Re: bisected crash due to "x86: Move swiotlb initialization beforedma32_free_bootmem"

[FUJITA Tomonori]

On Tue, 15 Dec 2009 01:11:05 -0800
Yinghai Lu <yinghai@xxxxxxxxxx> wrote:

> FUJITA Tomonori wrote:
> > On Mon, 14 Dec 2009 23:47:07 -0800
> > Roland Dreier <rdreier@xxxxxxxxx> wrote:
> > 
> >> I have a big box (64 threads, 256GB memory) that is crashing early in
> >> boot as below.  I bisected it down to f4780ca0 ("x86: Move swiotlb
> >> initialization before dma32_free_bootmem"); reverting just this commit
> >> from the latest git (3ea6b3d0 is what I tested) fixes things.
> > 
> > Ah, really sorry about that.
> > 
> > 
> >> I haven't tried to debug this yet, but I guess on such a huge box there
> >> is not enough memory below 4GB for swiotlb if we don't free the 
> > 
> > Yeah, Yinghai also hit this (his box has more memory than yours).
> > 
> > 
> >> stuff allocated earlier?  I don't know why that would be, since the
> >> bootmem is grabbing 512MB and I have pretty close to 4GB below 4GB.
> >> Anyway, I'm going to go to bed soon, but if you need more information or
> >> have anything you want me to try, I will do it tomorrow morning.
> > 
> > http://www.kernel.org/pub/linux/kernel/people/tomo/misc/0001-x86-two-stage-swiotlb-initialization.patch
> > 
> > It makes the swiotlb initialization into two stages. I don't like it
> > much since I like to avoid complicating the initialization.
> > 
> > dma32_reserve_bootmem() allocates 128MB for broken GART IOMMU but I
> > think 64MB should be enough since broken GART IOMMU allocates
> > 64MB. The following simple patch might work too because swiotlb uses
> > 64MB.
> > 
> > With coming huge memory boxes, we might need to work on ZONE_DMA32
> > shortage issue anyway (sparse-vmemmap, anything else)?
> 
> maybe just revert f4780ca0... for now

As I wrote, I think that the following patch works.

http://www.kernel.org/pub/linux/kernel/people/tomo/misc/0001-x86-two-stage-swiotlb-initialization.patch

There are a few people who hit this. How many people use a box with
over 256GB memory?

And you can work around this with "dma32_size" kernel boot option.


> actually dma32_free_bootmem will also make sure it will give some buffer to pci_swiotlb_init...

But unlike broken GART BIOS, swiotlb doesn't set the goal. swiotlb had
been fine without dma32_reserve_bootmem(). We use more DMA32_ZONE than
we did though.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/