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 = (sector_t)blocks[block_idx + i] << sector_bits;
 293
 294                         /* Additional contiguous file blocks? */
 295                         while (i + nblocks < blocks_per_page &&
 296                                (sector + (nblocks << sector_bits)) ==
 297                                ((sector_t)blocks[block_idx + i + nblocks] <<
 298                                 sector_bits))
 299                                 nblocks++;
 300
 301                         if (bio != NULL &&
 302                             can_be_merged(bio, sector) &&
 303                             bio_add_page(bio, page,
 304                                          blocksize * nblocks, page_offset) != 0)
 305                                 continue;       /* added this frag OK */
 306
 307                         if (bio != NULL) {
 308                                 request_queue_t *q =
 309                                         bdev_get_queue(bio->bi_bdev);
 310
 311                                 /* Dang! I have to fragment this I/O */
 312                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 313                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 314                                        bio->bi_size,
 315                                        bio->bi_vcnt, bio->bi_max_vecs,
 316                                        bio->bi_size >> 9, q->max_sectors,
 317                                        bio_phys_segments(q, bio),
 318                                        q->max_phys_segments,
 319                                        bio_hw_segments(q, bio),
 320                                        q->max_hw_segments);
 321
 322                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 323                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 324                                 if (rc < 0) {
 325                                         CERROR("Can't send bio: %d\n", rc);
 326                                         record_finish_io(iobuf, rw, rc);
 327                                         goto out;
 328                                 }
 329                                 frags++;
 330                         }
 331
 332                         /* allocate new bio */
 333                         bio = bio_alloc(GFP_NOIO,
 334                                         (npages - page_idx) * blocks_per_page);
 335                         if (bio == NULL) {
 336                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 337                                        (npages - page_idx), blocks_per_page,
 338                                        (npages - page_idx) * blocks_per_page);
 339                                 rc = -ENOMEM;
 340                                 goto out;
 341                         }
 342
 343                         bio->bi_bdev = inode->i_sb->s_bdev;
 344                         bio->bi_sector = sector;
 345                         bio->bi_end_io = dio_complete_routine;
 346                         bio->bi_private = iobuf;
 347
 348                         rc = bio_add_page(bio, page,
 349                                           blocksize * nblocks, page_offset);
 350                         LASSERT (rc != 0);
 351                 }
 352         }
 353
 354         if (bio != NULL) {
 355                 record_start_io(iobuf, rw, bio->bi_size, exp);
 356                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 357                 if (rc >= 0) {
 358                         frags++;
 359                         rc = 0;
 360                 } else {
 361                         CERROR("Can't send bio: %d\n", rc);
 362                         record_finish_io(iobuf, rw, rc);
 363                 }
 364         }
 365
 366  out:
 367         wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0);
 368
 369         if (rw == OBD_BRW_READ) {
 370                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_R_DIO_FRAGS],
 371                                  frags);
 372                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_DIO_FRAGS],
 373                                  frags);
 374                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_R_IO_TIME],
 375                                       jiffies - start_time);
 376                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_IO_TIME], jiffies - start_time);
 377         } else {
 378                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_W_DIO_FRAGS],
 379                                  frags);
 380                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_DIO_FRAGS],
 381                                  frags);
 382                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_W_IO_TIME],
 383                                       jiffies - start_time);
 384                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_IO_TIME], jiffies - start_time);
 385         }
 386
 387         if (rc == 0)
 388                 rc = iobuf->dr_error;
 389         RETURN(rc);
 390 }
 391
 392 /* These are our hacks to keep our directio/bh IO coherent with ext3's
 393  * page cache use.  Most notably ext3 reads file data into the page
 394  * cache when it is zeroing the tail of partial-block truncates and
 395  * leaves it there, sometimes generating io from it at later truncates.
 396  * This removes the partial page and its buffers from the page cache,
 397  * so it should only ever cause a wait in rare cases, as otherwise we
 398  * always do full-page IO to the OST.
 399  *
 400  * The call to truncate_complete_page() will call journal_invalidatepage()
 401  * to free the buffers and drop the page from cache.  The buffers should
 402  * not be dirty, because we already called fdatasync/fdatawait on them.
 403  */
 404 static int filter_sync_inode_data(struct inode *inode, int locked)
 405 {
 406         int rc = 0;
 407
 408         /* This is nearly do_fsync(), without the waiting on the inode */
 409         /* XXX: in 2.6.16 (at least) we don't need to hold i_mutex over
 410          * filemap_fdatawrite() and filemap_fdatawait(), so we may no longer
 411          * need this lock here at all. */
 412         if (!locked)
 413                 LOCK_INODE_MUTEX(inode);
 414         if (inode->i_mapping->nrpages) {
 415 #ifdef PF_SYNCWRITE
 416                 current->flags |= PF_SYNCWRITE;
 417 #endif
 418                 rc = filemap_fdatawrite(inode->i_mapping);
 419                 if (rc == 0)
 420                         rc = filemap_fdatawait(inode->i_mapping);
 421 #ifdef PF_SYNCWRITE
 422                 current->flags &= ~PF_SYNCWRITE;
 423 #endif
 424         }
 425         if (!locked)
 426                 UNLOCK_INODE_MUTEX(inode);
 427
 428         return rc;
 429 }
 430 /* Clear pages from the mapping before we do direct IO to that offset.
 431  * Now that the only source of such pages in the truncate path flushes
 432  * these pages to disk and then discards them, this is error condition.
 433  * If add back read cache this will happen again.  This could be disabled
 434  * until that time if we never see the below error. */
 435 static int filter_clear_page_cache(struct inode *inode,
 436                                    struct filter_iobuf *iobuf)
 437 {
 438         struct page *page;
 439         int i, rc;
 440
 441         rc = filter_sync_inode_data(inode, 0);
 442         if (rc != 0)
 443                 RETURN(rc);
 444
 445         /* be careful to call this after fsync_inode_data_buffers has waited
 446          * for IO to complete before we evict it from the cache */
 447         for (i = 0; i < iobuf->dr_npages; i++) {
 448                 page = find_lock_page(inode->i_mapping,
 449                                       iobuf->dr_pages[i]->index);
 450                 if (page == NULL)
 451                         continue;
 452                 if (page->mapping != NULL) {
 453                         CERROR("page %lu (%d/%d) in page cache during write!\n",
 454                                page->index, i, iobuf->dr_npages);
 455                         wait_on_page_writeback(page);
 456                         ll_truncate_complete_page(page);
 457                 }
 458
 459                 unlock_page(page);
 460                 page_cache_release(page);
 461         }
 462
 463         return 0;
 464 }
 465
 466 int filter_clear_truncated_page(struct inode *inode)
 467 {
 468         struct page *page;
 469         int rc;
 470
 471         /* Truncate on page boundary, so nothing to flush? */
 472         if (!(i_size_read(inode) & ~CFS_PAGE_MASK))
 473                 return 0;
 474
 475         rc = filter_sync_inode_data(inode, 1);
 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         page = find_lock_page(inode->i_mapping,
 482                               i_size_read(inode) >> CFS_PAGE_SHIFT);
 483         if (page) {
 484                 if (page->mapping != NULL) {
 485                         wait_on_page_writeback(page);
 486                         ll_truncate_complete_page(page);
 487                 }
 488                 unlock_page(page);
 489                 page_cache_release(page);
 490         }
 491
 492         return 0;
 493 }
 494
 495 /* Must be called with i_mutex taken for writes; this will drop it */
 496 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 497                      struct obd_export *exp, struct iattr *attr,
 498                      struct obd_trans_info *oti, void **wait_handle)
 499 {
 500         struct obd_device *obd = exp->exp_obd;
 501         struct inode *inode = dchild->d_inode;
 502         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 503         int rc, rc2, create;
 504         struct semaphore *sem;
 505         ENTRY;
 506
 507         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 508                  iobuf->dr_npages, iobuf->dr_max_pages);
 509         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 510
 511         if (rw == OBD_BRW_READ) {
 512                 if (iobuf->dr_npages == 0)
 513                         RETURN(0);
 514                 create = 0;
 515                 sem = NULL;
 516         } else {
 517                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 518                 LASSERT(iobuf->dr_npages > 0);
 519                 create = 1;
 520                 sem = &obd->u.filter.fo_alloc_lock;
 521
 522                 lquota_enforce(filter_quota_interface_ref, obd, iobuf->dr_ignore_quota);
 523         }
 524
 525         rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 526                                     iobuf->dr_npages, iobuf->dr_blocks,
 527                                     obdfilter_created_scratchpad, create, sem);
 528
 529         if (rw == OBD_BRW_WRITE) {
 530                 if (rc == 0) {
 531                         filter_tally(exp, iobuf->dr_pages,
 532                                      iobuf->dr_npages, iobuf->dr_blocks,
 533                                      blocks_per_page, 1);
 534                         if (attr->ia_size > i_size_read(inode))
 535                                 attr->ia_valid |= ATTR_SIZE;
 536                         rc = fsfilt_setattr(obd, dchild,
 537                                             oti->oti_handle, attr, 0);
 538                 }
 539
 540                 UNLOCK_INODE_MUTEX(inode);
 541
 542                 rc2 = filter_finish_transno(exp, oti, 0, 0);
 543                 if (rc2 != 0) {
 544                         CERROR("can't close transaction: %d\n", rc2);
 545                         if (rc == 0)
 546                                 rc = rc2;
 547                 }
 548
 549                 rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
 550                                           wait_handle);
 551                 if (rc == 0)
 552                         rc = rc2;
 553                 if (rc != 0)
 554                         RETURN(rc);
 555         } else if (rc == 0) {
 556                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 557                              iobuf->dr_blocks, blocks_per_page, 0);
 558         }
 559
 560         rc = filter_clear_page_cache(inode, iobuf);
 561         if (rc != 0)
 562                 RETURN(rc);
 563
 564         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 565 }
 566
 567 /* See if there are unallocated parts in given file region */
 568 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 569 {
 570         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 571                 inode->i_mapping->a_ops->bmap;
 572         int j;
 573
 574         /* We can't know if we are overwriting or not */
 575         if (fs_bmap == NULL)
 576                 return 0;
 577
 578         offset >>= inode->i_blkbits;
 579         len >>= inode->i_blkbits;
 580
 581         for (j = 0; j <= len; j++)
 582                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 583                         return 0;
 584
 585         return 1;
 586 }
 587
 588 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 589                           int objcount, struct obd_ioobj *obj, int niocount,
 590                           struct niobuf_local *res, struct obd_trans_info *oti,
 591                           int rc)
 592 {
 593         struct niobuf_local *lnb;
 594         struct filter_iobuf *iobuf = NULL;
 595         struct lvfs_run_ctxt saved;
 596         struct fsfilt_objinfo fso;
 597         struct iattr iattr = { 0 };
 598         struct inode *inode = NULL;
 599         unsigned long now = jiffies;
 600         int i, err, cleanup_phase = 0;
 601         struct obd_device *obd = exp->exp_obd;
 602         void *wait_handle;
 603         int   total_size = 0, rc2;
 604         unsigned int qcids[MAXQUOTAS] = {0, 0};
 605         ENTRY;
 606
 607         LASSERT(oti != NULL);
 608         LASSERT(objcount == 1);
 609         LASSERT(current->journal_info == NULL);
 610
 611         if (rc != 0)
 612                 GOTO(cleanup, rc);
 613
 614         /* Unfortunately, if quota master is too busy to handle the
 615          * pre-dqacq in time and quota hash on ost is used up, we
 616          * have to wait for the completion of in flight dqacq/dqrel,
 617          * then try again */
 618         if ((rc2 = lquota_chkquota(filter_quota_interface_ref, obd, oa->o_uid,
 619                                    oa->o_gid, niocount)) == QUOTA_RET_ACQUOTA) {
 620                 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_HOLD_WRITE_RPC, 90);
 621                 lquota_acquire(filter_quota_interface_ref, obd, oa->o_uid,
 622                                oa->o_gid);
 623         }
 624
 625         if (rc2 < 0) {
 626                 rc = rc2;
 627                 GOTO(cleanup, rc);
 628         }
 629
 630         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 631         if (IS_ERR(iobuf))
 632                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 633         cleanup_phase = 1;
 634
 635         fso.fso_dentry = res->dentry;
 636         fso.fso_bufcnt = obj->ioo_bufcnt;
 637         inode = res->dentry->d_inode;
 638
 639         iobuf->dr_ignore_quota = 0;
 640         for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
 641                 loff_t this_size;
 642
 643                 /* If overwriting an existing block, we don't need a grant */
 644                 if (!(lnb->flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC &&
 645                     filter_range_is_mapped(inode, lnb->offset, lnb->len))
 646                         lnb->rc = 0;
 647
 648                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 649                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 650                         continue;
 651                 }
 652
 653                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 654                 LASSERT (err == 0);
 655
 656                 total_size += lnb->len;
 657
 658                 /* we expect these pages to be in offset order, but we'll
 659                  * be forgiving */
 660                 this_size = lnb->offset + lnb->len;
 661                 if (this_size > iattr.ia_size)
 662                         iattr.ia_size = this_size;
 663
 664                 /* if one page is a write-back page from client cache, or it's
 665                  * written by root, then mark the whole io request as ignore
 666                  * quota request */
 667                 if (lnb->flags & (OBD_BRW_FROM_GRANT | OBD_BRW_NOQUOTA))
 668                         iobuf->dr_ignore_quota = 1;
 669         }
 670
 671         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 672         cleanup_phase = 2;
 673
 674         DQUOT_INIT(inode);
 675
 676         LOCK_INODE_MUTEX(inode);
 677         fsfilt_check_slow(obd, now, obd_timeout, "i_mutex");
 678         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 679                                            oti);
 680         if (IS_ERR(oti->oti_handle)) {
 681                 UNLOCK_INODE_MUTEX(inode);
 682                 rc = PTR_ERR(oti->oti_handle);
 683                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 684                        "error starting transaction: rc = %d\n", rc);
 685                 oti->oti_handle = NULL;
 686                 GOTO(cleanup, rc);
 687         }
 688         /* have to call fsfilt_commit() from this point on */
 689
 690         fsfilt_check_slow(obd, now, obd_timeout, "brw_start");
 691
 692         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 693
 694         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 695          * then we will accept the UID+GID if sent by the client for
 696          * initializing the ownership of this inode.  We only allow this to
 697          * happen once (so clear these bits) and later only allow setattr. */
 698         if (inode->i_mode & S_ISUID)
 699                 i |= OBD_MD_FLUID;
 700         if (inode->i_mode & S_ISGID)
 701                 i |= OBD_MD_FLGID;
 702
 703         iattr_from_obdo(&iattr, oa, i);
 704         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 705                 unsigned int save;
 706
 707                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 708                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 709
 710                 cap_raise(current->cap_effective, CAP_SYS_RESOURCE);
 711
 712                 iattr.ia_valid |= ATTR_MODE;
 713                 iattr.ia_mode = inode->i_mode;
 714                 if (iattr.ia_valid & ATTR_UID)
 715                         iattr.ia_mode &= ~S_ISUID;
 716                 if (iattr.ia_valid & ATTR_GID)
 717                         iattr.ia_mode &= ~S_ISGID;
 718
 719                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 720
 721                 /* To avoid problems with quotas, UID and GID must be set
 722                  * in the inode before filter_direct_io() - see bug 10357. */
 723                 save = iattr.ia_valid;
 724                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 725                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle, &iattr, 0);
 726                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 727                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 728                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 729         }
 730
 731         /* filter_direct_io drops i_mutex */
 732         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 733                               oti, &wait_handle);
 734         if (rc == 0)
 735                 obdo_from_inode(oa, inode,
 736                                 FILTER_VALID_FLAGS |OBD_MD_FLUID |OBD_MD_FLGID);
 737         else
 738                 obdo_from_inode(oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
 739
 740         lquota_getflag(filter_quota_interface_ref, obd, oa);
 741
 742         fsfilt_check_slow(obd, now, obd_timeout, "direct_io");
 743
 744         err = fsfilt_commit_wait(obd, inode, wait_handle);
 745         if (err) {
 746                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 747                 rc = err;
 748         }
 749
 750         if (obd->obd_replayable && !rc)
 751                 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
 752                          "oti_transno "LPU64" last_committed "LPU64"\n",
 753                          oti->oti_transno, obd->obd_last_committed);
 754
 755         fsfilt_check_slow(obd, now, obd_timeout, "commitrw commit");
 756
 757 cleanup:
 758         filter_grant_commit(exp, niocount, res);
 759
 760         switch (cleanup_phase) {
 761         case 2:
 762                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 763                 LASSERT(current->journal_info == NULL);
 764         case 1:
 765                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 766         case 0:
 767                 /*
 768                  * lnb->page automatically returns back into per-thread page
 769                  * pool (bug 5137)
 770                  */
 771                 f_dput(res->dentry);
 772         }
 773
 774         /* trigger quota pre-acquire */
 775         qcids[USRQUOTA] = oa->o_uid;
 776         qcids[GRPQUOTA] = oa->o_gid;
 777         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 778                             FSFILT_OP_CREATE);
 779         CDEBUG(err ? D_ERROR : D_QUOTA,
 780                "filter adjust qunit! (rc:%d)\n", err);
 781
 782         RETURN(rc);
 783 }