Re: [PATCH] Avoid mask based num_possible_cpus and num_online_cpus -v3

["Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com>]

On 01/24/2012 08:04 AM, Venkatesh Pallipadi wrote:

> Kernel's notion of possible cpus (from include/linux/cpumask.h)
>  *     cpu_possible_mask- has bit 'cpu' set iff cpu is populatable
> 
>  *  The cpu_possible_mask is fixed at boot time, as the set of CPU id's
>  *  that it is possible might ever be plugged in at anytime during the
>  *  life of that system boot.
> 
>  #define num_possible_cpus()     cpumask_weight(cpu_possible_mask)
> 
> and on x86 cpumask_weight() calls hweight64 and hweight64 (on older kernels
> and systems with !X86_FEATURE_POPCNT) or a popcnt based alternative.
> 
> i.e, We needlessly go through this mask based calculation everytime
> num_possible_cpus() is called.
> 
> The problem is there with cpu_online_mask() as well, which is fixed value at
> boot time in !CONFIG_HOTPLUG_CPU case and should not change that often even
> in HOTPLUG case.
> 
> Though most of the callers of these two routines are init time (with few
> exceptions of runtime calls), it is cleaner to use variables
> and not go through this repeated mask based calculation.
> 
> Signed-off-by: Venkatesh Pallipadi <venki@google.com>
> ---
>  arch/x86/kernel/smpboot.c |    2 +-
>  include/linux/cpumask.h   |   10 ++++++++--
>  kernel/cpu.c              |    5 +++++
>  kernel/smp.c              |    4 ++++
>  4 files changed, 18 insertions(+), 3 deletions(-)
> 
> diff --git a/arch/x86/kernel/smpboot.c b/arch/x86/kernel/smpboot.c
> index 66d250c..f87fcde 100644
> --- a/arch/x86/kernel/smpboot.c
> +++ b/arch/x86/kernel/smpboot.c
> @@ -947,7 +947,7 @@ static int __init smp_sanity_check(unsigned max_cpus)
>  			nr++;
>  		}
> 
> -		nr_cpu_ids = 8;
> +		setup_nr_cpu_ids();
>  	}
>  #endif
> 
> diff --git a/include/linux/cpumask.h b/include/linux/cpumask.h
> index 4f7a632..ac3113b 100644
> --- a/include/linux/cpumask.h
> +++ b/include/linux/cpumask.h
> @@ -23,10 +23,14 @@ typedef struct cpumask { DECLARE_BITMAP(bits, NR_CPUS); } cpumask_t;
> 
>  #if NR_CPUS == 1
>  #define nr_cpu_ids		1
> +#define nr_possible_cpus	1
>  #else
>  extern int nr_cpu_ids;
> +extern int nr_possible_cpus;
>  #endif
> 
> +extern int nr_online_cpus;
> +
>  #ifdef CONFIG_CPUMASK_OFFSTACK
>  /* Assuming NR_CPUS is huge, a runtime limit is more efficient.  Also,
>   * not all bits may be allocated. */
> @@ -81,8 +85,10 @@ extern const struct cpumask *const cpu_present_mask;
>  extern const struct cpumask *const cpu_active_mask;
> 
>  #if NR_CPUS > 1
> -#define num_online_cpus()	cpumask_weight(cpu_online_mask)
> -#define num_possible_cpus()	cpumask_weight(cpu_possible_mask)
> +
> +#define num_online_cpus()	(nr_online_cpus)
> +#define num_possible_cpus()	(nr_possible_cpus)
> +
>  #define num_present_cpus()	cpumask_weight(cpu_present_mask)
>  #define num_active_cpus()	cpumask_weight(cpu_active_mask)
>  #define cpu_online(cpu)		cpumask_test_cpu((cpu), cpu_online_mask)
> diff --git a/kernel/cpu.c b/kernel/cpu.c
> index 2060c6e..f179baa 100644
> --- a/kernel/cpu.c
> +++ b/kernel/cpu.c
> @@ -622,6 +622,9 @@ static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
>  const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
>  EXPORT_SYMBOL(cpu_active_mask);
> 
> +int nr_online_cpus __read_mostly;
> +EXPORT_SYMBOL(nr_online_cpus);
> +
>  void set_cpu_possible(unsigned int cpu, bool possible)
>  {
>  	if (possible)
> @@ -644,6 +647,8 @@ void set_cpu_online(unsigned int cpu, bool online)
>  		cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
>  	else
>  		cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
> +
> +	nr_online_cpus = cpumask_weight(cpu_online_mask);
>  }
> 
>  void set_cpu_active(unsigned int cpu, bool active)
> diff --git a/kernel/smp.c b/kernel/smp.c
> index db197d6..106e519 100644
> --- a/kernel/smp.c
> +++ b/kernel/smp.c
> @@ -658,10 +658,14 @@ early_param("maxcpus", maxcpus);
>  int nr_cpu_ids __read_mostly = NR_CPUS;
>  EXPORT_SYMBOL(nr_cpu_ids);
> 
> +int nr_possible_cpus __read_mostly = NR_CPUS;
> +EXPORT_SYMBOL(nr_possible_cpus);
> +
>  /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
>  void __init setup_nr_cpu_ids(void)
>  {
>  	nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
> +	nr_possible_cpus = cpumask_weight(cpu_possible_mask);
>  }
> 
>  /* Called by boot processor to activate the rest. */



This patch still has problems, IMHO.

A quick grep on the source for set_cpu_possible() revealed the following
problematic areas:
1. arch/alpha/kernel/process.c: common_shutdown_1()
   set_cpu_possible() with second parameter as 'false' is used here.
   Though this is the shutdown code, a disparity between what
   num_possible_cpus() reports and what is actually there in
   cpu_possible_mask is not such a good idea, at any point in time.

2. arch/cris/arch-v32/kernel/smp.c: smp_prepare_boot_cpu()
   smp_prepare_boot_cpu() is called after setup_nr_cpu_ids().

   (Btw, I think things are really messed up in cris as it is, if I am not
   totally mistaken).

3. arch/mn10300/kernel/smp.c: smp_prepare_cpus()
   This function calls set_cpu_possible(). smp_prepare_cpus() is called in
   kernel_init(), which is much later than setup_nr_cpu_ids().

4. arch/um/kernel/smp.c: smp_prepare_cpus()
   Same as in 3.

5. As far as arch/x86/xen is concerned, I can see code such as the following,
   among other things:

    xen_smp_prepare_cpus():

    /* Restrict the possible_map according to max_cpus. */
    while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
            for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
                    continue;
            set_cpu_possible(cpu, false);
    }

If I am not missing anything obvious, applying this patch can effectively
convert the above code into an infinite loop, among other damages!

I still feel it would be safer to edit set_cpu_possible() such that
nr_possible_cpus is updated whenever cpu_possible_mask is altered.

Regards,
Srivatsa S. Bhat
IBM Linux Technology Center