lustre/obdfilter/filter_io_26.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdfilter/filter_io_26.c
  37  *
  38  * Author: Peter Braam <braam@clusterfs.com>
  39  * Author: Andreas Dilger <adilger@clusterfs.com>
  40  * Author: Phil Schwan <phil@clusterfs.com>
  41  */
  42
  43 #ifndef AUTOCONF_INCLUDED
  44 #include <linux/config.h>
  45 #endif
  46 #include <linux/module.h>
  47 #include <linux/pagemap.h> // XXX kill me soon
  48 #include <linux/version.h>
  49 #include <linux/buffer_head.h>
  50
  51 #define DEBUG_SUBSYSTEM S_FILTER
  52
  53 #include <obd_class.h>
  54 #include <lustre_fsfilt.h>
  55 #include <lustre_quota.h>
  56 #include "filter_internal.h"
  57
  58 /* 512byte block min */
  59 #define MAX_BLOCKS_PER_PAGE (CFS_PAGE_SIZE / 512)
  60 struct filter_iobuf {
  61         atomic_t          dr_numreqs;  /* number of reqs being processed */
  62         wait_queue_head_t dr_wait;
  63         int               dr_max_pages;
  64         int               dr_npages;
  65         int               dr_error;
  66         struct page     **dr_pages;
  67         unsigned long    *dr_blocks;
  68         unsigned int      dr_ignore_quota:1;
  69         struct filter_obd *dr_filter;
  70 };
  71
  72 static void record_start_io(struct filter_iobuf *iobuf, int rw, int size,
  73                             struct obd_export *exp)
  74 {
  75         struct filter_obd *filter = iobuf->dr_filter;
  76
  77         atomic_inc(&iobuf->dr_numreqs);
  78
  79         if (rw == OBD_BRW_READ) {
  80                 atomic_inc(&filter->fo_r_in_flight);
  81                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST],
  82                                  atomic_read(&filter->fo_r_in_flight));
  83                 lprocfs_oh_tally_log2(&filter->
  84                                        fo_filter_stats.hist[BRW_R_DISK_IOSIZE],
  85                                       size);
  86                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
  87                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
  88                                           hist[BRW_R_RPC_HIST],
  89                                          atomic_read(&filter->fo_r_in_flight));
  90                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
  91                                          nid_brw_stats->hist[BRW_R_DISK_IOSIZE],
  92                                               size);
  93                 }
  94         } else {
  95                 atomic_inc(&filter->fo_w_in_flight);
  96                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
  97                                  atomic_read(&filter->fo_w_in_flight));
  98                 lprocfs_oh_tally_log2(&filter->
  99                                        fo_filter_stats.hist[BRW_W_DISK_IOSIZE],
 100                                       size);
 101                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 102                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 103                                           hist[BRW_W_RPC_HIST],
 104                                          atomic_read(&filter->fo_r_in_flight));
 105                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 106                                         nid_brw_stats->hist[BRW_W_DISK_IOSIZE],
 107                                               size);
 108                 }
 109         }
 110 }
 111
 112 static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
 113 {
 114         struct filter_obd *filter = iobuf->dr_filter;
 115
 116         /* CAVEAT EMPTOR: possibly in IRQ context
 117          * DO NOT record procfs stats here!!! */
 118
 119         if (rw == OBD_BRW_READ)
 120                 atomic_dec(&filter->fo_r_in_flight);
 121         else
 122                 atomic_dec(&filter->fo_w_in_flight);
 123
 124         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 125                 wake_up(&iobuf->dr_wait);
 126 }
 127
 128 #ifdef HAVE_BIO_ENDIO_2ARG
 129 #define DIO_RETURN(a)   return
 130 static void dio_complete_routine(struct bio *bio, int error)
 131 #else
 132 #define DIO_RETURN(a)   return a
 133 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 134 #endif
 135 {
 136         struct filter_iobuf *iobuf = bio->bi_private;
 137         struct bio_vec *bvl;
 138         int i;
 139
 140         /* CAVEAT EMPTOR: possibly in IRQ context
 141          * DO NOT record procfs stats here!!! */
 142
 143         if (bio->bi_size)                       /* Not complete */
 144                 DIO_RETURN(1);
 145
 146         if (unlikely(iobuf == NULL)) {
 147                 CERROR("***** bio->bi_private is NULL!  This should never "
 148                        "happen.  Normally, I would crash here, but instead I "
 149                        "will dump the bio contents to the console.  Please "
 150                        "report this to <http://bugzilla.lustre.org/> , along "
 151                        "with any interesting messages leading up to this point "
 152                        "(like SCSI errors, perhaps).  Because bi_private is "
 153                        "NULL, I can't wake up the thread that initiated this "
 154                        "IO - you will probably have to reboot this node.\n");
 155                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 156                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 157                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 158                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 159                        bio->bi_end_io, atomic_read(&bio->bi_cnt),
 160                        bio->bi_private);
 161                 DIO_RETURN(0);
 162         }
 163
 164         /* the check is outside of the cycle for performance reason -bzzz */
 165         if (!test_bit(BIO_RW, &bio->bi_rw)) {
 166                 bio_for_each_segment(bvl, bio, i) {
 167                         if (likely(error == 0))
 168                                 SetPageUptodate(bvl->bv_page);
 169                         LASSERT(PageLocked(bvl->bv_page));
 170                         ClearPageConstant(bvl->bv_page);
 171                 }
 172                 record_finish_io(iobuf, OBD_BRW_READ, error);
 173         } else {
 174                 if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){
 175                         bio_for_each_segment(bvl, bio, i) {
 176                                 ClearPageConstant(bvl->bv_page);
 177                         }
 178                 }
 179                 record_finish_io(iobuf, OBD_BRW_WRITE, error);
 180         }
 181
 182         /* any real error is good enough -bzzz */
 183         if (error != 0 && iobuf->dr_error == 0)
 184                 iobuf->dr_error = error;
 185
 186         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 187          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 188          * mempool when serious on-disk fragmentation was encountered,
 189          * deadlocking the OST.  The bios are now released as soon as complete
 190          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 191         bio_put(bio);
 192         DIO_RETURN(0);
 193 }
 194
 195 static int can_be_merged(struct bio *bio, sector_t sector)
 196 {
 197         unsigned int size;
 198
 199         if (!bio)
 200                 return 0;
 201
 202         size = bio->bi_size >> 9;
 203         return bio->bi_sector + size == sector ? 1 : 0;
 204 }
 205
 206 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 207                                         int rw, int num_pages)
 208 {
 209         struct filter_iobuf *iobuf;
 210
 211         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 212
 213         OBD_ALLOC(iobuf, sizeof(*iobuf));
 214         if (iobuf == NULL)
 215                 goto failed_0;
 216
 217         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 218         if (iobuf->dr_pages == NULL)
 219                 goto failed_1;
 220
 221         OBD_ALLOC(iobuf->dr_blocks,
 222                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 223         if (iobuf->dr_blocks == NULL)
 224                 goto failed_2;
 225
 226         iobuf->dr_filter = filter;
 227         init_waitqueue_head(&iobuf->dr_wait);
 228         atomic_set(&iobuf->dr_numreqs, 0);
 229         iobuf->dr_max_pages = num_pages;
 230         iobuf->dr_npages = 0;
 231         iobuf->dr_error = 0;
 232
 233         RETURN(iobuf);
 234
 235  failed_2:
 236         OBD_FREE(iobuf->dr_pages,
 237                  num_pages * sizeof(*iobuf->dr_pages));
 238  failed_1:
 239         OBD_FREE(iobuf, sizeof(*iobuf));
 240  failed_0:
 241         RETURN(ERR_PTR(-ENOMEM));
 242 }
 243
 244 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
 245 {
 246         iobuf->dr_npages = 0;
 247         iobuf->dr_error = 0;
 248         atomic_set(&iobuf->dr_numreqs, 0);
 249 }
 250
 251 void filter_free_iobuf(struct filter_iobuf *iobuf)
 252 {
 253         int num_pages = iobuf->dr_max_pages;
 254
 255         filter_clear_iobuf(iobuf);
 256
 257         OBD_FREE(iobuf->dr_blocks,
 258                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 259         OBD_FREE(iobuf->dr_pages,
 260                  num_pages * sizeof(*iobuf->dr_pages));
 261         OBD_FREE_PTR(iobuf);
 262 }
 263
 264 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 265                       struct obd_trans_info *oti)
 266 {
 267         int thread_id = (oti && oti->oti_thread) ?
 268                         oti->oti_thread->t_id : -1;
 269
 270         if (unlikely(thread_id < 0)) {
 271                 filter_free_iobuf(iobuf);
 272                 return;
 273         }
 274
 275         LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
 276                  "iobuf mismatch for thread %d: pool %p iobuf %p\n",
 277                  thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
 278         filter_clear_iobuf(iobuf);
 279 }
 280
 281 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 282                           struct inode *inode, struct page *page)
 283 {
 284         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 285         iobuf->dr_pages[iobuf->dr_npages++] = page;
 286
 287         return 0;
 288 }
 289
 290 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 291                   struct filter_iobuf *iobuf, int rw)
 292 {
 293         struct obd_device *obd = exp->exp_obd;
 294         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 295         struct page  **pages = iobuf->dr_pages;
 296         int            npages = iobuf->dr_npages;
 297         unsigned long *blocks = iobuf->dr_blocks;
 298         int            total_blocks = npages * blocks_per_page;
 299         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 300         unsigned int   blocksize = inode->i_sb->s_blocksize;
 301         struct bio    *bio = NULL;
 302         int            frags = 0;
 303         unsigned long  start_time = jiffies;
 304         struct page   *page;
 305         unsigned int   page_offset;
 306         sector_t       sector;
 307         int            nblocks;
 308         int            block_idx;
 309         int            page_idx;
 310         int            i;
 311         int            rc = 0;
 312         ENTRY;
 313
 314         LASSERT(iobuf->dr_npages == npages);
 315         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 316
 317         for (page_idx = 0, block_idx = 0;
 318              page_idx < npages;
 319              page_idx++, block_idx += blocks_per_page) {
 320
 321                 page = pages[page_idx];
 322                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 323
 324                 for (i = 0, page_offset = 0;
 325                      i < blocks_per_page;
 326                      i += nblocks, page_offset += blocksize * nblocks) {
 327
 328                         nblocks = 1;
 329
 330                         if (blocks[block_idx + i] == 0) {  /* hole */
 331                                 LASSERT(rw == OBD_BRW_READ);
 332                                 memset(kmap(page) + page_offset, 0, blocksize);
 333                                 kunmap(page);
 334                                 continue;
 335                         }
 336
 337                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 338
 339                         /* Additional contiguous file blocks? */
 340                         while (i + nblocks < blocks_per_page &&
 341                                (sector + (nblocks << sector_bits)) ==
 342                                ((sector_t)blocks[block_idx + i + nblocks] <<
 343                                 sector_bits))
 344                                 nblocks++;
 345
 346                         /* I only set the page to be constant only if it
 347                          * is mapped to a contiguous underlying disk block(s).
 348                          * It will then make sure the corresponding device
 349                          * cache of raid5 will be overwritten by this page.
 350                          * - jay */
 351                         if ((rw == OBD_BRW_WRITE) &&
 352                             (nblocks == blocks_per_page) &&
 353                             mapping_cap_page_constant_write(inode->i_mapping))
 354                                SetPageConstant(page);
 355
 356                         if (bio != NULL &&
 357                             can_be_merged(bio, sector) &&
 358                             bio_add_page(bio, page,
 359                                          blocksize * nblocks, page_offset) != 0)
 360                                 continue;       /* added this frag OK */
 361
 362                         if (bio != NULL) {
 363                                 struct request_queue *q =
 364                                        bdev_get_queue(bio->bi_bdev);
 365
 366                                 /* Dang! I have to fragment this I/O */
 367                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 368                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 369                                        bio->bi_size,
 370                                        bio->bi_vcnt, bio->bi_max_vecs,
 371                                        bio->bi_size >> 9, q->max_sectors,
 372                                        bio_phys_segments(q, bio),
 373                                        q->max_phys_segments,
 374                                        bio_hw_segments(q, bio),
 375                                        q->max_hw_segments);
 376
 377                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 378                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 379                                 if (rc < 0) {
 380                                         CERROR("Can't send bio: %d\n", rc);
 381                                         record_finish_io(iobuf, rw, rc);
 382                                         goto out;
 383                                 }
 384                                 frags++;
 385                         }
 386
 387                         /* allocate new bio, limited by max BIO size, b=9945 */
 388                         bio = bio_alloc(GFP_NOIO, max(BIO_MAX_PAGES,
 389                                                       (npages - page_idx) *
 390                                                       blocks_per_page));
 391                         if (bio == NULL) {
 392                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 393                                        (npages - page_idx), blocks_per_page,
 394                                        (npages - page_idx) * blocks_per_page);
 395                                 rc = -ENOMEM;
 396                                 goto out;
 397                         }
 398
 399                         bio->bi_bdev = inode->i_sb->s_bdev;
 400                         bio->bi_sector = sector;
 401                         bio->bi_end_io = dio_complete_routine;
 402                         bio->bi_private = iobuf;
 403
 404                         rc = bio_add_page(bio, page,
 405                                           blocksize * nblocks, page_offset);
 406                         LASSERT (rc != 0);
 407                 }
 408         }
 409
 410         if (bio != NULL) {
 411                 record_start_io(iobuf, rw, bio->bi_size, exp);
 412                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 413                 if (rc >= 0) {
 414                         frags++;
 415                         rc = 0;
 416                 } else {
 417                         CERROR("Can't send bio: %d\n", rc);
 418                         record_finish_io(iobuf, rw, rc);
 419                 }
 420         }
 421
 422  out:
 423         wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0);
 424
 425         if (rw == OBD_BRW_READ) {
 426                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 427                                   hist[BRW_R_DIO_FRAGS],
 428                                  frags);
 429                 lprocfs_oh_tally_log2(&obd->u.filter.
 430                                        fo_filter_stats.hist[BRW_R_IO_TIME],
 431                                       jiffies - start_time);
 432                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 433                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 434                                           hist[BRW_R_DIO_FRAGS],
 435                                          frags);
 436                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 437                                              nid_brw_stats->hist[BRW_R_IO_TIME],
 438                                               jiffies - start_time);
 439                 }
 440         } else {
 441                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 442                                   hist[BRW_W_DIO_FRAGS], frags);
 443                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.
 444                                        hist[BRW_W_IO_TIME],
 445                                       jiffies - start_time);
 446                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 447                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 448                                           hist[BRW_W_DIO_FRAGS],
 449                                          frags);
 450                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 451                                              nid_brw_stats->hist[BRW_W_IO_TIME],
 452                                               jiffies - start_time);
 453                 }
 454         }
 455
 456         if (rc == 0)
 457                 rc = iobuf->dr_error;
 458         RETURN(rc);
 459 }
 460
 461 /* Must be called with i_mutex taken for writes; this will drop it */
 462 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 463                      struct obd_export *exp, struct iattr *attr,
 464                      struct obd_trans_info *oti, void **wait_handle)
 465 {
 466         struct obd_device *obd = exp->exp_obd;
 467         struct inode *inode = dchild->d_inode;
 468         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 469         int rc, rc2, create;
 470         struct semaphore *sem;
 471         ENTRY;
 472
 473         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 474                  iobuf->dr_npages, iobuf->dr_max_pages);
 475         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 476
 477         if (rw == OBD_BRW_READ) {
 478                 if (iobuf->dr_npages == 0)
 479                         RETURN(0);
 480                 create = 0;
 481                 sem = NULL;
 482         } else {
 483                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 484                 LASSERT(iobuf->dr_npages > 0);
 485                 create = 1;
 486                 sem = &obd->u.filter.fo_alloc_lock;
 487
 488                 lquota_enforce(filter_quota_interface_ref, obd,
 489                                iobuf->dr_ignore_quota);
 490         }
 491
 492         rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 493                                     iobuf->dr_npages, iobuf->dr_blocks,
 494                                     obdfilter_created_scratchpad, create, sem);
 495
 496         if (rw == OBD_BRW_WRITE) {
 497                 if (rc == 0) {
 498                         filter_tally(exp, iobuf->dr_pages,
 499                                      iobuf->dr_npages, iobuf->dr_blocks,
 500                                      blocks_per_page, 1);
 501                         if (attr->ia_size > i_size_read(inode))
 502                                 attr->ia_valid |= ATTR_SIZE;
 503                         rc = fsfilt_setattr(obd, dchild,
 504                                             oti->oti_handle, attr, 0);
 505                 }
 506
 507                 UNLOCK_INODE_MUTEX(inode);
 508
 509                 rc2 = filter_finish_transno(exp, inode, oti, 0, 0);
 510                 if (rc2 != 0) {
 511                         CERROR("can't close transaction: %d\n", rc2);
 512                         if (rc == 0)
 513                                 rc = rc2;
 514                 }
 515
 516                 if (wait_handle)
 517                         rc2 = fsfilt_commit_async(obd, inode, oti->oti_handle,
 518                                                   wait_handle);
 519                 else
 520                         rc2 = fsfilt_commit(obd, inode, oti->oti_handle, 0);
 521                 if (rc == 0)
 522                         rc = rc2;
 523                 if (rc != 0)
 524                         RETURN(rc);
 525         } else if (rc == 0) {
 526                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 527                              iobuf->dr_blocks, blocks_per_page, 0);
 528         }
 529
 530         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 531 }
 532
 533 /* See if there are unallocated parts in given file region */
 534 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 535 {
 536         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 537                 inode->i_mapping->a_ops->bmap;
 538         int j;
 539
 540         /* We can't know if we are overwriting or not */
 541         if (fs_bmap == NULL)
 542                 return 0;
 543
 544         offset >>= inode->i_blkbits;
 545         len >>= inode->i_blkbits;
 546
 547         for (j = 0; j <= len; j++)
 548                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 549                         return 0;
 550
 551         return 1;
 552 }
 553
 554 /*
 555  * interesting use cases on how it interacts with VM:
 556  *
 557  * - vm writeout -- shouldn't see our pages as we don't mark them dirty
 558  *   though vm can find partial page left dirty by truncate. in this
 559  *   usual writeout is used unless our write rewrite that page - then we
 560  *   drop PG_dirty with PG_lock held.
 561  *
 562  * - else?
 563  *
 564  */
 565 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 566                           int objcount, struct obd_ioobj *obj,
 567                           struct niobuf_remote *nb, int niocount,
 568                           struct niobuf_local *res, struct obd_trans_info *oti,
 569                           int rc)
 570 {
 571         struct niobuf_local *lnb;
 572         struct filter_iobuf *iobuf = NULL;
 573         struct lvfs_run_ctxt saved;
 574         struct fsfilt_objinfo fso;
 575         struct iattr iattr = { 0 };
 576         struct inode *inode = res->dentry->d_inode;
 577         unsigned long now = jiffies;
 578         int i, err, cleanup_phase = 0;
 579         struct obd_device *obd = exp->exp_obd;
 580         struct filter_obd *fo = &obd->u.filter;
 581         void *wait_handle = NULL;
 582         int total_size = 0;
 583         int quota_pending[2] = {0, 0}, quota_pages = 0;
 584         unsigned int qcids[MAXQUOTAS] = {0, 0};
 585         int sync_journal_commit = obd->u.filter.fo_syncjournal;
 586         ENTRY;
 587
 588         LASSERT(oti != NULL);
 589         LASSERT(objcount == 1);
 590         LASSERT(current->journal_info == NULL);
 591
 592         if (rc != 0)
 593                 GOTO(cleanup, rc);
 594
 595         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 596         if (IS_ERR(iobuf))
 597                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 598         cleanup_phase = 1;
 599
 600         fso.fso_dentry = res->dentry;
 601         fso.fso_bufcnt = obj->ioo_bufcnt;
 602
 603         iobuf->dr_ignore_quota = 0;
 604         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 605                 loff_t this_size;
 606
 607
 608                 if (filter_range_is_mapped(inode, lnb->offset, lnb->len)) {
 609                         /* If overwriting an existing block,
 610                          * we don't need a grant */
 611                         if (!(lnb->flags & OBD_BRW_GRANTED) &&
 612                             lnb->rc == -ENOSPC)
 613                                 lnb->rc = 0;
 614                 } else {
 615                         quota_pages++;
 616                 }
 617
 618                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 619                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 620                         continue;
 621                 }
 622
 623                 LASSERT(PageLocked(lnb->page));
 624                 LASSERT(!PageWriteback(lnb->page));
 625
 626                 /* preceding filemap_write_and_wait() should have clean pages */
 627                 if (fo->fo_writethrough_cache)
 628                         clear_page_dirty_for_io(lnb->page);
 629                 LASSERT(!PageDirty(lnb->page));
 630
 631                 SetPageUptodate(lnb->page);
 632
 633                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 634                 LASSERT (err == 0);
 635
 636                 total_size += lnb->len;
 637
 638                 /* we expect these pages to be in offset order, but we'll
 639                  * be forgiving */
 640                 this_size = lnb->offset + lnb->len;
 641                 if (this_size > iattr.ia_size)
 642                         iattr.ia_size = this_size;
 643
 644                 /* if one page is a write-back page from client cache and
 645                  * not from direct_io, or it's written by root, then mark
 646                  * the whole io request as ignore quota request */
 647                 if (lnb->flags & OBD_BRW_NOQUOTA ||
 648                     (lnb->flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 649                     OBD_BRW_FROM_GRANT)
 650                         iobuf->dr_ignore_quota = 1;
 651
 652                 if (!(lnb->flags & OBD_BRW_ASYNC)) {
 653                         sync_journal_commit = 1;
 654                 }
 655         }
 656
 657         /* we try to get enough quota to write here, and let ldiskfs
 658          * decide if it is out of quota or not b=14783 */
 659         lquota_chkquota(filter_quota_interface_ref, obd, oa->o_uid, oa->o_gid,
 660                         quota_pages, quota_pending, oti, inode,
 661                         obj->ioo_bufcnt);
 662
 663         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 664         cleanup_phase = 2;
 665
 666         DQUOT_INIT(inode);
 667
 668         LOCK_INODE_MUTEX(inode);
 669         fsfilt_check_slow(obd, now, "i_mutex");
 670         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 671                                            oti);
 672         if (IS_ERR(oti->oti_handle)) {
 673                 UNLOCK_INODE_MUTEX(inode);
 674                 rc = PTR_ERR(oti->oti_handle);
 675                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 676                        "error starting transaction: rc = %d\n", rc);
 677                 oti->oti_handle = NULL;
 678                 GOTO(cleanup, rc);
 679         }
 680         /* have to call fsfilt_commit() from this point on */
 681
 682         fsfilt_check_slow(obd, now, "brw_start");
 683
 684         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 685
 686         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 687          * then we will accept the UID+GID if sent by the client for
 688          * initializing the ownership of this inode.  We only allow this to
 689          * happen once (so clear these bits) and later only allow setattr. */
 690         if (inode->i_mode & S_ISUID)
 691                 i |= OBD_MD_FLUID;
 692         if (inode->i_mode & S_ISGID)
 693                 i |= OBD_MD_FLGID;
 694
 695         iattr_from_obdo(&iattr, oa, i);
 696         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 697                 unsigned int save;
 698
 699                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 700                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 701
 702                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 703
 704                 iattr.ia_valid |= ATTR_MODE;
 705                 iattr.ia_mode = inode->i_mode;
 706                 if (iattr.ia_valid & ATTR_UID)
 707                         iattr.ia_mode &= ~S_ISUID;
 708                 if (iattr.ia_valid & ATTR_GID)
 709                         iattr.ia_mode &= ~S_ISGID;
 710
 711                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 712
 713                 /* To avoid problems with quotas, UID and GID must be set
 714                  * in the inode before filter_direct_io() - see bug 10357. */
 715                 save = iattr.ia_valid;
 716                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 717                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle,&iattr,0);
 718                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 719                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 720                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 721         }
 722
 723         /* filter_direct_io drops i_mutex */
 724         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 725                               oti, sync_journal_commit ? &wait_handle : NULL);
 726         if (rc == 0)
 727                 obdo_from_inode(oa, inode,
 728                                 FILTER_VALID_FLAGS |OBD_MD_FLUID |OBD_MD_FLGID);
 729         else
 730                 obdo_from_inode(oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
 731
 732         lquota_getflag(filter_quota_interface_ref, obd, oa);
 733
 734         fsfilt_check_slow(obd, now, "direct_io");
 735
 736         if (wait_handle)
 737                 err = fsfilt_commit_wait(obd, inode, wait_handle);
 738         else
 739                 err = 0;
 740
 741         if (err) {
 742                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 743                 if (rc == 0)
 744                         rc = err;
 745         }
 746
 747         if (obd->obd_replayable && !rc && wait_handle)
 748                 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
 749                          "oti_transno "LPU64" last_committed "LPU64"\n",
 750                          oti->oti_transno, obd->obd_last_committed);
 751
 752         fsfilt_check_slow(obd, now, "commitrw commit");
 753
 754 cleanup:
 755         if (quota_pending[0] || quota_pending[1])
 756                 lquota_pending_commit(filter_quota_interface_ref, obd,
 757                                       oa->o_uid, oa->o_gid, quota_pending);
 758
 759         filter_grant_commit(exp, niocount, res);
 760
 761         switch (cleanup_phase) {
 762         case 2:
 763                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 764                 LASSERT(current->journal_info == NULL);
 765         case 1:
 766                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 767         case 0:
 768                 /*
 769                  * lnb->page automatically returns back into per-thread page
 770                  * pool (bug 5137)
 771                  */
 772                 f_dput(res->dentry);
 773         }
 774
 775         /* trigger quota pre-acquire */
 776         qcids[USRQUOTA] = oa->o_uid;
 777         qcids[GRPQUOTA] = oa->o_gid;
 778         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 779                             FSFILT_OP_CREATE);
 780         CDEBUG(err ? D_ERROR : D_QUOTA,
 781                "filter adjust qunit! (rc:%d)\n", err);
 782
 783         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 784                 if (lnb->page == NULL)
 785                         continue;
 786
 787                 LASSERT(PageLocked(lnb->page));
 788                 unlock_page(lnb->page);
 789
 790                 page_cache_release(lnb->page);
 791                 lnb->page = NULL;
 792         }
 793
 794         if (inode) {
 795                 if (fo->fo_writethrough_cache == 0 ||
 796                     i_size_read(inode) > fo->fo_readcache_max_filesize)
 797                         filter_release_cache(obd, obj, nb, inode);
 798                 up_read(&inode->i_alloc_sem);
 799         }
 800
 801         RETURN(rc);
 802 }