Re: [PATCH v2 1/2] drivers: dma-coherent: Fix dev->cma_area vs dev->dma_mem breakage

From: Robin Murphy
Date: Fri Jul 07 2017 - 12:10:32 EST


On 07/07/17 16:40, Vladimir Murzin wrote:
> Christoph,
>
> On 07/07/17 15:27, Christoph Hellwig wrote:
>> Vladimir,
>>
>> this is why I really didn't like overloading the current
>> dma coherent infrastructure with the global pool.
>>
>> And this new patch seems like piling hacks over hacks. I think we
>> should go back and make sure allocations from the global coherent
>> pool are done by the dma ops implementation, and not before calling
>> into them - preferably still reusing the common code for it.
>>
>> Vladimir or Vitaly - can you look into that?
>>
>
> It is really sad that Vitaly and George did not join to discussions earlier,
> so we could avoid being in situation like this.
>
> Likely I'm missing something, but what should happen if device relies on
> dma_contiguous_default_area?
>
> Originally, intention behind dma-default was to simplify things, so instead of
>
> reserved-memory {
> #address-cells = <1>;
> #size-cells = <1>;
> ranges;
>
> coherent_dma: linux,dma {
> compatible = "shared-dma-pool";
> no-map;
> reg = <0x78000000 0x800000>;
> };
> };
>
>
> dev0: dev@12300000 {
> memory-region = <&coherent_dma>;
> /* ... */
> };
>
> dev1: dev@12500000 {
> memory-region = <&coherent_dma>;
> /* ... */
> };
>
> dev2: dev@12600000 {
> memory-region = <&coherent_dma>;
> /* ... */
> };
>
> in device tree we could simply have
>
> reserved-memory {
> #address-cells = <1>;
> #size-cells = <1>;
> ranges;
>
> coherent_dma: linux,dma {
> compatible = "shared-dma-pool";
> no-map;
> reg = <0x78000000 0x800000>;
> linux,dma-default;
> };
> };
>
> and that just work in my (NOMMU) case because there is no CMA there...
>
> However, given that dma-default is being overloaded and there are no device
> tree users merged yet, I would not object stepping back, reverting "drivers:
> dma-coherent: Introduce default DMA pool" and cooperatively rethinking
> design/implementation, so every party gets happy.

I don't think we need to go that far, I reckon it would be clear enough
to just split the per-device vs. global pool interfaces, something like
I've sketched out below (such that the ops->alloc implementation calls
dma_alloc_from_global_coherent() if dma_alloc_from_contiguous() fails).

If anyone wants to take that and run with it, feel free.

Robin.

----->8-----
diff --git a/drivers/base/dma-coherent.c b/drivers/base/dma-coherent.c
index 640a7e63c453..e6393c6d8359 100644
--- a/drivers/base/dma-coherent.c
+++ b/drivers/base/dma-coherent.c
@@ -143,6 +143,44 @@ void *dma_mark_declared_memory_occupied(struct
device *dev,
}
EXPORT_SYMBOL(dma_mark_declared_memory_occupied);

+static void *__dma_alloc_from_coherent(struct dma_coherent_mem *mem,
ssize_t size,
+ dma_addr_t *dma_handle)
+{
+ int order = get_order(size);
+ unsigned long flags;
+ int pageno;
+ int dma_memory_map;
+ void *ret;
+
+ spin_lock_irqsave(&mem->spinlock, flags);
+
+ if (unlikely(size > (mem->size << PAGE_SHIFT)))
+ goto err;
+
+ pageno = bitmap_find_free_region(mem->bitmap, mem->size, order);
+ if (unlikely(pageno < 0))
+ goto err;
+
+ /*
+ * Memory was found in the coherent area.
+ */
+ *dma_handle = mem->device_base + (pageno << PAGE_SHIFT);
+ ret = mem->virt_base + (pageno << PAGE_SHIFT);
+ dma_memory_map = (mem->flags & DMA_MEMORY_MAP);
+ spin_unlock_irqrestore(&mem->spinlock, flags);
+ if (dma_memory_map)
+ memset(ret, 0, size);
+ else
+ memset_io(ret, 0, size);
+
+ return ret;
+
+err:
+ spin_unlock_irqrestore(&mem->spinlock, flags);
+ return NULL;
+}
+EXPORT_SYMBOL(dma_alloc_from_coherent);
+
/**
* dma_alloc_from_coherent() - try to allocate memory from the
per-device coherent area
*
@@ -162,10 +200,6 @@ int dma_alloc_from_coherent(struct device *dev,
ssize_t size,
dma_addr_t *dma_handle, void **ret)
{
struct dma_coherent_mem *mem;
- int order = get_order(size);
- unsigned long flags;
- int pageno;
- int dma_memory_map;

if (!dev)
return 0;
@@ -173,32 +207,10 @@ int dma_alloc_from_coherent(struct device *dev,
ssize_t size,
if (!mem)
return 0;

- *ret = NULL;
- spin_lock_irqsave(&mem->spinlock, flags);
+ *ret = __dma_alloc_from_coherent(mem, size, dma_handle);
+ if (*ret)
+ return 1;

- if (unlikely(size > (mem->size << PAGE_SHIFT)))
- goto err;
-
- pageno = bitmap_find_free_region(mem->bitmap, mem->size, order);
- if (unlikely(pageno < 0))
- goto err;
-
- /*
- * Memory was found in the per-device area.
- */
- *dma_handle = mem->device_base + (pageno << PAGE_SHIFT);
- *ret = mem->virt_base + (pageno << PAGE_SHIFT);
- dma_memory_map = (mem->flags & DMA_MEMORY_MAP);
- spin_unlock_irqrestore(&mem->spinlock, flags);
- if (dma_memory_map)
- memset(*ret, 0, size);
- else
- memset_io(*ret, 0, size);
-
- return 1;
-
-err:
- spin_unlock_irqrestore(&mem->spinlock, flags);
/*
* In the case where the allocation can not be satisfied from the
* per-device area, try to fall back to generic memory if the
@@ -208,6 +220,15 @@ int dma_alloc_from_coherent(struct device *dev,
ssize_t size,
}
EXPORT_SYMBOL(dma_alloc_from_coherent);

+void *dma_alloc_from_global_coherent(ssize_t size, dma_addr_t *dma_handle)
+{
+ if (!dma_coherent_default_memory)
+ return NULL;
+
+ return __dma_alloc_from_coherent(dma_coherent_default_memory, size,
+ handle);
+}
+
/**
* dma_release_from_coherent() - try to free the memory allocated from
per-device coherent memory pool
* @dev: device from which the memory was allocated