lustre/obdfilter/filter_io_24.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
  32 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  33
  34 #define DEBUG_SUBSYSTEM S_FILTER
  35
  36 #include <linux/iobuf.h>
  37 #include <linux/locks.h>
  38
  39 #include <linux/obd_class.h>
  40 #include <linux/lustre_fsfilt.h>
  41 #include "filter_internal.h"
  42
  43
  44 /* We should only change the file mtime (and not the ctime, like
  45  * update_inode_times() in generic_file_write()) when we only change data. */
  46 void inode_update_time(struct inode *inode, int ctime_too)
  47 {
  48         time_t now = CURRENT_TIME;
  49         if (inode->i_mtime == now && (!ctime_too || inode->i_ctime == now))
  50                 return;
  51         inode->i_mtime = now;
  52         if (ctime_too)
  53                 inode->i_ctime = now;
  54         mark_inode_dirty_sync(inode);
  55 }
  56
  57 /* Bug 2254 -- this is better done in ext3_map_inode_page, but this
  58  * workaround will suffice until everyone has upgraded their kernels */
  59 static void check_pending_bhs(unsigned long *blocks, int nr_pages, dev_t dev,
  60                               int size)
  61 {
  62 #if (LUSTRE_KERNEL_VERSION < 32)
  63         struct buffer_head *bh;
  64         int i;
  65
  66         for (i = 0; i < nr_pages; i++) {
  67                 bh = get_hash_table(dev, blocks[i], size);
  68                 if (bh == NULL)
  69                         continue;
  70                 if (!buffer_dirty(bh)) {
  71                         put_bh(bh);
  72                         continue;
  73                 }
  74                 mark_buffer_clean(bh);
  75                 wait_on_buffer(bh);
  76                 clear_bit(BH_Req, &bh->b_state);
  77                 __brelse(bh);
  78         }
  79 #endif
  80 }
  81
  82 /* Must be called with i_sem taken; this will drop it */
  83 static int filter_direct_io(int rw, struct dentry *dchild, struct kiobuf *iobuf,
  84                             struct obd_export *exp, struct iattr *attr,
  85                             struct obd_trans_info *oti, void **wait_handle)
  86 {
  87         struct obd_device *obd = exp->exp_obd;
  88         struct inode *inode = dchild->d_inode;
  89         struct page *page;
  90         unsigned long *b = iobuf->blocks;
  91         int rc, i, create = (rw == OBD_BRW_WRITE), blocks_per_page;
  92         int *cr, cleanup_phase = 0, *created = NULL;
  93         int committed = 0;
  94         ENTRY;
  95
  96         blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
  97         if (iobuf->nr_pages * blocks_per_page > KIO_MAX_SECTORS)
  98                 GOTO(cleanup, rc = -EINVAL);
  99
 100         OBD_ALLOC(created, sizeof(*created) * iobuf->nr_pages*blocks_per_page);
 101         if (created == NULL)
 102                 GOTO(cleanup, rc = -ENOMEM);
 103         cleanup_phase = 1;
 104
 105         rc = lock_kiovec(1, &iobuf, 1);
 106         if (rc < 0)
 107                 GOTO(cleanup, rc);
 108         cleanup_phase = 2;
 109
 110         down(&exp->exp_obd->u.filter.fo_alloc_lock);
 111         for (i = 0, cr = created, b = iobuf->blocks; i < iobuf->nr_pages; i++){
 112                 page = iobuf->maplist[i];
 113
 114                 rc = fsfilt_map_inode_page(obd, inode, page, b, cr, create);
 115                 if (rc) {
 116                         CERROR("ino %lu, blk %lu cr %u create %d: rc %d\n",
 117                                inode->i_ino, *b, *cr, create, rc);
 118                         up(&exp->exp_obd->u.filter.fo_alloc_lock);
 119                         GOTO(cleanup, rc);
 120                 }
 121
 122                 b += blocks_per_page;
 123                 cr += blocks_per_page;
 124         }
 125         up(&exp->exp_obd->u.filter.fo_alloc_lock);
 126
 127         filter_tally_write(&obd->u.filter, iobuf->maplist, iobuf->nr_pages,
 128                            iobuf->blocks, blocks_per_page);
 129
 130         if (attr->ia_size > inode->i_size)
 131                 attr->ia_valid |= ATTR_SIZE;
 132         rc = fsfilt_setattr(obd, dchild, oti->oti_handle, attr, 0);
 133         if (rc)
 134                 GOTO(cleanup, rc);
 135
 136         up(&inode->i_sem);
 137         cleanup_phase = 3;
 138
 139         rc = filter_finish_transno(exp, oti, 0);
 140         if (rc)
 141                 GOTO(cleanup, rc);
 142
 143         rc = fsfilt_commit_async(obd, inode, oti->oti_handle, wait_handle);
 144         oti->oti_handle = NULL;
 145         committed = 1;
 146         if (rc)
 147                 GOTO(cleanup, rc);
 148
 149         check_pending_bhs(iobuf->blocks, iobuf->nr_pages, inode->i_dev,
 150                           1 << inode->i_blkbits);
 151
 152         rc = filemap_fdatasync(inode->i_mapping);
 153         if (rc == 0)
 154                 rc = fsync_inode_data_buffers(inode);
 155         if (rc == 0)
 156                 rc = filemap_fdatawait(inode->i_mapping);
 157         if (rc < 0)
 158                 GOTO(cleanup, rc);
 159
 160         rc = brw_kiovec(WRITE, 1, &iobuf, inode->i_dev, iobuf->blocks,
 161                         1 << inode->i_blkbits);
 162         CDEBUG(D_INFO, "tried to write %d pages, rc = %d\n",
 163                iobuf->nr_pages, rc);
 164         if (rc != (1 << inode->i_blkbits) * iobuf->nr_pages * blocks_per_page)
 165                 CERROR("short write?  expected %d, wrote %d\n",
 166                        (1 << inode->i_blkbits) * iobuf->nr_pages *
 167                        blocks_per_page, rc);
 168         if (rc > 0)
 169                 rc = 0;
 170
 171         EXIT;
 172 cleanup:
 173         if (!committed) {
 174                 int err = fsfilt_commit_async(obd, inode,
 175                                               oti->oti_handle, wait_handle);
 176                 oti->oti_handle = NULL;
 177                 if (err)
 178                         CERROR("can't close transaction: %d\n", err);
 179                 /*
 180                  * this is error path, so we prefer to return
 181                  * original error, not this one
 182                  */
 183         }
 184
 185         switch(cleanup_phase) {
 186         case 3:
 187         case 2:
 188                 unlock_kiovec(1, &iobuf);
 189         case 1:
 190                 OBD_FREE(created, sizeof(*created) *
 191                          iobuf->nr_pages*blocks_per_page);
 192         case 0:
 193                 if (cleanup_phase == 3)
 194                         break;
 195                 up(&inode->i_sem);
 196                 break;
 197         default:
 198                 CERROR("corrupt cleanup_phase (%d)?\n", cleanup_phase);
 199                 LBUG();
 200                 break;
 201         }
 202         return rc;
 203 }
 204
 205 /* See if there are unallocated parts in given file region */
 206 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 207 {
 208         int (*fs_bmap)(struct address_space *, long) =
 209                 inode->i_mapping->a_ops->bmap;
 210         int j;
 211
 212         /* We can't know if the range is mapped already or not */
 213         if (fs_bmap == NULL)
 214                 return 0;
 215
 216         offset >>= inode->i_blkbits;
 217         len >>= inode->i_blkbits;
 218
 219         for (j = 0; j <= len; j++)
 220                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 221                         return 0;
 222
 223         return 1;
 224 }
 225
 226 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa, int objcount,
 227                           struct obd_ioobj *obj, int niocount,
 228                           struct niobuf_local *res, struct obd_trans_info *oti)
 229 {
 230         struct obd_device *obd = exp->exp_obd;
 231         struct obd_run_ctxt saved;
 232         struct niobuf_local *lnb;
 233         struct fsfilt_objinfo fso;
 234         struct iattr iattr = { 0 };
 235         struct kiobuf *iobuf;
 236         struct inode *inode = NULL;
 237         int rc = 0, i, n, cleanup_phase = 0, err;
 238         unsigned long now = jiffies; /* DEBUGGING OST TIMEOUTS */
 239         void *wait_handle;
 240         ENTRY;
 241         LASSERT(oti != NULL);
 242         LASSERT(objcount == 1);
 243         LASSERT(current->journal_info == NULL);
 244
 245         rc = alloc_kiovec(1, &iobuf);
 246         if (rc)
 247                 GOTO(cleanup, rc);
 248         cleanup_phase = 1;
 249
 250 #if (LINUX_VERSION_CODE == KERNEL_VERSION(2,4,18))
 251         iobuf->dovary = 0; /* this prevents corruption, not present in 2.4.20 */
 252 #endif
 253         rc = expand_kiobuf(iobuf, obj->ioo_bufcnt);
 254         if (rc)
 255                 GOTO(cleanup, rc);
 256
 257         iobuf->offset = 0;
 258         iobuf->length = 0;
 259         iobuf->nr_pages = 0;
 260
 261         cleanup_phase = 1;
 262         fso.fso_dentry = res->dentry;
 263         fso.fso_bufcnt = obj->ioo_bufcnt;
 264         inode = res->dentry->d_inode;
 265
 266         for (i = 0, lnb = res, n = 0; i < obj->ioo_bufcnt; i++, lnb++) {
 267                 loff_t this_size;
 268
 269                 /* If overwriting an existing block, we don't need a grant */
 270                 if (!(lnb->flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC &&
 271                     filter_range_is_mapped(inode, lnb->offset, lnb->len))
 272                         lnb->rc = 0;
 273
 274                 if (lnb->rc) /* ENOSPC, network RPC error */
 275                         continue;
 276
 277                 iobuf->maplist[n++] = lnb->page;
 278                 iobuf->length += PAGE_SIZE;
 279                 iobuf->nr_pages++;
 280
 281                 /* We expect these pages to be in offset order, but we'll
 282                  * be forgiving */
 283                 this_size = lnb->offset + lnb->len;
 284                 if (this_size > iattr.ia_size)
 285                         iattr.ia_size = this_size;
 286         }
 287
 288         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 289         cleanup_phase = 2;
 290
 291         down(&inode->i_sem);
 292         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 293                                            oti);
 294         if (IS_ERR(oti->oti_handle)) {
 295                 rc = PTR_ERR(oti->oti_handle);
 296                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 297                        "error starting transaction: rc = %d\n", rc);
 298                 oti->oti_handle = NULL;
 299                 GOTO(cleanup, rc);
 300         }
 301
 302         if (time_after(jiffies, now + 15 * HZ))
 303                 CERROR("slow brw_start %lus\n", (jiffies - now) / HZ);
 304
 305         iattr_from_obdo(&iattr,oa,OBD_MD_FLATIME|OBD_MD_FLMTIME|OBD_MD_FLCTIME);
 306         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 307                               oti, &wait_handle);
 308         if (rc == 0)
 309                 obdo_from_inode(oa, inode, FILTER_VALID_FLAGS);
 310
 311         if (time_after(jiffies, now + 15 * HZ))
 312                 CERROR("slow direct_io %lus\n", (jiffies - now) / HZ);
 313
 314         filter_grant_commit(exp, niocount, res);
 315         err = fsfilt_commit_wait(obd, inode, wait_handle);
 316         if (err)
 317                 rc = err;
 318         if (obd_sync_filter)
 319                 LASSERT(oti->oti_transno <= obd->obd_last_committed);
 320         if (time_after(jiffies, now + 15 * HZ))
 321                 CERROR("slow commitrw commit %lus\n", (jiffies - now) / HZ);
 322
 323 cleanup:
 324         switch (cleanup_phase) {
 325         case 2:
 326                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 327                 LASSERT(current->journal_info == NULL);
 328         case 1:
 329                 free_kiovec(1, &iobuf);
 330         case 0:
 331                 for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
 332                         /* flip_.. gets a ref, while free_page only frees
 333                          * when it decrefs to 0 */
 334                         if (rc == 0)
 335                                 flip_into_page_cache(inode, lnb->page);
 336                         __free_page(lnb->page);
 337                 }
 338                 f_dput(res->dentry);
 339         }
 340
 341         RETURN(rc);
 342 }
 343
 344 #endif
 345