Skip to contentSkip to conversation

2.6.5 speedstep on P4Ms

PATCH In review Christian Hoelbling 5 participants 21 messages Jun 8, 2004 – Jun 11, 2004

lore
Share & tools
lore
#1 lore raw
In reply to Christian Hoelbling
Quoted 9 lines · Christian Hoelbling

hi list,

i am trying to do a small rewrite of the speedstep-ich driver to address the following issues:

1.) detect all P4M's via the model_id string
2.) correctly register drivers on hyperthreading CPU's (and stop freeze-ups i occasionally had)
3.) do P4-clockmod on top of speedstep on P4-Ms

It would be great if you send 3 differents patches...

Also, I don't think 3) is how thinks should be done.
It would be much more usefull to allow 2 cpufreq drivers to be loaded. I think it's planed for 2.7 btw.

Cheers,

Signature · 4 lines
Bruno Ducrot

--  Which is worse:  ignorance or apathy?
--  Don't know.  Don't care.
#0 lore raw

PATCH 1/1 2.6.5 speedstep on P4Ms

hi list,

i am trying to do a small rewrite of the speedstep-ich driver to address the following issues:

1.) detect all P4M's via the model_id string
2.) correctly register drivers on hyperthreading CPU's (and stop freeze-ups i occasionally had) 3.) do P4-clockmod on top of speedstep on P4-Ms

the patched driver works on my machine, but there seem to be 2 problems: sometimes the clockmod state is not set correctly and the powersave governor seems to miss the lowest frequency state (it instead switches to the second lowest).

it would be great if someone could test it on there machines and send me some reports. also any suggestions/tips are very welcome (since this is my first attempt at kernel coding).

best,

Christian Hoelbling

arch/i386/kernel/cpu/cpufreq/speedstep-ich.c +333 -5
diff --unified --recursive --new-file linux-2.6.5/arch/i386/kernel/cpu/cpufreq/speedstep-ich.c linux-2.6.5.speedstep/arch/i386/kernel/cpu/cpufreq/speedstep-ich.c
@@ -11,6 +11,9 @@
  *  for extensive testing.
  *
  *  BIG FAT DISCLAIMER: Work in progress code. Possibly *dangerous*
+ *
+ *  Added SMT and p4-clockmod support on P4-M's
+ *  Christian Hoelbling, 2004
  */
 
 
@@ -27,6 +30,10 @@
 
 #include "speedstep-lib.h"
 
+/* invalidate high voltage states slower than fastest
+ * low voltage state
+ */
+#define INVALIDATE_SLOW_HI_STATES
 
 /* speedstep_chipset:
  *   It is necessary to know which chipset is used. As accesses to 
@@ -51,6 +58,44 @@
 	{0,			CPUFREQ_TABLE_END},
 };
 
+/*
+ * Duty Cycle (3bits) from p4-clockmod
+ * for each of the voltages
+ */
+enum {
+        DC_RESV, LO_DC_DFLT, LO_DC_25PT, LO_DC_38PT, LO_DC_50PT,
+	LO_DC_64PT, LO_DC_75PT, LO_DC_88PT, LO_DC_DISABLE,
+        HI_DC_DFLT, HI_DC_25PT, HI_DC_38PT, HI_DC_50PT,
+	HI_DC_64PT, HI_DC_75PT, HI_DC_88PT, HI_DC_DISABLE
+};
+
+/* 
+ * Extended speedstep + p4-clockmod table
+ * for P4-M's
+ * Values are in kHz for the time being.
+ */
+static struct cpufreq_frequency_table speedstep_p4_freqs[] = {
+	{DC_RESV, CPUFREQ_ENTRY_INVALID},
+	{LO_DC_DFLT, 0},
+	{LO_DC_25PT, 0},
+	{LO_DC_38PT, 0},
+	{LO_DC_50PT, 0},
+	{LO_DC_64PT, 0},
+	{LO_DC_75PT, 0},
+	{LO_DC_88PT, 0},
+	{LO_DC_DISABLE, 0},
+	{HI_DC_DFLT, 0},
+	{HI_DC_25PT, 0},
+	{HI_DC_38PT, 0},
+	{HI_DC_50PT, 0},
+	{HI_DC_64PT, 0},
+	{HI_DC_75PT, 0},
+	{HI_DC_88PT, 0},
+	{HI_DC_DISABLE, 0},
+	{DC_RESV, CPUFREQ_TABLE_END},
+};
+
+static unsigned int stock_freq[2];
 
 /* DEBUG
  *   Define it if you want verbose debug output, e.g. for bug reporting
@@ -149,6 +194,190 @@
 }
 
 
+static int speedstep_p4_set_state(unsigned int cpu, unsigned int newstate, unsigned int notify)
+{
+	u32 l, h, pmbase;
+	u8 pm2_blk, value;
+	cpumask_t cpus_allowed, affected_cpu_map;
+	unsigned long flags;
+	struct cpufreq_freqs freqs;
+	int hyperthreading = 0;
+	int sibling = 0;
+	unsigned int dc_newstate,speedstep_newstate,oldstate,dc_oldstate,speedstep_oldstate;
+	unsigned int speedstep_trans,dc_trans;
+
+	if (!cpu_online(cpu) || (newstate > HI_DC_DISABLE) || 
+		(newstate == DC_RESV))
+		return -EINVAL;
+
+	/* disentangle speedstep and clockmod */
+	dc_newstate=((newstate-1) & 0x07)+1;
+	speedstep_newstate=(newstate>LO_DC_DISABLE);
+
+	/* switch to physical CPU where state is to be changed*/
+	cpus_allowed = current->cpus_allowed;
+
+	/* only run on CPU to be set, or on its sibling */
+       affected_cpu_map = cpumask_of_cpu(cpu);
+#ifdef CONFIG_X86_HT
+	hyperthreading = ((cpu_has_ht) && (smp_num_siblings == 2));
+	if (hyperthreading) {
+		sibling = cpu_sibling_map[cpu];
+                cpu_set(sibling, affected_cpu_map);
+	}
+#endif
+	set_cpus_allowed(current, affected_cpu_map);
+        BUG_ON(!cpu_isset(smp_processor_id(), affected_cpu_map));
+
+	/* get current clockmod state */
+	rdmsr(MSR_IA32_THERM_CONTROL, l, h);
+	dprintk(KERN_INFO "cpufreq: CPU#%d read 0x%x from control reg.0x%x\n", cpu, l, h); 
+	if (l & 0x10) {
+		dc_oldstate = l >> 1;
+		dc_oldstate &= 0x7;
+	} else
+	        dc_oldstate = 0x08; /* dc disabled */
+
+	if (dc_oldstate == DC_RESV) {
+	        printk(KERN_ERR "cpufreq: BIG FAT WARNING: currently in invalid setting\n");
+	}
+	dc_trans=(dc_oldstate != dc_newstate);
+
+	/* get currect speedstep state */
+	speedstep_oldstate=(speedstep_get_processor_frequency(speedstep_processor)==speedstep_p4_freqs[HI_DC_DISABLE].frequency);
+	speedstep_trans=(speedstep_oldstate != speedstep_newstate);
+	oldstate  = ((dc_oldstate-1) + (speedstep_oldstate << 3)) + 1;
+
+	if (notify) {
+	        freqs.old = speedstep_p4_freqs[oldstate].frequency; /* relies on indices for all allowed states in frequency table being in sequential order starting from 1 */
+		freqs.cpu = cpu;
+		dprintk(KERN_DEBUG "cpufreq: os: 0x%x ns: 0x%x ssos: 0x%x ssns: 0x%x dcos: 0x%x dcns: 0x%x\n", oldstate,newstate,speedstep_oldstate,speedstep_newstate,dc_oldstate,dc_newstate);
+
+		/* notifiers */
+		freqs.new = speedstep_p4_freqs[newstate].frequency; /* relies on indices for all allowed states in frequency table being in sequential order starting from 1 */
+		dprintk(KERN_DEBUG "cpufreq: preparing transition on cpu %i from %i kHz to %ikHz \n",cpu ,freqs.old ,freqs.new);
+		cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
+		if (hyperthreading) {
+		        freqs.cpu = sibling;
+			cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
+		}
+	}
+
+	/* speedstep state transition */
+	if (speedstep_trans) {
+
+	        /* get PMBASE */
+	        pci_read_config_dword(speedstep_chipset_dev, 0x40, &pmbase);
+		if (!(pmbase & 0x01)) {
+		        printk(KERN_ERR "cpufreq: could not find speedstep register\n");
+			return 0;
+		}
+		
+		pmbase &= 0xFFFFFFFE;
+		if (!pmbase) {
+		       printk(KERN_ERR "cpufreq: could not find speedstep register\n");
+		       return 0;
+		}
+		
+		/* Disable IRQs */
+		local_irq_save(flags);
+
+		/* read state */
+		value = inb(pmbase + 0x50);
+
+		dprintk(KERN_DEBUG "cpufreq: read at pmbase 0x%x + 0x50 returned 0x%x\n", pmbase, value);
+
+		/* write new state */
+		value &= 0xFE;
+		value |= 1-speedstep_newstate;
+		
+		dprintk(KERN_DEBUG "cpufreq: writing 0x%x to pmbase 0x%x + 0x50\n", value, pmbase);
+
+		/* Disable bus master arbitration */
+		pm2_blk = inb(pmbase + 0x20);
+		pm2_blk |= 0x01;
+		outb(pm2_blk, (pmbase + 0x20));
+
+		/* Actual transition */
+		outb(value, (pmbase + 0x50));
+		
+		/* Restore bus master arbitration */
+		pm2_blk &= 0xfe;
+		outb(pm2_blk, (pmbase + 0x20));
+
+		/* check if transition was successful */
+		value = inb(pmbase + 0x50);
+
+		/* Enable IRQs */
+		local_irq_restore(flags);
+
+		dprintk(KERN_DEBUG "cpufreq: read at pmbase 0x%x + 0x50 returned 0x%x\n", pmbase, value);
+
+		if (speedstep_newstate != (value & 0x1)) {
+		        dprintk (KERN_INFO "cpufreq: change to %u kHz speedstep state succeeded\n", (speedstep_get_processor_frequency(speedstep_processor)));
+		} else {
+		        printk (KERN_ERR "cpufreq: change failed - I/O error\n");
+		}
+	}
+
+	/* clockmod transition */
+	if (dc_trans) {
+
+	        rdmsr(MSR_IA32_THERM_STATUS, l, h);
+#if 0
+		if (l & 0x01)
+		        dprintk(KERN_DEBUG "cpufreq: CPU#%d currently thermal throttled\n", cpu);
+#endif
+		rdmsr(MSR_IA32_THERM_CONTROL, l, h);
+		if (dc_newstate == 0x08) {
+		        dprintk(KERN_INFO "cpufreq: CPU#%d disabling modulation\n", cpu); 
+		        dprintk(KERN_INFO "cpufreq: CPU#%d writing 0x%x to control reg.\n", cpu, (l | 0x0e)& ~(1<<4)); 
+			wrmsr(MSR_IA32_THERM_CONTROL, (l | 0x0e) & ~(1<<4), h);
+		} else {
+      		        dprintk(KERN_INFO "cpufreq: CPU#%d setting duty cycle to %d%%\n",
+		        cpu, ((125 * dc_newstate) / 10)); 
+	              /* bits 63 - 5	: reserved 
+		       * bit  4	: enable/disable
+		       * bits 3-1	: duty cycle
+		       * bit  0	: reserved
+		       */
+			l = (l & ~14);
+			l = l | (1<<4) | ((dc_newstate & 0x7)<<1);
+		        dprintk(KERN_INFO "cpufreq: CPU#%d writing 0x%x to control reg.0x%x\n", cpu, l, h); 
+			wrmsr(MSR_IA32_THERM_CONTROL, l, h);
+		}
+
+#ifdef SPEEDSTEP_DEBUG
+		/* get current clockmod state */
+		rdmsr(MSR_IA32_THERM_CONTROL, l, h);
+		dprintk(KERN_INFO "cpufreq: CPU#%d read 0x%x from control reg.0x%x\n", cpu, l, h); 
+		if (l & 0x10) {
+		        l = l >> 1;
+			l &= 0x7;
+		} else
+		        l = 0x08; /* dc disabled */
+		dprintk(KERN_DEBUG "cpufreq: new dc state: 0x%x\n", l);
+#endif
+
+	}
+
+	set_cpus_allowed(current, cpus_allowed);
+
+	if (notify) {
+	        /* notifiers */
+	        cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
+	        if (hyperthreading) {
+		        freqs.cpu = cpu;
+			cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
+		}
+		dprintk(KERN_DEBUG "cpufreq: transition on cpu %i from %i kHz to %ikHz completed\n",cpu ,freqs.old ,freqs.new);
+		
+	}
+
+	return 0;
+}
+
+
 /**
  * speedstep_activate - activate SpeedStep control in the chipset
  *
@@ -164,6 +393,7 @@
 
 	pci_read_config_word(speedstep_chipset_dev, 
 			     0x00A0, &value);
+	dprintk(KERN_DEBUG "cpufreq: speedstep registers: 0x%x \n",value);
 	if (!(value & 0x08)) {
 		value |= 0x08;
 		dprintk(KERN_DEBUG "cpufreq: activating SpeedStep (TM) registers\n");
@@ -259,6 +489,21 @@
 }
 
 
+static int speedstep_p4_target (struct cpufreq_policy *policy,
+			     unsigned int target_freq,
+			     unsigned int relation)
+{
+	unsigned int	newstate = DC_RESV;
+
+	if (cpufreq_frequency_table_target(policy, &speedstep_p4_freqs[0], target_freq, relation, &newstate))
+		return -EINVAL;
+
+	speedstep_p4_set_state(policy->cpu, speedstep_p4_freqs[newstate].index, 1);
+
+	return 0;
+}
+
+
 /**
  * speedstep_verify - verifies a new CPUFreq policy
  * @freq: new policy
@@ -272,6 +517,12 @@
 }
 
 
+static int speedstep_p4_verify (struct cpufreq_policy *policy)
+{
+	return cpufreq_frequency_table_verify(policy, &speedstep_p4_freqs[0]);
+}
+
+
 static int speedstep_cpu_init(struct cpufreq_policy *policy)
 {
 	int		result = 0;
@@ -312,6 +563,67 @@
 	return 0;
 }
 
+static int speedstep_P4_cpu_init(struct cpufreq_policy *policy)
+{
+	int		result = 0;
+	unsigned int    i,max,min,max_to_min_ratio,idx_lo;
+
+	/* detect low and high frequency */
+	result=speedstep_p4_set_state(policy->cpu, speedstep_p4_freqs[LO_DC_DISABLE].index, 0);
+	stock_freq[0]=speedstep_get_processor_frequency(speedstep_processor);
+	if (result)
+		return result;
+	result=speedstep_p4_set_state(policy->cpu, speedstep_p4_freqs[HI_DC_DISABLE].index, 0);
+	stock_freq[1]=speedstep_get_processor_frequency(speedstep_processor);
+	if (result)
+	        return result;
+
+	speedstep_p4_freqs[LO_DC_DISABLE].frequency=stock_freq[0];
+	speedstep_p4_freqs[HI_DC_DISABLE].frequency=stock_freq[1];
+
+	/* init low and high clockmod tables */
+	for (i=1; i<8; i++) {
+	        speedstep_p4_freqs[i].frequency = (speedstep_p4_freqs[LO_DC_DISABLE].frequency * i)/8;
+	        speedstep_p4_freqs[8+i].frequency = (speedstep_p4_freqs[HI_DC_DISABLE].frequency * i)/8;
+	}
+
+	/* check, if there are frequencies which are possible in both voltage states */
+	max=stock_freq[1];
+	min=stock_freq[0]/8;
+	max_to_min_ratio=max/min;
+	if ((min*max_to_min_ratio)==max) {
+	       dprintk(KERN_INFO "cpufreq: non-unique frequencies may appear\n");
+	       /* if yes, subtract 1 from higher voltage frequency - ugly but effective */
+	       for (i=1; i<9; i++) {
+		       idx_lo=(max*i)/(8*min);
+		       if ((idx_lo<=8) && (speedstep_p4_freqs[idx_lo].frequency == speedstep_p4_freqs[8+i].frequency)) 
+		               speedstep_p4_freqs[8+i].frequency--;
+	       }
+	}
+
+#ifdef INVALIDATE_SLOW_HI_STATES
+
+	/*  invalidate high voltage states slower than fastest low voltage state */
+	for (i=9; (speedstep_p4_freqs[i].frequency<stock_freq[0]) && (speedstep_p4_freqs[i].frequency!=CPUFREQ_TABLE_END) ; i++) {
+	        dprintk(KERN_INFO "cpufreq: invalidating mode %i with %i kHz\n", 
+			i,speedstep_p4_freqs[i].frequency);	        
+		speedstep_p4_freqs[i].frequency=CPUFREQ_ENTRY_INVALID;
+	}
+
+#endif
+
+	dprintk(KERN_INFO "cpufreq: detected mobile P4 currently at %i MHz\n", 
+		(stock_freq[1] / 1000));
+
+        cpufreq_frequency_table_get_attr(speedstep_p4_freqs, policy->cpu);
+
+	/* cpuinfo and default policy values */
+	policy->governor = CPUFREQ_DEFAULT_GOVERNOR;
+	policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL+1000000; /* assumed */
+	policy->cur = stock_freq[1];
+
+	return cpufreq_frequency_table_cpuinfo(policy, &speedstep_p4_freqs[0]);
+}
 
 static int speedstep_cpu_exit(struct cpufreq_policy *policy)
 {
@@ -336,6 +648,16 @@
 	.attr		= speedstep_attr,
 };
 
+static struct cpufreq_driver speedstep_P4_driver = {
+	.name		= "speedstep-ich",
+	.verify 	= speedstep_p4_verify,
+	.target 	= speedstep_p4_target,
+	.init		= speedstep_P4_cpu_init,
+	.exit		= speedstep_cpu_exit,    /* same as speedstep only */
+	.owner		= THIS_MODULE,
+	.attr		= speedstep_attr,        /* dummy - use same as speedstep only */
+};
+
 
 /**
  * speedstep_init - initializes the SpeedStep CPUFreq driver
@@ -359,11 +681,14 @@
 
 	/* activate speedstep support */
 	if (speedstep_activate())
-		return -EINVAL;
+	        return -EINVAL;
 
-	return cpufreq_register_driver(&speedstep_driver);
-}
+	if (speedstep_processor==SPEEDSTEP_PROCESSOR_P4M)
+	        return cpufreq_register_driver(&speedstep_P4_driver);
+	else
+	        return cpufreq_register_driver(&speedstep_driver);
 
+}
 
 /**
  * speedstep_exit - unregisters SpeedStep support
@@ -372,11 +697,14 @@
  */
 static void __exit speedstep_exit(void)
 {
-	cpufreq_unregister_driver(&speedstep_driver);
+	if (speedstep_processor==SPEEDSTEP_PROCESSOR_P4M)
+	        cpufreq_unregister_driver(&speedstep_P4_driver);
+	else
+	        cpufreq_unregister_driver(&speedstep_driver);
 }
 
 
-MODULE_AUTHOR ("Dave Jones <[email protected]>, Dominik Brodowski <[email protected]>");
+MODULE_AUTHOR ("Dave Jones <[email protected]>, Dominik Brodowski <[email protected]>, Christian Hoelbling <[email protected]>");
 MODULE_DESCRIPTION ("Speedstep driver for Intel mobile processors on chipsets with ICH-M southbridges.");
 MODULE_LICENSE ("GPL");
 
arch/i386/kernel/cpu/cpufreq/speedstep-lib.c +14 -4
diff --unified --recursive --new-file linux-2.6.5/arch/i386/kernel/cpu/cpufreq/speedstep-lib.c linux-2.6.5.speedstep/arch/i386/kernel/cpu/cpufreq/speedstep-lib.c
@@ -210,8 +210,17 @@
 		ebx = cpuid_ebx(0x00000001);
 		ebx &= 0x000000FF;
 
-		dprintk(KERN_INFO "ebx value is %x, x86_mask is %x\n", ebx, c->86_mask);
+		dprintk(KERN_INFO "ebx value is %x\n", ebx);
 
+		dprintk(KERN_INFO "model_id is %s\n", c->x86_model_id);
+		
+		/*
+		 * If the x86_model_id string contais "Mobile Intel(R) Pentium(R) 4"
+                 * omit all other checks and treat the CPU as a M-P4-M
+		 */
+		if (strstr(c->x86_model_id,"Mobile Intel(R) Pentium(R) 4") != NULL)
+		       return SPEEDSTEP_PROCESSOR_P4M;
+		
 		switch (c->x86_mask) {
 		case 4: 
 			/*
@@ -248,10 +257,11 @@
 			 * So, how to distinguish all those processors with
 			 * ebx=0xf? I don't know. Sort them out, and wait
 			 * for someone to complain.
+			 * also, M-P4M HTs actually have ebx=0x8, too
 			 */
-			if (ebx == 0x0e)
-				return SPEEDSTEP_PROCESSOR_P4M;
-			break;
+		         if (ebx == 0x0e)
+		                return SPEEDSTEP_PROCESSOR_P4M;
+			 break;
 		default:
 			break;
 		}

_______________________________________________
Cpufreq mailing list
[email protected]
http://www.linux.org.uk/mailman/listinfo/cpufreq

#2 lore raw
In reply to Bruno Ducrot
Quoted 13 lines · Bruno Ducrot
Quoted 9 lines · Christian Hoelbling

hi list,

i am trying to do a small rewrite of the speedstep-ich driver to address the following issues:

1.) detect all P4M's via the model_id string
2.) correctly register drivers on hyperthreading CPU's (and stop freeze-ups i occasionally had) 3.) do P4-clockmod on top of speedstep on P4-Ms

Quoted 10 lines

It would be great if you send 3 differents patches...

Also, I don't think 3) is how thinks should be done.
It would be much more usefull to allow 2 cpufreq drivers to be loaded. I think it's planed for 2.7 btw.

Cheers,

thanks alot for your help. here are the patches for 1) and 2) seperately.

for 3): what you say sounds much better than my brutal hack, so i won't post it again. i'd definitely want to try to implement 2 cpufreq drivers, but i fear this is way beyond me at the moment and probably involves design choices i can't make.

cheers,

chris

arch/i386/kernel/cpu/cpufreq/speedstep-lib.c +11 -1
diff --unified --recursive --new-file linux-2.6.5/arch/i386/kernel/cpu/cpufreq/speedstep-lib.c linux-2.6.5.speedstep/arch/i386/kernel/cpu/cpufreq/speedstep-lib.c
@@ -210,8 +210,17 @@
 		ebx = cpuid_ebx(0x00000001);
 		ebx &= 0x000000FF;
 
-		dprintk(KERN_INFO "ebx value is %x, x86_mask is %x\n", ebx, c->86_mask);
+		dprintk(KERN_INFO "ebx value is %x\n", ebx);
 
+		dprintk(KERN_INFO "model_id is %s\n", c->x86_model_id);
+		
+		/*
+		 * If the x86_model_id string contais "Mobile Intel(R) Pentium(R) 4"
+                 * omit all other checks and treat the CPU as a M-P4-M
+		 */
+		if (strstr(c->x86_model_id,"Mobile Intel(R) Pentium(R) 4") != NULL)
+		       return SPEEDSTEP_PROCESSOR_P4M;
+		
 		switch (c->x86_mask) {
 		case 4: 
 			/*
@@ -248,6 +257,7 @@
 			 * So, how to distinguish all those processors with
 			 * ebx=0xf? I don't know. Sort them out, and wait
 			 * for someone to complain.
+			 * also, M-P4M HTs actually have ebx=0x8, too
 			 */
 			if (ebx == 0x0e)
 				return SPEEDSTEP_PROCESSOR_P4M;
arch/i386/kernel/cpu/cpufreq/speedstep-ich.c +46 -12

97 lines not on this page · raw message

arch/i386/kernel/cpu/cpufreq/speedstep-lib.c +26 -5
diff --unified --recursive --new-file linux-2.6.5/arch/i386/kernel/cpu/cpufreq/speedstep-lib.c linux-2.6.5.speedstep/arch/i386/kernel/cpu/cpufreq/speedstep-lib.c
@@ -310,20 +310,40 @@
  *                     DETECT SPEEDSTEP SPEEDS                       *
  *********************************************************************/
 
-unsigned int speedstep_get_freqs(unsigned int processor,
+unsigned int speedstep_get_freqs(unsigned int cpu,
+				  unsigned int processor,
 				  unsigned int *low_speed,
 				  unsigned int *high_speed,
-				  void (*set_state) (unsigned int state,
+				  void (*set_state) (unsigned int cpu,
+						     unsigned int state,
 						     unsigned int notify)
 				 )
 {
 	unsigned int prev_speed;
 	unsigned int ret = 0;
 	unsigned long flags;
+	cpumask_t cpus_allowed, affected_cpu_map;
+	int hyperthreading = 0;
+	int sibling = 0;
 
 	if ((!processor) || (!low_speed) || (!high_speed) || (!set_state))
 		return -EINVAL;
 
+	/* switch to physical CPU where state is to be changed*/
+	cpus_allowed = current->cpus_allowed;
+
+	/* only run on CPU to be set, or on its sibling */
+       affected_cpu_map = cpumask_of_cpu(cpu);
+#ifdef CONFIG_X86_HT
+	hyperthreading = ((cpu_has_ht) && (smp_num_siblings == 2));
+	if (hyperthreading) {
+		sibling = cpu_sibling_map[cpu];
+                cpu_set(sibling, affected_cpu_map);
+	}
+#endif
+	set_cpus_allowed(current, affected_cpu_map);
+        BUG_ON(!cpu_isset(smp_processor_id(), affected_cpu_map));
+
 	/* get current speed */
 	prev_speed = speedstep_get_processor_frequency(processor);
 	if (!prev_speed)
@@ -332,7 +352,7 @@
 	local_irq_save(flags);
 
 	/* switch to low state */
-	set_state(SPEEDSTEP_LOW, 0);
+	set_state(cpu, SPEEDSTEP_LOW, 0);
 	*low_speed = speedstep_get_processor_frequency(processor);
 	if (!*low_speed) {
 		ret = -EIO;
@@ -340,7 +360,7 @@
 	}
 
 	/* switch to high state */
-	set_state(SPEEDSTEP_HIGH, 0);
+	set_state(cpu, SPEEDSTEP_HIGH, 0);
 	*high_speed = speedstep_get_processor_frequency(processor);
 	if (!*high_speed) {
 		ret = -EIO;
@@ -354,10 +374,11 @@
 
 	/* switch to previous state, if necessary */
 	if (*high_speed != prev_speed)
-		set_state(SPEEDSTEP_LOW, 0);
+		set_state(cpu, SPEEDSTEP_LOW, 0);
 
  out:
 	local_irq_restore(flags);
+	set_cpus_allowed(current, cpus_allowed);
 	return (ret);
 }
 EXPORT_SYMBOL_GPL(speedstep_get_freqs);
arch/i386/kernel/cpu/cpufreq/speedstep-lib.h +2 -2
diff --unified --recursive --new-file linux-2.6.5/arch/i386/kernel/cpu/cpufreq/speedstep-lib.h linux-2.6.5.speedstep/arch/i386/kernel/cpu/cpufreq/speedstep-lib.h
@@ -41,7 +41,7 @@
  * SPEEDSTEP_LOW; the second argument is zero so that no 
  * cpufreq_notify_transition calls are initiated.
  */
-extern unsigned int speedstep_get_freqs(unsigned int processor,
+extern unsigned int speedstep_get_freqs(unsigned int cpu, unsigned int processor,
 	  unsigned int *low_speed,
 	  unsigned int *high_speed,
-	  void (*set_state) (unsigned int state, unsigned int notify));
+	  void (*set_state) (unsigned int cpu, unsigned int state, unsigned int notify));

_______________________________________________
Cpufreq mailing list
[email protected]
http://www.linux.org.uk/mailman/listinfo/cpufreq

#3 lore raw

Hi Chris,

Quoted 4 lines

for 3): what you say sounds much better than my brutal hack, so i won't post it again. i'd definitely want to try to implement 2 cpufreq drivers, but i fear this is way beyond me at the moment and probably involves design choices i can't make.

well, somebody _does_ have to do design choices... However, I'm not yet convinced of the necessity of supporting throttling and frequency scaling on the same system. Even more, under normal circumstances [i.e. idling works, which it does on most systems], I see no _technical_ reason to do throttling, or even non-temperature related _dynamic throttling_ like other OSes do it. For those new to this list, there's some maths in the archives which proves that you need more energy == battery power if you do throttling for any given task... So it's something we need to discuss during 2.7.

Now, about your patch 1):

Quoted 9 lines

diff --unified --recursive --new-file linux-2.6.5/arch/i386/kernel/cpu/cpufreq/speedstep-lib.c linux-2.6.5.speedstep/arch/i386/kernel/cpu/cpufreq/speedstep-lib.c

arch/i386/kernel/cpu/cpufreq/speedstep-lib.c +1 -1
@@ -210,8 +210,17 @@
 		ebx = cpuid_ebx(0x00000001);
 		ebx &= 0x000000FF;
 
-		dprintk(KERN_INFO "ebx value is %x, x86_mask is %x\n", ebx, c->86_mask);
+		dprintk(KERN_INFO "ebx value is %x\n", ebx);

arch/i386/kernel/cpu/cpufreq/speedstep-lib.c line 223

Quoted 10 lines
arch/i386/kernel/cpu/cpufreq/speedstep-lib.c +9 -0
 
+		dprintk(KERN_INFO "model_id is %s\n", c->x86_model_id);
+		
+		/*
+		 * If the x86_model_id string contais "Mobile Intel(R) Pentium(R) 4"
+                 * omit all other checks and treat the CPU as a M-P4-M
+		 */
+		if (strstr(c->x86_model_id,"Mobile Intel(R) Pentium(R) 4") != NULL)
+		       return SPEEDSTEP_PROCESSOR_P4M;
+		

I'm not sure I like this approach: Intel is too relaxed with regard to their branding of mobile intel pentiums. There are pentium 3-m, mobile pentium 3-m, pentium 4-m, mobile pentium 4, and so on. And some -- which even have "mobile" in their name" -- do not support SpeedStep. THe CPUID [and some msr's for p3s] have proven to be more accurate in detecting speedstep capabilities. So I'd like to keep it that way, especially as the CPUID tends to change if features are changed; but the model_id string may stay the same for (future) processors which do no longer do speedstep in the way it is currently implemented in speedstep-ich.

arch/i386/kernel/cpu/cpufreq/speedstep-lib.c line 263

Quoted 11 lines
arch/i386/kernel/cpu/cpufreq/speedstep-lib.c +1 -0
 		switch (c->x86_mask) {
 		case 4: 
 			/*
@@ -248,6 +257,7 @@
 			 * So, how to distinguish all those processors with
 			 * ebx=0xf? I don't know. Sort them out, and wait
 			 * for someone to complain.
+			 * also, M-P4M HTs actually have ebx=0x8, too
 			 */
 			if (ebx == 0x0e)
 				return SPEEDSTEP_PROCESSOR_P4M;

So I'd prefer, due to the nastiness of the CPUID in the specific 0x0F29 case, to move the strstr check to this specific case -- e.g. have

			if (ebx == 0x0e)
				return SPEEDSTEP_PROCESSOR_P4M;
			if (strstr(c->x86_model_id,"Mobile Intel(R) Pentium(R) 4") != NULL)
				return SPEEDSTEP_PROCESSOR_P4M;

Could you update your patch accordingly, please?

Now, about your patch 2) [slightly re-ordered]:

arch/i386/kernel/cpu/cpufreq/speedstep-lib.c line 319

Quoted 9 lines
arch/i386/kernel/cpu/cpufreq/speedstep-lib.c +4 -2
-unsigned int speedstep_get_freqs(unsigned int processor,
+unsigned int speedstep_get_freqs(unsigned int cpu,
+				  unsigned int processor,
 				  unsigned int *low_speed,
 				  unsigned int *high_speed,
-				  void (*set_state) (unsigned int state,
+				  void (*set_state) (unsigned int cpu,
+						     unsigned int state,
 						     unsigned int notify)

This will break speedstep-smi, as it uses speedstep_get_freqs too...

Quoted 4 lines
arch/i386/kernel/cpu/cpufreq/speedstep-lib.c +4 -0
+	/* switch to physical CPU where state is to be changed*/
+	cpus_allowed = current->cpus_allowed;
+
+	/* only run on CPU to be set, or on its sibling */

arch/i386/kernel/cpu/cpufreq/speedstep-lib.c

Quoted 10 lines
arch/i386/kernel/cpu/cpufreq/speedstep-lib.c +10 -0
+       affected_cpu_map = cpumask_of_cpu(cpu);
+#ifdef CONFIG_X86_HT
+	hyperthreading = ((cpu_has_ht) && (smp_num_siblings == 2));
+	if (hyperthreading) {
+		sibling = cpu_sibling_map[cpu];
+                cpu_set(sibling, affected_cpu_map);
+	}
+#endif
+	set_cpus_allowed(current, affected_cpu_map);
+        BUG_ON(!cpu_isset(smp_processor_id(), affected_cpu_map));

Due to changes to cpu_sibling_map [it's a cpumask_t now], this won't generate appropriate code [if it even compiles] on 2.6.7-rc3. See in p4-clockmod.c how "easy" it becomes now.

arch/i386/kernel/cpu/cpufreq/speedstep-lib.c

Quoted 10 lines
arch/i386/kernel/cpu/cpufreq/speedstep-lib.c +10 -0
+       affected_cpu_map = cpumask_of_cpu(cpu);
+#ifdef CONFIG_X86_HT
+	hyperthreading = ((cpu_has_ht) && (smp_num_siblings == 2));
+	if (hyperthreading) {
+		sibling = cpu_sibling_map[cpu];
+                cpu_set(sibling, affected_cpu_map);
+	}
+#endif
+	set_cpus_allowed(current, affected_cpu_map);
+        BUG_ON(!cpu_isset(smp_processor_id(), affected_cpu_map));

Same here.

arch/i386/kernel/cpu/cpufreq/speedstep-lib.c

Quoted 3 lines
arch/i386/kernel/cpu/cpufreq/speedstep-lib.c +0 -3
-	/* capability check */
-	if (policy->cpu != 0)
-		return -ENODEV;

For the time being, this should be kept. Else you can load two different governors, or set two different frequencies, on both siblings. Obviously, it'll break. I'll try to fix it in the 2.6.8 timeframe...

Thanks,
Dominik

_______________________________________________
Cpufreq mailing list
[email protected]
http://www.linux.org.uk/mailman/listinfo/cpufreq

#4 lore raw
Quoted 7 lines · Dominik Brodowski

Hi Chris,

Quoted 4 lines

for 3): what you say sounds much better than my brutal hack, so i won't post it again. i'd definitely want to try to implement 2 cpufreq drivers, but i fear this is way beyond me at the moment and probably involves design choices i can't make.

Quoted 9 lines

well, somebody _does_ have to do design choices... However, I'm not yet convinced of the necessity of supporting throttling and frequency scaling on the same system. Even more, under normal circumstances [i.e. idling works, which it does on most systems], I see no _technical_ reason to do throttling, or even non-temperature related _dynamic throttling_ like other OSes do it. For those new to this list, there's some maths in the archives which proves that you need more energy == battery power if you do throttling for any given task... So it's something we need to discuss during 2.7.

In fact, I was more thinking about the FSB stuff which generate some noise recently I believe (and that have to be done during 2.7), though I'm pretty sure changing FSB is more helpfull for overclocker.

Signature · 4 lines
Bruno Ducrot

--  Which is worse:  ignorance or apathy?
--  Don't know.  Don't care.
#5 lore raw
In reply to Dominik Brodowski
Quoted 7 lines · Dominik Brodowski

Hi Chris,

Quoted 4 lines

for 3): what you say sounds much better than my brutal hack, so i won't post it again. i'd definitely want to try to implement 2 cpufreq drivers, but i fear this is way beyond me at the moment and probably involves design choices i can't make.

Quoted 9 lines

well, somebody _does_ have to do design choices... However, I'm not yet convinced of the necessity of supporting throttling and frequency scaling on the same system. Even more, under normal circumstances [i.e. idling works, which it does on most systems], I see no _technical_ reason to do throttling, or even non-temperature related _dynamic throttling_ like other OSes do it. For those new to this list, there's some maths in the archives which proves that you need more energy == battery power if you do throttling for any given task... So it's something we need to discuss during 2.7.

I'm actually inclined to kill off p4-clockmod one day. It doesn't seem to do anything useful whatsoever. It certainly doesn't lower the temperature on my xeon.

Dave

#6 lore raw
Quoted 18 lines · Dave Jones
Quoted 7 lines · Dominik Brodowski

Hi Chris,

Quoted 4 lines

for 3): what you say sounds much better than my brutal hack, so i won't post it again. i'd definitely want to try to implement 2 cpufreq drivers, but i fear this is way beyond me at the moment and probably involves design choices i can't make.

Quoted 9 lines

well, somebody _does_ have to do design choices... However, I'm not yet convinced of the necessity of supporting throttling and frequency scaling on the same system. Even more, under normal circumstances [i.e. idling works, which it does on most systems], I see no _technical_ reason to do throttling, or even non-temperature related _dynamic throttling_ like other OSes do it. For those new to this list, there's some maths in the archives which proves that you need more energy == battery power if you do throttling for any given task... So it's something we need to discuss during 2.7.

Quoted 3 lines

I'm actually inclined to kill off p4-clockmod one day. It doesn't seem to do anything useful whatsoever. It certainly doesn't lower the temperature on my xeon.

No surprise on that. It only helps if a) idling is broken, or b) you want to limit the CPU power to avoid overheating. So, I won't oppose removing p4-clockmod, even though I see some _very_ _limitied_ usefulness of it...

Dominik

#7 lore raw
Quoted 2 lines · Dominik Brodowski
Quoted 1 line · Dave Jones

On Wed, 2004-06-09 at 16:47, Dominik Brodowski wrote:

[...]

Quoted 7 lines
Quoted 3 lines

I'm actually inclined to kill off p4-clockmod one day. It doesn't seem to do anything useful whatsoever. It certainly doesn't lower the temperature on my xeon.

No surprise on that. It only helps if a) idling is broken, or b) you want to limit the CPU power to avoid overheating. So, I won't oppose removing p4-clockmod, even though I see some _very_ _limitied_ usefulness of it...

Please don't remove it!
I use it for (b). It seems my BIOS (or me?) is not too smart and sometimes fans run too slow too cool down the processor:

Jun  5 16:53:16 inferi-2 kernel: CPU#0: Temperature above threshold
Jun  5 16:53:16 inferi-2 kernel: CPU#0: Running in modulated clock mode
Jun  5 16:53:16 inferi-2 kernel: CPU#0: Temperature/speed normal
Jun  5 16:53:18 inferi-2 kernel: CPU#0: Temperature above threshold
Jun  5 16:53:18 inferi-2 kernel: CPU#0: Running in modulated clock mode

...

the only solution is to be to decrease processor speed through cpufreq. Maybe my workaround hides some misconfiguration but it works for now.

Signature · 2 lines
mattia
:wq!
#8 lore raw
In reply to Dominik Brodowski

hi dominik,

thanks alot for your detailed reply and patience with a newbie. i have attached 2 patches w.r.t. 2.6.7-rc3 that should adress your suggestions.

i also wonder about one thing, maybe you can help me on that. with the old speedstep driver, i would have occational hard freezups of my machine. this seems to be gone (just experimental evidence, since i didn't have one in a few days now where before they occured about hourly). so my guess is that something in this patch must have corrected it. since my machine just has one physical cpu, i really don't understand what went wrong before. do you have any idea on that?

Quoted 4 lines · Dominik Brodowski

Hi Chris,

<snip>

Quoted 8 lines

Even more, under normal circumstances [i.e. idling works,
which it does on most systems], I see no _technical_ reason to do throttling, or even non-temperature related _dynamic throttling_ like other OSes do it. For those new to this list, there's some maths in the archives which proves that you need more energy == battery power if you do throttling for any given task... So it's something we need to discuss during 2.7.

thanks, i didn't know that

Quoted 3 lines

Now, about your patch 1):

<snip>

Quoted 10 lines

So I'd prefer, due to the nastiness of the CPUID in the specific 0x0F29 case, to move the strstr check to this specific case -- e.g. have

			if (ebx == 0x0e)
				return SPEEDSTEP_PROCESSOR_P4M;
			if (strstr(c->x86_model_id,"Mobile Intel(R) Pentium(R) 4") != NULL)
				return SPEEDSTEP_PROCESSOR_P4M;

Could you update your patch accordingly, please?

thanks, done it. i was somehow anticipating that since reading the version string feels like a cheat.

Quoted 4 lines

Now, about your patch 2) [slightly re-ordered]:

Quoted 16 lines
Quoted 11 lines
(quoted diff) +2 -1
-unsigned int speedstep_get_freqs(unsigned int processor,
+unsigned int speedstep_get_freqs(unsigned int cpu,
+				  unsigned int processor,

unsigned int *low_speed, unsigned int *high_speed,

(quoted diff) +2 -1
-				  void (*set_state) (unsigned int state,
+				  void (*set_state) (unsigned int cpu,
+						     unsigned int state,

unsigned int notify)

This will break speedstep-smi, as it uses speedstep_get_freqs too...

ok, i chopped out the notify and cpu switching part from speedstep_set_state

Quoted 12 lines
Quoted 12 lines
(quoted diff) +12 -0
+	/* switch to physical CPU where state is to be changed*/
+	cpus_allowed = current->cpus_allowed;
+
+	/* only run on CPU to be set, or on its sibling */
+       affected_cpu_map = cpumask_of_cpu(cpu);
+#ifdef CONFIG_X86_HT
+	hyperthreading = ((cpu_has_ht) && (smp_num_siblings == 2));
+	if (hyperthreading) {
+		sibling = cpu_sibling_map[cpu];
+                cpu_set(sibling, affected_cpu_map);
+	}
+#endif
Quoted 10 lines

Review

Quoted 4 lines
(quoted diff) +2 -0
+	set_cpus_allowed(current, affected_cpu_map);
+        BUG_ON(!cpu_isset(smp_processor_id(), affected_cpu_map));

Due to changes to cpu_sibling_map [it's a cpumask_t now], this won't generate appropriate code [if it even compiles] on 2.6.7-rc3. See in p4-clockmod.c how "easy" it becomes now.

thanks, done it

<snip>

Quoted 10 lines

Review

Quoted 5 lines
(quoted diff) +0 -3
-	/* capability check */
-	if (policy->cpu != 0)
-		return -ENODEV;

For the time being, this should be kept. Else you can load two different governors, or set two different frequencies, on both siblings. Obviously, it'll break.

as far as i understand, the frequencies are always the same, because the siblings are explicitly notified. about the 2 different governors, the same problem exists in the current p4-clockmod.

Quoted 3 lines

I'll try to fix it in the 2.6.8 timeframe...

you mean that a change of governors gets communicated to the siblings?

thanks again for your help!

chris

arch/i386/kernel/cpu/cpufreq/speedstep-lib.c +3 -4
diff --unified --recursive --new-file linux-2.6.7-rc3/arch/i386/kernel/cpu/cpufreq/speedstep-lib.c linux-2.6.7-rc3.speedstep/arch/i386/kernel/cpu/cpufreq/speedstep-lib.c
@@ -252,11 +252,10 @@
 			 * specific.
 			 * M-P4-Ms may have either ebx=0xe or 0xf [see above]
 			 * M-P4/533 have either ebx=0xe or 0xf. [25317607.pdf]
-			 * So, how to distinguish all those processors with
-			 * ebx=0xf? I don't know. Sort them out, and wait
-			 * for someone to complain.
+			 * also, M-P4M HTs have ebx=0x8, too
+			 * For now, they are distinguished by the model_id string
 			 */
-			if (ebx == 0x0e)
+		        if ((ebx == 0x0e) || (strstr(c->x86_model_id,"Mobile Intel(R) Pentium(R) 4") != NULL)) 
 				return SPEEDSTEP_PROCESSOR_P4M;
 			break;
 		default:
arch/i386/kernel/cpu/cpufreq/speedstep-ich.c +59 -16

120 lines not on this page · raw message

_______________________________________________
Cpufreq mailing list
[email protected]
http://www.linux.org.uk/mailman/listinfo/cpufreq

#9 lore raw
Quoted 2 lines

i also wonder about one thing, maybe you can help me on that. with the old speedstep driver,

Do you mean speedstep-smi? Or what driver did you use?

Quoted 6 lines

i would have occational hard freezups of my
machine. this seems to be gone (just experimental evidence, since i didn't have one in a few days now where before they occured about hourly). so my guess is that something in this patch must have corrected it. since my machine just has one physical cpu, i really don't understand what went wrong before. do you have any idea on that?

Depends on the driver you were using :-)

Quoted 8 lines

as far as i understand, the frequencies are always the same, because the siblings are explicitly notified. about the 2 different governors, the same problem exists in the current p4-clockmod.

Quoted 3 lines

I'll try to fix it in the 2.6.8 timeframe...

you mean that a change of governors gets communicated to the siblings?

I mean that only one instance of a governor can be loaded on one physical CPU. The same problem exists in current p4-clockmod and was a mis-design of the cpufreq core [by me...], so it needs to be fixed quite soon.

Both patches look good to me -- Dave, could you apply them, please?

Thanks,
Dominik

_______________________________________________
Cpufreq mailing list
[email protected]
http://www.linux.org.uk/mailman/listinfo/cpufreq

#11 lore raw
In reply to Christian Hoelbling
Quoted 5 lines · Christian Hoelbling

hi dominik,

thanks alot for your detailed reply and patience with a newbie. i have attached 2 patches w.r.t. 2.6.7-rc3 that should adress your suggestions.

Quoted 8 lines

i also wonder about one thing, maybe you can help me on that. with the old speedstep driver, i would have occational hard freezups of my machine. this seems to be gone (just experimental evidence, since i didn't have one in a few days now where before they occured about hourly). so my guess is that something in this patch must have corrected it. since my machine just has one physical cpu, i really don't understand what went wrong before. do you have any idea on that?

IMHO the lockup is due to the fact we use local_irq_disable() but it should be done more globally. I believe you may still have some lockup still.

Cheers,

Signature · 4 lines
Bruno Ducrot

--  Which is worse:  ignorance or apathy?
--  Don't know.  Don't care.
#10 lore raw
In reply to Dominik Brodowski
Quoted 5 lines · Dominik Brodowski

I mean that only one instance of a governor can be loaded on one physical CPU. The same problem exists in current p4-clockmod and was a mis-design of the cpufreq core [by me...], so it needs to be fixed quite soon.

Both patches look good to me -- Dave, could you apply them, please?

I fell behind again on the cpufreq backlog. I'll try to get around to clearing it today/tomorrow.

Dave

#12 lore raw
In reply to Christian Hoelbling

Hi,

arch/i386/kernel/cpu/cpufreq/speedstep-ich.c line 73

Quoted 8 lines · Christian Hoelbling

diff --unified --recursive --new-file linux-2.6.7-rc3/arch/i386/kernel/cpu/cpufreq/speedstep-ich.c linux-2.6.7-rc3.speedstep/arch/i386/kernel/cpu/cpufreq/speedstep-ich.c

arch/i386/kernel/cpu/cpufreq/speedstep-ich.c +1 -0
@@ -70,6 +70,7 @@
  * @notify: whether to call cpufreq_notify_transition for CPU speed changes
  *
  *   Tries to change the SpeedStep state. 
+ * Note: notify is a dummy argument. The routine never notifies.

Maybe it's better to kill notify, and modify a little speedstep-smi and speedstep-lib

Dominik, your though, please?

...

arch/i386/kernel/cpu/cpufreq/speedstep-ich.c line 279

Quoted 7 lines
arch/i386/kernel/cpu/cpufreq/speedstep-ich.c +7 -0
+
+	for_each_cpu(i) {
+	        if (cpu_isset(i, affected_cpu_map)) {
+	                freqs.cpu = i;
+			cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
+		}
+	}

What about

	for_each_cpu_mask(i, affected_cpu_map) {
		...
	}

instead?

arch/i386/kernel/cpu/cpufreq/speedstep-ich.c line 342

Quoted 10 lines
arch/i386/kernel/cpu/cpufreq/speedstep-ich.c +3 -0
 	/* detect low and high frequency */
 	result = speedstep_get_freqs(speedstep_processor,
@@ -297,6 +337,9 @@
 	if (!speed)
 		return -EIO;
 
+	/* allow to run on any CPU */
+	set_cpus_allowed(current, cpus_allowed);
+

Something wrong. I mean, there is a return but the set_cpu_allowed() is done after...

Cheers,

Signature · 4 lines
Bruno Ducrot

--  Which is worse:  ignorance or apathy?
--  Don't know.  Don't care.
#13 lore raw
Quoted 1 line · Bruno Ducrot

Dominik, your though, please?

I'd agree with this change.

Quoted 5 lines

What about

	for_each_cpu_mask(i, affected_cpu_map) {
		...
	}

instead?

That's indeed shorter -- don't know why whoever converted p4-clockmod to the new sibling mask did use for_each_cpu_mask...

Quoted 10 lines

arch/i386/kernel/cpu/cpufreq/speedstep-ich.c line 342

Quoted 7 lines
arch/i386/kernel/cpu/cpufreq/speedstep-ich.c +3 -0
@@ -297,6 +337,9 @@
 	if (!speed)
 		return -EIO;
 
+	/* allow to run on any CPU */
+	set_cpus_allowed(current, cpus_allowed);
+

Something wrong. I mean, there is a return but the set_cpu_allowed() is done after...

Thanks for catching this -- it definitely needs to be fixed.

Dominik

#14 lore raw

PATCH 1/1 Remove notify in speedstep_set_state [1/2]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Remove 'notify' in speedstep_set_state for speedstep-lib

 arch/i386/kernel/cpu/cpufreq/speedstep-lib.c |   10 ++++------
 arch/i386/kernel/cpu/cpufreq/speedstep-lib.h |    2 +-
 2 files changed, 5 insertions(+), 7 deletions(-)

- --- linux-2.6.7-rc3-mm1/arch/i386/kernel/cpu/cpufreq/speedstep-lib.c 2004/06/10 18:28:39 1.1
+++ linux-2.6.7-rc3-mm1/arch/i386/kernel/cpu/cpufreq/speedstep-lib.c 2004/06/10 18:30:43
@@ -321,9 +321,7 @@ EXPORT_SYMBOL_GPL(speedstep_detect_proce
unsigned int speedstep_get_freqs(unsigned int processor,
unsigned int *low_speed, unsigned int *high_speed,
- - void (*set_state) (unsigned int state,
- - unsigned int notify)
- - )
+ void (*set_state) (unsigned int state))
{
unsigned int prev_speed;
unsigned int ret = 0;
@@ -340,7 +338,7 @@ unsigned int speedstep_get_freqs(unsigne
local_irq_save(flags);

/* switch to low state */
- - set_state(SPEEDSTEP_LOW, 0);
+ set_state(SPEEDSTEP_LOW);
*low_speed = speedstep_get_processor_frequency(processor); if (!*low_speed) {
ret = -EIO;
@@ -348,7 +346,7 @@ unsigned int speedstep_get_freqs(unsigne
}

/* switch to high state */
- - set_state(SPEEDSTEP_HIGH, 0);
+ set_state(SPEEDSTEP_HIGH);
*high_speed = speedstep_get_processor_frequency(processor); if (!*high_speed) {
ret = -EIO;
@@ -362,7 +360,7 @@ unsigned int speedstep_get_freqs(unsigne

/* switch to previous state, if necessary */ if (*high_speed != prev_speed)
- - set_state(SPEEDSTEP_LOW, 0);
+ set_state(SPEEDSTEP_LOW);

out:
local_irq_restore(flags);
- --- linux-2.6.7-rc3-mm1/arch/i386/kernel/cpu/cpufreq/speedstep-lib.h 2004/06/10 18:28:29 1.1
+++ linux-2.6.7-rc3-mm1/arch/i386/kernel/cpu/cpufreq/speedstep-lib.h 2004/06/10 18:47:50
@@ -44,4 +44,4 @@ extern unsigned int speedstep_get_proces
extern unsigned int speedstep_get_freqs(unsigned int processor,
unsigned int *low_speed,
unsigned int *high_speed,
- - void (*set_state) (unsigned int state, unsigned int notify));
+ void (*set_state) (unsigned int state));

- --
Bruno Ducrot

- -- Which is worse: ignorance or apathy?
- -- Don't know. Don't care.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (GNU/Linux)

iD8DBQFAyLXwiZtNKH9GMYcRArLaAJ97vEFTzsoFXBIqFNXsOeOI4nyr1QCeJwod ewJTE56F+yqKoyN5kx9SxWU=
=+VHa
-----END PGP SIGNATURE-----

#15 lore raw

PATCH 1/1 Remove notify in speedstep_set_state [2/2]

In reply to Dominik Brodowski

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Remove notify in speedstep_set_state [2/2]

This is for speedstep-smi.

 arch/i386/kernel/cpu/cpufreq/speedstep-smi.c |   32 ++++++++++-----------------
 1 files changed, 13 insertions(+), 19 deletions(-)

- --- linux-2.6.7-rc3-mm1/arch/i386/kernel/cpu/cpufreq/speedstep-smi.c 2004/06/10 19:13:22 1.2
+++ linux-2.6.7-rc3-mm1/arch/i386/kernel/cpu/cpufreq/speedstep-smi.c 2004/06/10 19:17:17
@@ -145,31 +145,18 @@ static int speedstep_get_state (void)
/**
* speedstep_set_state - set the SpeedStep state
* @state: new processor frequency state (SPEEDSTEP_LOW or SPEEDSTEP_HIGH)
- - * @notify: whether to call cpufreq_notify_transition
*
*/
- -static void speedstep_set_state (unsigned int state, unsigned int notify) +static void speedstep_set_state (unsigned int state)
{
- - unsigned int old_state, result = 0, command, new_state;
+ unsigned int result = 0, command, new_state;
unsigned long flags;
- - struct cpufreq_freqs freqs;
unsigned int function=SET_SPEEDSTEP_STATE;
unsigned int retry = 0;

if (state > 0x1)
return;

- - old_state = speedstep_get_state();
- - freqs.old = speedstep_freqs[old_state].frequency;
- - freqs.new = speedstep_freqs[state].frequency;
- - freqs.cpu = 0; /* speedstep.c is UP only driver */
- -
- - if (old_state == state)
- - return;
- -
- - if (notify)
- - cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
- -
/* Disable IRQs */
local_irq_save(flags);

@@ -198,9 +185,6 @@ static void speedstep_set_state (unsigne
printk(KERN_ERR "cpufreq: change failed with new_state %u and result %u\n", new_state, result);
}

- - if (notify)
- - cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
- -
return;
}

@@ -217,11 +201,21 @@ static int speedstep_target (struct cpuf
unsigned int target_freq, unsigned int relation)
{
unsigned int newstate = 0;
+ struct cpufreq_freqs freqs;

if (cpufreq_frequency_table_target(policy, &speedstep_freqs[0], target_freq, relation, &newstate))
return -EINVAL;

- - speedstep_set_state(newstate, 1);
+ freqs.old = speedstep_freqs[speedstep_get_state()].frequency;
+ freqs.new = speedstep_freqs[newstate].frequency;
+ freqs.cpu = 0; /* speedstep.c is UP only driver */
+
+ if (freqs.old == freqs.new)
+ return 0;
+
+ cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
+ speedstep_set_state(newstate);
+ cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);

return 0;
}

- --
Bruno Ducrot

- -- Which is worse: ignorance or apathy?
- -- Don't know. Don't care.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (GNU/Linux)

iD8DBQFAyLZBiZtNKH9GMYcRAv6eAKCJINQlfzWXPEHXri8v3zc+VhdklQCfXiXT PnuBJapbMHvGslvKRIVLJuU=
=VHyT
-----END PGP SIGNATURE-----

#16 lore raw

PATCH 1/1 security fix for speedstep-smi

In reply to Dominik Brodowski

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Looking around speedstep-smi, there is a lot of call to speedstep_get_state() which is used as an index for the array speedstep_freqs. But speedstep_get_state() will return the value of ebx after a call to a SMI handler. It's not safe IMHO.

 arch/i386/kernel/cpu/cpufreq/speedstep-smi.c |    2 +-
 1 files changed, 1 insertion(+), 1 deletion(-)

- --- linux-2.6.7-rc3-mm1/arch/i386/kernel/cpu/cpufreq/speedstep-smi.c 2004/06/10 18:30:54 1.1
+++ linux-2.6.7-rc3-mm1/arch/i386/kernel/cpu/cpufreq/speedstep-smi.c 2004/06/10 19:12:49
@@ -139,7 +139,7 @@ static int speedstep_get_state (void)
: "a" (command), "b" (function), "c" (0), "d" (smi_port), "S" (0)
);

- - return state;
+ return (state & 1);
}

/**

- --
Bruno Ducrot

- -- Which is worse: ignorance or apathy?
- -- Don't know. Don't care.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (GNU/Linux)

iD8DBQFAyLfwiZtNKH9GMYcRAhlbAJ42SVKb1w+0sv0Ov2/Jbp8NB+J34wCfcfSv 57zWVNSxfyl5fQzUpzPEYa8=
=J4Ym
-----END PGP SIGNATURE-----

#17 lore raw
In reply to Dominik Brodowski
Quoted 4 lines · Dominik Brodowski
Quoted 1 line · Bruno Ducrot

Dominik, your though, please?

I'd agree with this change.

I will do the change for the speedstep-ich tomorrow unless someone else (Christian ?) beat me.

Cheers,

Signature · 4 lines
Bruno Ducrot

--  Which is worse:  ignorance or apathy?
--  Don't know.  Don't care.
#18 lore raw

PATCH 1/1 replace for_each_cpu with for_each_cpu_mask (was Re: [PATCH] 2.6.5 speedstep on P4Ms)

In reply to Dominik Brodowski

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Quoted 9 lines · Dominik Brodowski
Quoted 5 lines · Bruno Ducrot

What about

	for_each_cpu_mask(i, affected_cpu_map) {
		...
	}

instead?

That's indeed shorter -- don't know why whoever converted p4-clockmod to the new sibling mask did use for_each_cpu_mask...

Something like that should be ok, I guess.

- --- linux-2.6.7-rc3-mm1/arch/i386/kernel/cpu/cpufreq/p4-clockmod.c 2004/06/10 19:54:23 1.1
+++ linux-2.6.7-rc3-mm1/arch/i386/kernel/cpu/cpufreq/p4-clockmod.c 2004/06/10 19:57:00
@@ -132,34 +132,28 @@ static int cpufreq_p4_target(struct cpuf
#endif

/* notifiers */
- - for_each_cpu(i) {
- - if (cpu_isset(i, affected_cpu_map)) {
- - freqs.cpu = i;
- - cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
- - }
+ for_each_cpu_mask(i, affected_cpu_map) {
+ freqs.cpu = i;
+ cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
}

/* run on each logical CPU, see section 13.15.3 of IA32 Intel Architecture Software
* Developer's Manual, Volume 3
*/
- - for_each_cpu(i) {
- - if (cpu_isset(i, affected_cpu_map)) {
- - cpumask_t this_cpu = cpumask_of_cpu(i);
+ for_each_cpu_mask(i, affected_cpu_map) {
+ cpumask_t this_cpu = cpumask_of_cpu(i);

- - set_cpus_allowed(current, this_cpu);
- - BUG_ON(smp_processor_id() != i);
+ set_cpus_allowed(current, this_cpu);
+ BUG_ON(smp_processor_id() != i);

- - cpufreq_p4_setdc(i, p4clockmod_table[newstate].index);
- - }
+ cpufreq_p4_setdc(i, p4clockmod_table[newstate].index);
}
set_cpus_allowed(current, cpus_allowed);

/* notifiers */
- - for_each_cpu(i) {
- - if (cpu_isset(i, affected_cpu_map)) {
- - freqs.cpu = i;
- - cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
- - }
+ for_each_cpu_mask(i, affected_cpu_map) {
+ freqs.cpu = i;
+ cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
}

return 0;

- --
Bruno Ducrot

- -- Which is worse: ignorance or apathy?
- -- Don't know. Don't care.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (GNU/Linux)

iD8DBQFAyL7LiZtNKH9GMYcRAoMXAJ0d7AHn7kWiG8FOsw2BR/goNb4/igCgjws8 DVnDvwv5x+kdIPyMTS7RIAI=
=ZtPZ
-----END PGP SIGNATURE-----

#19 lore raw

Re: [PATCH] replace for_each_cpu with for_each_cpu_mask (was Re: [PATCH] 2.6.5 speedstep on P4Ms)

Quoted 5 lines · Bruno Ducrot

- --- linux-2.6.7-rc3-mm1/arch/i386/kernel/cpu/cpufreq/p4-clockmod.c 2004/06/10 19:54:23 1.1
+++ linux-2.6.7-rc3-mm1/arch/i386/kernel/cpu/cpufreq/p4-clockmod.c 2004/06/10 19:57:00

(quoted diff) +0 -0
@@ -132,34 +132,28 @@ static int cpufreq_p4_target(struct cpuf
 #endif
 

Review static int cpufreq_p4_target(struct cpuf · line 138

Quoted 10 lines
(quoted diff) +3 -5
 	/* notifiers */
- -	for_each_cpu(i) {
- -		if (cpu_isset(i, affected_cpu_map)) {
- -			freqs.cpu = i;
- -			cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
- -		}
+	for_each_cpu_mask(i, affected_cpu_map) {
+		freqs.cpu = i;
+		cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
 	}

That's a really funky looking patch. Care to rediff ? 8-)

Dave

#20 lore raw

Re: [PATCH] replace for_each_cpu with for_each_cpu_mask (was Re: [PATCH] 2.6.5 speedstep on P4Ms)

Quoted 7 lines · Dave Jones
Quoted 5 lines · Bruno Ducrot

- --- linux-2.6.7-rc3-mm1/arch/i386/kernel/cpu/cpufreq/p4-clockmod.c 2004/06/10 19:54:23 1.1
+++ linux-2.6.7-rc3-mm1/arch/i386/kernel/cpu/cpufreq/p4-clockmod.c 2004/06/10 19:57:00

(quoted diff) +0 -0
@@ -132,34 +132,28 @@ static int cpufreq_p4_target(struct cpuf
 #endif
 
Quoted 12 lines

Review static int cpufreq_p4_target(struct cpuf · line 138

Quoted 10 lines
(quoted diff) +3 -5
 	/* notifiers */
- -	for_each_cpu(i) {
- -		if (cpu_isset(i, affected_cpu_map)) {
- -			freqs.cpu = i;
- -			cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
- -		}
+	for_each_cpu_mask(i, affected_cpu_map) {
+		freqs.cpu = i;
+		cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
 	}

That's a really funky looking patch. Care to rediff ? 8-)

Something wrong how I configure mutt I guess. Will resend.

Signature · 4 lines
Bruno Ducrot

--  Which is worse:  ignorance or apathy?
--  Don't know.  Don't care.