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         cfs_atomic_t       dr_numreqs;  /* number of reqs being processed */
  62         cfs_waitq_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         cfs_atomic_inc(&iobuf->dr_numreqs);
  78
  79         if (rw == OBD_BRW_READ) {
  80                 cfs_atomic_inc(&filter->fo_r_in_flight);
  81                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST],
  82                                  cfs_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                                          cfs_atomic_read(&filter-> \
  90                                          fo_r_in_flight));
  91                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
  92                                          nid_brw_stats->hist[BRW_R_DISK_IOSIZE],
  93                                               size);
  94                 }
  95         } else {
  96                 cfs_atomic_inc(&filter->fo_w_in_flight);
  97                 lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST],
  98                                  cfs_atomic_read(&filter->fo_w_in_flight));
  99                 lprocfs_oh_tally_log2(&filter->
 100                                        fo_filter_stats.hist[BRW_W_DISK_IOSIZE],
 101                                       size);
 102                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 103                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 104                                           hist[BRW_W_RPC_HIST],
 105                                          cfs_atomic_read(&filter-> \
 106                                          fo_r_in_flight));
 107                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 108                                         nid_brw_stats->hist[BRW_W_DISK_IOSIZE],
 109                                               size);
 110                 }
 111         }
 112 }
 113
 114 static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
 115 {
 116         struct filter_obd *filter = iobuf->dr_filter;
 117
 118         /* CAVEAT EMPTOR: possibly in IRQ context
 119          * DO NOT record procfs stats here!!! */
 120
 121         if (rw == OBD_BRW_READ)
 122                 cfs_atomic_dec(&filter->fo_r_in_flight);
 123         else
 124                 cfs_atomic_dec(&filter->fo_w_in_flight);
 125
 126         if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs))
 127                 cfs_waitq_signal(&iobuf->dr_wait);
 128 }
 129
 130 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 131 {
 132         struct filter_iobuf *iobuf = bio->bi_private;
 133         struct bio_vec *bvl;
 134         int i;
 135
 136         /* CAVEAT EMPTOR: possibly in IRQ context
 137          * DO NOT record procfs stats here!!! */
 138
 139 #ifdef HAVE_BIO_ENDIO_2ARG
 140        /* The "bi_size" check was needed for kernels < 2.6.24 in order to
 141         * handle the case where a SCSI request error caused this callback
 142         * to be called before all of the biovecs had been processed.
 143         * Without this check the server thread will hang.  In newer kernels
 144         * the bio_end_io routine is never called for partial completions,
 145         * so this check is no longer needed. */
 146         if (bio->bi_size)                      /* Not complete */
 147                 DIO_RETURN(1);
 148 #endif
 149
 150         if (unlikely(iobuf == NULL)) {
 151                 CERROR("***** bio->bi_private is NULL!  This should never "
 152                        "happen.  Normally, I would crash here, but instead I "
 153                        "will dump the bio contents to the console.  Please "
 154                        "report this to <http://bugzilla.lustre.org/> , along "
 155                        "with any interesting messages leading up to this point "
 156                        "(like SCSI errors, perhaps).  Because bi_private is "
 157                        "NULL, I can't wake up the thread that initiated this "
 158                        "IO - you will probably have to reboot this node.\n");
 159                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 160                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 161                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 162                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 163                        bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
 164                        bio->bi_private);
 165                 return 0;
 166         }
 167
 168         /* the check is outside of the cycle for performance reason -bzzz */
 169         if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
 170                 bio_for_each_segment(bvl, bio, i) {
 171                         if (likely(error == 0))
 172                                 SetPageUptodate(bvl->bv_page);
 173                         LASSERT(PageLocked(bvl->bv_page));
 174                         ClearPageConstant(bvl->bv_page);
 175                 }
 176                 record_finish_io(iobuf, OBD_BRW_READ, error);
 177         } else {
 178                 if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){
 179                         bio_for_each_segment(bvl, bio, i) {
 180                                 ClearPageConstant(bvl->bv_page);
 181                         }
 182                 }
 183                 record_finish_io(iobuf, OBD_BRW_WRITE, error);
 184         }
 185
 186         /* any real error is good enough -bzzz */
 187         if (error != 0 && iobuf->dr_error == 0)
 188                 iobuf->dr_error = error;
 189
 190         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 191          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 192          * mempool when serious on-disk fragmentation was encountered,
 193          * deadlocking the OST.  The bios are now released as soon as complete
 194          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 195         bio_put(bio);
 196         return 0;
 197 }
 198
 199 static int can_be_merged(struct bio *bio, sector_t sector)
 200 {
 201         unsigned int size;
 202
 203         if (!bio)
 204                 return 0;
 205
 206         size = bio->bi_size >> 9;
 207         return bio->bi_sector + size == sector ? 1 : 0;
 208 }
 209
 210 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 211                                         int rw, int num_pages)
 212 {
 213         struct filter_iobuf *iobuf;
 214
 215         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 216
 217         OBD_ALLOC(iobuf, sizeof(*iobuf));
 218         if (iobuf == NULL)
 219                 goto failed_0;
 220
 221         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 222         if (iobuf->dr_pages == NULL)
 223                 goto failed_1;
 224
 225         OBD_ALLOC(iobuf->dr_blocks,
 226                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 227         if (iobuf->dr_blocks == NULL)
 228                 goto failed_2;
 229
 230         iobuf->dr_filter = filter;
 231         cfs_waitq_init(&iobuf->dr_wait);
 232         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 233         iobuf->dr_max_pages = num_pages;
 234         iobuf->dr_npages = 0;
 235         iobuf->dr_error = 0;
 236
 237         RETURN(iobuf);
 238
 239  failed_2:
 240         OBD_FREE(iobuf->dr_pages,
 241                  num_pages * sizeof(*iobuf->dr_pages));
 242  failed_1:
 243         OBD_FREE(iobuf, sizeof(*iobuf));
 244  failed_0:
 245         RETURN(ERR_PTR(-ENOMEM));
 246 }
 247
 248 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
 249 {
 250         iobuf->dr_npages = 0;
 251         iobuf->dr_error = 0;
 252         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 253 }
 254
 255 void filter_free_iobuf(struct filter_iobuf *iobuf)
 256 {
 257         int num_pages = iobuf->dr_max_pages;
 258
 259         filter_clear_iobuf(iobuf);
 260
 261         OBD_FREE(iobuf->dr_blocks,
 262                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 263         OBD_FREE(iobuf->dr_pages,
 264                  num_pages * sizeof(*iobuf->dr_pages));
 265         OBD_FREE_PTR(iobuf);
 266 }
 267
 268 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 269                       struct obd_trans_info *oti)
 270 {
 271         int thread_id = (oti && oti->oti_thread) ?
 272                         oti->oti_thread->t_id : -1;
 273
 274         if (unlikely(thread_id < 0)) {
 275                 filter_free_iobuf(iobuf);
 276                 return;
 277         }
 278
 279         LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
 280                  "iobuf mismatch for thread %d: pool %p iobuf %p\n",
 281                  thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
 282         filter_clear_iobuf(iobuf);
 283 }
 284
 285 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 286                           struct inode *inode, struct page *page)
 287 {
 288         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 289         iobuf->dr_pages[iobuf->dr_npages++] = page;
 290
 291         return 0;
 292 }
 293
 294 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 295                   struct filter_iobuf *iobuf, int rw)
 296 {
 297         struct obd_device *obd = exp->exp_obd;
 298         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 299         struct page  **pages = iobuf->dr_pages;
 300         int            npages = iobuf->dr_npages;
 301         unsigned long *blocks = iobuf->dr_blocks;
 302         int            total_blocks = npages * blocks_per_page;
 303         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 304         unsigned int   blocksize = inode->i_sb->s_blocksize;
 305         struct bio    *bio = NULL;
 306         int            frags = 0;
 307         unsigned long  start_time = jiffies;
 308         struct page   *page;
 309         unsigned int   page_offset;
 310         sector_t       sector;
 311         int            nblocks;
 312         int            block_idx;
 313         int            page_idx;
 314         int            i;
 315         int            rc = 0;
 316         ENTRY;
 317
 318         LASSERT(iobuf->dr_npages == npages);
 319         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 320
 321         for (page_idx = 0, block_idx = 0;
 322              page_idx < npages;
 323              page_idx++, block_idx += blocks_per_page) {
 324
 325                 page = pages[page_idx];
 326                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 327
 328                 for (i = 0, page_offset = 0;
 329                      i < blocks_per_page;
 330                      i += nblocks, page_offset += blocksize * nblocks) {
 331
 332                         nblocks = 1;
 333
 334                         if (blocks[block_idx + i] == 0) {  /* hole */
 335                                 LASSERT(rw == OBD_BRW_READ);
 336                                 memset(kmap(page) + page_offset, 0, blocksize);
 337                                 kunmap(page);
 338                                 continue;
 339                         }
 340
 341                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 342
 343                         /* Additional contiguous file blocks? */
 344                         while (i + nblocks < blocks_per_page &&
 345                                (sector + (nblocks << sector_bits)) ==
 346                                ((sector_t)blocks[block_idx + i + nblocks] <<
 347                                 sector_bits))
 348                                 nblocks++;
 349
 350                         /* I only set the page to be constant only if it
 351                          * is mapped to a contiguous underlying disk block(s).
 352                          * It will then make sure the corresponding device
 353                          * cache of raid5 will be overwritten by this page.
 354                          * - jay */
 355                         if ((rw == OBD_BRW_WRITE) &&
 356                             (nblocks == blocks_per_page) &&
 357                             mapping_cap_page_constant_write(inode->i_mapping))
 358                                SetPageConstant(page);
 359
 360                         if (bio != NULL &&
 361                             can_be_merged(bio, sector) &&
 362                             bio_add_page(bio, page,
 363                                          blocksize * nblocks, page_offset) != 0)
 364                                 continue;       /* added this frag OK */
 365
 366                         if (bio != NULL) {
 367                                 request_queue_t *q =
 368                                         bdev_get_queue(bio->bi_bdev);
 369
 370                                 /* Dang! I have to fragment this I/O */
 371                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 372                                        "sectors %d(%d) psg %d(%d) hsg %d(%d) "
 373                                        "sector %llu next %llu\n",
 374                                        bio->bi_size,
 375                                        bio->bi_vcnt, bio->bi_max_vecs,
 376                                        bio->bi_size >> 9, q->max_sectors,
 377                                        bio_phys_segments(q, bio),
 378                                        q->max_phys_segments,
 379                                        bio_hw_segments(q, bio),
 380                                        q->max_hw_segments,
 381                                        (unsigned long long)bio->bi_sector,
 382                                        (unsigned long long)sector);
 383
 384                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 385                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 386                                 if (rc < 0) {
 387                                         CERROR("Can't send bio: %d\n", rc);
 388                                         record_finish_io(iobuf, rw, rc);
 389                                         goto out;
 390                                 }
 391                                 frags++;
 392                         }
 393
 394                         /* allocate new bio, limited by max BIO size, b=9945 */
 395                         bio = bio_alloc(GFP_NOIO, max(BIO_MAX_PAGES,
 396                                                       (npages - page_idx) *
 397                                                       blocks_per_page));
 398                         if (bio == NULL) {
 399                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 400                                        (npages - page_idx), blocks_per_page,
 401                                        (npages - page_idx) * blocks_per_page);
 402                                 rc = -ENOMEM;
 403                                 goto out;
 404                         }
 405
 406                         bio->bi_bdev = inode->i_sb->s_bdev;
 407                         bio->bi_sector = sector;
 408                         bio->bi_end_io = dio_complete_routine;
 409                         bio->bi_private = iobuf;
 410
 411                         rc = bio_add_page(bio, page,
 412                                           blocksize * nblocks, page_offset);
 413                         LASSERT (rc != 0);
 414                 }
 415         }
 416
 417         if (bio != NULL) {
 418                 record_start_io(iobuf, rw, bio->bi_size, exp);
 419                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 420                 if (rc >= 0) {
 421                         frags++;
 422                         rc = 0;
 423                 } else {
 424                         CERROR("Can't send bio: %d\n", rc);
 425                         record_finish_io(iobuf, rw, rc);
 426                 }
 427         }
 428
 429  out:
 430         cfs_wait_event(iobuf->dr_wait,
 431                        cfs_atomic_read(&iobuf->dr_numreqs) == 0);
 432
 433         if (rw == OBD_BRW_READ) {
 434                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 435                                   hist[BRW_R_DIO_FRAGS],
 436                                  frags);
 437                 lprocfs_oh_tally_log2(&obd->u.filter.
 438                                        fo_filter_stats.hist[BRW_R_IO_TIME],
 439                                       jiffies - start_time);
 440                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 441                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 442                                           hist[BRW_R_DIO_FRAGS],
 443                                          frags);
 444                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 445                                              nid_brw_stats->hist[BRW_R_IO_TIME],
 446                                               jiffies - start_time);
 447                 }
 448         } else {
 449                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 450                                   hist[BRW_W_DIO_FRAGS], frags);
 451                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.
 452                                        hist[BRW_W_IO_TIME],
 453                                       jiffies - start_time);
 454                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 455                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 456                                           hist[BRW_W_DIO_FRAGS],
 457                                          frags);
 458                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 459                                              nid_brw_stats->hist[BRW_W_IO_TIME],
 460                                               jiffies - start_time);
 461                 }
 462         }
 463
 464         if (rc == 0)
 465                 rc = iobuf->dr_error;
 466         RETURN(rc);
 467 }
 468
 469 /* Must be called with i_mutex taken for writes; this will drop it */
 470 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 471                      struct obd_export *exp, struct iattr *attr,
 472                      struct obd_trans_info *oti, void **wait_handle)
 473 {
 474         struct obd_device *obd = exp->exp_obd;
 475         struct inode *inode = dchild->d_inode;
 476         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 477         int rc, rc2, create;
 478         cfs_semaphore_t *sem;
 479         ENTRY;
 480
 481         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 482                  iobuf->dr_npages, iobuf->dr_max_pages);
 483         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 484
 485         if (rw == OBD_BRW_READ) {
 486                 if (iobuf->dr_npages == 0)
 487                         RETURN(0);
 488                 create = 0;
 489                 sem = NULL;
 490         } else {
 491                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 492                 LASSERT(iobuf->dr_npages > 0);
 493                 create = 1;
 494                 sem = &obd->u.filter.fo_alloc_lock;
 495
 496                 lquota_enforce(filter_quota_interface_ref, obd,
 497                                iobuf->dr_ignore_quota);
 498         }
 499
 500         rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 501                                     iobuf->dr_npages, iobuf->dr_blocks,
 502                                     obdfilter_created_scratchpad, create, sem);
 503
 504         if (rw == OBD_BRW_WRITE) {
 505                 if (rc == 0) {
 506                         filter_tally(exp, iobuf->dr_pages,
 507                                      iobuf->dr_npages, iobuf->dr_blocks,
 508                                      blocks_per_page, 1);
 509                         if (attr->ia_size > i_size_read(inode))
 510                                 attr->ia_valid |= ATTR_SIZE;
 511                         rc = fsfilt_setattr(obd, dchild,
 512                                             oti->oti_handle, attr, 0);
 513                 }
 514
 515                 UNLOCK_INODE_MUTEX(inode);
 516
 517                 rc2 = filter_finish_transno(exp, inode, oti, 0, 0);
 518                 if (rc2 != 0) {
 519                         CERROR("can't close transaction: %d\n", rc2);
 520                         if (rc == 0)
 521                                 rc = rc2;
 522                 }
 523
 524                 if (wait_handle)
 525                         rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
 526                                                   wait_handle);
 527                 else
 528                         rc2 = fsfilt_commit(obd, inode, oti->oti_handle, 0);
 529                 if (rc == 0)
 530                         rc = rc2;
 531                 if (rc != 0)
 532                         RETURN(rc);
 533         } else if (rc == 0) {
 534                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 535                              iobuf->dr_blocks, blocks_per_page, 0);
 536         }
 537
 538         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 539 }
 540
 541 /* See if there are unallocated parts in given file region */
 542 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 543 {
 544         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 545                 inode->i_mapping->a_ops->bmap;
 546         int j;
 547
 548         /* We can't know if we are overwriting or not */
 549         if (fs_bmap == NULL)
 550                 return 0;
 551
 552         offset >>= inode->i_blkbits;
 553         len >>= inode->i_blkbits;
 554
 555         for (j = 0; j <= len; j++)
 556                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 557                         return 0;
 558
 559         return 1;
 560 }
 561
 562 /*
 563  * interesting use cases on how it interacts with VM:
 564  *
 565  * - vm writeout -- shouldn't see our pages as we don't mark them dirty
 566  *   though vm can find partial page left dirty by truncate. in this
 567  *   usual writeout is used unless our write rewrite that page - then we
 568  *   drop PG_dirty with PG_lock held.
 569  *
 570  * - else?
 571  *
 572  */
 573 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 574                           int objcount, struct obd_ioobj *obj,
 575                           struct niobuf_remote *nb, int niocount,
 576                           struct niobuf_local *res, struct obd_trans_info *oti,
 577                           int rc)
 578 {
 579         struct niobuf_local *lnb;
 580         struct filter_iobuf *iobuf = NULL;
 581         struct lvfs_run_ctxt saved;
 582         struct fsfilt_objinfo fso;
 583         struct iattr iattr = { 0 };
 584         struct inode *inode = res->dentry->d_inode;
 585         unsigned long now = jiffies;
 586         int i, err, cleanup_phase = 0;
 587         struct obd_device *obd = exp->exp_obd;
 588         struct filter_obd *fo = &obd->u.filter;
 589         void *wait_handle = NULL;
 590         int total_size = 0;
 591         unsigned int qcids[MAXQUOTAS] = { oa->o_uid, oa->o_gid };
 592         int rec_pending[MAXQUOTAS] = { 0, 0 }, quota_pages = 0;
 593         int sync_journal_commit = obd->u.filter.fo_syncjournal;
 594         ENTRY;
 595
 596         LASSERT(oti != NULL);
 597         LASSERT(objcount == 1);
 598         LASSERT(current->journal_info == NULL);
 599
 600         if (rc != 0)
 601                 GOTO(cleanup, rc);
 602
 603         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 604         if (IS_ERR(iobuf))
 605                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 606         cleanup_phase = 1;
 607
 608         fso.fso_dentry = res->dentry;
 609         fso.fso_bufcnt = obj->ioo_bufcnt;
 610
 611         iobuf->dr_ignore_quota = 0;
 612         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 613                 loff_t this_size;
 614                 __u32 flags = lnb->flags;
 615
 616                 if (filter_range_is_mapped(inode, lnb->offset, lnb->len)) {
 617                         /* If overwriting an existing block,
 618                          * we don't need a grant */
 619                         if (!(flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC)
 620                                 lnb->rc = 0;
 621                 } else {
 622                         quota_pages++;
 623                 }
 624
 625                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 626                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 627                         continue;
 628                 }
 629
 630                 LASSERT(PageLocked(lnb->page));
 631                 LASSERT(!PageWriteback(lnb->page));
 632
 633                 /* since write & truncate are serialized by the i_alloc_sem,
 634                  * even partial truncate should not leave dirty pages in
 635                  * the page cache */
 636                 LASSERT(!PageDirty(lnb->page));
 637
 638                 SetPageUptodate(lnb->page);
 639
 640                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 641                 LASSERT (err == 0);
 642
 643                 total_size += lnb->len;
 644
 645                 /* we expect these pages to be in offset order, but we'll
 646                  * be forgiving */
 647                 this_size = lnb->offset + lnb->len;
 648                 if (this_size > iattr.ia_size)
 649                         iattr.ia_size = this_size;
 650
 651                 /* if one page is a write-back page from client cache and
 652                  * not from direct_io, or it's written by root, then mark
 653                  * the whole io request as ignore quota request, remote
 654                  * client can not break through quota. */
 655                 if (exp_connect_rmtclient(exp))
 656                         flags &= ~OBD_BRW_NOQUOTA;
 657                 if ((flags & OBD_BRW_NOQUOTA) ||
 658                     (flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 659                      OBD_BRW_FROM_GRANT)
 660                         iobuf->dr_ignore_quota = 1;
 661
 662                 if (!(lnb->flags & OBD_BRW_ASYNC)) {
 663                         sync_journal_commit = 1;
 664                 }
 665         }
 666
 667         /* we try to get enough quota to write here, and let ldiskfs
 668          * decide if it is out of quota or not b=14783 */
 669         rc = lquota_chkquota(filter_quota_interface_ref, obd, exp, qcids,
 670                              rec_pending, quota_pages, oti, LQUOTA_FLAGS_BLK,
 671                              (void *)inode, obj->ioo_bufcnt);
 672         if (rc == -ENOTCONN)
 673                 GOTO(cleanup, rc);
 674
 675         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 676         cleanup_phase = 2;
 677
 678         ll_vfs_dq_init(inode);
 679         fsfilt_check_slow(obd, now, "quota init");
 680
 681         LOCK_INODE_MUTEX(inode);
 682         fsfilt_check_slow(obd, now, "i_mutex");
 683         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 684                                            oti);
 685         if (IS_ERR(oti->oti_handle)) {
 686                 UNLOCK_INODE_MUTEX(inode);
 687                 rc = PTR_ERR(oti->oti_handle);
 688                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 689                        "error starting transaction: rc = %d\n", rc);
 690                 oti->oti_handle = NULL;
 691                 GOTO(cleanup, rc);
 692         }
 693         /* have to call fsfilt_commit() from this point on */
 694
 695         fsfilt_check_slow(obd, now, "brw_start");
 696
 697         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 698
 699         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 700          * then we will accept the UID+GID if sent by the client for
 701          * initializing the ownership of this inode.  We only allow this to
 702          * happen once (so clear these bits) and later only allow setattr. */
 703         if (inode->i_mode & S_ISUID)
 704                 i |= OBD_MD_FLUID;
 705         if (inode->i_mode & S_ISGID)
 706                 i |= OBD_MD_FLGID;
 707
 708         iattr_from_obdo(&iattr, oa, i);
 709         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 710                 unsigned int save;
 711
 712                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 713                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 714
 715                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 716
 717                 iattr.ia_valid |= ATTR_MODE;
 718                 iattr.ia_mode = inode->i_mode;
 719                 if (iattr.ia_valid & ATTR_UID)
 720                         iattr.ia_mode &= ~S_ISUID;
 721                 if (iattr.ia_valid & ATTR_GID)
 722                         iattr.ia_mode &= ~S_ISGID;
 723
 724                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 725
 726                 /* To avoid problems with quotas, UID and GID must be set
 727                  * in the inode before filter_direct_io() - see bug 10357. */
 728                 save = iattr.ia_valid;
 729                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 730                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle,&iattr,0);
 731                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 732                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 733                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 734         }
 735
 736         /* filter_direct_io drops i_mutex */
 737         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 738                               oti, sync_journal_commit ? &wait_handle : NULL);
 739
 740         obdo_from_inode(oa, inode, NULL, rc == 0 ? FILTER_VALID_FLAGS : 0 |
 741                                                    OBD_MD_FLUID |OBD_MD_FLGID);
 742
 743         lquota_getflag(filter_quota_interface_ref, obd, oa);
 744
 745         fsfilt_check_slow(obd, now, "direct_io");
 746
 747         if (wait_handle)
 748                 err = fsfilt_commit_wait(obd, inode, wait_handle);
 749         else
 750                 err = 0;
 751
 752         if (err) {
 753                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 754                 if (rc == 0)
 755                         rc = err;
 756         }
 757
 758         if (obd->obd_replayable && !rc && wait_handle)
 759                 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
 760                          "oti_transno "LPU64" last_committed "LPU64"\n",
 761                          oti->oti_transno, obd->obd_last_committed);
 762
 763         fsfilt_check_slow(obd, now, "commitrw commit");
 764
 765 cleanup:
 766         lquota_pending_commit(filter_quota_interface_ref, obd, qcids,
 767                               rec_pending, 1);
 768
 769         filter_grant_commit(exp, niocount, res);
 770
 771         switch (cleanup_phase) {
 772         case 2:
 773                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 774                 LASSERT(current->journal_info == NULL);
 775         case 1:
 776                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 777         case 0:
 778                 /*
 779                  * lnb->page automatically returns back into per-thread page
 780                  * pool (bug 5137)
 781                  */
 782                  break;
 783         }
 784
 785         /* trigger quota pre-acquire */
 786         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 787                             FSFILT_OP_CREATE);
 788         CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 789                "(rc:%d, uid:%u, gid:%u)\n",
 790                err, qcids[USRQUOTA], qcids[GRPQUOTA]);
 791         if (qcids[USRQUOTA] != oa->o_uid || qcids[GRPQUOTA] != oa->o_gid) {
 792                 qcids[USRQUOTA] = oa->o_uid;
 793                 qcids[GRPQUOTA] = oa->o_gid;
 794                 err = lquota_adjust(filter_quota_interface_ref, obd, qcids,
 795                                     NULL, rc, FSFILT_OP_CREATE);
 796                 CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 797                        "(rc:%d, uid:%u, gid:%u)\n",
 798                        err, qcids[USRQUOTA], qcids[GRPQUOTA]);
 799         }
 800
 801         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 802                 if (lnb->page == NULL)
 803                         continue;
 804
 805                 if (rc)
 806                         /* If the write has failed, the page cache may
 807                          * not be consitent with what is on disk, so
 808                          * force pages to be reread next time it is
 809                          * accessed */
 810                         ClearPageUptodate(lnb->page);
 811
 812                 LASSERT(PageLocked(lnb->page));
 813                 unlock_page(lnb->page);
 814
 815                 page_cache_release(lnb->page);
 816                 lnb->page = NULL;
 817         }
 818         f_dput(res->dentry);
 819
 820         if (inode) {
 821                 if (fo->fo_writethrough_cache == 0 ||
 822                     i_size_read(inode) > fo->fo_readcache_max_filesize)
 823                         filter_release_cache(obd, obj, nb, inode);
 824                 up_read(&inode->i_alloc_sem);
 825         }
 826
 827         RETURN(rc);
 828 }