[ACPI] Reduce acpi-cpufreq switching latency by 50%
authorVenkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Thu, 25 Aug 2005 19:59:00 +0000 (15:59 -0400)
committerLen Brown <len.brown@intel.com>
Sat, 27 Aug 2005 02:22:33 +0000 (22:22 -0400)
The acpi-cpufreq driver does a P-state get after a P-state set
to verify whether set went through successfully. This test
is kind of redundant as set goes throught most of the times,
and the test is also expensive as a get of P-states can
take a lot of time (same as a set operation) as it goes
through SMM mode. Effectively, we are doubling the P-state
latency due to this get opertion.

momdule parameter "acpi_pstate_strict" restores orginal paranoia.

http://bugzilla.kernel.org/show_bug.cgi?id=5129

Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Len Brown <len.brown@intel.com>
arch/i386/kernel/cpu/cpufreq/acpi-cpufreq.c

index 60a9e54..822c8ce 100644 (file)
@@ -31,6 +31,7 @@
 #include <linux/cpufreq.h>
 #include <linux/proc_fs.h>
 #include <linux/seq_file.h>
+#include <linux/compiler.h>
 #include <asm/io.h>
 #include <asm/delay.h>
 #include <asm/uaccess.h>
@@ -57,6 +58,8 @@ static struct cpufreq_acpi_io *acpi_io_data[NR_CPUS];
 
 static struct cpufreq_driver acpi_cpufreq_driver;
 
+static unsigned int acpi_pstate_strict;
+
 static int
 acpi_processor_write_port(
        u16     port,
@@ -163,34 +166,44 @@ acpi_processor_set_performance (
        }
 
        /*
-        * Then we read the 'status_register' and compare the value with the
-        * target state's 'status' to make sure the transition was successful.
-        * Note that we'll poll for up to 1ms (100 cycles of 10us) before
-        * giving up.
+        * Assume the write went through when acpi_pstate_strict is not used.
+        * As read status_register is an expensive operation and there 
+        * are no specific error cases where an IO port write will fail.
         */
-
-       port = data->acpi_data.status_register.address;
-       bit_width = data->acpi_data.status_register.bit_width;
-
-       dprintk("Looking for 0x%08x from port 0x%04x\n",
-               (u32) data->acpi_data.states[state].status, port);
-
-       for (i=0; i<100; i++) {
-               ret = acpi_processor_read_port(port, bit_width, &value);
-               if (ret) {      
-                       dprintk("Invalid port width 0x%04x\n", bit_width);
-                       retval = ret;
-                       goto migrate_end;
+       if (acpi_pstate_strict) {
+               /* Then we read the 'status_register' and compare the value 
+                * with the target state's 'status' to make sure the 
+                * transition was successful.
+                * Note that we'll poll for up to 1ms (100 cycles of 10us) 
+                * before giving up.
+                */
+
+               port = data->acpi_data.status_register.address;
+               bit_width = data->acpi_data.status_register.bit_width;
+
+               dprintk("Looking for 0x%08x from port 0x%04x\n",
+                       (u32) data->acpi_data.states[state].status, port);
+
+               for (i=0; i<100; i++) {
+                       ret = acpi_processor_read_port(port, bit_width, &value);
+                       if (ret) {      
+                               dprintk("Invalid port width 0x%04x\n", bit_width);
+                               retval = ret;
+                               goto migrate_end;
+                       }
+                       if (value == (u32) data->acpi_data.states[state].status)
+                               break;
+                       udelay(10);
                }
-               if (value == (u32) data->acpi_data.states[state].status)
-                       break;
-               udelay(10);
+       } else {
+               i = 0;
+               value = (u32) data->acpi_data.states[state].status;
        }
 
        /* notify cpufreq */
        cpufreq_notify_transition(&cpufreq_freqs, CPUFREQ_POSTCHANGE);
 
-       if (value != (u32) data->acpi_data.states[state].status) {
+       if (unlikely(value != (u32) data->acpi_data.states[state].status)) {
                unsigned int tmp = cpufreq_freqs.new;
                cpufreq_freqs.new = cpufreq_freqs.old;
                cpufreq_freqs.old = tmp;
@@ -537,6 +550,8 @@ acpi_cpufreq_exit (void)
        return;
 }
 
+module_param(acpi_pstate_strict, uint, 0644);
+MODULE_PARM_DESC(acpi_pstate_strict, "value 0 or non-zero. non-zero -> strict ACPI checks are performed during frequency changes.");
 
 late_initcall(acpi_cpufreq_init);
 module_exit(acpi_cpufreq_exit);