*/
#include "includes.h"
-#include "lib/events/events.h"
+#include "lib/tevent/tevent.h"
#include "lib/tdb/include/tdb.h"
#include "system/network.h"
#include "system/filesys.h"
#include "../include/ctdb_private.h"
#include "db_wrap.h"
#include "lib/util/dlinklist.h"
-#include "lib/events/events.h"
+#include "lib/tevent/tevent.h"
#include "../include/ctdb_private.h"
#include "../common/rb_tree.h"
enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
struct ctdb_vacuum_child_context {
+ struct ctdb_vacuum_child_context *next, *prev;
struct ctdb_vacuum_handle *vacuum_handle;
+ /* fd child writes status to */
int fd[2];
pid_t child_pid;
enum vacuum_child_status status;
struct ctdb_vacuum_handle {
struct ctdb_db_context *ctdb_db;
struct ctdb_vacuum_child_context *child_ctx;
+ uint32_t fast_path_count;
};
uint32_t total;
uint32_t vacuumed;
uint32_t copied;
+ uint32_t fast_added_to_vacuum_fetch_list;
+ uint32_t fast_added_to_delete_tree;
+ uint32_t fast_deleted;
+ uint32_t fast_skipped;
+ uint32_t fast_error;
+ uint32_t fast_total;
+ uint32_t full_added_to_vacuum_fetch_list;
+ uint32_t full_added_to_delete_tree;
+ uint32_t full_skipped;
+ uint32_t full_error;
+ uint32_t full_total;
};
/* tuning information stored for every db */
struct ctdb_marshall_buffer *records;
};
-static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
- struct timeval t, void *private_data);
+/**
+ * Store key and header in a tree, indexed by the key hash.
+ */
+static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
+ struct ctdb_db_context *ctdb_db,
+ trbt_tree_t *tree,
+ const struct ctdb_ltdb_header *hdr,
+ TDB_DATA key)
+{
+ struct delete_record_data *dd;
+ uint32_t hash;
+ dd = talloc_zero(tree, struct delete_record_data);
+ if (dd == NULL) {
+ DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
+ return -1;
+ }
-/*
- * traverse function for gathering the records that can be deleted
- */
-static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
+ dd->ctdb = ctdb;
+ dd->ctdb_db = ctdb_db;
+ dd->key.dsize = key.dsize;
+ dd->key.dptr = talloc_memdup(dd, key.dptr, key.dsize);
+ if (dd->key.dptr == NULL) {
+ DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
+ return -1;
+ }
+
+ dd->hdr = *hdr;
+
+ hash = ctdb_hash(&key);
+
+ trbt_insert32(tree, hash, dd);
+
+ return 0;
+}
+
+static int add_record_to_delete_tree(struct vacuum_data *vdata, TDB_DATA key,
+ struct ctdb_ltdb_header *hdr)
{
- struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
struct ctdb_context *ctdb = vdata->ctdb;
struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
- uint32_t lmaster;
- struct ctdb_ltdb_header *hdr;
- struct ctdb_rec_data *rec;
- size_t old_size;
-
- lmaster = ctdb_lmaster(ctdb, &key);
- if (lmaster >= ctdb->vnn_map->size) {
- return 0;
- }
+ uint32_t hash;
+ int ret;
- if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
- /* its not a deleted record */
+ hash = ctdb_hash(&key);
+
+ if (trbt_lookup32(vdata->delete_tree, hash)) {
+ DEBUG(DEBUG_INFO, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
return 0;
}
- hdr = (struct ctdb_ltdb_header *)data.dptr;
-
- if (hdr->dmaster != ctdb->pnn) {
- return 0;
+ ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
+ vdata->delete_tree,
+ hdr, key);
+ if (ret != 0) {
+ return -1;
}
- /* is this a records we could possibly delete? I.e.
- if the record is empty and also we are both lmaster
- and dmaster for the record we should be able to delete it
- */
- if (lmaster == ctdb->pnn) {
- uint32_t hash;
-
- hash = ctdb_hash(&key);
- if (trbt_lookup32(vdata->delete_tree, hash)) {
- DEBUG(DEBUG_INFO, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
- }
- else {
- struct delete_record_data *dd;
-
- /* store key and header indexed by the key hash */
- dd = talloc_zero(vdata->delete_tree, struct delete_record_data);
- if (dd == NULL) {
- DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
- return -1;
- }
- dd->ctdb = ctdb;
- dd->ctdb_db = ctdb_db;
- dd->key.dsize = key.dsize;
- dd->key.dptr = talloc_memdup(dd, key.dptr, key.dsize);
- if (dd->key.dptr == NULL) {
- DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
- return -1;
- }
+ vdata->delete_count++;
- dd->hdr = *hdr;
-
- trbt_insert32(vdata->delete_tree, hash, dd);
+ return 0;
+}
- vdata->delete_count++;
- }
- }
+/**
+ * Add a record to the list of records to be sent
+ * to their lmaster with VACUUM_FETCH.
+ */
+static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
+ TDB_DATA key)
+{
+ struct ctdb_context *ctdb = vdata->ctdb;
+ struct ctdb_rec_data *rec;
+ uint32_t lmaster;
+ size_t old_size;
+
+ lmaster = ctdb_lmaster(ctdb, &key);
- /* add the record to the blob ready to send to the nodes */
rec = ctdb_marshall_record(vdata->list[lmaster], ctdb->pnn, key, NULL, tdb_null);
if (rec == NULL) {
DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
vdata->traverse_error = true;
return -1;
}
+
old_size = talloc_get_size(vdata->list[lmaster]);
- vdata->list[lmaster] = talloc_realloc_size(NULL, vdata->list[lmaster],
+ vdata->list[lmaster] = talloc_realloc_size(NULL, vdata->list[lmaster],
old_size + rec->length);
if (vdata->list[lmaster] == NULL) {
DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
vdata->traverse_error = true;
return -1;
}
+
vdata->list[lmaster]->count++;
memcpy(old_size+(uint8_t *)vdata->list[lmaster], rec, rec->length);
talloc_free(rec);
return 0;
}
+
+static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
+ struct timeval t, void *private_data);
+
+
+/*
+ * traverse function for gathering the records that can be deleted
+ */
+static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
+{
+ struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
+ struct ctdb_context *ctdb = vdata->ctdb;
+ uint32_t lmaster;
+ struct ctdb_ltdb_header *hdr;
+ int res = 0;
+
+ vdata->full_total++;
+
+ lmaster = ctdb_lmaster(ctdb, &key);
+ if (lmaster >= ctdb->num_nodes) {
+ vdata->full_error++;
+ DEBUG(DEBUG_CRIT, (__location__
+ " lmaster[%u] >= ctdb->num_nodes[%u] for key"
+ " with hash[%u]!\n",
+ (unsigned)lmaster,
+ (unsigned)ctdb->num_nodes,
+ (unsigned)ctdb_hash(&key)));
+ return -1;
+ }
+
+ if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
+ /* its not a deleted record */
+ vdata->full_skipped++;
+ return 0;
+ }
+
+ hdr = (struct ctdb_ltdb_header *)data.dptr;
+
+ if (hdr->dmaster != ctdb->pnn) {
+ vdata->full_skipped++;
+ return 0;
+ }
+
+ if (lmaster == ctdb->pnn) {
+ /*
+ * We are both lmaster and dmaster, and the record * is empty.
+ * So we should be able to delete it.
+ */
+ res = add_record_to_delete_tree(vdata, key, hdr);
+ if (res != 0) {
+ vdata->full_error++;
+ } else {
+ vdata->full_added_to_delete_tree++;
+ }
+ } else {
+ /*
+ * We are not lmaster.
+ * Add the record to the blob ready to send to the nodes.
+ */
+ res = add_record_to_vacuum_fetch_list(vdata, key);
+ if (res != 0) {
+ vdata->full_error++;
+ } else {
+ vdata->full_added_to_vacuum_fetch_list++;
+ }
+ }
+
+ return res;
+}
+
/*
* traverse the tree of records to delete and marshall them into
* a blob
memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
}
+/**
+ * traverse function for the traversal of the delete_queue,
+ * the fast-path vacuuming list.
+ *
+ * - If the record has been migrated off the node
+ * or has been revived (filled with data) on the node,
+ * then skip the record.
+ *
+ * - If the current node is the record's lmaster and it is
+ * a record that has never been migrated with data, then
+ * delete the record from the local tdb.
+ *
+ * - If the current node is the record's lmaster and it has
+ * been migrated with data, then schedule it for the normal
+ * vacuuming procedure (i.e. add it to the delete_list).
+ *
+ * - If the current node is NOT the record's lmaster then
+ * add it to the list of records that are to be sent to
+ * the lmaster with the VACUUM_FETCH message.
+ */
+static void delete_queue_traverse(void *param, void *data)
+{
+ struct delete_record_data *dd =
+ talloc_get_type(data, struct delete_record_data);
+ struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
+ struct ctdb_db_context *ctdb_db = dd->ctdb_db;
+ struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
+ int res;
+ struct ctdb_ltdb_header *header;
+ TDB_DATA tdb_data;
+ uint32_t lmaster;
+
+ vdata->fast_total++;
+
+ res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
+ if (res != 0) {
+ DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
+ vdata->fast_error++;
+ return;
+ }
+
+ tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
+ if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
+ /* Does not exist or not a ctdb record. Skip. */
+ goto skipped;
+ }
+
+ if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
+ /* The record has been recycled (filled with data). Skip. */
+ goto skipped;
+ }
+
+ header = (struct ctdb_ltdb_header *)tdb_data.dptr;
+
+ if (header->dmaster != ctdb->pnn) {
+ /* The record has been migrated off the node. Skip. */
+ goto skipped;
+ }
+
+
+ if (header->rsn != dd->hdr.rsn) {
+ /*
+ * The record has been migrated off the node and back again.
+ * But not requeued for deletion. Skip it.
+ */
+ goto skipped;
+ }
+
+ /*
+ * We are dmaster, and the record has no data, and it has
+ * not been migrated after it has been queued for deletion.
+ *
+ * At this stage, the record could still have been revived locally
+ * and last been written with empty data. This can only be
+ * fixed with the addition of an active or delete flag. (TODO)
+ */
+
+ lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
+
+ if (lmaster != ctdb->pnn) {
+ res = add_record_to_vacuum_fetch_list(vdata, dd->key);
+
+ if (res != 0) {
+ DEBUG(DEBUG_ERR,
+ (__location__ " Error adding record to list "
+ "of records to send to lmaster.\n"));
+ vdata->fast_error++;
+ } else {
+ vdata->fast_added_to_vacuum_fetch_list++;
+ }
+ goto done;
+ }
+
+ /* use header->flags or dd->hdr.flags ?? */
+ if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
+ res = add_record_to_delete_tree(vdata, dd->key, &dd->hdr);
+
+ if (res != 0) {
+ DEBUG(DEBUG_ERR,
+ (__location__ " Error adding record to list "
+ "of records for deletion on lmaster.\n"));
+ vdata->fast_error++;
+ } else {
+ vdata->fast_added_to_delete_tree++;
+ }
+ } else {
+ res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
+
+ if (res != 0) {
+ DEBUG(DEBUG_ERR,
+ (__location__ " Error deleting record from local "
+ "data base.\n"));
+ vdata->fast_error++;
+ } else {
+ vdata->fast_deleted++;
+ }
+ }
+
+ goto done;
+
+skipped:
+ vdata->fast_skipped++;
+
+done:
+ if (tdb_data.dptr != NULL) {
+ free(tdb_data.dptr);
+ }
+ tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
+
+ return;
+}
+
/*
* read-only traverse the database in order to find
* records that can be deleted and try to delete these
* records on the other nodes
* this executes in the child context
*/
-static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db, struct vacuum_data *vdata)
+static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
+ struct vacuum_data *vdata,
+ bool full_vacuum_run)
{
struct ctdb_context *ctdb = ctdb_db->ctdb;
const char *name = ctdb_db->db_name;
int ret, i, pnn;
+ DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
+ "%s db_id[0x%08x]\n",
+ full_vacuum_run ? "full" : "fast",
+ ctdb_db->db_name, ctdb_db->db_id));
+
ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
if (ret != 0) {
DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
}
ctdb->pnn = pnn;
+
+ vdata->fast_added_to_delete_tree = 0;
+ vdata->fast_added_to_vacuum_fetch_list = 0;
+ vdata->fast_deleted = 0;
+ vdata->fast_skipped = 0;
+ vdata->fast_error = 0;
+ vdata->fast_total = 0;
+ vdata->full_added_to_delete_tree = 0;
+ vdata->full_added_to_vacuum_fetch_list = 0;
+ vdata->full_skipped = 0;
+ vdata->full_error = 0;
+ vdata->full_total = 0;
+
/* the list needs to be of length num_nodes */
- vdata->list = talloc_array(vdata, struct ctdb_marshall_buffer *, ctdb->vnn_map->size);
+ vdata->list = talloc_array(vdata, struct ctdb_marshall_buffer *, ctdb->num_nodes);
if (vdata->list == NULL) {
DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
return -1;
}
- for (i = 0; i < ctdb->vnn_map->size; i++) {
+ for (i = 0; i < ctdb->num_nodes; i++) {
vdata->list[i] = (struct ctdb_marshall_buffer *)
talloc_zero_size(vdata->list,
offsetof(struct ctdb_marshall_buffer, data));
vdata->list[i]->db_id = ctdb_db->db_id;
}
- /* read-only traverse, looking for records that might be able to be vacuumed */
- if (tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata) == -1 ||
- vdata->traverse_error) {
- DEBUG(DEBUG_ERR,(__location__ " Traverse error in vacuuming '%s'\n", name));
- return -1;
+ /*
+ * Traverse the delete_queue.
+ * This builds the same lists as the db traverse.
+ */
+ trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
+
+ if (vdata->fast_total > 0) {
+ DEBUG(DEBUG_INFO,
+ (__location__
+ " fast vacuuming delete_queue traverse statistics: "
+ "db[%s] "
+ "total[%u] "
+ "del[%u] "
+ "skp[%u] "
+ "err[%u] "
+ "adt[%u] "
+ "avf[%u]\n",
+ ctdb_db->db_name,
+ (unsigned)vdata->fast_total,
+ (unsigned)vdata->fast_deleted,
+ (unsigned)vdata->fast_skipped,
+ (unsigned)vdata->fast_error,
+ (unsigned)vdata->fast_added_to_delete_tree,
+ (unsigned)vdata->fast_added_to_vacuum_fetch_list));
}
- for ( i = 0; i < ctdb->vnn_map->size; i++) {
- if (vdata->list[i]->count == 0) {
+ /*
+ * read-only traverse of the database, looking for records that
+ * might be able to be vacuumed.
+ *
+ * This is not done each time but only every tunable
+ * VacuumFastPathCount times.
+ */
+ if (full_vacuum_run) {
+ ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
+ if (ret == -1 || vdata->traverse_error) {
+ DEBUG(DEBUG_ERR,(__location__ " Traverse error in vacuuming '%s'\n", name));
+ return -1;
+ }
+ if (vdata->full_total > 0) {
+ DEBUG(DEBUG_INFO,
+ (__location__
+ " full vacuuming db traverse statistics: "
+ "db[%s] "
+ "total[%u] "
+ "skp[%u] "
+ "err[%u] "
+ "adt[%u] "
+ "avf[%u]\n",
+ ctdb_db->db_name,
+ (unsigned)vdata->full_total,
+ (unsigned)vdata->full_skipped,
+ (unsigned)vdata->full_error,
+ (unsigned)vdata->full_added_to_delete_tree,
+ (unsigned)vdata->full_added_to_vacuum_fetch_list));
+ }
+ }
+
+ /*
+ * For records where we are not the lmaster,
+ * tell the lmaster to fetch the record.
+ */
+ for (i = 0; i < ctdb->num_nodes; i++) {
+ TDB_DATA data;
+
+ if (ctdb->nodes[i]->pnn == ctdb->pnn) {
continue;
}
- /* for records where we are not the lmaster, tell the lmaster to fetch the record */
- if (ctdb->vnn_map->map[i] != ctdb->pnn) {
- TDB_DATA data;
- DEBUG(DEBUG_NOTICE,("Found %u records for lmaster %u in '%s'\n",
- vdata->list[i]->count, i, name));
-
- data.dsize = talloc_get_size(vdata->list[i]);
- data.dptr = (void *)vdata->list[i];
- if (ctdb_send_message(ctdb, ctdb->vnn_map->map[i], CTDB_SRVID_VACUUM_FETCH, data) != 0) {
- DEBUG(DEBUG_ERR,(__location__ " Failed to send vacuum fetch message to %u\n",
- ctdb->vnn_map->map[i]));
- return -1;
- }
+ if (vdata->list[i]->count == 0) {
continue;
}
+
+ DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
+ vdata->list[i]->count, ctdb->nodes[i]->pnn,
+ name));
+
+ data.dsize = talloc_get_size(vdata->list[i]);
+ data.dptr = (void *)vdata->list[i];
+ if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn, CTDB_SRVID_VACUUM_FETCH, data) != 0) {
+ DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
+ "fetch message to %u\n",
+ ctdb->nodes[i]->pnn));
+ return -1;
+ }
}
/* Process all records we can delete (if any) */
struct delete_records_list *recs;
TDB_DATA indata, outdata;
int32_t res;
+ struct ctdb_node_map *nodemap;
+ uint32_t *active_nodes;
+ int num_active_nodes;
recs = talloc_zero(vdata, struct delete_records_list);
if (recs == NULL) {
indata.dptr = (void *)recs->records;
/*
- * now tell all the other nodes to delete all these records
+ * now tell all the active nodes to delete all these records
* (if possible)
*/
- for (i = 0; i < ctdb->vnn_map->size; i++) {
+
+ ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
+ CTDB_CURRENT_NODE,
+ recs, /* talloc context */
+ &nodemap);
+ if (ret != 0) {
+ DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
+ return -1;
+ }
+
+ active_nodes = list_of_active_nodes(ctdb, nodemap,
+ nodemap, /* talloc context */
+ false /* include self */);
+ /* yuck! ;-) */
+ num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
+
+ for (i = 0; i < num_active_nodes; i++) {
struct ctdb_marshall_buffer *records;
struct ctdb_rec_data *rec;
- if (ctdb->vnn_map->map[i] == ctdb->pnn) {
- /* we dont delete the records on the local node just yet */
- continue;
- }
-
- ret = ctdb_control(ctdb, ctdb->vnn_map->map[i], 0,
+ ret = ctdb_control(ctdb, active_nodes[i], 0,
CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
indata, recs, &outdata, &res,
NULL, NULL);
if (ret != 0 || res != 0) {
- DEBUG(DEBUG_ERR,("Failed to delete records on node %u\n", ctdb->vnn_map->map[i]));
+ DEBUG(DEBUG_ERR, ("Failed to delete records on "
+ "node %u: ret[%d] res[%d]\n",
+ active_nodes[i], ret, res));
return -1;
}
}
}
+ /* free nodemap and active_nodes */
+ talloc_free(nodemap);
+
/*
* The only records remaining in the tree would be those
* records where all other nodes could successfully
* there might be hash collisions so we have to compare the keys here to be sure
*/
if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
- vdata->vacuumed++;
- return 0;
+ struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
+ /*
+ * we have to check if the record hasn't changed in the meantime in order to
+ * savely remove it from the database
+ */
+ if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
+ hdr->dmaster == kd->ctdb->pnn &&
+ ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
+ kd->hdr.rsn == hdr->rsn) {
+ vdata->vacuumed++;
+ return 0;
+ }
}
}
if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
return -1;
}
- tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb), TDB_INTERNAL, O_RDWR|O_CREAT, 0);
+ tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
+ TDB_INTERNAL|TDB_DISALLOW_NESTING,
+ O_RDWR|O_CREAT, 0);
if (tmp_db == NULL) {
DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
tdb_transaction_cancel(tdb);
return -1;
}
- DEBUG(DEBUG_NOTICE,(__location__ " %u records vacuumed\n", vdata->vacuumed));
+ DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
if (vdata->traverse_error) {
DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
return -1;
}
- DEBUG(DEBUG_NOTICE,(__location__ " %u records copied\n", vdata->copied));
+ DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
return 0;
}
struct vacuum_tuning_data tdata;
struct vacuum_tuning_data *tptr;
char *vac_dbname;
+ int flags;
vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u",
- ctdb_db->ctdb->db_directory,
- TUNINGDBNAME, ctdb_db->ctdb->pnn);
+ ctdb_db->ctdb->db_directory_state,
+ TUNINGDBNAME, ctdb_db->ctdb->pnn);
if (vac_dbname == NULL) {
DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
talloc_free(tmp_ctx);
return -1;
}
- tune_tdb = tdb_open(vac_dbname, 0, 0, O_RDWR|O_CREAT, 0600);
+ flags = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
+ flags |= TDB_DISALLOW_NESTING;
+ tune_tdb = tdb_open(vac_dbname, 0,
+ flags,
+ O_RDWR|O_CREAT, 0600);
if (tune_tdb == NULL) {
DEBUG(DEBUG_ERR,(__location__ " Failed to create/open %s\n", TUNINGDBNAME));
talloc_free(tmp_ctx);
vdata->delete_count < vdata->vacuum_limit) {
if (tdata.last_interval < ctdb_db->ctdb->tunable.vacuum_max_interval) {
tdata.new_interval = tdata.last_interval * 110 / 100;
- DEBUG(DEBUG_NOTICE,("Increasing vacuum interval %u -> %u for %s\n",
+ DEBUG(DEBUG_INFO,("Increasing vacuum interval %u -> %u for %s\n",
tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
}
} else {
tdata.new_interval > ctdb_db->ctdb->tunable.vacuum_max_interval) {
tdata.new_interval = ctdb_db->ctdb->tunable.vacuum_min_interval;
}
- DEBUG(DEBUG_ERR,("Decreasing vacuum interval %u -> %u for %s\n",
+ DEBUG(DEBUG_INFO,("Decreasing vacuum interval %u -> %u for %s\n",
tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
}
tdata.last_interval = tdata.new_interval;
} else {
- DEBUG(DEBUG_ERR,(__location__ " Cannot find tunedb record for %s. Using default interval\n", ctdb_db->db_name));
+ DEBUG(DEBUG_DEBUG,(__location__ " Cannot find tunedb record for %s. Using default interval\n", ctdb_db->db_name));
tdata.last_num_repack = freelist;
tdata.last_num_empty = vdata->delete_count;
tdata.last_interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
* repack and vaccum a db
* called from the child context
*/
-static int ctdb_repack_db(struct ctdb_db_context *ctdb_db, TALLOC_CTX *mem_ctx)
+static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
+ TALLOC_CTX *mem_ctx,
+ bool full_vacuum_run)
{
uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
vdata->vacuum_limit = vacuum_limit;
vdata->repack_limit = repack_limit;
vdata->delete_tree = trbt_create(vdata, 0);
+ vdata->ctdb_db = ctdb_db;
if (vdata->delete_tree == NULL) {
DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
talloc_free(vdata);
/*
* gather all records that can be deleted in vdata
*/
- if (ctdb_vacuum_db(ctdb_db, vdata) != 0) {
+ if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
}
* decide if a repack is necessary
*/
if (size < repack_limit && vdata->delete_count < vacuum_limit) {
- talloc_free(vdata);
update_tuning_db(ctdb_db, vdata, size);
+ talloc_free(vdata);
return 0;
}
- DEBUG(DEBUG_NOTICE,("Repacking %s with %u freelist entries and %u records to delete\n",
+ DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
name, size, vdata->delete_count));
/*
char *vac_dbname;
uint interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
struct ctdb_context *ctdb = ctdb_db->ctdb;
+ int flags;
vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u", ctdb->db_directory, TUNINGDBNAME, ctdb->pnn);
if (vac_dbname == NULL) {
return interval;
}
- tdb = tdb_open(vac_dbname, 0, 0, O_RDONLY, 0600);
+ flags = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
+ flags |= TDB_DISALLOW_NESTING;
+ tdb = tdb_open(vac_dbname, 0,
+ flags,
+ O_RDWR|O_CREAT, 0600);
if (!tdb) {
- DEBUG(DEBUG_ERR,("Unable to open database %s using default interval\n", vac_dbname));
+ DEBUG(DEBUG_ERR,("Unable to open/create database %s using default interval. Errno : %s (%d)\n", vac_dbname, strerror(errno), errno));
talloc_free(tmp_ctx);
return interval;
}
}
}
free(value.dptr);
-
- DEBUG(DEBUG_NOTICE,("Using new interval %u for database %s\n", interval, ctdb_db->db_name));
}
tdb_close(tdb);
struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
struct ctdb_context *ctdb = ctdb_db->ctdb;
- DEBUG(DEBUG_ERR,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
+ DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
if (child_ctx->child_pid != -1) {
kill(child_ctx->child_pid, SIGKILL);
+ } else {
+ /* Bump the number of successful fast-path runs. */
+ child_ctx->vacuum_handle->fast_path_count++;
}
+ DLIST_REMOVE(ctdb->vacuumers, child_ctx);
+
event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
ctdb_vacuum_event, child_ctx->vacuum_handle);
char c = 0;
int ret;
- DEBUG(DEBUG_NOTICE,("Vacuuming child finished for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
+ DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
child_ctx->child_pid = -1;
ret = read(child_ctx->fd[0], &c, 1);
struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
struct ctdb_context *ctdb = ctdb_db->ctdb;
struct ctdb_vacuum_child_context *child_ctx;
+ struct tevent_fd *fde;
int ret;
- /* we dont vacuum if we are in recovery mode */
- if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE) {
+ /* we dont vacuum if we are in recovery mode, or db frozen */
+ if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
+ ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
+ DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
+ ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
+ : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
+ ? "freeze pending"
+ : "frozen"));
event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
return;
}
- DEBUG(DEBUG_NOTICE,("Start a vacuuming child process for db %s\n", ctdb_db->db_name));
-
child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
if (child_ctx == NULL) {
DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
return;
}
- child_ctx->child_pid = fork();
+ if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
+ vacuum_handle->fast_path_count = 0;
+ }
+
+ child_ctx->child_pid = ctdb_fork(ctdb);
if (child_ctx->child_pid == (pid_t)-1) {
close(child_ctx->fd[0]);
close(child_ctx->fd[1]);
if (child_ctx->child_pid == 0) {
char cc = 0;
+ bool full_vacuum_run = false;
close(child_ctx->fd[0]);
- if (switch_from_server_to_client(ctdb) != 0) {
+ DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
+
+ if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
_exit(1);
}
/*
* repack the db
*/
- cc = ctdb_repack_db(ctdb_db, child_ctx);
+ if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
+ (vacuum_handle->fast_path_count == 0))
+ {
+ full_vacuum_run = true;
+ }
+ cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
+ full_vacuum_run);
write(child_ctx->fd[1], &cc, 1);
_exit(0);
child_ctx->status = VACUUM_RUNNING;
child_ctx->start_time = timeval_current();
+ DLIST_ADD(ctdb->vacuumers, child_ctx);
talloc_set_destructor(child_ctx, vacuum_child_destructor);
+ /*
+ * Clear the fastpath vacuuming list in the parent.
+ */
+ talloc_free(ctdb_db->delete_queue);
+ ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
+ if (ctdb_db->delete_queue == NULL) {
+ /* fatal here? ... */
+ ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
+ "in parent context. Shutting down\n");
+ }
+
event_add_timed(ctdb->ev, child_ctx,
timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
vacuum_child_timeout, child_ctx);
- event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
- EVENT_FD_READ|EVENT_FD_AUTOCLOSE,
- vacuum_child_handler,
- child_ctx);
+ DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
+
+ fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
+ EVENT_FD_READ, vacuum_child_handler, child_ctx);
+ tevent_fd_set_auto_close(fde);
vacuum_handle->child_ctx = child_ctx;
child_ctx->vacuum_handle = vacuum_handle;
}
+void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
+{
+ /* Simply free them all. */
+ while (ctdb->vacuumers) {
+ DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
+ ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
+ (int)ctdb->vacuumers->child_pid));
+ /* vacuum_child_destructor kills it, removes from list */
+ talloc_free(ctdb->vacuumers);
+ }
+}
/* this function initializes the vacuuming context for a database
* starts the vacuuming events
*/
int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
{
+ if (ctdb_db->persistent != 0) {
+ DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
+ return 0;
+ }
+
ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
return 0;
}
+
+/**
+ * Insert a record into the ctdb_db context's delete queue,
+ * handling hash collisions.
+ */
+static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
+ const struct ctdb_ltdb_header *hdr,
+ TDB_DATA key)
+{
+ struct delete_record_data *kd;
+ uint32_t hash;
+ int ret;
+
+ hash = (uint32_t)ctdb_hash(&key);
+
+ DEBUG(DEBUG_INFO, (__location__ " Schedule for deletion: db[%s] "
+ "db_id[0x%08x] "
+ "key_hash[0x%08x] "
+ "lmaster[%u] "
+ "migrated_with_data[%s]\n",
+ ctdb_db->db_name, ctdb_db->db_id,
+ hash,
+ ctdb_lmaster(ctdb_db->ctdb, &key),
+ hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
+
+ kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
+ if (kd != NULL) {
+ if ((kd->key.dsize != key.dsize) ||
+ (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
+ {
+ DEBUG(DEBUG_INFO,
+ ("schedule for deletion: Hash collision (0x%08x)."
+ " Skipping the record.\n", hash));
+ return 0;
+ } else {
+ DEBUG(DEBUG_DEBUG,
+ ("schedule for deletetion: Overwriting entry for "
+ "key with hash 0x%08x.\n", hash));
+ }
+ }
+
+ ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
+ ctdb_db->delete_queue,
+ hdr, key);
+ if (ret != 0) {
+ return -1;
+ }
+
+ return 0;
+}
+
+/**
+ * Schedule a record for deletetion.
+ * Called from the parent context.
+ */
+int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
+ TDB_DATA indata)
+{
+ struct ctdb_control_schedule_for_deletion *dd;
+ struct ctdb_db_context *ctdb_db;
+ int ret;
+ TDB_DATA key;
+
+ dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
+
+ ctdb_db = find_ctdb_db(ctdb, dd->db_id);
+ if (ctdb_db == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
+ dd->db_id));
+ return -1;
+ }
+
+ key.dsize = dd->keylen;
+ key.dptr = dd->key;
+
+ ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
+
+ return ret;
+}
+
+int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
+ const struct ctdb_ltdb_header *hdr,
+ TDB_DATA key)
+{
+ int ret;
+ struct ctdb_control_schedule_for_deletion *dd;
+ TDB_DATA indata;
+ int32_t status;
+
+ if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
+ /* main daemon - directly queue */
+ ret = insert_delete_record_data_into_tree(ctdb_db->ctdb,
+ ctdb_db,
+ ctdb_db->delete_queue,
+ hdr, key);
+ if (ret != 0) {
+ return -1;
+ }
+ return 0;
+ }
+
+ /* child process: send the main daemon a control */
+
+ indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
+ indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
+ if (indata.dptr == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
+ return -1;
+ }
+ dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
+ dd->db_id = ctdb_db->db_id;
+ dd->hdr = *hdr;
+ dd->keylen = key.dsize;
+ memcpy(dd->key, key.dptr, key.dsize);
+
+ ret = ctdb_control(ctdb_db->ctdb,
+ CTDB_CURRENT_NODE,
+ ctdb_db->db_id,
+ CTDB_CONTROL_SCHEDULE_FOR_DELETION,
+ CTDB_CTRL_FLAG_NOREPLY, /* flags */
+ indata,
+ NULL, /* mem_ctx */
+ NULL, /* outdata */
+ &status,
+ NULL, /* timeout : NULL == wait forever */
+ NULL); /* error message */
+
+ talloc_free(indata.dptr);
+
+ if (ret != 0 || status != 0) {
+ DEBUG(DEBUG_ERR, (__location__ " Error sending "
+ "SCHEDULE_FOR_DELETION "
+ "control.\n"));
+ if (status != 0) {
+ ret = -1;
+ }
+ }
+
+ return ret;
+}