x86/topology: Improve CPUID.1F handling
cover x86/topology: Improve CPUID.1F handling
On Intel AlderLake-N platforms where there are Ecores only, the Ecore Module topology is enumerated via CPUID.1F Module level, which has not been supported by Linux kernel yet.
This exposes two issues in current CPUID.1F handling code.
1. Linux interprets the Module id bits as package id and erroneously reports a multi module system as a multi-package system.
2. Linux excludes the unknown Module id bits from the core_id, and results in duplicate core_id’s shown in a package after the first issue solved.
Plus that, a third problem is observed on Intel Hybrid ADL-S/P platforms. The return value of CPUID.1F SMT level EBX (number of siblings) differs on Pcore CPUs and Ecore CPUs, and results in inconsistent smp_num_siblings value based on the Pcore/Ecore CPU enumeration order. This could bring some potential issues although we have not observed any functionalities issues so far.
Patch 1/7 and 2/7 fix the first two issues. And at the same time, it reveals a reality that the core_id could be sparse on platforms with CPUID.1F support.
Patch 3/7 improves coretemp driver code to be able to handle sparse core id, which is the only driver that uses core_id as array index and run on platforms with CPUID.1F support.
Patch 4/7 to 7/7 propose a fix for the third problem and update the related Documents.
The patch series have been tested on Intel ICL/CLX servers, SKL/KBL/ADL clients.
thanks,
-rui
PATCH 1/7 x86/topology: Fix multiple packages shown on a single-package system
CPUID.1F/B does not emumerate Package level explicitly, instead, all the APICID bits above the enumerated levels are assumed to be package id bits.
Current code gets package id by shifting out all the APICID bits that Linux supports, rather than shifting out all the APICID bits that CPUID.1F enumerates. This introduces problems when CPUID.1F enumerates a level that Linux does not support.
For example, on an AlderLake-N platform, there are 2 Ecore Modules, which has 4 atom cores in each module, in a single package.
Linux does not support Module level and interprets the Module id bits as package id and erroneously reports a multi module system as a multi-package system.
Fix this by using APICID bits above all the CPUID.1F enumerated levels as package id.
--- arch/x86/kernel/cpu/topology.c | 14 ++++++++++---- 1 file changed, 10 insertions(+), 4 deletions(-)
arch/x86/kernel/cpu/topology.c +10 -4
@@ -96,6 +96,7 @@ int detect_extended_topology(struct cpuinfo_x86 *c)unsigned int ht_mask_width, core_plus_mask_width, die_plus_mask_width;unsigned int core_select_mask, core_level_siblings;unsigned int die_select_mask, die_level_siblings; +unsigned int pkg_mask_width;bool die_level_present = false;int leaf;
@@ -111,10 +112,10 @@ int detect_extended_topology(struct cpuinfo_x86 *c) core_level_siblings = smp_num_siblings =LEVEL_MAX_SIBLINGS (ebx); core_plus_mask_width = ht_mask_width =BITS_SHIFT_NEXT_LEVEL (eax); die_level_siblings =LEVEL_MAX_SIBLINGS (ebx); - die_plus_mask_width =BITS_SHIFT_NEXT_LEVEL (eax); + pkg_mask_width = die_plus_mask_width =BITS_SHIFT_NEXT_LEVEL (eax); sub_index =1 ; - do { +while (true) {cpuid_count (leaf, sub_index, &eax, &ebx, &ecx, &edx);/*
@@ -132,8 +133,13 @@ int detect_extended_topology(struct cpuinfo_x86 *c) die_plus_mask_width =BITS_SHIFT_NEXT_LEVEL (eax); } +if (LEAFB_SUBTYPE (ecx) !=INVALID_TYPE ) + pkg_mask_width =BITS_SHIFT_NEXT_LEVEL (eax); +else +break ; + sub_index++; - }while (LEAFB_SUBTYPE (ecx) !=INVALID_TYPE ); + } core_select_mask = (~(-1 << core_plus_mask_width)) >> ht_mask_width; die_select_mask = (~(-1 << die_plus_mask_width)) >>
@@ -148,7 +154,7 @@ int detect_extended_topology(struct cpuinfo_x86 *c) } c->phys_proc_id = apic->phys_pkg_id (c->initial_apicid, - die_plus_mask_width); + pkg_mask_width);/* * Reinit the apicid, now that we have extended initial_apicid. */
Signature · 1 line
2.34.1
PATCH 2/7 x86/topology: Fix duplicated core_id within a package
In reply to Zhang RuiToday, core_id is assumed to be unique within each package.
But an AlderLake-N platform adds a Module level between core and package, Linux excludes the unknown modules bits from the core_id, resulting in duplicate core_id's.
To keep core_id unique within a package, Linux must include all APIC-id bits for known or un-known levels above the core and below the package in the core_id.
It is important to understand that core_id's have always come directly from the APIC-id encoding, which comes from the BIOS. Thus there is no guarantee that they start at 0, or that they are contiguous. As such, using them for array indexes can be problematic.
--- arch/x86/kernel/cpu/topology.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kernel/cpu/topology.c +1 -1
@@ -141,7 +141,7 @@ int detect_extended_topology(struct cpuinfo_x86 *c) sub_index++; } - core_select_mask = (~(-1 << core_plus_mask_width)) >> ht_mask_width; + core_select_mask = (~(-1 << pkg_mask_width)) >> ht_mask_width; die_select_mask = (~(-1 << die_plus_mask_width)) >> core_plus_mask_width;
Signature · 1 line
2.34.1
PATCH 3/7 hwmon/coretemp: Handle large core id value
In reply to Zhang RuiThe coretemp driver supports up to a hard-coded limit of 128 cores.
Today, the driver can not support a core with an id above that limit. Yet, the encoding of core_id's is arbitrary (BIOS APIC-id) and so they may be sparse and they may be large.
Update the driver to map arbitrary core_id numbers into appropriate array indexes so that 128 cores can be supported, no matter the encoding of core_ids's.
--- drivers/hwmon/coretemp.c | 55 +++++++++++++++++++++++++++++----------- 1 file changed, 40 insertions(+), 15 deletions(-)
drivers/hwmon/coretemp.c +40 -15
@@ -46,9 +46,6 @@ MODULE_PARM_DESC(tjmax, "TjMax value in degrees Celsius");#define TOTAL_ATTRS (MAX_CORE_ATTRS + 1)#define MAX_CORE_DATA (NUM_REAL_CORES + BASE_SYSFS_ATTR_NO) -#define TO_CORE_ID (cpu) (cpu_data(cpu).cpu_core_id) -#define TO_ATTR_NO (cpu) (TO_CORE_ID(cpu) + BASE_SYSFS_ATTR_NO) -#ifdef CONFIG_SMP #define for_each_sibling (i, cpu) \ for_each_cpu(i, topology_sibling_cpumask(cpu))
@@ -91,6 +88,8 @@ struct temp_data {struct platform_data {struct device *hwmon_dev;u16 pkg_id; +u16 cpu_map[NUM_REAL_CORES ]; +struct ida ida;struct cpumask cpumask;struct temp_data *core_data[MAX_CORE_DATA ];struct device_attribute name_attr;
@@ -441,7 +440,7 @@ static struct temp_data *init_temp_data(unsigned int cpu, int pkg_flag)MSR_IA32_THERM_STATUS ; tdata->is_pkg_data = pkg_flag; tdata->cpu = cpu; - tdata->cpu_core_id =TO_CORE_ID (cpu); + tdata->cpu_core_id =topology_core_id (cpu); tdata->attr_size =MAX_CORE_ATTRS ;mutex_init (&tdata->update_lock);return tdata;
@@ -454,7 +453,7 @@ static int create_core_data(struct platform_device *pdev, unsigned int cpu,struct platform_data *pdata =platform_get_drvdata (pdev);struct cpuinfo_x86 *c = &cpu_data (cpu);u32 eax, edx; -int err, attr_no; +int err, index, attr_no;/* * Find attr number for sysfs:
@@ -462,14 +461,26 @@ static int create_core_data(struct platform_device *pdev, unsigned int cpu,* The attr number is always core id + 2 * The Pkgtemp will always show up as temp1_*, if available */ - attr_no = pkg_flag ?PKG_SYSFS_ATTR_NO :TO_ATTR_NO (cpu); +if (pkg_flag) + attr_no =PKG_SYSFS_ATTR_NO ; +else { + index =ida_alloc (&pdata->ida,GFP_KERNEL ); +if (index <0 ) +return index; + pdata->cpu_map[index] =topology_core_id (cpu); + attr_no = index +BASE_SYSFS_ATTR_NO ; + } -if (attr_no >MAX_CORE_DATA -1 ) -return -ERANGE ; +if (attr_no >MAX_CORE_DATA -1 ) { + err = -ERANGE ; + goto ida_free; + } tdata =init_temp_data (cpu, pkg_flag); -if (!tdata) -return -ENOMEM ; +if (!tdata) { + err = -ENOMEM ; + goto ida_free; + }/* Test if we can access the status register */ err =rdmsr_safe_on_cpu (cpu, tdata->status_reg, &eax, &edx);
@@ -505,6 +516,9 @@ static int create_core_data(struct platform_device *pdev, unsigned int cpu, exit_free: pdata->core_data[attr_no] =NULL ;kfree (tdata); +ida_free: +if (!pkg_flag) +ida_free (&pdata->ida, index);return err; }
@@ -524,6 +538,8 @@ static void coretemp_remove_core(struct platform_data *pdata, int indx)kfree (pdata->core_data[indx]); pdata->core_data[indx] =NULL ; + +ida_free (&pdata->ida, indx -BASE_SYSFS_ATTR_NO ); }static int coretemp_probe (struct platform_device *pdev)
@@ -537,6 +553,7 @@ static int coretemp_probe(struct platform_device *pdev)return -ENOMEM ; pdata->pkg_id = pdev->id; +ida_init (&pdata->ida);platform_set_drvdata (pdev, pdata); pdata->hwmon_dev = devm_hwmon_device_register_with_groups(dev, DRVNAME,
@@ -553,6 +570,7 @@ static int coretemp_remove(struct platform_device *pdev)if (pdata->core_data[i])coretemp_remove_core (pdata, i); +ida_destroy (&pdata->ida);return 0 ; }
@@ -647,7 +665,7 @@ static int coretemp_cpu_offline(unsigned int cpu)struct platform_device *pdev =coretemp_get_pdev (cpu);struct platform_data *pd;struct temp_data *tdata; -int indx, target; +int i, indx =-1 , target;/* * Don't execute this on suspend as the device remove locks
@@ -660,12 +678,19 @@ static int coretemp_cpu_offline(unsigned int cpu)if (!pdev)return 0 ; -/* The core id is too big, just return */ - indx =TO_ATTR_NO (cpu); -if (indx >MAX_CORE_DATA -1 ) + pd =platform_get_drvdata (pdev); + +for (i =0 ; i <NUM_REAL_CORES ; i++) { +if (pd->cpu_map[i] ==topology_core_id (cpu)) { + indx = i +BASE_SYSFS_ATTR_NO ; +break ; + } + } + +/* Too many cores and this core is not pupolated, just return */ +if (indx <0 )return 0 ; - pd =platform_get_drvdata (pdev); tdata = pd->core_data[indx];cpumask_clear_cpu (cpu, &pd->cpumask);
Signature · 1 line
2.34.1
Re: [PATCH 3/7] hwmon/coretemp: Handle large core id value
Quoted 140 lines · Zhang Rui
The coretemp driver supports up to a hard-coded limit of 128 cores.
Today, the driver can not support a core with an id above that limit. Yet, the encoding of core_id's is arbitrary (BIOS APIC-id) and so they may be sparse and they may be large.
Update the driver to map arbitrary
core_idnumbers into appropriate array indexes so that 128 cores can be supported, no matter the encoding of core_ids's.Acked-by Len Brown <[email protected]>Signed-off-by Zhang Rui <[email protected]>--- drivers/hwmon/coretemp.c | 55 +++++++++++++++++++++++++++++----------- 1 file changed, 40 insertions(+), 15 deletions(-)drivers/hwmon/coretemp.c +29
-14@@ -46,9 +46,6 @@ MODULE_PARM_DESC(tjmax, "TjMax value in degrees Celsius");#define TOTAL_ATTRS (MAX_CORE_ATTRS + 1)#define MAX_CORE_DATA (NUM_REAL_CORES + BASE_SYSFS_ATTR_NO) -#define TO_CORE_ID (cpu) (cpu_data(cpu).cpu_core_id) -#define TO_ATTR_NO (cpu) (TO_CORE_ID(cpu) + BASE_SYSFS_ATTR_NO) -#ifdef CONFIG_SMP #define for_each_sibling (i, cpu) \ for_each_cpu(i, topology_sibling_cpumask(cpu))@@ -91,6 +88,8 @@ struct temp_data {struct platform_data {struct device *hwmon_dev;u16 pkg_id; +u16 cpu_map[NUM_REAL_CORES ]; +struct ida ida;struct cpumask cpumask;struct temp_data *core_data[MAX_CORE_DATA ];struct device_attribute name_attr;@@ -441,7 +440,7 @@ static struct temp_data *init_temp_data(unsigned int cpu, int pkg_flag)MSR_IA32_THERM_STATUS ; tdata->is_pkg_data = pkg_flag; tdata->cpu = cpu; - tdata->cpu_core_id =TO_CORE_ID (cpu); + tdata->cpu_core_id =topology_core_id (cpu); tdata->attr_size =MAX_CORE_ATTRS ;mutex_init (&tdata->update_lock);return tdata;@@ -454,7 +453,7 @@ static int create_core_data(struct platform_device *pdev, unsigned int cpu,struct platform_data *pdata =platform_get_drvdata (pdev);struct cpuinfo_x86 *c = &cpu_data (cpu);u32 eax, edx; -int err, attr_no; +int err, index, attr_no;/* * Find attr number for sysfs: @@ -462,14 +461,26 @@ static int create_core_data(struct platform_device *pdev, unsigned int cpu,* The attr number is always core id + 2 * The Pkgtemp will always show up as temp1_*, if available */ - attr_no = pkg_flag ?PKG_SYSFS_ATTR_NO :TO_ATTR_NO (cpu); +if (pkg_flag) + attr_no =PKG_SYSFS_ATTR_NO ; +else { + index =ida_alloc (&pdata->ida,GFP_KERNEL ); +if (index <0 ) +return index; + pdata->cpu_map[index] =topology_core_id (cpu); + attr_no = index +BASE_SYSFS_ATTR_NO ; + } -if (attr_no >MAX_CORE_DATA -1 ) -return -ERANGE ; +if (attr_no >MAX_CORE_DATA -1 ) { + err = -ERANGE ; + goto ida_free; + } tdata =init_temp_data (cpu, pkg_flag); -if (!tdata) -return -ENOMEM ; +if (!tdata) { + err = -ENOMEM ; + goto ida_free; + }/* Test if we can access the status register */ err =rdmsr_safe_on_cpu (cpu, tdata->status_reg, &eax, &edx);@@ -505,6 +516,9 @@ static int create_core_data(struct platform_device *pdev, unsigned int cpu, exit_free: pdata->core_data[attr_no] =NULL ;kfree (tdata); +ida_free: +if (!pkg_flag) +ida_free (&pdata->ida, index);return err; }@@ -524,6 +538,8 @@ static void coretemp_remove_core(struct platform_data *pdata, int indx)kfree (pdata->core_data[indx]); pdata->core_data[indx] =NULL ; + +ida_free (&pdata->ida, indx -BASE_SYSFS_ATTR_NO ); }static int coretemp_probe (struct platform_device *pdev)@@ -537,6 +553,7 @@ static int coretemp_probe(struct platform_device *pdev)return -ENOMEM ; pdata->pkg_id = pdev->id; +ida_init (&pdata->ida);platform_set_drvdata (pdev, pdata); pdata->hwmon_dev = devm_hwmon_device_register_with_groups(dev, DRVNAME,@@ -553,6 +570,7 @@ static int coretemp_remove(struct platform_device *pdev)if (pdata->core_data[i])coretemp_remove_core (pdata, i); +ida_destroy (&pdata->ida);return 0 ; }@@ -647,7 +665,7 @@ static int coretemp_cpu_offline(unsigned int cpu)struct platform_device *pdev =coretemp_get_pdev (cpu);struct platform_data *pd;struct temp_data *tdata; -int indx, target; +int i, indx =-1 , target;/* * Don't execute this on suspend as the device remove locks @@ -660,12 +678,19 @@ static int coretemp_cpu_offline(unsigned int cpu)if (!pdev)return 0 ; -/* The core id is too big, just return */ - indx =TO_ATTR_NO (cpu); -if (indx >MAX_CORE_DATA -1 )
drivers/hwmon/coretemp.c static int coretemp_cpu_offline(unsigned int cpu) · line 690
Quoted 10 lines
drivers/hwmon/coretemp.c +10
-0+ pd =platform_get_drvdata (pdev); + +for (i =0 ; i <NUM_REAL_CORES ; i++) { +if (pd->cpu_map[i] ==topology_core_id (cpu)) { + indx = i +BASE_SYSFS_ATTR_NO ; +break ; + } + } + +/* Too many cores and this core is not pupolated, just return */
populated
Other than that looks good.
Quoted 10 lines
drivers/hwmon/coretemp.c +1
-1+if (indx <0 )return 0 ; - pd =platform_get_drvdata (pdev); tdata = pd->core_data[indx];cpumask_clear_cpu (cpu, &pd->cpumask);Signature · 1 line
2.34.1
Re: [PATCH 3/7] hwmon/coretemp: Handle large core id value
Hi, Guenter,
Quoted 18 lines · Guenter Roeck
Quoted 14 lines
- /* The core id is too big, just return */
- indx =TO_ATTR_NO(cpu);
- if (indx >MAX_CORE_DATA- 1)
+ pd =platform_get_drvdata(pdev);drivers/hwmon/coretemp.c +9
-0+ + for (i = 0; i < NUM_REAL_CORES; i++) { + if (pd->cpu_map[i] == topology_core_id(cpu)) { + indx = i + BASE_SYSFS_ATTR_NO; + break; + } + } + + /* Too many cores and this core is not pupolated, justQuoted 1 line
return */
populated
Quoted 3 lines
Other than that looks good.
Acked-by Guenter Roeck <[email protected]>
Thanks for reviewing, will fix in next version.
thanks,
rui
Re: [PATCH 3/7] hwmon/coretemp: Handle large core id value
In reply to Zhang RuiQuoted 9 lines · Zhang Rui
The coretemp driver supports up to a hard-coded limit of 128 cores.
Today, the driver can not support a core with an id above that limit. Yet, the encoding of core_id's is arbitrary (BIOS APIC-id) and so they may be sparse and they may be large.
Update the driver to map arbitrary
core_idnumbers into appropriate array indexes so that 128 cores can be supported, no matter the encoding of core_ids's.
Please capitalize 'ID' consistently throughout the series.
drivers/hwmon/coretemp.c static int create_core_data(struct platform_device *pdev, unsigned int cpu, · line 472
Quoted 10 lines
drivers/hwmon/coretemp.c +9
-1- attr_no = pkg_flag ?PKG_SYSFS_ATTR_NO :TO_ATTR_NO (cpu); +if (pkg_flag) + attr_no =PKG_SYSFS_ATTR_NO ; +else { + index =ida_alloc (&pdata->ida,GFP_KERNEL ); +if (index <0 ) +return index; + pdata->cpu_map[index] =topology_core_id (cpu); + attr_no = index +BASE_SYSFS_ATTR_NO ; + }
Unbalanced curly braces.
drivers/hwmon/coretemp.c static int create_core_data(struct platform_device *pdev, unsigned int cpu, · line 456
Quoted 2 lines
drivers/hwmon/coretemp.c +1
-1-int err, attr_no; +int err, index, attr_no;
So it's 'index' here.
Review static void coretemp_remove_core(struct platform_data *pdata, int indx)
Quoted 1 line
(quoted diff) +0
-0@@ -524,6 +538,8 @@ static void coretemp_remove_core(struct platform_data *pdata, int indx)
But 'indx' here.
drivers/hwmon/coretemp.c static int coretemp_cpu_offline(unsigned int cpu) · line 668
Quoted 2 lines
drivers/hwmon/coretemp.c +1
-1-int indx, target; +int i, indx =-1 , target;
And 'indx' again. Did we run out of the letter 'e'? Either use 'index' naming consistently, or 'idx' if it has to be abbreviated.
Thanks,
Ingo
Re: [PATCH 3/7] hwmon/coretemp: Handle large core id value
Quoted 29 lines · Ingo Molnar
Quoted 8 lines · Zhang Rui
The coretemp driver supports up to a hard-coded limit of 128 cores.
Today, the driver can not support a core with an id above that limit.
Yet, the encoding of core_id's is arbitrary (BIOS APIC-id) and so they
may be sparse and they may be large.Quoted 4 lines
Update the driver to map arbitrary
core_idnumbers into appropriate array indexes so that 128 cores can be supported, no matter the encoding
of core_ids's.Please capitalize 'ID' consistently throughout the series.
Quoted 10 lines
- attr_no =
pkg_flag?PKG_SYSFS_ATTR_NO:TO_ATTR_NO(cpu); + if (pkg_flag)
+ attr_no =PKG_SYSFS_ATTR_NO;
+ else {
+ index =ida_alloc(&pdata->ida, GFP_KERNEL); + if (index < 0)
+ return index;
+ pdata->cpu_map[index] =topology_core_id(cpu); + attr_no = index +BASE_SYSFS_ATTR_NO;
+ }
Quoted 1 line
Unbalanced curly braces.
Sure, will fix these two issues in next version.
Quoted 11 lines
Quoted 3 lines
- int err,
attr_no;
+ int err, index,attr_no;So it's 'index' here.
Review static void coretemp_remove_core(struct · line 538
Quoted 2 lines
(quoted diff) +0
-0@@ -524,6 +538,8 @@ static void coretemp_remove_core(struct platform_data *pdata, int indx)But 'indx' here.
Quoted 6 lines
drivers/hwmon/coretemp.c static int coretemp_cpu_offline(unsigned int cpu) · line 668
Quoted 2 lines
drivers/hwmon/coretemp.c +1
-1- int indx, target; + int i, indx = -1, target;And 'indx' again. Did we run out of the letter 'e'? Either use 'index'
naming consistently, or 'idx' if it has to be abbreviated.
I'd prefer 'index', but here, this 'indx' is from previous code and this patch just initializes it to -1.
thanks,
rui
Quoted 4 lines
Thanks,
Ingo
Re: [PATCH 3/7] hwmon/coretemp: Handle large core id value
Quoted 48 lines · Zhang Rui
Quoted 29 lines · Ingo Molnar
Quoted 8 lines · Zhang Rui
The coretemp driver supports up to a hard-coded limit of 128 cores.
Today, the driver can not support a core with an id above that limit.
Yet, the encoding of core_id's is arbitrary (BIOS APIC-id) and so they
may be sparse and they may be large.Quoted 4 lines
Update the driver to map arbitrary
core_idnumbers into appropriate array indexes so that 128 cores can be supported, no matter the encoding
of core_ids's.Please capitalize 'ID' consistently throughout the series.
Quoted 10 lines
- attr_no =
pkg_flag?PKG_SYSFS_ATTR_NO:TO_ATTR_NO(cpu); + if (pkg_flag)
+ attr_no =PKG_SYSFS_ATTR_NO;
+ else {
+ index =ida_alloc(&pdata->ida, GFP_KERNEL); + if (index < 0)
+ return index;
+ pdata->cpu_map[index] =topology_core_id(cpu); + attr_no = index +BASE_SYSFS_ATTR_NO;
+ }Quoted 1 line
Unbalanced curly braces.
Sure, will fix these two issues in next version.
Quoted 14 lines
Quoted 3 lines
- int err,
attr_no;
+ int err, index,attr_no;So it's 'index' here.
Review static void coretemp_remove_core(struct · line 538
Quoted 2 lines
(quoted diff) +0
-0@@ -524,6 +538,8 @@ static void coretemp_remove_core(struct platform_data *pdata, int indx)But 'indx' here.
Quoted 2 lines
drivers/hwmon/coretemp.c +1
-1- int indx, target; + int i, indx = -1, target;
Quoted 6 lines
Quoted 3 lines
And 'indx' again. Did we run out of the letter 'e'? Either use 'index'
naming consistently, or 'idx' if it has to be abbreviated.I'd prefer 'index', but here, this 'indx' is from previous code and this patch just initializes it to -1.
Then queue up a cleanup patch as patch #1 - but idiosyncratic noise like that makes review harder.
Thanks,
Ingo
PATCH 4/7 x86/topology: Fix max_siblings calculation
In reply to Zhang RuiThe max siblings value returned by CPUID.1F SMT level EBX differs among CPUs on Intel Hybrid platforms like ADL-S/P.
It returns 2 for Pcore CPUs which have HT sibling and 1 for Ecore CPUs which do not.
Today, CPUID SMT level EBX sets the global variable smp_num_siblings. Thus, smp_num_siblings is overridden to different values based on the CPU Pcore/Ecore enumeration order.
For example,
[ 0.201005] detect_extended_topology: CPU APICID 0x0, smp_num_siblings 2, x86_max_cores 10 [ 0.201117] start_kernel->check_bugs->cpu_smt_check_topology: smp_num_siblings 2 ... [ 0.010146] detect_extended_topology: CPU APICID 0x8, smp_num_siblings 2, x86_max_cores 10 ... [ 0.010146] detect_extended_topology: CPU APICID 0x39, smp_num_siblings 2, x86_max_cores 10 [ 0.010146] detect_extended_topology: CPU APICID 0x48, smp_num_siblings 1, x86_max_cores 20 ... [ 0.010146] detect_extended_topology: CPU APICID 0x4e, smp_num_siblings 1, x86_max_cores 20 [ 2.583800] sched_set_itmt_core_prio: smp_num_siblings 1
This inconsistency brings several potential issues:
1. some kernel configuration like cpu_smt_control, as set in start_kernel()->check_bugs()->cpu_smt_check_topology(), depends on smp_num_siblings set by cpu0.
It is pure luck that all the current hybrid platforms use Pcore as cpu0 and hide this problem.
2. some per CPU data like cpuinfo_x86.x86_max_cores that depends on smp_num_siblings becomes inconsistent and bogus.
3. the final smp_num_siblings value after boot depends on the last CPU enumerated, which could either be Pcore or Ecore CPU.
The solution is to use CPUID EAX bits_shift to get the maximum number of addressable logical processors, and use this to determin max siblings. Because:
1. the CPUID EAX bits_shift values are consistent among CPUs as far as observed.
2. some code already uses smp_num_siblings value to isolate the SMT ID bits in APIC-ID, like apic_id_is_primary_thread().
--- arch/x86/kernel/cpu/topology.c | 17 +++++++++++------ 1 file changed, 11 insertions(+), 6 deletions(-)
arch/x86/kernel/cpu/topology.c +11 -6
@@ -23,7 +23,12 @@#define LEAFB_SUBTYPE (ecx) (((ecx) >> 8) & 0xff)#define BITS_SHIFT_NEXT_LEVEL (eax) ((eax) & 0x1f) -#define LEVEL_MAX_SIBLINGS (ebx) ((ebx) & 0xffff) + +/* +* Use EAX bit_shift to calculate the maximum number of addressable logical +* processors sharing the current level. +*/ +#define LEVEL_MAX_SIBLINGS (eax) (1 << BITS_SHIFT_NEXT_LEVEL(eax))unsigned int __max_die_per_package__read_mostly =1 ;EXPORT_SYMBOL (__max_die_per_package);
@@ -79,7 +84,7 @@ int detect_extended_topology_early(struct cpuinfo_x86 *c)* initial apic id, which also represents 32-bit extended x2apic id. */ c->initial_apicid = edx; - smp_num_siblings =LEVEL_MAX_SIBLINGS (ebx); + smp_num_siblings =LEVEL_MAX_SIBLINGS (eax);#endif return 0 ; }
@@ -109,9 +114,9 @@ int detect_extended_topology(struct cpuinfo_x86 *c)*/ cpuid_count (leaf,SMT_LEVEL , &eax, &ebx, &ecx, &edx); c->initial_apicid = edx; - core_level_siblings = smp_num_siblings =LEVEL_MAX_SIBLINGS (ebx); + core_level_siblings = smp_num_siblings =LEVEL_MAX_SIBLINGS (eax); core_plus_mask_width = ht_mask_width =BITS_SHIFT_NEXT_LEVEL (eax); - die_level_siblings =LEVEL_MAX_SIBLINGS (ebx); + die_level_siblings =LEVEL_MAX_SIBLINGS (eax); pkg_mask_width = die_plus_mask_width =BITS_SHIFT_NEXT_LEVEL (eax); sub_index =1 ;
@@ -122,14 +127,14 @@ int detect_extended_topology(struct cpuinfo_x86 *c)* Check for the Core type in the implemented sub leaves. */ if (LEAFB_SUBTYPE (ecx) ==CORE_TYPE ) { - core_level_siblings =LEVEL_MAX_SIBLINGS (ebx); + core_level_siblings =LEVEL_MAX_SIBLINGS (eax); core_plus_mask_width =BITS_SHIFT_NEXT_LEVEL (eax); die_level_siblings = core_level_siblings; die_plus_mask_width =BITS_SHIFT_NEXT_LEVEL (eax); }if (LEAFB_SUBTYPE (ecx) ==DIE_TYPE ) { die_level_present = true; - die_level_siblings =LEVEL_MAX_SIBLINGS (ebx); + die_level_siblings =LEVEL_MAX_SIBLINGS (eax); die_plus_mask_width =BITS_SHIFT_NEXT_LEVEL (eax); }
Signature · 1 line
2.34.1
PATCH 5/7 Documentation: x86: Update smp_num_siblings/x86_max_cores description
In reply to Zhang Ruismp_num_siblings/cpuinfo_x86.x86_max_cores are retrieved via CPUID EAX bit_shift value, and they represent the maximum possible number of threads in a core, and the maximum possible number of cores in a package.
Update the smp_num_siblings/cpuinfo_x86.x86_max_cores description in the documentation.
--- Documentation/x86/topology.rst | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-)
Documentation/x86/topology.rst +5 -4
@@ -49,7 +49,8 @@ AMD nomenclature for package is 'Node'. - cpuinfo_x86.x86_max_cores: - The number of cores in a package. This information is retrieved via CPUID. + The maximum possible number of cores in a package. This information is + retrieved via CPUID. - cpuinfo_x86.x86_max_dies:
@@ -102,10 +103,10 @@ AMDs nomenclature for a CMT core is "Compute Unit". The kernel always uses - smp_num_siblings: - The number of threads in a core. The number of threads in a package can be - calculated by:: + The maximum possible number of threads in a core. The maximum possible + number of threads in a package can be calculated by:: - threads_per_package = cpuinfo_x86.x86_max_cores * smp_num_siblings + maximum_threads_per_package = cpuinfo_x86.x86_max_cores * smp_num_siblings Threads
Signature · 1 line
2.34.1
PATCH 6/7 Documentation: x86: Remove obsolete x86_max_dies description
In reply to Zhang Ruicpuinfo_x86.x86_max_dies is introduced in commit 7745f03eb395 ("x86/topology: Add CPUID.1F multi-die/package support") and then removed in commit 14d96d6c06b5
("x86/topology: Create topology_max_die_per_package()").
Remove the obsolete cpuinfo_x86.x86_max_dies description.
--- Documentation/x86/topology.rst | 4 ---- 1 file changed, 4 deletions(-)
Documentation/x86/topology.rst +0 -4
@@ -52,10 +52,6 @@ AMD nomenclature for package is 'Node'. The maximum possible number of cores in a package. This information is retrieved via CPUID. - - cpuinfo_x86.x86_max_dies: - - The number of dies in a package. This information is retrieved via CPUID. - - cpuinfo_x86.cpu_die_id: The physical ID of the die. This information is retrieved via CPUID.
Signature · 1 line
2.34.1
PATCH 7/7 perf/x86/intel/P4: Fix smp_num_siblings usage
In reply to Zhang Ruismp_num_siblings can be larger than 2.
Any value larger than 1 suggests HT is supported.
--- arch/x86/include/asm/perf_event_p4.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/include/asm/perf_event_p4.h +1 -1
@@ -189,7 +189,7 @@ static inline int p4_ht_active(void)static inline int p4_ht_thread (int cpu) {#ifdef CONFIG_SMP -if (smp_num_siblings ==2 ) +if (smp_num_siblings >1 )return cpu !=cpumask_first (this_cpu_cpumask_var_ptr (cpu_sibling_map));#endif return 0 ;
Signature · 1 line
2.34.1
Re: [PATCH 7/7] perf/x86/intel/P4: Fix smp_num_siblings usage
Quoted 15 lines · Zhang Rui
smp_num_siblingscan be larger than 2.Any value larger than 1 suggests HT is supported.
Reviewed-by Len Brown <[email protected]>Signed-off-by Zhang Rui <[email protected]>--- arch/x86/include/asm/perf_event_p4.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)arch/x86/include/asm/perf_event_p4.h +0
-0
arch/x86/include/asm/perf_event_p4.h static inline int p4_ht_active(void) · line 193
Quoted 6 lines
arch/x86/include/asm/perf_event_p4.h +1
-1@@ -189,7 +189,7 @@ static inline int p4_ht_active(void)static inline int p4_ht_thread (int cpu) {#ifdef CONFIG_SMP -if (smp_num_siblings ==2 ) +if (smp_num_siblings >1 )return cpu !=cpumask_first (this_cpu_cpumask_var_ptr (cpu_sibling_map));
This fix too should probably come before all the other changes.
(Not that Pentium 4 code is expected to ever see such SMT thread values.)
Thanks,
Ingo
Re: [PATCH 7/7] perf/x86/intel/P4: Fix smp_num_siblings usage
Hi, Ingo,
Quoted 29 lines · Ingo Molnar
Quoted 21 lines · Zhang Rui
smp_num_siblingscan be larger than 2.Any value larger than 1 suggests HT is supported.
Reviewed-by Len Brown <[email protected]>Signed-off-by Zhang Rui <[email protected]>---
arch/x86/include/asm/perf_event_p4.h | 2 +-1 file changed, 1 insertion(+), 1 deletion(-)arch/x86/include/asm/perf_event_p4.h +1
-1@@ -189,7 +189,7 @@ static inline int p4_ht_active(void) static inline int p4_ht_thread(int cpu) { #ifdef CONFIG_SMP - if (smp_num_siblings == 2) + if (smp_num_siblings > 1)Quoted 2 lines
return cpu != cpumask_first(this_cpu_cpumask_var_ptr(cpu_sibling_map));
This fix too should probably come before all the other changes.
Quoted 3 lines
(Not that Pentium 4 code is expected to ever see such SMT thread values.)
Do you mean that this is a clean fix, and there is no reason for this patch to be blocked by any of the other patches in this series?
thanks,
rui
Re: [PATCH 7/7] perf/x86/intel/P4: Fix smp_num_siblings usage
In reply to Zhang RuiQuoted 1 line · Zhang Rui
smp_num_siblingscan be larger than 2.
Not on a P4 it can't ;-)
Quoted 14 lines
Any value larger than 1 suggests HT is supported.
Reviewed-by Len Brown <[email protected]>Signed-off-by Zhang Rui <[email protected]>--- arch/x86/include/asm/perf_event_p4.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)arch/x86/include/asm/perf_event_p4.h +0
-0
arch/x86/include/asm/perf_event_p4.h static inline int p4_ht_active(void) · line 195
Quoted 8 lines
arch/x86/include/asm/perf_event_p4.h +1
-1@@ -189,7 +189,7 @@ static inline int p4_ht_active(void)static inline int p4_ht_thread (int cpu) {#ifdef CONFIG_SMP -if (smp_num_siblings ==2 ) +if (smp_num_siblings >1 )return cpu !=cpumask_first (this_cpu_cpumask_var_ptr (cpu_sibling_map));#endif return 0 ;
Unless Intel plans to respin an P4 with extra siblings on, I don't think this qualifies for the word 'fix'.
Re: [PATCH 7/7] perf/x86/intel/P4: Fix smp_num_siblings usage
Quoted 4 lines · Peter Zijlstra
Quoted 1 line · Zhang Rui
smp_num_siblingscan be larger than 2.Not on a P4 it can't ;-)
Kernel code doesn't prevent this from happening, so it just depends on how SMT ID is encoded in APICID.
Checking for smp_num_sibling > 1 is the right logic to detect HT support, which is followed by all other kernel code except this one. :)
thanks,
rui
Quoted 28 lines
Quoted 23 lines
Any value larger than 1 suggests HT is supported.
Reviewed-by Len Brown <[email protected]>Signed-off-by Zhang Rui <[email protected]>--- arch/x86/include/asm/perf_event_p4.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)arch/x86/include/asm/perf_event_p4.h +1
-1@@ -189,7 +189,7 @@ static inline int p4_ht_active(void)static inline int p4_ht_thread (int cpu) {#ifdef CONFIG_SMP -if (smp_num_siblings ==2 ) +if (smp_num_siblings >1 )return cpu !=cpumask_first (this_cpu_cpumask_var_ptr (cpu_sibling_map));#endif Quoted 1 line
return 0;
Unless Intel plans to respin an P4 with extra siblings on, I don't think
this qualifies for the word 'fix'.
Re: [PATCH 7/7] perf/x86/intel/P4: Fix smp_num_siblings usage
Quoted 8 lines · Zhang Rui
Quoted 4 lines · Peter Zijlstra
Quoted 1 line · Zhang Rui
smp_num_siblingscan be larger than 2.Not on a P4 it can't ;-)
Kernel code doesn't prevent this from happening, so it just depends on how SMT ID is encoded in APICID.
No P4 ever encountered had it different and since no P4 will ever be made again (I sincerely hope) we have a complete case.
Quoted 2 lines
Checking for
smp_num_sibling> 1 is the right logic to detect HT support, which is followed by all other kernel code except this one. :)
I'm fine with making it consistent, I'm arguing with the subject calling this a fix, it is not, it's a functional no-op.
Quoted 24 lines · Zhang Rui
On Intel AlderLake-N platforms where there are Ecores only, the Ecore Module topology is enumerated via CPUID.1F Module level, which has not been supported by Linux kernel yet.
This exposes two issues in current CPUID.1F handling code.
1. Linux interprets the Module id bits as package id and erroneously reports a multi module system as a multi-package system.
2. Linux excludes the unknown Module id bits from thecore_id, and results in duplicate core_id’s shown in a package after the first issue solved.Plus that, a third problem is observed on Intel Hybrid ADL-S/P platforms. The return value of CPUID.1F SMT level EBX (number of siblings) differs on Pcore CPUs and Ecore CPUs, and results in inconsistent
smp_num_siblingsvalue based on the Pcore/Ecore CPU enumeration order. This could bring some potential issues although we have not observed any functionalities issues so far.Patch 1/7 and 2/7 fix the first two issues. And at the same time, it reveals a reality that the
core_idcould be sparse on platforms with CPUID.1F support.
Patch 3/7 improves coretemp driver code to be able to handle sparse core id, which is the only driver that usescore_idas array index and run on platforms with CPUID.1F support.
Quoted 2 lines
Patch 4/7 to 7/7 propose a fix for the third problem and update the related Documents.
Yeah, so patch 3/7 probably needs to come first - otherwise there's a window for bisection breakage.
Thanks,
Ingo
Hi, Ingo,
Thanks for reviewing this patch series.
Quoted 41 lines · Ingo Molnar
Quoted 38 lines · Zhang Rui
On Intel AlderLake-N platforms where there are Ecores only, the Ecore
Module topology is enumerated via CPUID.1F Module level, which has not
been supported by Linux kernel yet.This exposes two issues in current CPUID.1F handling code.
1. Linux interprets the Module id bits as package id and erroneously
reports a multi module system as a multi-package system.
2. Linux excludes the unknown Module id bits from thecore_id, and results
in duplicate core_id’s shown in a package after the first issue solved.Plus that, a third problem is observed on Intel Hybrid ADL-S/P platforms.
The return value of CPUID.1F SMT level EBX (number of siblings) differs on
Pcore CPUs and Ecore CPUs, and results in inconsistentsmp_num_siblings
value based on the Pcore/Ecore CPU enumeration order. This could bring
some potential issues although we have not observed any functionalities
issues so far.Patch 1/7 and 2/7 fix the first two issues. And at the same time, it
reveals a reality that thecore_idcould be sparse on platforms with
CPUID.1F support.
Patch 3/7 improves coretemp driver code to be able to handle sparse core
id, which is the only driver that usescore_idas array index and run on
platforms with CPUID.1F support.
Quoted 5 lines
Quoted 2 lines
Patch 4/7 to 7/7 propose a fix for the third problem and update the related Documents.
Yeah, so patch 3/7 probably needs to come first - otherwise there's a window for bisection breakage.
Sure, I will re-arrange this.
thanks,
rui