Re: [PATCH 00/11] st: remove scsi_execute_async usage (the first half)

[Boaz Harrosh <bharrosh@panasas.com>]

Vladislav Bolkhovitin wrote:
> FUJITA Tomonori wrote:
>> This patchset removes the majority of scsi_execute_async in st
>> driver. IOW, this converts st driver to use the block layer functions.
>>
>> We are in the process of removing scsi_execute_async and
>> scsi_req_map_sg. scsi_execute_async in sg driver were removed in
>> 2.6.27. Now only st and osst drivers use scsi_execute_async().
>>
>> st driver uses scsi_execute_async for two purposes, performing sort
>> management SCSI commands internally and data transfer between user and
>> kernel space (st_read and st_write). This patchset converts the
>> former.
>>
>> The former performs SCSI commands synchronously. It uses a liner
>> in-kernel buffer (not scatter gather) for data transfer. To perform
>> such commands, other upper layer drivers (e.g. sd) use
>> scsi_execute_req (internally uses blk_rq_map_kern and and
>> blk_execute_rq). scsi_execute_req is not fully fit for st so I adds a
>> helper function similar to scsi_execute_req and replace
>> scsi_execute_async in st driver (I might modify scsi_execute_req for
>> st and remove the helper function later).
>>
>> I'm still working on converting scsi_execute_async for data transfer
>> between user and kernel space but I want to get this first half be
>> merged.
> 
> May I ask a possibly stupid question? Is amount of overall code in Linux 
> SCSI subsystem after your conversion is fully done and 
> scsi_execute_async() with scsi_req_map_sg() completely removed going to 
> get smaller or bigger?
> 
>  From diffstats of your patches seems it's is going to get considerably 
> bigger (171 insertion vs 61 deletion). And this is only in one user of 
> scsi_execute_async() of 4. What's the point then in scsi_execute_async() 
> removal if the resulting code is going to get bigger, not smaller? And, 
> frankly, it doesn't look as it's going to get clearer too..
> 
> P.S. Scsi_execute_async() is just ~50 lines long, scsi_req_map_sg() is 
> just ~80 lines long and both look like nice helper functions, which are 
> not worse than st_scsi_kern_execute().
> 
> Regards,
> Vlad
> 

No! you are not looking at the full picture. See previous conversion of
sg.c driver. What you see here is only the rq_map_kern  side which is
fine. But the ugliness starts with when you need sg-lists.

With Scsi_execute_async you go from:
	memory=>sg-lists=>bios=>sg-lists=>driver
going directly to blk you save the allocation of the first sg-list and
translation to BIOs. And the ugly self-made preparation of the sg-list
by each driver again and again. Not talking about correctness. Like when
the block layer takes care of alignment, draining, sizes and all that funny
stuff all kind of drivers need (sa*cough*ta)

The net effect is a very big gain. both in code and specially in resources
and performance.

Note that all users of block layer have user-memory, kernel-pointer, page*.
Only target driver happens to be on the other side and actually hold an sg-list
at hand. That's because, in principle, it is down-side-up.

Boaz