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 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 141 static void dio_complete_routine(struct bio *bio)
 142 {
 143         int error = bio->bi_error;
 144 #else
 145 static void dio_complete_routine(struct bio *bio, int error)
 146 {
 147 #endif
 148         struct osd_iobuf *iobuf = bio->bi_private;
 149         int iter;
 150         struct bio_vec *bvl;
 151
 152         /* CAVEAT EMPTOR: possibly in IRQ context
 153          * DO NOT record procfs stats here!!! */
 154
 155         if (unlikely(iobuf == NULL)) {
 156                 CERROR("***** bio->bi_private is NULL!  This should never "
 157                        "happen.  Normally, I would crash here, but instead I "
 158                        "will dump the bio contents to the console.  Please "
 159                        "report this to <https://jira.whamcloud.com/> , along "
 160                        "with any interesting messages leading up to this point "
 161                        "(like SCSI errors, perhaps).  Because bi_private is "
 162                        "NULL, I can't wake up the thread that initiated this "
 163                        "IO - you will probably have to reboot this node.\n");
 164                 CERROR("bi_next: %p, bi_flags: %lx, "
 165 #ifdef HAVE_BI_RW
 166                        "bi_rw: %lu,"
 167 #else
 168                        "bi_opf: %u,"
 169 #endif
 170                        "bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p,"
 171                        "bi_cnt: %d, bi_private: %p\n", bio->bi_next,
 172                         (unsigned long)bio->bi_flags,
 173 #ifdef HAVE_BI_RW
 174                         bio->bi_rw,
 175 #else
 176                         bio->bi_opf,
 177 #endif
 178                         bio->bi_vcnt, bio_idx(bio),
 179                         bio_sectors(bio) << 9, bio->bi_end_io,
 180 #ifdef HAVE_BI_CNT
 181                         atomic_read(&bio->bi_cnt),
 182 #else
 183                         atomic_read(&bio->__bi_cnt),
 184 #endif
 185                         bio->bi_private);
 186                 return;
 187         }
 188
 189         /* the check is outside of the cycle for performance reason -bzzz */
 190         if (!bio_data_dir(bio)) {
 191                 bio_for_each_segment_all(bvl, bio, iter) {
 192                         if (likely(error == 0))
 193                                 SetPageUptodate(bvl_to_page(bvl));
 194                         LASSERT(PageLocked(bvl_to_page(bvl)));
 195                 }
 196                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 197         } else {
 198                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 199         }
 200
 201         /* any real error is good enough -bzzz */
 202         if (error != 0 && iobuf->dr_error == 0)
 203                 iobuf->dr_error = error;
 204
 205         /*
 206          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 207          * it's possible that service thread will see dr_numreqs
 208          * is zero, but dr_elapsed is not set yet, leading to lost
 209          * data in this processing and an assertion in a subsequent
 210          * call to OSD.
 211          */
 212         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 213                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
 214                 iobuf->dr_elapsed_valid = 1;
 215         }
 216         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 217                 wake_up(&iobuf->dr_wait);
 218
 219         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 220          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 221          * mempool when serious on-disk fragmentation was encountered,
 222          * deadlocking the OST.  The bios are now released as soon as complete
 223          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 224         bio_put(bio);
 225 }
 226
 227 static void record_start_io(struct osd_iobuf *iobuf, int size)
 228 {
 229         struct osd_device    *osd = iobuf->dr_dev;
 230         struct obd_histogram *h = osd->od_brw_stats.hist;
 231
 232         iobuf->dr_frags++;
 233         atomic_inc(&iobuf->dr_numreqs);
 234
 235         if (iobuf->dr_rw == 0) {
 236                 atomic_inc(&osd->od_r_in_flight);
 237                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 238                                  atomic_read(&osd->od_r_in_flight));
 239                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 240         } else if (iobuf->dr_rw == 1) {
 241                 atomic_inc(&osd->od_w_in_flight);
 242                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 243                                  atomic_read(&osd->od_w_in_flight));
 244                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 245         } else {
 246                 LBUG();
 247         }
 248 }
 249
 250 static void osd_submit_bio(int rw, struct bio *bio)
 251 {
 252         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 253 #ifdef HAVE_SUBMIT_BIO_2ARGS
 254         if (rw == 0)
 255                 submit_bio(READ, bio);
 256         else
 257                 submit_bio(WRITE, bio);
 258 #else
 259         bio->bi_opf |= rw;
 260         submit_bio(bio);
 261 #endif
 262 }
 263
 264 static int can_be_merged(struct bio *bio, sector_t sector)
 265 {
 266         if (bio == NULL)
 267                 return 0;
 268
 269         return bio_end_sector(bio) == sector ? 1 : 0;
 270 }
 271
 272 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 273                       struct osd_iobuf *iobuf)
 274 {
 275         int            blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 276         struct page  **pages = iobuf->dr_pages;
 277         int            npages = iobuf->dr_npages;
 278         sector_t      *blocks = iobuf->dr_blocks;
 279         int            total_blocks = npages * blocks_per_page;
 280         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 281         unsigned int   blocksize = inode->i_sb->s_blocksize;
 282         struct bio    *bio = NULL;
 283         struct page   *page;
 284         unsigned int   page_offset;
 285         sector_t       sector;
 286         int            nblocks;
 287         int            block_idx;
 288         int            page_idx;
 289         int            i;
 290         int            rc = 0;
 291         DECLARE_PLUG(plug);
 292         ENTRY;
 293
 294         LASSERT(iobuf->dr_npages == npages);
 295
 296         osd_brw_stats_update(osd, iobuf);
 297         iobuf->dr_start_time = cfs_time_current();
 298
 299         blk_start_plug(&plug);
 300         for (page_idx = 0, block_idx = 0;
 301              page_idx < npages;
 302              page_idx++, block_idx += blocks_per_page) {
 303
 304                 page = pages[page_idx];
 305                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 306
 307                 for (i = 0, page_offset = 0;
 308                      i < blocks_per_page;
 309                      i += nblocks, page_offset += blocksize * nblocks) {
 310
 311                         nblocks = 1;
 312
 313                         if (blocks[block_idx + i] == 0) {  /* hole */
 314                                 LASSERTF(iobuf->dr_rw == 0,
 315                                          "page_idx %u, block_idx %u, i %u\n",
 316                                          page_idx, block_idx, i);
 317                                 memset(kmap(page) + page_offset, 0, blocksize);
 318                                 kunmap(page);
 319                                 continue;
 320                         }
 321
 322                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 323
 324                         /* Additional contiguous file blocks? */
 325                         while (i + nblocks < blocks_per_page &&
 326                                (sector + (nblocks << sector_bits)) ==
 327                                ((sector_t)blocks[block_idx + i + nblocks] <<
 328                                 sector_bits))
 329                                 nblocks++;
 330
 331                         if (bio != NULL &&
 332                             can_be_merged(bio, sector) &&
 333                             bio_add_page(bio, page,
 334                                          blocksize * nblocks, page_offset) != 0)
 335                                 continue;       /* added this frag OK */
 336
 337                         if (bio != NULL) {
 338                                 struct request_queue *q =
 339                                         bdev_get_queue(bio->bi_bdev);
 340                                 unsigned int bi_size = bio_sectors(bio) << 9;
 341
 342                                 /* Dang! I have to fragment this I/O */
 343                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 344                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 345                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 346                                        bio_sectors(bio),
 347                                        queue_max_sectors(q),
 348                                        bio_phys_segments(q, bio),
 349                                        queue_max_phys_segments(q),
 350                                        0, queue_max_hw_segments(q));
 351                                 record_start_io(iobuf, bi_size);
 352                                 osd_submit_bio(iobuf->dr_rw, bio);
 353                         }
 354
 355                         /* allocate new bio */
 356                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 357                                                       (npages - page_idx) *
 358                                                       blocks_per_page));
 359                         if (bio == NULL) {
 360                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 361                                        (npages - page_idx), blocks_per_page,
 362                                        (npages - page_idx) * blocks_per_page);
 363                                 rc = -ENOMEM;
 364                                 goto out;
 365                         }
 366
 367                         bio->bi_bdev = inode->i_sb->s_bdev;
 368                         bio_set_sector(bio, sector);
 369 #ifdef HAVE_BI_RW
 370                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
 371 #else
 372                         bio->bi_opf = (iobuf->dr_rw == 0) ? READ : WRITE;
 373 #endif
 374                         bio->bi_end_io = dio_complete_routine;
 375                         bio->bi_private = iobuf;
 376
 377                         rc = bio_add_page(bio, page,
 378                                           blocksize * nblocks, page_offset);
 379                         LASSERT(rc != 0);
 380                 }
 381         }
 382
 383         if (bio != NULL) {
 384                 record_start_io(iobuf, bio_sectors(bio) << 9);
 385                 osd_submit_bio(iobuf->dr_rw, bio);
 386                 rc = 0;
 387         }
 388
 389 out:
 390         blk_finish_plug(&plug);
 391
 392         /* in order to achieve better IO throughput, we don't wait for writes
 393          * completion here. instead we proceed with transaction commit in
 394          * parallel and wait for IO completion once transaction is stopped
 395          * see osd_trans_stop() for more details -bzzz */
 396         if (iobuf->dr_rw == 0) {
 397                 wait_event(iobuf->dr_wait,
 398                            atomic_read(&iobuf->dr_numreqs) == 0);
 399                 osd_fini_iobuf(osd, iobuf);
 400         }
 401
 402         if (rc == 0)
 403                 rc = iobuf->dr_error;
 404         RETURN(rc);
 405 }
 406
 407 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 408                                    struct niobuf_local *lnb)
 409 {
 410         ENTRY;
 411
 412         *nrpages = 0;
 413
 414         while (len > 0) {
 415                 int poff = offset & (PAGE_SIZE - 1);
 416                 int plen = PAGE_SIZE - poff;
 417
 418                 if (plen > len)
 419                         plen = len;
 420                 lnb->lnb_file_offset = offset;
 421                 lnb->lnb_page_offset = poff;
 422                 lnb->lnb_len = plen;
 423                 /* lnb->lnb_flags = rnb->rnb_flags; */
 424                 lnb->lnb_flags = 0;
 425                 lnb->lnb_page = NULL;
 426                 lnb->lnb_rc = 0;
 427
 428                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 429                          (long long) len);
 430                 offset += plen;
 431                 len -= plen;
 432                 lnb++;
 433                 (*nrpages)++;
 434         }
 435
 436         RETURN(0);
 437 }
 438
 439 static struct page *osd_get_page(struct dt_object *dt, loff_t offset,
 440                                  gfp_t gfp_mask)
 441 {
 442         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 443         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 444         struct page *page;
 445
 446         LASSERT(inode);
 447
 448         page = find_or_create_page(inode->i_mapping, offset >> PAGE_SHIFT,
 449                                    gfp_mask);
 450
 451         if (unlikely(page == NULL))
 452                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 453
 454         return page;
 455 }
 456
 457 /*
 458  * there are following "locks":
 459  * journal_start
 460  * i_mutex
 461  * page lock
 462  *
 463  * osd write path:
 464  *  - lock page(s)
 465  *  - journal_start
 466  *  - truncate_sem
 467  *
 468  * ext4 vmtruncate:
 469  *  - lock pages, unlock
 470  *  - journal_start
 471  *  - lock partial page
 472  *  - i_data_sem
 473  *
 474  */
 475
 476 /**
 477  * Unlock and release pages loaded by osd_bufs_get()
 478  *
 479  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 480  *
 481  * \param env           thread execution environment
 482  * \param dt            dt object undergoing IO (OSD object + methods)
 483  * \param lnb           array of pages undergoing IO
 484  * \param npages        number of pages in \a lnb
 485  *
 486  * \retval 0            always
 487  */
 488 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 489                         struct niobuf_local *lnb, int npages)
 490 {
 491         int i;
 492
 493         for (i = 0; i < npages; i++) {
 494                 if (lnb[i].lnb_page == NULL)
 495                         continue;
 496                 LASSERT(PageLocked(lnb[i].lnb_page));
 497                 unlock_page(lnb[i].lnb_page);
 498                 put_page(lnb[i].lnb_page);
 499                 dt_object_put(env, dt);
 500                 lnb[i].lnb_page = NULL;
 501         }
 502
 503         RETURN(0);
 504 }
 505
 506 /**
 507  * Load and lock pages undergoing IO
 508  *
 509  * Pages as described in the \a lnb array are fetched (from disk or cache)
 510  * and locked for IO by the caller.
 511  *
 512  * DLM locking protects us from write and truncate competing for same region,
 513  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 514  * It's possible the writeout on a such a page is in progress when we access
 515  * it. It's also possible that during this writeout we put new (partial) data
 516  * into the page, but won't be able to proceed in filter_commitrw_write().
 517  * Therefore, just wait for writeout completion as it should be rare enough.
 518  *
 519  * \param env           thread execution environment
 520  * \param dt            dt object undergoing IO (OSD object + methods)
 521  * \param pos           byte offset of IO start
 522  * \param len           number of bytes of IO
 523  * \param lnb           array of extents undergoing IO
 524  * \param rw            read or write operation, and other flags
 525  * \param capa          capabilities
 526  *
 527  * \retval pages        (zero or more) loaded successfully
 528  * \retval -ENOMEM      on memory/page allocation error
 529  */
 530 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 531                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 532                         enum dt_bufs_type rw)
 533 {
 534         struct osd_object *obj = osd_dt_obj(dt);
 535         int npages, i, rc = 0;
 536         gfp_t gfp_mask;
 537
 538         LASSERT(obj->oo_inode);
 539
 540         osd_map_remote_to_local(pos, len, &npages, lnb);
 541
 542         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 543         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 544                                              GFP_HIGHUSER;
 545         for (i = 0; i < npages; i++, lnb++) {
 546                 lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset,
 547                                              gfp_mask);
 548                 if (lnb->lnb_page == NULL)
 549                         GOTO(cleanup, rc = -ENOMEM);
 550
 551                 wait_on_page_writeback(lnb->lnb_page);
 552                 BUG_ON(PageWriteback(lnb->lnb_page));
 553
 554                 lu_object_get(&dt->do_lu);
 555         }
 556
 557         RETURN(i);
 558
 559 cleanup:
 560         if (i > 0)
 561                 osd_bufs_put(env, dt, lnb - i, i);
 562         return rc;
 563 }
 564
 565 #ifndef HAVE_LDISKFS_MAP_BLOCKS
 566
 567 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 568 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 569 #endif
 570
 571 struct bpointers {
 572         sector_t *blocks;
 573         unsigned long start;
 574         int num;
 575         int init_num;
 576         int create;
 577 };
 578
 579 static long ldiskfs_ext_find_goal(struct inode *inode,
 580                                   struct ldiskfs_ext_path *path,
 581                                   unsigned long block, int *aflags)
 582 {
 583         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 584         unsigned long bg_start;
 585         unsigned long colour;
 586         int depth;
 587
 588         if (path) {
 589                 struct ldiskfs_extent *ex;
 590                 depth = path->p_depth;
 591
 592                 /* try to predict block placement */
 593                 if ((ex = path[depth].p_ext))
 594                         return ldiskfs_ext_pblock(ex) +
 595                                 (block - le32_to_cpu(ex->ee_block));
 596
 597                 /* it looks index is empty
 598                  * try to find starting from index itself */
 599                 if (path[depth].p_bh)
 600                         return path[depth].p_bh->b_blocknr;
 601         }
 602
 603         /* OK. use inode's group */
 604         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 605                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 606         colour = (current->pid % 16) *
 607                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 608         return bg_start + colour + block;
 609 }
 610
 611 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 612                                 struct ldiskfs_ext_path *path,
 613                                 unsigned long block, unsigned long *count,
 614                                 int *err)
 615 {
 616         struct ldiskfs_allocation_request ar;
 617         unsigned long pblock;
 618         int aflags;
 619
 620         /* find neighbour allocated blocks */
 621         ar.lleft = block;
 622         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 623         if (*err)
 624                 return 0;
 625         ar.lright = block;
 626         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 627         if (*err)
 628                 return 0;
 629
 630         /* allocate new block */
 631         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 632         ar.inode = inode;
 633         ar.logical = block;
 634         ar.len = *count;
 635         ar.flags = LDISKFS_MB_HINT_DATA;
 636         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 637         *count = ar.len;
 638         return pblock;
 639 }
 640
 641 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 642                                      struct ldiskfs_ext_path *path,
 643                                      struct ldiskfs_ext_cache *cex,
 644 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 645                                      struct ldiskfs_extent *ex,
 646 #endif
 647                                      void *cbdata)
 648 {
 649         struct bpointers *bp = cbdata;
 650         struct ldiskfs_extent nex;
 651         unsigned long pblock = 0;
 652         unsigned long tgen;
 653         int err, i;
 654         unsigned long count;
 655         handle_t *handle;
 656
 657 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 658         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 659 #else
 660         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 661 #endif
 662                 err = EXT_CONTINUE;
 663                 goto map;
 664         }
 665
 666         if (bp->create == 0) {
 667                 i = 0;
 668                 if (cex->ec_block < bp->start)
 669                         i = bp->start - cex->ec_block;
 670                 if (i >= cex->ec_len)
 671                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 672                                         i, cex->ec_len);
 673                 for (; i < cex->ec_len && bp->num; i++) {
 674                         *(bp->blocks) = 0;
 675                         bp->blocks++;
 676                         bp->num--;
 677                         bp->start++;
 678                 }
 679
 680                 return EXT_CONTINUE;
 681         }
 682
 683         tgen = LDISKFS_I(inode)->i_ext_generation;
 684         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
 685
 686         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
 687                                    count + LDISKFS_ALLOC_NEEDED + 1);
 688         if (IS_ERR(handle)) {
 689                 return PTR_ERR(handle);
 690         }
 691
 692         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
 693                 /* the tree has changed. so path can be invalid at moment */
 694                 ldiskfs_journal_stop(handle);
 695                 return EXT_REPEAT;
 696         }
 697
 698         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
 699          * protected by i_data_sem as whole. so we patch it to store
 700          * generation to path and now verify the tree hasn't changed */
 701         down_write((&LDISKFS_I(inode)->i_data_sem));
 702
 703         /* validate extent, make sure the extent tree does not changed */
 704         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
 705                 /* cex is invalid, try again */
 706                 up_write(&LDISKFS_I(inode)->i_data_sem);
 707                 ldiskfs_journal_stop(handle);
 708                 return EXT_REPEAT;
 709         }
 710
 711         count = cex->ec_len;
 712         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
 713         if (!pblock)
 714                 goto out;
 715         BUG_ON(count > cex->ec_len);
 716
 717         /* insert new extent */
 718         nex.ee_block = cpu_to_le32(cex->ec_block);
 719         ldiskfs_ext_store_pblock(&nex, pblock);
 720         nex.ee_len = cpu_to_le16(count);
 721         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
 722         if (err) {
 723                 /* free data blocks we just allocated */
 724                 /* not a good idea to call discard here directly,
 725                  * but otherwise we'd need to call it every free() */
 726                 ldiskfs_discard_preallocations(inode);
 727 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
 728                 ldiskfs_free_blocks(handle, inode, NULL,
 729                                     ldiskfs_ext_pblock(&nex),
 730                                     le16_to_cpu(nex.ee_len), 0);
 731 #else
 732                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
 733                                     le16_to_cpu(nex.ee_len), 0);
 734 #endif
 735                 goto out;
 736         }
 737
 738         /*
 739          * Putting len of the actual extent we just inserted,
 740          * we are asking ldiskfs_ext_walk_space() to continue
 741          * scaning after that block
 742          */
 743         cex->ec_len = le16_to_cpu(nex.ee_len);
 744         cex->ec_start = ldiskfs_ext_pblock(&nex);
 745         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
 746         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
 747
 748 out:
 749         up_write((&LDISKFS_I(inode)->i_data_sem));
 750         ldiskfs_journal_stop(handle);
 751 map:
 752         if (err >= 0) {
 753                 /* map blocks */
 754                 if (bp->num == 0) {
 755                         CERROR("hmm. why do we find this extent?\n");
 756                         CERROR("initial space: %lu:%u\n",
 757                                 bp->start, bp->init_num);
 758 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 759                         CERROR("current extent: %u/%u/%llu %d\n",
 760                                 cex->ec_block, cex->ec_len,
 761                                 (unsigned long long)cex->ec_start,
 762                                 cex->ec_type);
 763 #else
 764                         CERROR("current extent: %u/%u/%llu\n",
 765                                 cex->ec_block, cex->ec_len,
 766                                 (unsigned long long)cex->ec_start);
 767 #endif
 768                 }
 769                 i = 0;
 770                 if (cex->ec_block < bp->start)
 771                         i = bp->start - cex->ec_block;
 772                 if (i >= cex->ec_len)
 773                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 774                                         i, cex->ec_len);
 775                 for (; i < cex->ec_len && bp->num; i++) {
 776                         *(bp->blocks) = cex->ec_start + i;
 777                         if (pblock != 0) {
 778                                 /* unmap any possible underlying metadata from
 779                                  * the block device mapping.  bug 6998. */
 780 #ifndef HAVE_CLEAN_BDEV_ALIASES
 781                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
 782                                                           *(bp->blocks));
 783 #else
 784                                 clean_bdev_aliases(inode->i_sb->s_bdev,
 785                                                    *(bp->blocks), 1);
 786 #endif
 787                         }
 788                         bp->blocks++;
 789                         bp->num--;
 790                         bp->start++;
 791                 }
 792         }
 793         return err;
 794 }
 795
 796 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
 797                                    int clen, sector_t *blocks, int create)
 798 {
 799         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 800         struct bpointers bp;
 801         int err;
 802
 803         if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
 804                 return -EFBIG;
 805
 806         bp.blocks = blocks;
 807         bp.start = index * blocks_per_page;
 808         bp.init_num = bp.num = clen * blocks_per_page;
 809         bp.create = create;
 810
 811         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
 812                bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
 813
 814         err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
 815                                      ldiskfs_ext_new_extent_cb, &bp);
 816         ldiskfs_ext_invalidate_cache(inode);
 817
 818         return err;
 819 }
 820
 821 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
 822                                           struct page **page, int pages,
 823                                           sector_t *blocks, int create)
 824 {
 825         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 826         pgoff_t bitmap_max_page_index;
 827         sector_t *b;
 828         int rc = 0, i;
 829
 830         bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
 831                                 PAGE_SHIFT;
 832         for (i = 0, b = blocks; i < pages; i++, page++) {
 833                 if ((*page)->index + 1 >= bitmap_max_page_index) {
 834                         rc = -EFBIG;
 835                         break;
 836                 }
 837                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
 838                 if (rc) {
 839                         CERROR("ino %lu, blk %llu create %d: rc %d\n",
 840                                inode->i_ino,
 841                                (unsigned long long)*b, create, rc);
 842                         break;
 843                 }
 844                 b += blocks_per_page;
 845         }
 846         return rc;
 847 }
 848
 849 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
 850                                            struct page **page,
 851                                            int pages, sector_t *blocks,
 852                                            int create)
 853 {
 854         int rc = 0, i = 0, clen = 0;
 855         struct page *fp = NULL;
 856
 857         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 858                 inode->i_ino, pages, (*page)->index);
 859
 860         /* pages are sorted already. so, we just have to find
 861          * contig. space and process them properly */
 862         while (i < pages) {
 863                 if (fp == NULL) {
 864                         /* start new extent */
 865                         fp = *page++;
 866                         clen = 1;
 867                         i++;
 868                         continue;
 869                 } else if (fp->index + clen == (*page)->index) {
 870                         /* continue the extent */
 871                         page++;
 872                         clen++;
 873                         i++;
 874                         continue;
 875                 }
 876
 877                 /* process found extent */
 878                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 879                                              blocks, create);
 880                 if (rc)
 881                         GOTO(cleanup, rc);
 882
 883                 /* look for next extent */
 884                 fp = NULL;
 885                 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
 886         }
 887
 888         if (fp)
 889                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
 890                                              blocks, create);
 891
 892 cleanup:
 893         return rc;
 894 }
 895
 896 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 897                                        int pages, sector_t *blocks,
 898                                        int create)
 899 {
 900         int rc;
 901
 902         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 903                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
 904                                                      blocks, create);
 905                 return rc;
 906         }
 907         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
 908
 909         return rc;
 910 }
 911 #else
 912 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 913                                        int pages, sector_t *blocks,
 914                                        int create)
 915 {
 916         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 917         int rc = 0, i = 0;
 918         struct page *fp = NULL;
 919         int clen = 0;
 920         pgoff_t max_page_index;
 921         handle_t *handle = NULL;
 922
 923         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
 924
 925         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 926                 inode->i_ino, pages, (*page)->index);
 927
 928         if (create) {
 929                 create = LDISKFS_GET_BLOCKS_CREATE;
 930                 handle = ldiskfs_journal_current_handle();
 931                 LASSERT(handle != NULL);
 932                 rc = osd_attach_jinode(inode);
 933                 if (rc)
 934                         return rc;
 935         }
 936         /* pages are sorted already. so, we just have to find
 937          * contig. space and process them properly */
 938         while (i < pages) {
 939                 long blen, total = 0;
 940                 struct ldiskfs_map_blocks map = { 0 };
 941
 942                 if (fp == NULL) { /* start new extent */
 943                         fp = *page++;
 944                         clen = 1;
 945                         if (++i != pages)
 946                                 continue;
 947                 } else if (fp->index + clen == (*page)->index) {
 948                         /* continue the extent */
 949                         page++;
 950                         clen++;
 951                         if (++i != pages)
 952                                 continue;
 953                 }
 954                 if (fp->index + clen >= max_page_index)
 955                         GOTO(cleanup, rc = -EFBIG);
 956                 /* process found extent */
 957                 map.m_lblk = fp->index * blocks_per_page;
 958                 map.m_len = blen = clen * blocks_per_page;
 959 cont_map:
 960                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 961                 if (rc >= 0) {
 962                         int c = 0;
 963                         for (; total < blen && c < map.m_len; c++, total++) {
 964                                 if (rc == 0) {
 965                                         *(blocks + total) = 0;
 966                                         total++;
 967                                         break;
 968                                 } else {
 969                                         *(blocks + total) = map.m_pblk + c;
 970                                         /* unmap any possible underlying
 971                                          * metadata from the block device
 972                                          * mapping.  bug 6998. */
 973                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
 974                                             create)
 975 #ifndef HAVE_CLEAN_BDEV_ALIASES
 976                                                 unmap_underlying_metadata(
 977                                                         inode->i_sb->s_bdev,
 978                                                         map.m_pblk + c);
 979 #else
 980                                                 clean_bdev_aliases(
 981                                                         inode->i_sb->s_bdev,
 982                                                         map.m_pblk + c, 1);
 983 #endif
 984                                 }
 985                         }
 986                         rc = 0;
 987                 }
 988                 if (rc == 0 && total < blen) {
 989                         map.m_lblk = fp->index * blocks_per_page + total;
 990                         map.m_len = blen - total;
 991                         goto cont_map;
 992                 }
 993                 if (rc != 0)
 994                         GOTO(cleanup, rc);
 995
 996                 /* look for next extent */
 997                 fp = NULL;
 998                 blocks += blocks_per_page * clen;
 999         }
1000 cleanup:
1001         return rc;
1002 }
1003 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
1004
1005 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1006                           struct niobuf_local *lnb, int npages)
1007 {
1008         struct osd_thread_info *oti   = osd_oti_get(env);
1009         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1010         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
1011         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
1012         ktime_t start;
1013         ktime_t end;
1014         s64 timediff;
1015         ssize_t                 isize;
1016         __s64                   maxidx;
1017         int                     rc = 0;
1018         int                     i;
1019         int                     cache = 0;
1020
1021         LASSERT(inode);
1022
1023         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1024         if (unlikely(rc != 0))
1025                 RETURN(rc);
1026
1027         isize = i_size_read(inode);
1028         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1029
1030         if (osd->od_writethrough_cache)
1031                 cache = 1;
1032         if (isize > osd->od_readcache_max_filesize)
1033                 cache = 0;
1034
1035         start = ktime_get();
1036         for (i = 0; i < npages; i++) {
1037
1038                 if (cache == 0)
1039                         generic_error_remove_page(inode->i_mapping,
1040                                                   lnb[i].lnb_page);
1041
1042                 /*
1043                  * till commit the content of the page is undefined
1044                  * we'll set it uptodate once bulk is done. otherwise
1045                  * subsequent reads can access non-stable data
1046                  */
1047                 ClearPageUptodate(lnb[i].lnb_page);
1048
1049                 if (lnb[i].lnb_len == PAGE_SIZE)
1050                         continue;
1051
1052                 if (maxidx >= lnb[i].lnb_page->index) {
1053                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1054                 } else {
1055                         long off;
1056                         char *p = kmap(lnb[i].lnb_page);
1057
1058                         off = lnb[i].lnb_page_offset;
1059                         if (off)
1060                                 memset(p, 0, off);
1061                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1062                               ~PAGE_MASK;
1063                         if (off)
1064                                 memset(p + off, 0, PAGE_SIZE - off);
1065                         kunmap(lnb[i].lnb_page);
1066                 }
1067         }
1068         end = ktime_get();
1069         timediff = ktime_us_delta(end, start);
1070         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1071
1072         if (iobuf->dr_npages) {
1073                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1074                                                  iobuf->dr_npages,
1075                                                  iobuf->dr_blocks, 0);
1076                 if (likely(rc == 0)) {
1077                         rc = osd_do_bio(osd, inode, iobuf);
1078                         /* do IO stats for preparation reads */
1079                         osd_fini_iobuf(osd, iobuf);
1080                 }
1081         }
1082         RETURN(rc);
1083 }
1084
1085 struct osd_fextent {
1086         sector_t        start;
1087         sector_t        end;
1088         unsigned int    mapped:1;
1089 };
1090
1091 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1092                          struct osd_fextent *cached_extent)
1093 {
1094         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1095         sector_t block = offset >> inode->i_blkbits;
1096         sector_t start;
1097         struct fiemap_extent_info fei = { 0 };
1098         struct fiemap_extent fe = { 0 };
1099         mm_segment_t saved_fs;
1100         int rc;
1101
1102         if (block >= cached_extent->start && block < cached_extent->end)
1103                 return cached_extent->mapped;
1104
1105         if (i_size_read(inode) == 0)
1106                 return 0;
1107
1108         /* Beyond EOF, must not be mapped */
1109         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1110                 return 0;
1111
1112         fei.fi_extents_max = 1;
1113         fei.fi_extents_start = &fe;
1114
1115         saved_fs = get_fs();
1116         set_fs(get_ds());
1117         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1118         set_fs(saved_fs);
1119         if (rc != 0)
1120                 return 0;
1121
1122         start = fe.fe_logical >> inode->i_blkbits;
1123
1124         if (start > block) {
1125                 cached_extent->start = block;
1126                 cached_extent->end = start;
1127                 cached_extent->mapped = 0;
1128         } else {
1129                 cached_extent->start = start;
1130                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1131                                       inode->i_blkbits;
1132                 cached_extent->mapped = 1;
1133         }
1134
1135         return cached_extent->mapped;
1136 }
1137
1138 static int osd_declare_write_commit(const struct lu_env *env,
1139                                     struct dt_object *dt,
1140                                     struct niobuf_local *lnb, int npages,
1141                                     struct thandle *handle)
1142 {
1143         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1144         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1145         struct osd_thandle      *oh;
1146         int                     extents = 1;
1147         int                     depth;
1148         int                     i;
1149         int                     newblocks;
1150         int                     rc = 0;
1151         int                     flags = 0;
1152         int                     credits = 0;
1153         long long               quota_space = 0;
1154         struct osd_fextent      extent = { 0 };
1155         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1156         ENTRY;
1157
1158         LASSERT(handle != NULL);
1159         oh = container_of0(handle, struct osd_thandle, ot_super);
1160         LASSERT(oh->ot_handle == NULL);
1161
1162         newblocks = npages;
1163
1164         /* calculate number of extents (probably better to pass nb) */
1165         for (i = 0; i < npages; i++) {
1166                 if (i && lnb[i].lnb_file_offset !=
1167                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1168                         extents++;
1169
1170                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1171                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1172                 else
1173                         quota_space += PAGE_SIZE;
1174
1175                 /* ignore quota for the whole request if any page is from
1176                  * client cache or written by root.
1177                  *
1178                  * XXX once we drop the 1.8 client support, the checking
1179                  * for whether page is from cache can be simplified as:
1180                  * !(lnb[i].flags & OBD_BRW_SYNC)
1181                  *
1182                  * XXX we could handle this on per-lnb basis as done by
1183                  * grant. */
1184                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1185                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1186                     OBD_BRW_FROM_GRANT)
1187                         declare_flags |= OSD_QID_FORCE;
1188         }
1189
1190         /*
1191          * each extent can go into new leaf causing a split
1192          * 5 is max tree depth: inode + 4 index blocks
1193          * with blockmaps, depth is 3 at most
1194          */
1195         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1196                 /*
1197                  * many concurrent threads may grow tree by the time
1198                  * our transaction starts. so, consider 2 is a min depth
1199                  */
1200                 depth = ext_depth(inode);
1201                 depth = max(depth, 1) + 1;
1202                 newblocks += depth;
1203                 credits++; /* inode */
1204                 credits += depth * 2 * extents;
1205         } else {
1206                 depth = 3;
1207                 newblocks += depth;
1208                 credits++; /* inode */
1209                 credits += depth * extents;
1210         }
1211
1212         /* quota space for metadata blocks */
1213         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1214
1215         /* quota space should be reported in 1K blocks */
1216         quota_space = toqb(quota_space);
1217
1218         /* each new block can go in different group (bitmap + gd) */
1219
1220         /* we can't dirty more bitmap blocks than exist */
1221         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1222                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1223         else
1224                 credits += newblocks;
1225
1226         /* we can't dirty more gd blocks than exist */
1227         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1228                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1229         else
1230                 credits += newblocks;
1231
1232         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1233
1234         /* make sure the over quota flags were not set */
1235         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1236
1237         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1238                                    i_projid_read(inode), quota_space, oh,
1239                                    osd_dt_obj(dt), &flags, declare_flags);
1240
1241         /* we need only to store the overquota flags in the first lnb for
1242          * now, once we support multiple objects BRW, this code needs be
1243          * revised. */
1244         if (flags & QUOTA_FL_OVER_USRQUOTA)
1245                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1246         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1247                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1248         if (flags & QUOTA_FL_OVER_PRJQUOTA)
1249                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1250
1251         RETURN(rc);
1252 }
1253
1254 /* Check if a block is allocated or not */
1255 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1256                             struct niobuf_local *lnb, int npages,
1257                             struct thandle *thandle)
1258 {
1259         struct osd_thread_info *oti = osd_oti_get(env);
1260         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1261         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1262         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1263         loff_t isize;
1264         int rc = 0, i;
1265
1266         LASSERT(inode);
1267
1268         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1269         if (unlikely(rc != 0))
1270                 RETURN(rc);
1271
1272         isize = i_size_read(inode);
1273         ll_vfs_dq_init(inode);
1274
1275         for (i = 0; i < npages; i++) {
1276                 if (lnb[i].lnb_rc == -ENOSPC &&
1277                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1278                         /* Allow the write to proceed if overwriting an
1279                          * existing block */
1280                         lnb[i].lnb_rc = 0;
1281                 }
1282
1283                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1284                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1285                                lnb[i].lnb_rc);
1286                         LASSERT(lnb[i].lnb_page);
1287                         generic_error_remove_page(inode->i_mapping,
1288                                                   lnb[i].lnb_page);
1289                         continue;
1290                 }
1291
1292                 LASSERT(PageLocked(lnb[i].lnb_page));
1293                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1294
1295                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1296                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1297
1298                 /*
1299                  * Since write and truncate are serialized by oo_sem, even
1300                  * partial-page truncate should not leave dirty pages in the
1301                  * page cache.
1302                  */
1303                 LASSERT(!PageDirty(lnb[i].lnb_page));
1304
1305                 SetPageUptodate(lnb[i].lnb_page);
1306
1307                 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1308         }
1309
1310         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1311
1312         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1313                 rc = -ENOSPC;
1314         } else if (iobuf->dr_npages > 0) {
1315                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1316                                                  iobuf->dr_npages,
1317                                                  iobuf->dr_blocks, 1);
1318         } else {
1319                 /* no pages to write, no transno is needed */
1320                 thandle->th_local = 1;
1321         }
1322
1323         if (likely(rc == 0)) {
1324                 spin_lock(&inode->i_lock);
1325                 if (isize > i_size_read(inode)) {
1326                         i_size_write(inode, isize);
1327                         LDISKFS_I(inode)->i_disksize = isize;
1328                         spin_unlock(&inode->i_lock);
1329                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1330                 } else {
1331                         spin_unlock(&inode->i_lock);
1332                 }
1333
1334                 rc = osd_do_bio(osd, inode, iobuf);
1335                 /* we don't do stats here as in read path because
1336                  * write is async: we'll do this in osd_put_bufs() */
1337         } else {
1338                 osd_fini_iobuf(osd, iobuf);
1339         }
1340
1341         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1342
1343         if (unlikely(rc != 0)) {
1344                 /* if write fails, we should drop pages from the cache */
1345                 for (i = 0; i < npages; i++) {
1346                         if (lnb[i].lnb_page == NULL)
1347                                 continue;
1348                         LASSERT(PageLocked(lnb[i].lnb_page));
1349                         generic_error_remove_page(inode->i_mapping,
1350                                                   lnb[i].lnb_page);
1351                 }
1352         }
1353
1354         RETURN(rc);
1355 }
1356
1357 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1358                          struct niobuf_local *lnb, int npages)
1359 {
1360         struct osd_thread_info *oti = osd_oti_get(env);
1361         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1362         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1363         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1364         int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1365         ktime_t start, end;
1366         s64 timediff;
1367         loff_t isize;
1368
1369         LASSERT(inode);
1370
1371         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1372         if (unlikely(rc != 0))
1373                 RETURN(rc);
1374
1375         isize = i_size_read(inode);
1376
1377         if (osd->od_read_cache)
1378                 cache = 1;
1379         if (isize > osd->od_readcache_max_filesize)
1380                 cache = 0;
1381
1382         start = ktime_get();
1383         for (i = 0; i < npages; i++) {
1384
1385                 if (isize <= lnb[i].lnb_file_offset)
1386                         /* If there's no more data, abort early.
1387                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1388                         break;
1389
1390                 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1391                         lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1392                 else
1393                         lnb[i].lnb_rc = lnb[i].lnb_len;
1394
1395                 /* Bypass disk read if fail_loc is set properly */
1396                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1397                         SetPageUptodate(lnb[i].lnb_page);
1398
1399                 if (PageUptodate(lnb[i].lnb_page)) {
1400                         cache_hits++;
1401                 } else {
1402                         cache_misses++;
1403                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1404                 }
1405
1406                 if (cache == 0)
1407                         generic_error_remove_page(inode->i_mapping,
1408                                                   lnb[i].lnb_page);
1409         }
1410         end = ktime_get();
1411         timediff = ktime_us_delta(end, start);
1412         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1413
1414         if (cache_hits != 0)
1415                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1416                                     cache_hits);
1417         if (cache_misses != 0)
1418                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1419                                     cache_misses);
1420         if (cache_hits + cache_misses != 0)
1421                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1422                                     cache_hits + cache_misses);
1423
1424         if (iobuf->dr_npages) {
1425                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1426                                                  iobuf->dr_npages,
1427                                                  iobuf->dr_blocks, 0);
1428                 rc = osd_do_bio(osd, inode, iobuf);
1429
1430                 /* IO stats will be done in osd_bufs_put() */
1431         }
1432
1433         RETURN(rc);
1434 }
1435
1436 /*
1437  * XXX: Another layering violation for now.
1438  *
1439  * We don't want to use ->f_op->read methods, because generic file write
1440  *
1441  *         - serializes on ->i_sem, and
1442  *
1443  *         - does a lot of extra work like balance_dirty_pages(),
1444  *
1445  * which doesn't work for globally shared files like /last_rcvd.
1446  */
1447 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1448 {
1449         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1450
1451         memcpy(buffer, (char *)ei->i_data, buflen);
1452
1453         return  buflen;
1454 }
1455
1456 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1457 {
1458         struct buffer_head *bh;
1459         unsigned long block;
1460         int osize;
1461         int blocksize;
1462         int csize;
1463         int boffs;
1464
1465         /* prevent reading after eof */
1466         spin_lock(&inode->i_lock);
1467         if (i_size_read(inode) < *offs + size) {
1468                 loff_t diff = i_size_read(inode) - *offs;
1469                 spin_unlock(&inode->i_lock);
1470                 if (diff < 0) {
1471                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1472                                i_size_read(inode), *offs);
1473                         return -EBADR;
1474                 } else if (diff == 0) {
1475                         return 0;
1476                 } else {
1477                         size = diff;
1478                 }
1479         } else {
1480                 spin_unlock(&inode->i_lock);
1481         }
1482
1483         blocksize = 1 << inode->i_blkbits;
1484         osize = size;
1485         while (size > 0) {
1486                 block = *offs >> inode->i_blkbits;
1487                 boffs = *offs & (blocksize - 1);
1488                 csize = min(blocksize - boffs, size);
1489                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1490                 if (IS_ERR(bh)) {
1491                         CERROR("%s: can't read %u@%llu on ino %lu: "
1492                                "rc = %ld\n", osd_ino2name(inode),
1493                                csize, *offs, inode->i_ino,
1494                                PTR_ERR(bh));
1495                         return PTR_ERR(bh);
1496                 }
1497
1498                 if (bh != NULL) {
1499                         memcpy(buf, bh->b_data + boffs, csize);
1500                         brelse(bh);
1501                 } else {
1502                         memset(buf, 0, csize);
1503                 }
1504
1505                 *offs += csize;
1506                 buf += csize;
1507                 size -= csize;
1508         }
1509         return osize;
1510 }
1511
1512 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1513                         struct lu_buf *buf, loff_t *pos)
1514 {
1515         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1516         int           rc;
1517
1518         /* Read small symlink from inode body as we need to maintain correct
1519          * on-disk symlinks for ldiskfs.
1520          */
1521         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1522             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1523                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1524         else
1525                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1526
1527         return rc;
1528 }
1529
1530 static inline int osd_extents_enabled(struct super_block *sb,
1531                                       struct inode *inode)
1532 {
1533         if (inode != NULL) {
1534                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1535                         return 1;
1536         } else if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
1537                                 LDISKFS_FEATURE_INCOMPAT_EXTENTS)) {
1538                 return 1;
1539         }
1540         return 0;
1541 }
1542
1543 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1544                            const loff_t size, const loff_t pos,
1545                            const int blocks)
1546 {
1547         int credits, bits, bs, i;
1548
1549         bits = sb->s_blocksize_bits;
1550         bs = 1 << bits;
1551
1552         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1553          * we do not expect blockmaps on the large files,
1554          * so let's shrink it to 2 levels (4GB files) */
1555
1556         /* this is default reservation: 2 levels */
1557         credits = (blocks + 2) * 3;
1558
1559         /* actual offset is unknown, hard to optimize */
1560         if (pos == -1)
1561                 return credits;
1562
1563         /* now check for few specific cases to optimize */
1564         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1565                 /* no indirects */
1566                 credits = blocks;
1567                 /* allocate if not allocated */
1568                 if (inode == NULL) {
1569                         credits += blocks * 2;
1570                         return credits;
1571                 }
1572                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1573                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1574                         if (LDISKFS_I(inode)->i_data[i] == 0)
1575                                 credits += 2;
1576                 }
1577         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1578                 /* single indirect */
1579                 credits = blocks * 3;
1580                 if (inode == NULL ||
1581                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1582                         credits += 3;
1583                 else
1584                         /* The indirect block may be modified. */
1585                         credits += 1;
1586         }
1587
1588         return credits;
1589 }
1590
1591 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1592                                  const struct lu_buf *buf, loff_t _pos,
1593                                  struct thandle *handle)
1594 {
1595         struct osd_object  *obj  = osd_dt_obj(dt);
1596         struct inode       *inode = obj->oo_inode;
1597         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1598         struct osd_thandle *oh;
1599         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1600         int                 bits, bs;
1601         int                 depth, size;
1602         loff_t              pos;
1603         ENTRY;
1604
1605         LASSERT(buf != NULL);
1606         LASSERT(handle != NULL);
1607
1608         oh = container_of0(handle, struct osd_thandle, ot_super);
1609         LASSERT(oh->ot_handle == NULL);
1610
1611         size = buf->lb_len;
1612         bits = sb->s_blocksize_bits;
1613         bs = 1 << bits;
1614
1615         if (_pos == -1) {
1616                 /* if this is an append, then we
1617                  * should expect cross-block record */
1618                 pos = 0;
1619         } else {
1620                 pos = _pos;
1621         }
1622
1623         /* blocks to modify */
1624         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1625         LASSERT(blocks > 0);
1626
1627         if (inode != NULL && _pos != -1) {
1628                 /* object size in blocks */
1629                 est = (i_size_read(inode) + bs - 1) >> bits;
1630                 allocated = inode->i_blocks >> (bits - 9);
1631                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1632                         /* looks like an overwrite, no need to modify tree */
1633                         credits = blocks;
1634                         /* no need to modify i_size */
1635                         goto out;
1636                 }
1637         }
1638
1639         if (osd_extents_enabled(sb, inode)) {
1640                 /*
1641                  * many concurrent threads may grow tree by the time
1642                  * our transaction starts. so, consider 2 is a min depth
1643                  * for every level we may need to allocate a new block
1644                  * and take some entries from the old one. so, 3 blocks
1645                  * to allocate (bitmap, gd, itself) + old block - 4 per
1646                  * level.
1647                  */
1648                 depth = inode != NULL ? ext_depth(inode) : 0;
1649                 depth = max(depth, 1) + 1;
1650                 credits = depth;
1651                 /* if not append, then split may need to modify
1652                  * existing blocks moving entries into the new ones */
1653                 if (_pos != -1)
1654                         credits += depth;
1655                 /* blocks to store data: bitmap,gd,itself */
1656                 credits += blocks * 3;
1657         } else {
1658                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1659         }
1660         /* if inode is created as part of the transaction,
1661          * then it's counted already by the creation method */
1662         if (inode != NULL)
1663                 credits++;
1664
1665 out:
1666
1667         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1668
1669         /* dt_declare_write() is usually called for system objects, such
1670          * as llog or last_rcvd files. We needn't enforce quota on those
1671          * objects, so always set the lqi_space as 0. */
1672         if (inode != NULL)
1673                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1674                                            i_gid_read(inode),
1675                                            i_projid_read(inode), 0,
1676                                            oh, obj, NULL, OSD_QID_BLK);
1677         RETURN(rc);
1678 }
1679
1680 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1681 {
1682         /* LU-2634: clear the extent format for fast symlink */
1683         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1684
1685         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1686         spin_lock(&inode->i_lock);
1687         LDISKFS_I(inode)->i_disksize = buflen;
1688         i_size_write(inode, buflen);
1689         spin_unlock(&inode->i_lock);
1690         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1691
1692         return 0;
1693 }
1694
1695 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1696                              int write_NUL, loff_t *offs, handle_t *handle)
1697 {
1698         struct buffer_head *bh        = NULL;
1699         loff_t              offset    = *offs;
1700         loff_t              new_size  = i_size_read(inode);
1701         unsigned long       block;
1702         int                 blocksize = 1 << inode->i_blkbits;
1703         int                 err = 0;
1704         int                 size;
1705         int                 boffs;
1706         int                 dirty_inode = 0;
1707
1708         if (write_NUL) {
1709                 /*
1710                  * long symlink write does not count the NUL terminator in
1711                  * bufsize, we write it, and the inode's file size does not
1712                  * count the NUL terminator as well.
1713                  */
1714                 ((char *)buf)[bufsize] = '\0';
1715                 ++bufsize;
1716         }
1717
1718         while (bufsize > 0) {
1719                 int credits = handle->h_buffer_credits;
1720
1721                 if (bh)
1722                         brelse(bh);
1723
1724                 block = offset >> inode->i_blkbits;
1725                 boffs = offset & (blocksize - 1);
1726                 size = min(blocksize - boffs, bufsize);
1727                 bh = __ldiskfs_bread(handle, inode, block, 1);
1728                 if (IS_ERR_OR_NULL(bh)) {
1729                         if (bh == NULL) {
1730                                 err = -EIO;
1731                         } else {
1732                                 err = PTR_ERR(bh);
1733                                 bh = NULL;
1734                         }
1735
1736                         CERROR("%s: error reading offset %llu (block %lu, "
1737                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
1738                                inode->i_sb->s_id, offset, block, bufsize, *offs,
1739                                credits, handle->h_buffer_credits, err);
1740                         break;
1741                 }
1742
1743                 err = ldiskfs_journal_get_write_access(handle, bh);
1744                 if (err) {
1745                         CERROR("journal_get_write_access() returned error %d\n",
1746                                err);
1747                         break;
1748                 }
1749                 LASSERTF(boffs + size <= bh->b_size,
1750                          "boffs %d size %d bh->b_size %lu\n",
1751                          boffs, size, (unsigned long)bh->b_size);
1752                 memcpy(bh->b_data + boffs, buf, size);
1753                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1754                 if (err)
1755                         break;
1756
1757                 if (offset + size > new_size)
1758                         new_size = offset + size;
1759                 offset += size;
1760                 bufsize -= size;
1761                 buf += size;
1762         }
1763         if (bh)
1764                 brelse(bh);
1765
1766         if (write_NUL)
1767                 --new_size;
1768         /* correct in-core and on-disk sizes */
1769         if (new_size > i_size_read(inode)) {
1770                 spin_lock(&inode->i_lock);
1771                 if (new_size > i_size_read(inode))
1772                         i_size_write(inode, new_size);
1773                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1774                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1775                         dirty_inode = 1;
1776                 }
1777                 spin_unlock(&inode->i_lock);
1778                 if (dirty_inode)
1779                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1780         }
1781
1782         if (err == 0)
1783                 *offs = offset;
1784         return err;
1785 }
1786
1787 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1788                          const struct lu_buf *buf, loff_t *pos,
1789                          struct thandle *handle, int ignore_quota)
1790 {
1791         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1792         struct osd_thandle      *oh;
1793         ssize_t                 result;
1794         int                     is_link;
1795
1796         LASSERT(dt_object_exists(dt));
1797
1798         LASSERT(handle != NULL);
1799         LASSERT(inode != NULL);
1800         ll_vfs_dq_init(inode);
1801
1802         /* XXX: don't check: one declared chunk can be used many times */
1803         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1804
1805         oh = container_of(handle, struct osd_thandle, ot_super);
1806         LASSERT(oh->ot_handle->h_transaction != NULL);
1807         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1808
1809         /* Write small symlink to inode body as we need to maintain correct
1810          * on-disk symlinks for ldiskfs.
1811          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1812          * does not count it in.
1813          */
1814         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1815         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1816                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1817         else
1818                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1819                                                   buf->lb_len, is_link, pos,
1820                                                   oh->ot_handle);
1821         if (result == 0)
1822                 result = buf->lb_len;
1823
1824         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1825
1826         return result;
1827 }
1828
1829 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1830                              __u64 start, __u64 end, struct thandle *th)
1831 {
1832         struct osd_thandle *oh;
1833         struct inode       *inode;
1834         int                 rc;
1835         ENTRY;
1836
1837         LASSERT(th);
1838         oh = container_of(th, struct osd_thandle, ot_super);
1839
1840         /*
1841          * we don't need to reserve credits for whole truncate
1842          * it's not possible as truncate may need to free too many
1843          * blocks and that won't fit a single transaction. instead
1844          * we reserve credits to change i_size and put inode onto
1845          * orphan list. if needed truncate will extend or restart
1846          * transaction
1847          */
1848         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1849                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1850
1851         inode = osd_dt_obj(dt)->oo_inode;
1852         LASSERT(inode);
1853
1854         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1855                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1856                                    NULL, OSD_QID_BLK);
1857         RETURN(rc);
1858 }
1859
1860 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1861                      __u64 start, __u64 end, struct thandle *th)
1862 {
1863         struct osd_thandle *oh;
1864         struct osd_object  *obj = osd_dt_obj(dt);
1865         struct inode       *inode = obj->oo_inode;
1866         handle_t           *h;
1867         tid_t               tid;
1868         int                rc = 0, rc2 = 0;
1869         ENTRY;
1870
1871         LASSERT(end == OBD_OBJECT_EOF);
1872         LASSERT(dt_object_exists(dt));
1873         LASSERT(osd_invariant(obj));
1874         LASSERT(inode != NULL);
1875         ll_vfs_dq_init(inode);
1876
1877         LASSERT(th);
1878         oh = container_of(th, struct osd_thandle, ot_super);
1879         LASSERT(oh->ot_handle->h_transaction != NULL);
1880
1881         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1882
1883         tid = oh->ot_handle->h_transaction->t_tid;
1884
1885         spin_lock(&inode->i_lock);
1886         i_size_write(inode, start);
1887         spin_unlock(&inode->i_lock);
1888         ll_truncate_pagecache(inode, start);
1889 #ifdef HAVE_INODEOPS_TRUNCATE
1890         if (inode->i_op->truncate) {
1891                 inode->i_op->truncate(inode);
1892         } else
1893 #endif
1894                 ldiskfs_truncate(inode);
1895
1896         /*
1897          * For a partial-page truncate, flush the page to disk immediately to
1898          * avoid data corruption during direct disk write.  b=17397
1899          */
1900         if ((start & ~PAGE_MASK) != 0)
1901                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1902
1903         h = journal_current_handle();
1904         LASSERT(h != NULL);
1905         LASSERT(h == oh->ot_handle);
1906
1907         /* do not check credits with osd_trans_exec_check() as the truncate
1908          * can restart the transaction internally and we restart the
1909          * transaction in this case */
1910
1911         if (tid != h->h_transaction->t_tid) {
1912                 int credits = oh->ot_credits;
1913                 /*
1914                  * transaction has changed during truncate
1915                  * we need to restart the handle with our credits
1916                  */
1917                 if (h->h_buffer_credits < credits) {
1918                         if (ldiskfs_journal_extend(h, credits))
1919                                 rc2 = ldiskfs_journal_restart(h, credits);
1920                 }
1921         }
1922
1923         RETURN(rc == 0 ? rc2 : rc);
1924 }
1925
1926 static int fiemap_check_ranges(struct inode *inode,
1927                                u64 start, u64 len, u64 *new_len)
1928 {
1929         loff_t maxbytes;
1930
1931         *new_len = len;
1932
1933         if (len == 0)
1934                 return -EINVAL;
1935
1936         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1937                 maxbytes = inode->i_sb->s_maxbytes;
1938         else
1939                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1940
1941         if (start > maxbytes)
1942                 return -EFBIG;
1943
1944         /*
1945          * Shrink request scope to what the fs can actually handle.
1946          */
1947         if (len > maxbytes || (maxbytes - len) < start)
1948                 *new_len = maxbytes - start;
1949
1950         return 0;
1951 }
1952
1953 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1954 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
1955
1956 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1957                           struct fiemap *fm)
1958 {
1959         struct fiemap_extent_info fieinfo = {0, };
1960         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1961         u64 len;
1962         int rc;
1963
1964
1965         LASSERT(inode);
1966         if (inode->i_op->fiemap == NULL)
1967                 return -EOPNOTSUPP;
1968
1969         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
1970                 return -EINVAL;
1971
1972         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
1973         if (rc)
1974                 return rc;
1975
1976         fieinfo.fi_flags = fm->fm_flags;
1977         fieinfo.fi_extents_max = fm->fm_extent_count;
1978         fieinfo.fi_extents_start = fm->fm_extents;
1979
1980         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
1981                 filemap_write_and_wait(inode->i_mapping);
1982
1983         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
1984         fm->fm_flags = fieinfo.fi_flags;
1985         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
1986
1987         return rc;
1988 }
1989
1990 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
1991                        __u64 start, __u64 end, enum lu_ladvise_type advice)
1992 {
1993         int              rc = 0;
1994         struct inode    *inode = osd_dt_obj(dt)->oo_inode;
1995         ENTRY;
1996
1997         switch (advice) {
1998         case LU_LADVISE_DONTNEED:
1999                 if (end == 0)
2000                         break;
2001                 invalidate_mapping_pages(inode->i_mapping,
2002                                          start >> PAGE_CACHE_SHIFT,
2003                                          (end - 1) >> PAGE_CACHE_SHIFT);
2004                 break;
2005         default:
2006                 rc = -ENOTSUPP;
2007                 break;
2008         }
2009
2010         RETURN(rc);
2011 }
2012
2013 /*
2014  * in some cases we may need declare methods for objects being created
2015  * e.g., when we create symlink
2016  */
2017 const struct dt_body_operations osd_body_ops_new = {
2018         .dbo_declare_write = osd_declare_write,
2019 };
2020
2021 const struct dt_body_operations osd_body_ops = {
2022         .dbo_read                       = osd_read,
2023         .dbo_declare_write              = osd_declare_write,
2024         .dbo_write                      = osd_write,
2025         .dbo_bufs_get                   = osd_bufs_get,
2026         .dbo_bufs_put                   = osd_bufs_put,
2027         .dbo_write_prep                 = osd_write_prep,
2028         .dbo_declare_write_commit       = osd_declare_write_commit,
2029         .dbo_write_commit               = osd_write_commit,
2030         .dbo_read_prep                  = osd_read_prep,
2031         .dbo_declare_punch              = osd_declare_punch,
2032         .dbo_punch                      = osd_punch,
2033         .dbo_fiemap_get                 = osd_fiemap_get,
2034         .dbo_ladvise                    = osd_ladvise,
2035 };