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