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 #ifdef HAVE_PAGE_CONSTANT
 110         struct bio_vec *bvl;
 111         int i;
 112 #endif
 113
 114         /* CAVEAT EMPTOR: possibly in IRQ context
 115          * DO NOT record procfs stats here!!! */
 116
 117         if (bio->bi_size)                       /* Not complete */
 118                 return 1;
 119
 120         if (iobuf == NULL) {
 121                 CERROR("***** bio->bi_private is NULL!  This should never "
 122                        "happen.  Normally, I would crash here, but instead I "
 123                        "will dump the bio contents to the console.  Please "
 124                        "report this to CFS, along with any interesting "
 125                        "messages leading up to this point (like SCSI errors, "
 126                        "perhaps).  Because bi_private is NULL, I can't wake up "
 127                        "the thread that initiated this I/O -- so you will "
 128                        "probably have to reboot this node.\n");
 129                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 130                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 131                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 132                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 133                        bio->bi_end_io, atomic_read(&bio->bi_cnt),
 134                        bio->bi_private);
 135                 return 0;
 136         }
 137
 138 #ifdef HAVE_PAGE_CONSTANT
 139         bio_for_each_segment(bvl, bio, i)
 140                 ClearPageConstant(bvl->bv_page);
 141 #endif
 142
 143         spin_lock_irqsave(&iobuf->dr_lock, flags);
 144         if (iobuf->dr_error == 0)
 145                 iobuf->dr_error = error;
 146         spin_unlock_irqrestore(&iobuf->dr_lock, flags);
 147
 148         record_finish_io(iobuf, test_bit(BIO_RW, &bio->bi_rw) ?
 149                          OBD_BRW_WRITE : OBD_BRW_READ, error);
 150
 151         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 152          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 153          * mempool when serious on-disk fragmentation was encountered,
 154          * deadlocking the OST.  The bios are now released as soon as complete
 155          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 156         bio_put(bio);
 157         return 0;
 158 }
 159
 160 static int can_be_merged(struct bio *bio, sector_t sector)
 161 {
 162         unsigned int size;
 163
 164         if (!bio)
 165                 return 0;
 166
 167         size = bio->bi_size >> 9;
 168         return bio->bi_sector + size == sector ? 1 : 0;
 169 }
 170
 171 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 172                                         int rw, int num_pages)
 173 {
 174         struct filter_iobuf *iobuf;
 175
 176         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 177
 178         OBD_ALLOC(iobuf, sizeof(*iobuf));
 179         if (iobuf == NULL)
 180                 goto failed_0;
 181
 182         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 183         if (iobuf->dr_pages == NULL)
 184                 goto failed_1;
 185
 186         OBD_ALLOC(iobuf->dr_blocks,
 187                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 188         if (iobuf->dr_blocks == NULL)
 189                 goto failed_2;
 190
 191         iobuf->dr_filter = filter;
 192         init_waitqueue_head(&iobuf->dr_wait);
 193         atomic_set(&iobuf->dr_numreqs, 0);
 194         spin_lock_init(&iobuf->dr_lock);
 195         iobuf->dr_max_pages = num_pages;
 196         iobuf->dr_npages = 0;
 197         iobuf->dr_error = 0;
 198
 199         RETURN(iobuf);
 200
 201  failed_2:
 202         OBD_FREE(iobuf->dr_pages,
 203                  num_pages * sizeof(*iobuf->dr_pages));
 204  failed_1:
 205         OBD_FREE(iobuf, sizeof(*iobuf));
 206  failed_0:
 207         RETURN(ERR_PTR(-ENOMEM));
 208 }
 209
 210 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
 211 {
 212         iobuf->dr_npages = 0;
 213         iobuf->dr_error = 0;
 214         atomic_set(&iobuf->dr_numreqs, 0);
 215 }
 216
 217 void filter_free_iobuf(struct filter_iobuf *iobuf)
 218 {
 219         int num_pages = iobuf->dr_max_pages;
 220
 221         filter_clear_iobuf(iobuf);
 222
 223         OBD_FREE(iobuf->dr_blocks,
 224                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 225         OBD_FREE(iobuf->dr_pages,
 226                  num_pages * sizeof(*iobuf->dr_pages));
 227         OBD_FREE_PTR(iobuf);
 228 }
 229
 230 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 231                       struct obd_trans_info *oti)
 232 {
 233         int thread_id = oti ? oti->oti_thread_id : -1;
 234
 235         if (unlikely(thread_id < 0)) {
 236                 filter_free_iobuf(iobuf);
 237                 return;
 238         }
 239
 240         LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
 241                  "iobuf mismatch for thread %d: pool %p iobuf %p\n",
 242                  thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
 243         filter_clear_iobuf(iobuf);
 244 }
 245
 246 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 247                           struct inode *inode, struct page *page)
 248 {
 249         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 250         iobuf->dr_pages[iobuf->dr_npages++] = page;
 251
 252         return 0;
 253 }
 254
 255 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 256                   struct filter_iobuf *iobuf, int rw)
 257 {
 258         struct obd_device *obd = exp->exp_obd;
 259         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 260         struct page  **pages = iobuf->dr_pages;
 261         int            npages = iobuf->dr_npages;
 262         unsigned long *blocks = iobuf->dr_blocks;
 263         int            total_blocks = npages * blocks_per_page;
 264         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 265         unsigned int   blocksize = inode->i_sb->s_blocksize;
 266         struct bio    *bio = NULL;
 267         int            frags = 0;
 268         unsigned long  start_time = jiffies;
 269         struct page   *page;
 270         unsigned int   page_offset;
 271         sector_t       sector;
 272         int            nblocks;
 273         int            block_idx;
 274         int            page_idx;
 275         int            i;
 276         int            rc = 0;
 277         ENTRY;
 278
 279         LASSERT(iobuf->dr_npages == npages);
 280         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 281
 282         for (page_idx = 0, block_idx = 0;
 283              page_idx < npages;
 284              page_idx++, block_idx += blocks_per_page) {
 285
 286                 page = pages[page_idx];
 287                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 288
 289                 for (i = 0, page_offset = 0;
 290                      i < blocks_per_page;
 291                      i += nblocks, page_offset += blocksize * nblocks) {
 292
 293                         nblocks = 1;
 294
 295                         if (blocks[block_idx + i] == 0) {  /* hole */
 296                                 LASSERT(rw == OBD_BRW_READ);
 297                                 memset(kmap(page) + page_offset, 0, blocksize);
 298                                 kunmap(page);
 299                                 continue;
 300                         }
 301
 302                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 303
 304                         /* Additional contiguous file blocks? */
 305                         while (i + nblocks < blocks_per_page &&
 306                                (sector + (nblocks << sector_bits)) ==
 307                                ((sector_t)blocks[block_idx + i + nblocks] <<
 308                                 sector_bits))
 309                                 nblocks++;
 310
 311 #ifdef HAVE_PAGE_CONSTANT
 312                         /* I only set the page to be constant only if it
 313                          * is mapped to a contiguous underlying disk block(s).
 314                          * It will then make sure the corresponding device
 315                          * cache of raid5 will be overwritten by this page.
 316                          * - jay */
 317                         if ((rw == OBD_BRW_WRITE) &&
 318                             (nblocks == blocks_per_page) &&
 319                             mapping_cap_page_constant_write(inode->i_mapping))
 320                                SetPageConstant(page);
 321 #endif
 322
 323                         if (bio != NULL &&
 324                             can_be_merged(bio, sector) &&
 325                             bio_add_page(bio, page,
 326                                          blocksize * nblocks, page_offset) != 0)
 327                                 continue;       /* added this frag OK */
 328
 329                         if (bio != NULL) {
 330                                 request_queue_t *q =
 331                                         bdev_get_queue(bio->bi_bdev);
 332
 333                                 /* Dang! I have to fragment this I/O */
 334                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 335                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 336                                        bio->bi_size,
 337                                        bio->bi_vcnt, bio->bi_max_vecs,
 338                                        bio->bi_size >> 9, q->max_sectors,
 339                                        bio_phys_segments(q, bio),
 340                                        q->max_phys_segments,
 341                                        bio_hw_segments(q, bio),
 342                                        q->max_hw_segments);
 343
 344                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 345                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 346                                 if (rc < 0) {
 347                                         CERROR("Can't send bio: %d\n", rc);
 348                                         record_finish_io(iobuf, rw, rc);
 349                                         goto out;
 350                                 }
 351                                 frags++;
 352                         }
 353
 354                         /* allocate new bio */
 355                         bio = bio_alloc(GFP_NOIO,
 356                                         (npages - page_idx) * blocks_per_page);
 357                         if (bio == NULL) {
 358                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 359                                        (npages - page_idx), blocks_per_page,
 360                                        (npages - page_idx) * blocks_per_page);
 361                                 rc = -ENOMEM;
 362                                 goto out;
 363                         }
 364
 365                         bio->bi_bdev = inode->i_sb->s_bdev;
 366                         bio->bi_sector = sector;
 367                         bio->bi_end_io = dio_complete_routine;
 368                         bio->bi_private = iobuf;
 369
 370                         rc = bio_add_page(bio, page,
 371                                           blocksize * nblocks, page_offset);
 372                         LASSERT (rc != 0);
 373                 }
 374         }
 375
 376         if (bio != NULL) {
 377                 record_start_io(iobuf, rw, bio->bi_size, exp);
 378                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 379                 if (rc >= 0) {
 380                         frags++;
 381                         rc = 0;
 382                 } else {
 383                         CERROR("Can't send bio: %d\n", rc);
 384                         record_finish_io(iobuf, rw, rc);
 385                 }
 386         }
 387
 388  out:
 389         wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0);
 390
 391         if (rw == OBD_BRW_READ) {
 392                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_R_DIO_FRAGS],
 393                                  frags);
 394                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_DIO_FRAGS],
 395                                  frags);
 396                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_R_IO_TIME],
 397                                       jiffies - start_time);
 398                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_IO_TIME], jiffies - start_time);
 399                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 400                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_R_DIO_FRAGS],
 401                                          frags);
 402                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_R_IO_TIME],
 403                                               jiffies - start_time);
 404                 }
 405         } else {
 406                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_W_DIO_FRAGS],
 407                                  frags);
 408                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_DIO_FRAGS],
 409                                  frags);
 410                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_W_IO_TIME],
 411                                       jiffies - start_time);
 412                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_IO_TIME], jiffies - start_time);
 413                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 414                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_W_DIO_FRAGS],
 415                                          frags);
 416                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_W_IO_TIME],
 417                                               jiffies - start_time);
 418                 }
 419         }
 420
 421         if (rc == 0)
 422                 rc = iobuf->dr_error;
 423         RETURN(rc);
 424 }
 425
 426 /* These are our hacks to keep our directio/bh IO coherent with ext3's
 427  * page cache use.  Most notably ext3 reads file data into the page
 428  * cache when it is zeroing the tail of partial-block truncates and
 429  * leaves it there, sometimes generating io from it at later truncates.
 430  * This removes the partial page and its buffers from the page cache,
 431  * so it should only ever cause a wait in rare cases, as otherwise we
 432  * always do full-page IO to the OST.
 433  *
 434  * The call to truncate_complete_page() will call journal_invalidatepage()
 435  * to free the buffers and drop the page from cache.  The buffers should
 436  * not be dirty, because we already called fdatasync/fdatawait on them.
 437  */
 438 static int filter_sync_inode_data(struct inode *inode, int locked)
 439 {
 440         int rc = 0;
 441
 442         /* This is nearly do_fsync(), without the waiting on the inode */
 443         /* XXX: in 2.6.16 (at least) we don't need to hold i_mutex over
 444          * filemap_fdatawrite() and filemap_fdatawait(), so we may no longer
 445          * need this lock here at all. */
 446         if (!locked)
 447                 LOCK_INODE_MUTEX(inode);
 448         if (inode->i_mapping->nrpages) {
 449 #ifdef PF_SYNCWRITE
 450                 current->flags |= PF_SYNCWRITE;
 451 #endif
 452                 rc = filemap_fdatawrite(inode->i_mapping);
 453                 if (rc == 0)
 454                         rc = filemap_fdatawait(inode->i_mapping);
 455 #ifdef PF_SYNCWRITE
 456                 current->flags &= ~PF_SYNCWRITE;
 457 #endif
 458         }
 459         if (!locked)
 460                 UNLOCK_INODE_MUTEX(inode);
 461
 462         return rc;
 463 }
 464 /* Clear pages from the mapping before we do direct IO to that offset.
 465  * Now that the only source of such pages in the truncate path flushes
 466  * these pages to disk and then discards them, this is error condition.
 467  * If add back read cache this will happen again.  This could be disabled
 468  * until that time if we never see the below error. */
 469 static int filter_clear_page_cache(struct inode *inode,
 470                                    struct filter_iobuf *iobuf)
 471 {
 472         struct page *page;
 473         int i, rc;
 474
 475         rc = filter_sync_inode_data(inode, 0);
 476         if (rc != 0)
 477                 RETURN(rc);
 478
 479         /* be careful to call this after fsync_inode_data_buffers has waited
 480          * for IO to complete before we evict it from the cache */
 481         for (i = 0; i < iobuf->dr_npages; i++) {
 482                 page = find_lock_page(inode->i_mapping,
 483                                       iobuf->dr_pages[i]->index);
 484                 if (page == NULL)
 485                         continue;
 486                 if (page->mapping != NULL) {
 487                         CERROR("page %lu (%d/%d) in page cache during write!\n",
 488                                page->index, i, iobuf->dr_npages);
 489                         wait_on_page_writeback(page);
 490                         ll_truncate_complete_page(page);
 491                 }
 492
 493                 unlock_page(page);
 494                 page_cache_release(page);
 495         }
 496
 497         return 0;
 498 }
 499
 500 int filter_clear_truncated_page(struct inode *inode)
 501 {
 502         struct page *page;
 503         int rc;
 504
 505         /* Truncate on page boundary, so nothing to flush? */
 506         if (!(i_size_read(inode) & ~CFS_PAGE_MASK))
 507                 return 0;
 508
 509         rc = filter_sync_inode_data(inode, 1);
 510         if (rc != 0)
 511                 RETURN(rc);
 512
 513         /* be careful to call this after fsync_inode_data_buffers has waited
 514          * for IO to complete before we evict it from the cache */
 515         page = find_lock_page(inode->i_mapping,
 516                               i_size_read(inode) >> CFS_PAGE_SHIFT);
 517         if (page) {
 518                 if (page->mapping != NULL) {
 519                         wait_on_page_writeback(page);
 520                         ll_truncate_complete_page(page);
 521                 }
 522                 unlock_page(page);
 523                 page_cache_release(page);
 524         }
 525
 526         return 0;
 527 }
 528
 529 /* Must be called with i_mutex taken for writes; this will drop it */
 530 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 531                      struct obd_export *exp, struct iattr *attr,
 532                      struct obd_trans_info *oti, void **wait_handle)
 533 {
 534         struct obd_device *obd = exp->exp_obd;
 535         struct inode *inode = dchild->d_inode;
 536         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 537         int rc, rc2, create;
 538         struct semaphore *sem;
 539         ENTRY;
 540
 541         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 542                  iobuf->dr_npages, iobuf->dr_max_pages);
 543         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 544
 545         if (rw == OBD_BRW_READ) {
 546                 if (iobuf->dr_npages == 0)
 547                         RETURN(0);
 548                 create = 0;
 549                 sem = NULL;
 550         } else {
 551                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 552                 LASSERT(iobuf->dr_npages > 0);
 553                 create = 1;
 554                 sem = &obd->u.filter.fo_alloc_lock;
 555
 556                 lquota_enforce(filter_quota_interface_ref, obd, iobuf->dr_ignore_quota);
 557         }
 558
 559         rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 560                                     iobuf->dr_npages, iobuf->dr_blocks,
 561                                     obdfilter_created_scratchpad, create, sem);
 562
 563         if (rw == OBD_BRW_WRITE) {
 564                 if (rc == 0) {
 565                         filter_tally(exp, iobuf->dr_pages,
 566                                      iobuf->dr_npages, iobuf->dr_blocks,
 567                                      blocks_per_page, 1);
 568                         if (attr->ia_size > i_size_read(inode))
 569                                 attr->ia_valid |= ATTR_SIZE;
 570                         rc = fsfilt_setattr(obd, dchild,
 571                                             oti->oti_handle, attr, 0);
 572                 }
 573
 574                 UNLOCK_INODE_MUTEX(inode);
 575
 576                 rc2 = filter_finish_transno(exp, oti, 0, 0);
 577                 if (rc2 != 0) {
 578                         CERROR("can't close transaction: %d\n", rc2);
 579                         if (rc == 0)
 580                                 rc = rc2;
 581                 }
 582
 583                 rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
 584                                           wait_handle);
 585                 if (rc == 0)
 586                         rc = rc2;
 587                 if (rc != 0)
 588                         RETURN(rc);
 589         } else if (rc == 0) {
 590                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 591                              iobuf->dr_blocks, blocks_per_page, 0);
 592         }
 593
 594         rc = filter_clear_page_cache(inode, iobuf);
 595         if (rc != 0)
 596                 RETURN(rc);
 597
 598         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 599 }
 600
 601 /* See if there are unallocated parts in given file region */
 602 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 603 {
 604         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 605                 inode->i_mapping->a_ops->bmap;
 606         int j;
 607
 608         /* We can't know if we are overwriting or not */
 609         if (fs_bmap == NULL)
 610                 return 0;
 611
 612         offset >>= inode->i_blkbits;
 613         len >>= inode->i_blkbits;
 614
 615         for (j = 0; j <= len; j++)
 616                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 617                         return 0;
 618
 619         return 1;
 620 }
 621
 622 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 623                           int objcount, struct obd_ioobj *obj, int niocount,
 624                           struct niobuf_local *res, struct obd_trans_info *oti,
 625                           int rc)
 626 {
 627         struct niobuf_local *lnb;
 628         struct filter_iobuf *iobuf = NULL;
 629         struct lvfs_run_ctxt saved;
 630         struct fsfilt_objinfo fso;
 631         struct iattr iattr = { 0 };
 632         struct inode *inode = NULL;
 633         unsigned long now = jiffies;
 634         int i, err, cleanup_phase = 0;
 635         struct obd_device *obd = exp->exp_obd;
 636         void *wait_handle;
 637         int   total_size = 0, rc2;
 638         unsigned int qcids[MAXQUOTAS] = {0, 0};
 639         ENTRY;
 640
 641         LASSERT(oti != NULL);
 642         LASSERT(objcount == 1);
 643         LASSERT(current->journal_info == NULL);
 644
 645         if (rc != 0)
 646                 GOTO(cleanup, rc);
 647
 648         /* Unfortunately, if quota master is too busy to handle the
 649          * pre-dqacq in time and quota hash on ost is used up, we
 650          * have to wait for the completion of in flight dqacq/dqrel,
 651          * then try again */
 652         if ((rc2 = lquota_chkquota(filter_quota_interface_ref, obd, oa->o_uid,
 653                                    oa->o_gid, niocount)) == QUOTA_RET_ACQUOTA) {
 654                 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_HOLD_WRITE_RPC, 90);
 655                 lquota_acquire(filter_quota_interface_ref, obd, oa->o_uid,
 656                                oa->o_gid);
 657         }
 658
 659         if (rc2 < 0) {
 660                 rc = rc2;
 661                 GOTO(cleanup, rc);
 662         }
 663
 664         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 665         if (IS_ERR(iobuf))
 666                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 667         cleanup_phase = 1;
 668
 669         fso.fso_dentry = res->dentry;
 670         fso.fso_bufcnt = obj->ioo_bufcnt;
 671         inode = res->dentry->d_inode;
 672
 673         iobuf->dr_ignore_quota = 0;
 674         for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
 675                 loff_t this_size;
 676
 677                 /* If overwriting an existing block, we don't need a grant */
 678                 if (!(lnb->flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC &&
 679                     filter_range_is_mapped(inode, lnb->offset, lnb->len))
 680                         lnb->rc = 0;
 681
 682                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 683                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 684                         continue;
 685                 }
 686
 687                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 688                 LASSERT (err == 0);
 689
 690                 total_size += lnb->len;
 691
 692                 /* we expect these pages to be in offset order, but we'll
 693                  * be forgiving */
 694                 this_size = lnb->offset + lnb->len;
 695                 if (this_size > iattr.ia_size)
 696                         iattr.ia_size = this_size;
 697
 698                 /* if one page is a write-back page from client cache, or it's
 699                  * written by root, then mark the whole io request as ignore
 700                  * quota request */
 701                 if (lnb->flags & (OBD_BRW_FROM_GRANT | OBD_BRW_NOQUOTA))
 702                         iobuf->dr_ignore_quota = 1;
 703         }
 704
 705         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 706         cleanup_phase = 2;
 707
 708         DQUOT_INIT(inode);
 709
 710         LOCK_INODE_MUTEX(inode);
 711         fsfilt_check_slow(obd, now, obd_timeout, "i_mutex");
 712         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 713                                            oti);
 714         if (IS_ERR(oti->oti_handle)) {
 715                 UNLOCK_INODE_MUTEX(inode);
 716                 rc = PTR_ERR(oti->oti_handle);
 717                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 718                        "error starting transaction: rc = %d\n", rc);
 719                 oti->oti_handle = NULL;
 720                 GOTO(cleanup, rc);
 721         }
 722         /* have to call fsfilt_commit() from this point on */
 723
 724         fsfilt_check_slow(obd, now, obd_timeout, "brw_start");
 725
 726         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 727
 728         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 729          * then we will accept the UID+GID if sent by the client for
 730          * initializing the ownership of this inode.  We only allow this to
 731          * happen once (so clear these bits) and later only allow setattr. */
 732         if (inode->i_mode & S_ISUID)
 733                 i |= OBD_MD_FLUID;
 734         if (inode->i_mode & S_ISGID)
 735                 i |= OBD_MD_FLGID;
 736
 737         iattr_from_obdo(&iattr, oa, i);
 738         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 739                 unsigned int save;
 740
 741                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 742                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 743
 744                 cap_raise(current->cap_effective, CAP_SYS_RESOURCE);
 745
 746                 iattr.ia_valid |= ATTR_MODE;
 747                 iattr.ia_mode = inode->i_mode;
 748                 if (iattr.ia_valid & ATTR_UID)
 749                         iattr.ia_mode &= ~S_ISUID;
 750                 if (iattr.ia_valid & ATTR_GID)
 751                         iattr.ia_mode &= ~S_ISGID;
 752
 753                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 754
 755                 /* To avoid problems with quotas, UID and GID must be set
 756                  * in the inode before filter_direct_io() - see bug 10357. */
 757                 save = iattr.ia_valid;
 758                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 759                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle, &iattr, 0);
 760                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 761                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 762                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 763         }
 764
 765         /* filter_direct_io drops i_mutex */
 766         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 767                               oti, &wait_handle);
 768         if (rc == 0)
 769                 obdo_from_inode(oa, inode,
 770                                 FILTER_VALID_FLAGS |OBD_MD_FLUID |OBD_MD_FLGID);
 771         else
 772                 obdo_from_inode(oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
 773
 774         lquota_getflag(filter_quota_interface_ref, obd, oa);
 775
 776         fsfilt_check_slow(obd, now, obd_timeout, "direct_io");
 777
 778         err = fsfilt_commit_wait(obd, inode, wait_handle);
 779         if (err) {
 780                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 781                 rc = err;
 782         }
 783
 784         if (obd->obd_replayable && !rc)
 785                 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
 786                          "oti_transno "LPU64" last_committed "LPU64"\n",
 787                          oti->oti_transno, obd->obd_last_committed);
 788
 789         fsfilt_check_slow(obd, now, obd_timeout, "commitrw commit");
 790
 791 cleanup:
 792         filter_grant_commit(exp, niocount, res);
 793
 794         switch (cleanup_phase) {
 795         case 2:
 796                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 797                 LASSERT(current->journal_info == NULL);
 798         case 1:
 799                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 800         case 0:
 801                 /*
 802                  * lnb->page automatically returns back into per-thread page
 803                  * pool (bug 5137)
 804                  */
 805                 f_dput(res->dentry);
 806         }
 807
 808         /* trigger quota pre-acquire */
 809         qcids[USRQUOTA] = oa->o_uid;
 810         qcids[GRPQUOTA] = oa->o_gid;
 811         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 812                             FSFILT_OP_CREATE);
 813         CDEBUG(err ? D_ERROR : D_QUOTA,
 814                "filter adjust qunit! (rc:%d)\n", err);
 815
 816         RETURN(rc);
 817 }