Re: more iommu sg merging fallout

[FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>]

On Tue, 05 Feb 2008 20:41:38 -0800 (PST)
David Miller <davem@davemloft.net> wrote:

> 
> The sparc64 change:
> 
>     commit fde6a3c82d67f592eb587be4d12222b0ae6d4321
>     Author: FUJITA Tomonori <tomof@acm.org>
>     Date:   Mon Feb 4 22:28:02 2008 -0800
> 
>     iommu sg merging: sparc64: make iommu respect the segment size limits
>     
>     This patch makes iommu respect segment size limits when merging sg
>     lists.
>     
>     Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
>     Cc: Jeff Garzik <jeff@garzik.org>
>     Cc: James Bottomley <James.Bottomley@steeleye.com>
>     Acked-by: Jens Axboe <jens.axboe@oracle.com>
>     Cc: "David S. Miller" <davem@davemloft.net>
>     Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
>     Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
> 
> has significant errors and is going to eat people's disks, as it just
> nearly did to mine.

Really sorry about it.


> Typically what you'll see are NULL pointer derefernces in
> dma_4v_map_sg() and dma_4u_map_sg() and then the kernel usually craps
> on your superblock very shortly thereafter.
> 
> The changeset above modified only prepare_sg() but that is only the
> first pass of the SG mapping algorithm of the sparc64 IOMMU layer.
> 
> The second pass that fills in the entries depends upon how the first
> pass does things.  So if you change the first pass decision making you
> have to update the second pass's as well.
>
> That second pass is implemented in fill_sg() (there is a version in
> both arch/sparc64/kernel/iommu.c and arch/sparc64/kernel/pci_sun4v.c),
> which probably needs new logic as was added to prepare_sg() to handle
> dma_get_max_seg_size().

PARISC, Alpha, and IA64 IOMMUs use the two-pass algorithm like SPARC
but their first pass decides how to merge sg entires (and stores that
information in the sg entries), then the second pass simpliy follows
it (Hopefully I understand these IOMMUs correctly, or else I break
them too).

I'll try this work again for 2.6.26. The SPARC IOMMUs are the most
complicated and seems that few people tests the -mm SPARC code. So I
guess that I need to find a SPARC box...

Thanks,