Re: [PATCH] remoteproc: Add module parameter 'auto_boot'

[Bjorn Andersson]

On Sat 21 Nov 12:38 CST 2020, Paul Cercueil wrote:

> Hi Mathieu,
> 
> Le ven. 20 nov. 2020 à 15:37, Mathieu Poirier <mathieu.poirier@xxxxxxxxxx> a
> écrit :
> > Hi Paul,
> > 
> > On Sun, Nov 15, 2020 at 11:50:56AM +0000, Paul Cercueil wrote:
> > >  Until now the remoteproc core would always default to trying to
> > > boot the
> > >  remote processor at startup. The various remoteproc drivers could
> > >  however override that setting.
> > > 
> > >  Whether or not we want the remote processor to boot, really depends
> > > on
> > >  the nature of the processor itself - a processor built into a WiFi
> > > chip
> > >  will need to be booted for the WiFi hardware to be usable, for
> > > instance,
> > >  but a general-purpose co-processor does not have any predeterminated
> > >  function, and as such we cannot assume that the OS will want the
> > >  processor to be booted - yet alone that we have a single do-it-all
> > >  firmware to load.
> > > 
> > 
> > If I understand correctly you have various remote processors that use
> > the same firmware
> > but are serving different purposes - is this correct?
> 
> That's the opposite actually. I have one remote processor which is
> general-purpose, and as such userspace may or may not want it started at
> boot time - depending on what it wants to do with it. The kernel shouldn't
> decide itself whether or not the remote processor should be started, because
> that's policy.
> 
> > 
> > >  Add a 'auto_boot' module parameter that instructs the remoteproc
> > > whether
> > >  or not it should auto-boot the remote processor, which will default
> > > to
> > >  "true" to respect the previous behaviour.
> > > 
> > 
> > Given that the core can't be a module I wonder if this isn't something
> > that
> > would be better off in the specific platform driver or the device
> > tree...  Other
> > people might have an opinion as well.
> 
> Hardcoded in the platform driver or flagged in the device tree, doesn't
> change the fundamental problem - it should be up to the userspace to decide
> whether or not the remote processor should boot.
> 

Unfortunately it depends on what you're using your remoteprocs for. And
in a system with multiple remoteproc instances I don't think a single
global parameter is sufficient - not even a per-driver setting is.

I do agree with you that there are types of systems where the decision
to auto boot things would happen after the kernel and/or DT has been
written.

Regards,
Bjorn

> Cheers,
> -Paul
> 
> > 
> > >  Signed-off-by: Paul Cercueil <paul@xxxxxxxxxxxxxxx>
> > >  ---
> > >   drivers/remoteproc/remoteproc_core.c | 7 ++++++-
> > >   1 file changed, 6 insertions(+), 1 deletion(-)
> > > 
> > >  diff --git a/drivers/remoteproc/remoteproc_core.c
> > > b/drivers/remoteproc/remoteproc_core.c
> > >  index dab2c0f5caf0..687b1bfd49db 100644
> > >  --- a/drivers/remoteproc/remoteproc_core.c
> > >  +++ b/drivers/remoteproc/remoteproc_core.c
> > >  @@ -44,6 +44,11 @@
> > > 
> > >   #define HIGH_BITS_MASK 0xFFFFFFFF00000000ULL
> > > 
> > >  +static bool auto_boot = true;
> > >  +module_param(auto_boot, bool, 0400);
> > >  +MODULE_PARM_DESC(auto_boot,
> > >  +		 "Auto-boot the remote processor [default=true]");
> > >  +
> > >   static DEFINE_MUTEX(rproc_list_mutex);
> > >   static LIST_HEAD(rproc_list);
> > >   static struct notifier_block rproc_panic_nb;
> > >  @@ -2176,7 +2181,7 @@ struct rproc *rproc_alloc(struct device *dev,
> > > const char *name,
> > >   		return NULL;
> > > 
> > >   	rproc->priv = &rproc[1];
> > >  -	rproc->auto_boot = true;
> > >  +	rproc->auto_boot = auto_boot;
> > >   	rproc->elf_class = ELFCLASSNONE;
> > >   	rproc->elf_machine = EM_NONE;
> > > 
> > >  --
> > >  2.29.2
> > > 
> 
>