[PATCH 07/16] soundwire: cadence: mask Slave interrupt before stopping clock - Pierre-Louis Bossart

From: Pierre-Louis Bossart <pierre-louis.bossart@linux.intel.com>
To: alsa-devel@alsa-project.org
Cc: linux-kernel@vger.kernel.org, tiwai@suse.de, broonie@kernel.org,
	vkoul@kernel.org, gregkh@linuxfoundation.org, jank@cadence.com,
	srinivas.kandagatla@linaro.org, slawomir.blauciak@intel.com,
	Bard liao <yung-chuan.liao@linux.intel.com>,
	Rander Wang <rander.wang@linux.intel.com>,
	Ranjani Sridharan <ranjani.sridharan@linux.intel.com>,
	Hui Wang <hui.wang@canonical.com>,
	Pierre-Louis Bossart <pierre-louis.bossart@linux.intel.com>,
	Sanyog Kale <sanyog.r.kale@intel.com>
Subject: [PATCH 07/16] soundwire: cadence: mask Slave interrupt before stopping clock
Date: Wed, 11 Mar 2020 13:41:19 -0500	[thread overview]
Message-ID: <20200311184128.4212-8-pierre-louis.bossart@linux.intel.com> (raw)
In-Reply-To: <20200311184128.4212-1-pierre-louis.bossart@linux.intel.com>

Intel QA reported a very rare case, possibly hardware-dependent, where
a Slave can become UNATTACHED during a clock stop sequence, which
leads to timeouts and failed suspend sequences.

This patch suppresses the handling of all Slave events while this
transition happens. The two cases that matter are:

a) alerts: if the Slave wants to signal an alert condition, it can do
so using the in-band wake, so there's almost no impact with this
patch.

b) sync loss or imp-def reset: in those cases, bringing back the Slave
to functional state requires a complete re-enumeration. It's better to
just ignore this case and restart cleanly, rather than attempt a
'clean' suspend.

Validation results show the timeouts no longer visible with this patch.

GitHub issue: https://github.com/thesofproject/linux/issues/1678
Signed-off-by: Pierre-Louis Bossart <pierre-louis.bossart@linux.intel.com>
---
 drivers/soundwire/cadence_master.c | 28 ++++++++++++++++++++++++++++
 1 file changed, 28 insertions(+)

diff --git a/drivers/soundwire/cadence_master.c b/drivers/soundwire/cadence_master.c
index 02f18ce6b7e7..a4ba57f44c1f 100644
--- a/drivers/soundwire/cadence_master.c
+++ b/drivers/soundwire/cadence_master.c
@@ -865,6 +865,24 @@ int sdw_cdns_exit_reset(struct sdw_cdns *cdns)
 }
 EXPORT_SYMBOL(sdw_cdns_exit_reset);
 
+/**
+ * sdw_cdns_enable_slave_interrupt() - Enable SDW slave interrupts
+ * @cdns: Cadence instance
+ * @state: boolean for true/false
+ */
+static void cdns_enable_slave_interrupts(struct sdw_cdns *cdns, bool state)
+{
+	u32 mask;
+
+	mask = cdns_readl(cdns, CDNS_MCP_INTMASK);
+	if (state)
+		mask |= CDNS_MCP_INT_SLAVE_MASK;
+	else
+		mask &= ~CDNS_MCP_INT_SLAVE_MASK;
+
+	cdns_writel(cdns, CDNS_MCP_INTMASK, mask);
+}
+
 /**
  * sdw_cdns_enable_interrupt() - Enable SDW interrupts
  * @cdns: Cadence instance
@@ -1272,6 +1290,13 @@ int sdw_cdns_clock_stop(struct sdw_cdns *cdns, bool block_wake)
 		return 1;
 	}
 
+	/*
+	 * Before entering clock stop we mask the Slave
+	 * interrupts. This helps avoid having to deal with e.g. a
+	 * Slave becoming UNATTACHED while the clock is being stopped
+	 */
+	cdns_enable_slave_interrupts(cdns, false);
+
 	/*
 	 * For specific platforms, it is required to be able to put
 	 * master into a state in which it ignores wake-up trials
@@ -1349,6 +1374,9 @@ int sdw_cdns_clock_restart(struct sdw_cdns *cdns, bool bus_reset)
 {
 	int ret;
 
+	/* unmask Slave interrupts that were masked when stopping the clock */
+	cdns_enable_slave_interrupts(cdns, true);
+
 	ret = cdns_clear_bit(cdns, CDNS_MCP_CONTROL,
 			     CDNS_MCP_CONTROL_CLK_STOP_CLR);
 	if (ret < 0) {
-- 
2.20.1

