lustre/osd-ldiskfs/osd_io.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  *
  38  * lustre/osd/osd_io.c
  39  *
  40  * body operations
  41  *
  42  * Author: Nikita Danilov <nikita@clusterfs.com>
  43  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  44  *
  45  */
  46
  47 /* LUSTRE_VERSION_CODE */
  48 #include <lustre_ver.h>
  49 /* prerequisite for linux/xattr.h */
  50 #include <linux/types.h>
  51 /* prerequisite for linux/xattr.h */
  52 #include <linux/fs.h>
  53
  54 /* ext_depth() */
  55 #include <ldiskfs/ldiskfs.h>
  56 #include <ldiskfs/ldiskfs_jbd2.h>
  57 #include <ldiskfs/ldiskfs_extents.h>
  58
  59 /*
  60  * struct OBD_{ALLOC,FREE}*()
  61  * OBD_FAIL_CHECK
  62  */
  63 #include <obd_support.h>
  64
  65 #include "osd_internal.h"
  66
  67 #ifndef HAVE_PAGE_CONSTANT
  68 #define mapping_cap_page_constant_write(mapping) 0
  69 #define SetPageConstant(page) do {} while (0)
  70 #define ClearPageConstant(page) do {} while (0)
  71 #endif
  72
  73 #ifndef HAS_GENERIC_ERROR_REMOVE_PAGE
  74 int generic_error_remove_page(struct address_space *mapping, struct page *page)
  75 {
  76         if (mapping == NULL)
  77                 return -EINVAL;
  78
  79         if (mapping != page->mapping)
  80                 return -EIO;
  81         /*
  82          * Only punch for normal data pages for now.
  83          * Handling other types like directories would need more auditing.
  84          */
  85         if (!S_ISREG(mapping->host->i_mode))
  86                 return -EIO;
  87
  88         if (page_mapped(page)) {
  89                 unmap_mapping_range(mapping,
  90                                     (loff_t)page->index << PAGE_CACHE_SHIFT,
  91                                     PAGE_CACHE_SIZE, 0);
  92         }
  93         truncate_complete_page(mapping, page);
  94         return 0;
  95 }
  96 #endif
  97
  98 static void osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,int rw)
  99 {
 100         cfs_waitq_init(&iobuf->dr_wait);
 101         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 102         iobuf->dr_max_pages = PTLRPC_MAX_BRW_PAGES;
 103         iobuf->dr_npages = 0;
 104         iobuf->dr_error = 0;
 105         iobuf->dr_dev = d;
 106         iobuf->dr_frags = 0;
 107         iobuf->dr_elapsed = 0;
 108         /* must be counted before, so assert */
 109         LASSERT(iobuf->dr_elapsed_valid == 0);
 110         iobuf->dr_rw = rw;
 111 }
 112
 113 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
 114 {
 115         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 116         iobuf->dr_pages[iobuf->dr_npages++] = page;
 117 }
 118
 119 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 120 {
 121         int rw = iobuf->dr_rw;
 122
 123         if (iobuf->dr_elapsed_valid) {
 124                 iobuf->dr_elapsed_valid = 0;
 125                 LASSERT(iobuf->dr_dev == d);
 126                 LASSERT(iobuf->dr_frags > 0);
 127                 lprocfs_oh_tally(&d->od_brw_stats.
 128                                  hist[BRW_R_DIO_FRAGS+rw],
 129                                  iobuf->dr_frags);
 130                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 131                                       iobuf->dr_elapsed);
 132         }
 133 }
 134
 135 #ifdef HAVE_BIO_ENDIO_2ARG
 136 #define DIO_RETURN(a)
 137 static void dio_complete_routine(struct bio *bio, int error)
 138 #else
 139 #define DIO_RETURN(a)   return(a)
 140 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 141 #endif
 142 {
 143         struct osd_iobuf *iobuf = bio->bi_private;
 144         struct bio_vec *bvl;
 145         int i;
 146
 147         /* CAVEAT EMPTOR: possibly in IRQ context
 148          * DO NOT record procfs stats here!!! */
 149
 150         if (unlikely(iobuf == NULL)) {
 151                 CERROR("***** bio->bi_private is NULL!  This should never "
 152                        "happen.  Normally, I would crash here, but instead I "
 153                        "will dump the bio contents to the console.  Please "
 154                        "report this to <http://jira.whamcloud.com/> , along "
 155                        "with any interesting messages leading up to this point "
 156                        "(like SCSI errors, perhaps).  Because bi_private is "
 157                        "NULL, I can't wake up the thread that initiated this "
 158                        "IO - you will probably have to reboot this node.\n");
 159                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 160                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 161                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 162                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 163                        bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
 164                        bio->bi_private);
 165                 DIO_RETURN(0);
 166         }
 167
 168         /* the check is outside of the cycle for performance reason -bzzz */
 169         if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
 170                 bio_for_each_segment(bvl, bio, i) {
 171                         if (likely(error == 0))
 172                                 SetPageUptodate(bvl->bv_page);
 173                         LASSERT(PageLocked(bvl->bv_page));
 174                         ClearPageConstant(bvl->bv_page);
 175                 }
 176                 cfs_atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 177         } else {
 178                 struct page *p = iobuf->dr_pages[0];
 179                 if (p->mapping) {
 180                         if (mapping_cap_page_constant_write(p->mapping)) {
 181                                 bio_for_each_segment(bvl, bio, i) {
 182                                         ClearPageConstant(bvl->bv_page);
 183                                 }
 184                         }
 185                 }
 186                 cfs_atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 187         }
 188
 189         /* any real error is good enough -bzzz */
 190         if (error != 0 && iobuf->dr_error == 0)
 191                 iobuf->dr_error = error;
 192
 193         if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs)) {
 194                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
 195                 iobuf->dr_elapsed_valid = 1;
 196                 cfs_waitq_signal(&iobuf->dr_wait);
 197         }
 198
 199         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 200          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 201          * mempool when serious on-disk fragmentation was encountered,
 202          * deadlocking the OST.  The bios are now released as soon as complete
 203          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 204         bio_put(bio);
 205         DIO_RETURN(0);
 206 }
 207
 208 static void record_start_io(struct osd_iobuf *iobuf, int size)
 209 {
 210         struct osd_device    *osd = iobuf->dr_dev;
 211         struct obd_histogram *h = osd->od_brw_stats.hist;
 212
 213         iobuf->dr_frags++;
 214         cfs_atomic_inc(&iobuf->dr_numreqs);
 215
 216         if (iobuf->dr_rw == 0) {
 217                 cfs_atomic_inc(&osd->od_r_in_flight);
 218                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 219                                  cfs_atomic_read(&osd->od_r_in_flight));
 220                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 221         } else if (iobuf->dr_rw == 1) {
 222                 cfs_atomic_inc(&osd->od_w_in_flight);
 223                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 224                                  cfs_atomic_read(&osd->od_w_in_flight));
 225                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 226         } else {
 227                 LBUG();
 228         }
 229 }
 230
 231 static void osd_submit_bio(int rw, struct bio *bio)
 232 {
 233         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 234         if (rw == 0)
 235                 submit_bio(READ, bio);
 236         else
 237                 submit_bio(WRITE, bio);
 238 }
 239
 240 static int can_be_merged(struct bio *bio, sector_t sector)
 241 {
 242         unsigned int size;
 243
 244         if (!bio)
 245                 return 0;
 246
 247         size = bio->bi_size >> 9;
 248         return bio->bi_sector + size == sector ? 1 : 0;
 249 }
 250
 251 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 252                       struct osd_iobuf *iobuf)
 253 {
 254         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 255         struct page  **pages = iobuf->dr_pages;
 256         int            npages = iobuf->dr_npages;
 257         unsigned long *blocks = iobuf->dr_blocks;
 258         int            total_blocks = npages * blocks_per_page;
 259         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 260         unsigned int   blocksize = inode->i_sb->s_blocksize;
 261         struct bio    *bio = NULL;
 262         struct page   *page;
 263         unsigned int   page_offset;
 264         sector_t       sector;
 265         int            nblocks;
 266         int            block_idx;
 267         int            page_idx;
 268         int            i;
 269         int            rc = 0;
 270         ENTRY;
 271
 272         LASSERT(iobuf->dr_npages == npages);
 273
 274         osd_brw_stats_update(osd, iobuf);
 275         iobuf->dr_start_time = cfs_time_current();
 276
 277         for (page_idx = 0, block_idx = 0;
 278              page_idx < npages;
 279              page_idx++, block_idx += blocks_per_page) {
 280
 281                 page = pages[page_idx];
 282                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 283
 284                 for (i = 0, page_offset = 0;
 285                      i < blocks_per_page;
 286                      i += nblocks, page_offset += blocksize * nblocks) {
 287
 288                         nblocks = 1;
 289
 290                         if (blocks[block_idx + i] == 0) {  /* hole */
 291                                 LASSERTF(iobuf->dr_rw == 0,
 292                                          "page_idx %u, block_idx %u, i %u\n",
 293                                          page_idx, block_idx, i);
 294                                 memset(kmap(page) + page_offset, 0, blocksize);
 295                                 kunmap(page);
 296                                 continue;
 297                         }
 298
 299                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 300
 301                         /* Additional contiguous file blocks? */
 302                         while (i + nblocks < blocks_per_page &&
 303                                (sector + (nblocks << sector_bits)) ==
 304                                ((sector_t)blocks[block_idx + i + nblocks] <<
 305                                 sector_bits))
 306                                 nblocks++;
 307
 308                         /* I only set the page to be constant only if it
 309                          * is mapped to a contiguous underlying disk block(s).
 310                          * It will then make sure the corresponding device
 311                          * cache of raid5 will be overwritten by this page.
 312                          * - jay */
 313                         if (iobuf->dr_rw && (nblocks == blocks_per_page) &&
 314                             mapping_cap_page_constant_write(inode->i_mapping))
 315                                 SetPageConstant(page);
 316
 317                         if (bio != NULL &&
 318                             can_be_merged(bio, sector) &&
 319                             bio_add_page(bio, page,
 320                                          blocksize * nblocks, page_offset) != 0)
 321                                 continue;       /* added this frag OK */
 322
 323                         if (bio != NULL) {
 324                                 struct request_queue *q =
 325                                         bdev_get_queue(bio->bi_bdev);
 326
 327                                 /* Dang! I have to fragment this I/O */
 328                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 329                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 330                                        bio->bi_size,
 331                                        bio->bi_vcnt, bio->bi_max_vecs,
 332                                        bio->bi_size >> 9, queue_max_sectors(q),
 333                                        bio_phys_segments(q, bio),
 334                                        queue_max_phys_segments(q),
 335                                        bio_hw_segments(q, bio),
 336                                        queue_max_hw_segments(q));
 337
 338                                 record_start_io(iobuf, bio->bi_size);
 339                                 osd_submit_bio(iobuf->dr_rw, bio);
 340                         }
 341
 342                         /* allocate new bio, limited by max BIO size, b=9945 */
 343                         bio = bio_alloc(GFP_NOIO, max(BIO_MAX_PAGES,
 344                                                       (npages - page_idx) *
 345                                                       blocks_per_page));
 346                         if (bio == NULL) {
 347                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 348                                        (npages - page_idx), blocks_per_page,
 349                                        (npages - page_idx) * blocks_per_page);
 350                                 rc = -ENOMEM;
 351                                 goto out;
 352                         }
 353
 354                         bio->bi_bdev = inode->i_sb->s_bdev;
 355                         bio->bi_sector = sector;
 356                         bio->bi_end_io = dio_complete_routine;
 357                         bio->bi_private = iobuf;
 358
 359                         rc = bio_add_page(bio, page,
 360                                           blocksize * nblocks, page_offset);
 361                         LASSERT(rc != 0);
 362                 }
 363         }
 364
 365         if (bio != NULL) {
 366                 record_start_io(iobuf, bio->bi_size);
 367                 osd_submit_bio(iobuf->dr_rw, bio);
 368                 rc = 0;
 369         }
 370
 371  out:
 372         /* in order to achieve better IO throughput, we don't wait for writes
 373          * completion here. instead we proceed with transaction commit in
 374          * parallel and wait for IO completion once transaction is stopped
 375          * see osd_trans_stop() for more details -bzzz */
 376         if (iobuf->dr_rw == 0) {
 377                 cfs_wait_event(iobuf->dr_wait,
 378                                cfs_atomic_read(&iobuf->dr_numreqs) == 0);
 379         }
 380
 381         if (rc == 0)
 382                 rc = iobuf->dr_error;
 383         RETURN(rc);
 384 }
 385
 386 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 387                                    struct niobuf_local *lnb)
 388 {
 389         ENTRY;
 390
 391         *nrpages = 0;
 392
 393         while (len > 0) {
 394                 int poff = offset & (CFS_PAGE_SIZE - 1);
 395                 int plen = CFS_PAGE_SIZE - poff;
 396
 397                 if (plen > len)
 398                         plen = len;
 399                 lnb->offset = offset;
 400                 /* lnb->lnb_page_offset = poff; */
 401                 lnb->len = plen;
 402                 /* lb->flags = rnb->flags; */
 403                 lnb->flags = 0;
 404                 lnb->page = NULL;
 405                 lnb->rc = 0;
 406
 407                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 408                          (long long) len);
 409                 offset += plen;
 410                 len -= plen;
 411                 lnb++;
 412                 (*nrpages)++;
 413         }
 414
 415         RETURN(0);
 416 }
 417
 418 struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
 419 {
 420         struct inode      *inode = osd_dt_obj(dt)->oo_inode;
 421         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 422         struct page       *page;
 423
 424         LASSERT(inode);
 425
 426         page = find_or_create_page(inode->i_mapping, offset >> CFS_PAGE_SHIFT,
 427                                    GFP_NOFS | __GFP_HIGHMEM);
 428         if (unlikely(page == NULL))
 429                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 430
 431         return page;
 432 }
 433
 434 /*
 435  * there are following "locks":
 436  * journal_start
 437  * i_alloc_sem
 438  * i_mutex
 439  * page lock
 440
 441  * osd write path
 442     * lock page(s)
 443     * journal_start
 444     * truncate_sem
 445
 446  * ext4 vmtruncate:
 447     * lock pages, unlock
 448     * journal_start
 449     * lock partial page
 450     * i_data_sem
 451
 452 */
 453 int osd_bufs_get(const struct lu_env *env, struct dt_object *d, loff_t pos,
 454                  ssize_t len, struct niobuf_local *lnb, int rw,
 455                  struct lustre_capa *capa)
 456 {
 457         struct osd_object   *obj    = osd_dt_obj(d);
 458         int npages, i, rc = 0;
 459
 460         LASSERT(obj->oo_inode);
 461
 462         osd_map_remote_to_local(pos, len, &npages, lnb);
 463
 464         for (i = 0; i < npages; i++, lnb++) {
 465
 466                 /* We still set up for ungranted pages so that granted pages
 467                  * can be written to disk as they were promised, and portals
 468                  * needs to keep the pages all aligned properly. */
 469                 lnb->dentry = (void *) obj;
 470
 471                 lnb->page = osd_get_page(d, lnb->offset, rw);
 472                 if (lnb->page == NULL)
 473                         GOTO(cleanup, rc = -ENOMEM);
 474
 475                 /* DLM locking protects us from write and truncate competing
 476                  * for same region, but truncate can leave dirty page in the
 477                  * cache. it's possible the writeout on a such a page is in
 478                  * progress when we access it. it's also possible that during
 479                  * this writeout we put new (partial) data, but then won't
 480                  * be able to proceed in filter_commitrw_write(). thus let's
 481                  * just wait for writeout completion, should be rare enough.
 482                  * -bzzz */
 483                 wait_on_page_writeback(lnb->page);
 484                 BUG_ON(PageWriteback(lnb->page));
 485
 486                 lu_object_get(&d->do_lu);
 487         }
 488         rc = i;
 489
 490 cleanup:
 491         RETURN(rc);
 492 }
 493
 494 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 495                         struct niobuf_local *lnb, int npages)
 496 {
 497         struct osd_thread_info *oti = osd_oti_get(env);
 498         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 499         struct osd_device      *d = osd_obj2dev(osd_dt_obj(dt));
 500         int                     i;
 501
 502         /* to do IO stats, notice we do this here because
 503          * osd_do_bio() doesn't wait for write to complete */
 504         osd_fini_iobuf(d, iobuf);
 505
 506         for (i = 0; i < npages; i++) {
 507                 if (lnb[i].page == NULL)
 508                         continue;
 509                 LASSERT(PageLocked(lnb[i].page));
 510                 unlock_page(lnb[i].page);
 511                 page_cache_release(lnb[i].page);
 512                 lu_object_put(env, &dt->do_lu);
 513                 lnb[i].page = NULL;
 514         }
 515         RETURN(0);
 516 }
 517
 518 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 519                           struct niobuf_local *lnb, int npages)
 520 {
 521         struct osd_thread_info *oti   = osd_oti_get(env);
 522         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 523         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 524         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 525         struct timeval          start;
 526         struct timeval          end;
 527         unsigned long           timediff;
 528         ssize_t                 isize;
 529         __s64                   maxidx;
 530         int                     rc = 0;
 531         int                     i;
 532         int                     cache = 0;
 533
 534         LASSERT(inode);
 535
 536         osd_init_iobuf(osd, iobuf, 0);
 537
 538         isize = i_size_read(inode);
 539         maxidx = ((isize + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT) - 1;
 540
 541         if (osd->od_writethrough_cache)
 542                 cache = 1;
 543         if (isize > osd->od_readcache_max_filesize)
 544                 cache = 0;
 545
 546         cfs_gettimeofday(&start);
 547         for (i = 0; i < npages; i++) {
 548
 549                 if (cache == 0)
 550                         generic_error_remove_page(inode->i_mapping,
 551                                                   lnb[i].page);
 552
 553                 /*
 554                  * till commit the content of the page is undefined
 555                  * we'll set it uptodate once bulk is done. otherwise
 556                  * subsequent reads can access non-stable data
 557                  */
 558                 ClearPageUptodate(lnb[i].page);
 559
 560                 if (lnb[i].len == CFS_PAGE_SIZE)
 561                         continue;
 562
 563                 if (maxidx >= lnb[i].page->index) {
 564                         osd_iobuf_add_page(iobuf, lnb[i].page);
 565                 } else {
 566                         long off;
 567                         char *p = kmap(lnb[i].page);
 568
 569                         off = lnb[i].offset;
 570                         if (off)
 571                                 memset(p, 0, off);
 572                         off = lnb[i].offset + lnb[i].len;
 573                         off &= ~CFS_PAGE_MASK;
 574                         if (off)
 575                                 memset(p + off, 0, CFS_PAGE_SIZE - off);
 576                         kunmap(lnb[i].page);
 577                 }
 578         }
 579         cfs_gettimeofday(&end);
 580         timediff = cfs_timeval_sub(&end, &start, NULL);
 581         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 582
 583         if (iobuf->dr_npages) {
 584                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 585                                                        iobuf->dr_npages,
 586                                                        iobuf->dr_blocks,
 587                                                        oti->oti_created,
 588                                                        0, NULL);
 589                 if (likely(rc == 0)) {
 590                         rc = osd_do_bio(osd, inode, iobuf);
 591                         /* do IO stats for preparation reads */
 592                         osd_fini_iobuf(osd, iobuf);
 593                 }
 594         }
 595         RETURN(rc);
 596 }
 597
 598 static int osd_declare_write_commit(const struct lu_env *env,
 599                                     struct dt_object *dt,
 600                                     struct niobuf_local *lnb, int npages,
 601                                     struct thandle *handle)
 602 {
 603         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 604         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
 605         struct osd_thandle      *oh;
 606         int                      extents = 1;
 607         int                      depth;
 608         int                      i;
 609         int                      newblocks;
 610         int                      old;
 611
 612         LASSERT(handle != NULL);
 613         oh = container_of0(handle, struct osd_thandle, ot_super);
 614         LASSERT(oh->ot_handle == NULL);
 615
 616         old = oh->ot_credits;
 617         newblocks = npages;
 618
 619         /* calculate number of extents (probably better to pass nb) */
 620         for (i = 1; i < npages; i++)
 621                 if (lnb[i].offset !=
 622                     lnb[i - 1].offset + lnb[i - 1].len)
 623                         extents++;
 624
 625         /*
 626          * each extent can go into new leaf causing a split
 627          * 5 is max tree depth: inode + 4 index blocks
 628          * with blockmaps, depth is 3 at most
 629          */
 630         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
 631                 /*
 632                  * many concurrent threads may grow tree by the time
 633                  * our transaction starts. so, consider 2 is a min depth
 634                  */
 635                 depth = ext_depth(inode);
 636                 depth = max(depth, 1) + 1;
 637                 newblocks += depth;
 638                 oh->ot_credits++; /* inode */
 639                 oh->ot_credits += depth * 2 * extents;
 640         } else {
 641                 depth = 3;
 642                 newblocks += depth;
 643                 oh->ot_credits++; /* inode */
 644                 oh->ot_credits += depth * extents;
 645         }
 646
 647         /* each new block can go in different group (bitmap + gd) */
 648
 649         /* we can't dirty more bitmap blocks than exist */
 650         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
 651                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
 652         else
 653                 oh->ot_credits += newblocks;
 654
 655         /* we can't dirty more gd blocks than exist */
 656         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
 657                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
 658         else
 659                 oh->ot_credits += newblocks;
 660
 661         RETURN(0);
 662 }
 663
 664 /* Check if a block is allocated or not */
 665 static int osd_is_mapped(struct inode *inode, obd_size offset)
 666 {
 667         sector_t (*fs_bmap)(struct address_space *, sector_t);
 668
 669         fs_bmap = inode->i_mapping->a_ops->bmap;
 670
 671         /* We can't know if we are overwriting or not */
 672         if (fs_bmap == NULL)
 673                 return 0;
 674
 675         if (fs_bmap(inode->i_mapping, offset >> inode->i_blkbits) == 0)
 676                 return 0;
 677
 678         return 1;
 679 }
 680
 681 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
 682                             struct niobuf_local *lnb, int npages,
 683                             struct thandle *thandle)
 684 {
 685         struct osd_thread_info *oti = osd_oti_get(env);
 686         struct osd_iobuf *iobuf = &oti->oti_iobuf;
 687         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 688         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
 689         loff_t isize;
 690         int rc = 0, i;
 691
 692         LASSERT(inode);
 693
 694         osd_init_iobuf(osd, iobuf, 1);
 695         isize = i_size_read(inode);
 696
 697         for (i = 0; i < npages; i++) {
 698                 if (lnb[i].rc == -ENOSPC &&
 699                     osd_is_mapped(inode, lnb[i].offset)) {
 700                         /* Allow the write to proceed if overwriting an
 701                          * existing block */
 702                         lnb[i].rc = 0;
 703                 }
 704
 705                 if (lnb[i].rc) { /* ENOSPC, network RPC error, etc. */
 706                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
 707                                lnb[i].rc);
 708                         LASSERT(lnb[i].page);
 709                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 710                         continue;
 711                 }
 712
 713                 LASSERT(PageLocked(lnb[i].page));
 714                 LASSERT(!PageWriteback(lnb[i].page));
 715
 716                 if (lnb[i].offset + lnb[i].len > isize)
 717                         isize = lnb[i].offset + lnb[i].len;
 718
 719                 /*
 720                  * Since write and truncate are serialized by oo_sem, even
 721                  * partial-page truncate should not leave dirty pages in the
 722                  * page cache.
 723                  */
 724                 LASSERT(!PageDirty(lnb[i].page));
 725
 726                 SetPageUptodate(lnb[i].page);
 727
 728                 osd_iobuf_add_page(iobuf, lnb[i].page);
 729         }
 730
 731         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
 732                 rc = -ENOSPC;
 733         } else if (iobuf->dr_npages > 0) {
 734                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 735                                                        iobuf->dr_npages,
 736                                                        iobuf->dr_blocks,
 737                                                        oti->oti_created,
 738                                                        1, NULL);
 739         } else {
 740                 /* no pages to write, no transno is needed */
 741                 thandle->th_local = 1;
 742         }
 743
 744         if (likely(rc == 0)) {
 745                 if (isize > i_size_read(inode)) {
 746                         i_size_write(inode, isize);
 747                         LDISKFS_I(inode)->i_disksize = isize;
 748                         inode->i_sb->s_op->dirty_inode(inode);
 749                 }
 750
 751                 rc = osd_do_bio(osd, inode, iobuf);
 752                 /* we don't do stats here as in read path because
 753                  * write is async: we'll do this in osd_put_bufs() */
 754         }
 755
 756         if (unlikely(rc != 0)) {
 757                 /* if write fails, we should drop pages from the cache */
 758                 for (i = 0; i < npages; i++) {
 759                         if (lnb[i].page == NULL)
 760                                 continue;
 761                         LASSERT(PageLocked(lnb[i].page));
 762                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 763                 }
 764         }
 765
 766         RETURN(rc);
 767 }
 768
 769 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
 770                          struct niobuf_local *lnb, int npages)
 771 {
 772         struct osd_thread_info *oti = osd_oti_get(env);
 773         struct osd_iobuf *iobuf = &oti->oti_iobuf;
 774         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 775         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
 776         struct timeval start, end;
 777         unsigned long timediff;
 778         int rc = 0, i, m = 0, cache = 0;
 779
 780         LASSERT(inode);
 781
 782         osd_init_iobuf(osd, iobuf, 0);
 783
 784         if (osd->od_read_cache)
 785                 cache = 1;
 786         if (i_size_read(inode) > osd->od_readcache_max_filesize)
 787                 cache = 0;
 788
 789         cfs_gettimeofday(&start);
 790         for (i = 0; i < npages; i++) {
 791
 792                 if (i_size_read(inode) <= lnb[i].offset)
 793                         /* If there's no more data, abort early.
 794                          * lnb->rc == 0, so it's easy to detect later. */
 795                         break;
 796
 797                 if (i_size_read(inode) <
 798                     lnb[i].offset + lnb[i].len - 1)
 799                         lnb[i].rc = i_size_read(inode) - lnb[i].offset;
 800                 else
 801                         lnb[i].rc = lnb[i].len;
 802                 m += lnb[i].len;
 803
 804                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS, 1);
 805                 if (PageUptodate(lnb[i].page)) {
 806                         lprocfs_counter_add(osd->od_stats,
 807                                             LPROC_OSD_CACHE_HIT, 1);
 808                 } else {
 809                         lprocfs_counter_add(osd->od_stats,
 810                                             LPROC_OSD_CACHE_MISS, 1);
 811                         osd_iobuf_add_page(iobuf, lnb[i].page);
 812                 }
 813                 if (cache == 0)
 814                         generic_error_remove_page(inode->i_mapping,lnb[i].page);
 815         }
 816         cfs_gettimeofday(&end);
 817         timediff = cfs_timeval_sub(&end, &start, NULL);
 818         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
 819
 820         if (iobuf->dr_npages) {
 821                 rc = osd->od_fsops->fs_map_inode_pages(inode, iobuf->dr_pages,
 822                                                        iobuf->dr_npages,
 823                                                        iobuf->dr_blocks,
 824                                                        oti->oti_created,
 825                                                        0, NULL);
 826                 rc = osd_do_bio(osd, inode, iobuf);
 827
 828                 /* IO stats will be done in osd_bufs_put() */
 829         }
 830
 831         RETURN(rc);
 832 }
 833
 834 /*
 835  * XXX: Another layering violation for now.
 836  *
 837  * We don't want to use ->f_op->read methods, because generic file write
 838  *
 839  *         - serializes on ->i_sem, and
 840  *
 841  *         - does a lot of extra work like balance_dirty_pages(),
 842  *
 843  * which doesn't work for globally shared files like /last_rcvd.
 844  */
 845 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
 846 {
 847         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 848
 849         memcpy(buffer, (char *)ei->i_data, buflen);
 850
 851         return  buflen;
 852 }
 853
 854 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
 855 {
 856         struct buffer_head *bh;
 857         unsigned long block;
 858         int osize = size;
 859         int blocksize;
 860         int csize;
 861         int boffs;
 862         int err;
 863
 864         /* prevent reading after eof */
 865         cfs_spin_lock(&inode->i_lock);
 866         if (i_size_read(inode) < *offs + size) {
 867                 size = i_size_read(inode) - *offs;
 868                 cfs_spin_unlock(&inode->i_lock);
 869                 if (size < 0) {
 870                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
 871                                i_size_read(inode), *offs);
 872                         return -EBADR;
 873                 } else if (size == 0) {
 874                         return 0;
 875                 }
 876         } else {
 877                 cfs_spin_unlock(&inode->i_lock);
 878         }
 879
 880         blocksize = 1 << inode->i_blkbits;
 881         while (size > 0) {
 882                 block = *offs >> inode->i_blkbits;
 883                 boffs = *offs & (blocksize - 1);
 884                 csize = min(blocksize - boffs, size);
 885                 bh = ldiskfs_bread(NULL, inode, block, 0, &err);
 886                 if (!bh) {
 887                         CERROR("%s: error reading offset %llu (block %lu): "
 888                                "rc = %d\n",
 889                                inode->i_sb->s_id, *offs, block, err);
 890                         return err;
 891                 }
 892
 893                 memcpy(buf, bh->b_data + boffs, csize);
 894                 brelse(bh);
 895
 896                 *offs += csize;
 897                 buf += csize;
 898                 size -= csize;
 899         }
 900         return osize;
 901 }
 902
 903 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
 904                         struct lu_buf *buf, loff_t *pos,
 905                         struct lustre_capa *capa)
 906 {
 907         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 908         int           rc;
 909
 910         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_READ))
 911                 RETURN(-EACCES);
 912
 913         /* Read small symlink from inode body as we need to maintain correct
 914          * on-disk symlinks for ldiskfs.
 915          */
 916         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
 917             (buf->lb_len <= sizeof(LDISKFS_I(inode)->i_data)))
 918                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
 919         else
 920                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
 921
 922         return rc;
 923 }
 924
 925 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 926                                  const loff_t size, loff_t pos,
 927                                  struct thandle *handle)
 928 {
 929         struct osd_thandle *oh;
 930         int                 credits;
 931
 932         LASSERT(handle != NULL);
 933
 934         oh = container_of0(handle, struct osd_thandle, ot_super);
 935         LASSERT(oh->ot_handle == NULL);
 936
 937         /* XXX: size == 0 or INT_MAX indicating a catalog header update or
 938          *      llog write, see comment in mdd_declare_llog_record().
 939          *
 940          *      This hack will be removed with llog over OSD landing
 941          */
 942         if (size == DECLARE_LLOG_REWRITE)
 943                 credits = 2;
 944         else if (size == DECLARE_LLOG_WRITE)
 945                 credits = 6;
 946         else
 947                 credits = osd_dto_credits_noquota[DTO_WRITE_BLOCK];
 948
 949         OSD_DECLARE_OP(oh, write);
 950         oh->ot_credits += credits;
 951
 952         if (osd_dt_obj(dt)->oo_inode == NULL)
 953                 return 0;
 954
 955         osd_declare_qid(dt, oh, USRQUOTA, osd_dt_obj(dt)->oo_inode->i_uid,
 956                         osd_dt_obj(dt)->oo_inode);
 957         osd_declare_qid(dt, oh, GRPQUOTA, osd_dt_obj(dt)->oo_inode->i_gid,
 958                         osd_dt_obj(dt)->oo_inode);
 959         return 0;
 960 }
 961
 962 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
 963 {
 964
 965         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
 966         LDISKFS_I(inode)->i_disksize = buflen;
 967         i_size_write(inode, buflen);
 968         inode->i_sb->s_op->dirty_inode(inode);
 969
 970         return 0;
 971 }
 972
 973 static int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
 974                                     loff_t *offs, handle_t *handle)
 975 {
 976         struct buffer_head *bh        = NULL;
 977         loff_t              offset    = *offs;
 978         loff_t              new_size  = i_size_read(inode);
 979         unsigned long       block;
 980         int                 blocksize = 1 << inode->i_blkbits;
 981         int                 err = 0;
 982         int                 size;
 983         int                 boffs;
 984         int                 dirty_inode = 0;
 985
 986         while (bufsize > 0) {
 987                 if (bh != NULL)
 988                         brelse(bh);
 989
 990                 block = offset >> inode->i_blkbits;
 991                 boffs = offset & (blocksize - 1);
 992                 size = min(blocksize - boffs, bufsize);
 993                 bh = ldiskfs_bread(handle, inode, block, 1, &err);
 994                 if (!bh) {
 995                         CERROR("%s: error reading offset %llu (block %lu): "
 996                                "rc = %d\n",
 997                                inode->i_sb->s_id, offset, block, err);
 998                         break;
 999                 }
1000
1001                 err = ldiskfs_journal_get_write_access(handle, bh);
1002                 if (err) {
1003                         CERROR("journal_get_write_access() returned error %d\n",
1004                                err);
1005                         break;
1006                 }
1007                 LASSERTF(boffs + size <= bh->b_size,
1008                          "boffs %d size %d bh->b_size %lu",
1009                          boffs, size, (unsigned long)bh->b_size);
1010                 memcpy(bh->b_data + boffs, buf, size);
1011                 err = ldiskfs_journal_dirty_metadata(handle, bh);
1012                 if (err)
1013                         break;
1014
1015                 if (offset + size > new_size)
1016                         new_size = offset + size;
1017                 offset += size;
1018                 bufsize -= size;
1019                 buf += size;
1020         }
1021         if (bh)
1022                 brelse(bh);
1023
1024         /* correct in-core and on-disk sizes */
1025         if (new_size > i_size_read(inode)) {
1026                 cfs_spin_lock(&inode->i_lock);
1027                 if (new_size > i_size_read(inode))
1028                         i_size_write(inode, new_size);
1029                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1030                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1031                         dirty_inode = 1;
1032                 }
1033                 cfs_spin_unlock(&inode->i_lock);
1034                 if (dirty_inode)
1035                         inode->i_sb->s_op->dirty_inode(inode);
1036         }
1037
1038         if (err == 0)
1039                 *offs = offset;
1040         return err;
1041 }
1042
1043 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1044                          const struct lu_buf *buf, loff_t *pos,
1045                          struct thandle *handle, struct lustre_capa *capa,
1046                          int ignore_quota)
1047 {
1048         struct inode       *inode = osd_dt_obj(dt)->oo_inode;
1049         struct osd_thandle *oh;
1050         ssize_t             result;
1051 #ifdef HAVE_QUOTA_SUPPORT
1052         cfs_cap_t           save = cfs_curproc_cap_pack();
1053 #endif
1054
1055         LASSERT(dt_object_exists(dt));
1056
1057         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_WRITE))
1058                 return -EACCES;
1059
1060         LASSERT(handle != NULL);
1061
1062         /* XXX: don't check: one declared chunk can be used many times */
1063         /* OSD_EXEC_OP(handle, write); */
1064
1065         oh = container_of(handle, struct osd_thandle, ot_super);
1066         LASSERT(oh->ot_handle->h_transaction != NULL);
1067 #ifdef HAVE_QUOTA_SUPPORT
1068         if (ignore_quota)
1069                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
1070         else
1071                 cfs_cap_lower(CFS_CAP_SYS_RESOURCE);
1072 #endif
1073         /* Write small symlink to inode body as we need to maintain correct
1074          * on-disk symlinks for ldiskfs.
1075          */
1076         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1077             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1078                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1079         else
1080                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1081                                                   buf->lb_len, pos,
1082                                                   oh->ot_handle);
1083 #ifdef HAVE_QUOTA_SUPPORT
1084         cfs_curproc_cap_unpack(save);
1085 #endif
1086         if (result == 0)
1087                 result = buf->lb_len;
1088         return result;
1089 }
1090
1091 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1092                              __u64 start, __u64 end, struct thandle *th)
1093 {
1094         struct osd_thandle *oh;
1095         ENTRY;
1096
1097         LASSERT(th);
1098         oh = container_of(th, struct osd_thandle, ot_super);
1099
1100         OSD_DECLARE_OP(oh, punch);
1101
1102         /*
1103          * we don't need to reserve credits for whole truncate
1104          * it's not possible as truncate may need to free too many
1105          * blocks and that won't fit a single transaction. instead
1106          * we reserve credits to change i_size and put inode onto
1107          * orphan list. if needed truncate will extend or restart
1108          * transaction
1109          */
1110         oh->ot_credits += osd_dto_credits_noquota[DTO_ATTR_SET_BASE];
1111         oh->ot_credits += 3;
1112
1113         RETURN(0);
1114 }
1115
1116 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1117                      __u64 start, __u64 end, struct thandle *th,
1118                      struct lustre_capa *capa)
1119 {
1120         struct osd_thandle *oh;
1121         struct osd_object  *obj = osd_dt_obj(dt);
1122         struct inode       *inode = obj->oo_inode;
1123         handle_t           *h;
1124         tid_t               tid;
1125         int                 rc, rc2 = 0;
1126         ENTRY;
1127
1128         LASSERT(end == OBD_OBJECT_EOF);
1129         LASSERT(dt_object_exists(dt));
1130         LASSERT(osd_invariant(obj));
1131
1132         LASSERT(th);
1133         oh = container_of(th, struct osd_thandle, ot_super);
1134         LASSERT(oh->ot_handle->h_transaction != NULL);
1135
1136         OSD_EXEC_OP(th, punch);
1137
1138         tid = oh->ot_handle->h_transaction->t_tid;
1139
1140         rc = vmtruncate(inode, start);
1141
1142         /*
1143          * For a partial-page truncate, flush the page to disk immediately to
1144          * avoid data corruption during direct disk write.  b=17397
1145          */
1146         if (rc == 0 && (start & ~CFS_PAGE_MASK) != 0)
1147                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1148
1149         h = journal_current_handle();
1150         LASSERT(h != NULL);
1151         LASSERT(h == oh->ot_handle);
1152
1153         if (tid != h->h_transaction->t_tid) {
1154                 int credits = oh->ot_credits;
1155                 /*
1156                  * transaction has changed during truncate
1157                  * we need to restart the handle with our credits
1158                  */
1159                 if (h->h_buffer_credits < credits) {
1160                         if (ldiskfs_journal_extend(h, credits))
1161                                 rc2 = ldiskfs_journal_restart(h, credits);
1162                 }
1163         }
1164
1165         RETURN(rc == 0 ? rc2 : rc);
1166 }
1167
1168 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1169                           struct ll_user_fiemap *fm)
1170 {
1171         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1172         struct osd_thread_info *info   = osd_oti_get(env);
1173         struct dentry          *dentry = &info->oti_obj_dentry;
1174         struct file            *file   = &info->oti_file;
1175         mm_segment_t            saved_fs;
1176         int rc;
1177
1178         LASSERT(inode);
1179         dentry->d_inode = inode;
1180         file->f_dentry = dentry;
1181         file->f_mapping = inode->i_mapping;
1182         file->f_op = inode->i_fop;
1183
1184         saved_fs = get_fs();
1185         set_fs(get_ds());
1186         /* ldiskfs_ioctl does not have a inode argument */
1187         if (inode->i_fop->unlocked_ioctl)
1188                 rc = inode->i_fop->unlocked_ioctl(file, FSFILT_IOC_FIEMAP,
1189                                                   (long)fm);
1190         else
1191                 rc = -ENOTTY;
1192         set_fs(saved_fs);
1193         return rc;
1194 }
1195
1196 /*
1197  * in some cases we may need declare methods for objects being created
1198  * e.g., when we create symlink
1199  */
1200 const struct dt_body_operations osd_body_ops_new = {
1201         .dbo_declare_write = osd_declare_write,
1202 };
1203
1204 const struct dt_body_operations osd_body_ops = {
1205         .dbo_read                 = osd_read,
1206         .dbo_declare_write        = osd_declare_write,
1207         .dbo_write                = osd_write,
1208         .dbo_bufs_get             = osd_bufs_get,
1209         .dbo_bufs_put             = osd_bufs_put,
1210         .dbo_write_prep           = osd_write_prep,
1211         .dbo_declare_write_commit = osd_declare_write_commit,
1212         .dbo_write_commit         = osd_write_commit,
1213         .dbo_read_prep            = osd_read_prep,
1214         .do_declare_punch         = osd_declare_punch,
1215         .do_punch                 = osd_punch,
1216         .dbo_fiemap_get           = osd_fiemap_get,
1217 };
1218