Re: [RFC 2/4] cenalloc: Constraint-Enabled Allocation helpers for dma-buf
From: Greg Kroah-Hartman
Date: Fri Oct 10 2014 - 19:10:32 EST
On Sat, Oct 11, 2014 at 01:37:56AM +0530, Sumit Semwal wrote:
> Devices sharing buffers using dma-buf could benefit from sharing their
> constraints via struct device, and dma-buf framework would manage the
> common constraints for all attached devices per buffer.
> With that information, we could have a 'generic' allocator helper in
> the form of a central dma-buf exporter, which can create dma-bufs, and
> allocate backing storage at the time of first call to
> This allocation would utilise the constraint-mask by matching it to
> the right allocator from a pool of allocators, and then allocating
> buffer backing storage from this allocator.
> The pool of allocators could be platform-dependent, allowing for
> platforms to hide the specifics of these allocators from the devices
> that access the dma-buf buffers.
> A sample sequence could be:
> - get handle to cenalloc_device,
> - create a dmabuf using cenalloc_buffer_create;
> - use this dmabuf to attach each device, which has its constraints
> set in the constraints mask (dev->dma_params->access_constraints_mask)
> - at each dma_buf_attach() call, dma-buf will check to see if the constraint
> mask for the device requesting attachment is compatible with the constraints
> of devices already attached to the dma-buf; returns an error if it isn't.
> - after all devices have attached, the first call to dma_buf_map_attachment()
> will allocate the backing storage for the buffer.
> - follow the dma-buf api for map / unmap etc usage.
> - detach all attachments,
> - call cenalloc_buffer_free to free the buffer if refcount reaches zero;
> ** IMPORTANT**
> This mechanism of delayed allocation based on constraint-enablement will work
> *ONLY IF* the first map_attachment() call is made AFTER all attach() calls are
> Signed-off-by: Sumit Semwal <sumit.semwal@xxxxxxxxxx>
> Cc: linux-kernel@xxxxxxxxxxxxxxx
> Cc: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
> Cc: linux-media@xxxxxxxxxxxxxxx
> Cc: dri-devel@xxxxxxxxxxxxxxxxxxxxx
> Cc: linaro-mm-sig@xxxxxxxxxxxxxxxx
> MAINTAINERS | 1 +
> drivers/cenalloc/cenalloc.c | 597 +++++++++++++++++++++++++++++++++++++++
> drivers/cenalloc/cenalloc.h | 99 +++++++
> drivers/cenalloc/cenalloc_priv.h | 188 ++++++++++++
> 4 files changed, 885 insertions(+)
> create mode 100644 drivers/cenalloc/cenalloc.c
> create mode 100644 drivers/cenalloc/cenalloc.h
> create mode 100644 drivers/cenalloc/cenalloc_priv.h
> diff --git a/MAINTAINERS b/MAINTAINERS
> index 40d4796..e88ac81 100644
> --- a/MAINTAINERS
> +++ b/MAINTAINERS
> @@ -3039,6 +3039,7 @@ L: linux-media@xxxxxxxxxxxxxxx
> L: dri-devel@xxxxxxxxxxxxxxxxxxxxx
> L: linaro-mm-sig@xxxxxxxxxxxxxxxx
> F: drivers/dma-buf/
> +F: drivers/cenalloc/
> F: include/linux/dma-buf*
> F: include/linux/reservation.h
> F: include/linux/*fence.h
> diff --git a/drivers/cenalloc/cenalloc.c b/drivers/cenalloc/cenalloc.c
> new file mode 100644
> index 0000000..f278056
> --- /dev/null
> +++ b/drivers/cenalloc/cenalloc.c
> @@ -0,0 +1,597 @@
> + * Allocator helper framework for constraints-aware dma-buf backing storage
> + * allocation.
> + * This allows constraint-sharing devices to deferred-allocate buffers shared
> + * via dma-buf.
> + *
> + * Copyright(C) 2014 Linaro Limited. All rights reserved.
> + * Author: Sumit Semwal <sumit.semwal@xxxxxxxxxx>
> + *
> + * Structure for management of clients, buffers etc heavily derived from
> + * Android's ION framework.
Does that mean we can drop ION after this gets merged?
Anyway, why a new directory? Why not just put it in drivers/dma-buf ?
Or a subdir below there?
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/