/*
Copyright 1990, Daniel J. Bernstein. All rights reserved.

Please address any questions or comments to the author at brnstnd@acf10.nyu.edu.
*/

The #include's should be rewritten.

All functions return 0 on success.

Environment variables: KRB5RCACHETYPE, KRB5RCACHENAME, KRB5RCACHEDIR,
and TMPDIR. Obsolete: KRB5RCACHE.

All header files are both ANSI-compatible and K&R-compatible. The .c files
are only ANSI compatible. Everything passes gcc -Wall -ansi -pedantic.

Strings are freed using FREE(), which is defined in terms of free(). 

The error header files should be redone.

The header files don't use __ because that's reserved.

Each .c file assumes <malloc.h>. rc_io.c assumes fsync() and a gaggle of
error codes. These assumptions are not as portable as the code itself.


rcache.c:

The rcache.c compatibility interface's type registration is a no-op; it
simply passes the type name on to rc_base.h. rcache.h is obsolete; use
rc_base.h if possible.

There are some slight differences between rcache.c and the prototypes I
saw in krb/func-proto.h. Don't look at me, it's your interface.

rcache.c's get_name doesn't fill with zeros unless strncpy does.


rc_base.c:

If SEMAPHORE is defined and <semaphore.h> exists when rc_base.c is
compiled, all access to the type list will be protected by appropriate
semaphore ups and downs. This is untested.

It doesn't take linker magic to preregister types. Just change the
typehead initialization in rc_base.c, with an appropriate include file
setting the ops.


rc_dfl.c:

If NOIOSTUFF is defined when rc_dfl.c is compiled, all dfl rcaches will
be per-process. This is untested.

Provided that separate threads use separate rcaches, rc_dfl.c is safe
for multithreading.

Getting the name of a cache is only valid after it is created and before
it is closed. Recovering a cache is only valid after it has been created.

krb5_unparse_name had better produce a zero-terminated string.

rc_dfl.c isn't smart enough to try expunge/retry upon a malloc error.
Then again, such an error indicates that the whole system's about to die;
without real memory management there's no good solution.

HASHSIZE can be defined at compile time. It defaults to 997 in rc_dfl.c.
EXCESSREPS can be defined at compile time. It defaults to 30 in rc_dfl.c.

Hopefully adding a deltat to a time to compare to another time cannot
overflow.

In rc_dfl's struct dfl_data, the name field is never freed, even though
it may be malloced by io_creat on a generate-name call. This should not
be a problem: a single process should not be opening and closing many
rcaches. One fix would be another field to indicate whether the string
was malloced or not; normally this is an unstated characteristic of a
char pointer, but here it would have to be explicit.


rc_io.c:

rc_io.c assumes that siginterrupt() is not set. If siginterrupt() is set
and a signal occurs during, say, close(), then the close will fail.

On a machine without fsync() you might as well not use the disk at all.