lustre/osd-ldiskfs/osd_io.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/osd/osd_io.c
  33  *
  34  * body operations
  35  *
  36  * Author: Nikita Danilov <nikita@clusterfs.com>
  37  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  38  *
  39  */
  40
  41 /* LUSTRE_VERSION_CODE */
  42 #include <lustre_ver.h>
  43 /* prerequisite for linux/xattr.h */
  44 #include <linux/types.h>
  45 /* prerequisite for linux/xattr.h */
  46 #include <linux/fs.h>
  47
  48 /*
  49  * struct OBD_{ALLOC,FREE}*()
  50  * OBD_FAIL_CHECK
  51  */
  52 #include <obd_support.h>
  53
  54 #include "osd_internal.h"
  55
  56 /* ext_depth() */
  57 #include <ldiskfs/ldiskfs_extents.h>
  58
  59 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  60                             int rw, int line, int pages)
  61 {
  62         int blocks, i;
  63
  64         LASSERTF(iobuf->dr_elapsed_valid == 0,
  65                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  66                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  67                  iobuf->dr_init_at);
  68         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  69
  70         init_waitqueue_head(&iobuf->dr_wait);
  71         atomic_set(&iobuf->dr_numreqs, 0);
  72         iobuf->dr_npages = 0;
  73         iobuf->dr_error = 0;
  74         iobuf->dr_dev = d;
  75         iobuf->dr_frags = 0;
  76         iobuf->dr_elapsed = 0;
  77         /* must be counted before, so assert */
  78         iobuf->dr_rw = rw;
  79         iobuf->dr_init_at = line;
  80
  81         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  82         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  83                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  84                         pages * sizeof(iobuf->dr_pages[0]));
  85                 return 0;
  86         }
  87
  88         /* start with 1MB for 4K blocks */
  89         i = 256;
  90         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
  91                 i <<= 1;
  92
  93         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
  94                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
  95         pages = i;
  96         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  97         iobuf->dr_max_pages = 0;
  98         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
  99                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 100
 101         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 102         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 103         if (unlikely(iobuf->dr_blocks == NULL))
 104                 return -ENOMEM;
 105
 106         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 107         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 108         if (unlikely(iobuf->dr_pages == NULL))
 109                 return -ENOMEM;
 110
 111         iobuf->dr_max_pages = pages;
 112
 113         return 0;
 114 }
 115 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 116         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 117
 118 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
 119 {
 120         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 121         iobuf->dr_pages[iobuf->dr_npages++] = page;
 122 }
 123
 124 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 125 {
 126         int rw = iobuf->dr_rw;
 127
 128         if (iobuf->dr_elapsed_valid) {
 129                 iobuf->dr_elapsed_valid = 0;
 130                 LASSERT(iobuf->dr_dev == d);
 131                 LASSERT(iobuf->dr_frags > 0);
 132                 lprocfs_oh_tally(&d->od_brw_stats.
 133                                  hist[BRW_R_DIO_FRAGS+rw],
 134                                  iobuf->dr_frags);
 135                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 136                                       iobuf->dr_elapsed);
 137         }
 138 }
 139
 140 #ifndef REQ_WRITE /* pre-2.6.35 */
 141 #define __REQ_WRITE BIO_RW
 142 #endif
 143
 144 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 145 static void dio_complete_routine(struct bio *bio)
 146 {
 147         int error = bio->bi_error;
 148 #else
 149 static void dio_complete_routine(struct bio *bio, int error)
 150 {
 151 #endif
 152         struct osd_iobuf *iobuf = bio->bi_private;
 153         int iter;
 154         struct bio_vec *bvl;
 155
 156         /* CAVEAT EMPTOR: possibly in IRQ context
 157          * DO NOT record procfs stats here!!! */
 158
 159         if (unlikely(iobuf == NULL)) {
 160                 CERROR("***** bio->bi_private is NULL!  This should never "
 161                        "happen.  Normally, I would crash here, but instead I "
 162                        "will dump the bio contents to the console.  Please "
 163                        "report this to <https://jira.hpdd.intel.com/> , along "
 164                        "with any interesting messages leading up to this point "
 165                        "(like SCSI errors, perhaps).  Because bi_private is "
 166                        "NULL, I can't wake up the thread that initiated this "
 167                        "IO - you will probably have to reboot this node.\n");
 168                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 169                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 170                        "bi_private: %p\n", bio->bi_next,
 171                         (unsigned long)bio->bi_flags,
 172                         bio->bi_rw, bio->bi_vcnt, bio_idx(bio),
 173                         bio_sectors(bio) << 9, bio->bi_end_io,
 174 #ifdef HAVE_BI_CNT
 175                         atomic_read(&bio->bi_cnt),
 176 #else
 177                         atomic_read(&bio->__bi_cnt),
 178 #endif
 179                         bio->bi_private);
 180                 return;
 181         }
 182
 183         /* the check is outside of the cycle for performance reason -bzzz */
 184         if (!test_bit(__REQ_WRITE, &bio->bi_rw)) {
 185                 bio_for_each_segment_all(bvl, bio, iter) {
 186                         if (likely(error == 0))
 187                                 SetPageUptodate(bvl_to_page(bvl));
 188                         LASSERT(PageLocked(bvl_to_page(bvl)));
 189                 }
 190                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 191         } else {
 192                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 193         }
 194
 195         /* any real error is good enough -bzzz */
 196         if (error != 0 && iobuf->dr_error == 0)
 197                 iobuf->dr_error = error;
 198
 199         /*
 200          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 201          * it's possible that service thread will see dr_numreqs
 202          * is zero, but dr_elapsed is not set yet, leading to lost
 203          * data in this processing and an assertion in a subsequent
 204          * call to OSD.
 205          */
 206         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 207                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
 208                 iobuf->dr_elapsed_valid = 1;
 209         }
 210         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 211                 wake_up(&iobuf->dr_wait);
 212
 213         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 214          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 215          * mempool when serious on-disk fragmentation was encountered,
 216          * deadlocking the OST.  The bios are now released as soon as complete
 217          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 218         bio_put(bio);
 219 }
 220
 221 static void record_start_io(struct osd_iobuf *iobuf, int size)
 222 {
 223         struct osd_device    *osd = iobuf->dr_dev;
 224         struct obd_histogram *h = osd->od_brw_stats.hist;
 225
 226         iobuf->dr_frags++;
 227         atomic_inc(&iobuf->dr_numreqs);
 228
 229         if (iobuf->dr_rw == 0) {
 230                 atomic_inc(&osd->od_r_in_flight);
 231                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 232                                  atomic_read(&osd->od_r_in_flight));
 233                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 234         } else if (iobuf->dr_rw == 1) {
 235                 atomic_inc(&osd->od_w_in_flight);
 236                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 237                                  atomic_read(&osd->od_w_in_flight));
 238                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 239         } else {
 240                 LBUG();
 241         }
 242 }
 243
 244 static void osd_submit_bio(int rw, struct bio *bio)
 245 {
 246         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 247         if (rw == 0)
 248                 submit_bio(READ, bio);
 249         else
 250                 submit_bio(WRITE, bio);
 251 }
 252
 253 static int can_be_merged(struct bio *bio, sector_t sector)
 254 {
 255         if (bio == NULL)
 256                 return 0;
 257
 258         return bio_end_sector(bio) == sector ? 1 : 0;
 259 }
 260
 261 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 262                       struct osd_iobuf *iobuf)
 263 {
 264         int            blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 265         struct page  **pages = iobuf->dr_pages;
 266         int            npages = iobuf->dr_npages;
 267         sector_t      *blocks = iobuf->dr_blocks;
 268         int            total_blocks = npages * blocks_per_page;
 269         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 270         unsigned int   blocksize = inode->i_sb->s_blocksize;
 271         struct bio    *bio = NULL;
 272         struct page   *page;
 273         unsigned int   page_offset;
 274         sector_t       sector;
 275         int            nblocks;
 276         int            block_idx;
 277         int            page_idx;
 278         int            i;
 279         int            rc = 0;
 280         DECLARE_PLUG(plug);
 281         ENTRY;
 282
 283         LASSERT(iobuf->dr_npages == npages);
 284
 285         osd_brw_stats_update(osd, iobuf);
 286         iobuf->dr_start_time = cfs_time_current();
 287
 288         blk_start_plug(&plug);
 289         for (page_idx = 0, block_idx = 0;
 290              page_idx < npages;
 291              page_idx++, block_idx += blocks_per_page) {
 292
 293                 page = pages[page_idx];
 294                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 295
 296                 for (i = 0, page_offset = 0;
 297                      i < blocks_per_page;
 298                      i += nblocks, page_offset += blocksize * nblocks) {
 299
 300                         nblocks = 1;
 301
 302                         if (blocks[block_idx + i] == 0) {  /* hole */
 303                                 LASSERTF(iobuf->dr_rw == 0,
 304                                          "page_idx %u, block_idx %u, i %u\n",
 305                                          page_idx, block_idx, i);
 306                                 memset(kmap(page) + page_offset, 0, blocksize);
 307                                 kunmap(page);
 308                                 continue;
 309                         }
 310
 311                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 312
 313                         /* Additional contiguous file blocks? */
 314                         while (i + nblocks < blocks_per_page &&
 315                                (sector + (nblocks << sector_bits)) ==
 316                                ((sector_t)blocks[block_idx + i + nblocks] <<
 317                                 sector_bits))
 318                                 nblocks++;
 319
 320                         if (bio != NULL &&
 321                             can_be_merged(bio, sector) &&
 322                             bio_add_page(bio, page,
 323                                          blocksize * nblocks, page_offset) != 0)
 324                                 continue;       /* added this frag OK */
 325
 326                         if (bio != NULL) {
 327                                 struct request_queue *q =
 328                                         bdev_get_queue(bio->bi_bdev);
 329                                 unsigned int bi_size = bio_sectors(bio) << 9;
 330
 331                                 /* Dang! I have to fragment this I/O */
 332                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 333                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 334                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 335                                        bio_sectors(bio),
 336                                        queue_max_sectors(q),
 337                                        bio_phys_segments(q, bio),
 338                                        queue_max_phys_segments(q),
 339                                        0, queue_max_hw_segments(q));
 340                                 record_start_io(iobuf, bi_size);
 341                                 osd_submit_bio(iobuf->dr_rw, bio);
 342                         }
 343
 344                         /* allocate new bio */
 345                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 346                                                       (npages - page_idx) *
 347                                                       blocks_per_page));
 348                         if (bio == NULL) {
 349                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 350                                        (npages - page_idx), blocks_per_page,
 351                                        (npages - page_idx) * blocks_per_page);
 352                                 rc = -ENOMEM;
 353                                 goto out;
 354                         }
 355
 356                         bio->bi_bdev = inode->i_sb->s_bdev;
 357                         bio_set_sector(bio, sector);
 358                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 359                         bio->bi_end_io = dio_complete_routine;
 360                         bio->bi_private = iobuf;
 361
 362                         rc = bio_add_page(bio, page,
 363                                           blocksize * nblocks, page_offset);
 364                         LASSERT(rc != 0);
 365                 }
 366         }
 367
 368         if (bio != NULL) {
 369                 record_start_io(iobuf, bio_sectors(bio) << 9);
 370                 osd_submit_bio(iobuf->dr_rw, bio);
 371                 rc = 0;
 372         }
 373
 374 out:
 375         blk_finish_plug(&plug);
 376
 377         /* in order to achieve better IO throughput, we don't wait for writes
 378          * completion here. instead we proceed with transaction commit in
 379          * parallel and wait for IO completion once transaction is stopped
 380          * see osd_trans_stop() for more details -bzzz */
 381         if (iobuf->dr_rw == 0) {
 382                 wait_event(iobuf->dr_wait,
 383                            atomic_read(&iobuf->dr_numreqs) == 0);
 384                 osd_fini_iobuf(osd, iobuf);
 385         }
 386
 387         if (rc == 0)
 388                 rc = iobuf->dr_error;
 389         RETURN(rc);
 390 }
 391
 392 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 393                                    struct niobuf_local *lnb)
 394 {
 395         ENTRY;
 396
 397         *nrpages = 0;
 398
 399         while (len > 0) {
 400                 int poff = offset & (PAGE_SIZE - 1);
 401                 int plen = PAGE_SIZE - poff;
 402
 403                 if (plen > len)
 404                         plen = len;
 405                 lnb->lnb_file_offset = offset;
 406                 lnb->lnb_page_offset = poff;
 407                 lnb->lnb_len = plen;
 408                 /* lnb->lnb_flags = rnb->rnb_flags; */
 409                 lnb->lnb_flags = 0;
 410                 lnb->lnb_page = NULL;
 411                 lnb->lnb_rc = 0;
 412
 413                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 414                          (long long) len);
 415                 offset += plen;
 416                 len -= plen;
 417                 lnb++;
 418                 (*nrpages)++;
 419         }
 420
 421         RETURN(0);
 422 }
 423
 424 static struct page *osd_get_page(struct dt_object *dt, loff_t offset,
 425                                  gfp_t gfp_mask)
 426 {
 427         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 428         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 429         struct page *page;
 430
 431         LASSERT(inode);
 432
 433         page = find_or_create_page(inode->i_mapping, offset >> PAGE_SHIFT,
 434                                    gfp_mask);
 435
 436         if (unlikely(page == NULL))
 437                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 438
 439         return page;
 440 }
 441
 442 /*
 443  * there are following "locks":
 444  * journal_start
 445  * i_mutex
 446  * page lock
 447  *
 448  * osd write path:
 449  *  - lock page(s)
 450  *  - journal_start
 451  *  - truncate_sem
 452  *
 453  * ext4 vmtruncate:
 454  *  - lock pages, unlock
 455  *  - journal_start
 456  *  - lock partial page
 457  *  - i_data_sem
 458  *
 459  */
 460
 461 /**
 462  * Unlock and release pages loaded by osd_bufs_get()
 463  *
 464  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 465  *
 466  * \param env           thread execution environment
 467  * \param dt            dt object undergoing IO (OSD object + methods)
 468  * \param lnb           array of pages undergoing IO
 469  * \param npages        number of pages in \a lnb
 470  *
 471  * \retval 0            always
 472  */
 473 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 474                         struct niobuf_local *lnb, int npages)
 475 {
 476         int i;
 477
 478         for (i = 0; i < npages; i++) {
 479                 if (lnb[i].lnb_page == NULL)
 480                         continue;
 481                 LASSERT(PageLocked(lnb[i].lnb_page));
 482                 unlock_page(lnb[i].lnb_page);
 483                 put_page(lnb[i].lnb_page);
 484                 dt_object_put(env, dt);
 485                 lnb[i].lnb_page = NULL;
 486         }
 487
 488         RETURN(0);
 489 }
 490
 491 /**
 492  * Load and lock pages undergoing IO
 493  *
 494  * Pages as described in the \a lnb array are fetched (from disk or cache)
 495  * and locked for IO by the caller.
 496  *
 497  * DLM locking protects us from write and truncate competing for same region,
 498  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 499  * It's possible the writeout on a such a page is in progress when we access
 500  * it. It's also possible that during this writeout we put new (partial) data
 501  * into the page, but won't be able to proceed in filter_commitrw_write().
 502  * Therefore, just wait for writeout completion as it should be rare enough.
 503  *
 504  * \param env           thread execution environment
 505  * \param dt            dt object undergoing IO (OSD object + methods)
 506  * \param pos           byte offset of IO start
 507  * \param len           number of bytes of IO
 508  * \param lnb           array of extents undergoing IO
 509  * \param rw            read or write operation, and other flags
 510  * \param capa          capabilities
 511  *
 512  * \retval pages        (zero or more) loaded successfully
 513  * \retval -ENOMEM      on memory/page allocation error
 514  */
 515 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 516                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 517                         enum dt_bufs_type rw)
 518 {
 519         struct osd_object *obj = osd_dt_obj(dt);
 520         int npages, i, rc = 0;
 521         gfp_t gfp_mask;
 522
 523         LASSERT(obj->oo_inode);
 524
 525         osd_map_remote_to_local(pos, len, &npages, lnb);
 526
 527         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 528         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 529                                              GFP_HIGHUSER;
 530         for (i = 0; i < npages; i++, lnb++) {
 531                 lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset,
 532                                              gfp_mask);
 533                 if (lnb->lnb_page == NULL)
 534                         GOTO(cleanup, rc = -ENOMEM);
 535
 536                 wait_on_page_writeback(lnb->lnb_page);
 537                 BUG_ON(PageWriteback(lnb->lnb_page));
 538
 539                 lu_object_get(&dt->do_lu);
 540         }
 541
 542         RETURN(i);
 543
 544 cleanup:
 545         if (i > 0)
 546                 osd_bufs_put(env, dt, lnb - i, i);
 547         return rc;
 548 }
 549
 550 #ifndef HAVE_LDISKFS_MAP_BLOCKS
 551
 552 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 553 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 554 #endif
 555
 556 struct bpointers {
 557         sector_t *blocks;
 558         unsigned long start;
 559         int num;
 560         int init_num;
 561         int create;
 562 };
 563
 564 static long ldiskfs_ext_find_goal(struct inode *inode,
 565                                   struct ldiskfs_ext_path *path,
 566                                   unsigned long block, int *aflags)
 567 {
 568         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 569         unsigned long bg_start;
 570         unsigned long colour;
 571         int depth;
 572
 573         if (path) {
 574                 struct ldiskfs_extent *ex;
 575                 depth = path->p_depth;
 576
 577                 /* try to predict block placement */
 578                 if ((ex = path[depth].p_ext))
 579                         return ldiskfs_ext_pblock(ex) +
 580                                 (block - le32_to_cpu(ex->ee_block));
 581
 582                 /* it looks index is empty
 583                  * try to find starting from index itself */
 584                 if (path[depth].p_bh)
 585                         return path[depth].p_bh->b_blocknr;
 586         }
 587
 588         /* OK. use inode's group */
 589         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 590                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 591         colour = (current->pid % 16) *
 592                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 593         return bg_start + colour + block;
 594 }
 595
 596 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 597                                 struct ldiskfs_ext_path *path,
 598                                 unsigned long block, unsigned long *count,
 599                                 int *err)
 600 {
 601         struct ldiskfs_allocation_request ar;
 602         unsigned long pblock;
 603         int aflags;
 604
 605         /* find neighbour allocated blocks */
 606         ar.lleft = block;
 607         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 608         if (*err)
 609                 return 0;
 610         ar.lright = block;
 611         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 612         if (*err)
 613                 return 0;
 614
 615         /* allocate new block */
 616         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 617         ar.inode = inode;
 618         ar.logical = block;
 619         ar.len = *count;
 620         ar.flags = LDISKFS_MB_HINT_DATA;
 621         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 622         *count = ar.len;
 623         return pblock;
 624 }
 625
 626 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 627                                      struct ldiskfs_ext_path *path,
 628                                      struct ldiskfs_ext_cache *cex,
 629 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 630                                      struct ldiskfs_extent *ex,
 631 #endif
 632                                      void *cbdata)
 633 {
 634         struct bpointers *bp = cbdata;
 635         struct ldiskfs_extent nex;
 636         unsigned long pblock = 0;
 637         unsigned long tgen;
 638         int err, i;
 639         unsigned long count;
 640         handle_t *handle;
 641
 642 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 643         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 644 #else
 645         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 646 #endif
 647                 err = EXT_CONTINUE;
 648                 goto map;
 649         }
 650
 651         if (bp->create == 0) {
 652                 i = 0;
 653                 if (cex->ec_block < bp->start)
 654                         i = bp->start - cex->ec_block;
 655                 if (i >= cex->ec_len)
 656                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 657                                         i, cex->ec_len);
 658                 for (; i < cex->ec_len && bp->num; i++) {
 659                         *(bp->blocks) = 0;
 660                         bp->blocks++;
 661                         bp->num--;
 662                         bp->start++;
 663                 }
 664
 665                 return EXT_CONTINUE;
 666         }
 667
 668         tgen = LDISKFS_I(inode)->i_ext_generation;
 669         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
 670
 671         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
 672                                    count + LDISKFS_ALLOC_NEEDED + 1);
 673         if (IS_ERR(handle)) {
 674                 return PTR_ERR(handle);
 675         }
 676
 677         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
 678                 /* the tree has changed. so path can be invalid at moment */
 679                 ldiskfs_journal_stop(handle);
 680                 return EXT_REPEAT;
 681         }
 682
 683         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
 684          * protected by i_data_sem as whole. so we patch it to store
 685          * generation to path and now verify the tree hasn't changed */
 686         down_write((&LDISKFS_I(inode)->i_data_sem));
 687
 688         /* validate extent, make sure the extent tree does not changed */
 689         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
 690                 /* cex is invalid, try again */
 691                 up_write(&LDISKFS_I(inode)->i_data_sem);
 692                 ldiskfs_journal_stop(handle);
 693                 return EXT_REPEAT;
 694         }
 695
 696         count = cex->ec_len;
 697         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
 698         if (!pblock)
 699                 goto out;
 700         BUG_ON(count > cex->ec_len);
 701
 702         /* insert new extent */
 703         nex.ee_block = cpu_to_le32(cex->ec_block);
 704         ldiskfs_ext_store_pblock(&nex, pblock);
 705         nex.ee_len = cpu_to_le16(count);
 706         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
 707         if (err) {
 708                 /* free data blocks we just allocated */
 709                 /* not a good idea to call discard here directly,
 710                  * but otherwise we'd need to call it every free() */
 711                 ldiskfs_discard_preallocations(inode);
 712 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
 713                 ldiskfs_free_blocks(handle, inode, NULL,
 714                                     ldiskfs_ext_pblock(&nex),
 715                                     le16_to_cpu(nex.ee_len), 0);
 716 #else
 717                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
 718                                     le16_to_cpu(nex.ee_len), 0);
 719 #endif
 720                 goto out;
 721         }
 722
 723         /*
 724          * Putting len of the actual extent we just inserted,
 725          * we are asking ldiskfs_ext_walk_space() to continue
 726          * scaning after that block
 727          */
 728         cex->ec_len = le16_to_cpu(nex.ee_len);
 729         cex->ec_start = ldiskfs_ext_pblock(&nex);
 730         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
 731         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
 732
 733 out:
 734         up_write((&LDISKFS_I(inode)->i_data_sem));
 735         ldiskfs_journal_stop(handle);
 736 map:
 737         if (err >= 0) {
 738                 /* map blocks */
 739                 if (bp->num == 0) {
 740                         CERROR("hmm. why do we find this extent?\n");
 741                         CERROR("initial space: %lu:%u\n",
 742                                 bp->start, bp->init_num);
 743 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 744                         CERROR("current extent: %u/%u/%llu %d\n",
 745                                 cex->ec_block, cex->ec_len,
 746                                 (unsigned long long)cex->ec_start,
 747                                 cex->ec_type);
 748 #else
 749                         CERROR("current extent: %u/%u/%llu\n",
 750                                 cex->ec_block, cex->ec_len,
 751                                 (unsigned long long)cex->ec_start);
 752 #endif
 753                 }
 754                 i = 0;
 755                 if (cex->ec_block < bp->start)
 756                         i = bp->start - cex->ec_block;
 757                 if (i >= cex->ec_len)
 758                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 759                                         i, cex->ec_len);
 760                 for (; i < cex->ec_len && bp->num; i++) {
 761                         *(bp->blocks) = cex->ec_start + i;
 762                         if (pblock != 0) {
 763                                 /* unmap any possible underlying metadata from
 764                                  * the block device mapping.  bug 6998. */
 765                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
 766                                                           *(bp->blocks));
 767                         }
 768                         bp->blocks++;
 769                         bp->num--;
 770                         bp->start++;
 771                 }
 772         }
 773         return err;
 774 }
 775
 776 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
 777                                    int clen, sector_t *blocks, int create)
 778 {
 779         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 780         struct bpointers bp;
 781         int err;
 782
 783         if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
 784                 return -EFBIG;
 785
 786         bp.blocks = blocks;
 787         bp.start = index * blocks_per_page;
 788         bp.init_num = bp.num = clen * blocks_per_page;
 789         bp.create = create;
 790
 791         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 792                bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
 793
 794         err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
 795                                      ldiskfs_ext_new_extent_cb, &bp);
 796         ldiskfs_ext_invalidate_cache(inode);
 797
 798         return err;
 799 }
 800
 801 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
 802                                           struct page **page, int pages,
 803                                           sector_t *blocks, int create)
 804 {
 805         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 806         pgoff_t bitmap_max_page_index;
 807         sector_t *b;
 808         int rc = 0, i;
 809
 810         bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
 811                                 PAGE_SHIFT;
 812         for (i = 0, b = blocks; i < pages; i++, page++) {
 813                 if ((*page)->index + 1 >= bitmap_max_page_index) {
 814                         rc = -EFBIG;
 815                         break;
 816                 }
 817                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
 818                 if (rc) {
 819                         CERROR("ino %lu, blk %llu create %d: rc %d\n",
 820                                inode->i_ino,
 821                                (unsigned long long)*b, create, rc);
 822                         break;
 823                 }
 824                 b += blocks_per_page;
 825         }
 826         return rc;
 827 }
 828
 829 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
 830                                            struct page **page,
 831                                            int pages, sector_t *blocks,
 832                                            int create)
 833 {
 834         int rc = 0, i = 0, clen = 0;
 835         struct page *fp = NULL;
 836
 837         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 838                 inode->i_ino, pages, (*page)->index);
 839
 840         /* pages are sorted already. so, we just have to find
 841          * contig. space and process them properly */
 842         while (i < pages) {
 843                 if (fp == NULL) {
 844                         /* start new extent */
 845                         fp = *page++;
 846                         clen = 1;
 847                         i++;
 848                         continue;
 849                 } else if (fp->index + clen == (*page)->index) {
 850                         /* continue the extent */
 851                         page++;
 852                         clen++;
 853                         i++;
 854                         continue;
 855                 }
 856
 857                 /* process found extent */
 858                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 859                                              blocks, create);
 860                 if (rc)
 861                         GOTO(cleanup, rc);
 862
 863                 /* look for next extent */
 864                 fp = NULL;
 865                 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
 866         }
 867
 868         if (fp)
 869                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 870                                              blocks, create);
 871
 872 cleanup:
 873         return rc;
 874 }
 875
 876 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 877                                        int pages, sector_t *blocks,
 878                                        int create)
 879 {
 880         int rc;
 881
 882         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 883                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
 884                                                      blocks, create);
 885                 return rc;
 886         }
 887         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
 888
 889         return rc;
 890 }
 891 #else
 892 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 893                                        int pages, sector_t *blocks,
 894                                        int create)
 895 {
 896         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 897         int rc = 0, i = 0;
 898         struct page *fp = NULL;
 899         int clen = 0;
 900         pgoff_t max_page_index;
 901         handle_t *handle = NULL;
 902
 903         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
 904
 905         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 906                 inode->i_ino, pages, (*page)->index);
 907
 908         if (create) {
 909                 create = LDISKFS_GET_BLOCKS_CREATE;
 910                 handle = ldiskfs_journal_current_handle();
 911                 LASSERT(handle != NULL);
 912                 rc = osd_attach_jinode(inode);
 913                 if (rc)
 914                         return rc;
 915         }
 916         /* pages are sorted already. so, we just have to find
 917          * contig. space and process them properly */
 918         while (i < pages) {
 919                 long blen, total = 0;
 920                 struct ldiskfs_map_blocks map = { 0 };
 921
 922                 if (fp == NULL) { /* start new extent */
 923                         fp = *page++;
 924                         clen = 1;
 925                         if (++i != pages)
 926                                 continue;
 927                 } else if (fp->index + clen == (*page)->index) {
 928                         /* continue the extent */
 929                         page++;
 930                         clen++;
 931                         if (++i != pages)
 932                                 continue;
 933                 }
 934                 if (fp->index + clen >= max_page_index)
 935                         GOTO(cleanup, rc = -EFBIG);
 936                 /* process found extent */
 937                 map.m_lblk = fp->index * blocks_per_page;
 938                 map.m_len = blen = clen * blocks_per_page;
 939 cont_map:
 940                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 941                 if (rc >= 0) {
 942                         int c = 0;
 943                         for (; total < blen && c < map.m_len; c++, total++) {
 944                                 if (rc == 0) {
 945                                         *(blocks + total) = 0;
 946                                         total++;
 947                                         break;
 948                                 } else {
 949                                         *(blocks + total) = map.m_pblk + c;
 950                                         /* unmap any possible underlying
 951                                          * metadata from the block device
 952                                          * mapping.  bug 6998. */
 953                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
 954                                             create)
 955                                                 unmap_underlying_metadata(
 956                                                         inode->i_sb->s_bdev,
 957                                                         map.m_pblk + c);
 958                                 }
 959                         }
 960                         rc = 0;
 961                 }
 962                 if (rc == 0 && total < blen) {
 963                         map.m_lblk = fp->index * blocks_per_page + total;
 964                         map.m_len = blen - total;
 965                         goto cont_map;
 966                 }
 967                 if (rc != 0)
 968                         GOTO(cleanup, rc);
 969
 970                 /* look for next extent */
 971                 fp = NULL;
 972                 blocks += blocks_per_page * clen;
 973         }
 974 cleanup:
 975         return rc;
 976 }
 977 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
 978
 979 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 980                           struct niobuf_local *lnb, int npages)
 981 {
 982         struct osd_thread_info *oti   = osd_oti_get(env);
 983         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 984         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 985         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 986         ktime_t start;
 987         ktime_t end;
 988         s64 timediff;
 989         ssize_t                 isize;
 990         __s64                   maxidx;
 991         int                     rc = 0;
 992         int                     i;
 993         int                     cache = 0;
 994
 995         LASSERT(inode);
 996
 997         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 998         if (unlikely(rc != 0))
 999                 RETURN(rc);
1000
1001         isize = i_size_read(inode);
1002         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1003
1004         if (osd->od_writethrough_cache)
1005                 cache = 1;
1006         if (isize > osd->od_readcache_max_filesize)
1007                 cache = 0;
1008
1009         start = ktime_get();
1010         for (i = 0; i < npages; i++) {
1011
1012                 if (cache == 0)
1013                         generic_error_remove_page(inode->i_mapping,
1014                                                   lnb[i].lnb_page);
1015
1016                 /*
1017                  * till commit the content of the page is undefined
1018                  * we'll set it uptodate once bulk is done. otherwise
1019                  * subsequent reads can access non-stable data
1020                  */
1021                 ClearPageUptodate(lnb[i].lnb_page);
1022
1023                 if (lnb[i].lnb_len == PAGE_SIZE)
1024                         continue;
1025
1026                 if (maxidx >= lnb[i].lnb_page->index) {
1027                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1028                 } else {
1029                         long off;
1030                         char *p = kmap(lnb[i].lnb_page);
1031
1032                         off = lnb[i].lnb_page_offset;
1033                         if (off)
1034                                 memset(p, 0, off);
1035                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1036                               ~PAGE_MASK;
1037                         if (off)
1038                                 memset(p + off, 0, PAGE_SIZE - off);
1039                         kunmap(lnb[i].lnb_page);
1040                 }
1041         }
1042         end = ktime_get();
1043         timediff = ktime_us_delta(end, start);
1044         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1045
1046         if (iobuf->dr_npages) {
1047                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1048                                                  iobuf->dr_npages,
1049                                                  iobuf->dr_blocks, 0);
1050                 if (likely(rc == 0)) {
1051                         rc = osd_do_bio(osd, inode, iobuf);
1052                         /* do IO stats for preparation reads */
1053                         osd_fini_iobuf(osd, iobuf);
1054                 }
1055         }
1056         RETURN(rc);
1057 }
1058
1059 struct osd_fextent {
1060         sector_t        start;
1061         sector_t        end;
1062         unsigned int    mapped:1;
1063 };
1064
1065 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1066                          struct osd_fextent *cached_extent)
1067 {
1068         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1069         sector_t block = offset >> inode->i_blkbits;
1070         sector_t start;
1071         struct fiemap_extent_info fei = { 0 };
1072         struct fiemap_extent fe = { 0 };
1073         mm_segment_t saved_fs;
1074         int rc;
1075
1076         if (block >= cached_extent->start && block < cached_extent->end)
1077                 return cached_extent->mapped;
1078
1079         if (i_size_read(inode) == 0)
1080                 return 0;
1081
1082         /* Beyond EOF, must not be mapped */
1083         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1084                 return 0;
1085
1086         fei.fi_extents_max = 1;
1087         fei.fi_extents_start = &fe;
1088
1089         saved_fs = get_fs();
1090         set_fs(get_ds());
1091         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1092         set_fs(saved_fs);
1093         if (rc != 0)
1094                 return 0;
1095
1096         start = fe.fe_logical >> inode->i_blkbits;
1097
1098         if (start > block) {
1099                 cached_extent->start = block;
1100                 cached_extent->end = start;
1101                 cached_extent->mapped = 0;
1102         } else {
1103                 cached_extent->start = start;
1104                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1105                                       inode->i_blkbits;
1106                 cached_extent->mapped = 1;
1107         }
1108
1109         return cached_extent->mapped;
1110 }
1111
1112 static int osd_declare_write_commit(const struct lu_env *env,
1113                                     struct dt_object *dt,
1114                                     struct niobuf_local *lnb, int npages,
1115                                     struct thandle *handle)
1116 {
1117         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1118         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1119         struct osd_thandle      *oh;
1120         int                     extents = 1;
1121         int                     depth;
1122         int                     i;
1123         int                     newblocks;
1124         int                     rc = 0;
1125         int                     flags = 0;
1126         int                     credits = 0;
1127         long long               quota_space = 0;
1128         struct osd_fextent      extent = { 0 };
1129         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1130         ENTRY;
1131
1132         LASSERT(handle != NULL);
1133         oh = container_of0(handle, struct osd_thandle, ot_super);
1134         LASSERT(oh->ot_handle == NULL);
1135
1136         newblocks = npages;
1137
1138         /* calculate number of extents (probably better to pass nb) */
1139         for (i = 0; i < npages; i++) {
1140                 if (i && lnb[i].lnb_file_offset !=
1141                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1142                         extents++;
1143
1144                 if (!osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1145                         quota_space += PAGE_SIZE;
1146
1147                 /* ignore quota for the whole request if any page is from
1148                  * client cache or written by root.
1149                  *
1150                  * XXX once we drop the 1.8 client support, the checking
1151                  * for whether page is from cache can be simplified as:
1152                  * !(lnb[i].flags & OBD_BRW_SYNC)
1153                  *
1154                  * XXX we could handle this on per-lnb basis as done by
1155                  * grant. */
1156                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1157                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1158                     OBD_BRW_FROM_GRANT)
1159                         declare_flags |= OSD_QID_FORCE;
1160         }
1161
1162         /*
1163          * each extent can go into new leaf causing a split
1164          * 5 is max tree depth: inode + 4 index blocks
1165          * with blockmaps, depth is 3 at most
1166          */
1167         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1168                 /*
1169                  * many concurrent threads may grow tree by the time
1170                  * our transaction starts. so, consider 2 is a min depth
1171                  */
1172                 depth = ext_depth(inode);
1173                 depth = max(depth, 1) + 1;
1174                 newblocks += depth;
1175                 credits++; /* inode */
1176                 credits += depth * 2 * extents;
1177         } else {
1178                 depth = 3;
1179                 newblocks += depth;
1180                 credits++; /* inode */
1181                 credits += depth * extents;
1182         }
1183
1184         /* quota space for metadata blocks */
1185         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1186
1187         /* quota space should be reported in 1K blocks */
1188         quota_space = toqb(quota_space);
1189
1190         /* each new block can go in different group (bitmap + gd) */
1191
1192         /* we can't dirty more bitmap blocks than exist */
1193         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1194                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1195         else
1196                 credits += newblocks;
1197
1198         /* we can't dirty more gd blocks than exist */
1199         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1200                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1201         else
1202                 credits += newblocks;
1203
1204         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1205
1206         /* make sure the over quota flags were not set */
1207         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1208
1209         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1210                                    i_projid_read(inode), quota_space, oh,
1211                                    osd_dt_obj(dt), &flags, declare_flags);
1212
1213         /* we need only to store the overquota flags in the first lnb for
1214          * now, once we support multiple objects BRW, this code needs be
1215          * revised. */
1216         if (flags & QUOTA_FL_OVER_USRQUOTA)
1217                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1218         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1219                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1220         if (flags & QUOTA_FL_OVER_PRJQUOTA)
1221                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1222
1223         RETURN(rc);
1224 }
1225
1226 /* Check if a block is allocated or not */
1227 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1228                             struct niobuf_local *lnb, int npages,
1229                             struct thandle *thandle)
1230 {
1231         struct osd_thread_info *oti = osd_oti_get(env);
1232         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1233         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1234         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1235         loff_t isize;
1236         int rc = 0, i;
1237         struct osd_fextent extent = { 0 };
1238
1239         LASSERT(inode);
1240
1241         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1242         if (unlikely(rc != 0))
1243                 RETURN(rc);
1244
1245         isize = i_size_read(inode);
1246         ll_vfs_dq_init(inode);
1247
1248         for (i = 0; i < npages; i++) {
1249                 if (lnb[i].lnb_rc == -ENOSPC &&
1250                     osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent)) {
1251                         /* Allow the write to proceed if overwriting an
1252                          * existing block */
1253                         lnb[i].lnb_rc = 0;
1254                 }
1255
1256                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1257                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1258                                lnb[i].lnb_rc);
1259                         LASSERT(lnb[i].lnb_page);
1260                         generic_error_remove_page(inode->i_mapping,
1261                                                   lnb[i].lnb_page);
1262                         continue;
1263                 }
1264
1265                 LASSERT(PageLocked(lnb[i].lnb_page));
1266                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1267
1268                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1269                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1270
1271                 /*
1272                  * Since write and truncate are serialized by oo_sem, even
1273                  * partial-page truncate should not leave dirty pages in the
1274                  * page cache.
1275                  */
1276                 LASSERT(!PageDirty(lnb[i].lnb_page));
1277
1278                 SetPageUptodate(lnb[i].lnb_page);
1279
1280                 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1281         }
1282
1283         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1284
1285         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1286                 rc = -ENOSPC;
1287         } else if (iobuf->dr_npages > 0) {
1288                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1289                                                  iobuf->dr_npages,
1290                                                  iobuf->dr_blocks, 1);
1291         } else {
1292                 /* no pages to write, no transno is needed */
1293                 thandle->th_local = 1;
1294         }
1295
1296         if (likely(rc == 0)) {
1297                 spin_lock(&inode->i_lock);
1298                 if (isize > i_size_read(inode)) {
1299                         i_size_write(inode, isize);
1300                         LDISKFS_I(inode)->i_disksize = isize;
1301                         spin_unlock(&inode->i_lock);
1302                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1303                 } else {
1304                         spin_unlock(&inode->i_lock);
1305                 }
1306
1307                 rc = osd_do_bio(osd, inode, iobuf);
1308                 /* we don't do stats here as in read path because
1309                  * write is async: we'll do this in osd_put_bufs() */
1310         } else {
1311                 osd_fini_iobuf(osd, iobuf);
1312         }
1313
1314         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1315
1316         if (unlikely(rc != 0)) {
1317                 /* if write fails, we should drop pages from the cache */
1318                 for (i = 0; i < npages; i++) {
1319                         if (lnb[i].lnb_page == NULL)
1320                                 continue;
1321                         LASSERT(PageLocked(lnb[i].lnb_page));
1322                         generic_error_remove_page(inode->i_mapping,
1323                                                   lnb[i].lnb_page);
1324                 }
1325         }
1326
1327         RETURN(rc);
1328 }
1329
1330 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1331                          struct niobuf_local *lnb, int npages)
1332 {
1333         struct osd_thread_info *oti = osd_oti_get(env);
1334         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1335         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1336         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1337         int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1338         ktime_t start, end;
1339         s64 timediff;
1340         loff_t isize;
1341
1342         LASSERT(inode);
1343
1344         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1345         if (unlikely(rc != 0))
1346                 RETURN(rc);
1347
1348         isize = i_size_read(inode);
1349
1350         if (osd->od_read_cache)
1351                 cache = 1;
1352         if (isize > osd->od_readcache_max_filesize)
1353                 cache = 0;
1354
1355         start = ktime_get();
1356         for (i = 0; i < npages; i++) {
1357
1358                 if (isize <= lnb[i].lnb_file_offset)
1359                         /* If there's no more data, abort early.
1360                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1361                         break;
1362
1363                 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1364                         lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1365                 else
1366                         lnb[i].lnb_rc = lnb[i].lnb_len;
1367
1368                 /* Bypass disk read if fail_loc is set properly */
1369                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1370                         SetPageUptodate(lnb[i].lnb_page);
1371
1372                 if (PageUptodate(lnb[i].lnb_page)) {
1373                         cache_hits++;
1374                 } else {
1375                         cache_misses++;
1376                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1377                 }
1378
1379                 if (cache == 0)
1380                         generic_error_remove_page(inode->i_mapping,
1381                                                   lnb[i].lnb_page);
1382         }
1383         end = ktime_get();
1384         timediff = ktime_us_delta(end, start);
1385         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1386
1387         if (cache_hits != 0)
1388                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1389                                     cache_hits);
1390         if (cache_misses != 0)
1391                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1392                                     cache_misses);
1393         if (cache_hits + cache_misses != 0)
1394                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1395                                     cache_hits + cache_misses);
1396
1397         if (iobuf->dr_npages) {
1398                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1399                                                  iobuf->dr_npages,
1400                                                  iobuf->dr_blocks, 0);
1401                 rc = osd_do_bio(osd, inode, iobuf);
1402
1403                 /* IO stats will be done in osd_bufs_put() */
1404         }
1405
1406         RETURN(rc);
1407 }
1408
1409 /*
1410  * XXX: Another layering violation for now.
1411  *
1412  * We don't want to use ->f_op->read methods, because generic file write
1413  *
1414  *         - serializes on ->i_sem, and
1415  *
1416  *         - does a lot of extra work like balance_dirty_pages(),
1417  *
1418  * which doesn't work for globally shared files like /last_rcvd.
1419  */
1420 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1421 {
1422         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1423
1424         memcpy(buffer, (char *)ei->i_data, buflen);
1425
1426         return  buflen;
1427 }
1428
1429 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1430 {
1431         struct buffer_head *bh;
1432         unsigned long block;
1433         int osize;
1434         int blocksize;
1435         int csize;
1436         int boffs;
1437
1438         /* prevent reading after eof */
1439         spin_lock(&inode->i_lock);
1440         if (i_size_read(inode) < *offs + size) {
1441                 loff_t diff = i_size_read(inode) - *offs;
1442                 spin_unlock(&inode->i_lock);
1443                 if (diff < 0) {
1444                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1445                                i_size_read(inode), *offs);
1446                         return -EBADR;
1447                 } else if (diff == 0) {
1448                         return 0;
1449                 } else {
1450                         size = diff;
1451                 }
1452         } else {
1453                 spin_unlock(&inode->i_lock);
1454         }
1455
1456         blocksize = 1 << inode->i_blkbits;
1457         osize = size;
1458         while (size > 0) {
1459                 block = *offs >> inode->i_blkbits;
1460                 boffs = *offs & (blocksize - 1);
1461                 csize = min(blocksize - boffs, size);
1462                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1463                 if (IS_ERR(bh)) {
1464                         CERROR("%s: can't read %u@%llu on ino %lu: "
1465                                "rc = %ld\n", osd_ino2name(inode),
1466                                csize, *offs, inode->i_ino,
1467                                PTR_ERR(bh));
1468                         return PTR_ERR(bh);
1469                 }
1470
1471                 if (bh != NULL) {
1472                         memcpy(buf, bh->b_data + boffs, csize);
1473                         brelse(bh);
1474                 } else {
1475                         memset(buf, 0, csize);
1476                 }
1477
1478                 *offs += csize;
1479                 buf += csize;
1480                 size -= csize;
1481         }
1482         return osize;
1483 }
1484
1485 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1486                         struct lu_buf *buf, loff_t *pos)
1487 {
1488         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1489         int           rc;
1490
1491         /* Read small symlink from inode body as we need to maintain correct
1492          * on-disk symlinks for ldiskfs.
1493          */
1494         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1495             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1496                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1497         else
1498                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1499
1500         return rc;
1501 }
1502
1503 static inline int osd_extents_enabled(struct super_block *sb,
1504                                       struct inode *inode)
1505 {
1506         if (inode != NULL) {
1507                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1508                         return 1;
1509         } else if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
1510                                 LDISKFS_FEATURE_INCOMPAT_EXTENTS)) {
1511                 return 1;
1512         }
1513         return 0;
1514 }
1515
1516 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1517                            const loff_t size, const loff_t pos,
1518                            const int blocks)
1519 {
1520         int credits, bits, bs, i;
1521
1522         bits = sb->s_blocksize_bits;
1523         bs = 1 << bits;
1524
1525         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1526          * we do not expect blockmaps on the large files,
1527          * so let's shrink it to 2 levels (4GB files) */
1528
1529         /* this is default reservation: 2 levels */
1530         credits = (blocks + 2) * 3;
1531
1532         /* actual offset is unknown, hard to optimize */
1533         if (pos == -1)
1534                 return credits;
1535
1536         /* now check for few specific cases to optimize */
1537         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1538                 /* no indirects */
1539                 credits = blocks;
1540                 /* allocate if not allocated */
1541                 if (inode == NULL) {
1542                         credits += blocks * 2;
1543                         return credits;
1544                 }
1545                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1546                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1547                         if (LDISKFS_I(inode)->i_data[i] == 0)
1548                                 credits += 2;
1549                 }
1550         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1551                 /* single indirect */
1552                 credits = blocks * 3;
1553                 if (inode == NULL ||
1554                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1555                         credits += 3;
1556                 else
1557                         /* The indirect block may be modified. */
1558                         credits += 1;
1559         }
1560
1561         return credits;
1562 }
1563
1564 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1565                                  const struct lu_buf *buf, loff_t _pos,
1566                                  struct thandle *handle)
1567 {
1568         struct osd_object  *obj  = osd_dt_obj(dt);
1569         struct inode       *inode = obj->oo_inode;
1570         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1571         struct osd_thandle *oh;
1572         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1573         int                 bits, bs;
1574         int                 depth, size;
1575         loff_t              pos;
1576         ENTRY;
1577
1578         LASSERT(buf != NULL);
1579         LASSERT(handle != NULL);
1580
1581         oh = container_of0(handle, struct osd_thandle, ot_super);
1582         LASSERT(oh->ot_handle == NULL);
1583
1584         size = buf->lb_len;
1585         bits = sb->s_blocksize_bits;
1586         bs = 1 << bits;
1587
1588         if (_pos == -1) {
1589                 /* if this is an append, then we
1590                  * should expect cross-block record */
1591                 pos = 0;
1592         } else {
1593                 pos = _pos;
1594         }
1595
1596         /* blocks to modify */
1597         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1598         LASSERT(blocks > 0);
1599
1600         if (inode != NULL && _pos != -1) {
1601                 /* object size in blocks */
1602                 est = (i_size_read(inode) + bs - 1) >> bits;
1603                 allocated = inode->i_blocks >> (bits - 9);
1604                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1605                         /* looks like an overwrite, no need to modify tree */
1606                         credits = blocks;
1607                         /* no need to modify i_size */
1608                         goto out;
1609                 }
1610         }
1611
1612         if (osd_extents_enabled(sb, inode)) {
1613                 /*
1614                  * many concurrent threads may grow tree by the time
1615                  * our transaction starts. so, consider 2 is a min depth
1616                  * for every level we may need to allocate a new block
1617                  * and take some entries from the old one. so, 3 blocks
1618                  * to allocate (bitmap, gd, itself) + old block - 4 per
1619                  * level.
1620                  */
1621                 depth = inode != NULL ? ext_depth(inode) : 0;
1622                 depth = max(depth, 1) + 1;
1623                 credits = depth;
1624                 /* if not append, then split may need to modify
1625                  * existing blocks moving entries into the new ones */
1626                 if (_pos != -1)
1627                         credits += depth;
1628                 /* blocks to store data: bitmap,gd,itself */
1629                 credits += blocks * 3;
1630         } else {
1631                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1632         }
1633         /* if inode is created as part of the transaction,
1634          * then it's counted already by the creation method */
1635         if (inode != NULL)
1636                 credits++;
1637
1638 out:
1639
1640         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1641
1642         /* dt_declare_write() is usually called for system objects, such
1643          * as llog or last_rcvd files. We needn't enforce quota on those
1644          * objects, so always set the lqi_space as 0. */
1645         if (inode != NULL)
1646                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1647                                            i_gid_read(inode),
1648                                            i_projid_read(inode), 0,
1649                                            oh, obj, NULL, OSD_QID_BLK);
1650         RETURN(rc);
1651 }
1652
1653 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1654 {
1655         /* LU-2634: clear the extent format for fast symlink */
1656         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1657
1658         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1659         spin_lock(&inode->i_lock);
1660         LDISKFS_I(inode)->i_disksize = buflen;
1661         i_size_write(inode, buflen);
1662         spin_unlock(&inode->i_lock);
1663         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1664
1665         return 0;
1666 }
1667
1668 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1669                              int write_NUL, loff_t *offs, handle_t *handle)
1670 {
1671         struct buffer_head *bh        = NULL;
1672         loff_t              offset    = *offs;
1673         loff_t              new_size  = i_size_read(inode);
1674         unsigned long       block;
1675         int                 blocksize = 1 << inode->i_blkbits;
1676         int                 err = 0;
1677         int                 size;
1678         int                 boffs;
1679         int                 dirty_inode = 0;
1680
1681         if (write_NUL) {
1682                 /*
1683                  * long symlink write does not count the NUL terminator in
1684                  * bufsize, we write it, and the inode's file size does not
1685                  * count the NUL terminator as well.
1686                  */
1687                 ((char *)buf)[bufsize] = '\0';
1688                 ++bufsize;
1689         }
1690
1691         while (bufsize > 0) {
1692                 int credits = handle->h_buffer_credits;
1693
1694                 if (bh)
1695                         brelse(bh);
1696
1697                 block = offset >> inode->i_blkbits;
1698                 boffs = offset & (blocksize - 1);
1699                 size = min(blocksize - boffs, bufsize);
1700                 bh = __ldiskfs_bread(handle, inode, block, 1);
1701                 if (IS_ERR_OR_NULL(bh)) {
1702                         if (bh == NULL) {
1703                                 err = -EIO;
1704                         } else {
1705                                 err = PTR_ERR(bh);
1706                                 bh = NULL;
1707                         }
1708
1709                         CERROR("%s: error reading offset %llu (block %lu, "
1710                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
1711                                inode->i_sb->s_id, offset, block, bufsize, *offs,
1712                                credits, handle->h_buffer_credits, err);
1713                         break;
1714                 }
1715
1716                 err = ldiskfs_journal_get_write_access(handle, bh);
1717                 if (err) {
1718                         CERROR("journal_get_write_access() returned error %d\n",
1719                                err);
1720                         break;
1721                 }
1722                 LASSERTF(boffs + size <= bh->b_size,
1723                          "boffs %d size %d bh->b_size %lu\n",
1724                          boffs, size, (unsigned long)bh->b_size);
1725                 memcpy(bh->b_data + boffs, buf, size);
1726                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1727                 if (err)
1728                         break;
1729
1730                 if (offset + size > new_size)
1731                         new_size = offset + size;
1732                 offset += size;
1733                 bufsize -= size;
1734                 buf += size;
1735         }
1736         if (bh)
1737                 brelse(bh);
1738
1739         if (write_NUL)
1740                 --new_size;
1741         /* correct in-core and on-disk sizes */
1742         if (new_size > i_size_read(inode)) {
1743                 spin_lock(&inode->i_lock);
1744                 if (new_size > i_size_read(inode))
1745                         i_size_write(inode, new_size);
1746                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1747                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1748                         dirty_inode = 1;
1749                 }
1750                 spin_unlock(&inode->i_lock);
1751                 if (dirty_inode)
1752                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1753         }
1754
1755         if (err == 0)
1756                 *offs = offset;
1757         return err;
1758 }
1759
1760 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1761                          const struct lu_buf *buf, loff_t *pos,
1762                          struct thandle *handle, int ignore_quota)
1763 {
1764         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1765         struct osd_thandle      *oh;
1766         ssize_t                 result;
1767         int                     is_link;
1768
1769         LASSERT(dt_object_exists(dt));
1770
1771         LASSERT(handle != NULL);
1772         LASSERT(inode != NULL);
1773         ll_vfs_dq_init(inode);
1774
1775         /* XXX: don't check: one declared chunk can be used many times */
1776         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1777
1778         oh = container_of(handle, struct osd_thandle, ot_super);
1779         LASSERT(oh->ot_handle->h_transaction != NULL);
1780         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1781
1782         /* Write small symlink to inode body as we need to maintain correct
1783          * on-disk symlinks for ldiskfs.
1784          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1785          * does not count it in.
1786          */
1787         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1788         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1789                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1790         else
1791                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1792                                                   buf->lb_len, is_link, pos,
1793                                                   oh->ot_handle);
1794         if (result == 0)
1795                 result = buf->lb_len;
1796
1797         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1798
1799         return result;
1800 }
1801
1802 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1803                              __u64 start, __u64 end, struct thandle *th)
1804 {
1805         struct osd_thandle *oh;
1806         struct inode       *inode;
1807         int                 rc;
1808         ENTRY;
1809
1810         LASSERT(th);
1811         oh = container_of(th, struct osd_thandle, ot_super);
1812
1813         /*
1814          * we don't need to reserve credits for whole truncate
1815          * it's not possible as truncate may need to free too many
1816          * blocks and that won't fit a single transaction. instead
1817          * we reserve credits to change i_size and put inode onto
1818          * orphan list. if needed truncate will extend or restart
1819          * transaction
1820          */
1821         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1822                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1823
1824         inode = osd_dt_obj(dt)->oo_inode;
1825         LASSERT(inode);
1826
1827         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1828                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1829                                    NULL, OSD_QID_BLK);
1830         RETURN(rc);
1831 }
1832
1833 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1834                      __u64 start, __u64 end, struct thandle *th)
1835 {
1836         struct osd_thandle *oh;
1837         struct osd_object  *obj = osd_dt_obj(dt);
1838         struct inode       *inode = obj->oo_inode;
1839         handle_t           *h;
1840         tid_t               tid;
1841         int                rc = 0, rc2 = 0;
1842         ENTRY;
1843
1844         LASSERT(end == OBD_OBJECT_EOF);
1845         LASSERT(dt_object_exists(dt));
1846         LASSERT(osd_invariant(obj));
1847         LASSERT(inode != NULL);
1848         ll_vfs_dq_init(inode);
1849
1850         LASSERT(th);
1851         oh = container_of(th, struct osd_thandle, ot_super);
1852         LASSERT(oh->ot_handle->h_transaction != NULL);
1853
1854         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1855
1856         tid = oh->ot_handle->h_transaction->t_tid;
1857
1858         spin_lock(&inode->i_lock);
1859         i_size_write(inode, start);
1860         spin_unlock(&inode->i_lock);
1861         ll_truncate_pagecache(inode, start);
1862 #ifdef HAVE_INODEOPS_TRUNCATE
1863         if (inode->i_op->truncate) {
1864                 inode->i_op->truncate(inode);
1865         } else
1866 #endif
1867                 ldiskfs_truncate(inode);
1868
1869         /*
1870          * For a partial-page truncate, flush the page to disk immediately to
1871          * avoid data corruption during direct disk write.  b=17397
1872          */
1873         if ((start & ~PAGE_MASK) != 0)
1874                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1875
1876         h = journal_current_handle();
1877         LASSERT(h != NULL);
1878         LASSERT(h == oh->ot_handle);
1879
1880         /* do not check credits with osd_trans_exec_check() as the truncate
1881          * can restart the transaction internally and we restart the
1882          * transaction in this case */
1883
1884         if (tid != h->h_transaction->t_tid) {
1885                 int credits = oh->ot_credits;
1886                 /*
1887                  * transaction has changed during truncate
1888                  * we need to restart the handle with our credits
1889                  */
1890                 if (h->h_buffer_credits < credits) {
1891                         if (ldiskfs_journal_extend(h, credits))
1892                                 rc2 = ldiskfs_journal_restart(h, credits);
1893                 }
1894         }
1895
1896         RETURN(rc == 0 ? rc2 : rc);
1897 }
1898
1899 static int fiemap_check_ranges(struct inode *inode,
1900                                u64 start, u64 len, u64 *new_len)
1901 {
1902         loff_t maxbytes;
1903
1904         *new_len = len;
1905
1906         if (len == 0)
1907                 return -EINVAL;
1908
1909         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1910                 maxbytes = inode->i_sb->s_maxbytes;
1911         else
1912                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1913
1914         if (start > maxbytes)
1915                 return -EFBIG;
1916
1917         /*
1918          * Shrink request scope to what the fs can actually handle.
1919          */
1920         if (len > maxbytes || (maxbytes - len) < start)
1921                 *new_len = maxbytes - start;
1922
1923         return 0;
1924 }
1925
1926 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1927 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
1928
1929 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1930                           struct fiemap *fm)
1931 {
1932         struct fiemap_extent_info fieinfo = {0, };
1933         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1934         u64 len;
1935         int rc;
1936
1937
1938         LASSERT(inode);
1939         if (inode->i_op->fiemap == NULL)
1940                 return -EOPNOTSUPP;
1941
1942         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1943                 return -EINVAL;
1944
1945         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1946         if (rc)
1947                 return rc;
1948
1949         fieinfo.fi_flags = fm->fm_flags;
1950         fieinfo.fi_extents_max = fm->fm_extent_count;
1951         fieinfo.fi_extents_start = fm->fm_extents;
1952
1953         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1954                 filemap_write_and_wait(inode->i_mapping);
1955
1956         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
1957         fm->fm_flags = fieinfo.fi_flags;
1958         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
1959
1960         return rc;
1961 }
1962
1963 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1964                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1965 {
1966         int              rc = 0;
1967         struct inode    *inode = osd_dt_obj(dt)->oo_inode;
1968         ENTRY;
1969
1970         switch (advice) {
1971         case LU_LADVISE_DONTNEED:
1972                 if (end == 0)
1973                         break;
1974                 invalidate_mapping_pages(inode->i_mapping,
1975                                          start >> PAGE_CACHE_SHIFT,
1976                                          (end - 1) >> PAGE_CACHE_SHIFT);
1977                 break;
1978         default:
1979                 rc = -ENOTSUPP;
1980                 break;
1981         }
1982
1983         RETURN(rc);
1984 }
1985
1986 /*
1987  * in some cases we may need declare methods for objects being created
1988  * e.g., when we create symlink
1989  */
1990 const struct dt_body_operations osd_body_ops_new = {
1991         .dbo_declare_write = osd_declare_write,
1992 };
1993
1994 const struct dt_body_operations osd_body_ops = {
1995         .dbo_read                       = osd_read,
1996         .dbo_declare_write              = osd_declare_write,
1997         .dbo_write                      = osd_write,
1998         .dbo_bufs_get                   = osd_bufs_get,
1999         .dbo_bufs_put                   = osd_bufs_put,
2000         .dbo_write_prep                 = osd_write_prep,
2001         .dbo_declare_write_commit       = osd_declare_write_commit,
2002         .dbo_write_commit               = osd_write_commit,
2003         .dbo_read_prep                  = osd_read_prep,
2004         .dbo_declare_punch              = osd_declare_punch,
2005         .dbo_punch                      = osd_punch,
2006         .dbo_fiemap_get                 = osd_fiemap_get,
2007         .dbo_ladvise                    = osd_ladvise,
2008 };