[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Xen-devel] Re: [PATCH R4 6/7] mm: Extend memory hotplug API to allow memory hotplug in virtual guests



On Tue, 2011-03-08 at 22:50 +0100, Daniel Kiper wrote:
> This patch extends memory hotplug API to allow easy memory hotplug
> in virtual guests. It contains:
>   - generic section aligment macro,
>   - online_page_chain and apropriate functions for registering/unregistering
>     online page notifiers,
>   - add_virtual_memory(u64 *size) function which adds memory region
>     of size >= *size above max_pfn; new region is section aligned
>     and size is modified to be multiple of section size.

Usually, when you can list stuff out like this, it's a good sign that
they belong in separate patches.  I think it's true here as well.

But, these are looking a lot better.  It looks like much less code, and
it's quite a bit simpler.

> +/*
> + * online_page_chain contains chain of notifiers called when page is onlined.
> + * When kernel is booting native_online_page_notifier() is registered with
> + * priority 0 as default notifier. Custom notifier should be registered with
> + * pririty > 0. It could be terminal (it should return NOTIFY_STOP on 
> success)

"pririty"?

> + * or not (it should return NOTIFY_DONE or NOTIFY_OK on success; for full 
> list
> + * of return codes look into include/linux/notifier.h).
> + *
> + * Working example of usage: drivers/xen/balloon.c
> + */
> +
> +static RAW_NOTIFIER_HEAD(online_page_chain);
> +
>  DEFINE_MUTEX(mem_hotplug_mutex);
> 
>  void lock_memory_hotplug(void)
> @@ -361,8 +375,33 @@ int __remove_pages(struct zone *zone, unsigned long 
> phys_start_pfn,
>  }
>  EXPORT_SYMBOL_GPL(__remove_pages);
> 
> -void online_page(struct page *page)
> +int register_online_page_notifier(struct notifier_block *nb)
> +{
> +     int rc;
> +
> +     lock_memory_hotplug();
> +     rc = raw_notifier_chain_register(&online_page_chain, nb);
> +     unlock_memory_hotplug();
> +
> +     return rc;
> +}
> +EXPORT_SYMBOL_GPL(register_online_page_notifier);
> +
> +int unregister_online_page_notifier(struct notifier_block *nb)
> +{
> +     int rc;
> +
> +     lock_memory_hotplug();
> +     rc = raw_notifier_chain_unregister(&online_page_chain, nb);
> +     unlock_memory_hotplug();
> +
> +     return rc;
> +}
> +EXPORT_SYMBOL_GPL(unregister_online_page_notifier);

The whole "native" thing really is Xen terminology.  Could we call this
"generic_online_page_notifier()" perhaps?  This really isn't even
"native" either since some hypervisors actually do use this code.

> +static int native_online_page_notifier(struct notifier_block *nb, unsigned 
> long val, void *v)
>  {
> +     struct page *page = v;
>       unsigned long pfn = page_to_pfn(page);
> 
>       totalram_pages++;
> @@ -375,12 +414,30 @@ void online_page(struct page *page)
>  #endif
> 
>  #ifdef CONFIG_FLATMEM
> -     max_mapnr = max(page_to_pfn(page), max_mapnr);
> +     max_mapnr = max(pfn, max_mapnr);
>  #endif

This is another tidbit that's probably good to do, but it's superfluous
here.  

>       ClearPageReserved(page);
>       init_page_count(page);
>       __free_page(page);
> +
> +     return NOTIFY_OK;
> +}
> +
> +static struct notifier_block native_online_page_nb = {
> +     .notifier_call = native_online_page_notifier,
> +     .priority = 0
> +};

That comment about priority really belongs here.  

 /*
  * 0 priority makes this the fallthrough default.  All
  * architectures wanting to override this should set a
  * higher priority and return NOTIFY_STOP to keep this
  * from running.
  */

> +static int __init init_online_page_chain(void)
> +{
> +     return register_online_page_notifier(&native_online_page_nb);
> +}
> +pure_initcall(init_online_page_chain);
> +
> +static void online_page(struct page *page)
> +{
> +     raw_notifier_call_chain(&online_page_chain, 0, page);
>  }
> 
>  static int online_pages_range(unsigned long start_pfn, unsigned long 
> nr_pages,
> @@ -591,6 +648,36 @@ out:
>  }
>  EXPORT_SYMBOL_GPL(add_memory);
> 
> +/*
> + * add_virtual_memory() adds memory region of size >= *size above max_pfn.
> + * New region is section aligned and size is modified to be multiple of
> + * section size.

Aligned up or down?  Why did you choose up or down?

> Those features allow optimal use of address space and
> + * establish proper aligment when this function is called first time after

                      ^^^^^^^^ alignment?

> + * boot (last section not fully populated at boot time may contains unused
> + * memory pages with PG_reserved bit not set; online_pages() does not allow
> + * page onlining in whole section if first page does not have PG_reserved
> + * bit set). Real size of added memory should be established at page onlining
> + * stage.
> + *
> + * This function is often used in virtual guests because mainly they do not
> + * care about new memory region address.

Remember, you're touching generic memory hotplug code here.  I really
don't know what a "virtual guest" is or how it relates to this code.
How about something like this?

        This code is expected to be used in cases where a certain amount
        of memory needs to get added, but when the hardware or
        hypervisor does not dictate where it will be placed.

> + * Working example of usage: drivers/xen/balloon.c

Please pull this out.  It'll probably become stale before anyone uses
it.  I trust people to know how to use cscope. :)

> +int add_virtual_memory(u64 *size)
> +{
> +     int nid;
> +     u64 start;
> +
> +     start = PFN_PHYS(SECTION_ALIGN(max_pfn));
> +     *size = (((*size >> PAGE_SHIFT) & PAGE_SECTION_MASK) + 
> PAGES_PER_SECTION) << PAGE_SHIFT;

Why use PFN_PHYS() in one case but not the other?

I'd also highly suggest using the ALIGN() macro in cases like this.  It
makes it much more readable:

        *size = PFN_PHYS(ALIGN(*size, SECTION_SIZE)));  

> +     nid = memory_add_physaddr_to_nid(start);
> +
> +     return add_memory(nid, start, *size);
> +}

Could you talk a little bit more about how 'size' gets used?  Also, are
we sure we want an interface where we're so liberal with 'size'?  It
seems like requiring that it be section-aligned is a fair burden to
place on the caller.  That way, we're not in a position of _guessing_
what the caller wants (aligning up or down).

-- Dave


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.