dbo_read                 = osd_read,
1808         .dbo_declare_write        = osd_declare_write,
1809         .dbo_write                = osd_write,
1810         .dbo_bufs_get             = osd_bufs_get,
1811         .dbo_bufs_put             = osd_bufs_put,
1812         .dbo_write_prep           = osd_write_prep,
1813         .dbo_declare_write_commit = osd_declare_write_commit,
1814         .dbo_write_commit         = osd_write_commit,
1815         .dbo_read_prep            = osd_read_prep,
1816         .dbo_declare_punch         = osd_declare_punch,
1817         .dbo_punch                 = osd_punch,
1818         .dbo_fiemap_get           = osd_fiemap_get,
1819 };
1820