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
  32 #define DEBUG_SUBSYSTEM S_FILTER
  33
  34 #include <linux/obd_class.h>
  35 #include <linux/lustre_fsfilt.h>
  36 #include "filter_internal.h"
  37
  38 #warning "implement writeback mode -bzzz"
  39
  40 /* 512byte block min */
  41 #define MAX_BLOCKS_PER_PAGE (PAGE_SIZE / 512)
  42 struct dio_request {
  43         atomic_t numreqs;       /* number of reqs being processed */
  44         struct bio *bio_list;   /* list of completed bios */
  45         wait_queue_head_t wait;
  46         int created[MAX_BLOCKS_PER_PAGE];
  47         unsigned long blocks[MAX_BLOCKS_PER_PAGE];
  48         spinlock_t lock;
  49 };
  50
  51 static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
  52 {
  53         struct dio_request *dreq = bio->bi_private;
  54         unsigned long flags;
  55
  56         spin_lock_irqsave(&dreq->lock, flags);
  57         bio->bi_private = dreq->bio_list;
  58         dreq->bio_list = bio;
  59         spin_unlock_irqrestore(&dreq->lock, flags);
  60         if (atomic_dec_and_test(&dreq->numreqs))
  61                 wake_up(&dreq->wait);
  62
  63         return 0;
  64 }
  65
  66 static int can_be_merged(struct bio *bio, sector_t sector)
  67 {
  68         int size;
  69
  70         if (!bio)
  71                 return 0;
  72
  73         size = bio->bi_size >> 9;
  74         return bio->bi_sector + size == sector ? 1 : 0;
  75 }
  76
  77 /* See if there are unallocated parts in given file region */
  78 static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len)
  79 {
  80         sector_t (*fs_bmap)(struct address_space *, sector_t) =
  81                 inode->i_mapping->a_ops->bmap;
  82         int j;
  83
  84         /* We can't know if we are overwriting or not */
  85         if (fs_bmap == NULL)
  86                 return 0;
  87
  88         offset >>= inode->i_blkbits;
  89         len >>= inode->i_blkbits;
  90
  91         for (j = 0; j <= len; j++)
  92                 if (fs_bmap(inode->i_mapping, offset + j) == 0)
  93                         return 0;
  94
  95         return 1;
  96 }
  97
  98 int filter_commitrw_write(struct obd_export *exp, struct obdo *oa, int objcount,
  99                           struct obd_ioobj *obj, int niocount,
 100                           struct niobuf_local *res, struct obd_trans_info *oti)
 101 {
 102         struct obd_device *obd = exp->exp_obd;
 103         struct obd_run_ctxt saved;
 104         struct niobuf_local *lnb;
 105         struct fsfilt_objinfo fso;
 106         struct iattr iattr = { .ia_valid = ATTR_SIZE, .ia_size = 0, };
 107         struct inode *inode = NULL;
 108         int rc = 0, i, k, cleanup_phase = 0, err;
 109         unsigned long now = jiffies; /* DEBUGGING OST TIMEOUTS */
 110         int blocks_per_page;
 111         struct dio_request *dreq;
 112         struct bio *bio = NULL;
 113         ENTRY;
 114         LASSERT(oti != NULL);
 115         LASSERT(objcount == 1);
 116         LASSERT(current->journal_info == NULL);
 117
 118         inode = res->dentry->d_inode;
 119         blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 120         LASSERT(blocks_per_page <= MAX_BLOCKS_PER_PAGE);
 121
 122         OBD_ALLOC(dreq, sizeof(*dreq));
 123         if (dreq == NULL)
 124                 RETURN(-ENOMEM);
 125         dreq->bio_list = NULL;
 126         init_waitqueue_head(&dreq->wait);
 127         atomic_set(&dreq->numreqs, 0);
 128         spin_lock_init(&dreq->lock);
 129
 130         cleanup_phase = 1;
 131         fso.fso_dentry = res->dentry;
 132         fso.fso_bufcnt = obj->ioo_bufcnt;
 133
 134         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 135         cleanup_phase = 2;
 136
 137         oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res, oti);
 138         if (IS_ERR(oti->oti_handle)) {
 139                 rc = PTR_ERR(oti->oti_handle);
 140                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 141                        "error starting transaction: rc = %d\n", rc);
 142                 oti->oti_handle = NULL;
 143                 GOTO(cleanup, rc);
 144         }
 145
 146         if (time_after(jiffies, now + 15 * HZ))
 147                 CERROR("slow brw_start %lus\n", (jiffies - now) / HZ);
 148
 149         iattr_from_obdo(&iattr,oa,OBD_MD_FLATIME|OBD_MD_FLMTIME|OBD_MD_FLCTIME);
 150         for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
 151                 loff_t this_size;
 152                 sector_t sector;
 153                 int offs;
 154
 155                 /* If overwriting an existing block, we don't need a grant */
 156                 if (!(lnb->flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC &&
 157                     filter_range_is_mapped(inode, lnb->offset, lnb->len))
 158                         lnb->rc = 0;
 159
 160                 if (lnb->rc) /* ENOSPC, network RPC error */
 161                         continue;
 162
 163                 /* get block number for next page */
 164                 rc = fsfilt_map_inode_page(obd, inode, lnb->page, dreq->blocks,
 165                                            dreq->created, 1);
 166                 if (rc)
 167                         GOTO(cleanup, rc);
 168
 169                 for (k = 0; k < blocks_per_page; k++) {
 170                         sector = dreq->blocks[k] *(inode->i_sb->s_blocksize>>9);
 171                         offs = k * inode->i_sb->s_blocksize;
 172
 173                         if (!bio || !can_be_merged(bio, sector) ||
 174                             !bio_add_page(bio, lnb->page, lnb->len, offs)) {
 175                                 if (bio) {
 176                                         atomic_inc(&dreq->numreqs);
 177                                         submit_bio(WRITE, bio);
 178                                         bio = NULL;
 179                                 }
 180                                 /* allocate new bio */
 181                                 bio = bio_alloc(GFP_NOIO, obj->ioo_bufcnt);
 182                                 bio->bi_bdev = inode->i_sb->s_bdev;
 183                                 bio->bi_sector = sector;
 184                                 bio->bi_end_io = dio_complete_routine;
 185                                 bio->bi_private = dreq;
 186
 187                                 if (!bio_add_page(bio, lnb->page, lnb->len, 0))
 188                                         LBUG();
 189                         }
 190                 }
 191
 192                 /* We expect these pages to be in offset order, but we'll
 193                  * be forgiving */
 194                 this_size = lnb->offset + lnb->len;
 195                 if (this_size > iattr.ia_size)
 196                         iattr.ia_size = this_size;
 197         }
 198
 199 #warning This probably needs filemap_fdatasync() like filter_io_24 (bug 2366)
 200         if (bio) {
 201                 atomic_inc(&dreq->numreqs);
 202                 submit_bio(WRITE, bio);
 203         }
 204
 205         /* time to wait for I/O completion */
 206         wait_event(dreq->wait, atomic_read(&dreq->numreqs) == 0);
 207
 208         /* free all bios */
 209         while (dreq->bio_list) {
 210                 bio = dreq->bio_list;
 211                 dreq->bio_list = bio->bi_private;
 212                 bio_put(bio);
 213         }
 214
 215         if (rc == 0) {
 216                 down(&inode->i_sem);
 217                 if (iattr.ia_size > inode->i_size) {
 218                         CDEBUG(D_INFO, "setting i_size to "LPU64"\n",
 219                                iattr.ia_size);
 220                         fsfilt_setattr(obd, res->dentry, oti->oti_handle,
 221                                        &iattr, 0);
 222                 }
 223                 up(&inode->i_sem);
 224         }
 225
 226         if (time_after(jiffies, now + 15 * HZ))
 227                 CERROR("slow direct_io %lus\n", (jiffies - now) / HZ);
 228
 229         rc = filter_finish_transno(exp, oti, rc);
 230         err = fsfilt_commit(obd, inode, oti->oti_handle, obd_sync_filter);
 231         if (err)
 232                 rc = err;
 233         if (obd_sync_filter)
 234                 LASSERT(oti->oti_transno <= obd->obd_last_committed);
 235         if (time_after(jiffies, now + 15 * HZ))
 236                 CERROR("slow commitrw commit %lus\n", (jiffies - now) / HZ);
 237
 238 cleanup:
 239         filter_grant_commit(exp, niocount, res);
 240
 241         switch (cleanup_phase) {
 242         case 2:
 243                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 244                 LASSERT(current->journal_info == NULL);
 245         case 1:
 246                 OBD_FREE(dreq, sizeof(*dreq));
 247         case 0:
 248                 for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
 249                         /* flip_.. gets a ref, while free_page only frees
 250                          * when it decrefs to 0 */
 251                         if (rc == 0)
 252                                 flip_into_page_cache(inode, lnb->page);
 253                         __free_page(lnb->page);
 254                 }
 255                 f_dput(res->dentry);
 256         }
 257
 258         RETURN(rc);
 259 }