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 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdfilter/filter_io_26.c
  37  *
  38  * Author: Peter Braam <braam@clusterfs.com>
  39  * Author: Andreas Dilger <adilger@clusterfs.com>
  40  * Author: Phil Schwan <phil@clusterfs.com>
  41  */
  42
  43 #ifndef AUTOCONF_INCLUDED
  44 #include <linux/config.h>
  45 #endif
  46 #include <linux/module.h>
  47 #include <linux/pagemap.h> // XXX kill me soon
  48 #include <linux/version.h>
  49 #include <linux/buffer_head.h>
  50
  51 #define DEBUG_SUBSYSTEM S_FILTER
  52
  53 #include <obd_class.h>
  54 #include <lustre_fsfilt.h>
  55 #include <lustre_quota.h>
  56 #include "filter_internal.h"
  57
  58 /* 512byte block min */
  59 #define MAX_BLOCKS_PER_PAGE (CFS_PAGE_SIZE / 512)
  60 struct filter_iobuf {
  61         atomic_t          dr_numreqs;  /* number of reqs being processed */
  62         wait_queue_head_t dr_wait;
  63         int               dr_max_pages;
  64         int               dr_npages;
  65         int               dr_error;
  66         struct page     **dr_pages;
  67         unsigned long    *dr_blocks;
  68         unsigned int      dr_ignore_quota:1;
  69         struct filter_obd *dr_filter;
  70 };
  71
  72 static void record_start_io(struct filter_iobuf *iobuf, int rw, int size,
  73                             struct obd_export *exp)
  74 {
  75         struct filter_obd *filter = iobuf->dr_filter;
  76
  77         atomic_inc(&iobuf->dr_numreqs);
  78
  79         if (rw == OBD_BRW_READ) {
  80                 atomic_inc(&filter->fo_r_in_flight);
  81                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST],
  82                                  atomic_read(&filter->fo_r_in_flight));
  83                 lprocfs_oh_tally_log2(&filter->
  84                                        fo_filter_stats.hist[BRW_R_DISK_IOSIZE],
  85                                       size);
  86                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
  87                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
  88                                           hist[BRW_R_RPC_HIST],
  89                                          atomic_read(&filter->fo_r_in_flight));
  90                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
  91                                          nid_brw_stats->hist[BRW_R_DISK_IOSIZE],
  92                                               size);
  93                 }
  94         } else {
  95                 atomic_inc(&filter->fo_w_in_flight);
  96                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
  97                                  atomic_read(&filter->fo_w_in_flight));
  98                 lprocfs_oh_tally_log2(&filter->
  99                                        fo_filter_stats.hist[BRW_W_DISK_IOSIZE],
 100                                       size);
 101                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 102                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 103                                           hist[BRW_W_RPC_HIST],
 104                                          atomic_read(&filter->fo_w_in_flight));
 105                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 106                                         nid_brw_stats->hist[BRW_W_DISK_IOSIZE],
 107                                               size);
 108                 }
 109         }
 110 }
 111
 112 static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
 113 {
 114         struct filter_obd *filter = iobuf->dr_filter;
 115
 116         /* CAVEAT EMPTOR: possibly in IRQ context
 117          * DO NOT record procfs stats here!!! */
 118
 119         if (rw == OBD_BRW_READ)
 120                 atomic_dec(&filter->fo_r_in_flight);
 121         else
 122                 atomic_dec(&filter->fo_w_in_flight);
 123
 124         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 125                 wake_up(&iobuf->dr_wait);
 126 }
 127
 128 #ifdef HAVE_BIO_ENDIO_2ARG
 129 #define DIO_RETURN(a)
 130 static void dio_complete_routine(struct bio *bio, int error)
 131 #else
 132 #define DIO_RETURN(a)   a
 133 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 134 #endif
 135 {
 136         struct filter_iobuf *iobuf = bio->bi_private;
 137         struct bio_vec *bvl;
 138         int i;
 139
 140         /* CAVEAT EMPTOR: possibly in IRQ context
 141          * DO NOT record procfs stats here!!! */
 142
 143 #ifndef HAVE_BIO_ENDIO_2ARG
 144         /* The "bi_size" check was needed for kernels < 2.6.24 in order to
 145          * handle the case where a SCSI request error caused this callback
 146          * to be called before all of the biovecs had been processed.
 147          * Without this check the server thread will hang.  In newer kernels
 148          * the bio_end_io routine is never called for partial completions,
 149          * so this check is no longer needed. */
 150         if (bio->bi_size)                      /* Not complete */
 151                 return DIO_RETURN(1);
 152 #endif
 153
 154         if (unlikely(iobuf == NULL)) {
 155                 LCONSOLE_ERROR("bio->bi_private is NULL!  This should never "
 156                        "happen.  Normally, I would crash here, but instead I "
 157                        "will dump the bio contents to the console.  Please "
 158                        "report this to <http://bugs.whamcloud.com/> , along "
 159                        "with any interesting messages leading up to this point "
 160                        "(like SCSI errors, perhaps).  Because bi_private is "
 161                        "NULL, I can't wake up the thread that initiated this "
 162                        "IO - you will probably have to reboot this node.\n");
 163                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 164                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 165                        "bi_private: %p, error: %d\n",
 166                        bio->bi_next, bio->bi_flags,
 167                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 168                        bio->bi_end_io, atomic_read(&bio->bi_cnt),
 169                        bio->bi_private, error);
 170                 return DIO_RETURN(0);
 171         }
 172
 173         /* the check is outside of the cycle for performance reason -bzzz */
 174         if (!test_bit(BIO_RW, &bio->bi_rw)) {
 175                 bio_for_each_segment(bvl, bio, i) {
 176                         if (likely(error == 0))
 177                                 SetPageUptodate(bvl->bv_page);
 178                         LASSERT(PageLocked(bvl->bv_page));
 179                         ClearPageConstant(bvl->bv_page);
 180                 }
 181                 record_finish_io(iobuf, OBD_BRW_READ, error);
 182         } else {
 183                 if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){
 184                         bio_for_each_segment(bvl, bio, i) {
 185                                 ClearPageConstant(bvl->bv_page);
 186                         }
 187                 }
 188                 record_finish_io(iobuf, OBD_BRW_WRITE, error);
 189         }
 190
 191         /* any real error is good enough -bzzz */
 192         if (error != 0 && iobuf->dr_error == 0)
 193                 iobuf->dr_error = error;
 194
 195         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 196          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 197          * mempool when serious on-disk fragmentation was encountered,
 198          * deadlocking the OST.  The bios are now released as soon as complete
 199          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 200         bio_put(bio);
 201         return DIO_RETURN(0);
 202 }
 203
 204 static int can_be_merged(struct bio *bio, sector_t sector)
 205 {
 206         unsigned int size;
 207
 208         if (!bio)
 209                 return 0;
 210
 211         size = bio->bi_size >> 9;
 212         return bio->bi_sector + size == sector ? 1 : 0;
 213 }
 214
 215 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 216                                         int rw, int num_pages)
 217 {
 218         struct filter_iobuf *iobuf;
 219
 220         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 221
 222         OBD_ALLOC(iobuf, sizeof(*iobuf));
 223         if (iobuf == NULL)
 224                 goto failed_0;
 225
 226         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 227         if (iobuf->dr_pages == NULL)
 228                 goto failed_1;
 229
 230         OBD_ALLOC(iobuf->dr_blocks,
 231                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 232         if (iobuf->dr_blocks == NULL)
 233                 goto failed_2;
 234
 235         iobuf->dr_filter = filter;
 236         init_waitqueue_head(&iobuf->dr_wait);
 237         atomic_set(&iobuf->dr_numreqs, 0);
 238         iobuf->dr_max_pages = num_pages;
 239         iobuf->dr_npages = 0;
 240         iobuf->dr_error = 0;
 241
 242         RETURN(iobuf);
 243
 244  failed_2:
 245         OBD_FREE(iobuf->dr_pages,
 246                  num_pages * sizeof(*iobuf->dr_pages));
 247  failed_1:
 248         OBD_FREE(iobuf, sizeof(*iobuf));
 249  failed_0:
 250         RETURN(ERR_PTR(-ENOMEM));
 251 }
 252
 253 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
 254 {
 255         iobuf->dr_npages = 0;
 256         iobuf->dr_error = 0;
 257         atomic_set(&iobuf->dr_numreqs, 0);
 258 }
 259
 260 void filter_free_iobuf(struct filter_iobuf *iobuf)
 261 {
 262         int num_pages = iobuf->dr_max_pages;
 263
 264         filter_clear_iobuf(iobuf);
 265
 266         OBD_FREE(iobuf->dr_blocks,
 267                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 268         OBD_FREE(iobuf->dr_pages,
 269                  num_pages * sizeof(*iobuf->dr_pages));
 270         OBD_FREE_PTR(iobuf);
 271 }
 272
 273 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 274                       struct obd_trans_info *oti)
 275 {
 276         int thread_id = (oti && oti->oti_thread) ?
 277                         oti->oti_thread->t_id : -1;
 278
 279         if (unlikely(thread_id < 0)) {
 280                 filter_free_iobuf(iobuf);
 281                 return;
 282         }
 283
 284         LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
 285                  "iobuf mismatch for thread %d: pool %p iobuf %p\n",
 286                  thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
 287         filter_clear_iobuf(iobuf);
 288 }
 289
 290 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 291                           struct inode *inode, struct page *page)
 292 {
 293         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 294         iobuf->dr_pages[iobuf->dr_npages++] = page;
 295
 296         return 0;
 297 }
 298
 299 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 300                   struct filter_iobuf *iobuf, int rw)
 301 {
 302         struct obd_device *obd = exp->exp_obd;
 303         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 304         struct page  **pages = iobuf->dr_pages;
 305         int            npages = iobuf->dr_npages;
 306         unsigned long *blocks = iobuf->dr_blocks;
 307         int            total_blocks = npages * blocks_per_page;
 308         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 309         unsigned int   blocksize = inode->i_sb->s_blocksize;
 310         struct bio    *bio = NULL;
 311         int            frags = 0;
 312         unsigned long  start_time = jiffies;
 313         struct page   *page;
 314         unsigned int   page_offset;
 315         sector_t       sector;
 316         int            nblocks;
 317         int            block_idx;
 318         int            page_idx;
 319         int            i;
 320         int            rc = 0;
 321         ENTRY;
 322
 323         LASSERT(iobuf->dr_npages == npages);
 324         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 325
 326         for (page_idx = 0, block_idx = 0;
 327              page_idx < npages;
 328              page_idx++, block_idx += blocks_per_page) {
 329
 330                 page = pages[page_idx];
 331                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 332
 333                 for (i = 0, page_offset = 0;
 334                      i < blocks_per_page;
 335                      i += nblocks, page_offset += blocksize * nblocks) {
 336
 337                         nblocks = 1;
 338
 339                         if (blocks[block_idx + i] == 0) {  /* hole */
 340                                 LASSERT(rw == OBD_BRW_READ);
 341                                 memset(kmap(page) + page_offset, 0, blocksize);
 342                                 kunmap(page);
 343                                 continue;
 344                         }
 345
 346                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 347
 348                         /* Additional contiguous file blocks? */
 349                         while (i + nblocks < blocks_per_page &&
 350                                (sector + (nblocks << sector_bits)) ==
 351                                ((sector_t)blocks[block_idx + i + nblocks] <<
 352                                 sector_bits))
 353                                 nblocks++;
 354
 355                         /* I only set the page to be constant only if it
 356                          * is mapped to a contiguous underlying disk block(s).
 357                          * It will then make sure the corresponding device
 358                          * cache of raid5 will be overwritten by this page.
 359                          * - jay */
 360                         if ((rw == OBD_BRW_WRITE) &&
 361                             (nblocks == blocks_per_page) &&
 362                             mapping_cap_page_constant_write(inode->i_mapping))
 363                                SetPageConstant(page);
 364
 365                         if (bio != NULL &&
 366                             can_be_merged(bio, sector) &&
 367                             bio_add_page(bio, page,
 368                                          blocksize * nblocks, page_offset) != 0)
 369                                 continue;       /* added this frag OK */
 370
 371                         if (bio != NULL) {
 372                                 struct request_queue *q =
 373                                        bdev_get_queue(bio->bi_bdev);
 374
 375                                 /* Dang! I have to fragment this I/O */
 376                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 377                                        "sectors %d(%d) psg %d(%d) max hsg %d "
 378                                        "sector "LPU64" next "LPU64"\n",
 379                                        bio->bi_size,
 380                                        bio->bi_vcnt, bio->bi_max_vecs,
 381                                        bio->bi_size >> 9, queue_max_sectors(q),
 382                                        bio_phys_segments(q, bio),
 383                                        queue_max_phys_segments(q),
 384                                        queue_max_hw_segments(q),
 385                                        (u64)bio->bi_sector,
 386                                        (u64)sector);
 387
 388                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 389                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 390                                 if (rc < 0) {
 391                                         CERROR("Can't send bio: %d\n", rc);
 392                                         record_finish_io(iobuf, rw, rc);
 393                                         goto out;
 394                                 }
 395                                 frags++;
 396                         }
 397
 398                         /* allocate new bio, limited by max BIO size, b=9945 */
 399                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 400                                                       (npages - page_idx) *
 401                                                       blocks_per_page));
 402                         if (bio == NULL) {
 403                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 404                                        (npages - page_idx), blocks_per_page,
 405                                        (npages - page_idx) * blocks_per_page);
 406                                 rc = -ENOMEM;
 407                                 goto out;
 408                         }
 409
 410                         bio->bi_bdev = inode->i_sb->s_bdev;
 411                         bio->bi_sector = sector;
 412                         bio->bi_end_io = dio_complete_routine;
 413                         bio->bi_private = iobuf;
 414
 415                         rc = bio_add_page(bio, page,
 416                                           blocksize * nblocks, page_offset);
 417                         LASSERT (rc != 0);
 418                 }
 419         }
 420
 421         if (bio != NULL) {
 422                 record_start_io(iobuf, rw, bio->bi_size, exp);
 423                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 424                 if (rc >= 0) {
 425                         frags++;
 426                         rc = 0;
 427                 } else {
 428                         CERROR("Can't send bio: %d\n", rc);
 429                         record_finish_io(iobuf, rw, rc);
 430                 }
 431         }
 432
 433  out:
 434         wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0);
 435
 436         if (rw == OBD_BRW_READ) {
 437                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 438                                   hist[BRW_R_DIO_FRAGS],
 439                                  frags);
 440                 lprocfs_oh_tally_log2(&obd->u.filter.
 441                                        fo_filter_stats.hist[BRW_R_IO_TIME],
 442                                       jiffies - start_time);
 443                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 444                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 445                                           hist[BRW_R_DIO_FRAGS],
 446                                          frags);
 447                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 448                                              nid_brw_stats->hist[BRW_R_IO_TIME],
 449                                               jiffies - start_time);
 450                 }
 451         } else {
 452                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 453                                   hist[BRW_W_DIO_FRAGS], frags);
 454                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.
 455                                        hist[BRW_W_IO_TIME],
 456                                       jiffies - start_time);
 457                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 458                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 459                                           hist[BRW_W_DIO_FRAGS],
 460                                          frags);
 461                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 462                                              nid_brw_stats->hist[BRW_W_IO_TIME],
 463                                               jiffies - start_time);
 464                 }
 465         }
 466
 467         if (rc == 0)
 468                 rc = iobuf->dr_error;
 469         RETURN(rc);
 470 }
 471
 472 /* Must be called with i_mutex taken for writes; this will drop it */
 473 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 474                      struct obd_export *exp, struct iattr *attr,
 475                      struct obd_trans_info *oti, void **wait_handle)
 476 {
 477         struct obd_device *obd = exp->exp_obd;
 478         struct inode *inode = dchild->d_inode;
 479         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 480         int rc, rc2, create;
 481         struct semaphore *sem;
 482         ENTRY;
 483
 484         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 485                  iobuf->dr_npages, iobuf->dr_max_pages);
 486         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 487
 488         if (rw == OBD_BRW_READ) {
 489                 if (iobuf->dr_npages == 0)
 490                         RETURN(0);
 491                 create = 0;
 492                 sem = NULL;
 493         } else {
 494                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 495                 LASSERT(iobuf->dr_npages > 0);
 496                 create = 1;
 497                 sem = &obd->u.filter.fo_alloc_lock;
 498
 499                 lquota_enforce(filter_quota_interface_ref, obd,
 500                                iobuf->dr_ignore_quota);
 501         }
 502
 503         if (rw == OBD_BRW_WRITE &&
 504             OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
 505                 rc = -ENOSPC;
 506         } else {
 507                 rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 508                                     iobuf->dr_npages, iobuf->dr_blocks,
 509                                     obdfilter_created_scratchpad, create, sem);
 510         }
 511
 512         if (rw == OBD_BRW_WRITE) {
 513                 if (rc == 0) {
 514                         filter_tally(exp, iobuf->dr_pages,
 515                                      iobuf->dr_npages, iobuf->dr_blocks,
 516                                      blocks_per_page, 1);
 517                         if (attr->ia_size > i_size_read(inode))
 518                                 attr->ia_valid |= ATTR_SIZE;
 519                         rc = fsfilt_setattr(obd, dchild,
 520                                             oti->oti_handle, attr, 0);
 521                 }
 522
 523                 UNLOCK_INODE_MUTEX(inode);
 524
 525                 /* Force commit to make the just-deleted blocks
 526                  * reusable. LU-456 */
 527                 if (rc == -ENOSPC) {
 528                         fsfilt_commit(obd, inode, oti->oti_handle, 1);
 529                         RETURN(rc);
 530                 }
 531
 532                 rc2 = filter_finish_transno(exp, inode, oti, 0, 0);
 533                 if (rc2 != 0) {
 534                         CERROR("can't close transaction: %d\n", rc2);
 535                         if (rc == 0)
 536                                 rc = rc2;
 537                 }
 538
 539                 if (wait_handle)
 540                         rc2 = fsfilt_commit_async(obd, inode, oti->oti_handle,
 541                                                   wait_handle);
 542                 else
 543                         rc2 = fsfilt_commit(obd, inode, oti->oti_handle, 0);
 544                 if (rc == 0)
 545                         rc = rc2;
 546                 if (rc != 0)
 547                         RETURN(rc);
 548         } else if (rc == 0) {
 549                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 550                              iobuf->dr_blocks, blocks_per_page, 0);
 551         }
 552
 553         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 554 }
 555
 556 /* See if there are unallocated parts in given file region */
 557 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 558 {
 559         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 560                 inode->i_mapping->a_ops->bmap;
 561         int j;
 562
 563         /* We can't know if we are overwriting or not */
 564         if (fs_bmap == NULL)
 565                 return 0;
 566
 567         offset >>= inode->i_blkbits;
 568         len >>= inode->i_blkbits;
 569
 570         for (j = 0; j <= len; j++)
 571                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 572                         return 0;
 573
 574         return 1;
 575 }
 576
 577 /*
 578  * interesting use cases on how it interacts with VM:
 579  *
 580  * - vm writeout -- shouldn't see our pages as we don't mark them dirty
 581  *   though vm can find partial page left dirty by truncate. in this
 582  *   usual writeout is used unless our write rewrite that page - then we
 583  *   drop PG_dirty with PG_lock held.
 584  *
 585  * - else?
 586  *
 587  */
 588 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 589                           int objcount, struct obd_ioobj *obj,
 590                           struct niobuf_remote *nb, int niocount,
 591                           struct niobuf_local *res, struct obd_trans_info *oti,
 592                           int rc)
 593 {
 594         struct niobuf_local *lnb;
 595         struct filter_iobuf *iobuf = NULL;
 596         struct lvfs_run_ctxt saved;
 597         struct fsfilt_objinfo fso;
 598         struct iattr iattr = { 0 };
 599         struct inode *inode = res->dentry->d_inode;
 600         unsigned long now = jiffies;
 601         int i, err, cleanup_phase = 0;
 602         struct obd_device *obd = exp->exp_obd;
 603         struct filter_obd *fo = &obd->u.filter;
 604         void *wait_handle = NULL;
 605         int total_size = 0;
 606         int quota_pending[2] = {0, 0}, quota_pages = 0;
 607         unsigned int qcids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
 608         int sync_journal_commit = obd->u.filter.fo_syncjournal;
 609         int retries = 0;
 610         ENTRY;
 611
 612         LASSERT(oti != NULL);
 613         LASSERT(objcount == 1);
 614         LASSERT(current->journal_info == NULL);
 615
 616         if (rc != 0)
 617                 GOTO(cleanup, rc);
 618
 619         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 620         if (IS_ERR(iobuf))
 621                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 622         cleanup_phase = 1;
 623
 624         fso.fso_dentry = res->dentry;
 625         fso.fso_bufcnt = obj->ioo_bufcnt;
 626
 627         iobuf->dr_ignore_quota = 0;
 628         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 629                 loff_t this_size;
 630
 631
 632                 if (filter_range_is_mapped(inode, lnb->offset, lnb->len)) {
 633                         /* If overwriting an existing block,
 634                          * we don't need a grant */
 635                         if (!(lnb->flags & OBD_BRW_GRANTED) &&
 636                             lnb->rc == -ENOSPC)
 637                                 lnb->rc = 0;
 638                 } else {
 639                         quota_pages++;
 640                 }
 641
 642                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 643                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 644                         continue;
 645                 }
 646
 647                 LASSERT(PageLocked(lnb->page));
 648                 LASSERT(!PageWriteback(lnb->page));
 649
 650                 /* Calling clear_page_dirty_for_io() is no longer needed since
 651                  * write and truncate are serialized by i_alloc_sem and
 652                  * filter_setattr_internal() flushes any partial-page
 653                  * truncates before releasing i_alloc_sem, so there should
 654                  * never be any dirty pages at this time. */
 655                 LASSERT(!PageDirty(lnb->page));
 656
 657                 SetPageUptodate(lnb->page);
 658
 659                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 660                 LASSERT (err == 0);
 661
 662                 total_size += lnb->len;
 663
 664                 /* we expect these pages to be in offset order, but we'll
 665                  * be forgiving */
 666                 this_size = lnb->offset + lnb->len;
 667                 if (this_size > iattr.ia_size)
 668                         iattr.ia_size = this_size;
 669
 670                 /* if one page is a write-back page from client cache and
 671                  * not from direct_io, or it's written by root, then mark
 672                  * the whole io request as ignore quota request */
 673                 if (lnb->flags & OBD_BRW_NOQUOTA ||
 674                     (lnb->flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 675                     OBD_BRW_FROM_GRANT)
 676                         iobuf->dr_ignore_quota = 1;
 677
 678                 if (!(lnb->flags & OBD_BRW_ASYNC)) {
 679                         sync_journal_commit = 1;
 680                 }
 681         }
 682
 683         /* we try to get enough quota to write here, and let ldiskfs
 684          * decide if it is out of quota or not b=14783 */
 685         rc = lquota_chkquota(filter_quota_interface_ref, exp, qcids[0],
 686                              qcids[1], quota_pages, quota_pending, oti,
 687                              inode, obj->ioo_bufcnt);
 688         if (rc == -ENOTCONN)
 689                 GOTO(cleanup, rc);
 690
 691         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 692         cleanup_phase = 2;
 693
 694         fsfilt_check_slow(obd, now, "quota init");
 695
 696 retry:
 697         LOCK_INODE_MUTEX(inode);
 698         fsfilt_check_slow(obd, now, "i_mutex");
 699         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 700                                            oti);
 701         if (IS_ERR(oti->oti_handle)) {
 702                 UNLOCK_INODE_MUTEX(inode);
 703                 rc = PTR_ERR(oti->oti_handle);
 704                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 705                        "error starting transaction: rc = %d\n", rc);
 706                 oti->oti_handle = NULL;
 707                 GOTO(cleanup, rc);
 708         }
 709         /* have to call fsfilt_commit() from this point on */
 710
 711         fsfilt_check_slow(obd, now, "brw_start");
 712
 713         /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
 714         ll_vfs_dq_init(inode);
 715
 716         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 717
 718         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 719          * then we will accept the UID+GID if sent by the client for
 720          * initializing the ownership of this inode.  We only allow this to
 721          * happen once (so clear these bits) and later only allow setattr. */
 722         if (inode->i_mode & S_ISUID)
 723                 i |= OBD_MD_FLUID;
 724         if (inode->i_mode & S_ISGID)
 725                 i |= OBD_MD_FLGID;
 726
 727         iattr_from_obdo(&iattr, oa, i);
 728         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 729                 unsigned int save;
 730
 731                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 732                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 733
 734                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 735
 736                 iattr.ia_valid |= ATTR_MODE;
 737                 iattr.ia_mode = inode->i_mode;
 738                 if (iattr.ia_valid & ATTR_UID)
 739                         iattr.ia_mode &= ~S_ISUID;
 740                 if (iattr.ia_valid & ATTR_GID)
 741                         iattr.ia_mode &= ~S_ISGID;
 742
 743                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 744
 745                 /* To avoid problems with quotas, UID and GID must be set
 746                  * in the inode before filter_direct_io() - see bug 10357. */
 747                 save = iattr.ia_valid;
 748                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 749                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle,&iattr,0);
 750                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 751                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 752                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 753         }
 754
 755         /* filter_direct_io drops i_mutex */
 756         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 757                               oti, sync_journal_commit ? &wait_handle : NULL);
 758         if (rc == -ENOSPC && retries++ < 3) {
 759                 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
 760                        retries);
 761                 oti->oti_handle = NULL;
 762                 fsfilt_check_slow(obd, now, "direct_io");
 763                 goto retry;
 764         }
 765
 766         if (rc == 0)
 767                 obdo_from_inode(oa, inode,
 768                                 FILTER_VALID_FLAGS |OBD_MD_FLUID |OBD_MD_FLGID);
 769         else
 770                 obdo_from_inode(oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
 771
 772         lquota_getflag(filter_quota_interface_ref, obd, oa);
 773
 774         fsfilt_check_slow(obd, now, "direct_io");
 775
 776         if (wait_handle)
 777                 err = fsfilt_commit_wait(obd, inode, wait_handle);
 778         else
 779                 err = 0;
 780
 781         if (err) {
 782                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 783                 if (rc == 0)
 784                         rc = err;
 785         }
 786
 787         if (obd->obd_replayable && !rc && wait_handle)
 788                 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
 789                          "oti_transno "LPU64" last_committed "LPU64"\n",
 790                          oti->oti_transno, obd->obd_last_committed);
 791
 792         fsfilt_check_slow(obd, now, "commitrw commit");
 793
 794 cleanup:
 795         if (quota_pending[0] || quota_pending[1])
 796                 lquota_pending_commit(filter_quota_interface_ref, obd,
 797                                       qcids[0], qcids[1], quota_pending);
 798
 799         filter_grant_commit(exp, niocount, res);
 800
 801         switch (cleanup_phase) {
 802         case 2:
 803                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 804                 LASSERT(current->journal_info == NULL);
 805         case 1:
 806                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 807         case 0:
 808                 /*
 809                  * lnb->page automatically returns back into per-thread page
 810                  * pool (bug 5137)
 811                  */
 812                  break;
 813         }
 814
 815         /* trigger quota pre-acquire */
 816         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 817                             FSFILT_OP_CREATE);
 818         CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 819                "(rc:%d, uid:%u, gid:%u)\n", err, qcids[0], qcids[1]);
 820         if (qcids[USRQUOTA] != oa->o_uid || qcids[GRPQUOTA] != oa->o_gid) {
 821                 qcids[0] = oa->o_uid;
 822                 qcids[1] = oa->o_gid;
 823                 err = lquota_adjust(filter_quota_interface_ref, obd, qcids,
 824                                     NULL, rc, FSFILT_OP_CREATE);
 825                 CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 826                        "(rc:%d, uid:%u, gid:%u)\n", err, qcids[0], qcids[1]);
 827         }
 828
 829         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 830                 if (lnb->page == NULL)
 831                         continue;
 832
 833                 if (rc)
 834                         /* If the write has failed, the page cache may
 835                          * not be consitent with what is on disk, so
 836                          * force pages to be reread next time it is
 837                          * accessed */
 838                         ClearPageUptodate(lnb->page);
 839
 840                 LASSERT(PageLocked(lnb->page));
 841                 unlock_page(lnb->page);
 842
 843                 page_cache_release(lnb->page);
 844                 lnb->page = NULL;
 845         }
 846
 847         if (inode) {
 848                 if (fo->fo_writethrough_cache == 0 ||
 849                     i_size_read(inode) > fo->fo_readcache_max_filesize)
 850                         filter_release_cache(obd, obj, nb, inode);
 851                 up_read(&inode->i_alloc_sem);
 852         }
 853         f_dput(res->dentry);
 854
 855         RETURN(rc);
 856 }