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], 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 #ifdef HAVE_PAGE_CONSTANT
 108         struct bio_vec *bvl;
 109         int i;
 110 #endif
 111
 112         /* CAVEAT EMPTOR: possibly in IRQ context
 113          * DO NOT record procfs stats here!!! */
 114
 115         if (bio->bi_size)                       /* Not complete */
 116                 return 1;
 117
 118         if (iobuf == NULL) {
 119                 CERROR("***** bio->bi_private is NULL!  This should never "
 120                        "happen.  Normally, I would crash here, but instead I "
 121                        "will dump the bio contents to the console.  Please "
 122                        "report this to CFS, along with any interesting "
 123                        "messages leading up to this point (like SCSI errors, "
 124                        "perhaps).  Because bi_private is NULL, I can't wake up "
 125                        "the thread that initiated this I/O -- so you will "
 126                        "probably have to reboot this node.\n");
 127                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 128                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 129                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 130                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 131                        bio->bi_end_io, atomic_read(&bio->bi_cnt),
 132                        bio->bi_private);
 133                 return 0;
 134         }
 135
 136 #ifdef HAVE_PAGE_CONSTANT
 137         bio_for_each_segment(bvl, bio, i)
 138                 ClearPageConstant(bvl->bv_page);
 139 #endif
 140
 141         spin_lock_irqsave(&iobuf->dr_lock, flags);
 142         if (iobuf->dr_error == 0)
 143                 iobuf->dr_error = error;
 144         spin_unlock_irqrestore(&iobuf->dr_lock, flags);
 145
 146         record_finish_io(iobuf, test_bit(BIO_RW, &bio->bi_rw) ?
 147                          OBD_BRW_WRITE : OBD_BRW_READ, error);
 148
 149         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 150          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 151          * mempool when serious on-disk fragmentation was encountered,
 152          * deadlocking the OST.  The bios are now released as soon as complete
 153          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 154         bio_put(bio);
 155         return 0;
 156 }
 157
 158 static int can_be_merged(struct bio *bio, sector_t sector)
 159 {
 160         unsigned int size;
 161
 162         if (!bio)
 163                 return 0;
 164
 165         size = bio->bi_size >> 9;
 166         return bio->bi_sector + size == sector ? 1 : 0;
 167 }
 168
 169 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 170                                         int rw, int num_pages)
 171 {
 172         struct filter_iobuf *iobuf;
 173
 174         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 175
 176         OBD_ALLOC(iobuf, sizeof(*iobuf));
 177         if (iobuf == NULL)
 178                 goto failed_0;
 179
 180         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 181         if (iobuf->dr_pages == NULL)
 182                 goto failed_1;
 183
 184         OBD_ALLOC(iobuf->dr_blocks,
 185                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 186         if (iobuf->dr_blocks == NULL)
 187                 goto failed_2;
 188
 189         iobuf->dr_filter = filter;
 190         init_waitqueue_head(&iobuf->dr_wait);
 191         atomic_set(&iobuf->dr_numreqs, 0);
 192         spin_lock_init(&iobuf->dr_lock);
 193         iobuf->dr_max_pages = num_pages;
 194         iobuf->dr_npages = 0;
 195         iobuf->dr_error = 0;
 196
 197         RETURN(iobuf);
 198
 199  failed_2:
 200         OBD_FREE(iobuf->dr_pages,
 201                  num_pages * sizeof(*iobuf->dr_pages));
 202  failed_1:
 203         OBD_FREE(iobuf, sizeof(*iobuf));
 204  failed_0:
 205         RETURN(ERR_PTR(-ENOMEM));
 206 }
 207
 208 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
 209 {
 210         iobuf->dr_npages = 0;
 211         iobuf->dr_error = 0;
 212         atomic_set(&iobuf->dr_numreqs, 0);
 213 }
 214
 215 void filter_free_iobuf(struct filter_iobuf *iobuf)
 216 {
 217         int num_pages = iobuf->dr_max_pages;
 218
 219         filter_clear_iobuf(iobuf);
 220
 221         OBD_FREE(iobuf->dr_blocks,
 222                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 223         OBD_FREE(iobuf->dr_pages,
 224                  num_pages * sizeof(*iobuf->dr_pages));
 225         OBD_FREE_PTR(iobuf);
 226 }
 227
 228 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 229                       struct obd_trans_info *oti)
 230 {
 231         int thread_id = oti ? oti->oti_thread_id : -1;
 232
 233         if (unlikely(thread_id < 0)) {
 234                 filter_free_iobuf(iobuf);
 235                 return;
 236         }
 237
 238         LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
 239                  "iobuf mismatch for thread %d: pool %p iobuf %p\n",
 240                  thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
 241         filter_clear_iobuf(iobuf);
 242 }
 243
 244 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 245                           struct inode *inode, struct page *page)
 246 {
 247         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 248         iobuf->dr_pages[iobuf->dr_npages++] = page;
 249
 250         return 0;
 251 }
 252
 253 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 254                   struct filter_iobuf *iobuf, int rw)
 255 {
 256         struct obd_device *obd = exp->exp_obd;
 257         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 258         struct page  **pages = iobuf->dr_pages;
 259         int            npages = iobuf->dr_npages;
 260         unsigned long *blocks = iobuf->dr_blocks;
 261         int            total_blocks = npages * blocks_per_page;
 262         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 263         unsigned int   blocksize = inode->i_sb->s_blocksize;
 264         struct bio    *bio = NULL;
 265         int            frags = 0;
 266         unsigned long  start_time = jiffies;
 267         struct page   *page;
 268         unsigned int   page_offset;
 269         sector_t       sector;
 270         int            nblocks;
 271         int            block_idx;
 272         int            page_idx;
 273         int            i;
 274         int            rc = 0;
 275         ENTRY;
 276
 277         LASSERT(iobuf->dr_npages == npages);
 278         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 279
 280         for (page_idx = 0, block_idx = 0;
 281              page_idx < npages;
 282              page_idx++, block_idx += blocks_per_page) {
 283
 284                 page = pages[page_idx];
 285                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 286
 287                 for (i = 0, page_offset = 0;
 288                      i < blocks_per_page;
 289                      i += nblocks, page_offset += blocksize * nblocks) {
 290
 291                         nblocks = 1;
 292
 293                         if (blocks[block_idx + i] == 0) {  /* hole */
 294                                 LASSERT(rw == OBD_BRW_READ);
 295                                 memset(kmap(page) + page_offset, 0, blocksize);
 296                                 kunmap(page);
 297                                 continue;
 298                         }
 299
 300                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 301
 302                         /* Additional contiguous file blocks? */
 303                         while (i + nblocks < blocks_per_page &&
 304                                (sector + (nblocks << sector_bits)) ==
 305                                ((sector_t)blocks[block_idx + i + nblocks] <<
 306                                 sector_bits))
 307                                 nblocks++;
 308
 309 #ifdef HAVE_PAGE_CONSTANT
 310                         /* I only set the page to be constant only if it
 311                          * is mapped to a contiguous underlying disk block(s).
 312                          * It will then make sure the corresponding device
 313                          * cache of raid5 will be overwritten by this page.
 314                          * - jay */
 315                         if ((rw == OBD_BRW_WRITE) &&
 316                             (nblocks == blocks_per_page) &&
 317                             mapping_cap_page_constant_write(inode->i_mapping))
 318                                 SetPageConstant(page);
 319 #endif
 320
 321                         if (bio != NULL &&
 322                             can_be_merged(bio, sector) &&
 323                             bio_add_page(bio, page,
 324                                          blocksize * nblocks, page_offset) != 0)
 325                                 continue;       /* added this frag OK */
 326
 327                         if (bio != NULL) {
 328                                 request_queue_t *q =
 329                                         bdev_get_queue(bio->bi_bdev);
 330
 331                                 /* Dang! I have to fragment this I/O */
 332                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 333                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 334                                        bio->bi_size,
 335                                        bio->bi_vcnt, bio->bi_max_vecs,
 336                                        bio->bi_size >> 9, q->max_sectors,
 337                                        bio_phys_segments(q, bio),
 338                                        q->max_phys_segments,
 339                                        bio_hw_segments(q, bio),
 340                                        q->max_hw_segments);
 341
 342                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 343                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 344                                 if (rc < 0) {
 345                                         CERROR("Can't send bio: %d\n", rc);
 346                                         record_finish_io(iobuf, rw, rc);
 347                                         goto out;
 348                                 }
 349                                 frags++;
 350                         }
 351
 352                         /* allocate new bio */
 353                         bio = bio_alloc(GFP_NOIO,
 354                                         (npages - page_idx) * blocks_per_page);
 355                         if (bio == NULL) {
 356                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 357                                        (npages - page_idx), blocks_per_page,
 358                                        (npages - page_idx) * blocks_per_page);
 359                                 rc = -ENOMEM;
 360                                 goto out;
 361                         }
 362
 363                         bio->bi_bdev = inode->i_sb->s_bdev;
 364                         bio->bi_sector = sector;
 365                         bio->bi_end_io = dio_complete_routine;
 366                         bio->bi_private = iobuf;
 367
 368                         rc = bio_add_page(bio, page,
 369                                           blocksize * nblocks, page_offset);
 370                         LASSERT (rc != 0);
 371                 }
 372         }
 373
 374         if (bio != NULL) {
 375                 record_start_io(iobuf, rw, bio->bi_size, exp);
 376                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 377                 if (rc >= 0) {
 378                         frags++;
 379                         rc = 0;
 380                 } else {
 381                         CERROR("Can't send bio: %d\n", rc);
 382                         record_finish_io(iobuf, rw, rc);
 383                 }
 384         }
 385
 386  out:
 387         wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0);
 388
 389         if (rw == OBD_BRW_READ) {
 390                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_R_DIO_FRAGS], frags);
 391                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_DIO_FRAGS],
 392                                  frags);
 393                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_R_IO_TIME],
 394                                       jiffies - start_time);
 395                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_IO_TIME],
 396                                  jiffies - start_time);
 397                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 398                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_R_DIO_FRAGS],
 399                                          frags);
 400                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_R_IO_TIME],
 401                                               jiffies - start_time);
 402                 }
 403         } else {
 404                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_W_DIO_FRAGS], frags);
 405                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_DIO_FRAGS],
 406                                  frags);
 407                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_W_IO_TIME],
 408                                       jiffies - start_time);
 409                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_IO_TIME],
 410                                  jiffies - start_time);
 411                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 412                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_W_DIO_FRAGS],
 413                                          frags);
 414                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_W_IO_TIME],
 415                                               jiffies - start_time);
 416                 }
 417         }
 418
 419         if (rc == 0)
 420                 rc = iobuf->dr_error;
 421         RETURN(rc);
 422 }
 423
 424 /* These are our hacks to keep our directio/bh IO coherent with ext3's
 425  * page cache use.  Most notably ext3 reads file data into the page
 426  * cache when it is zeroing the tail of partial-block truncates and
 427  * leaves it there, sometimes generating io from it at later truncates.
 428  * This removes the partial page and its buffers from the page cache,
 429  * so it should only ever cause a wait in rare cases, as otherwise we
 430  * always do full-page IO to the OST.
 431  *
 432  * The call to truncate_complete_page() will call journal_invalidatepage()
 433  * to free the buffers and drop the page from cache.  The buffers should
 434  * not be dirty, because we already called fdatasync/fdatawait on them.
 435  */
 436 static int filter_sync_inode_data(struct inode *inode, int locked)
 437 {
 438         int rc = 0;
 439
 440         /* This is nearly do_fsync(), without the waiting on the inode */
 441         /* XXX: in 2.6.16 (at least) we don't need to hold i_mutex over
 442          * filemap_fdatawrite() and filemap_fdatawait(), so we may no longer
 443          * need this lock here at all. */
 444         if (!locked)
 445                 LOCK_INODE_MUTEX(inode);
 446         if (inode->i_mapping->nrpages) {
 447 #ifdef PF_SYNCWRITE
 448                 current->flags |= PF_SYNCWRITE;
 449 #endif
 450                 rc = filemap_fdatawrite(inode->i_mapping);
 451                 if (rc == 0)
 452                         rc = filemap_fdatawait(inode->i_mapping);
 453 #ifdef PF_SYNCWRITE
 454                 current->flags &= ~PF_SYNCWRITE;
 455 #endif
 456         }
 457         if (!locked)
 458                 UNLOCK_INODE_MUTEX(inode);
 459
 460         return rc;
 461 }
 462
 463 /* Clear pages from the mapping before we do direct IO to that offset.
 464  * Now that the only source of such pages in the truncate path flushes
 465  * these pages to disk and then discards them, this is error condition.
 466  * If add back read cache this will happen again.  This could be disabled
 467  * until that time if we never see the below error. */
 468 static int filter_clear_page_cache(struct inode *inode,
 469                                    struct filter_iobuf *iobuf)
 470 {
 471         struct page *page;
 472         int i, rc;
 473
 474         rc = filter_sync_inode_data(inode, 0);
 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         for (i = 0; i < iobuf->dr_npages; i++) {
 481                 page = find_lock_page(inode->i_mapping,
 482                                       iobuf->dr_pages[i]->index);
 483                 if (page == NULL)
 484                         continue;
 485                 if (page->mapping != NULL) {
 486                         CERROR("page %lu (%d/%d) in page cache during write!\n",
 487                                page->index, i, iobuf->dr_npages);
 488                         wait_on_page_writeback(page);
 489                         ll_truncate_complete_page(page);
 490                 }
 491
 492                 unlock_page(page);
 493                 page_cache_release(page);
 494         }
 495
 496         return 0;
 497 }
 498
 499 int filter_clear_truncated_page(struct inode *inode)
 500 {
 501         struct page *page;
 502         int rc;
 503
 504         /* Truncate on page boundary, so nothing to flush? */
 505         if (!(i_size_read(inode) & ~CFS_PAGE_MASK))
 506                 return 0;
 507
 508         rc = filter_sync_inode_data(inode, 1);
 509         if (rc != 0)
 510                 RETURN(rc);
 511
 512         /* be careful to call this after fsync_inode_data_buffers has waited
 513          * for IO to complete before we evict it from the cache */
 514         page = find_lock_page(inode->i_mapping,
 515                               i_size_read(inode) >> CFS_PAGE_SHIFT);
 516         if (page) {
 517                 if (page->mapping != NULL) {
 518                         wait_on_page_writeback(page);
 519                         ll_truncate_complete_page(page);
 520                 }
 521                 unlock_page(page);
 522                 page_cache_release(page);
 523         }
 524
 525         return 0;
 526 }
 527
 528 /* Must be called with i_mutex taken for writes; this will drop it */
 529 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 530                      struct obd_export *exp, struct iattr *attr,
 531                      struct obd_trans_info *oti, void **wait_handle)
 532 {
 533         struct obd_device *obd = exp->exp_obd;
 534         struct inode *inode = dchild->d_inode;
 535         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 536         int rc, rc2, create;
 537         struct semaphore *sem;
 538         ENTRY;
 539
 540         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 541                  iobuf->dr_npages, iobuf->dr_max_pages);
 542         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 543
 544         if (rw == OBD_BRW_READ) {
 545                 if (iobuf->dr_npages == 0)
 546                         RETURN(0);
 547                 create = 0;
 548                 sem = NULL;
 549         } else {
 550                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 551                 LASSERT(iobuf->dr_npages > 0);
 552                 create = 1;
 553                 sem = &obd->u.filter.fo_alloc_lock;
 554
 555                 lquota_enforce(filter_quota_interface_ref, obd,
 556                                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,wait_handle);
 584                 if (rc == 0)
 585                         rc = rc2;
 586                 if (rc != 0)
 587                         RETURN(rc);
 588         } else if (rc == 0) {
 589                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 590                              iobuf->dr_blocks, blocks_per_page, 0);
 591         }
 592
 593         rc = filter_clear_page_cache(inode, iobuf);
 594         if (rc != 0)
 595                 RETURN(rc);
 596
 597         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 598 }
 599
 600 /* See if there are unallocated parts in given file region */
 601 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 602 {
 603         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 604                 inode->i_mapping->a_ops->bmap;
 605         int j;
 606
 607         /* We can't know if we are overwriting or not */
 608         if (fs_bmap == NULL)
 609                 return 0;
 610
 611         offset >>= inode->i_blkbits;
 612         len >>= inode->i_blkbits;
 613
 614         for (j = 0; j <= len; j++)
 615                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 616                         return 0;
 617
 618         return 1;
 619 }
 620
 621 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 622                           int objcount, struct obd_ioobj *obj, int niocount,
 623                           struct niobuf_local *res, struct obd_trans_info *oti,
 624                           int rc)
 625 {
 626         struct niobuf_local *lnb;
 627         struct filter_iobuf *iobuf = NULL;
 628         struct lvfs_run_ctxt saved;
 629         struct fsfilt_objinfo fso;
 630         struct iattr iattr = { 0 };
 631         struct inode *inode = NULL;
 632         unsigned long now = jiffies;
 633         int i, err, cleanup_phase = 0;
 634         struct obd_device *obd = exp->exp_obd;
 635         void *wait_handle;
 636         int total_size = 0;
 637         int rec_pending = 0;
 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         /* we try to get enough quota to write here, and let ldiskfs
 649          * decide if it is out of quota or not b=14783 */
 650         lquota_chkquota(filter_quota_interface_ref, obd, oa->o_uid,
 651                         oa->o_gid, niocount, &rec_pending);
 652
 653         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 654         if (IS_ERR(iobuf))
 655                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 656         cleanup_phase = 1;
 657
 658         fso.fso_dentry = res->dentry;
 659         fso.fso_bufcnt = obj->ioo_bufcnt;
 660         inode = res->dentry->d_inode;
 661
 662         iobuf->dr_ignore_quota = 0;
 663         for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
 664                 loff_t this_size;
 665
 666                 /* If overwriting an existing block, we don't need a grant */
 667                 if (!(lnb->flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC &&
 668                     filter_range_is_mapped(inode, lnb->offset, lnb->len))
 669                         lnb->rc = 0;
 670
 671                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 672                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 673                         continue;
 674                 }
 675
 676                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 677                 LASSERT (err == 0);
 678
 679                 total_size += lnb->len;
 680
 681                 /* we expect these pages to be in offset order, but we'll
 682                  * be forgiving */
 683                 this_size = lnb->offset + lnb->len;
 684                 if (this_size > iattr.ia_size)
 685                         iattr.ia_size = this_size;
 686
 687                 /* if one page is a write-back page from client cache, or it's
 688                  * written by root, then mark the whole io request as ignore
 689                  * quota request */
 690                 if (lnb->flags & (OBD_BRW_FROM_GRANT | OBD_BRW_NOQUOTA))
 691                         iobuf->dr_ignore_quota = 1;
 692         }
 693
 694         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 695         cleanup_phase = 2;
 696
 697         DQUOT_INIT(inode);
 698
 699         LOCK_INODE_MUTEX(inode);
 700         fsfilt_check_slow(obd, now, "i_mutex");
 701         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 702                                            oti);
 703         if (IS_ERR(oti->oti_handle)) {
 704                 UNLOCK_INODE_MUTEX(inode);
 705                 rc = PTR_ERR(oti->oti_handle);
 706                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 707                        "error starting transaction: rc = %d\n", rc);
 708                 oti->oti_handle = NULL;
 709                 GOTO(cleanup, rc);
 710         }
 711         /* have to call fsfilt_commit() from this point on */
 712
 713         fsfilt_check_slow(obd, now, "brw_start");
 714
 715         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 716
 717         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 718          * then we will accept the UID+GID if sent by the client for
 719          * initializing the ownership of this inode.  We only allow this to
 720          * happen once (so clear these bits) and later only allow setattr. */
 721         if (inode->i_mode & S_ISUID)
 722                 i |= OBD_MD_FLUID;
 723         if (inode->i_mode & S_ISGID)
 724                 i |= OBD_MD_FLGID;
 725
 726         iattr_from_obdo(&iattr, oa, i);
 727         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 728                 unsigned int save;
 729
 730                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 731                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 732
 733                 cap_raise(current->cap_effective, CAP_SYS_RESOURCE);
 734
 735                 iattr.ia_valid |= ATTR_MODE;
 736                 iattr.ia_mode = inode->i_mode;
 737                 if (iattr.ia_valid & ATTR_UID)
 738                         iattr.ia_mode &= ~S_ISUID;
 739                 if (iattr.ia_valid & ATTR_GID)
 740                         iattr.ia_mode &= ~S_ISGID;
 741
 742                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 743
 744                 /* To avoid problems with quotas, UID and GID must be set
 745                  * in the inode before filter_direct_io() - see bug 10357. */
 746                 save = iattr.ia_valid;
 747                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 748                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle, &iattr, 0);
 749                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 750                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 751                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 752         }
 753
 754         /* filter_direct_io drops i_mutex */
 755         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 756                               oti, &wait_handle);
 757         if (rc == 0)
 758                 obdo_from_inode(oa, inode,
 759                                 FILTER_VALID_FLAGS |OBD_MD_FLUID |OBD_MD_FLGID);
 760         else
 761                 obdo_from_inode(oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
 762
 763         lquota_getflag(filter_quota_interface_ref, obd, oa);
 764
 765         fsfilt_check_slow(obd, now, "direct_io");
 766
 767         err = fsfilt_commit_wait(obd, inode, wait_handle);
 768         if (err) {
 769                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 770                 rc = err;
 771         }
 772
 773         if (obd->obd_replayable && !rc)
 774                 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
 775                          "oti_transno "LPU64" last_committed "LPU64"\n",
 776                          oti->oti_transno, obd->obd_last_committed);
 777
 778         fsfilt_check_slow(obd, now, "commitrw commit");
 779
 780 cleanup:
 781         if (rec_pending)
 782                 lquota_pending_commit(filter_quota_interface_ref, obd, oa->o_uid,
 783                                       oa->o_gid, niocount);
 784
 785         filter_grant_commit(exp, niocount, res);
 786
 787         switch (cleanup_phase) {
 788         case 2:
 789                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 790                 LASSERT(current->journal_info == NULL);
 791         case 1:
 792                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 793         case 0:
 794                 /*
 795                  * lnb->page automatically returns back into per-thread page
 796                  * pool (bug 5137)
 797                  */
 798                 f_dput(res->dentry);
 799         }
 800
 801         /* trigger quota pre-acquire */
 802         qcids[USRQUOTA] = oa->o_uid;
 803         qcids[GRPQUOTA] = oa->o_gid;
 804         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 805                             FSFILT_OP_CREATE);
 806         CDEBUG(err ? D_ERROR : D_QUOTA,
 807                "filter adjust qunit! (rc:%d)\n", err);
 808
 809         RETURN(rc);
 810 }