On Wed, May 22, 2024 at 09:26:00AM +0200, Krzysztof Kozlowski wrote: > On 21/05/2024 21:17, Bjorn Andersson wrote: > >>> > >>> Hi Krzysztof, > >>> > >>> Sorry for the confusion, I dont think I meant that the smem driver will > >>> ever crash. The referred to crash in the cover letter is a crash in the > >>> firmware running on the remoteproc. The remoteproc could crash for any > >>> unexpected reason, related or unrelated to smem, while holding the tcsr > >>> mutex. I want to ensure that all resources that a remoteproc might be > >>> using are released as part of remoteproc stop. > >>> > >>> The SMEM driver manages the lock/unlock operations on the tcsr mutex > >>> from the Linux CPU's perspective. This case is for cleaning up from the > >>> remote side's perspective. > >>> > >>> In this case it's the hwspinlock used to synchronize SMEM, but it's > >>> conceivable that firmware running on the remoteproc has additional locks > >>> that need to be busted in order for the system to continue executing > >>> until the firmware is reinitialized. > >>> > >>> We did consider tying this to the SMEM instance, but the entitiy > >>> relating to firmware is the remoteproc instance. > >> > >> I still do not understand why you have to add hwlock to remoteproc, even > >> though it is not directly used. Your driver problem looks like lack of > >> proper driver architecture - you want to control the locks not from the > >> layer took the lock, but one layer up. Sorry, no, fix the driver > >> architecture. > >> > > > > No, it is the firmware's reference to the lock that is represented in > > the remoteproc node, while SMEM deals with Linux's reference to the lock. > > > > This reference would be used to release the lock - on behalf of the > > firmware - in the event that the firmware held it when it > > stopped/crashed. > > I understood, but the remoteproc driver did not acquire the hardware > lock. It was taken by smem, if I got it correctly, so you should poke > smem to bust the spinlock. > The remoteproc instance is the closest representation of the entity that took the lock (i.e. the firmware). SMEM here is just another consumer of the same lock. > The hwlock is not a property of remote proc, because remote proc does > not care, right? Other device cares... and now for every smem user you > will add new binding property? > Right, the issue seen relates to SMEM, because the remote processor (not the remoteproc driver) took the lock. > No, you are adding a binding based on your driver solution. Similar to how hwspinlocks are used in other platforms (e.g. TI) the firmware could take multiple locks, e.g. to synchronize access to other shared memory mechanism (i.e. not SMEM). While I am not aware of such use case today, my expectation was that in such case we just list all the hwlocks related to the firmware and bust those from the remoteproc instance. Having to export APIs from each one of such drivers and make the remoteproc identify the relevant instances and call those APIs does indeed seem inconvenient. SMEM is special here because it's singleton, but this would not necessarily be true for other cases. Regards, Bjorn
