[V3] PM / OPP: Don't remove the first cpu in the mask before removing others

Message ID 29c512d06de99663c3c8acc2e04a898831879a6c.1480398079.git.viresh.kumar@linaro.org
State New
Headers show

Commit Message

Viresh Kumar Nov. 29, 2016, 5:42 a.m.
Joonyoung Shim reported an interesting problem on his ARM octa-core
Odoroid-XU3 platform. During system suspend, dev_pm_opp_put_regulator()
was failing for a struct device for which dev_pm_opp_set_regulator() is
called earlier.

This happened because an earlier call to
dev_pm_opp_of_cpumask_remove_table() function (from cpufreq-dt.c file)
removed all the entries from opp_table->dev_list apart from the last CPU
device in the cpumask of CPUs sharing the OPP.

But both dev_pm_opp_set_regulator() and dev_pm_opp_put_regulator()
routines get CPU device for the first CPU in the cpumask. And so the OPP
core failed to find the OPP table for the struct device.

In order to fix that up properly, we need to revisit APIs like
dev_pm_opp_set_regulator() and make them talk in terms of cookies
provided by the OPP core. But such a solution will be hard to backport
to stable kernels.

This patch attempts to fix this problem (in a Hacky way) by specially
handling the first cpu in the mask. A FIXME is also added to make sure
that this Hack doesn't get unnoticed later on.

Cc:  # v4.4+ <stable@vger.kernel.org>
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>

---
V2->V3:
- Fix $Subject, which carried text from V1.

V1->V2:
- A completely different approach, more of hack so that backport to
  stable kernels can be done easily.
- A more comprehensive solution is required to fix the design flaws.

 drivers/base/power/opp/cpu.c | 50 +++++++++++++++++++++++++++++++-------------
 1 file changed, 35 insertions(+), 15 deletions(-)

-- 
2.7.1.410.g6faf27b

--
To unsubscribe from this list: send the line "unsubscribe stable" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Comments

Rafael J. Wysocki Nov. 29, 2016, 12:49 p.m. | #1
On Tue, Nov 29, 2016 at 6:42 AM, Viresh Kumar <viresh.kumar@linaro.org> wrote:
> Joonyoung Shim reported an interesting problem on his ARM octa-core

> Odoroid-XU3 platform. During system suspend, dev_pm_opp_put_regulator()

> was failing for a struct device for which dev_pm_opp_set_regulator() is

> called earlier.

>

> This happened because an earlier call to

> dev_pm_opp_of_cpumask_remove_table() function (from cpufreq-dt.c file)

> removed all the entries from opp_table->dev_list apart from the last CPU

> device in the cpumask of CPUs sharing the OPP.

>

> But both dev_pm_opp_set_regulator() and dev_pm_opp_put_regulator()

> routines get CPU device for the first CPU in the cpumask. And so the OPP

> core failed to find the OPP table for the struct device.

>

> In order to fix that up properly, we need to revisit APIs like

> dev_pm_opp_set_regulator() and make them talk in terms of cookies

> provided by the OPP core. But such a solution will be hard to backport

> to stable kernels.

>

> This patch attempts to fix this problem (in a Hacky way) by specially

> handling the first cpu in the mask. A FIXME is also added to make sure

> that this Hack doesn't get unnoticed later on.

>

> Cc:  # v4.4+ <stable@vger.kernel.org>

> Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>

> ---

> V2->V3:

> - Fix $Subject, which carried text from V1.

>

> V1->V2:

> - A completely different approach, more of hack so that backport to

>   stable kernels can be done easily.

> - A more comprehensive solution is required to fix the design flaws.


I'll wait for a response from Stephen this time.

Thanks,
Rafael
--
To unsubscribe from this list: send the line "unsubscribe stable" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch

diff --git a/drivers/base/power/opp/cpu.c b/drivers/base/power/opp/cpu.c
index 8c3434bdb26d..5d1b0f98bcb0 100644
--- a/drivers/base/power/opp/cpu.c
+++ b/drivers/base/power/opp/cpu.c
@@ -118,26 +118,46 @@  void dev_pm_opp_free_cpufreq_table(struct device *dev,
 EXPORT_SYMBOL_GPL(dev_pm_opp_free_cpufreq_table);
 #endif	/* CONFIG_CPU_FREQ */
 
+void _cpu_remove_table(unsigned int cpu, bool of)
+{
+	struct device *cpu_dev = get_cpu_device(cpu);
+
+	if (!cpu_dev) {
+		pr_err("%s: failed to get cpu%d device\n", __func__, cpu);
+		return;
+	}
+
+	if (of)
+		dev_pm_opp_of_remove_table(cpu_dev);
+	else
+		dev_pm_opp_remove_table(cpu_dev);
+}
+
 void _dev_pm_opp_cpumask_remove_table(const struct cpumask *cpumask, bool of)
 {
-	struct device *cpu_dev;
-	int cpu;
+	struct cpumask tmpmask;
+	int cpu, first_cpu;
 
 	WARN_ON(cpumask_empty(cpumask));
 
-	for_each_cpu(cpu, cpumask) {
-		cpu_dev = get_cpu_device(cpu);
-		if (!cpu_dev) {
-			pr_err("%s: failed to get cpu%d device\n", __func__,
-			       cpu);
-			continue;
-		}
-
-		if (of)
-			dev_pm_opp_of_remove_table(cpu_dev);
-		else
-			dev_pm_opp_remove_table(cpu_dev);
-	}
+	/*
+	 * The first cpu in the cpumask is important as that is used to create
+	 * the opp-table initially and routines like dev_pm_opp_put_regulator()
+	 * will expect the list-dev for the first CPU to be present while such
+	 * routines are called, otherwise we will fail to find the opp-table for
+	 * such devices.
+	 *
+	 * FIXME: Cleanup this mess and implement cookie based solutions instead
+	 * of working on the device pointer.
+	 */
+	first_cpu = cpumask_first(cpumask);
+	cpumask_copy(&tmpmask, cpumask);
+	cpumask_clear_cpu(first_cpu, &tmpmask);
+
+	for_each_cpu(cpu, &tmpmask)
+		_cpu_remove_table(cpu, of);
+
+	_cpu_remove_table(first_cpu, of);
 }
 
 /**