Security in Condor is a broad issue, with many aspects to consider. Because Condor's main purpose is to allow users to run arbitrary code on large numbers of computers, it is important to try to limit who can access a Condor pool and what privileges they have when using the pool. This section covers these topics.
There is a distinction between the kinds of resource attacks Condor can defeat, and the kinds of attacks Condor cannot defeat. Condor cannot prevent security breaches of users that can elevate their privilege to the root or administrator account. Condor does not run user jobs in sandboxes (standard universe jobs are a partial exception to this), so Condor cannot defeat all malicious actions by user jobs. An example of a malicious job is one that launches a distributed denial of service attack. Condor assumes that users are trustworthy. Condor can prevent unauthorized access to the Condor pool, to help ensure that only trusted users have access to the pool. In addition, Condor provides encryption and integrity checking, to ensure that data (both Condor's data and user jobs' data) has not been examined or tampered with while in transit.
Broadly speaking, the aspects of security in Condor may be categorized and described:
At the heart of Condor's security model is the notion that communications are subject to various security checks. A request from one Condor daemon to another may require authentication to prevent subversion of the system. A request from a user of Condor may need to be denied due to the confidential nature of the request. The security model handles these example situations and many more.
Requests to Condor are categorized into groups of access levels, based on the type of operation requested. The user of a specific request must be authorized at the required access level. For example, executing the condor_status command requires the READ access level. Actions that accomplish management tasks, such as shutting down or restarting of a daemon require an ADMINISTRATOR access level. See Section 3.6.7 for a full list of Condor's access levels and their meanings.
There are two sides to any communication or command invocation in Condor. One side is identified as the client, and the other side is identified as the daemon. The client is the party that initiates the command, and the daemon is the party that processes the command and responds. In some cases it is easy to distinguish the client from the daemon, while in other cases it is not as easy. Condor tools such as condor_submit and condor_config_val are clients. They send commands to daemons and act as clients in all their communications. For example, the condor_submit command communicates with the condor_schedd. Behind the scenes, Condor daemons also communicate with each other; in this case the daemon initiating the command plays the role of the client. For instance, the condor_negotiator daemon acts as a client when contacting the condor_schedd daemon to initiate matchmaking. Once a match has been found, the condor_schedd daemon acts as a client and contacts the condor_startd daemon.
Condor's security model is implemented using configuration. Commands in Condor are executed over TCP/IP network connections. While network communication enables Condor to manage resources that are distributed across an organization (or beyond), it also brings in security challenges. Condor must have ways of ensuring that commands are being sent by trustworthy users. Jobs that are operating on sensitive data must be allowed to use encryption such that the data is not seen by outsiders. Jobs may need assurance that data has not been tampered with. These issues can be addressed with Condor's authentication, encryption, and integrity features.
The following is a list of registered commands that daemons will accept. The list is ordered by daemon. For each daemon, the commands are grouped by the access level required for a daemon to accept the command from a given machine.
The command sent as a result of condor_reconfig to reconfigure a daemon.
All commands that relate to a condor_schedd daemon claiming a machine, starting jobs there, or stopping those jobs.
The command that condor_checkpoint sends to periodically checkpoint all running jobs.
The command that condor_preen sends to request the current state of the condor_startd daemon.
The commands which write information into the job queue (such as condor_submit and condor_hold). Note that for most commands which attempt to write to the job queue, Condor will perform an additional user-level authentication step. This additional user-level authentication prevents, for example, an ordinary user from removing a different user's jobs.
The commands that a condor_startd sends to the condor_schedd when the condor_schedd daemon's claim is being preempted and also when the lease on the claim is renewed. These operations only require READ access, rather than DAEMON in order to limit the level of trust that the condor_schedd must have for the condor_startd. Success of these commands is only possible if the condor_startd knows the secret claim id, so effectively, authorization for these commands is more specific than Condor's general security model implies. The condor_schedd automatically grants the condor_startd READ access for the duration of the claim. Therefore, if one desires to only authorize specific execute machines to run jobs, one must either limit which machines are allowed to advertise themselves to the pool (most common) or configure the condor_schedd's ALLOW_CLIENT setting to only allow connections from the condor_schedd to the trusted execute machines.
MASTER: All commands are registered with ADMINISTRATOR access:
Because of the wide range of environments and security demands necessary, Condor must be flexible. Configuration provides this flexibility. The process by which Condor determines the security settings that will be used when a connection is established is called security negotiation. Security negotiation's primary purpose is to determine which of the features of authentication, encryption, and integrity checking will be enabled for a connection. In addition, since Condor supports multiple technologies for authentication and encryption, security negotiation also determines which technology is chosen for the connection.
Security negotiation is a completely separate process from matchmaking, and should not be confused with any specific function of the condor_negotiator daemon. Security negotiation occurs when one Condor daemon or tool initiates communication with another Condor daemon, to determine the security settings by which the communication will be ruled. The condor_negotiator daemon does negotiation, whereby queued jobs and available machines within a pool go through the process of matchmaking (deciding out which machines will run which jobs).
The configuration macro names that determine what features will be used during client-daemon communication follow the pattern:
portion of the macro name determines which security feature's
policy is being set.
<feature> may be any one of
AUTHENTICATION ENCRYPTION INTEGRITY NEGOTIATION
<context> component of the security policy macros can be
used to craft a fine-grained security policy based on the type of
communication taking place.
<context> may be any one of
CLIENT READ WRITE ADMINISTRATOR CONFIG OWNER DAEMON NEGOTIATOR ADVERTISE_MASTER ADVERTISE_STARTD ADVERTISE_SCHEDD DEFAULT
Any of these constructed configuration macros may be set to any of the following values:
REQUIRED PREFERRED OPTIONAL NEVER
Security negotiation resolves various client-daemon combinations of desired security features in order to set a policy.
As an example, consider Frida the scientist. Frida wants to avoid authentication when possible. She sets
SEC_DEFAULT_AUTHENTICATION = OPTIONALThe machine running the condor_schedd to which Frida will remotely submit jobs, however, is operated by a security-conscious system administrator who dutifully sets:
SEC_DEFAULT_AUTHENTICATION = REQUIREDWhen Frida submits her jobs, Condor's security negotiation determines that authentication will be used, and allows the command to continue. This example illustrates the point that the most restrictive security policy sets the levels of security enforced. There is actually more to the understanding of this scenario. Some Condor commands, such as the use of condor_submit to submit jobs always require authentication of the submitter, no matter what the policy says. This is because the identity of the submitter needs to be known in order to carry out the operation. Others commands, such as condor_q, do not always require authentication, so in the above example, the server's policy would force Frida's condor_q queries to be authenticated, whereas a different policy could allow condor_q to happen without any authentication.
Whether or not security negotiation occurs depends on the setting at both the client and daemon side of the configuration variable(s) defined by SEC_*_NEGOTIATION. SEC_DEFAULT_NEGOTIATION is a variable representing the entire set of configuration variables for NEGOTIATION. For the client side setting, the only definitions that make sense are REQUIRED and NEVER. For the daemon side setting, the PREFERRED value makes no sense. Table 3.1 shows how security negotiation resolves various client-daemon combinations of security negotiation policy settings. Within the table, Yes means the security negotiation will take place. No means it will not. Fail means that the policy settings are incompatible and the communication cannot continue.
Enabling authentication, encryption, and integrity checks is dependent on security negotiation taking place. The enabled security negotiation further sets the policy for these other features. Table 3.2 shows how security features are resolved for client-daemon combinations of security feature policy settings. Like Table 3.1, Yes means the feature will be utilized. No means it will not. Fail implies incompatibility and the feature cannot be resolved.
The enabling of encryption and/or integrity checks is dependent on authentication taking place. The authentication provides a key exchange. The key is needed for both encryption and integrity checks.
SEC_CLIENT_<feature> determines the policy for all
The policy for incoming commands
(the daemon side of the communication)
takes a more fine-grained approach that implements a set of
access levels for the received command.
For example, it is desirable to have all incoming administrative
requests require authentication.
Inquiries on pool status may not be so restrictive.
To implement this, the administrator configures the policy:
SEC_ADMINISTRATOR_AUTHENTICATION = REQUIRED SEC_READ_AUTHENTICATION = OPTIONAL
DEFAULT value for
<context> provides a way
to set a policy for all access levels
that do not have a specific configuration variable defined.
In addition, some access levels will default to the settings specified
for other access levels. For example, ADVERTISE_STARTD
defaults to DAEMON, and DAEMON defaults to
WRITE, which then defaults to the general
Authentication and encryption can each be accomplished by a variety of methods or technologies. Which method is utilized is determined during security negotiation.
The configuration macros that determine the methods to use for authentication and/or encryption are
These macros are defined by a comma or space delimited list of possible methods to use. Section 3.6.3 lists all implemented authentication methods. Section 3.6.5 lists all implemented encryption methods.
The client side of any communication uses one of two macros to specify whether authentication is to occur:
For the daemon side, there are a larger number of macros to specify whether authentication is to take place, based upon the necessary access level:
SEC_DEFAULT_AUTHENTICATION SEC_READ_AUTHENTICATION SEC_WRITE_AUTHENTICATION SEC_ADMINISTRATOR_AUTHENTICATION SEC_CONFIG_AUTHENTICATION SEC_OWNER_AUTHENTICATION SEC_DAEMON_AUTHENTICATION SEC_NEGOTIATOR_AUTHENTICATION SEC_ADVERTISE_MASTER_AUTHENTICATION SEC_ADVERTISE_STARTD_AUTHENTICATION SEC_ADVERTISE_SCHEDD_AUTHENTICATION
As an example, the macro defined in the configuration file for a daemon as
SEC_WRITE_AUTHENTICATION = REQUIREDsignifies that the daemon must authenticate the client for any communication that requires the WRITE access level. If the daemon's configuration contains
SEC_DEFAULT_AUTHENTICATION = REQUIREDand does not contain any other security configuration for
AUTHENTICATION, then this default defines the daemon's needs for authentication over all access levels. Where a specific macro is defined, the more specific value takes precedence over the default definition.
If authentication is to be done, then the communicating parties must negotiate a mutually acceptable method of authentication to be used. A list of acceptable methods may be provided by the client, using the macros
SEC_DEFAULT_AUTHENTICATION_METHODS SEC_CLIENT_AUTHENTICATION_METHODSA list of acceptable methods may be provided by the daemon, using the macros
SEC_DEFAULT_AUTHENTICATION_METHODS SEC_READ_AUTHENTICATION_METHODS SEC_WRITE_AUTHENTICATION_METHODS SEC_ADMINISTRATOR_AUTHENTICATION_METHODS SEC_CONFIG_AUTHENTICATION_METHODS SEC_OWNER_AUTHENTICATION_METHODS SEC_DAEMON_AUTHENTICATION_METHODS SEC_NEGOTIATOR_AUTHENTICATION_METHODS SEC_ADVERTISE_MASTER_AUTHENTICATION_METHODS SEC_ADVERTISE_STARTD_AUTHENTICATION_METHODS SEC_ADVERTISE_SCHEDD_AUTHENTICATION_METHODSThe methods are given as a comma-separated list of acceptable values. These variables list the authentication methods that are available to be used. The ordering of the list defines preference; the first item in the list indicates the highest preference. Defined values are
GSI SSL KERBEROS PASSWORD FS FS_REMOTE NTSSPI CLAIMTOBE ANONYMOUS
For example, a client may be configured with:
SEC_CLIENT_AUTHENTICATION_METHODS = FS, GSIand a daemon the client is trying to contact with:
SEC_DEFAULT_AUTHENTICATION_METHODS = GSI
Security negotiation will determine that GSI authentication is the only compatible choice. If there are multiple compatible authentication methods, security negotiation will make a list of acceptable methods and they will be tried in order until one succeeds.
As another example, the macro
SEC_DEFAULT_AUTHENTICATION_METHODS = KERBEROS, NTSSPIindicates that either Kerberos or Windows authentication may be used, but Kerberos is preferred over Windows. Note that if the client and daemon agree that multiple authentication methods may be used, then they are tried in turn. For instance, if they both agree that Kerberos or NTSSPI may be used, then Kerberos will be tried first, and if there is a failure for any reason, then NTSSPI will be tried.
An additional specialized method of authentication exists for communication between the condor_schedd and condor_startd. It is especially useful when operating at large scale over high latency networks or in situations where it is inconvenient to set up one of the other methods of strong authentication between the submit and execute daemons. See the description of SEC_ENABLE_MATCH_PASSWORD_AUTHENTICATION on for details.
If the configuration for a machine does not define any variable
then Condor uses a default value of
Authentication will be required for
any operation which modifies the job queue,
such as condor_qedit and condor_rm.
If the configuration for a machine does not define any variable
the default value for a Unix machine is
This default value for a Windows machine is
A simple introduction to this type of authentication defines Condor's use of terminology, and it illuminates the needed items that Condor must access to do this authentication. Assume that A authenticates to B. In this example, A is the client, and B is the daemon within their communication. This example's one-way authentication implies that B is verifying the identity of A, using the certificate A provides, and utilizing B's own set of trusted CAs (Certification Authorities). Client A provides its certificate (or proxy) to daemon B. B does two things: B checks that the certificate is valid, and B checks to see that the CA that signed A's certificate is one that B trusts.
For the GSI authentication protocol, an X.509 certificate is required. Files with predetermined names hold a certificate, a key, and optionally, a proxy. A separate directory has one or more files that become the list of trusted CAs.
Allowing Condor to do this GSI authentication requires knowledge of the locations of the client A's certificate and the daemon B's list of trusted CAs. When one side of the communication (as either client A or daemon B) is a Condor daemon, these locations are determined by configuration or by default locations. When one side of the communication (as a client A) is a user of Condor (the process owner of a Condor tool, for example condor_submit), these locations are determined by the pre-set values of environment variables or by default locations.
For a Condor daemon, the certificate may be a single host certificate, and all Condor daemons on the same machine may share the same certificate. In some cases, the certificate can also be copied to other machines, where local copies are necessary. This may occur only in cases where a single host certificate can match multiple host names, something that is beyond the scope of this manual. The certificates must be protected by access rights to files, since the password file is not encrypted.
The specification of the location of the necessary files through configuration uses the following precedence.
GSI_DAEMON_CERT = $(GSI_DAEMON_DIRECTORY)/hostcert.pem GSI_DAEMON_KEY = $(GSI_DAEMON_DIRECTORY)/hostkey.pem GSI_DAEMON_TRUSTED_CA_DIR = $(GSI_DAEMON_DIRECTORY)/certificates
Note that no proxy is assumed in this case.
When a daemon acts as the client within authentication, the daemon needs a listing of those from which it will accept certificates. This is done with GSI_DAEMON_NAME. This name is specified with the following format
GSI_DAEMON_NAME = /X.509/name/of/server/1,/X.509/name/of/server/2,...
Condor will also need a way to map an X.509 distinguished
name to a Condor user id.
There are two ways to accomplish this mapping.
For a first way to specify the mapping, see
to use Condor's unified map file.
The second way to do the mapping is within
an administrator-maintained GSI-specific file called an X.509 map file,
mapping from X.509 Distinguished Name (DN) to Condor user id.
It is similar to a Globus grid map file, except that it is only used for
mapping to a user id, not for authorization.
If the user names in the
map file do not specify a domain for the user
(specification would appear as
then the value of UID_DOMAIN is used.
Entries (lines) in the file each contain two items.
The first item in an entry is the
X.509 certificate subject name, and it is enclosed in double quote marks
(using the character
The second item is the Condor user id.
The two items in an entry are separated by tab or space character(s).
Here is an example of an entry in an X.509 map file.
Entries must be on a single line; this example is broken
onto two lines for formatting reasons.
"/C=US/O=Globus/O=University of Wisconsin/ OU=Computer Sciences Department/CN=Alice Smith" asmith
Condor finds the map file in one of three ways. If the configuration variable GRIDMAP is defined, it gives the full path name to the map file. When not defined, Condor looks for the map file in
$(GSI_DAEMON_DIRECTORY)/grid-mapfileIf GSI_DAEMON_DIRECTORY is not defined, then the third place Condor looks for the map file is given by
The user specifies the location of a certificate, proxy, etc. in one of two ways:
X509_USER_PROXY gives the path and file name of the proxy. This proxy will have been created using the grid-proxy-init program, which will place the proxy in the /tmp directory with the file name being determined by the format:
/tmp/x509up_uXXXXThe specific file name is given by substituting the
XXXXcharacters with the UID of the user. Note that when a valid proxy is used, the certificate and key locations are not needed.
X509_USER_CERT gives the path and file name of the certificate. It is also used if a proxy location has been checked, but the proxy is no longer valid.
X509_USER_KEY gives the path and file name of the key. Note that most keys are password encrypted, such that knowing the location could not lead to using the key.
X509_CERT_DIR gives the path to the directory containing the list of trusted CAs.
Here is an example portion of the configuration file that would enable and require GSI authentication, along with a minimal set of other variables to make it work.
SEC_DEFAULT_AUTHENTICATION = REQUIRED SEC_DEFAULT_AUTHENTICATION_METHODS = GSI SEC_DEFAULT_INTEGRITY = REQUIRED GSI_DAEMON_DIRECTORY = /etc/grid-security GRIDMAP = /etc/grid-security/grid-mapfile # authorize based on user names produced by the map file ALLOW_READ = *@cs.wisc.edu/*.cs.wisc.edu ALLOW_DAEMON = email@example.com/*.cs.wisc.edu ALLOW_NEGOTIATOR = firstname.lastname@example.org/condor.cs.wisc.edu, \ email@example.com/condor2.cs.wisc.edu ALLOW_ADMINISTRATOR = firstname.lastname@example.org/*.cs.wisc.edu # condor daemon certificate(s) trusted by condor tools and daemons # when connecting to other condor daemons GSI_DAEMON_NAME = /C=US/O=Condor/O=UW/OU=CS/CNemail@example.com # clear out any host-based authorizations # (unnecessary if you leave authentication REQUIRED, # but useful if you make it optional and want to # allow some unauthenticated operations, such as # ALLOW_READ = */*.cs.wisc.edu) HOSTALLOW_READ = HOSTALLOW_WRITE = HOSTALLOW_NEGOTIATOR = HOSTALLOW_ADMINISTRATOR =
The SEC_DEFAULT_AUTHENTICATION macro specifies that authentication is required for all communications. This single macro covers all communications, but could be replaced with a set of macros that require authentication for only specific communications.
The macro GSI_DAEMON_DIRECTORY is specified to give Condor a single place to find the daemon's certificate. This path may be a directory on a shared file system such as AFS. Alternatively, this path name can point to local copies of the certificate stored in a local file system.
The macro GRIDMAP specifies the file to use for mapping GSI names to user names within Condor. For example, it might look like this:
Additional mappings would be needed for the users who submit jobs to the pool or who issue administrative commands.
All SSL authentication is mutual authentication in Condor. This means that when SSL authentication is used and when one process communicates with another, each process must be able to verify the signature on the certificate presented by the other process. The process that initiates the connection is the client, and the process that receives the connection is the server. For example, when a condor_startd daemon authenticates with a condor_collector daemon to provide a machine ClassAd, the condor_startd daemon initiates the connection and acts as the client, and the condor_collector daemon acts as the server.
The names and locations of keys and certificates for clients, servers, and the files used to specify trusted certificate authorities (CAs) are defined by settings in the configuration files. The contents of the files are identical in format and interpretation to those used by other systems which use SSL, such as Apache httpd.
The configuration variables AUTH_SSL_CLIENT_CERTFILE and AUTH_SSL_SERVER_CERTFILE specify the file location for the certificate file for the initiator and recipient of connections, respectively. Similarly, the configuration variables AUTH_SSL_CLIENT_KEYFILE and AUTH_SSL_SERVER_KEYFILE specify the locations for keys.
The configuration variables AUTH_SSL_SERVER_CAFILE and AUTH_SSL_CLIENT_CAFILE each specify a path and file name, providing the location of a file containing one or more certificates issued by trusted certificate authorities. Similarly, AUTH_SSL_SERVER_CADIR and AUTH_SSL_CLIENT_CADIR each specify a directory with one or more files, each which may contain a single CA certificate. The directories must be prepared using the OpenSSL c_rehash utility.
If Kerberos is used for authentication, then a mapping from a Kerberos domain (called a realm) to a Condor UID domain is necessary. There are two ways to accomplish this mapping. For a first way to specify the mapping, see section 3.6.4 to use Condor's unified map file. A second way to specify the mapping defines the configuration variable KERBEROS_MAP_FILE to define a path to an administrator-maintained Kerberos-specific map file. The configuration syntax is
KERBEROS_MAP_FILE = /path/to/etc/condor.kmap
Lines within this map file have the syntax
KERB.REALM = UID.domain.name
Here are two lines from a map file to use as an example:
CS.WISC.EDU = cs.wisc.edu ENGR.WISC.EDU = ee.wisc.edu
If a KERBEROS_MAP_FILE configuration variable is defined and set, then all permitted realms must be explicitly mapped. If no map file is specified, then Condor assumes that the Kerberos realm is the same as the Condor UID domain.
The configuration variable
defines the name of a Kerberos principal.
If KERBEROS_SERVER_PRINCIPAL is not defined,
then the default value used is
A principal specifies a unique name to which a set of credentials
may be assigned.
Condor takes the specified (or default) principal and appends a slash character, the host name, an '@' (at sign character), and the Kerberos realm. As an example, the configuration
KERBEROS_SERVER_PRINCIPAL = condor-daemonresults in Condor's use of
condor-daemon/the.host.name@YOUR.KERB.REALMas the server principal.
Here is an example of configuration settings that use Kerberos for authentication and require authentication of all communications of the write or administrator access level.
SEC_WRITE_AUTHENTICATION = REQUIRED SEC_WRITE_AUTHENTICATION_METHODS = KERBEROS SEC_ADMINISTRATOR_AUTHENTICATION = REQUIRED SEC_ADMINISTRATOR_AUTHENTICATION_METHODS = KERBEROS
Kerberos authentication on Unix platforms requires access to various files that usually are only accessible by the root user. At this time, the only supported way to use KERBEROS authentication on Unix platforms is to start daemons Condor as user root.
The password method provides mutual authentication through the use of a shared secret. This is often a good choice when strong security is desired, but an existing Kerberos or X.509 infrastructure is not in place. Password authentication is available on both Unix and Windows. It currently can only be used for daemon-to-daemon authentication. The shared secret in this context is referred to as the pool password.
Before a daemon can use password authentication, the pool password must be stored on the daemon's local machine. On Unix, the password will be placed in a file defined by the configuration variable SEC_PASSWORD_FILE . This file will be accessible only by the UID that Condor is started as. On Windows, the same secure password store that is used for user passwords will be used for the pool password (see section 6.2.3).
Under Unix, the password file can be generated by using the following command to write directly to the password file:
condor_store_cred -f /path/to/password/file
Under Windows (or under Unix), storing the pool password is done with the -c option when using to condor_store_cred add. Running
condor_store_cred -c addprompts for the pool password and store it on the local machine, making it available for daemons to use in authentication. The condor_master must be running for this command to work.
In addition, storing the pool password to a given machine requires
CONFIG-level access. For example, if the pool password should
only be set locally, and only by root, the following would be placed in
the global configuration file.
ALLOW_CONFIG = root@mydomain/$(IP_ADDRESS)
It is also possible to set the pool password remotely, but this is
recommended only if it can be done over an encrypted channel. This is
possible on Windows, for example, in an environment where common
accounts exist across all the machines in the pool. In this case,
ALLOW_CONFIG can be set to allow the Condor administrator (who
in this example has an account condor common to all machines in
the pool) to set the password from the central manager as follows.
ALLOW_CONFIG = condor@mydomain/$(CONDOR_HOST)The Condor administrator then executes
condor_store_cred -c -n host.mydomain addfrom the central manager to store the password to a given machine. Since the condor account exists on both the central manager and
host.mydomain, the NTSSPI authentication method can be used to authenticate and encrypt the connection. condor_store_cred will warn and prompt for cancellation, if the channel is not encrypted for whatever reason (typically because common accounts do not exist or Condor's security is misconfigured).
When a daemon is authenticated using a pool password, its security
$(UID_DOMAIN) is taken from the daemon's configuration. The
variables for authorization
should restrict access using this name. For example,
ALLOW_DAEMON = condor_pool@mydomain/*, condor@mydomain/$(IP_ADDRESS) ALLOW_NEGOTIATOR = condor_pool@mydomain/$(CONDOR_HOST)This configuration allows remote
NEGOTIATOR-level access, if the pool password is known. Local daemons authenticated as
condor@mydomainare also allowed access. This is done so local authentication can be done using another method such as
SEC_PASSWORD_FILE = $(LOCK)/pool_password SEC_DAEMON_AUTHENTICATION = REQUIRED SEC_DAEMON_INTEGRITY = REQUIRED SEC_DAEMON_AUTHENTICATION_METHODS = PASSWORD SEC_NEGOTIATOR_AUTHENTICATION = REQUIRED SEC_NEGOTIATOR_INTEGRITY = REQUIRED SEC_NEGOTIATOR_AUTHENTICATION_METHODS = PASSWORD SEC_CLIENT_AUTHENTICATION_METHODS = FS, PASSWORD, KERBEROS, GSI ALLOW_DAEMON = condor_pool@$(UID_DOMAIN)/*.cs.wisc.edu, \ condor@$(UID_DOMAIN)/$(IP_ADDRESS) ALLOW_NEGOTIATOR = condor_pool@$(UID_DOMAIN)/negotiator.machine.name
One problem with the pool password method of authentication is that it involves a single, shared secret. This does not scale well with the addition of remote users who flock to the local pool. However, the pool password may still be used for authenticating portions of the local pool, while others (such as the remote condor_schedd daemons involved in flocking) are authenticated by other means.
In this example, only the condor_startd daemons in the local pool are required to have the pool password when they advertise themselves to the condor_collector daemon.
SEC_PASSWORD_FILE = $(LOCK)/pool_password SEC_ADVERTISE_STARTD_AUTHENTICATION = REQUIRED SEC_ADVERTISE_STARTD_INTEGRITY = REQUIRED SEC_ADVERTISE_STARTD_AUTHENTICATION_METHODS = PASSWORD SEC_CLIENT_AUTHENTICATION_METHODS = FS, PASSWORD, KERBEROS, GSI ALLOW_ADVERTISE_STARTD = condor_pool@$(UID_DOMAIN)/*.cs.wisc.edu
This form of authentication utilizes the ownership of a file in the identity verification of a client. A daemon authenticating a client requires the client to write a file in a specific location (/tmp). The daemon then checks the ownership of the file. The file's ownership verifies the identity of the client. In this way, the file system becomes the trusted authority. This authentication method is only appropriate for clients and daemons that are on the same computer.
Condor's unified map file allows the mappings from authenticated names to a Condor canonical user name to be specified as a single list within a single file. The location of the unified map file is defined by the configuration variable CERTIFICATE_MAPFILE ; it specifies the path and file name of the unified map file. Each mapping is on its own line of the unified map file. Each line contains 3 fields, separated by white space (space or tab characters):
Allowable authentication method names are the same as used to define any of the configuration variables SEC_*_AUTHENTICATION_METHODS, as repeated here:
GSI SSL KERBEROS PASSWORD FS FS_REMOTE NTSSPI CLAIMTOBE ANONYMOUS
The fields that represent an authenticated name and the canonical Condor user name may utilize regular expressions as defined by PCRE (Perl-Compatible Regular Expressions). Due to this, more than one line (mapping) within the unified map file may match. Look ups are therefore defined to use the first mapping that matches.
A regular expression may need to contain spaces, and in this case the entire expression can be surrounded by double quote marks. If a double quote character also needs to appear in such an expression, it is preceded by a backslash.
The default behavior of Condor when no map file is specified is to do the following mappings, with some additional logic noted below:
FS (.*) \1 FS_REMOTE (.*) \1 GSI (.*) GSS_ASSIST_GRIDMAP SSL (.*) ssl@unmapped KERBEROS ([^/]*)/?[^@]*@(.*) \1@\2 NTSSPI (.*) \1 CLAIMTOBE (.*) \1 PASSWORD (.*) \1
For GSI (or SSL), the special name GSS_ASSIST_GRIDMAP instructs
Condor to use the GSI grid map file (configured with GRIDMAP
as shown in section 3.6.3) to do the mapping.
If no mapping can be found for GSI (with or without the use of
GSS_ASSIST_GRIDMAP), the user is mapped to
For Kerberos, if KERBEROS_MAP_FILE is specified, the domain portion of the name is obtained by mapping the Kerberos realm to the value specified in the map file, rather than just using the realm verbatim as the domain portion of the condor user name. See section 3.6.3 for details.
If authentication did not happen or failed and was not required, then
the user is given the name
With the integration of VOMS for GSI authentication, the interpretation of the regular expression representing the authenticated name may change. First, the full serialized DN and FQAN are used in attempting a match. If no match is found using the full DN and FQAN, then the DN is then used on its own without the FQAN. Using this, roles or user names from the VOMS attributes may be extracted to be used as the target for mapping. And, in this case the FQAN are verified, permitting reliance on their authenticity.
The client uses one of two macros to enable or disable encryption:
For the daemon, there are seven macros to enable or disable encryption:
SEC_DEFAULT_ENCRYPTION SEC_READ_ENCRYPTION SEC_WRITE_ENCRYPTION SEC_ADMINISTRATOR_ENCRYPTION SEC_CONFIG_ENCRYPTION SEC_OWNER_ENCRYPTION SEC_DAEMON_ENCRYPTION SEC_NEGOTIATOR_ENCRYPTION SEC_ADVERTISE_MASTER_ENCRYPTION SEC_ADVERTISE_STARTD_ENCRYPTION SEC_ADVERTISE_SCHEDD_ENCRYPTION
As an example, the macro defined in the configuration file for a daemon as
SEC_CONFIG_ENCRYPTION = REQUIREDsignifies that any communication that changes a daemon's configuration must be encrypted. If a daemon's configuration contains
SEC_DEFAULT_ENCRYPTION = REQUIREDand does not contain any other security configuration for ENCRYPTION, then this default defines the daemon's needs for encryption over all access levels. Where a specific macro is present, its value takes precedence over any default given.
If encryption is to be done, then the communicating parties must find (negotiate) a mutually acceptable method of encryption to be used. A list of acceptable methods may be provided by the client, using the macros
SEC_DEFAULT_CRYPTO_METHODS SEC_CLIENT_CRYPTO_METHODSA list of acceptable methods may be provided by the daemon, using the macros
SEC_DEFAULT_CRYPTO_METHODS SEC_READ_CRYPTO_METHODS SEC_WRITE_CRYPTO_METHODS SEC_ADMINISTRATOR_CRYPTO_METHODS SEC_CONFIG_CRYPTO_METHODS SEC_OWNER_CRYPTO_METHODS SEC_DAEMON_CRYPTO_METHODS SEC_NEGOTIATOR_CRYPTO_METHODS SEC_ADVERTISE_MASTER_CRYPTO_METHODS SEC_ADVERTISE_STARTD_CRYPTO_METHODS SEC_ADVERTISE_SCHEDD_CRYPTO_METHODS
The methods are given as a comma-separated list of acceptable values. These variables list the encryption methods that are available to be used. The ordering of the list gives preference; the first item in the list indicates the highest preference. Possible values are
An integrity check assures that the messages between communicating parties have not been tampered with. Any change, such as addition, modification, or deletion can be detected. Through configuration macros, both the client and the daemon can specify whether an integrity check is required of further communication.
The client uses one of two macros to enable or disable an integrity check:
For the daemon, there are seven macros to enable or disable an integrity check:
SEC_DEFAULT_INTEGRITY SEC_READ_INTEGRITY SEC_WRITE_INTEGRITY SEC_ADMINISTRATOR_INTEGRITY SEC_CONFIG_INTEGRITY SEC_OWNER_INTEGRITY SEC_DAEMON_INTEGRITY SEC_NEGOTIATOR_INTEGRITY SEC_ADVERTISE_MASTER_INTEGRITY SEC_ADVERTISE_STARTD_INTEGRITY SEC_ADVERTISE_SCHEDD_INTEGRITY
As an example, the macro defined in the configuration file for a daemon as
SEC_CONFIG_INTEGRITY = REQUIREDsignifies that any communication that changes a daemon's configuration must have its integrity assured. If a daemon's configuration contains
SEC_DEFAULT_INTEGRITY = REQUIREDand does not contain any other security configuration for
INTEGRITY, then this default defines the daemon's needs for integrity checks over all access levels. Where a specific macro is present, its value takes precedence over any default given.
A signed MD5 check sum is currently the only available method for integrity checking. Its use is implied whenever integrity checks occur. If more methods are implemented, then there will be further macros to allow both the client and the daemon to specify which methods are acceptable.
Authorization protects resource usage by granting or denying access requests made to the resources. It defines who is allowed to do what.
The authorization portion of the security of a Condor pool is based on a set of configuration macros. The macros list which user will be authorized to issue what request given a specific access level. When a daemon is to be authorized, its user name is the login under which the daemon is executed.
These configuration macros define a set of users that will be allowed to (or denied from) carrying out various Condor commands. Each access level may have its own list of authorized users. A complete list of the authorization macros:
ALLOW_READ ALLOW_WRITE ALLOW_ADMINISTRATOR ALLOW_CONFIG ALLOW_SOAP ALLOW_OWNER ALLOW_NEGOTIATOR ALLOW_DAEMON DENY_READ DENY_WRITE DENY_ADMINISTRATOR DENY_SOAP DENY_CONFIG DENY_OWNER DENY_NEGOTIATOR DENY_DAEMON
In addition, the following are used to control authorization of specific types of Condor daemons when advertising themselves to the pool. If unspecified, these default to the broader ALLOW_DAEMON and DENY_DAEMON settings.
ALLOW_ADVERTISE_MASTER ALLOW_ADVERTISE_STARTD ALLOW_ADVERTISE_SCHEDD DENY_ADVERTISE_MASTER DENY_ADVERTISE_STARTD DENY_ADVERTISE_SCHEDD
Each client side of a connection may also specify its own list of trusted servers. This is done using the following settings. Note that the FS and CLAIMTOBE authentication methods are not symmetric. The client is authenticated by the server, but the server is not authenticated by the client. When the server is not authenticated to the client, only the network address of the host may be authorized and not the specific identity of the server.
The names ALLOW_CLIENT and DENY_CLIENT should be thought of as ``when I am acting as a client, these are the servers I allow or deny.'' It should not be confused with the incorrect thought ``when I am the server, these are the clients I allow or deny.''
All authorization settings are defined by a comma-separated list of fully qualified users. Each fully qualified user is described using the following format:
username@domain/hostnameThe information to the left of the slash character describes a user within a domain. The information to the right of the slash character describes one or more machines from which the user would be issuing a command. This host name may take the form of either a fully qualified host name of the form
bird.cs.wisc.eduor an IP address of the form
An example is
Within the format, wild card characters (the asterisk, *) are allowed. The use of wild cards is limited to one wild card on either side of the slash character. A wild card character used in the host name is further limited to come at the beginning of a fully qualified host name or at the end of an IP address. For example,
*@cs.wisc.edu/bird.cs.wisc.edurefers to any user that comes from
cs.wisc.edu, where the command is originating from the machine
bird.cs.wisc.edu. Another valid example,
firstname.lastname@example.org/*.cs.wisc.edurefers to commands coming from any machine within the
cs.wisc.edudomain, and issued by
zmiller. A third valid example,
*@cs.wisc.edu/*refers to commands coming from any user within the
cs.wisc.edudomain where the command is issued from any machine. A fourth valid example,
*@cs.wisc.edu/128.105.*refers to commands coming from any user within the
cs.wisc.edudomain where the command is issued from machines within the network that match the first two octets of the IP address.
If the set of machines is specified by an IP address, then further specification using a net mask identifies a physical set (subnet) of machines. This physical set of machines is specified using the form
networkis an IP address. The net mask takes one of two forms. It may be a decimal number which refers to the number of leading bits of the IP address that are used in describing a subnet. Or, the net mask may take the form of
dare decimal numbers that each specify an 8-bit mask. An example net mask is
255.255.192.0which specifies the bit mask
A single complete example of a configuration variable that uses a net mask is
ALLOW_WRITE = email@example.com/18.104.22.168/17User
cs.wisc.edudomain is given write authorization when originating from machines that match their leftmost 17 bits of the IP address.
This flexible set of configuration macros could used to define conflicting authorization. Therefore, the following protocol defines the precedence of the configuration macros.
In addition, there are some hard-coded authorization rules that cannot be modified by configuration.
*@unmappedare not allowed to do any job management commands (e.g. submitting, removing, or modifying jobs). This prevents these operations from being done by unauthenticated users and users who are authenticated but lacking a name in the map file.
execute-side@matchsessionis automatically granted READ access to the condor_schedd and DAEMON access to the condor_shadow.
An example of the configuration variables for the user-side authorization is derived from the necessary access levels as described in Section 3.6.1.
ALLOW_READ = *@cs.wisc.edu/* ALLOW_WRITE = *@cs.wisc.edu/*.cs.wisc.edu ALLOW_ADMINISTRATOR = firstname.lastname@example.org/*.cs.wisc.edu ALLOW_CONFIG = email@example.com/*.cs.wisc.edu ALLOW_NEGOTIATOR = firstname.lastname@example.org/condor.cs.wisc.edu, \ email@example.com/condor2.cs.wisc.edu ALLOW_DAEMON = firstname.lastname@example.org/*.cs.wisc.edu # Clear out any old-style HOSTALLOW settings: HOSTALLOW_READ = HOSTALLOW_WRITE = HOSTALLOW_DAEMON = HOSTALLOW_NEGOTIATOR = HOSTALLOW_ADMINISTRATOR = HOSTALLOW_OWNER =
This example configuration authorizes
any authenticated user in the
cs.wisc.edu domain to
carry out a request that requires the
READ access level
from any machine.
Any user in the
cs.wisc.edu domain may
carry out a request that requires the
WRITE access level
from any machine in the
Only the user called
carry out a request that requires the
ADMINISTRATOR access level
from any machine in the
The administrator, logged into any machine within
cs.wisc.edu domain is authorized at the
CONFIG access level.
Only the negotiator daemon, running as
condor on the two central managers
NEGOTIATOR access level.
And, the last line of the example presumes that there is a
user called condor, and that the daemons have all been started
up as this user.
It authorizes only programs (which will be the daemons)
carry out requests that require the
DAEMON access level,
where the commands originate from
any machine in the
In the local configuration file for each host, the host's owner should be authorized as the owner of the machine. An example of the entry in the local configuration file:
ALLOW_OWNER = email@example.com/hostname.cs.wisc.eduIn this example the owner has a login of
username, and the machine's name is represented by
If the authorization policy denies a network request, an explanation
of why the request was denied is printed in the log file of the daemon
that denied the request. The line in the log file contains the words
To get Condor to generate a similar explanation of why requests are
accepted, add D_SECURITY to the daemon's debug options (and
restart or reconfig the daemon). The line in the log file for these
cases will contain the words
PERMISSION GRANTED. If you do not
want to see a full explanation but just want to see when requests
are made, add D_COMMAND to the daemon's debug options.
If the authorization policy makes use of host or domain names, then be aware that Condor depends on DNS to map IP addresses to names. The security and accuracy of your DNS service is therefore a requirement. Typos in DNS mappings are an occasional source of unexpected behavior. If the authorization policy is not behaving as expected, carefully compare the names in the policy with the host names Condor mentions in the explanations of why requests are granted or denied.
To set up and configure secure communications in Condor, authentication, encryption, and integrity checks can be used. However, these come at a cost: performing strong authentication can take a significant amount of time, and generating the cryptographic keys for encryption and integrity checks can take a significant amount of processing power.
The Condor system makes many network connections between different daemons. If each one of these was to be authenticated, and new keys were generated for each connection, Condor would not be able to scale well. Therefore, Condor uses the concept of sessions to cache relevant security information for future use and greatly speed up the establishment of secure communications between the various Condor daemons.
A new session is established the first time a connection is made from one daemon to another. Each session has a fixed lifetime after which it will expire and a new session will need to be created again. But while a valid session exists, it can be re-used as many times as needed, thereby preventing the need to continuously re-establish secure connections. Each entity of a connection will have access to a session key that proves the identity of the other entity on the opposing side of the connection. This session key is exchanged securely using a strong authentication method, such as Kerberos or GSI. Other authentication methods, such as NTSSPI, FS_REMOTE, CLAIMTOBE, and ANONYMOUS, do not support secure key exchange. An entity listening on the wire may be able to impersonate the client or server in a session that does not use a strong authentication method.
Establishing a secure session requires that either the encryption or the integrity options be enabled. If the encryption capability is enabled, then the session will be restarted using the session key as the encryption key. If integrity capability is enabled, then the check sum includes the session key even though it is not transmitted. Without either of these two methods enabled, it is possible for an attacker to use an open session to make a connection to a daemon and use that connection for nefarious purposes. It is strongly recommended that if you have authentication turned on, you should also turn on integrity and/or encryption.
The configuration parameter SEC_DEFAULT_NEGOTIATION will allow
a user to set the default level of secure sessions in Condor.
Like other security settings, the possible values for this parameter can be
If you disable sessions and you have authentication turned
on, then most authentication (other than commands like
condor_submit) will fail because Condor requires sessions when you
have security turned on.
On the other hand, if you are not using strong security in Condor, but
you are relying on the default host-based security, turning off
sessions may be useful in certain situations. These might include debugging problems
with the security session management or slightly decreasing the memory
consumption of the daemons, which keep track of the sessions in use.
Session lifetimes for specific daemons are already properly configured in the default installation of Condor. Condor tools such as condor_q and condor_status create a session that expires after one minute. Theoretically they should not create a session at all, because the session cannot be reused between program invocations, but this is difficult to do in the general case. This allows a very small window of time for any possible attack, and it helps keep the memory footprint of running daemons down, because they are not keeping track of all of the sessions. The session durations may be manually tuned by using macros in the configuration file, but this is not recommended.
This section describes the mechanisms for setting up Condor's host-based security. This is now an outdated form of implementing security levels for machine access. It remains available and documented for purposes of backward compatibility. If used at the same time as the user-based authorization, the two specifications are merged together.
The host-based security paradigm allows control over which machines can join a Condor pool, which machines can find out information about your pool, and which machines within a pool can perform administrative commands. By default, Condor is configured to allow anyone to view or join a pool. It is recommended that this parameter is changed to only allow access from machines that you trust.
This section discusses how the host-based security works inside Condor. It lists the different levels of access and what parts of Condor use which levels. There is a description of how to configure a pool to grant or deny certain levels of access to various machines. Configuration examples and the settings of configuration variables using the condor_config_val command complete this section.
Inside the Condor daemons or tools that use DaemonCore (see section 3.9 for details), most tasks are accomplished by sending commands to another Condor daemon. These commands are represented by an integer value to specify which command is being requested, followed by any optional information that the protocol requires at that point (such as a ClassAd, capability string, etc). When the daemons start up, they will register which commands they are willing to accept, what to do with arriving commands, and the access level required for each command. When a command request is received by a daemon, Condor identifies the access level required and checks the IP address of the sender to verify that it satisfies the allow/deny settings from the configuration file. If permission is granted, the command request is honored; otherwise, the request will be aborted.
Settings for the access levels in the global configuration file will affect all the machines in the pool. Settings in a local configuration file will only affect the specific machine. The settings for a given machine determine what other hosts can send commands to that machine. If a machine foo is to be given administrator access on machine bar, place foo in bar's configuration file access list (not the other way around).
The following are the various access levels that commands within Condor can be registered with:
IMPORTANT: For a machine to join a Condor pool, the machine must have both WRITE permission AND READ permission. WRITE permission is not enough.
IMPORTANT: Giving ADMINISTRATOR privileges to a machine grants administrator access for the pool to ANY USER on that machine. This includes any users who can run Condor jobs on that machine. It is recommended that ADMINISTRATOR access is granted with due diligence.
ADMINISTRATOR and NEGOTIATOR access default to the central manager machine. OWNER access defaults to the local machine, as well as any machines given with ADMINISTRATOR access. CONFIG access is not granted to any machine as its default. These defaults are sufficient for most pools, and should not be changed without a compelling reason. If machines other than the default are to have to have OWNER access, they probably should also have ADMINISTRATOR access. By granting machines ADMINISTRATOR access, they will automatically have OWNER access, given how OWNER access is set within the configuration.
Here is a sample security configuration:
ALLOW_ADMINISTRATOR = $(CONDOR_HOST) ALLOW_OWNER = $(FULL_HOSTNAME), $(ALLOW_ADMINISTRATOR) ALLOW_READ = * ALLOW_WRITE = * ALLOW_NEGOTIATOR = $(COLLECTOR_HOST) ALLOW_NEGOTIATOR_SCHEDD = $(COLLECTOR_HOST), $(FLOCK_NEGOTIATOR_HOSTS) ALLOW_WRITE_COLLECTOR = $(ALLOW_WRITE), $(FLOCK_FROM) ALLOW_WRITE_STARTD = $(ALLOW_WRITE), $(FLOCK_FROM) ALLOW_READ_COLLECTOR = $(ALLOW_READ), $(FLOCK_FROM) ALLOW_READ_STARTD = $(ALLOW_READ), $(FLOCK_FROM) ALLOW_CLIENT = *
This example configuration presumes that the condor_collector and condor_negotiator daemons are running on the same machine.
For each access level, an ALLOW or a DENY may be added.
Multiple machine entries in the configuration files may be separated by either a space or a comma. The machines may be listed by
To resolve an entry that falls into both allow and deny: individual machines have a higher order of precedence than wild card entries, and host names with a wild card have a higher order of precedence than IP subnets. Otherwise, DENY has a higher order of precedence than ALLOW. This is how most people would intuitively expect it to work.
In addition, the above access levels may be specified on a per-daemon basis, instead of machine-wide for all daemons. Do this with the subsystem string (described in section 3.3.1 on Subsystem Names), which is one of: STARTD, SCHEDD, MASTER, NEGOTIATOR, or COLLECTOR. For example, to grant different read access for the condor_schedd:
ALLOW_READ_SCHEDD = <list of machines>
Here are more examples of configuration settings. Notice that ADMINISTRATOR access is only granted through an ALLOW setting to explicitly grant access to a small number of machines. We recommend this.
ALLOW_ADMINISTRATOR = $(CONDOR_HOST) ALLOW_OWNER = $(FULL_HOSTNAME), $(ALLOW_ADMINISTRATOR)
ALLOW_READ = *.ncsa.uiuc.edu ALLOW_WRITE = *.ncsa.uiuc.edu ALLOW_ADMINISTRATOR = $(CONDOR_HOST) ALLOW_OWNER = $(FULL_HOSTNAME), $(ALLOW_ADMINISTRATOR)
ALLOW_WRITE = *.ncsa.uiuc.edu, *.math.uiuc.edu DENY_WRITE = lab-*.edu, *.lab.uiuc.edu, 177.55.* ALLOW_ADMINISTRATOR = bigcheese.ncsa.uiuc.edu ALLOW_OWNER = $(FULL_HOSTNAME), $(ALLOW_ADMINISTRATOR)
\, to continue a long list of machines onto multiple lines, making it more readable. This works for all configuration file entries, not just host access entries.
ALLOW_READ = *.ncsa.uiuc.edu, *.cs.wisc.edu ALLOW_WRITE = *.ncsa.uiuc.edu, raven.cs.wisc.edu ALLOW_ADMINISTRATOR = $(CONDOR_HOST), bigcheese.ncsa.uiuc.edu, \ biggercheese.uiuc.edu ALLOW_OWNER = $(FULL_HOSTNAME), $(ALLOW_ADMINISTRATOR)
DENY_READ = *.mil ALLOW_READ_SCHEDD = *.ncsa.uiuc.edu ALLOW_WRITE = *.ncsa.uiuc.edu ALLOW_ADMINISTRATOR = $(CONDOR_HOST), bigcheese.ncsa.uiuc.edu, \ biggercheese.uiuc.edu ALLOW_ADMINISTRATOR_NEGOTIATOR = biggercheese.uiuc.edu ALLOW_OWNER = $(FULL_HOSTNAME), $(ALLOW_ADMINISTRATOR)
A new security feature introduced in Condor version 6.3.2 enables more fine-grained control over the configuration settings that can be modified remotely with the condor_config_val command. The manual page for condor_config_val on page details how to use condor_config_val to modify configuration settings remotely. Since certain configuration attributes can have a large impact on the functioning of the Condor system and the security of the machines in a Condor pool, it is important to restrict the ability to change attributes remotely.
For each security access level described, the Condor administrator can define which configuration settings a host at that access level is allowed to change. Optionally, the administrator can define separate lists of settable attributes for each Condor daemon, or the administrator can define one list that is used by all daemons.
For each command that requests a change in configuration setting, Condor searches all the different possible security access levels to see which, if any, the request satisfies. (Some hosts can qualify for multiple access levels. For example, any host with ADMINISTRATOR permission probably has WRITE permission also). Within the qualified access level, Condor searches for the list of attributes that may be modified. If the request is covered by the list, the request will be granted. If not covered, the request will be refused.
The default configuration shipped with Condor is exceedingly restrictive. Condor users or administrators cannot set configuration values from remote hosts with condor_config_val. Enabling this feature requires a change to the settings in the configuration file. Use this security feature carefully. Grant access only for attributes which you need to be able to modify in this manner, and grant access only at the most restrictive security level possible.
The most secure use of this feature allows Condor users to set attributes in the configuration file which are not used by Condor directly. These are custom attributes published by various Condor daemons with the <SUBSYS>_ATTRS setting described in section 3.3.5 on page . It is secure to grant access only to modify attributes that are used by Condor to publish information. Granting access to modify settings used to control the behavior of Condor is not secure. The goal is to ensure no one can use the power to change configuration attributes to compromise the security of your Condor pool.
The control lists are defined by configuration settings that contain SETTABLE_ATTRS in their name. The name of the control lists have the following form:
The two parts of this name that can vary are
<PERMISSION-LEVEL> and the
<PERMISSION-LEVEL> can be any of the security access levels
described earlier in this section.
Examples include WRITE, OWNER, and CONFIG.
<SUBSYS> is an optional portion of the name.
It can be used to
define separate rules for which configuration attributes can be set
for each kind of Condor daemon (for example,
STARTD, SCHEDD, and MASTER).
There are many configuration settings that can be defined differently
for each daemon that use this
<SUBSYS> naming convention.
See section 3.3.1 on
page for a list.
If there is no daemon-specific value for a given daemon, Condor will
look for SETTABLE_ATTRS_<PERMISSION-LEVEL> .
Each control list is defined by a comma-separated list of attribute
names which should be allowed to be modified.
The lists can contain wild cards characters (
Some examples of valid definitions of control lists with explanations:
SETTABLE_ATTRS_CONFIG = *Grant unlimited access to modify configuration attributes to any request that came from a machine in the CONFIG access level. This was the default behavior before Condor version 6.3.2.
SETTABLE_ATTRS_ADMINISTRATOR = *_DEBUG, MAX_*_LOGGrant access to change any configuration setting that ended with
_DEBUG(for example, STARTD_DEBUG) and any attribute that matched
MAX_*_LOG(for example, MAX_SCHEDD_LOG) to any host with ADMINISTRATOR access.
STARTD.SETTABLE_ATTRS_OWNER = HasDataSetAllows any request to modify the HasDataSet attribute that came from a host with OWNER access. By default, OWNER covers any request originating from the local host, plus any machines listed in the ADMINISTRATOR level. Therefore, any Condor job would qualify for OWNER access to the machine where it is running. So, this setting would allow any process running on a given host, including a Condor job, to modify the HasDataSet variable for that host. HasDataSet is not used by Condor, it is an invented attribute included in the STARTD_ATTRS setting in order for this example to make sense.
This topic is now addressed in more detail in section 3.7, which explains network communication in Condor.
On a Unix system, UIDs (User IDentification numbers) form part of an operating system's tools for maintaining access control. Each executing program has a UID, a unique identifier of a user executing the program. This is also called the real UID. A common situation has one user executing the program owned by another user. Many system commands work this way, with a user (corresponding to a person) executing a program belonging to (owned by) root. Since the program may require privileges that root has which the user does not have, a special bit in the program's protection specification (a setuid bit) allows the program to run with the UID of the program's owner, instead of the user that executes the program. This UID of the program's owner is called an effective UID.
Condor works most smoothly when its daemons run as root. The daemons then have the ability to switch their effective UIDs at will. When the daemons run as root, they normally leave their effective UID and GID (Group IDentification) to be those of user and group condor. This allows access to the log files without changing the ownership of the log files. It also allows access to these files when the user condor's home directory resides on an NFS server. root can not normally access NFS files.
If there is no condor user and group on the system, an administrator can specify which UID and GID the Condor daemons should use when they do not need root privileges in two ways: either with the CONDOR_IDS environment variable or the CONDOR_IDS configuration variable. In either case, the value should be the UID integer, followed by a period, followed by the GID integer. For example, if a Condor administrator does not want to create a condor user, and instead wants their Condor daemons to run as the daemon user (a common non-root user for system daemons to execute as), the daemon user's UID was 2, and group daemon had a GID of 2, the corresponding setting in the Condor configuration file would be CONDOR_IDS = 2.2.
On a machine where a job is submitted, the condor_schedd daemon changes its effective UID to root such that it has the capability to start up a condor_shadow daemon for the job. Before a condor_shadow daemon is created, the condor_schedd daemon switches back to root, so that it can start up the condor_shadow daemon with the (real) UID of the user who submitted the job. Since the condor_shadow runs as the owner of the job, all remote system calls are performed under the owner's UID and GID. This ensures that as the job executes, it can access only files that its owner could access if the job were running locally, without Condor.
On the machine where the job executes, the job runs either as the submitting user or as user nobody, to help ensure that the job cannot access local resources or do harm. If the UID_DOMAIN matches, and the user exists as the same UID in password files on both the submitting machine and on the execute machine, the job will run as the submitting user. If the user does not exist in the execute machine's password file and SOFT_UID_DOMAIN is True, then the job will run under the submitting user's UID anyway (as defined in the submitting machine's password file). If SOFT_UID_DOMAIN is False, and UID_DOMAIN matches, and the user is not in the execute machine's password file, then the job execution attempt will be aborted.
While we strongly recommend starting up the Condor daemons as root, we understand that it is not always possible to do so. The main problems of not running Condor daemons as root appear when one Condor installation is shared by many users on a single machine, or if machines are set up to only execute Condor jobs. With a submit-only installation for a single user, there is no need for or benefit from running as root.
The effects of Condor of running both with and without root access are classified for each daemon:
In addition, some system information cannot be obtained without root access on some platforms. As a result, when running without root access, the condor_startd must call other programs such as uptime, to get this information. This is much less efficient than getting the information directly from the kernel, as is done when running as root. On Linux, this information is available without root access, so it is not a concern on those platforms.
If all of Condor cannot be run as root, at least consider installing the condor_startd as setuid root. That would solve both problems. Barring that, install it as a setgid sys or kmem program, depending on whatever group has read access to /dev/kmem on the system. That would solve the system information problem.
Consider installing condor_submit as a setgid condor program so that at least the stdout, stderr and user log files get created with the right permissions. If condor_submit is a setgid program, it will automatically set its umask to 002 and create group-writable files. This way, the simple case of a job that only writes to stdout and stderr will work. If users have programs that open their own files, they will need to know and set the proper permissions on the directories they submit from.
If Condor is not run as root, then choose almost any user name. A common choice is to set up and use the condor user; this simplifies the setup, because Condor will look for its configuration files in the condor user's directory. If condor is not selected, then the configuration must be placed properly such that Condor can find its configuration files.
If users will be submitting jobs as a user different than the user Condor is running as (perhaps you are running as the condor user and users are submitting as themselves), then users have to be careful to only have file permissions properly set up to be accessible by the user Condor is using. In practice, this means creating world-writable directories for output from Condor jobs. This creates a potential security risk, in that any user on the machine where the job is submitted can alter the data, remove it, or do other undesirable things. It is only acceptable in an environment where users can trust other users.
Normally, users without root access who wish to use Condor on their machines create a condor home directory somewhere within their own accounts and start up the daemons (to run with the UID of the user). As in the case where the daemons run as user condor, there is no ability to switch UIDs or GIDs. The daemons run as the UID and GID of the user who started them. On a machine where jobs are submitted, the condor_shadow daemons all run as this same user. But, if other users are using Condor on the machine in this environment, the condor_shadow daemons for these other users' jobs execute with the UID of the user who started the daemons. This is a security risk, since the Condor job of the other user has access to all the files and directories of the user who started the daemons. Some installations have this level of trust, but others do not. Where this level of trust does not exist, it is best to set up a condor account and group, or to have each user start up their own Personal Condor submit installation.
When a machine is an execution site for a Condor job, the Condor job executes with the UID of the user who started the condor_startd daemon. This is also potentially a security risk, which is why we do not recommend starting up the execution site daemons as a regular user. Use either root or a user such as condor that exists only to run Condor jobs.
Under Unix, Condor runs jobs either as the user that submitted the jobs, or as the user called nobody. Condor uses user nobody if the value of the UID_DOMAIN configuration variable of the submitting and executing machines are different or if STARTER_ALLOW_RUNAS_OWNER is false or if the job ClassAd contains RunAsOwner=False. Under Windows, Condor by default runs jobs under a dynamically created local account that exists for the duration of the job, but it can optionally run the job as the user account that owns the job if STARTER_ALLOW_RUNAS_OWNER is True and the job contains RunAsOwner=True.
When Condor cleans up after executing a vanilla universe job, it does the best that it can by deleting all of the processes started by the job. During the life of the job, it also does its best to track the CPU usage of all processes created by the job. There are a variety of mechanisms used by Condor to detect all such processes, but, in general, the only foolproof mechanism is for the job to run under a dedicated execution account (as it does under Windows by default). With all other mechanisms, it is possible to fool Condor, and leave processes behind after Condor has cleaned up. In the case of a shared account, such as the Unix user nobody, it is possible for the job to leave a lurker process lying in wait for the next job run as nobody. The lurker process may prey maliciously on the next nobody user job, wreaking havoc.
Condor could prevent this problem by simply killing all processes run by the nobody user, but this would annoy many system administrators. The nobody user is often used for non-Condor system processes. It may also be used by other Condor jobs running on the same machine, if it is a multi-processor machine.
Condor provides a two-part solution to this difficulty. First, create user accounts specifically for Condor to use instead of user nobody. These can be low-privilege accounts, as the nobody user is. Create one of these accounts for each job execution slot per computer, so that distinct users can be used for concurrent processes. This prevents malicious behavior between processes running on distinct slots. Section 3.13.9 details slots. For a sample machine with two compute slots, create two users that are intended only to be used by Condor. As an example, call them cndrusr1 and cndrusr2. Tell Condor about these users with the SLOT<N>_USER configuration variables, where <N> is replaced with the slot number. In this example:
SLOT1_USER = cndrusr1 SLOT2_USER = cndrusr2
Then tell Condor that these accounts are intended only to be used by Condor, so Condor can kill all the processes belonging to these users upon job completion. The configuration variable DEDICATED_EXECUTE_ACCOUNT_REGEXP is introduced and set to a regular expression that matches the account names we have just created.
DEDICATED_EXECUTE_ACCOUNT_REGEXP = cndrusr[0-9]+
Finally, tell Condor not to run jobs as the job owner:
STARTER_ALLOW_RUNAS_OWNER = False
Tracking process family by login "cndrusr1"
Every executing process has a notion of its current working directory. This is the directory that acts as the base for all file system access. There are two current working directories for any Condor job: one where the job is submitted and a second where the job executes. When a user submits a job, the submit-side current working directory is the same as for the user when the condor_submit command is issued. The initialdir submit command may change this, thereby allowing different jobs to have different working directories. This is useful when submitting large numbers of jobs. This submit-side current working directory remains unchanged for the entire life of a job. The submit-side current working directory is also the working directory of the condor_shadow daemon. This is particularly relevant for standard universe jobs, since file system access for the job goes through the condor_shadow daemon, and therefore all accesses behave as if they were executing without Condor.
There is also an execute-side current working directory. For standard universe jobs, it is set to the execute subdirectory of Condor's home directory. This directory is world-writable, since a Condor job usually runs as user nobody. Normally, standard universe jobs would never access this directory, since all I/O system calls are passed back to the condor_shadow daemon on the submit machine. In the event, however, that a job crashes and creates a core dump file, the execute-side current working directory needs to be accessible by the job so that it can write the core file. The core file is moved back to the submit machine, and the condor_shadow daemon is informed. The condor_shadow daemon sends e-mail to the job owner announcing the core file, and provides a pointer to where the core file resides in the submit-side current working directory.
Section 3.6.13 discusses why, under most circumstances, it is beneficial to run the Condor daemons as root. In situations where multiple users are involved or where Condor is responsible for enforcing a machine owner's policy, running as root is the only way for Condor to do its job correctly and securely.
Unfortunately, this requirement of running Condor as root is at odds with a well-established goal of security-conscious administrators: keeping the amount of software that runs with superuser privileges to a minimum. Condor's nature as a large distributed system that routinely communicates with potentially untrusted components over the network further aggravates this goal.
The privilege separation (PrivSep) effort in Condor aims to minimize the amount of code that needs root-level access, while still giving Condor the tools it needs to work properly. Note that PrivSep is currently only available for execute side functionality, and is not implemented on Windows.
In the PrivSep model, all logic in Condor that requires superuser privilege is contained in a small component called the PrivSep Kernel. The Condor daemons execute as an unprivileged account. They explicitly request action from the PrivSep Kernel whenever root-level operations are needed.
The PrivSep model then prevents the following attack scenario. In the attack scenario, an attacker has found an exploit in the condor_startd that allows for execution of arbitrary code on that daemon's behalf. This gives the attacker root access and therefore control over any machine on which the condor_startd is running as root and the exploit can be exercised. Under the PrivSep model, the condor_startd no longer runs as root. This prevents the attacker from taking arbitrary action as root. Further, limits on requested actions from the PrivSep Kernel contain and restrict the attacker's sphere of influence.
The following section describes the configuration necessary to enable PrivSep for an execute-side Condor installation. After this is a detailed description of the services that the PrivSep Kernel provides to Condor, and how it limits the allowed root-level actions.
The PrivSep Kernel is implemented as two programs: the condor_root_switchboard and the condor_procd. Both are contained in the sbin directory of the Condor distribution. When Condor is running in PrivSep mode, these are to be the only two Condor daemons that run with root privilege.
Each of these binaries must be accessible on the file system via a trusted path. A trusted path ensures that no user (other than root) can alter the binary or path to the binary referred to. To ensure that the paths to these binaries are trusted, use only root-owned directories, and set the permissions on these directories to deny write access to all but root. The binaries themselves must also be owned by root and not writable by any other. The condor_root_switchboard program additionally is installed with the setuid bit set. The following command properly sets the permissions on the condor_root_switchboard binary:
chmod 4755 /opt/condor/release/sbin/condor_root_switchboard
The PrivSep Kernel has its own configuration file. This file must be /etc/condor/privsep_config. The format of this file is different than a Condor configuration file. It consists of lines with ``key = value'' pairs. Lines with only white space or lines with ``#'' as the first non-white space character are ignored.
In the PrivSep Kernel configuration file, some configuration settings are interpreted as single values, while others are interpreted as lists. To populate a list with multiple values, use multiple lines with the same key. For example, the following configures the valid-dirs setting as a list with two entries:
valid-dirs = /opt/condor/execute_1 valid-dirs = /opt/condor/execute_2It is an error to have multiple lines with the same key for a setting that is not interpreted as a list.
Some PrivSep Kernel configuration file settings require a list of UIDs or GIDs, and these allow for a more specialized syntax. User and group IDs can be specified either numerically or textually. Multiple list entries may be given on a single line using the : (colon) character as a delimiter. In addition, list entries may specify a range of IDs using a - (dash) character to separate the minimum and maximum IDs included. The * (asterisk) character on the right-hand side of such a range indicates that the range extends to the maximum possible ID. The following example builds a complex list of IDs:
valid-target-uids = nobody : nfsuser1 : nfsuser2 valid-target-uids = condor_run_1 - condor_run_8 valid-target-uids = 800 - *
If condor_run_1 maps to UID 701, and condor_run_8 maps to UID 708, then this range specifies the 8 UIDs of 701 through 708 (inclusive).
The following settings are required to configure the PrivSep Kernel:
Here is an example of a full privsep_config file. This file gives the condor account access to the PrivSep Kernel. Condor's use of this execute machine will be restricted to a set of eight dedicated accounts, along with the users group. Condor's $(EXECUTE) directory and the condor_procd executable are also specified, as required.
valid-caller-uids = condor valid-caller-gids = condor valid-target-uids = condor_run_1 - condor_run_8 valid-target-gids = users : condor_run_1 - condor_run_8 valid-dirs = /opt/condor/local/execute procd-executable = /opt/condor/release/sbin/condor_procd
Once the PrivSep Kernel is properly installed and configured, Condor's configuration must be updated to specify that PrivSep should be used. The Condor configuration variable PRIVSEP_ENABLED is a boolean flag serving this purpose. In addition, Condor must be told where the condor_root_switchboard binary is located using the PRIVSEP_SWITCHBOARD setting. The following example illustrates:
PRIVSEP_ENABLED = True PRIVSEP_SWITCHBOARD = $(SBIN)/condor_root_switchboard
Finally, note that while the condor_procd is in general an optional component of Condor, it is required when PrivSep is in use. If PRIVSEP_ENABLED is True, the condor_procd will be used regardless of the USE_PROCD setting. Details on these Condor configuration variables are in section 3.3.28 for PrivSep variables and section 3.3.18 for condor_procd variables.
This section describes the root-enabled operations that the PrivSep Kernel makes available to Condor. The PrivSep Kernel's interface is designed to provide only operations needed by Condor in order to function properly. Each operation is further restricted based on the PrivSep Kernel's configuration settings.
The following list describes each action that can be performed via the PrivSep Kernel, along with the limitations enforced on how it may be used. The terms valid target users, valid target groups, and valid directories refer respectively to the settings for valid-target-uids, valid-target-gids, and valid-dirs from the PrivSep Kernel's configuration.
After launching a program as a valid target user, the PrivSep Kernel allows Condor limited control over its execution. The following operations are supported on a program executed via the PrivSep Kernel:
By sufficiently constraining the valid target accounts and valid directories to which the PrivSep Kernel allows access, the ability of a compromised Condor daemon to do damage can be considerably reduced.
glexec is a tool that provides a sudo-like capability in a grid environment. glexec takes an X.509 proxy and a command to run as inputs, and maps the proxy to a local identity (that is, a Unix UID), which it then uses to execute the command. Like the condor_root_switchboard command, which provides similar functionality for Condor's PrivSep mode (see section 3.6.14), glexec must be installed as a root-owned setuid program. See http://www.nikhef.nl/grid/lcaslcmaps/glexec/ for more information about glexec.
Condor can interoperate with glexec, using it in a similar way to how the condor_root_switchboard is used when running Condor in PrivSep mode. The condor_starter uses glexec when launching a job, in order to give the job a separate UID from that of the Condor daemons. glexec is also used when performing maintenance actions such as cleaning up a job's files and processes, which cannot be done well directly under the Condor daemons' UID due to permissions. A consequence of this type of integration with glexec is that the execution of a single Condor job results in several glexec invocations, and each must map the proxy to the same UID. It is thus important to ensure that glexec is configured to provide this guarantee.
Configuration for glexec support is straightforward. The boolean configuration variable GLEXEC_JOB must be set True on execute machines where glexec is to be used. Condor also must be given the full path to the glexec binary using the GLEXEC configuration variable. Note that Condor must be started as a non-root user when glexec is used. This is because when Condor runs as root, it can perform actions as other UIDs arbitrarily, and glexec's services are not needed. Finally, for a job to execute properly in the mode utilizing glexec, the job must be submitted with a proxy specified via the x509userproxy command in its submit description file, since a proxy is needed as input to glexec.
Earlier versions of Condor employed a different form of glexec support, where the condor_starter daemon ran under the same UID as the job. This feature was enabled using the GLEXEC_STARTER configuration variable. This configuration variable is no longer used, and it is replaced by the GLEXEC_JOB configuration variable, to enable usage of glexec.