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 #ifndef AUTOCONF_INCLUDED
  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],
  72                                       size);
  73                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_RPC_HIST],
  74                                  atomic_read(&filter->fo_r_in_flight));
  75                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_DISK_IOSIZE], size);
  76         } else {
  77                 atomic_inc(&filter->fo_w_in_flight);
  78                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
  79                                  atomic_read(&filter->fo_w_in_flight));
  80                 lprocfs_oh_tally_log2(&filter->fo_filter_stats.hist[BRW_W_DISK_IOSIZE],
  81                                       size);
  82                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_RPC_HIST],
  83                                  atomic_read(&filter->fo_w_in_flight));
  84                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_DISK_IOSIZE], size);
  85         }
  86 }
  87
  88 static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
  89 {
  90         struct filter_obd *filter = iobuf->dr_filter;
  91
  92         /* CAVEAT EMPTOR: possibly in IRQ context
  93          * DO NOT record procfs stats here!!! */
  94
  95         if (rw == OBD_BRW_READ)
  96                 atomic_dec(&filter->fo_r_in_flight);
  97         else
  98                 atomic_dec(&filter->fo_w_in_flight);
  99
 100         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 101                 wake_up(&iobuf->dr_wait);
 102 }
 103
 104 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 105 {
 106         struct filter_iobuf *iobuf = bio->bi_private;
 107         unsigned long        flags;
 108
 109         /* CAVEAT EMPTOR: possibly in IRQ context
 110          * DO NOT record procfs stats here!!! */
 111
 112         if (bio->bi_size)                       /* Not complete */
 113                 return 1;
 114
 115         if (iobuf == NULL) {
 116                 CERROR("***** bio->bi_private is NULL!  This should never "
 117                        "happen.  Normally, I would crash here, but instead I "
 118                        "will dump the bio contents to the console.  Please "
 119                        "report this to CFS, along with any interesting "
 120                        "messages leading up to this point (like SCSI errors, "
 121                        "perhaps).  Because bi_private is NULL, I can't wake up "
 122                        "the thread that initiated this I/O -- so you will "
 123                        "probably have to reboot this node.\n");
 124                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 125                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 126                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 127                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 128                        bio->bi_end_io, atomic_read(&bio->bi_cnt),
 129                        bio->bi_private);
 130                 return 0;
 131         }
 132
 133         spin_lock_irqsave(&iobuf->dr_lock, flags);
 134         if (iobuf->dr_error == 0)
 135                 iobuf->dr_error = error;
 136         spin_unlock_irqrestore(&iobuf->dr_lock, flags);
 137
 138         record_finish_io(iobuf, test_bit(BIO_RW, &bio->bi_rw) ?
 139                          OBD_BRW_WRITE : OBD_BRW_READ, error);
 140
 141         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 142          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 143          * mempool when serious on-disk fragmentation was encountered,
 144          * deadlocking the OST.  The bios are now released as soon as complete
 145          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 146         bio_put(bio);
 147         return 0;
 148 }
 149
 150 static int can_be_merged(struct bio *bio, sector_t sector)
 151 {
 152         unsigned int size;
 153
 154         if (!bio)
 155                 return 0;
 156
 157         size = bio->bi_size >> 9;
 158         return bio->bi_sector + size == sector ? 1 : 0;
 159 }
 160
 161 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 162                                         int rw, int num_pages)
 163 {
 164         struct filter_iobuf *iobuf;
 165
 166         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 167
 168         OBD_ALLOC(iobuf, sizeof(*iobuf));
 169         if (iobuf == NULL)
 170                 goto failed_0;
 171
 172         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 173         if (iobuf->dr_pages == NULL)
 174                 goto failed_1;
 175
 176         OBD_ALLOC(iobuf->dr_blocks,
 177                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 178         if (iobuf->dr_blocks == NULL)
 179                 goto failed_2;
 180
 181         iobuf->dr_filter = filter;
 182         init_waitqueue_head(&iobuf->dr_wait);
 183         atomic_set(&iobuf->dr_numreqs, 0);
 184         spin_lock_init(&iobuf->dr_lock);
 185         iobuf->dr_max_pages = num_pages;
 186         iobuf->dr_npages = 0;
 187         iobuf->dr_error = 0;
 188
 189         RETURN(iobuf);
 190
 191  failed_2:
 192         OBD_FREE(iobuf->dr_pages,
 193                  num_pages * sizeof(*iobuf->dr_pages));
 194  failed_1:
 195         OBD_FREE(iobuf, sizeof(*iobuf));
 196  failed_0:
 197         RETURN(ERR_PTR(-ENOMEM));
 198 }
 199
 200 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
 201 {
 202         iobuf->dr_npages = 0;
 203         iobuf->dr_error = 0;
 204         atomic_set(&iobuf->dr_numreqs, 0);
 205 }
 206
 207 void filter_free_iobuf(struct filter_iobuf *iobuf)
 208 {
 209         int num_pages = iobuf->dr_max_pages;
 210
 211         filter_clear_iobuf(iobuf);
 212
 213         OBD_FREE(iobuf->dr_blocks,
 214                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 215         OBD_FREE(iobuf->dr_pages,
 216                  num_pages * sizeof(*iobuf->dr_pages));
 217         OBD_FREE_PTR(iobuf);
 218 }
 219
 220 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 221                       struct obd_trans_info *oti)
 222 {
 223         int thread_id = oti ? oti->oti_thread_id : -1;
 224
 225         if (unlikely(thread_id < 0)) {
 226                 filter_free_iobuf(iobuf);
 227                 return;
 228         }
 229
 230         LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
 231                  "iobuf mismatch for thread %d: pool %p iobuf %p\n",
 232                  thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
 233         filter_clear_iobuf(iobuf);
 234 }
 235
 236 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 237                           struct inode *inode, struct page *page)
 238 {
 239         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 240         iobuf->dr_pages[iobuf->dr_npages++] = page;
 241
 242         return 0;
 243 }
 244
 245 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 246                   struct filter_iobuf *iobuf, int rw)
 247 {
 248         struct obd_device *obd = exp->exp_obd;
 249         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 250         struct page  **pages = iobuf->dr_pages;
 251         int            npages = iobuf->dr_npages;
 252         unsigned long *blocks = iobuf->dr_blocks;
 253         int            total_blocks = npages * blocks_per_page;
 254         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 255         unsigned int   blocksize = inode->i_sb->s_blocksize;
 256         struct bio    *bio = NULL;
 257         int            frags = 0;
 258         unsigned long  start_time = jiffies;
 259         struct page   *page;
 260         unsigned int   page_offset;
 261         sector_t       sector;
 262         int            nblocks;
 263         int            block_idx;
 264         int            page_idx;
 265         int            i;
 266         int            rc = 0;
 267         ENTRY;
 268
 269         LASSERT(iobuf->dr_npages == npages);
 270         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 271
 272         for (page_idx = 0, block_idx = 0;
 273              page_idx < npages;
 274              page_idx++, block_idx += blocks_per_page) {
 275
 276                 page = pages[page_idx];
 277                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 278
 279                 for (i = 0, page_offset = 0;
 280                      i < blocks_per_page;
 281                      i += nblocks, page_offset += blocksize * nblocks) {
 282
 283                         nblocks = 1;
 284
 285                         if (blocks[block_idx + i] == 0) {  /* hole */
 286                                 LASSERT(rw == OBD_BRW_READ);
 287                                 memset(kmap(page) + page_offset, 0, blocksize);
 288                                 kunmap(page);
 289                                 continue;
 290                         }
 291
 292                         sector = blocks[block_idx + i] << sector_bits;
 293
 294                         /* Additional contiguous file blocks? */
 295                         while (i + nblocks < blocks_per_page &&
 296                                (sector + nblocks*(blocksize>>9)) ==
 297                                (blocks[block_idx + i + nblocks] << sector_bits))
 298                                 nblocks++;
 299
 300                         if (bio != NULL &&
 301                             can_be_merged(bio, sector) &&
 302                             bio_add_page(bio, page,
 303                                          blocksize * nblocks, page_offset) != 0)
 304                                 continue;       /* added this frag OK */
 305
 306                         if (bio != NULL) {
 307                                 request_queue_t *q =
 308                                         bdev_get_queue(bio->bi_bdev);
 309
 310                                 /* Dang! I have to fragment this I/O */
 311                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 312                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 313                                        bio->bi_size,
 314                                        bio->bi_vcnt, bio->bi_max_vecs,
 315                                        bio->bi_size >> 9, q->max_sectors,
 316                                        bio_phys_segments(q, bio),
 317                                        q->max_phys_segments,
 318                                        bio_hw_segments(q, bio),
 319                                        q->max_hw_segments);
 320
 321                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 322                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 323                                 if (rc < 0) {
 324                                         CERROR("Can't send bio: %d\n", rc);
 325                                         record_finish_io(iobuf, rw, rc);
 326                                         goto out;
 327                                 }
 328                                 frags++;
 329                         }
 330
 331                         /* allocate new bio */
 332                         bio = bio_alloc(GFP_NOIO,
 333                                         (npages - page_idx) * blocks_per_page);
 334                         if (bio == NULL) {
 335                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 336                                        (npages - page_idx), blocks_per_page,
 337                                        (npages - page_idx) * blocks_per_page);
 338                                 rc = -ENOMEM;
 339                                 goto out;
 340                         }
 341
 342                         bio->bi_bdev = inode->i_sb->s_bdev;
 343                         bio->bi_sector = sector;
 344                         bio->bi_end_io = dio_complete_routine;
 345                         bio->bi_private = iobuf;
 346
 347                         rc = bio_add_page(bio, page,
 348                                           blocksize * nblocks, page_offset);
 349                         LASSERT (rc != 0);
 350                 }
 351         }
 352
 353         if (bio != NULL) {
 354                 record_start_io(iobuf, rw, bio->bi_size, exp);
 355                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 356                 if (rc >= 0) {
 357                         frags++;
 358                         rc = 0;
 359                 } else {
 360                         CERROR("Can't send bio: %d\n", rc);
 361                         record_finish_io(iobuf, rw, rc);
 362                 }
 363         }
 364
 365  out:
 366         wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0);
 367
 368         if (rw == OBD_BRW_READ) {
 369                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_R_DIO_FRAGS],
 370                                  frags);
 371                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_DIO_FRAGS],
 372                                  frags);
 373                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_R_IO_TIME],
 374                                       jiffies - start_time);
 375                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_IO_TIME], jiffies - start_time);
 376         } else {
 377                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_W_DIO_FRAGS],
 378                                  frags);
 379                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_DIO_FRAGS],
 380                                  frags);
 381                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_W_IO_TIME],
 382                                       jiffies - start_time);
 383                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_IO_TIME], jiffies - start_time);
 384         }
 385
 386         if (rc == 0)
 387                 rc = iobuf->dr_error;
 388         RETURN(rc);
 389 }
 390
 391 /* These are our hacks to keep our directio/bh IO coherent with ext3's
 392  * page cache use.  Most notably ext3 reads file data into the page
 393  * cache when it is zeroing the tail of partial-block truncates and
 394  * leaves it there, sometimes generating io from it at later truncates.
 395  * This removes the partial page and its buffers from the page cache,
 396  * so it should only ever cause a wait in rare cases, as otherwise we
 397  * always do full-page IO to the OST.
 398  *
 399  * The call to truncate_complete_page() will call journal_invalidatepage()
 400  * to free the buffers and drop the page from cache.  The buffers should
 401  * not be dirty, because we already called fdatasync/fdatawait on them.
 402  */
 403 static int filter_sync_inode_data(struct inode *inode, int locked)
 404 {
 405         int rc = 0;
 406
 407         /* This is nearly do_fsync(), without the waiting on the inode */
 408         /* XXX: in 2.6.16 (at least) we don't need to hold i_mutex over
 409          * filemap_fdatawrite() and filemap_fdatawait(), so we may no longer
 410          * need this lock here at all. */
 411         if (!locked)
 412                 LOCK_INODE_MUTEX(inode);
 413         if (inode->i_mapping->nrpages) {
 414 #ifdef PF_SYNCWRITE
 415                 current->flags |= PF_SYNCWRITE;
 416 #endif
 417                 rc = filemap_fdatawrite(inode->i_mapping);
 418                 if (rc == 0)
 419                         rc = filemap_fdatawait(inode->i_mapping);
 420 #ifdef PF_SYNCWRITE
 421                 current->flags &= ~PF_SYNCWRITE;
 422 #endif
 423         }
 424         if (!locked)
 425                 UNLOCK_INODE_MUTEX(inode);
 426
 427         return rc;
 428 }
 429 /* Clear pages from the mapping before we do direct IO to that offset.
 430  * Now that the only source of such pages in the truncate path flushes
 431  * these pages to disk and then discards them, this is error condition.
 432  * If add back read cache this will happen again.  This could be disabled
 433  * until that time if we never see the below error. */
 434 static int filter_clear_page_cache(struct inode *inode,
 435                                    struct filter_iobuf *iobuf)
 436 {
 437         struct page *page;
 438         int i, rc;
 439
 440         rc = filter_sync_inode_data(inode, 0);
 441         if (rc != 0)
 442                 RETURN(rc);
 443
 444         /* be careful to call this after fsync_inode_data_buffers has waited
 445          * for IO to complete before we evict it from the cache */
 446         for (i = 0; i < iobuf->dr_npages; i++) {
 447                 page = find_lock_page(inode->i_mapping,
 448                                       iobuf->dr_pages[i]->index);
 449                 if (page == NULL)
 450                         continue;
 451                 if (page->mapping != NULL) {
 452                         CERROR("page %lu (%d/%d) in page cache during write!\n",
 453                                page->index, i, iobuf->dr_npages);
 454                         wait_on_page_writeback(page);
 455                         ll_truncate_complete_page(page);
 456                 }
 457
 458                 unlock_page(page);
 459                 page_cache_release(page);
 460         }
 461
 462         return 0;
 463 }
 464
 465 int filter_clear_truncated_page(struct inode *inode)
 466 {
 467         struct page *page;
 468         int rc;
 469
 470         /* Truncate on page boundary, so nothing to flush? */
 471         if (!(i_size_read(inode) & ~CFS_PAGE_MASK))
 472                 return 0;
 473
 474         rc = filter_sync_inode_data(inode, 1);
 475         if (rc != 0)
 476                 RETURN(rc);
 477
 478         /* be careful to call this after fsync_inode_data_buffers has waited
 479          * for IO to complete before we evict it from the cache */
 480         page = find_lock_page(inode->i_mapping,
 481                               i_size_read(inode) >> CFS_PAGE_SHIFT);
 482         if (page) {
 483                 if (page->mapping != NULL) {
 484                         wait_on_page_writeback(page);
 485                         ll_truncate_complete_page(page);
 486                 }
 487                 unlock_page(page);
 488                 page_cache_release(page);
 489         }
 490
 491         return 0;
 492 }
 493
 494 /* Must be called with i_mutex taken for writes; this will drop it */
 495 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 496                      struct obd_export *exp, struct iattr *attr,
 497                      struct obd_trans_info *oti, void **wait_handle)
 498 {
 499         struct obd_device *obd = exp->exp_obd;
 500         struct inode *inode = dchild->d_inode;
 501         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 502         int rc, rc2, create;
 503         struct semaphore *sem;
 504         ENTRY;
 505
 506         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 507                  iobuf->dr_npages, iobuf->dr_max_pages);
 508         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 509
 510         if (rw == OBD_BRW_READ) {
 511                 if (iobuf->dr_npages == 0)
 512                         RETURN(0);
 513                 create = 0;
 514                 sem = NULL;
 515         } else {
 516                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 517                 LASSERT(iobuf->dr_npages > 0);
 518                 create = 1;
 519                 sem = &obd->u.filter.fo_alloc_lock;
 520
 521                 lquota_enforce(filter_quota_interface_ref, obd, iobuf->dr_ignore_quota);
 522         }
 523
 524         rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 525                                     iobuf->dr_npages, iobuf->dr_blocks,
 526                                     obdfilter_created_scratchpad, create, sem);
 527
 528         if (rw == OBD_BRW_WRITE) {
 529                 if (rc == 0) {
 530                         filter_tally(exp, iobuf->dr_pages,
 531                                      iobuf->dr_npages, iobuf->dr_blocks,
 532                                      blocks_per_page, 1);
 533                         if (attr->ia_size > i_size_read(inode))
 534                                 attr->ia_valid |= ATTR_SIZE;
 535                         rc = fsfilt_setattr(obd, dchild,
 536                                             oti->oti_handle, attr, 0);
 537                 }
 538
 539                 UNLOCK_INODE_MUTEX(inode);
 540
 541                 rc2 = filter_finish_transno(exp, oti, 0, 0);
 542                 if (rc2 != 0) {
 543                         CERROR("can't close transaction: %d\n", rc2);
 544                         if (rc == 0)
 545                                 rc = rc2;
 546                 }
 547
 548                 rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
 549                                           wait_handle);
 550                 if (rc == 0)
 551                         rc = rc2;
 552                 if (rc != 0)
 553                         RETURN(rc);
 554         } else if (rc == 0) {
 555                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 556                              iobuf->dr_blocks, blocks_per_page, 0);
 557         }
 558
 559         rc = filter_clear_page_cache(inode, iobuf);
 560         if (rc != 0)
 561                 RETURN(rc);
 562
 563         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 564 }
 565
 566 /* See if there are unallocated parts in given file region */
 567 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 568 {
 569         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 570                 inode->i_mapping->a_ops->bmap;
 571         int j;
 572
 573         /* We can't know if we are overwriting or not */
 574         if (fs_bmap == NULL)
 575                 return 0;
 576
 577         offset >>= inode->i_blkbits;
 578         len >>= inode->i_blkbits;
 579
 580         for (j = 0; j <= len; j++)
 581                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 582                         return 0;
 583
 584         return 1;
 585 }
 586
 587 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 588                           int objcount, struct obd_ioobj *obj, int niocount,
 589                           struct niobuf_local *res, struct obd_trans_info *oti,
 590                           int rc)
 591 {
 592         struct niobuf_local *lnb;
 593         struct filter_iobuf *iobuf = NULL;
 594         struct lvfs_run_ctxt saved;
 595         struct fsfilt_objinfo fso;
 596         struct iattr iattr = { 0 };
 597         struct inode *inode = NULL;
 598         unsigned long now = jiffies;
 599         int i, err, cleanup_phase = 0;
 600         struct obd_device *obd = exp->exp_obd;
 601         void *wait_handle;
 602         int   total_size = 0, rc2;
 603         unsigned int qcids[MAXQUOTAS] = {0, 0};
 604         ENTRY;
 605
 606         LASSERT(oti != NULL);
 607         LASSERT(objcount == 1);
 608         LASSERT(current->journal_info == NULL);
 609
 610         if (rc != 0)
 611                 GOTO(cleanup, rc);
 612
 613         /* Unfortunately, if quota master is too busy to handle the
 614          * pre-dqacq in time and quota hash on ost is used up, we
 615          * have to wait for the completion of in flight dqacq/dqrel,
 616          * then try again */
 617         if ((rc2 = lquota_chkquota(filter_quota_interface_ref, obd, oa->o_uid,
 618                                    oa->o_gid, niocount)) == QUOTA_RET_ACQUOTA) {
 619                 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_HOLD_WRITE_RPC, 90);
 620                 lquota_acquire(filter_quota_interface_ref, obd, oa->o_uid,
 621                                oa->o_gid);
 622         }
 623
 624         if (rc2 < 0) {
 625                 rc = rc2;
 626                 GOTO(cleanup, rc);
 627         }
 628
 629         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 630         if (IS_ERR(iobuf))
 631                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 632         cleanup_phase = 1;
 633
 634         fso.fso_dentry = res->dentry;
 635         fso.fso_bufcnt = obj->ioo_bufcnt;
 636         inode = res->dentry->d_inode;
 637
 638         iobuf->dr_ignore_quota = 0;
 639         for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
 640                 loff_t this_size;
 641
 642                 /* If overwriting an existing block, we don't need a grant */
 643                 if (!(lnb->flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC &&
 644                     filter_range_is_mapped(inode, lnb->offset, lnb->len))
 645                         lnb->rc = 0;
 646
 647                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 648                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 649                         continue;
 650                 }
 651
 652                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 653                 LASSERT (err == 0);
 654
 655                 total_size += lnb->len;
 656
 657                 /* we expect these pages to be in offset order, but we'll
 658                  * be forgiving */
 659                 this_size = lnb->offset + lnb->len;
 660                 if (this_size > iattr.ia_size)
 661                         iattr.ia_size = this_size;
 662
 663                 /* if one page is a write-back page from client cache, or it's
 664                  * written by root, then mark the whole io request as ignore
 665                  * quota request */
 666                 if (lnb->flags & (OBD_BRW_FROM_GRANT | OBD_BRW_NOQUOTA))
 667                         iobuf->dr_ignore_quota = 1;
 668         }
 669
 670         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 671         cleanup_phase = 2;
 672
 673         DQUOT_INIT(inode);
 674
 675         LOCK_INODE_MUTEX(inode);
 676         fsfilt_check_slow(obd, now, obd_timeout, "i_mutex");
 677         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 678                                            oti);
 679         if (IS_ERR(oti->oti_handle)) {
 680                 UNLOCK_INODE_MUTEX(inode);
 681                 rc = PTR_ERR(oti->oti_handle);
 682                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 683                        "error starting transaction: rc = %d\n", rc);
 684                 oti->oti_handle = NULL;
 685                 GOTO(cleanup, rc);
 686         }
 687         /* have to call fsfilt_commit() from this point on */
 688
 689         fsfilt_check_slow(obd, now, obd_timeout, "brw_start");
 690
 691         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 692
 693         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 694          * then we will accept the UID+GID if sent by the client for
 695          * initializing the ownership of this inode.  We only allow this to
 696          * happen once (so clear these bits) and later only allow setattr. */
 697         if (inode->i_mode & S_ISUID)
 698                 i |= OBD_MD_FLUID;
 699         if (inode->i_mode & S_ISGID)
 700                 i |= OBD_MD_FLGID;
 701
 702         iattr_from_obdo(&iattr, oa, i);
 703         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 704                 unsigned int save;
 705
 706                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 707                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 708
 709                 cap_raise(current->cap_effective, CAP_SYS_RESOURCE);
 710
 711                 iattr.ia_valid |= ATTR_MODE;
 712                 iattr.ia_mode = inode->i_mode;
 713                 if (iattr.ia_valid & ATTR_UID)
 714                         iattr.ia_mode &= ~S_ISUID;
 715                 if (iattr.ia_valid & ATTR_GID)
 716                         iattr.ia_mode &= ~S_ISGID;
 717
 718                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 719
 720                 /* To avoid problems with quotas, UID and GID must be set
 721                  * in the inode before filter_direct_io() - see bug 10357. */
 722                 save = iattr.ia_valid;
 723                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 724                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle, &iattr, 0);
 725                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 726                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 727                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 728         }
 729
 730         /* filter_direct_io drops i_mutex */
 731         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 732                               oti, &wait_handle);
 733         if (rc == 0)
 734                 obdo_from_inode(oa, inode,
 735                                 FILTER_VALID_FLAGS |OBD_MD_FLUID |OBD_MD_FLGID);
 736         else
 737                 obdo_from_inode(oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
 738
 739         lquota_getflag(filter_quota_interface_ref, obd, oa);
 740
 741         fsfilt_check_slow(obd, now, obd_timeout, "direct_io");
 742
 743         err = fsfilt_commit_wait(obd, inode, wait_handle);
 744         if (err) {
 745                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 746                 rc = err;
 747         }
 748
 749         if (obd->obd_replayable && !rc)
 750                 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
 751                          "oti_transno "LPU64" last_committed "LPU64"\n",
 752                          oti->oti_transno, obd->obd_last_committed);
 753
 754         fsfilt_check_slow(obd, now, obd_timeout, "commitrw commit");
 755
 756 cleanup:
 757         filter_grant_commit(exp, niocount, res);
 758
 759         switch (cleanup_phase) {
 760         case 2:
 761                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 762                 LASSERT(current->journal_info == NULL);
 763         case 1:
 764                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 765         case 0:
 766                 /*
 767                  * lnb->page automatically returns back into per-thread page
 768                  * pool (bug 5137)
 769                  */
 770                 f_dput(res->dentry);
 771         }
 772
 773         /* trigger quota pre-acquire */
 774         qcids[USRQUOTA] = oa->o_uid;
 775         qcids[GRPQUOTA] = oa->o_gid;
 776         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 777                             FSFILT_OP_CREATE);
 778         CDEBUG(err ? D_ERROR : D_QUOTA,
 779                "filter adjust qunit! (rc:%d)\n", err);
 780
 781         RETURN(rc);
 782 }