Re: adding swap workarounds oom - was: Re: Out of Memory: Killed process 16498 (java).

[Andi Kleen <ak@suse.de>]

On Wednesday 01 March 2006 15:18, Jens Axboe wrote:
> On Wed, Mar 01 2006, Andi Kleen wrote:
> > On Wednesday 01 March 2006 14:41, Jens Axboe wrote:
> > > On Wed, Mar 01 2006, Andy Chittenden wrote:
> > > > with revised patch that does:
> > > > 
> > > >                 printk("sg%d: dma=%llx, dma_len=%u/%u, pfn=%lu\n", i,
> > > > (unsigned long long) sg->dma_address, sg->dma_length, sg->offset,
> > > > page_to_pfn(sg->page));
> > > 
> > > That is correct, thanks!
> > > 
> > > > hda: DMA table too small
> > > > ide dma table, 255 entries, bounce pfn 1310720
> > > > sg0: dma=81c8800, dma_len=4096/0, pfn=1296369
> > > 
> > > Still the same badness here, it's 2kb into a page so straddles two pages
> > > for one entry.
> > 
> > That's normal if it was in the IOMMU and a merged entry. 
> > 
> > You can try iommu=nomerge.
> > 
> > Or maybe the higher layers are passing in physically continuous pages
> > that get merged? Not too unlikely at boot.
> 
> But that would have to be 1kb or 512b io going in

Yes. Why not?

> , I would expect 4kb to 
> be the base entries for any normal type of setup (with a 4kb fs). The
> 8kb could be one such merged entry, I agree.

Probably yes.

> > > Andi, any idea what is going on here? Why is this throwing up all of a
> > > sudden??
> > 
> > What is throwing up exactly?
> >
> > There was a change recently in the merging algorithm, but it shouldn't
> > cause any bad side effects for correct users of *_map_sg()
> 
> That the request we end up passing to blk_rq_map_sg() and then to
> pci_map_sg() ends up with more entries than the driver advertised. So
> far I think only Andy reported this, and then only with your
> blk_queue_bounce_limit() patch applied.

It shouldn't end up with more, only with less.

Does iommu=nomerge make a difference?

-Andi