Re: [PATCH] mailbox: txdone_method shouldn't always be reset

From: Sudeep Holla
Date: Thu Nov 16 2017 - 12:25:00 EST

(- Remove Alexey Klimov's ARM id as he has left the company, adding his
personal id instead)

On 16/11/17 16:51, Jassi Brar wrote:
> On Thu, Nov 16, 2017 at 11:01 AM, Bjorn Andersson
> <bjorn.andersson@xxxxxxxxxx> wrote:
>> A client that knows how to drive txdone would temporarily "upgrade" the
>> method to TXDONE_BY_ACK. But with the introduction of commit 33cd7123ac0ba
>> ("mailbox: reset txdone_method TXDONE_BY_POLL if client knows_txdone")
>> there is no longer a distinction between a channel in "upgraded" state
>> or a channel for a controller that only supports TXDONE_BY_ACK. So upon
>> freeing the channel it will be "downgraded" to TXDONE_BY_POLL.
>> But a channel that operates with the txdone method of TXDONE_BY_POLL
>> requires that the controller implements the last_tx_done callback and
>> that the associated hrtimer was initialized when the controller was
>> registered.
>> So the core now relies on the fact that subsequent calls to
>> mbox_request_channel() "upgrades" the channel to TXDONE_BY_ACK or it
>> will dereference the non-initialized hrtimer.
> I think we just need avoid the channel not running under POLL
> Does this fix your issue?

Unfortunately only Alexey had fast reproducer of the issue reported with
original commit. I will try to run on my setup, but it was hard to
reproduce(mainly lack of firmware support for multiple virtual channels).
I will try my best to test this.

Sorry for not considering temporary "upgrade" use case.