Notes: 2001-09-24
-----------------

This "description" (if one chooses to call it that) needed some major updating
so here goes. This update addresses a change being made at the same time to
OpenSSL, and it pretty much completely restructures the underlying mechanics of
the "ENGINE" code. So it serves a double purpose of being a "ENGINE internals
for masochists" document *and* a rather extensive commit log message. (I'd get
lynched for sticking all this in CHANGES or the commit mails :-).

ENGINE_TABLE underlies this restructuring, as described in the internal header
"eng_int.h", implemented in eng_table.c, and used in each of the "class" files;
tb_rsa.c, tb_dsa.c, etc.

However, "EVP_CIPHER" underlies the motivation and design of ENGINE_TABLE so
I'll mention a bit about that first. EVP_CIPHER (and most of this applies
equally to EVP_MD for digests) is both a "method" and a algorithm/mode
identifier that, in the current API, "lingers". These cipher description +
implementation structures can be defined or obtained directly by applications,
or can be loaded "en masse" into EVP storage so that they can be catalogued and
searched in various ways, ie. two ways of encrypting with the "des_cbc"
algorithm/mode pair are;

(i) directly;
     const EVP_CIPHER *cipher = EVP_des_cbc();
     EVP_EncryptInit(&ctx, cipher, key, iv);
     [ ... use EVP_EncryptUpdate() and EVP_EncryptFinal() ...]

(ii) indirectly; 
     OpenSSL_add_all_ciphers();
     cipher = EVP_get_cipherbyname("des_cbc");
     EVP_EncryptInit(&ctx, cipher, key, iv);
     [ ... etc ... ]

The latter is more generally used because it also allows ciphers/digests to be
looked up based on other identifiers which can be useful for automatic cipher
selection, eg. in SSL/TLS, or by user-controllable configuration.

The important point about this is that EVP_CIPHER definitions and structures are
passed around with impunity and there is no safe way, without requiring massive
rewrites of many applications, to assume that EVP_CIPHERs can be reference
counted. One an EVP_CIPHER is exposed to the caller, neither it nor anything it
comes from can "safely" be destroyed. Unless of course the way of getting to
such ciphers is via entirely distinct API calls that didn't exist before.
However existing API usage cannot be made to understand when an EVP_CIPHER
pointer, that has been passed to the caller, is no longer being used.

The other problem with the existing API w.r.t. to hooking EVP_CIPHER support
into ENGINE is storage - the OBJ_NAME-based storage used by EVP to register
ciphers simultaneously registers cipher *types* and cipher *implementations* -
they are effectively the same thing, an "EVP_CIPHER" pointer. The problem with
hooking in ENGINEs is that multiple ENGINEs may implement the same ciphers. The
solution is necessarily that ENGINE-provided ciphers simply are not registered,
stored, or exposed to the caller in the same manner as existing ciphers. This is
especially necessary considering the fact ENGINE uses reference counts to allow
for cleanup, modularity, and DSO support - yet EVP_CIPHERs, as exposed to
callers in the current API, support no such controls.

Another sticking point for integrating cipher support into ENGINE is linkage.
Already there is a problem with the way ENGINE supports RSA, DSA, etc whereby
they are available *because* they're part of a giant ENGINE called "openssl".
Ie. all implementations *have* to come from an ENGINE, but we get round that by
having a giant ENGINE with all the software support encapsulated. This creates
linker hassles if nothing else - linking a 1-line application that calls 2 basic
RSA functions (eg. "RSA_free(RSA_new());") will result in large quantities of
ENGINE code being linked in *and* because of that DSA, DH, and RAND also. If we
continue with this approach for EVP_CIPHER support (even if it *was* possible)
we would lose our ability to link selectively by selectively loading certain
implementations of certain functionality. Touching any part of any kind of
crypto would result in massive static linkage of everything else. So the
solution is to change the way ENGINE feeds existing "classes", ie. how the
hooking to ENGINE works from RSA, DSA, DH, RAND, as well as adding new hooking
for EVP_CIPHER, and EVP_MD.

The way this is now being done is by mostly reverting back to how things used to
work prior to ENGINE :-). Ie. RSA now has a "RSA_METHOD" pointer again - this
was previously replaced by an "ENGINE" pointer and all RSA code that required
the RSA_METHOD would call ENGINE_get_RSA() each time on its ENGINE handle to
temporarily get and use the ENGINE's RSA implementation. Apart from being more
efficient, switching back to each RSA having an RSA_METHOD pointer also allows
us to conceivably operate with *no* ENGINE. As we'll see, this removes any need
for a fallback ENGINE that encapsulates default implementations - we can simply
have our RSA structure pointing its RSA_METHOD pointer to the software
implementation and have its ENGINE pointer set to NULL.

A look at the EVP_CIPHER hooking is most explanatory, the RSA, DSA (etc) cases
turn out to be degenerate forms of the same thing. The EVP storage of ciphers,
and the existing EVP API functions that return "software" implementations and
descriptions remain untouched. However, the storage takes more meaning in terms
of "cipher description" and less meaning in terms of "implementation". When an
EVP_CIPHER_CTX is actually initialised with an EVP_CIPHER method and is about to
begin en/decryption, the hooking to ENGINE comes into play. What happens is that
cipher-specific ENGINE code is asked for an ENGINE pointer (a functional
reference) for any ENGINE that is registered to perform the algo/mode that the
provided EVP_CIPHER structure represents. Under normal circumstances, that
ENGINE code will return NULL because no ENGINEs will have had any cipher
implementations *registered*. As such, a NULL ENGINE pointer is stored in the
EVP_CIPHER_CTX context, and the EVP_CIPHER structure is left hooked into the
context and so is used as the implementation. Pretty much how things work now
except we'd have a redundant ENGINE pointer set to NULL and doing nothing.

Conversely, if an ENGINE *has* been registered to perform the algorithm/mode
combination represented by the provided EVP_CIPHER, then a functional reference
to that ENGINE will be returned to the EVP_CIPHER_CTX during initialisation.
That functional reference will be stored in the context (and released on
cleanup) - and having that reference provides a *safe* way to use an EVP_CIPHER
definition that is private to the ENGINE. Ie. the EVP_CIPHER provided by the
application will actually be replaced by an EVP_CIPHER from the registered
ENGINE - it will support the same algorithm/mode as the original but will be a
completely different implementation. Because this EVP_CIPHER isn't stored in the
EVP storage, nor is it returned to applications from traditional API functions,
there is no associated problem with it not having reference counts. And of
course, when one of these "private" cipher implementations is hooked into
EVP_CIPHER_CTX, it is done whilst the EVP_CIPHER_CTX holds a functional
reference to the ENGINE that owns it, thus the use of the ENGINE's EVP_CIPHER is
safe.

The "cipher-specific ENGINE code" I mentioned is implemented in tb_cipher.c but
in essence it is simply an instantiation of "ENGINE_TABLE" code for use by
EVP_CIPHER code. tb_digest.c is virtually identical but, of course, it is for
use by EVP_MD code. Ditto for tb_rsa.c, tb_dsa.c, etc. These instantiations of
ENGINE_TABLE essentially provide linker-separation of the classes so that even
if ENGINEs implement *all* possible algorithms, an application using only
EVP_CIPHER code will link at most code relating to EVP_CIPHER, tb_cipher.c, core
ENGINE code that is independant of class, and of course the ENGINE
implementation that the application loaded. It will *not* however link any
class-specific ENGINE code for digests, RSA, etc nor will it bleed over into
other APIs, such as the RSA/DSA/etc library code.

ENGINE_TABLE is a little more complicated than may seem necessary but this is
mostly to avoid a lot of "init()"-thrashing on ENGINEs (that may have to load
DSOs, and other expensive setup that shouldn't be thrashed unnecessarily) *and*
to duplicate "default" behaviour. Basically an ENGINE_TABLE instantiation, for
example tb_cipher.c, implements a hash-table keyed by integer "nid" values.
These nids provide the uniquenness of an algorithm/mode - and each nid will hash
to a potentially NULL "ENGINE_PILE". An ENGINE_PILE is essentially a list of
pointers to ENGINEs that implement that particular 'nid'. Each "pile" uses some
caching tricks such that requests on that 'nid' will be cached and all future
requests will return immediately (well, at least with minimal operation) unless
a change is made to the pile, eg. perhaps an ENGINE was unloaded. The reason is
that an application could have support for 10 ENGINEs statically linked
in, and the machine in question may not have any of the hardware those 10
ENGINEs support. If each of those ENGINEs has a "des_cbc" implementation, we
want to avoid every EVP_CIPHER_CTX setup from trying (and failing) to initialise
each of those 10 ENGINEs. Instead, the first such request will try to do that
and will either return (and cache) a NULL ENGINE pointer or will return a
functional reference to the first that successfully initialised. In the latter
case it will also cache an extra functional reference to the ENGINE as a
"default" for that 'nid'. The caching is acknowledged by a 'uptodate' variable
that is unset only if un/registration takes place on that pile. Ie. if
implementations of "des_cbc" are added or removed. This behaviour can be
tweaked; the ENGINE_TABLE_FLAG_NOINIT value can be passed to
ENGINE_set_table_flags(), in which case the only ENGINEs that tb_cipher.c will
try to initialise from the "pile" will be those that are already initialised
(ie. it's simply an increment of the functional reference count, and no real
"initialisation" will take place).

RSA, DSA, DH, and RAND all have their own ENGINE_TABLE code as well, and the
difference is that they all use an implicit 'nid' of 1. Whereas EVP_CIPHERs are
actually qualitatively different depending on 'nid' (the "des_cbc" EVP_CIPHER is
not an interoperable implementation of "aes_256_cbc"), RSA_METHODs are
necessarily interoperable and don't have different flavours, only different
implementations. In other words, the ENGINE_TABLE for RSA will either be empty,
or will have a single ENGING_PILE hashed to by the 'nid' 1 and that pile
represents ENGINEs that implement the single "type" of RSA there is.

Cleanup - the registration and unregistration may pose questions about how
cleanup works with the ENGINE_PILE doing all this caching nonsense (ie. when the
application or EVP_CIPHER code releases its last reference to an ENGINE, the
ENGINE_PILE code may still have references and thus those ENGINEs will stay
hooked in forever). The way this is handled is via "unregistration". With these
new ENGINE changes, an abstract ENGINE can be loaded and initialised, but that
is an algorithm-agnostic process. Even if initialised, it will not have
registered any of its implementations (to do so would link all class "table"
code despite the fact the application may use only ciphers, for example). This
is deliberately a distinct step. Moreover, registration and unregistration has
nothing to do with whether an ENGINE is *functional* or not (ie. you can even
register an ENGINE and its implementations without it being operational, you may
not even have the drivers to make it operate). What actually happens with
respect to cleanup is managed inside eng_lib.c with the "engine_cleanup_***"
functions. These functions are internal-only and each part of ENGINE code that
could require cleanup will, upon performing its first allocation, register a
callback with the "engine_cleanup" code. The other part of this that makes it
tick is that the ENGINE_TABLE instantiations (tb_***.c) use NULL as their
initialised state. So if RSA code asks for an ENGINE and no ENGINE has
registered an implementation, the code will simply return NULL and the tb_rsa.c
state will be unchanged. Thus, no cleanup is required unless registration takes
place. ENGINE_cleanup() will simply iterate across a list of registered cleanup
callbacks calling each in turn, and will then internally delete its own storage
(a STACK). When a cleanup callback is next registered (eg. if the cleanup() is
part of a gracefull restart and the application wants to cleanup all state then
start again), the internal STACK storage will be freshly allocated. This is much
the same as the situation in the ENGINE_TABLE instantiations ... NULL is the
initialised state, so only modification operations (not queries) will cause that
code to have to register a cleanup.

What else? The bignum callbacks and associated ENGINE functions have been
removed for two obvious reasons; (i) there was no way to generalise them to the
mechanism now used by RSA/DSA/..., because there's no such thing as a BIGNUM
method, and (ii) because of (i), there was no meaningful way for library or
application code to automatically hook and use ENGINE supplied bignum functions
anyway. Also, ENGINE_cpy() has been removed (although an internal-only version
exists) - the idea of providing an ENGINE_cpy() function probably wasn't a good
one and now certainly doesn't make sense in any generalised way. Some of the
RSA, DSA, DH, and RAND functions that were fiddled during the original ENGINE
changes have now, as a consequence, been reverted back. This is because the
hooking of ENGINE is now automatic (and passive, it can interally use a NULL
ENGINE pointer to simply ignore ENGINE from then on).

Hell, that should be enough for now ... comments welcome: geoff@openssl.org