Re: Moving files from lisp/gnus/ to lisp/net/?

[Simon Josefsson]

There are risks associated with caching passwords (i.e., the password is
stored in memory and may be swapped out to disk), and possibly it should
not be the default.  However, I think the risks are somewhat overblown.
The garbage-collecting nature of elisp exposes this risk anyway.  All
passwords entered into emacs may be swapped out to disk.

There are solutions intended to address the real problem here: GnuPG
uses a helper program pin-entry that asks the user for passwords.  I
believe it can cache passwords too.  However, this is not widely used,
and definitely not used or known except to GnuPG users.  I'm not
convinced it is mature enough to be suggested as a solution for password
input in emacs.  And anyway, you'd still have the gc-problem in emacs if
the password is transfered into an emacs elisp variable.

Another point of view is that several parts of Emacs already contains
code to cache passwords.  Reducing that code duplication is what
triggered me to write password.el.

The existing `read-passwd' API is not suitable for password.el, because
each password needs to be associated with an application-dependent
'key'.  There is no parameter for that in `read-passwd'.  Do you think
it is worth adding one?

Alternatively, and what I consider the best idea (but it was some time
since this was discussed and I may very well have forgotten some
important point): let's make `read-passwd' a more lower-level primitive,
used by `password-read'.  New elisp applications, that want to allow
users to cache passwords, can start to use the `password-read' API.
Existing elisp applications with their own ad-hoc code to cache
passwords could be modified to use password-read.  Users can modify the
caching behaviour by setting password-* variables.

Regarding naming, possibly `password-read-and-cache' should be removed.
The problem with it is that it enters passwords into the cache without
allowing applications to verify that the password is correct.  However,
applications could detect this and use `password-cache-remove', but it
seems less reliable to me.

/Simon

Richard Stallman <address@hidden> writes:

> Here's what I said about read-passwd and password.el in 2004.  Setting
> aside what I said about the short term, the main issue seems to be
> whether to cache passwords.
>
> Simon, what do you think about this now?
> What do others think?
>
>
> First message:
>
> It occurs to me that paranoid people might be worried about saving
> passwords in a cache like this.  What do people know about that issue?
>
> The text added to the manual is clear enough.  I have some issues
> about the substance:
>
>      -- Function: read-passwd prompt &optional confirm default
>        This function reads a password, prompting with PROMPT.  It does
>        not echo the password as the user types it; instead, it echoes `.'
>        for each character in the password.
>
> This ought to use the cache now, but the text does not say so.
>
> (Later): It looks like all you did was move read-passwd into
> password.el without changing it.  What I suggested was to integrate
> read-passwd into this file--which means, make it use the cache.
>
> There is no sense in moving read-passwd into the new file without
> making it use the new file's facilities.  That change only causes
> password.el to be loaded for programs that don't use the cache.
> It provides no benefit.
>
> So if its definition is to remain unchanged, it should stay in subr.el
> where it is preloaded.  But I would rather see it move to password.el
> and *take advantage of the facilities of password.el*.
>
> Can this be done?
>
>      -- Function: password-read prompt key
>        Read a password from the user, using `read-passwd', prompting with
>        PROMPT.  If a password has been stored in the password cache,
>        using `password-cache-add' on the same KEY, it is returned
>        directly, without querying the user.
>
> Once read-passwd uses the cache, won't password-read be obsolete?
>
> Why have both password-read and password-read-and-add?
> Why not always add?  Is the idea that for some purposes
> it is ok to cache, but for others it is too risky?
>
>
> Second message:
>
>     Other applications typically ask the user whether they want to
>     remember the password in memory.  If read-passwd is changed to cache
>     passwords (however, to use the cache, callers of read-passwd must be
>     updated, to provide the "key" into the hash table), it could ask the
>     user this.  Opinions on this welcome.
>
> Having it always ask would be too annoying, I think.  So that would
> need to be a new argument, which means the feature is no benefit
> unless we change the callers.
>
> That is the wrong thing to do at present.  So I think we should put
> read-passwd back where it was and remove the new file for now.
>
> Looking ahead to the future,
>
>     > Why have both password-read and password-read-and-add?
>     > Why not always add?  Is the idea that for some purposes
>     > it is ok to cache, but for others it is too risky?
>
>     No, the reason was this: if the user entered an incorrect password, it
>     should not be cached.  If an incorrect password is cached, the code
>     might infloop trying the incorrect password automatically over and
>     over again.  It was considered safer to first read the password, then
>     try to use it, and if successful then it is cached.
>
> In that case, password-read-and-add makes no sense, right?
> Why add a shortcut that in best practice should not be used?
>
>     I'm not sure my argument is good, it may be simpler to always cache,
>     and have the calling code invoke password-cache-remove whenever there
>     is a password failure.
>
> That is a reasonable alternative, I guess, but then password-read
> should add the password and we should not have password-read-and-add
> as a separate entry point.
>
> But it doesn't make sense to discuss this without dealing with the
> question of whether caching of correct passwords is desirable.