Re: cdrom: use kmalloced buffers instead of buffers on stack

[Jens Axboe <jens.axboe@oracle.com>]

On Tue, Apr 22 2008, FUJITA Tomonori wrote:
> On Tue, 22 Apr 2008 07:48:58 +0200
> tsbogend@alpha.franken.de (Thomas Bogendoerfer) wrote:
> 
> > On Mon, Apr 21, 2008 at 10:01:26PM -0400, Jeff Garzik wrote:
> > > >    cdrom: use kmalloced buffers instead of buffers on stack
> > > >    
> > > >    If cdrom commands are issued to a scsi drive in most cases the buffer 
> > > >    will be
> > > >    filled via dma.  This leads to bad stack corruption on non coherent 
> > > >    platforms,
> > > >    because the buffers are neither cache line aligned nor is the size a 
> > > >    multiple
> > > >    of the cache line size.  Using kmalloced buffers avoids this.
> > > >    
> > > >    Signed-off-by: Thomas Bogendoerfer <tsbogend@alpha.franken.de>
> > > >    Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
> > > >---
> > > > drivers/cdrom/cdrom.c |  274 
> > > > ++++++++++++++++++++++++++++++++-----------------
> > > > 1 files changed, 181 insertions(+), 93 deletions(-)
> > > 
> > > Eh... AFAICS this is only really useful in two of the cases converted.
> > > 
> > > For all the other cases (<= 32 bytes), it is _far_ less complex, far 
> > > less code to simply communicate the additional alignment requirements to 
> > > the compiler.
> > > 
> > > What about __attribute__ __aligned__?  Was that tried?
> > 
> > I used that while narrowing down the bug. But not only the alignment is
> > important, but also size needs to be a multiple of the cache line size.
> > Which means it needs to be 128 bytes for most SGI machines. That
> > and the following in DMA-mapping.txt
> > 
> > "This rule also means that you may use neither kernel image addresses
> > (items in data/text/bss segments), nor module image addresses, nor
> > stack addresses for DMA."
> > 
> > let me choose the kmalloc() solution.
> 
> Can we advertise such architecture's dma restrictions? For example, if
> we can update dma_pad_mask and dma_alignment in request_queue,
> blk_rq_map_kern uses a proper bounce buffer for such
> architectures. Then we can avoid putting extra complexity in uppper
> drivers such as cdrom.c

That would work fine, if cdrom was then also updated to get rid of
->generic_packet() and use the regular queue transport instead. Which
would be a VERY nice thing to do anyway, so I'd welcome the effort :-)

-- 
Jens Axboe