Ejemplo n.º 1
0
/*
 * __wt_log_scan --
 *	Scan the logs, calling a function on each record found.
 */
int
__wt_log_scan(WT_SESSION_IMPL *session, WT_LSN *lsnp, uint32_t flags,
    int (*func)(WT_SESSION_IMPL *session,
    WT_ITEM *record, WT_LSN *lsnp, void *cookie), void *cookie)
{
	WT_CONNECTION_IMPL *conn;
	WT_ITEM buf;
	WT_DECL_RET;
	WT_FH *log_fh;
	WT_LOG *log;
	WT_LOG_RECORD *logrec;
	WT_LSN end_lsn, rd_lsn, start_lsn;
	off_t log_size;
	uint32_t allocsize, cksum, firstlog, lastlog, lognum, rdup_len, reclen;
	u_int i, logcount;
	int eol;
	char **logfiles;

	conn = S2C(session);
	log = conn->log;
	log_fh = NULL;
	logcount = 0;
	logfiles = NULL;
	eol = 0;
	WT_CLEAR(buf);

	/*
	 * If the caller did not give us a callback function there is nothing
	 * to do.
	 */
	if (func == NULL)
		return (0);

	if (LF_ISSET(WT_LOGSCAN_RECOVER))
		WT_RET(__wt_verbose(session, WT_VERB_LOG,
		    "__wt_log_scan truncating to %u/%" PRIuMAX,
		    log->trunc_lsn.file, (uintmax_t)log->trunc_lsn.offset));

	if (log != NULL) {
		allocsize = log->allocsize;

		if (lsnp == NULL) {
			if (LF_ISSET(WT_LOGSCAN_FIRST))
				start_lsn = log->first_lsn;
			else if (LF_ISSET(WT_LOGSCAN_FROM_CKP))
				start_lsn = log->ckpt_lsn;
			else
				return (WT_ERROR);	/* Illegal usage */
		} else {
			if (LF_ISSET(WT_LOGSCAN_FIRST|WT_LOGSCAN_FROM_CKP))
				WT_RET_MSG(session, WT_ERROR,
			    "choose either a start LSN or a start flag");

			/* Offsets must be on allocation boundaries. */
			if (lsnp->offset % allocsize != 0 ||
			    lsnp->file > log->fileid)
				return (WT_NOTFOUND);

			/*
			 * Log cursors may not know the starting LSN.  If an
			 * LSN pointer is passed in, but it is the INIT_LSN,
			 * start from the first_lsn.
			 */
			start_lsn = *lsnp;
			if (IS_INIT_LSN(&start_lsn))
				start_lsn = log->first_lsn;
		}
		end_lsn = log->alloc_lsn;
	} else {
		/*
		 * If logging is not configured, we can still print out the log
		 * if log files exist.  We just need to set the LSNs from what
		 * is in the files versus what is in the live connection.
		 */
		/*
		 * Set allocsize to the minimum alignment it could be.  Larger
		 * records and larger allocation boundaries should always be
		 * a multiple of this.
		 */
		allocsize = LOG_ALIGN;
		lastlog = 0;
		firstlog = UINT32_MAX;
		WT_RET(__wt_log_get_files(session, &logfiles, &logcount));
		if (logcount == 0)
			/*
			 * Return it is not supported if none don't exist.
			 */
			return (ENOTSUP);
		for (i = 0; i < logcount; i++) {
			WT_ERR(__wt_log_extract_lognum(session, logfiles[i],
			    &lognum));
			lastlog = WT_MAX(lastlog, lognum);
			firstlog = WT_MIN(firstlog, lognum);
		}
		start_lsn.file = firstlog;
		end_lsn.file = lastlog;
		start_lsn.offset = end_lsn.offset = 0;
		__wt_log_files_free(session, logfiles, logcount);
		logfiles = NULL;
	}
	WT_ERR(__log_openfile(session, 0, &log_fh, start_lsn.file));
	WT_ERR(__log_filesize(session, log_fh, &log_size));
	rd_lsn = start_lsn;
	WT_ERR(__wt_buf_initsize(session, &buf, LOG_ALIGN));
	for (;;) {
		if (rd_lsn.offset + allocsize > log_size) {
advance:
			/*
			 * If we read the last record, go to the next file.
			 */
			WT_ERR(__wt_close(session, log_fh));
			log_fh = NULL;
			eol = 1;
			/*
			 * Truncate this log file before we move to the next.
			 */
			if (LF_ISSET(WT_LOGSCAN_RECOVER))
				WT_ERR(__log_truncate(session, &rd_lsn, 1));
			rd_lsn.file++;
			rd_lsn.offset = 0;
			/*
			 * Avoid an error message when we reach end of log
			 * by checking here.
			 */
			if (rd_lsn.file > end_lsn.file)
				break;
			WT_ERR(__log_openfile(
			    session, 0, &log_fh, rd_lsn.file));
			WT_ERR(__log_filesize(session, log_fh, &log_size));
			continue;
		}
		/*
		 * Read the minimum allocation size a record could be.
		 */
		WT_ASSERT(session, buf.memsize >= allocsize);
		WT_ERR(__wt_read(session,
		    log_fh, rd_lsn.offset, (size_t)allocsize, buf.mem));
		/*
		 * First 8 bytes is the real record length.  See if we
		 * need to read more than the allocation size.  We expect
		 * that we rarely will have to read more.  Most log records
		 * will be fairly small.
		 */
		reclen = *(uint32_t *)buf.mem;
		/*
		 * Log files are pre-allocated.  We never expect a zero length
		 * unless we've reached the end of the log.  The log can be
		 * written out of order, so when recovery finds the end of
		 * the log, truncate the file and remove any later log files
		 * that may exist.
		 */
		if (reclen == 0) {
			/* This LSN is the end. */
			break;
		}
		rdup_len = __wt_rduppo2(reclen, allocsize);
		if (reclen > allocsize) {
			/*
			 * The log file end could be the middle of this
			 * log record.
			 */
			if (rd_lsn.offset + rdup_len > log_size)
				goto advance;
			/*
			 * We need to round up and read in the full padded
			 * record, especially for direct I/O.
			 */
			WT_ERR(__wt_buf_grow(session, &buf, rdup_len));
			WT_ERR(__wt_read(session,
			    log_fh, rd_lsn.offset, (size_t)rdup_len, buf.mem));
			WT_STAT_FAST_CONN_INCR(session, log_scan_rereads);
		}
		/*
		 * We read in the record, verify checksum.
		 */
		buf.size = reclen;
		logrec = (WT_LOG_RECORD *)buf.mem;
		cksum = logrec->checksum;
		logrec->checksum = 0;
		logrec->checksum = __wt_cksum(logrec, logrec->len);
		if (logrec->checksum != cksum) {
			/*
			 * A checksum mismatch means we have reached the end of
			 * the useful part of the log.  This should be found on
			 * the first pass through recovery.  In the second pass
			 * where we truncate the log, this is where it should
			 * end.
			 */
			if (log != NULL)
				log->trunc_lsn = rd_lsn;
			break;
		}

		/*
		 * We have a valid log record.  If it is not the log file
		 * header, invoke the callback.
		 */
		WT_STAT_FAST_CONN_INCR(session, log_scan_records);
		if (rd_lsn.offset != 0) {
			WT_ERR((*func)(session, &buf, &rd_lsn, cookie));
			if (LF_ISSET(WT_LOGSCAN_ONE))
				break;
		}
		rd_lsn.offset += (off_t)rdup_len;
	}

	/* Truncate if we're in recovery. */
	if (LF_ISSET(WT_LOGSCAN_RECOVER) &&
	    LOG_CMP(&rd_lsn, &log->trunc_lsn) < 0)
		WT_ERR(__log_truncate(session, &rd_lsn, 0));

err:	WT_STAT_FAST_CONN_INCR(session, log_scans);
	if (logfiles != NULL)
		__wt_log_files_free(session, logfiles, logcount);
	__wt_buf_free(session, &buf);
	/*
	 * If the caller wants one record and it is at the end of log,
	 * return WT_NOTFOUND.
	 */
	if (LF_ISSET(WT_LOGSCAN_ONE) && eol && ret == 0)
		ret = WT_NOTFOUND;
	if (ret == ENOENT)
		ret = 0;
	if (log_fh != NULL)
		WT_TRET(__wt_close(session, log_fh));
	return (ret);
}
Ejemplo n.º 2
0
/*
 * __wt_txn_recover --
 *	Run recovery.
 */
int
__wt_txn_recover(WT_CONNECTION_IMPL *conn)
{
	WT_CURSOR *metac;
	WT_DECL_RET;
	WT_RECOVERY r;
	WT_SESSION_IMPL *session;
	struct WT_RECOVERY_FILE *metafile;
	char *config;
	int was_backup;

	WT_CLEAR(r);
	INIT_LSN(&r.ckpt_lsn);
	was_backup = F_ISSET(conn, WT_CONN_WAS_BACKUP) ? 1 : 0;

	/* We need a real session for recovery. */
	WT_RET(__wt_open_session(conn, NULL, NULL, &session));
	F_SET(session, WT_SESSION_NO_LOGGING);
	r.session = session;

	WT_ERR(__wt_metadata_search(session, WT_METAFILE_URI, &config));
	WT_ERR(__recovery_setup_file(&r, WT_METAFILE_URI, config));
	WT_ERR(__wt_metadata_cursor(session, NULL, &metac));
	metafile = &r.files[WT_METAFILE_ID];
	metafile->c = metac;

	/*
	 * First, do a pass through the log to recover the metadata, and
	 * establish the last checkpoint LSN.  Skip this when opening a hot
	 * backup: we already have the correct metadata in that case.
	 */
	if (!was_backup) {
		r.metadata_only = 1;
		if (IS_INIT_LSN(&metafile->ckpt_lsn))
			WT_ERR(__wt_log_scan(session,
			    NULL, WT_LOGSCAN_FIRST, __txn_log_recover, &r));
		else
			WT_ERR(__wt_log_scan(session,
			    &metafile->ckpt_lsn, 0, __txn_log_recover, &r));

		WT_ASSERT(session,
		    LOG_CMP(&r.ckpt_lsn, &conn->log->first_lsn) >= 0);
	}

	/* Scan the metadata to find the live files and their IDs. */
	WT_ERR(__recovery_file_scan(&r));

	/*
	 * We no longer need the metadata cursor: close it to avoid pinning any
	 * resources that could block eviction during recovery.
	 */
	r.files[0].c = NULL;
	WT_ERR(metac->close(metac));

	/*
	 * Now, recover all the files apart from the metadata.
	 * Pass WT_LOGSCAN_RECOVER so that old logs get truncated.
	 */
	r.metadata_only = 0;
	WT_ERR(__wt_verbose(session, WT_VERB_RECOVERY,
	    "Main recovery loop: starting at %u/%" PRIuMAX,
	    r.ckpt_lsn.file, (uintmax_t)r.ckpt_lsn.offset));
	if (IS_INIT_LSN(&r.ckpt_lsn))
		WT_ERR(__wt_log_scan(session, NULL,
		    WT_LOGSCAN_FIRST | WT_LOGSCAN_RECOVER,
		    __txn_log_recover, &r));
	else
		WT_ERR(__wt_log_scan(session, &r.ckpt_lsn,
		    WT_LOGSCAN_RECOVER,
		    __txn_log_recover, &r));

	conn->next_file_id = r.max_fileid;

	/*
	 * If recovery ran successfully forcibly log a checkpoint so the next
	 * open is fast and keep the metadata up to date with the checkpoint
	 * LSN and archiving.
	 */
	WT_ERR(session->iface.checkpoint(&session->iface, "force=1"));

err:	WT_TRET(__recovery_free(&r));
	__wt_free(session, config);
	WT_TRET(session->iface.close(&session->iface, NULL));

	return (ret);
}