/*
* __txn_log_recover --
* Roll the log forward to recover committed changes.
*/
static int
__txn_log_recover(WT_SESSION_IMPL *session,
WT_ITEM *logrec, WT_LSN *lsnp, WT_LSN *next_lsnp,
void *cookie, int firstrecord)
{
WT_RECOVERY *r;
const uint8_t *end, *p;
uint64_t txnid;
uint32_t rectype;
WT_UNUSED(next_lsnp);
r = cookie;
p = WT_LOG_SKIP_HEADER(logrec->data);
end = (const uint8_t *)logrec->data + logrec->size;
WT_UNUSED(firstrecord);
/* First, peek at the log record type. */
WT_RET(__wt_logrec_read(session, &p, end, &rectype));
switch (rectype) {
case WT_LOGREC_CHECKPOINT:
if (r->metadata_only)
WT_RET(__wt_txn_checkpoint_logread(
session, &p, end, &r->ckpt_lsn));
break;
case WT_LOGREC_COMMIT:
WT_RET(__wt_vunpack_uint(&p, WT_PTRDIFF(end, p), &txnid));
WT_UNUSED(txnid);
WT_RET(__txn_commit_apply(r, lsnp, &p, end));
break;
}
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;
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);
}
/*
* __txn_log_recover --
* Roll the log forward to recover committed changes.
*/
static int
__txn_log_recover(WT_SESSION_IMPL *session,
WT_ITEM *logrec, WT_LSN *lsnp, WT_LSN *next_lsnp,
void *cookie, int firstrecord)
{
WT_DECL_RET;
WT_RECOVERY *r;
uint64_t txnid_unused;
uint32_t rectype;
const uint8_t *end, *p;
r = cookie;
p = WT_LOG_SKIP_HEADER(logrec->data);
end = (const uint8_t *)logrec->data + logrec->size;
WT_UNUSED(firstrecord);
/* First, peek at the log record type. */
WT_RET(__wt_logrec_read(session, &p, end, &rectype));
/*
* Record the highest LSN we process during the metadata phase.
* If not the metadata phase, then stop at that LSN.
*/
if (r->metadata_only)
r->max_rec_lsn = *next_lsnp;
else if (__wt_log_cmp(lsnp, &r->max_rec_lsn) >= 0)
return (0);
switch (rectype) {
case WT_LOGREC_CHECKPOINT:
if (r->metadata_only)
WT_RET(__wt_txn_checkpoint_logread(
session, &p, end, &r->ckpt_lsn));
break;
case WT_LOGREC_COMMIT:
if ((ret = __wt_vunpack_uint(
&p, WT_PTRDIFF(end, p), &txnid_unused)) != 0)
WT_RET_MSG(
session, ret, "txn_log_recover: unpack failure");
WT_RET(__txn_commit_apply(r, lsnp, &p, end));
break;
}
return (0);
}
/*
* __curlog_logrec --
* Callback function from log_scan to get a log record.
*/
static int
__curlog_logrec(WT_SESSION_IMPL *session,
WT_ITEM *logrec, WT_LSN *lsnp, WT_LSN *next_lsnp,
void *cookie, int firstrecord)
{
WT_CURSOR_LOG *cl;
cl = cookie;
WT_UNUSED(firstrecord);
/* Set up the LSNs and take a copy of the log record for the cursor. */
*cl->cur_lsn = *lsnp;
*cl->next_lsn = *next_lsnp;
WT_RET(__wt_buf_set(session, cl->logrec, logrec->data, logrec->size));
/*
* Read the log header. Set up the step pointers to walk the
* operations inside the record. Get the record type.
*/
cl->stepp = WT_LOG_SKIP_HEADER(cl->logrec->data);
cl->stepp_end = (uint8_t *)cl->logrec->data + logrec->size;
WT_RET(__wt_logrec_read(session, &cl->stepp, cl->stepp_end,
&cl->rectype));
/* A step count of 0 means the entire record. */
cl->step_count = 0;
/*
* Unpack the txnid so that we can return each
* individual operation for this txnid.
*/
if (cl->rectype == WT_LOGREC_COMMIT)
WT_RET(__wt_vunpack_uint(&cl->stepp,
WT_PTRDIFF(cl->stepp_end, cl->stepp), &cl->txnid));
else {
/*
* Step over anything else.
* Setting stepp to NULL causes the next()
* method to read a new record on the next call.
*/
cl->stepp = NULL;
cl->txnid = 0;
}
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;
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_printlog --
* Print a log record in a human-readable format.
*/
static int
__txn_printlog(WT_SESSION_IMPL *session,
WT_ITEM *rawrec, WT_LSN *lsnp, WT_LSN *next_lsnp,
void *cookie, int firstrecord)
{
WT_LOG_RECORD *logrec;
WT_TXN_PRINTLOG_ARGS *args;
const uint8_t *end, *p;
const char *msg;
uint64_t txnid;
uint32_t fileid, lsnfile, lsnoffset, rectype;
int32_t start;
bool compressed;
WT_UNUSED(next_lsnp);
args = cookie;
p = WT_LOG_SKIP_HEADER(rawrec->data);
end = (const uint8_t *)rawrec->data + rawrec->size;
logrec = (WT_LOG_RECORD *)rawrec->data;
compressed = F_ISSET(logrec, WT_LOG_RECORD_COMPRESSED);
/* First, peek at the log record type. */
WT_RET(__wt_logrec_read(session, &p, end, &rectype));
if (!firstrecord)
WT_RET(__wt_fprintf(session, WT_STDOUT(session), ",\n"));
WT_RET(__wt_fprintf(session, WT_STDOUT(session),
" { \"lsn\" : [%" PRIu32 ",%" PRIu32 "],\n",
lsnp->l.file, lsnp->l.offset));
WT_RET(__wt_fprintf(session, WT_STDOUT(session),
" \"hdr_flags\" : \"%s\",\n", compressed ? "compressed" : ""));
WT_RET(__wt_fprintf(session, WT_STDOUT(session),
" \"rec_len\" : %" PRIu32 ",\n", logrec->len));
WT_RET(__wt_fprintf(session, WT_STDOUT(session),
" \"mem_len\" : %" PRIu32 ",\n",
compressed ? logrec->mem_len : logrec->len));
switch (rectype) {
case WT_LOGREC_CHECKPOINT:
WT_RET(__wt_struct_unpack(session, p, WT_PTRDIFF(end, p),
WT_UNCHECKED_STRING(II), &lsnfile, &lsnoffset));
WT_RET(__wt_fprintf(session, WT_STDOUT(session),
" \"type\" : \"checkpoint\",\n"));
WT_RET(__wt_fprintf(session, WT_STDOUT(session),
" \"ckpt_lsn\" : [%" PRIu32 ",%" PRIu32 "]\n",
lsnfile, lsnoffset));
break;
case WT_LOGREC_COMMIT:
WT_RET(__wt_vunpack_uint(&p, WT_PTRDIFF(end, p), &txnid));
WT_RET(__wt_fprintf(session, WT_STDOUT(session),
" \"type\" : \"commit\",\n"));
WT_RET(__wt_fprintf(session, WT_STDOUT(session),
" \"txnid\" : %" PRIu64 ",\n", txnid));
WT_RET(__txn_oplist_printlog(session, &p, end, args->flags));
break;
case WT_LOGREC_FILE_SYNC:
WT_RET(__wt_struct_unpack(session, p, WT_PTRDIFF(end, p),
WT_UNCHECKED_STRING(Ii), &fileid, &start));
WT_RET(__wt_fprintf(session, WT_STDOUT(session),
" \"type\" : \"file_sync\",\n"));
WT_RET(__wt_fprintf(session, WT_STDOUT(session),
" \"fileid\" : %" PRIu32 ",\n", fileid));
WT_RET(__wt_fprintf(session, WT_STDOUT(session),
" \"start\" : %" PRId32 "\n", start));
break;
case WT_LOGREC_MESSAGE:
WT_RET(__wt_struct_unpack(session, p, WT_PTRDIFF(end, p),
WT_UNCHECKED_STRING(S), &msg));
WT_RET(__wt_fprintf(session, WT_STDOUT(session),
" \"type\" : \"message\",\n"));
WT_RET(__wt_fprintf(session, WT_STDOUT(session),
" \"message\" : \"%s\"\n", msg));
break;
case WT_LOGREC_SYSTEM:
WT_RET(__wt_struct_unpack(session, p, WT_PTRDIFF(end, p),
WT_UNCHECKED_STRING(II), &lsnfile, &lsnoffset));
WT_RET(__wt_fprintf(session, WT_STDOUT(session),
" \"type\" : \"system\",\n"));
WT_RET(__txn_oplist_printlog(session, &p, end, args->flags));
break;
}
WT_RET(__wt_fprintf(session, WT_STDOUT(session), " }"));
return (0);
}