/*
* __curlog_kv --
* Set the key and value of the log cursor to return to the user.
*/
static int
__curlog_kv(WT_SESSION_IMPL *session, WT_CURSOR *cursor)
{
WT_CURSOR_LOG *cl;
WT_DECL_RET;
uint32_t fileid, key_count, opsize, optype, raw;
cl = (WT_CURSOR_LOG *)cursor;
/* Temporarily turn off raw so we can do direct cursor operations. */
raw = F_MASK(cursor, WT_CURSTD_RAW);
F_CLR(cursor, WT_CURSTD_RAW);
/*
* If it is a commit and we have stepped over the header, peek to get
* the size and optype and read out any key/value from this operation.
*/
if ((key_count = cl->step_count++) > 0) {
WT_ERR(__wt_logop_read(session,
&cl->stepp, cl->stepp_end, &optype, &opsize));
WT_ERR(__curlog_op_read(session, cl, optype, opsize, &fileid));
/* Position on the beginning of the next record part. */
cl->stepp += opsize;
} else {
optype = WT_LOGOP_INVALID;
fileid = 0;
cl->opkey->data = NULL;
cl->opkey->size = 0;
/*
* Non-commit records we want to return the record without the
* header and the adjusted size. Add one to skip over the type
* which is normally consumed by __wt_logrec_read.
*/
cl->opvalue->data = WT_LOG_SKIP_HEADER(cl->logrec->data) + 1;
cl->opvalue->size = WT_LOG_REC_SIZE(cl->logrec->size) - 1;
}
/*
* The log cursor sets the LSN and step count as the cursor key and
* and log record related data in the value. The data in the value
* contains any operation key/value that was in the log record.
*/
__wt_cursor_set_key(cursor, cl->cur_lsn->l.file, cl->cur_lsn->l.offset,
key_count);
__wt_cursor_set_value(cursor, cl->txnid, cl->rectype, optype, fileid,
cl->opkey, cl->opvalue);
err: F_SET(cursor, raw);
return (ret);
}
int
__wt_txn_op_printlog(
WT_SESSION_IMPL *session, const uint8_t **pp, const uint8_t *end, FILE *out)
{
uint32_t optype, opsize;
/* Peek at the size and the type. */
WT_RET(__wt_logop_read(session, pp, end, &optype, &opsize));
end = *pp + opsize;
switch (optype) {
case WT_LOGOP_COL_PUT:
WT_RET(__wt_logop_col_put_print(session, pp, end, out));
break;
case WT_LOGOP_COL_REMOVE:
WT_RET(__wt_logop_col_remove_print(session, pp, end, out));
break;
case WT_LOGOP_COL_TRUNCATE:
WT_RET(__wt_logop_col_truncate_print(session, pp, end, out));
break;
case WT_LOGOP_ROW_PUT:
WT_RET(__wt_logop_row_put_print(session, pp, end, out));
break;
case WT_LOGOP_ROW_REMOVE:
WT_RET(__wt_logop_row_remove_print(session, pp, end, out));
break;
case WT_LOGOP_ROW_TRUNCATE:
WT_RET(__wt_logop_row_truncate_print(session, pp, end, out));
break;
WT_ILLEGAL_VALUE(session);
}
return (0);
}
/*
* __curlog_kv --
* Set the key and value of the log cursor to return to the user.
*/
static int
__curlog_kv(WT_SESSION_IMPL *session, WT_CURSOR *cursor)
{
WT_CURSOR_LOG *cl;
uint32_t fileid, key_count, opsize, optype;
cl = (WT_CURSOR_LOG *)cursor;
/*
* If it is a commit and we have stepped over the header, peek to get
* the size and optype and read out any key/value from this operation.
*/
if ((key_count = cl->step_count++) > 0) {
WT_RET(__wt_logop_read(session,
&cl->stepp, cl->stepp_end, &optype, &opsize));
WT_RET(__curlog_op_read(session, cl, optype, opsize, &fileid));
/* Position on the beginning of the next record part. */
cl->stepp += opsize;
} else {
optype = WT_LOGOP_INVALID;
fileid = 0;
cl->opkey->data = NULL;
cl->opkey->size = 0;
cl->opvalue->data = cl->logrec->data;
cl->opvalue->size = cl->logrec->size;
}
/*
* The log cursor sets the LSN and step count as the cursor key and
* and log record related data in the value. The data in the value
* contains any operation key/value that was in the log record.
*/
__wt_cursor_set_key(cursor, cl->cur_lsn->file, cl->cur_lsn->offset,
key_count);
__wt_cursor_set_value(cursor, cl->txnid, cl->rectype, optype,
fileid, cl->opkey, cl->opvalue);
return (0);
}
/*
* __txn_op_apply --
* Apply a transactional operation during recovery.
*/
static int
__txn_op_apply(
WT_RECOVERY *r, WT_LSN *lsnp, const uint8_t **pp, const uint8_t *end)
{
WT_CURSOR *cursor, *start, *stop;
WT_DECL_RET;
WT_ITEM key, start_key, stop_key, value;
WT_SESSION_IMPL *session;
uint64_t recno, start_recno, stop_recno;
uint32_t fileid, mode, optype, opsize;
session = r->session;
cursor = NULL;
/* Peek at the size and the type. */
WT_ERR(__wt_logop_read(session, pp, end, &optype, &opsize));
end = *pp + opsize;
switch (optype) {
case WT_LOGOP_COL_PUT:
WT_ERR(__wt_logop_col_put_unpack(session, pp, end,
&fileid, &recno, &value));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
cursor->set_key(cursor, recno);
__wt_cursor_set_raw_value(cursor, &value);
WT_ERR(cursor->insert(cursor));
break;
case WT_LOGOP_COL_REMOVE:
WT_ERR(__wt_logop_col_remove_unpack(session, pp, end,
&fileid, &recno));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
cursor->set_key(cursor, recno);
WT_ERR(cursor->remove(cursor));
break;
case WT_LOGOP_COL_TRUNCATE:
WT_ERR(__wt_logop_col_truncate_unpack(session, pp, end,
&fileid, &start_recno, &stop_recno));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
/* Set up the cursors. */
if (start_recno == WT_RECNO_OOB) {
start = NULL;
stop = cursor;
} else if (stop_recno == WT_RECNO_OOB) {
start = cursor;
stop = NULL;
} else {
start = cursor;
WT_ERR(__recovery_cursor(
session, r, lsnp, fileid, true, &stop));
}
/* Set the keys. */
if (start != NULL)
start->set_key(start, start_recno);
if (stop != NULL)
stop->set_key(stop, stop_recno);
WT_TRET(session->iface.truncate(&session->iface, NULL,
start, stop, NULL));
/* If we opened a duplicate cursor, close it now. */
if (stop != NULL && stop != cursor)
WT_TRET(stop->close(stop));
WT_ERR(ret);
break;
case WT_LOGOP_ROW_PUT:
WT_ERR(__wt_logop_row_put_unpack(session, pp, end,
&fileid, &key, &value));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
__wt_cursor_set_raw_key(cursor, &key);
__wt_cursor_set_raw_value(cursor, &value);
WT_ERR(cursor->insert(cursor));
break;
case WT_LOGOP_ROW_REMOVE:
WT_ERR(__wt_logop_row_remove_unpack(session, pp, end,
&fileid, &key));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
__wt_cursor_set_raw_key(cursor, &key);
WT_ERR(cursor->remove(cursor));
break;
case WT_LOGOP_ROW_TRUNCATE:
WT_ERR(__wt_logop_row_truncate_unpack(session, pp, end,
&fileid, &start_key, &stop_key, &mode));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
/* Set up the cursors. */
start = stop = NULL;
switch (mode) {
case WT_TXN_TRUNC_ALL:
/* Both cursors stay NULL. */
break;
case WT_TXN_TRUNC_BOTH:
start = cursor;
WT_ERR(__recovery_cursor(
session, r, lsnp, fileid, true, &stop));
break;
case WT_TXN_TRUNC_START:
start = cursor;
break;
case WT_TXN_TRUNC_STOP:
stop = cursor;
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/* Set the keys. */
if (start != NULL)
__wt_cursor_set_raw_key(start, &start_key);
if (stop != NULL)
__wt_cursor_set_raw_key(stop, &stop_key);
WT_TRET(session->iface.truncate(&session->iface, NULL,
start, stop, NULL));
/* If we opened a duplicate cursor, close it now. */
if (stop != NULL && stop != cursor)
WT_TRET(stop->close(stop));
WT_ERR(ret);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/* Reset the cursor so it doesn't block eviction. */
if (cursor != NULL)
WT_ERR(cursor->reset(cursor));
err: if (ret != 0)
__wt_err(session, ret, "Operation failed during recovery");
return (ret);
}
/*
* __curlog_kv --
* Set the key and value of the log cursor to return to the user.
*/
static int
__curlog_kv(WT_SESSION_IMPL *session, WT_CURSOR *cursor)
{
WT_CURSOR_LOG *cl;
WT_ITEM item;
uint32_t fileid, key_count, opsize, optype;
cl = (WT_CURSOR_LOG *)cursor;
/*
* If it is a commit and we have stepped over the header, peek to get
* the size and optype and read out any key/value from this operation.
*/
if ((key_count = cl->step_count++) > 0) {
WT_RET(__wt_logop_read(session,
&cl->stepp, cl->stepp_end, &optype, &opsize));
WT_RET(__curlog_op_read(session, cl, optype, opsize, &fileid));
/* Position on the beginning of the next record part. */
cl->stepp += opsize;
} else {
optype = WT_LOGOP_INVALID;
fileid = 0;
cl->opkey->data = NULL;
cl->opkey->size = 0;
/*
* Non-commit records we want to return the record without the
* header and the adjusted size. Add one to skip over the type
* which is normally consumed by __wt_logrec_read.
*/
cl->opvalue->data = WT_LOG_SKIP_HEADER(cl->logrec->data) + 1;
cl->opvalue->size = WT_LOG_REC_SIZE(cl->logrec->size) - 1;
}
/*
* The log cursor sets the LSN and step count as the cursor key and
* and log record related data in the value. The data in the value
* contains any operation key/value that was in the log record.
* For the special case that the caller needs the result in raw form,
* we create packed versions of the key/value.
*/
if (FLD_ISSET(cursor->flags, WT_CURSTD_RAW)) {
memset(&item, 0, sizeof(item));
WT_RET(wiredtiger_struct_size((WT_SESSION *)session,
&item.size, WT_LOGC_KEY_FORMAT, cl->cur_lsn->l.file,
cl->cur_lsn->l.offset, key_count));
WT_RET(__wt_realloc(session, NULL, item.size, &cl->packed_key));
item.data = cl->packed_key;
WT_RET(wiredtiger_struct_pack((WT_SESSION *)session,
cl->packed_key, item.size, WT_LOGC_KEY_FORMAT,
cl->cur_lsn->l.file, cl->cur_lsn->l.offset, key_count));
__wt_cursor_set_key(cursor, &item);
WT_RET(wiredtiger_struct_size((WT_SESSION *)session,
&item.size, WT_LOGC_VALUE_FORMAT, cl->txnid, cl->rectype,
optype, fileid, cl->opkey, cl->opvalue));
WT_RET(__wt_realloc(session, NULL, item.size,
&cl->packed_value));
item.data = cl->packed_value;
WT_RET(wiredtiger_struct_pack((WT_SESSION *)session,
cl->packed_value, item.size, WT_LOGC_VALUE_FORMAT,
cl->txnid, cl->rectype, optype, fileid, cl->opkey,
cl->opvalue));
__wt_cursor_set_value(cursor, &item);
} else {
__wt_cursor_set_key(cursor, cl->cur_lsn->l.file,
cl->cur_lsn->l.offset, key_count);
__wt_cursor_set_value(cursor, cl->txnid, cl->rectype, optype,
fileid, cl->opkey, cl->opvalue);
}
return (0);
}
/*
* __txn_op_apply --
* Apply a transactional operation during recovery.
*/
static int
__txn_op_apply(
WT_RECOVERY *r, WT_LSN *lsnp, const uint8_t **pp, const uint8_t *end)
{
WT_CURSOR *cursor, *start, *stop;
WT_DECL_RET;
WT_ITEM key, start_key, stop_key, value;
WT_SESSION_IMPL *session;
uint64_t recno, start_recno, stop_recno;
uint32_t fileid, mode, optype, opsize;
session = r->session;
cursor = NULL;
/* Peek at the size and the type. */
WT_ERR(__wt_logop_read(session, pp, end, &optype, &opsize));
end = *pp + opsize;
switch (optype) {
case WT_LOGOP_COL_MODIFY:
WT_ERR(__wt_logop_col_modify_unpack(session, pp, end,
&fileid, &recno, &value));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
cursor->set_key(cursor, recno);
if ((ret = cursor->search(cursor)) != 0)
WT_ERR_NOTFOUND_OK(ret);
else {
/*
* Build/insert a complete value during recovery rather
* than using cursor modify to create a partial update
* (for no particular reason than simplicity).
*/
WT_ERR(__wt_modify_apply(session, cursor, value.data));
WT_ERR(cursor->insert(cursor));
}
break;
case WT_LOGOP_COL_PUT:
WT_ERR(__wt_logop_col_put_unpack(session, pp, end,
&fileid, &recno, &value));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
cursor->set_key(cursor, recno);
__wt_cursor_set_raw_value(cursor, &value);
WT_ERR(cursor->insert(cursor));
break;
case WT_LOGOP_COL_REMOVE:
WT_ERR(__wt_logop_col_remove_unpack(session, pp, end,
&fileid, &recno));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
cursor->set_key(cursor, recno);
WT_ERR(cursor->remove(cursor));
break;
case WT_LOGOP_COL_TRUNCATE:
WT_ERR(__wt_logop_col_truncate_unpack(session, pp, end,
&fileid, &start_recno, &stop_recno));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
/* Set up the cursors. */
if (start_recno == WT_RECNO_OOB) {
start = NULL;
stop = cursor;
} else if (stop_recno == WT_RECNO_OOB) {
start = cursor;
stop = NULL;
} else {
start = cursor;
WT_ERR(__recovery_cursor(
session, r, lsnp, fileid, true, &stop));
}
/* Set the keys. */
if (start != NULL)
start->set_key(start, start_recno);
if (stop != NULL)
stop->set_key(stop, stop_recno);
WT_TRET(session->iface.truncate(&session->iface, NULL,
start, stop, NULL));
/* If we opened a duplicate cursor, close it now. */
if (stop != NULL && stop != cursor)
WT_TRET(stop->close(stop));
WT_ERR(ret);
break;
case WT_LOGOP_ROW_MODIFY:
WT_ERR(__wt_logop_row_modify_unpack(session, pp, end,
&fileid, &key, &value));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
__wt_cursor_set_raw_key(cursor, &key);
if ((ret = cursor->search(cursor)) != 0)
WT_ERR_NOTFOUND_OK(ret);
else {
/*
* Build/insert a complete value during recovery rather
* than using cursor modify to create a partial update
* (for no particular reason than simplicity).
*/
WT_ERR(__wt_modify_apply(session, cursor, value.data));
WT_ERR(cursor->insert(cursor));
}
break;
case WT_LOGOP_ROW_PUT:
WT_ERR(__wt_logop_row_put_unpack(session, pp, end,
&fileid, &key, &value));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
__wt_cursor_set_raw_key(cursor, &key);
__wt_cursor_set_raw_value(cursor, &value);
WT_ERR(cursor->insert(cursor));
break;
case WT_LOGOP_ROW_REMOVE:
WT_ERR(__wt_logop_row_remove_unpack(session, pp, end,
&fileid, &key));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
__wt_cursor_set_raw_key(cursor, &key);
WT_ERR(cursor->remove(cursor));
break;
case WT_LOGOP_ROW_TRUNCATE:
WT_ERR(__wt_logop_row_truncate_unpack(session, pp, end,
&fileid, &start_key, &stop_key, &mode));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
/* Set up the cursors. */
start = stop = NULL;
switch (mode) {
case WT_TXN_TRUNC_ALL:
/* Both cursors stay NULL. */
break;
case WT_TXN_TRUNC_BOTH:
start = cursor;
WT_ERR(__recovery_cursor(
session, r, lsnp, fileid, true, &stop));
break;
case WT_TXN_TRUNC_START:
start = cursor;
break;
case WT_TXN_TRUNC_STOP:
stop = cursor;
break;
WT_ILLEGAL_VALUE_ERR(session, mode);
}
/* Set the keys. */
if (start != NULL)
__wt_cursor_set_raw_key(start, &start_key);
if (stop != NULL)
__wt_cursor_set_raw_key(stop, &stop_key);
WT_TRET(session->iface.truncate(&session->iface, NULL,
start, stop, NULL));
/* If we opened a duplicate cursor, close it now. */
if (stop != NULL && stop != cursor)
WT_TRET(stop->close(stop));
WT_ERR(ret);
break;
WT_ILLEGAL_VALUE_ERR(session, optype);
}
/* Reset the cursor so it doesn't block eviction. */
if (cursor != NULL)
WT_ERR(cursor->reset(cursor));
return (0);
err: __wt_err(session, ret,
"operation apply failed during recovery: operation type %"
PRIu32 " at LSN %" PRIu32 "/%" PRIu32,
optype, lsnp->l.file, lsnp->l.offset);
return (ret);
}
