On Wed, 21 Oct 2015 17:30:50 +0200 Arvin Schnell wrote:

On Wed, Oct 21, 2015 at 04:36:19PM +0200, Josef Reidinger wrote:

...

Now lets move to examples. From my user POV there is some confusing API calls and parameters. For example lets use find1.cc as example which I comment (I can do this for all examples).

// creating some global storage is fine Devicegraph* devicegraph = new Devicegraph();

// this looks strange for me // 1) how disk can be created? I expect disk is detected or proposed

The probing code must also be able to create the object representing a disk. With the Image/Kiwi mode you can create a disk which is backed by a loop device.

What about using name "Disk::propose" for creating such disk? I think it is just confusion caused by naming, as said I create in library disk a bit confusing.

...

// 2) why saying that disk have "/dev/sda" what if I am on qemu which // have "/dev/vda"? Disk* sda = Disk::create(devicegraph, "/dev/sda");

The probing code uses the parameter. Currently you are basically looking at functions to manipulate the device graph.

ok, if using propose, then it make sense to me as it propose disk with given name, which is fine.

...

// this looks fine for me, creating partition table on disk PartitionTable* gpt = sda->create_partition_table(PtType::GPT);

// here I do not get why I need to pass "/dev/sda1" ? Let me say it // this way, if I have generic code that generate partition for passed // partition table how I can know if it is hda, sda or vda? and why // I need to know number of partition? cannot it by default create next // available one and pass number as optional parameter? gpt->create_partition("/dev/sda1", PRIMARY); Partition* sda2 = gpt->create_partition("/dev/sda2", PRIMARY);

Use PartitionTable.get_unused_partition_slots() to get the name, type and possible region. The interface will change to allow also passing the region.

Great, so it already have such functionality. I just expect that create partition can have reasonable defaults and use this method inside for default parameters. So you have gpt->create_partition(); and it inside get available ones with above mentioned get_unused_partition_slots. Of course it should support passing manual names, region and so on. This way API will be simple for common usage and powerfull if needed.

...

// here I create top level container, looks fine for me LvmVg* system = LvmVg::create(devicegraph, "/dev/system");

// this is very confusing, why I need devicegraph here? Why exposing it // to user? From doc I know that both sda2 and system have reference to // it so why it need to be passed? // And also why it need User::create? Why it is not simple // `system.add(sda2)` call which is more intuitive for me? User::create(devicegraph, sda2, system);

User::create is a low-level function to manipulate the device graph. Functions like system.add_physcial_volume will be added.

Great.

...

// in general it looks fine for me, just is there reason to pass // whole device name? who not having thing like // `system->create_lvm_lv("root")` LvmLv* system_root = system->create_lvm_lv("/dev/system/root");

This will change. For LVM there are only mostly empty classes currently.

...

// creation of filesystem, intuitive and easy Filesystem* filesystem = system_root->create_filesystem(EXT4); // quite confusing, what is adding mountpoint to filesystem? // filesystem do not know about mount points, I expect something like // `devicegraph.mountpoint.add("/", filesystem)` filesystem->add_mountpoint("/");

Only filesystems have mount-points so to me it looks natural to handle them there. Alternatively we could add objects for mount-points in the graph. BTW: E.g. ext4 knows its last mount point (try dumpe2fs).

...

// some debug output, nothing to comment, not sure if it is needed to // be public methods cout << "num_devices: " << devicegraph->num_devices()<< endl; cout << "num_holders: " << devicegraph->num_holders() << endl; cout << endl;

// validation of storage, easy and intuitive... does it raise exception // if failed? devicegraph->check();

My idea are exceptions for fatal errors where the graph is so broken that it cannot be used anymore, e.g. duplicate storage id. Such an error indicates a libstorage internal problem. For other things that the user can repair I think a list of issues should be returned, e.g. the user created a raid with level 6 but only added two drives. Here the user just has to add more drives, change the level or remove the raid.

...

// printing of object, fine and intuitive C++ code cout << devicegraph << endl; // printing graphiz image, easy and intuitive API call devicegraph->write_graphviz("test1.gv");

// looking for all filesystems that are mounted as root mountpoint // is confusing. why it do not return single filesystem? for (const Filesystem* filesystem : Filesystem::find_by_mountpoint(devicegraph, "/")) {

You can mount a device at several mount-points. With btrfs this is even standard, see https://github.com/openSUSE/libstorage/blob/master/doc/status-current-code.m....

Yes, I know that device can have multiple mountpoints, but here is logic reverse. It said that one mount point can have multiple devices ( or filesystem like used here ) which is at least for common usage without overlapper layered mount points not possible. So on "/" is one btrfs...of course you can get same btrfs also for /boot/grub2 etc.

...

// I am not sure if I can imagine what is ancestor of filesystem? // I hope there is better name for it // second note is boolean parameter, it is really hard to read it and // hard to remember what such parameter mean. see e.g //http://programmers.stackexchange.com/questions/147977/is-it-wrong-to-use-a-b... for (const Device* device : filesystem->get_ancestors(false)) {

get_ancestors is the base function in Device and thus is not specially named for filesystems. Otherwise the name is standard, e.g. https://en.wikipedia.org/wiki/Tree_%28data_structure%29.

Yes, name make sense if you know that it is tree structure, but it is inside POV, as user I try to read what api said to me and it said "get me ancestors for given filesystem" which is a bit confusing and require to know that filesystem is device and devices is stored in tree structure and then you understand what code want you to say. So more specialized call like filesystem->get_device() or get_partition ( even if I know that disk can be partitionless) can be more friendly for reader of code.

The bool parameter was already mentioned several times here and will be replaced by an enum.

good

...

// this a bit break polymophysm, but it is explained in design // document, so fine for me if (dynamic_cast(device)) cout << "mount point \"/\" somehow uses a logical volume" << endl; } }

delete devicegraph;

...

So in general, I think that we should now more focus on API and its usability as it is hard to change it in future. When having good API, some cleaning or refactoring of implementation later is easier ( current code in new libstorage is short and easy, but I worry in future with more features, we need to be prepared to clean it also).

We are focused on the API at the moment.

Good, that brings me two ideas. 1) I can do such API review for all examples in example directory if it make sense for you 2) maybe it make sense to write down expected use cases for library usage and write example code for such use cases which can be discussed. Josef

Regards, Arvin

-- To unsubscribe, e-mail: yast-devel+unsubscribe@opensuse.org To contact the owner, e-mail: yast-devel+owner@opensuse.org

On Thu, Oct 22, 2015 at 08:59:52AM +0200, Josef Reidinger wrote:

...

You can mount a device at several mount-points. With btrfs this is even standard, see https://github.com/openSUSE/libstorage/blob/master/doc/status-current-code.m....

Yes, I know that device can have multiple mountpoints, but here is logic reverse. It said that one mount point can have multiple devices ( or filesystem like used here ) which is at least for common usage without overlapper layered mount points not possible.

It is possible, although likely stupid: cat /proc/mounts [...] /dev/mapper/system-abuild /mnt ext4 rw,relatime,commit=600 0 0 /dev/mapper/system-giant--xfs /mnt xfs rw,relatime,attr2,inode64,noquota 0 0 So the API must be able to handle the situation. Or do you prefer an exception during probing?

...

get_ancestors is the base function in Device and thus is not specially named for filesystems. Otherwise the name is standard, e.g. https://en.wikipedia.org/wiki/Tree_%28data_structure%29.

Yes, name make sense if you know that it is tree structure, but it is inside POV, as user I try to read what api said to me and it said "get me ancestors for given filesystem" which is a bit confusing and require to know that filesystem is device and devices is stored in tree structure and then you understand what code want you to say. So more specialized call like filesystem->get_device() or get_partition ( even if I know that disk can be partitionless) can be more friendly for reader of code.

The example in find1.cc does something special: It checks whether the filesystem *somehow* is located on a LVM logical volume. This is still the case even if first on the logical volume is an encryption device and then the filesystem. If you just want to know which block devices the filesystem is directly located on use filesystem->get_blkdevices().

1) I can do such API review for all examples in example directory if it make sense for you

Not yet or in other words the examples are not use-cases. Regards, Arvin -- Arvin Schnell, Senior Software Engineer, Research & Development SUSE Linux GmbH, GF: Felix Imendörffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nürnberg) Maxfeldstraße 5 90409 Nürnberg Germany -- To unsubscribe, e-mail: yast-devel+unsubscribe@opensuse.org To contact the owner, e-mail: yast-devel+owner@opensuse.org

Dne 23.10.2015 v 10:24 Josef Reidinger napsal(a): [...]

2) common query should return only later as first one is overlaped, of course there can be specialized call to manipulate mount points knowing also about overlap or knowing that removing first mount point return to game previous one

It seems that Libzypp has something similar for repository URLs. Each repository can have assigned multiple URLs and you can get all URLs via baseUrls() methods [1] or a single URL via url() method [2] (it's the first one from the list). The most common use case is just a simple url(), the list of URLs is usually not needed. [1] https://github.com/openSUSE/libzypp/blob/master/zypp/RepoInfo.h#L140 [2] https://github.com/openSUSE/libzypp/blob/master/zypp/RepoInfo.h#L127 -- Ladislav Slezák Appliance department / YaST Developer Lihovarská 1060/12 190 00 Prague 9 / Czech Republic tel: +420 284 028 960 lslezak@suse.com SUSE -- To unsubscribe, e-mail: yast-devel+unsubscribe@opensuse.org To contact the owner, e-mail: yast-devel+owner@opensuse.org

Am 23.10.2015 um 10:24 schrieb Josef Reidinger:

I think it is two different things.

1) probing should handle such situation

2) common query should return only later as first one is overlaped, of course there can be specialized call to manipulate mount points knowing also about overlap or knowing that removing first mount point return to game previous one

Not sure if the two of you are talking about the same things here. Maybe this helps to clarify: In general, it is not a good idea to enforce a consistency check after each low-level call. There should be such a consistency check, of course, but when is it triggered, and what consequences does this have? Use case: The user is in the expert partitioner and wants to change mount points because he discovered he made a mistake: He really wants /home on his blazingly fast SSD /dev/sda and /data on the slower rotating disk /dev/sdb, but he accidentially put /data on /dev/sda3 and /home on /dev/sdb1. In effect, this is a classic "swap" operation. So, from a user's point of view, I want to be able to: - Change the mount point of /dev/sda3 from /data to /home. So, for the moment, I have a /home mount point on both /dev/sda3 and /dev/sdb1. - Change the mount point of /dev/sdb1 from /home to /data. Voila, only one /home mount point left, everything okay. If we force the consistency check after the first step, it will complain about the two identical mount points, so we force the user to either change the mount point of one of the two partitions to nothing temporarily (maybe causing another error popup if this is checked, too) or to invent a temporary fake mountpoint. One way or the other, this is another operation the user has to do, and it's not very intuitive. This serves to illustrate the point: We need to keep low-level operations and consistency checks apart. Consistency checks should be invoked at strategic points - either just before commiting / executing all the changes, or explicitly triggered from app code when the logical transaction is done - i.e. when the user tries to exit the expert partitioner. This would be a good place to do the checks and to display any warning or error dialogs to the user. Any time before that and we are limiting the user's workflow, forcing him into detours and thus distracting him from what he really wants to do. This affects the real end user as well as any application code using libstorage. HTH -- Stefan Hundhammer YaST Developer SUSE Linux GmbH GF: Felix Imendörffer, Jane Smithard, Graham Norton; HRB 21284 (AG Nürnberg) Maxfeldstr. 5, 90409 Nürnberg, Germany -- To unsubscribe, e-mail: yast-devel+unsubscribe@opensuse.org To contact the owner, e-mail: yast-devel+owner@opensuse.org

Dne 21.10.2015 v 17:52 Arvin Schnell napsal(a):

...

As libstorage basically use CLI of other programs, so question is why not use more high level language.

Why should C++11 not be a "very" (not in the term of https://en.wikipedia.org/wiki/Very_high-level_programming_language) high level language? I fear an endless discussion coming.

Um, if I would start a complete new library I'd very likely use Ruby (unless another language would fit way better, I'd be willing to sacrifice some features to avoid learning yet another language...) But I do not want to start a flamewar, if libstorage developers are skilled in C++ and want to use it I'm fine with that. (It could be even worse, I know some worse languages than C++ ;-)) My concerns are mainly: - the code should be readable even for people who are not C++ experts or used C++ long time ago and forgot some C++ tricks - the code should be easily testable -- Ladislav Slezák Appliance department / YaST Developer Lihovarská 1060/12 190 00 Prague 9 / Czech Republic tel: +420 284 028 960 lslezak@suse.com SUSE -- To unsubscribe, e-mail: yast-devel+unsubscribe@opensuse.org To contact the owner, e-mail: yast-devel+owner@opensuse.org

On Wed, 21 Oct 2015 17:52:40 +0200 Arvin Schnell wrote:

On Wed, Oct 21, 2015 at 04:36:19PM +0200, Josef Reidinger wrote:

...

Now first note. I do not see in decisions why C++ is used as language for this library. I think it would be nice to document it why it is needed as currently from document I see that planned users are machinery ( which use ruby ), Yast (ruby) and kiwi (perl). Is there any performance reasons, availability of bindings, libraries or any other reason to use it?

Sure, I'll write something to the documentation.

...

As libstorage basically use CLI of other programs, so question is why not use more high level language.

Why should C++11 not be a "very" (not in the term of https://en.wikipedia.org/wiki/Very_high-level_programming_language) high level language? I fear an endless discussion coming.

I do not want to start flame and I am fine with such explanation, I just said it should be documented ( and it will as you write above ). During my presentations I often get such question for yast. Why it is ruby and not python, js, go, whatever ( funny that noone mention C++ or perl, probably not enough sexy :). Josef

Regards, Arvin

-- To unsubscribe, e-mail: yast-devel+unsubscribe@opensuse.org To contact the owner, e-mail: yast-devel+owner@opensuse.org