log_get.c

这是国外的resip协议栈

		for (p = dblp->bufp + (lp->b_off - lp->len);;) {
			memcpy(hdr, p, hdr->size);
			if (hdr->prev == lsn->offset) {
				b_region = (u_int32_t)(p - dblp->bufp);
				break;
			}
			p = dblp->bufp + (hdr->prev - lp->w_off);
		}

	/*
	 * If we don't have enough room for the record, we have to allocate
	 * space.  We have to do it while holding the region lock, which is
	 * truly annoying, but there's no way around it.  This call is why
	 * we allocate cursor buffer space when allocating the cursor instead
	 * of waiting.
	 */
	if (logc->bp_size <= b_region + b_disk) {
		len = (size_t)DB_ALIGN((b_region + b_disk) * 2, 128);
		if ((ret = __os_realloc(logc->dbenv, len, &logc->bp)) != 0)
			return (ret);
		logc->bp_size = (u_int32_t)len;
	}

	/* Copy the region's bytes to the end of the cursor's buffer. */
	p = (logc->bp + logc->bp_size) - b_region;
	memcpy(p, dblp->bufp, b_region);

	/* Release the region lock. */
	if (*rlockp == L_ACQUIRED) {
		*rlockp = L_NONE;
		R_UNLOCK(dbenv, &dblp->reginfo);
	}

	/*
	 * Read the rest of the information from disk.  Neither short reads
	 * or EOF are acceptable, the bytes we want had better be there.
	 */
	if (b_disk != 0) {
		p -= b_disk;
		nr = b_disk;
		if ((ret = __log_c_io(
		    logc, lsn->file, lsn->offset, p, &nr, NULL)) != 0)
			return (ret);
		if (nr < b_disk)
			return (__log_c_shortread(logc, lsn, 0));
	}

	/* Copy the header information into the caller's structure. */
	memcpy(hdr, p, hdr->size);

	*pp = p;
	return (0);
}

/*
 * __log_c_ondisk --
 *	Read a record off disk.
 */
static int
__log_c_ondisk(logc, lsn, last_lsn, flags, hdr, pp, eofp)
	DB_LOGC *logc;
	DB_LSN *lsn, *last_lsn;
	u_int32_t flags;
	int *eofp;
	HDR *hdr;
	u_int8_t **pp;
{
	DB_ENV *dbenv;
	size_t len, nr;
	u_int32_t offset;
	int ret;

	dbenv = logc->dbenv;
	*eofp = 0;

	nr = hdr->size;
	if ((ret =
	    __log_c_io(logc, lsn->file, lsn->offset, hdr, &nr, eofp)) != 0)
		return (ret);
	if (*eofp)
		return (0);

	/*
	 * If the read was successful, but we can't read a full header, assume
	 * we've hit EOF.  We can't check that the header has been partially
	 * zeroed out, but it's unlikely that this is caused by a write failure
	 * since the header is written as a single write call and it's less
	 * than sector.
	 */
	if (nr < hdr->size) {
		*eofp = 1;
		return (0);
	}

	/* Check the HDR. */
	if ((ret = __log_c_hdrchk(logc, lsn, hdr, eofp)) != 0)
		return (ret);
	if (*eofp)
		return (0);

	/*
	 * Regardless of how we return, the previous contents of the cursor's
	 * buffer are useless -- trash it.
	 */
	logc->bp_rlen = 0;

	/*
	 * Otherwise, we now (finally!) know how big the record is.  (Maybe
	 * we should have just stuck the length of the record into the LSN!?)
	 * Make sure we have enough space.
	 */
	if (logc->bp_size <= hdr->len) {
		len = (size_t)DB_ALIGN(hdr->len * 2, 128);
		if ((ret = __os_realloc(dbenv, len, &logc->bp)) != 0)
			return (ret);
		logc->bp_size = (u_int32_t)len;
	}

	/*
	 * If we're moving forward in the log file, read this record in at the
	 * beginning of the buffer.  Otherwise, read this record in at the end
	 * of the buffer, making sure we don't try and read before the start
	 * of the file.  (We prefer positioning at the end because transaction
	 * aborts use DB_SET to move backward through the log and we might get
	 * lucky.)
	 *
	 * Read a buffer's worth, without reading past the logical EOF.  The
	 * last_lsn may be a zero LSN, but that's OK, the test works anyway.
	 */
	if (flags == DB_FIRST || flags == DB_NEXT)
		offset = lsn->offset;
	else if (lsn->offset + hdr->len < logc->bp_size)
		offset = 0;
	else
		offset = (lsn->offset + hdr->len) - logc->bp_size;

	nr = logc->bp_size;
	if (lsn->file == last_lsn->file && offset + nr >= last_lsn->offset)
		nr = last_lsn->offset - offset;

	if ((ret =
	    __log_c_io(logc, lsn->file, offset, logc->bp, &nr, eofp)) != 0)
		return (ret);

	/*
	 * We should have at least gotten the bytes up-to-and-including the
	 * record we're reading.
	 */
	if (nr < (lsn->offset + hdr->len) - offset)
		return (__log_c_shortread(logc, lsn, 1));

	/*
	 * Set up the return information.
	 *
	 * !!!
	 * No need to set the bp_lsn.file field, __log_c_io set it for us.
	 */
	logc->bp_rlen = (u_int32_t)nr;
	logc->bp_lsn.offset = offset;

	*pp = logc->bp + (lsn->offset - offset);

	return (0);
}

/*
 * __log_c_hdrchk --
 *
 * Check for corrupted HDRs before we use them to allocate memory or find
 * records.
 *
 * If the log files were pre-allocated, a zero-filled HDR structure is the
 * logical file end.  However, we can see buffers filled with 0's during
 * recovery, too (because multiple log buffers were written asynchronously,
 * and one made it to disk before a different one that logically precedes
 * it in the log file.
 *
 * Check for impossibly large records.  The malloc should fail later, but we
 * have customers that run mallocs that treat all allocation failures as fatal
 * errors.
 *
 * Note that none of this is necessarily something awful happening.  We let
 * the application hand us any LSN they want, and it could be a pointer into
 * the middle of a log record, there's no way to tell.
 */
static int
__log_c_hdrchk(logc, lsn, hdr, eofp)
	DB_LOGC *logc;
	DB_LSN *lsn;
	HDR *hdr;
	int *eofp;
{
	DB_ENV *dbenv;
	int ret;

	dbenv = logc->dbenv;

	/*
	 * Check EOF before we do any other processing.
	 */
	if (eofp != NULL) {
		if (hdr->prev == 0 && hdr->chksum[0] == 0 && hdr->len == 0) {
			*eofp = 1;
			return (0);
		}
		*eofp = 0;
	}

	/*
	 * Sanity check the log record's size.
	 * We must check it after "virtual" EOF above.
	 */
	if (hdr->len <= hdr->size)
		goto err;

	/*
	 * If the cursor's max-record value isn't yet set, it means we aren't
	 * reading these records from a log file and no check is necessary.
	 */
	if (logc->bp_maxrec != 0 && hdr->len > logc->bp_maxrec) {
		/*
		 * If we fail the check, there's the pathological case that
		 * we're reading the last file, it's growing, and our initial
		 * check information was wrong.  Get it again, to be sure.
		 */
		if ((ret = __log_c_set_maxrec(logc, NULL)) != 0) {
			__db_err(dbenv, "DB_LOGC->get: %s", db_strerror(ret));
			return (ret);
		}
		if (logc->bp_maxrec != 0 && hdr->len > logc->bp_maxrec)
			goto err;
	}
	return (0);

err:	if (!F_ISSET(logc, DB_LOG_SILENT_ERR))
		__db_err(dbenv,
		    "DB_LOGC->get: LSN %lu/%lu: invalid log record header",
		    (u_long)lsn->file, (u_long)lsn->offset);
	return (EIO);
}

/*
 * __log_c_io --
 *	Read records from a log file.
 */
static int
__log_c_io(logc, fnum, offset, p, nrp, eofp)
	DB_LOGC *logc;
	u_int32_t fnum, offset;
	void *p;
	size_t *nrp;
	int *eofp;
{
	DB_ENV *dbenv;
	DB_LOG *dblp;
	int ret;
	char *np;

	dbenv = logc->dbenv;
	dblp = dbenv->lg_handle;

	/*
	 * If we've switched files, discard the current file handle and acquire
	 * a new one.
	 */
	if (logc->c_fhp != NULL && logc->bp_lsn.file != fnum) {
		ret = __os_closehandle(dbenv, logc->c_fhp);
		logc->c_fhp = NULL;
		logc->bp_lsn.file = 0;

		if (ret != 0)
			return (ret);
	}
	if (logc->c_fhp == NULL) {
		if ((ret = __log_name(dblp, fnum,
		    &np, &logc->c_fhp, DB_OSO_RDONLY | DB_OSO_SEQ)) != 0) {
			/*
			 * If we're allowed to return EOF, assume that's the
			 * problem, set the EOF status flag and return 0.
			 */
			if (eofp != NULL) {
				*eofp = 1;
				ret = 0;
			} else if (!F_ISSET(logc, DB_LOG_SILENT_ERR))
				__db_err(dbenv, "DB_LOGC->get: %s: %s",
				    np, db_strerror(ret));
			__os_free(dbenv, np);
			return (ret);
		}

		if ((ret = __log_c_set_maxrec(logc, np)) != 0) {
			__db_err(dbenv,
			    "DB_LOGC->get: %s: %s", np, db_strerror(ret));
			__os_free(dbenv, np);
			return (ret);
		}
		__os_free(dbenv, np);

		logc->bp_lsn.file = fnum;
	}

	/* Seek to the record's offset. */
	if ((ret = __os_seek(dbenv,
	    logc->c_fhp, 0, 0, offset, 0, DB_OS_SEEK_SET)) != 0) {
		if (!F_ISSET(logc, DB_LOG_SILENT_ERR))
			__db_err(dbenv,
			    "DB_LOGC->get: LSN: %lu/%lu: seek: %s",
			    (u_long)fnum, (u_long)offset, db_strerror(ret));
		return (ret);
	}

	/* Read the data. */
	if ((ret = __os_read(dbenv, logc->c_fhp, p, *nrp, nrp)) != 0) {
		if (!F_ISSET(logc, DB_LOG_SILENT_ERR))
			__db_err(dbenv,
			    "DB_LOGC->get: LSN: %lu/%lu: read: %s",
			    (u_long)fnum, (u_long)offset, db_strerror(ret));
		return (ret);
	}

	return (0);
}

/*
 * __log_c_shortread --
 *	Read was short -- return a consistent error message and error.
 */
static int
__log_c_shortread(logc, lsn, check_silent)
	DB_LOGC *logc;
	DB_LSN *lsn;
	int check_silent;
{
	if (!check_silent || !F_ISSET(logc, DB_LOG_SILENT_ERR))
		__db_err(logc->dbenv, "DB_LOGC->get: LSN: %lu/%lu: short read",
		    (u_long)lsn->file, (u_long)lsn->offset);
	return (EIO);
}

/*
 * __log_c_set_maxrec --
 *	Bound the maximum log record size in a log file.
 */
static int
__log_c_set_maxrec(logc, np)
	DB_LOGC *logc;
	char *np;
{
	DB_ENV *dbenv;
	DB_LOG *dblp;
	LOG *lp;
	u_int32_t mbytes, bytes;
	int ret;

	dbenv = logc->dbenv;
	dblp = dbenv->lg_handle;

	/*
	 * We don't want to try and allocate huge chunks of memory because
	 * applications with error-checking malloc's often consider that a
	 * hard failure.  If we're about to look at a corrupted record with
	 * a bizarre size, we need to know before trying to allocate space
	 * to hold it.  We could read the persistent data at the beginning
	 * of the file but that's hard -- we may have to decrypt it, checksum
	 * it and so on.  Stat the file instead.
	 */
	if (logc->c_fhp != NULL) {
		if ((ret = __os_ioinfo(dbenv, np, logc->c_fhp,
		    &mbytes, &bytes, NULL)) != 0)
			return (ret);
		if (logc->bp_maxrec < (mbytes * MEGABYTE + bytes))
			logc->bp_maxrec = mbytes * MEGABYTE + bytes;
	}

	/*
	 * If reading from the log file currently being written, we could get
	 * an incorrect size, that is, if the cursor was opened on the file
	 * when it had only a few hundred bytes, and then the cursor used to
	 * move forward in the file, after more log records were written, the
	 * original stat value would be wrong.  Use the maximum of the current
	 * log file size and the size of the buffer -- that should represent
	 * the max of any log record currently in the file.
	 *
	 * The log buffer size is set when the environment is opened and never
	 * changed, we don't need a lock on it.
	 */
	lp = dblp->reginfo.primary;
	if (logc->bp_maxrec < lp->buffer_size)
		logc->bp_maxrec = lp->buffer_size;

	return (0);
}