server/ctdb_update_record.c

   1 /*
   2    implementation of the update record control
   3
   4    Copyright (C) Andrew Tridgell  2007
   5    Copyright (C) Ronnie Sahlberg  2007
   6
   7    This program is free software; you can redistribute it and/or modify
   8    it under the terms of the GNU General Public License as published by
   9    the Free Software Foundation; either version 3 of the License, or
  10    (at your option) any later version.
  11
  12    This program is distributed in the hope that it will be useful,
  13    but WITHOUT ANY WARRANTY; without even the implied warranty of
  14    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15    GNU General Public License for more details.
  16
  17    You should have received a copy of the GNU General Public License
  18    along with this program; if not, see <http://www.gnu.org/licenses/>.
  19 */
  20
  21 #include "includes.h"
  22 #include "db_wrap.h"
  23 #include "lib/tdb/include/tdb.h"
  24 #include "ctdb_private.h"
  25
  26 struct ctdb_persistent_write_state {
  27         struct ctdb_db_context *ctdb_db;
  28         struct ctdb_marshall_buffer *m;
  29         struct ctdb_req_control *c;
  30         uint32_t flags;
  31 };
  32
  33 /* dont create/update records that does not exist locally */
  34 #define UPDATE_FLAGS_REPLACE_ONLY       1
  35
  36 /*
  37   called from a child process to write the data
  38  */
  39 static int ctdb_persistent_store(struct ctdb_persistent_write_state *state)
  40 {
  41         int ret, i;
  42         struct ctdb_rec_data *rec = NULL;
  43         struct ctdb_marshall_buffer *m = state->m;
  44
  45         ret = tdb_transaction_start(state->ctdb_db->ltdb->tdb);
  46         if (ret == -1) {
  47                 DEBUG(DEBUG_ERR,("Failed to start transaction for db_id 0x%08x in ctdb_persistent_store\n",
  48                                  state->ctdb_db->db_id));
  49                 return -1;
  50         }
  51
  52         for (i=0;i<m->count;i++) {
  53                 struct ctdb_ltdb_header oldheader;
  54                 struct ctdb_ltdb_header header;
  55                 TDB_DATA key, data, olddata;
  56                 TALLOC_CTX *tmp_ctx = talloc_new(state);
  57
  58                 rec = ctdb_marshall_loop_next(m, rec, NULL, &header, &key, &data);
  59
  60                 if (rec == NULL) {
  61                         DEBUG(DEBUG_ERR,("Failed to get next record %d for db_id 0x%08x in ctdb_persistent_store\n",
  62                                          i, state->ctdb_db->db_id));
  63                         talloc_free(tmp_ctx);
  64                         goto failed;
  65                 }
  66
  67                 /* we must check if the record exists or not because
  68                    ctdb_ltdb_fetch will unconditionally create a record
  69                  */
  70                 if (state->flags & UPDATE_FLAGS_REPLACE_ONLY) {
  71                         TDB_DATA trec;
  72                         trec = tdb_fetch(state->ctdb_db->ltdb->tdb, key);
  73                         if (trec.dsize == 0) {
  74                                 talloc_free(tmp_ctx);
  75                                 continue;
  76                         }
  77                         free(trec.dptr);
  78                 }
  79
  80                 /* fetch the old header and ensure the rsn is less than the new rsn */
  81                 ret = ctdb_ltdb_fetch(state->ctdb_db, key, &oldheader, tmp_ctx, &olddata);
  82                 if (ret != 0) {
  83                         DEBUG(DEBUG_ERR,("Failed to fetch old record for db_id 0x%08x in ctdb_persistent_store\n",
  84                                          state->ctdb_db->db_id));
  85                         talloc_free(tmp_ctx);
  86                         goto failed;
  87                 }
  88
  89                 if (oldheader.rsn >= header.rsn &&
  90                     (olddata.dsize != data.dsize ||
  91                      memcmp(olddata.dptr, data.dptr, data.dsize) != 0)) {
  92                         DEBUG(DEBUG_CRIT,("existing header for db_id 0x%08x has larger RSN %llu than new RSN %llu in ctdb_persistent_store\n",
  93                                           state->ctdb_db->db_id,
  94                                           (unsigned long long)oldheader.rsn, (unsigned long long)header.rsn));
  95                         talloc_free(tmp_ctx);
  96                         goto failed;
  97                 }
  98
  99                 talloc_free(tmp_ctx);
 100
 101                 ret = ctdb_ltdb_store(state->ctdb_db, key, &header, data);
 102                 if (ret != 0) {
 103                         DEBUG(DEBUG_CRIT,("Failed to store record for db_id 0x%08x in ctdb_persistent_store\n",
 104                                           state->ctdb_db->db_id));
 105                         goto failed;
 106                 }
 107         }
 108
 109         ret = tdb_transaction_commit(state->ctdb_db->ltdb->tdb);
 110         if (ret == -1) {
 111                 DEBUG(DEBUG_ERR,("Failed to commit transaction for db_id 0x%08x in ctdb_persistent_store\n",
 112                                  state->ctdb_db->db_id));
 113                 return -1;
 114         }
 115
 116         return 0;
 117
 118 failed:
 119         tdb_transaction_cancel(state->ctdb_db->ltdb->tdb);
 120         return -1;
 121 }
 122
 123
 124 /*
 125   called when we the child has completed the persistent write
 126   on our behalf
 127  */
 128 static void ctdb_persistent_write_callback(int status, void *private_data)
 129 {
 130         struct ctdb_persistent_write_state *state = talloc_get_type(private_data,
 131                                                                    struct ctdb_persistent_write_state);
 132
 133
 134         ctdb_request_control_reply(state->ctdb_db->ctdb, state->c, NULL, status, NULL);
 135
 136         talloc_free(state);
 137 }
 138
 139 /*
 140   called if our lockwait child times out
 141  */
 142 static void ctdb_persistent_lock_timeout(struct event_context *ev, struct timed_event *te,
 143                                          struct timeval t, void *private_data)
 144 {
 145         struct ctdb_persistent_write_state *state = talloc_get_type(private_data,
 146                                                                    struct ctdb_persistent_write_state);
 147         ctdb_request_control_reply(state->ctdb_db->ctdb, state->c, NULL, -1, "timeout in ctdb_persistent_lock");
 148         talloc_free(state);
 149 }
 150
 151 struct childwrite_handle {
 152         struct ctdb_context *ctdb;
 153         struct ctdb_db_context *ctdb_db;
 154         struct fd_event *fde;
 155         int fd[2];
 156         pid_t child;
 157         void *private_data;
 158         void (*callback)(int, void *);
 159         struct timeval start_time;
 160 };
 161
 162 static int childwrite_destructor(struct childwrite_handle *h)
 163 {
 164         CTDB_DECREMENT_STAT(h->ctdb, pending_childwrite_calls);
 165         ctdb_kill(h->ctdb, h->child, SIGKILL);
 166         return 0;
 167 }
 168
 169 /* called when the child process has finished writing the record to the
 170    database
 171 */
 172 static void childwrite_handler(struct event_context *ev, struct fd_event *fde,
 173                              uint16_t flags, void *private_data)
 174 {
 175         struct childwrite_handle *h = talloc_get_type(private_data,
 176                                                      struct childwrite_handle);
 177         void *p = h->private_data;
 178         void (*callback)(int, void *) = h->callback;
 179         pid_t child = h->child;
 180         TALLOC_CTX *tmp_ctx = talloc_new(ev);
 181         int ret;
 182         char c;
 183
 184         CTDB_UPDATE_LATENCY(h->ctdb, h->ctdb_db, "persistent", childwrite_latency, h->start_time);
 185         CTDB_DECREMENT_STAT(h->ctdb, pending_childwrite_calls);
 186
 187         /* the handle needs to go away when the context is gone - when
 188            the handle goes away this implicitly closes the pipe, which
 189            kills the child */
 190         talloc_steal(tmp_ctx, h);
 191
 192         talloc_set_destructor(h, NULL);
 193
 194         ret = read(h->fd[0], &c, 1);
 195         if (ret < 1) {
 196                 DEBUG(DEBUG_ERR, (__location__ " Read returned %d. Childwrite failed\n", ret));
 197                 c = 1;
 198         }
 199
 200         callback(c, p);
 201
 202         ctdb_kill(h->ctdb, child, SIGKILL);
 203         talloc_free(tmp_ctx);
 204 }
 205
 206 /* this creates a child process which will take out a tdb transaction
 207    and write the record to the database.
 208 */
 209 static struct childwrite_handle *ctdb_childwrite(
 210                                 struct ctdb_db_context *ctdb_db,
 211                                 void (*callback)(int, void *private_data),
 212                                 struct ctdb_persistent_write_state *state)
 213 {
 214         struct childwrite_handle *result;
 215         int ret;
 216         pid_t parent = getpid();
 217
 218         CTDB_INCREMENT_STAT(ctdb_db->ctdb, childwrite_calls);
 219         CTDB_INCREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 220
 221         if (!(result = talloc_zero(state, struct childwrite_handle))) {
 222                 CTDB_DECREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 223                 return NULL;
 224         }
 225
 226         ret = pipe(result->fd);
 227
 228         if (ret != 0) {
 229                 talloc_free(result);
 230                 CTDB_DECREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 231                 return NULL;
 232         }
 233
 234         result->child = ctdb_fork(ctdb_db->ctdb);
 235
 236         if (result->child == (pid_t)-1) {
 237                 close(result->fd[0]);
 238                 close(result->fd[1]);
 239                 talloc_free(result);
 240                 CTDB_DECREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 241                 return NULL;
 242         }
 243
 244         result->callback = callback;
 245         result->private_data = state;
 246         result->ctdb = ctdb_db->ctdb;
 247         result->ctdb_db = ctdb_db;
 248
 249         if (result->child == 0) {
 250                 char c = 0;
 251
 252                 close(result->fd[0]);
 253                 debug_extra = talloc_asprintf(NULL, "childwrite-%s:", ctdb_db->db_name);
 254                 ret = ctdb_persistent_store(state);
 255                 if (ret != 0) {
 256                         DEBUG(DEBUG_ERR, (__location__ " Failed to write persistent data\n"));
 257                         c = 1;
 258                 }
 259
 260                 write(result->fd[1], &c, 1);
 261
 262                 /* make sure we die when our parent dies */
 263                 while (ctdb_kill(ctdb_db->ctdb, parent, 0) == 0 || errno != ESRCH) {
 264                         sleep(5);
 265                 }
 266                 _exit(0);
 267         }
 268
 269         close(result->fd[1]);
 270         set_close_on_exec(result->fd[0]);
 271
 272         talloc_set_destructor(result, childwrite_destructor);
 273
 274         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d for ctdb_childwrite\n", result->fd[0]));
 275
 276         result->fde = event_add_fd(ctdb_db->ctdb->ev, result, result->fd[0],
 277                                    EVENT_FD_READ, childwrite_handler,
 278                                    (void *)result);
 279         if (result->fde == NULL) {
 280                 talloc_free(result);
 281                 CTDB_DECREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 282                 return NULL;
 283         }
 284         tevent_fd_set_auto_close(result->fde);
 285
 286         result->start_time = timeval_current();
 287
 288         return result;
 289 }
 290
 291 /*
 292    update a record on this node if the new record has a higher rsn than the
 293    current record
 294  */
 295 int32_t ctdb_control_update_record(struct ctdb_context *ctdb,
 296                                    struct ctdb_req_control *c, TDB_DATA recdata,
 297                                    bool *async_reply)
 298 {
 299         struct ctdb_db_context *ctdb_db;
 300         struct ctdb_persistent_write_state *state;
 301         struct childwrite_handle *handle;
 302         struct ctdb_marshall_buffer *m = (struct ctdb_marshall_buffer *)recdata.dptr;
 303
 304         if (ctdb->recovery_mode != CTDB_RECOVERY_NORMAL) {
 305                 DEBUG(DEBUG_INFO,("rejecting ctdb_control_update_record when recovery active\n"));
 306                 return -1;
 307         }
 308
 309         ctdb_db = find_ctdb_db(ctdb, m->db_id);
 310         if (ctdb_db == NULL) {
 311                 DEBUG(DEBUG_ERR,("Unknown database 0x%08x in ctdb_control_update_record\n", m->db_id));
 312                 return -1;
 313         }
 314
 315         if (ctdb_db->unhealthy_reason) {
 316                 DEBUG(DEBUG_ERR,("db(%s) unhealty in ctdb_control_update_record: %s\n",
 317                                  ctdb_db->db_name, ctdb_db->unhealthy_reason));
 318                 return -1;
 319         }
 320
 321         state = talloc(ctdb, struct ctdb_persistent_write_state);
 322         CTDB_NO_MEMORY(ctdb, state);
 323
 324         state->ctdb_db = ctdb_db;
 325         state->c       = c;
 326         state->m       = m;
 327         state->flags   = 0;
 328         if (!ctdb_db->persistent) {
 329                 state->flags   = UPDATE_FLAGS_REPLACE_ONLY;
 330         }
 331
 332         /* create a child process to take out a transaction and
 333            write the data.
 334         */
 335         handle = ctdb_childwrite(ctdb_db, ctdb_persistent_write_callback, state);
 336         if (handle == NULL) {
 337                 DEBUG(DEBUG_ERR,("Failed to setup childwrite handler in ctdb_control_update_record\n"));
 338                 talloc_free(state);
 339                 return -1;
 340         }
 341
 342         /* we need to wait for the replies */
 343         *async_reply = true;
 344
 345         /* need to keep the control structure around */
 346         talloc_steal(state, c);
 347
 348         /* but we won't wait forever */
 349         event_add_timed(ctdb->ev, state, timeval_current_ofs(ctdb->tunable.control_timeout, 0),
 350                         ctdb_persistent_lock_timeout, state);
 351
 352         return 0;
 353 }
 354