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 /* 512byte block min */
  40 #define MAX_BLOCKS_PER_PAGE (PAGE_SIZE / 512)
  41 struct dio_request {
  42         atomic_t          dr_numreqs;  /* number of reqs being processed */
  43         struct bio       *dr_bios;     /* list of completed bios */
  44         wait_queue_head_t dr_wait;
  45         int               dr_max_pages;
  46         int               dr_npages;
  47         int               dr_error;
  48         struct page     **dr_pages;
  49         unsigned long    *dr_blocks;
  50         spinlock_t        dr_lock;
  51 };
  52
  53 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
  54 {
  55         struct dio_request *dreq = bio->bi_private;
  56         unsigned long flags;
  57
  58         if (bio->bi_size) {
  59                 CWARN("gets called against non-complete bio 0x%p: %d/%d/%d\n",
  60                       bio, bio->bi_size, done, error);
  61                 return 1;
  62         }
  63
  64         if (dreq == NULL) {
  65                 CERROR("***** bio->bi_private is NULL!  This should never "
  66                        "happen.  Normally, I would crash here, but instead I "
  67                        "will dump the bio contents to the console.  Please "
  68                        "report this to CFS, along with any interesting messages "
  69                        "leading up to this point (like SCSI errors, perhaps).  "
  70                        "Because bi_private is NULL, I can't wake up the thread "
  71                        "that initiated this I/O -- so you will probably have to "
  72                        "reboot this node.");
  73                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
  74                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
  75                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
  76                        bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
  77                        bio->bi_end_io, atomic_read(&bio->bi_cnt),
  78                        bio->bi_private);
  79                 return 0;
  80         }
  81
  82         spin_lock_irqsave(&dreq->dr_lock, flags);
  83         bio->bi_private = dreq->dr_bios;
  84         dreq->dr_bios = bio;
  85         if (dreq->dr_error == 0)
  86                 dreq->dr_error = error;
  87         spin_unlock_irqrestore(&dreq->dr_lock, flags);
  88
  89         if (atomic_dec_and_test(&dreq->dr_numreqs))
  90                 wake_up(&dreq->dr_wait);
  91
  92         return 0;
  93 }
  94
  95 static int can_be_merged(struct bio *bio, sector_t sector)
  96 {
  97         unsigned int size;
  98         if (!bio)
  99                 return 0;
 100
 101         size = bio->bi_size >> 9;
 102         return bio->bi_sector + size == sector ? 1 : 0;
 103 }
 104
 105
 106 int filter_alloc_iobuf(int rw, int num_pages, void **ret)
 107 {
 108         struct dio_request *dreq;
 109
 110         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 111
 112         OBD_ALLOC(dreq, sizeof(*dreq));
 113         if (dreq == NULL)
 114                 goto failed_0;
 115
 116         OBD_ALLOC(dreq->dr_pages, num_pages * sizeof(*dreq->dr_pages));
 117         if (dreq->dr_pages == NULL)
 118                 goto failed_1;
 119
 120         OBD_ALLOC(dreq->dr_blocks,
 121                   MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*dreq->dr_blocks));
 122         if (dreq->dr_blocks == NULL)
 123                 goto failed_2;
 124
 125         dreq->dr_bios = NULL;
 126         init_waitqueue_head(&dreq->dr_wait);
 127         atomic_set(&dreq->dr_numreqs, 0);
 128         spin_lock_init(&dreq->dr_lock);
 129         dreq->dr_max_pages = num_pages;
 130         dreq->dr_npages = 0;
 131
 132         *ret = dreq;
 133         RETURN(0);
 134
 135  failed_2:
 136         OBD_FREE(dreq->dr_pages,
 137                  num_pages * sizeof(*dreq->dr_pages));
 138  failed_1:
 139         OBD_FREE(dreq, sizeof(*dreq));
 140  failed_0:
 141         RETURN(-ENOMEM);
 142 }
 143
 144 void filter_free_iobuf(void *iobuf)
 145 {
 146         struct dio_request *dreq = iobuf;
 147         int                 num_pages = dreq->dr_max_pages;
 148
 149         /* free all bios */
 150         while (dreq->dr_bios) {
 151                 struct bio *bio = dreq->dr_bios;
 152                 dreq->dr_bios = bio->bi_private;
 153                 bio_put(bio);
 154         }
 155
 156         OBD_FREE(dreq->dr_blocks,
 157                  MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*dreq->dr_blocks));
 158         OBD_FREE(dreq->dr_pages,
 159                  num_pages * sizeof(*dreq->dr_pages));
 160         OBD_FREE(dreq, sizeof(*dreq));
 161 }
 162
 163 int filter_iobuf_add_page(struct obd_device *obd, void *iobuf,
 164                           struct inode *inode, struct page *page)
 165 {
 166         struct dio_request *dreq = iobuf;
 167
 168         LASSERT (dreq->dr_npages < dreq->dr_max_pages);
 169         dreq->dr_pages[dreq->dr_npages++] = page;
 170
 171         return 0;
 172 }
 173
 174 int filter_do_bio(struct obd_device *obd, struct inode *inode,
 175                   struct dio_request *dreq, int rw)
 176 {
 177         int            blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 178         struct page  **pages = dreq->dr_pages;
 179         int            npages = dreq->dr_npages;
 180         unsigned long *blocks = dreq->dr_blocks;
 181         int            total_blocks = npages * blocks_per_page;
 182         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
 183         unsigned int   blocksize = inode->i_sb->s_blocksize;
 184         struct bio    *bio = NULL;
 185         struct page   *page;
 186         unsigned int   page_offset;
 187         sector_t       sector;
 188         int            nblocks;
 189         int            block_idx;
 190         int            page_idx;
 191         int            i;
 192         int            rc = 0;
 193         ENTRY;
 194
 195         LASSERT(dreq->dr_npages == npages);
 196         LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES);
 197
 198         for (page_idx = 0, block_idx = 0;
 199              page_idx < npages;
 200              page_idx++, block_idx += blocks_per_page) {
 201
 202                 page = pages[page_idx];
 203                 LASSERT (block_idx + blocks_per_page <= total_blocks);
 204
 205                 for (i = 0, page_offset = 0;
 206                      i < blocks_per_page;
 207                      i += nblocks, page_offset += blocksize * nblocks) {
 208
 209                         nblocks = 1;
 210
 211                         if (blocks[block_idx + i] == 0) {  /* hole */
 212                                 LASSERT(rw == OBD_BRW_READ);
 213                                 memset(kmap(page) + page_offset, 0, blocksize);
 214                                 kunmap(page);
 215                                 continue;
 216                         }
 217
 218                         sector = blocks[block_idx + i] << sector_bits;
 219
 220                         /* Additional contiguous file blocks? */
 221                         while (i + nblocks < blocks_per_page &&
 222                                (sector + nblocks*(blocksize>>9)) ==
 223                                (blocks[block_idx + i + nblocks] << sector_bits))
 224                                 nblocks++;
 225
 226                         if (bio != NULL &&
 227                             can_be_merged(bio, sector) &&
 228                             bio_add_page(bio, page,
 229                                          blocksize * nblocks, page_offset) != 0)
 230                                 continue;       /* added this frag OK */
 231
 232                         if (bio != NULL) {
 233                                 request_queue_t *q = bdev_get_queue(bio->bi_bdev);
 234
 235                                 /* Dang! I have to fragment this I/O */
 236                                 CDEBUG(D_INODE|D_ERROR, "bio++ sz %d vcnt %d(%d) "
 237                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 238                                        bio->bi_size,
 239                                        bio->bi_vcnt, bio->bi_max_vecs,
 240                                        bio->bi_size >> 9, q->max_sectors,
 241                                        bio_phys_segments(q, bio),
 242                                        q->max_phys_segments,
 243                                        bio_hw_segments(q, bio),
 244                                        q->max_hw_segments);
 245
 246                                 atomic_inc(&dreq->dr_numreqs);
 247                                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 248                                 if (rc < 0) {
 249                                         CERROR("Can't send bio: %d\n", rc);
 250                                         /* OK do dec; we do the waiting */
 251                                         atomic_dec(&dreq->dr_numreqs);
 252                                         goto out;
 253                                 }
 254                                 rc = 0;
 255
 256                                 bio = NULL;
 257                         }
 258
 259                         /* allocate new bio */
 260                         bio = bio_alloc(GFP_NOIO,
 261                                         (npages - page_idx) * blocks_per_page);
 262                         if (bio == NULL) {
 263                                 CERROR ("Can't allocate bio\n");
 264                                 rc = -ENOMEM;
 265                                 goto out;
 266                         }
 267
 268                         bio->bi_bdev = inode->i_sb->s_bdev;
 269                         bio->bi_sector = sector;
 270                         bio->bi_end_io = dio_complete_routine;
 271                         bio->bi_private = dreq;
 272
 273                         rc = bio_add_page(bio, page,
 274                                           blocksize * nblocks, page_offset);
 275                         LASSERT (rc != 0);
 276                 }
 277         }
 278
 279         if (bio != NULL) {
 280                 atomic_inc(&dreq->dr_numreqs);
 281                 rc = fsfilt_send_bio(rw, obd, inode, bio);
 282                 if (rc >= 0) {
 283                         rc = 0;
 284                 } else {
 285                         CERROR("Can't send bio: %d\n", rc);
 286                         /* OK do dec; we do the waiting */
 287                         atomic_dec(&dreq->dr_numreqs);
 288                 }
 289         }
 290
 291  out:
 292         wait_event(dreq->dr_wait, atomic_read(&dreq->dr_numreqs) == 0);
 293
 294         if (rc == 0)
 295                 rc = dreq->dr_error;
 296         RETURN(rc);
 297 }
 298
 299 /* These are our hacks to keep our directio/bh IO coherent with ext3's
 300  * page cache use.  Most notably ext3 reads file data into the page
 301  * cache when it is zeroing the tail of partial-block truncates and
 302  * leaves it there, sometimes generating io from it at later truncates.
 303  * This removes the partial page and its buffers from the page cache,
 304  * so it should only ever cause a wait in rare cases, as otherwise we
 305  * always do full-page IO to the OST.
 306  *
 307  * The call to truncate_complete_page() will call journal_invalidatepage()
 308  * to free the buffers and drop the page from cache.  The buffers should
 309  * not be dirty, because we already called fdatasync/fdatawait on them.
 310  */
 311 static int filter_clear_page_cache(struct inode *inode,
 312                                    struct dio_request *iobuf)
 313 {
 314         struct page *page;
 315         int i, rc, rc2;
 316
 317         /* This is nearly generic_osync_inode, without the waiting on the inode
 318         rc = generic_osync_inode(inode, inode->i_mapping,
 319                                   OSYNC_DATA|OSYNC_METADATA);
 320         */
 321         rc = filemap_fdatawrite(inode->i_mapping);
 322         rc2 = sync_mapping_buffers(inode->i_mapping);
 323         if (rc == 0)
 324                 rc = rc2;
 325         rc2 = filemap_fdatawait(inode->i_mapping);
 326         if (rc == 0)
 327                 rc = rc2;
 328         if (rc != 0)
 329                 RETURN(rc);
 330
 331         /* be careful to call this after fsync_inode_data_buffers has waited
 332          * for IO to complete before we evict it from the cache */
 333         for (i = 0; i < iobuf->dr_npages; i++) {
 334                 page = find_lock_page(inode->i_mapping,
 335                                        iobuf->dr_pages[i]->index);
 336                 if (page == NULL)
 337                        continue;
 338                 if (page->mapping != NULL) {
 339                        wait_on_page_writeback(page);
 340                        ll_truncate_complete_page(page);
 341                 }
 342
 343                 unlock_page(page);
 344                 page_cache_release(page);
 345         }
 346         return 0;
 347 }
 348 /* Must be called with i_sem taken for writes; this will drop it */
 349 int filter_direct_io(int rw, struct dentry *dchild, void *iobuf,
 350                      struct obd_export *exp, struct iattr *attr,
 351                      struct obd_trans_info *oti, void **wait_handle)
 352 {
 353         struct obd_device *obd = exp->exp_obd;
 354         struct inode *inode = dchild->d_inode;
 355         struct dio_request *dreq = iobuf;
 356         struct semaphore *sem = NULL;
 357         int rc, rc2, create = 0;
 358         ENTRY;
 359
 360         LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw);
 361         LASSERTF(dreq->dr_npages <= dreq->dr_max_pages, "%d,%d\n",
 362                  dreq->dr_npages, dreq->dr_max_pages);
 363
 364         if (dreq->dr_npages == 0)
 365                 RETURN(0);
 366
 367         if (dreq->dr_npages > OBDFILTER_CREATED_SCRATCHPAD_ENTRIES)
 368                 RETURN(-EINVAL);
 369
 370         if (rw == OBD_BRW_WRITE) {
 371                 create = 1;
 372                 //sem = &obd->u.filter.fo_alloc_lock;
 373         }
 374
 375         rc = fsfilt_map_inode_pages(obd, inode,
 376                                     dreq->dr_pages, dreq->dr_npages,
 377                                     dreq->dr_blocks,
 378                                     obdfilter_created_scratchpad,
 379                                     create, sem);
 380
 381         if (rw == OBD_BRW_WRITE) {
 382                 if (rc == 0) {
 383                         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 384                         filter_tally_write(&obd->u.filter,  dreq->dr_pages,
 385                                            dreq->dr_npages, dreq->dr_blocks,
 386                                            blocks_per_page);
 387                         if (attr->ia_size > inode->i_size)
 388                                 attr->ia_valid |= ATTR_SIZE;
 389                         rc = fsfilt_setattr(obd, dchild,
 390                                             oti->oti_handle, attr, 0);
 391                 }
 392
 393                 up(&inode->i_sem);
 394
 395                 rc2 = filter_finish_transno(exp, oti, 0);
 396                 if (rc2 != 0)
 397                         CERROR("can't close transaction: %d\n", rc);
 398                 rc = (rc == 0) ? rc2 : rc;
 399
 400                 rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,wait_handle);
 401                 rc = (rc == 0) ? rc2 : rc;
 402
 403                 if (rc != 0)
 404                         RETURN(rc);
 405         }
 406
 407         rc = filter_clear_page_cache(inode, dreq);
 408         if (rc != 0)
 409                 RETURN(rc);
 410
 411         RETURN(filter_do_bio(obd, inode, dreq, rw));
 412 }
 413
 414 /* See if there are unallocated parts in given file region */
 415 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 416 {
 417         sector_t (*fs_bmap)(struct address_space *, sector_t) =
 418                 inode->i_mapping->a_ops->bmap;
 419         int j;
 420
 421         /* We can't know if we are overwriting or not */
 422         if (fs_bmap == NULL)
 423                 return 0;
 424
 425         offset >>= inode->i_blkbits;
 426         len >>= inode->i_blkbits;
 427
 428         for (j = 0; j <= len; j++)
 429                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 430                         return 0;
 431
 432         return 1;
 433 }
 434
 435 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
 436                           int objcount, struct obd_ioobj *obj, int niocount,
 437                           struct niobuf_local *res, struct obd_trans_info *oti,
 438                           int rc)
 439 {
 440         struct niobuf_local *lnb;
 441         struct dio_request *dreq = NULL;
 442         struct lvfs_run_ctxt saved;
 443         struct fsfilt_objinfo fso;
 444         struct iattr iattr = { 0 };
 445         struct inode *inode = NULL;
 446         unsigned long now = jiffies;
 447         int i, err, cleanup_phase = 0;
 448         struct obd_device *obd = exp->exp_obd;
 449         void *wait_handle = NULL;
 450         int   total_size = 0;
 451         loff_t old_size;
 452         ENTRY;
 453
 454         LASSERT(oti != NULL);
 455         LASSERT(objcount == 1);
 456         LASSERT(current->journal_info == NULL);
 457
 458         if (rc != 0)
 459                 GOTO(cleanup, rc);
 460
 461         rc = filter_alloc_iobuf(OBD_BRW_WRITE, obj->ioo_bufcnt, (void **)&dreq);
 462         if (rc)
 463                 GOTO(cleanup, rc);
 464         cleanup_phase = 1;
 465
 466         fso.fso_dentry = res->dentry;
 467         fso.fso_bufcnt = obj->ioo_bufcnt;
 468         inode = res->dentry->d_inode;
 469
 470         for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
 471                 loff_t this_size;
 472
 473                 /* If overwriting an existing block, we don't need a grant */
 474                 if (!(lnb->flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC &&
 475                     filter_range_is_mapped(inode, lnb->offset, lnb->len))
 476                         lnb->rc = 0;
 477
 478                 if (lnb->rc) { /* ENOSPC, network RPC error, etc. */
 479                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc);
 480                         continue;
 481                 }
 482
 483                 err = filter_iobuf_add_page(obd, dreq, inode, lnb->page);
 484                 LASSERT (err == 0);
 485
 486                 total_size += lnb->len;
 487
 488                 /* we expect these pages to be in offset order, but we'll
 489                  * be forgiving */
 490                 this_size = lnb->offset + lnb->len;
 491                 if (this_size > iattr.ia_size)
 492                         iattr.ia_size = this_size;
 493         }
 494 #if 0
 495         /* I use this when I'm checking our lovely 1M I/Os reach the disk -eeb */
 496         if (total_size != (1<<20))
 497                 CWARN("total size %d (%d pages)\n",
 498                       total_size, total_size/PAGE_SIZE);
 499 #endif
 500         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 501         cleanup_phase = 2;
 502
 503         down(&inode->i_sem);
 504         old_size = inode->i_size;
 505         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 506                                            oti);
 507         if (IS_ERR(oti->oti_handle)) {
 508                 up(&inode->i_sem);
 509                 rc = PTR_ERR(oti->oti_handle);
 510                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 511                        "error starting transaction: rc = %d\n", rc);
 512                 oti->oti_handle = NULL;
 513                 GOTO(cleanup, rc);
 514         }
 515         /* have to call fsfilt_commit() from this point on */
 516
 517         fsfilt_check_slow(now, obd_timeout, "brw_start");
 518
 519         iattr_from_obdo(&iattr,oa,OBD_MD_FLATIME|OBD_MD_FLMTIME|OBD_MD_FLCTIME);
 520         /* filter_direct_io drops i_sem */
 521         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, dreq, exp, &iattr,
 522                               oti, &wait_handle);
 523
 524 #if 0
 525         if (inode->i_size != old_size) {
 526                 struct llog_cookie *cookie = obdo_logcookie(oa);
 527                 struct lustre_id *id = obdo_id(oa);
 528                 filter_log_sz_change(obd, id, oa->o_easize, cookie, inode);
 529         }
 530 #endif
 531
 532         if (rc == 0)
 533                 obdo_from_inode(oa, inode, FILTER_VALID_FLAGS);
 534
 535         fsfilt_check_slow(now, obd_timeout, "direct_io");
 536
 537         err = fsfilt_commit_wait(obd, inode, wait_handle);
 538         if (rc == 0)
 539                 rc = err;
 540
 541         fsfilt_check_slow(now, obd_timeout, "commitrw commit");
 542
 543 cleanup:
 544         filter_grant_commit(exp, niocount, res);
 545
 546         switch (cleanup_phase) {
 547         case 2:
 548                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 549                 LASSERT(current->journal_info == NULL);
 550         case 1:
 551                 filter_free_iobuf(dreq);
 552         case 0:
 553                 filter_free_dio_pages(objcount, obj, niocount, res);
 554                 f_dput(res->dentry);
 555         }
 556
 557         RETURN(rc);
 558 }