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