Andrew,
This patch updates the performance-monitoring counters extensions'
ppc32 low-level driver as follows:
- Replace ugly /proc/cpuinfo parsing code (for detecting core
clock frequency) with code that asks Open Firmware instead.
This way we also get the important timebase frequency.
- Using new OF interface, detect and support generic chips having
a timebase register. This is useful because that's enough for
user-space to implement hrvtime().
- Using new OF interface, finally be able to derive the correct
timebase-to-core multiplier on old 604 (not 604e) chips.
- Avoid having to do partial processor detection in ppc_tests.c.
- Some code cleanups.
Signed-off-by: Mikael Pettersson <[email protected]>
diff -ruN linux-2.6.7-mm1/drivers/perfctr/ppc.c linux-2.6.7-mm1.perfctr-ppc32-update/drivers/perfctr/ppc.c
--- linux-2.6.7-mm1/drivers/perfctr/ppc.c 2004-06-21 19:33:30.000000000 +0200
+++ linux-2.6.7-mm1.perfctr-ppc32-update/drivers/perfctr/ppc.c 2004-06-21 21:13:57.685020000 +0200
@@ -8,8 +8,7 @@
#include <linux/sched.h>
#include <linux/fs.h>
#include <linux/perfctr.h>
-#include <linux/seq_file.h>
-#include <asm/machdep.h>
+#include <asm/prom.h>
#include <asm/time.h> /* tb_ticks_per_jiffy, get_tbl() */
#include "ppc_tests.h"
@@ -23,6 +22,7 @@
unsigned int ppc_mmcr[3];
} ____cacheline_aligned;
static DEFINE_PER_CPU(struct per_cpu_cache, per_cpu_cache);
+#define get_cpu_cache() __get_cpu_var(per_cpu_cache)
/* Structure for counter snapshots, as 32-bit values. */
struct perfctr_low_ctrs {
@@ -31,11 +31,12 @@
};
enum pm_type {
- PM_604,
- PM_604e,
- PM_750, /* XXX: Minor event set diffs between IBM and Moto. */
- PM_7400,
- PM_7450,
+ PM_NONE,
+ PM_604,
+ PM_604e,
+ PM_750, /* XXX: Minor event set diffs between IBM and Moto. */
+ PM_7400,
+ PM_7450,
};
static enum pm_type pm_type;
@@ -244,7 +245,7 @@
if (state->cstatus & (1<<30)) {
unsigned int mmcr0 = mfspr(SPRN_MMCR0);
state->ppc_mmcr[0] = mmcr0;
- __get_cpu_var(per_cpu_cache).ppc_mmcr[0] = mmcr0;
+ get_cpu_cache().ppc_mmcr[0] = mmcr0;
}
}
@@ -276,9 +277,10 @@
case PM_750:
case PM_604e:
return 4;
- default: /* impossible, but silences gcc warning */
case PM_604:
return 2;
+ default: /* PM_NONE, but silences gcc warning */
+ return 0;
}
}
@@ -365,7 +367,7 @@
struct per_cpu_cache *cache;
unsigned int value;
- cache = &__get_cpu_var(per_cpu_cache);
+ cache = &get_cpu_cache();
if (cache->k1.id == state->k1.id)
return;
/*
@@ -411,6 +413,8 @@
mtspr(SPRN_MMCR1, 0);
case PM_604:
mtspr(SPRN_MMCR0, 0);
+ case PM_NONE:
+ ;
}
switch (pm_type) {
case PM_7450:
@@ -424,6 +428,8 @@
case PM_604:
mtspr(SPRN_PMC2, 0);
mtspr(SPRN_PMC1, 0);
+ case PM_NONE:
+ ;
}
}
@@ -460,7 +466,7 @@
#ifdef CONFIG_SMP
clear_isuspend_cpu(state);
#else
- __get_cpu_var(per_cpu_cache).k1.id = 0;
+ get_cpu_cache().k1.id = 0;
#endif
}
@@ -603,7 +609,7 @@
{
struct per_cpu_cache *cache;
- cache = &__get_cpu_var(per_cpu_cache);
+ cache = &get_cpu_cache();
memset(cache, 0, sizeof *cache);
cache->k1.id = -1;
@@ -715,8 +721,7 @@
pll_cfg = (hid1 >> shift) & mask;
ratio = cfg_ratio[pll_cfg];
if (!ratio)
- printk(KERN_WARNING "perfctr/%s: unknown PLL_CFG 0x%x\n",
- __FILE__, pll_cfg);
+ printk(KERN_WARNING "perfctr: unknown PLL_CFG 0x%x\n", pll_cfg);
return (4/2) * ratio;
}
@@ -727,102 +732,59 @@
return tb_ticks_per_jiffy * tb_to_core * (HZ/10) / (1000/10);
}
-/* Extract the CPU clock frequency from /proc/cpuinfo. */
+/* Extract core and timebase frequencies from Open Firmware. */
-static unsigned int __init parse_clock_khz(struct seq_file *m)
+static unsigned int __init of_to_core_khz(void)
{
- /* "/proc/cpuinfo" formats:
- *
- * "core clock\t: %d MHz\n" // 8260 (show_percpuinfo)
- * "clock\t\t: %ldMHz\n" // 4xx (show_percpuinfo)
- * "clock\t\t: %dMHz\n" // oak (show_percpuinfo)
- * "clock\t\t: %ldMHz\n" // prep (show_percpuinfo)
- * "clock\t\t: %dMHz\n" // pmac (show_percpuinfo)
- * "clock\t\t: %dMHz\n" // gemini (show_cpuinfo!)
- */
- char *p;
- unsigned int mhz;
+ struct device_node *cpu;
+ unsigned int *fp, core, tb;
- p = m->buf;
- p[m->count] = '\0';
-
- for(;;) { /* for each line */
- if (strncmp(p, "core ", 5) == 0)
- p += 5;
- do {
- if (strncmp(p, "clock\t", 6) != 0)
- break;
- p += 6;
- while (*p == '\t')
- ++p;
- if (*p != ':')
- break;
- do {
- ++p;
- } while (*p == ' ');
- mhz = simple_strtoul(p, 0, 10);
- if (mhz)
- return mhz * 1000;
- } while (0);
- for(;;) { /* skip to next line */
- switch (*p++) {
- case '\n':
- break;
- case '\0':
- return 0;
- default:
- continue;
- }
- break;
- }
- }
+ cpu = find_type_devices("cpu");
+ if (!cpu)
+ return 0;
+ fp = (unsigned int*)get_property(cpu, "clock-frequency", NULL);
+ if (!fp || !(core = *fp))
+ return 0;
+ fp = (unsigned int*)get_property(cpu, "timebase-frequency", NULL);
+ if (!fp || !(tb = *fp))
+ return 0;
+ perfctr_info.tsc_to_cpu_mult = core / tb;
+ return core / 1000;
}
static unsigned int __init detect_cpu_khz(enum pll_type pll_type)
{
- char buf[512];
- struct seq_file m;
unsigned int khz;
khz = pll_to_core_khz(pll_type);
if (khz)
return khz;
- memset(&m, 0, sizeof m);
- m.buf = buf;
- m.size = (sizeof buf)-1;
-
- m.count = 0;
- if (ppc_md.show_percpuinfo != 0 &&
- ppc_md.show_percpuinfo(&m, 0) == 0 &&
- (khz = parse_clock_khz(&m)) != 0)
- return khz;
-
- m.count = 0;
- if (ppc_md.show_cpuinfo != 0 &&
- ppc_md.show_cpuinfo(&m) == 0 &&
- (khz = parse_clock_khz(&m)) != 0)
+ khz = of_to_core_khz();
+ if (khz)
return khz;
- printk(KERN_WARNING "perfctr/%s: unable to determine CPU speed\n",
- __FILE__);
+ printk(KERN_WARNING "perfctr: unable to determine CPU speed\n");
return 0;
}
-static int __init generic_init(void)
+static int __init known_init(void)
{
- static char generic_name[] __initdata = "PowerPC 60x/7xx/74xx";
+ static char known_name[] __initdata = "PowerPC 60x/7xx/74xx";
unsigned int features;
enum pll_type pll_type;
unsigned int pvr;
+ int have_mmcr1;
features = PERFCTR_FEATURE_RDTSC | PERFCTR_FEATURE_RDPMC;
+ have_mmcr1 = 1;
pvr = mfspr(SPRN_PVR);
switch (PVR_VER(pvr)) {
case 0x0004: /* 604 */
pm_type = PM_604;
pll_type = PLL_NONE;
features = PERFCTR_FEATURE_RDTSC;
+ have_mmcr1 = 0;
break;
case 0x0009: /* 604e; */
case 0x000A: /* 604ev */
@@ -860,15 +822,29 @@
pll_type = PLL_7457;
break;
default:
- printk(KERN_WARNING "perfctr/%s: unknown PowerPC with "
- "PVR 0x%08x -- bailing out\n", __FILE__, pvr);
return -ENODEV;
}
perfctr_info.cpu_features = features;
perfctr_info.cpu_type = 0; /* user-space should inspect PVR */
- perfctr_cpu_name = generic_name;
+ perfctr_cpu_name = known_name;
perfctr_info.cpu_khz = detect_cpu_khz(pll_type);
- perfctr_ppc_init_tests();
+ perfctr_ppc_init_tests(have_mmcr1);
+ return 0;
+}
+
+static int __init unknown_init(void)
+{
+ static char unknown_name[] __initdata = "Generic PowerPC with TB";
+ unsigned int khz;
+
+ khz = detect_cpu_khz(PLL_NONE);
+ if (!khz)
+ return -ENODEV;
+ perfctr_info.cpu_features = PERFCTR_FEATURE_RDTSC;
+ perfctr_info.cpu_type = 0;
+ perfctr_cpu_name = unknown_name;
+ perfctr_info.cpu_khz = khz;
+ pm_type = PM_NONE;
return 0;
}
@@ -896,9 +872,12 @@
perfctr_info.cpu_features = 0;
- err = generic_init();
- if (err)
- goto out;
+ err = known_init();
+ if (err) {
+ err = unknown_init();
+ if (err)
+ goto out;
+ }
perfctr_cpu_init_one(NULL);
smp_call_function(perfctr_cpu_init_one, NULL, 1, 1);
diff -ruN linux-2.6.7-mm1/drivers/perfctr/ppc_tests.c linux-2.6.7-mm1.perfctr-ppc32-update/drivers/perfctr/ppc_tests.c
--- linux-2.6.7-mm1/drivers/perfctr/ppc_tests.c 2004-06-21 19:33:30.000000000 +0200
+++ linux-2.6.7-mm1.perfctr-ppc32-update/drivers/perfctr/ppc_tests.c 2004-06-21 21:13:54.085020000 +0200
@@ -280,7 +280,7 @@
check_trigger(1);
}
-void __init perfctr_ppc_init_tests(void)
+void __init perfctr_ppc_init_tests(int have_mmcr1)
{
- measure_overheads(PVR_VER(mfspr(SPRN_PVR)) != 0x0004);
+ measure_overheads(have_mmcr1);
}
diff -ruN linux-2.6.7-mm1/drivers/perfctr/ppc_tests.h linux-2.6.7-mm1.perfctr-ppc32-update/drivers/perfctr/ppc_tests.h
--- linux-2.6.7-mm1/drivers/perfctr/ppc_tests.h 2004-06-21 19:33:30.000000000 +0200
+++ linux-2.6.7-mm1.perfctr-ppc32-update/drivers/perfctr/ppc_tests.h 2004-06-21 21:13:54.085020000 +0200
@@ -6,7 +6,7 @@
*/
#ifdef CONFIG_PERFCTR_INIT_TESTS
-extern void perfctr_ppc_init_tests(void);
+extern void perfctr_ppc_init_tests(int have_mmcr1);
#else
-#define perfctr_ppc_init_tests()
+static inline void perfctr_ppc_init_tests(int have_mmcr1) { }
#endif
diff -ruN linux-2.6.7-mm1/include/asm-ppc/perfctr.h linux-2.6.7-mm1.perfctr-ppc32-update/include/asm-ppc/perfctr.h
--- linux-2.6.7-mm1/include/asm-ppc/perfctr.h 2004-06-21 19:33:30.000000000 +0200
+++ linux-2.6.7-mm1.perfctr-ppc32-update/include/asm-ppc/perfctr.h 2004-06-21 21:13:54.085020000 +0200
@@ -7,6 +7,7 @@
#define _ASM_PPC_PERFCTR_H
/* perfctr_info.cpu_type values */
+#define PERFCTR_PPC_GENERIC 0
#define PERFCTR_PPC_604 1
#define PERFCTR_PPC_604e 2
#define PERFCTR_PPC_750 3
Hrm... your code will not work with externally clocked timebases
(like the G5) and I'm not sure you get the core freq. right with
CPU that can do clock slewing or machines that can switch the
core/bus ratio (laptops).
We should rather define an arch API to return those infos...
Ben.
Benjamin Herrenschmidt writes:
> Hrm... your code will not work with externally clocked timebases
> (like the G5) and I'm not sure you get the core freq. right with
> CPU that can do clock slewing or machines that can switch the
> core/bus ratio (laptops).
Do you mean the PLL_CFG code that's been in -mm for the last couple
of weeks, or just the recently posted update? The update replaced
in-kernel /proc/cpuinfo parsing (gross) with OF queries taken straight
from the pmac code in arch/ppc/platform/.
I'm ignoring 970/G5 until IBM releases the damn documentation.
> We should rather define an arch API to return those infos...
No argument there.
On Tue, 2004-06-22 at 15:09, Mikael Pettersson wrote:
> Benjamin Herrenschmidt writes:
> > Hrm... your code will not work with externally clocked timebases
> > (like the G5) and I'm not sure you get the core freq. right with
> > CPU that can do clock slewing or machines that can switch the
> > core/bus ratio (laptops).
>
> Do you mean the PLL_CFG code that's been in -mm for the last couple
> of weeks, or just the recently posted update? The update replaced
> in-kernel /proc/cpuinfo parsing (gross) with OF queries taken straight
> from the pmac code in arch/ppc/platform/.
>
> I'm ignoring 970/G5 until IBM releases the damn documentation.
Well, the G5 can have it's own tb but can also be externally clocked and
that's how Apple does. I'm not sure about all G4 models.
> > We should rather define an arch API to return those infos...
>
> No argument there.
--
Benjamin Herrenschmidt <[email protected]>
Benjamin Herrenschmidt writes:
> On Tue, 2004-06-22 at 15:09, Mikael Pettersson wrote:
> > Benjamin Herrenschmidt writes:
> > > Hrm... your code will not work with externally clocked timebases
> > > (like the G5) and I'm not sure you get the core freq. right with
> > > CPU that can do clock slewing or machines that can switch the
> > > core/bus ratio (laptops).
> >
> > Do you mean the PLL_CFG code that's been in -mm for the last couple
> > of weeks, or just the recently posted update? The update replaced
> > in-kernel /proc/cpuinfo parsing (gross) with OF queries taken straight
> > from the pmac code in arch/ppc/platform/.
> >
> > I'm ignoring 970/G5 until IBM releases the damn documentation.
>
> Well, the G5 can have it's own tb but can also be externally clocked and
> that's how Apple does. I'm not sure about all G4 models.
So what you're saying is that PLL_CFG may not reflect the true
relationship between the TB frequency and the core frequency?
That shouldn't be a problem as long as there's _some_ in-kernel
interface for finding that out. If querying OF isn't the correct
approach, then what is?
/Mikael
> So what you're saying is that PLL_CFG may not reflect the true
> relationship between the TB frequency and the core frequency?
Right.
> That shouldn't be a problem as long as there's _some_ in-kernel
> interface for finding that out. If querying OF isn't the correct
> approach, then what is?
What do you need exactly ? The TB one or the core one ? the core
I suppose ? Well, we should probably define an ppc_get_cpu_core_frequency
or something like that that uses the cpufreq callback like the pmac code
when cpufreq is enabled or default to the old parsing when not. Look at
the pmac code. You may also want to install a cpufreq notifier callback
to be informed of core frequency changes.
Ben.
Benjamin Herrenschmidt writes:
>
> > So what you're saying is that PLL_CFG may not reflect the true
> > relationship between the TB frequency and the core frequency?
>
> Right.
>
> > That shouldn't be a problem as long as there's _some_ in-kernel
> > interface for finding that out. If querying OF isn't the correct
> > approach, then what is?
>
> What do you need exactly ? The TB one or the core one ? the core
> I suppose ? Well, we should probably define an ppc_get_cpu_core_frequency
> or something like that that uses the cpufreq callback like the pmac code
> when cpufreq is enabled or default to the old parsing when not. Look at
> the pmac code. You may also want to install a cpufreq notifier callback
> to be informed of core frequency changes.
I want both TB and core speeds, but TB is more important.
Originally, I used and sampled the x86 time-stamp counter to
provide an additional clock-like register for measurements.
On x86, observable TSC freq == core freq, so the CPU speed
I report to users coincides with TSC speed. This is then
used to derive high-resolution actual time from TSC counts.
When adding PPC32 support, I used TB for the clock-like
entity, but had to introduce a TB-to-core multiplier to
maintain the notion that clock*multiplier == core speed.
If TB speed is constant but core speed is not, then this
was a mistake and I should instead report TB speed only,
and let user-space convert TB counts to time directly
without conversion via core speed.
So can I assume constant TB speed? In that case I don't
really care about core speed changes.
/Mikael
On Tue, 2004-06-22 at 17:18, Mikael Pettersson wrote:
> So can I assume constant TB speed? In that case I don't
> really care about core speed changes.
So far, tb speed is constant yes, I haven't seen a case where it's
not. What we should really do is to expose the tb speed to userspace
generically on ppc, possibly via /proc/cpuinfo like we do on ppc64,
it definitely has other uses than just profiling.
Ben.