lustre/obdclass/llog.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2001-2003 Cluster File Systems, Inc.
   5  *   Author: Andreas Dilger <adilger@clusterfs.com>
   6  *
   7  *   This file is part of Lustre, http://www.lustre.org.
   8  *
   9  *   Lustre is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Lustre is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Lustre; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  *
  22  * OST<->MDS recovery logging infrastructure.
  23  *
  24  * Invariants in implementation:
  25  * - we do not share logs among different OST<->MDS connections, so that
  26  *   if an OST or MDS fails it need only look at log(s) relevant to itself
  27  */
  28
  29 #define DEBUG_SUBSYSTEM S_LOG
  30
  31 #ifndef EXPORT_SYMTAB
  32 #define EXPORT_SYMTAB
  33 #endif
  34
  35 #ifdef __KERNEL__
  36 #include <linux/fs.h>
  37 #else
  38 #include <liblustre.h>
  39 #endif
  40
  41 #include <linux/obd_class.h>
  42 #include <linux/lustre_log.h>
  43 #include <libcfs/list.h>
  44
  45 /* Allocate a new log or catalog handle */
  46 struct llog_handle *llog_alloc_handle(void)
  47 {
  48         struct llog_handle *loghandle;
  49         ENTRY;
  50
  51         OBD_ALLOC(loghandle, sizeof(*loghandle));
  52         if (loghandle == NULL)
  53                 RETURN(ERR_PTR(-ENOMEM));
  54
  55         init_rwsem(&loghandle->lgh_lock);
  56
  57         RETURN(loghandle);
  58 }
  59 EXPORT_SYMBOL(llog_alloc_handle);
  60
  61
  62 void llog_free_handle(struct llog_handle *loghandle)
  63 {
  64         if (!loghandle)
  65                 return;
  66
  67         if (!loghandle->lgh_hdr)
  68                 goto out;
  69         if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN)
  70                 list_del_init(&loghandle->u.phd.phd_entry);
  71         if (loghandle->lgh_hdr->llh_flags & LLOG_F_IS_CAT)
  72                 LASSERT(list_empty(&loghandle->u.chd.chd_head));
  73         OBD_FREE(loghandle->lgh_hdr, LLOG_CHUNK_SIZE);
  74
  75  out:
  76         OBD_FREE(loghandle, sizeof(*loghandle));
  77 }
  78 EXPORT_SYMBOL(llog_free_handle);
  79
  80 /* returns negative on error; 0 if success; 1 if success & log destroyed */
  81 int llog_cancel_rec(struct llog_handle *loghandle, int index)
  82 {
  83         struct llog_log_hdr *llh = loghandle->lgh_hdr;
  84         int rc = 0;
  85         ENTRY;
  86
  87         CDEBUG(D_HA, "canceling %d in log "LPX64"\n",
  88                index, loghandle->lgh_id.lgl_oid);
  89
  90         if (index == 0) {
  91                 CERROR("cannot cancel index 0 (which is header)\n");
  92                 RETURN(-EINVAL);
  93         }
  94
  95         if (!ext2_clear_bit(index, llh->llh_bitmap)) {
  96                 CDEBUG(D_HA, "catalog index %u already clear?\n", index);
  97                 RETURN(-EINVAL);
  98         }
  99
 100         llh->llh_count--;
 101
 102         if ((llh->llh_flags & LLOG_F_ZAP_WHEN_EMPTY) &&
 103             (llh->llh_count == 1) &&
 104             (loghandle->lgh_last_idx == (LLOG_BITMAP_BYTES * 8) - 1)) {
 105                 rc = llog_destroy(loghandle);
 106                 if (rc) {
 107                         CERROR("failure destroying log after last cancel: %d\n",
 108                                rc);
 109                         ext2_set_bit(index, llh->llh_bitmap);
 110                         llh->llh_count++;
 111                 } else {
 112                         rc = 1;
 113                 }
 114                 RETURN(rc);
 115         }
 116
 117         rc = llog_write_rec(loghandle, &llh->llh_hdr, NULL, 0, NULL, 0);
 118         if (rc) {
 119                 CERROR("failure re-writing header %d\n", rc);
 120                 ext2_set_bit(index, llh->llh_bitmap);
 121                 llh->llh_count++;
 122         }
 123         RETURN(rc);
 124 }
 125 EXPORT_SYMBOL(llog_cancel_rec);
 126
 127 int llog_init_handle(struct llog_handle *handle, int flags,
 128                      struct obd_uuid *uuid)
 129 {
 130         int rc;
 131         struct llog_log_hdr *llh;
 132         ENTRY;
 133         LASSERT(handle->lgh_hdr == NULL);
 134
 135         OBD_ALLOC(llh, sizeof(*llh));
 136         if (llh == NULL)
 137                 RETURN(-ENOMEM);
 138         handle->lgh_hdr = llh;
 139         /* first assign flags to use llog_client_ops */
 140         llh->llh_flags = flags;
 141         rc = llog_read_header(handle);
 142         if (rc == 0) {
 143                 flags = llh->llh_flags;
 144                 if (uuid)
 145                         LASSERT(obd_uuid_equals(uuid, &llh->llh_tgtuuid));
 146                 GOTO(out, rc);
 147         } else if (rc != LLOG_EEMPTY || !flags) {
 148                 /* set a pesudo flag for initialization */
 149                 flags = LLOG_F_IS_CAT;
 150                 GOTO(out, rc);
 151         }
 152         rc = 0;
 153
 154         handle->lgh_last_idx = 0; /* header is record with index 0 */
 155         llh->llh_count = 1;         /* for the header record */
 156         llh->llh_hdr.lrh_type = LLOG_HDR_MAGIC;
 157         llh->llh_hdr.lrh_len = llh->llh_tail.lrt_len = LLOG_CHUNK_SIZE;
 158         llh->llh_hdr.lrh_index = llh->llh_tail.lrt_index = 0;
 159         llh->llh_timestamp = CURRENT_SECONDS;
 160         if (uuid)
 161                 memcpy(&llh->llh_tgtuuid, uuid, sizeof(llh->llh_tgtuuid));
 162         llh->llh_bitmap_offset = offsetof(typeof(*llh),llh_bitmap);
 163         ext2_set_bit(0, llh->llh_bitmap);
 164
 165 out:
 166         if (flags & LLOG_F_IS_CAT) {
 167                 INIT_LIST_HEAD(&handle->u.chd.chd_head);
 168                 llh->llh_size = sizeof(struct llog_logid_rec);
 169         } else if (flags & LLOG_F_IS_PLAIN) {
 170                 INIT_LIST_HEAD(&handle->u.phd.phd_entry);
 171         } else {
 172                 CERROR("Unknown flags: %#x (Expected %#x or %#x\n",
 173                        flags, LLOG_F_IS_CAT, LLOG_F_IS_PLAIN);
 174                 LBUG();
 175         }
 176
 177         if (rc) {
 178                 OBD_FREE(llh, sizeof(*llh));
 179                 handle->lgh_hdr = NULL;
 180         }
 181         RETURN(rc);
 182 }
 183 EXPORT_SYMBOL(llog_init_handle);
 184
 185 int llog_close(struct llog_handle *loghandle)
 186 {
 187         struct llog_operations *lop;
 188         int rc;
 189         ENTRY;
 190
 191         rc = llog_handle2ops(loghandle, &lop);
 192         if (rc)
 193                 GOTO(out, rc);
 194         if (lop->lop_close == NULL)
 195                 GOTO(out, -EOPNOTSUPP);
 196         rc = lop->lop_close(loghandle);
 197  out:
 198         llog_free_handle(loghandle);
 199         RETURN(rc);
 200 }
 201 EXPORT_SYMBOL(llog_close);
 202
 203 int llog_process(struct llog_handle *loghandle, llog_cb_t cb,
 204                  void *data, void *catdata)
 205 {
 206         struct llog_log_hdr *llh = loghandle->lgh_hdr;
 207         struct llog_process_cat_data *cd = catdata;
 208         char *buf;
 209         __u64 cur_offset = LLOG_CHUNK_SIZE;
 210         int rc = 0, index = 1, last_index;
 211         int saved_index = 0;
 212         ENTRY;
 213
 214         OBD_ALLOC(buf, LLOG_CHUNK_SIZE);
 215         if (!buf)
 216                 RETURN(-ENOMEM);
 217
 218         if (cd != NULL)
 219                 index = cd->first_idx + 1;
 220         if (cd != NULL && cd->last_idx)
 221                 last_index = cd->last_idx;
 222         else
 223                 last_index = LLOG_BITMAP_BYTES * 8 - 1;
 224
 225         while (rc == 0) {
 226                 struct llog_rec_hdr *rec;
 227
 228                 /* skip records not set in bitmap */
 229                 while (index <= last_index &&
 230                        !ext2_test_bit(index, llh->llh_bitmap))
 231                         ++index;
 232
 233                 LASSERT(index <= last_index + 1);
 234                 if (index == last_index + 1)
 235                         break;
 236
 237                 CDEBUG(D_OTHER, "index: %d last_index %d\n",
 238                        index, last_index);
 239
 240                 /* get the buf with our target record; avoid old garbage */
 241                 memset(buf, 0, LLOG_CHUNK_SIZE);
 242                 rc = llog_next_block(loghandle, &saved_index, index,
 243                                      &cur_offset, buf, LLOG_CHUNK_SIZE);
 244                 if (rc)
 245                         GOTO(out, rc);
 246
 247                 /* NB: when rec->lrh_len is accessed it is already swabbed
 248                  * since it is used at the "end" of the loop and the rec
 249                  * swabbing is done at the beginning of the loop. */
 250                 for (rec = (struct llog_rec_hdr *)buf;
 251                      (char *)rec < buf + LLOG_CHUNK_SIZE;
 252                      rec = (struct llog_rec_hdr *)((char *)rec + rec->lrh_len)){
 253
 254                         CDEBUG(D_OTHER, "processing rec 0x%p type %#x\n",
 255                                rec, rec->lrh_type);
 256
 257                         if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
 258                                 lustre_swab_llog_rec(rec, NULL);
 259
 260                         CDEBUG(D_OTHER, "after swabbing, type: %#x\n",
 261                                rec->lrh_type);
 262
 263                         if (rec->lrh_index == 0)
 264                                 GOTO(out, 0); /* no more records */
 265
 266                         if (rec->lrh_index < index) {
 267                                 CDEBUG(D_OTHER, "skipping lrh_index %d\n",
 268                                        rec->lrh_index);
 269                                 continue;
 270                         }
 271
 272                         CDEBUG(D_OTHER,
 273                                "lrh_index: %d lrh_len: %d (%d remains)\n",
 274                                rec->lrh_index, rec->lrh_len,
 275                                (int)(buf + LLOG_CHUNK_SIZE - (char *)rec));
 276
 277                         /* if set, process the callback on this record */
 278                         if (ext2_test_bit(index, llh->llh_bitmap)) {
 279                                 rc = cb(loghandle, rec, data);
 280                                 if (rc == LLOG_PROC_BREAK) {
 281                                         CWARN("recovery from log: "LPX64":%x"
 282                                               " stopped\n",
 283                                               loghandle->lgh_id.lgl_oid,
 284                                               loghandle->lgh_id.lgl_ogen);
 285                                         GOTO(out, rc);
 286                                 }
 287                                 if (rc)
 288                                         GOTO(out, rc);
 289                         } else {
 290                                 CDEBUG(D_OTHER, "Skipped index %d\n", index);
 291                         }
 292
 293                         /* next record, still in buffer? */
 294                         ++index;
 295                         if (index > last_index)
 296                                 GOTO(out, rc = 0);
 297                 }
 298         }
 299
 300  out:
 301         if (buf)
 302                 OBD_FREE(buf, LLOG_CHUNK_SIZE);
 303         RETURN(rc);
 304 }
 305 EXPORT_SYMBOL(llog_process);