Independent of success, the delay time is reset to its zero default value when the PAM service module returns control to the application. The delay occurs after all authentication modules have been called, but before control is returned to the service application.
When using this function the programmer should check if it is available with:
#ifdef HAVE_PAM_FAIL_DELAY .... #endif /* HAVE_PAM_FAIL_DELAY */
For applications written with a single thread that are event driven in nature, generating this delay may be undesirable. Instead, the application may want to register the delay in some other way. For example, in a single threaded server that serves multiple authentication requests from a single event loop, the application might want to simply mark a given connection as blocked until an application timer expires. For this reason the delay function can be changed with the PAM_FAIL_DELAY item. It can be queried and set with pam_get_item(3) and pam_set_item(3) respectively. The value used to set it should be a function pointer of the following prototype:
void (*delay_fn)(int retval, unsigned usec_delay, void *appdata_ptr);
The arguments being the retval return code of the module stack, the usec_delay micro-second delay that libpam is requesting and the appdata_ptr that the application has associated with the current pamh. This last value was set by the application when it called pam_start(3) or explicitly with pam_set_item(3).
Note that the PAM_FAIL_DELAY item is set to NULL by default. This indicates that PAM should perform a random delay as described above when authentication fails and a delay has been suggested. If an application does not want the PAM library to perform any delay on authentication failure, then the application must define a custom delay function that executes no statements and set the PAM_FAIL_DELAY item to point to this function.
It is often possible to attack an authentication scheme by exploiting the time it takes the scheme to deny access to an applicant user. In cases of short timeouts, it may prove possible to attempt a brute force dictionary attack -- with an automated process, the attacker tries all possible passwords to gain access to the system. In other cases, where individual failures can take measurable amounts of time (indicating the nature of the failure), an attacker can obtain useful information about the authentication process. These latter attacks make use of procedural delays that constitute a covert channel of useful information.
To minimize the effectiveness of such attacks, it is desirable to introduce a random delay in a failed authentication process. Preferable this value should be set by the application or a special PAM module. Standard PAM modules should not modify the delay unconditional.
For example, a login application may require a failure delay of roughly 3 seconds. It will contain the following code:
pam_fail_delay (pamh, 3000000 /* micro-seconds */ ); pam_authenticate (pamh, 0);
if the modules do not request a delay, the failure delay will be between 1.5 and 4.5 seconds.
However, the modules, invoked in the authentication process, may also request delays:
module #1: pam_fail_delay (pamh, 2000000); module #2: pam_fail_delay (pamh, 4000000);
The pam_fail_delay function is an Linux-PAM extension.