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         unsigned int i;
  57         int ret;
  58         struct ctdb_rec_data_old *rec = NULL;
  59         struct ctdb_marshall_buffer *m = state->m;
  60
  61         ret = tdb_transaction_start(state->ctdb_db->ltdb->tdb);
  62         if (ret == -1) {
  63                 DEBUG(DEBUG_ERR,("Failed to start transaction for db_id 0x%08x in ctdb_persistent_store\n",
  64                                  state->ctdb_db->db_id));
  65                 return -1;
  66         }
  67
  68         for (i=0;i<m->count;i++) {
  69                 struct ctdb_ltdb_header oldheader;
  70                 struct ctdb_ltdb_header header;
  71                 TDB_DATA key, data, olddata;
  72                 TALLOC_CTX *tmp_ctx = talloc_new(state);
  73
  74                 rec = ctdb_marshall_loop_next(m, rec, NULL, &header, &key, &data);
  75
  76                 if (rec == NULL) {
  77                         D_ERR("Failed to get next record %u for db_id 0x%08x "
  78                               "in ctdb_persistent_store\n",
  79                               i,
  80                               state->ctdb_db->db_id);
  81                         talloc_free(tmp_ctx);
  82                         goto failed;
  83                 }
  84
  85                 /* we must check if the record exists or not because
  86                    ctdb_ltdb_fetch will unconditionally create a record
  87                  */
  88                 if (state->flags & UPDATE_FLAGS_REPLACE_ONLY) {
  89                         TDB_DATA trec;
  90                         trec = tdb_fetch(state->ctdb_db->ltdb->tdb, key);
  91                         if (trec.dsize == 0) {
  92                                 talloc_free(tmp_ctx);
  93                                 continue;
  94                         }
  95                         free(trec.dptr);
  96                 }
  97
  98                 /* fetch the old header and ensure the rsn is less than the new rsn */
  99                 ret = ctdb_ltdb_fetch(state->ctdb_db, key, &oldheader, tmp_ctx, &olddata);
 100                 if (ret != 0) {
 101                         DEBUG(DEBUG_ERR,("Failed to fetch old record for db_id 0x%08x in ctdb_persistent_store\n",
 102                                          state->ctdb_db->db_id));
 103                         talloc_free(tmp_ctx);
 104                         goto failed;
 105                 }
 106
 107                 if (oldheader.rsn >= header.rsn &&
 108                     (olddata.dsize != data.dsize ||
 109                      memcmp(olddata.dptr, data.dptr, data.dsize) != 0)) {
 110                         DEBUG(DEBUG_CRIT,("existing header for db_id 0x%08x has larger RSN %llu than new RSN %llu in ctdb_persistent_store\n",
 111                                           state->ctdb_db->db_id,
 112                                           (unsigned long long)oldheader.rsn, (unsigned long long)header.rsn));
 113                         talloc_free(tmp_ctx);
 114                         goto failed;
 115                 }
 116
 117                 talloc_free(tmp_ctx);
 118
 119                 ret = ctdb_ltdb_store(state->ctdb_db, key, &header, data);
 120                 if (ret != 0) {
 121                         DEBUG(DEBUG_CRIT,("Failed to store record for db_id 0x%08x in ctdb_persistent_store\n",
 122                                           state->ctdb_db->db_id));
 123                         goto failed;
 124                 }
 125         }
 126
 127         ret = tdb_transaction_commit(state->ctdb_db->ltdb->tdb);
 128         if (ret == -1) {
 129                 DEBUG(DEBUG_ERR,("Failed to commit transaction for db_id 0x%08x in ctdb_persistent_store\n",
 130                                  state->ctdb_db->db_id));
 131                 return -1;
 132         }
 133
 134         return 0;
 135
 136 failed:
 137         tdb_transaction_cancel(state->ctdb_db->ltdb->tdb);
 138         return -1;
 139 }
 140
 141
 142 /*
 143   called when we the child has completed the persistent write
 144   on our behalf
 145  */
 146 static void ctdb_persistent_write_callback(int status, void *private_data)
 147 {
 148         struct ctdb_persistent_write_state *state = talloc_get_type(private_data,
 149                                                                    struct ctdb_persistent_write_state);
 150
 151
 152         ctdb_request_control_reply(state->ctdb_db->ctdb, state->c, NULL, status, NULL);
 153
 154         talloc_free(state);
 155 }
 156
 157 /*
 158   called if our lockwait child times out
 159  */
 160 static void ctdb_persistent_lock_timeout(struct tevent_context *ev,
 161                                          struct tevent_timer *te,
 162                                          struct timeval t, void *private_data)
 163 {
 164         struct ctdb_persistent_write_state *state = talloc_get_type(private_data,
 165                                                                    struct ctdb_persistent_write_state);
 166         ctdb_request_control_reply(state->ctdb_db->ctdb, state->c, NULL, -1, "timeout in ctdb_persistent_lock");
 167         talloc_free(state);
 168 }
 169
 170 struct childwrite_handle {
 171         struct ctdb_context *ctdb;
 172         struct ctdb_db_context *ctdb_db;
 173         struct tevent_fd *fde;
 174         int fd[2];
 175         pid_t child;
 176         void *private_data;
 177         void (*callback)(int, void *);
 178         struct timeval start_time;
 179 };
 180
 181 static int childwrite_destructor(struct childwrite_handle *h)
 182 {
 183         CTDB_DECREMENT_STAT(h->ctdb, pending_childwrite_calls);
 184         ctdb_kill(h->ctdb, h->child, SIGKILL);
 185         return 0;
 186 }
 187
 188 /* called when the child process has finished writing the record to the
 189    database
 190 */
 191 static void childwrite_handler(struct tevent_context *ev,
 192                                struct tevent_fd *fde,
 193                                uint16_t flags, void *private_data)
 194 {
 195         struct childwrite_handle *h = talloc_get_type(private_data,
 196                                                      struct childwrite_handle);
 197         void *p = h->private_data;
 198         void (*callback)(int, void *) = h->callback;
 199         pid_t child = h->child;
 200         TALLOC_CTX *tmp_ctx = talloc_new(ev);
 201         int ret;
 202         char c;
 203
 204         CTDB_UPDATE_LATENCY(h->ctdb, h->ctdb_db, "persistent", childwrite_latency, h->start_time);
 205         CTDB_DECREMENT_STAT(h->ctdb, pending_childwrite_calls);
 206
 207         /* the handle needs to go away when the context is gone - when
 208            the handle goes away this implicitly closes the pipe, which
 209            kills the child */
 210         talloc_steal(tmp_ctx, h);
 211
 212         talloc_set_destructor(h, NULL);
 213
 214         ret = sys_read(h->fd[0], &c, 1);
 215         if (ret < 1) {
 216                 DEBUG(DEBUG_ERR, (__location__ " Read returned %d. Childwrite failed\n", ret));
 217                 c = 1;
 218         }
 219
 220         callback(c, p);
 221
 222         ctdb_kill(h->ctdb, child, SIGKILL);
 223         talloc_free(tmp_ctx);
 224 }
 225
 226 /* this creates a child process which will take out a tdb transaction
 227    and write the record to the database.
 228 */
 229 static struct childwrite_handle *ctdb_childwrite(
 230                                 struct ctdb_db_context *ctdb_db,
 231                                 void (*callback)(int, void *private_data),
 232                                 struct ctdb_persistent_write_state *state)
 233 {
 234         struct childwrite_handle *result;
 235         int ret;
 236         pid_t parent = getpid();
 237
 238         CTDB_INCREMENT_STAT(ctdb_db->ctdb, childwrite_calls);
 239         CTDB_INCREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 240
 241         if (!(result = talloc_zero(state, struct childwrite_handle))) {
 242                 CTDB_DECREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 243                 return NULL;
 244         }
 245
 246         ret = pipe(result->fd);
 247
 248         if (ret != 0) {
 249                 talloc_free(result);
 250                 CTDB_DECREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 251                 return NULL;
 252         }
 253
 254         result->child = ctdb_fork(ctdb_db->ctdb);
 255
 256         if (result->child == (pid_t)-1) {
 257                 close(result->fd[0]);
 258                 close(result->fd[1]);
 259                 talloc_free(result);
 260                 CTDB_DECREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 261                 return NULL;
 262         }
 263
 264         result->callback = callback;
 265         result->private_data = state;
 266         result->ctdb = ctdb_db->ctdb;
 267         result->ctdb_db = ctdb_db;
 268
 269         if (result->child == 0) {
 270                 char c = 0;
 271
 272                 close(result->fd[0]);
 273                 prctl_set_comment("ctdb_write_persistent");
 274                 ret = ctdb_persistent_store(state);
 275                 if (ret != 0) {
 276                         DEBUG(DEBUG_ERR, (__location__ " Failed to write persistent data\n"));
 277                         c = 1;
 278                 }
 279
 280                 sys_write(result->fd[1], &c, 1);
 281
 282                 ctdb_wait_for_process_to_exit(parent);
 283                 _exit(0);
 284         }
 285
 286         close(result->fd[1]);
 287         set_close_on_exec(result->fd[0]);
 288
 289         talloc_set_destructor(result, childwrite_destructor);
 290
 291         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d for ctdb_childwrite\n", result->fd[0]));
 292
 293         result->fde = tevent_add_fd(ctdb_db->ctdb->ev, result, result->fd[0],
 294                                     TEVENT_FD_READ, childwrite_handler,
 295                                     (void *)result);
 296         if (result->fde == NULL) {
 297                 talloc_free(result);
 298                 CTDB_DECREMENT_STAT(ctdb_db->ctdb, pending_childwrite_calls);
 299                 return NULL;
 300         }
 301         tevent_fd_set_auto_close(result->fde);
 302
 303         result->start_time = timeval_current();
 304
 305         return result;
 306 }
 307
 308 /*
 309    update a record on this node if the new record has a higher rsn than the
 310    current record
 311  */
 312 int32_t ctdb_control_update_record(struct ctdb_context *ctdb,
 313                                    struct ctdb_req_control_old *c, TDB_DATA recdata,
 314                                    bool *async_reply)
 315 {
 316         struct ctdb_db_context *ctdb_db;
 317         struct ctdb_persistent_write_state *state;
 318         struct childwrite_handle *handle;
 319         struct ctdb_marshall_buffer *m = (struct ctdb_marshall_buffer *)recdata.dptr;
 320
 321         if (ctdb->recovery_mode != CTDB_RECOVERY_NORMAL) {
 322                 DEBUG(DEBUG_INFO,("rejecting ctdb_control_update_record when recovery active\n"));
 323                 return -1;
 324         }
 325
 326         ctdb_db = find_ctdb_db(ctdb, m->db_id);
 327         if (ctdb_db == NULL) {
 328                 DEBUG(DEBUG_ERR,("Unknown database 0x%08x in ctdb_control_update_record\n", m->db_id));
 329                 return -1;
 330         }
 331
 332         if (ctdb_db->unhealthy_reason) {
 333                 DEBUG(DEBUG_ERR,("db(%s) unhealty in ctdb_control_update_record: %s\n",
 334                                  ctdb_db->db_name, ctdb_db->unhealthy_reason));
 335                 return -1;
 336         }
 337
 338         state = talloc(ctdb, struct ctdb_persistent_write_state);
 339         CTDB_NO_MEMORY(ctdb, state);
 340
 341         state->ctdb_db = ctdb_db;
 342         state->c       = c;
 343         state->m       = m;
 344         state->flags   = 0;
 345         if (ctdb_db_volatile(ctdb_db)) {
 346                 state->flags   = UPDATE_FLAGS_REPLACE_ONLY;
 347         }
 348
 349         /* create a child process to take out a transaction and
 350            write the data.
 351         */
 352         handle = ctdb_childwrite(ctdb_db, ctdb_persistent_write_callback, state);
 353         if (handle == NULL) {
 354                 DEBUG(DEBUG_ERR,("Failed to setup childwrite handler in ctdb_control_update_record\n"));
 355                 talloc_free(state);
 356                 return -1;
 357         }
 358
 359         /* we need to wait for the replies */
 360         *async_reply = true;
 361
 362         /* need to keep the control structure around */
 363         talloc_steal(state, c);
 364
 365         /* but we won't wait forever */
 366         tevent_add_timer(ctdb->ev, state,
 367                          timeval_current_ofs(ctdb->tunable.control_timeout, 0),
 368                          ctdb_persistent_lock_timeout, state);
 369
 370         return 0;
 371 }
 372