Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Subject: Re: dma_mask limited to 32-bits with OF platform device
To:     Roger Quadros <rogerq@ti.com>, Christoph Hellwig <hch@lst.de>,
        =?UTF-8?Q?P=c3=a9ter_Ujfalusi?= <peter.ujfalusi@ti.com>,
        Murali Karicheri <m-karicheri2@ti.com>,
        "Nori, Sekhar" <nsekhar@ti.com>, "Anna, Suman" <s-anna@ti.com>
Cc:     stefan.wahren@i2se.com, afaerber@suse.de, hverkuil@xs4all.nl,
        Rob Herring <robh+dt@kernel.org>, devicetree@vger.kernel.org,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        Nishanth Menon <nm@ti.com>
References: <c1c75923-3094-d3fc-fe8e-ee44f17b1a0a@ti.com>
From:   Robin Murphy <robin.murphy@arm.com>
Message-ID: <3a91f306-f544-a63c-dfe2-7eae7b32bcca@arm.com>
Date:   Wed, 12 Feb 2020 11:37:06 +0000
User-Agent: Mozilla/5.0 (Windows NT 10.0; rv:68.0) Gecko/20100101
 Thunderbird/68.4.2
MIME-Version: 1.0
In-Reply-To: <c1c75923-3094-d3fc-fe8e-ee44f17b1a0a@ti.com>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Language: en-GB
Content-Transfer-Encoding: 8bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 2020-02-12 10:49 am, Roger Quadros wrote:
> Hi,
> 
> I'd like to understand why of_dma_configure() is limiting the dma and 
> coherent masks
> instead of overriding them.
> 
> see commits
> a5516219b102 ("of/platform: Initialise default DMA masks")
> ee7b1f31200d ("of: fix DMA mask generation")
> 
> In of_platform_device_create_pdata(), we initialize both masks to 
> 32-bits unconditionally,
> which is fine to support legacy cases.
> 
>      dev->dev.coherent_dma_mask = DMA_BIT_MASK(32);
>          if (!dev->dev.dma_mask)
>                  dev->dev.dma_mask = &dev->dev.coherent_dma_mask;
> 
> Then in of_dma_configure() we limit it like so.
> 
>          dev->coherent_dma_mask &= mask;
>          *dev->dma_mask &= mask;
> 
> This way, legitimate devices which correctly set dma-ranges in DT
> will never get a dma_mask above 32-bits at all. How is this expected to 
> work?

Because these are still just the *default* masks - although drivers are 
all expected to call dma_set_mask() and friends explicitly nowadays, 
there are sure to be some out there for 32-bit devices still assuming 
the default 32-bit masks are in place. Thus if platform code secretly 
makes them larger, Bad Things ensue. Making them *smaller* where there 
are platform limitations shouldn't really matter now that we have the 
bus_dma_limit mechanism working well, but also doesn't do any harm, so 
it was left in for good measure.

The current paradigm is that the device masks represent the inherent 
capability of the device as far as the driver knows, and external 
interconnect constraints are kept separately as private DMA API 
internals via the bus limit.

> For a test, I added this in dra7.dtsi sata node. (NOTE: CONFIG_ARM_LPAE=y)
> 
> diff --git a/arch/arm/boot/dts/dra7.dtsi b/arch/arm/boot/dts/dra7.dtsi
> index 93aa65c75b45..cd8c6cea23d5 100644
> --- a/arch/arm/boot/dts/dra7.dtsi
> +++ b/arch/arm/boot/dts/dra7.dtsi
> @@ -571,6 +571,8 @@
>           sata: sata@4a141100 {
>               compatible = "snps,dwc-ahci";
>               reg = <0x4a140000 0x1100>, <0x4a141100 0x7>;
> +            #size-cells = <2>;
> +            dma-ranges = <0x00000000 0x00000000 0x10 0x00000000>;
>               interrupts = <GIC_SPI 49 IRQ_TYPE_LEVEL_HIGH>;
>               phys = <&sata_phy>;
>               phy-names = "sata-phy";
> 
> ----------------------------- drivers/of/device.c 
> -----------------------------
> index e9127db7b067..1072cebad57a 100644
> @@ -95,6 +95,7 @@ int of_dma_configure(struct device *dev, struct 
> device_node *np, bool force_dma)
>       const struct iommu_ops *iommu;
>       u64 mask, end;
> 
> +    dev_info(dev, "of_dma_configure\n");
>       ret = of_dma_get_range(np, &dma_addr, &paddr, &size);
>       if (ret < 0) {
>           /*
> @@ -123,7 +124,8 @@ int of_dma_configure(struct device *dev, struct 
> device_node *np, bool force_dma)
>               dev_err(dev, "Adjusted size 0x%llx invalid\n", size);
>               return -EINVAL;
>           }
> -        dev_dbg(dev, "dma_pfn_offset(%#08lx)\n", offset);
> +        dev_info(dev, "dma %llx paddr %llx size %llx\n", dma_addr, 
> paddr, size);
> +        dev_info(dev, "dma_pfn_offset(%#08lx)\n", offset);
>       }
> 
>       /*
> @@ -152,6 +154,8 @@ int of_dma_configure(struct device *dev, struct 
> device_node *np, bool force_dma)
>       mask = DMA_BIT_MASK(ilog2(end) + 1);
>       dev->coherent_dma_mask &= mask;
>       *dev->dma_mask &= mask;
> +
> +    dev_info(dev, "end %llx, mask %llx\n", end, mask);
>       /* ...but only set bus limit if we found valid dma-ranges earlier */
>       if (!ret)
>           dev->bus_dma_limit = end;
> 
> And I see.
> 
> [    1.134294]  4a140000.sata: of_platform
> [   13.203917] ahci 4a140000.sata: of_dma_configure
> [   13.225635] ahci 4a140000.sata: dma 0 paddr 0 size 1000000000
> [   13.266178] ahci 4a140000.sata: dma_pfn_offset(0x000000)
> [   13.297621] ahci 4a140000.sata: end fffffffff, mask fffffffff
> [   13.585499] ahci 4a140000.sata: dma_mask 0xffffffff, coherent_mask 
> 0xffffffff
> [   13.599082] ahci 4a140000.sata: setting 64-bit mask ffffffffffffffff
> 
> Truncation of dma_mask and coherent_mask is undesired in this case.

Again, it's only the initial default masks that are truncated, and the 
driver appropriately replaces them with its 64-bit masks anyway once it 
probes. However, bus_dma_limit should still reflect that 36-bit upstream 
constraint, and that's what really matters. If you've found a path 
through a DMA API implementation which is subsequently failing to 
respect that limit, that's a bug in that implementation.

> How about fixing it like so?
> 
> -     dev->coherent_dma_mask &= mask;
> -    *dev->dma_mask &= mask;
> +     dev->coherent_dma_mask = mask;
> +     *dev->dma_mask = mask;

As above, making the "32-bit default" larger than 32 bits stands to 
break 32-bit devices with old drivers, and there's nothing to "fix" at 
this point anyway.

> Also this comment doesn't make sense anymore?
> 
>          /*
>           * Limit coherent and dma mask based on size and default mask
>           * set by the driver.
>           */

TBH that's never made much sense, unless "driver" was supposed to refer 
to bus code. Its continued presence is down to inertia more than any 
other reason :)

Robin.