log_get.c

这是linux下运行的mysql软件包,可用于linux 下安装 php + mysql + apach 的网络配置

	if (logc->bp_lsn.file != lsn->file)
		return (0);
	if (logc->bp_lsn.offset > lsn->offset)
		return (0);
	if (logc->bp_lsn.offset + logc->bp_rlen <= lsn->offset + hdr->size)
		return (0);

	/*
	 * Read the record's header and check if the record is entirely held
	 * in the buffer.  If the record is not entirely held, get it again.
	 * (The only advantage in having part of the record locally is that
	 * we might avoid a system call because we already have the HDR in
	 * memory.)
	 *
	 * If the header check fails for any reason, it must be because the
	 * LSN is bogus.  Fail hard.
	 */
	p = logc->bp + (lsn->offset - logc->bp_lsn.offset);
	memcpy(hdr, p, hdr->size);
	if (__log_c_hdrchk(logc, hdr, NULL))
		return (DB_NOTFOUND);
	if (logc->bp_lsn.offset + logc->bp_rlen <= lsn->offset + hdr->len)
		return (0);

	*pp = p;				/* Success. */

	return (0);
}

/*
 * __log_c_inregion --
 *	Check to see if the requested record is in the region's buffer.
 */
static int
__log_c_inregion(logc, lsn, rlockp, last_lsn, hdr, pp)
	DB_LOGC *logc;
	DB_LSN *lsn, *last_lsn;
	RLOCK *rlockp;
	HDR *hdr;
	u_int8_t **pp;
{
	DB_ENV *dbenv;
	DB_LOG *dblp;
	LOG *lp;
	size_t len, nr;
	u_int32_t b_disk, b_region;
	int ret;
	u_int8_t *p;

	dbenv = logc->dbenv;
	dblp = dbenv->lg_handle;
	lp = ((DB_LOG *)logc->dbenv->lg_handle)->reginfo.primary;

	ret = 0;
	*pp = NULL;

	/* If we haven't yet acquired the log region lock, do so. */
	if (*rlockp == L_NONE) {
		*rlockp = L_ACQUIRED;
		R_LOCK(dbenv, &dblp->reginfo);
	}

	/*
	 * The routines to read from disk must avoid reading past the logical
	 * end of the log, so pass that information back to it.
	 *
	 * Since they're reading directly from the disk, they must also avoid
	 * reading past the offset we've written out.  If the log was
	 * truncated, it's possible that there are zeroes or garbage on
	 * disk after this offset, and the logical end of the log can
	 * come later than this point if the log buffer isn't empty.
	 */
	*last_lsn = lp->lsn;
	if (last_lsn->offset > lp->w_off)
		last_lsn->offset = lp->w_off;

	/*
	 * Test to see if the requested LSN could be part of the region's
	 * buffer.
	 *
	 * During recovery, we read the log files getting the information to
	 * initialize the region.  In that case, the region's lsn field will
	 * not yet have been filled in, use only the disk.
	 *
	 * The record must not start at a byte offset after the region buffer's
	 * end, since that means the request is for a record after the end of
	 * the log.  Do this test even if the region's buffer is empty -- after
	 * recovery, the log files may continue past the declared end-of-log,
	 * and the disk reading routine will incorrectly attempt to read the
	 * remainder of the log.
	 *
	 * Otherwise, test to see if the region's buffer actually has what we
	 * want:
	 *
	 * The buffer must have some useful content.
	 * The record must be in the same file as the region's buffer and must
	 * start at a byte offset equal to or greater than the region's buffer.
	 */
	if (IS_ZERO_LSN(lp->lsn))
		return (0);
	if (lsn->file > lp->lsn.file ||
	    (lsn->file == lp->lsn.file && lsn->offset >= lp->lsn.offset))
		return (DB_NOTFOUND);
	if (lp->b_off == 0)
		return (0);
	if (lsn->file < lp->f_lsn.file || lsn->offset < lp->f_lsn.offset)
		return (0);

	/*
	 * The current contents of the cursor's buffer will be useless for a
	 * future call -- trash it rather than try and make it look correct.
	 */
	ZERO_LSN(logc->bp_lsn);

	/*
	 * If the requested LSN is greater than the region buffer's first
	 * byte, we know the entire record is in the buffer.
	 *
	 * If the header check fails for any reason, it must be because the
	 * LSN is bogus.  Fail hard.
	 */
	if (lsn->offset > lp->f_lsn.offset) {
		p = dblp->bufp + (lsn->offset - lp->w_off);
		memcpy(hdr, p, hdr->size);
		if (__log_c_hdrchk(logc, hdr, NULL))
			return (DB_NOTFOUND);
		if (logc->bp_size <= hdr->len) {
			len = ALIGN(hdr->len * 2, 128);
			if ((ret =
			    __os_realloc(logc->dbenv, len, &logc->bp)) != 0)
				 return (ret);
			logc->bp_size = (u_int32_t)len;
		}
		memcpy(logc->bp, p, hdr->len);
		*pp = logc->bp;
		return (0);
	}

	/*
	 * There's a partial record, that is, the requested record starts
	 * in a log file and finishes in the region buffer.  We have to
	 * find out how many bytes of the record are in the region buffer
	 * so we can copy them out into the cursor buffer.  First, check
	 * to see if the requested record is the only record in the region
	 * buffer, in which case we should copy the entire region buffer.
	 *
	 * Else, walk back through the region's buffer to find the first LSN
	 * after the record that crosses the buffer boundary -- we can detect
	 * that LSN, because its "prev" field will reference the record we
	 * want.  The bytes we need to copy from the region buffer are the
	 * bytes up to the record we find.  The bytes we'll need to allocate
	 * to hold the log record are the bytes between the two offsets.
	 */
	b_disk = lp->w_off - lsn->offset;
	if (lp->b_off <= lp->len)
		b_region = (u_int32_t)lp->b_off;
	else
		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 = 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, 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;
	int flags, *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 we read 0 bytes, assume we've hit EOF. */
	if (nr == 0) {
		*eofp = 1;
		return (0);
	}

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

	/* Otherwise, we should have gotten the bytes we wanted. */
	if (nr < hdr->size)
		return (__log_c_shortread(logc, 0));

	/*
	 * Regardless of how we return, the previous contents of the cursor's
	 * buffer are useless -- trash it.
	 */
	ZERO_LSN(logc->bp_lsn);

	/*
	 * 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 = 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, 1));

	/* Set up the return information. */
	logc->bp_rlen = (u_int32_t)nr;
	logc->bp_lsn.file = lsn->file;
	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.
 *
 * XXX
 * I think there's a potential pre-allocation recovery flaw here -- if we
 * fail to write a buffer at the end of a log file (by scheduling its
 * write asynchronously, and it never making it to disk), then succeed in
 * writing a log file block to a subsequent log file, I don't think we will
 * detect that the buffer of 0's should have marked the end of the log files
 * during recovery.  I think we may need to always write some garbage after
 * each block write if we pre-allocate log files.  (At the moment, we do not
 * pre-allocate, so this isn't currently an issue.)
 *
 * 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, hdr, eofp)
	DB_LOGC *logc;
	HDR *hdr;
	int *eofp;
{
	DB_ENV *dbenv;
	int ret;

	dbenv = logc->dbenv;

	/* Sanity check the log record's size. */
	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;
	}

	if (eofp != NULL) {
		if (hdr->prev == 0 && hdr->chksum[0] == 0 && hdr->len == 0) {
			*eofp = 1;
			return (0);
		}
		*eofp = 0;
	}
	return (0);

err:	if (!F_ISSET(logc, DB_LOG_SILENT_ERR))
		__db_err(dbenv, "DB_LOGC->get: invalid log record header");
	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 (F_ISSET(logc->c_fh, DB_FH_VALID) && logc->bp_lsn.file != fnum)
		if ((ret = __os_closehandle(dbenv, logc->c_fh)) != 0)
			return (ret);
	if (!F_ISSET(logc->c_fh, DB_FH_VALID)) {
		if ((ret = __log_name(dblp, fnum,
		    &np, logc->c_fh, 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);
	}

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

	/* Read the data. */
	if ((ret = __os_read(dbenv, logc->c_fh, p, *nrp, nrp)) != 0) {
		if (!F_ISSET(logc, DB_LOG_SILENT_ERR))
			__db_err(dbenv,
			    "DB_LOGC->get: read: %s", 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, silent)
	DB_LOGC *logc;
	int silent;
{
	if (!silent || !F_ISSET(logc, DB_LOG_SILENT_ERR))
		__db_err(logc->dbenv, "DB_LOGC->get: short read");
	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 ((ret =
	    __os_ioinfo(dbenv, np, logc->c_fh, &mbytes, &bytes, NULL)) != 0)
		return (ret);

	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;
	logc->bp_maxrec += lp->buffer_size;

	return (0);
}