page zone 分配策略如何理解

zylthinking

1. 
   
2. struct zonelist {
   
3.     struct zone *zones[MAX_NUMNODES * MAX_NR_ZONES + 1]; // NULL delimited
   
4. };
   
5. 
   
6. typedef struct pglist_data {
   
7.     struct zone node_zones[MAX_NR_ZONES];
   
8.     struct zonelist node_zonelists[MAX_ZONELISTS];
   
9.     int nr_zones;
   
10. #ifdef CONFIG_FLAT_NODE_MEM_MAP    /* means !SPARSEMEM */
    
11.     struct page *node_mem_map;
    
12. #ifdef CONFIG_CGROUP_MEM_RES_CTLR
    
13.     struct page_cgroup *node_page_cgroup;
    
14. #endif
    
15. #endif
    
16. #ifndef CONFIG_NO_BOOTMEM
    
17.     struct bootmem_data *bdata;
    
18. #endif
    
19. #ifdef CONFIG_MEMORY_HOTPLUG
    
20.     /*
    
21.          * Must be held any time you expect node_start_pfn, node_present_pages
    
22.          * or node_spanned_pages stay constant.  Holding this will also
    
23.          * guarantee that any pfn_valid() stays that way.
    
24.          *
    
25.          * Nests above zone->lock and zone->size_seqlock.
    
26.          */
    
27.     spinlock_t node_size_lock;
    
28. #endif
    
29.     unsigned long node_start_pfn;
    
30.     unsigned long node_present_pages; /* total number of physical pages */
    
31.     unsigned long node_spanned_pages; /* total size of physical page
    
32.                                              range, including holes */
    
33.     int node_id;
    
34.     wait_queue_head_t kswapd_wait;
    
35.     struct task_struct *kswapd;
    
36.     int kswapd_max_order;
    
37.     enum zone_type classzone_idx;
    
38. } pg_data_t;
    
39. 
    
40. 
    

复制代码

如上的 pglist_data 结构， 一直以为 node_zones 和 node_zonelists 关系如下：
node_zones 是实际上的 zone 结构， 每个 pglist_data 存在最多三个 zone
node_zonelists 代表分配策略， node_zonelists 是 zonelist 数组， 每个zonelist 又是 zone 的一个链表；
因此形成如下的关系：
pglist_data_0 = {
    .node_zones = {.node_zones = {zone_0_dma, zone_0_normal, zone_0_high},}
    .node_zonelists = {.zones = {&zone_0_high, &zone_1_high, &zone_0_normal, &zone_1_normal, 0}}
}

pglist_data_0 = {
    .node_zones = {.node_zones = {zone_1_dma, zone_1_normal, zone_1_high},}
    .node_zonelists = {.zones = {&zone_1_high, &zone_0_high, &zone_1_normal, &zone_0_normal, 0}}
}

但看以下代码有点傻眼：

1. static unsigned int nr_free_zone_pages(int offset)
   
2. {
   
3.     pg_data_t *pgdat;
   
4.     unsigned int sum = 0;
   
5. 
   
6.     for_each_pgdat(pgdat) {
   
7.         struct zonelist *zonelist = pgdat->node_zonelists + offset;
   
8.         struct zone **zonep = zonelist->zones;
   
9.         struct zone *zone;
   
10. 
    
11.         for (zone = *zonep++; zone; zone = *zonep++) {
    
12.             unsigned long size = zone->present_pages;
    
13.             unsigned long high = zone->pages_high;
    
14.             if (size > high)
    
15.                 sum += size - high;
    
16.         }
    
17.     }
    
18. 
    
19.     return sum;
    
20. }
    

复制代码

如果按照我的理解， 很显然 sum 值是实际值的2倍， 因为计算 pglist_data_0 时其实将 zone_1_high， zone_1_normal 也计算进去了， pglist_data_1 时同理将 zone_0_high， zone_0_normal 计算进去了

不知实际上node_zones 和 node_zonelists 关系到底是什么样的

George_lz

node_zones 是pglist里面的所包含对zone， node_zonelists是分配时的优先级链。例如：对于 某个zonelist： zone1->zone2->zone3, 分配zone1内存时，如果在zone1上分配失败，再去zone2上分配，依次向后。
对node_zonelists source code 上的注释很清楚：
* One allocation request operates on a zonelist. A zonelist
* is a list of zones, the first one is the 'goal' of the
* allocation, the other zones are fallback zones, in decreasing
* priority.

zylthinking

George_lz 发表于 2012-08-14 17:11
node_zones 是pglist里面的所包含对zone， node_zonelists是分配时的优先级链。例如：对于 某个zonelist：  ...

非所问。

George_lz

回复 3# zylthinking

理解错了。
我用的35上的代码刚好解除了你的疑问：
static unsigned int nr_free_zone_pages(int offset)
{
        struct zoneref *z;
        struct zone *zone;

        /* Just pick one node, since fallback list is circular */
        unsigned int sum = 0;

        struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);

        for_each_zone_zonelist(zone, z, zonelist, offset) {
                unsigned long size = zone->present_pages;
                unsigned long high = high_wmark_pages(zone);
                if (size > high)
                        sum += size - high;
        }

        return sum;
}

   

zylthinking

George_lz 发表于 2012-08-14 18:22
回复 3# zylthinking

理解错了。

为什么解除了？ 我还是感觉照样

George_lz

回复 5# zylthinking


Don't feel confusing, it's just a bug.
The following is the information for this commit.


commit e310fd43256b3cf4d37f6447b8f7413ca744657a
Author: Martin J. Bligh <mbligh@mbligh.org>
Date:   Fri Jul 29 22:59:18 2005 -0700

    [PATCH] Fix NUMA node sizing in nr_free_zone_pages
   
    We are iterating over all nodes in nr_free_zone_pages().  Because the
    fallback zonelists contain all nodes in the system, and we walk all the
    zonelists, we're counting memory multiple times (once for each node).  This
    caused us to make a size estimate of 32GB for an 8GB AMD64 box, which makes
    all the dirty ratio calculations, etc incorrect.
   
    There's still a further bug to fix from e820 holes causing overestimation
    as well, but this fix is separate, and good as is, and fixes one class of
    problems.  Problem found by Badari, and tested by Ram Pai - thanks!
   
    Signed-off-by:  Martin J. Bligh <mbligh@mbligh.org>
    Signed-off-by:  Matt Dobson <colpatch@us.ibm.com>
    Signed-off-by: Andrew Morton <akpm@osdl.org>
    Signed-off-by: Linus Torvalds <torvalds@osdl.org>

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 42bccfb..8d08837 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1061,20 +1061,19 @@ unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)

static unsigned int nr_free_zone_pages(int offset)
{
-        pg_data_t *pgdat;
+        /* Just pick one node, since fallback list is circular */
+        pg_data_t *pgdat = NODE_DATA(numa_node_id());
        unsigned int sum = 0;

-        for_each_pgdat(pgdat) {
-                struct zonelist *zonelist = pgdat->node_zonelists + offset;
-                struct zone **zonep = zonelist->zones;
-                struct zone *zone;
+        struct zonelist *zonelist = pgdat->node_zonelists + offset;
+        struct zone **zonep = zonelist->zones;
+        struct zone *zone;

-                for (zone = *zonep++; zone; zone = *zonep++) {
-                        unsigned long size = zone->present_pages;
-                        unsigned long high = zone->pages_high;
-                        if (size > high)
-                                sum += size - high;
-                }
+        for (zone = *zonep++; zone; zone = *zonep++) {
+                unsigned long size = zone->present_pages;
+                unsigned long high = zone->pages_high;
+                if (size > high)
+                        sum += size - high;
        }

        return sum;

zylthinking

那就知道了， 3.x 的变化比如 fallback list 缩减成一个 这些我也知道， 因此 3.x 上我也没觉得有问题， 就是之前的代码让我很迷惑。
话说找到这些 commit 有什么窍门么， 经常看到一些人拿出 commit commence 来说事， 但找到这些 commence 莫不是有什么技巧？ 我的意思是说， 除了 git 那个二分查找搜索某个改变在哪个commit引入的命令

embeddedlwp

建议楼上大哥直接看最新代码，

embeddedlwp

git log --grep='nr_free_zone_pages'