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