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 (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  *
  38  * lustre/obdfilter/filter_io_26.c
  39  *
  40  * Author: Peter Braam <braam@clusterfs.com>
  41  * Author: Andreas Dilger <adilger@clusterfs.com>
  42  * Author: Phil Schwan <phil@clusterfs.com>
  43  */
  44
  45 #ifndef AUTOCONF_INCLUDED
  46 #include <linux/config.h>
  47 #endif
  48 #include <linux/module.h>
  49 #include <linux/pagemap.h> // XXX kill me soon
  50 #include <linux/version.h>
  51 #include <linux/buffer_head.h>
  52
  53 #define DEBUG_SUBSYSTEM S_FILTER
  54
  55 #include <obd_class.h>
  56 #include <lustre_fsfilt.h>
  57 #include <lustre_quota.h>
  58 #include "filter_internal.h"
  59
  60 /* 512byte block min */
  61 #define MAX_BLOCKS_PER_PAGE (CFS_PAGE_SIZE / 512)
  62 struct filter_iobuf {
  63         cfs_atomic_t       dr_numreqs;  /* number of reqs being processed */
  64         cfs_waitq_t        dr_wait;
  65         int                dr_max_pages;
  66         int                dr_npages;
  67         int                dr_error;
  68         struct page      **dr_pages;
  69         unsigned long     *dr_blocks;
  70         unsigned int       dr_ignore_quota:1;
  71         struct filter_obd *dr_filter;
  72 };
  73
  74 static void record_start_io(struct filter_iobuf *iobuf, int rw, int size,
  75                             struct obd_export *exp)
  76 {
  77         struct filter_obd *filter = iobuf->dr_filter;
  78
  79         cfs_atomic_inc(&iobuf->dr_numreqs);
  80
  81         if (rw == OBD_BRW_READ) {
  82                 cfs_atomic_inc(&filter->fo_r_in_flight);
  83                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST],
  84                                  cfs_atomic_read(&filter->fo_r_in_flight));
  85                 lprocfs_oh_tally_log2(&filter->
  86                                        fo_filter_stats.hist[BRW_R_DISK_IOSIZE],
  87                                       size);
  88                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
  89                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
  90                                          hist[BRW_R_RPC_HIST],
  91                                          cfs_atomic_read(&filter-> \
  92                                          fo_r_in_flight));
  93                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
  94                                          nid_brw_stats->hist[BRW_R_DISK_IOSIZE],
  95                                               size);
  96                 }
  97         } else {
  98                 cfs_atomic_inc(&filter->fo_w_in_flight);
  99                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
 100                                  cfs_atomic_read(&filter->fo_w_in_flight));
 101                 lprocfs_oh_tally_log2(&filter->
 102                                        fo_filter_stats.hist[BRW_W_DISK_IOSIZE],
 103                                       size);
 104                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 105                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 106                                           hist[BRW_W_RPC_HIST],
 107                                          cfs_atomic_read(&filter-> \
 108                                          fo_w_in_flight));
 109                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 110                                         nid_brw_stats->hist[BRW_W_DISK_IOSIZE],
 111                                               size);
 112                 }
 113         }
 114 }
 115
 116 static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
 117 {
 118         struct filter_obd *filter = iobuf->dr_filter;
 119
 120         /* CAVEAT EMPTOR: possibly in IRQ context
 121          * DO NOT record procfs stats here!!! */
 122
 123         if (rw == OBD_BRW_READ)
 124                 cfs_atomic_dec(&filter->fo_r_in_flight);
 125         else
 126                 cfs_atomic_dec(&filter->fo_w_in_flight);
 127
 128         if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs))
 129                 cfs_waitq_signal(&iobuf->dr_wait);
 130 }
 131
 132 #ifdef HAVE_BIO_ENDIO_2ARG
 133 #define DIO_RETURN(a)
 134 static void dio_complete_routine(struct bio *bio, int error)
 135 #else
 136 #define DIO_RETURN(a)   return(a)
 137 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 138 #endif
 139 {
 140         struct filter_iobuf *iobuf = bio->bi_private;
 141         struct bio_vec *bvl;
 142         int i;
 143
 144         /* CAVEAT EMPTOR: possibly in IRQ context
 145          * DO NOT record procfs stats here!!! */
 146
 147 #ifdef HAVE_BIO_ENDIO_2ARG
 148        /* The "bi_size" check was needed for kernels < 2.6.24 in order to
 149         * handle the case where a SCSI request error caused this callback
 150         * to be called before all of the biovecs had been processed.
 151         * Without this check the server thread will hang.  In newer kernels
 152         * the bio_end_io routine is never called for partial completions,
 153         * so this check is no longer needed. */
 154         if (bio->bi_size)                      /* Not complete */
 155                 DIO_RETURN(1);
 156 #endif
 157
 158         if (unlikely(iobuf == NULL)) {
 159                 CERROR("***** bio->bi_private is NULL!  This should never "
 160                        "happen.  Normally, I would crash here, but instead I "
 161                        "will dump the bio contents to the console.  Please "
 162                        "report this to <http://bugs.whamcloud.com/>, along "
 163                        "with any interesting messages leading up to this point "
 164                        "(like SCSI errors, perhaps).  Because bi_private is "
 165                        "NULL, I can't wake up the thread that initiated this "
 166                        "IO - you will probably have to reboot this node.\n");
 167                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 168                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 169                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 170                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 171                        bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
 172                        bio->bi_private);
 173                 DIO_RETURN(0);
 174         }
 175
 176         /* the check is outside of the cycle for performance reason -bzzz */
 177         if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
 178                 bio_for_each_segment(bvl, bio, i) {
 179                         if (likely(error == 0))
 180                                 SetPageUptodate(bvl->bv_page);
 181                         LASSERT(PageLocked(bvl->bv_page));
 182                         ClearPageConstant(bvl->bv_page);
 183                 }
 184                 record_finish_io(iobuf, OBD_BRW_READ, error);
 185         } else {
 186                 if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){
 187                         bio_for_each_segment(bvl, bio, i) {
 188                                 ClearPageConstant(bvl->bv_page);
 189                         }
 190                 }
 191                 record_finish_io(iobuf, OBD_BRW_WRITE, error);
 192         }
 193
 194         /* any real error is good enough -bzzz */
 195         if (error != 0 && iobuf->dr_error == 0)
 196                 iobuf->dr_error = error;
 197
 198         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 199          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 200          * mempool when serious on-disk fragmentation was encountered,
 201          * deadlocking the OST.  The bios are now released as soon as complete
 202          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 203         bio_put(bio);
 204         DIO_RETURN(0);
 205 }
 206
 207 static int can_be_merged(struct bio *bio, sector_t sector)
 208 {
 209         unsigned int size;
 210
 211         if (!bio)
 212                 return 0;
 213
 214         size = bio->bi_size >> 9;
 215         return bio->bi_sector + size == sector ? 1 : 0;
 216 }
 217
 218 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 219                                         int rw, int num_pages)
 220 {
 221         struct filter_iobuf *iobuf;
 222
 223         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 224
 225         OBD_ALLOC(iobuf, sizeof(*iobuf));
 226         if (iobuf == NULL)
 227                 goto failed_0;
 228
 229         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 230         if (iobuf->dr_pages == NULL)
 231                 goto failed_1;
 232
 233         OBD_ALLOC(iobuf->dr_blocks,
 234                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 235         if (iobuf->dr_blocks == NULL)
 236                 goto failed_2;
 237
 238         iobuf->dr_filter = filter;
 239         cfs_waitq_init(&iobuf->dr_wait);
 240         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 241         iobuf->dr_max_pages = num_pages;
 242         iobuf->dr_npages = 0;
 243         iobuf->dr_error = 0;
 244
 245         RETURN(iobuf);
 246
 247  failed_2:
 248         OBD_FREE(iobuf->dr_pages,
 249                  num_pages * sizeof(*iobuf->dr_pages));
 250  failed_1:
 251         OBD_FREE(iobuf, sizeof(*iobuf));
 252  failed_0:
 253         RETURN(ERR_PTR(-ENOMEM));
 254 }
 255
 256 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
 257 {
 258         iobuf->dr_npages = 0;
 259         iobuf->dr_error = 0;
 260         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 261 }
 262
 263 void filter_free_iobuf(struct filter_iobuf *iobuf)
 264 {
 265         int num_pages = iobuf->dr_max_pages;
 266
 267         filter_clear_iobuf(iobuf);
 268
 269         OBD_FREE(iobuf->dr_blocks,
 270                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 271         OBD_FREE(iobuf->dr_pages,
 272                  num_pages * sizeof(*iobuf->dr_pages));
 273         OBD_FREE_PTR(iobuf);
 274 }
 275
 276 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 277                       struct obd_trans_info *oti)
 278 {
 279         int thread_id = (oti && oti->oti_thread) ?
 280                         oti->oti_thread->t_id : -1;
 281
 282         if (unlikely(thread_id < 0)) {
 283                 filter_free_iobuf(iobuf);
 284                 return;
 285         }
 286
 287         LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
 288                  "iobuf mismatch for thread %d: pool %p iobuf %p\n",
 289                  thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
 290         filter_clear_iobuf(iobuf);
 291 }
 292
 293 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 294                           struct inode *inode, struct page *page)
 295 {
 296         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 297         iobuf->dr_pages[iobuf->dr_npages++] = page;
 298
 299         return 0;
 300 }
 301
 302 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 303                   struct filter_iobuf *iobuf, int rw)
 304 {
 305         struct obd_device *obd = exp->exp_obd;
 306         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 307         struct page  **pages = iobuf->dr_pages;
 308         int            npages = iobuf->dr_npages;
 309         unsigned long *blocks = iobuf->dr_blocks;
 310         int            total_blocks = npages * blocks_per_page;
 311         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 312         unsigned int   blocksize = inode->i_sb->s_blocksize;
 313         struct bio    *bio = NULL;
 314         int            frags = 0;
 315         unsigned long  start_time = jiffies;
 316         struct page   *page;
 317         unsigned int   page_offset;
 318         sector_t       sector;
 319         int            nblocks;
 320         int            block_idx;
 321         int            page_idx;
 322         int            i;
 323         int            rc = 0;
 324         ENTRY;
 325
 326         LASSERT(iobuf->dr_npages == npages);
 327         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 328
 329         for (page_idx = 0, block_idx = 0;
 330              page_idx < npages;
 331              page_idx++, block_idx += blocks_per_page) {
 332
 333                 page = pages[page_idx];
 334                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 335
 336                 for (i = 0, page_offset = 0;
 337                      i < blocks_per_page;
 338                      i += nblocks, page_offset += blocksize * nblocks) {
 339
 340                         nblocks = 1;
 341
 342                         if (blocks[block_idx + i] == 0) {  /* hole */
 343                                 LASSERT(rw == OBD_BRW_READ);
 344                                 memset(kmap(page) + page_offset, 0, blocksize);
 345                                 kunmap(page);
 346                                 continue;
 347                         }
 348
 349                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 350
 351                         /* Additional contiguous file blocks? */
 352                         while (i + nblocks < blocks_per_page &&
 353                                (sector + (nblocks << sector_bits)) ==
 354                                ((sector_t)blocks[block_idx + i + nblocks] <<
 355                                 sector_bits))
 356                                 nblocks++;
 357
 358                         /* I only set the page to be constant only if it
 359                          * is mapped to a contiguous underlying disk block(s).
 360                          * It will then make sure the corresponding device
 361                          * cache of raid5 will be overwritten by this page.
 362                          * - jay */
 363                         if ((rw == OBD_BRW_WRITE) &&
 364                             (nblocks == blocks_per_page) &&
 365                             mapping_cap_page_constant_write(inode->i_mapping))
 366                                SetPageConstant(page);
 367
 368                         if (bio != NULL &&
 369                             can_be_merged(bio, sector) &&
 370                             bio_add_page(bio, page,
 371                                          blocksize * nblocks, page_offset) != 0)
 372                                 continue;       /* added this frag OK */
 373
 374                         if (bio != NULL) {
 375                                 struct request_queue *q =
 376                                         bdev_get_queue(bio->bi_bdev);
 377
 378                                 /* Dang! I have to fragment this I/O */
 379                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 380                                        "sectors %d(%d) psg %d(%d) hsg %d(%d) "
 381                                        "sector %llu next %llu\n",
 382                                        bio->bi_size,
 383                                        bio->bi_vcnt, bio->bi_max_vecs,
 384                                        bio->bi_size >> 9, queue_max_sectors(q),
 385                                        bio_phys_segments(q, bio),
 386                                        queue_max_phys_segments(q),
 387                                        bio_hw_segments(q, bio),
 388                                        queue_max_hw_segments(q),
 389                                        (unsigned long long)bio->bi_sector,
 390                                        (unsigned long long)sector);
 391
 392                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 393                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 394                                 if (rc < 0) {
 395                                         CERROR("Can't send bio: %d\n", rc);
 396                                         record_finish_io(iobuf, rw, rc);
 397                                         goto out;
 398                                 }
 399                                 frags++;
 400                         }
 401
 402                         /* allocate new bio, limited by max BIO size, b=9945 */
 403                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 404                                                       (npages - page_idx) *
 405                                                       blocks_per_page));
 406                         if (bio == NULL) {
 407                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 408                                        (npages - page_idx), blocks_per_page,
 409                                        (npages - page_idx) * blocks_per_page);
 410                                 rc = -ENOMEM;
 411                                 goto out;
 412                         }
 413
 414                         bio->bi_bdev = inode->i_sb->s_bdev;
 415                         bio->bi_sector = sector;
 416                         bio->bi_end_io = dio_complete_routine;
 417                         bio->bi_private = iobuf;
 418
 419                         rc = bio_add_page(bio, page,
 420                                           blocksize * nblocks, page_offset);
 421                         LASSERT (rc != 0);
 422                 }
 423         }
 424
 425         if (bio != NULL) {
 426                 record_start_io(iobuf, rw, bio->bi_size, exp);
 427                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 428                 if (rc >= 0) {
 429                         frags++;
 430                         rc = 0;
 431                 } else {
 432                         CERROR("Can't send bio: %d\n", rc);
 433                         record_finish_io(iobuf, rw, rc);
 434                 }
 435         }
 436
 437  out:
 438         cfs_wait_event(iobuf->dr_wait,
 439                        cfs_atomic_read(&iobuf->dr_numreqs) == 0);
 440
 441         if (rw == OBD_BRW_READ) {
 442                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 443                                   hist[BRW_R_DIO_FRAGS],
 444                                  frags);
 445                 lprocfs_oh_tally_log2(&obd->u.filter.
 446                                        fo_filter_stats.hist[BRW_R_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_R_DIO_FRAGS],
 451                                          frags);
 452                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 453                                              nid_brw_stats->hist[BRW_R_IO_TIME],
 454                                               jiffies - start_time);
 455                 }
 456         } else {
 457                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 458                                   hist[BRW_W_DIO_FRAGS], frags);
 459                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.
 460                                        hist[BRW_W_IO_TIME],
 461                                       jiffies - start_time);
 462                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 463                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 464                                           hist[BRW_W_DIO_FRAGS],
 465                                          frags);
 466                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 467                                              nid_brw_stats->hist[BRW_W_IO_TIME],
 468                                               jiffies - start_time);
 469                 }
 470         }
 471
 472         if (rc == 0)
 473                 rc = iobuf->dr_error;
 474         RETURN(rc);
 475 }
 476
 477 /* Must be called with i_mutex taken for writes; this will drop it */
 478 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 479                      struct obd_export *exp, struct iattr *attr,
 480                      struct obd_trans_info *oti, void **wait_handle)
 481 {
 482         struct obd_device *obd = exp->exp_obd;
 483         struct inode *inode = dchild->d_inode;
 484         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 485         int rc, rc2, create;
 486         cfs_mutex_t *mutex;
 487         ENTRY;
 488
 489         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 490                  iobuf->dr_npages, iobuf->dr_max_pages);
 491         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 492
 493         if (rw == OBD_BRW_READ) {
 494                 if (iobuf->dr_npages == 0)
 495                         RETURN(0);
 496                 create = 0;
 497                 mutex = NULL;
 498         } else {
 499                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 500                 LASSERT(iobuf->dr_npages > 0);
 501                 create = 1;
 502                 mutex = &obd->u.filter.fo_alloc_lock;
 503
 504                 lquota_enforce(filter_quota_interface_ref, obd,
 505                                iobuf->dr_ignore_quota);
 506         }
 507
 508         if (rw == OBD_BRW_WRITE &&
 509             OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
 510                 rc = -ENOSPC;
 511         } else {
 512                 rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 513                                     iobuf->dr_npages, iobuf->dr_blocks,
 514                                     obdfilter_created_scratchpad, create, mutex);
 515         }
 516
 517         if (rw == OBD_BRW_WRITE) {
 518                 if (rc == 0) {
 519                         filter_tally(exp, iobuf->dr_pages,
 520                                      iobuf->dr_npages, iobuf->dr_blocks,
 521                                      blocks_per_page, 1);
 522                         if (attr->ia_size > i_size_read(inode))
 523                                 attr->ia_valid |= ATTR_SIZE;
 524                         rc = fsfilt_setattr(obd, dchild,
 525                                             oti->oti_handle, attr, 0);
 526                 }
 527
 528                 UNLOCK_INODE_MUTEX(inode);
 529
 530                 /* Force commit to make the just-deleted blocks
 531                  * reusable. LU-456 */
 532                 if (rc == -ENOSPC) {
 533                         fsfilt_commit(obd, inode, oti->oti_handle, 1);
 534                         RETURN(rc);
 535                 }
 536
 537                 rc2 = filter_finish_transno(exp, inode, oti, 0, 0);
 538                 if (rc2 != 0) {
 539                         CERROR("can't close transaction: %d\n", rc2);
 540                         if (rc == 0)
 541                                 rc = rc2;
 542                 }
 543
 544                 if (wait_handle)
 545                         rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
 546                                                   wait_handle);
 547                 else
 548                         rc2 = fsfilt_commit(obd, inode, oti->oti_handle, 0);
 549                 if (rc == 0)
 550                         rc = rc2;
 551                 if (rc != 0)
 552                         RETURN(rc);
 553         } else if (rc == 0) {
 554                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 555                              iobuf->dr_blocks, blocks_per_page, 0);
 556         }
 557
 558         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 559 }
 560
 561 /* See if there are unallocated parts in given file region */
 562 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 563 {
 564         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 565                 inode->i_mapping->a_ops->bmap;
 566         int j;
 567
 568         /* We can't know if we are overwriting or not */
 569         if (fs_bmap == NULL)
 570                 return 0;
 571
 572         offset >>= inode->i_blkbits;
 573         len >>= inode->i_blkbits;
 574
 575         for (j = 0; j <= len; j++)
 576                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 577                         return 0;
 578
 579         return 1;
 580 }
 581
 582 /*
 583  * interesting use cases on how it interacts with VM:
 584  *
 585  * - vm writeout -- shouldn't see our pages as we don't mark them dirty
 586  *   though vm can find partial page left dirty by truncate. in this
 587  *   usual writeout is used unless our write rewrite that page - then we
 588  *   drop PG_dirty with PG_lock held.
 589  *
 590  * - else?
 591  *
 592  */
 593 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 594                           int objcount, struct obd_ioobj *obj,
 595                           struct niobuf_remote *nb, int niocount,
 596                           struct niobuf_local *res, struct obd_trans_info *oti,
 597                           int rc)
 598 {
 599         struct niobuf_local *lnb;
 600         struct filter_iobuf *iobuf = NULL;
 601         struct lvfs_run_ctxt saved;
 602         struct fsfilt_objinfo fso;
 603         struct iattr iattr = { 0 };
 604         struct inode *inode = res->dentry->d_inode;
 605         unsigned long now = jiffies;
 606         int i, err, cleanup_phase = 0;
 607         struct obd_device *obd = exp->exp_obd;
 608         struct filter_obd *fo = &obd->u.filter;
 609         void *wait_handle = NULL;
 610         int total_size = 0;
 611         unsigned int qcids[MAXQUOTAS] = { oa->o_uid, oa->o_gid };
 612         int rec_pending[MAXQUOTAS] = { 0, 0 }, quota_pages = 0;
 613         int sync_journal_commit = obd->u.filter.fo_syncjournal;
 614         int retries = 0;
 615         ENTRY;
 616
 617         LASSERT(oti != NULL);
 618         LASSERT(objcount == 1);
 619         LASSERT(current->journal_info == NULL);
 620
 621         if (rc != 0)
 622                 GOTO(cleanup, rc);
 623
 624         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 625         if (IS_ERR(iobuf))
 626                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 627         cleanup_phase = 1;
 628
 629         fso.fso_dentry = res->dentry;
 630         fso.fso_bufcnt = obj->ioo_bufcnt;
 631
 632         iobuf->dr_ignore_quota = 0;
 633         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 634                 loff_t this_size;
 635                 __u32 flags = lnb->flags;
 636
 637                 if (filter_range_is_mapped(inode, lnb->offset, lnb->len)) {
 638                         /* If overwriting an existing block,
 639                          * we don't need a grant */
 640                         if (!(flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC)
 641                                 lnb->rc = 0;
 642                 } else {
 643                         quota_pages++;
 644                 }
 645
 646                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 647                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 648                         continue;
 649                 }
 650
 651                 LASSERT(PageLocked(lnb->page));
 652                 LASSERT(!PageWriteback(lnb->page));
 653
 654                 /* since write & truncate are serialized by the i_alloc_sem,
 655                  * even partial truncate should not leave dirty pages in
 656                  * the page cache */
 657                 LASSERT(!PageDirty(lnb->page));
 658
 659                 SetPageUptodate(lnb->page);
 660
 661                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 662                 LASSERT (err == 0);
 663
 664                 total_size += lnb->len;
 665
 666                 /* we expect these pages to be in offset order, but we'll
 667                  * be forgiving */
 668                 this_size = lnb->offset + lnb->len;
 669                 if (this_size > iattr.ia_size)
 670                         iattr.ia_size = this_size;
 671
 672                 /* if one page is a write-back page from client cache and
 673                  * not from direct_io, or it's written by root, then mark
 674                  * the whole io request as ignore quota request, remote
 675                  * client can not break through quota. */
 676                 if (exp_connect_rmtclient(exp))
 677                         flags &= ~OBD_BRW_NOQUOTA;
 678                 if ((flags & OBD_BRW_NOQUOTA) ||
 679                     (flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 680                      OBD_BRW_FROM_GRANT)
 681                         iobuf->dr_ignore_quota = 1;
 682
 683                 if (!(lnb->flags & OBD_BRW_ASYNC)) {
 684                         sync_journal_commit = 1;
 685                 }
 686         }
 687
 688         /* we try to get enough quota to write here, and let ldiskfs
 689          * decide if it is out of quota or not b=14783 */
 690         rc = lquota_chkquota(filter_quota_interface_ref, obd, exp, qcids,
 691                              rec_pending, quota_pages, oti, LQUOTA_FLAGS_BLK,
 692                              (void *)inode, obj->ioo_bufcnt);
 693         if (rc == -ENOTCONN)
 694                 GOTO(cleanup, rc);
 695
 696         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 697         cleanup_phase = 2;
 698
 699         fsfilt_check_slow(obd, now, "quota init");
 700
 701 retry:
 702         LOCK_INODE_MUTEX(inode);
 703         fsfilt_check_slow(obd, now, "i_mutex");
 704         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 705                                            oti);
 706         if (IS_ERR(oti->oti_handle)) {
 707                 UNLOCK_INODE_MUTEX(inode);
 708                 rc = PTR_ERR(oti->oti_handle);
 709                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 710                        "error starting transaction: rc = %d\n", rc);
 711                 oti->oti_handle = NULL;
 712                 GOTO(cleanup, rc);
 713         }
 714         /* have to call fsfilt_commit() from this point on */
 715
 716         fsfilt_check_slow(obd, now, "brw_start");
 717
 718         /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
 719         ll_vfs_dq_init(inode);
 720
 721         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 722
 723         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 724          * then we will accept the UID+GID if sent by the client for
 725          * initializing the ownership of this inode.  We only allow this to
 726          * happen once (so clear these bits) and later only allow setattr. */
 727         if (inode->i_mode & S_ISUID)
 728                 i |= OBD_MD_FLUID;
 729         if (inode->i_mode & S_ISGID)
 730                 i |= OBD_MD_FLGID;
 731
 732         iattr_from_obdo(&iattr, oa, i);
 733         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 734                 unsigned int save;
 735
 736                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 737                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 738
 739                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 740
 741                 iattr.ia_valid |= ATTR_MODE;
 742                 iattr.ia_mode = inode->i_mode;
 743                 if (iattr.ia_valid & ATTR_UID)
 744                         iattr.ia_mode &= ~S_ISUID;
 745                 if (iattr.ia_valid & ATTR_GID)
 746                         iattr.ia_mode &= ~S_ISGID;
 747
 748                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 749
 750                 /* To avoid problems with quotas, UID and GID must be set
 751                  * in the inode before filter_direct_io() - see bug 10357. */
 752                 save = iattr.ia_valid;
 753                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 754                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle,&iattr,0);
 755                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 756                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 757                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 758         }
 759
 760         /* filter_direct_io drops i_mutex */
 761         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 762                               oti, sync_journal_commit ? &wait_handle : NULL);
 763         if (rc == -ENOSPC && retries++ < 3) {
 764                 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
 765                        retries);
 766                 oti->oti_handle = NULL;
 767                 fsfilt_check_slow(obd, now, "direct_io");
 768                 goto retry;
 769         }
 770
 771         obdo_from_inode(oa, inode, (rc == 0 ? FILTER_VALID_FLAGS : 0) |
 772                                    OBD_MD_FLUID | OBD_MD_FLGID);
 773
 774         lquota_getflag(filter_quota_interface_ref, obd, oa);
 775
 776         fsfilt_check_slow(obd, now, "direct_io");
 777
 778         if (wait_handle)
 779                 err = fsfilt_commit_wait(obd, inode, wait_handle);
 780         else
 781                 err = 0;
 782
 783         if (err) {
 784                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 785                 if (rc == 0)
 786                         rc = err;
 787         }
 788
 789         /* In rare cases fsfilt_commit_wait() will wake up and return after
 790          * the transaction has finished its work and updated j_commit_sequence
 791          * but the commit callbacks have not been run yet.  Wait here until
 792          * that is finished so that clients requesting sync IO don't see the
 793          * reply transno < last_committed.  LU-753 */
 794         if (unlikely(obd->obd_replayable && !rc && wait_handle &&
 795                      oti->oti_transno > obd->obd_last_committed)) {
 796                 cfs_waitq_t wq;
 797                 struct l_wait_info lwi =
 798                         LWI_TIMEOUT_INTERVAL(cfs_time_seconds(5),
 799                                              (cfs_duration_t)((HZ + 4)/5),
 800                                              NULL, NULL);
 801                 cfs_waitq_init(&wq);
 802                 l_wait_event(wq,
 803                              oti->oti_transno <= obd->obd_last_committed,
 804                              &lwi);
 805
 806                 /* commit callback isn't done after waiting for 5 secs ? */
 807                 if (unlikely(oti->oti_transno > obd->obd_last_committed))
 808                         CERROR("transno:"LPU64" > last_committed:"LPU64"\n",
 809                                oti->oti_transno, obd->obd_last_committed);
 810         }
 811
 812         fsfilt_check_slow(obd, now, "commitrw commit");
 813
 814 cleanup:
 815         lquota_pending_commit(filter_quota_interface_ref, obd, qcids,
 816                               rec_pending, 1);
 817
 818         filter_grant_commit(exp, niocount, res);
 819
 820         switch (cleanup_phase) {
 821         case 2:
 822                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 823                 LASSERT(current->journal_info == NULL);
 824         case 1:
 825                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 826         case 0:
 827                 /*
 828                  * lnb->page automatically returns back into per-thread page
 829                  * pool (bug 5137)
 830                  */
 831                  break;
 832         }
 833
 834         /* trigger quota pre-acquire */
 835         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 836                             FSFILT_OP_CREATE);
 837         CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 838                "(rc:%d, uid:%u, gid:%u)\n",
 839                err, qcids[USRQUOTA], qcids[GRPQUOTA]);
 840         if (qcids[USRQUOTA] != oa->o_uid || qcids[GRPQUOTA] != oa->o_gid) {
 841                 qcids[USRQUOTA] = oa->o_uid;
 842                 qcids[GRPQUOTA] = oa->o_gid;
 843                 err = lquota_adjust(filter_quota_interface_ref, obd, qcids,
 844                                     NULL, rc, FSFILT_OP_CREATE);
 845                 CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 846                        "(rc:%d, uid:%u, gid:%u)\n",
 847                        err, qcids[USRQUOTA], qcids[GRPQUOTA]);
 848         }
 849
 850         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 851                 if (lnb->page == NULL)
 852                         continue;
 853
 854                 if (rc)
 855                         /* If the write has failed, the page cache may
 856                          * not be consitent with what is on disk, so
 857                          * force pages to be reread next time it is
 858                          * accessed */
 859                         ClearPageUptodate(lnb->page);
 860
 861                 LASSERT(PageLocked(lnb->page));
 862                 unlock_page(lnb->page);
 863
 864                 page_cache_release(lnb->page);
 865                 lnb->page = NULL;
 866         }
 867         f_dput(res->dentry);
 868
 869         if (inode) {
 870                 if (fo->fo_writethrough_cache == 0 ||
 871                     i_size_read(inode) > fo->fo_readcache_max_filesize)
 872                         filter_release_cache(obd, obj, nb, inode);
 873                 up_read(&inode->i_alloc_sem);
 874         }
 875
 876         RETURN(rc);
 877 }