tech-kern archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: fs-independent quotas



On Thu, Nov 17, 2011 at 01:02:33PM +0000, David Holland wrote:
>  > > Neither is good enough if you're providing backwards compatibility;
>  > 
>  > An error is still better than a crash.
> 
> ...and neither is acceptable for the quotactl system call, either in
> the kernel processing it or in userlevel processing the returned data.
> The code has to be right; consequently, what happens if it's wrong is
> less important or even entirely unimportant.

and I still believe it's easier to get it right with something like
prolib (where you have to pass the data you got through problib
functions), than using ad-hoc checks, that may or may not be used
by all involved parties.

> 
>  > > it
>  > > has to *work*. This is the standard we're committed to, and I continue
>  > > to think there's no particular advantage for proplib in this regard,
>  > > particularly for this particular kind of data.
>  > 
>  > And I still think a text-based format is better than a binary format for
>  > backward compatibility.
> 
> And I still don't agree, or even really see where the substance of the
> argument is.

So we agree to disagree :)

> 
>  > > (I don't think any semistructured or "self-describing" data model,
>  > > including the perfect one I'd replace proplib with if I could wave a
>  > > wand to do so, provides any particular advantage for procedure call
>  > > compatibility. Sure, you can tag data bundles with version codes and
>  > > such, but we can and do already do that by tagging the call itself and
>  > > have lots of support architecture in place for doing it that way. The
>  > > advantages appear when you're dealing with irregularly structured
>  > > material, like when there are large numbers of optional fields or
>  > > optional parameters and so forth.)
>  > 
>  > Or when you send different commands which takes different arguments,
>  > which is the case for the quota syscall and VNOPS.
> 
> That's true but each command code is associated with a specific type,
> in the same way that functions are associated with a specific type.
> 
> Well, in a similar way. Is what you're objecting to here that the old
> quotactl interface used void pointers and wasn't (even halfway)
> typesafe? That seems like a stronger point but it has nothing
> whatsoever to do with compatibility...

the old quotactl interfaces has more issues that that, but yes, this
is one thing I didn't like.
If you have different quotactl commands, you'll likely have different
arguments so you'll end up using a void pointer (or something equivalent,
like a pointer to a union).


> 
>  > > I don't do Perl. I might be persuaded to do Python bindings, but it
>  > > would probably be more effective to enlist someone who already knows
>  > > how to do this and won't therefore make newbie mistakes with the
>  > > interpreter. Anyway, the hard part is making the library interface
>  > > available; wrapping Perl or Python around it should be entirely
>  > > trivial.
>  > 
>  > So you propose to remplace something that we can work with using existing
>  > modules it various language, by something that needs a specific module (to
>  > be written) for each langage. Clearly that's not a win.
> 
> That line of reasoning can be used to justify almost any bad
> architectural decision (ranging from HTML email to using msdosfs as an
> application save format...) so I'm not going to accept it.
> 
> Writing language bindings for a simple and straightforward library is
> a simple and straightforward undertaking.

OK, so prove it by writing a perl binding format :)
I've never written a language binding, so it's not going to be
straightforward for me anyway.

> [...]
>  > > I don't think that feature is worth preserving either, since it's
>  > > purely a side-effect of having visible quota files. And I don't see
>  > > the point; it's not like mounting the volume to run edquota will cause
>  > > a catastrophe.
>  > 
>  > I dropped this functionnality for ufs-quota2; I used it for ufs-quota1
>  > because otherwise things would go wrong (something to do with offsets:
>  > a uid would get the quotas for another uid - of course this is a bug, but
>  > I never managed to track it down).
> 
> That is definitely a bug. And it's likely to still exist. Do you have
> a way to reproduce it, or to try to reproduce it? Did it affect all
> tools or only some/one? (And where's the PR number? :-p )

No, I dond't know how to reproduce it (or I would have debugged it),
and it's not clear to me if the problem is in the kernel or userland.
I noticed it on 2 different servers after a reboot, but I'm not sure
how I ran into this situation.

BTW, quotacheck always do quota update offline, because it's run before
quotaon.


On Thu, Nov 17, 2011 at 01:13:26PM +0000, David Holland wrote:
> On Tue, Nov 15, 2011 at 07:56:09PM +0100, Manuel Bouyer wrote:
> 
>  > > Do we expect NetBSD developers to be checking in handlers for
>  > > functionally equivalent binary formats that are *more* likely to
>  > > cause a crash if invalid data are presented?  That does not seem
>  > > good.
>  > 
>  > If you pass e.g. a string to the legacy quotactl syscall instead of
>  > the required data structure, the kernel won't notice and will use
>  > the bogus data.
> 
> That's no different from any other system call, though.

s/any/some/ because for some of them, doing a sanity check is easy
(it's part of copyin/copyout for example).

> 
>  > > Whether text or binary, the data have to be validated before being
>  > > used.  Text or binary, code that doesn't do that is simply buggy.
>  > 
>  > But it's easier to validate a text format than a binary one.
> 
> But we don't have any schema handling or validation code for proplib
> bundles. One can write checks by hand, but then they're likely to be
> incomplete and/or wrong, and it ends up not being a net win. Also, the
> more validation you do the less you can use the approach below.

that's why having a schema validation is not a strong requirement for me.

> 
>  > > I think that if we aren't _actually_ going to provide backward
>  > > compatibility in some way _now_, it doesn't matter.  Either way, to
>  > > actually get sane semantics, versioning is required.
>  > 
>  > The actual quotactl interface has a version number embeeded, for
>  > this reason.  But, for example, some fields can be added to the
>  > strucure without changing the version number. The consumer will
>  > just notice if the new field is present or not and real with it;
>  > this provides backward and forward compat (older consumers will
>  > just ignore the new fields).  This is somethig you can't do with a
>  > binary format.
> 
> Well, yes it is (see uvmexp_sysctl for example) but besides that, as
> I've pointed out before, this only works if you can predict in advance
> what kinds of changes you expect. Otherwise, when you get a proplib
> bundle that contains unexpected data, it fails the validation checks
> and you die with an error.

You don't need to predict in advance, you just make sure to ignore
what you don't need in the data. If a key/value pair you need is not there,
it's an error. If you get an extra key/value pair that you don't need/don't
know about (this is the same thing), you just ignore it.

> 
> Also, without schema handling for proplib do you really expect this
> kind of ad hoc approach to be workable over a span of years, involving
> multiple releases and possibly 3rd-party code?

If the code follow the rules of ignoring unknown/uneeded key/values, yes.
This is how e.g. html has done it, and it's not that bad.

> 
> (While I see the code you put in netatalk uses only struct
> ufs_quota_entry and not the proplib bundles directly, you've been
> arguing in this thread that struct ufs_quota_entry (after being
> renamed) isn't good enough in the long term. Even though I don't
> believe this, I do expect your arguments to be self-consistent...)

the netatalk code is strongly tied to ufs-like quota semantics; 
if some new fields shows up it wouldn't know how to use it without a
deep rework anyway. But if the changes don't remove key/value
needed by ufs_quota_entry, then it'll work out of the box, even if the
dictionary it got from the kenrel is larger.

-- 
Manuel Bouyer <bouyer%antioche.eu.org@localhost>
     NetBSD: 26 ans d'experience feront toujours la difference
--


Home | Main Index | Thread Index | Old Index