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  *  linux/fs/obdfilter/filter_io.c
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *   Author: Peter Braam <braam@clusterfs.com>
   8  *   Author: Andreas Dilger <adilger@clusterfs.com>
   9  *   Author: Phil Schwan <phil@clusterfs.com>
  10  *
  11  *   This file is part of Lustre, http://www.lustre.org.
  12  *
  13  *   Lustre is free software; you can redistribute it and/or
  14  *   modify it under the terms of version 2 of the GNU General Public
  15  *   License as published by the Free Software Foundation.
  16  *
  17  *   Lustre is distributed in the hope that it will be useful,
  18  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  *   GNU General Public License for more details.
  21  *
  22  *   You should have received a copy of the GNU General Public License
  23  *   along with Lustre; if not, write to the Free Software
  24  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  25  */
  26
  27 #include <linux/config.h>
  28 #include <linux/module.h>
  29 #include <linux/pagemap.h> // XXX kill me soon
  30 #include <linux/version.h>
  31 #include <linux/buffer_head.h>
  32
  33 #define DEBUG_SUBSYSTEM S_FILTER
  34
  35 #include <linux/obd_class.h>
  36 #include <linux/lustre_fsfilt.h>
  37 #include "filter_internal.h"
  38
  39 #warning "implement writeback mode -bzzz"
  40
  41 /* 512byte block min */
  42 #define MAX_BLOCKS_PER_PAGE (PAGE_SIZE / 512)
  43 struct dio_request {
  44         atomic_t          dr_numreqs;  /* number of reqs being processed */
  45         struct bio       *dr_bios;     /* list of completed bios */
  46         wait_queue_head_t dr_wait;
  47         int               dr_max_pages;
  48         int               dr_npages;
  49         int               dr_error;
  50         struct page     **dr_pages;
  51         unsigned long    *dr_blocks;
  52         spinlock_t        dr_lock;
  53 };
  54
  55 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
  56 {
  57         struct dio_request *dreq = bio->bi_private;
  58         unsigned long flags;
  59
  60         if (bio->bi_size) {
  61                 CWARN("gets called against non-complete bio 0x%p: %d/%d/%d\n",
  62                       bio, bio->bi_size, done, error);
  63                 return 1;
  64         }
  65
  66         if (dreq == NULL) {
  67                 CERROR("***** bio->bi_private is NULL!  This should never "
  68                        "happen.  Normally, I would crash here, but instead I "
  69                        "will dump the bio contents to the console.  Please "
  70                        "report this to CFS, along with any interesting messages "
  71                        "leading up to this point (like SCSI errors, perhaps).  "
  72                        "Because bi_private is NULL, I can't wake up the thread "
  73                        "that initiated this I/O -- so you will probably have to "
  74                        "reboot this node.");
  75                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
  76                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
  77                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
  78                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
  79                        bio->bi_end_io, atomic_read(&bio->bi_cnt),
  80                        bio->bi_private);
  81                 return 0;
  82         }
  83
  84         spin_lock_irqsave(&dreq->dr_lock, flags);
  85         bio->bi_private = dreq->dr_bios;
  86         dreq->dr_bios = bio;
  87         if (dreq->dr_error == 0)
  88                 dreq->dr_error = error;
  89         spin_unlock_irqrestore(&dreq->dr_lock, flags);
  90
  91         if (atomic_dec_and_test(&dreq->dr_numreqs))
  92                 wake_up(&dreq->dr_wait);
  93
  94         return 0;
  95 }
  96
  97 static int can_be_merged(struct bio *bio, sector_t sector)
  98 {
  99         unsigned int size;
 100         if (!bio)
 101                 return 0;
 102
 103         size = bio->bi_size >> 9;
 104         return bio->bi_sector + size == sector ? 1 : 0;
 105 }
 106
 107
 108 int filter_alloc_iobuf(int rw, int num_pages, void **ret)
 109 {
 110         struct dio_request *dreq;
 111
 112         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 113
 114         OBD_ALLOC(dreq, sizeof(*dreq));
 115         if (dreq == NULL)
 116                 goto failed_0;
 117
 118         OBD_ALLOC(dreq->dr_pages, num_pages * sizeof(*dreq->dr_pages));
 119         if (dreq->dr_pages == NULL)
 120                 goto failed_1;
 121
 122         OBD_ALLOC(dreq->dr_blocks,
 123                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*dreq->dr_blocks));
 124         if (dreq->dr_blocks == NULL)
 125                 goto failed_2;
 126
 127         dreq->dr_bios = NULL;
 128         init_waitqueue_head(&dreq->dr_wait);
 129         atomic_set(&dreq->dr_numreqs, 0);
 130         spin_lock_init(&dreq->dr_lock);
 131         dreq->dr_max_pages = num_pages;
 132         dreq->dr_npages = 0;
 133
 134         *ret = dreq;
 135         RETURN(0);
 136
 137  failed_2:
 138         OBD_FREE(dreq->dr_pages,
 139                  num_pages * sizeof(*dreq->dr_pages));
 140  failed_1:
 141         OBD_FREE(dreq, sizeof(*dreq));
 142  failed_0:
 143         RETURN(-ENOMEM);
 144 }
 145
 146 void filter_free_iobuf(void *iobuf)
 147 {
 148         struct dio_request *dreq = iobuf;
 149         int                 num_pages = dreq->dr_max_pages;
 150
 151         /* free all bios */
 152         while (dreq->dr_bios) {
 153                 struct bio *bio = dreq->dr_bios;
 154                 dreq->dr_bios = bio->bi_private;
 155                 bio_put(bio);
 156         }
 157
 158         OBD_FREE(dreq->dr_blocks,
 159                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*dreq->dr_blocks));
 160         OBD_FREE(dreq->dr_pages,
 161                  num_pages * sizeof(*dreq->dr_pages));
 162         OBD_FREE(dreq, sizeof(*dreq));
 163 }
 164
 165 int filter_iobuf_add_page(struct obd_device *obd, void *iobuf,
 166                           struct inode *inode, struct page *page)
 167 {
 168         struct dio_request *dreq = iobuf;
 169
 170         LASSERT (dreq->dr_npages < dreq->dr_max_pages);
 171         dreq->dr_pages[dreq->dr_npages++] = page;
 172
 173         return 0;
 174 }
 175
 176 int filter_do_bio(struct obd_device *obd, struct inode *inode,
 177                   struct dio_request *dreq, int rw)
 178 {
 179         int            blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 180         struct page  **pages = dreq->dr_pages;
 181         int            npages = dreq->dr_npages;
 182         unsigned long *blocks = dreq->dr_blocks;
 183         int            total_blocks = npages * blocks_per_page;
 184         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 185         unsigned int   blocksize = inode->i_sb->s_blocksize;
 186         struct bio    *bio = NULL;
 187         struct page   *page;
 188         unsigned int   page_offset;
 189         sector_t       sector;
 190         int            nblocks;
 191         int            block_idx;
 192         int            page_idx;
 193         int            i;
 194         int            rc = 0;
 195         ENTRY;
 196
 197         LASSERT(dreq->dr_npages == npages);
 198         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 199
 200         for (page_idx = 0, block_idx = 0;
 201              page_idx < npages;
 202              page_idx++, block_idx += blocks_per_page) {
 203
 204                 page = pages[page_idx];
 205                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 206
 207                 for (i = 0, page_offset = 0;
 208                      i < blocks_per_page;
 209                      i += nblocks, page_offset += blocksize * nblocks) {
 210
 211                         nblocks = 1;
 212
 213                         if (blocks[block_idx + i] == 0) {  /* hole */
 214                                 LASSERT(rw == OBD_BRW_READ);
 215                                 memset(kmap(page) + page_offset, 0, blocksize);
 216                                 kunmap(page);
 217                                 continue;
 218                         }
 219
 220                         sector = blocks[block_idx + i] << sector_bits;
 221
 222                         /* Additional contiguous file blocks? */
 223                         while (i + nblocks < blocks_per_page &&
 224                                (sector + nblocks*(blocksize>>9)) ==
 225                                (blocks[block_idx + i + nblocks] << sector_bits))
 226                                 nblocks++;
 227
 228                         if (bio != NULL &&
 229                             can_be_merged(bio, sector) &&
 230                             bio_add_page(bio, page,
 231                                          blocksize * nblocks, page_offset) != 0)
 232                                 continue;       /* added this frag OK */
 233
 234                         if (bio != NULL) {
 235                                 request_queue_t *q = bdev_get_queue(bio->bi_bdev);
 236
 237                                 /* Dang! I have to fragment this I/O */
 238                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 239                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 240                                        bio->bi_size,
 241                                        bio->bi_vcnt, bio->bi_max_vecs,
 242                                        bio->bi_size >> 9, q->max_sectors,
 243                                        bio_phys_segments(q, bio),
 244                                        q->max_phys_segments,
 245                                        bio_hw_segments(q, bio),
 246                                        q->max_hw_segments);
 247
 248                                 atomic_inc(&dreq->dr_numreqs);
 249                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 250                                 if (rc < 0) {
 251                                         CERROR("Can't send bio: %d\n", rc);
 252                                         /* OK do dec; we do the waiting */
 253                                         atomic_dec(&dreq->dr_numreqs);
 254                                         goto out;
 255                                 }
 256                                 rc = 0;
 257
 258                                 bio = NULL;
 259                         }
 260
 261                         /* allocate new bio */
 262                         bio = bio_alloc(GFP_NOIO,
 263                                         (npages - page_idx) * blocks_per_page);
 264                         if (bio == NULL) {
 265                                 CERROR ("Can't allocate bio\n");
 266                                 rc = -ENOMEM;
 267                                 goto out;
 268                         }
 269
 270                         bio->bi_bdev = inode->i_sb->s_bdev;
 271                         bio->bi_sector = sector;
 272                         bio->bi_end_io = dio_complete_routine;
 273                         bio->bi_private = dreq;
 274
 275                         rc = bio_add_page(bio, page,
 276                                           blocksize * nblocks, page_offset);
 277                         LASSERT (rc != 0);
 278                 }
 279         }
 280
 281         if (bio != NULL) {
 282                 atomic_inc(&dreq->dr_numreqs);
 283                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 284                 if (rc >= 0) {
 285                         rc = 0;
 286                 } else {
 287                         CERROR("Can't send bio: %d\n", rc);
 288                         /* OK do dec; we do the waiting */
 289                         atomic_dec(&dreq->dr_numreqs);
 290                 }
 291         }
 292
 293  out:
 294         wait_event(dreq->dr_wait, atomic_read(&dreq->dr_numreqs) == 0);
 295
 296         if (rc == 0)
 297                 rc = dreq->dr_error;
 298         RETURN(rc);
 299 }
 300
 301 static void filter_clear_page_cache(struct inode *inode,
 302                                     struct dio_request *iobuf)
 303 {
 304         struct page *page;
 305         int i;
 306
 307         for (i = 0; i < iobuf->dr_npages ; i++) {
 308                 page = find_lock_page(inode->i_mapping,
 309                                       iobuf->dr_pages[i]->index);
 310                 if (page == NULL)
 311                         continue;
 312                 if (page->mapping != NULL) {
 313                         wait_on_page_writeback(page);
 314                         block_invalidatepage(page, 0);
 315                         ll_truncate_complete_page(page);
 316                 }
 317                 unlock_page(page);
 318                 page_cache_release(page);
 319         }
 320 }
 321
 322 /* Must be called with i_sem taken for writes; this will drop it */
 323 int filter_direct_io(int rw, struct dentry *dchild, void *iobuf,
 324                      struct obd_export *exp, struct iattr *attr,
 325                      struct obd_trans_info *oti, void **wait_handle)
 326 {
 327         struct obd_device *obd = exp->exp_obd;
 328         struct inode *inode = dchild->d_inode;
 329         struct dio_request *dreq = iobuf;
 330         int rc, rc2;
 331         ENTRY;
 332
 333         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 334         LASSERTF(dreq->dr_npages <= dreq->dr_max_pages, "%d,%d\n",
 335                  dreq->dr_npages, dreq->dr_max_pages);
 336
 337         if (dreq->dr_npages == 0)
 338                 RETURN(0);
 339
 340         if (dreq->dr_npages > OBDFILTER_CREATED_SCRATCHPAD_ENTRIES)
 341                 RETURN(-EINVAL);
 342
 343         rc = fsfilt_map_inode_pages(obd, inode,
 344                                     dreq->dr_pages, dreq->dr_npages,
 345                                     dreq->dr_blocks,
 346                                     obdfilter_created_scratchpad,
 347                                     rw == OBD_BRW_WRITE, NULL);
 348
 349         if (rw == OBD_BRW_WRITE) {
 350                 if (rc == 0) {
 351 #if 0
 352                         filter_tally_write(&obd->u.filter,
 353                                            dreq->dr_pages,
 354                                            dreq->dr_page_idx,
 355                                            dreq->dr_blocks,
 356                                            blocks_per_page);
 357 #endif
 358                         if (attr->ia_size > inode->i_size)
 359                                 attr->ia_valid |= ATTR_SIZE;
 360                         rc = fsfilt_setattr(obd, dchild,
 361                                             oti->oti_handle, attr, 0);
 362                 }
 363
 364                 up(&inode->i_sem);
 365
 366                 rc2 = filter_finish_transno(exp, oti, 0);
 367                 if (rc2 != 0)
 368                         CERROR("can't close transaction: %d\n", rc);
 369
 370                 if (rc == 0)
 371                         rc = rc2;
 372                 if (rc != 0)
 373                         RETURN(rc);
 374         }
 375
 376         /* This is nearly osync_inode, without the waiting
 377         rc = generic_osync_inode(inode, inode->i_mapping,
 378                                  OSYNC_DATA|OSYNC_METADATA); */
 379         rc = filemap_fdatawrite(inode->i_mapping);
 380         rc2 = sync_mapping_buffers(inode->i_mapping);
 381         if (rc == 0)
 382                 rc = rc2;
 383         rc2 = filemap_fdatawait(inode->i_mapping);
 384         if (rc == 0)
 385                 rc = rc2;
 386
 387         if (rc != 0)
 388                 RETURN(rc);
 389
 390         /* be careful to call this after fsync_inode_data_buffers has waited
 391          * for IO to complete before we evict it from the cache */
 392         filter_clear_page_cache(inode, dreq);
 393
 394         RETURN(filter_do_bio(obd, inode, dreq, rw));
 395 }
 396
 397 /* See if there are unallocated parts in given file region */
 398 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 399 {
 400         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 401                 inode->i_mapping->a_ops->bmap;
 402         int j;
 403
 404         /* We can't know if we are overwriting or not */
 405         if (fs_bmap == NULL)
 406                 return 0;
 407
 408         offset >>= inode->i_blkbits;
 409         len >>= inode->i_blkbits;
 410
 411         for (j = 0; j <= len; j++)
 412                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 413                         return 0;
 414
 415         return 1;
 416 }
 417
 418 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 419                           int objcount, struct obd_ioobj *obj, int niocount,
 420                           struct niobuf_local *res, struct obd_trans_info *oti,
 421                           int rc)
 422 {
 423         struct niobuf_local *lnb;
 424         struct dio_request *dreq = NULL;
 425         struct lvfs_run_ctxt saved;
 426         struct fsfilt_objinfo fso;
 427         struct iattr iattr = { 0 };
 428         struct inode *inode = NULL;
 429         unsigned long now = jiffies;
 430         int i, err, cleanup_phase = 0;
 431         struct obd_device *obd = exp->exp_obd;
 432         int   total_size = 0;
 433         ENTRY;
 434
 435         LASSERT(oti != NULL);
 436         LASSERT(objcount == 1);
 437         LASSERT(current->journal_info == NULL);
 438
 439         if (rc != 0)
 440                 GOTO(cleanup, rc);
 441
 442         rc = filter_alloc_iobuf(OBD_BRW_WRITE, obj->ioo_bufcnt, (void **)&dreq);
 443         if (rc)
 444                 GOTO(cleanup, rc);
 445         cleanup_phase = 1;
 446
 447         fso.fso_dentry = res->dentry;
 448         fso.fso_bufcnt = obj->ioo_bufcnt;
 449         inode = res->dentry->d_inode;
 450
 451         for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
 452                 loff_t this_size;
 453
 454                 /* If overwriting an existing block, we don't need a grant */
 455                 if (!(lnb->flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC &&
 456                     filter_range_is_mapped(inode, lnb->offset, lnb->len))
 457                         lnb->rc = 0;
 458
 459                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 460                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 461                         continue;
 462                 }
 463
 464                 err = filter_iobuf_add_page(obd, dreq, inode, lnb->page);
 465                 LASSERT (err == 0);
 466
 467                 total_size += lnb->len;
 468
 469                 /* we expect these pages to be in offset order, but we'll
 470                  * be forgiving */
 471                 this_size = lnb->offset + lnb->len;
 472                 if (this_size > iattr.ia_size)
 473                         iattr.ia_size = this_size;
 474         }
 475 #if 0
 476         /* I use this when I'm checking our lovely 1M I/Os reach the disk -eeb */
 477         if (total_size != (1<<20))
 478                 CWARN("total size %d (%d pages)\n",
 479                       total_size, total_size/PAGE_SIZE);
 480 #endif
 481         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 482         cleanup_phase = 2;
 483
 484         down(&inode->i_sem);
 485         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 486                                            oti);
 487         if (IS_ERR(oti->oti_handle)) {
 488                 up(&inode->i_sem);
 489                 rc = PTR_ERR(oti->oti_handle);
 490                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 491                        "error starting transaction: rc = %d\n", rc);
 492                 oti->oti_handle = NULL;
 493                 GOTO(cleanup, rc);
 494         }
 495         /* have to call fsfilt_commit() from this point on */
 496
 497         fsfilt_check_slow(now, obd_timeout, "brw_start");
 498
 499         iattr_from_obdo(&iattr,oa,OBD_MD_FLATIME|OBD_MD_FLMTIME|OBD_MD_FLCTIME);
 500         /* filter_direct_io drops i_sem */
 501         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, dreq, exp, &iattr,
 502                               oti, NULL);
 503         if (rc == 0)
 504                 obdo_from_inode(oa, inode, FILTER_VALID_FLAGS);
 505
 506         fsfilt_check_slow(now, obd_timeout, "direct_io");
 507
 508         err = fsfilt_commit(obd, obd->u.filter.fo_sb, inode, oti->oti_handle,
 509                             obd_sync_filter);
 510         if (err)
 511                 rc = err;
 512
 513         if (obd_sync_filter && !err)
 514                 LASSERT(oti->oti_transno <= obd->obd_last_committed);
 515
 516         fsfilt_check_slow(now, obd_timeout, "commitrw commit");
 517
 518 cleanup:
 519         filter_grant_commit(exp, niocount, res);
 520
 521         switch (cleanup_phase) {
 522         case 2:
 523                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 524                 LASSERT(current->journal_info == NULL);
 525         case 1:
 526                 filter_free_iobuf(dreq);
 527         case 0:
 528                 filter_free_dio_pages(objcount, obj, niocount, res);
 529                 f_dput(res->dentry);
 530         }
 531
 532         RETURN(rc);
 533 }