Hello Mathieu,

On Tue, Nov 03, 2026 at 09:40:22AM -0600, Mathieu Poirier wrote:
> On Tue, Oct 27, 2026 at 05:25:20PM +0100, Uwe Kleine-König wrote:
> > diff --git a/drivers/remoteproc/remoteproc_elf_loader.c b/drivers/remoteproc/remoteproc_elf_loader.c
> > index da3cddbe7d4c..3c749a39e5da 100644
> > --- a/drivers/remoteproc/remoteproc_elf_loader.c
> > +++ b/drivers/remoteproc/remoteproc_elf_loader.c
> > @@ -227,35 +227,45 @@ int rproc_elf_load_segments(struct rproc *rproc, const struct firmware *fw)
> >  			break;
> >  		}
> >  
> > -		/* grab the kernel address for this device address */
> > -		ptr = rproc_da_to_va(rproc, da, memsz, &is_iomem);
> > -		if (!ptr) {
> > -			dev_err(dev, "bad phdr da 0x%llx mem 0x%llx\n", da,
> > -				memsz);
> > -			ret = -EINVAL;
> > -			break;
> > -		}
> > +		if (rproc->ops->memcpy_tova) {
> > +			ret = rproc->ops->memcpy_tova(rproc, da, elf_data + offset, filesz, memsz);
> > +			if (ret) {
> > +				dev_err(dev, "Failed to copy segment %d to device address 0x%llx: %pe\n",
> > +					i, da, ERR_PTR(ret));
> > +				break;
> > +			}
> 
> I really like your patch, but I'm wondering why you decided to move
> rproc_da_to_va() in ops->memcpy_tova().  Right now it is called in both path,
> i.e with and without ops->memcpy_tova().  Wouldn't it be better to
> unconditionally leave rproc_da_to_va() in rproc_elf_load_segments()?

The reason is that on ADI's SC846 the memory for the remote processor
isn't directly accessible by Linux. So there is no virtual address for
the given device address and thus the implementation of .memcpy_tova()
on that machine won't use rproc_da_to_va(). 

Best regards
Uwe