Skip to contentSkip to conversation

Hot plug CPU to support physical add of new processors (i386)

PATCH In review [email protected] 5 participants 6 messages Sep 1, 2005 CPU hotplug

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Hot plug CPU to support physical add of new processors (i386)

Current IA32 CPU hotplug code doesn't allow bringing up processors that were not present in the boot configuration. To make existing hot plug facility more practical for physical hot plug, possible processors should be encountered during boot for potentual hot add/replace/remove. On ES7000, ACPI marks all the sockets that are empty or not assigned to the partitionas as "disabled". The patch allows arrays/masks with APIC info for disabled processors to be initialized. Then the OS can bring up a processor that was inserted in the socked and brought into configuration during runtime.
To test the code, one can boot the system with maxcpu=1 and then bring the rest of the processors up, which was not possible so far (only maxcpus number of nodes were created).
The patch also makes proc entry for interrupts dynamically change to only show current onlined processors.

---

 arch/i386/kernel/acpi/boot.c      |    2 ++
 arch/i386/kernel/irq.c            |    8 ++++----
 arch/i386/kernel/mpparse.c        |    4 ++++
 arch/i386/kernel/smpboot.c        |    3 ++-
 arch/i386/mach-default/topology.c |    4 ++--
 kernel/cpu.c                      |    4 ++++
 6 files changed, 18 insertions(+), 7 deletions(-)
arch/i386/kernel/acpi/boot.c renamed from arch/i386/kernel/acpi/boot.c~hotcpu-i386 +2 -0
diff -puN arch/i386/kernel/acpi/boot.c~hotcpu-i386 arch/i386/kernel/acpi/boot.c
@@ -255,9 +255,11 @@ acpi_parse_lapic(acpi_table_entry_header
 
 	acpi_table_print_madt_entry(header);
 
+#ifndef CONFIG_HOTPLUG_CPU
 	/* no utility in registering a disabled processor */
 	if (processor->flags.enabled == 0)
 		return 0;
+#endif
 
 	x86_acpiid_to_apicid[processor->acpi_id] = processor->id;
 
arch/i386/kernel/mpparse.c renamed from arch/i386/kernel/mpparse.c~hotcpu-i386 +4 -0
diff -puN arch/i386/kernel/mpparse.c~hotcpu-i386 arch/i386/kernel/mpparse.c
@@ -125,8 +125,10 @@ static void __init MP_processor_info (st
  	int ver, apicid, cpu, found_bsp = 0;
 	physid_mask_t tmp;
  	
+#ifndef CONFIG_HOTPLUG_CPU
 	if (!(m->mpc_cpuflag & CPU_ENABLED))
 		return;
+#endif
 
 	apicid = mpc_apic_id(m, translation_table[mpc_record]);
 
@@ -190,11 +192,13 @@ static void __init MP_processor_info (st
 		return;
 	}
 
+#ifndef CONFIG_HOTPLUG_CPU
 	if (num_processors >= maxcpus) {
 		printk(KERN_WARNING "WARNING: maxcpus limit of %i reached."
 			" Processor ignored.\n", maxcpus); 
 		return;
 	}
+#endif
 	num_processors++;
 	ver = m->mpc_apicver;
 
arch/i386/kernel/smpboot.c renamed from arch/i386/kernel/smpboot.c~hotcpu-i386 +2 -1
diff -puN arch/i386/kernel/smpboot.c~hotcpu-i386 arch/i386/kernel/smpboot.c
@@ -1003,9 +1003,10 @@ int __devinit smp_prepare_cpu(int cpu)
 	struct warm_boot_cpu_info info;
 	struct work_struct task;
 	int	apicid, ret;
+	extern u8 bios_cpu_apicid[NR_CPUS];
 
 	lock_cpu_hotplug();
-	apicid = x86_cpu_to_apicid[cpu];
+	apicid = bios_cpu_apicid[cpu];
 	if (apicid == BAD_APICID) {
 		ret = -ENODEV;
 		goto exit;
arch/i386/mach-default/topology.c renamed from arch/i386/mach-default/topology.c~hotcpu-i386 +2 -2
diff -puN arch/i386/mach-default/topology.c~hotcpu-i386 arch/i386/mach-default/topology.c
@@ -76,7 +76,7 @@ static int __init topology_init(void)
 	for_each_online_node(i)
 		arch_register_node(i);
 
-	for_each_present_cpu(i)
+	for_each_cpu(i)
 		arch_register_cpu(i);
 	return 0;
 }
@@ -87,7 +87,7 @@ static int __init topology_init(void)
 {
 	int i;
 
-	for_each_present_cpu(i)
+	for_each_cpu(i)
 		arch_register_cpu(i);
 	return 0;
 }
kernel/cpu.c renamed from kernel/cpu.c~hotcpu-i386 +4 -0
diff -puN kernel/cpu.c~hotcpu-i386 kernel/cpu.c
@@ -158,7 +158,11 @@ int __devinit cpu_up(unsigned int cpu)
 	if ((ret = down_interruptible(&cpucontrol)) != 0)
 		return ret;
 
+#ifdef CONFIG_HOTPLUG_CPU
+	if (cpu_online(cpu)) {
+#else
 	if (cpu_online(cpu) || !cpu_present(cpu)) {
+#endif
 		ret = -EINVAL;
 		goto out;
 	}
arch/i386/kernel/irq.c renamed from arch/i386/kernel/irq.c~hotcpu-i386 +4 -4
diff -puN arch/i386/kernel/irq.c~hotcpu-i386 arch/i386/kernel/irq.c
@@ -248,7 +248,7 @@ int show_interrupts(struct seq_file *p, 
 
 	if (i == 0) {
 		seq_printf(p, "           ");
-		for_each_cpu(j)
+		for_each_online_cpu(j)
 			seq_printf(p, "CPU%d       ",j);
 		seq_putc(p, '\n');
 	}
@@ -262,7 +262,7 @@ int show_interrupts(struct seq_file *p, 
 #ifndef CONFIG_SMP
 		seq_printf(p, "%10u ", kstat_irqs(i));
 #else
-		for_each_cpu(j)
+		for_each_online_cpu(j)
 			seq_printf(p, "%10u ", kstat_cpu(j).irqs[i]);
 #endif
 		seq_printf(p, " %14s", irq_desc[i].handler->typename);
@@ -276,12 +276,12 @@ skip:
 		spin_unlock_irqrestore(&irq_desc[i].lock, flags);
 	} else if (i == NR_IRQS) {
 		seq_printf(p, "NMI: ");
-		for_each_cpu(j)
+		for_each_online_cpu(j)
 			seq_printf(p, "%10u ", nmi_count(j));
 		seq_putc(p, '\n');
 #ifdef CONFIG_X86_LOCAL_APIC
 		seq_printf(p, "LOC: ");
-		for_each_cpu(j)
+		for_each_online_cpu(j)
 			seq_printf(p, "%10u ",
 				per_cpu(irq_stat,j).apic_timer_irqs);
 		seq_putc(p, '\n');

_

#1 lore raw

Hi,

Quoted 8 lines · On Wed, 2005-08-31 at 20:13 +0800, [email protected] wrote:

Current IA32 CPU hotplug code doesn't allow bringing up processors that were not present in the boot configuration.
To make existing hot plug facility more practical for physical hot plug, possible processors should be encountered
during boot for potentual hot add/replace/remove. On ES7000, ACPI marks all the sockets that are empty or not assigned
to the partitionas as "disabled". The patch allows arrays/masks with APIC info for disabled processors to be

Quoted 8 lines

initialized. Then the OS can bring up a processor that was inserted in the socked and brought into configuration
during runtime.
To test the code, one can boot the system with maxcpu=1 and then bring the rest of the processors up, which was not
possible so far (only maxcpus number of nodes were created). The patch also makes proc entry for interrupts dynamically change to only show current onlined processors.

Could we clean up the cpu_present_map a bit, like IA64 does? Eg, if a new CPU is inserted, we allocated cpu id for it and set cpu_present_map. Current alloc_cpu_id is just a workaround for suspend/resume use, but isn't ok for physical cpu hotplug to me.

Thanks,
Shaohua

#2 lore raw
In reply to [email protected]

Hallo Natalie,

Quoted 6 lines · On Wednesday 31 August 2005 14:13, [email protected] wrote:

Current IA32 CPU hotplug code doesn't allow bringing up processors that were not present in the boot configuration. To make existing hot plug facility more practical for physical hot plug, possible processors should be encountered during boot for potentual hot add/replace/remove. On ES7000, ACPI marks all the sockets that are empty or not assigned to the partitionas as "disabled".

Good idea. In fact I always hated the behaviour of the existing hotplug code that assumes all possible CPUs can be hotplugged. It would be much nicer to be told be the firmware what CPUs are hotpluggable. It would be great if all ia32/x86-64 hotplug capable BIOS behaved like your.

arch/i386/kernel/smpboot.c int __devinit smp_prepare_cpu(int cpu) · line 1006

Quoted 4 lines
arch/i386/kernel/smpboot.c renamed from arch/i386/kernel/smpboot.c~hotcpu-i386 +1 -0
 	struct warm_boot_cpu_info info;
 	struct work_struct task;
 	int	apicid, ret;
+	extern u8 bios_cpu_apicid[NR_CPUS];

This should be in some header.

arch/i386/kernel/smpboot.c int __devinit smp_prepare_cpu(int cpu) · line 1009

Quoted 4 lines
arch/i386/kernel/smpboot.c renamed from arch/i386/kernel/smpboot.c~hotcpu-i386 +1 -1
 	lock_cpu_hotplug();
-	apicid = x86_cpu_to_apicid[cpu];
+	apicid = bios_cpu_apicid[cpu];

Why this change? It seems unrelated.

Quoted 10 lines
arch/i386/kernel/smpboot.c renamed from arch/i386/kernel/smpboot.c~hotcpu-i386 +0 -0
 	if (apicid == BAD_APICID) {
 		ret = -ENODEV;
 		goto exit;

diff -puN arch/i386/mach-default/topology.c~hotcpu-i386 arch/i386/mach-default/topology.c ---
linux-2.6.13-rc6-mm2/arch/i386/mach-default/topology.c~hotcpu-i386 2005-08-
31 04:17:20.957019600 -0700 +++
linux-2.6.13-rc6-mm2-root/arch/i386/mach-default/topology.c 2005-08-31
04:22:13.020619184 -0700 @@ -76,7 +76,7 @@ static int __init topology_init(void)

arch/i386/mach-default/topology.c static int __init topology_init(void) · line 79

Quoted 5 lines
arch/i386/mach-default/topology.c renamed from arch/i386/mach-default/topology.c~hotcpu-i386 +1 -1
 	for_each_online_node(i)
 		arch_register_node(i);
 
-	for_each_present_cpu(i)
+	for_each_cpu(i)

This looks wrong. The CPUs should be in the present mask if it's present. Followup code similar.

BTW shouldn't there be some attribute in sysfs that says "CPU disabled"?

-Andi

#3 lore raw
Quoted 10 lines · Andi Kleen

Hallo Natalie,

Quoted 6 lines · On Wednesday 31 August 2005 14:13, [email protected] wrote:

Current IA32 CPU hotplug code doesn't allow bringing up processors that were not present in the boot configuration. To make existing hot plug facility more practical for physical hot plug, possible processors should be encountered during boot for potentual hot add/replace/remove. On ES7000, ACPI marks all the sockets that are empty or not assigned to the partitionas as "disabled".

Quoted 6 lines

Good idea. In fact I always hated the behaviour of the existing hotplug code that assumes all possible CPUs can be hotplugged. It would be much nicer to be told be the firmware what CPUs are hotpluggable. It would be great if all ia32/x86-64 hotplug capable BIOS behaved like your.

Andi, you are getting mixed up with software only ability to offline with hardware eject capability. ACPI indicates ability to hotplug by the presence of _EJD in the appropriate scope of the object. So ACPI does have ability to do what you mention above precicely, but the entire namespace is not known upfront since some could be dynamically loaded. Which is why we need to show the entire NR_CPUS as hotpluggable.

Possibly we can keep cpu_possible_map as NR_CPUS only when support for PHYSICAL_CPU_HOTPLUG is present, otherwise we can keep it cloned as cpu_present_map. (we dont have a generic PHYSICAL hotplug CONFIG option today)

What CONFIG_HOTPLUG_CPU=y indicates is ability to offline a processor from the kernel. It DOES NOT indicate physical hotpluggablity. So we dont need any hardware support (apart arch/kernel support) for this to work. Support for physical hotplug is indicated via CONFIG_ACPI_HOTPLUG_CPU.

Be aware that suspend/resume folks using CPU hotplug to offline CPUS except BSP need just the kernel support to offline. BIOS has nothing to do with being able to offline a CPU (preferably called as soft-removal).

Cheers,
ashok

#4 lore raw
In reply to [email protected]

kernel/cpu.c int __devinit cpu_up(unsigned int cpu) · line 168

Quoted 8 lines · [email protected] wrote:
kernel/cpu.c renamed from kernel/cpu.c~hotcpu-i386 +4 -0
+#ifdef CONFIG_HOTPLUG_CPU
+	if (cpu_online(cpu)) {
+#else
 	if (cpu_online(cpu) || !cpu_present(cpu)) {
+#endif
 		ret = -EINVAL;
 		goto out;
 	}

Why this change? I think the cpu_present check is needed for ppc64 since it has non-present cpus in sysfs.

#5 lore raw
In reply to [email protected]

On Wed, Aug 31, 2005 at 05:13:10AM -0700, [email protected] wrote:

Hi Natalie,

sorry for the late catchup...

Just to make sure we use the cpu maps properly here is the current definition iam not sure if you are breaking any of these assumptions, cause if you do you will break existing implementations... so pardon for the verbosity.

cpu_possible_map - must include all possible cpus that can ever exist in this
instance of boot. This map is static, i.e you can grow or remove bits without causing confution to the rest of the subsystem. We attempted to change, but there was not much in return for the savings, for the complexity we would add to make this map dynamic.

cpu_present_map - indicates that these are present and operational. OS
can decide if they should be brought up or not.

when CONFIG_HOTPLUG_CPU is *not* defined, then cpu_possible_map and cpu_present_map are the same. When CONFIG_HOTPLUG_CPU=y, then we extend cpu_possible_map to cover the range of NR_CPUS, but cpu_present_map still shows what is real.

When a physical hotplug happens, cpu_present_map is updated by who ever manages the physical hotplug.

Quoted 5 lines

Current IA32 CPU hotplug code doesn't allow bringing up processors that were not present in the boot configuration. To make existing hot plug facility more practical for physical hot plug, possible processors should be encountered during boot for potentual hot add/replace/remove. On ES7000, ACPI marks all the sockets that are empty or not assigned to the partitionas as "disabled". The patch allows arrays/masks with APIC info for disabled processors to be

This sounds like a cluge to me. The correct implementation would be you would need some sysmgmt deamon or something that works with the kernel to notify of new cpus and populate apicid and grow cpu_present_map. Making an assumption that disabled APICID are valid for ES7000 sake is not a safe assumption.

Quoted 21 lines

initialized. Then the OS can bring up a processor that was inserted in the socked and brought into configuration during runtime.
To test the code, one can boot the system with maxcpu=1 and then bring the rest of the processors up, which was not possible so far (only maxcpus number of nodes were created).
The patch also makes proc entry for interrupts dynamically change to only show current onlined processors.

---

 arch/i386/kernel/acpi/boot.c      |    2 ++
 arch/i386/kernel/irq.c            |    8 ++++----
 arch/i386/kernel/mpparse.c        |    4 ++++
 arch/i386/kernel/smpboot.c        |    3 ++-
 arch/i386/mach-default/topology.c |    4 ++--
 kernel/cpu.c                      |    4 ++++
 6 files changed, 18 insertions(+), 7 deletions(-)

diff -puN arch/i386/kernel/acpi/boot.c~hotcpu-i386 arch/i386/kernel/acpi/boot.c

arch/i386/kernel/acpi/boot.c acpi_parse_lapic(acpi_table_entry_header · line 262

Quoted 8 lines
arch/i386/kernel/acpi/boot.c renamed from arch/i386/kernel/acpi/boot.c~hotcpu-i386 +2 -0
@@ -255,9 +255,11 @@ acpi_parse_lapic(acpi_table_entry_header
 
 	acpi_table_print_madt_entry(header);
 
+#ifndef CONFIG_HOTPLUG_CPU
 	/* no utility in registering a disabled processor */
 	if (processor->flags.enabled == 0)
 		return 0;
+#endif

On same lines, the above config code seems not appropriate. When your platform needs to hot-add a cpu, find a way to set these flags, you cant break current behaviour as for a platform quirk!

Quoted 7 lines
arch/i386/kernel/acpi/boot.c renamed from arch/i386/kernel/acpi/boot.c~hotcpu-i386 +0 -0
 	x86_acpiid_to_apicid[processor->acpi_id] = processor->id;
 

diff -puN arch/i386/kernel/mpparse.c~hotcpu-i386 arch/i386/kernel/mpparse.c

arch/i386/kernel/mpparse.c static void __init MP_processor_info (st · line 131

Quoted 7 lines
arch/i386/kernel/mpparse.c renamed from arch/i386/kernel/mpparse.c~hotcpu-i386 +2 -0
@@ -125,8 +125,10 @@ static void __init MP_processor_info (st
  	int ver, apicid, cpu, found_bsp = 0;
 	physid_mask_t tmp;
 
+#ifndef CONFIG_HOTPLUG_CPU
 	if (!(m->mpc_cpuflag & CPU_ENABLED))
 		return;
+#endif

Same here... platform quirk!

Quoted 36 lines
arch/i386/kernel/mpparse.c renamed from arch/i386/kernel/mpparse.c~hotcpu-i386 +2 -0
 	apicid = mpc_apic_id(m, translation_table[mpc_record]);
 
@@ -190,11 +192,13 @@ static void __init MP_processor_info (st
 		return;
 	}
 
+#ifndef CONFIG_HOTPLUG_CPU
 	if (num_processors >= maxcpus) {
 		printk(KERN_WARNING "WARNING: maxcpus limit of %i reached."
 			" Processor ignored.\n", maxcpus); 
 		return;
 	}
+#endif
 	num_processors++;
 	ver = m->mpc_apicver;
 

diff -puN arch/i386/kernel/smpboot.c~hotcpu-i386 arch/i386/kernel/smpboot.c

arch/i386/kernel/smpboot.c renamed from arch/i386/kernel/smpboot.c~hotcpu-i386 +2 -1
@@ -1003,9 +1003,10 @@ int __devinit smp_prepare_cpu(int cpu)
 	struct warm_boot_cpu_info info;
 	struct work_struct task;
 	int	apicid, ret;
+	extern u8 bios_cpu_apicid[NR_CPUS];
 
 	lock_cpu_hotplug();
-	apicid = x86_cpu_to_apicid[cpu];
+	apicid = bios_cpu_apicid[cpu];
 	if (apicid == BAD_APICID) {
 		ret = -ENODEV;
 		goto exit;

diff -puN arch/i386/mach-default/topology.c~hotcpu-i386 arch/i386/mach-default/topology.c

arch/i386/mach-default/topology.c static int __init topology_init(void) · line 79

Quoted 5 lines
arch/i386/mach-default/topology.c renamed from arch/i386/mach-default/topology.c~hotcpu-i386 +1 -1
@@ -76,7 +76,7 @@ static int __init topology_init(void)
 	for_each_online_node(i)
 		arch_register_node(i);
 
-	for_each_present_cpu(i)
+	for_each_cpu(i)

Nope. Should still be for present_cpus. with NR_CPUS large we would see way too many files in sysfs than what is really available.

Quoted 16 lines
arch/i386/mach-default/topology.c renamed from arch/i386/mach-default/topology.c~hotcpu-i386 +1 -1
 		arch_register_cpu(i);
 	return 0;
 }
@@ -87,7 +87,7 @@ static int __init topology_init(void)
 {
 	int i;
 
-	for_each_present_cpu(i)
+	for_each_cpu(i)
 		arch_register_cpu(i);
 	return 0;
 }

diff -puN kernel/cpu.c~hotcpu-i386 kernel/cpu.c

kernel/cpu.c int __devinit cpu_up(unsigned int cpu) · line 165

Quoted 8 lines
kernel/cpu.c renamed from kernel/cpu.c~hotcpu-i386 +4 -0
@@ -158,7 +158,11 @@ int __devinit cpu_up(unsigned int cpu)
 	if ((ret = down_interruptible(&cpucontrol)) != 0)
 		return ret;
 
+#ifdef CONFIG_HOTPLUG_CPU
+	if (cpu_online(cpu)) {
+#else
 	if (cpu_online(cpu) || !cpu_present(cpu)) {
+#endif

ditto.....

Quoted 40 lines
kernel/cpu.c renamed from kernel/cpu.c~hotcpu-i386 +0 -0
 		ret = -EINVAL;
 		goto out;
 	}

diff -puN arch/i386/kernel/irq.c~hotcpu-i386 arch/i386/kernel/irq.c

arch/i386/kernel/irq.c renamed from arch/i386/kernel/irq.c~hotcpu-i386 +4 -4
@@ -248,7 +248,7 @@ int show_interrupts(struct seq_file *p, 
 
 	if (i == 0) {
 		seq_printf(p, "           ");
-		for_each_cpu(j)
+		for_each_online_cpu(j)
 			seq_printf(p, "CPU%d       ",j);
 		seq_putc(p, '\n');
 	}
@@ -262,7 +262,7 @@ int show_interrupts(struct seq_file *p, 
 #ifndef CONFIG_SMP
 		seq_printf(p, "%10u ", kstat_irqs(i));
 #else
-		for_each_cpu(j)
+		for_each_online_cpu(j)
 			seq_printf(p, "%10u ", kstat_cpu(j).irqs[i]);
 #endif
 		seq_printf(p, " %14s", irq_desc[i].handler->typename);
@@ -276,12 +276,12 @@ skip:
 		spin_unlock_irqrestore(&irq_desc[i].lock, flags);
 	} else if (i == NR_IRQS) {
 		seq_printf(p, "NMI: ");
-		for_each_cpu(j)
+		for_each_online_cpu(j)
 			seq_printf(p, "%10u ", nmi_count(j));
 		seq_putc(p, '\n');
 #ifdef CONFIG_X86_LOCAL_APIC
 		seq_printf(p, "LOC: ");
-		for_each_cpu(j)
+		for_each_online_cpu(j)
 			seq_printf(p, "%10u ",
 				per_cpu(irq_stat,j).apic_timer_irqs);
 		seq_putc(p, '\n');

_

Signature · 3 lines
Cheers,
Ashok Raj
- Open Source Technology Center