Re: [Scst-devel] Discussion about SCST sysfs layout and implementation.

[Daniel Debonzi <debonzi@linux.vnet.ibm.com>]

Vladislav Bolkhovitin wrote:
> Daniel Debonzi, on 04/16/2009 10:23 PM wrote:
>> Vladislav Bolkhovitin wrote:
>>> Hi All,
>>>
>>> Below is proposal for the SCST sysfs layout, which will replace 
>>> existing procfs-based infrastructure. Any comments, questions and 
>>> suggestions are welcome!
>>>
>>> I. SCST sysfs layout.
>>>
>>> Root would be /sys/scsi_tgt.
>>>
>>> In the root there would be the following files and subdirectories:
>>>
>>>  - targets - subdirectory listing names of all registered target 
>>> drivers.
>>>
>>>  - devices - subdirectory listing all registered backend devices.
>>>
>>>  - sgv - subdirectory listing all existing SGV pools.
>>>
>>>  - drivers - subdirectory listing all loaded target and backend 
>>> drivers (dev handlers).
>>>
>>>  - threads - RW file listing number of global SCST threads. Writing 
>>> to that file would allow to change that value.
>>>
>>>  - trace_level - RW file listing SCST core logging level. Writing to 
>>> that file would allow to change that. Example content: "out_of_mem | 
>>> minor | pid | line | function | special | mgmt | mgmt_minor | 
>>> mgmt_dbg | retry". See current procfs implementation of this file for 
>>> more info.
>>>
>>>  - version - RO file listing version of SCST core and enabled compile 
>>> time features. Example content: "1.0.2, EXTRACHECKS, 
>>> DEBUG"                                                                                                                          
>>
>>
>> Based on all I read this last days, I believe we are not allowed to 
>> include the directory scsi_tgt on /sys root. I think it has to be in a 
>> existent directory reserved for this sort of application. I just 
>> didn't figured out which one it would be.
> 
> /sys/class? It already has scsi_device, scsi_disk, scsi_generic and 
> scsi_host.

I don't think so because all the directories on /sys/class have symlinks 
to the files somewhere else. However I noticed that many of them on my 
system are on /sys/device/virtual

>>> III. Implementation considerations
>>>
>>> 1. Top level subdirectories and files should be created by init_scst().
>>>
>>> 2. targets/target_driver_name drivers/target_driver_name and should 
>>> be created by scst_register_target_template() and removed by 
>>> scst_unregister_target_template().
>>>
>>> 3. targets/target_driver_name/target with sessions, luns and 
>>> ini_groups subdirectories should be created by scst_register() and 
>>> removed by scst_unregister().
>>>
>>> 4. targets/target_driver_name/target/sessions/session and below 
>>> should be created by scst_init_session() and removed by 
>>> scst_free_session().
>>>
>>> 5. Pass-through devices should be added to devices/ by 
>>> scst_register_device() and removed by scst_unregister_device(). 
>>> Initially they should have "handler" link pointed to "none".
>>>
>>> 6. Virtual devices should be added to devices/ by 
>>> scst_register_virtual_device() and removed by 
>>> scst_unregister_virtual_device().
>>>
>>> 7. drivers/dev_handler_name should be added by 
>>> scst_register_dev_driver() or scst_register_virtual_dev_driver() and 
>>> removed by scst_unregister_dev_driver() or 
>>> scst_unregister_virtual_dev_driver().
>>>
>>> 8. It isn't clear how to best combine standard and custom entries in 
>>> targets/target_driver_name/target, devices/device, 
>>> drivers/target_driver_name and drivers/dev_handler_name, I don't know 
>>> sysfs interfaces sufficiently well. I can only suggest places, where 
>>> it should be done:
>>>
>>>  - For targets/target_driver_name/target a target driver after 
>>> scst_register() register call should call new function 
>>> scst_get_target_root() and add there new entries. Before 
>>> scst_unregister() call the target driver should remove them at first. 
>>> Similarly it should be done for drivers/target_driver_name and 
>>> drivers/dev_handler_name.
>>>
>>>  - For devices/device a dev handler should do it in attach() callback 
>>> and remove in detach() callback. Similarly to scst_get_target_root(), 
>>> the dev handler should get its sysfs root by calling 
>>> scst_get_dev_root().
>>>
>>> Both should be made in some generic way to minimize duplicated code 
>>> between target drivers and between dev handlers.
>>
>> Also based on what I read, the way to have data structures reflected 
>> on sysfs is using kobjecs. I feel that the expected approach to have 
>> it is to include a kobject (or kset depending on the case) on the 
>> existent structures which will be reflected on the sysfs. I notice 
>> that on the actual implementation all the proc interface is 
>> implemented on scst_proc.c and I don't know if it will be possible to 
>> keep having the access interfaces on a separated source file. We could 
>> possible have the callback functions on a separated file but I can't 
>> visualize it without start to touch it more deeply. Probably you guys 
>> have a better view of this implementation possibilities.
> 
> All the above sysfs layout reflects internal SCST objects:
> 
> 1. targets/target_driver_name and drivers/target_driver_name -> struct 
> scst_tgt_template
> 
> 2. targets/target_driver_name/target -> struct scst_tgt
> 
> 3. targets/target_driver_name/target/sessions/session -> struct 
> scst_session
> 
> 4. targets/target_driver_name/target/ini_groups/group -> struct 
> scst_acg. (For targets/target_driver_name/target/luns each struct 
> scst_tgt would have internal struct scst_acg.)
> 
> 5. 
> targets/target_driver_name/target/ini_groups/group/initiators/initiator 
> -> struct scst_acn
> 
> 6. targets/target_driver_name/target/ini_groups/group/luns/lun and 
> targets/target_driver_name/target/luns/lun -> struct scst_acg_dev
> 
> 7. devices/device -> struct scst_device
> 
> 8. drivers/dev_handler_name -> struct scst_dev_type
> 
> 9. sgv/pool -> struct sgv_pool
> 
> So, we should simply add kobjects in them. To manipulate with then 
> already there is an API, used by procfs, and should be also used by 
> sysfs. This API is completely out of scst_proc.c
> 

great, this will save some time.

> 
> LAYOUT UPDATE.
> 
> Looks like it would be better to move entries from 
> drivers/target_driver_name to targets/target_driver_name to keep all the 
> related entries in one place. Then to reflect new state rename drivers/ 
> to back_drivers/.

Ack.

Thanks,
Daniel Debonzi