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  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdfilter/filter_io_26.c
  37  *
  38  * Author: Peter Braam <braam@clusterfs.com>
  39  * Author: Andreas Dilger <adilger@clusterfs.com>
  40  * Author: Phil Schwan <phil@clusterfs.com>
  41  */
  42
  43 #ifndef AUTOCONF_INCLUDED
  44 #include <linux/config.h>
  45 #endif
  46 #include <linux/module.h>
  47 #include <linux/pagemap.h> // XXX kill me soon
  48 #include <linux/version.h>
  49 #include <linux/buffer_head.h>
  50
  51 #define DEBUG_SUBSYSTEM S_FILTER
  52
  53 #include <obd_class.h>
  54 #include <lustre_fsfilt.h>
  55 #include <lustre_quota.h>
  56 #include "filter_internal.h"
  57
  58 /* 512byte block min */
  59 #define MAX_BLOCKS_PER_PAGE (CFS_PAGE_SIZE / 512)
  60 struct filter_iobuf {
  61         cfs_atomic_t       dr_numreqs;  /* number of reqs being processed */
  62         cfs_waitq_t        dr_wait;
  63         int                dr_max_pages;
  64         int                dr_npages;
  65         int                dr_error;
  66         struct page      **dr_pages;
  67         unsigned long     *dr_blocks;
  68         unsigned int       dr_ignore_quota:1;
  69         struct filter_obd *dr_filter;
  70 };
  71
  72 static void record_start_io(struct filter_iobuf *iobuf, int rw, int size,
  73                             struct obd_export *exp)
  74 {
  75         struct filter_obd *filter = iobuf->dr_filter;
  76
  77         cfs_atomic_inc(&iobuf->dr_numreqs);
  78
  79         if (rw == OBD_BRW_READ) {
  80                 cfs_atomic_inc(&filter->fo_r_in_flight);
  81                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST],
  82                                  cfs_atomic_read(&filter->fo_r_in_flight));
  83                 lprocfs_oh_tally_log2(&filter->
  84                                        fo_filter_stats.hist[BRW_R_DISK_IOSIZE],
  85                                       size);
  86                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
  87                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
  88                                          hist[BRW_R_RPC_HIST],
  89                                          cfs_atomic_read(&filter-> \
  90                                          fo_r_in_flight));
  91                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
  92                                          nid_brw_stats->hist[BRW_R_DISK_IOSIZE],
  93                                               size);
  94                 }
  95         } else {
  96                 cfs_atomic_inc(&filter->fo_w_in_flight);
  97                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
  98                                  cfs_atomic_read(&filter->fo_w_in_flight));
  99                 lprocfs_oh_tally_log2(&filter->
 100                                        fo_filter_stats.hist[BRW_W_DISK_IOSIZE],
 101                                       size);
 102                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 103                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 104                                           hist[BRW_W_RPC_HIST],
 105                                          cfs_atomic_read(&filter-> \
 106                                          fo_r_in_flight));
 107                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 108                                         nid_brw_stats->hist[BRW_W_DISK_IOSIZE],
 109                                               size);
 110                 }
 111         }
 112 }
 113
 114 static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
 115 {
 116         struct filter_obd *filter = iobuf->dr_filter;
 117
 118         /* CAVEAT EMPTOR: possibly in IRQ context
 119          * DO NOT record procfs stats here!!! */
 120
 121         if (rw == OBD_BRW_READ)
 122                 cfs_atomic_dec(&filter->fo_r_in_flight);
 123         else
 124                 cfs_atomic_dec(&filter->fo_w_in_flight);
 125
 126         if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs))
 127                 cfs_waitq_signal(&iobuf->dr_wait);
 128 }
 129
 130 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 131 {
 132         struct filter_iobuf *iobuf = bio->bi_private;
 133         struct bio_vec *bvl;
 134         int i;
 135
 136         /* CAVEAT EMPTOR: possibly in IRQ context
 137          * DO NOT record procfs stats here!!! */
 138
 139         if (bio->bi_size)                       /* Not complete */
 140                 return 1;
 141
 142         if (unlikely(iobuf == NULL)) {
 143                 CERROR("***** bio->bi_private is NULL!  This should never "
 144                        "happen.  Normally, I would crash here, but instead I "
 145                        "will dump the bio contents to the console.  Please "
 146                        "report this to <http://bugzilla.lustre.org/> , along "
 147                        "with any interesting messages leading up to this point "
 148                        "(like SCSI errors, perhaps).  Because bi_private is "
 149                        "NULL, I can't wake up the thread that initiated this "
 150                        "IO - you will probably have to reboot this node.\n");
 151                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 152                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 153                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 154                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 155                        bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
 156                        bio->bi_private);
 157                 return 0;
 158         }
 159
 160         /* the check is outside of the cycle for performance reason -bzzz */
 161         if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
 162                 bio_for_each_segment(bvl, bio, i) {
 163                         if (likely(error == 0))
 164                                 SetPageUptodate(bvl->bv_page);
 165                         LASSERT(PageLocked(bvl->bv_page));
 166                         ClearPageConstant(bvl->bv_page);
 167                 }
 168                 record_finish_io(iobuf, OBD_BRW_READ, error);
 169         } else {
 170                 if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){
 171                         bio_for_each_segment(bvl, bio, i) {
 172                                 ClearPageConstant(bvl->bv_page);
 173                         }
 174                 }
 175                 record_finish_io(iobuf, OBD_BRW_WRITE, error);
 176         }
 177
 178         /* any real error is good enough -bzzz */
 179         if (error != 0 && iobuf->dr_error == 0)
 180                 iobuf->dr_error = error;
 181
 182         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 183          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 184          * mempool when serious on-disk fragmentation was encountered,
 185          * deadlocking the OST.  The bios are now released as soon as complete
 186          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 187         bio_put(bio);
 188         return 0;
 189 }
 190
 191 static int can_be_merged(struct bio *bio, sector_t sector)
 192 {
 193         unsigned int size;
 194
 195         if (!bio)
 196                 return 0;
 197
 198         size = bio->bi_size >> 9;
 199         return bio->bi_sector + size == sector ? 1 : 0;
 200 }
 201
 202 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 203                                         int rw, int num_pages)
 204 {
 205         struct filter_iobuf *iobuf;
 206
 207         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 208
 209         OBD_ALLOC(iobuf, sizeof(*iobuf));
 210         if (iobuf == NULL)
 211                 goto failed_0;
 212
 213         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 214         if (iobuf->dr_pages == NULL)
 215                 goto failed_1;
 216
 217         OBD_ALLOC(iobuf->dr_blocks,
 218                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 219         if (iobuf->dr_blocks == NULL)
 220                 goto failed_2;
 221
 222         iobuf->dr_filter = filter;
 223         cfs_waitq_init(&iobuf->dr_wait);
 224         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 225         iobuf->dr_max_pages = num_pages;
 226         iobuf->dr_npages = 0;
 227         iobuf->dr_error = 0;
 228
 229         RETURN(iobuf);
 230
 231  failed_2:
 232         OBD_FREE(iobuf->dr_pages,
 233                  num_pages * sizeof(*iobuf->dr_pages));
 234  failed_1:
 235         OBD_FREE(iobuf, sizeof(*iobuf));
 236  failed_0:
 237         RETURN(ERR_PTR(-ENOMEM));
 238 }
 239
 240 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
 241 {
 242         iobuf->dr_npages = 0;
 243         iobuf->dr_error = 0;
 244         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 245 }
 246
 247 void filter_free_iobuf(struct filter_iobuf *iobuf)
 248 {
 249         int num_pages = iobuf->dr_max_pages;
 250
 251         filter_clear_iobuf(iobuf);
 252
 253         OBD_FREE(iobuf->dr_blocks,
 254                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 255         OBD_FREE(iobuf->dr_pages,
 256                  num_pages * sizeof(*iobuf->dr_pages));
 257         OBD_FREE_PTR(iobuf);
 258 }
 259
 260 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 261                       struct obd_trans_info *oti)
 262 {
 263         int thread_id = (oti && oti->oti_thread) ?
 264                         oti->oti_thread->t_id : -1;
 265
 266         if (unlikely(thread_id < 0)) {
 267                 filter_free_iobuf(iobuf);
 268                 return;
 269         }
 270
 271         LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
 272                  "iobuf mismatch for thread %d: pool %p iobuf %p\n",
 273                  thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
 274         filter_clear_iobuf(iobuf);
 275 }
 276
 277 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 278                           struct inode *inode, struct page *page)
 279 {
 280         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 281         iobuf->dr_pages[iobuf->dr_npages++] = page;
 282
 283         return 0;
 284 }
 285
 286 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 287                   struct filter_iobuf *iobuf, int rw)
 288 {
 289         struct obd_device *obd = exp->exp_obd;
 290         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 291         struct page  **pages = iobuf->dr_pages;
 292         int            npages = iobuf->dr_npages;
 293         unsigned long *blocks = iobuf->dr_blocks;
 294         int            total_blocks = npages * blocks_per_page;
 295         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 296         unsigned int   blocksize = inode->i_sb->s_blocksize;
 297         struct bio    *bio = NULL;
 298         int            frags = 0;
 299         unsigned long  start_time = jiffies;
 300         struct page   *page;
 301         unsigned int   page_offset;
 302         sector_t       sector;
 303         int            nblocks;
 304         int            block_idx;
 305         int            page_idx;
 306         int            i;
 307         int            rc = 0;
 308         ENTRY;
 309
 310         LASSERT(iobuf->dr_npages == npages);
 311         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 312
 313         for (page_idx = 0, block_idx = 0;
 314              page_idx < npages;
 315              page_idx++, block_idx += blocks_per_page) {
 316
 317                 page = pages[page_idx];
 318                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 319
 320                 for (i = 0, page_offset = 0;
 321                      i < blocks_per_page;
 322                      i += nblocks, page_offset += blocksize * nblocks) {
 323
 324                         nblocks = 1;
 325
 326                         if (blocks[block_idx + i] == 0) {  /* hole */
 327                                 LASSERT(rw == OBD_BRW_READ);
 328                                 memset(kmap(page) + page_offset, 0, blocksize);
 329                                 kunmap(page);
 330                                 continue;
 331                         }
 332
 333                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 334
 335                         /* Additional contiguous file blocks? */
 336                         while (i + nblocks < blocks_per_page &&
 337                                (sector + (nblocks << sector_bits)) ==
 338                                ((sector_t)blocks[block_idx + i + nblocks] <<
 339                                 sector_bits))
 340                                 nblocks++;
 341
 342                         /* I only set the page to be constant only if it
 343                          * is mapped to a contiguous underlying disk block(s).
 344                          * It will then make sure the corresponding device
 345                          * cache of raid5 will be overwritten by this page.
 346                          * - jay */
 347                         if ((rw == OBD_BRW_WRITE) &&
 348                             (nblocks == blocks_per_page) &&
 349                             mapping_cap_page_constant_write(inode->i_mapping))
 350                                SetPageConstant(page);
 351
 352                         if (bio != NULL &&
 353                             can_be_merged(bio, sector) &&
 354                             bio_add_page(bio, page,
 355                                          blocksize * nblocks, page_offset) != 0)
 356                                 continue;       /* added this frag OK */
 357
 358                         if (bio != NULL) {
 359                                 request_queue_t *q =
 360                                         bdev_get_queue(bio->bi_bdev);
 361
 362                                 /* Dang! I have to fragment this I/O */
 363                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 364                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 365                                        bio->bi_size,
 366                                        bio->bi_vcnt, bio->bi_max_vecs,
 367                                        bio->bi_size >> 9, q->max_sectors,
 368                                        bio_phys_segments(q, bio),
 369                                        q->max_phys_segments,
 370                                        bio_hw_segments(q, bio),
 371                                        q->max_hw_segments);
 372
 373                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 374                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 375                                 if (rc < 0) {
 376                                         CERROR("Can't send bio: %d\n", rc);
 377                                         record_finish_io(iobuf, rw, rc);
 378                                         goto out;
 379                                 }
 380                                 frags++;
 381                         }
 382
 383                         /* allocate new bio, limited by max BIO size, b=9945 */
 384                         bio = bio_alloc(GFP_NOIO, max(BIO_MAX_PAGES,
 385                                                       (npages - page_idx) *
 386                                                       blocks_per_page));
 387                         if (bio == NULL) {
 388                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 389                                        (npages - page_idx), blocks_per_page,
 390                                        (npages - page_idx) * blocks_per_page);
 391                                 rc = -ENOMEM;
 392                                 goto out;
 393                         }
 394
 395                         bio->bi_bdev = inode->i_sb->s_bdev;
 396                         bio->bi_sector = sector;
 397                         bio->bi_end_io = dio_complete_routine;
 398                         bio->bi_private = iobuf;
 399
 400                         rc = bio_add_page(bio, page,
 401                                           blocksize * nblocks, page_offset);
 402                         LASSERT (rc != 0);
 403                 }
 404         }
 405
 406         if (bio != NULL) {
 407                 record_start_io(iobuf, rw, bio->bi_size, exp);
 408                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 409                 if (rc >= 0) {
 410                         frags++;
 411                         rc = 0;
 412                 } else {
 413                         CERROR("Can't send bio: %d\n", rc);
 414                         record_finish_io(iobuf, rw, rc);
 415                 }
 416         }
 417
 418  out:
 419         cfs_wait_event(iobuf->dr_wait,
 420                        cfs_atomic_read(&iobuf->dr_numreqs) == 0);
 421
 422         if (rw == OBD_BRW_READ) {
 423                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 424                                   hist[BRW_R_DIO_FRAGS],
 425                                  frags);
 426                 lprocfs_oh_tally_log2(&obd->u.filter.
 427                                        fo_filter_stats.hist[BRW_R_IO_TIME],
 428                                       jiffies - start_time);
 429                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 430                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 431                                           hist[BRW_R_DIO_FRAGS],
 432                                          frags);
 433                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 434                                              nid_brw_stats->hist[BRW_R_IO_TIME],
 435                                               jiffies - start_time);
 436                 }
 437         } else {
 438                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 439                                   hist[BRW_W_DIO_FRAGS], frags);
 440                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.
 441                                        hist[BRW_W_IO_TIME],
 442                                       jiffies - start_time);
 443                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 444                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 445                                           hist[BRW_W_DIO_FRAGS],
 446                                          frags);
 447                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 448                                              nid_brw_stats->hist[BRW_W_IO_TIME],
 449                                               jiffies - start_time);
 450                 }
 451         }
 452
 453         if (rc == 0)
 454                 rc = iobuf->dr_error;
 455         RETURN(rc);
 456 }
 457
 458 /* Must be called with i_mutex taken for writes; this will drop it */
 459 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 460                      struct obd_export *exp, struct iattr *attr,
 461                      struct obd_trans_info *oti, void **wait_handle)
 462 {
 463         struct obd_device *obd = exp->exp_obd;
 464         struct inode *inode = dchild->d_inode;
 465         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 466         int rc, rc2, create;
 467         cfs_semaphore_t *sem;
 468         ENTRY;
 469
 470         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 471                  iobuf->dr_npages, iobuf->dr_max_pages);
 472         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 473
 474         if (rw == OBD_BRW_READ) {
 475                 if (iobuf->dr_npages == 0)
 476                         RETURN(0);
 477                 create = 0;
 478                 sem = NULL;
 479         } else {
 480                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 481                 LASSERT(iobuf->dr_npages > 0);
 482                 create = 1;
 483                 sem = &obd->u.filter.fo_alloc_lock;
 484
 485                 lquota_enforce(filter_quota_interface_ref, obd,
 486                                iobuf->dr_ignore_quota);
 487         }
 488
 489         rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 490                                     iobuf->dr_npages, iobuf->dr_blocks,
 491                                     obdfilter_created_scratchpad, create, sem);
 492
 493         if (rw == OBD_BRW_WRITE) {
 494                 if (rc == 0) {
 495                         filter_tally(exp, iobuf->dr_pages,
 496                                      iobuf->dr_npages, iobuf->dr_blocks,
 497                                      blocks_per_page, 1);
 498                         if (attr->ia_size > i_size_read(inode))
 499                                 attr->ia_valid |= ATTR_SIZE;
 500                         rc = fsfilt_setattr(obd, dchild,
 501                                             oti->oti_handle, attr, 0);
 502                 }
 503
 504                 UNLOCK_INODE_MUTEX(inode);
 505
 506                 rc2 = filter_finish_transno(exp, inode, oti, 0, 0);
 507                 if (rc2 != 0) {
 508                         CERROR("can't close transaction: %d\n", rc2);
 509                         if (rc == 0)
 510                                 rc = rc2;
 511                 }
 512
 513                 rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
 514                                           wait_handle);
 515                 if (rc == 0)
 516                         rc = rc2;
 517                 if (rc != 0)
 518                         RETURN(rc);
 519         } else if (rc == 0) {
 520                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 521                              iobuf->dr_blocks, blocks_per_page, 0);
 522         }
 523
 524         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 525 }
 526
 527 /* See if there are unallocated parts in given file region */
 528 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 529 {
 530         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 531                 inode->i_mapping->a_ops->bmap;
 532         int j;
 533
 534         /* We can't know if we are overwriting or not */
 535         if (fs_bmap == NULL)
 536                 return 0;
 537
 538         offset >>= inode->i_blkbits;
 539         len >>= inode->i_blkbits;
 540
 541         for (j = 0; j <= len; j++)
 542                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 543                         return 0;
 544
 545         return 1;
 546 }
 547
 548 /*
 549  * interesting use cases on how it interacts with VM:
 550  *
 551  * - vm writeout -- shouldn't see our pages as we don't mark them dirty
 552  *   though vm can find partial page left dirty by truncate. in this
 553  *   usual writeout is used unless our write rewrite that page - then we
 554  *   drop PG_dirty with PG_lock held.
 555  *
 556  * - else?
 557  *
 558  */
 559 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 560                           int objcount, struct obd_ioobj *obj,
 561                           struct niobuf_remote *nb, int niocount,
 562                           struct niobuf_local *res, struct obd_trans_info *oti,
 563                           int rc)
 564 {
 565         struct niobuf_local *lnb;
 566         struct filter_iobuf *iobuf = NULL;
 567         struct lvfs_run_ctxt saved;
 568         struct fsfilt_objinfo fso;
 569         struct iattr iattr = { 0 };
 570         struct inode *inode = res->dentry->d_inode;
 571         unsigned long now = jiffies;
 572         int i, err, cleanup_phase = 0;
 573         struct obd_device *obd = exp->exp_obd;
 574         struct filter_obd *fo = &obd->u.filter;
 575         void *wait_handle;
 576         int total_size = 0;
 577         unsigned int qcids[MAXQUOTAS] = { oa->o_uid, oa->o_gid };
 578         int rec_pending[MAXQUOTAS] = { 0, 0 }, quota_pages = 0;
 579         ENTRY;
 580
 581         LASSERT(oti != NULL);
 582         LASSERT(objcount == 1);
 583         LASSERT(current->journal_info == NULL);
 584
 585         if (rc != 0)
 586                 GOTO(cleanup, rc);
 587
 588         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 589         if (IS_ERR(iobuf))
 590                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 591         cleanup_phase = 1;
 592
 593         fso.fso_dentry = res->dentry;
 594         fso.fso_bufcnt = obj->ioo_bufcnt;
 595
 596         iobuf->dr_ignore_quota = 0;
 597         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 598                 loff_t this_size;
 599                 __u32 flags = lnb->flags;
 600
 601                 if (filter_range_is_mapped(inode, lnb->offset, lnb->len)) {
 602                         /* If overwriting an existing block,
 603                          * we don't need a grant */
 604                         if (!(flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC)
 605                                 lnb->rc = 0;
 606                 } else {
 607                         quota_pages++;
 608                 }
 609
 610                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 611                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 612                         continue;
 613                 }
 614
 615                 LASSERT(PageLocked(lnb->page));
 616                 LASSERT(!PageWriteback(lnb->page));
 617
 618                 /* since write & truncate are serialized by the i_alloc_sem,
 619                  * even partial truncate should not leave dirty pages in
 620                  * the page cache */
 621                 LASSERT(!PageDirty(lnb->page));
 622
 623                 SetPageUptodate(lnb->page);
 624
 625                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 626                 LASSERT (err == 0);
 627
 628                 total_size += lnb->len;
 629
 630                 /* we expect these pages to be in offset order, but we'll
 631                  * be forgiving */
 632                 this_size = lnb->offset + lnb->len;
 633                 if (this_size > iattr.ia_size)
 634                         iattr.ia_size = this_size;
 635
 636                 /* if one page is a write-back page from client cache and
 637                  * not from direct_io, or it's written by root, then mark
 638                  * the whole io request as ignore quota request, remote
 639                  * client can not break through quota. */
 640                 if (exp_connect_rmtclient(exp))
 641                         flags &= ~OBD_BRW_NOQUOTA;
 642                 if ((flags & OBD_BRW_NOQUOTA) ||
 643                     (flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 644                      OBD_BRW_FROM_GRANT)
 645                         iobuf->dr_ignore_quota = 1;
 646         }
 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         rc = lquota_chkquota(filter_quota_interface_ref, obd, exp, qcids,
 651                              rec_pending, quota_pages, oti, LQUOTA_FLAGS_BLK,
 652                              (void *)inode, obj->ioo_bufcnt);
 653         if (rc == -ENOTCONN)
 654                 GOTO(cleanup, rc);
 655
 656         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 657         cleanup_phase = 2;
 658
 659         DQUOT_INIT(inode);
 660
 661         LOCK_INODE_MUTEX(inode);
 662         fsfilt_check_slow(obd, now, "i_mutex");
 663         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 664                                            oti);
 665         if (IS_ERR(oti->oti_handle)) {
 666                 UNLOCK_INODE_MUTEX(inode);
 667                 rc = PTR_ERR(oti->oti_handle);
 668                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 669                        "error starting transaction: rc = %d\n", rc);
 670                 oti->oti_handle = NULL;
 671                 GOTO(cleanup, rc);
 672         }
 673         /* have to call fsfilt_commit() from this point on */
 674
 675         fsfilt_check_slow(obd, now, "brw_start");
 676
 677         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 678
 679         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 680          * then we will accept the UID+GID if sent by the client for
 681          * initializing the ownership of this inode.  We only allow this to
 682          * happen once (so clear these bits) and later only allow setattr. */
 683         if (inode->i_mode & S_ISUID)
 684                 i |= OBD_MD_FLUID;
 685         if (inode->i_mode & S_ISGID)
 686                 i |= OBD_MD_FLGID;
 687
 688         iattr_from_obdo(&iattr, oa, i);
 689         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 690                 unsigned int save;
 691
 692                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 693                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 694
 695                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 696
 697                 iattr.ia_valid |= ATTR_MODE;
 698                 iattr.ia_mode = inode->i_mode;
 699                 if (iattr.ia_valid & ATTR_UID)
 700                         iattr.ia_mode &= ~S_ISUID;
 701                 if (iattr.ia_valid & ATTR_GID)
 702                         iattr.ia_mode &= ~S_ISGID;
 703
 704                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 705
 706                 /* To avoid problems with quotas, UID and GID must be set
 707                  * in the inode before filter_direct_io() - see bug 10357. */
 708                 save = iattr.ia_valid;
 709                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 710                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle,&iattr,0);
 711                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 712                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 713                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 714         }
 715
 716         /* filter_direct_io drops i_mutex */
 717         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 718                               oti, &wait_handle);
 719         if (rc == 0)
 720                 obdo_from_inode(oa, inode,
 721                                 FILTER_VALID_FLAGS |OBD_MD_FLUID |OBD_MD_FLGID);
 722         else
 723                 obdo_from_inode(oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
 724
 725         lquota_getflag(filter_quota_interface_ref, obd, oa);
 726
 727         fsfilt_check_slow(obd, now, "direct_io");
 728
 729         err = fsfilt_commit_wait(obd, inode, wait_handle);
 730         if (err) {
 731                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 732                 rc = err;
 733         }
 734
 735         if (obd->obd_replayable && !rc)
 736                 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
 737                          "oti_transno "LPU64" last_committed "LPU64"\n",
 738                          oti->oti_transno, obd->obd_last_committed);
 739
 740         fsfilt_check_slow(obd, now, "commitrw commit");
 741
 742 cleanup:
 743         lquota_pending_commit(filter_quota_interface_ref, obd, qcids,
 744                               rec_pending, 1);
 745
 746         filter_grant_commit(exp, niocount, res);
 747
 748         switch (cleanup_phase) {
 749         case 2:
 750                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 751                 LASSERT(current->journal_info == NULL);
 752         case 1:
 753                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 754         case 0:
 755                 /*
 756                  * lnb->page automatically returns back into per-thread page
 757                  * pool (bug 5137)
 758                  */
 759                  break;
 760         }
 761
 762         /* trigger quota pre-acquire */
 763         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 764                             FSFILT_OP_CREATE);
 765         CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 766                "(rc:%d, uid:%u, gid:%u)\n",
 767                err, qcids[USRQUOTA], qcids[GRPQUOTA]);
 768         if (qcids[USRQUOTA] != oa->o_uid || qcids[GRPQUOTA] != oa->o_gid) {
 769                 qcids[USRQUOTA] = oa->o_uid;
 770                 qcids[GRPQUOTA] = oa->o_gid;
 771                 err = lquota_adjust(filter_quota_interface_ref, obd, qcids,
 772                                     NULL, rc, FSFILT_OP_CREATE);
 773                 CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 774                        "(rc:%d, uid:%u, gid:%u)\n",
 775                        err, qcids[USRQUOTA], qcids[GRPQUOTA]);
 776         }
 777
 778         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 779                 if (lnb->page == NULL)
 780                         continue;
 781
 782                 if (rc)
 783                         /* If the write has failed, the page cache may
 784                          * not be consitent with what is on disk, so
 785                          * force pages to be reread next time it is
 786                          * accessed */
 787                         ClearPageUptodate(lnb->page);
 788
 789                 LASSERT(PageLocked(lnb->page));
 790                 unlock_page(lnb->page);
 791
 792                 page_cache_release(lnb->page);
 793                 lnb->page = NULL;
 794         }
 795         f_dput(res->dentry);
 796
 797         if (inode) {
 798                 if (fo->fo_writethrough_cache == 0 ||
 799                     i_size_read(inode) > fo->fo_readcache_max_filesize)
 800                         filter_release_cache(obd, obj, nb, inode);
 801                 up_read(&inode->i_alloc_sem);
 802         }
 803
 804         RETURN(rc);
 805 }