lustre/obdfilter/filter_io_26.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/fs/obdfilter/filter_io.c
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *   Author: Peter Braam <braam@clusterfs.com>
   8  *   Author: Andreas Dilger <adilger@clusterfs.com>
   9  *   Author: Phil Schwan <phil@clusterfs.com>
  10  *
  11  *   This file is part of the Lustre file system, http://www.lustre.org
  12  *   Lustre is a trademark of Cluster File Systems, Inc.
  13  *
  14  *   You may have signed or agreed to another license before downloading
  15  *   this software.  If so, you are bound by the terms and conditions
  16  *   of that agreement, and the following does not apply to you.  See the
  17  *   LICENSE file included with this distribution for more information.
  18  *
  19  *   If you did not agree to a different license, then this copy of Lustre
  20  *   is open source software; you can redistribute it and/or modify it
  21  *   under the terms of version 2 of the GNU General Public License as
  22  *   published by the Free Software Foundation.
  23  *
  24  *   In either case, Lustre is distributed in the hope that it will be
  25  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  26  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  27  *   license text for more details.
  28  */
  29
  30 #ifdef HAVE_KERNEL_CONFIG_H
  31 #include <linux/config.h>
  32 #endif
  33 #include <linux/module.h>
  34 #include <linux/pagemap.h> // XXX kill me soon
  35 #include <linux/version.h>
  36 #include <linux/buffer_head.h>
  37
  38 #define DEBUG_SUBSYSTEM S_FILTER
  39
  40 #include <obd_class.h>
  41 #include <lustre_fsfilt.h>
  42 #include <lustre_quota.h>
  43 #include "filter_internal.h"
  44
  45 /* 512byte block min */
  46 #define MAX_BLOCKS_PER_PAGE (CFS_PAGE_SIZE / 512)
  47 struct filter_iobuf {
  48         atomic_t          dr_numreqs;  /* number of reqs being processed */
  49         wait_queue_head_t dr_wait;
  50         int               dr_max_pages;
  51         int               dr_npages;
  52         int               dr_error;
  53         struct page     **dr_pages;
  54         unsigned long    *dr_blocks;
  55         spinlock_t        dr_lock;              /* IRQ lock */
  56         unsigned int      dr_ignore_quota:1;
  57         struct filter_obd *dr_filter;
  58 };
  59
  60 static void record_start_io(struct filter_iobuf *iobuf, int rw, int size,
  61                             struct obd_export *exp)
  62 {
  63         struct filter_obd *filter = iobuf->dr_filter;
  64
  65         atomic_inc(&iobuf->dr_numreqs);
  66
  67         if (rw == OBD_BRW_READ) {
  68                 atomic_inc(&filter->fo_r_in_flight);
  69                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST],
  70                                  atomic_read(&filter->fo_r_in_flight));
  71                 lprocfs_oh_tally_log2(&filter->fo_filter_stats.hist[BRW_R_DISK_IOSIZE], size);
  72                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_RPC_HIST],
  73                                  atomic_read(&filter->fo_r_in_flight));
  74                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_DISK_IOSIZE], size);
  75         } else {
  76                 atomic_inc(&filter->fo_w_in_flight);
  77                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
  78                                  atomic_read(&filter->fo_w_in_flight));
  79                 lprocfs_oh_tally_log2(&filter->fo_filter_stats.hist[BRW_W_DISK_IOSIZE], size);
  80                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_RPC_HIST],
  81                                  atomic_read(&filter->fo_w_in_flight));
  82                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_DISK_IOSIZE], size);
  83         }
  84 }
  85
  86 static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
  87 {
  88         struct filter_obd *filter = iobuf->dr_filter;
  89
  90         /* CAVEAT EMPTOR: possibly in IRQ context
  91          * DO NOT record procfs stats here!!! */
  92
  93         if (rw == OBD_BRW_READ)
  94                 atomic_dec(&filter->fo_r_in_flight);
  95         else
  96                 atomic_dec(&filter->fo_w_in_flight);
  97
  98         if (atomic_dec_and_test(&iobuf->dr_numreqs))
  99                 wake_up(&iobuf->dr_wait);
 100 }
 101
 102 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 103 {
 104         struct filter_iobuf *iobuf = bio->bi_private;
 105         unsigned long        flags;
 106
 107         /* CAVEAT EMPTOR: possibly in IRQ context
 108          * DO NOT record procfs stats here!!! */
 109
 110         if (bio->bi_size)                       /* Not complete */
 111                 return 1;
 112
 113         if (iobuf == NULL) {
 114                 CERROR("***** bio->bi_private is NULL!  This should never "
 115                        "happen.  Normally, I would crash here, but instead I "
 116                        "will dump the bio contents to the console.  Please "
 117                        "report this to CFS, along with any interesting "
 118                        "messages leading up to this point (like SCSI errors, "
 119                        "perhaps).  Because bi_private is NULL, I can't wake up "
 120                        "the thread that initiated this I/O -- so you will "
 121                        "probably have to reboot this node.\n");
 122                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 123                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 124                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 125                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 126                        bio->bi_end_io, atomic_read(&bio->bi_cnt),
 127                        bio->bi_private);
 128                 return 0;
 129         }
 130
 131         spin_lock_irqsave(&iobuf->dr_lock, flags);
 132         if (iobuf->dr_error == 0)
 133                 iobuf->dr_error = error;
 134         spin_unlock_irqrestore(&iobuf->dr_lock, flags);
 135
 136         record_finish_io(iobuf, test_bit(BIO_RW, &bio->bi_rw) ?
 137                          OBD_BRW_WRITE : OBD_BRW_READ, error);
 138
 139         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 140          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 141          * mempool when serious on-disk fragmentation was encountered,
 142          * deadlocking the OST.  The bios are now released as soon as complete
 143          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 144         bio_put(bio);
 145         return 0;
 146 }
 147
 148 static int can_be_merged(struct bio *bio, sector_t sector)
 149 {
 150         unsigned int size;
 151
 152         if (!bio)
 153                 return 0;
 154
 155         size = bio->bi_size >> 9;
 156         return bio->bi_sector + size == sector ? 1 : 0;
 157 }
 158
 159 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 160                                         int rw, int num_pages)
 161 {
 162         struct filter_iobuf *iobuf;
 163
 164         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 165
 166         OBD_ALLOC(iobuf, sizeof(*iobuf));
 167         if (iobuf == NULL)
 168                 goto failed_0;
 169
 170         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 171         if (iobuf->dr_pages == NULL)
 172                 goto failed_1;
 173
 174         OBD_ALLOC(iobuf->dr_blocks,
 175                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 176         if (iobuf->dr_blocks == NULL)
 177                 goto failed_2;
 178
 179         iobuf->dr_filter = filter;
 180         init_waitqueue_head(&iobuf->dr_wait);
 181         atomic_set(&iobuf->dr_numreqs, 0);
 182         spin_lock_init(&iobuf->dr_lock);
 183         iobuf->dr_max_pages = num_pages;
 184         iobuf->dr_npages = 0;
 185
 186         RETURN(iobuf);
 187
 188  failed_2:
 189         OBD_FREE(iobuf->dr_pages,
 190                  num_pages * sizeof(*iobuf->dr_pages));
 191  failed_1:
 192         OBD_FREE(iobuf, sizeof(*iobuf));
 193  failed_0:
 194         RETURN(ERR_PTR(-ENOMEM));
 195 }
 196
 197 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
 198 {
 199         iobuf->dr_npages = 0;
 200         atomic_set(&iobuf->dr_numreqs, 0);
 201 }
 202
 203 void filter_free_iobuf(struct filter_iobuf *iobuf)
 204 {
 205         int num_pages = iobuf->dr_max_pages;
 206
 207         filter_clear_iobuf(iobuf);
 208
 209         OBD_FREE(iobuf->dr_blocks,
 210                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 211         OBD_FREE(iobuf->dr_pages,
 212                  num_pages * sizeof(*iobuf->dr_pages));
 213         OBD_FREE_PTR(iobuf);
 214 }
 215
 216 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 217                       struct obd_trans_info *oti)
 218 {
 219         int thread_id = oti ? oti->oti_thread_id : -1;
 220
 221         if (unlikely(thread_id < 0)) {
 222                 filter_free_iobuf(iobuf);
 223                 return;
 224         }
 225
 226         LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
 227                  "iobuf mismatch for thread %d: pool %p iobuf %p\n",
 228                  thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
 229         filter_clear_iobuf(iobuf);
 230 }
 231
 232 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 233                           struct inode *inode, struct page *page)
 234 {
 235         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 236         iobuf->dr_pages[iobuf->dr_npages++] = page;
 237
 238         return 0;
 239 }
 240
 241 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 242                   struct filter_iobuf *iobuf, int rw)
 243 {
 244         struct obd_device *obd = exp->exp_obd;
 245         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 246         struct page  **pages = iobuf->dr_pages;
 247         int            npages = iobuf->dr_npages;
 248         unsigned long *blocks = iobuf->dr_blocks;
 249         int            total_blocks = npages * blocks_per_page;
 250         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 251         unsigned int   blocksize = inode->i_sb->s_blocksize;
 252         struct bio    *bio = NULL;
 253         int            frags = 0;
 254         unsigned long  start_time = jiffies;
 255         struct page   *page;
 256         unsigned int   page_offset;
 257         sector_t       sector;
 258         int            nblocks;
 259         int            block_idx;
 260         int            page_idx;
 261         int            i;
 262         int            rc = 0;
 263         ENTRY;
 264
 265         LASSERT(iobuf->dr_npages == npages);
 266         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 267
 268         for (page_idx = 0, block_idx = 0;
 269              page_idx < npages;
 270              page_idx++, block_idx += blocks_per_page) {
 271
 272                 page = pages[page_idx];
 273                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 274
 275                 for (i = 0, page_offset = 0;
 276                      i < blocks_per_page;
 277                      i += nblocks, page_offset += blocksize * nblocks) {
 278
 279                         nblocks = 1;
 280
 281                         if (blocks[block_idx + i] == 0) {  /* hole */
 282                                 LASSERT(rw == OBD_BRW_READ);
 283                                 memset(kmap(page) + page_offset, 0, blocksize);
 284                                 kunmap(page);
 285                                 continue;
 286                         }
 287
 288                         sector = blocks[block_idx + i] << sector_bits;
 289
 290                         /* Additional contiguous file blocks? */
 291                         while (i + nblocks < blocks_per_page &&
 292                                (sector + nblocks*(blocksize>>9)) ==
 293                                (blocks[block_idx + i + nblocks] << sector_bits))
 294                                 nblocks++;
 295
 296                         if (bio != NULL &&
 297                             can_be_merged(bio, sector) &&
 298                             bio_add_page(bio, page,
 299                                          blocksize * nblocks, page_offset) != 0)
 300                                 continue;       /* added this frag OK */
 301
 302                         if (bio != NULL) {
 303                                 request_queue_t *q =
 304                                         bdev_get_queue(bio->bi_bdev);
 305
 306                                 /* Dang! I have to fragment this I/O */
 307                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 308                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 309                                        bio->bi_size,
 310                                        bio->bi_vcnt, bio->bi_max_vecs,
 311                                        bio->bi_size >> 9, q->max_sectors,
 312                                        bio_phys_segments(q, bio),
 313                                        q->max_phys_segments,
 314                                        bio_hw_segments(q, bio),
 315                                        q->max_hw_segments);
 316
 317                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 318                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 319                                 if (rc < 0) {
 320                                         CERROR("Can't send bio: %d\n", rc);
 321                                         record_finish_io(iobuf, rw, rc);
 322                                         goto out;
 323                                 }
 324                                 frags++;
 325                         }
 326
 327                         /* allocate new bio */
 328                         bio = bio_alloc(GFP_NOIO,
 329                                         (npages - page_idx) * blocks_per_page);
 330                         if (bio == NULL) {
 331                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 332                                        (npages - page_idx), blocks_per_page,
 333                                        (npages - page_idx) * blocks_per_page);
 334                                 rc = -ENOMEM;
 335                                 goto out;
 336                         }
 337
 338                         bio->bi_bdev = inode->i_sb->s_bdev;
 339                         bio->bi_sector = sector;
 340                         bio->bi_end_io = dio_complete_routine;
 341                         bio->bi_private = iobuf;
 342
 343                         rc = bio_add_page(bio, page,
 344                                           blocksize * nblocks, page_offset);
 345                         LASSERT (rc != 0);
 346                 }
 347         }
 348
 349         if (bio != NULL) {
 350                 record_start_io(iobuf, rw, bio->bi_size, exp);
 351                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 352                 if (rc >= 0) {
 353                         frags++;
 354                         rc = 0;
 355                 } else {
 356                         CERROR("Can't send bio: %d\n", rc);
 357                         record_finish_io(iobuf, rw, rc);
 358                 }
 359         }
 360
 361  out:
 362         wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0);
 363
 364         if (rw == OBD_BRW_READ) {
 365                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_R_DIO_FRAGS], frags);
 366                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_DIO_FRAGS],
 367                                  frags);
 368                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_R_IO_TIME],
 369                                       jiffies - start_time);
 370                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_IO_TIME],
 371                                  jiffies - start_time);
 372         } else {
 373                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_W_DIO_FRAGS], frags);
 374                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_DIO_FRAGS],
 375                                  frags);
 376                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_W_IO_TIME],
 377                                       jiffies - start_time);
 378                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_IO_TIME],
 379                                  jiffies - start_time);
 380         }
 381
 382         if (rc == 0)
 383                 rc = iobuf->dr_error;
 384         RETURN(rc);
 385 }
 386
 387 /* These are our hacks to keep our directio/bh IO coherent with ext3's
 388  * page cache use.  Most notably ext3 reads file data into the page
 389  * cache when it is zeroing the tail of partial-block truncates and
 390  * leaves it there, sometimes generating io from it at later truncates.
 391  * This removes the partial page and its buffers from the page cache,
 392  * so it should only ever cause a wait in rare cases, as otherwise we
 393  * always do full-page IO to the OST.
 394  *
 395  * The call to truncate_complete_page() will call journal_invalidatepage()
 396  * to free the buffers and drop the page from cache.  The buffers should
 397  * not be dirty, because we already called fdatasync/fdatawait on them.
 398  */
 399 static int filter_sync_inode_data(struct inode *inode, int locked)
 400 {
 401         int rc = 0;
 402
 403         /* This is nearly do_fsync(), without the waiting on the inode */
 404         /* XXX: in 2.6.16 (at least) we don't need to hold i_mutex over
 405          * filemap_fdatawrite() and filemap_fdatawait(), so we may no longer
 406          * need this lock here at all. */
 407         if (!locked)
 408                 LOCK_INODE_MUTEX(inode);
 409         if (inode->i_mapping->nrpages) {
 410                 current->flags |= PF_SYNCWRITE;
 411                 rc = filemap_fdatawrite(inode->i_mapping);
 412                 if (rc == 0)
 413                         rc = filemap_fdatawait(inode->i_mapping);
 414                 current->flags &= ~PF_SYNCWRITE;
 415         }
 416         if (!locked)
 417                 UNLOCK_INODE_MUTEX(inode);
 418
 419         return rc;
 420 }
 421
 422 /* Clear pages from the mapping before we do direct IO to that offset.
 423  * Now that the only source of such pages in the truncate path flushes
 424  * these pages to disk and then discards them, this is error condition.
 425  * If add back read cache this will happen again.  This could be disabled
 426  * until that time if we never see the below error. */
 427 static int filter_clear_page_cache(struct inode *inode,
 428                                    struct filter_iobuf *iobuf)
 429 {
 430         struct page *page;
 431         int i, rc;
 432
 433         rc = filter_sync_inode_data(inode, 0);
 434         if (rc != 0)
 435                 RETURN(rc);
 436
 437         /* be careful to call this after fsync_inode_data_buffers has waited
 438          * for IO to complete before we evict it from the cache */
 439         for (i = 0; i < iobuf->dr_npages; i++) {
 440                 page = find_lock_page(inode->i_mapping,
 441                                       iobuf->dr_pages[i]->index);
 442                 if (page == NULL)
 443                         continue;
 444                 if (page->mapping != NULL) {
 445                         CERROR("page %lu (%d/%d) in page cache during write!\n",
 446                                page->index, i, iobuf->dr_npages);
 447                         wait_on_page_writeback(page);
 448                         ll_truncate_complete_page(page);
 449                 }
 450
 451                 unlock_page(page);
 452                 page_cache_release(page);
 453         }
 454
 455         return 0;
 456 }
 457
 458 int filter_clear_truncated_page(struct inode *inode)
 459 {
 460         struct page *page;
 461         int rc;
 462
 463         /* Truncate on page boundary, so nothing to flush? */
 464         if (!(inode->i_size & ~CFS_PAGE_MASK))
 465                 return 0;
 466
 467         rc = filter_sync_inode_data(inode, 1);
 468         if (rc != 0)
 469                 RETURN(rc);
 470
 471         /* be careful to call this after fsync_inode_data_buffers has waited
 472          * for IO to complete before we evict it from the cache */
 473         page = find_lock_page(inode->i_mapping,
 474                               inode->i_size >> CFS_PAGE_SHIFT);
 475         if (page) {
 476                 if (page->mapping != NULL) {
 477                         wait_on_page_writeback(page);
 478                         ll_truncate_complete_page(page);
 479                 }
 480                 unlock_page(page);
 481                 page_cache_release(page);
 482         }
 483
 484         return 0;
 485 }
 486
 487 /* Must be called with i_mutex taken for writes; this will drop it */
 488 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 489                      struct obd_export *exp, struct iattr *attr,
 490                      struct obd_trans_info *oti, void **wait_handle)
 491 {
 492         struct obd_device *obd = exp->exp_obd;
 493         struct inode *inode = dchild->d_inode;
 494         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 495         int rc, rc2, create;
 496         struct semaphore *sem;
 497         ENTRY;
 498
 499         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 500                  iobuf->dr_npages, iobuf->dr_max_pages);
 501         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 502
 503         if (rw == OBD_BRW_READ) {
 504                 if (iobuf->dr_npages == 0)
 505                         RETURN(0);
 506                 create = 0;
 507                 sem = NULL;
 508         } else {
 509                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 510                 LASSERT(iobuf->dr_npages > 0);
 511                 create = 1;
 512                 sem = &obd->u.filter.fo_alloc_lock;
 513
 514                 lquota_enforce(filter_quota_interface_ref, obd,
 515                                iobuf->dr_ignore_quota);
 516         }
 517
 518         rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 519                                     iobuf->dr_npages, iobuf->dr_blocks,
 520                                     obdfilter_created_scratchpad, create, sem);
 521
 522         if (rw == OBD_BRW_WRITE) {
 523                 if (rc == 0) {
 524                         filter_tally_write(exp, iobuf->dr_pages,
 525                                            iobuf->dr_npages, iobuf->dr_blocks,
 526                                            blocks_per_page);
 527                         if (attr->ia_size > inode->i_size)
 528                                 attr->ia_valid |= ATTR_SIZE;
 529                         rc = fsfilt_setattr(obd, dchild,
 530                                             oti->oti_handle, attr, 0);
 531                 }
 532
 533                 UNLOCK_INODE_MUTEX(inode);
 534
 535                 rc2 = filter_finish_transno(exp, oti, 0);
 536                 if (rc2 != 0) {
 537                         CERROR("can't close transaction: %d\n", rc2);
 538                         if (rc == 0)
 539                                 rc = rc2;
 540                 }
 541
 542                 rc2 =fsfilt_commit_async(obd,inode,oti->oti_handle,wait_handle);
 543                 if (rc == 0)
 544                         rc = rc2;
 545                 if (rc != 0)
 546                         RETURN(rc);
 547         } else if (rc == 0) {
 548                 filter_tally_read(exp, iobuf->dr_pages,
 549                                   iobuf->dr_npages, iobuf->dr_blocks,
 550                                   blocks_per_page);
 551         }
 552
 553         rc = filter_clear_page_cache(inode, iobuf);
 554         if (rc != 0)
 555                 RETURN(rc);
 556
 557         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 558 }
 559
 560 /* See if there are unallocated parts in given file region */
 561 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 562 {
 563         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 564                 inode->i_mapping->a_ops->bmap;
 565         int j;
 566
 567         /* We can't know if we are overwriting or not */
 568         if (fs_bmap == NULL)
 569                 return 0;
 570
 571         offset >>= inode->i_blkbits;
 572         len >>= inode->i_blkbits;
 573
 574         for (j = 0; j <= len; j++)
 575                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 576                         return 0;
 577
 578         return 1;
 579 }
 580
 581 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 582                           int objcount, struct obd_ioobj *obj, int niocount,
 583                           struct niobuf_local *res, struct obd_trans_info *oti,
 584                           int rc)
 585 {
 586         struct niobuf_local *lnb;
 587         struct filter_iobuf *iobuf = NULL;
 588         struct lvfs_run_ctxt saved;
 589         struct fsfilt_objinfo fso;
 590         struct iattr iattr = { 0 };
 591         struct inode *inode = NULL;
 592         unsigned long now = jiffies;
 593         int i, err, cleanup_phase = 0;
 594         struct obd_device *obd = exp->exp_obd;
 595         void *wait_handle;
 596         int   total_size = 0, rc2;
 597         unsigned int qcids[MAXQUOTAS] = {0, 0};
 598         ENTRY;
 599
 600         LASSERT(oti != NULL);
 601         LASSERT(objcount == 1);
 602         LASSERT(current->journal_info == NULL);
 603
 604         if (rc != 0)
 605                 GOTO(cleanup, rc);
 606
 607         /* Unfortunately, if quota master is too busy to handle the
 608          * pre-dqacq in time and quota hash on ost is used up, we
 609          * have to wait for the completion of in flight dqacq/dqrel,
 610          * then try again */
 611         if ((rc2 = lquota_chkquota(filter_quota_interface_ref, obd, oa->o_uid,
 612                                    oa->o_gid, niocount)) == QUOTA_RET_ACQUOTA) {
 613                 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_HOLD_WRITE_RPC, 90);
 614                 lquota_acquire(filter_quota_interface_ref, obd, oa->o_uid,
 615                                oa->o_gid);
 616         }
 617
 618         if (rc2 < 0) {
 619                 rc = rc2;
 620                 GOTO(cleanup, rc);
 621         }
 622
 623         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 624         if (IS_ERR(iobuf))
 625                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 626         cleanup_phase = 1;
 627
 628         fso.fso_dentry = res->dentry;
 629         fso.fso_bufcnt = obj->ioo_bufcnt;
 630         inode = res->dentry->d_inode;
 631
 632         iobuf->dr_ignore_quota = 0;
 633         for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
 634                 loff_t this_size;
 635
 636                 /* If overwriting an existing block, we don't need a grant */
 637                 if (!(lnb->flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC &&
 638                     filter_range_is_mapped(inode, lnb->offset, lnb->len))
 639                         lnb->rc = 0;
 640
 641                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 642                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 643                         continue;
 644                 }
 645
 646                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 647                 LASSERT (err == 0);
 648
 649                 total_size += lnb->len;
 650
 651                 /* we expect these pages to be in offset order, but we'll
 652                  * be forgiving */
 653                 this_size = lnb->offset + lnb->len;
 654                 if (this_size > iattr.ia_size)
 655                         iattr.ia_size = this_size;
 656
 657                 /* if one page is a write-back page from client cache, or it's
 658                  * written by root, then mark the whole io request as ignore
 659                  * quota request */
 660                 if (lnb->flags & (OBD_BRW_FROM_GRANT | OBD_BRW_NOQUOTA))
 661                         iobuf->dr_ignore_quota = 1;
 662         }
 663
 664         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 665         cleanup_phase = 2;
 666
 667         DQUOT_INIT(inode);
 668
 669         LOCK_INODE_MUTEX(inode);
 670         fsfilt_check_slow(obd, now, obd_timeout, "i_mutex");
 671         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 672                                            oti);
 673         if (IS_ERR(oti->oti_handle)) {
 674                 UNLOCK_INODE_MUTEX(inode);
 675                 rc = PTR_ERR(oti->oti_handle);
 676                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 677                        "error starting transaction: rc = %d\n", rc);
 678                 oti->oti_handle = NULL;
 679                 GOTO(cleanup, rc);
 680         }
 681         /* have to call fsfilt_commit() from this point on */
 682
 683         fsfilt_check_slow(obd, now, obd_timeout, "brw_start");
 684
 685         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 686
 687         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 688          * then we will accept the UID+GID if sent by the client for
 689          * initializing the ownership of this inode.  We only allow this to
 690          * happen once (so clear these bits) and later only allow setattr. */
 691         if (inode->i_mode & S_ISUID)
 692                 i |= OBD_MD_FLUID;
 693         if (inode->i_mode & S_ISGID)
 694                 i |= OBD_MD_FLGID;
 695
 696         iattr_from_obdo(&iattr, oa, i);
 697         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 698                 unsigned int save;
 699
 700                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 701                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 702
 703                 cap_raise(current->cap_effective, CAP_SYS_RESOURCE);
 704
 705                 iattr.ia_valid |= ATTR_MODE;
 706                 iattr.ia_mode = inode->i_mode;
 707                 if (iattr.ia_valid & ATTR_UID)
 708                         iattr.ia_mode &= ~S_ISUID;
 709                 if (iattr.ia_valid & ATTR_GID)
 710                         iattr.ia_mode &= ~S_ISGID;
 711
 712                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 713
 714                 /* To avoid problems with quotas, UID and GID must be set
 715                  * in the inode before filter_direct_io() - see bug 10357. */
 716                 save = iattr.ia_valid;
 717                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 718                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle, &iattr, 0);
 719                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 720                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 721                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 722         }
 723
 724         /* filter_direct_io drops i_mutex */
 725         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 726                               oti, &wait_handle);
 727         if (rc == 0)
 728                 obdo_from_inode(oa, inode,
 729                                 FILTER_VALID_FLAGS |OBD_MD_FLUID |OBD_MD_FLGID);
 730         else
 731                 obdo_from_inode(oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
 732
 733         lquota_getflag(filter_quota_interface_ref, obd, oa);
 734
 735         fsfilt_check_slow(obd, now, obd_timeout, "direct_io");
 736
 737         err = fsfilt_commit_wait(obd, inode, wait_handle);
 738         if (err) {
 739                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 740                 rc = err;
 741         }
 742
 743         if (obd->obd_replayable && !rc)
 744                 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
 745                          "oti_transno "LPU64" last_committed "LPU64"\n",
 746                          oti->oti_transno, obd->obd_last_committed);
 747
 748         fsfilt_check_slow(obd, now, obd_timeout, "commitrw commit");
 749
 750 cleanup:
 751         filter_grant_commit(exp, niocount, res);
 752
 753         switch (cleanup_phase) {
 754         case 2:
 755                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 756                 LASSERT(current->journal_info == NULL);
 757         case 1:
 758                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 759         case 0:
 760                 /*
 761                  * lnb->page automatically returns back into per-thread page
 762                  * pool (bug 5137)
 763                  */
 764                 f_dput(res->dentry);
 765         }
 766
 767         /* trigger quota pre-acquire */
 768         qcids[USRQUOTA] = oa->o_uid;
 769         qcids[GRPQUOTA] = oa->o_gid;
 770         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 771                             FSFILT_OP_CREATE);
 772         CDEBUG(err ? D_ERROR : D_QUOTA,
 773                "filter adjust qunit! (rc:%d)\n", err);
 774
 775         RETURN(rc);
 776 }