[PATCH 3.6-rc6 2/2] cpufreq/powernow-k8: workqueue user shouldn't migrate the kworker to another CPU

[Tejun Heo <tj@kernel.org>]

powernowk8_target() runs off a per-cpu work item and if the
cpufreq_policy->cpu is different from the current one, it migrates the
kworker to the target CPU by manipulating current->cpus_allowed.  The
function migrates the kworker back to the original CPU but this is
still broken.  Workqueue concurrency management requires the kworkers
to stay on the same CPU and powernowk8_target() ends up triggerring
BUG_ON(rq != this_rq()) in try_to_wake_up_local() if it contends on
fidvid_mutex and sleeps.

It is unclear why this bug is being reported now.  Duncan says it
appeared to be a regression of 3.6-rc1 and couldn't reproduce it on
3.5.  Bisection seemed to point to 63d95a91 "workqueue: use @pool
instead of @gcwq or @cpu where applicable" which is an non-functional
change.  Given that the reproduce case sometimes took upto days to
trigger, it's easy to be misled while bisecting.  Maybe something made
contention on fidvid_mutex more likely?  I don't know.

This patch fixes the bug by using work_on_cpu() instead if @pol->cpu
isn't the same as the current one.  The code assumes that
cpufreq_policy->cpu is kept online by the caller, which Rafael tells
me is the case.

stable: This needs the patch "workqueue: reimplement work_on_cpu()
        using system_wq"; otherwise, the behavior could be horrible.

Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Duncan <1i5t5.duncan@cox.net>
Cc: Rafael J. Wysocki <rjw@sisk.pl>
Cc: Andreas Herrmann <andreas.herrmann3@amd.com>
Cc: stable@vger.kernel.org
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=47301
---
 drivers/cpufreq/powernow-k8.c |   63 ++++++++++++++++++++++--------------------
 1 file changed, 34 insertions(+), 29 deletions(-)

--- a/drivers/cpufreq/powernow-k8.c
+++ b/drivers/cpufreq/powernow-k8.c
@@ -35,7 +35,6 @@
 #include <linux/slab.h>
 #include <linux/string.h>
 #include <linux/cpumask.h>
-#include <linux/sched.h>	/* for current / set_cpus_allowed() */
 #include <linux/io.h>
 #include <linux/delay.h>
 
@@ -1139,16 +1138,23 @@ static int transition_frequency_pstate(s
 	return res;
 }
 
-/* Driver entry point to switch to the target frequency */
-static int powernowk8_target(struct cpufreq_policy *pol,
-		unsigned targfreq, unsigned relation)
+struct powernowk8_target_arg {
+	struct cpufreq_policy		*pol;
+	unsigned			targfreq;
+	unsigned			relation;
+};
+
+static long powernowk8_target_fn(void *arg)
 {
-	cpumask_var_t oldmask;
+	struct powernowk8_target_arg *pta = arg;
+	struct cpufreq_policy *pol = pta->pol;
+	unsigned targfreq = pta->targfreq;
+	unsigned relation = pta->relation;
 	struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
 	u32 checkfid;
 	u32 checkvid;
 	unsigned int newstate;
-	int ret = -EIO;
+	int ret;
 
 	if (!data)
 		return -EINVAL;
@@ -1156,29 +1162,16 @@ static int powernowk8_target(struct cpuf
 	checkfid = data->currfid;
 	checkvid = data->currvid;
 
-	/* only run on specific CPU from here on. */
-	/* This is poor form: use a workqueue or smp_call_function_single */
-	if (!alloc_cpumask_var(&oldmask, GFP_KERNEL))
-		return -ENOMEM;
-
-	cpumask_copy(oldmask, tsk_cpus_allowed(current));
-	set_cpus_allowed_ptr(current, cpumask_of(pol->cpu));
-
-	if (smp_processor_id() != pol->cpu) {
-		printk(KERN_ERR PFX "limiting to cpu %u failed\n", pol->cpu);
-		goto err_out;
-	}
-
 	if (pending_bit_stuck()) {
 		printk(KERN_ERR PFX "failing targ, change pending bit set\n");
-		goto err_out;
+		return -EIO;
 	}
 
 	pr_debug("targ: cpu %d, %d kHz, min %d, max %d, relation %d\n",
 		pol->cpu, targfreq, pol->min, pol->max, relation);
 
 	if (query_current_values_with_pending_wait(data))
-		goto err_out;
+		return -EIO;
 
 	if (cpu_family != CPU_HW_PSTATE) {
 		pr_debug("targ: curr fid 0x%x, vid 0x%x\n",
@@ -1196,7 +1189,7 @@ static int powernowk8_target(struct cpuf
 
 	if (cpufreq_frequency_table_target(pol, data->powernow_table,
 				targfreq, relation, &newstate))
-		goto err_out;
+		return -EIO;
 
 	mutex_lock(&fidvid_mutex);
 
@@ -1209,9 +1202,8 @@ static int powernowk8_target(struct cpuf
 		ret = transition_frequency_fidvid(data, newstate);
 	if (ret) {
 		printk(KERN_ERR PFX "transition frequency failed\n");
-		ret = 1;
 		mutex_unlock(&fidvid_mutex);
-		goto err_out;
+		return 1;
 	}
 	mutex_unlock(&fidvid_mutex);
 
@@ -1220,12 +1212,25 @@ static int powernowk8_target(struct cpuf
 				data->powernow_table[newstate].index);
 	else
 		pol->cur = find_khz_freq_from_fid(data->currfid);
-	ret = 0;
 
-err_out:
-	set_cpus_allowed_ptr(current, oldmask);
-	free_cpumask_var(oldmask);
-	return ret;
+	return 0;
+}
+
+/* Driver entry point to switch to the target frequency */
+static int powernowk8_target(struct cpufreq_policy *pol,
+		unsigned targfreq, unsigned relation)
+{
+	struct powernowk8_target_arg pta = { .pol = pol, .targfreq = targfreq,
+					     .relation = relation };
+
+	/*
+	 * Must run on @pol->cpu.  cpufreq core is responsible for ensuring
+	 * that we're bound to the current CPU and pol->cpu stays online.
+	 */
+	if (smp_processor_id() == pol->cpu)
+		return powernowk8_target_fn(&pta);
+	else
+		return work_on_cpu(pol->cpu, powernowk8_target_fn, &pta);
 }
 
 /* Driver entry point to verify the policy and range of frequencies */