lustre/obdfilter/filter_io_26.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdfilter/filter_io_26.c
  37  *
  38  * Author: Peter Braam <braam@clusterfs.com>
  39  * Author: Andreas Dilger <adilger@clusterfs.com>
  40  * Author: Phil Schwan <phil@clusterfs.com>
  41  */
  42
  43 #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
  61 static void record_start_io(struct filter_iobuf *iobuf, int rw, int size,
  62                             struct obd_export *exp)
  63 {
  64         struct filter_obd *filter = iobuf->dr_filter;
  65
  66         cfs_atomic_inc(&iobuf->dr_numreqs);
  67
  68         if (rw == OBD_BRW_READ) {
  69                 cfs_atomic_inc(&filter->fo_r_in_flight);
  70                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST],
  71                                  cfs_atomic_read(&filter->fo_r_in_flight));
  72                 lprocfs_oh_tally_log2(&filter->
  73                                        fo_filter_stats.hist[BRW_R_DISK_IOSIZE],
  74                                       size);
  75                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
  76                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
  77                                          hist[BRW_R_RPC_HIST],
  78                                          cfs_atomic_read(&filter-> \
  79                                          fo_r_in_flight));
  80                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
  81                                          nid_brw_stats->hist[BRW_R_DISK_IOSIZE],
  82                                               size);
  83                 }
  84         } else {
  85                 cfs_atomic_inc(&filter->fo_w_in_flight);
  86                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
  87                                  cfs_atomic_read(&filter->fo_w_in_flight));
  88                 lprocfs_oh_tally_log2(&filter->
  89                                        fo_filter_stats.hist[BRW_W_DISK_IOSIZE],
  90                                       size);
  91                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
  92                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
  93                                           hist[BRW_W_RPC_HIST],
  94                                          cfs_atomic_read(&filter-> \
  95                                          fo_w_in_flight));
  96                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
  97                                         nid_brw_stats->hist[BRW_W_DISK_IOSIZE],
  98                                               size);
  99                 }
 100         }
 101 }
 102
 103 static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
 104 {
 105         struct filter_obd *filter = iobuf->dr_filter;
 106
 107         /* CAVEAT EMPTOR: possibly in IRQ context
 108          * DO NOT record procfs stats here!!! */
 109
 110         if (rw == OBD_BRW_READ)
 111                 cfs_atomic_dec(&filter->fo_r_in_flight);
 112         else
 113                 cfs_atomic_dec(&filter->fo_w_in_flight);
 114
 115         if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs))
 116                 cfs_waitq_signal(&iobuf->dr_wait);
 117 }
 118
 119 #ifdef HAVE_BIO_ENDIO_2ARG
 120 #define DIO_RETURN(a)
 121 static void dio_complete_routine(struct bio *bio, int error)
 122 #else
 123 #define DIO_RETURN(a)   return(a)
 124 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 125 #endif
 126 {
 127         struct filter_iobuf *iobuf = bio->bi_private;
 128         struct bio_vec *bvl;
 129         int i;
 130
 131         /* CAVEAT EMPTOR: possibly in IRQ context
 132          * DO NOT record procfs stats here!!! */
 133
 134 #ifdef HAVE_BIO_ENDIO_2ARG
 135        /* The "bi_size" check was needed for kernels < 2.6.24 in order to
 136         * handle the case where a SCSI request error caused this callback
 137         * to be called before all of the biovecs had been processed.
 138         * Without this check the server thread will hang.  In newer kernels
 139         * the bio_end_io routine is never called for partial completions,
 140         * so this check is no longer needed. */
 141         if (bio->bi_size)                      /* Not complete */
 142                 DIO_RETURN(1);
 143 #endif
 144
 145         if (unlikely(iobuf == NULL)) {
 146                 CERROR("***** bio->bi_private is NULL!  This should never "
 147                        "happen.  Normally, I would crash here, but instead I "
 148                        "will dump the bio contents to the console.  Please "
 149                        "report this to <http://bugs.whamcloud.com/>, along "
 150                        "with any interesting messages leading up to this point "
 151                        "(like SCSI errors, perhaps).  Because bi_private is "
 152                        "NULL, I can't wake up the thread that initiated this "
 153                        "IO - you will probably have to reboot this node.\n");
 154                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 155                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 156                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 157                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 158                        bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
 159                        bio->bi_private);
 160                 DIO_RETURN(0);
 161         }
 162
 163         /* the check is outside of the cycle for performance reason -bzzz */
 164         if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
 165                 bio_for_each_segment(bvl, bio, i) {
 166                         if (likely(error == 0))
 167                                 SetPageUptodate(bvl->bv_page);
 168                         LASSERT(PageLocked(bvl->bv_page));
 169                         ClearPageConstant(bvl->bv_page);
 170                 }
 171                 record_finish_io(iobuf, OBD_BRW_READ, error);
 172         } else {
 173                 if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){
 174                         bio_for_each_segment(bvl, bio, i) {
 175                                 ClearPageConstant(bvl->bv_page);
 176                         }
 177                 }
 178                 record_finish_io(iobuf, OBD_BRW_WRITE, error);
 179         }
 180
 181         /* any real error is good enough -bzzz */
 182         if (error != 0 && iobuf->dr_error == 0)
 183                 iobuf->dr_error = error;
 184
 185         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 186          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 187          * mempool when serious on-disk fragmentation was encountered,
 188          * deadlocking the OST.  The bios are now released as soon as complete
 189          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 190         bio_put(bio);
 191         DIO_RETURN(0);
 192 }
 193
 194 static int can_be_merged(struct bio *bio, sector_t sector)
 195 {
 196         unsigned int size;
 197
 198         if (!bio)
 199                 return 0;
 200
 201         size = bio->bi_size >> 9;
 202         return bio->bi_sector + size == sector ? 1 : 0;
 203 }
 204
 205 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 206                                         int rw, int num_pages)
 207 {
 208         struct filter_iobuf *iobuf;
 209
 210         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 211
 212         OBD_ALLOC(iobuf, sizeof(*iobuf));
 213         if (iobuf == NULL)
 214                 goto failed_0;
 215
 216         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 217         if (iobuf->dr_pages == NULL)
 218                 goto failed_1;
 219
 220         OBD_ALLOC(iobuf->dr_blocks,
 221                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 222         if (iobuf->dr_blocks == NULL)
 223                 goto failed_2;
 224
 225         CFS_INIT_HLIST_NODE(&iobuf->dr_hlist);
 226         iobuf->dr_filter = filter;
 227         cfs_waitq_init(&iobuf->dr_wait);
 228         cfs_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         cfs_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         LASSERT(cfs_hlist_unhashed(&iobuf->dr_hlist));
 258         OBD_FREE(iobuf->dr_blocks,
 259                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 260         OBD_FREE(iobuf->dr_pages,
 261                  num_pages * sizeof(*iobuf->dr_pages));
 262         OBD_FREE_PTR(iobuf);
 263 }
 264
 265 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 266                       struct obd_trans_info *oti)
 267 {
 268         int thread_id = (oti && oti->oti_thread) ?
 269                         oti->oti_thread->t_id : -1;
 270
 271         if (unlikely(thread_id < 0)) {
 272                 filter_free_iobuf(iobuf);
 273                 return;
 274         }
 275
 276         filter_clear_iobuf(iobuf);
 277 }
 278
 279 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 280                           struct inode *inode, struct page *page)
 281 {
 282         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 283         iobuf->dr_pages[iobuf->dr_npages++] = page;
 284
 285         return 0;
 286 }
 287
 288 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 289                   struct filter_iobuf *iobuf, int rw)
 290 {
 291         struct obd_device *obd = exp->exp_obd;
 292         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 293         struct page  **pages = iobuf->dr_pages;
 294         int            npages = iobuf->dr_npages;
 295         unsigned long *blocks = iobuf->dr_blocks;
 296         int            total_blocks = npages * blocks_per_page;
 297         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 298         unsigned int   blocksize = inode->i_sb->s_blocksize;
 299         struct bio    *bio = NULL;
 300         int            frags = 0;
 301         unsigned long  start_time = jiffies;
 302         struct page   *page;
 303         unsigned int   page_offset;
 304         sector_t       sector;
 305         int            nblocks;
 306         int            block_idx;
 307         int            page_idx;
 308         int            i;
 309         int            rc = 0;
 310         ENTRY;
 311
 312         LASSERT(iobuf->dr_npages == npages);
 313         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 314
 315         for (page_idx = 0, block_idx = 0;
 316              page_idx < npages;
 317              page_idx++, block_idx += blocks_per_page) {
 318
 319                 page = pages[page_idx];
 320                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 321
 322                 for (i = 0, page_offset = 0;
 323                      i < blocks_per_page;
 324                      i += nblocks, page_offset += blocksize * nblocks) {
 325
 326                         nblocks = 1;
 327
 328                         if (blocks[block_idx + i] == 0) {  /* hole */
 329                                 LASSERT(rw == OBD_BRW_READ);
 330                                 memset(kmap(page) + page_offset, 0, blocksize);
 331                                 kunmap(page);
 332                                 continue;
 333                         }
 334
 335                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 336
 337                         /* Additional contiguous file blocks? */
 338                         while (i + nblocks < blocks_per_page &&
 339                                (sector + (nblocks << sector_bits)) ==
 340                                ((sector_t)blocks[block_idx + i + nblocks] <<
 341                                 sector_bits))
 342                                 nblocks++;
 343
 344                         /* I only set the page to be constant only if it
 345                          * is mapped to a contiguous underlying disk block(s).
 346                          * It will then make sure the corresponding device
 347                          * cache of raid5 will be overwritten by this page.
 348                          * - jay */
 349                         if ((rw == OBD_BRW_WRITE) &&
 350                             (nblocks == blocks_per_page) &&
 351                             mapping_cap_page_constant_write(inode->i_mapping))
 352                                SetPageConstant(page);
 353
 354                         if (bio != NULL &&
 355                             can_be_merged(bio, sector) &&
 356                             bio_add_page(bio, page,
 357                                          blocksize * nblocks, page_offset) != 0)
 358                                 continue;       /* added this frag OK */
 359
 360                         if (bio != NULL) {
 361                                 struct request_queue *q =
 362                                         bdev_get_queue(bio->bi_bdev);
 363
 364                                 /* Dang! I have to fragment this I/O */
 365                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 366                                        "sectors %d(%d) psg %d(%d) hsg %d(%d) "
 367                                        "sector %llu next %llu\n",
 368                                        bio->bi_size,
 369                                        bio->bi_vcnt, bio->bi_max_vecs,
 370                                        bio->bi_size >> 9, queue_max_sectors(q),
 371                                        bio_phys_segments(q, bio),
 372                                        queue_max_phys_segments(q),
 373                                        bio_hw_segments(q, bio),
 374                                        queue_max_hw_segments(q),
 375                                        (unsigned long long)bio->bi_sector,
 376                                        (unsigned long long)sector);
 377
 378                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 379                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 380                                 if (rc < 0) {
 381                                         CERROR("Can't send bio: %d\n", rc);
 382                                         record_finish_io(iobuf, rw, rc);
 383                                         goto out;
 384                                 }
 385                                 frags++;
 386                         }
 387
 388                         /* allocate new bio, limited by max BIO size, b=9945 */
 389                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 390                                                       (npages - page_idx) *
 391                                                       blocks_per_page));
 392                         if (bio == NULL) {
 393                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 394                                        (npages - page_idx), blocks_per_page,
 395                                        (npages - page_idx) * blocks_per_page);
 396                                 rc = -ENOMEM;
 397                                 goto out;
 398                         }
 399
 400                         bio->bi_bdev = inode->i_sb->s_bdev;
 401                         bio->bi_sector = sector;
 402                         bio->bi_end_io = dio_complete_routine;
 403                         bio->bi_private = iobuf;
 404
 405                         rc = bio_add_page(bio, page,
 406                                           blocksize * nblocks, page_offset);
 407                         LASSERT (rc != 0);
 408                 }
 409         }
 410
 411         if (bio != NULL) {
 412                 record_start_io(iobuf, rw, bio->bi_size, exp);
 413                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 414                 if (rc >= 0) {
 415                         frags++;
 416                         rc = 0;
 417                 } else {
 418                         CERROR("Can't send bio: %d\n", rc);
 419                         record_finish_io(iobuf, rw, rc);
 420                 }
 421         }
 422
 423  out:
 424         cfs_wait_event(iobuf->dr_wait,
 425                        cfs_atomic_read(&iobuf->dr_numreqs) == 0);
 426
 427         if (rw == OBD_BRW_READ) {
 428                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 429                                   hist[BRW_R_DIO_FRAGS],
 430                                  frags);
 431                 lprocfs_oh_tally_log2(&obd->u.filter.
 432                                        fo_filter_stats.hist[BRW_R_IO_TIME],
 433                                       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->
 436                                           hist[BRW_R_DIO_FRAGS],
 437                                          frags);
 438                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 439                                              nid_brw_stats->hist[BRW_R_IO_TIME],
 440                                               jiffies - start_time);
 441                 }
 442         } else {
 443                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 444                                   hist[BRW_W_DIO_FRAGS], frags);
 445                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.
 446                                        hist[BRW_W_IO_TIME],
 447                                       jiffies - start_time);
 448                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 449                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 450                                           hist[BRW_W_DIO_FRAGS],
 451                                          frags);
 452                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 453                                              nid_brw_stats->hist[BRW_W_IO_TIME],
 454                                               jiffies - start_time);
 455                 }
 456         }
 457
 458         if (rc == 0)
 459                 rc = iobuf->dr_error;
 460         RETURN(rc);
 461 }
 462
 463 /* Must be called with i_mutex taken for writes; this will drop it */
 464 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 465                      struct obd_export *exp, struct iattr *attr,
 466                      struct obd_trans_info *oti, void **wait_handle)
 467 {
 468         struct obd_device *obd = exp->exp_obd;
 469         struct inode *inode = dchild->d_inode;
 470         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 471         int rc, rc2, create;
 472         cfs_mutex_t *mutex;
 473         ENTRY;
 474
 475         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 476                  iobuf->dr_npages, iobuf->dr_max_pages);
 477         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 478
 479         if (rw == OBD_BRW_READ) {
 480                 if (iobuf->dr_npages == 0)
 481                         RETURN(0);
 482                 create = 0;
 483                 mutex = NULL;
 484         } else {
 485                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 486                 LASSERT(iobuf->dr_npages > 0);
 487                 create = 1;
 488                 mutex = &obd->u.filter.fo_alloc_lock;
 489
 490                 lquota_enforce(filter_quota_interface_ref, obd,
 491                                iobuf->dr_ignore_quota);
 492         }
 493
 494         if (rw == OBD_BRW_WRITE &&
 495             OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
 496                 rc = -ENOSPC;
 497         } else {
 498                 rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 499                                     iobuf->dr_npages, iobuf->dr_blocks,
 500                                     obdfilter_created_scratchpad, create, mutex);
 501         }
 502
 503         if (rw == OBD_BRW_WRITE) {
 504                 if (rc == 0) {
 505                         filter_tally(exp, iobuf->dr_pages,
 506                                      iobuf->dr_npages, iobuf->dr_blocks,
 507                                      blocks_per_page, 1);
 508                         if (attr->ia_size > i_size_read(inode))
 509                                 attr->ia_valid |= ATTR_SIZE;
 510                         rc = fsfilt_setattr(obd, dchild,
 511                                             oti->oti_handle, attr, 0);
 512                 }
 513
 514                 UNLOCK_INODE_MUTEX(inode);
 515
 516                 /* Force commit to make the just-deleted blocks
 517                  * reusable. LU-456 */
 518                 if (rc == -ENOSPC) {
 519                         fsfilt_commit(obd, inode, oti->oti_handle, 1);
 520                         RETURN(rc);
 521                 }
 522
 523                 rc2 = filter_finish_transno(exp, inode, oti, 0, 0);
 524                 if (rc2 != 0) {
 525                         CERROR("can't close transaction: %d\n", rc2);
 526                         if (rc == 0)
 527                                 rc = rc2;
 528                 }
 529
 530                 if (wait_handle)
 531                         rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
 532                                                   wait_handle);
 533                 else
 534                         rc2 = fsfilt_commit(obd, inode, oti->oti_handle, 0);
 535                 if (rc == 0)
 536                         rc = rc2;
 537                 if (rc != 0)
 538                         RETURN(rc);
 539         } else if (rc == 0) {
 540                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 541                              iobuf->dr_blocks, blocks_per_page, 0);
 542         }
 543
 544         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 545 }
 546
 547 /* See if there are unallocated parts in given file region */
 548 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 549 {
 550         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 551                 inode->i_mapping->a_ops->bmap;
 552         int j;
 553
 554         /* We can't know if we are overwriting or not */
 555         if (fs_bmap == NULL)
 556                 return 0;
 557
 558         offset >>= inode->i_blkbits;
 559         len >>= inode->i_blkbits;
 560
 561         for (j = 0; j <= len; j++)
 562                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 563                         return 0;
 564
 565         return 1;
 566 }
 567
 568 /*
 569  * interesting use cases on how it interacts with VM:
 570  *
 571  * - vm writeout -- shouldn't see our pages as we don't mark them dirty
 572  *   though vm can find partial page left dirty by truncate. in this
 573  *   usual writeout is used unless our write rewrite that page - then we
 574  *   drop PG_dirty with PG_lock held.
 575  *
 576  * - else?
 577  *
 578  */
 579 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 580                           int objcount, struct obd_ioobj *obj,
 581                           struct niobuf_remote *nb, int niocount,
 582                           struct niobuf_local *res, struct obd_trans_info *oti,
 583                           int rc)
 584 {
 585         struct niobuf_local *lnb;
 586         struct filter_iobuf *iobuf = NULL;
 587         struct lvfs_run_ctxt saved;
 588         struct fsfilt_objinfo fso;
 589         struct iattr iattr = { 0 };
 590         struct inode *inode = res->dentry->d_inode;
 591         unsigned long now = jiffies;
 592         int i, err, cleanup_phase = 0;
 593         struct obd_device *obd = exp->exp_obd;
 594         struct filter_obd *fo = &obd->u.filter;
 595         void *wait_handle = NULL;
 596         int total_size = 0;
 597         unsigned int qcids[MAXQUOTAS] = { oa->o_uid, oa->o_gid };
 598         int rec_pending[MAXQUOTAS] = { 0, 0 }, quota_pages = 0;
 599         int sync_journal_commit = obd->u.filter.fo_syncjournal;
 600         int retries = 0;
 601         ENTRY;
 602
 603         LASSERT(oti != NULL);
 604         LASSERT(objcount == 1);
 605         LASSERT(current->journal_info == NULL);
 606
 607         if (rc != 0)
 608                 GOTO(cleanup, rc);
 609
 610         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 611         if (IS_ERR(iobuf))
 612                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 613         cleanup_phase = 1;
 614
 615         fso.fso_dentry = res->dentry;
 616         fso.fso_bufcnt = obj->ioo_bufcnt;
 617
 618         iobuf->dr_ignore_quota = 0;
 619         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 620                 loff_t this_size;
 621                 __u32 flags = lnb->flags;
 622
 623                 if (filter_range_is_mapped(inode, lnb->offset, lnb->len)) {
 624                         /* If overwriting an existing block,
 625                          * we don't need a grant */
 626                         if (!(flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC)
 627                                 lnb->rc = 0;
 628                 } else {
 629                         quota_pages++;
 630                 }
 631
 632                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 633                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 634                         continue;
 635                 }
 636
 637                 LASSERT(PageLocked(lnb->page));
 638                 LASSERT(!PageWriteback(lnb->page));
 639
 640                 /* since write & truncate are serialized by the i_alloc_sem,
 641                  * even partial truncate should not leave dirty pages in
 642                  * the page cache */
 643                 LASSERT(!PageDirty(lnb->page));
 644
 645                 SetPageUptodate(lnb->page);
 646
 647                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 648                 LASSERT (err == 0);
 649
 650                 total_size += lnb->len;
 651
 652                 /* we expect these pages to be in offset order, but we'll
 653                  * be forgiving */
 654                 this_size = lnb->offset + lnb->len;
 655                 if (this_size > iattr.ia_size)
 656                         iattr.ia_size = this_size;
 657
 658                 /* if one page is a write-back page from client cache and
 659                  * not from direct_io, or it's written by root, then mark
 660                  * the whole io request as ignore quota request, remote
 661                  * client can not break through quota. */
 662                 if (exp_connect_rmtclient(exp))
 663                         flags &= ~OBD_BRW_NOQUOTA;
 664                 if ((flags & OBD_BRW_NOQUOTA) ||
 665                     (flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 666                      OBD_BRW_FROM_GRANT)
 667                         iobuf->dr_ignore_quota = 1;
 668
 669                 if (!(lnb->flags & OBD_BRW_ASYNC)) {
 670                         sync_journal_commit = 1;
 671                 }
 672         }
 673
 674         /* we try to get enough quota to write here, and let ldiskfs
 675          * decide if it is out of quota or not b=14783 */
 676         rc = lquota_chkquota(filter_quota_interface_ref, obd, exp, qcids,
 677                              rec_pending, quota_pages, oti, LQUOTA_FLAGS_BLK,
 678                              (void *)inode, obj->ioo_bufcnt);
 679         if (rc == -ENOTCONN)
 680                 GOTO(cleanup, rc);
 681
 682         if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
 683                 GOTO(cleanup, rc = -EINPROGRESS);
 684
 685         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 686         cleanup_phase = 2;
 687
 688         fsfilt_check_slow(obd, now, "quota init");
 689
 690 retry:
 691         LOCK_INODE_MUTEX(inode);
 692         fsfilt_check_slow(obd, now, "i_mutex");
 693         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 694                                            oti);
 695         if (IS_ERR(oti->oti_handle)) {
 696                 UNLOCK_INODE_MUTEX(inode);
 697                 rc = PTR_ERR(oti->oti_handle);
 698                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 699                        "error starting transaction: rc = %d\n", rc);
 700                 oti->oti_handle = NULL;
 701                 GOTO(cleanup, rc);
 702         }
 703         /* have to call fsfilt_commit() from this point on */
 704
 705         fsfilt_check_slow(obd, now, "brw_start");
 706
 707         /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
 708         ll_vfs_dq_init(inode);
 709
 710         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 711
 712         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 713          * then we will accept the UID+GID if sent by the client for
 714          * initializing the ownership of this inode.  We only allow this to
 715          * happen once (so clear these bits) and later only allow setattr. */
 716         if (inode->i_mode & S_ISUID)
 717                 i |= OBD_MD_FLUID;
 718         if (inode->i_mode & S_ISGID)
 719                 i |= OBD_MD_FLGID;
 720
 721         iattr_from_obdo(&iattr, oa, i);
 722         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 723                 unsigned int save;
 724
 725                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 726                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 727
 728                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 729
 730                 iattr.ia_valid |= ATTR_MODE;
 731                 iattr.ia_mode = inode->i_mode;
 732                 if (iattr.ia_valid & ATTR_UID)
 733                         iattr.ia_mode &= ~S_ISUID;
 734                 if (iattr.ia_valid & ATTR_GID)
 735                         iattr.ia_mode &= ~S_ISGID;
 736
 737                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 738
 739                 /* To avoid problems with quotas, UID and GID must be set
 740                  * in the inode before filter_direct_io() - see bug 10357. */
 741                 save = iattr.ia_valid;
 742                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 743                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle,&iattr,0);
 744                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 745                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 746                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 747         }
 748
 749         CDEBUG(D_INODE, "FID "DFID" to write: s="LPU64" m="LPU64" a="LPU64
 750                         " c="LPU64" b="LPU64"\n",
 751                         oa->o_id, (__u32)oa->o_seq, (__u32)oa->o_seq,
 752                         oa->o_size, oa->o_mtime, oa->o_atime, oa->o_ctime,
 753                         oa->o_blocks);
 754         /* filter_direct_io drops i_mutex */
 755         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 756                               oti, sync_journal_commit ? &wait_handle : NULL);
 757         if (rc == -ENOSPC && retries++ < 3) {
 758                 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
 759                        retries);
 760                 oti->oti_handle = NULL;
 761                 fsfilt_check_slow(obd, now, "direct_io");
 762                 goto retry;
 763         }
 764
 765         obdo_from_inode(oa, inode, (rc == 0 ? FILTER_VALID_FLAGS : 0) |
 766                                    OBD_MD_FLUID | OBD_MD_FLGID);
 767
 768         CDEBUG(D_INODE, "FID "DFID" after write: s="LPU64" m="LPU64" a="LPU64
 769                         " c="LPU64" b="LPU64"\n",
 770                         oa->o_id, (__u32)oa->o_seq, (__u32)oa->o_seq,
 771                         oa->o_size, oa->o_mtime, oa->o_atime, oa->o_ctime,
 772                         oa->o_blocks);
 773         lquota_getflag(filter_quota_interface_ref, obd, oa);
 774
 775         fsfilt_check_slow(obd, now, "direct_io");
 776
 777         if (wait_handle)
 778                 err = fsfilt_commit_wait(obd, inode, wait_handle);
 779         else
 780                 err = 0;
 781
 782         if (err) {
 783                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 784                 if (rc == 0)
 785                         rc = err;
 786         }
 787
 788         /* In rare cases fsfilt_commit_wait() will wake up and return after
 789          * the transaction has finished its work and updated j_commit_sequence
 790          * but the commit callbacks have not been run yet.  Wait here until
 791          * that is finished so that clients requesting sync IO don't see the
 792          * reply transno < last_committed.  LU-753 */
 793         if (unlikely(obd->obd_replayable && !rc && wait_handle &&
 794                      oti->oti_transno > obd->obd_last_committed)) {
 795                 cfs_waitq_t wq;
 796                 struct l_wait_info lwi =
 797                         LWI_TIMEOUT_INTERVAL(cfs_time_seconds(5),
 798                                              (cfs_duration_t)((HZ + 4)/5),
 799                                              NULL, NULL);
 800                 cfs_waitq_init(&wq);
 801                 l_wait_event(wq,
 802                              oti->oti_transno <= obd->obd_last_committed,
 803                              &lwi);
 804
 805                 /* commit callback isn't done after waiting for 5 secs ? */
 806                 if (unlikely(oti->oti_transno > obd->obd_last_committed))
 807                         CERROR("transno:"LPU64" > last_committed:"LPU64"\n",
 808                                oti->oti_transno, obd->obd_last_committed);
 809         }
 810
 811         fsfilt_check_slow(obd, now, "commitrw commit");
 812
 813 cleanup:
 814         lquota_pending_commit(filter_quota_interface_ref, obd, qcids,
 815                               rec_pending, 1);
 816
 817         filter_grant_commit(exp, niocount, res);
 818
 819         switch (cleanup_phase) {
 820         case 2:
 821                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 822                 LASSERT(current->journal_info == NULL);
 823         case 1:
 824                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 825         case 0:
 826                 /*
 827                  * lnb->page automatically returns back into per-thread page
 828                  * pool (bug 5137)
 829                  */
 830                  break;
 831         }
 832
 833         /* trigger quota pre-acquire */
 834         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 835                             FSFILT_OP_CREATE);
 836         CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 837                "(rc:%d, uid:%u, gid:%u)\n",
 838                err, qcids[USRQUOTA], qcids[GRPQUOTA]);
 839         if (qcids[USRQUOTA] != oa->o_uid || qcids[GRPQUOTA] != oa->o_gid) {
 840                 qcids[USRQUOTA] = oa->o_uid;
 841                 qcids[GRPQUOTA] = oa->o_gid;
 842                 err = lquota_adjust(filter_quota_interface_ref, obd, qcids,
 843                                     NULL, rc, FSFILT_OP_CREATE);
 844                 CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 845                        "(rc:%d, uid:%u, gid:%u)\n",
 846                        err, qcids[USRQUOTA], qcids[GRPQUOTA]);
 847         }
 848
 849         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 850                 if (lnb->page == NULL)
 851                         continue;
 852
 853                 if (rc)
 854                         /* If the write has failed, the page cache may
 855                          * not be consitent with what is on disk, so
 856                          * force pages to be reread next time it is
 857                          * accessed */
 858                         ClearPageUptodate(lnb->page);
 859
 860                 LASSERT(PageLocked(lnb->page));
 861                 unlock_page(lnb->page);
 862
 863                 page_cache_release(lnb->page);
 864                 lnb->page = NULL;
 865         }
 866         f_dput(res->dentry);
 867
 868         if (inode) {
 869                 if (fo->fo_writethrough_cache == 0 ||
 870                     i_size_read(inode) > fo->fo_readcache_max_filesize)
 871                         filter_release_cache(obd, obj, nb, inode);
 872                 up_read(&inode->i_alloc_sem);
 873         }
 874
 875         RETURN(rc);
 876 }