lustre/obdfilter/filter_io_26.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdfilter/filter_io_26.c
  37  *
  38  * Author: Peter Braam <braam@clusterfs.com>
  39  * Author: Andreas Dilger <adilger@clusterfs.com>
  40  * Author: Phil Schwan <phil@clusterfs.com>
  41  */
  42
  43 #ifndef AUTOCONF_INCLUDED
  44 #include <linux/config.h>
  45 #endif
  46 #include <linux/module.h>
  47 #include <linux/pagemap.h> // XXX kill me soon
  48 #include <linux/version.h>
  49 #include <linux/buffer_head.h>
  50
  51 #define DEBUG_SUBSYSTEM S_FILTER
  52
  53 #include <obd_class.h>
  54 #include <lustre_fsfilt.h>
  55 #include <lustre_quota.h>
  56 #include "filter_internal.h"
  57
  58 /* 512byte block min */
  59 #define MAX_BLOCKS_PER_PAGE (CFS_PAGE_SIZE / 512)
  60 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_w_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 #ifdef HAVE_BIO_ENDIO_2ARG
 131 #define DIO_RETURN(a)
 132 static void dio_complete_routine(struct bio *bio, int error)
 133 #else
 134 #define DIO_RETURN(a)   return(a)
 135 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
 136 #endif
 137 {
 138         struct filter_iobuf *iobuf = bio->bi_private;
 139         struct bio_vec *bvl;
 140         int i;
 141
 142         /* CAVEAT EMPTOR: possibly in IRQ context
 143          * DO NOT record procfs stats here!!! */
 144
 145 #ifdef HAVE_BIO_ENDIO_2ARG
 146        /* The "bi_size" check was needed for kernels < 2.6.24 in order to
 147         * handle the case where a SCSI request error caused this callback
 148         * to be called before all of the biovecs had been processed.
 149         * Without this check the server thread will hang.  In newer kernels
 150         * the bio_end_io routine is never called for partial completions,
 151         * so this check is no longer needed. */
 152         if (bio->bi_size)                      /* Not complete */
 153                 DIO_RETURN(1);
 154 #endif
 155
 156         if (unlikely(iobuf == NULL)) {
 157                 CERROR("***** bio->bi_private is NULL!  This should never "
 158                        "happen.  Normally, I would crash here, but instead I "
 159                        "will dump the bio contents to the console.  Please "
 160                        "report this to <http://bugs.whamcloud.com/>, along "
 161                        "with any interesting messages leading up to this point "
 162                        "(like SCSI errors, perhaps).  Because bi_private is "
 163                        "NULL, I can't wake up the thread that initiated this "
 164                        "IO - you will probably have to reboot this node.\n");
 165                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 166                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 167                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 168                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 169                        bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
 170                        bio->bi_private);
 171                 DIO_RETURN(0);
 172         }
 173
 174         /* the check is outside of the cycle for performance reason -bzzz */
 175         if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
 176                 bio_for_each_segment(bvl, bio, i) {
 177                         if (likely(error == 0))
 178                                 SetPageUptodate(bvl->bv_page);
 179                         LASSERT(PageLocked(bvl->bv_page));
 180                         ClearPageConstant(bvl->bv_page);
 181                 }
 182                 record_finish_io(iobuf, OBD_BRW_READ, error);
 183         } else {
 184                 if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){
 185                         bio_for_each_segment(bvl, bio, i) {
 186                                 ClearPageConstant(bvl->bv_page);
 187                         }
 188                 }
 189                 record_finish_io(iobuf, OBD_BRW_WRITE, error);
 190         }
 191
 192         /* any real error is good enough -bzzz */
 193         if (error != 0 && iobuf->dr_error == 0)
 194                 iobuf->dr_error = error;
 195
 196         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 197          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 198          * mempool when serious on-disk fragmentation was encountered,
 199          * deadlocking the OST.  The bios are now released as soon as complete
 200          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 201         bio_put(bio);
 202         DIO_RETURN(0);
 203 }
 204
 205 static int can_be_merged(struct bio *bio, sector_t sector)
 206 {
 207         unsigned int size;
 208
 209         if (!bio)
 210                 return 0;
 211
 212         size = bio->bi_size >> 9;
 213         return bio->bi_sector + size == sector ? 1 : 0;
 214 }
 215
 216 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 217                                         int rw, int num_pages)
 218 {
 219         struct filter_iobuf *iobuf;
 220
 221         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 222
 223         OBD_ALLOC(iobuf, sizeof(*iobuf));
 224         if (iobuf == NULL)
 225                 goto failed_0;
 226
 227         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 228         if (iobuf->dr_pages == NULL)
 229                 goto failed_1;
 230
 231         OBD_ALLOC(iobuf->dr_blocks,
 232                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 233         if (iobuf->dr_blocks == NULL)
 234                 goto failed_2;
 235
 236         iobuf->dr_filter = filter;
 237         cfs_waitq_init(&iobuf->dr_wait);
 238         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 239         iobuf->dr_max_pages = num_pages;
 240         iobuf->dr_npages = 0;
 241         iobuf->dr_error = 0;
 242
 243         RETURN(iobuf);
 244
 245  failed_2:
 246         OBD_FREE(iobuf->dr_pages,
 247                  num_pages * sizeof(*iobuf->dr_pages));
 248  failed_1:
 249         OBD_FREE(iobuf, sizeof(*iobuf));
 250  failed_0:
 251         RETURN(ERR_PTR(-ENOMEM));
 252 }
 253
 254 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
 255 {
 256         iobuf->dr_npages = 0;
 257         iobuf->dr_error = 0;
 258         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 259 }
 260
 261 void filter_free_iobuf(struct filter_iobuf *iobuf)
 262 {
 263         int num_pages = iobuf->dr_max_pages;
 264
 265         filter_clear_iobuf(iobuf);
 266
 267         OBD_FREE(iobuf->dr_blocks,
 268                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 269         OBD_FREE(iobuf->dr_pages,
 270                  num_pages * sizeof(*iobuf->dr_pages));
 271         OBD_FREE_PTR(iobuf);
 272 }
 273
 274 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 275                       struct obd_trans_info *oti)
 276 {
 277         int thread_id = (oti && oti->oti_thread) ?
 278                         oti->oti_thread->t_id : -1;
 279
 280         if (unlikely(thread_id < 0)) {
 281                 filter_free_iobuf(iobuf);
 282                 return;
 283         }
 284
 285         LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
 286                  "iobuf mismatch for thread %d: pool %p iobuf %p\n",
 287                  thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
 288         filter_clear_iobuf(iobuf);
 289 }
 290
 291 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 292                           struct inode *inode, struct page *page)
 293 {
 294         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 295         iobuf->dr_pages[iobuf->dr_npages++] = page;
 296
 297         return 0;
 298 }
 299
 300 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 301                   struct filter_iobuf *iobuf, int rw)
 302 {
 303         struct obd_device *obd = exp->exp_obd;
 304         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 305         struct page  **pages = iobuf->dr_pages;
 306         int            npages = iobuf->dr_npages;
 307         unsigned long *blocks = iobuf->dr_blocks;
 308         int            total_blocks = npages * blocks_per_page;
 309         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 310         unsigned int   blocksize = inode->i_sb->s_blocksize;
 311         struct bio    *bio = NULL;
 312         int            frags = 0;
 313         unsigned long  start_time = jiffies;
 314         struct page   *page;
 315         unsigned int   page_offset;
 316         sector_t       sector;
 317         int            nblocks;
 318         int            block_idx;
 319         int            page_idx;
 320         int            i;
 321         int            rc = 0;
 322         ENTRY;
 323
 324         LASSERT(iobuf->dr_npages == npages);
 325         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 326
 327         for (page_idx = 0, block_idx = 0;
 328              page_idx < npages;
 329              page_idx++, block_idx += blocks_per_page) {
 330
 331                 page = pages[page_idx];
 332                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 333
 334                 for (i = 0, page_offset = 0;
 335                      i < blocks_per_page;
 336                      i += nblocks, page_offset += blocksize * nblocks) {
 337
 338                         nblocks = 1;
 339
 340                         if (blocks[block_idx + i] == 0) {  /* hole */
 341                                 LASSERT(rw == OBD_BRW_READ);
 342                                 memset(kmap(page) + page_offset, 0, blocksize);
 343                                 kunmap(page);
 344                                 continue;
 345                         }
 346
 347                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 348
 349                         /* Additional contiguous file blocks? */
 350                         while (i + nblocks < blocks_per_page &&
 351                                (sector + (nblocks << sector_bits)) ==
 352                                ((sector_t)blocks[block_idx + i + nblocks] <<
 353                                 sector_bits))
 354                                 nblocks++;
 355
 356                         /* I only set the page to be constant only if it
 357                          * is mapped to a contiguous underlying disk block(s).
 358                          * It will then make sure the corresponding device
 359                          * cache of raid5 will be overwritten by this page.
 360                          * - jay */
 361                         if ((rw == OBD_BRW_WRITE) &&
 362                             (nblocks == blocks_per_page) &&
 363                             mapping_cap_page_constant_write(inode->i_mapping))
 364                                SetPageConstant(page);
 365
 366                         if (bio != NULL &&
 367                             can_be_merged(bio, sector) &&
 368                             bio_add_page(bio, page,
 369                                          blocksize * nblocks, page_offset) != 0)
 370                                 continue;       /* added this frag OK */
 371
 372                         if (bio != NULL) {
 373                                 struct request_queue *q =
 374                                         bdev_get_queue(bio->bi_bdev);
 375
 376                                 /* Dang! I have to fragment this I/O */
 377                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 378                                        "sectors %d(%d) psg %d(%d) hsg %d(%d) "
 379                                        "sector %llu next %llu\n",
 380                                        bio->bi_size,
 381                                        bio->bi_vcnt, bio->bi_max_vecs,
 382                                        bio->bi_size >> 9, queue_max_sectors(q),
 383                                        bio_phys_segments(q, bio),
 384                                        queue_max_phys_segments(q),
 385                                        bio_hw_segments(q, bio),
 386                                        queue_max_hw_segments(q),
 387                                        (unsigned long long)bio->bi_sector,
 388                                        (unsigned long long)sector);
 389
 390                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 391                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 392                                 if (rc < 0) {
 393                                         CERROR("Can't send bio: %d\n", rc);
 394                                         record_finish_io(iobuf, rw, rc);
 395                                         goto out;
 396                                 }
 397                                 frags++;
 398                         }
 399
 400                         /* allocate new bio, limited by max BIO size, b=9945 */
 401                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 402                                                       (npages - page_idx) *
 403                                                       blocks_per_page));
 404                         if (bio == NULL) {
 405                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 406                                        (npages - page_idx), blocks_per_page,
 407                                        (npages - page_idx) * blocks_per_page);
 408                                 rc = -ENOMEM;
 409                                 goto out;
 410                         }
 411
 412                         bio->bi_bdev = inode->i_sb->s_bdev;
 413                         bio->bi_sector = sector;
 414                         bio->bi_end_io = dio_complete_routine;
 415                         bio->bi_private = iobuf;
 416
 417                         rc = bio_add_page(bio, page,
 418                                           blocksize * nblocks, page_offset);
 419                         LASSERT (rc != 0);
 420                 }
 421         }
 422
 423         if (bio != NULL) {
 424                 record_start_io(iobuf, rw, bio->bi_size, exp);
 425                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 426                 if (rc >= 0) {
 427                         frags++;
 428                         rc = 0;
 429                 } else {
 430                         CERROR("Can't send bio: %d\n", rc);
 431                         record_finish_io(iobuf, rw, rc);
 432                 }
 433         }
 434
 435  out:
 436         cfs_wait_event(iobuf->dr_wait,
 437                        cfs_atomic_read(&iobuf->dr_numreqs) == 0);
 438
 439         if (rw == OBD_BRW_READ) {
 440                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 441                                   hist[BRW_R_DIO_FRAGS],
 442                                  frags);
 443                 lprocfs_oh_tally_log2(&obd->u.filter.
 444                                        fo_filter_stats.hist[BRW_R_IO_TIME],
 445                                       jiffies - start_time);
 446                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 447                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 448                                           hist[BRW_R_DIO_FRAGS],
 449                                          frags);
 450                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 451                                              nid_brw_stats->hist[BRW_R_IO_TIME],
 452                                               jiffies - start_time);
 453                 }
 454         } else {
 455                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 456                                   hist[BRW_W_DIO_FRAGS], frags);
 457                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.
 458                                        hist[BRW_W_IO_TIME],
 459                                       jiffies - start_time);
 460                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 461                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 462                                           hist[BRW_W_DIO_FRAGS],
 463                                          frags);
 464                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 465                                              nid_brw_stats->hist[BRW_W_IO_TIME],
 466                                               jiffies - start_time);
 467                 }
 468         }
 469
 470         if (rc == 0)
 471                 rc = iobuf->dr_error;
 472         RETURN(rc);
 473 }
 474
 475 /* Must be called with i_mutex taken for writes; this will drop it */
 476 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 477                      struct obd_export *exp, struct iattr *attr,
 478                      struct obd_trans_info *oti, void **wait_handle)
 479 {
 480         struct obd_device *obd = exp->exp_obd;
 481         struct inode *inode = dchild->d_inode;
 482         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 483         int rc, rc2, create;
 484         cfs_mutex_t *mutex;
 485         ENTRY;
 486
 487         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 488                  iobuf->dr_npages, iobuf->dr_max_pages);
 489         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 490
 491         if (rw == OBD_BRW_READ) {
 492                 if (iobuf->dr_npages == 0)
 493                         RETURN(0);
 494                 create = 0;
 495                 mutex = NULL;
 496         } else {
 497                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 498                 LASSERT(iobuf->dr_npages > 0);
 499                 create = 1;
 500                 mutex = &obd->u.filter.fo_alloc_lock;
 501
 502                 lquota_enforce(filter_quota_interface_ref, obd,
 503                                iobuf->dr_ignore_quota);
 504         }
 505
 506         if (rw == OBD_BRW_WRITE &&
 507             OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
 508                 rc = -ENOSPC;
 509         } else {
 510                 rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 511                                     iobuf->dr_npages, iobuf->dr_blocks,
 512                                     obdfilter_created_scratchpad, create, mutex);
 513         }
 514
 515         if (rw == OBD_BRW_WRITE) {
 516                 if (rc == 0) {
 517                         filter_tally(exp, iobuf->dr_pages,
 518                                      iobuf->dr_npages, iobuf->dr_blocks,
 519                                      blocks_per_page, 1);
 520                         if (attr->ia_size > i_size_read(inode))
 521                                 attr->ia_valid |= ATTR_SIZE;
 522                         rc = fsfilt_setattr(obd, dchild,
 523                                             oti->oti_handle, attr, 0);
 524                 }
 525
 526                 UNLOCK_INODE_MUTEX(inode);
 527
 528                 /* Force commit to make the just-deleted blocks
 529                  * reusable. LU-456 */
 530                 if (rc == -ENOSPC) {
 531                         fsfilt_commit(obd, inode, oti->oti_handle, 1);
 532                         RETURN(rc);
 533                 }
 534
 535                 rc2 = filter_finish_transno(exp, inode, oti, 0, 0);
 536                 if (rc2 != 0) {
 537                         CERROR("can't close transaction: %d\n", rc2);
 538                         if (rc == 0)
 539                                 rc = rc2;
 540                 }
 541
 542                 if (wait_handle)
 543                         rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
 544                                                   wait_handle);
 545                 else
 546                         rc2 = fsfilt_commit(obd, inode, oti->oti_handle, 0);
 547                 if (rc == 0)
 548                         rc = rc2;
 549                 if (rc != 0)
 550                         RETURN(rc);
 551         } else if (rc == 0) {
 552                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 553                              iobuf->dr_blocks, blocks_per_page, 0);
 554         }
 555
 556         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 557 }
 558
 559 /* See if there are unallocated parts in given file region */
 560 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 561 {
 562         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 563                 inode->i_mapping->a_ops->bmap;
 564         int j;
 565
 566         /* We can't know if we are overwriting or not */
 567         if (fs_bmap == NULL)
 568                 return 0;
 569
 570         offset >>= inode->i_blkbits;
 571         len >>= inode->i_blkbits;
 572
 573         for (j = 0; j <= len; j++)
 574                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 575                         return 0;
 576
 577         return 1;
 578 }
 579
 580 /*
 581  * interesting use cases on how it interacts with VM:
 582  *
 583  * - vm writeout -- shouldn't see our pages as we don't mark them dirty
 584  *   though vm can find partial page left dirty by truncate. in this
 585  *   usual writeout is used unless our write rewrite that page - then we
 586  *   drop PG_dirty with PG_lock held.
 587  *
 588  * - else?
 589  *
 590  */
 591 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 592                           int objcount, struct obd_ioobj *obj,
 593                           struct niobuf_remote *nb, int niocount,
 594                           struct niobuf_local *res, struct obd_trans_info *oti,
 595                           int rc)
 596 {
 597         struct niobuf_local *lnb;
 598         struct filter_iobuf *iobuf = NULL;
 599         struct lvfs_run_ctxt saved;
 600         struct fsfilt_objinfo fso;
 601         struct iattr iattr = { 0 };
 602         struct inode *inode = res->dentry->d_inode;
 603         unsigned long now = jiffies;
 604         int i, err, cleanup_phase = 0;
 605         struct obd_device *obd = exp->exp_obd;
 606         struct filter_obd *fo = &obd->u.filter;
 607         void *wait_handle = NULL;
 608         int total_size = 0;
 609         unsigned int qcids[MAXQUOTAS] = { oa->o_uid, oa->o_gid };
 610         int rec_pending[MAXQUOTAS] = { 0, 0 }, quota_pages = 0;
 611         int sync_journal_commit = obd->u.filter.fo_syncjournal;
 612         int retries = 0;
 613         ENTRY;
 614
 615         LASSERT(oti != NULL);
 616         LASSERT(objcount == 1);
 617         LASSERT(current->journal_info == NULL);
 618
 619         if (rc != 0)
 620                 GOTO(cleanup, rc);
 621
 622         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 623         if (IS_ERR(iobuf))
 624                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 625         cleanup_phase = 1;
 626
 627         fso.fso_dentry = res->dentry;
 628         fso.fso_bufcnt = obj->ioo_bufcnt;
 629
 630         iobuf->dr_ignore_quota = 0;
 631         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 632                 loff_t this_size;
 633                 __u32 flags = lnb->flags;
 634
 635                 if (filter_range_is_mapped(inode, lnb->offset, lnb->len)) {
 636                         /* If overwriting an existing block,
 637                          * we don't need a grant */
 638                         if (!(flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC)
 639                                 lnb->rc = 0;
 640                 } else {
 641                         quota_pages++;
 642                 }
 643
 644                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 645                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 646                         continue;
 647                 }
 648
 649                 LASSERT(PageLocked(lnb->page));
 650                 LASSERT(!PageWriteback(lnb->page));
 651
 652                 /* since write & truncate are serialized by the i_alloc_sem,
 653                  * even partial truncate should not leave dirty pages in
 654                  * the page cache */
 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, remote
 673                  * client can not break through quota. */
 674                 if (exp_connect_rmtclient(exp))
 675                         flags &= ~OBD_BRW_NOQUOTA;
 676                 if ((flags & OBD_BRW_NOQUOTA) ||
 677                     (flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 678                      OBD_BRW_FROM_GRANT)
 679                         iobuf->dr_ignore_quota = 1;
 680
 681                 if (!(lnb->flags & OBD_BRW_ASYNC)) {
 682                         sync_journal_commit = 1;
 683                 }
 684         }
 685
 686         /* we try to get enough quota to write here, and let ldiskfs
 687          * decide if it is out of quota or not b=14783 */
 688         rc = lquota_chkquota(filter_quota_interface_ref, obd, exp, qcids,
 689                              rec_pending, quota_pages, oti, LQUOTA_FLAGS_BLK,
 690                              (void *)inode, obj->ioo_bufcnt);
 691         if (rc == -ENOTCONN)
 692                 GOTO(cleanup, rc);
 693
 694         if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
 695                 GOTO(cleanup, rc = -EINPROGRESS);
 696
 697         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 698         cleanup_phase = 2;
 699
 700         fsfilt_check_slow(obd, now, "quota init");
 701
 702 retry:
 703         LOCK_INODE_MUTEX(inode);
 704         fsfilt_check_slow(obd, now, "i_mutex");
 705         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 706                                            oti);
 707         if (IS_ERR(oti->oti_handle)) {
 708                 UNLOCK_INODE_MUTEX(inode);
 709                 rc = PTR_ERR(oti->oti_handle);
 710                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 711                        "error starting transaction: rc = %d\n", rc);
 712                 oti->oti_handle = NULL;
 713                 GOTO(cleanup, rc);
 714         }
 715         /* have to call fsfilt_commit() from this point on */
 716
 717         fsfilt_check_slow(obd, now, "brw_start");
 718
 719         /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
 720         ll_vfs_dq_init(inode);
 721
 722         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 723
 724         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 725          * then we will accept the UID+GID if sent by the client for
 726          * initializing the ownership of this inode.  We only allow this to
 727          * happen once (so clear these bits) and later only allow setattr. */
 728         if (inode->i_mode & S_ISUID)
 729                 i |= OBD_MD_FLUID;
 730         if (inode->i_mode & S_ISGID)
 731                 i |= OBD_MD_FLGID;
 732
 733         iattr_from_obdo(&iattr, oa, i);
 734         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 735                 unsigned int save;
 736
 737                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 738                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 739
 740                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 741
 742                 iattr.ia_valid |= ATTR_MODE;
 743                 iattr.ia_mode = inode->i_mode;
 744                 if (iattr.ia_valid & ATTR_UID)
 745                         iattr.ia_mode &= ~S_ISUID;
 746                 if (iattr.ia_valid & ATTR_GID)
 747                         iattr.ia_mode &= ~S_ISGID;
 748
 749                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 750
 751                 /* To avoid problems with quotas, UID and GID must be set
 752                  * in the inode before filter_direct_io() - see bug 10357. */
 753                 save = iattr.ia_valid;
 754                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 755                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle,&iattr,0);
 756                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 757                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 758                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 759         }
 760
 761         /* filter_direct_io drops i_mutex */
 762         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 763                               oti, sync_journal_commit ? &wait_handle : NULL);
 764         if (rc == -ENOSPC && retries++ < 3) {
 765                 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
 766                        retries);
 767                 oti->oti_handle = NULL;
 768                 fsfilt_check_slow(obd, now, "direct_io");
 769                 goto retry;
 770         }
 771
 772         obdo_from_inode(oa, inode, (rc == 0 ? FILTER_VALID_FLAGS : 0) |
 773                                    OBD_MD_FLUID | OBD_MD_FLGID);
 774
 775         lquota_getflag(filter_quota_interface_ref, obd, oa);
 776
 777         fsfilt_check_slow(obd, now, "direct_io");
 778
 779         if (wait_handle)
 780                 err = fsfilt_commit_wait(obd, inode, wait_handle);
 781         else
 782                 err = 0;
 783
 784         if (err) {
 785                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 786                 if (rc == 0)
 787                         rc = err;
 788         }
 789
 790         /* In rare cases fsfilt_commit_wait() will wake up and return after
 791          * the transaction has finished its work and updated j_commit_sequence
 792          * but the commit callbacks have not been run yet.  Wait here until
 793          * that is finished so that clients requesting sync IO don't see the
 794          * reply transno < last_committed.  LU-753 */
 795         if (unlikely(obd->obd_replayable && !rc && wait_handle &&
 796                      oti->oti_transno > obd->obd_last_committed)) {
 797                 cfs_waitq_t wq;
 798                 struct l_wait_info lwi =
 799                         LWI_TIMEOUT_INTERVAL(cfs_time_seconds(5),
 800                                              (cfs_duration_t)((HZ + 4)/5),
 801                                              NULL, NULL);
 802                 cfs_waitq_init(&wq);
 803                 l_wait_event(wq,
 804                              oti->oti_transno <= obd->obd_last_committed,
 805                              &lwi);
 806
 807                 /* commit callback isn't done after waiting for 5 secs ? */
 808                 if (unlikely(oti->oti_transno > obd->obd_last_committed))
 809                         CERROR("transno:"LPU64" > last_committed:"LPU64"\n",
 810                                oti->oti_transno, obd->obd_last_committed);
 811         }
 812
 813         fsfilt_check_slow(obd, now, "commitrw commit");
 814
 815 cleanup:
 816         lquota_pending_commit(filter_quota_interface_ref, obd, qcids,
 817                               rec_pending, 1);
 818
 819         filter_grant_commit(exp, niocount, res);
 820
 821         switch (cleanup_phase) {
 822         case 2:
 823                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 824                 LASSERT(current->journal_info == NULL);
 825         case 1:
 826                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 827         case 0:
 828                 /*
 829                  * lnb->page automatically returns back into per-thread page
 830                  * pool (bug 5137)
 831                  */
 832                  break;
 833         }
 834
 835         /* trigger quota pre-acquire */
 836         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 837                             FSFILT_OP_CREATE);
 838         CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 839                "(rc:%d, uid:%u, gid:%u)\n",
 840                err, qcids[USRQUOTA], qcids[GRPQUOTA]);
 841         if (qcids[USRQUOTA] != oa->o_uid || qcids[GRPQUOTA] != oa->o_gid) {
 842                 qcids[USRQUOTA] = oa->o_uid;
 843                 qcids[GRPQUOTA] = oa->o_gid;
 844                 err = lquota_adjust(filter_quota_interface_ref, obd, qcids,
 845                                     NULL, rc, FSFILT_OP_CREATE);
 846                 CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 847                        "(rc:%d, uid:%u, gid:%u)\n",
 848                        err, qcids[USRQUOTA], qcids[GRPQUOTA]);
 849         }
 850
 851         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 852                 if (lnb->page == NULL)
 853                         continue;
 854
 855                 if (rc)
 856                         /* If the write has failed, the page cache may
 857                          * not be consitent with what is on disk, so
 858                          * force pages to be reread next time it is
 859                          * accessed */
 860                         ClearPageUptodate(lnb->page);
 861
 862                 LASSERT(PageLocked(lnb->page));
 863                 unlock_page(lnb->page);
 864
 865                 page_cache_release(lnb->page);
 866                 lnb->page = NULL;
 867         }
 868         f_dput(res->dentry);
 869
 870         if (inode) {
 871                 if (fo->fo_writethrough_cache == 0 ||
 872                     i_size_read(inode) > fo->fo_readcache_max_filesize)
 873                         filter_release_cache(obd, obj, nb, inode);
 874                 up_read(&inode->i_alloc_sem);
 875         }
 876
 877         RETURN(rc);
 878 }