Date:   Sat, 6 Aug 2022 15:21:51 +0800
From:   Jianmin Lv <lvjianmin@...ngson.cn>
To:     Robin Murphy <robin.murphy@....com>, lpieralisi@...nel.org,
        guohanjun@...wei.com, sudeep.holla@....com, rafael@...nel.org,
        lenb@...nel.org, robert.moore@...el.com
Cc:     linux-kernel@...r.kernel.org, linux-acpi@...r.kernel.org
Subject: Re: [PATCH V2 1/2] ACPI / scan: Support multiple dma windows with
 different offsets



On 2022/8/6 下午2:13, Jianmin Lv wrote:
> 
> 
> On 2022/8/5 下午8:46, Robin Murphy wrote:
>> On 2022-08-05 08:31, Jianmin Lv wrote:
>>> For DT, of_dma_get_range returns bus_dma_region typed dma regions,
>>> which makes multiple dma windows with different offset available
>>> for translation between dma address and cpu address.
>>>
>>> But for ACPI, acpi_dma_get_range doesn't return similar dma regions,
>>> causing no path for setting dev->dma_range_map conveniently. So the
>>> patch changes acpi_dma_get_range and returns bus_dma_region typed
>>> dma regions according to of_dma_get_range.
>>>
>>> After changing acpi_dma_get_range, acpi_arch_dma_setup is changed for
>>> ARM64, where original dma_addr and size are removed, and pass 0 and
>>> U64_MAX for dma_base and size of arch_setup_dma_ops.
>>
>> Hmm, I realise the reasoning behind this exists largely in my head, so 
>> for the benefit of everyone else we should perhaps clarify that these 
>> arguments are now redundant, so this is a simplification consistent 
>> with what other ACPI architectures also pass to iommu_setup_dma_ops().
>>
> 
> Ok, I'll add reason of changing it into commit log for everyone else.
> 
> 
>>> Signed-off-by: Jianmin Lv <lvjianmin@...ngson.cn>
>>> ---
>>>   drivers/acpi/arm64/dma.c | 44 
>>> +++++++++++++++++++++++++++++++++-----------
>>>   drivers/acpi/scan.c      | 48 
>>> ++++++++++++++++++++++++------------------------
>>>   include/acpi/acpi_bus.h  |  3 +--
>>>   include/linux/acpi.h     |  7 +++----
>>>   4 files changed, 61 insertions(+), 41 deletions(-)
>>>
>>> diff --git a/drivers/acpi/arm64/dma.c b/drivers/acpi/arm64/dma.c
>>> index f16739a..64a41b2 100644
>>> --- a/drivers/acpi/arm64/dma.c
>>> +++ b/drivers/acpi/arm64/dma.c
>>> @@ -4,11 +4,12 @@
>>>   #include <linux/device.h>
>>>   #include <linux/dma-direct.h>
>>> -void acpi_arch_dma_setup(struct device *dev, u64 *dma_addr, u64 
>>> *dma_size)
>>> +void acpi_arch_dma_setup(struct device *dev)
>>>   {
>>>       int ret;
>>>       u64 end, mask;
>>> -    u64 dmaaddr = 0, size = 0, offset = 0;
>>> +    u64 size = 0;
>>> +    const struct bus_dma_region *map = NULL;
>>>       /*
>>>        * If @dev is expected to be DMA-capable then the bus code that 
>>> created
>>> @@ -26,7 +27,33 @@ void acpi_arch_dma_setup(struct device *dev, u64 
>>> *dma_addr, u64 *dma_size)
>>>       else
>>>           size = 1ULL << 32;
>>> -    ret = acpi_dma_get_range(dev, &dmaaddr, &offset, &size);
>>> +    ret = acpi_dma_get_range(dev, &map);
>>> +    if (!ret) {
>>> +        const struct bus_dma_region *r = map;
>>> +        u64 dma_start, dma_end = 0;
>>> +
>>> +        /* determine the overall bounds of all dma regions */
>>> +        for (dma_start = U64_MAX; r->size; r++) {
>>> +
>>> +            /* Take lower and upper limits */
>>> +            if (r->dma_start < dma_start)
>>> +                dma_start = r->dma_start;
>>
>> dma_start now only serves as an overcomplicated way to tell if the 
>> list was empty, so we don't need it...
> 
> Ok,thanks.
> 
>>
>>> +            if (r->dma_start + r->size - 1 > dma_end)
>>> +                dma_end = r->dma_start + r->size - 1;
>>> +        }
>>> +
>>> +        if (dma_start >= dma_end) {
>>> +            dev_dbg(dev, "Invalid DMA regions configuration\n");
>>> +            return;
>>
>> ...but we also don't need this check anway - either there was at least 
>> one valid resource, so dma_end is now set to something sensible, or 
>> there were none, in which case the implementation of 
>> acpi_dma_get_range() here returns success while leaving map == NULL, 
>> so we've already crashed dereferencing r->size.
>>
> 
> Ok, thanks.
> 
> 
>>> +        }
>>> +
>>> +        mask = DMA_BIT_MASK(ilog2(dma_end) + 1);
>>> +        dev->bus_dma_limit = dma_end;
>>> +        dev->dma_range_map = map;
>>> +        dev->coherent_dma_mask = min(dev->coherent_dma_mask, mask);
>>> +        *dev->dma_mask = min(*dev->dma_mask, mask);
>>
>> Don't duplicate all this, just set "size = dma_end + 1" so the 
>> existing code below picks it up.
>>
> 
> TBH, I'm not familiar with ARM64's DMA stuff, but from the code here, I 
> think there are two ways to get dma ranges: acpi_dma_get_range or 
> iort_dma_get_ranges.
> 
> Before the patch, after obtaining the dmaaddr, offset and size in any 
> way of the both, the bus_dma_limit, mask and dma_range_map will be 
> initialized, and dma_range_map is created a *new* one for *both* ways.
> 
> But after changing acpi_dma_get_range, when we are ok to obtain 
> dma_range_map by acpi_dma_get_range, we'll use it instead of creating a
> new one like before. For iort_dma_get_ranges, we have to still create a 
> new dma_range_map. So, I have to handle separately each way here.
> 
> It seems that it's enough to handle dma_range_map separately, and others 
> can be still initialized in one place.
> 

Maybe I got what you mean, I looked into dma_direct_set_offset, which is 
only meaningful when acpi_dma_get_range return 0 with correct offset and 
dmaaddr. iort_dma_get_ranges is only used to retrieve size. Right? If 
so, code here should be:

dev->dma_range_map = map;
size = dma_end + 1;

>>> +    }
>>> +
>>>       if (ret == -ENODEV)
>>>           ret = iort_dma_get_ranges(dev, &size);
>>>       if (!ret) {
>>> @@ -34,17 +61,12 @@ void acpi_arch_dma_setup(struct device *dev, u64 
>>> *dma_addr, u64 *dma_size)
>>>            * Limit coherent and dma mask based on size retrieved from
>>>            * firmware.
>>>            */
>>> -        end = dmaaddr + size - 1;
>>> +        end = size - 1;
>>>           mask = DMA_BIT_MASK(ilog2(end) + 1);
>>>           dev->bus_dma_limit = end;
>>>           dev->coherent_dma_mask = min(dev->coherent_dma_mask, mask);
>>>           *dev->dma_mask = min(*dev->dma_mask, mask);
>>> -    }
>>> -    *dma_addr = dmaaddr;
>>> -    *dma_size = size;
>>> -
>>> -    ret = dma_direct_set_offset(dev, dmaaddr + offset, dmaaddr, size);
>>> -
>>> -    dev_dbg(dev, "dma_offset(%#08llx)%s\n", offset, ret ? " failed!" 
>>> : "");
>>> +        ret = dma_direct_set_offset(dev, 0, 0, size);
>>
>> Don't call this - we're setting dev->dma_range_map now, so it will 
>> only scream and fail (or do nothing if there's no offset).
>>
> 
> Same as above.
> 

Same, If what I understand is right above, dma_direct_set_offset is 
should be removed as you suggested.

>>> +    }
>>>   }
>>> diff --git a/drivers/acpi/scan.c b/drivers/acpi/scan.c
>>> index 762b61f..736517e 100644
>>> --- a/drivers/acpi/scan.c
>>> +++ b/drivers/acpi/scan.c
>>> @@ -20,6 +20,7 @@
>>>   #include <linux/platform_data/x86/apple.h>
>>>   #include <linux/pgtable.h>
>>>   #include <linux/crc32.h>
>>> +#include <linux/dma-direct.h>
>>>   #include "internal.h"
>>> @@ -1492,15 +1493,15 @@ enum dev_dma_attr acpi_get_dma_attr(struct 
>>> acpi_device *adev)
>>>    *
>>>    * Return 0 on success, < 0 on failure.
>>>    */
>>
>> Don't forget to update the kerneldoc as well.
>>
> 
> Ok, thanks.
> 
> 
>>> -int acpi_dma_get_range(struct device *dev, u64 *dma_addr, u64 *offset,
>>> -               u64 *size)
>>> +int acpi_dma_get_range(struct device *dev, const struct 
>>> bus_dma_region **map)
>>>   {
>>>       struct acpi_device *adev;
>>>       LIST_HEAD(list);
>>>       struct resource_entry *rentry;
>>>       int ret;
>>>       struct device *dma_dev = dev;
>>> -    u64 len, dma_start = U64_MAX, dma_end = 0, dma_offset = 0;
>>> +    int num_ranges = 0;
>>> +    struct bus_dma_region *r;
>>>       /*
>>>        * Walk the device tree chasing an ACPI companion with a _DMA
>>> @@ -1525,31 +1526,31 @@ int acpi_dma_get_range(struct device *dev, 
>>> u64 *dma_addr, u64 *offset,
>>>       ret = acpi_dev_get_dma_resources(adev, &list);
>>>       if (ret > 0) {
>>> +        list_for_each_entry(rentry, &list, node)
>>> +            num_ranges++;
>>
>> We already have the number of resources in ret.
>>
> 
> Ok, thanks.
> 
> 
>> Looking at this, I also now wonder if we're doing the right thing if 
>> the object is present but contains no resources. The spec isn't clear 
>> whether that's even really valid, but if it is, is it meaningful? It 
>> seems we'd currently consider an empty object equivalent to no object, 
>> but if anything it should perhaps be interpreted as the opposite, i.e. 
>> that no DMA is possible because the bus does not decode any ranges. Is 
>> anyone more familiar with the intent of the spec here?
>>
>> Thanks,
>> Robin.
>>
> 
> According to _DMA of ACPI spec, _DMA is optional, if without it(no _DMA 
> object configured in DSDT), kernel assumes that any address range placed 
> on bus by a child device is valid and can be decoded. So I think it is 
> still legal if _DMA is present and no resources contained in it. I mean 
> an empty _DMA object equals no _DMA object. But obviously it's not 
> meaningful for an empty _DMA object, anybody should not pass an empty 
> _DMA object.
> 
>>> +
>>> +        r = kcalloc(num_ranges + 1, sizeof(*r), GFP_KERNEL);
>>> +        if (!r) {
>>> +            ret = -ENOMEM;
>>> +            goto out;
>>> +        }
>>> +
>>> +        *map = r;
>>> +
>>>           list_for_each_entry(rentry, &list, node) {
>>> -            if (dma_offset && rentry->offset != dma_offset) {
>>> +            if (rentry->res->start >= rentry->res->end) {
>>>                   ret = -EINVAL;
>>> -                dev_warn(dma_dev, "Can't handle multiple windows 
>>> with different offsets\n");
>>> +                dev_dbg(dma_dev, "Invalid DMA regions 
>>> configuration\n");
>>>                   goto out;
>>>               }
>>> -            dma_offset = rentry->offset;
>>> -
>>> -            /* Take lower and upper limits */
>>> -            if (rentry->res->start < dma_start)
>>> -                dma_start = rentry->res->start;
>>> -            if (rentry->res->end > dma_end)
>>> -                dma_end = rentry->res->end;
>>> -        }
>>> -        if (dma_start >= dma_end) {
>>> -            ret = -EINVAL;
>>> -            dev_dbg(dma_dev, "Invalid DMA regions configuration\n");
>>> -            goto out;
>>> +            r->cpu_start = rentry->res->start;
>>> +            r->dma_start = rentry->res->start - rentry->offset;
>>> +            r->size = rentry->res->end - rentry->res->start + 1;
>>> +            r->offset = rentry->offset;
>>> +            r++;
>>>           }
>>> -        *dma_addr = dma_start - dma_offset;
>>> -        len = dma_end - dma_start;
>>> -        *size = max(len, len + 1);
>>> -        *offset = dma_offset;
>>>       }
>>>    out:
>>>       acpi_dev_free_resource_list(&list);
>>> @@ -1639,20 +1640,19 @@ int acpi_dma_configure_id(struct device *dev, 
>>> enum dev_dma_attr attr,
>>>                 const u32 *input_id)
>>>   {
>>>       const struct iommu_ops *iommu;
>>> -    u64 dma_addr = 0, size = 0;
>>>       if (attr == DEV_DMA_NOT_SUPPORTED) {
>>>           set_dma_ops(dev, &dma_dummy_ops);
>>>           return 0;
>>>       }
>>> -    acpi_arch_dma_setup(dev, &dma_addr, &size);
>>> +    acpi_arch_dma_setup(dev);
>>>       iommu = acpi_iommu_configure_id(dev, input_id);
>>>       if (PTR_ERR(iommu) == -EPROBE_DEFER)
>>>           return -EPROBE_DEFER;
>>> -    arch_setup_dma_ops(dev, dma_addr, size,
>>> +    arch_setup_dma_ops(dev, 0, U64_MAX,
>>>                   iommu, attr == DEV_DMA_COHERENT);
>>>       return 0;
>>> diff --git a/include/acpi/acpi_bus.h b/include/acpi/acpi_bus.h
>>> index 0dc1ea0b..e106073 100644
>>> --- a/include/acpi/acpi_bus.h
>>> +++ b/include/acpi/acpi_bus.h
>>> @@ -611,8 +611,7 @@ struct acpi_pci_root {
>>>   int acpi_iommu_fwspec_init(struct device *dev, u32 id,
>>>                  struct fwnode_handle *fwnode,
>>>                  const struct iommu_ops *ops);
>>> -int acpi_dma_get_range(struct device *dev, u64 *dma_addr, u64 *offset,
>>> -               u64 *size);
>>> +int acpi_dma_get_range(struct device *dev, const struct 
>>> bus_dma_region **map);
>>>   int acpi_dma_configure_id(struct device *dev, enum dev_dma_attr attr,
>>>                  const u32 *input_id);
>>>   static inline int acpi_dma_configure(struct device *dev,
>>> diff --git a/include/linux/acpi.h b/include/linux/acpi.h
>>> index 44975c1..34e0545 100644
>>> --- a/include/linux/acpi.h
>>> +++ b/include/linux/acpi.h
>>> @@ -280,12 +280,12 @@ int acpi_table_parse_madt(enum acpi_madt_type id,
>>>   #ifdef CONFIG_ARM64
>>>   void acpi_numa_gicc_affinity_init(struct acpi_srat_gicc_affinity *pa);
>>> -void acpi_arch_dma_setup(struct device *dev, u64 *dma_addr, u64 
>>> *dma_size);
>>> +void acpi_arch_dma_setup(struct device *dev);
>>>   #else
>>>   static inline void
>>>   acpi_numa_gicc_affinity_init(struct acpi_srat_gicc_affinity *pa) { }
>>>   static inline void
>>> -acpi_arch_dma_setup(struct device *dev, u64 *dma_addr, u64 
>>> *dma_size) { }
>>> +acpi_arch_dma_setup(struct device *dev) { }
>>>   #endif
>>>   int acpi_numa_memory_affinity_init (struct acpi_srat_mem_affinity 
>>> *ma);
>>> @@ -974,8 +974,7 @@ static inline enum dev_dma_attr 
>>> acpi_get_dma_attr(struct acpi_device *adev)
>>>       return DEV_DMA_NOT_SUPPORTED;
>>>   }
>>> -static inline int acpi_dma_get_range(struct device *dev, u64 *dma_addr,
>>> -                     u64 *offset, u64 *size)
>>> +static inline int acpi_dma_get_range(struct device *dev, const 
>>> struct bus_dma_region **map)
>>>   {
>>>       return -ENODEV;
>>>   }

Powered by blists - more mailing lists