[PATCH 6/8] iommu/arm-smmu: add support for non-pci devices

[thunder.leizhen@huawei.com (leizhen)]

On 2015/6/30 19:26, Robin Murphy wrote:
> On 26/06/15 09:33, Zhen Lei wrote:
>> Now, we only support a master with only one stream id. It will cover most
>> hardware platforms and coding so easy.
>>
>> Please refer Documentation\devicetree\bindings\iommu\iommu.txt on how to
>> bind device tree.
>>
>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
>> ---
>>   arch/arm64/include/asm/device.h |  2 +
>>   drivers/iommu/arm-smmu-v3.c     | 88 +++++++++++++++++++++++++++++++++++++++--
>>   2 files changed, 87 insertions(+), 3 deletions(-)
>>
>> diff --git a/arch/arm64/include/asm/device.h b/arch/arm64/include/asm/device.h
>> index 243ef25..225e4f9 100644
>> --- a/arch/arm64/include/asm/device.h
>> +++ b/arch/arm64/include/asm/device.h
>> @@ -20,6 +20,8 @@ struct dev_archdata {
>>       struct dma_map_ops *dma_ops;
>>   #ifdef CONFIG_IOMMU_API
>>       void *iommu;            /* private IOMMU data */
> 
> There's already a perfectly good place to store driver-private data right here.
> 
>> +    struct device_node *of_smmu;
>> +    u32 sid;
> 
> This looks far too specific to be in core code. It doesn't seem extensible for e.g. ACPI platform devices; it doesn't seem extensible for platform devices with multiple stream IDs e.g. PL330; it also
> doesn't seem (sensibly) extensible for IOMMUs with #iommu-cells > 1.
> 
> Just allocate an SMMU-private struct for this and stash it in archdata.iommu - that way we can change things as much as we like in the driver with zero churn in core code.

OK. If platform devices with multiple stream IDs exist, I will support it in patch V2.

> 
>>   #endif
>>       bool dma_coherent;
>>   };
>> diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c
>> index 483c918..87c3d9b 100644
>> --- a/drivers/iommu/arm-smmu-v3.c
>> +++ b/drivers/iommu/arm-smmu-v3.c
>> @@ -30,9 +30,14 @@
>>   #include <linux/of_address.h>
>>   #include <linux/pci.h>
>>   #include <linux/platform_device.h>
>> +#include <linux/amba/bus.h>
>> +#include <linux/of_iommu.h>
>>
>>   #include "io-pgtable.h"
>>
>> +/* Maximum number of stream IDs assigned to a single device */
>> +#define MAX_MASTER_STREAMIDS        1
>> +
>>   /* MMIO registers */
>>   #define ARM_SMMU_IDR0            0x0
>>   #define IDR0_ST_LVL_SHIFT        27
>> @@ -608,6 +613,22 @@ static struct arm_smmu_domain *to_smmu_domain(struct iommu_domain *dom)
>>       return container_of(dom, struct arm_smmu_domain, domain);
>>   }
>>
>> +static struct arm_smmu_device *find_smmu_for_device(struct device *dev)
>> +{
>> +    struct arm_smmu_device *smmu;
>> +
>> +    spin_lock(&arm_smmu_devices_lock);
>> +    list_for_each_entry(smmu, &arm_smmu_devices, list) {
>> +        if (smmu->dev->of_node == dev->archdata.of_smmu) {
>> +            spin_unlock(&arm_smmu_devices_lock);
>> +            return smmu;
>> +        }
>> +    }
>> +    spin_unlock(&arm_smmu_devices_lock);
>> +
>> +    return NULL;
>> +}
>> +
> 
> This should be unnecessary with the right probe order, see below...
> 
>>   /* Low-level queue manipulation functions */
>>   static bool queue_full(struct arm_smmu_queue *q)
>>   {
>> @@ -1760,9 +1781,36 @@ static int arm_smmu_add_device(struct device *dev)
>>       struct arm_smmu_group *smmu_group;
>>       struct arm_smmu_device *smmu;
>>
>> -    /* We only support PCI, for now */
>> -    if (!dev_is_pci(dev))
>> -        return -ENODEV;
>> +    if (!dev_is_pci(dev)) {
>> +        smmu = find_smmu_for_device(dev);
>> +        if (!smmu)
>> +            return -ENODEV;
>> +
>> +        group = iommu_group_alloc();
>> +        if (IS_ERR(group)) {
>> +            dev_err(dev, "Failed to allocate IOMMU group\n");
>> +            return PTR_ERR(group);
>> +        }
>> +
>> +        ret = iommu_group_add_device(group, dev);
>> +        if (ret)
>> +            goto out_put_group;
>> +
>> +        smmu_group = kzalloc(sizeof(*smmu_group), GFP_KERNEL);
>> +        if (!smmu_group) {
>> +            ret = -ENOMEM;
>> +            goto out_put_group;
>> +        }
>> +
>> +        smmu_group->ste.valid = true;
>> +        smmu_group->smmu = smmu;
>> +        iommu_group_set_iommudata(group, smmu_group,
>> +                        __arm_smmu_release_iommudata);
>> +
>> +        sid = dev->archdata.sid;
>> +
>> +        goto handle_stream_id;
>> +    }
>>
>>       pdev = to_pci_dev(dev);
>>       group = iommu_group_get_for_dev(dev);
>> @@ -1793,6 +1841,8 @@ static int arm_smmu_add_device(struct device *dev)
>>
>>       /* Assume SID == RID until firmware tells us otherwise */
>>       pci_for_each_dma_alias(pdev, __arm_smmu_get_pci_sid, &sid);
>> +
>> +handle_stream_id:
> 
> This is going to get messy quickly - how about breaking out the "platform device" and "PCI device" specifics above into their own functions that return the group and sid data to the common code here?
> 
>>       for (i = 0; i < smmu_group->num_sids; ++i) {
>>           /* If we already know about this SID, then we're done */
>>           if (smmu_group->sids[i] == sid)
>> @@ -1881,7 +1931,23 @@ out_unlock:
>>       return ret;
>>   }
>>
>> +static int arm_smmu_of_xlate(struct device *dev, struct of_phandle_args *args)
>> +{
>> +    if (args->args_count > MAX_MASTER_STREAMIDS) {
>> +        dev_err(dev,
>> +            "reached maximum number (%d) of stream IDs for master device %s\n",
>> +            MAX_MASTER_STREAMIDS, dev->of_node->name);
>> +        return -ENOSPC;
>> +    }
>> +
>> +    dev->archdata.of_smmu = args->np;
>> +    dev->archdata.sid = args->args[0];
>> +
>> +    return 0;
>> +}
> 
> This isn't going to work the way you expect: the way the binding is defined, a master with multiple stream IDs should look like so:
> 
>     iommus = <&smmu 0>, <&smmu 1>,...
> 
> so you'd get multiple calls, never hit the warning, and just end up with whichever ID came last.
> 
> Secondly, as mentioned above, it would be nicer to just associate the arm_smmu_device directly here and obviate the indirect lookup. That would depend on having correct probe ordering, but you need to
> enforce that anyway, since any add_device callbacks before the SMMU itself has been probed will break. Laurent's probe deferral series that Will pointed to is the ultimate goal, but for a stop-gap
> solution which works with the current code I'd suggest taking a look at patches 16 and 17 of Marek's Exynos SysMMU series[1]

OK, thanks. I will consider these in patch v2.

> 
> Robin.
> 
> [1]:http://thread.gmane.org/gmane.linux.kernel.samsung-soc/45416
> 
>> +
>>   static struct iommu_ops arm_smmu_ops = {
>> +    .of_xlate        = arm_smmu_of_xlate,
>>       .capable        = arm_smmu_capable,
>>       .domain_alloc        = arm_smmu_domain_alloc,
>>       .domain_free        = arm_smmu_domain_free,
>> @@ -2655,6 +2721,14 @@ static int __init arm_smmu_init(void)
>>       if (ret)
>>           return ret;
>>
>> +    if (!iommu_present(&platform_bus_type))
>> +        bus_set_iommu(&platform_bus_type, &arm_smmu_ops);
>> +
>> +#ifdef CONFIG_ARM_AMBA
>> +    if (!iommu_present(&amba_bustype))
>> +        bus_set_iommu(&amba_bustype, &arm_smmu_ops);
>> +#endif
>> +
>>       return bus_set_iommu(&pci_bus_type, &arm_smmu_ops);
>>   }
>>
>> @@ -2666,6 +2740,14 @@ static void __exit arm_smmu_exit(void)
>>   subsys_initcall(arm_smmu_init);
>>   module_exit(arm_smmu_exit);
>>
>> +static int arm_smmu_of_iommu_init(struct device_node *np)
>> +{
>> +    of_iommu_set_ops(np, &arm_smmu_ops);
>> +
>> +    return 0;
>> +}
>> +IOMMU_OF_DECLARE(arm_smmu_v3, "arm,smmu-v3", arm_smmu_of_iommu_init);
>> +
>>   MODULE_DESCRIPTION("IOMMU API for ARM architected SMMUv3 implementations");
>>   MODULE_AUTHOR("Will Deacon <will.deacon@arm.com>");
>>   MODULE_LICENSE("GPL v2");
>> -- 
>> 1.8.0
>>
>>
>> _______________________________________________
>> iommu mailing list
>> iommu at lists.linux-foundation.org
>> https://lists.linuxfoundation.org/mailman/listinfo/iommu
>>
> 
> 
> .
>