Re: [PATCH v4 00/17] remoteproc: Add support for detaching a rproc

Hi Mathieu

On 12/18/20 6:32 PM, Mathieu Poirier wrote:

Following the work done here [1], this set provides support for the
remoteproc core to release resources associated with a remote processor
without having to switch it off. That way a platform driver can be removed
or the application processor power cycled while the remote processor is
still operating.

Of special interest in this series are patches 5 and 6 where getting the
address of the resource table installed by an eternal entity if moved to
the core.  This is to support scenarios where a remote process has been
booted by the core but is being detached.  To re-attach the remote
processor, the address of the resource table needs to be known at a later
time than the platform driver's probe() function.

Applies cleanly on v5.10

Thanks,
Mathieu

[1]. https://lkml.org/lkml/2020/7/14/1600

----
New for v4:
- Made binding description OS agnostic (Rob)
- Added functionality to set the external resource table in the core
- Fixed a crash when detaching (Arnaud)
- Fixed error code propagation in rproc_cdev_relase() and rproc_del() (Arnaud)
- Added RB tags

I tested you series, attach and  detach is working well.

Then I faced issue when tried to re-attach after a detach.

But I don't know if this feature has to be supported in this step.

The 2 issues I found are:

1) memory carveouts are released on detach so need to be reinitialized.
The use of prepare/unprepare for the attach and detach would solve the issue but
probably need to add parameter to differentiate a start/stop from a attach/detach.

2) The vrings in the loaded resource table (so no cached) has to be properly
reinitialized. In rproc_free_vring  the vring da is set to 0 that is then
considered as a fixed address.

Here is a fix which works on the stm32 platform

@@ -425,7 +425,7 @@ void rproc_free_vring(struct rproc_vring *rvring)
 	 */
 	if (rproc->table_ptr) {
 		rsc = (void *)rproc->table_ptr + rvring->rvdev->rsc_offset;
-		rsc->vring[idx].da = 0;
+		rsc->vring[idx].da = FW_RSC_ADDR_ANY;
 		rsc->vring[idx].notifyid = -1;
 	}
 }

Here, perhaps a better alternative would be to make a cached copy on attach
before updating it. On the next attach, the cached copy would be copied as it is
done in rproc_start.

Thanks,
Arnaud

Mathieu Poirier (17):
  dt-bindings: remoteproc: Add bindind to support autonomous processors
  remoteproc: Re-check state in rproc_shutdown()
  remoteproc: Remove useless check in rproc_del()
  remoteproc: Rename function rproc_actuate()
  remoteproc: Add new get_loaded_rsc_table() remoteproc operation
  remoteproc: stm32: Move resource table setup to rproc_ops
  remoteproc: Add new RPROC_ATTACHED state
  remoteproc: Properly represent the attached state
  remoteproc: Properly deal with a kernel panic when attached
  remoteproc: Add new detach() remoteproc operation
  remoteproc: Introduce function __rproc_detach()
  remoteproc: Introduce function rproc_detach()
  remoteproc: Add return value to function rproc_shutdown()
  remoteproc: Properly deal with a stop request when attached
  remoteproc: Properly deal with a start request when attached
  remoteproc: Properly deal with detach request
  remoteproc: Refactor rproc delete and cdev release path

 .../bindings/remoteproc/remoteproc-core.yaml  |  27 +++
 drivers/remoteproc/remoteproc_cdev.c          |  32 ++-
 drivers/remoteproc/remoteproc_core.c          | 211 +++++++++++++++---
 drivers/remoteproc/remoteproc_internal.h      |   8 +
 drivers/remoteproc/remoteproc_sysfs.c         |  20 +-
 drivers/remoteproc/stm32_rproc.c              | 147 ++++++------
 include/linux/remoteproc.h                    |  24 +-
 7 files changed, 344 insertions(+), 125 deletions(-)
 create mode 100644 Documentation/devicetree/bindings/remoteproc/remoteproc-core.yaml