Re: regression: CD burning (k3b) went broke

[Jeff Garzik <jeff@garzik.org>]

Andrew Morton wrote:
> On Tue, 26 Feb 2008 14:36:43 +0100 Mike Galbraith <efault@gmx.de> wrote:
> 
>> On Tue, 2008-02-26 at 10:48 +0100, Mike Galbraith wrote:
>>> Greetings,
>>>
>>> I straced both a good and a bad kernel (good being .git with attached
>>> revert patch applied) and filtered/diffed/merged the output.  Scroll
>>> down to "HERE" to see the problem (resid).
>>>
>>> I'm poking around, but not having much luck.
> 
> cc's added.
> 
> I'm told this is part of "Tejun's DMA drain handling".

Correct.


>> Seems the problem is data_len changes, but raw_data_len doesn't.  I've
>> not the foggiest IO-land clue, but k3b works again, so the below may
>> have some diagnostic value.
> 
> So this change fixes a bug?  Can we have a recap of how it does this?
> 
>> diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
>> index ba21d97..7a6f784 100644
>> --- a/drivers/scsi/scsi_lib.c
>> +++ b/drivers/scsi/scsi_lib.c
>> @@ -871,7 +871,7 @@ void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
>>  			scsi_end_bidi_request(cmd);
>>  			return;
>>  		}
>> -		req->data_len = scsi_get_resid(cmd);
>> +		req->data_len = req->raw_data_len = scsi_get_resid(cmd);
>>  	}

I would love to get an answer as to what data_len (and of course 
raw_data_len) should be set to AFTER the command completes, which is 
what is going on here.

I can see the above being correct -- scsi_get_resid() returns the length 
of the left-over data after the command is processed -- but I am mainly 
curious why setting [raw_]data_len matters after I/O completion.

	Jeff