[PATCH 2/2][v3] powerpc: Make the CMM memory hotplug aware

[Robert Jennings]

* Benjamin Herrenschmidt (benh at kernel.crashing.org) wrote:
> On Fri, 2009-10-09 at 15:41 -0500, Robert Jennings wrote:
> > The Collaborative Memory Manager (CMM) module allocates individual pages
> > over time that are not migratable.  On a long running system this can
> > severely impact the ability to find enough pages to support a hotplug
> > memory remove operation.
> > 
> > This patch adds a memory isolation notifier and a memory hotplug notifier.
> > The memory isolation notifier will return the number of pages found
> > in the range specified.  This is used to determine if all of the used
> > pages in a pageblock are owned by the balloon (or other entities in
> > the notifier chain).  The hotplug notifier will free pages in the range
> > which is to be removed.  The priority of this hotplug notifier is low
> > so that it will be called near last, this helps avoids removing loaned
> > pages in operations that fail due to other handlers.
> > 
> > CMM activity will be halted when hotplug remove operations are active
> > and resume activity after a delay period to allow the hypervisor time
> > to adjust.
> > 
> > Signed-off-by: Robert Jennings <rcj at linux.vnet.ibm.com>
> 
> Do you need me to merge that via the powerpc tree after the relevant
> generic parts go in ? This is 2.6.33 material ?

I didn't know how this part works honestly, this is the first time I've
pushed patches with dependencies like this.  Andrew Morton had pulled
an earlier version of this and the mm hotplug related changes for 2.6.32.

> > +module_param_named(hotplug_delay, hotplug_delay, uint, S_IRUGO | S_IWUSR);
> > +MODULE_PARM_DESC(delay, "Delay (in seconds) after memory hotplug remove "
> > +		 "before activity resumes. "
> > +		 "[Default=" __stringify(CMM_HOTPLUG_DELAY) "]");
> 
> What is the above ? That sounds scary :-)

I'm changing this to read "Delay (in seconds) after memory hotplug
remove before loaning resumes." in order to clear this us.  This is a
period where loaning from the balloon is paused after the hotplug
completes.  This gives the userspace tools time to mark the sections
as isolated and unusable with the hypervisor and the hypervisor to take
this into account regarding the loaning levels it requests of the OS.

> >  module_param_named(oom_kb, oom_kb, uint, S_IRUGO | S_IWUSR);
> >  MODULE_PARM_DESC(oom_kb, "Amount of memory in kb to free on OOM. "
> >  		 "[Default=" __stringify(CMM_OOM_KB) "]");
> > @@ -88,6 +101,8 @@ struct cmm_page_array {
> >  static unsigned long loaned_pages;
> >  static unsigned long loaned_pages_target;
> >  static unsigned long oom_freed_pages;
> > +static atomic_t hotplug_active = ATOMIC_INIT(0);
> > +static atomic_t hotplug_occurred = ATOMIC_INIT(0);
> 
> That sounds like a hand made lock with atomics... rarely a good idea,
> tends to miss appropriate barriers etc...
> 

I have changes this so that we have a mutex held during the memory
hotplug remove.  The hotplug_occurred flag is now and integer protected
by the mutex; it is used to provide the delay after the hotplug remove
completes.

> >  static struct cmm_page_array *cmm_page_list;
> >  static DEFINE_SPINLOCK(cmm_lock);
> > @@ -110,6 +125,9 @@ static long cmm_alloc_pages(long nr)
> >  	cmm_dbg("Begin request for %ld pages\n", nr);
> >  
> >  	while (nr) {
> > +		if (atomic_read(&hotplug_active))
> > +			break;
> > +
> 
> Ok so I'm not familiar with that whole memory hotplug stuff, so the code
> might be right, but wouldn't the above be racy anyways in case hotplug
> just becomes active after this statement ?
> 
> Shouldn't you use a mutex_trylock instead ? That has clearer semantics
> and will provide the appropriate memory barriers.

I have changed this to use a mutex in the same location.  This allows
the allocation of pages to terminate early during a hotplug remove
operation.

If hotplug becomes active after this check in cmm_alloc_pages() one page
will be allocated to the balloon, but this page will not belong to the
memory range going offline because the pageblock will have already been
isolated.  After allocating the page we might need to wait on the lock to
add the page to the list if hotplug is removing pages from the balloon.
After one page is added, cmm_alloc_pages() will exit early when it checks
to see if hotplug is active or if it has occurred.

I wanted to keep the section locked by cmm_lock small, so that we can
safely respond to the hotplug notifier as quickly as possible while
minimizing memory pressure.

There are no checks in cmm_free_pages() to have it abort early during
hotplug memory remove with the rationale that it's good for hotplug
to allow the balloon to shrink even if it means holding the cmm_lock a
bit longer.

> >  		addr = __get_free_page(GFP_NOIO | __GFP_NOWARN |
> >  				       __GFP_NORETRY | __GFP_NOMEMALLOC);
> >  		if (!addr)
> > @@ -119,8 +137,10 @@ static long cmm_alloc_pages(long nr)
> >  		if (!pa || pa->index >= CMM_NR_PAGES) {
> >  			/* Need a new page for the page list. */
> >  			spin_unlock(&cmm_lock);
> > -			npa = (struct cmm_page_array *)__get_free_page(GFP_NOIO | __GFP_NOWARN |
> > -								       __GFP_NORETRY | __GFP_NOMEMALLOC);
> > +			npa = (struct cmm_page_array *)__get_free_page(
> > +					GFP_NOIO | __GFP_NOWARN |
> > +					__GFP_NORETRY | __GFP_NOMEMALLOC |
> > +					__GFP_MOVABLE);
> >  			if (!npa) {
> >  				pr_info("%s: Can not allocate new page list\n", __func__);
> >  				free_page(addr);
> > @@ -273,9 +293,23 @@ static int cmm_thread(void *dummy)
> >  	while (1) {
> >  		timeleft = msleep_interruptible(delay * 1000);
> > 
> > -		if (kthread_should_stop() || timeleft) {
> > -			loaned_pages_target = loaned_pages;
> > +		if (kthread_should_stop() || timeleft)
> >  			break;
> > +
> > +		if (atomic_read(&hotplug_active)) {
> > +			cmm_dbg("Hotplug operation in progress, activity "
> > +					"suspended\n");
> > +			continue;
> > +		}
> > +
> > +		if (atomic_dec_if_positive(&hotplug_occurred) >= 0) {
> > +			cmm_dbg("Hotplug operation has occurred, loaning "
> > +					"activity suspended for %d seconds.\n",
> > +					hotplug_delay);
> > +			timeleft = msleep_interruptible(hotplug_delay * 1000);
> > +			if (kthread_should_stop() || timeleft)
> > +				break;
> > +			continue;
> >  		}
> 
> I have less problems with hotplug_occured but if you use a
> mutex_trylock, overall, you can turn the above into a normal int instead
> of an atomic.

Changed to a mutex to indicate that a hotplug operation is active and an
int protected by the mutex to indicate that a hotplug operation has
occurred.

>  ../..
> 
> > +static int cmm_memory_cb(struct notifier_block *self,
> > +			unsigned long action, void *arg)
> > +{
> > +	int ret = 0;
> > +
> > +	switch (action) {
> > +	case MEM_GOING_OFFLINE:
> > +		atomic_set(&hotplug_active, 1);
> 
> So that would become a mutex_lock(). Added advantage is that
> it would wait for a current CMM operation to complete.

I've added the mutex but the scope will not prevent hotplug from
starting before the current CMM operation has completed.  This allows us
to abort the allocation.  The important globals for managing the list of
pages are still covered by the cmm_lock spinlock.

I've tested the patch and I'll send it out shortly.