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         spinlock_t        dr_lock;              /* IRQ lock */
  69         unsigned int      dr_ignore_quota:1;
  70         struct filter_obd *dr_filter;
  71 };
  72
  73 static void record_start_io(struct filter_iobuf *iobuf, int rw, int size,
  74                             struct obd_export *exp)
  75 {
  76         struct filter_obd *filter = iobuf->dr_filter;
  77
  78         atomic_inc(&iobuf->dr_numreqs);
  79
  80         if (rw == OBD_BRW_READ) {
  81                 atomic_inc(&filter->fo_r_in_flight);
  82                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST],
  83                                  atomic_read(&filter->fo_r_in_flight));
  84                 lprocfs_oh_tally_log2(&filter->fo_filter_stats.hist[BRW_R_DISK_IOSIZE],
  85                                       size);
  86                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_RPC_HIST],
  87                                  atomic_read(&filter->fo_r_in_flight));
  88                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_DISK_IOSIZE], size);
  89         } else {
  90                 atomic_inc(&filter->fo_w_in_flight);
  91                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
  92                                  atomic_read(&filter->fo_w_in_flight));
  93                 lprocfs_oh_tally_log2(&filter->fo_filter_stats.hist[BRW_W_DISK_IOSIZE],
  94                                       size);
  95                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_RPC_HIST],
  96                                  atomic_read(&filter->fo_w_in_flight));
  97                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_DISK_IOSIZE], size);
  98         }
  99 }
 100
 101 static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
 102 {
 103         struct filter_obd *filter = iobuf->dr_filter;
 104
 105         /* CAVEAT EMPTOR: possibly in IRQ context
 106          * DO NOT record procfs stats here!!! */
 107
 108         if (rw == OBD_BRW_READ)
 109                 atomic_dec(&filter->fo_r_in_flight);
 110         else
 111                 atomic_dec(&filter->fo_w_in_flight);
 112
 113         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 114                 wake_up(&iobuf->dr_wait);
 115 }
 116
 117 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 118 {
 119         struct filter_iobuf *iobuf = bio->bi_private;
 120         unsigned long        flags;
 121
 122 #ifdef HAVE_PAGE_CONSTANT
 123         struct bio_vec *bvl;
 124         int i;
 125 #endif
 126
 127         /* CAVEAT EMPTOR: possibly in IRQ context
 128          * DO NOT record procfs stats here!!! */
 129
 130         if (bio->bi_size)                       /* Not complete */
 131                 return 1;
 132
 133         if (iobuf == NULL) {
 134                 CERROR("***** bio->bi_private is NULL!  This should never "
 135                        "happen.  Normally, I would crash here, but instead I "
 136                        "will dump the bio contents to the console.  Please "
 137                        "report this to <http://bugzilla.lustre.org/> , along "
 138                        "with any interesting messages leading up to this point "
 139                        "(like SCSI errors, perhaps).  Because bi_private is "
 140                        "NULL, I can't wake up the thread that initiated this "
 141                        "I/O -- so you will probably have to reboot this node.\n");
 142                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 143                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 144                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 145                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 146                        bio->bi_end_io, atomic_read(&bio->bi_cnt),
 147                        bio->bi_private);
 148                 return 0;
 149         }
 150
 151 #ifdef HAVE_PAGE_CONSTANT
 152         bio_for_each_segment(bvl, bio, i)
 153                 ClearPageConstant(bvl->bv_page);
 154 #endif
 155
 156         spin_lock_irqsave(&iobuf->dr_lock, flags);
 157         if (iobuf->dr_error == 0)
 158                 iobuf->dr_error = error;
 159         spin_unlock_irqrestore(&iobuf->dr_lock, flags);
 160
 161         record_finish_io(iobuf, test_bit(BIO_RW, &bio->bi_rw) ?
 162                          OBD_BRW_WRITE : OBD_BRW_READ, error);
 163
 164         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 165          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 166          * mempool when serious on-disk fragmentation was encountered,
 167          * deadlocking the OST.  The bios are now released as soon as complete
 168          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 169         bio_put(bio);
 170         return 0;
 171 }
 172
 173 static int can_be_merged(struct bio *bio, sector_t sector)
 174 {
 175         unsigned int size;
 176
 177         if (!bio)
 178                 return 0;
 179
 180         size = bio->bi_size >> 9;
 181         return bio->bi_sector + size == sector ? 1 : 0;
 182 }
 183
 184 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 185                                         int rw, int num_pages)
 186 {
 187         struct filter_iobuf *iobuf;
 188
 189         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 190
 191         OBD_ALLOC(iobuf, sizeof(*iobuf));
 192         if (iobuf == NULL)
 193                 goto failed_0;
 194
 195         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 196         if (iobuf->dr_pages == NULL)
 197                 goto failed_1;
 198
 199         OBD_ALLOC(iobuf->dr_blocks,
 200                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 201         if (iobuf->dr_blocks == NULL)
 202                 goto failed_2;
 203
 204         iobuf->dr_filter = filter;
 205         init_waitqueue_head(&iobuf->dr_wait);
 206         atomic_set(&iobuf->dr_numreqs, 0);
 207         spin_lock_init(&iobuf->dr_lock);
 208         iobuf->dr_max_pages = num_pages;
 209         iobuf->dr_npages = 0;
 210         iobuf->dr_error = 0;
 211
 212         RETURN(iobuf);
 213
 214  failed_2:
 215         OBD_FREE(iobuf->dr_pages,
 216                  num_pages * sizeof(*iobuf->dr_pages));
 217  failed_1:
 218         OBD_FREE(iobuf, sizeof(*iobuf));
 219  failed_0:
 220         RETURN(ERR_PTR(-ENOMEM));
 221 }
 222
 223 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
 224 {
 225         iobuf->dr_npages = 0;
 226         iobuf->dr_error = 0;
 227         atomic_set(&iobuf->dr_numreqs, 0);
 228 }
 229
 230 void filter_free_iobuf(struct filter_iobuf *iobuf)
 231 {
 232         int num_pages = iobuf->dr_max_pages;
 233
 234         filter_clear_iobuf(iobuf);
 235
 236         OBD_FREE(iobuf->dr_blocks,
 237                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 238         OBD_FREE(iobuf->dr_pages,
 239                  num_pages * sizeof(*iobuf->dr_pages));
 240         OBD_FREE_PTR(iobuf);
 241 }
 242
 243 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 244                       struct obd_trans_info *oti)
 245 {
 246         int thread_id = oti ? oti->oti_thread_id : -1;
 247
 248         if (unlikely(thread_id < 0)) {
 249                 filter_free_iobuf(iobuf);
 250                 return;
 251         }
 252
 253         LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
 254                  "iobuf mismatch for thread %d: pool %p iobuf %p\n",
 255                  thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
 256         filter_clear_iobuf(iobuf);
 257 }
 258
 259 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 260                           struct inode *inode, struct page *page)
 261 {
 262         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 263         iobuf->dr_pages[iobuf->dr_npages++] = page;
 264
 265         return 0;
 266 }
 267
 268 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 269                   struct filter_iobuf *iobuf, int rw)
 270 {
 271         struct obd_device *obd = exp->exp_obd;
 272         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 273         struct page  **pages = iobuf->dr_pages;
 274         int            npages = iobuf->dr_npages;
 275         unsigned long *blocks = iobuf->dr_blocks;
 276         int            total_blocks = npages * blocks_per_page;
 277         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 278         unsigned int   blocksize = inode->i_sb->s_blocksize;
 279         struct bio    *bio = NULL;
 280         int            frags = 0;
 281         unsigned long  start_time = jiffies;
 282         struct page   *page;
 283         unsigned int   page_offset;
 284         sector_t       sector;
 285         int            nblocks;
 286         int            block_idx;
 287         int            page_idx;
 288         int            i;
 289         int            rc = 0;
 290         ENTRY;
 291
 292         LASSERT(iobuf->dr_npages == npages);
 293         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 294
 295         for (page_idx = 0, block_idx = 0;
 296              page_idx < npages;
 297              page_idx++, block_idx += blocks_per_page) {
 298
 299                 page = pages[page_idx];
 300                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 301
 302                 for (i = 0, page_offset = 0;
 303                      i < blocks_per_page;
 304                      i += nblocks, page_offset += blocksize * nblocks) {
 305
 306                         nblocks = 1;
 307
 308                         if (blocks[block_idx + i] == 0) {  /* hole */
 309                                 LASSERT(rw == OBD_BRW_READ);
 310                                 memset(kmap(page) + page_offset, 0, blocksize);
 311                                 kunmap(page);
 312                                 continue;
 313                         }
 314
 315                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 316
 317                         /* Additional contiguous file blocks? */
 318                         while (i + nblocks < blocks_per_page &&
 319                                (sector + (nblocks << sector_bits)) ==
 320                                ((sector_t)blocks[block_idx + i + nblocks] <<
 321                                 sector_bits))
 322                                 nblocks++;
 323
 324 #ifdef HAVE_PAGE_CONSTANT
 325                         /* I only set the page to be constant only if it
 326                          * is mapped to a contiguous underlying disk block(s).
 327                          * It will then make sure the corresponding device
 328                          * cache of raid5 will be overwritten by this page.
 329                          * - jay */
 330                         if ((rw == OBD_BRW_WRITE) &&
 331                             (nblocks == blocks_per_page) &&
 332                             mapping_cap_page_constant_write(inode->i_mapping))
 333                                SetPageConstant(page);
 334 #endif
 335
 336                         if (bio != NULL &&
 337                             can_be_merged(bio, sector) &&
 338                             bio_add_page(bio, page,
 339                                          blocksize * nblocks, page_offset) != 0)
 340                                 continue;       /* added this frag OK */
 341
 342                         if (bio != NULL) {
 343                                 request_queue_t *q =
 344                                         bdev_get_queue(bio->bi_bdev);
 345
 346                                 /* Dang! I have to fragment this I/O */
 347                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 348                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 349                                        bio->bi_size,
 350                                        bio->bi_vcnt, bio->bi_max_vecs,
 351                                        bio->bi_size >> 9, q->max_sectors,
 352                                        bio_phys_segments(q, bio),
 353                                        q->max_phys_segments,
 354                                        bio_hw_segments(q, bio),
 355                                        q->max_hw_segments);
 356
 357                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 358                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 359                                 if (rc < 0) {
 360                                         CERROR("Can't send bio: %d\n", rc);
 361                                         record_finish_io(iobuf, rw, rc);
 362                                         goto out;
 363                                 }
 364                                 frags++;
 365                         }
 366
 367                         /* allocate new bio */
 368                         bio = bio_alloc(GFP_NOIO,
 369                                         (npages - page_idx) * blocks_per_page);
 370                         if (bio == NULL) {
 371                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 372                                        (npages - page_idx), blocks_per_page,
 373                                        (npages - page_idx) * blocks_per_page);
 374                                 rc = -ENOMEM;
 375                                 goto out;
 376                         }
 377
 378                         bio->bi_bdev = inode->i_sb->s_bdev;
 379                         bio->bi_sector = sector;
 380                         bio->bi_end_io = dio_complete_routine;
 381                         bio->bi_private = iobuf;
 382
 383                         rc = bio_add_page(bio, page,
 384                                           blocksize * nblocks, page_offset);
 385                         LASSERT (rc != 0);
 386                 }
 387         }
 388
 389         if (bio != NULL) {
 390                 record_start_io(iobuf, rw, bio->bi_size, exp);
 391                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 392                 if (rc >= 0) {
 393                         frags++;
 394                         rc = 0;
 395                 } else {
 396                         CERROR("Can't send bio: %d\n", rc);
 397                         record_finish_io(iobuf, rw, rc);
 398                 }
 399         }
 400
 401  out:
 402         wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0);
 403
 404         if (rw == OBD_BRW_READ) {
 405                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_R_DIO_FRAGS],
 406                                  frags);
 407                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_DIO_FRAGS],
 408                                  frags);
 409                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_R_IO_TIME],
 410                                       jiffies - start_time);
 411                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_IO_TIME], jiffies - start_time);
 412                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 413                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_R_DIO_FRAGS],
 414                                          frags);
 415                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_R_IO_TIME],
 416                                               jiffies - start_time);
 417                 }
 418         } else {
 419                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_W_DIO_FRAGS],
 420                                  frags);
 421                 lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_DIO_FRAGS],
 422                                  frags);
 423                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_W_IO_TIME],
 424                                       jiffies - start_time);
 425                 lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_IO_TIME], 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->hist[BRW_W_DIO_FRAGS],
 428                                          frags);
 429                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->nid_brw_stats->hist[BRW_W_IO_TIME],
 430                                               jiffies - start_time);
 431                 }
 432         }
 433
 434         if (rc == 0)
 435                 rc = iobuf->dr_error;
 436         RETURN(rc);
 437 }
 438
 439 /* These are our hacks to keep our directio/bh IO coherent with ext3's
 440  * page cache use.  Most notably ext3 reads file data into the page
 441  * cache when it is zeroing the tail of partial-block truncates and
 442  * leaves it there, sometimes generating io from it at later truncates.
 443  * This removes the partial page and its buffers from the page cache,
 444  * so it should only ever cause a wait in rare cases, as otherwise we
 445  * always do full-page IO to the OST.
 446  *
 447  * The call to truncate_complete_page() will call journal_invalidatepage()
 448  * to free the buffers and drop the page from cache.  The buffers should
 449  * not be dirty, because we already called fdatasync/fdatawait on them.
 450  */
 451 static int filter_sync_inode_data(struct inode *inode, int locked)
 452 {
 453         int rc = 0;
 454
 455         /* This is nearly do_fsync(), without the waiting on the inode */
 456         /* XXX: in 2.6.16 (at least) we don't need to hold i_mutex over
 457          * filemap_fdatawrite() and filemap_fdatawait(), so we may no longer
 458          * need this lock here at all. */
 459         if (!locked)
 460                 LOCK_INODE_MUTEX(inode);
 461         if (inode->i_mapping->nrpages) {
 462 #ifdef PF_SYNCWRITE
 463                 current->flags |= PF_SYNCWRITE;
 464 #endif
 465                 rc = filemap_fdatawrite(inode->i_mapping);
 466                 if (rc == 0)
 467                         rc = filemap_fdatawait(inode->i_mapping);
 468 #ifdef PF_SYNCWRITE
 469                 current->flags &= ~PF_SYNCWRITE;
 470 #endif
 471         }
 472         if (!locked)
 473                 UNLOCK_INODE_MUTEX(inode);
 474
 475         return rc;
 476 }
 477 /* Clear pages from the mapping before we do direct IO to that offset.
 478  * Now that the only source of such pages in the truncate path flushes
 479  * these pages to disk and then discards them, this is error condition.
 480  * If add back read cache this will happen again.  This could be disabled
 481  * until that time if we never see the below error. */
 482 static int filter_clear_page_cache(struct inode *inode,
 483                                    struct filter_iobuf *iobuf)
 484 {
 485         struct page *page;
 486         int i, rc;
 487
 488         rc = filter_sync_inode_data(inode, 0);
 489         if (rc != 0)
 490                 RETURN(rc);
 491
 492         /* be careful to call this after fsync_inode_data_buffers has waited
 493          * for IO to complete before we evict it from the cache */
 494         for (i = 0; i < iobuf->dr_npages; i++) {
 495                 page = find_lock_page(inode->i_mapping,
 496                                       iobuf->dr_pages[i]->index);
 497                 if (page == NULL)
 498                         continue;
 499                 if (page->mapping != NULL) {
 500                         CERROR("page %lu (%d/%d) in page cache during write!\n",
 501                                page->index, i, iobuf->dr_npages);
 502                         wait_on_page_writeback(page);
 503                         ll_truncate_complete_page(page);
 504                 }
 505
 506                 unlock_page(page);
 507                 page_cache_release(page);
 508         }
 509
 510         return 0;
 511 }
 512
 513 int filter_clear_truncated_page(struct inode *inode)
 514 {
 515         struct page *page;
 516         int rc;
 517
 518         /* Truncate on page boundary, so nothing to flush? */
 519         if (!(i_size_read(inode) & ~CFS_PAGE_MASK))
 520                 return 0;
 521
 522         rc = filter_sync_inode_data(inode, 1);
 523         if (rc != 0)
 524                 RETURN(rc);
 525
 526         /* be careful to call this after fsync_inode_data_buffers has waited
 527          * for IO to complete before we evict it from the cache */
 528         page = find_lock_page(inode->i_mapping,
 529                               i_size_read(inode) >> CFS_PAGE_SHIFT);
 530         if (page) {
 531                 if (page->mapping != NULL) {
 532                         wait_on_page_writeback(page);
 533                         ll_truncate_complete_page(page);
 534                 }
 535                 unlock_page(page);
 536                 page_cache_release(page);
 537         }
 538
 539         return 0;
 540 }
 541
 542 /* Must be called with i_mutex taken for writes; this will drop it */
 543 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 544                      struct obd_export *exp, struct iattr *attr,
 545                      struct obd_trans_info *oti, void **wait_handle)
 546 {
 547         struct obd_device *obd = exp->exp_obd;
 548         struct inode *inode = dchild->d_inode;
 549         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 550         int rc, rc2, create;
 551         struct semaphore *sem;
 552         ENTRY;
 553
 554         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 555                  iobuf->dr_npages, iobuf->dr_max_pages);
 556         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 557
 558         if (rw == OBD_BRW_READ) {
 559                 if (iobuf->dr_npages == 0)
 560                         RETURN(0);
 561                 create = 0;
 562                 sem = NULL;
 563         } else {
 564                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 565                 LASSERT(iobuf->dr_npages > 0);
 566                 create = 1;
 567                 sem = &obd->u.filter.fo_alloc_lock;
 568
 569                 lquota_enforce(filter_quota_interface_ref, obd, iobuf->dr_ignore_quota);
 570         }
 571
 572         rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 573                                     iobuf->dr_npages, iobuf->dr_blocks,
 574                                     obdfilter_created_scratchpad, create, sem);
 575
 576         if (rw == OBD_BRW_WRITE) {
 577                 if (rc == 0) {
 578                         filter_tally(exp, iobuf->dr_pages,
 579                                      iobuf->dr_npages, iobuf->dr_blocks,
 580                                      blocks_per_page, 1);
 581                         if (attr->ia_size > i_size_read(inode))
 582                                 attr->ia_valid |= ATTR_SIZE;
 583                         rc = fsfilt_setattr(obd, dchild,
 584                                             oti->oti_handle, attr, 0);
 585                 }
 586
 587                 UNLOCK_INODE_MUTEX(inode);
 588
 589                 rc2 = filter_finish_transno(exp, oti, 0, 0);
 590                 if (rc2 != 0) {
 591                         CERROR("can't close transaction: %d\n", rc2);
 592                         if (rc == 0)
 593                                 rc = rc2;
 594                 }
 595
 596                 rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
 597                                           wait_handle);
 598                 if (rc == 0)
 599                         rc = rc2;
 600                 if (rc != 0)
 601                         RETURN(rc);
 602         } else if (rc == 0) {
 603                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 604                              iobuf->dr_blocks, blocks_per_page, 0);
 605         }
 606
 607         rc = filter_clear_page_cache(inode, iobuf);
 608         if (rc != 0)
 609                 RETURN(rc);
 610
 611         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 612 }
 613
 614 /* See if there are unallocated parts in given file region */
 615 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 616 {
 617         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 618                 inode->i_mapping->a_ops->bmap;
 619         int j;
 620
 621         /* We can't know if we are overwriting or not */
 622         if (fs_bmap == NULL)
 623                 return 0;
 624
 625         offset >>= inode->i_blkbits;
 626         len >>= inode->i_blkbits;
 627
 628         for (j = 0; j <= len; j++)
 629                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 630                         return 0;
 631
 632         return 1;
 633 }
 634
 635 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 636                           int objcount, struct obd_ioobj *obj, int niocount,
 637                           struct niobuf_local *res, struct obd_trans_info *oti,
 638                           int rc)
 639 {
 640         struct niobuf_local *lnb;
 641         struct filter_iobuf *iobuf = NULL;
 642         struct lvfs_run_ctxt saved;
 643         struct fsfilt_objinfo fso;
 644         struct iattr iattr = { 0 };
 645         struct inode *inode = NULL;
 646         unsigned long now = jiffies;
 647         int i, err, cleanup_phase = 0;
 648         struct obd_device *obd = exp->exp_obd;
 649         void *wait_handle;
 650         int   total_size = 0, rc2;
 651         unsigned int qcids[MAXQUOTAS] = {0, 0};
 652         ENTRY;
 653
 654         LASSERT(oti != NULL);
 655         LASSERT(objcount == 1);
 656         LASSERT(current->journal_info == NULL);
 657
 658         if (rc != 0)
 659                 GOTO(cleanup, rc);
 660
 661         /* Unfortunately, if quota master is too busy to handle the
 662          * pre-dqacq in time and quota hash on ost is used up, we
 663          * have to wait for the completion of in flight dqacq/dqrel,
 664          * then try again */
 665         if ((rc2 = lquota_chkquota(filter_quota_interface_ref, obd, oa->o_uid,
 666                                    oa->o_gid, niocount)) == QUOTA_RET_ACQUOTA) {
 667                 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_HOLD_WRITE_RPC, 90);
 668                 lquota_acquire(filter_quota_interface_ref, obd, oa->o_uid,
 669                                oa->o_gid);
 670         }
 671
 672         if (rc2 < 0) {
 673                 rc = rc2;
 674                 GOTO(cleanup, rc);
 675         }
 676
 677         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 678         if (IS_ERR(iobuf))
 679                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 680         cleanup_phase = 1;
 681
 682         fso.fso_dentry = res->dentry;
 683         fso.fso_bufcnt = obj->ioo_bufcnt;
 684         inode = res->dentry->d_inode;
 685
 686         iobuf->dr_ignore_quota = 0;
 687         for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
 688                 loff_t this_size;
 689
 690                 /* If overwriting an existing block, we don't need a grant */
 691                 if (!(lnb->flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC &&
 692                     filter_range_is_mapped(inode, lnb->offset, lnb->len))
 693                         lnb->rc = 0;
 694
 695                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 696                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 697                         continue;
 698                 }
 699
 700                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 701                 LASSERT (err == 0);
 702
 703                 total_size += lnb->len;
 704
 705                 /* we expect these pages to be in offset order, but we'll
 706                  * be forgiving */
 707                 this_size = lnb->offset + lnb->len;
 708                 if (this_size > iattr.ia_size)
 709                         iattr.ia_size = this_size;
 710
 711                 /* if one page is a write-back page from client cache, or it's
 712                  * written by root, then mark the whole io request as ignore
 713                  * quota request */
 714                 if (lnb->flags & (OBD_BRW_FROM_GRANT | OBD_BRW_NOQUOTA))
 715                         iobuf->dr_ignore_quota = 1;
 716         }
 717
 718         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 719         cleanup_phase = 2;
 720
 721         DQUOT_INIT(inode);
 722
 723         LOCK_INODE_MUTEX(inode);
 724         fsfilt_check_slow(obd, now, "i_mutex");
 725         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 726                                            oti);
 727         if (IS_ERR(oti->oti_handle)) {
 728                 UNLOCK_INODE_MUTEX(inode);
 729                 rc = PTR_ERR(oti->oti_handle);
 730                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 731                        "error starting transaction: rc = %d\n", rc);
 732                 oti->oti_handle = NULL;
 733                 GOTO(cleanup, rc);
 734         }
 735         /* have to call fsfilt_commit() from this point on */
 736
 737         fsfilt_check_slow(obd, now, "brw_start");
 738
 739         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 740
 741         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 742          * then we will accept the UID+GID if sent by the client for
 743          * initializing the ownership of this inode.  We only allow this to
 744          * happen once (so clear these bits) and later only allow setattr. */
 745         if (inode->i_mode & S_ISUID)
 746                 i |= OBD_MD_FLUID;
 747         if (inode->i_mode & S_ISGID)
 748                 i |= OBD_MD_FLGID;
 749
 750         iattr_from_obdo(&iattr, oa, i);
 751         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 752                 unsigned int save;
 753
 754                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 755                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 756
 757                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 758
 759                 iattr.ia_valid |= ATTR_MODE;
 760                 iattr.ia_mode = inode->i_mode;
 761                 if (iattr.ia_valid & ATTR_UID)
 762                         iattr.ia_mode &= ~S_ISUID;
 763                 if (iattr.ia_valid & ATTR_GID)
 764                         iattr.ia_mode &= ~S_ISGID;
 765
 766                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 767
 768                 /* To avoid problems with quotas, UID and GID must be set
 769                  * in the inode before filter_direct_io() - see bug 10357. */
 770                 save = iattr.ia_valid;
 771                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 772                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle, &iattr, 0);
 773                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 774                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 775                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 776         }
 777
 778         /* filter_direct_io drops i_mutex */
 779         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 780                               oti, &wait_handle);
 781         if (rc == 0)
 782                 obdo_from_inode(oa, inode,
 783                                 FILTER_VALID_FLAGS |OBD_MD_FLUID |OBD_MD_FLGID);
 784         else
 785                 obdo_from_inode(oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
 786
 787         lquota_getflag(filter_quota_interface_ref, obd, oa);
 788
 789         fsfilt_check_slow(obd, now, "direct_io");
 790
 791         err = fsfilt_commit_wait(obd, inode, wait_handle);
 792         if (err) {
 793                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 794                 rc = err;
 795         }
 796
 797         if (obd->obd_replayable && !rc)
 798                 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
 799                          "oti_transno "LPU64" last_committed "LPU64"\n",
 800                          oti->oti_transno, obd->obd_last_committed);
 801
 802         fsfilt_check_slow(obd, now, "commitrw commit");
 803
 804 cleanup:
 805         filter_grant_commit(exp, niocount, res);
 806
 807         switch (cleanup_phase) {
 808         case 2:
 809                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 810                 LASSERT(current->journal_info == NULL);
 811         case 1:
 812                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 813         case 0:
 814                 /*
 815                  * lnb->page automatically returns back into per-thread page
 816                  * pool (bug 5137)
 817                  */
 818                 f_dput(res->dentry);
 819         }
 820
 821         /* trigger quota pre-acquire */
 822         qcids[USRQUOTA] = oa->o_uid;
 823         qcids[GRPQUOTA] = oa->o_gid;
 824         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 825                             FSFILT_OP_CREATE);
 826         CDEBUG(err ? D_ERROR : D_QUOTA,
 827                "filter adjust qunit! (rc:%d)\n", err);
 828
 829         RETURN(rc);
 830 }