Re: audio cd writing causes massive swap and crash

[Ed Sweetman <safemode@comcast.net>]

Martin Schlemmer wrote:

>On Thu, 2004-07-22 at 14:54, Jens Axboe wrote:
>  
>
>>On Thu, Jul 22 2004, Ed Sweetman wrote:
>>    
>>
>>>I've had other people test writing.  It appears that scsi-emu is not 
>>>effected by this memory leak when writing audio cds.  So it would appear 
>>>that ide-cd along with any of the dependent ide source files is the 
>>>culprit. But I cannot find anywhere in ide-cd that is apparent to being 
>>>a mem leak.  There are various conditions in ide_do_drive_cmd that state 
>>>that the cdrom driver has to be very careful about handling but without 
>>>intimate knowledge of the driver, I can't be sure that it's sufficiently 
>>>handling those situations.  
>>>
>>>Surprisingly, it's very hard to find anyone who's used the native atapi 
>>>mode to write an audio cd in 2.6.  Which is partly why this problem 
>>>hasn't generated more mail traffic here I would guess. 
>>>      
>>>
>>That's not true, lots of people use it. But, oddly, the leak isn't
>>reproducable on any machine I've tested.
>>    
>>
>
>I seem to remember he noted a patch about dma during audio writing,
>and his 'testing' if it might be the cause was to just disable dma
>on the drive ...
>  
>


The patch is now part of vanilla 2.6.   It was an audio dma api that was 
a workaround for the broken dma api that only allows ide commands 
512bytes long.  Schilling mentioned something about this in an earlier 
lkml posting around january.  

I've asked some other people in an irc channel to test out the problem.  
Basically You need to be using the native atapi method for recording 
audio.   Also, it appears that the mem leak is just what happens to some 
people, other people like some of those who tested in the irc channel 
experienced random program sigsevs and no mem leak at all.  It would 
appear from this that what may be happening is the cdrom ide module is 
mangling a pointer and either can't free it due to it possibly randomly 
pointing to null or cause crashes due to it randomly pointing to other 
anonymous/shared memory regions as kfree is called. That's obviously 
just a guess.  But some sort of memory mangling is apparently 
happening.  This is not just happening on my computer. 


My drive is using udma33 and not mdma so maybe that makes the problem 
occur more quickly.    I disabled dma to test it writing the old way but 
it appears to be occuring in ide-cd or somewhere between cdrom.o and the 
block device layer.  ide-scsi people dont have the problem at all 
apparently.