cpufreq: OMAP: Add SMP support for OMAP4+
Russell King [Wed, 21 Sep 2011 23:53:00 +0000 (16:53 -0700)]
On OMAP SMP configuartion, both processors share the voltage
and clock. So both CPUs needs to be scaled together and hence
needs software co-ordination.

Also, update lpj with reference value to avoid progressive error.

Adjust _both_ the per-cpu loops_per_jiffy and global lpj. Calibrate
them with with reference to the initial values to avoid a
progressively bigger and bigger error in the value over time.

While at this, re-use the notifiers for UP/SMP since on UP machine or
UP_ON_SMP policy->cpus mask would contain only the boot CPU.

Based on initial SMP support by Santosh Shilimkar.

Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
[khilman@ti.com: due to overlap/rework, combined original Santosh patch
                 and Russell's rework]
Signed-off-by: Kevin Hilman <khilman@ti.com>

drivers/cpufreq/omap-cpufreq.c

index a6b2be7..1953f9d 100644 (file)
 #include <linux/clk.h>
 #include <linux/io.h>
 #include <linux/opp.h>
+#include <linux/cpu.h>
 
 #include <asm/system.h>
 #include <asm/smp_plat.h>
+#include <asm/cpu.h>
 
 #include <plat/clock.h>
 #include <plat/omap-pm.h>
 
 #define VERY_HI_RATE   900000000
 
+#ifdef CONFIG_SMP
+struct lpj_info {
+       unsigned long   ref;
+       unsigned int    freq;
+};
+
+static DEFINE_PER_CPU(struct lpj_info, lpj_ref);
+static struct lpj_info global_lpj_ref;
+#endif
+
 static struct cpufreq_frequency_table *freq_table;
 static struct clk *mpu_clk;
 
@@ -60,7 +72,7 @@ static unsigned int omap_getspeed(unsigned int cpu)
 {
        unsigned long rate;
 
-       if (cpu)
+       if (cpu >= NR_CPUS)
                return 0;
 
        rate = clk_get_rate(mpu_clk) / 1000;
@@ -71,7 +83,7 @@ static int omap_target(struct cpufreq_policy *policy,
                       unsigned int target_freq,
                       unsigned int relation)
 {
-       int ret = 0;
+       int i, ret = 0;
        struct cpufreq_freqs freqs;
 
        /* Ensure desired rate is within allowed range.  Some govenors
@@ -81,22 +93,57 @@ static int omap_target(struct cpufreq_policy *policy,
        if (target_freq > policy->max)
                target_freq = policy->max;
 
-       freqs.old = omap_getspeed(0);
+       freqs.old = omap_getspeed(policy->cpu);
        freqs.new = clk_round_rate(mpu_clk, target_freq * 1000) / 1000;
-       freqs.cpu = 0;
+       freqs.cpu = policy->cpu;
 
        if (freqs.old == freqs.new)
                return ret;
 
-       cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
+       /* notifiers */
+       for_each_cpu(i, policy->cpus) {
+               freqs.cpu = i;
+               cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
+       }
 
 #ifdef CONFIG_CPU_FREQ_DEBUG
        pr_info("cpufreq-omap: transition: %u --> %u\n", freqs.old, freqs.new);
 #endif
 
        ret = clk_set_rate(mpu_clk, freqs.new * 1000);
+       freqs.new = omap_getspeed(policy->cpu);
+
+#ifdef CONFIG_SMP
+       /*
+        * Note that loops_per_jiffy is not updated on SMP systems in
+        * cpufreq driver. So, update the per-CPU loops_per_jiffy value
+        * on frequency transition. We need to update all dependent CPUs.
+        */
+       for_each_cpu(i, policy->cpus) {
+               struct lpj_info *lpj = &per_cpu(lpj_ref, i);
+               if (!lpj->freq) {
+                       lpj->ref = per_cpu(cpu_data, i).loops_per_jiffy;
+                       lpj->freq = freqs.old;
+               }
+
+               per_cpu(cpu_data, i).loops_per_jiffy =
+                       cpufreq_scale(lpj->ref, lpj->freq, freqs.new);
+       }
 
-       cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
+       /* And don't forget to adjust the global one */
+       if (!global_lpj_ref.freq) {
+               global_lpj_ref.ref = loops_per_jiffy;
+               global_lpj_ref.freq = freqs.old;
+       }
+       loops_per_jiffy = cpufreq_scale(global_lpj_ref.ref, global_lpj_ref.freq,
+                                       freqs.new);
+#endif
+
+       /* notifiers */
+       for_each_cpu(i, policy->cpus) {
+               freqs.cpu = i;
+               cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
+       }
 
        return ret;
 }
@@ -105,6 +152,7 @@ static int __cpuinit omap_cpu_init(struct cpufreq_policy *policy)
 {
        int result = 0;
        struct device *mpu_dev;
+       static cpumask_var_t cpumask;
 
        if (cpu_is_omap24xx())
                mpu_clk = clk_get(NULL, "virt_prcm_set");
@@ -116,12 +164,12 @@ static int __cpuinit omap_cpu_init(struct cpufreq_policy *policy)
        if (IS_ERR(mpu_clk))
                return PTR_ERR(mpu_clk);
 
-       if (policy->cpu != 0)
+       if (policy->cpu >= NR_CPUS)
                return -EINVAL;
 
-       policy->cur = policy->min = policy->max = omap_getspeed(0);
-
+       policy->cur = policy->min = policy->max = omap_getspeed(policy->cpu);
        mpu_dev = omap2_get_mpuss_device();
+
        if (!mpu_dev) {
                pr_warning("%s: unable to get the mpu device\n", __func__);
                return -EINVAL;
@@ -141,7 +189,20 @@ static int __cpuinit omap_cpu_init(struct cpufreq_policy *policy)
 
        policy->min = policy->cpuinfo.min_freq;
        policy->max = policy->cpuinfo.max_freq;
-       policy->cur = omap_getspeed(0);
+       policy->cur = omap_getspeed(policy->cpu);
+
+       /*
+        * On OMAP SMP configuartion, both processors share the voltage
+        * and clock. So both CPUs needs to be scaled together and hence
+        * needs software co-ordination. Use cpufreq affected_cpus
+        * interface to handle this scenario. Additional is_smp() check
+        * is to keep SMP_ON_UP build working.
+        */
+       if (is_smp()) {
+               policy->shared_type = CPUFREQ_SHARED_TYPE_ANY;
+               cpumask_or(cpumask, cpumask_of(policy->cpu), cpumask);
+               cpumask_copy(policy->cpus, cpumask);
+       }
 
        /* FIXME: what's the actual transition time? */
        policy->cpuinfo.transition_latency = 300 * 1000;