Date:	Fri, 26 Oct 2012 09:39:22 +0800
From:	Lai Jiangshan <laijs@...fujitsu.com>
To:	KOSAKI Motohiro <kosaki.motohiro@...il.com>
CC:	linux-kernel@...r.kernel.org, Rob Landley <rob@...dley.net>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Jiang Liu <jiang.liu@...wei.com>,
	Jianguo Wu <wujianguo@...wei.com>,
	Kay Sievers <kay.sievers@...y.org>,
	Greg Kroah-Hartman <gregkh@...e.de>,
	Xishi Qiu <qiuxishi@...wei.com>, Mel Gorman <mgorman@...e.de>,
	linux-doc@...r.kernel.org, linux-mm@...ck.org
Subject: Re: [PATCH 1/3] memory_hotplug: fix stale node_states[N_NORMAL_MEMORY]

Hi, KOSAKI 

On 09/28/2012 06:03 AM, KOSAKI Motohiro wrote:
> (9/27/12 2:47 AM), Lai Jiangshan wrote:
>> Currently memory_hotplug only manages the node_states[N_HIGH_MEMORY],
>> it forgets to manage node_states[N_NORMAL_MEMORY]. it causes
>> node_states[N_NORMAL_MEMORY] becomes stale.
> 
> What's mean 'stale'? I guess
> 
> : Currently memory_hotplug doesn't turn on/off node_states[N_NORMAL_MEMORY]


Right.

> and
> : then it will be invalid if the platform has highmem. Luckily, almost memory 
> : hotplug aware platform don't have highmem, but are not all.
> 
> right?

Some platforms(32 bit) support logic-memory-hotplug.
Some platforms have movable memory.
They are all considered.

> I supporse this patch only meaningful on ARM platform practically.
> 

any platform whic supports memory-hotplug.

> 
> 
>> We add check_nodemasks_changes_online() and check_nodemasks_changes_offline()
>> to detect whether node_states[N_HIGH_MEMORY] and node_states[N_NORMAL_MEMORY]
>> are changed while hotpluging.
> 
> 
>> Also add @status_change_nid_normal to struct memory_notify, thus
>> the memory hotplug callbacks know whether the node_states[N_NORMAL_MEMORY]
>> are changed.
> 
> status_change_nid_normal is very ugly to me. When status_change_nid and 
> status_change_nid_normal has positive value, they are always the same.
> nid and flags value are more natual to me.

If we use flags, the semantic of "status_change_nid" is changed and we need to
change more current code, and we will add complicated to the memory hotplug
callbacks.

like this:

-		node = arg->status_change_nid;
+		if (arg->status_change_flags & (1UL << N_HIGH_MEMORY))
+			node = arg->status_change_nid;
+		else
+			node = -1;

> 
> 
> 
>>
>> Signed-off-by: Lai Jiangshan <laijs@...fujitsu.com>
>> ---
>>  Documentation/memory-hotplug.txt |    5 ++-
>>  include/linux/memory.h           |    1 +
>>  mm/memory_hotplug.c              |   94 +++++++++++++++++++++++++++++++------
>>  3 files changed, 83 insertions(+), 17 deletions(-)
>>
>> diff --git a/Documentation/memory-hotplug.txt b/Documentation/memory-hotplug.txt
>> index 6d0c251..6e6cbc7 100644
>> --- a/Documentation/memory-hotplug.txt
>> +++ b/Documentation/memory-hotplug.txt
>> @@ -377,15 +377,18 @@ The third argument is passed by pointer of struct memory_notify.
>>  struct memory_notify {
>>         unsigned long start_pfn;
>>         unsigned long nr_pages;
>> +       int status_change_nid_normal;
>>         int status_change_nid;
>>  }
>>  
>>  start_pfn is start_pfn of online/offline memory.
>>  nr_pages is # of pages of online/offline memory.
>> +status_change_nid_normal is set node id when N_NORMAL_MEMORY of nodemask
>> +is (will be) set/clear, if this is -1, then nodemask status is not changed.
>>  status_change_nid is set node id when N_HIGH_MEMORY of nodemask is (will be)
>>  set/clear. It means a new(memoryless) node gets new memory by online and a
>>  node loses all memory. If this is -1, then nodemask status is not changed.
>> -If status_changed_nid >= 0, callback should create/discard structures for the
>> +If status_changed_nid* >= 0, callback should create/discard structures for the
>>  node if necessary.
>>  
>>  --------------
>> diff --git a/include/linux/memory.h b/include/linux/memory.h
>> index ff9a9f8..a09216d 100644
>> --- a/include/linux/memory.h
>> +++ b/include/linux/memory.h
>> @@ -53,6 +53,7 @@ int arch_get_memory_phys_device(unsigned long start_pfn);
>>  struct memory_notify {
>>  	unsigned long start_pfn;
>>  	unsigned long nr_pages;
>> +	int status_change_nid_normal;
>>  	int status_change_nid;
>>  };
>>  
>> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
>> index 6a5b90d..b62d429b 100644
>> --- a/mm/memory_hotplug.c
>> +++ b/mm/memory_hotplug.c
>> @@ -460,6 +460,34 @@ static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
>>  	return 0;
>>  }
>>  
>> +static void check_nodemasks_changes_online(unsigned long nr_pages,
>> +	struct zone *zone, struct memory_notify *arg)
>> +{
>> +	int nid = zone_to_nid(zone);
>> +	enum zone_type zone_last = ZONE_NORMAL;
>> +
>> +	if (N_HIGH_MEMORY == N_NORMAL_MEMORY)
>> +		zone_last = ZONE_MOVABLE;
> 
> This is very strange (or ugly) code. ZONE_MOVABLE don't depend on high mem.

If we don't have HIGHMEM,
any node of N_NORMAL_MEMORY has 0...ZONE_MOVABLE

if we have HIGHMEM,
any node of N_NORMAL_MEMORY has 0...ZONE_NORMAL

> 
> 
>> +
>> +	if (zone_idx(zone) <= zone_last && !node_state(nid, N_NORMAL_MEMORY))
>> +		arg->status_change_nid_normal = nid;
>> +	else
>> +		arg->status_change_nid_normal = -1;
> 
> Wrong. The onlined node may only have high mem zone. IOW, think fake numa case etc.

"zone_idx(zone) <= zone_last" checks this case. the result is "else" branch.

> 
> 
>> +
>> +	if (!node_state(nid, N_HIGH_MEMORY))
>> +		arg->status_change_nid = nid;
>> +	else
>> +		arg->status_change_nid = -1;
>> +}
>> +
>> +static void set_nodemasks(int node, struct memory_notify *arg)
> 
> Too ugly. just remove this and use node_set_state() directly.
> 
>> +{
>> +	if (arg->status_change_nid_normal >= 0)
>> +		node_set_state(node, N_NORMAL_MEMORY);
>> +
>> +	node_set_state(node, N_HIGH_MEMORY);
>> +}
>> +
>>  
>>  int __ref online_pages(unsigned long pfn, unsigned long nr_pages)
>>  {
>> @@ -471,13 +499,18 @@ int __ref online_pages(unsigned long pfn, unsigned long nr_pages)
>>  	struct memory_notify arg;
>>  
>>  	lock_memory_hotplug();
>> +	/*
>> +	 * This doesn't need a lock to do pfn_to_page().
>> +	 * The section can't be removed here because of the
>> +	 * memory_block->state_mutex.
>> +	 */
> 
> Please explain the intention of this comment. We think lock_memory_hotplug() close
> a race against memory offline directly.

I move old code up. the context does not changed.

> 
> 
>> +	zone = page_zone(pfn_to_page(pfn));
>> +
>>  	arg.start_pfn = pfn;
>>  	arg.nr_pages = nr_pages;
>> -	arg.status_change_nid = -1;
>> +	check_nodemasks_changes_online(nr_pages, zone, &arg);
>>  
>>  	nid = page_to_nid(pfn_to_page(pfn));
>> -	if (node_present_pages(nid) == 0)
>> -		arg.status_change_nid = nid;
>>  
>>  	ret = memory_notify(MEM_GOING_ONLINE, &arg);
>>  	ret = notifier_to_errno(ret);
>> @@ -487,12 +520,6 @@ int __ref online_pages(unsigned long pfn, unsigned long nr_pages)
>>  		return ret;
>>  	}
>>  	/*
>> -	 * This doesn't need a lock to do pfn_to_page().
>> -	 * The section can't be removed here because of the
>> -	 * memory_block->state_mutex.
>> -	 */
>> -	zone = page_zone(pfn_to_page(pfn));
>> -	/*
>>  	 * If this zone is not populated, then it is not in zonelist.
>>  	 * This means the page allocator ignores this zone.
>>  	 * So, zonelist must be updated after online.
>> @@ -517,7 +544,7 @@ int __ref online_pages(unsigned long pfn, unsigned long nr_pages)
>>  	zone->present_pages += onlined_pages;
>>  	zone->zone_pgdat->node_present_pages += onlined_pages;
>>  	if (onlined_pages) {
>> -		node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
>> +		set_nodemasks(zone_to_nid(zone), &arg);
>>  		if (need_zonelists_rebuild)
>>  			build_all_zonelists(NULL, zone);
>>  		else
>> @@ -870,6 +897,44 @@ check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
>>  	return offlined;
>>  }
>>  
>> +static void check_nodemasks_changes_offline(unsigned long nr_pages,
>> +		struct zone *zone, struct memory_notify *arg)
>> +{
> 
> This should remove ugly memory_notify argument and should be fold 
> check_nodemasks_changes_online().
> 
> 
>> +	struct pglist_data *pgdat = zone->zone_pgdat;
>> +	unsigned long present_pages = 0;
>> +	enum zone_type zt, zone_last = ZONE_NORMAL;
>> +
>> +	if (N_HIGH_MEMORY == N_NORMAL_MEMORY)
>> +		zone_last = ZONE_MOVABLE;
>> +
>> +	for (zt = 0; zt <= zone_last; zt++)
>> +		present_pages += pgdat->node_zones[zt].present_pages;
>> +	if (zone_idx(zone) <= zone_last && nr_pages >= present_pages)
>> +		arg->status_change_nid_normal = zone_to_nid(zone);
>> +	else
>> +		arg->status_change_nid_normal = -1;
> 
> Wrong. Think, a zone has both normal and highmem zone and when admin 
> only remove highmem area. you should check pfn too.

I checked the pfn via "zone_idx(zone) <= zone_last".

"""zone = page_zone(pfn_to_page(pfn))"""

> 
> 
>> +
>> +	zone_last = ZONE_MOVABLE;
>> +	for (; zt <= zone_last; zt++)
>> +		present_pages += pgdat->node_zones[zt].present_pages;
>> +	if (nr_pages >= present_pages)
>> +		arg->status_change_nid = zone_to_nid(zone);
>> +	else
>> +		arg->status_change_nid = -1;
>> +}
>> +
>> +static void clear_nodemasks(int node, struct memory_notify *arg)
>> +{
>> +	if (arg->status_change_nid_normal >= 0)
>> +		node_clear_state(node, N_NORMAL_MEMORY);
>> +
>> +	if (N_HIGH_MEMORY == N_NORMAL_MEMORY)
>> +		return;
> 
> You should use #ifdef. otherwise we face unreachable statement warning.

fixed.

> 
>> +
>> +	if (arg->status_change_nid >= 0)
>> +		node_clear_state(node, N_HIGH_MEMORY);
>> +}
>> +
>>  static int __ref offline_pages(unsigned long start_pfn,
>>  		  unsigned long end_pfn, unsigned long timeout)
>>  {
>> @@ -903,9 +968,7 @@ static int __ref offline_pages(unsigned long start_pfn,
>>  
>>  	arg.start_pfn = start_pfn;
>>  	arg.nr_pages = nr_pages;
>> -	arg.status_change_nid = -1;
>> -	if (nr_pages >= node_present_pages(node))
>> -		arg.status_change_nid = node;
>> +	check_nodemasks_changes_offline(nr_pages, zone, &arg);
>>  
>>  	ret = memory_notify(MEM_GOING_OFFLINE, &arg);
>>  	ret = notifier_to_errno(ret);
>> @@ -973,10 +1036,9 @@ repeat:
>>  	if (!populated_zone(zone))
>>  		zone_pcp_reset(zone);
>>  
>> -	if (!node_present_pages(node)) {
>> -		node_clear_state(node, N_HIGH_MEMORY);
>> +	clear_nodemasks(node, &arg);
>> +	if (arg.status_change_nid >= 0)
>>  		kswapd_stop(node);
>> -	}
>>  
>>  	vm_total_pages = nr_free_pagecache_pages();
>>  	writeback_set_ratelimit();
>>

Thank you very much for the review.
And sorry for so late respond.

Thank,
Lai

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

Powered by blists - more mailing lists