Re: [PATCH v8 3/7] crash: add generic infrastructure for crash hotplug support

[Eric DeVolder]

David,
Great questions! See inline responses below.
eric

On 5/12/22 03:52, David Hildenbrand wrote:
> On 05.05.22 20:45, Eric DeVolder wrote:
> > CPU and memory change notifications are received in order to
> > regenerate the elfcorehdr.
> >
> > To support cpu hotplug, a callback is registered to capture the
> > CPUHP_AP_ONLINE_DYN online and offline events via
> > cpuhp_setup_state_nocalls().
> >
> > To support memory hotplug, a notifier is registered to capture the
> > MEM_ONLINE and MEM_OFFLINE events via register_memory_notifier().
> >
> > The cpu callback and memory notifiers call handle_hotplug_event()
> > to handle the hot plug/unplug event. Then handle_hotplug_event()
> > dispatches the event to the architecture specific
> > arch_crash_handle_hotplug_event(). During the process, the
> > kexec_mutex is held.
> >
> > Signed-off-by: Eric DeVolder <eric.devolder@xxxxxxxxxx>
>
> [...]
>
> > +
> > +#if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_MEMORY_HOTPLUG)
> > +void __weak arch_crash_handle_hotplug_event(struct kimage *image,
> > +							unsigned int hp_action, unsigned int cpu)
> > +{
> > +	WARN(1, "crash hotplug handler not implemented");
>
>
> Won't that trigger on any arch that has CONFIG_HOTPLUG_CPU and CONFIG_MEMORY_HOTPLUG?
> I mean, you only implement it for x86 later in this series. Or what else stops this WARN from
> triggering?
You're correct. What about: printk_once(KERN_DEBUG "...") ?

> > +}
> > +
> > +static void handle_hotplug_event(unsigned int hp_action, unsigned int cpu)
> > +{
> > +	/* Obtain lock while changing crash information */
> > +	if (!mutex_trylock(&kexec_mutex))
> > +		return;
>
> This looks wrong. What if you offline memory but for some reason the mutex
> is currently locked? You'd miss updating the vmcore, which would be broken.
>
> Why is this trylock in place? Some workaround for potential locking issues,
> or what is the rationale?

I took this from kernel/kexec.c:do_my_load(), but you are right, this is not
right.

     *
     * KISS: always take the mutex.
     */
    if (!mutex_trylock(&kexec_mutex))
        return -EBUSY;

This should simply be mutex_lock(&kexec_mutex).

> > +
> > +	/* Check kdump is loaded */
> > +	if (kexec_crash_image) {
> > +		pr_debug("crash hp: hp_action %u, cpu %u", hp_action, cpu);
> > +
> > +		/* Needed in order for the segments to be updated */
> > +		arch_kexec_unprotect_crashkres();
> > +
> > +		/* Flag to differentiate between normal load and hotplug */
> > +		kexec_crash_image->hotplug_event = true;
>
> 1. Why is that required? Why can't arch_crash_handle_hotplug_event() forward that
> information? I mean, *hotplug* in the anme implies that the function should be
> aware what's happening.
Member .hotplug_event is needed in crash_prepare_elf64_headers(). To date, it has made
sense (to me) to pass this parameter to crash_prepare_elf64_headers() via the
struct kimage, rather than directly. The patch "crash: prototype change for 
crash_prepare_elf64_headers" changes crash_prepare_elf64_headers() to accept the
struct kimage. If it is so desired, it could just as well be changed to pass just
this one parameter; though struct kimage seems a more useful way to go.

> 2. Why can't the unprotect+reprotect not be done inside
> arch_crash_handle_hotplug_event() ? It's all arch specific either way.
>
> IMHO, this code here should be as simple as
>
> if (kexec_crash_image)
> 	arch_crash_handle_hotplug_event(kexec_crash_image, hp_action, cpu);

The intent of this code was to be generic infrastructure. Just invoking the
arch_crash_handle_hotplug_event() would certainly be as generic as it gets.
But there were a series of steps that seemed to be common, so those I hoisted
into this bit of code.

> 3. Why do we have to forward the CPU for CPU onlining/offlining but not the
> memory block id (or similar) when onlining/offlining a memory block?
From patch "kexec: exclude hot remove cpu from elfcorehdr notes" commit message:

Due to use of CPUHP_AP_ONLINE_DYN, upon CPU unplug, the CPU is
still in the for_each_present_cpu() list when within the
handle_hotplug_event(). Thus the CPU must be explicitly excluded
when building the new list of CPUs.

This change identifies in handle_hotplug_event() the CPU to be
excluded, and the check for excluding the CPU in
crash_prepare_elf64_headers().

If there is a better CPUHP_ to use than _DYN, I'd be all for that!

> > +
> > +		/* Now invoke arch-specific update handler */
> > +		arch_crash_handle_hotplug_event(kexec_crash_image, hp_action, cpu);
> > +
> > +		/* No longer handling a hotplug event */
> > +		kexec_crash_image->hotplug_event = false;
> > +
> > +		/* Change back to read-only */
> > +		arch_kexec_protect_crashkres();
> > +	}
> > +
> > +	/* Release lock now that update complete */
> > +	mutex_unlock(&kexec_mutex);
> > +}
> > +
> > +static int crash_memhp_notifier(struct notifier_block *nb, unsigned long val, void *v)
> > +{
> > +	switch (val) {
> > +	case MEM_ONLINE:
> > +		handle_hotplug_event(KEXEC_CRASH_HP_ADD_MEMORY, 0);
> > +		break;
> > +
> > +	case MEM_OFFLINE:
> > +		handle_hotplug_event(KEXEC_CRASH_HP_REMOVE_MEMORY, 0);
> > +		break;
> > +	}
> > +	return NOTIFY_OK;
> > +}
> > +
> > +static struct notifier_block crash_memhp_nb = {
> > +	.notifier_call = crash_memhp_notifier,
> > +	.priority = 0
> > +};
> > +
> > +static int crash_cpuhp_online(unsigned int cpu)
> > +{
> > +	handle_hotplug_event(KEXEC_CRASH_HP_ADD_CPU, cpu);
> > +	return 0;
> > +}
> > +
> > +static int crash_cpuhp_offline(unsigned int cpu)
> > +{
> > +	handle_hotplug_event(KEXEC_CRASH_HP_REMOVE_CPU, cpu);
> > +	return 0;
> > +}
> > +
> > +static int __init crash_hotplug_init(void)
> > +{
> > +	int result = 0;
> > +
> > +	if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG))
> > +		register_memory_notifier(&crash_memhp_nb);
> > +
> > +	if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
> > +		result = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
> > +									"crash/cpuhp",
> > +									crash_cpuhp_online,
> > +									crash_cpuhp_offline);
>
> Ehm, this indentation looks very weird.
>
> > +
> > +	return result;
> > +}
> > +
> > +subsys_initcall(crash_hotplug_init);
> > +#endif