RE: [RFC PATCH] iommu/dma: check pci host bridge dma_mask for IOVA allocation

[Oza Oza]

My responses inline:

-----Original Message-----
From: Robin Murphy [mailto:robin.murphy@xxxxxxx]
Sent: Tuesday, March 14, 2017 4:27 PM
To: Oza Pawandeep; Joerg Roedel
Cc: iommu@xxxxxxxxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
linux-arm-kernel@xxxxxxxxxxxxxxxxxxx;
bcm-kernel-feedback-list@xxxxxxxxxxxx; arnd@xxxxxxxx; Nikita Yushchenko
Subject: Re: [RFC PATCH] iommu/dma: check pci host bridge dma_mask for
IOVA allocation

On 14/03/17 08:48, Oza Pawandeep wrote:
> It is possible that PCI device supports 64-bit DMA addressing, and
> thus it's driver sets device's dma_mask to DMA_BIT_MASK(64), however
> PCI host bridge may have limitations on the inbound transaction
> addressing. As an example, consider NVME SSD device connected to
> iproc-PCIe controller.

Aw heck, not another one :(

> Currently, the IOMMU DMA ops only considers PCI device dma_mask when
> allocating an IOVA. This is particularly problematic on
> ARM/ARM64 SOCs where the IOMMU (i.e. SMMU) translates IOVA to PA for
> in-bound transactions only after PCI Host has forwarded these
> transactions on SOC IO bus. This means on such ARM/ARM64 SOCs the IOVA
> of in-bound transactions has to honor the addressing restrictions of
> the PCI Host.

Depending on whether this most closely matches the R-Car situation[1] or
the X-Gene situation[2], this may not address the real problem at all (if
the non-IOMMU case is also affected). Either way it also fails to help
non-PCI devices which can face the exact same problem.

I am not fully aware of R-car and X-gene situation, but for iproc based
SOCs, our pcie host floats addresses in higher order dma_mask is set.

> This patch tries to solve above described IOVA allocation problem by:
> 1. Adding iommu_get_dma_mask() to get dma_mask of any device 2. For
> PCI device, iommu_get_dma_mask() compare dma_mask of PCI device and
> corresponding PCI Host dma_mask (if set).
> 3. Use iommu_get_dma_mask() in IOMMU DMA ops implementation instead of
> dma_get_mask()

Sorry, but NAK to this approach - not only is the implementation rather
ugly, and incomplete as above, but the fact that it starts with a clear
description of the underlying problem and then goes on to bodge around it
down the line instead of actually fixing it is the kicker.

:)
that is why this is RFC patch, and I had something similar in mind to go
and poke around which you have provided.. [1] and [3].
Let me try this on our platform, and see if that solves the issue.

I've got a simple arm64 patch for proper DMA mask inheritance based on the
various discussions at [1] which I will try to clean up and send out for
4.12 if nobody else sends anything soon, but it does depend on getting [3]
merged first to avoid regressing certain platforms.

Robin.

[1]:http://www.mail-archive.com/linux-kernel@xxxxxxxxxxxxxxx/msg1306545.ht
ml
[2]:http://www.spinics.net/lists/arm-kernel/msg552422.html
[3]:http://www.spinics.net/lists/arm-kernel/msg566947.html



> Signed-off-by: Oza Pawandeep <oza.oza@xxxxxxxxxxxx>
> Reviewed-by: Anup Patel <anup.patel@xxxxxxxxxxxx>
> Reviewed-by: Scott Branden <scott.branden@xxxxxxxxxxxx>
> ---
>  drivers/iommu/dma-iommu.c | 44
> ++++++++++++++++++++++++++++++++++++++++----
>  1 file changed, 40 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
> index 48d36ce..e93e536 100644
> --- a/drivers/iommu/dma-iommu.c
> +++ b/drivers/iommu/dma-iommu.c
> @@ -108,6 +108,42 @@ int iommu_get_dma_cookie(struct iommu_domain
> *domain)  }  EXPORT_SYMBOL(iommu_get_dma_cookie);
>
> +static u64 __iommu_dma_mask(struct device *dev, bool is_coherent) {
> +#ifdef CONFIG_PCI
> +	u64 pci_hb_dma_mask;
> +
> +	if (dev_is_pci(dev)) {
> +		struct pci_dev *pdev = to_pci_dev(dev);
> +		struct pci_host_bridge *br =
pci_find_host_bridge(pdev->bus);
> +
> +		if ((!is_coherent) && !(br->dev.dma_mask))
> +			goto default_dev_dma_mask;
> +
> +		/* pci host bridge dma-mask. */
> +		pci_hb_dma_mask = (!is_coherent) ? *br->dev.dma_mask :
> +				   br->dev.coherent_dma_mask;
> +
> +		if (pci_hb_dma_mask && ((pci_hb_dma_mask) <
(*dev->dma_mask)))
> +			return pci_hb_dma_mask;
> +	}
> +default_dev_dma_mask:
> +#endif
> +	return (!is_coherent) ? dma_get_mask(dev) :
> +		dev->coherent_dma_mask;
> +}
> +
> +static u64 __iommu_dma_get_coherent_mask(struct device *dev) {
> +	return __iommu_dma_mask(dev, true);
> +}
> +
> +static u64 __iommu_dma_get_mask(struct device *dev) {
> +	return __iommu_dma_mask(dev, false); }
> +
> +
>  /**
>   * iommu_get_msi_cookie - Acquire just MSI remapping resources
>   * @domain: IOMMU domain to prepare
> @@ -461,7 +497,7 @@ struct page **iommu_dma_alloc(struct device *dev,
size_t size, gfp_t gfp,
>  	if (!pages)
>  		return NULL;
>
> -	iova = __alloc_iova(domain, size, dev->coherent_dma_mask, dev);
> +	iova = __alloc_iova(domain, size,
> +__iommu_dma_get_coherent_mask(dev), dev);
>  	if (!iova)
>  		goto out_free_pages;
>
> @@ -532,7 +568,7 @@ static dma_addr_t __iommu_dma_map(struct device
*dev, phys_addr_t phys,
>  	struct iova_domain *iovad = cookie_iovad(domain);
>  	size_t iova_off = iova_offset(iovad, phys);
>  	size_t len = iova_align(iovad, size + iova_off);
> -	struct iova *iova = __alloc_iova(domain, len, dma_get_mask(dev),
dev);
> +	struct iova *iova = __alloc_iova(domain, len,
> +__iommu_dma_get_mask(dev), dev);
>
>  	if (!iova)
>  		return DMA_ERROR_CODE;
> @@ -690,7 +726,7 @@ int iommu_dma_map_sg(struct device *dev, struct
scatterlist *sg,
>  		prev = s;
>  	}
>
> -	iova = __alloc_iova(domain, iova_len, dma_get_mask(dev), dev);
> +	iova = __alloc_iova(domain, iova_len, __iommu_dma_get_mask(dev),
> +dev);
>  	if (!iova)
>  		goto out_restore_sg;
>
> @@ -760,7 +796,7 @@ static struct iommu_dma_msi_page
> *iommu_dma_get_msi_page(struct device *dev,
>
>  	msi_page->phys = msi_addr;
>  	if (iovad) {
> -		iova = __alloc_iova(domain, size, dma_get_mask(dev), dev);
> +		iova = __alloc_iova(domain, size,
__iommu_dma_get_mask(dev), dev);
>  		if (!iova)
>  			goto out_free_page;
>  		msi_page->iova = iova_dma_addr(iovad, iova);
>