Re: 2.6.24-rc1-54866f032307063776b4eff7eadb131d47f9f9b4 fails to boot: kernel BUG at include/linux/scatterlist.h:49!

["Torsten Kaiser" <just.for.lkml@googlemail.com>]

On 11/2/07, Stefan Richter <stefanr@s5r6.in-berlin.de> wrote:
> Torsten Kaiser wrote:
> > Both 2.6.24-rc1 and the current git version fail to boot for me:
> >
> > [   57.182205] ohci1394: fw-host0: OHCI-1394 1.1 (PCI): IRQ=[16]
> > MMIO=[effff800-efffffff]  Max Packet=
> > [2048]  IR/IT contexts=[4/8]
> > [   57.194032] eth1394: eth2: IPv4 over IEEE 1394 (fw-host0)
> > [   57.199527] ------------[ cut here ]------------
> > [   57.204154] kernel BUG at include/linux/scatterlist.h:49!
>
> To which extent do you need IEEE 1394 drivers?

Using eth1394 as primary network connection on this computer.
So switching to the new stack is currently not an option, or did I
missing something?
(I also have a firewire disk, but that is not used right now)

> > from gdb:
> > 0xffffffff80463512 is in dma_region_alloc (include/linux/scatterlist.h:49).
> > 44               * In order for the low bit stealing approach to work, pages
> > 45               * must be aligned at a 32-bit boundary as a minimum.
> > 46               */
> > 47              BUG_ON((unsigned long) page & 0x03);
> > 48      #ifdef CONFIG_DEBUG_SG
> > 49              BUG_ON(sg->sg_magic != SG_MAGIC);
> > 50      #endif
> > 51              sg->page_link = page_link | (unsigned long) page;
> > 52      }
> > 53
>
> Uh oh...

Looking that calltrace upwards, it seems replacing the
memset(dma->sglist,...) with sg_init_table(...) would fix the BUG_ON()
as that inits the SG_MAGIC. But I do not trust myself to fixing all
the iterators correctly to convert this completely to the new API.

What I noticed trying to find the definition of sg_dma_address() is
that all architectures seem to implement identical macros for
sg_dma_address() and sg_dma_length(). The only difference is that some
archs call their fields different.

But I stopped looking if unifiying all names was possible when I found
swiotlb_map_sg() in lib/swiotlb.c...

Why is there a comment about using "sg_dma_{address,length}(SG)" to
obtain the "appropriate dma address and length" when I see no instance
of this macros there? Instead there are direct accesses to sg->length
and sg->dma_length. And the field dma_length only exists in some
arches! (alpha, ia64, powerpc only if __powerpc64__ is defined,
sparc64 and the 64bit variant of x86)
On the other hand I can't find any other "iova_length" in the hole
tree apart from include/asm-parisc/scatterlist.h...

... so I can't really make sense of the sg mechanism and abstain from
trying to fix this myself.

Torsten