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@xxxxxxxxxxxxxxxxxxx> 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@xxxxxxxxxx>
> > > ---
> > >   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