On 17/10/16 22:13, Hauke Mehrtens wrote:
On 10/11/2016 03:42 PM, Matt Redfearn wrote:
The MIPS remote processor driver allows non-Linux firmware to take
control of and execute on one of the systems VPEs. The CPU must be
offlined from Linux first. A sysfs interface is created which allows
firmware to be loaded and changed at runtime. A full description is
available at . An example firmware that can be used with this driver
is available at .
This is useful to allow running bare metal code, or an RTOS, on one or
more CPUs while allowing Linux to continue running on those remaining.
The remote processor framework allows for firmwares to provide any
virtio device for communication between the firmware running on the
remote VP and Linux. For example  demonstrates a simple firmware
which provides a virtual serial port. Any string sent to the port is
case inverted and returned.
This is conceptually similar to the VPE loader functionality, but is
more standard as it fits into the remoteproc subsystem.
The first patches in this series lay the groundwork for the driver
before it is added. The last series deprecates the VPE loader.
This functionality is supported on:
- MIPS32r2 devices implementing the MIPS MT ASE for multithreading, such
- MIPS32r6 devices implementing VPs, such as I6400.
- The remoteproc core supports only 32bit ELFs. Therefore it is only
possible to run 32bit firmware on the remote processor. Also, for
virtio communication, pointers are passed from the kernel to firmware.
There can be no mismatch in pointer sizes between the kernel and
firmware, so this limits the host kernel to 32bit as well.
The functionality has been tested on the Ci40 board which has a 2 core 2
Does this also work with big firmware binaries, like 2 MB firmware size?
It looks like dma_alloc_coherent() is used in rproc_handle_carveout() to
allocate the memory which will probably not work after full boot.
In my testing, a 2Mb firmware carveout does work fine on a Creator Ci40
with 256Mb RAM. But like you say, using the DMA api will limit the size
of the firmware. In our driver's case, we end up setting up wired TLB
entries mapping the carveout into the offline CPUs virtual memory, so in
principle there is no reason (for our driver) that the memory has to be
contiguous since we could map smaller pages. But this is code in the
generic remoteproc framework so must work for other processors which
could not deal with that.
Is it possible to configure the cache line in a way that it will not be
removed from the cache to avoid high latency loads from main memory?
This is only needed for small firmware binaries, but allows even shorter
reaction times and better real time behavior.
The firmware running on the offline CPU has full access to the hardware
so could do whatever it likes with it's cache lines - it's not something
that the remote processor driver ought to do as a matter of course
though I don't think.
In general I am in favor of having this as a standard Linux interface.