Issues to be addressed for src/util/et: -*- text -*-

Non-thread-safe aspects:

error_message uses a static buffer for "unknown error code" messages;

a per-thread buffer may be better, but that depends on dynamic

allocation working.  A caller-provided buffer would be best, but

that's a API change.

initialize_foo_error_table uses a global linked list hung off an

unprotected variable in the library.  {add,remove}_error_table do

likewise, but can be changed without externally visible effect.

Workaround: Use a global lock for all calls to error_message and

com_err, and when adding or removing error tables.

API divergence:

Transarc and Heimdal both have APIs that are different from this

version.  (Specifics?)

Karl Ramm has offered to try to combine them.

Workaround:

Reference counting:

If libraries are dynamically loaded and unloaded, and the init/fini

functions add and remove error tables for *other* libraries they

depend on (e.g., if a dynamically loadable Zephyr library's fini

function removes the krb4 library error table and then dlcloses the

krb4 library, while another dlopen reference keeps the krb4 library

around), the error table is kept; a table must be removed the same

number of times it was added before the library itself can be

discarded.

It's not implemented as a reference count, but the effect is the same.

Fix needed: Update documentation.

64-bit support:

Values are currently computed as 32-bit values, sign-extended to

"long", and output with "L" suffixes.  Type errcode_t is "long".

Kerberos uses a separately chosen signed type of at least 32 bits for

error codes.  The com_err library only look at the low 32 bits, so

this is mostly just an issue for application code -- if anything

truncates to 32 bits, and then widens without sign-extending, the

value may not compare equal to the macro defined in the .h file.  This

isn't anything unusual for signed constants, of course, just something

to keep in mind....

Workaround: Always use signed types of at least 32 bits for error

codes.

man page:

No documentation on add_error_table/remove_error_table interfaces,

even though they're the new, preferred interface.