The xrootd Protocol

Version 2.7.0

Andrew Hanushevsky

Stanford Linear Accelerator Center

25-Jan-2007

Produced under contract DE-AC02-76-SFO0515 with the Department of Energy

This code is available under a BSD-style license allowing minimally restricted use.

2 Request/Response Protocol

2.1 Format of Client-Server Initial Handshake

When a client first connects to the XRootd server, it must perform a special handshake. This handshake will determine whether the client is communicating with an XRootd server or a rootd server.

The handshake consists of the client sending 20 bytes, as follows:

kXR_int32 0

kXR_int32 4 (network byte order)

kXR_int32 2012 (network byte order)

The first twelve bytes are zero. The next eight bytes correspond to a standard rootd server protocol request (i.e., kROOTD_PROTOCOL). Both, rootd and XRootd, servers will respond, as follows:

rootd Response XRootd Response

streamid: kXR_char smid[2]

status: kXR_unt16 0

msglen: kXR_int32 8 msglen: kXR_int32 rlen

msgtype: kXR_int32 2012 msgval1: kXR_int32 pval

msgval: kXR_int32 pval msgval2: kXR_int32 flag

Where:

smid is the initial streamid. The smid for the initial response is always two null characters (i.e., ‘\0’);

rlen is the binary response length (e.g., 8 for the indicated response).

pval is the binary protocol version number.

flag is additional bit-encoded information about the server; as follows:

kXR_DataServer - This is a data server.

KXR_LBalServer - This is a load-balancing server.

Notes

1) All binary fields are transmitted in network byte order using an explicit length. The kXR_char and kXR_unt16 data types are treated as unsigned values. All reserved fields must be initialized to binary zero.

2) The first four bytes of the reply determine whether a client is communicating with rootd (has a value of 8) or XRootd (has a value of 0).

3) All twenty bytes must be received by the server at one time. All known TCP implementations will guarantee that the first message is sent intact if all twenty bytes are sent in a single system call. Using multiple system calls for the first message may cause unpredictable results.

2.2 Data Serialization

All data sent and received is serialized (i.e., marshaled) in three ways:

1. Bytes are sent unaligned without any padding,

2. Data type characteristics are predefined (see table below), and

3. All integer quantities are sent in network byte order (i.e, big endian).

XRootd Type	Sign	Bit Length	Bit Alignment	Typical Host Type
kXR_char8	unsigned	8	8	unsigned char
kXR_unt16	unsigned	16	16	unsigned short
kXR_int32	signed	32	32	long[1]
kXR_int64	signed	64	64	long long

Table 1: XRootd Protocol Data Types

Network byte order is defined by the Unix htons() and htonl() macros for host to network short and host to network long, respectively. The reverse is defined by the ntohs() and ntohl() macros. Many systems do not define the long long versions of these macros. XRootd protocol requires that the POSIX version of long long serialization be used, as defined in the following figures. The OS-dependent isLittleEndian() function returns true if the underlying hardware using little endian integer representation.

unsigned long long htonll(unsigned long long x)

{unsigned long long ret_val;

if (isLittleEndian())

{*( (unsigned long *)(&ret_val) + 1) =

htonl(*( (unsigned long *)(&x)));

*(((unsigned long *)(&ret_val))) =

htonl(*( ((unsigned long *)(&x))+1) );

} else {

*( (unsigned long *)(&ret_val)) =

htonl(*( (unsigned long *)(&x)));

*(((unsigned long *)(&ret_val)) + 1) =

htonl(*( ((unsigned long *)(&x))+1) );

}

return ret_val;

};

Figure 1: POSIX Host to Network Byte Order Serialization

unsigned long long ntohll(unsigned long long x)

{unsigned long long ret_val;

if (isLittleEndian())

{*( (unsigned long *)(&ret_val) + 1) =

ntohl(*( (unsigned long *)(&x)));

*(((unsigned long *)(&ret_val))) =

ntohl(*(((unsigned long *)(&x))+1));

} else {

*( (unsigned long *)(&ret_val)) =

ntohl(*( (unsigned long*)(&x)));

*(((unsigned long*)(&ret_val)) + 1) =

ntohl(*(((unsigned long*)(&x))+1));

}

return ret_val;

};

Figure 2: POSIX Network to Host Byte Order Serialization

More compact and efficient, though OS restricted (i.e., Solaris and Linux), versions of 64-bit network byte ordering routines are given in the following figure.

#if defined(__sparc) || __BYTE_ORDER==__BIG_ENDIAN

#ifndef htonll

#define htonll(x) x

#endif

#ifndef ntohll

#define ntohll(x) x

#endif

#else

#ifndef htonll

#define htonll(x) __bswap_64(x)

#endif

#ifndef ntohll

#define ntohll(x) __bswap_64(x)

#endif

Figure 3: Network and Host Byte Ordering Macros

2.3 Client Request Format

Requests sent to the server are a mixture of ASCII and binary. All requests, other than the initial handshake request, have the same format, as follows:

kXR_char streamid[2]

kXR_unt16 requestid

kXR_char parms[16]

kXR_int32 dlen

kXR_char data[dlen]

Where:

streamid

is the binary identifier that is associated with this request stream. This identifier will be echoed along with any response to the request.

requestid

is the binary identifier of the operation to be performed by the server.

parms are parameters specific to the requestid.

dlen is the binary length of the data portion of the message. If no data is present, then the value is zero.

data are data specific to the requestid. Not all requests have associated data. If the request does have data, the length of this field is recorded in the dlen field.

Notes

2) All XRootd client requests consist of a standard 24-byte fixed length message. The 24-byte header may then be optionally followed by request specific data.

3) Stream id’s are arbitrary and are assigned by the client. Typically these id’s correspond to logical connections multiplexed over a physical connection established to a particular server.

4) The client may send any number of requests to the same server. The order in which requests are performed is undefined. Therefore, each request should have a different streamid so that returned results may be paired up with associated requests.

5) Requests sent by a client over a single physical connection may be processed in an arbitrary order. Therefore the client is responsible for serializing requests, as needed.

2.3.1 Valid Client Requests

The following table lists all possible requests and their arguments. Grayed rows represent requests that are not currently supported.

Requestid	Login?	Auth?	Redirect?	Arguments
kXR_admin	yes	yes	no	args
kXR_auth	y	n	n	authtype, authinfo
KXR_bind	n	n	n	sessid
kXR_chmod	y	y	yes	mode, path
kXR_close	y	y	n	fdnum
KXR_dirlist	y	y	y	path
KXR_endsess	y	y	n	sessid
kXR_getfile	y	y	y*	path
kXR_login	n	n	n	userid, token
kXR_ls	y	y	y	options path
kXR_mkdir	y	y	y	mode, path
kXR_mv	y	y	y	old_name, new_name
kXR_open	y	y	y*	mode, flags, path
kXR_ping	y	n	n
kXR_prepare	y	y	n	paths
kXR_protocol	n	n	n
kXR_putfile	y	y	y*	mode, flags, path
kXR_query	y	y	n	args
kXR_read	y	y	y	fdnum, length, offset
kXR_rm	y	y	y	path
kXR_rmdir	y	y	y	path
kXR_set	y	y	n	info
kXR_stat	y	y	y	path
kXR_statx	y	y	n	pathlist
kXR_write	y	y	y	fdnum, length, offset, data
Table 2: Valid Client Requests

2.3.2 Valid Client Paths

The XRootd server accepts only absolute paths where a path may be specified. Relative paths must be resolved by the client interface prior to sending them to XRootd. This means that the interface must handle a virtual “current working directory” to resolve relative paths should they arise.

Path names are restricted to the following set of characters:

Letters (upper or lower case),
Digits (0-9), and
Special characters: !@#%^_-+=:./

In general, paths may not contain shell meta-characters or imbedded spaces.

2.3.3 Client Recovery From Server Failures

A server failure should be recognized when the server unexpectedly closes it’s TCP/IP connection. Should this happen, the client may recover all operations by treating the termination of the connection as a redirection request (see page 30) to the initial XRootd server for all streams associated with the closed TCP/IP connections.

Because many clients are likely to be affected by a server failure, it is important that clients pace their reconnection to the initial XRootd server. One effective way to do this is to use the last three bits of the client’s IP address as the number of seconds to wait before attempting a reconnection. It is up to the client to determine either the number of times or the time window in which reconnections should be attempted before failure is declared. Typical values are 16 attempts or 3 minutes, whichever is longer.

Note that it may not be possible to recover in this way for files that were opened in update mode. Clients who do not provide proper transactional support generally cannot recover via redirection for any read/write resources.

2.3.4 Client Recovery From File Location Failures

A file location failure should be recognized when a server returns kXR_error status code with a kXR_NotFound error code and is either a load balancing server or the target of a redirect. The recovery steps are:

The client should contact the last load balancer used.
For kXR_Open requests, the request should be re-issued with the kXR_refresh option set. For other requests, no recovery is currently possible.
If the same result is encountered again, the client should consider the file missing and not attempt any further recovery actions.

1 Contents