[PATCH 2/2] nvme-pci: use dma_alloc_noncontigous if possible

[Christoph Hellwig]

On Mon, Dec 02, 2024 at 09:05:41PM +0200, Leon Romanovsky wrote:
> On Fri, Nov 01, 2024 at 05:40:05AM +0100, Christoph Hellwig wrote:
> > Use dma_alloc_noncontigous to allocate a single IOVA-contigous segment
> > when backed by an IOMMU.  This allow to easily use bigger segments and
> > avoids running into segment limits if we can avoid it.
> > 
> > Signed-off-by: Christoph Hellwig <hch at lst.de>
> > ---
> >  drivers/nvme/host/pci.c | 58 +++++++++++++++++++++++++++++++++++++----
> >  1 file changed, 53 insertions(+), 5 deletions(-)
> 
> <...>
> 
> > +static int nvme_alloc_host_mem_multi(struct nvme_dev *dev, u64 preferred,
> >  		u32 chunk_size)
> >  {
> >  	struct nvme_host_mem_buf_desc *descs;
> > @@ -2049,9 +2086,18 @@ static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
> >  	u64 hmminds = max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
> >  	u64 chunk_size;
> >  
> > +	/*
> > +	 * If there is an IOMMU that can merge pages, try a virtually
> > +	 * non-contiguous allocation for a single segment first.
> > +	 */
> > +	if (!(PAGE_SIZE & dma_get_merge_boundary(dev->dev))) {
> > +		if (!nvme_alloc_host_mem_single(dev, preferred))
> > +			return 0;
> > +	}
> 
> We assume that the addition of the lines above are the root cause of the
> following panic during boot. It is happening when we are trying to allocate
> 61 MiB chunk.

We should not hit this path for dma-direct, but only dma-iommu that can
do discotigous allocations.  If we hit it with dma-direct I got my math
on the boundary check above wrong.  I'll try to figure out what is
wrong, but I'm actually in Japan for meetings, so things will not only
be delayed but also in a weird time zone.

> -       if (!(PAGE_SIZE & dma_get_merge_boundary(dev->dev))) {
> +       if (!(PAGE_SIZE & dma_get_merge_boundary(dev->dev)) && preferred < max_chunk) {

That would basically be a revert of the new funtionality..