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 = brw_kiovec(WRITE, 1, &iobuf, inode->i_dev, iobuf->blocks,
 159                         1 << inode->i_blkbits);
 160         CDEBUG(D_INFO, "tried to write %d pages, rc = %d\n",
 161                iobuf->nr_pages, rc);
 162         if (rc != (1 << inode->i_blkbits) * iobuf->nr_pages * blocks_per_page)
 163                 CERROR("short write?  expected %d, wrote %d\n",
 164                        (1 << inode->i_blkbits) * iobuf->nr_pages *
 165                        blocks_per_page, rc);
 166         if (rc > 0)
 167                 rc = 0;
 168
 169         EXIT;
 170 cleanup:
 171         if (!committed) {
 172                 int err = fsfilt_commit_async(obd, inode,
 173                                               oti->oti_handle, wait_handle);
 174                 oti->oti_handle = NULL;
 175                 if (err)
 176                         CERROR("can't close transaction: %d\n", err);
 177                 /*
 178                  * this is error path, so we prefer to return
 179                  * original error, not this one
 180                  */
 181         }
 182
 183         switch(cleanup_phase) {
 184         case 3:
 185         case 2:
 186                 unlock_kiovec(1, &iobuf);
 187         case 1:
 188                 OBD_FREE(created, sizeof(*created) *
 189                          iobuf->nr_pages*blocks_per_page);
 190         case 0:
 191                 if (cleanup_phase == 3)
 192                         break;
 193                 up(&inode->i_sem);
 194                 break;
 195         default:
 196                 CERROR("corrupt cleanup_phase (%d)?\n", cleanup_phase);
 197                 LBUG();
 198                 break;
 199         }
 200         return rc;
 201 }
 202
 203 /* See if there are unallocated parts in given file region */
 204 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
 205 {
 206         int (*fs_bmap)(struct address_space *, long) =
 207                 inode->i_mapping->a_ops->bmap;
 208         int j;
 209
 210         /* We can't know if the range is mapped already or not */
 211         if (fs_bmap == NULL)
 212                 return 0;
 213
 214         offset >>= inode->i_blkbits;
 215         len >>= inode->i_blkbits;
 216
 217         for (j = 0; j <= len; j++)
 218                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
 219                         return 0;
 220
 221         return 1;
 222 }
 223
 224 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa, int objcount,
 225                           struct obd_ioobj *obj, int niocount,
 226                           struct niobuf_local *res, struct obd_trans_info *oti,
 227                           int rc)
 228 {
 229         struct obd_device *obd = exp->exp_obd;
 230         struct obd_run_ctxt saved;
 231         struct niobuf_local *lnb;
 232         struct fsfilt_objinfo fso;
 233         struct iattr iattr = { 0 };
 234         struct kiobuf *iobuf;
 235         struct inode *inode = NULL;
 236         int i, n, cleanup_phase = 0, err;
 237         unsigned long now = jiffies; /* DEBUGGING OST TIMEOUTS */
 238         void *wait_handle;
 239         ENTRY;
 240         LASSERT(oti != NULL);
 241         LASSERT(objcount == 1);
 242         LASSERT(current->journal_info == NULL);
 243
 244         if (rc != 0)
 245                 GOTO(cleanup, rc);
 246
 247         rc = alloc_kiovec(1, &iobuf);
 248         if (rc)
 249                 GOTO(cleanup, rc);
 250         cleanup_phase = 1;
 251
 252 #ifdef HAVE_KIOBUF_DOVARY
 253         iobuf->dovary = 0; /* this prevents corruption, not present in 2.4.20 */
 254 #endif
 255         rc = expand_kiobuf(iobuf, obj->ioo_bufcnt);
 256         if (rc)
 257                 GOTO(cleanup, rc);
 258
 259         iobuf->offset = 0;
 260         iobuf->length = 0;
 261         iobuf->nr_pages = 0;
 262
 263         cleanup_phase = 1;
 264         fso.fso_dentry = res->dentry;
 265         fso.fso_bufcnt = obj->ioo_bufcnt;
 266         inode = res->dentry->d_inode;
 267
 268         for (i = 0, lnb = res, n = 0; i < obj->ioo_bufcnt; i++, lnb++) {
 269                 loff_t this_size;
 270
 271                 /* If overwriting an existing block, we don't need a grant */
 272                 if (!(lnb->flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC &&
 273                     filter_range_is_mapped(inode, lnb->offset, lnb->len))
 274                         lnb->rc = 0;
 275
 276                 if (lnb->rc) /* ENOSPC, network RPC error */
 277                         continue;
 278
 279                 iobuf->maplist[n++] = lnb->page;
 280                 iobuf->length += PAGE_SIZE;
 281                 iobuf->nr_pages++;
 282
 283                 /* We expect these pages to be in offset order, but we'll
 284                  * be forgiving */
 285                 this_size = lnb->offset + lnb->len;
 286                 if (this_size > iattr.ia_size)
 287                         iattr.ia_size = this_size;
 288         }
 289
 290         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 291         cleanup_phase = 2;
 292
 293         down(&inode->i_sem);
 294         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
 295                                            oti);
 296         if (IS_ERR(oti->oti_handle)) {
 297                 rc = PTR_ERR(oti->oti_handle);
 298                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 299                        "error starting transaction: rc = %d\n", rc);
 300                 oti->oti_handle = NULL;
 301                 GOTO(cleanup, rc);
 302         }
 303
 304         if (time_after(jiffies, now + 15 * HZ))
 305                 CERROR("slow brw_start %lus\n", (jiffies - now) / HZ);
 306
 307         iattr_from_obdo(&iattr,oa,OBD_MD_FLATIME|OBD_MD_FLMTIME|OBD_MD_FLCTIME);
 308         rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
 309                               oti, &wait_handle);
 310         if (rc == 0)
 311                 obdo_from_inode(oa, inode, FILTER_VALID_FLAGS);
 312
 313         if (time_after(jiffies, now + 15 * HZ))
 314                 CERROR("slow direct_io %lus\n", (jiffies - now) / HZ);
 315
 316         err = fsfilt_commit_wait(obd, inode, wait_handle);
 317         if (err)
 318                 rc = err;
 319         if (obd_sync_filter)
 320                 LASSERT(oti->oti_transno <= obd->obd_last_committed);
 321         if (time_after(jiffies, now + 15 * HZ))
 322                 CERROR("slow commitrw commit %lus\n", (jiffies - now) / HZ);
 323
 324 cleanup:
 325         filter_grant_commit(exp, niocount, res);
 326
 327         switch (cleanup_phase) {
 328         case 2:
 329                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 330                 LASSERT(current->journal_info == NULL);
 331         case 1:
 332                 free_kiovec(1, &iobuf);
 333         case 0:
 334                 for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
 335                         /* flip_.. gets a ref, while free_page only frees
 336                          * when it decrefs to 0 */
 337                         if (rc == 0)
 338                                 flip_into_page_cache(inode, lnb->page);
 339                         __free_page(lnb->page);
 340                 }
 341                 f_dput(res->dentry);
 342         }
 343
 344         RETURN(rc);
 345 }