lustre/obdfilter/filter_io_26.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  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 #include <linux/module.h>
  44 #include <linux/pagemap.h> // XXX kill me soon
  45 #include <linux/version.h>
  46 #include <linux/buffer_head.h>
  47
  48 #define DEBUG_SUBSYSTEM S_FILTER
  49
  50 #include <obd_class.h>
  51 #include <lustre_fsfilt.h>
  52 #include "filter_internal.h"
  53
  54 /* 512byte block min */
  55 #define MAX_BLOCKS_PER_PAGE (CFS_PAGE_SIZE / 512)
  56
  57 static void record_start_io(struct filter_iobuf *iobuf, int rw, int size,
  58                             struct obd_export *exp)
  59 {
  60         struct filter_obd *filter = iobuf->dr_filter;
  61
  62         cfs_atomic_inc(&iobuf->dr_numreqs);
  63
  64         if (rw == OBD_BRW_READ) {
  65                 cfs_atomic_inc(&filter->fo_r_in_flight);
  66                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST],
  67                                  cfs_atomic_read(&filter->fo_r_in_flight));
  68                 lprocfs_oh_tally_log2(&filter->
  69                                        fo_filter_stats.hist[BRW_R_DISK_IOSIZE],
  70                                       size);
  71                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
  72                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
  73                                          hist[BRW_R_RPC_HIST],
  74                                          cfs_atomic_read(&filter-> \
  75                                          fo_r_in_flight));
  76                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
  77                                          nid_brw_stats->hist[BRW_R_DISK_IOSIZE],
  78                                               size);
  79                 }
  80         } else {
  81                 cfs_atomic_inc(&filter->fo_w_in_flight);
  82                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
  83                                  cfs_atomic_read(&filter->fo_w_in_flight));
  84                 lprocfs_oh_tally_log2(&filter->
  85                                        fo_filter_stats.hist[BRW_W_DISK_IOSIZE],
  86                                       size);
  87                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
  88                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
  89                                           hist[BRW_W_RPC_HIST],
  90                                          cfs_atomic_read(&filter-> \
  91                                          fo_w_in_flight));
  92                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
  93                                         nid_brw_stats->hist[BRW_W_DISK_IOSIZE],
  94                                               size);
  95                 }
  96         }
  97 }
  98
  99 static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
 100 {
 101         struct filter_obd *filter = iobuf->dr_filter;
 102
 103         /* CAVEAT EMPTOR: possibly in IRQ context
 104          * DO NOT record procfs stats here!!! */
 105
 106         if (rw == OBD_BRW_READ)
 107                 cfs_atomic_dec(&filter->fo_r_in_flight);
 108         else
 109                 cfs_atomic_dec(&filter->fo_w_in_flight);
 110
 111         if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs))
 112                 cfs_waitq_signal(&iobuf->dr_wait);
 113 }
 114
 115 #ifdef HAVE_BIO_ENDIO_2ARG
 116 #define DIO_RETURN(a)
 117 static void dio_complete_routine(struct bio *bio, int error)
 118 #else
 119 #define DIO_RETURN(a)   return(a)
 120 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 121 #endif
 122 {
 123         struct filter_iobuf *iobuf = bio->bi_private;
 124         struct bio_vec *bvl;
 125         int i;
 126
 127         /* CAVEAT EMPTOR: possibly in IRQ context
 128          * DO NOT record procfs stats here!!! */
 129
 130 #ifdef HAVE_BIO_ENDIO_2ARG
 131        /* The "bi_size" check was needed for kernels < 2.6.24 in order to
 132         * handle the case where a SCSI request error caused this callback
 133         * to be called before all of the biovecs had been processed.
 134         * Without this check the server thread will hang.  In newer kernels
 135         * the bio_end_io routine is never called for partial completions,
 136         * so this check is no longer needed. */
 137         if (bio->bi_size)                      /* Not complete */
 138                 DIO_RETURN(1);
 139 #endif
 140
 141         if (unlikely(iobuf == NULL)) {
 142                 CERROR("***** bio->bi_private is NULL!  This should never "
 143                        "happen.  Normally, I would crash here, but instead I "
 144                        "will dump the bio contents to the console.  Please "
 145                        "report this to <http://bugs.whamcloud.com/>, along "
 146                        "with any interesting messages leading up to this point "
 147                        "(like SCSI errors, perhaps).  Because bi_private is "
 148                        "NULL, I can't wake up the thread that initiated this "
 149                        "IO - you will probably have to reboot this node.\n");
 150                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 151                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 152                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 153                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 154                        bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
 155                        bio->bi_private);
 156                 DIO_RETURN(0);
 157         }
 158
 159         /* the check is outside of the cycle for performance reason -bzzz */
 160         if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
 161                 bio_for_each_segment(bvl, bio, i) {
 162                         if (likely(error == 0))
 163                                 SetPageUptodate(bvl->bv_page);
 164                         LASSERT(PageLocked(bvl->bv_page));
 165                         ClearPageConstant(bvl->bv_page);
 166                 }
 167                 record_finish_io(iobuf, OBD_BRW_READ, error);
 168         } else {
 169                 if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){
 170                         bio_for_each_segment(bvl, bio, i) {
 171                                 ClearPageConstant(bvl->bv_page);
 172                         }
 173                 }
 174                 record_finish_io(iobuf, OBD_BRW_WRITE, error);
 175         }
 176
 177         /* any real error is good enough -bzzz */
 178         if (error != 0 && iobuf->dr_error == 0)
 179                 iobuf->dr_error = error;
 180
 181         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 182          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 183          * mempool when serious on-disk fragmentation was encountered,
 184          * deadlocking the OST.  The bios are now released as soon as complete
 185          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 186         bio_put(bio);
 187         DIO_RETURN(0);
 188 }
 189
 190 static int can_be_merged(struct bio *bio, sector_t sector)
 191 {
 192         unsigned int size;
 193
 194         if (!bio)
 195                 return 0;
 196
 197         size = bio->bi_size >> 9;
 198         return bio->bi_sector + size == sector ? 1 : 0;
 199 }
 200
 201 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 202                                         int rw, int num_pages)
 203 {
 204         struct filter_iobuf *iobuf;
 205
 206         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 207
 208         OBD_ALLOC(iobuf, sizeof(*iobuf));
 209         if (iobuf == NULL)
 210                 goto failed_0;
 211
 212         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 213         if (iobuf->dr_pages == NULL)
 214                 goto failed_1;
 215
 216         OBD_ALLOC(iobuf->dr_blocks,
 217                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 218         if (iobuf->dr_blocks == NULL)
 219                 goto failed_2;
 220
 221         CFS_INIT_HLIST_NODE(&iobuf->dr_hlist);
 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         LASSERT(cfs_hlist_unhashed(&iobuf->dr_hlist));
 254         OBD_FREE(iobuf->dr_blocks,
 255                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 256         OBD_FREE(iobuf->dr_pages,
 257                  num_pages * sizeof(*iobuf->dr_pages));
 258         OBD_FREE_PTR(iobuf);
 259 }
 260
 261 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 262                       struct obd_trans_info *oti)
 263 {
 264         int thread_id = (oti && oti->oti_thread) ?
 265                         oti->oti_thread->t_id : -1;
 266
 267         if (unlikely(thread_id < 0)) {
 268                 filter_free_iobuf(iobuf);
 269                 return;
 270         }
 271
 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                                 struct request_queue *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) "
 363                                        "sector %llu next %llu\n",
 364                                        bio->bi_size,
 365                                        bio->bi_vcnt, bio->bi_max_vecs,
 366                                        bio->bi_size >> 9, queue_max_sectors(q),
 367                                        bio_phys_segments(q, bio),
 368                                        queue_max_phys_segments(q),
 369                                        bio_hw_segments(q, bio),
 370                                        queue_max_hw_segments(q),
 371                                        (unsigned long long)bio->bi_sector,
 372                                        (unsigned long long)sector);
 373
 374                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 375                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 376                                 if (rc < 0) {
 377                                         CERROR("Can't send bio: %d\n", rc);
 378                                         record_finish_io(iobuf, rw, rc);
 379                                         goto out;
 380                                 }
 381                                 frags++;
 382                         }
 383
 384                         /* allocate new bio, limited by max BIO size, b=9945 */
 385                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 386                                                       (npages - page_idx) *
 387                                                       blocks_per_page));
 388                         if (bio == NULL) {
 389                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 390                                        (npages - page_idx), blocks_per_page,
 391                                        (npages - page_idx) * blocks_per_page);
 392                                 rc = -ENOMEM;
 393                                 goto out;
 394                         }
 395
 396                         bio->bi_bdev = inode->i_sb->s_bdev;
 397                         bio->bi_sector = sector;
 398                         bio->bi_rw = (rw == OBD_BRW_READ ? READ : WRITE);
 399                         bio->bi_end_io = dio_complete_routine;
 400                         bio->bi_private = iobuf;
 401
 402                         rc = bio_add_page(bio, page,
 403                                           blocksize * nblocks, page_offset);
 404                         LASSERT (rc != 0);
 405                 }
 406         }
 407
 408         if (bio != NULL) {
 409                 record_start_io(iobuf, rw, bio->bi_size, exp);
 410                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 411                 if (rc >= 0) {
 412                         frags++;
 413                         rc = 0;
 414                 } else {
 415                         CERROR("Can't send bio: %d\n", rc);
 416                         record_finish_io(iobuf, rw, rc);
 417                 }
 418         }
 419
 420  out:
 421         cfs_wait_event(iobuf->dr_wait,
 422                        cfs_atomic_read(&iobuf->dr_numreqs) == 0);
 423
 424         if (rw == OBD_BRW_READ) {
 425                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 426                                   hist[BRW_R_DIO_FRAGS],
 427                                  frags);
 428                 lprocfs_oh_tally_log2(&obd->u.filter.
 429                                        fo_filter_stats.hist[BRW_R_IO_TIME],
 430                                       jiffies - start_time);
 431                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 432                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 433                                           hist[BRW_R_DIO_FRAGS],
 434                                          frags);
 435                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 436                                              nid_brw_stats->hist[BRW_R_IO_TIME],
 437                                               jiffies - start_time);
 438                 }
 439         } else {
 440                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 441                                   hist[BRW_W_DIO_FRAGS], frags);
 442                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.
 443                                        hist[BRW_W_IO_TIME],
 444                                       jiffies - start_time);
 445                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 446                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 447                                           hist[BRW_W_DIO_FRAGS],
 448                                          frags);
 449                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 450                                              nid_brw_stats->hist[BRW_W_IO_TIME],
 451                                               jiffies - start_time);
 452                 }
 453         }
 454
 455         if (rc == 0)
 456                 rc = iobuf->dr_error;
 457         RETURN(rc);
 458 }
 459
 460 /* Must be called with i_mutex taken for writes; this will drop it */
 461 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 462                      struct obd_export *exp, struct iattr *attr,
 463                      struct obd_trans_info *oti, void **wait_handle)
 464 {
 465         struct obd_device *obd = exp->exp_obd;
 466         struct inode *inode = dchild->d_inode;
 467         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 468         int rc, rc2, create;
 469         cfs_mutex_t *mutex;
 470         ENTRY;
 471
 472         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 473                  iobuf->dr_npages, iobuf->dr_max_pages);
 474         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 475
 476         if (rw == OBD_BRW_READ) {
 477                 if (iobuf->dr_npages == 0)
 478                         RETURN(0);
 479                 create = 0;
 480                 mutex = NULL;
 481         } else {
 482                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 483                 LASSERT(iobuf->dr_npages > 0);
 484                 create = 1;
 485                 mutex = &obd->u.filter.fo_alloc_lock;
 486         }
 487
 488         if (rw == OBD_BRW_WRITE &&
 489             OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
 490                 rc = -ENOSPC;
 491         } else {
 492                 rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 493                                     iobuf->dr_npages, iobuf->dr_blocks,
 494                                     obdfilter_created_scratchpad, create, mutex);
 495         }
 496
 497         if (rw == OBD_BRW_WRITE) {
 498                 if (rc == 0) {
 499                         filter_tally(exp, iobuf->dr_pages,
 500                                      iobuf->dr_npages, iobuf->dr_blocks,
 501                                      blocks_per_page, 1);
 502                         if (attr->ia_size > i_size_read(inode))
 503                                 attr->ia_valid |= ATTR_SIZE;
 504                         rc = fsfilt_setattr(obd, dchild,
 505                                             oti->oti_handle, attr, 0);
 506                 }
 507
 508                 mutex_unlock(&inode->i_mutex);
 509
 510                 /* Force commit to make the just-deleted blocks
 511                  * reusable. LU-456 */
 512                 if (rc == -ENOSPC) {
 513                         fsfilt_commit(obd, inode, oti->oti_handle, 1);
 514                         RETURN(rc);
 515                 }
 516
 517                 rc2 = filter_finish_transno(exp, inode, oti, 0, 0);
 518                 if (rc2 != 0) {
 519                         CERROR("can't close transaction: %d\n", rc2);
 520                         if (rc == 0)
 521                                 rc = rc2;
 522                 }
 523
 524                 if (wait_handle)
 525                         rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
 526                                                   wait_handle);
 527                 else
 528                         rc2 = fsfilt_commit(obd, inode, oti->oti_handle, 0);
 529                 if (rc == 0)
 530                         rc = rc2;
 531                 if (rc != 0)
 532                         RETURN(rc);
 533         } else if (rc == 0) {
 534                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 535                              iobuf->dr_blocks, blocks_per_page, 0);
 536         }
 537
 538         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 539 }
 540
 541 /* See if there are unallocated parts in given file region */
 542 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 543 {
 544         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 545                 inode->i_mapping->a_ops->bmap;
 546         int j;
 547
 548         /* We can't know if we are overwriting or not */
 549         if (fs_bmap == NULL)
 550                 return 0;
 551
 552         offset >>= inode->i_blkbits;
 553         len >>= inode->i_blkbits;
 554
 555         for (j = 0; j <= len; j++)
 556                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 557                         return 0;
 558
 559         return 1;
 560 }
 561
 562 /*
 563  * interesting use cases on how it interacts with VM:
 564  *
 565  * - vm writeout -- shouldn't see our pages as we don't mark them dirty
 566  *   though vm can find partial page left dirty by truncate. in this
 567  *   usual writeout is used unless our write rewrite that page - then we
 568  *   drop PG_dirty with PG_lock held.
 569  *
 570  * - else?
 571  *
 572  */
 573 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 574                           int objcount, struct obd_ioobj *obj,
 575                           struct niobuf_remote *nb, int niocount,
 576                           struct niobuf_local *res, struct obd_trans_info *oti,
 577                           int rc)
 578 {
 579         struct niobuf_local *lnb;
 580         struct filter_iobuf *iobuf = NULL;
 581         struct lvfs_run_ctxt saved;
 582         struct fsfilt_objinfo fso;
 583         struct iattr iattr = { 0 };
 584         struct inode *inode = res->dentry->d_inode;
 585         unsigned long now = jiffies;
 586         int i, err, cleanup_phase = 0;
 587         struct obd_device *obd = exp->exp_obd;
 588         struct filter_obd *fo = &obd->u.filter;
 589         void *wait_handle = NULL;
 590         int total_size = 0;
 591         int sync_journal_commit = obd->u.filter.fo_syncjournal;
 592         int retries = 0;
 593         ENTRY;
 594
 595         LASSERT(oti != NULL);
 596         LASSERT(objcount == 1);
 597         LASSERT(current->journal_info == NULL);
 598
 599         if (rc != 0)
 600                 GOTO(cleanup, rc);
 601
 602         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 603         if (IS_ERR(iobuf))
 604                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 605         cleanup_phase = 1;
 606
 607         fso.fso_dentry = res->dentry;
 608         fso.fso_bufcnt = obj->ioo_bufcnt;
 609
 610         iobuf->dr_ignore_quota = 0;
 611         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 612                 loff_t this_size;
 613                 __u32 flags = lnb->flags;
 614
 615                 if (filter_range_is_mapped(inode, lnb->lnb_file_offset,
 616                                            lnb->len)) {
 617                         /* If overwriting an existing block,
 618                          * we don't need a grant */
 619                         if (!(flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC)
 620                                 lnb->rc = 0;
 621                 }
 622
 623                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 624                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 625                         continue;
 626                 }
 627
 628                 LASSERT(PageLocked(lnb->page));
 629                 LASSERT(!PageWriteback(lnb->page));
 630
 631                 /* since write & truncate are serialized by the i_alloc_sem,
 632                  * even partial truncate should not leave dirty pages in
 633                  * the page cache */
 634                 LASSERT(!PageDirty(lnb->page));
 635
 636                 SetPageUptodate(lnb->page);
 637
 638                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 639                 LASSERT (err == 0);
 640
 641                 total_size += lnb->len;
 642
 643                 /* we expect these pages to be in offset order, but we'll
 644                  * be forgiving */
 645                 this_size = lnb->lnb_file_offset + lnb->len;
 646                 if (this_size > iattr.ia_size)
 647                         iattr.ia_size = this_size;
 648
 649                 /* if one page is a write-back page from client cache and
 650                  * not from direct_io, or it's written by root, then mark
 651                  * the whole io request as ignore quota request, remote
 652                  * client can not break through quota. */
 653                 if (exp_connect_rmtclient(exp))
 654                         flags &= ~OBD_BRW_NOQUOTA;
 655                 if ((flags & OBD_BRW_NOQUOTA) ||
 656                     (flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 657                      OBD_BRW_FROM_GRANT)
 658                         iobuf->dr_ignore_quota = 1;
 659
 660                 if (!(lnb->flags & OBD_BRW_ASYNC)) {
 661                         sync_journal_commit = 1;
 662                 }
 663         }
 664
 665         if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
 666                 GOTO(cleanup, rc = -EINPROGRESS);
 667
 668         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 669         cleanup_phase = 2;
 670
 671 retry:
 672         mutex_lock(&inode->i_mutex);
 673         fsfilt_check_slow(obd, now, "i_mutex");
 674         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 675                                            oti);
 676         if (IS_ERR(oti->oti_handle)) {
 677                 mutex_unlock(&inode->i_mutex);
 678                 rc = PTR_ERR(oti->oti_handle);
 679                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 680                        "error starting transaction: rc = %d\n", rc);
 681                 oti->oti_handle = NULL;
 682                 GOTO(cleanup, rc);
 683         }
 684         /* have to call fsfilt_commit() from this point on */
 685
 686         fsfilt_check_slow(obd, now, "brw_start");
 687
 688         /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
 689         ll_vfs_dq_init(inode);
 690
 691         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 692
 693         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 694          * then we will accept the UID+GID if sent by the client for
 695          * initializing the ownership of this inode.  We only allow this to
 696          * happen once (so clear these bits) and later only allow setattr. */
 697         if (inode->i_mode & S_ISUID)
 698                 i |= OBD_MD_FLUID;
 699         if (inode->i_mode & S_ISGID)
 700                 i |= OBD_MD_FLGID;
 701
 702         iattr_from_obdo(&iattr, oa, i);
 703         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 704                 unsigned int save;
 705
 706                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 707                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 708
 709                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 710
 711                 iattr.ia_valid |= ATTR_MODE;
 712                 iattr.ia_mode = inode->i_mode;
 713                 if (iattr.ia_valid & ATTR_UID)
 714                         iattr.ia_mode &= ~S_ISUID;
 715                 if (iattr.ia_valid & ATTR_GID)
 716                         iattr.ia_mode &= ~S_ISGID;
 717
 718                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 719
 720                 /* To avoid problems with quotas, UID and GID must be set
 721                  * in the inode before filter_direct_io() - see bug 10357. */
 722                 save = iattr.ia_valid;
 723                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 724                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle,&iattr,0);
 725                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 726                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 727                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 728         }
 729
 730         CDEBUG(D_INODE, "FID "DFID" to write: s="LPU64" m="LPU64" a="LPU64
 731                         " c="LPU64" b="LPU64"\n",
 732                         oa->o_id, (__u32)oa->o_seq, (__u32)oa->o_seq,
 733                         oa->o_size, oa->o_mtime, oa->o_atime, oa->o_ctime,
 734                         oa->o_blocks);
 735         /* filter_direct_io drops i_mutex */
 736         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 737                               oti, sync_journal_commit ? &wait_handle : NULL);
 738         if (rc == -ENOSPC && retries++ < 3) {
 739                 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
 740                        retries);
 741                 oti->oti_handle = NULL;
 742                 fsfilt_check_slow(obd, now, "direct_io");
 743                 goto retry;
 744         }
 745
 746         obdo_from_inode(oa, inode, (rc == 0 ? FILTER_VALID_FLAGS : 0) |
 747                                    OBD_MD_FLUID | OBD_MD_FLGID);
 748
 749         CDEBUG(D_INODE, "FID "DFID" after write: s="LPU64" m="LPU64" a="LPU64
 750                         " c="LPU64" b="LPU64"\n",
 751                         oa->o_id, (__u32)oa->o_seq, (__u32)oa->o_seq,
 752                         oa->o_size, oa->o_mtime, oa->o_atime, oa->o_ctime,
 753                         oa->o_blocks);
 754
 755         fsfilt_check_slow(obd, now, "direct_io");
 756
 757         if (wait_handle)
 758                 err = fsfilt_commit_wait(obd, inode, wait_handle);
 759         else
 760                 err = 0;
 761
 762         if (err) {
 763                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 764                 if (rc == 0)
 765                         rc = err;
 766         }
 767
 768         /* In rare cases fsfilt_commit_wait() will wake up and return after
 769          * the transaction has finished its work and updated j_commit_sequence
 770          * but the commit callbacks have not been run yet.  Wait here until
 771          * that is finished so that clients requesting sync IO don't see the
 772          * reply transno < last_committed.  LU-753 */
 773         if (unlikely(obd->obd_replayable && !rc && wait_handle &&
 774                      oti->oti_transno > obd->obd_last_committed)) {
 775                 cfs_waitq_t wq;
 776                 struct l_wait_info lwi =
 777                         LWI_TIMEOUT_INTERVAL(cfs_time_seconds(5),
 778                                              (cfs_duration_t)((HZ + 4)/5),
 779                                              NULL, NULL);
 780                 cfs_waitq_init(&wq);
 781                 l_wait_event(wq,
 782                              oti->oti_transno <= obd->obd_last_committed,
 783                              &lwi);
 784
 785                 /* commit callback isn't done after waiting for 5 secs ? */
 786                 if (unlikely(oti->oti_transno > obd->obd_last_committed))
 787                         CERROR("transno:"LPU64" > last_committed:"LPU64"\n",
 788                                oti->oti_transno, obd->obd_last_committed);
 789         }
 790
 791         fsfilt_check_slow(obd, now, "commitrw commit");
 792
 793 cleanup:
 794         filter_grant_commit(exp, niocount, res);
 795
 796         switch (cleanup_phase) {
 797         case 2:
 798                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 799                 LASSERT(current->journal_info == NULL);
 800         case 1:
 801                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 802         case 0:
 803                 /*
 804                  * lnb->page automatically returns back into per-thread page
 805                  * pool (bug 5137)
 806                  */
 807                  break;
 808         }
 809
 810         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 811                 if (lnb->page == NULL)
 812                         continue;
 813
 814                 if (rc)
 815                         /* If the write has failed, the page cache may
 816                          * not be consitent with what is on disk, so
 817                          * force pages to be reread next time it is
 818                          * accessed */
 819                         ClearPageUptodate(lnb->page);
 820
 821                 LASSERT(PageLocked(lnb->page));
 822                 unlock_page(lnb->page);
 823
 824                 page_cache_release(lnb->page);
 825                 lnb->page = NULL;
 826         }
 827         f_dput(res->dentry);
 828
 829         if (inode) {
 830                 if (fo->fo_writethrough_cache == 0 ||
 831                     i_size_read(inode) > fo->fo_readcache_max_filesize)
 832                         filter_release_cache(obd, obj, nb, inode);
 833                 up_read(&inode->i_alloc_sem);
 834         }
 835
 836         RETURN(rc);
 837 }