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         atomic_t          dr_numreqs;  /* number of reqs being processed */
  62         wait_queue_head_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         atomic_inc(&iobuf->dr_numreqs);
  78
  79         if (rw == OBD_BRW_READ) {
  80                 atomic_inc(&filter->fo_r_in_flight);
  81                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST],
  82                                  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                                          atomic_read(&filter->fo_r_in_flight));
  90                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
  91                                          nid_brw_stats->hist[BRW_R_DISK_IOSIZE],
  92                                               size);
  93                 }
  94         } else {
  95                 atomic_inc(&filter->fo_w_in_flight);
  96                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
  97                                  atomic_read(&filter->fo_w_in_flight));
  98                 lprocfs_oh_tally_log2(&filter->
  99                                        fo_filter_stats.hist[BRW_W_DISK_IOSIZE],
 100                                       size);
 101                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 102                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 103                                           hist[BRW_W_RPC_HIST],
 104                                          atomic_read(&filter->fo_r_in_flight));
 105                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 106                                         nid_brw_stats->hist[BRW_W_DISK_IOSIZE],
 107                                               size);
 108                 }
 109         }
 110 }
 111
 112 static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
 113 {
 114         struct filter_obd *filter = iobuf->dr_filter;
 115
 116         /* CAVEAT EMPTOR: possibly in IRQ context
 117          * DO NOT record procfs stats here!!! */
 118
 119         if (rw == OBD_BRW_READ)
 120                 atomic_dec(&filter->fo_r_in_flight);
 121         else
 122                 atomic_dec(&filter->fo_w_in_flight);
 123
 124         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 125                 wake_up(&iobuf->dr_wait);
 126 }
 127
 128 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 129 {
 130         struct filter_iobuf *iobuf = bio->bi_private;
 131         struct bio_vec *bvl;
 132         int i;
 133
 134         /* CAVEAT EMPTOR: possibly in IRQ context
 135          * DO NOT record procfs stats here!!! */
 136
 137         if (bio->bi_size)                       /* Not complete */
 138                 return 1;
 139
 140         if (unlikely(iobuf == NULL)) {
 141                 CERROR("***** bio->bi_private is NULL!  This should never "
 142                        "happen.  Normally, I would crash here, but instead I "
 143                        "will dump the bio contents to the console.  Please "
 144                        "report this to <http://bugzilla.lustre.org/> , along "
 145                        "with any interesting messages leading up to this point "
 146                        "(like SCSI errors, perhaps).  Because bi_private is "
 147                        "NULL, I can't wake up the thread that initiated this "
 148                        "IO - you will probably have to reboot this node.\n");
 149                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 150                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 151                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 152                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 153                        bio->bi_end_io, atomic_read(&bio->bi_cnt),
 154                        bio->bi_private);
 155                 return 0;
 156         }
 157
 158         /* the check is outside of the cycle for performance reason -bzzz */
 159         if (!test_bit(BIO_RW, &bio->bi_rw)) {
 160                 bio_for_each_segment(bvl, bio, i) {
 161                         if (likely(error == 0))
 162                                 SetPageUptodate(bvl->bv_page);
 163                         LASSERT(PageLocked(bvl->bv_page));
 164                         ClearPageConstant(bvl->bv_page);
 165                 }
 166                 record_finish_io(iobuf, OBD_BRW_READ, error);
 167         } else {
 168                 if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){
 169                         bio_for_each_segment(bvl, bio, i) {
 170                                 ClearPageConstant(bvl->bv_page);
 171                         }
 172                 }
 173                 record_finish_io(iobuf, OBD_BRW_WRITE, error);
 174         }
 175
 176         /* any real error is good enough -bzzz */
 177         if (error != 0 && iobuf->dr_error == 0)
 178                 iobuf->dr_error = error;
 179
 180         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 181          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 182          * mempool when serious on-disk fragmentation was encountered,
 183          * deadlocking the OST.  The bios are now released as soon as complete
 184          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 185         bio_put(bio);
 186         return 0;
 187 }
 188
 189 static int can_be_merged(struct bio *bio, sector_t sector)
 190 {
 191         unsigned int size;
 192
 193         if (!bio)
 194                 return 0;
 195
 196         size = bio->bi_size >> 9;
 197         return bio->bi_sector + size == sector ? 1 : 0;
 198 }
 199
 200 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 201                                         int rw, int num_pages)
 202 {
 203         struct filter_iobuf *iobuf;
 204
 205         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 206
 207         OBD_ALLOC(iobuf, sizeof(*iobuf));
 208         if (iobuf == NULL)
 209                 goto failed_0;
 210
 211         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 212         if (iobuf->dr_pages == NULL)
 213                 goto failed_1;
 214
 215         OBD_ALLOC(iobuf->dr_blocks,
 216                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 217         if (iobuf->dr_blocks == NULL)
 218                 goto failed_2;
 219
 220         iobuf->dr_filter = filter;
 221         init_waitqueue_head(&iobuf->dr_wait);
 222         atomic_set(&iobuf->dr_numreqs, 0);
 223         iobuf->dr_max_pages = num_pages;
 224         iobuf->dr_npages = 0;
 225         iobuf->dr_error = 0;
 226
 227         RETURN(iobuf);
 228
 229  failed_2:
 230         OBD_FREE(iobuf->dr_pages,
 231                  num_pages * sizeof(*iobuf->dr_pages));
 232  failed_1:
 233         OBD_FREE(iobuf, sizeof(*iobuf));
 234  failed_0:
 235         RETURN(ERR_PTR(-ENOMEM));
 236 }
 237
 238 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
 239 {
 240         iobuf->dr_npages = 0;
 241         iobuf->dr_error = 0;
 242         atomic_set(&iobuf->dr_numreqs, 0);
 243 }
 244
 245 void filter_free_iobuf(struct filter_iobuf *iobuf)
 246 {
 247         int num_pages = iobuf->dr_max_pages;
 248
 249         filter_clear_iobuf(iobuf);
 250
 251         OBD_FREE(iobuf->dr_blocks,
 252                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 253         OBD_FREE(iobuf->dr_pages,
 254                  num_pages * sizeof(*iobuf->dr_pages));
 255         OBD_FREE_PTR(iobuf);
 256 }
 257
 258 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 259                       struct obd_trans_info *oti)
 260 {
 261         int thread_id = (oti && oti->oti_thread) ?
 262                         oti->oti_thread->t_id : -1;
 263
 264         if (unlikely(thread_id < 0)) {
 265                 filter_free_iobuf(iobuf);
 266                 return;
 267         }
 268
 269         LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
 270                  "iobuf mismatch for thread %d: pool %p iobuf %p\n",
 271                  thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
 272         filter_clear_iobuf(iobuf);
 273 }
 274
 275 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 276                           struct inode *inode, struct page *page)
 277 {
 278         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 279         iobuf->dr_pages[iobuf->dr_npages++] = page;
 280
 281         return 0;
 282 }
 283
 284 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 285                   struct filter_iobuf *iobuf, int rw)
 286 {
 287         struct obd_device *obd = exp->exp_obd;
 288         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 289         struct page  **pages = iobuf->dr_pages;
 290         int            npages = iobuf->dr_npages;
 291         unsigned long *blocks = iobuf->dr_blocks;
 292         int            total_blocks = npages * blocks_per_page;
 293         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 294         unsigned int   blocksize = inode->i_sb->s_blocksize;
 295         struct bio    *bio = NULL;
 296         int            frags = 0;
 297         unsigned long  start_time = jiffies;
 298         struct page   *page;
 299         unsigned int   page_offset;
 300         sector_t       sector;
 301         int            nblocks;
 302         int            block_idx;
 303         int            page_idx;
 304         int            i;
 305         int            rc = 0;
 306         ENTRY;
 307
 308         LASSERT(iobuf->dr_npages == npages);
 309         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 310
 311         for (page_idx = 0, block_idx = 0;
 312              page_idx < npages;
 313              page_idx++, block_idx += blocks_per_page) {
 314
 315                 page = pages[page_idx];
 316                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 317
 318                 for (i = 0, page_offset = 0;
 319                      i < blocks_per_page;
 320                      i += nblocks, page_offset += blocksize * nblocks) {
 321
 322                         nblocks = 1;
 323
 324                         if (blocks[block_idx + i] == 0) {  /* hole */
 325                                 LASSERT(rw == OBD_BRW_READ);
 326                                 memset(kmap(page) + page_offset, 0, blocksize);
 327                                 kunmap(page);
 328                                 continue;
 329                         }
 330
 331                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 332
 333                         /* Additional contiguous file blocks? */
 334                         while (i + nblocks < blocks_per_page &&
 335                                (sector + (nblocks << sector_bits)) ==
 336                                ((sector_t)blocks[block_idx + i + nblocks] <<
 337                                 sector_bits))
 338                                 nblocks++;
 339
 340                         /* I only set the page to be constant only if it
 341                          * is mapped to a contiguous underlying disk block(s).
 342                          * It will then make sure the corresponding device
 343                          * cache of raid5 will be overwritten by this page.
 344                          * - jay */
 345                         if ((rw == OBD_BRW_WRITE) &&
 346                             (nblocks == blocks_per_page) &&
 347                             mapping_cap_page_constant_write(inode->i_mapping))
 348                                SetPageConstant(page);
 349
 350                         if (bio != NULL &&
 351                             can_be_merged(bio, sector) &&
 352                             bio_add_page(bio, page,
 353                                          blocksize * nblocks, page_offset) != 0)
 354                                 continue;       /* added this frag OK */
 355
 356                         if (bio != NULL) {
 357                                 request_queue_t *q =
 358                                         bdev_get_queue(bio->bi_bdev);
 359
 360                                 /* Dang! I have to fragment this I/O */
 361                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 362                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 363                                        bio->bi_size,
 364                                        bio->bi_vcnt, bio->bi_max_vecs,
 365                                        bio->bi_size >> 9, q->max_sectors,
 366                                        bio_phys_segments(q, bio),
 367                                        q->max_phys_segments,
 368                                        bio_hw_segments(q, bio),
 369                                        q->max_hw_segments);
 370
 371                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 372                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 373                                 if (rc < 0) {
 374                                         CERROR("Can't send bio: %d\n", rc);
 375                                         record_finish_io(iobuf, rw, rc);
 376                                         goto out;
 377                                 }
 378                                 frags++;
 379                         }
 380
 381                         /* allocate new bio, limited by max BIO size, b=9945 */
 382                         bio = bio_alloc(GFP_NOIO, max(BIO_MAX_PAGES,
 383                                                       (npages - page_idx) *
 384                                                       blocks_per_page));
 385                         if (bio == NULL) {
 386                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 387                                        (npages - page_idx), blocks_per_page,
 388                                        (npages - page_idx) * blocks_per_page);
 389                                 rc = -ENOMEM;
 390                                 goto out;
 391                         }
 392
 393                         bio->bi_bdev = inode->i_sb->s_bdev;
 394                         bio->bi_sector = sector;
 395                         bio->bi_end_io = dio_complete_routine;
 396                         bio->bi_private = iobuf;
 397
 398                         rc = bio_add_page(bio, page,
 399                                           blocksize * nblocks, page_offset);
 400                         LASSERT (rc != 0);
 401                 }
 402         }
 403
 404         if (bio != NULL) {
 405                 record_start_io(iobuf, rw, bio->bi_size, exp);
 406                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 407                 if (rc >= 0) {
 408                         frags++;
 409                         rc = 0;
 410                 } else {
 411                         CERROR("Can't send bio: %d\n", rc);
 412                         record_finish_io(iobuf, rw, rc);
 413                 }
 414         }
 415
 416  out:
 417         wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0);
 418
 419         if (rw == OBD_BRW_READ) {
 420                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 421                                   hist[BRW_R_DIO_FRAGS],
 422                                  frags);
 423                 lprocfs_oh_tally_log2(&obd->u.filter.
 424                                        fo_filter_stats.hist[BRW_R_IO_TIME],
 425                                       jiffies - start_time);
 426                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 427                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 428                                           hist[BRW_R_DIO_FRAGS],
 429                                          frags);
 430                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 431                                              nid_brw_stats->hist[BRW_R_IO_TIME],
 432                                               jiffies - start_time);
 433                 }
 434         } else {
 435                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 436                                   hist[BRW_W_DIO_FRAGS], frags);
 437                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.
 438                                        hist[BRW_W_IO_TIME],
 439                                       jiffies - start_time);
 440                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 441                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 442                                           hist[BRW_W_DIO_FRAGS],
 443                                          frags);
 444                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 445                                              nid_brw_stats->hist[BRW_W_IO_TIME],
 446                                               jiffies - start_time);
 447                 }
 448         }
 449
 450         if (rc == 0)
 451                 rc = iobuf->dr_error;
 452         RETURN(rc);
 453 }
 454
 455 /* Must be called with i_mutex taken for writes; this will drop it */
 456 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 457                      struct obd_export *exp, struct iattr *attr,
 458                      struct obd_trans_info *oti, void **wait_handle)
 459 {
 460         struct obd_device *obd = exp->exp_obd;
 461         struct inode *inode = dchild->d_inode;
 462         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 463         int rc, rc2, create;
 464         struct semaphore *sem;
 465         ENTRY;
 466
 467         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 468                  iobuf->dr_npages, iobuf->dr_max_pages);
 469         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 470
 471         if (rw == OBD_BRW_READ) {
 472                 if (iobuf->dr_npages == 0)
 473                         RETURN(0);
 474                 create = 0;
 475                 sem = NULL;
 476         } else {
 477                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 478                 LASSERT(iobuf->dr_npages > 0);
 479                 create = 1;
 480                 sem = &obd->u.filter.fo_alloc_lock;
 481
 482                 lquota_enforce(filter_quota_interface_ref, obd,
 483                                iobuf->dr_ignore_quota);
 484         }
 485
 486         rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 487                                     iobuf->dr_npages, iobuf->dr_blocks,
 488                                     obdfilter_created_scratchpad, create, sem);
 489
 490         if (rw == OBD_BRW_WRITE) {
 491                 if (rc == 0) {
 492                         filter_tally(exp, iobuf->dr_pages,
 493                                      iobuf->dr_npages, iobuf->dr_blocks,
 494                                      blocks_per_page, 1);
 495                         if (attr->ia_size > i_size_read(inode))
 496                                 attr->ia_valid |= ATTR_SIZE;
 497                         rc = fsfilt_setattr(obd, dchild,
 498                                             oti->oti_handle, attr, 0);
 499                 }
 500
 501                 UNLOCK_INODE_MUTEX(inode);
 502
 503                 rc2 = filter_finish_transno(exp, inode, oti, 0, 0);
 504                 if (rc2 != 0) {
 505                         CERROR("can't close transaction: %d\n", rc2);
 506                         if (rc == 0)
 507                                 rc = rc2;
 508                 }
 509
 510                 rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
 511                                           wait_handle);
 512                 if (rc == 0)
 513                         rc = rc2;
 514                 if (rc != 0)
 515                         RETURN(rc);
 516         } else if (rc == 0) {
 517                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 518                              iobuf->dr_blocks, blocks_per_page, 0);
 519         }
 520
 521         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 522 }
 523
 524 /* See if there are unallocated parts in given file region */
 525 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 526 {
 527         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 528                 inode->i_mapping->a_ops->bmap;
 529         int j;
 530
 531         /* We can't know if we are overwriting or not */
 532         if (fs_bmap == NULL)
 533                 return 0;
 534
 535         offset >>= inode->i_blkbits;
 536         len >>= inode->i_blkbits;
 537
 538         for (j = 0; j <= len; j++)
 539                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 540                         return 0;
 541
 542         return 1;
 543 }
 544
 545 /*
 546  * interesting use cases on how it interacts with VM:
 547  *
 548  * - vm writeout -- shouldn't see our pages as we don't mark them dirty
 549  *   though vm can find partial page left dirty by truncate. in this
 550  *   usual writeout is used unless our write rewrite that page - then we
 551  *   drop PG_dirty with PG_lock held.
 552  *
 553  * - else?
 554  *
 555  */
 556 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 557                           int objcount, struct obd_ioobj *obj,
 558                           struct niobuf_remote *nb, int niocount,
 559                           struct niobuf_local *res, struct obd_trans_info *oti,
 560                           int rc)
 561 {
 562         struct niobuf_local *lnb;
 563         struct filter_iobuf *iobuf = NULL;
 564         struct lvfs_run_ctxt saved;
 565         struct fsfilt_objinfo fso;
 566         struct iattr iattr = { 0 };
 567         struct inode *inode = res->dentry->d_inode;
 568         unsigned long now = jiffies;
 569         int i, err, cleanup_phase = 0;
 570         struct obd_device *obd = exp->exp_obd;
 571         struct filter_obd *fo = &obd->u.filter;
 572         void *wait_handle;
 573         int total_size = 0;
 574         int rec_pending = 0;
 575         unsigned int qcids[MAXQUOTAS] = {0, 0};
 576         ENTRY;
 577
 578         LASSERT(oti != NULL);
 579         LASSERT(objcount == 1);
 580         LASSERT(current->journal_info == NULL);
 581
 582         if (rc != 0)
 583                 GOTO(cleanup, rc);
 584
 585         /* we try to get enough quota to write here, and let ldiskfs
 586          * decide if it is out of quota or not b=14783 */
 587         lquota_chkquota(filter_quota_interface_ref, obd, oa->o_uid,
 588                         oa->o_gid, niocount, &rec_pending, oti,
 589                         LQUOTA_FLAGS_BLK, (void *)inode, obj->ioo_bufcnt);
 590
 591         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 592         if (IS_ERR(iobuf))
 593                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 594         cleanup_phase = 1;
 595
 596         fso.fso_dentry = res->dentry;
 597         fso.fso_bufcnt = obj->ioo_bufcnt;
 598
 599         iobuf->dr_ignore_quota = 0;
 600         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 601                 loff_t this_size;
 602                 __u32 flags = lnb->flags;
 603
 604                 /* If overwriting an existing block, we don't need a grant */
 605                 if (!(flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC &&
 606                     filter_range_is_mapped(inode, lnb->offset, lnb->len))
 607                         lnb->rc = 0;
 608
 609                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 610                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 611                         continue;
 612                 }
 613
 614                 LASSERT(PageLocked(lnb->page));
 615                 LASSERT(!PageWriteback(lnb->page));
 616
 617                 /* preceding filemap_write_and_wait() should have clean pages */
 618                 if (fo->fo_writethrough_cache)
 619                         clear_page_dirty_for_io(lnb->page);
 620                 LASSERT(!PageDirty(lnb->page));
 621
 622                 SetPageUptodate(lnb->page);
 623
 624                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 625                 LASSERT (err == 0);
 626
 627                 total_size += lnb->len;
 628
 629                 /* we expect these pages to be in offset order, but we'll
 630                  * be forgiving */
 631                 this_size = lnb->offset + lnb->len;
 632                 if (this_size > iattr.ia_size)
 633                         iattr.ia_size = this_size;
 634
 635                 /* if one page is a write-back page from client cache and
 636                  * not from direct_io, or it's written by root, then mark
 637                  * the whole io request as ignore quota request, remote
 638                  * client can not break through quota. */
 639                 if (exp_connect_rmtclient(exp))
 640                         flags &= ~OBD_BRW_NOQUOTA;
 641                 if ((flags & OBD_BRW_NOQUOTA) ||
 642                     (flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 643                      OBD_BRW_FROM_GRANT)
 644                         iobuf->dr_ignore_quota = 1;
 645         }
 646
 647         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 648         cleanup_phase = 2;
 649
 650         DQUOT_INIT(inode);
 651
 652         LOCK_INODE_MUTEX(inode);
 653         fsfilt_check_slow(obd, now, "i_mutex");
 654         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 655                                            oti);
 656         if (IS_ERR(oti->oti_handle)) {
 657                 UNLOCK_INODE_MUTEX(inode);
 658                 rc = PTR_ERR(oti->oti_handle);
 659                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 660                        "error starting transaction: rc = %d\n", rc);
 661                 oti->oti_handle = NULL;
 662                 GOTO(cleanup, rc);
 663         }
 664         /* have to call fsfilt_commit() from this point on */
 665
 666         fsfilt_check_slow(obd, now, "brw_start");
 667
 668         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 669
 670         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 671          * then we will accept the UID+GID if sent by the client for
 672          * initializing the ownership of this inode.  We only allow this to
 673          * happen once (so clear these bits) and later only allow setattr. */
 674         if (inode->i_mode & S_ISUID)
 675                 i |= OBD_MD_FLUID;
 676         if (inode->i_mode & S_ISGID)
 677                 i |= OBD_MD_FLGID;
 678
 679         iattr_from_obdo(&iattr, oa, i);
 680         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 681                 unsigned int save;
 682
 683                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 684                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 685
 686                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 687
 688                 iattr.ia_valid |= ATTR_MODE;
 689                 iattr.ia_mode = inode->i_mode;
 690                 if (iattr.ia_valid & ATTR_UID)
 691                         iattr.ia_mode &= ~S_ISUID;
 692                 if (iattr.ia_valid & ATTR_GID)
 693                         iattr.ia_mode &= ~S_ISGID;
 694
 695                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 696
 697                 /* To avoid problems with quotas, UID and GID must be set
 698                  * in the inode before filter_direct_io() - see bug 10357. */
 699                 save = iattr.ia_valid;
 700                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 701                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle,&iattr,0);
 702                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 703                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 704                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 705         }
 706
 707         /* filter_direct_io drops i_mutex */
 708         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 709                               oti, &wait_handle);
 710         if (rc == 0)
 711                 obdo_from_inode(oa, inode,
 712                                 FILTER_VALID_FLAGS |OBD_MD_FLUID |OBD_MD_FLGID);
 713         else
 714                 obdo_from_inode(oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
 715
 716         lquota_getflag(filter_quota_interface_ref, obd, oa);
 717
 718         fsfilt_check_slow(obd, now, "direct_io");
 719
 720         err = fsfilt_commit_wait(obd, inode, wait_handle);
 721         if (err) {
 722                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 723                 rc = err;
 724         }
 725
 726         if (obd->obd_replayable && !rc)
 727                 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
 728                          "oti_transno "LPU64" last_committed "LPU64"\n",
 729                          oti->oti_transno, obd->obd_last_committed);
 730
 731         fsfilt_check_slow(obd, now, "commitrw commit");
 732
 733 cleanup:
 734         if (rec_pending)
 735                 lquota_pending_commit(filter_quota_interface_ref, obd, oa->o_uid,
 736                                       oa->o_gid, rec_pending, 1);
 737
 738         filter_grant_commit(exp, niocount, res);
 739
 740         switch (cleanup_phase) {
 741         case 2:
 742                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 743                 LASSERT(current->journal_info == NULL);
 744         case 1:
 745                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 746         case 0:
 747                 /*
 748                  * lnb->page automatically returns back into per-thread page
 749                  * pool (bug 5137)
 750                  */
 751                 f_dput(res->dentry);
 752         }
 753
 754         /* trigger quota pre-acquire */
 755         qcids[USRQUOTA] = oa->o_uid;
 756         qcids[GRPQUOTA] = oa->o_gid;
 757         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 758                             FSFILT_OP_CREATE);
 759         CDEBUG(err ? D_ERROR : D_QUOTA,
 760                "filter adjust qunit! (rc:%d)\n", err);
 761
 762         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 763                 if (lnb->page == NULL)
 764                         continue;
 765
 766                 LASSERT(PageLocked(lnb->page));
 767                 unlock_page(lnb->page);
 768
 769                 page_cache_release(lnb->page);
 770                 lnb->page = NULL;
 771         }
 772
 773         if (inode && (fo->fo_writethrough_cache == 0 ||
 774                         i_size_read(inode) > fo->fo_readcache_max_filesize))
 775                 filter_invalidate_cache(obd, obj, nb, inode);
 776
 777         up_read(&inode->i_alloc_sem);
 778
 779         RETURN(rc);
 780 }