ctdb/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 "replace.h"
  22 #include "system/network.h"
  23 #include "system/time.h"
  24
  25 #include <talloc.h>
  26 #include <tevent.h>
  27
  28 #include "lib/tdb_wrap/tdb_wrap.h"
  29 #include "lib/util/debug.h"
  30 #include "lib/util/samba_util.h"
  31 #include "lib/util/sys_rw.h"
  32 #include "lib/util/util_process.h"
  33
  34 #include "ctdb_private.h"
  35 #include "ctdb_client.h"
  36
  37 #include "common/system.h"
  38 #include "common/common.h"
  39 #include "common/logging.h"
  40
  41 struct ctdb_persistent_write_state {
  42         struct ctdb_db_context *ctdb_db;
  43         struct ctdb_marshall_buffer *m;
  44         struct ctdb_req_control_old *c;
  45         uint32_t flags;
  46 };
  47
  48 /* don't create/update records that does not exist locally */
  49 #define UPDATE_FLAGS_REPLACE_ONLY       1
  50
  51 /*
  52   called from a child process to write the data
  53  */
  54 static int ctdb_persistent_store(struct ctdb_persistent_write_state *state)
  55 {
  56         int ret, i;
  57         struct ctdb_rec_data_old *rec = NULL;
  58         struct ctdb_marshall_buffer *m = state->m;
  59
  60         ret = tdb_transaction_start(state->ctdb_db->ltdb->tdb);
  61         if (ret == -1) {
  62                 DEBUG(DEBUG_ERR,("Failed to start transaction for db_id 0x%08x in ctdb_persistent_store\n",
  63                                  state->ctdb_db->db_id));
  64                 return -1;
  65         }
  66
  67         for (i=0;i<m->count;i++) {
  68                 struct ctdb_ltdb_header oldheader;
  69                 struct ctdb_ltdb_header header;
  70                 TDB_DATA key, data, olddata;
  71                 TALLOC_CTX *tmp_ctx = talloc_new(state);
  72
  73                 rec = ctdb_marshall_loop_next(m, rec, NULL, &header, &key, &data);
  74
  75                 if (rec == NULL) {
  76                         DEBUG(DEBUG_ERR,("Failed to get next record %d for db_id 0x%08x in ctdb_persistent_store\n",
  77                                          i, state->ctdb_db->db_id));
  78                         talloc_free(tmp_ctx);
  79                         goto failed;
  80                 }
  81
  82                 /* we must check if the record exists or not because
  83                    ctdb_ltdb_fetch will unconditionally create a record
  84                  */
  85                 if (state->flags & UPDATE_FLAGS_REPLACE_ONLY) {
  86                         TDB_DATA trec;
  87                         trec = tdb_fetch(state->ctdb_db->ltdb->tdb, key);
  88                         if (trec.dsize == 0) {
  89                                 talloc_free(tmp_ctx);
  90                                 continue;
  91                         }
  92                         free(trec.dptr);
  93                 }
  94
  95                 /* fetch the old header and ensure the rsn is less than the new rsn */
  96                 ret = ctdb_ltdb_fetch(state->ctdb_db, key, &oldheader, tmp_ctx, &olddata);
  97                 if (ret != 0) {
  98                         DEBUG(DEBUG_ERR,("Failed to fetch old record for db_id 0x%08x in ctdb_persistent_store\n",
  99                                          state->ctdb_db->db_id));
 100                         talloc_free(tmp_ctx);
 101                         goto failed;
 102                 }
 103
 104                 if (oldheader.rsn >= header.rsn &&
 105                     (olddata.dsize != data.dsize ||
 106                      memcmp(olddata.dptr, data.dptr, data.dsize) != 0)) {
 107                         DEBUG(DEBUG_CRIT,("existing header for db_id 0x%08x has larger RSN %llu than new RSN %llu in ctdb_persistent_store\n",
 108                                           state->ctdb_db->db_id,
 109                                           (unsigned long long)oldheader.rsn, (unsigned long long)header.rsn));
 110                         talloc_free(tmp_ctx);
 111                         goto failed;
 112                 }
 113
 114                 talloc_free(tmp_ctx);
 115
 116                 ret = ctdb_ltdb_store(state->ctdb_db, key, &header, data);
 117                 if (ret != 0) {
 118                         DEBUG(DEBUG_CRIT,("Failed to store record for db_id 0x%08x in ctdb_persistent_store\n",
 119                                           state->ctdb_db->db_id));
 120                         goto failed;
 121                 }
 122         }
 123
 124         ret = tdb_transaction_commit(state->ctdb_db->ltdb->tdb);
 125         if (ret == -1) {
 126                 DEBUG(DEBUG_ERR,("Failed to commit transaction for db_id 0x%08x in ctdb_persistent_store\n",
 127                                  state->ctdb_db->db_id));
 128                 return -1;
 129         }
 130
 131         return 0;
 132
 133 failed:
 134         tdb_transaction_cancel(state->ctdb_db->ltdb->tdb);
 135         return -1;
 136 }
 137
 138
 139 /*
 140   called when we the child has completed the persistent write
 141   on our behalf
 142  */
 143 static void ctdb_persistent_write_callback(int status, void *private_data)
 144 {
 145         struct ctdb_persistent_write_state *state = talloc_get_type(private_data,
 146                                                                    struct ctdb_persistent_write_state);
 147
 148
 149         ctdb_request_control_reply(state->ctdb_db->ctdb, state->c, NULL, status, NULL);
 150
 151         talloc_free(state);
 152 }
 153
 154 /*
 155   called if our lockwait child times out
 156  */
 157 static void ctdb_persistent_lock_timeout(struct tevent_context *ev,
 158                                          struct tevent_timer *te,
 159                                          struct timeval t, void *private_data)
 160 {
 161         struct ctdb_persistent_write_state *state = talloc_get_type(private_data,
 162                                                                    struct ctdb_persistent_write_state);
 163         ctdb_request_control_reply(state->ctdb_db->ctdb, state->c, NULL, -1, "timeout in ctdb_persistent_lock");
 164         talloc_free(state);
 165 }
 166
 167 struct childwrite_handle {
 168         struct ctdb_context *ctdb;
 169         struct ctdb_db_context *ctdb_db;
 170         struct tevent_fd *fde;
 171         int fd[2];
 172         pid_t child;
 173         void *private_data;
 174         void (*callback)(int, void *);
 175         struct timeval start_time;
 176 };
 177
 178 static int childwrite_destructor(struct childwrite_handle *h)
 179 {
 180         CTDB_DECREMENT_STAT(h->ctdb, pending_childwrite_calls);
 181         ctdb_kill(h->ctdb, h->child, SIGKILL);
 182         return 0;
 183 }
 184
 185 /* called when the child process has finished writing the record to the
 186    database
 187 */
 188 static void childwrite_handler(struct tevent_context *ev,
 189                                struct tevent_fd *fde,
 190                                uint16_t flags, void *private_data)
 191 {
 192         struct childwrite_handle *h = talloc_get_type(private_data,
 193                                                      struct childwrite_handle);
 194         void *p = h->private_data;
 195         void (*callback)(int, void *) = h->callback;
 196         pid_t child = h->child;
 197         TALLOC_CTX *tmp_ctx = talloc_new(ev);
 198         int ret;
 199         char c;
 200
 201         CTDB_UPDATE_LATENCY(h->ctdb, h->ctdb_db, "persistent", childwrite_latency, h->start_time);
 202         CTDB_DECREMENT_STAT(h->ctdb, pending_childwrite_calls);
 203
 204         /* the handle needs to go away when the context is gone - when
 205            the handle goes away this implicitly closes the pipe, which
 206            kills the child */
 207         talloc_steal(tmp_ctx, h);
 208
 209         talloc_set_destructor(h, NULL);
 210
 211         ret = sys_read(h->fd[0], &c, 1);
 212         if (ret < 1) {
 213                 DEBUG(DEBUG_ERR, (__location__ " Read returned %d. Childwrite failed\n", ret));
 214                 c = 1;
 215         }
 216
 217         callback(c, p);
 218
 219         ctdb_kill(h->ctdb, child, SIGKILL);
 220         talloc_free(tmp_ctx);
 221 }
 222
 223 /* this creates a child process which will take out a tdb transaction
 224    and write the record to the database.
 225 */
 226 static struct childwrite_handle *ctdb_childwrite(
 227                                 struct ctdb_db_context *ctdb_db,
 228                                 void (*callback)(int, void *private_data),
 229                                 struct ctdb_persistent_write_state *state)
 230 {
 231         struct childwrite_handle *result;
 232         int ret;
 233         pid_t parent = getpid();
 234
 235         CTDB_INCREMENT_STAT(ctdb_db->ctdb, childwrite_calls);
 236         CTDB_INCREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 237
 238         if (!(result = talloc_zero(state, struct childwrite_handle))) {
 239                 CTDB_DECREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 240                 return NULL;
 241         }
 242
 243         ret = pipe(result->fd);
 244
 245         if (ret != 0) {
 246                 talloc_free(result);
 247                 CTDB_DECREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 248                 return NULL;
 249         }
 250
 251         result->child = ctdb_fork(ctdb_db->ctdb);
 252
 253         if (result->child == (pid_t)-1) {
 254                 close(result->fd[0]);
 255                 close(result->fd[1]);
 256                 talloc_free(result);
 257                 CTDB_DECREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 258                 return NULL;
 259         }
 260
 261         result->callback = callback;
 262         result->private_data = state;
 263         result->ctdb = ctdb_db->ctdb;
 264         result->ctdb_db = ctdb_db;
 265
 266         if (result->child == 0) {
 267                 char c = 0;
 268
 269                 close(result->fd[0]);
 270                 prctl_set_comment("ctdb_write_persistent");
 271                 ret = ctdb_persistent_store(state);
 272                 if (ret != 0) {
 273                         DEBUG(DEBUG_ERR, (__location__ " Failed to write persistent data\n"));
 274                         c = 1;
 275                 }
 276
 277                 sys_write(result->fd[1], &c, 1);
 278
 279                 ctdb_wait_for_process_to_exit(parent);
 280                 _exit(0);
 281         }
 282
 283         close(result->fd[1]);
 284         set_close_on_exec(result->fd[0]);
 285
 286         talloc_set_destructor(result, childwrite_destructor);
 287
 288         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d for ctdb_childwrite\n", result->fd[0]));
 289
 290         result->fde = tevent_add_fd(ctdb_db->ctdb->ev, result, result->fd[0],
 291                                     TEVENT_FD_READ, childwrite_handler,
 292                                     (void *)result);
 293         if (result->fde == NULL) {
 294                 talloc_free(result);
 295                 CTDB_DECREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 296                 return NULL;
 297         }
 298         tevent_fd_set_auto_close(result->fde);
 299
 300         result->start_time = timeval_current();
 301
 302         return result;
 303 }
 304
 305 /*
 306    update a record on this node if the new record has a higher rsn than the
 307    current record
 308  */
 309 int32_t ctdb_control_update_record(struct ctdb_context *ctdb,
 310                                    struct ctdb_req_control_old *c, TDB_DATA recdata,
 311                                    bool *async_reply)
 312 {
 313         struct ctdb_db_context *ctdb_db;
 314         struct ctdb_persistent_write_state *state;
 315         struct childwrite_handle *handle;
 316         struct ctdb_marshall_buffer *m = (struct ctdb_marshall_buffer *)recdata.dptr;
 317
 318         if (ctdb->recovery_mode != CTDB_RECOVERY_NORMAL) {
 319                 DEBUG(DEBUG_INFO,("rejecting ctdb_control_update_record when recovery active\n"));
 320                 return -1;
 321         }
 322
 323         ctdb_db = find_ctdb_db(ctdb, m->db_id);
 324         if (ctdb_db == NULL) {
 325                 DEBUG(DEBUG_ERR,("Unknown database 0x%08x in ctdb_control_update_record\n", m->db_id));
 326                 return -1;
 327         }
 328
 329         if (ctdb_db->unhealthy_reason) {
 330                 DEBUG(DEBUG_ERR,("db(%s) unhealty in ctdb_control_update_record: %s\n",
 331                                  ctdb_db->db_name, ctdb_db->unhealthy_reason));
 332                 return -1;
 333         }
 334
 335         state = talloc(ctdb, struct ctdb_persistent_write_state);
 336         CTDB_NO_MEMORY(ctdb, state);
 337
 338         state->ctdb_db = ctdb_db;
 339         state->c       = c;
 340         state->m       = m;
 341         state->flags   = 0;
 342         if (ctdb_db_volatile(ctdb_db)) {
 343                 state->flags   = UPDATE_FLAGS_REPLACE_ONLY;
 344         }
 345
 346         /* create a child process to take out a transaction and
 347            write the data.
 348         */
 349         handle = ctdb_childwrite(ctdb_db, ctdb_persistent_write_callback, state);
 350         if (handle == NULL) {
 351                 DEBUG(DEBUG_ERR,("Failed to setup childwrite handler in ctdb_control_update_record\n"));
 352                 talloc_free(state);
 353                 return -1;
 354         }
 355
 356         /* we need to wait for the replies */
 357         *async_reply = true;
 358
 359         /* need to keep the control structure around */
 360         talloc_steal(state, c);
 361
 362         /* but we won't wait forever */
 363         tevent_add_timer(ctdb->ev, state,
 364                          timeval_current_ofs(ctdb->tunable.control_timeout, 0),
 365                          ctdb_persistent_lock_timeout, state);
 366
 367         return 0;
 368 }
 369