Hi Nicolas, On Tue, 5 May 2020 at 11:13, Nicolas Saenz Julienne <nsaenzjulienne@xxxxxxx> wrote: > > On Mon, 2020-05-04 at 21:46 +0100, Phil Elwell wrote: > > Hi Nicolas, > > > > On 04/05/2020 18:40, Nicolas Saenz Julienne wrote: > > > Hi Phil, Laurent, > > > > > > On Mon, 2020-05-04 at 12:26 +0300, Laurent Pinchart wrote: > > > > From: Phil Elwell <phil@xxxxxxxxxxxxxxx> > > > > > > > > Conditional on a new compatible string, change the pagelist encoding > > > > such that the top 24 bits are the pfn, leaving 8 bits for run length > > > > (-1). > > > > > > > > Signed-off-by: Phil Elwell <phil@xxxxxxxxxxxxxxx> > > > > Signed-off-by: Jacopo Mondi <jacopo@xxxxxxxxxx> > > > > --- > > > > .../interface/vchiq_arm/vchiq_2835_arm.c | 88 ++++++++++++++----- > > > > .../interface/vchiq_arm/vchiq_arm.c | 6 ++ > > > > .../interface/vchiq_arm/vchiq_arm.h | 1 + > > > > 3 files changed, 74 insertions(+), 21 deletions(-) > > > > > > > > diff --git > > > > a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_2835_arm.c > > > > b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_2835_arm.c > > > > index 3e422a7eb3f1..ecec84ad4345 100644 > > > > --- a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_2835_arm.c > > > > +++ b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_2835_arm.c > > > > @@ -16,6 +16,8 @@ > > > > #include <soc/bcm2835/raspberrypi-firmware.h> > > > > #define TOTAL_SLOTS (VCHIQ_SLOT_ZERO_SLOTS + 2 * 32) > > > > +#define VC_SAFE(x) (g_use_36bit_addrs ? ((u32)(x) | 0xc0000000) : > > > > (u32)(x)) > > > > +#define IS_VC_SAFE(x) (g_use_36bit_addrs ? !((x) & ~0x3fffffffull) : 1) > > > > > > As I commented earlier, this is a sign your dma-ranges are wrong, most of > > > the > > > code below reimplements what is already done by dma-direct (see > > > kernel/dma/direct.c). Once properly setup, you should be able to use > > > whatever > > > phys address dmam_alloc_coherent() provides and drop g_use_36bit_addrs. > > > > > > Note that on arm32+LPAE, dma-direct/swiotlb are the default dma_ops, so this > > > also applies there. > > > > As I explained in an offline email, the problem is that VCHIQ needs access to > > We discussed this privately, but I wanted to start from scratch, specially as I > hope I won't be the only one reviewing this :). > > > two > > kinds of "DMA" addresses - those suitable for the 40-bit DMA channel, and the > > 30-bit addresses that the VPU can use. Since each DT node only has access to a > > single set of DMA ranges, I can't see how to use dma-direct to calculate > > addreses > > for everything, but feel free to submit an alternative implementation showing > > how > > it should be done. > > How about this): > - Move vchiq to /soc > - Get a handle to the relevant dma controller device pointer (so 30-bit DMA > controller on old RPis, 40-bit on RPi4) > - Allocate slotmem/pagelist with dma_alloc_coherent(vpu_dev, ...) > - Map pages with dma_map_sg(dma_dev, ...) > > I hope I'm not missing anything, but short of implementing it and seeing the > end result, I think you'd be free of any rpi[1-3]/rpi4 distinction in the code. Thanks for the suggestion - I hadn't considered using a device pointer for the controller rather than the current (client) device. If that works then I propose to make the 40-bit dma property optional such that dma_dev defaults to vpu_dev. Phil
