Hi Vegard,

Vegard Nossum <vegard.nossum@gmail.com> writes:

quoted text
> From: Vegard Nossum <vegard.nossum@gmail.com>
> Date: Fri, 23 May 2008 15:53:03 +0200
> Subject: [PATCH] kmemcheck: SMP support
>
> This patch adds SMP support to kmemcheck, that is, the ability to boot
> more than one processor even when kmemcheck is enabled. (Previously,
> only one CPU would be booted even if more were present in the system.)
>
> On page fault, kmemcheck needs to pause all the other CPUs in the system
> in order to guarantee that no other CPU will modify the same memory
> location (which will otherwise be unprotected after we set the present
> bit of the PTE).
>
> Since the page fault can be taken with any irq state (i.e. enabled or
> disabled), we can't send a normal IPI broadcast since this can deadlock.
>
> Instead, we send an NMI. This is guaranteed to be received _except_ if
> the processor is already inside the NMI handler.
>
> This is of course not very efficient, and booting with maxcpus=1 is
> recommended, however, this allows the kernel to be configured with
> CONFIG_KMEMCHECK=y with close to zero overhead when kmemcheck is
> disabled. (It can still be runtime-enabled at any time, though.)
>
> The patch has been tested on real hardware.
>
> Signed-off-by: Vegard Nossum <vegard.nossum@gmail.com>
> ---
>  arch/x86/kernel/kmemcheck.c |  108 +++++++++++++++++++++++++++++++++++++++----
>  1 files changed, 98 insertions(+), 10 deletions(-)
>
> diff --git a/arch/x86/kernel/kmemcheck.c b/arch/x86/kernel/kmemcheck.c
> index c0045e8..fdf8acb 100644
> --- a/arch/x86/kernel/kmemcheck.c
> +++ b/arch/x86/kernel/kmemcheck.c
> @@ -16,6 +16,7 @@
>  #include <linux/kmemcheck.h>
>  #include <linux/mm.h>
>  #include <linux/module.h>
> +#include <linux/notifier.h>
>  #include <linux/page-flags.h>
>  #include <linux/percpu.h>
>  #include <linux/stacktrace.h>
> @@ -23,9 +24,12 @@
>  #include <asm/cacheflush.h>
>  #include <asm/kmemcheck.h>
>  #include <asm/pgtable.h>
> +#include <asm/smp.h>
>  #include <asm/string.h>
>  #include <asm/tlbflush.h>
>  
> +#include <mach_ipi.h>
> +
>  enum shadow {
>  	SHADOW_UNALLOCATED,
>  	SHADOW_UNINITIALIZED,
> @@ -240,18 +244,91 @@ static void do_wakeup(unsigned long data)
>  	}
>  }
>  
> +static atomic_t nmi_wait;
> +static atomic_t nmi_resume;
> +static atomic_t started;
> +static atomic_t finished;
> +
> +static int nmi_notifier(struct notifier_block *self,
> +	unsigned long val, void *data)
> +{
> +	struct die_args *args = (struct die_args *) data;
> +
> +	if (val != DIE_NMI_IPI || !atomic_read(&nmi_wait))
> +		return NOTIFY_DONE;
> +
> +	atomic_inc(&started);
> +
> +	/* Pause until the fault has been handled */
> +	while (!atomic_read(&nmi_resume))
> +		cpu_relax();
> +
> +	atomic_inc(&finished);
> +
> +	return NOTIFY_STOP;
> +}
> +
> +static void
> +pause_allbutself(void)
> +{
> +#ifdef CONFIG_SMP
> +	static spinlock_t nmi_spinlock;
> +
> +	int cpus;
> +	cpumask_t mask = cpu_online_map;
> +
> +	spin_lock(&nmi_spinlock);
> +
> +	cpus = num_online_cpus() - 1;
> +
> +	atomic_set(&started, 0);
> +	atomic_set(&finished, 0);
> +	atomic_set(&nmi_wait, 1);
> +	atomic_set(&nmi_resume, 0);
> +
> +	cpu_clear(safe_smp_processor_id(), mask);
> +	if (!cpus_empty(mask))
> +		send_IPI_mask(mask, NMI_VECTOR);
> +
> +	while (atomic_read(&started) != cpus)
> +		cpu_relax();
> +
> +	atomic_set(&nmi_wait, 0);
> +
> +	spin_unlock(&nmi_spinlock);
> +#endif
> +}
> +
> +static void
> +resume(void)
> +{
> +#ifdef CONFIG_SMP
> +	int cpus;
> +
> +	cpus = num_online_cpus() - 1;
> +
> +	atomic_set(&nmi_resume, 1);
> +
> +	while (atomic_read(&finished) != cpus)
> +		cpu_relax();
> +#endif
> +}

How about merging finished and started into one?  I.e. `paused'.

The notifiers increases `paused' before the waiting-loop and decreases
it again afterwards.

pause_allbutself() sends the IPIs and waits until `paused' reached the
number of CPUS.

resume() justs waits until `paused' reaches zero.

Would this work?  Will the NMI handler finish even when the CPU is
removed while the handler runs?

	Hannes
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/