Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Marc Zyngier <maz@kernel.org>
To:     linux-tegra@vger.kernel.org, linux-kernel@vger.kernel.org,
        linux-arm-kernel@lists.infradead.org
Cc:     Thierry Reding <thierry.reding@gmail.com>,
        Jonathan Hunter <jonathanh@nvidia.com>,
        Dmitry Osipenko <digetx@gmail.com>,
        Sowjanya Komatineni <skomatineni@nvidia.com>,
        Venkat Reddy Talla <vreddytalla@nvidia.com>,
        Thomas Gleixner <tglx@linutronix.de>, kernel-team@android.com
Subject: [PATCH 0/3] soc/tegra: Prevent the PMC driver from corrupting interrupt routing
Date:   Mon,  5 Oct 2020 12:14:40 +0100
Message-Id: <20201005111443.1390096-1-maz@kernel.org>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

Jon recently reported that one of the Tegra systems (Jetson TX2, aka
tegra186) stopped booting with the introduction of the "IPI as IRQs"
series. After a few weeks of head scratching and complete puzzlement,
I obtained a board and started looking at what was happening.

The interrupt hierarchy looks like this:

	[DEVICE] -A-> [PMC] -B-> [GIC]

which seems simple enough. However, not all the devices attached to
the PMC follow this hierarchy, and in some cases, the 'B' link isn't
present in the HW. In other cases, neither 'A' nor 'B' are present.
And yet the PMC driver creates such linkages using random hwirq values
for the non-existent links, potentially overriding existing mappings
in the process. "What could possibly go wrong?"

It turns out that for the 'B' link, the PMC driver uses hwirq 0, which
is SGI0 for the GIC, and used as the rescheduling IPI. Obviously, this
doesn't go very well, nor very far, as the IPI gets routed to random
drivers. Also, as the handling flow has been overridden, this
interrupt never gets deactivated and can't fire anymore. Yes, this is
bad.

The 'A' link is less problematic, but the hwirq value is still out of
the irqdomain range, and gets remapped every time a new 'A'-less
driver comes up.

Instead, let's trim the unused hierarchy levels as needed. This
requires some checks in the upper levels of the hierarchy as we now
have optional levels, but this looks a lot saner than what we
currently have. With this, tegra186 is back booting on -next.

I haven't tested any wake-up stuff, nor any other nvidia system (this
is the only one I have). If people agree to these changes, I can take
them via the irqchip tree so that they make it into the next merge
window.

	M.

Marc Zyngier (3):
  gpio: tegra186: Allow optional irq parent callbacks
  soc/tegra: pmc: Allow optional irq parent callbacks
  soc/tegra: pmc: Don't create fake interrupt hierarchy levels

 drivers/gpio/gpio-tegra186.c |  15 +++-
 drivers/soc/tegra/pmc.c      | 142 ++++++++++++++++++++---------------
 2 files changed, 95 insertions(+), 62 deletions(-)

-- 
2.28.0