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 (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011 Whamcloud, Inc.
  33  *
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * lustre/obdfilter/filter_io_26.c
  40  *
  41  * Author: Peter Braam <braam@clusterfs.com>
  42  * Author: Andreas Dilger <adilger@clusterfs.com>
  43  * Author: Phil Schwan <phil@clusterfs.com>
  44  */
  45
  46 #ifndef AUTOCONF_INCLUDED
  47 #include <linux/config.h>
  48 #endif
  49 #include <linux/module.h>
  50 #include <linux/pagemap.h> // XXX kill me soon
  51 #include <linux/version.h>
  52 #include <linux/buffer_head.h>
  53
  54 #define DEBUG_SUBSYSTEM S_FILTER
  55
  56 #include <obd_class.h>
  57 #include <lustre_fsfilt.h>
  58 #include <lustre_quota.h>
  59 #include "filter_internal.h"
  60
  61 /* 512byte block min */
  62 #define MAX_BLOCKS_PER_PAGE (CFS_PAGE_SIZE / 512)
  63 struct filter_iobuf {
  64         cfs_atomic_t       dr_numreqs;  /* number of reqs being processed */
  65         cfs_waitq_t        dr_wait;
  66         int                dr_max_pages;
  67         int                dr_npages;
  68         int                dr_error;
  69         struct page      **dr_pages;
  70         unsigned long     *dr_blocks;
  71         unsigned int       dr_ignore_quota:1;
  72         struct filter_obd *dr_filter;
  73 };
  74
  75 static void record_start_io(struct filter_iobuf *iobuf, int rw, int size,
  76                             struct obd_export *exp)
  77 {
  78         struct filter_obd *filter = iobuf->dr_filter;
  79
  80         cfs_atomic_inc(&iobuf->dr_numreqs);
  81
  82         if (rw == OBD_BRW_READ) {
  83                 cfs_atomic_inc(&filter->fo_r_in_flight);
  84                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST],
  85                                  cfs_atomic_read(&filter->fo_r_in_flight));
  86                 lprocfs_oh_tally_log2(&filter->
  87                                        fo_filter_stats.hist[BRW_R_DISK_IOSIZE],
  88                                       size);
  89                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
  90                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
  91                                          hist[BRW_R_RPC_HIST],
  92                                          cfs_atomic_read(&filter-> \
  93                                          fo_r_in_flight));
  94                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
  95                                          nid_brw_stats->hist[BRW_R_DISK_IOSIZE],
  96                                               size);
  97                 }
  98         } else {
  99                 cfs_atomic_inc(&filter->fo_w_in_flight);
 100                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
 101                                  cfs_atomic_read(&filter->fo_w_in_flight));
 102                 lprocfs_oh_tally_log2(&filter->
 103                                        fo_filter_stats.hist[BRW_W_DISK_IOSIZE],
 104                                       size);
 105                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 106                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 107                                           hist[BRW_W_RPC_HIST],
 108                                          cfs_atomic_read(&filter-> \
 109                                          fo_w_in_flight));
 110                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 111                                         nid_brw_stats->hist[BRW_W_DISK_IOSIZE],
 112                                               size);
 113                 }
 114         }
 115 }
 116
 117 static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
 118 {
 119         struct filter_obd *filter = iobuf->dr_filter;
 120
 121         /* CAVEAT EMPTOR: possibly in IRQ context
 122          * DO NOT record procfs stats here!!! */
 123
 124         if (rw == OBD_BRW_READ)
 125                 cfs_atomic_dec(&filter->fo_r_in_flight);
 126         else
 127                 cfs_atomic_dec(&filter->fo_w_in_flight);
 128
 129         if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs))
 130                 cfs_waitq_signal(&iobuf->dr_wait);
 131 }
 132
 133 #ifdef HAVE_BIO_ENDIO_2ARG
 134 #define DIO_RETURN(a)
 135 static void dio_complete_routine(struct bio *bio, int error)
 136 #else
 137 #define DIO_RETURN(a)   return(a)
 138 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 139 #endif
 140 {
 141         struct filter_iobuf *iobuf = bio->bi_private;
 142         struct bio_vec *bvl;
 143         int i;
 144
 145         /* CAVEAT EMPTOR: possibly in IRQ context
 146          * DO NOT record procfs stats here!!! */
 147
 148 #ifdef HAVE_BIO_ENDIO_2ARG
 149        /* The "bi_size" check was needed for kernels < 2.6.24 in order to
 150         * handle the case where a SCSI request error caused this callback
 151         * to be called before all of the biovecs had been processed.
 152         * Without this check the server thread will hang.  In newer kernels
 153         * the bio_end_io routine is never called for partial completions,
 154         * so this check is no longer needed. */
 155         if (bio->bi_size)                      /* Not complete */
 156                 DIO_RETURN(1);
 157 #endif
 158
 159         if (unlikely(iobuf == NULL)) {
 160                 CERROR("***** bio->bi_private is NULL!  This should never "
 161                        "happen.  Normally, I would crash here, but instead I "
 162                        "will dump the bio contents to the console.  Please "
 163                        "report this to <http://bugs.whamcloud.com/>, along "
 164                        "with any interesting messages leading up to this point "
 165                        "(like SCSI errors, perhaps).  Because bi_private is "
 166                        "NULL, I can't wake up the thread that initiated this "
 167                        "IO - you will probably have to reboot this node.\n");
 168                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 169                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 170                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 171                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 172                        bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
 173                        bio->bi_private);
 174                 DIO_RETURN(0);
 175         }
 176
 177         /* the check is outside of the cycle for performance reason -bzzz */
 178         if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
 179                 bio_for_each_segment(bvl, bio, i) {
 180                         if (likely(error == 0))
 181                                 SetPageUptodate(bvl->bv_page);
 182                         LASSERT(PageLocked(bvl->bv_page));
 183                         ClearPageConstant(bvl->bv_page);
 184                 }
 185                 record_finish_io(iobuf, OBD_BRW_READ, error);
 186         } else {
 187                 if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){
 188                         bio_for_each_segment(bvl, bio, i) {
 189                                 ClearPageConstant(bvl->bv_page);
 190                         }
 191                 }
 192                 record_finish_io(iobuf, OBD_BRW_WRITE, error);
 193         }
 194
 195         /* any real error is good enough -bzzz */
 196         if (error != 0 && iobuf->dr_error == 0)
 197                 iobuf->dr_error = error;
 198
 199         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 200          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 201          * mempool when serious on-disk fragmentation was encountered,
 202          * deadlocking the OST.  The bios are now released as soon as complete
 203          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 204         bio_put(bio);
 205         DIO_RETURN(0);
 206 }
 207
 208 static int can_be_merged(struct bio *bio, sector_t sector)
 209 {
 210         unsigned int size;
 211
 212         if (!bio)
 213                 return 0;
 214
 215         size = bio->bi_size >> 9;
 216         return bio->bi_sector + size == sector ? 1 : 0;
 217 }
 218
 219 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 220                                         int rw, int num_pages)
 221 {
 222         struct filter_iobuf *iobuf;
 223
 224         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 225
 226         OBD_ALLOC(iobuf, sizeof(*iobuf));
 227         if (iobuf == NULL)
 228                 goto failed_0;
 229
 230         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 231         if (iobuf->dr_pages == NULL)
 232                 goto failed_1;
 233
 234         OBD_ALLOC(iobuf->dr_blocks,
 235                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 236         if (iobuf->dr_blocks == NULL)
 237                 goto failed_2;
 238
 239         iobuf->dr_filter = filter;
 240         cfs_waitq_init(&iobuf->dr_wait);
 241         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 242         iobuf->dr_max_pages = num_pages;
 243         iobuf->dr_npages = 0;
 244         iobuf->dr_error = 0;
 245
 246         RETURN(iobuf);
 247
 248  failed_2:
 249         OBD_FREE(iobuf->dr_pages,
 250                  num_pages * sizeof(*iobuf->dr_pages));
 251  failed_1:
 252         OBD_FREE(iobuf, sizeof(*iobuf));
 253  failed_0:
 254         RETURN(ERR_PTR(-ENOMEM));
 255 }
 256
 257 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
 258 {
 259         iobuf->dr_npages = 0;
 260         iobuf->dr_error = 0;
 261         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 262 }
 263
 264 void filter_free_iobuf(struct filter_iobuf *iobuf)
 265 {
 266         int num_pages = iobuf->dr_max_pages;
 267
 268         filter_clear_iobuf(iobuf);
 269
 270         OBD_FREE(iobuf->dr_blocks,
 271                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 272         OBD_FREE(iobuf->dr_pages,
 273                  num_pages * sizeof(*iobuf->dr_pages));
 274         OBD_FREE_PTR(iobuf);
 275 }
 276
 277 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 278                       struct obd_trans_info *oti)
 279 {
 280         int thread_id = (oti && oti->oti_thread) ?
 281                         oti->oti_thread->t_id : -1;
 282
 283         if (unlikely(thread_id < 0)) {
 284                 filter_free_iobuf(iobuf);
 285                 return;
 286         }
 287
 288         LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
 289                  "iobuf mismatch for thread %d: pool %p iobuf %p\n",
 290                  thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
 291         filter_clear_iobuf(iobuf);
 292 }
 293
 294 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 295                           struct inode *inode, struct page *page)
 296 {
 297         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 298         iobuf->dr_pages[iobuf->dr_npages++] = page;
 299
 300         return 0;
 301 }
 302
 303 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 304                   struct filter_iobuf *iobuf, int rw)
 305 {
 306         struct obd_device *obd = exp->exp_obd;
 307         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 308         struct page  **pages = iobuf->dr_pages;
 309         int            npages = iobuf->dr_npages;
 310         unsigned long *blocks = iobuf->dr_blocks;
 311         int            total_blocks = npages * blocks_per_page;
 312         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 313         unsigned int   blocksize = inode->i_sb->s_blocksize;
 314         struct bio    *bio = NULL;
 315         int            frags = 0;
 316         unsigned long  start_time = jiffies;
 317         struct page   *page;
 318         unsigned int   page_offset;
 319         sector_t       sector;
 320         int            nblocks;
 321         int            block_idx;
 322         int            page_idx;
 323         int            i;
 324         int            rc = 0;
 325         ENTRY;
 326
 327         LASSERT(iobuf->dr_npages == npages);
 328         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 329
 330         for (page_idx = 0, block_idx = 0;
 331              page_idx < npages;
 332              page_idx++, block_idx += blocks_per_page) {
 333
 334                 page = pages[page_idx];
 335                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 336
 337                 for (i = 0, page_offset = 0;
 338                      i < blocks_per_page;
 339                      i += nblocks, page_offset += blocksize * nblocks) {
 340
 341                         nblocks = 1;
 342
 343                         if (blocks[block_idx + i] == 0) {  /* hole */
 344                                 LASSERT(rw == OBD_BRW_READ);
 345                                 memset(kmap(page) + page_offset, 0, blocksize);
 346                                 kunmap(page);
 347                                 continue;
 348                         }
 349
 350                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 351
 352                         /* Additional contiguous file blocks? */
 353                         while (i + nblocks < blocks_per_page &&
 354                                (sector + (nblocks << sector_bits)) ==
 355                                ((sector_t)blocks[block_idx + i + nblocks] <<
 356                                 sector_bits))
 357                                 nblocks++;
 358
 359                         /* I only set the page to be constant only if it
 360                          * is mapped to a contiguous underlying disk block(s).
 361                          * It will then make sure the corresponding device
 362                          * cache of raid5 will be overwritten by this page.
 363                          * - jay */
 364                         if ((rw == OBD_BRW_WRITE) &&
 365                             (nblocks == blocks_per_page) &&
 366                             mapping_cap_page_constant_write(inode->i_mapping))
 367                                SetPageConstant(page);
 368
 369                         if (bio != NULL &&
 370                             can_be_merged(bio, sector) &&
 371                             bio_add_page(bio, page,
 372                                          blocksize * nblocks, page_offset) != 0)
 373                                 continue;       /* added this frag OK */
 374
 375                         if (bio != NULL) {
 376                                 struct request_queue *q =
 377                                         bdev_get_queue(bio->bi_bdev);
 378
 379                                 /* Dang! I have to fragment this I/O */
 380                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 381                                        "sectors %d(%d) psg %d(%d) hsg %d(%d) "
 382                                        "sector %llu next %llu\n",
 383                                        bio->bi_size,
 384                                        bio->bi_vcnt, bio->bi_max_vecs,
 385                                        bio->bi_size >> 9, queue_max_sectors(q),
 386                                        bio_phys_segments(q, bio),
 387                                        queue_max_phys_segments(q),
 388                                        bio_hw_segments(q, bio),
 389                                        queue_max_hw_segments(q),
 390                                        (unsigned long long)bio->bi_sector,
 391                                        (unsigned long long)sector);
 392
 393                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 394                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 395                                 if (rc < 0) {
 396                                         CERROR("Can't send bio: %d\n", rc);
 397                                         record_finish_io(iobuf, rw, rc);
 398                                         goto out;
 399                                 }
 400                                 frags++;
 401                         }
 402
 403                         /* allocate new bio, limited by max BIO size, b=9945 */
 404                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 405                                                       (npages - page_idx) *
 406                                                       blocks_per_page));
 407                         if (bio == NULL) {
 408                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 409                                        (npages - page_idx), blocks_per_page,
 410                                        (npages - page_idx) * blocks_per_page);
 411                                 rc = -ENOMEM;
 412                                 goto out;
 413                         }
 414
 415                         bio->bi_bdev = inode->i_sb->s_bdev;
 416                         bio->bi_sector = sector;
 417                         bio->bi_end_io = dio_complete_routine;
 418                         bio->bi_private = iobuf;
 419
 420                         rc = bio_add_page(bio, page,
 421                                           blocksize * nblocks, page_offset);
 422                         LASSERT (rc != 0);
 423                 }
 424         }
 425
 426         if (bio != NULL) {
 427                 record_start_io(iobuf, rw, bio->bi_size, exp);
 428                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 429                 if (rc >= 0) {
 430                         frags++;
 431                         rc = 0;
 432                 } else {
 433                         CERROR("Can't send bio: %d\n", rc);
 434                         record_finish_io(iobuf, rw, rc);
 435                 }
 436         }
 437
 438  out:
 439         cfs_wait_event(iobuf->dr_wait,
 440                        cfs_atomic_read(&iobuf->dr_numreqs) == 0);
 441
 442         if (rw == OBD_BRW_READ) {
 443                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 444                                   hist[BRW_R_DIO_FRAGS],
 445                                  frags);
 446                 lprocfs_oh_tally_log2(&obd->u.filter.
 447                                        fo_filter_stats.hist[BRW_R_IO_TIME],
 448                                       jiffies - start_time);
 449                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 450                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 451                                           hist[BRW_R_DIO_FRAGS],
 452                                          frags);
 453                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 454                                              nid_brw_stats->hist[BRW_R_IO_TIME],
 455                                               jiffies - start_time);
 456                 }
 457         } else {
 458                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 459                                   hist[BRW_W_DIO_FRAGS], frags);
 460                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.
 461                                        hist[BRW_W_IO_TIME],
 462                                       jiffies - start_time);
 463                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 464                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 465                                           hist[BRW_W_DIO_FRAGS],
 466                                          frags);
 467                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 468                                              nid_brw_stats->hist[BRW_W_IO_TIME],
 469                                               jiffies - start_time);
 470                 }
 471         }
 472
 473         if (rc == 0)
 474                 rc = iobuf->dr_error;
 475         RETURN(rc);
 476 }
 477
 478 /* Must be called with i_mutex taken for writes; this will drop it */
 479 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 480                      struct obd_export *exp, struct iattr *attr,
 481                      struct obd_trans_info *oti, void **wait_handle)
 482 {
 483         struct obd_device *obd = exp->exp_obd;
 484         struct inode *inode = dchild->d_inode;
 485         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 486         int rc, rc2, create;
 487         cfs_semaphore_t *sem;
 488         ENTRY;
 489
 490         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 491                  iobuf->dr_npages, iobuf->dr_max_pages);
 492         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 493
 494         if (rw == OBD_BRW_READ) {
 495                 if (iobuf->dr_npages == 0)
 496                         RETURN(0);
 497                 create = 0;
 498                 sem = NULL;
 499         } else {
 500                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 501                 LASSERT(iobuf->dr_npages > 0);
 502                 create = 1;
 503                 sem = &obd->u.filter.fo_alloc_lock;
 504
 505                 lquota_enforce(filter_quota_interface_ref, obd,
 506                                iobuf->dr_ignore_quota);
 507         }
 508
 509         if (rw == OBD_BRW_WRITE &&
 510             OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
 511                 rc = -ENOSPC;
 512         } else {
 513                 rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 514                                     iobuf->dr_npages, iobuf->dr_blocks,
 515                                     obdfilter_created_scratchpad, create, sem);
 516         }
 517
 518         if (rw == OBD_BRW_WRITE) {
 519                 if (rc == 0) {
 520                         filter_tally(exp, iobuf->dr_pages,
 521                                      iobuf->dr_npages, iobuf->dr_blocks,
 522                                      blocks_per_page, 1);
 523                         if (attr->ia_size > i_size_read(inode))
 524                                 attr->ia_valid |= ATTR_SIZE;
 525                         rc = fsfilt_setattr(obd, dchild,
 526                                             oti->oti_handle, attr, 0);
 527                 }
 528
 529                 UNLOCK_INODE_MUTEX(inode);
 530
 531                 /* Force commit to make the just-deleted blocks
 532                  * reusable. LU-456 */
 533                 if (rc == -ENOSPC) {
 534                         fsfilt_commit(obd, inode, oti->oti_handle, 1);
 535                         RETURN(rc);
 536                 }
 537
 538                 rc2 = filter_finish_transno(exp, inode, oti, 0, 0);
 539                 if (rc2 != 0) {
 540                         CERROR("can't close transaction: %d\n", rc2);
 541                         if (rc == 0)
 542                                 rc = rc2;
 543                 }
 544
 545                 if (wait_handle)
 546                         rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
 547                                                   wait_handle);
 548                 else
 549                         rc2 = fsfilt_commit(obd, inode, oti->oti_handle, 0);
 550                 if (rc == 0)
 551                         rc = rc2;
 552                 if (rc != 0)
 553                         RETURN(rc);
 554         } else if (rc == 0) {
 555                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 556                              iobuf->dr_blocks, blocks_per_page, 0);
 557         }
 558
 559         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 560 }
 561
 562 /* See if there are unallocated parts in given file region */
 563 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 564 {
 565         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 566                 inode->i_mapping->a_ops->bmap;
 567         int j;
 568
 569         /* We can't know if we are overwriting or not */
 570         if (fs_bmap == NULL)
 571                 return 0;
 572
 573         offset >>= inode->i_blkbits;
 574         len >>= inode->i_blkbits;
 575
 576         for (j = 0; j <= len; j++)
 577                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 578                         return 0;
 579
 580         return 1;
 581 }
 582
 583 /*
 584  * interesting use cases on how it interacts with VM:
 585  *
 586  * - vm writeout -- shouldn't see our pages as we don't mark them dirty
 587  *   though vm can find partial page left dirty by truncate. in this
 588  *   usual writeout is used unless our write rewrite that page - then we
 589  *   drop PG_dirty with PG_lock held.
 590  *
 591  * - else?
 592  *
 593  */
 594 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 595                           int objcount, struct obd_ioobj *obj,
 596                           struct niobuf_remote *nb, int niocount,
 597                           struct niobuf_local *res, struct obd_trans_info *oti,
 598                           int rc)
 599 {
 600         struct niobuf_local *lnb;
 601         struct filter_iobuf *iobuf = NULL;
 602         struct lvfs_run_ctxt saved;
 603         struct fsfilt_objinfo fso;
 604         struct iattr iattr = { 0 };
 605         struct inode *inode = res->dentry->d_inode;
 606         unsigned long now = jiffies;
 607         int i, err, cleanup_phase = 0;
 608         struct obd_device *obd = exp->exp_obd;
 609         struct filter_obd *fo = &obd->u.filter;
 610         void *wait_handle = NULL;
 611         int total_size = 0;
 612         unsigned int qcids[MAXQUOTAS] = { oa->o_uid, oa->o_gid };
 613         int rec_pending[MAXQUOTAS] = { 0, 0 }, quota_pages = 0;
 614         int sync_journal_commit = obd->u.filter.fo_syncjournal;
 615         int retries = 0;
 616         ENTRY;
 617
 618         LASSERT(oti != NULL);
 619         LASSERT(objcount == 1);
 620         LASSERT(current->journal_info == NULL);
 621
 622         if (rc != 0)
 623                 GOTO(cleanup, rc);
 624
 625         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 626         if (IS_ERR(iobuf))
 627                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 628         cleanup_phase = 1;
 629
 630         fso.fso_dentry = res->dentry;
 631         fso.fso_bufcnt = obj->ioo_bufcnt;
 632
 633         iobuf->dr_ignore_quota = 0;
 634         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 635                 loff_t this_size;
 636                 __u32 flags = lnb->flags;
 637
 638                 if (filter_range_is_mapped(inode, lnb->offset, lnb->len)) {
 639                         /* If overwriting an existing block,
 640                          * we don't need a grant */
 641                         if (!(flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC)
 642                                 lnb->rc = 0;
 643                 } else {
 644                         quota_pages++;
 645                 }
 646
 647                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 648                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 649                         continue;
 650                 }
 651
 652                 LASSERT(PageLocked(lnb->page));
 653                 LASSERT(!PageWriteback(lnb->page));
 654
 655                 /* since write & truncate are serialized by the i_alloc_sem,
 656                  * even partial truncate should not leave dirty pages in
 657                  * the page cache */
 658                 LASSERT(!PageDirty(lnb->page));
 659
 660                 SetPageUptodate(lnb->page);
 661
 662                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 663                 LASSERT (err == 0);
 664
 665                 total_size += lnb->len;
 666
 667                 /* we expect these pages to be in offset order, but we'll
 668                  * be forgiving */
 669                 this_size = lnb->offset + lnb->len;
 670                 if (this_size > iattr.ia_size)
 671                         iattr.ia_size = this_size;
 672
 673                 /* if one page is a write-back page from client cache and
 674                  * not from direct_io, or it's written by root, then mark
 675                  * the whole io request as ignore quota request, remote
 676                  * client can not break through quota. */
 677                 if (exp_connect_rmtclient(exp))
 678                         flags &= ~OBD_BRW_NOQUOTA;
 679                 if ((flags & OBD_BRW_NOQUOTA) ||
 680                     (flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 681                      OBD_BRW_FROM_GRANT)
 682                         iobuf->dr_ignore_quota = 1;
 683
 684                 if (!(lnb->flags & OBD_BRW_ASYNC)) {
 685                         sync_journal_commit = 1;
 686                 }
 687         }
 688
 689         /* we try to get enough quota to write here, and let ldiskfs
 690          * decide if it is out of quota or not b=14783 */
 691         rc = lquota_chkquota(filter_quota_interface_ref, obd, exp, qcids,
 692                              rec_pending, quota_pages, oti, LQUOTA_FLAGS_BLK,
 693                              (void *)inode, obj->ioo_bufcnt);
 694         if (rc == -ENOTCONN)
 695                 GOTO(cleanup, rc);
 696
 697         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 698         cleanup_phase = 2;
 699
 700         ll_vfs_dq_init(inode);
 701         fsfilt_check_slow(obd, now, "quota init");
 702
 703 retry:
 704         LOCK_INODE_MUTEX(inode);
 705         fsfilt_check_slow(obd, now, "i_mutex");
 706         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 707                                            oti);
 708         if (IS_ERR(oti->oti_handle)) {
 709                 UNLOCK_INODE_MUTEX(inode);
 710                 rc = PTR_ERR(oti->oti_handle);
 711                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 712                        "error starting transaction: rc = %d\n", rc);
 713                 oti->oti_handle = NULL;
 714                 GOTO(cleanup, rc);
 715         }
 716         /* have to call fsfilt_commit() from this point on */
 717
 718         fsfilt_check_slow(obd, now, "brw_start");
 719
 720         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 721
 722         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 723          * then we will accept the UID+GID if sent by the client for
 724          * initializing the ownership of this inode.  We only allow this to
 725          * happen once (so clear these bits) and later only allow setattr. */
 726         if (inode->i_mode & S_ISUID)
 727                 i |= OBD_MD_FLUID;
 728         if (inode->i_mode & S_ISGID)
 729                 i |= OBD_MD_FLGID;
 730
 731         iattr_from_obdo(&iattr, oa, i);
 732         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 733                 unsigned int save;
 734
 735                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 736                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 737
 738                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 739
 740                 iattr.ia_valid |= ATTR_MODE;
 741                 iattr.ia_mode = inode->i_mode;
 742                 if (iattr.ia_valid & ATTR_UID)
 743                         iattr.ia_mode &= ~S_ISUID;
 744                 if (iattr.ia_valid & ATTR_GID)
 745                         iattr.ia_mode &= ~S_ISGID;
 746
 747                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 748
 749                 /* To avoid problems with quotas, UID and GID must be set
 750                  * in the inode before filter_direct_io() - see bug 10357. */
 751                 save = iattr.ia_valid;
 752                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 753                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle,&iattr,0);
 754                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 755                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 756                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 757         }
 758
 759         /* filter_direct_io drops i_mutex */
 760         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 761                               oti, sync_journal_commit ? &wait_handle : NULL);
 762         if (rc == -ENOSPC && retries++ < 3) {
 763                 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
 764                        retries);
 765                 oti->oti_handle = NULL;
 766                 fsfilt_check_slow(obd, now, "direct_io");
 767                 goto retry;
 768         }
 769
 770         obdo_from_inode(oa, inode, NULL, rc == 0 ? FILTER_VALID_FLAGS : 0 |
 771                                                    OBD_MD_FLUID |OBD_MD_FLGID);
 772
 773         lquota_getflag(filter_quota_interface_ref, obd, oa);
 774
 775         fsfilt_check_slow(obd, now, "direct_io");
 776
 777         if (wait_handle)
 778                 err = fsfilt_commit_wait(obd, inode, wait_handle);
 779         else
 780                 err = 0;
 781
 782         if (err) {
 783                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 784                 if (rc == 0)
 785                         rc = err;
 786         }
 787
 788         if (obd->obd_replayable && !rc && wait_handle)
 789                 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
 790                          "oti_transno "LPU64" last_committed "LPU64"\n",
 791                          oti->oti_transno, obd->obd_last_committed);
 792
 793         fsfilt_check_slow(obd, now, "commitrw commit");
 794
 795 cleanup:
 796         lquota_pending_commit(filter_quota_interface_ref, obd, qcids,
 797                               rec_pending, 1);
 798
 799         filter_grant_commit(exp, niocount, res);
 800
 801         switch (cleanup_phase) {
 802         case 2:
 803                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 804                 LASSERT(current->journal_info == NULL);
 805         case 1:
 806                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 807         case 0:
 808                 /*
 809                  * lnb->page automatically returns back into per-thread page
 810                  * pool (bug 5137)
 811                  */
 812                  break;
 813         }
 814
 815         /* trigger quota pre-acquire */
 816         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 817                             FSFILT_OP_CREATE);
 818         CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 819                "(rc:%d, uid:%u, gid:%u)\n",
 820                err, qcids[USRQUOTA], qcids[GRPQUOTA]);
 821         if (qcids[USRQUOTA] != oa->o_uid || qcids[GRPQUOTA] != oa->o_gid) {
 822                 qcids[USRQUOTA] = oa->o_uid;
 823                 qcids[GRPQUOTA] = oa->o_gid;
 824                 err = lquota_adjust(filter_quota_interface_ref, obd, qcids,
 825                                     NULL, rc, FSFILT_OP_CREATE);
 826                 CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 827                        "(rc:%d, uid:%u, gid:%u)\n",
 828                        err, qcids[USRQUOTA], qcids[GRPQUOTA]);
 829         }
 830
 831         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 832                 if (lnb->page == NULL)
 833                         continue;
 834
 835                 if (rc)
 836                         /* If the write has failed, the page cache may
 837                          * not be consitent with what is on disk, so
 838                          * force pages to be reread next time it is
 839                          * accessed */
 840                         ClearPageUptodate(lnb->page);
 841
 842                 LASSERT(PageLocked(lnb->page));
 843                 unlock_page(lnb->page);
 844
 845                 page_cache_release(lnb->page);
 846                 lnb->page = NULL;
 847         }
 848         f_dput(res->dentry);
 849
 850         if (inode) {
 851                 if (fo->fo_writethrough_cache == 0 ||
 852                     i_size_read(inode) > fo->fo_readcache_max_filesize)
 853                         filter_release_cache(obd, obj, nb, inode);
 854                 up_read(&inode->i_alloc_sem);
 855         }
 856
 857         RETURN(rc);
 858 }