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  2008 Sun Microsystems, Inc. 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         if (bio->bi_rw == WRITE &&
 137             unlikely(test_and_clear_bit(BIO_RDONLY, &bio->bi_flags))) {
 138                 struct block_device *bdev = bio->bi_bdev;
 139
 140                 CWARN("Write to readonly device %s (%#x) bi_flags: %lx, "
 141                       "bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_cnt: %d, "
 142                       "bi_private: %p, done: %u, error: %d\n",
 143                       bdev->bd_disk ? bdev->bd_disk->disk_name : "",
 144                       bdev->bd_dev, bio->bi_flags, bio->bi_vcnt, bio->bi_idx,
 145                       bio->bi_size, atomic_read(&bio->bi_cnt), bio->bi_private,
 146                       done, error);
 147         }
 148
 149         /* CAVEAT EMPTOR: possibly in IRQ context
 150          * DO NOT record procfs stats here!!! */
 151
 152         if (bio->bi_size)                       /* Not complete */
 153                 return 1;
 154
 155         if (unlikely(iobuf == NULL)) {
 156                 CERROR("***** bio->bi_private is NULL!  This should never "
 157                        "happen.  Normally, I would crash here, but instead I "
 158                        "will dump the bio contents to the console.  Please "
 159                        "report this to <http://bugzilla.lustre.org/> , along "
 160                        "with any interesting messages leading up to this point "
 161                        "(like SCSI errors, perhaps).  Because bi_private is "
 162                        "NULL, I can't wake up the thread that initiated this "
 163                        "IO - you will probably have to reboot this node.\n");
 164                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
 165                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
 166                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
 167                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
 168                        bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt),
 169                        bio->bi_private);
 170                 return 0;
 171         }
 172
 173         /* the check is outside of the cycle for performance reason -bzzz */
 174         if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
 175                 bio_for_each_segment(bvl, bio, i) {
 176                         if (likely(error == 0))
 177                                 SetPageUptodate(bvl->bv_page);
 178                         LASSERT(PageLocked(bvl->bv_page));
 179                         ClearPageConstant(bvl->bv_page);
 180                 }
 181                 record_finish_io(iobuf, OBD_BRW_READ, error);
 182         } else {
 183                 if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){
 184                         bio_for_each_segment(bvl, bio, i) {
 185                                 ClearPageConstant(bvl->bv_page);
 186                         }
 187                 }
 188                 record_finish_io(iobuf, OBD_BRW_WRITE, error);
 189         }
 190
 191         /* any real error is good enough -bzzz */
 192         if (error != 0 && iobuf->dr_error == 0)
 193                 iobuf->dr_error = error;
 194
 195         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 196          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 197          * mempool when serious on-disk fragmentation was encountered,
 198          * deadlocking the OST.  The bios are now released as soon as complete
 199          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 200         bio_put(bio);
 201         return 0;
 202 }
 203
 204 static int can_be_merged(struct bio *bio, sector_t sector)
 205 {
 206         unsigned int size;
 207
 208         if (!bio)
 209                 return 0;
 210
 211         size = bio->bi_size >> 9;
 212         return bio->bi_sector + size == sector ? 1 : 0;
 213 }
 214
 215 struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter,
 216                                         int rw, int num_pages)
 217 {
 218         struct filter_iobuf *iobuf;
 219
 220         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 221
 222         OBD_ALLOC(iobuf, sizeof(*iobuf));
 223         if (iobuf == NULL)
 224                 goto failed_0;
 225
 226         OBD_ALLOC(iobuf->dr_pages, num_pages * sizeof(*iobuf->dr_pages));
 227         if (iobuf->dr_pages == NULL)
 228                 goto failed_1;
 229
 230         OBD_ALLOC(iobuf->dr_blocks,
 231                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 232         if (iobuf->dr_blocks == NULL)
 233                 goto failed_2;
 234
 235         iobuf->dr_filter = filter;
 236         cfs_waitq_init(&iobuf->dr_wait);
 237         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 238         iobuf->dr_max_pages = num_pages;
 239         iobuf->dr_npages = 0;
 240         iobuf->dr_error = 0;
 241
 242         RETURN(iobuf);
 243
 244  failed_2:
 245         OBD_FREE(iobuf->dr_pages,
 246                  num_pages * sizeof(*iobuf->dr_pages));
 247  failed_1:
 248         OBD_FREE(iobuf, sizeof(*iobuf));
 249  failed_0:
 250         RETURN(ERR_PTR(-ENOMEM));
 251 }
 252
 253 static void filter_clear_iobuf(struct filter_iobuf *iobuf)
 254 {
 255         iobuf->dr_npages = 0;
 256         iobuf->dr_error = 0;
 257         cfs_atomic_set(&iobuf->dr_numreqs, 0);
 258 }
 259
 260 void filter_free_iobuf(struct filter_iobuf *iobuf)
 261 {
 262         int num_pages = iobuf->dr_max_pages;
 263
 264         filter_clear_iobuf(iobuf);
 265
 266         OBD_FREE(iobuf->dr_blocks,
 267                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*iobuf->dr_blocks));
 268         OBD_FREE(iobuf->dr_pages,
 269                  num_pages * sizeof(*iobuf->dr_pages));
 270         OBD_FREE_PTR(iobuf);
 271 }
 272
 273 void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf,
 274                       struct obd_trans_info *oti)
 275 {
 276         int thread_id = (oti && oti->oti_thread) ?
 277                         oti->oti_thread->t_id : -1;
 278
 279         if (unlikely(thread_id < 0)) {
 280                 filter_free_iobuf(iobuf);
 281                 return;
 282         }
 283
 284         LASSERTF(filter->fo_iobuf_pool[thread_id] == iobuf,
 285                  "iobuf mismatch for thread %d: pool %p iobuf %p\n",
 286                  thread_id, filter->fo_iobuf_pool[thread_id], iobuf);
 287         filter_clear_iobuf(iobuf);
 288 }
 289
 290 int filter_iobuf_add_page(struct obd_device *obd, struct filter_iobuf *iobuf,
 291                           struct inode *inode, struct page *page)
 292 {
 293         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 294         iobuf->dr_pages[iobuf->dr_npages++] = page;
 295
 296         return 0;
 297 }
 298
 299 int filter_do_bio(struct obd_export *exp, struct inode *inode,
 300                   struct filter_iobuf *iobuf, int rw)
 301 {
 302         struct obd_device *obd = exp->exp_obd;
 303         int            blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 304         struct page  **pages = iobuf->dr_pages;
 305         int            npages = iobuf->dr_npages;
 306         unsigned long *blocks = iobuf->dr_blocks;
 307         int            total_blocks = npages * blocks_per_page;
 308         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 309         unsigned int   blocksize = inode->i_sb->s_blocksize;
 310         struct bio    *bio = NULL;
 311         int            frags = 0;
 312         unsigned long  start_time = jiffies;
 313         struct page   *page;
 314         unsigned int   page_offset;
 315         sector_t       sector;
 316         int            nblocks;
 317         int            block_idx;
 318         int            page_idx;
 319         int            i;
 320         int            rc = 0;
 321         ENTRY;
 322
 323         LASSERT(iobuf->dr_npages == npages);
 324         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 325
 326         for (page_idx = 0, block_idx = 0;
 327              page_idx < npages;
 328              page_idx++, block_idx += blocks_per_page) {
 329
 330                 page = pages[page_idx];
 331                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 332
 333                 for (i = 0, page_offset = 0;
 334                      i < blocks_per_page;
 335                      i += nblocks, page_offset += blocksize * nblocks) {
 336
 337                         nblocks = 1;
 338
 339                         if (blocks[block_idx + i] == 0) {  /* hole */
 340                                 LASSERT(rw == OBD_BRW_READ);
 341                                 memset(kmap(page) + page_offset, 0, blocksize);
 342                                 kunmap(page);
 343                                 continue;
 344                         }
 345
 346                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 347
 348                         /* Additional contiguous file blocks? */
 349                         while (i + nblocks < blocks_per_page &&
 350                                (sector + (nblocks << sector_bits)) ==
 351                                ((sector_t)blocks[block_idx + i + nblocks] <<
 352                                 sector_bits))
 353                                 nblocks++;
 354
 355                         /* I only set the page to be constant only if it
 356                          * is mapped to a contiguous underlying disk block(s).
 357                          * It will then make sure the corresponding device
 358                          * cache of raid5 will be overwritten by this page.
 359                          * - jay */
 360                         if ((rw == OBD_BRW_WRITE) &&
 361                             (nblocks == blocks_per_page) &&
 362                             mapping_cap_page_constant_write(inode->i_mapping))
 363                                SetPageConstant(page);
 364
 365                         if (bio != NULL &&
 366                             can_be_merged(bio, sector) &&
 367                             bio_add_page(bio, page,
 368                                          blocksize * nblocks, page_offset) != 0)
 369                                 continue;       /* added this frag OK */
 370
 371                         if (bio != NULL) {
 372                                 request_queue_t *q =
 373                                         bdev_get_queue(bio->bi_bdev);
 374
 375                                 /* Dang! I have to fragment this I/O */
 376                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 377                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 378                                        bio->bi_size,
 379                                        bio->bi_vcnt, bio->bi_max_vecs,
 380                                        bio->bi_size >> 9, q->max_sectors,
 381                                        bio_phys_segments(q, bio),
 382                                        q->max_phys_segments,
 383                                        bio_hw_segments(q, bio),
 384                                        q->max_hw_segments);
 385
 386                                 record_start_io(iobuf, rw, bio->bi_size, exp);
 387                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 388                                 if (rc < 0) {
 389                                         CERROR("Can't send bio: %d\n", rc);
 390                                         record_finish_io(iobuf, rw, rc);
 391                                         goto out;
 392                                 }
 393                                 frags++;
 394                         }
 395
 396                         /* allocate new bio, limited by max BIO size, b=9945 */
 397                         bio = bio_alloc(GFP_NOIO, max(BIO_MAX_PAGES,
 398                                                       (npages - page_idx) *
 399                                                       blocks_per_page));
 400                         if (bio == NULL) {
 401                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 402                                        (npages - page_idx), blocks_per_page,
 403                                        (npages - page_idx) * blocks_per_page);
 404                                 rc = -ENOMEM;
 405                                 goto out;
 406                         }
 407
 408                         bio->bi_bdev = inode->i_sb->s_bdev;
 409                         bio->bi_sector = sector;
 410                         bio->bi_end_io = dio_complete_routine;
 411                         bio->bi_private = iobuf;
 412
 413                         rc = bio_add_page(bio, page,
 414                                           blocksize * nblocks, page_offset);
 415                         LASSERT (rc != 0);
 416                 }
 417         }
 418
 419         if (bio != NULL) {
 420                 record_start_io(iobuf, rw, bio->bi_size, exp);
 421                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 422                 if (rc >= 0) {
 423                         frags++;
 424                         rc = 0;
 425                 } else {
 426                         CERROR("Can't send bio: %d\n", rc);
 427                         record_finish_io(iobuf, rw, rc);
 428                 }
 429         }
 430
 431  out:
 432         cfs_wait_event(iobuf->dr_wait,
 433                        cfs_atomic_read(&iobuf->dr_numreqs) == 0);
 434
 435         if (rw == OBD_BRW_READ) {
 436                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 437                                   hist[BRW_R_DIO_FRAGS],
 438                                  frags);
 439                 lprocfs_oh_tally_log2(&obd->u.filter.
 440                                        fo_filter_stats.hist[BRW_R_IO_TIME],
 441                                       jiffies - start_time);
 442                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 443                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 444                                           hist[BRW_R_DIO_FRAGS],
 445                                          frags);
 446                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 447                                              nid_brw_stats->hist[BRW_R_IO_TIME],
 448                                               jiffies - start_time);
 449                 }
 450         } else {
 451                 lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.
 452                                   hist[BRW_W_DIO_FRAGS], frags);
 453                 lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.
 454                                        hist[BRW_W_IO_TIME],
 455                                       jiffies - start_time);
 456                 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) {
 457                         lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
 458                                           hist[BRW_W_DIO_FRAGS],
 459                                          frags);
 460                         lprocfs_oh_tally_log2(&exp->exp_nid_stats->
 461                                              nid_brw_stats->hist[BRW_W_IO_TIME],
 462                                               jiffies - start_time);
 463                 }
 464         }
 465
 466         if (rc == 0)
 467                 rc = iobuf->dr_error;
 468         RETURN(rc);
 469 }
 470
 471 /* Must be called with i_mutex taken for writes; this will drop it */
 472 int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
 473                      struct obd_export *exp, struct iattr *attr,
 474                      struct obd_trans_info *oti, void **wait_handle)
 475 {
 476         struct obd_device *obd = exp->exp_obd;
 477         struct inode *inode = dchild->d_inode;
 478         int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
 479         int rc, rc2, create;
 480         cfs_semaphore_t *sem;
 481         ENTRY;
 482
 483         LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
 484                  iobuf->dr_npages, iobuf->dr_max_pages);
 485         LASSERT(iobuf->dr_npages <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 486
 487         if (rw == OBD_BRW_READ) {
 488                 if (iobuf->dr_npages == 0)
 489                         RETURN(0);
 490                 create = 0;
 491                 sem = NULL;
 492         } else {
 493                 LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
 494                 LASSERT(iobuf->dr_npages > 0);
 495                 create = 1;
 496                 sem = &obd->u.filter.fo_alloc_lock;
 497
 498                 lquota_enforce(filter_quota_interface_ref, obd,
 499                                iobuf->dr_ignore_quota);
 500         }
 501
 502         rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages,
 503                                     iobuf->dr_npages, iobuf->dr_blocks,
 504                                     obdfilter_created_scratchpad, create, sem);
 505
 506         if (rw == OBD_BRW_WRITE) {
 507                 if (rc == 0) {
 508                         filter_tally(exp, iobuf->dr_pages,
 509                                      iobuf->dr_npages, iobuf->dr_blocks,
 510                                      blocks_per_page, 1);
 511                         if (attr->ia_size > i_size_read(inode))
 512                                 attr->ia_valid |= ATTR_SIZE;
 513                         rc = fsfilt_setattr(obd, dchild,
 514                                             oti->oti_handle, attr, 0);
 515                 }
 516
 517                 UNLOCK_INODE_MUTEX(inode);
 518
 519                 rc2 = filter_finish_transno(exp, inode, oti, 0, 0);
 520                 if (rc2 != 0) {
 521                         CERROR("can't close transaction: %d\n", rc2);
 522                         if (rc == 0)
 523                                 rc = rc2;
 524                 }
 525
 526                 rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,
 527                                           wait_handle);
 528                 if (rc == 0)
 529                         rc = rc2;
 530                 if (rc != 0)
 531                         RETURN(rc);
 532         } else if (rc == 0) {
 533                 filter_tally(exp, iobuf->dr_pages, iobuf->dr_npages,
 534                              iobuf->dr_blocks, blocks_per_page, 0);
 535         }
 536
 537         RETURN(filter_do_bio(exp, inode, iobuf, rw));
 538 }
 539
 540 /* See if there are unallocated parts in given file region */
 541 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 542 {
 543         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 544                 inode->i_mapping->a_ops->bmap;
 545         int j;
 546
 547         /* We can't know if we are overwriting or not */
 548         if (fs_bmap == NULL)
 549                 return 0;
 550
 551         offset >>= inode->i_blkbits;
 552         len >>= inode->i_blkbits;
 553
 554         for (j = 0; j <= len; j++)
 555                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 556                         return 0;
 557
 558         return 1;
 559 }
 560
 561 /*
 562  * interesting use cases on how it interacts with VM:
 563  *
 564  * - vm writeout -- shouldn't see our pages as we don't mark them dirty
 565  *   though vm can find partial page left dirty by truncate. in this
 566  *   usual writeout is used unless our write rewrite that page - then we
 567  *   drop PG_dirty with PG_lock held.
 568  *
 569  * - else?
 570  *
 571  */
 572 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 573                           int objcount, struct obd_ioobj *obj,
 574                           struct niobuf_remote *nb, int niocount,
 575                           struct niobuf_local *res, struct obd_trans_info *oti,
 576                           int rc)
 577 {
 578         struct niobuf_local *lnb;
 579         struct filter_iobuf *iobuf = NULL;
 580         struct lvfs_run_ctxt saved;
 581         struct fsfilt_objinfo fso;
 582         struct iattr iattr = { 0 };
 583         struct inode *inode = res->dentry->d_inode;
 584         unsigned long now = jiffies;
 585         int i, err, cleanup_phase = 0;
 586         struct obd_device *obd = exp->exp_obd;
 587         struct filter_obd *fo = &obd->u.filter;
 588         void *wait_handle;
 589         int total_size = 0;
 590         unsigned int qcids[MAXQUOTAS] = { oa->o_uid, oa->o_gid };
 591         int rec_pending[MAXQUOTAS] = { 0, 0 }, quota_pages = 0;
 592         ENTRY;
 593
 594         LASSERT(oti != NULL);
 595         LASSERT(objcount == 1);
 596         LASSERT(current->journal_info == NULL);
 597
 598         if (rc != 0)
 599                 GOTO(cleanup, rc);
 600
 601         iobuf = filter_iobuf_get(&obd->u.filter, oti);
 602         if (IS_ERR(iobuf))
 603                 GOTO(cleanup, rc = PTR_ERR(iobuf));
 604         cleanup_phase = 1;
 605
 606         fso.fso_dentry = res->dentry;
 607         fso.fso_bufcnt = obj->ioo_bufcnt;
 608
 609         iobuf->dr_ignore_quota = 0;
 610         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 611                 loff_t this_size;
 612                 __u32 flags = lnb->flags;
 613
 614                 if (filter_range_is_mapped(inode, lnb->offset, lnb->len)) {
 615                         /* If overwriting an existing block,
 616                          * we don't need a grant */
 617                         if (!(flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC)
 618                                 lnb->rc = 0;
 619                 } else {
 620                         quota_pages++;
 621                 }
 622
 623                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 624                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 625                         continue;
 626                 }
 627
 628                 LASSERT(PageLocked(lnb->page));
 629                 LASSERT(!PageWriteback(lnb->page));
 630
 631                 /* since write & truncate are serialized by the i_alloc_sem,
 632                  * even partial truncate should not leave dirty pages in
 633                  * the page cache */
 634                 LASSERT(!PageDirty(lnb->page));
 635
 636                 SetPageUptodate(lnb->page);
 637
 638                 err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page);
 639                 LASSERT (err == 0);
 640
 641                 total_size += lnb->len;
 642
 643                 /* we expect these pages to be in offset order, but we'll
 644                  * be forgiving */
 645                 this_size = lnb->offset + lnb->len;
 646                 if (this_size > iattr.ia_size)
 647                         iattr.ia_size = this_size;
 648
 649                 /* if one page is a write-back page from client cache and
 650                  * not from direct_io, or it's written by root, then mark
 651                  * the whole io request as ignore quota request, remote
 652                  * client can not break through quota. */
 653                 if (exp_connect_rmtclient(exp))
 654                         flags &= ~OBD_BRW_NOQUOTA;
 655                 if ((flags & OBD_BRW_NOQUOTA) ||
 656                     (flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
 657                      OBD_BRW_FROM_GRANT)
 658                         iobuf->dr_ignore_quota = 1;
 659         }
 660
 661         /* we try to get enough quota to write here, and let ldiskfs
 662          * decide if it is out of quota or not b=14783 */
 663         rc = lquota_chkquota(filter_quota_interface_ref, obd, exp, qcids,
 664                              rec_pending, quota_pages, oti, LQUOTA_FLAGS_BLK,
 665                              (void *)inode, obj->ioo_bufcnt);
 666         if (rc == -ENOTCONN)
 667                 GOTO(cleanup, rc);
 668
 669         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 670         cleanup_phase = 2;
 671
 672         DQUOT_INIT(inode);
 673
 674         LOCK_INODE_MUTEX(inode);
 675         fsfilt_check_slow(obd, now, "i_mutex");
 676         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 677                                            oti);
 678         if (IS_ERR(oti->oti_handle)) {
 679                 UNLOCK_INODE_MUTEX(inode);
 680                 rc = PTR_ERR(oti->oti_handle);
 681                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 682                        "error starting transaction: rc = %d\n", rc);
 683                 oti->oti_handle = NULL;
 684                 GOTO(cleanup, rc);
 685         }
 686         /* have to call fsfilt_commit() from this point on */
 687
 688         fsfilt_check_slow(obd, now, "brw_start");
 689
 690         i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 691
 692         /* If the inode still has SUID+SGID bits set (see filter_precreate())
 693          * then we will accept the UID+GID if sent by the client for
 694          * initializing the ownership of this inode.  We only allow this to
 695          * happen once (so clear these bits) and later only allow setattr. */
 696         if (inode->i_mode & S_ISUID)
 697                 i |= OBD_MD_FLUID;
 698         if (inode->i_mode & S_ISGID)
 699                 i |= OBD_MD_FLGID;
 700
 701         iattr_from_obdo(&iattr, oa, i);
 702         if (iattr.ia_valid & (ATTR_UID | ATTR_GID)) {
 703                 unsigned int save;
 704
 705                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
 706                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
 707
 708                 cfs_cap_raise(CFS_CAP_SYS_RESOURCE);
 709
 710                 iattr.ia_valid |= ATTR_MODE;
 711                 iattr.ia_mode = inode->i_mode;
 712                 if (iattr.ia_valid & ATTR_UID)
 713                         iattr.ia_mode &= ~S_ISUID;
 714                 if (iattr.ia_valid & ATTR_GID)
 715                         iattr.ia_mode &= ~S_ISGID;
 716
 717                 rc = filter_update_fidea(exp, inode, oti->oti_handle, oa);
 718
 719                 /* To avoid problems with quotas, UID and GID must be set
 720                  * in the inode before filter_direct_io() - see bug 10357. */
 721                 save = iattr.ia_valid;
 722                 iattr.ia_valid &= (ATTR_UID | ATTR_GID);
 723                 rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle,&iattr,0);
 724                 CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n",
 725                                 iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc);
 726                 iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID);
 727         }
 728
 729         /* filter_direct_io drops i_mutex */
 730         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 731                               oti, &wait_handle);
 732         if (rc == 0)
 733                 obdo_from_inode(oa, inode,
 734                                 FILTER_VALID_FLAGS |OBD_MD_FLUID |OBD_MD_FLGID);
 735         else
 736                 obdo_from_inode(oa, inode, OBD_MD_FLUID | OBD_MD_FLGID);
 737
 738         lquota_getflag(filter_quota_interface_ref, obd, oa);
 739
 740         fsfilt_check_slow(obd, now, "direct_io");
 741
 742         err = fsfilt_commit_wait(obd, inode, wait_handle);
 743         if (err) {
 744                 CERROR("Failure to commit OST transaction (%d)?\n", err);
 745                 rc = err;
 746         }
 747
 748         if (obd->obd_replayable && !rc)
 749                 LASSERTF(oti->oti_transno <= obd->obd_last_committed,
 750                          "oti_transno "LPU64" last_committed "LPU64"\n",
 751                          oti->oti_transno, obd->obd_last_committed);
 752
 753         fsfilt_check_slow(obd, now, "commitrw commit");
 754
 755 cleanup:
 756         lquota_pending_commit(filter_quota_interface_ref, obd, qcids,
 757                               rec_pending, 1);
 758
 759         filter_grant_commit(exp, niocount, res);
 760
 761         switch (cleanup_phase) {
 762         case 2:
 763                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 764                 LASSERT(current->journal_info == NULL);
 765         case 1:
 766                 filter_iobuf_put(&obd->u.filter, iobuf, oti);
 767         case 0:
 768                 /*
 769                  * lnb->page automatically returns back into per-thread page
 770                  * pool (bug 5137)
 771                  */
 772                  break;
 773         }
 774
 775         /* trigger quota pre-acquire */
 776         err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
 777                             FSFILT_OP_CREATE);
 778         CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 779                "(rc:%d, uid:%u, gid:%u)\n",
 780                err, qcids[USRQUOTA], qcids[GRPQUOTA]);
 781         if (qcids[USRQUOTA] != oa->o_uid || qcids[GRPQUOTA] != oa->o_gid) {
 782                 qcids[USRQUOTA] = oa->o_uid;
 783                 qcids[GRPQUOTA] = oa->o_gid;
 784                 err = lquota_adjust(filter_quota_interface_ref, obd, qcids,
 785                                     NULL, rc, FSFILT_OP_CREATE);
 786                 CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! "
 787                        "(rc:%d, uid:%u, gid:%u)\n",
 788                        err, qcids[USRQUOTA], qcids[GRPQUOTA]);
 789         }
 790
 791         for (i = 0, lnb = res; i < niocount; i++, lnb++) {
 792                 if (lnb->page == NULL)
 793                         continue;
 794
 795                 if (rc)
 796                         /* If the write has failed, the page cache may
 797                          * not be consitent with what is on disk, so
 798                          * force pages to be reread next time it is
 799                          * accessed */
 800                         ClearPageUptodate(lnb->page);
 801
 802                 LASSERT(PageLocked(lnb->page));
 803                 unlock_page(lnb->page);
 804
 805                 page_cache_release(lnb->page);
 806                 lnb->page = NULL;
 807         }
 808         f_dput(res->dentry);
 809
 810         if (inode) {
 811                 if (fo->fo_writethrough_cache == 0 ||
 812                     i_size_read(inode) > fo->fo_readcache_max_filesize)
 813                         filter_release_cache(obd, obj, nb, inode);
 814                 up_read(&inode->i_alloc_sem);
 815         }
 816
 817         RETURN(rc);
 818 }