On 03/03/2017 08:37 AM, Laurent Pinchart wrote:
> Hi Laura,
>
> Thank you for the patch.
>
> On Thursday 02 Mar 2017 13:44:36 Laura Abbott wrote:
>> Technically, calling dma_buf_map_attachment should return a buffer
>> properly dma_mapped. Add calls to dma_map_sg to begin_cpu_access to
>> ensure this happens. As a side effect, this lets Ion buffers take
>> advantage of the dma_buf sync ioctls.
>>
>> Signed-off-by: Laura Abbott <labbott@xxxxxxxxxx>
>> ---
>> drivers/staging/android/ion/ion.c | 101 +++++++++++++++++------------------
>> 1 file changed, 50 insertions(+), 51 deletions(-)
>>
>> diff --git a/drivers/staging/android/ion/ion.c
>> b/drivers/staging/android/ion/ion.c index ce4adac..a931b30 100644
>> --- a/drivers/staging/android/ion/ion.c
>> +++ b/drivers/staging/android/ion/ion.c
>> @@ -795,10 +795,6 @@ void ion_client_destroy(struct ion_client *client)
>> }
>> EXPORT_SYMBOL(ion_client_destroy);
>>
>> -static void ion_buffer_sync_for_device(struct ion_buffer *buffer,
>> - struct device *dev,
>> - enum dma_data_direction direction);
>> -
>> static struct sg_table *dup_sg_table(struct sg_table *table)
>> {
>> struct sg_table *new_table;
>> @@ -825,22 +821,43 @@ static struct sg_table *dup_sg_table(struct sg_table
>> *table) return new_table;
>> }
>>
>> +static void free_duped_table(struct sg_table *table)
>> +{
>> + sg_free_table(table);
>> + kfree(table);
>> +}
>> +
>> static struct sg_table *ion_map_dma_buf(struct dma_buf_attachment
>> *attachment, enum dma_data_direction direction)
>> {
>> struct dma_buf *dmabuf = attachment->dmabuf;
>> struct ion_buffer *buffer = dmabuf->priv;
>> + struct sg_table *table;
>> + int ret;
>> +
>> + /*
>> + * TODO: Need to sync wrt CPU or device completely owning?
>> + */
>> +
>> + table = dup_sg_table(buffer->sg_table);
>>
>> - ion_buffer_sync_for_device(buffer, attachment->dev, direction);
>> - return dup_sg_table(buffer->sg_table);
>> + if (!dma_map_sg(attachment->dev, table->sgl, table->nents,
>> + direction)){
>> + ret = -ENOMEM;
>> + goto err;
>> + }
>> +
>> +err:
>> + free_duped_table(table);
>> + return ERR_PTR(ret);
>> }
>>
>> static void ion_unmap_dma_buf(struct dma_buf_attachment *attachment,
>> struct sg_table *table,
>> enum dma_data_direction direction)
>> {
>> - sg_free_table(table);
>> - kfree(table);
>> + dma_unmap_sg(attachment->dev, table->sgl, table->nents, direction);
>> + free_duped_table(table);
>> }
>>
>> void ion_pages_sync_for_device(struct device *dev, struct page *page,
>> @@ -864,38 +881,6 @@ struct ion_vma_list {
>> struct vm_area_struct *vma;
>> };
>>
>> -static void ion_buffer_sync_for_device(struct ion_buffer *buffer,
>> - struct device *dev,
>> - enum dma_data_direction dir)
>> -{
>> - struct ion_vma_list *vma_list;
>> - int pages = PAGE_ALIGN(buffer->size) / PAGE_SIZE;
>> - int i;
>> -
>> - pr_debug("%s: syncing for device %s\n", __func__,
>> - dev ? dev_name(dev) : "null");
>> -
>> - if (!ion_buffer_fault_user_mappings(buffer))
>> - return;
>> -
>> - mutex_lock(&buffer->lock);
>> - for (i = 0; i < pages; i++) {
>> - struct page *page = buffer->pages[i];
>> -
>> - if (ion_buffer_page_is_dirty(page))
>> - ion_pages_sync_for_device(dev, ion_buffer_page(page),
>> - PAGE_SIZE, dir);
>> -
>> - ion_buffer_page_clean(buffer->pages + i);
>> - }
>> - list_for_each_entry(vma_list, &buffer->vmas, list) {
>> - struct vm_area_struct *vma = vma_list->vma;
>> -
>> - zap_page_range(vma, vma->vm_start, vma->vm_end - vma-
>> vm_start);
>> - }
>> - mutex_unlock(&buffer->lock);
>> -}
>> -
>> static int ion_vm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
>> {
>> struct ion_buffer *buffer = vma->vm_private_data;
>> @@ -1014,16 +999,24 @@ static int ion_dma_buf_begin_cpu_access(struct
>> dma_buf *dmabuf, struct ion_buffer *buffer = dmabuf->priv;
>> void *vaddr;
>>
>> - if (!buffer->heap->ops->map_kernel) {
>> - pr_err("%s: map kernel is not implemented by this heap.\n",
>> - __func__);
>> - return -ENODEV;
>> + /*
>> + * TODO: Move this elsewhere because we don't always need a vaddr
>> + */
>> + if (buffer->heap->ops->map_kernel) {
>> + mutex_lock(&buffer->lock);
>> + vaddr = ion_buffer_kmap_get(buffer);
>> + mutex_unlock(&buffer->lock);
>> }
>>
>> - mutex_lock(&buffer->lock);
>> - vaddr = ion_buffer_kmap_get(buffer);
>> - mutex_unlock(&buffer->lock);
>> - return PTR_ERR_OR_ZERO(vaddr);
>> + /*
>> + * Close enough right now? Flag to skip sync?
>> + */
>> + if (!dma_map_sg(buffer->dev->dev.this_device, buffer->sg_table->sgl,
>> + buffer->sg_table->nents,
>> + DMA_BIDIRECTIONAL))
>
> Aren't the dma_(un)map_* calls supposed to take a real, physical device as
> their first argument ? Beside, this doesn't seem to be the right place to
> create the mapping, as you mentioned in the commit message the buffer should
> be mapped in the dma_buf map handler. This is something that needs to be
> fixed, especially in the light of the comment in ion_buffer_create():
>
Yes, this might me a case of me getting the model incorrect again.
dma_buf_{begin,end}_cpu_access do not take a device structure and
from the comments:
/**
* dma_buf_begin_cpu_access - Must be called before accessing a dma_buf from the
* cpu in the kernel context. Calls begin_cpu_access to allow exporter-specific
* preparations. Coherency is only guaranteed in the specified range for the
* specified access direction.
* @dmabuf: [in] buffer to prepare cpu access for.
* @direction: [in] length of range for cpu access.
*
* Can return negative error values, returns 0 on success.
*/
If there are no buffer attachments, I guess the notion of 'coherency'
doesn't apply here so there is no need to do any kind of
syncing/mapping at all vs. trying to find a device out of nowhere.
I'll have to go back and re-think aligning sync/begin_cpu_access calls
and dma_buf_map calls, or more likely not overthink this.
> /*
> * this will set up dma addresses for the sglist -- it is not
> * technically correct as per the dma api -- a specific
> * device isn't really taking ownership here. However, in practice on
> * our systems the only dma_address space is physical addresses.
> * Additionally, we can't afford the overhead of invalidating every
> * allocation via dma_map_sg. The implicit contract here is that
> * memory coming from the heaps is ready for dma, ie if it has a
> * cached mapping that mapping has been invalidated
> */
>
> That's a showstopper in my opinion, the DMA address space can't be restricted
> to physical addresses, IOMMU have to be supported.
>
I missed a patch in this series to remove that. If Ion is going to exist outside
of staging it should not be making that assumption at all so I want to drop it.
Any performance implications should be fixed with the skip sync flag.
Thanks,
Laura
