[PATCH] dma-coherent: add support for multi coherent rmems per dev

[PATCH] dma-coherent: add support for multi coherent rmems per dev

[Howard Yen]

Add support for multiple coherent rmems per device. This patch addes
dma_mem_list to device structure to store multiple rmems.

These multiple rmems can be assigned to the device one by one by
of_reserved_mem_device_init_by_idx() with the memory-region
declaration in device tree as below and store the rmem to the dma_mem_list.

	device1@0 {
		...
		memory-region = <&reserved_mem0>, <&reserved_mem1>;
		...
	};

When driver tries to allocate memory from the rmems, looks for the first
available rmem and allocates the memory from this rmem.

Then if driver removed, of_reserved_mem_device_release() needs to be
invoked to release all the rmems assigned to the device.

Signed-off-by: Howard Yen <howardyen@google.com>
---
 include/linux/device.h |  1 +
 kernel/dma/coherent.c  | 66 +++++++++++++++++++++++++++++++++++-------
 2 files changed, 56 insertions(+), 11 deletions(-)

diff --git a/include/linux/device.h b/include/linux/device.h
index 97c4b046c09d..c8682ee507cf 100644
--- a/include/linux/device.h
+++ b/include/linux/device.h
@@ -751,6 +751,7 @@ struct device {
 #ifdef CONFIG_DMA_DECLARE_COHERENT
 	struct dma_coherent_mem	*dma_mem; /* internal for coherent mem
 					     override */
+	struct list_head	dma_mem_list;
 #endif
 #ifdef CONFIG_DMA_CMA
 	struct cma *cma_area;		/* contiguous memory area for dma
diff --git a/kernel/dma/coherent.c b/kernel/dma/coherent.c
index ff5683a57f77..f31befd2e6f8 100644
--- a/kernel/dma/coherent.c
+++ b/kernel/dma/coherent.c
@@ -18,6 +18,7 @@ struct dma_coherent_mem {
 	unsigned long	*bitmap;
 	spinlock_t	spinlock;
 	bool		use_dev_dma_pfn_offset;
+	struct list_head	node;
 };
 
 static inline struct dma_coherent_mem *dev_get_coherent_memory(struct device *dev)
@@ -61,6 +62,7 @@ static struct dma_coherent_mem *dma_init_coherent_memory(phys_addr_t phys_addr,
 	dma_mem->pfn_base = PFN_DOWN(phys_addr);
 	dma_mem->size = pages;
 	dma_mem->use_dev_dma_pfn_offset = use_dma_pfn_offset;
+	INIT_LIST_HEAD(&dma_mem->node);
 	spin_lock_init(&dma_mem->spinlock);
 
 	return dma_mem;
@@ -90,10 +92,13 @@ static int dma_assign_coherent_memory(struct device *dev,
 	if (!dev)
 		return -ENODEV;
 
-	if (dev->dma_mem)
-		return -EBUSY;
+	if (!dev->dma_mem) {
+		dev->dma_mem = mem;
+		INIT_LIST_HEAD(&dev->dma_mem_list);
+	}
+
+	list_add_tail(&mem->node, &dev->dma_mem_list);
 
-	dev->dma_mem = mem;
 	return 0;
 }
 
@@ -187,12 +192,17 @@ static void *__dma_alloc_from_coherent(struct device *dev,
 int dma_alloc_from_dev_coherent(struct device *dev, ssize_t size,
 		dma_addr_t *dma_handle, void **ret)
 {
-	struct dma_coherent_mem *mem = dev_get_coherent_memory(dev);
+	struct dma_coherent_mem *mem_tmp, *mem = dev_get_coherent_memory(dev);
 
 	if (!mem)
 		return 0;
 
-	*ret = __dma_alloc_from_coherent(dev, mem, size, dma_handle);
+	list_for_each_entry(mem_tmp, &dev->dma_mem_list, node) {
+		*ret = __dma_alloc_from_coherent(dev, mem_tmp, size, dma_handle);
+		if (*ret)
+			break;
+	}
+
 	return 1;
 }
 
@@ -226,9 +236,19 @@ static int __dma_release_from_coherent(struct dma_coherent_mem *mem,
  */
 int dma_release_from_dev_coherent(struct device *dev, int order, void *vaddr)
 {
-	struct dma_coherent_mem *mem = dev_get_coherent_memory(dev);
+	struct dma_coherent_mem *mem_tmp, *mem = dev_get_coherent_memory(dev);
+	int ret = 0;
 
-	return __dma_release_from_coherent(mem, order, vaddr);
+	if (!mem)
+		return ret;
+
+	list_for_each_entry(mem_tmp, &dev->dma_mem_list, node) {
+		ret = __dma_release_from_coherent(mem_tmp, order, vaddr);
+		if (ret == 1)
+			break;
+	}
+
+	return ret;
 }
 
 static int __dma_mmap_from_coherent(struct dma_coherent_mem *mem,
@@ -271,9 +291,19 @@ static int __dma_mmap_from_coherent(struct dma_coherent_mem *mem,
 int dma_mmap_from_dev_coherent(struct device *dev, struct vm_area_struct *vma,
 			   void *vaddr, size_t size, int *ret)
 {
-	struct dma_coherent_mem *mem = dev_get_coherent_memory(dev);
+	struct dma_coherent_mem *mem_tmp, *mem = dev_get_coherent_memory(dev);
+	int retval = 0;
 
-	return __dma_mmap_from_coherent(mem, vma, vaddr, size, ret);
+	if (!mem)
+		return retval;
+
+	list_for_each_entry(mem_tmp, &dev->dma_mem_list, node) {
+		retval = __dma_mmap_from_coherent(mem_tmp, vma, vaddr, size, ret);
+		if (retval == 1)
+			break;
+	}
+
+	return retval;
 }
 
 #ifdef CONFIG_DMA_GLOBAL_POOL
@@ -351,8 +381,22 @@ static int rmem_dma_device_init(struct reserved_mem *rmem, struct device *dev)
 static void rmem_dma_device_release(struct reserved_mem *rmem,
 				    struct device *dev)
 {
-	if (dev)
-		dev->dma_mem = NULL;
+	struct dma_coherent_mem *mem, *mem_tmp, *q;
+
+	if (dev) {
+		mem = dev_get_coherent_memory(dev);
+		if (rmem->priv != mem) {
+			list_for_each_entry_safe(mem_tmp, q, &dev->dma_mem_list, node) {
+				if (mem_tmp == rmem->priv) {
+					list_del_init(&mem_tmp->node);
+					break;
+				}
+			}
+		} else {
+			list_del_init(&mem->node);
+			dev->dma_mem = NULL;
+		}
+	}
 }
 
 static const struct reserved_mem_ops rmem_dma_ops = {
-- 
2.43.0.429.g432eaa2c6b-goog

Re: [PATCH] dma-coherent: add support for multi coherent rmems per dev

[Robin Murphy]

On 2024-02-01 9:35 am, Howard Yen wrote:
> Add support for multiple coherent rmems per device. This patch addes
> dma_mem_list to device structure to store multiple rmems.
> 
> These multiple rmems can be assigned to the device one by one by
> of_reserved_mem_device_init_by_idx() with the memory-region
> declaration in device tree as below and store the rmem to the dma_mem_list.
> 
> 	device1@0 {
> 		...
> 		memory-region = <&reserved_mem0>, <&reserved_mem1>;
> 		...
> 	};
> 
> When driver tries to allocate memory from the rmems, looks for the first
> available rmem and allocates the memory from this rmem.
> 
> Then if driver removed, of_reserved_mem_device_release() needs to be
> invoked to release all the rmems assigned to the device.
> 
> Signed-off-by: Howard Yen <howardyen@google.com>
> ---
>   include/linux/device.h |  1 +
>   kernel/dma/coherent.c  | 66 +++++++++++++++++++++++++++++++++++-------
>   2 files changed, 56 insertions(+), 11 deletions(-)
> 
> diff --git a/include/linux/device.h b/include/linux/device.h
> index 97c4b046c09d..c8682ee507cf 100644
> --- a/include/linux/device.h
> +++ b/include/linux/device.h
> @@ -751,6 +751,7 @@ struct device {
>   #ifdef CONFIG_DMA_DECLARE_COHERENT
>   	struct dma_coherent_mem	*dma_mem; /* internal for coherent mem
>   					     override */
> +	struct list_head	dma_mem_list;

I'm not necessarily against the idea, but only if it's implemented 
sensibly. If we're going to have a list of these it should *replace* the 
existing pointer, not do this weird thing with both.

Thanks,
Robin.

>   #endif
>   #ifdef CONFIG_DMA_CMA
>   	struct cma *cma_area;		/* contiguous memory area for dma
> diff --git a/kernel/dma/coherent.c b/kernel/dma/coherent.c
> index ff5683a57f77..f31befd2e6f8 100644
> --- a/kernel/dma/coherent.c
> +++ b/kernel/dma/coherent.c
> @@ -18,6 +18,7 @@ struct dma_coherent_mem {
>   	unsigned long	*bitmap;
>   	spinlock_t	spinlock;
>   	bool		use_dev_dma_pfn_offset;
> +	struct list_head	node;
>   };
>   
>   static inline struct dma_coherent_mem *dev_get_coherent_memory(struct device *dev)
> @@ -61,6 +62,7 @@ static struct dma_coherent_mem *dma_init_coherent_memory(phys_addr_t phys_addr,
>   	dma_mem->pfn_base = PFN_DOWN(phys_addr);
>   	dma_mem->size = pages;
>   	dma_mem->use_dev_dma_pfn_offset = use_dma_pfn_offset;
> +	INIT_LIST_HEAD(&dma_mem->node);
>   	spin_lock_init(&dma_mem->spinlock);
>   
>   	return dma_mem;
> @@ -90,10 +92,13 @@ static int dma_assign_coherent_memory(struct device *dev,
>   	if (!dev)
>   		return -ENODEV;
>   
> -	if (dev->dma_mem)
> -		return -EBUSY;
> +	if (!dev->dma_mem) {
> +		dev->dma_mem = mem;
> +		INIT_LIST_HEAD(&dev->dma_mem_list);
> +	}
> +
> +	list_add_tail(&mem->node, &dev->dma_mem_list);
>   
> -	dev->dma_mem = mem;
>   	return 0;
>   }
>   
> @@ -187,12 +192,17 @@ static void *__dma_alloc_from_coherent(struct device *dev,
>   int dma_alloc_from_dev_coherent(struct device *dev, ssize_t size,
>   		dma_addr_t *dma_handle, void **ret)
>   {
> -	struct dma_coherent_mem *mem = dev_get_coherent_memory(dev);
> +	struct dma_coherent_mem *mem_tmp, *mem = dev_get_coherent_memory(dev);
>   
>   	if (!mem)
>   		return 0;
>   
> -	*ret = __dma_alloc_from_coherent(dev, mem, size, dma_handle);
> +	list_for_each_entry(mem_tmp, &dev->dma_mem_list, node) {
> +		*ret = __dma_alloc_from_coherent(dev, mem_tmp, size, dma_handle);
> +		if (*ret)
> +			break;
> +	}
> +
>   	return 1;
>   }
>   
> @@ -226,9 +236,19 @@ static int __dma_release_from_coherent(struct dma_coherent_mem *mem,
>    */
>   int dma_release_from_dev_coherent(struct device *dev, int order, void *vaddr)
>   {
> -	struct dma_coherent_mem *mem = dev_get_coherent_memory(dev);
> +	struct dma_coherent_mem *mem_tmp, *mem = dev_get_coherent_memory(dev);
> +	int ret = 0;
>   
> -	return __dma_release_from_coherent(mem, order, vaddr);
> +	if (!mem)
> +		return ret;
> +
> +	list_for_each_entry(mem_tmp, &dev->dma_mem_list, node) {
> +		ret = __dma_release_from_coherent(mem_tmp, order, vaddr);
> +		if (ret == 1)
> +			break;
> +	}
> +
> +	return ret;
>   }
>   
>   static int __dma_mmap_from_coherent(struct dma_coherent_mem *mem,
> @@ -271,9 +291,19 @@ static int __dma_mmap_from_coherent(struct dma_coherent_mem *mem,
>   int dma_mmap_from_dev_coherent(struct device *dev, struct vm_area_struct *vma,
>   			   void *vaddr, size_t size, int *ret)
>   {
> -	struct dma_coherent_mem *mem = dev_get_coherent_memory(dev);
> +	struct dma_coherent_mem *mem_tmp, *mem = dev_get_coherent_memory(dev);
> +	int retval = 0;
>   
> -	return __dma_mmap_from_coherent(mem, vma, vaddr, size, ret);
> +	if (!mem)
> +		return retval;
> +
> +	list_for_each_entry(mem_tmp, &dev->dma_mem_list, node) {
> +		retval = __dma_mmap_from_coherent(mem_tmp, vma, vaddr, size, ret);
> +		if (retval == 1)
> +			break;
> +	}
> +
> +	return retval;
>   }
>   
>   #ifdef CONFIG_DMA_GLOBAL_POOL
> @@ -351,8 +381,22 @@ static int rmem_dma_device_init(struct reserved_mem *rmem, struct device *dev)
>   static void rmem_dma_device_release(struct reserved_mem *rmem,
>   				    struct device *dev)
>   {
> -	if (dev)
> -		dev->dma_mem = NULL;
> +	struct dma_coherent_mem *mem, *mem_tmp, *q;
> +
> +	if (dev) {
> +		mem = dev_get_coherent_memory(dev);
> +		if (rmem->priv != mem) {
> +			list_for_each_entry_safe(mem_tmp, q, &dev->dma_mem_list, node) {
> +				if (mem_tmp == rmem->priv) {
> +					list_del_init(&mem_tmp->node);
> +					break;
> +				}
> +			}
> +		} else {
> +			list_del_init(&mem->node);
> +			dev->dma_mem = NULL;
> +		}
> +	}
>   }
>   
>   static const struct reserved_mem_ops rmem_dma_ops = {

Re: [PATCH] dma-coherent: add support for multi coherent rmems per dev

[Andy Shevchenko]

On Thu, Feb 01, 2024 at 10:45:30AM +0000, Robin Murphy wrote:
> On 2024-02-01 9:35 am, Howard Yen wrote:

...

> >   	struct dma_coherent_mem	*dma_mem; /* internal for coherent mem
> >   					     override */
> > +	struct list_head	dma_mem_list;
> 
> I'm not necessarily against the idea, but only if it's implemented sensibly.
> If we're going to have a list of these it should *replace* the existing
> pointer, not do this weird thing with both.

+1 on the comment.

-- 
With Best Regards,
Andy Shevchenko

Re: [PATCH] dma-coherent: add support for multi coherent rmems per dev

[Greg KH]

On Thu, Feb 01, 2024 at 10:45:30AM +0000, Robin Murphy wrote:
> On 2024-02-01 9:35 am, Howard Yen wrote:
> > Add support for multiple coherent rmems per device. This patch addes
> > dma_mem_list to device structure to store multiple rmems.
> > 
> > These multiple rmems can be assigned to the device one by one by
> > of_reserved_mem_device_init_by_idx() with the memory-region
> > declaration in device tree as below and store the rmem to the dma_mem_list.
> > 
> > 	device1@0 {
> > 		...
> > 		memory-region = <&reserved_mem0>, <&reserved_mem1>;
> > 		...
> > 	};
> > 
> > When driver tries to allocate memory from the rmems, looks for the first
> > available rmem and allocates the memory from this rmem.
> > 
> > Then if driver removed, of_reserved_mem_device_release() needs to be
> > invoked to release all the rmems assigned to the device.
> > 
> > Signed-off-by: Howard Yen <howardyen@google.com>
> > ---
> >   include/linux/device.h |  1 +
> >   kernel/dma/coherent.c  | 66 +++++++++++++++++++++++++++++++++++-------
> >   2 files changed, 56 insertions(+), 11 deletions(-)
> > 
> > diff --git a/include/linux/device.h b/include/linux/device.h
> > index 97c4b046c09d..c8682ee507cf 100644
> > --- a/include/linux/device.h
> > +++ b/include/linux/device.h
> > @@ -751,6 +751,7 @@ struct device {
> >   #ifdef CONFIG_DMA_DECLARE_COHERENT
> >   	struct dma_coherent_mem	*dma_mem; /* internal for coherent mem
> >   					     override */
> > +	struct list_head	dma_mem_list;
> 
> I'm not necessarily against the idea, but only if it's implemented sensibly.
> If we're going to have a list of these it should *replace* the existing
> pointer, not do this weird thing with both.

Agreed, it should be one pointer max for this structure for this type of
thing.  Why not move it into the dma_coherent_mem structure?

thanks,

greg k-h

Re: [PATCH] dma-coherent: add support for multi coherent rmems per dev

[Howard Yen]

On Thu, Feb 1, 2024 at 11:41 PM Greg KH <gregkh@linuxfoundation.org> wrote:
>
> On Thu, Feb 01, 2024 at 10:45:30AM +0000, Robin Murphy wrote:
> > On 2024-02-01 9:35 am, Howard Yen wrote:
> > > Add support for multiple coherent rmems per device. This patch addes
> > > dma_mem_list to device structure to store multiple rmems.
> > >
> > > These multiple rmems can be assigned to the device one by one by
> > > of_reserved_mem_device_init_by_idx() with the memory-region
> > > declaration in device tree as below and store the rmem to the dma_mem_list.
> > >
> > >     device1@0 {
> > >             ...
> > >             memory-region = <&reserved_mem0>, <&reserved_mem1>;
> > >             ...
> > >     };
> > >
> > > When driver tries to allocate memory from the rmems, looks for the first
> > > available rmem and allocates the memory from this rmem.
> > >
> > > Then if driver removed, of_reserved_mem_device_release() needs to be
> > > invoked to release all the rmems assigned to the device.
> > >
> > > Signed-off-by: Howard Yen <howardyen@google.com>
> > > ---
> > >   include/linux/device.h |  1 +
> > >   kernel/dma/coherent.c  | 66 +++++++++++++++++++++++++++++++++++-------
> > >   2 files changed, 56 insertions(+), 11 deletions(-)
> > >
> > > diff --git a/include/linux/device.h b/include/linux/device.h
> > > index 97c4b046c09d..c8682ee507cf 100644
> > > --- a/include/linux/device.h
> > > +++ b/include/linux/device.h
> > > @@ -751,6 +751,7 @@ struct device {
> > >   #ifdef CONFIG_DMA_DECLARE_COHERENT
> > >     struct dma_coherent_mem *dma_mem; /* internal for coherent mem
> > >                                          override */
> > > +   struct list_head        dma_mem_list;
> >
> > I'm not necessarily against the idea, but only if it's implemented sensibly.
> > If we're going to have a list of these it should *replace* the existing
> > pointer, not do this weird thing with both.
>
> Agreed, it should be one pointer max for this structure for this type of
> thing.  Why not move it into the dma_coherent_mem structure?
>
> thanks,
>
> greg k-h

I'm considering to modify the change to

1. Move it into the dma_coherent_mem structure, like

     HEAD
mem0->node

This case, if I check list_empty(mem0->node), it would give me the
list is empty, but actually there is one rmem.

2. Replace the pointer to a list_head.

    HEAD
dma_mems ---> mem0->node

This case, if I check list_empty(dma_mems), it would give me the list
is non-empty, it matches the actual status.

So, the 2nd looks reasonable, I'm going to upload a v2 patch with the
2nd approach, does that make sense?

-- 
Regards,

Howard

Re: [PATCH] dma-coherent: add support for multi coherent rmems per dev

[Andy Shevchenko]

On Fri, Feb 02, 2024 at 12:40:59PM +0800, Howard Yen wrote:
> On Thu, Feb 1, 2024 at 11:41 PM Greg KH <gregkh@linuxfoundation.org> wrote:
> > On Thu, Feb 01, 2024 at 10:45:30AM +0000, Robin Murphy wrote:
> > > On 2024-02-01 9:35 am, Howard Yen wrote:

...

> I'm considering to modify the change to
> 
> 1. Move it into the dma_coherent_mem structure, like
> 
>      HEAD
> mem0->node
> 
> This case, if I check list_empty(mem0->node), it would give me the
> list is empty, but actually there is one rmem.
> 
> 2. Replace the pointer to a list_head.
> 
>     HEAD
> dma_mems ---> mem0->node
> 
> This case, if I check list_empty(dma_mems), it would give me the list
> is non-empty, it matches the actual status.
> 
> So, the 2nd looks reasonable, I'm going to upload a v2 patch with the
> 2nd approach, does that make sense?

I believe this is exactly what Robin told about "list to replace the pointer".

-- 
With Best Regards,
Andy Shevchenko