Discussion arising from [A template virtual filesystem]

[LV] Would it be beneficial a) to include this template with tclvfs and
b) to upgrade the existing tclvfs examples with the features your template
provide?

[SEH] I would love to have my work incorporated into the tcllib, but I don't think it's my decision.

[CMcC] With all respect to SEH, whose idea the template vfs is, I would  like to suggest that the template vfs either be rewritten in Snit, or a parallel version in [Snit] written and maintained as part of the same package.  I have rewritten Steve's template vfs in Snit, for the reasons I give below (although I'm not suggesting that my version is necessarily the best possible version.)

The reasons to use Snit for template vfs:
   1. Very often in [tclvfs] code one is forced to carry state around with each mount-point.  There's no easy/neat way to do this in pure templates (as code inspection will reveal :)
   2. Consistent -option handling, with defaults.
   3. Clearer interface definition and composition - Snit is designed to plug components together.
   4. OO paradigms are interconvertable more easily than rewriting something into OO.  Now's a good time to make the choice to go OO - before it's too late.
   5. Easier to extend a Snit type to extend functionality than to do the same with pure-namespace implementations.
   6. (add here)

Rubbish reasons not to use Snit for vfs:
   1. ''I don't know/use Snit'' ... <contra>either it's worth using or not, lack of training isn't an excuse.</contra>.
   2. ''I don't like Snit'' ... <contra>suggest another system, suggest some improvements - argue about it.</contra>
   3. ''I don't like OO'' / ''I've got my own OO style'' ... <contra>two words: code reuse</contra>

Real reasons not to use Snit for tclvfs:
   1. Snit's slower than pure template code aka ''Snit is no speed demon''... This is the hardest one to answer for tclvfs file systems which are speed critical - but what are you doing using tclvfs for a speed-critical facility? Snit's getting a lot of optimisation attention, and this will only improve the performance over time.

Finally, the best reason for using Snit for tclvfs:  Much of the need for this Template vfs is that the tclvfs raw interface has some anomalies and some lacks.

The template vfs is really a shim to provide:

   1. a consistent interface to tclvfs where there are some incongruities (smoothing out the rough spots.)
   2. genuine extensions of tclvfs functionality which could be better implemented in tcl than in C.

Some of the provisions in point 1 can eventually be moved into the tclvfs C core.  Doing this over an OO interface is (I believe) easier and less disruptive than doing it over a pure namespace implementation.

Therefore: I would like to argue, passionately, for at least *some* OO interface, and preferably Snit (because it seems to be the ostensive standard for pure tcl OO,) whether or not this is maintained in parallel with or alternative to a pure namespace template tclvfs shim.  I would also like to see the existing tclvfs' migrated to whatever system is adopted, and new tclvfs' adhering to whatever that system is.

Addendum:  Of course, if it's a choice between a raw namespace template vfs and a no template vfs at all, I would ''emphatically'' prefer the namespace template vfs.

[SEH] - 10/18/04 - My motivation for writing the template vfs was both to teach myself the Tclvfs API, and to verify that the Tclvfs package was in good enough shape to write software tools around, since I have been bitterly disappointed in the past by the gulf between claims of maturity in software and emergent reality.  I'm very enthusiastic about the vfs concept, and I think it's a solution looking for problems.  But the bottom line is the Tclvfs package is still officially in a beta state, and writing the template vfs did indeed reveal significant issues, some of which it appears will require changes in the core to address.  Given that, I find it hard to justify the extra time and effort necessary to increase the sophistication of the vfs's as you suggest.  No amount of polish in the vfs code will compensate for the inherent flaws in the underlying package.  If the maintainers of the package are willing to commit to bringing it out of beta and fixing the identified issues, I will revisit the possibility
of taking the vfs's to the next level.  Until then, I'm going to put my available time into solving the problems I face in the present that the Tclvfs package and the vfs's are capable of helping with in their current state.

On the other hand, I was also hoping to spark interest in the Tclvfs package, which seemed to be languishing, so I am heartened by your involvement and advocacy.  I hope the friction and noise will help to motivate the maintainers.

P.S.  By way of further motivation, I think there are enough bits floating around now to envision, if consolidated, a sophisticated content management/version control/code distribution and installation service that would knock Bitkeeper on its heels and blow the doors off anything offered by any competing scripting technology.
----
[LV] Have the tclvfs issues mentioned by [SEH] been communicated to the tclvfs project via the
bug report and feature request queues?

[SEH] -- They're all in there.
----
Let's take 'em one at a time:

'''[[1004273]] Can't read from channel in close callback''' [http://sourceforge.net/tracker/index.php?func=detail&aid=1004273&group_id=32692&atid=406308]

The problem is that the user has requested a read/only or write/only channel, but the tclvfs' close process absolutely requires fuller access to the channel.  For example: a user's write-only chan has to be read by close in order to be processed.  Another example: a user's read-only channel has to be initially seeded by a memchan.

This can be modelled by the creator of a channel creating it with full permissions, and casting them away before handing them to a user, then re-aquiring full channel permissions at the point where the channel needs to be closed - that is, immediately before the tclvfs' close callback is invoked.

TIP 220 was originally written to allow C code to modify permissions, but this makes the permissions system mean nothing, as a channel's permissions could then be freely modified in an ad hoc manner.

TIP 220 has been changed to specify a weaker modification that is still powerful enough to implemenent the desired channel semantics:  immediately prior to a vfs' close callback, a channel's permissions are set to the full set available.

   * escalation of privilege is controlled with the tcl core, unavailable to extensions.
   * privilege is only granted (via its close callback) to the entity which opened the channel.

Since the opening entity can in any case have complete access to any file it can create, and cannot successfully modify any file it couldn't have obtained from the o/s, and because the privileges are only available to the opener, and then at a point immediately before the tcl core is about to complete channel closing, re-escalation of privileges doesn't expose open channels to any additional risk of interference.