On Wed, 9 Dec 2020 at 08:00, Peng Fan <peng.fan@xxxxxxx> wrote:
>
> > Subject: RE: [PATCH V3 1/7] remoteproc: elf: support platform specific
> > memory hook
> >
> > Hi Bjorn,
> >
> > > Subject: Re: [PATCH V3 1/7] remoteproc: elf: support platform specific
> > > memory hook
> > >
> > > On Fri 04 Dec 01:40 CST 2020, Peng Fan (OSS) wrote:
> > >
> > > > From: Peng Fan <peng.fan@xxxxxxx>
> > > >
> > > > To arm64, "dc zva, dst" is used in memset.
> > > > Per ARM DDI 0487A.j, chapter C5.3.8 DC ZVA, Data Cache Zero by VA,
> > > >
> > > > "If the memory region being zeroed is any type of Device memory,
> > > > this instruction can give an alignment fault which is prioritized in
> > > > the same way as other alignment faults that are determined by the
> > > > memory type."
> > > >
> > > > On i.MX platforms, when elf is loaded to onchip TCM area, the region
> > > > is ioremapped, so "dc zva, dst" will trigger abort. And ioremap_wc()
> > > > on i.MX not able to write correct data to TCM area.
> > > >
> > > > So we need to use io helpers, and extend the elf loader to support
> > > > platform specific memory functions.
> > > >
> > > > Acked-by: Richard Zhu <hongxing.zhu@xxxxxxx>
> > > > Signed-off-by: Peng Fan <peng.fan@xxxxxxx>
> > > > Reviewed-by: Mathieu Poirier <mathieu.poirier@xxxxxxxxxx>
> > > > ---
> > > > drivers/remoteproc/remoteproc_elf_loader.c | 20
> > > ++++++++++++++++++--
> > > > include/linux/remoteproc.h | 4 ++++
> > > > 2 files changed, 22 insertions(+), 2 deletions(-)
> > > >
> > > > diff --git a/drivers/remoteproc/remoteproc_elf_loader.c
> > > > b/drivers/remoteproc/remoteproc_elf_loader.c
> > > > index df68d87752e4..6cb71fe47261 100644
> > > > --- a/drivers/remoteproc/remoteproc_elf_loader.c
> > > > +++ b/drivers/remoteproc/remoteproc_elf_loader.c
> > > > @@ -129,6 +129,22 @@ u64 rproc_elf_get_boot_addr(struct rproc
> > > > *rproc, const struct firmware *fw) }
> > > EXPORT_SYMBOL(rproc_elf_get_boot_addr);
> > > >
> > > > +static void rproc_elf_memcpy(struct rproc *rproc, void *dest, const
> > > > +void *src, size_t count) {
> > > > + if (!rproc->ops->elf_memcpy)
> > > > + memcpy(dest, src, count);
> > > > +
> > > > + rproc->ops->elf_memcpy(rproc, dest, src, count);
> > >
> > > Looking at the current set of remoteproc drivers I get a feeling that
> > > we'll end up with a while bunch of functions that all just wraps
> > > memcpy_toio(). And the reason for this is that we are we're "abusing"
> > > the carveout to carry the __iomem pointer without keeping track of it.
> > >
> > > And this is not the only time we're supposed to use an io-accessor,
> > > another example is rproc_copy_segment() in rproc_coredump.c
> > >
> > > It also means that if a platform driver for some reason where to
> > > support both ioremap and normal carveouts the elf_memcpy op would be
> > quite quirky.
> > >
> > >
> > > So I would prefer if we track the knowledge about void *va being a
> > > __iomem or not in the struct rproc_mem_entry and make rproc_da_to_va()
> > > return this information as well.
> > >
> > > Then instead of extending the ops we can make this simply call memcpy
> > > or
> > > memcpy_toio() depending on this.
> >
> > A draft proposal as below, are you ok with the approach?
>
> Mathieu, do you have any comments?
>
I will look into this on Monday.
