Skip to contentSkip to conversation

voyager: fix cpu bootmaps

PATCH In review James Bottomley 2 participants 4 messages Jan 30, 2009 – Feb 3, 2009

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 voyager: fix cpu bootmaps

commit 98a79d6a50181ca1ecf7400eda01d5dc1bc0dbf0

Author: Rusty Russell <[email protected]>
Date:   Sat Dec 13 21:19:41 2008 +1030

cpumask: centralize cpu_online_map and cpu_possible_map

Broke voyager largely because it currently initialises the possible_map by copying, which isn't possible in the new scheme. Fix this by using init_cpu_possible() instead and tidy up other of the cpumask declarations which now have global variables.

---
 arch/x86/mach-voyager/voyager_smp.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/mach-voyager/voyager_smp.c +1 -1
index 9840b7e..dcc07d5 100644
@@ -378,7 +378,7 @@ void __init find_smp_config(void)
 	cpus_addr(phys_cpu_present_map)[0] |=
 	    voyager_extended_cmos_read(VOYAGER_PROCESSOR_PRESENT_MASK +
 				       3) << 24;
-	cpu_possible_map = phys_cpu_present_map;
+	init_cpu_possible(&phys_cpu_present_map);
 	printk("VOYAGER SMP: phys_cpu_present_map = 0x%lx\n",
 	       cpus_addr(phys_cpu_present_map)[0]);
 	/* Here we set up the VIC to enable SMP */
Signature · 1 line
1.5.6.6
#1 lore raw
Quoted 21 lines · On Saturday 31 January 2009 05:01:57 James Bottomley wrote:

commit 98a79d6a50181ca1ecf7400eda01d5dc1bc0dbf0

Author: Rusty Russell <[email protected]>
Date:   Sat Dec 13 21:19:41 2008 +1030

cpumask: centralize cpu_online_map and cpu_possible_map

Broke voyager largely because it currently initialises the possible_map by copying, which isn't possible in the new scheme. Fix this by using init_cpu_possible() instead and tidy up other of the cpumask declarations which now have global variables.

---
 arch/x86/mach-voyager/voyager_smp.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/mach-voyager/voyager_smp.c +0 -0
index 9840b7e..dcc07d5 100644

arch/x86/mach-voyager/voyager_smp.c void __init find_smp_config(void) · line 384

Quoted 8 lines
arch/x86/mach-voyager/voyager_smp.c +1 -1
@@ -378,7 +378,7 @@ void __init find_smp_config(void)
 	cpus_addr(phys_cpu_present_map)[0] |=
 	    voyager_extended_cmos_read(VOYAGER_PROCESSOR_PRESENT_MASK +
 				       3) << 24;
-	cpu_possible_map = phys_cpu_present_map;
+	init_cpu_possible(&phys_cpu_present_map);
 	printk("VOYAGER SMP: phys_cpu_present_map = 0x%lx\n",
 	       cpus_addr(phys_cpu_present_map)[0]);
 	/* Here we set up the VIC to enable SMP */

Strange, the assignment should still work, even though this new method is preferred.

How do your patches normally get upstream? I'd normally just fwd this to Ingo...

Sorry I broke your platform...
Rusty.

#2 lore raw
Quoted 17 lines · Rusty Russell
Quoted 16 lines · On Saturday 31 January 2009 05:01:57 James Bottomley wrote:

commit 98a79d6a50181ca1ecf7400eda01d5dc1bc0dbf0

Author: Rusty Russell <[email protected]>
Date:   Sat Dec 13 21:19:41 2008 +1030

cpumask: centralize cpu_online_map and cpu_possible_map

Broke voyager largely because it currently initialises the possible_map by copying, which isn't possible in the new scheme. Fix this by using init_cpu_possible() instead and tidy up other of the cpumask declarations which now have global variables.

---
 arch/x86/mach-voyager/voyager_smp.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
Quoted 16 lines

arch/x86/mach-voyager/voyager_smp.c void __init find_smp_config(void) · line 384

Quoted 13 lines
arch/x86/mach-voyager/voyager_smp.c +1 -1
index 9840b7e..dcc07d5 100644
@@ -378,7 +378,7 @@ void __init find_smp_config(void)
 	cpus_addr(phys_cpu_present_map)[0] |=
 	    voyager_extended_cmos_read(VOYAGER_PROCESSOR_PRESENT_MASK +
 				       3) << 24;
-	cpu_possible_map = phys_cpu_present_map;
+	init_cpu_possible(&phys_cpu_present_map);
 	printk("VOYAGER SMP: phys_cpu_present_map = 0x%lx\n",
 	       cpus_addr(phys_cpu_present_map)[0]);
 	/* Here we set up the VIC to enable SMP */

Strange, the assignment should still work, even though this new method is preferred.

I know, it's weird ... it's like the compiler is copying to the lvalue but then discarding the result somehow ... with the direct assignment, we end up with nothing in cpu_possible_bits, then the smp alternatives switch to UP and all hell breaks loose when multiple CPUs are brought up.

I think the problem is that cpu_possible_mask isn't an alias for cpu_possible_bits, so when I update the former directly it breaks the correspondence set up in kernel/cpu.c ... your declaration of it as *const looks to be causing the compiler to do this.

Thus, I think direct assignment to cpu_possible_map (and hence cpu_possible_mask) should be forbidden.

Quoted 4 lines

How do your patches normally get upstream? I'd normally just fwd this to Ingo...

Sorry I broke your platform...

That's OK ... Voyager is one of those platforms that tends to come off worst on updates, so I'm used to sweeping up around it.

James

#3 lore raw
Quoted 4 lines · On Sunday 01 February 2009 01:34:53 James Bottomley wrote:
Quoted 3 lines · Rusty Russell

On Saturday 31 January 2009 05:01:57 James Bottomley wrote: 3) << 24;

Quoted 2 lines
(quoted diff) +1 -1
-	cpu_possible_map = phys_cpu_present_map;
+	init_cpu_possible(&phys_cpu_present_map);

...

Quoted 5 lines
Quoted 2 lines

Strange, the assignment should still work, even though this new method is preferred.

I know, it's weird ... it's like the compiler is copying to the lvalue but then discarding the result somehow

This makes me nervous. I may have broken other archs.

I'm testing the following patch now.

Thanks,
Rusty.

cpumask: avoid cast-away-const for deprecated cpu_*_map.

James Bottomley encountered a Voyager bug where gcc was discarding changes to cpu_possible_map, which is now #defined to cpu_possible_mask with the const cast away.

The modern way is to use init_cpu_possible or set_cpu_possible, but not everyone is converted yet. I'm concerned that other archs might have similar problems, so let's expose the underlying bits for the moment (and prepend those names with __).

(This also makes accessing cpu_*_mask slightly more efficient, rather than going via a pointer).

---
 include/linux/cpumask.h |   28 +++++++++++++++++++---------
 kernel/cpu.c            |   44 ++++++++++++++++++++------------------------
 2 files changed, 39 insertions(+), 33 deletions(-)
include/linux/cpumask.h +19 -9
@@ -482,16 +482,26 @@ int __next_cpu_nr(int n, const cpumask_t
  *    only one CPU.
  */
 
-extern const struct cpumask *const cpu_possible_mask;
-extern const struct cpumask *const cpu_online_mask;
-extern const struct cpumask *const cpu_present_mask;
-extern const struct cpumask *const cpu_active_mask;
+#define cpu_possible_mask \
+	((const struct cpumask *)to_cpumask(__cpu_possible_bits))
+#define cpu_online_mask \
+	((const struct cpumask *)to_cpumask(__cpu_online_bits))
+#define cpu_present_mask \
+	((const struct cpumask *)to_cpumask(__cpu_present_bits))
+#define cpu_active_mask \
+	((const struct cpumask *)to_cpumask(__cpu_active_bits))
 
-/* These strip const, as traditionally they weren't const. */
-#define cpu_possible_map	(*(cpumask_t *)cpu_possible_mask)
-#define cpu_online_map		(*(cpumask_t *)cpu_online_mask)
-#define cpu_present_map		(*(cpumask_t *)cpu_present_mask)
-#define cpu_active_map		(*(cpumask_t *)cpu_active_mask)
+/* Deprecated non-const versions. */
+#define cpu_possible_map	(*to_cpumask(__cpu_possible_bits))
+#define cpu_online_map		(*to_cpumask(__cpu_online_bits))
+#define cpu_present_map		(*to_cpumask(__cpu_present_bits))
+#define cpu_active_map		(*to_cpumask(__cpu_active_bits))
+
+/* Don't use these directly: use cpu_*_mask, set_cpu_* or init_cpu_*. */
+extern unsigned long __cpu_possible_bits[];
+extern unsigned long __cpu_online_bits[];
+extern unsigned long __cpu_present_bits[];
+extern unsigned long __cpu_active_bits[];
 
 #if NR_CPUS > 1
 #define num_online_cpus()	cpumask_weight(cpu_online_mask)
kernel/cpu.c +20 -24

89 lines not on this page · raw message