Re: Ongoing remoteproc discussions

[Suman Anna]

Hi Bjorn,

On 07/18/2016 06:10 PM, Bjorn Andersson wrote:
> During discussions with various people interested in moving their
> remoteproc-related out-of-tree patches towards mainline I have come
> across a set of topics common among various teams. The purpose of this
> email is to share some insight into these discussions and the current
> ideas for moving forward.

Thanks for putting this together - a very good summary. I have a few
comments/questions below.

> 
> == Auto-boot or always-on:
> There are cases where we want to achieve the current auto-boot mechanism
> without rpmsg and there are cases where we don't want to auto-boot a
> remoteproc just because its resource table contains rpmsg entries. So we
> need to decouple this logic from the vdev. I suggest that:

I am trying to understand the usecase where one doesn't want to
auto-boot with rpmsg entries, did you come across such an usecase?

> 
> * We introduce a property in the rproc struct where drivers can toggle
>   if they want always-on or not. We default this to true, as an estimate
>   of backwards compatibility.

Can this be made into a flags field rather than a boolean, especially to
dictate the behavior between firmwares with vdevs and without vdevs. I
do have some usecases that can have vdevs (so auto-boot) or no vdevs
(with dummy resource tables) based on firmware being loaded (rproc_boot
will be called by a client driver in that case).

Your current patch-set does break the wkup_m3_rproc driver as it now
auto-boots.

> 
> * We move the allocation of vrings to be done at the time of the other
>   allocations and follow that with the registration of virtio devices,
>   before actually booting the rproc. And we tear down the virtio devices
>   as part of shutdown.
> 
> * We remove the rproc_boot()/shutdown() calls from the
>   remoteproc_virtio.c and based on rproc->always_on we call these in the
>   async firmware loader callback, in rproc_trigger_recovery() and
>   rproc_del().
> 
> 
> A side effect of this is that the async firmware loading only purpose is
> to trigger the boot as the firmware becomes available (something we need
> to tweak further). I therefor see no reason to mandate the firmware is
> unchanged between boots, which simplifies the posed rproc_set_firmware()
> function - as long as it's done on a remoteproc that's not always_on.

I take it that you are talking about adding a new API down the road
which should allow client drivers to set firmwares.

+1 for that.

> 
> 
> == Amended resources:
> We need a way for the driver to amend resource definitions of the
> firmware with e.g. physical addresses and size constraints from
> devicetree, so I suggest that:
> 
> * We split the resource table parsing and allocation of resources in two
>   steps; where the parse step updates the lists of resources and then at
>   the time of boot we run through these and allocate the actual
>   resources.
> 
> * We expose an API to the drivers so that they can register resources,
>   as if they came from the table parser. We match each resource against
>   previously registered resources based on name and merge or reject the
>   updates. E.g. we merge a reference to the resource table offset and we
>   reject incompatible size changes.

So, all this to be invoked by a platform driver before rproc_add() even
before requesting and processing a firmware? Is this also only limited
for remoteproc platform drivers, or gonna be extended to client drivers
(in the case where they are setting the firmwares).

One question w.r.t Lee's series is also adding new resources - how does
that work with loaded resource tables given that you are dynamically
increasing the table size, and the firmware image is already pre-linked.
Or is the assumption here that there will not be a loaded resource table
in such a scenario, we will have a loaded resource table for firmwares
with vdevs for sure.

> 
> 
> == Resource-less firmware:
> To be able to use remoteproc with firmware either without a resource
> table or resource data in other forms we today provide a resource table
> stub in each driver, instead we could refactor the resource table
> parsing code.
> 
> * We replace the find_rsc_table operation in rproc_fw_ops with a parse
>   operation, that uses the newly created API (above) to register the
>   resources with the core; largely decoupling the resource table format
>   from the remoteproc core.
> 
> * We make the parse() function in rproc_fw_ops optional, allowing
>   remoteproc drivers to specify that there's no resource parsing to be
>   done (they can still provide resources programmatically between
>   rproc_alloc() and rproc_add()).
> 
> This setup allows custom resource building functions to be implemented,
> one such example is the Qualcomm firmware files where most resource data
> is a combination of static information (DT) and data from the ELF
> header.
> 
> 
> == Resource-less firmware with installed resource table:
> The now available list of resources that have been registered with the
> core can serve as input for a function that generates a resource table
> for the remote.  This gives us a mechanism for providing a remoteproc
> with information about resources in cases where the firmware lacks a
> resource table.
> 
> * We replace the rproc_find_loaded_rsc_table() with an new fw_op that is
>   tasked with installing the resource table and update rproc->table_ptr.
>   This op is made optional, for the remoteprocs with no installed
>   resource table.

So, kinda similar to above question, how do you find a suitable location
for this in the case of installed resource table?

> 
> * We provide a helper function in the core that can be used in the
>   fw_op, that builds a resource table in memory and copy this into the
>   appropriate address of the remote, and  update rproc->table_ptr to
>   this.
> 
> The install function will be tasked with making sure table_ptr is valid
> and we make sure that error paths out of there and upon shutdown the
> core will make table_ptr reference the core's version - which might be
> NULL if no resource table exist.
> 
> 
> == Supporting rpmsg alternatives:
> For systems with other communication mechanisms than rpmsg we still want
> to be able to instantiate and tear down devices representing these
> communication channels, in a way that follows the life cycle of the
> remoteproc. To do this I suggest that:
> 
> * Like other resources we split virtio device handling in the remoteproc
>   core into adding rproc_vdevs to rproc->rvdevs and actually calling
>   rproc_add_virtio_dev().
> 
> * We generalise the rproc->rvdev somewhat, to be a list of "subdevices"
>   that are registered with the remoteproc; each providing callbacks for
>   registering and unregistering themselves as we bring the remoteproc up
>   and down.

+1, this is a good direction.

regards
Suman

> 
> I made a quick hack of this with the Qualcomm SMD code and it turned out
> pretty good, but I believe it's too ugly to post until we get the
> cleanups from above tasks sorted out.
> 
> 
> == Ramdump:
> Being able to acquire core dumps from a miss-behaving remoteproc is an
> important feature in product development. I believe this snapshot should
> be taken between the shutdown of the remote and freeing of resources. As
> such I think it would be appropriate to:
> 
> * Split the inner working of rproc_shutdown() into the two steps of
>   shutting down the remote and cleaning up its resources. Giving us a
>   window of opportunity for snapshotting any resources.
> 
> In the generic case we will have to repopulate the resources with data
> from the firmware file (in case of corruptions), but we're expecting to
> load the same firmware again and as such I see no meaning in releasing
> and reacquiring carveouts etc. As such if we split the inner working of
> rproc_boot() into resource allocation, virtio device allocation and
> booting we can make rproc_trigger_recovery() do:
> 
>     * Shutdown the remoteproc
>     * Shutdown virtio devices
>     * Take snapshot
>     * Register virtio devices
>     * Start the remoteproc
> 
> (The order of the top two is opposite of todays implementation, but
> probably more appropriate for the case of getting an accurate snapshot
> of the device).
> 
> We need to discuss the requirements for the format of what comes out of
> this.  I've seen raw memory dumps of a fixed memory segments and I've
> seen ELF generators with segments matching those of the loaded ELF in
> the past.
> 
> 
> == Firmware from late mounted file systems (needs discussion):
> Typical for most remoteproc systems seems to be that the firmware files
> tend to be way to large for being stored in a initramfs. The two
> available mechanisms for dealing with this is to build remoteproc
> drivers as modules and relying on CONFIG_FW_LOADER_USER_HELPER_FALLBACK;
> the latter is being frowned upon by everyone but works fairly well for
> our purpose.
> 
> Forcing systems to compile remoteproc drivers as modules to achieve the
> delayed firmware load is not acceptable in my view.
> 
> So we need to come up with some mechanism for triggering auto-booting
> when firmware files are stored in a lately mounted file system.
> 
> 
> 
> Interwoven in these discussions are of course topics related to rpmsg
> and I will try to get back to these as things materialize.
> 
> Regards,
> Bjorn
>