5:20 p.m.
On 08/22/2017 11:34 AM, Peter Krempa wrote:
On Mon, Aug 21, 2017 at 11:27:17 -0400, John Ferlan wrote:
> On 08/21/2017 09:06 AM, Peter Krempa wrote:
>> On Fri, Aug 18, 2017 at 17:50:24 -0400, John Ferlan wrote:
[...]
> Oddly it's OK for virhash code to generically call something a "key"
and
> have a default search mechanism that uses a void * STREQ comparison, but
> yet when applied to virobject code used to combine the needs for the
virHash code also stores anything you give it and returns it back. In
your approach the code is not generic enough to take a virObject or a
void *.
> various driver/vir*obj modules to create, add, search, list, remove, and
> delete for a what amounts to be common code between each of them using
> hash table(s) to manage "uuid" or "name" based objects, then
it's not OK.
No, not entirely. As I've said, if you are going to add a "hash table on
steroids", it's fine if you call it anonymously.
What is not okay, that you are going to use the new object type and make
all (domain, storage, ...) objects inherit from it. I strongly dislike
this. Object members need to be clearly named, otherwise it will end up
only in confusion.
There's a difference between dislike and technically incorrect.
In my mind - if {k|K}ey1 was uuid and {k|K}ey2 was name and code was
altered to handle whether one exists or not, then there'd still be
issues with the approach. The effort expended to make those adjustments
and then have some other issue may not be worth it IMO.
Still I'd like to find some common ground using objects, but I have a
feeling that's just not going to be possible.
If LookupKeys were removed and LookupHash changed to rename objsKey1 and
objsKey2 to objsUUID and objsName, would that provide a path forward or
are you so against any of this that it's pointless to even attempt?
Trying to figure a way to constructively move forward.
You certainly can add some internal data structure, which can be an
object, to hold the data which allow this to work. But making domain
objects inherit from it is wrong.
> Ironically my first/initial pass at this took the path of a single
> src/conf/virpoolobj.c module, but it was suggested to use virobject
> instead to abstract things. When that's done - it's back to well I don't
> like that approach because it's too generic, so you should use a common
> module in order to make all those abstractions along with quite a bit
> more specific code in order to handle (known) differences between how
> each of the driver/vir*obj's needs to reference things.
Your first implementation was neither generic nor specific. It contained
generic parts (anonymized name and UUID) and then specific parts, where
you'd have to pass in VM/storage/nw objects. It couldn't really be used
for anything else.
The first one changed all the vir*Obj[List] to be virPoolObj[Table]. It
worked for the existing driver/vir*objs implementations. It was to a
essentially a proof of concept.
And this is mostly the same. You have a hybrid, where you need the
stored objects to inherit from the object you've created, and then
introduce fields which make it specific for this only use case "active".
I can remove active - it's not that important.
Using fields from the creator is mainly a shortcut. Removing them means
adjusting the Add/Remove API's a bit to take parameters.
Yes, it's a hybrid with new objects that have specific usages. I'm not
expecting other areas to use the objects like virObjectLockable is
"reused" in many places. If a new driver/vir*obj is created, then using
these objects I figure would make life easier.
With this you are basically attempting to move an implementation
detail
of the approach to map multiple keys to a virObject of some type to the
object properties themselves, which feels wrong. Especially since the
properties you are trying to extract can't be named uniformly for all
possible objects inheriting from such object.
And the alternative solution is to create a common vircommonobjlist.c
module that will essentially do something similar. There would be quite
a bit of shimmying going on though in order to FindEach and Search
ByUUID or ByName. I haven't put too much thought into that model though
since I focused on the initial review comments that pointed in the
direction of objects. Just because the object is "special" or
"directly"
purposed, doesn't invalidate it's usefulness.
> FWIW: Initial posting... Patch 2 is where the virobject
enhancement was
> suggested.
>
> https://www.redhat.com/archives/libvir-list/2017-February/msg00519.html
>
>>
>> This makes such helpers hard to use and hard to remember which
>> fields serves which purpose in the given use case. This also then
>> triggers custom wrappers for every single usage area where you put names
>> to those fields.
>>
>
> How is it hard to use/remember? I would think it's much harder to use
> and remember 8 different ways to do things as opposed to one, but maybe
> I'm the only one looking to simplify/unify.
If 'key1' is a name of the VM but 'uuid' for a secret, then it's
confusing and hard to remember. Confusing is worse than having different
I don't see it as confusing... What I do find confusing is the multiple
layers of domain XML post parse processing and validation via callbacks.
I find it easy to map name to key1, but trying to figure out and
understand all the rules of processing boggles my mind. Then there's the
hidden private data structures on top of that!
implementation. If you are going to argue that 'key1' should
ever be
interpreted in the context of the lookup functions, then you made it an
implementation detail of the lookup functions thus not requiring
everything to inherit from it.
So you prefer multiple cut-copy-n-paste functions that when one goes to
make a change to say which lock mechanism is used, then one must go
modify multiple files making multiple of the same changes. Essentially
what Michal would be doing to make rwlocks used for all the hash/list
processing instead of modifying one place.
Likewise when an issue pops up and is fixed in one place, then it would
(theoretically at least) need to be fixed in all places. Of course we
all know that's not how it usually happens unless a code reviewer asks,
wouldn't this also be a problem elsewhere or if by chance the person
making the changes actually does the same change across all drivers that
would have the same issue.
> If you really look at the code - you'll see for the most part
it's a
> wrapper around various virhash functions with specific needs for
> driver/vir*obj that includes the refcnt, locking, and smaller subset of
> driver/vir*obj specific code to "manage" the details of/for a vir*DefPtr.
Yes. That's what I object to. If
But this is where I figured the greatest gain is at least w/r/t code
reuse without needlessly adding switches to account for specific data.
>
> It works for nodedev, secret, network, and interface already as seen in
> later patches in the series.
>
> And again, usage of "key1/key2" vs. "uuid/name" is because:
>
> Nodedev uses "name" as it's key
> Secret uses "uuid" as it's key
> Network uses "uuid" and (with these patches) "name" as keys
> Interface uses "name" as it's key
> Nwfilter uses "name" as it's key
> Storage Pool uses "uuid" and "name" as keys
> Storage Volume uses "name" as it's key
> Domains use "uuid" and "name" as keys.
Yes they are different. And that's my point. If you are going to make an
object, which will agregate those properties, you need to be able to
name them. If you can't, don't try
A rose by any other name is still a rose.
> So, is that really that hard to remember? Do you really have a need to
Yes. Also duplicated. You still have the name in vm->def or pool->def.
It bugs me that we store the name in 'def' rather the object itself, but
I don't feel like trying to change it. (I tried and gave up few years
ago.)
Hmmm... so I create an object that stores the name and/or uuid and
that's a problem? Or did I misinterpret what you're trying to say.
> remember? Why do you feel it's so important that the object
knows it's
> managing a table of UUIDs or a table of names? The driver/vir*obj
No that's fine. As long as you don't force us to inherit from it. If you
inherit from it, the names need to be meaningful. If you treat the
object as an impl. Detail then they don't.
> implementation essentially wants at least 1 way to store things and
> possibly 2 so that lookup by a second key is quicker (lookup-by-keyid
> vs. search all for specific name). Yes, they are one or the other
> currently. In the long run, it's just a name.
>
> The driver/vir*obj.c code can choose which to use and manages it's own
> mapping. One need only look at the argument in the create/new to know
> which is which. But, it doesn't matter once you get to the Find/Search
> code as if there are two tables an object has to be in both tables. So,
> when using a Lookup function on some string both tables can be searched
> in order to perform lookup. When it comes to the ForEach or Search type
> functions, only one table needs to be searched in order to run through
> all the elements. So rather than need to make the (generic) object code
> need to check if one table or the other exists, just make at least one
> table required and define that to be the first table. It's just a design
> decision.
>
>>> I think that's the whole purpose of it. After some soul searching I feel
>>
>> Could you elaborate please on the purpose and final goal of this. I
>> was't able to piece together what you are trying to achieve. If you want
>> to unify the code that sections of libvirt use to hold lists of objects
>> I don't think you need a custom sub-type for them.
>>
>
> As if I haven't already elaborated the purpose and final goal. It's been
> stated multiple times and I think it's very obvious.
>
> The goal is to make things generic enough to be used by various
> driver/vir*obj modules in order to abstract away or combine alike code.
> If that's truly not desired that's fine. I think it's far better to have
> less cut/copy-n-paste code, but not everyone has that same viewpoint.
>
> Rather than take a myopic viewpoint of a domain or storage or network -
> once you consider that each of those subsystems essentially does the
> same thing - creates an object, stores that object, allows searches on
> that object, and eventually deletes that object. The details of the
> object is still handle-able by specific code, but there's much with that
> object that can and should be generically handled.
>
> The "reason" I went down this rabbit hole was some patches posted by
> Virtuozzo which essentially copied *a lot* from the storage driver in
> order to make another storage driver specific for their needs. I felt
> rather than cut-copy-n-paste that we should be able to "combine" a lot
> of code. As I dug into the code I found a mish-mash of object management
> throughout the drivers. Mostly forward linked lists, but each set of
> code just different enough to make things "difficult" as it relates to
> being able to support/maintain code over time. So rather than have
> multiple ways to do things, take a common approach so that when you're
> looking through code you don't have to know the differences between 8
> piles of code that all accomplished the same goal, but each in its own
> unique way.
Fair enough. So you want a better way to collect objects in lists.
That's a very worthwile goal.
[... trimmed the rest, since most points would repeat ... ]
>>
>> The wrapper object you are adding here doesn't seem to add any value,
>
> That's your opinion.
Well, you techincally asked for it by sending it for review. It would be
more worhtwhile to refute it rather than dismiss it.
I'm merely pointing out my opinion of the value of the code to oppose
your opinion. There is value since it's a worthwhile goal. This
mechanism is one way to achieve that.
John
I might have not responded to everything, but this message was rather
long. Feel free to reiterate points which you feel I've missed.