This is a note to let you know that we have just queued up the patch titled
Subject: CPUFREQ: powernow-k7: fix MHz rounding issue with perflib
to the 2.6.21-stable tree. Its filename is
cpufreq-powernow-k7-fix-mhz-rounding-issue-with-perflib.patch
A git repo of this tree can be found at
http://www.kernel.org/git/?p=linux/kernel/git/stable/stable-queue.git;a=summary
>From [email protected] Thu May 17 13:43:35 2007
Date: Tue, 15 May 2007 19:59:03 GMT
Message-Id: <[email protected]>
From: Linux Kernel Mailing List <[email protected]>
To: [email protected]
Resent-From: [email protected]
Subject: CPUFREQ: powernow-k7: fix MHz rounding issue with perflib
From: Daniel Drake <[email protected]>
When the PST tables are broken, powernow-k7 uses ACPI's processor_perflib to
deduce the available frequency multipliers from the _PSS tables.
Upon frequency change, processor_perflib performs some verification on the
frequency (checks that it's within allowable bounds).
powernow-k7 deals with absolute frequencies in KHz, whereas perflib only
deals with MHz values. When performing the above verification, perflib
multiplies the MHz values by 1000 to obtain the KHz value.
We then end up with situations like the following:
- powernow-k7 multiplies the multiplier by the FSB, and obtains a value
such as 1266768 KHz
- perflib belives the same state has frequency of 1266 MHz
- acpi_processor_ppc_notifier calls cpufreq_verify_within_limits to verify
that 1266768 is in the allowable range of 0 to 1266000 (i.e. 1266 * 1000)
- it's not, so that frequency is rejected
- the maximum CPU frequency is not reachable
This patch solves the problem by rounding up the MHz values stored in perflib's
tables. Additionally it corrects a broken URL.
It also fixes http://bugzilla.kernel.org/show_bug.cgi?id=8255 although this
case is a bit different: the frequencies in the _PSS tables are wildly wrong,
but we get better results if we force ACPI to respect the fsb * multiplier
calculations (even though it seems that the multiplier values aren't entirely
correct either).
Signed-off-by: Daniel Drake <[email protected]>
Signed-off-by: Dave Jones <[email protected]>
Signed-off-by: Chris Wright <[email protected]>
---
arch/i386/kernel/cpu/cpufreq/powernow-k7.c | 36 ++++++++++++++++++++++-------
1 file changed, 28 insertions(+), 8 deletions(-)
--- linux-2.6.21.1.orig/arch/i386/kernel/cpu/cpufreq/powernow-k7.c
+++ linux-2.6.21.1/arch/i386/kernel/cpu/cpufreq/powernow-k7.c
@@ -341,15 +341,17 @@ static int powernow_acpi_init(void)
pc.val = (unsigned long) acpi_processor_perf->states[0].control;
for (i = 0; i < number_scales; i++) {
u8 fid, vid;
- unsigned int speed;
+ struct acpi_processor_px *state =
+ &acpi_processor_perf->states[i];
+ unsigned int speed, speed_mhz;
- pc.val = (unsigned long) acpi_processor_perf->states[i].control;
+ pc.val = (unsigned long) state->control;
dprintk ("acpi: P%d: %d MHz %d mW %d uS control %08x SGTC %d\n",
i,
- (u32) acpi_processor_perf->states[i].core_frequency,
- (u32) acpi_processor_perf->states[i].power,
- (u32) acpi_processor_perf->states[i].transition_latency,
- (u32) acpi_processor_perf->states[i].control,
+ (u32) state->core_frequency,
+ (u32) state->power,
+ (u32) state->transition_latency,
+ (u32) state->control,
pc.bits.sgtc);
vid = pc.bits.vid;
@@ -360,6 +362,18 @@ static int powernow_acpi_init(void)
powernow_table[i].index |= (vid << 8); /* upper 8 bits */
speed = powernow_table[i].frequency;
+ speed_mhz = speed / 1000;
+
+ /* processor_perflib will multiply the MHz value by 1000 to
+ * get a KHz value (e.g. 1266000). However, powernow-k7 works
+ * with true KHz values (e.g. 1266768). To ensure that all
+ * powernow frequencies are available, we must ensure that
+ * ACPI doesn't restrict them, so we round up the MHz value
+ * to ensure that perflib's computed KHz value is greater than
+ * or equal to powernow's KHz value.
+ */
+ if (speed % 1000 > 0)
+ speed_mhz++;
if ((fid_codes[fid] % 10)==5) {
if (have_a0 == 1)
@@ -368,10 +382,16 @@ static int powernow_acpi_init(void)
dprintk (" FID: 0x%x (%d.%dx [%dMHz]) "
"VID: 0x%x (%d.%03dV)\n", fid, fid_codes[fid] / 10,
- fid_codes[fid] % 10, speed/1000, vid,
+ fid_codes[fid] % 10, speed_mhz, vid,
mobile_vid_table[vid]/1000,
mobile_vid_table[vid]%1000);
+ if (state->core_frequency != speed_mhz) {
+ state->core_frequency = speed_mhz;
+ dprintk(" Corrected ACPI frequency to %d\n",
+ speed_mhz);
+ }
+
if (latency < pc.bits.sgtc)
latency = pc.bits.sgtc;
@@ -602,7 +622,7 @@ static int __init powernow_cpu_init (str
result = powernow_acpi_init();
if (result) {
printk (KERN_INFO PFX "ACPI and legacy methods failed\n");
- printk (KERN_INFO PFX "See http://www.codemonkey.org.uk/projects/cpufreq/powernow-k7.shtml\n");
+ printk (KERN_INFO PFX "See http://www.codemonkey.org.uk/projects/cpufreq/powernow-k7.html\n");
}
} else {
/* SGTC use the bus clock as timer */
Patches currently in stable-queue which might be from [email protected] are
queue-2.6.21/cpufreq-powernow-k7-fix-mhz-rounding-issue-with-perflib.patch
queue-2.6.21/jfs-fix-race-waking-up-jfsio-kernel-thread.patch